Fault Diagnosis in Electric Generators: Methods, Trends and Challenges

Abstract

It has been more than a century since the day the first commercial generator was put into operation. Since then, our technical civilization has been dependent on the reliability of electric generators for electrical supply. The reliable and uninterruptible operation of power generators depends heavily on a proper maintenance strategy, and faults occurring during operation should be detected in a timely manner. In this work, a review of state-of-the-art fault diagnosis strategies is presented. Faults occurring in electric generators are presented and categorized, and the quantities utilized for their detection are provided. Traditional signal processing methods and machine learning (ML) approaches for their reliable detection are analyzed. Trends and challenges are discussed, and future directions are highlighted.

1. Introduction

Since the dawn of technology, maintenance has been recognized as a key factor in ensuring reliability when it comes to electric generators [1]. As the workhorses of the industry and of our technical civilization in general, generators need to achieve high reliability and ensure availability on a regular basis. The initially established concept of “time-based” scheduled maintenance often leads to high costs and sometimes unexpected interruptions to operation, so the need to search for new, more efficient tools has soon become evident [2,3]. The evolution of condition-monitoring techniques starting from the 1970s [4] has been a journey from manual to automated diagnosis, from isolated to connected systems and from descriptive to prescriptive maintenance, aiming to provide high reliability at a balanced cost.

Maintenance started to move to predictive concepts in the early 1990s [5], when the condition-based maintenance concept was introduced. It took about a decade to recognize the benefits of interconnection and integration of the monitored system, enabling online implementations, and from the 2010s onward, multiple industrial IoT sensors have been used for diagnosis [6]. Gradually, it became not enough to suspect that a fault had occurred; the fault should be diagnosed so that the exact cause and location of the fault could be identified.

Today, early fault diagnosis is especially important in the frame of the Industry 5.0 concept [7,8,9], as maintenance is gradually shifted from reactive fixes to proactive, predictive and intelligent interventions. Advanced signal processing techniques, artificial intelligence and digital twins are involved in this process, while the human operator retains the role of the expert who will make informed and strategic decisions for maintenance, instead of emergency repairs. This way, the possibility of catastrophic downtime is minimized, ensuring production resilience, stability of the power grid, energy efficiency, and sustainability.

The paper is organized as follows: In Section 2, a classification of the faults that may occur in electric generators is presented, analyzing their roots and causes. In Section 3, a review of the methods used for the diagnosis of the faults is conducted, followed by Section 4, where the condition-monitoring methods are classified in terms of the electrical or mechanical quantities used for each case. In Section 5 and Section 6, a review of signal processing and ML methods used for diagnosis of the faults occurring in electric generators is presented. In Section 7, important findings are highlighted, and future perspectives and conclusions are provided.

2. Classification of Faults in Electric Generators

2.1. Electric Generator Classification

An electric generator is a rotating machine which is used to convert mechanical power into electricity for transmission via the power grid. A generator can produce the electrical energy required for industrial and household applications and transportation systems. The mechanical power input is obtained from a rotating shaft and may derive from several sources such as turbines driven by steam generated from the heat of burning fossil fuels or nuclear fission; wind turbines; hydraulic turbines at waterfalls or dams; gas turbines; or gasoline and diesel engines. Mechanical power input usually determines the generator’s construction and speed.

Generators vary in shape and size and can be categorized according to their intended use, nominal power, and the type of generated current (AC or DC). They all operate according to the same fundamental principle: a rotating armature coil inside a permanent magnet field or a stationary coil inside a rotating magnetic field. AC generators, usually three-phase, are the most common in electric installations and grid applications. Based on their size, they include portable and domestic units, and mid-range and large industrial models. DC generators are still used in legacy applications, where their unique characteristics are needed (e.g., old historic electric trains), but they tend to be replaced by AC ones combined with rectifiers.

AC generators are categorized into synchronous and asynchronous. Synchronous generators (SGs) are the main component of large AC power systems, permitting continuous electrical energy provision [10]. The field winding current produces a magneto-motive force (MMF) which, in grid-connected generators, is regulated to control the produced reactive power. An additional generator, mechanically coupled with the main one, usually provides this excitation via two slip rings. In the modern power industry, synchronous generators are widely used in hydropower, thermal power, nuclear power, and diesel power generation. In certain cases, the magnetic field in a synchronous generator comes from permanent magnets, such as neodymium–boron–iron or samarium–cobalt. These generators do not require the provision of field current, and they cannot control the output voltage by field adjustment. Contrary to synchronous generators, in induction generators, the speed is slightly faster than the synchronous speed. This type of generator takes advantage of the rotating magnetic field in the air gap between the stator and rotor to interact with an induced current in the rotor windings. The output power depends on the slip rate, which means that the power delivered is dependent on how much the rotor speed differs from the rotating field speed. It cannot act as a standalone power source since it lacks an independent magnetic field [11]. Standalone asynchronous generators, nevertheless, require the use of loading capacitors to operate. Induction generators are frequently preferred over synchronous generators for small hydroelectric dams since they are not subject to loss of synchronism following transient changes in the power system. Lastly, a subcategory of induction generators is the Doubly Fed Induction Generators (DFIGs) that operate by connecting the stator directly to the grid and controlling the rotor circuit through power converters. These machines permit variable-speed operation while maintaining constant-frequency output to the grid and are used in medium- to large-scale wind energy systems.

2.2. Basic Types of Faults in Electric Generators

Generally, faults occurring in electrical machines can be distinguished into electrical, mechanical, magnet-related, and control faults. As illustrated in Figure 1, electrical faults found on the stator winding are the most prevalent in large-scale synchronous generators, while rotor winding faults are the second most common [12]. Stator winding interturn short-circuit faults caused by the aging of winding insulation are the most prevalent, accounting for more than a quarter of the faults occurring in electric generators [13]. In the wind turbine sector, the highest percentage of faults occur in the electrical system, followed by the control system and sensors [14]. A literature survey shows that bearing faults and stator insulation breakdown cause most of the machine failures in induction wind turbine generators [15,16,17]. In hydroelectric plants where salient-pole synchronous generators are mainly used, many faults may appear, with the major ones being eccentricity and electrical faults [18,19]. Electrical faults include stator core insulation failure, stator winding failure, and rotor winding failure. They occur due to insulation failure, short or open circuits, or other issues related to electrical machinery conditions such as overvoltage, undervoltage, phase reversal, improper connection of stator windings, interturn short circuits, or earth faults and stator defects that can cause a phase winding to open or short.

After electrical faults, the second most common category is mechanical faults [20]. In most cases, they occur on the rotor side of the electrical machine. Mechanical faults mainly include bearing failure, rotor or stator mechanical integrity failure, broken damper bars, shaft misalignment, and unbalanced load. Other types of mechanical faults include eccentricity faults and damaged rotor end rings or rotor bars. Another fault category, control faults, are due to improper machine operation or failure due to issues concerning the software or control system. These faults can be divided into power supply, power electronics or sensor failure and software malfunction. As far as magnet-related faults are concerned, permanent magnets can suffer irreparable demagnetization damage that causes a reduced or unbalanced rotor flux if combined with overloading, rotor defects or increased vibrations. In Figure 2, various types of faults occurring in different categories of rotating electrical machines, e.g., squirrel cage induction generators (SCIGs), wound rotor induction generators (WRIGs), permanent magnet synchronous generators (PMSGs), and wound field synchronous generators (WFSGs), are presented.

2.3. Main Factors Leading to Electric Generator Faults

Faults in electric generators are associated with electrical, mechanical, magnetic, and cooling components [10]. Electrical insulation destruction may progressively evolve into electric generator breakdown. Induced faults are the aftermath of stresses imposed on the generator stator. Unwanted alterations in physical and chemical properties of the insulation, linked with tensions during machinery operation, are called aging phenomena in the insulation system. Different stresses that act as inhibitors for stator winding health can broadly be classified into electrical, mechanical, thermal, and ambient stresses. Each of these stress types can lead to both short-term and long-term issues, depending on factors such as the effectiveness of the cooling system and the specific operating conditions.

Several factors contribute to thermal stress in generators, including extreme fluctuations in ambient temperature, malfunctions in the cooling system that hinder effective heat dissipation, voltage imbalances that raise copper losses, and overvoltage conditions that elevate the generator’s induced voltage, resulting in increased core losses. Generally, the AC winding insulation system of the generator is designed to withstand electrical, mechanical and thermal tensions over the generator’s predefined durability. As far as thermal faults are concerned, the machine’s internal temperature is raised during its on-load operation [21], which progressively leads to the insulation’s deterioration. As a result of the aging process, at some point the insulation system loses its mechanical integrity and may fracture under partial mechanical or electrical stress. Even without a shock, ongoing degradation of the insulation can eventually cause an interturn fault, which soon may evolve into a complete winding failure.

In terms of electrical stress, transient or DC voltages can cause insulation aging. Nevertheless, even under normal operating conditions, AC voltages place significant stress on the generator, while running under abnormal or faulty conditions accelerates electrical aging. Unlike thermal stresses, which arise from internal factors, electrical ones originate externally and are imposed on the generator by the grid. Several factors contribute to this type of stress, including the switching of circuit breakers, failures in power system insulation, the use of power factor correction capacitors, current-limiting fuses, frequent arcing or sparking, and short circuits occurring within the power grid [10].

Moreover, mechanical aging contributes directly and indirectly to insulation aging [22]. In generators, the magnetic field imposes electromagnetic forces on the current-carrying windings, causing them to displace and produce vibrations. These mechanical and electromagnetic forces, along with short-circuit shocks, rotor centrifugal forces, vibrations, resonance, wear, abrasion, bending, and an inadequate number of wedges, can lead to the relative movement of insulation components. Such movements place stress on both the generator and its insulation, accelerating the deterioration and eventual failure of the insulation system.

Lastly, ambient stresses can lead to mechanical and electrical defects. Generators typically have specific limits for acceptable temperature, humidity, and pollution tolerance levels. Any variation in these factors can impair the generator’s performance and shorten its expected durability. These stresses can be classified into several categories: environmental pollution, ambient humidity, abrasive particles, temperature fluctuations and exposure to cold weather, acidic vapors and corrosives in the environment, and radiation in nuclear plants [23]. Also, moisture-induced hydrolysis can break chemical bonds in many layered insulation materials, causing them to swell and making them more vulnerable to failure when exposed to other types of stress. Insulation materials and their chemical bonds can be degraded by exposure to oils, acids, alkalis, and solvents. Additionally, dust pollution and the accumulation of particles on the insulation surface pose further risks to its integrity [24].

3. Fault Analysis and Diagnostic Methods

In this section, analysis of faults is conducted according to their nature and location in an electric generator, for indicating causes and effects on the machine’s operating behavior. Based on this, typical diagnostic approaches are discussed for assessing the fault. In Figure 3, the typical framework for fault detection, identification, and severity estimation is presented. The process starts with sensor selection and installation for acquiring quantities that are related to potential faulty conditions, such as vibration, current, temperature or magnetic field. Next, signal processing along with feature extraction takes place. This stage includes filtering, conditioning, and extraction of useful fault-related indices. Based on the extracted features, deviations from the normal operating state are identified to detect the fault. The final step includes identification of the specific type of fault with a severity estimation.

3.1. Stator Winding Faults

As shown before in Figure 1, in large electric generators, a high percentage of faults are related to stator winding damage. Generally, this category of faults includes open-circuit faults, earth/ground faults, phase faults, coil faults, and turn-to-turn faults. An open-circuit fault in one of the windings turns the generator into a two-phase machine and this condition can be detected via protective relays. A phase-to-phase short-circuit (SC) fault (phase fault) arises when one or multiple points of a stator phase short circuit with other phases, causing a distinct imbalance and asymmetry in the generator that can be easily recognized. This type of fault can be very destructive for the winding; however, it happens infrequently compared to the earth fault (phase-to-earth SC fault). Earth faults can be very destructive for the generator, since the absence of resistance between high voltage and earth causes a significant current to flow through the fault path. SCs of coil-to-coil (coil faults) happen when sections of a coil come into contact with nearby coils belonging to the same phase. Compared to the aforementioned faults, they cause less imbalance and asymmetry in the generator.

Turn-to-turn faults (TTFs) are one of the most destructive faults and are more probable in small electrical synchronous generators than in large ones [12]. This fault can seriously damage the insulation and deform copper conductors. Interturn insulation damage is usually caused by small punctures, especially in the stator core end parts and in the initial coils of the end line. Since the coils are insulated both from each other and the stator using special insulating paper, TTFs are more likely to arise compared to other types of stator faults. In contrast to other generator faults that result in higher current levels, turn-to-turn faults are not easily identified early by conventional protection methods [10]. If TTFs are not detected in their initial stages, the high circulating current in the short-circuited turns and the intense heat it produces can severely damage the entire winding and may even harm the stator core [25]. Diagnosis of this type of fault, even though it is a difficult procedure, is highly important and has gained research interest. However, detecting an interturn fault in its early stages is challenging because it has an insignificant impact on the terminal currents. A widely held view is that turn-to-turn faults often go unnoticed at first, eventually evolving into severe phase-to-ground or phase-to-phase faults [12]. Stator winding fault diagnosis can be accomplished using both current/voltage methods and flux monitoring ones. The current/voltage-based ones are sensorless, less expensive and invasive to implement compared to flux monitoring methods, and therefore more popular [12].

As far as wind turbines (WT) are concerned, if the induction generator is not shut down for inspection and repaired in time, after the stator TTF occurs, it may cause a phase-to-phase SC or a ground SC fault [26]. In WT generators greater than 2 MW, stator failures have shown a significant increase, as a result of missing magnetic wedges which were employed to optimize the generator’s design and power output [27].

Open-circuit faults in induction generators allow the machine to operate with reduced torque while SCs of a few turns lead to a catastrophic failure in a short time. Induction machines typically use the stator’s negative sequence current to identify imbalances in supply voltage or internal asymmetries. When a significant fault like an open phase occurs, the resulting negative sequence current closely matches the positive one, allowing for easy detection and quick protection system response [28]. When an SC occurs inside an induction machine, typical protection systems may fail or the machine might continue operating while the shorted turns overheat, eventually causing severe insulation damage. If this fault goes unnoticed, it can worsen quickly—similar to what happens in synchronous generators—resulting in phase-to-ground or phase-to-phase faults. The resulting ground current can cause permanent damage to the core, requiring the machine to be taken out of service. All the available analytical models (winding function approach, analysis of MMFs, finite element approach) are helpful for predicting machine behavior and assisting in short-circuit detection, but identifying such faults through online measurements—without shutting down the machine—remains a challenge. Combining fault detection methods with periodic scheduled inspections of stator turn-to-turn insulation offers the most reliable means of identifying SC faults. The surge and offline partial discharge (PD) tests are the most widely used methods for evaluating stator turn-to-turn insulation [28]. In [17], the authors analyzed rotor-phase current and search-coil voltage to detect stator interturn faults in a DFIG. These signals can reveal interturn faults in the stator, even if only a small number of turns are shorted.

3.2. Rotor Winding Faults

In large, low-speed generators, the salient-pole synchronous machine rotor includes the rotor field winding and the damper winding. An SC fault in the rotor field winding may occur, and this can be divided into interturn SC fault and turn-to-ground fault. Comparing stator winding interturn SC fault and rotor field winding interturn SC, fault severity is low in the second case, because of the low magnitude of flowing DC current and the thermal tolerance of the insulation system [29,30]. As an interturn short-circuit fault worsens, it yields increases in temperature and vibration in the field winding. These effects lead to further insulation deterioration in nearby turns, making the fault more severe. This type of fault may develop into a rotor body SC. If the interturn SC occurs near the rotor body, it may damage the insulation between the field winding and the body, leading to a rotor ground fault. An SC in the field winding often results in magnetic field asymmetry across the air gap, which can generate shaft currents that pass through the bearings [10].

Rotor fault detection in an SG differs from that of the stator because of the DC flowing current in the rotor, making it difficult to apply detection techniques that rely on sinusoidal waveforms. The results of current and voltage spectrum analyses indicate that the rotor TTF amplifies the positive, negative and zero sequence components at the first right-sideband rotational frequency. Magnetic flux and vibration monitoring can additionally be implemented for rotor fault detection. In a faulty condition, certain frequency components present increased amplitudes in the flux and vibration signal spectra, compared with normal operating conditions [12]. To diagnose rotor winding faults in turbo SGs and hydro SGs, a flux probe can be installed in the stator slot, which is sensitive to air-gap magnetic field changes. The passage of each rotor slot past the flux probe leads to observable differences in the search coil’s induced voltage due to the magnetic poles. An interturn SC rotor fault in a coil reduces the peaks associated with the two opposite slots containing the faulty coil; thus, the presence of shorted turns can be detected.

In WFSGs, interturn short circuits in the rotor field winding are a frequent fault. Even though the machine may keep running with a small number of short-circuited turns, the heat generated by the fault can degrade adjacent insulation and eventually cause a sudden rotor winding failure. ITSC and eccentricity faults in WFSGs lead to changes in stator and rotor inductances and thereby impact the current and voltage signals. Harmonic fluctuations in stator terminal voltage due to ITSC and eccentricity faults are higher compared to those in the stator current. However, nonlinear loads interfere with fault-related harmonics in stator terminal voltage. This situation may reduce the accuracy of fault diagnosis [31]. Research indicates that monitoring techniques of the air gap and stray magnetic field (SMF)—the mirror of the air-gap magnetic field—are effective for identifying sensitive ITSC and eccentricity faults in WFSGs [32,33,34]. In [31], the authors propose the application of FFT (Fast Fourier Transform) to SMF measured on the stator backside of the WFSG (by comparing a faulty frequency spectrum with the normal operating state spectra), aiming to overcome the weaknesses of existing techniques for detecting faults at an early stage, while the WFSG works under open-circuit or grid-tied conditions.

Moreover, a new method to detect a turn-to-turn fault in the field windings of synchronous machines is proposed in [19] that does not require any additional sensor, by comparing the actual excitation current with the theoretical excitation current calculated under normal operating state conditions. If the excitation current is much higher than the theoretical value, it indicates that an interturn fault has occurred. In [35], the use of shaft voltages in synchronous turbine generators to diagnose field coil ITSC faults is proposed.

Concerning rotor ground faults, a single rotor ground fault typically does not directly influence the machine, as the excitation circuit is usually ungrounded. However, once the first fault occurs, the risk of a second ground fault increases. This is because the initial fault sets a voltage reference in the field winding, which raises the ground potential stress at other points. If a second ground fault takes place, it results in a reduction in healthy turns available in the field winding. In wind turbine generators (WTGs) [26], the main reason for rotor body failure is due to eddy current loss in the rotor caused by negative sequence voltage in the power supply. Regardless of the type, any electrical fault in the rotor of an induction generator causes rotor circuit asymmetry—either from impedance imbalance in wound rotor machines or structural defects like broken bars in squirrel-cage types. Signal-based methods typically rely on stator current measurements which are sensitive to rotor faults. This makes stator current a suitable source for extracting a diagnostic index and setting a threshold that distinguishes between a normal operating state and faulty conditions [36].

3.3. Bearing Faults

Bearing faults are the aftermath of inappropriate lubrication, stress, and installation on a weak foundation, which cause wear and fatigue of the bearings. In general, a bearing fault generates a rotor eccentricity fault, and if this fault is intensified, a collision between the rotor and the stator may occur. Most electrical machines commonly utilize ball- or rolling-element bearings, which are composed of inner and outer rings. Within these rings, balls or rollers move along raceways. Faults in bearings can arise from defects in the outer race, inner race, balls or the cage (train). Even under normal balanced conditions and with proper shaft alignment, fatigue-related faults may still develop. Factors such as vibrations, internal stresses, inherent eccentricities and bearing currents caused by power electronic systems significantly contribute to the progression of these faults [28].

In wind turbines, bearings support the generator and should have low vibrational properties [37,38]. Electrical pitting is a primary concern when designing these bearings. To address this issue, electrically insulated bearings are installed. There are two main types of electrically insulated bearings: coated bearings with insulating layers made of ceramic or conductive microfibers on the ring surface, and hybrid bearings that use ceramic balls with metal components. Additionally, certain strategies—such as applying conductive lubricants and grounding the bearings—can help minimize electrical damage to generator bearings [22]. In a WTG system, the generator bearing serves as the connection between the rotor and the stator. Excessive wear from prolonged operation or misalignment during installation can cause the bearing to produce an unbalanced force during operation, leading to abnormal friction. This condition is known as rotor sweep failure [26]. It typically results in a rise in the motor’s surface temperature and the occurrence of irregular noise. In WTGs, bearing failures account for 40% of total common faults. Common failures of bearings occur in four locations, namely the inner ring, outer ring, rolling elements and cage.

Bearing fault types can be divided into the following six subcategories [26]: fatigue spalling, fatigue wear, plastic deformation, gluing, cage damage, and corrosion. WTGs, often placed in high, windy locations, experience sudden stress changes in their bearings due to shifting wind speed and direction. This can cause tiny cracks, increased clearances between rolling elements and raceways, higher vibrations, and eventually surface spalling on the bearings. Fatigue wear in bearings may be caused by poor lubrication or dust intrusion and leads to uneven contact between rolling elements and raceways. This results in deformation and wear of the bearing surfaces. The severity of bearing failure increases with operating time, and if internal damage progresses too far, it can disrupt wind turbine operation. Aside from wear, rolling element deformation in WTG bearings can result from overheating, prolonged heavy loads, or inadequate lubrication. Inadequate lubrication in a wind turbine bearing leads to poor heat dissipation, causing surface bonding between internal components. Under heavy loads, this bonding worsens, accelerating bearing failure. Uneven force distribution can also damage the bearing, especially if the cage is weaker than the inner and outer rings. Lastly, WTG bearings are exposed to air, making them susceptible to corrosion from oxidation reactions with acidic substances. Bearing currents may also cause sparks between raceways, further contributing to surface corrosion.

3.4. Eccentricity Faults

In each generator, three axes can be distinguished: the rotor symmetrical axis, stator symmetrical axis, and rotor rotational axis, as shown in Figure 4. In an ideal generator operating normally, the aforementioned stator and rotor axes coincide with each other, which means that the air-gap distribution function is constant and ideally independent from rotor rotation. In a real generator, perfect alignment of these three axes is nearly impossible. Even in newly manufactured units, slight misalignment in manufacturing or assembly is common, known as inherent eccentricity. Several factors can worsen inherent eccentricity, including improper alignment between the stator and rotor, incorrect bearing placement, a bent rotor shaft, bearing wear, loose bolts, an elliptical shape of the stator’s inner surface, and a misaligned generator shaft [10]. In the literature, rotor eccentricity can be idealized as static eccentricity (SE), dynamic eccentricity (DE), or mixed eccentricity (ME). SE indicates that stator and rotor centers do not coincide, and the rotor revolves around the stator center. During an SE fault, as shown in Figure 4, the center of the rotor is shifted from the stator center, and it rotates around its center axis. In this case, the minimum air-gap length varies only by position, disregarding the inherent variations due to rotor pole saliency. The possibility of having a single SE fault is low because of high precision in the manufacturing process. A DE fault indicates that the rotor and the stator centers do not coincide, causing the rotor axis to revolve around both the stator axis and its own axis. As illustrated in Figure 4, in the case of a DE fault, the minimum air-gap length varies with both position and time. The ME contains both SE and DE situations, in which case the rotor revolves a third center other than the stator and the rotor. Hence, in the case of an ME fault, the rotor is displaced towards one side of the stator from the SE fault, while the DE fault produces a whirling motion that varies in time and position.

Eccentricity leads to bearing wear and escalates misalignment, which can cause severe rotor-stator damage, reduced efficiency, distorted electrical output, and increased vibration. The resulting higher vibrations can also cause noise and trigger short circuits due to insulation breakdown or localized overheating from magnetic saturation. Diagnostics for eccentricity faults include the assessment of the electrical quantities (current/voltage/power) and the internal inductance distortion. The finite element method (FEM), the analytical method, the magnetic equivalent circuit (MEC) method, and the modified winding function approach (MWFA) have been reported, and it was concluded that the FEM is an effective verification tool for both modeling and diagnostic methods [39].

In the WFSGs of hydropower plants located inside mountains, two kinds of mechanical faults are prevalent: static and dynamic eccentricity faults [31]. The increased vibration levels caused by these faults may be used to activate protection systems and shut down the wind turbine generator (WFSG) to prevent further damage. In some instances, operators may reduce the generator’s output power to lower vibrations, which leads to decreased overall power generation. Eccentricity fault detection using the stator current is only possible during on-load operation of the WFSG and the introduced harmonics for DE faults are the same as the harmonics in the power grid, namely, 7th, 11th, 13th, 17th, and 19th. Hence, false detections are likely to occur during grid-connected operation of the WFSG. Stator terminal voltage may also be used for DE fault detection. However, nonlinear loads may have negative impacts on the fault-related harmonics in the system, which may influence the accuracy and effectiveness of fault diagnosis. Lastly, the air-gap magnetic field can be measured using a search coil wound around the stator tooth or a Hall-effect sensor mounted on the stator tooth. The SMF can also offer valuable information for the early detection of DE faults. To overcome the aforementioned difficulties, in [31] a method for mixed ITSC and DE faults detection was proposed, during both open-circuit and grid-tied operations of a 22 MVA WFSG by applying FFT to SMF measured on the stator backside, indicating that the grid harmonics cannot distort the introduced pattern.

3.5. Misalignment Faults

Rotor misalignment faults are the most severe form of eccentricity in an electric generator, as they cause the air-gap length to vary along the generator’s longitudinal axis [10]. Misalignment faults are among the most critical mechanical issues, as they can lead to extensive damage across the entire machine—affecting bearings, stator or rotor cores, and their windings. Early detection is essential to avoid significant financial losses. These faults can arise from improper connection of the prime mover in turbine generators, unbalanced mechanical loads in motors, bent shafts, or manufacturing defects. Misalignment can be categorized, in analogy to eccentricity, into static, dynamic, and mixed. A second classification for misalignment faults, and the most critical, is parallel, angular, or combined misalignment, as explained in Figure 5. Parallel misalignment occurs when the shafts of connected motors are offset but still remain parallel, causing uneven loading on bearings and couplings, and angular misalignment arises when the shafts are at an angle with respect to each other, leading to cyclic forces which increase wear and vibration.

In large-scale wind turbines, misalignment occurs when the raceway edges are not aligned with the path of the ball contact, leading to issues such as overheating, increased vibration, and structural deformation. A common cause of this misalignment is the elastic bending of the crankshaft due to inertial forces acting on it [22]. Bearing misalignment is a frequent issue in transmission systems since the rubber bushings can shift under high torque and harsh environmental conditions [40]. In WRIGs, eccentricity faults are caused by an uneven air-gap distribution, which can result in unbalanced magnetic pull and potentially shaft bending. This unbalanced pull, along with misalignment, often leads to bearing damage, ultimately increasing vibration and noise levels [11].

3.6. Broken Damper Bar Faults

A salient-pole synchronous rotor includes, apart from the rotor field winding circuit, the damper bar circuit that is composed of copper bars in each rotor pole, of which the ends are short circuited by copper end rings. In practice, these bars help counteract the asynchronous air-gap flux that arises due to mechanical and electrical transients. They play a crucial role in limiting transient power and torque oscillations and assist in maintaining the synchronous generator at synchronous speed. However, faults in this type of bar are due to the deficient construction of damper cage or rigorous machinery usage. These faults arise in some SGs and can cause significant damage. Damper bars’ fault mechanism is very similar to that of induction motor cage bars, which means that the proposed methods for damper bar fault detection are similar to those used for induction motors [41,42]. Some diagnostic methods for damper bar breakage are the use of a flux probe to measure the air-gap flux in runup condition, temperature sensors (Resistance Temperature Detectors—RTD) mounted on the rotor to monitor the temperature increase in the abnormal operating state, and a combination of these measurements with advanced signal processing methods like the Hilbert–Huang Transform (HHT) [12].

3.7. Demagnetization Faults

Permanent magnet (PM) machines have recently attracted significant attention since they offer higher efficiency with reduced weight and size, greater power output, and lower maintenance requirements. They play a crucial role in green technologies like electric vehicles, sustainable transport, and renewable energy generation. When used as direct-drive generators in systems like wind and marine energy, mostly low-power generators, they eliminate the need for gearboxes, reducing maintenance and improving reliability [43,44]. Demagnetization faults arise when the working point on the magnet’s characteristic curve falls beneath the knee point. Demagnetization can occur due to factors like strong reaction magnetic fields, rising operating temperatures, or permanent magnet degradation. When it happens, the machine needs to draw more current to maintain the same power output [45]. This results in reduced efficiency and, more critically, speeds up the aging of insulating materials while worsening the demagnetization of the permanent magnets. If left undetected, this chain reaction can lead to total machine failure. Additionally, uneven demagnetization creates magnetic field imbalances, causing vibrations and noise that can damage the machine’s bearings. In contrast, uniform demagnetization does not cause such issues, as the magnetic field remains symmetrical.

Various methods have been suggested in the literature for detection of this type of fault during the machine’s operation [46,47,48] based on current, flux, torque, and voltage monitoring. Nevertheless, most studies focus on partial demagnetization in machines with a low number of poles, where the cancelation of signatures in the current spectrum is not an issue. This cancelation effect, which can hide fault indicators, depends on the relationship between the number of stator coils and magnetic poles. To overcome the cancelation effect, a single search coil—placed either in the air gap or externally—can be used. Additionally, for reliable detection of demagnetization, the diagnostic method must be able to distinguish it from other faults that generate similar harmonic patterns in the machine’s electromagnetic and mechanical signals [49]. In [50], after analyzing the effects of demagnetization in radial flux machines, results showed that MCSA is unreliable for detecting non-adjacent faults. It was also found that circulating currents negatively affect efficiency regardless of load, with the impact being more severe under low-load conditions. Additionally, torque monitoring is more effective than MCSA for fault detection, while air-gap flux monitoring emerges as the most reliable method overall.

In [51], the authors compare three different techniques for detecting demagnetization in a coreless synchronous generator. The three techniques evaluated in the paper were current signature analysis (CSA), Park’s vector approach (PVA), and extended Park’s vector approach (EPVA). The comparison results revealed that EPVA outcompeted, thanks to its clear fault identification and high sensitivity to even minor fault conditions. In [52], a new condition-monitoring method for air-cored axial-flux PM generators used in marine energy was introduced. This method uses the peak-to-peak value of the speed-normalized voltage derived from an auxiliary winding that serves as a fault indicator, which directly reflects fault severity. The signal’s amplitude-frequency components were analyzed, under both steady and variable speeds to detect faults. Unlike a single-pitch search coil, this approach requires no additional signal processing since the signal is zero when the machine is under normal operating state conditions. It also detects very low fault severities due to minimal signal ripples. Partial demagnetization induces oscillations in the search coil, with the peak-to-peak value directly related to the severity of the fault. However, since uniform demagnetization results in zero total magnetic flux through the coil, this method can only detect partial faults. In contrast, a single-pitch coil can detect uniform demagnetization because it measures voltage across each pole pitch.

4. Condition-Monitoring Methods

4.1. Vibration Monitoring

Monitoring mechanical vibrations helps with the online identification of mechanical faults. By monitoring the shaft and bearing vibrations, various failure mechanisms can be identified such as misalignment, unbalanced magnetic pull, looseness, bearing wear, rubbing, resonance, and eccentricity of rotating parts prior to their failure. Multiple sensors that are part of the monitoring process generate signals, and the combined signal processing with artificial intelligence methods can aid in the early detection of faults like bearing failure, rotor imbalance, and even electrical issues providing useful data for the predictive maintenance of turbines and generators. Vibration monitoring in electric generators plays a vital role in ensuring reliable and efficient operation. It enables early detection of mechanical faults such as imbalance, misalignment, and bearing wear, allowing operators to address issues before they lead to costly unplanned outages. By providing real-time insights into rotor and stator health, vibration monitoring supports predictive and condition-based maintenance strategies, reducing unnecessary service interventions and extending equipment durability. It also helps identify hidden electrical problems like rotor bar breakages and eccentricity. Overall, vibration analysis optimizes generator performance, minimizes downtime, and contributes to a more sustainable maintenance approach.

Vibration monitoring data can be acquired using accelerometers, velocity sensors, or displacement sensors, with the accelerometers being the most frequently used type since they present a wide frequency range. Vibration analysis is preferred for wind turbine bearing condition monitoring, given that its measurements demonstrate high precision due to their wide frequency range (often 1–30 kHz) and easy usage [22]. Most faults occurring in SGs have a significant impact on the air-gap flux density profile, which leads to air-gap field distortion and a change in vibration behavior. Thus, the vibration profile of an SG can aid in early fault detection. An accelerometer can also effectively measure vibrations caused by air-gap magnetic field asymmetry—a fault-related issue in electric machines—when mounted on accessible locations such as the machine frame or stator backside. These mounting points allow for sensor installation and monitoring during operation, eliminating the need for machine shutdown. A literature review on fault detection based on vibration signals shows that the majority of work on fault detection in electric machines based on vibration signals has been conducted on induction machines, whereas SGs have yet to be thoroughly studied [10].

4.2. Current/Voltage Monitoring

Current and voltage monitoring in electric generators is essential for ensuring safe and efficient operation. It helps maintain electrical parameters within safe limits, preventing damage from overcurrent or overvoltage. By detecting abnormalities such as phase imbalances, grounding faults, or short circuits, it enables early fault detection and protection. Monitoring current helps prevent overheating of windings due to spikes, while voltage monitoring ensures synchronization with grid systems. These measurements also aid in predictive maintenance by revealing gradual performance degradation. Additionally, they protect connected equipment from harmful voltage transients, support fault diagnostics, and enhance power quality by detecting harmonic distortion. Modern systems even allow remote monitoring and control based on these critical electrical parameters. For non-intrusive fault detection, current and voltage are typically employed, since they are easily accessible through current transformers, potential transformers, or measurement probes.

One very useful application of current/voltage monitoring is differential TTF protection in large synchronous generators with multi-branch winding [12], where the currents in various stator winding branches are compared. Under normal conditions, the current difference is nearly zero. However, in the event of a turn-to-turn fault in an SG, this difference increases, causing a differential current to flow between the windings. Another technique involves measuring residual voltage using an additional open-delta potential transformer, applicable to all SGs. In normal operating state conditions, the phase voltages are balanced, and the fundamental voltage harmonics cancel out, resulting in zero residual voltage. During a TTF, the voltage in the faulty winding drops, leading to a non-zero residual voltage. However, other harmonics—like the third—also appear and must be filtered out to avoid false tripping. While effective, this method requires an extra PT (potential transformer), which can be costly or difficult to install in some setups. To address the issue of requiring an additional potential transformer, a novel protection scheme has been proposed in [12]. Instead of directly measuring the zero-sequence voltage (ZSV) with an extra PT, this new method calculates the ZSV, eliminating the need for additional hardware and improving practicality and cost-effectiveness.

4.3. Magnetic Flux Monitoring

Magnetic flux monitoring refers to the monitoring of electromagnetic fields related to the operation of an electric generator. The reason behind monitoring magnetic flux in electric generators is that the condition of the motor is closely related to the distribution of air-gap magnetic flux [53]. Distortion of air-gap magnetic flux is related to abnormal or faulty operation of the motor and so it provides a very useful quantity for condition monitoring. Magnetic flux can be measured in numerous ways, both within the internal components and around the stator. Monitoring the flux outside the motor is often referred to as leakage or stray-flux monitoring. The different ways of measuring include search coils, hall-effect sensors, fluxgate sensors, and magnetoresistive sensors [53]. In the large synchronous rotors of turbine generators, flux monitoring is used to monitor the operating conditions of the windings in the rotor. Paper [54] is one of the first works to apply air-gap flux monitoring in synchronous generators. As described in this work, short circuits on round rotor windings can be detected with the use of a small search coil near the field winding of the rotor. The method was performed in offline mode, with stator phases shorted. Other approaches which have been proposed for online testing require different loading conditions of the generator [55]. Similar methods have been proposed for rotor-condition monitoring considering different failures, such as stator winding faults, bearing faults, and eccentricity faults. Aside from synchronous generators, air-gap and stray-flux monitoring have been applied in the detection of winding and broken rotor bar faults in asynchronous machines [55]. The main challenges of magnetic flux monitoring include primarily installation issues, the selection of the appropriate sensor, and restrictions due to air-gap size, especially in lower-power generators.

4.4. Temperature Monitoring

Temperature monitoring involves the measurement of temperature in crucial parts of an electric generator, such as the stator windings, rotor parts, bearings and cooling system. Multiple temperature monitoring devices are available, with the most commonly used being Resistance Temperature Detectors (RTD), thermocouples, thermistors, fiber optic temperature sensors, and infrared sensors. Windings are the most crucial part of an electric generator for temperature monitoring, as temperature directly reflects the condition of the winding insulation. Although in traditional methods, voltage and current are used to estimate the temperature of the windings, there are specific hot spots on a generator which this method fails to identify, so temperature sensing appears advantageous. Placing the temperature sensor is also a challenge since there are rotating or hard-to-reach parts (e.g., filed windings) [55]. Other than windings, temperature monitoring can reveal the health status of other rotor parts, such as the machine rotor bars. Today, with advancements in wireless monitoring, installation, and measurements, transmission issues can be solved. Temperature monitoring can be expanded in the near future in hard-to-reach components across various types of electric generators, but applications are still limited [56].

5. Signal Processing Methods

5.1. Frequency-Domain Analysis

Due to the advanced computational capabilities of today’s microcomputing devices, more and more sophisticated signal processing methods are being developed. The most traditional and widely used signal processing method for fault diagnosis of electric generators is the Fast Fourier Transform (FFT). FFT converts time-domain signals to the frequency domain, where fault signatures can be identified. FFT and its derivatives are the most well-established tools for frequency-domain analysis for vibration, current, and magnetic flux signal analysis. For a signal with N samples, FFT produces N frequency bins, while the spacing between them (frequency resolution) is equal to the sampling rate of the data acquisition system divided by the number of samples N. Alternatives to FFT include the periodogram and Welch’s method. A periodogram estimates Power Spectral Density (PSD), which is a statistical measure of how much power is distributed over frequencies and is calculated through the computation of FFT. Welch’s method provides an improved estimate of the PSD, with lower noise and variance. In [57], a comparison of FFT, periodogram, and Welch’s periodogram is addressed. It is shown that Welch’s method is more robust and can provide better frequency discrimination results. In fault diagnosis tasks of electric generators, certain frequency patterns or frequencies are of interest for fault detection and identification. In this direction, in [58], a Goertzel filter was employed to filter the signal and compute only the fault-related harmonics, reducing the computational cost of FFT. Moreover, Zoom FFT (ZFFT) was employed in [59]. The authors used statistical methods to calculate the supply frequency and the slip of an induction motor. Then, ZFFT was employed, in which higher frequency resolution and reduced computational cost can be achieved when compared with the traditional FFT. Another high-resolution frequency-domain analysis method is Multiple Signal Classification (MUSIC), which has been used in [60] for broken rotor bar fault detection in an induction machine. It is shown that MUSIC is a useful tool for signal spectral estimation when dealing with signals in noisy environments. Moreover, when combined with ZFFT, MUSIC can provide increased frequency resolution for the frequency range of interest, bypassing the limitations of ZFFT or FFT alone. In a similar approach, in [61], MUSIC, along with Maximum Likelihood Estimation (MLE), was combined for the detection of bearing faults in induction machines. The proposed method extends MUSIC to multidimensional spaces for supply frequency and fault-related frequency estimations. MLE is employed in combination with MUSIC for amplitude estimation of the specific frequencies of interest. This method offers the advantage of estimating the frequencies of interest without estimating operating speed or load, and it provides a measure of severity. An additional method that addresses frequency resolution and spectral leakage issues of FFT is the Sliding Discrete Fourier Transform (SDFT). Mousa et al. [62] propose the use of the sidelobe leakage phenomenon of SDFT for induction machine rotor bar fault detection. The results demonstrate that the proposed method overcomes limitations, such as rotor bar fault detection under no-load conditions, where methods like ZFFT and MUSIC typically fall short.

5.2. Time–Frequency-Domain Analysis

In many applications, like renewable energy generation, operating conditions are typically non-stationary. Speed and load changes impose non-stationarities in the acquired signals, where FFT cannot be applied. In this direction, time–frequency (TF) analysis tools are used. The extension of FFT to TF analysis is called Short-Time Fourier Transform (STFT). STFT relies on a window function that slides over the non-stationary signal in steps. In every step, FFT is computed, and finally, the TF representation is extracted by combining the computed FFTs. Panagiotou et al. [41] address the optimal selection problem of the STFT window for broken rotor bar diagnosis on induction machines. In this work, a lower limit is set for the length of the window required for the diagnosis of broken bars in an induction machine under a certain operating condition (speed or load). The same procedure can be followed for the selection of the window for other fault patterns. Considering PMSMs, STFT has been applied in the diagnosis of multiple faults, such as winding faults [63], static or dynamic eccentricity [64], partial or total demagnetization [65], and gear [66] faults. Similarly, in induction machines, Short-Time Fourier Transform (STFT) is widely used as a mainstream tool for detecting faults such as broken rotor bars [67], eccentricity [68], and winding issues [69] under non-stationary operating conditions.

The basic limitation of STFT is the fixed time–frequency resolution due to the use of fixed windows across all frequencies. This is addressed in wavelet analysis through multiresolution representations. Instead of using fixed windows, wavelets use short windows for high frequencies and long windows for low frequencies. This is especially useful for signals with short transients, where wavelets can provide good localization in both time and frequency. Wavelet-based methods commonly used for fault diagnosis include the Continuous Wavelet Transform (CWT), Discrete Wavelet Transform (DWT), Wavelet Packet Transform (WPT), and Empirical Wavelet Transform (EWT). In [69], the power output of a doubly fed induction generator is monitored, while wavelet analysis is performed for extracting specific fault-related frequencies. CWT is applied and successfully detects eccentricity faults in a real-world wind turbine application of a doubly fed induction generator. The authors emphasize that the detection of the eccentricity fault using the CWT occurred three months prior to a bearing failure, demonstrating the method’s effectiveness for early fault diagnosis. In [70], rotor winding interturn short-circuit fault detection and severity calculation are addressed for a DFIG. The authors successfully apply both FFT and DWT for analysis of the rotor current for ITSC fault detection. The main advantage of using the Discrete Wavelet Transform (DWT) in this study lies in its ability to combine specific wavelet coefficients with the RMS value of the rotor current to formulate a fault severity estimation index. In [71], incipient bearing fault detection for an induction generator is addressed. The authors propose a DWT-based filter to extract the energy of fault-related components and create a fault index. The method is validated through an experimental procedure, and it is also compared with the Wiener filter-based noise cancellation technique, yielding similar results. Therefore, it can serve as an effective alternative for detecting incipient bearing faults. EWT was applied to vibration signals for wind generator bearing fault diagnosis in [72]. The authors enhance the EWT-based fault feature extraction with the use of denoising methods before signal decomposition. It is shown that the proposed method can detect both weak fault signatures and compound fault scenarios.

One of the limitations of wavelet-based methods is the selection of the mother wavelet, which relies on the structure and the properties of the signal that is analyzed. In contrast, signal decomposition methods rely on the data structure without a pre-defined basis. They decompose the signal based on the oscillatory characteristics of it. The most widely used method is Empirical Mode Decomposition (EMD). The goal of signal decomposition methods is to decompose the signal into multiple Intrinsic Mode Functions (IMFs) and a residual signal. The advantages of EMD for rotating machinery fault diagnosis are highlighted in [73]. In this paper, EMD is applied to a power generator in a thermoelectric plant. The results show that EMD is capable of extracting rotor faults, such as rotor-bearing impacts. Also, a comparison is made with the discrete wavelet decomposition, where it is shown that the adaptability of EMD overcomes limitations of DWT, such as frequency overlapping. EMD has also been applied to gearbox fault detection through decomposition of vibration and acoustic signals. Amarnath et al. [74] applied EMD to gearbox acoustic and vibration signals along with statistical indices to identify gear faults. It has been shown that statistical parameters—such as kurtosis and crest factor—computed on the IMFs derived from EMD outperform the corresponding parameters calculated directly on the raw signals and thus can be used for reliable fault detection. The extracted IMFs from EMD can be further processed via Hilbert Transform to extract the instantaneous frequencies. The resulting method is called the Hilbert–Huang Transform (HHT). In [75], HHT was enhanced with wavelet packet transform and then was applied for fault diagnosis of rolling bearings in comparison with wavelet transform alone. The results showed that HHT performed better in terms of frequency and time resolution than the extracted scalograms, while it had better computational efficiency.

A commonly observed limitation of EMD that can also affect HHT is the occurrence of the mode-mixing phenomenon, which is caused due to the insufficient separation of the different IMFs. To address this issue, Ensemble EMD (EEMD) was proposed, where EMD data analysis was assisted with artificial noise. In [76], EEMD was applied to the wind turbine bearing failure detection to extract useful features from stator homopolar current. It is shown that specific extracted IMFs of the signal are energized when a failure in bearings occurs. Thus, they can be used as indicators of both fault occurrence and severity. In [77], Chen et al. employed EEMD combined with Hilbert square demodulation to analyze vibration signals for wind turbine gearbox fault diagnosis. EEMD was enhanced with a resampling criterion on raw signals, and then the IMF with the highest kurtosis value was demodulated through Hilbert square demodulation to extract fault information through spectral analysis. Variational Mode Decomposition (VMD) is an alternative approach that was proposed more recently and overcomes limitations of EMD, such as mode-mixing, end effects, noise sensitivity, and lack of theoretical foundation. In [78], VMD was applied for detecting rub-impact faults of rotating machinery. This method can detect multiple signatures of rub-impact faults and outperform other methods such as EWT, EEMD, and EMD. Moreover, in [79], VMD is combined with the refined composite multiscale dispersion entropy (RCMDE) for interturn short-circuit fault detection in DFIGs by monitoring the external leakage flux of the generator. The proposed method can detect and discriminate interturn short-circuit faults in both stator and rotor through the fault-sensitive selected IMFs of VMD and the proposed RCMDE index. The advantages of VMD versus EMD and wavelet methods have also been highlighted in [80] for the case of hydraulic generator bearing fault diagnosis and in [81] for demagnetization feature extraction and fault diagnosis in axial-flux permanent magnet synchronous generators.

Linear methods include FFT, STFT, and wavelets, where a linear operator is applied to the signal. Bilinear time–frequency representations provide higher joint time–frequency resolution by quadratic processing of the signal. Wigner–Ville distribution (WVD) is the most widely used method in time–frequency distribution analysis. In [81], WVD is applied for eccentricity and rotor asymmetries fault detection under startup conditions of an induction machine. Moreover, a quantification index based on WVD in the startup interval is introduced. It is highlighted that WVD provides better resolution compared to DWT, but artifact effects should be considered. These artifacts arise from the bilinear interaction of multiple signal components in the computation of the TF distribution. To reduce them, kernel methods have been proposed, such as Pseudo WVD, where a time-smoothing kernel window is used to suppress the cross-terms. Other approaches include Cohen class distributions or adaptive methods. In [82], an in-depth illustration of these approaches is examined.

5.3. Comparison of Methods

Multiple faults can occur in electric generators, which rely on both the application and the type of electrical machine used. No single signal processing method can be considered universally optimal for all fault detection scenarios; therefore, a comparative analysis and an assessment of their applicability are essential. In [83], a comparison of STFT, CWT, Pseudo WV, and HHT is performed for fault detection of a wind induction generator through stator current. STFT provides better computational cost and readability but lacks in resolution. CWT offers high readability and good resolution but requires high computational cost. Pseudo WV provides good resolution and medium computational cost, but it is restricted in readability. Finally, HHT provides very good resolution with good readability but is computationally intensive. In the review paper [84], different time, frequency, and time–frequency methods are compared for wind turbine condition monitoring. More specifically, Synchronous Sampling, Hilbert Transform, Envelope Analysis, Statistical Analysis, FFT, STFT, and wavelet methods are considered. The comparison is addressed by means of resolution, computational cost, non-stationary signal analysis capabilities, sampling rate requirements, and commercially available solutions. It is highlighted that FFT has the highest frequency-domain resolution, with medium complexity and sampling rate requirements, and it is used in commercially available systems, but cannot handle non-stationary signals. STFT and wavelet transform have medium resolution in time–frequency analysis, can handle non-stationary signals, and require a medium sampling rate; while STFT has higher complexity, it has not been applied in commercial systems. Synchronous Sampling method, Hilbert Transform, and Envelope Analysis provide high resolution in the time-domain analysis, have medium to low computational needs, can handle non-stationary signals, and need a medium to high sampling rate. From them, only the Synchronous Sampling method is not available commercially. Statistical methods vary in resolution in both time- and frequency-domain; due to their dependence on the signal, they have low computational cost, can handle non-stationarities, do not require high sampling rate, and are commercially available. In [85], current signals from a wind turbine induction generator are used for fault detection through different frequency and time–frequency-domain analysis techniques. First, FFT, along with Welch’s periodogram, is used. It is shown that Welch’s approach can provide better resolution in harmonics that are close to the fundamental of the current and are difficult to identify. Cepstrum analysis is also applied to identify periodicities in the frequency spectrum, which can provide better identification of the sideband harmonics near the fundamental of the current. To include time in the frequency-domain analysis, STFT and Wigner–Ville distribution are employed. Wigner–Ville distribution provides high resolution time–frequency analysis but requires high computational resources and suffers from cross-term interference. In addition, HHT and DWT are applied. The decomposition of the signal into the different modes provides better analysis of the fault-related components. The main challenges that are highlighted by the authors are the selection of the mother wavelet for the DWT case and the mode-mixing of IMFs in HHT.

A comparative study of periodogram, Welch’s periodogram, Spectrogram, and Scalogram is conducted in [86] for fault diagnosis in a Wind Turbine Induction Generator Drive Train. It is shown that Welch’s periodogram and periodogram have similar performance for all cases except the case of noisy data, where Welch’s one performs better in terms of signal-to-noise ratio. Spectrogram and Scalogram performed better in the case of time identification of the fault due to the time–frequency analysis. Another comparison is conducted in [87] for advanced signal processing methods for bearing fault detection in wind turbines. The different methods include time-domain statistics (RMS, peak-to-peak, crest factor, and kurtosis), frequency, order, envelope, and envelope order spectrum. From the above methods, time-domain indices have the lowest detectability, while envelope spectrum has the highest. In terms of parameter tuning, time-domain methods do not require any, whereas all the others need high-pass and band-pass filters. If the parameters are not selected appropriately, they affect all methods. Considering automated analysis for fault detection, time-domain indices and frequency spectrum can be extracted easily, while other methods present challenges due to the computational and configurability needs. In [88], TF methods (STFT, CWT, DWT, and time-series data mining) are compared in a Hydrogenerator Fault Detection task under noisy operating conditions. STFT can provide sufficient results for reliable fault detection for SNR below 40 dB, given the right selection of the length of the window. Considering wavelet methods, both CWT and DWT provide lower computational complexity than STFT, while they offer good noise immunity for data interpretation and fault detection. Finally, the time-series data mining approach appears to be highly sensitive to noise, which is not appropriate for a typical industrial application. In

Table 1. Comparison of signal processing methods for electric generators.

Methods	Characteristics	Computational Complexity	Advantages	Disadvantages
FFT	Non-parametric Fourier-based	$+$	Fast, Simple, well-established	Unsuitable for non-stationary signals
Welch’s Periodogram	Non-parametric Fourier-based	$++$	More robust and noise-resistant than FFT	Unsuitable for non-stationary signals
MUSIC	Parametric, High-resolution Spectral Estimation	$+++$	Very high frequency resolution	Sensitive to noise, high computational cost
STFT	Fourier-based Time–Frequency Analysis	$++$	Simple, well-established	Resolution limited due to fixed window
CWT, DWT	Multiresolution Time–Frequency Analysis	$++$	Muti-scale analysis, deals well with non- stationary and transient signals	Choice of mother wavelet
HHT	Adaptive, Data Driven	$++$	Data adaptive	Mode mixing, end effects
WVD	Quadratic Time–Frequency	$+++$	High time–frequency resolution	Cross-terms, high computational complexity

, a comparison of the different signal processing for electric generators that were examined is presented by means of the basic characteristics, computational complexity, advantages, and disadvantages.

6. Machine Learning Methods

In recent years, advancements in machine learning for fault diagnosis tasks—driven by enhancements in computational power, availability of large datasets, and algorithmic innovations—have accelerated rapidly. This increasing trend is illustrated in Figure 6, where publications on ML-based fault diagnosis in electric generators during recent years are presented. The data were extracted from the Web of Science (WoS) using keywords related to generator fault diagnosis and machine learning.

While ML methods are unnecessary for simple and straightforward faults or low-cost applications where typical fault diagnosis schemes are adequate, they can provide significant advantages in cases when automated, early, complex, and optimized fault diagnosis is needed. More specifically, ML methods offer automated fault diagnosis capabilities to the electric generator system, as historically acquired or continuously monitored data from sensors are employed for detection and identification of different failure modes [8]. That is useful for all electric generator applications, as both resource efficiency and operational safety are achieved. Considering early fault detection of electric generators, ML is used where fault signatures in the sensing quantities are weak or have small deviations from the normal operating state. Typical faults in the early stage are stator winding deterioration, bearing lubrication issues, or bearing wear [9]. Moreover, ML-based methods can be employed for recognition of complex relationships between different fault signatures, which traditional methods (Section 5) fail to identify [89]. That is also useful in cases where compound faults occur, like bearing and gear faults, where multiple fault signatures co-exist and correlate [90]. In electric generator applications, like wind and hydroelectric generators, operating conditions are non-stationary due to variable load torque and rotational speed. ML can adapt and handle better non-stationary and nonlinear measured signals than conventional methods [91]. As presented in the previous sections, to evaluate the condition of an electric generator, multiple sensing quantities can be monitored, and different signal processing methods are applied to extract useful indicators of faults. By using ML methods, important feature selection for electric generator fault diagnosis can be achieved, reducing the need for applying extensive data and signal processing [92]. Overall, credibility of fault diagnosis in electric generators can be enhanced through ML by providing accurate and early identification of nuanced, complex, and important fault signatures.

Machine learning methods can be divided into three basic categories based on the presence of labels in data. The categories are supervised, unsupervised, and semi-supervised learning. The categorization of the methods is presented in Figure 7.

6.1. Supervised Learning

Supervised learning involves the use of labeled data for training a model. The goal is to create a model that maps known inputs to outputs through learning the behavior of the modeled system. Classification and regression are the two types of learning problems, where the first’s output is a category or a class and the second’s output is a continuous quantity. In the context of fault diagnosis of electric generators, classification is chosen due to the need for identifying the existence and the type of a fault. Regression tasks can only be applied to fault severity estimation, but it is noted that this can also be approached as a classification task. Some of the most important methods in supervised learning are K-nearest neighbors, Naïve Bayes, decision trees, Support Vector Machines, and Neural Networks.

K-NN and Naïve Bayes are two of the basic machine learning methods. K-NN is a simple, non-parametric ML method that classifies data based on K most similar ones, which are called neighbors. The algorithm stores the labeled data and then, during prediction, it computes the Euclidean distance of new points to the stored ones in order to assign a class to them. Naïve Bayes is a probabilistic algorithm based on Bayes’ theorem. For each feature and class, the method learns the distribution of it. In prediction, the method calculates the probability of each class and assigns each point to the respective one. In [93], k-NN is applied for intelligent fault diagnosis of wind turbine blade damage. Microphone signals were acquired for condition monitoring, which were fed to the k-NN algorithm. The proposed method achieved accuracy of 98.9% while preserving simplicity and interpretation. K-NN has also been applied to rotor fault diagnosis of a synchronous machine turbogenerator in [92]. An experimental setup was created considering both normal operating state cases, eccentricity, turn-to-turn, and combined faults. The author used three-phase currents and voltages along with the radial flux density for condition monitoring. After feature selection, a decision rule based on k-NN was proposed. The classification accuracy for the case of combined faults reached 77.4% while the rotor faults classification accuracy reached 85.1%.

A decision tree is another fundamental ML method that uses a tree-based architecture. It builds a tree with branches and leaves, where the nodes represent features, branches are the outcomes of comparisons/tests, and leaves are classes. Based on decision tree architecture, multiple approaches have been proposed. Based on the tree structure and the ensemble concept, bagging and boosting methods are built. Random Forest (RF) is a bagging-based approach that uses random sampling of data and features to build multiple decision trees. In [94], RF, along with XGBoost, was proposed for wind turbine fault detection. RF was used for ranking the features based on feature importance. In this way, redundant information is not used, and important fault indicators are employed. Then, the selected features were fed to XGBoost to complete the fault detection framework. The proposed method is successfully applied to three different simulated fault conditions. Considering gearbox faults, in [95], the task is addressed with the use of RF. More specifically, the authors utilize acoustic and accelerometer measurements along with the wavelet packet transform. Then, deep Boltzmann machines are developed for deep representations of the WPT features. Finally, Random Forest is employed for fault diagnosis. The proposed method achieves 97.68% accuracy for 11 different operational conditions. In this case, Random Forest was used as a fusion approach for combining the outputs from the deep features. However, RF has some limitations, such as poor performance under noisy and redundant data. This challenge is addressed in [96], where kernel Principal Component Analysis (PCA) is applied before RF to reduce the training data of RF and keep the most important information. The method is applied to a wind energy conversion system equipped with an asynchronous generator, where it is shown in the paper that the proposed method can perform better than other classical ML methods.

Support Vector Machine (SVM) is another powerful ML method that has been applied to electric generator fault diagnosis. The method finds an optimal boundary, which is a hyperplane, and separates the data into classes. The closest points to this hyperplane are the so-called support vectors. Different SVM approaches for condition monitoring and fault diagnosis of multiple machine systems are presented in [97]. Rolling element bearings, induction machines, machine tools, pumps, compressors, valves, turbines, and HVAC systems are considered for SVM-based fault diagnosis. This work highlights the high performance of SVM in fault classification tasks, but also the need for SVM improvement. Limitations of SVM include the computational cost, hyperparameter tuning, and its performance on noisy or overlapping data. In this direction, SVM is often combined with other methods, such as in [98], where rough set theory is adopted and applied. The proposed method is able to describe complex and overlapping data relationships, which are suitable for fault diagnosis of hydroelectric generators.

In [99], Logistic Classifier; Linear SVM; and the tree-based methods of decision tree, CatBoost, XGBoost, and Random Forest were employed and compared for Wind Turbine Generator Shaft Misalignment fault diagnosis. Controller loop signals and mechanical signals were used for training and testing the aforementioned classifiers, reaching a validation accuracy near 100%. The methods, ranked from best to worst for the specific case, are the following: Logistic regression, SVM, CatBoost, XGBoost, Random Forest, and decision tree.

Artificial neural networks (ANNs) are a powerful class of ML methods that mimic the function of the human brain. They are increasingly adopted for fault diagnosis tasks in electric generators due to their ability to model complex and nonlinear relationships in data. The main categories of ANNs are the Feedforward Neural Networks (FNNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Radial Basis Function Networks (RBFNs), and combinations of the above. Based on the fault diagnosis task and the characteristics of the data or the features, the appropriate architecture is selected. For example, in [100], CNNs are selected due to their ability to handle infrared images for fault diagnosis. More specifically, three different CNN models (Inception, Xception, and MobileNet) are used for feature extraction from infrared images. Then, DWT is employed for simultaneous processing of features from all models. Finally, PCA, along with classical classifiers (KNN, SVM, QDA), is used for fault classification. The proposed method achieves 100% accuracy, outperforming end-to-end CNN models, with significantly lower training time. In [101], a 1D-CNN is utilized for wind turbine generator fault detection. SCADA data from a real-world operating farm in China are used to develop and validate the proposed method. Authors use ‘meta-learning’ principles to develop an effective fault detection method that overcomes limitations of other methods, such as early fault detection and high false positive and negative rates. In [102], RNNs are used for wind power generation system multi-fault diagnosis based on multiple time-series variables such as wind speed, rotor speed, generator speed, and power generation. Ten different fault cases are considered. The accuracy of the method reaches 98.9% and 99.8% with and without data interference, respectively. While traditional ANNs have been widely used for pattern recognition and fault diagnosis, challenges occur when it comes to bigger and more complex data.

Driven by computational power capabilities and the availability of large amounts of data today, deep learning (DL) neural networks were developed. Due to their importance and the growing applications, deep learning is examined separately in one of the following chapters.

6.2. Unsupervised Learning

Unsupervised learning does not rely on labeled data, and the basic goal is to discover patterns and groups in the examined data. Common tasks in unsupervised learning include clustering and dimensionality reduction. Some of the most important methods in clustering are K-Means Clustering and density-based spatial clustering of applications with noise (DBSCAN), while in dimensionality reduction, they are the Principal Component Analysis, t-Distributed Stochastic Neighbor Embedding (t-SNE), and autoencoders.

In fault diagnosis tasks, clustering is very useful for identifying different operational and faulty conditions in historical data, while it can be used for identifying previously unseen patterns and exploring the data. Based on the created models, newly acquired data can then be grouped into different clusters, indicating faulty or abnormal conditions. K-means creates clusters of unlabeled data based on the distances from computed centroids. In [103], an exploratory data analysis from SCADA measurements was applied using the K-Means method. The optimal number of clusters was chosen based on the “silhouette” factor. Wind speed, active power, and oil pressure were used for clustering the different groups of data. It was shown that K-means clusters can point out which wind turbines in a wind farm operate differently, leading the specialist to further analysis. Moreover, if a wind turbine behaves differently due to a faulty case, a cluster of the model will capture this behavior. In [104], K-means was applied to wind turbine gearbox oil temperature estimation. Clusters were selected according to previous knowledge, while univariate and multivariate variables were evaluated for clustering, such as wind speed, active power, impeller speed, and gearbox oil temperature. Moreover, an Extreme Learning Machine Neural Network was proposed along with K-means clustering. The total model was able to estimate the gearbox oil temperature with an absolute error and MAE (Mean Absolute Error) of 0.8 and 0.2 degrees Celsius, respectively. Another approach was proposed in [105] with application to SCADA data from a wind park in northwestern China. K-Means was used for creating clusters of data for the different time-varying operational conditions. Then, a SoftMax model was used for obtaining the operating condition of the wind turbine in real time, along with a health index based on Mahalanobis distance. The proposed method can recognize the running condition of the wind turbine and identify fault conditions. In [106], the importance of clustering in fault diagnosis on unlabeled data is demonstrated through a case study on rolling element bearings’ fault detection task. K-means is applied for clustering spectral-based features that contain important fault-related information. The combination of expert knowledge on feature sets and K-means clustering leads to 100% classification accuracy in multiple cases. In [107], wound rotor induction generator winding fault diagnosis is assessed through unsupervised learning. Short-circuit faults were emulated in stator and rotor windings, while rotor current, stator voltage, and speed were monitored under no-load and load conditions. K-means was applied to create clusters of the different operating conditions to identify and isolate faulty operating states. The best clustering results were observed when stator current harmonics were used as single-signal features, while excellent performance was also observed when multi-signal features were used.

Another clustering approach is DBSCAN. Data is grouped into clusters based on the density of the data points in the data space. Compared to K-means, DBSCAN does not need a specification of the number of clusters in advance. In [108], DBSCAN was applied along with Isolation Forest (iForest) for identification and isolation of anomaly data. Different datasets with SCADA data from wind farms were used, and the proposed method managed to identify different anomalies from the data, with lower computational resources compared to other approaches, such as Neural Networks. Similarly, in [109], DBSCAN was applied for normal and abnormal state data classification, where it successfully classified irregular performances of wind turbines from collected data from 31 wind turbines in Taiwan.

Dimensionality reduction refers to the reduction in the number of variables or features in a dataset while preserving the most important and essential information in them. In fault diagnosis tasks, dimensionality reduction serves as a feature extraction and selection method for selecting and keeping the most fault-relevant information from the dataset. That is particularly important for eliminating noise, improving diagnostic accuracy, and reducing computational complexity, especially when dealing with high-dimensional datasets. PCA is one of the most widely used dimensionality reduction methods. It transforms an original set of features into a smaller set of variables, which are the Principal Components, that retain most of the data’s original variation. In [110], a fault feature extraction method for a Gas Turbine Generator System was proposed. The authors used a set of wavelet-packet time-domain features from vibration signals. To reduce the redundant information and keep the most important one, they applied the kernel PCA. A set of experiments was conducted for the appropriate selection of the kernel. The extracted PCA features were then fed to an SVM model to highlight the efficacy of the method. Condition monitoring of electric generators, such as wind turbines, usually requires multiple sensors and a high sampling frequency. That leads to high amounts of data to be transferred and processed. To reduce the dataset and keep only important information, the use of PCA is proposed in [111] for optimal variable selection. The proposed approach was validated using simulation data, laboratory experimental data, and SCADA data collected from operating wind farms. It was shown that the proposed method was able to identify a set of variables with adequate information and low correlation between them. Moreover, the authors combined the selected variables with ANN and anomaly detection methods for identifying and quantifying different faults.

Another widely used method in dimensionality reduction is t-SNE. Compared to PCA, t-SNE is better suited for handling nonlinear dimensionality reduction tasks, while it is more efficient in uncovering clusters in data. In the first step of t-SNE, similarities are defined in the high-dimensional space, and then similarities are found in the low-dimensional space, using pairwise distances and distributions. Then, Kullback–Leibler (KL) divergence is employed to quantify the differences in high- and low-dimensional spaces. Like PCA, it is often applied in combination with other ML approaches for feature extraction and selection. In [112], t-SNE is employed for feature dimensionality reduction for a rolling bearing fault diagnosis task. Sigmoid-based refined composite multiscale fuzzy entropy method is proposed for fault information extraction from vibration signals. T-SNE for feature reduction is applied, and finally, the proposed variable predictive models-based class discrimination method is used for fault pattern recognition. The importance of the t-SNE method is reflected in the overall performance of the proposed framework, as it effectively captures critical feature information and reveals underlying patterns in the data. In [113], t-SNE was used in combination with a Generative Adversarial Network (GAN) method for intelligent fault diagnosis of rotating machinery. In this case, GAN was employed for dealing with limited data availability, which is common in many fault diagnosis tasks. T-SNE was used for feature reduction and visualization of the GAN-generated features. Clusters of normal operating state and faulty conditions were identified, indicating the efficacy of the GAN-proposed method. The same approach for evaluating the performance of GAN models is adopted in [114]. T-SNE is used for comparing three different GAN approaches by means of clustering and discrimination.

Autoencoders are a specific type of artificial neural network that were created for feature extraction, dimensionality reduction, and anomaly/fault detection. The architecture of an autoencoder consists of two main stages: an encoder, which compresses input to a lower-dimensional space, and a decoder, which reconstructs the original data from the latent space. In dimensionality reduction, the latent space of the autoencoder can be used for feature reduction. However, what makes autoencoders particularly important in fault detection is their ability to learn the overall characteristics of the normal operating state. Then, they can be used for fault detection on newly inserted data. Based on the architecture of an autoencoder, different variants exist, such as variational autoencoders, stacked denoising autoencoders, or convolutional autoencoders. In [115], a variational autoencoder is used for vibration signal-based bearings fault detection on the CWRU [116] dataset. The extracted features from the VAE are fed to a CNN classifier for further fault classification. The proposed method reaches 99.7% accuracy. In [117], a stacked denoising autoencoder is combined with a Long Short-Term Memory (LSTM) Network for wind turbine generator fault diagnosis. Multivariate data are used from the autoencoder-based model for accurate normal behavior learning. By using the proposed method, faults in the early stages can be detected, where traditional methods fall short. Moreover, the authors propose the use of an additional stacking ensemble method for fault classification in addition to fault detection. In many real-world scenarios, data are difficult to obtain or can be acquired only for specific operating conditions, like the normal operating state. In [118], gearbox fault diagnosis is addressed with limited and imbalanced samples by using a combination of a variational autoencoder and a GAN. The encoder part of the model is used for obtaining the distribution of fault samples. Then, the decoder part is used for fault sample generation. The trained model is finally used for intelligent fault diagnosis of a gearbox, where it performs better in terms of fault classification accuracy when compared with other methods.

6.3. Semi-Supervised Learning

Semi-supervised learning combines the two aforementioned categories. In cases where labeled data are hard to collect or limited, the model learning procedure starts with a small, labeled dataset and continues with the unlabeled data in order to improve the generalization of the model. A case using limited data is addressed in [119] for rolling bearing fault diagnosis. Authors use a three-stage semi-supervised approach, where data augmentation, K-means, and KL-divergence loss are used in each stage, respectively. Data augmentation expands the feature space for the limited dataset, while K-means is employed afterwards for obtaining cluster centers of the limited data samples. The unlabeled data are labeled based on the clusters and the feature distribution. Finally, Kullback–Leibler divergence loss is used for minimizing the distance between features and cluster centers. The proposed method reaches an average accuracy of 96%, overcoming performance and limitations of other approaches.

In a similar approach, authors in [120] apply a convolutional autoencoder that extracts discriminative features from unlabeled vibration signal measurements. Moreover, a SoftMax classifier is employed along with the encoder, which obtains the normal operating condition of the motor from labeled vibration samples. The proposed method is verified in the CWRU bearing dataset and on a hydro generator rotor fault diagnosis task. In both setups the proposed method reaches high accuracy with limited labeled training samples. A comprehensive study on semi-supervised ML for fault detection and diagnosis was conducted in [121]. The authors present a structured taxonomy of the methods, while they highlight important limitations and future trends. Some of the most important points, aligned with the findings of our study, include industry’s need for semi-supervised learning due to the limitation in acquiring labeled data and the need for deep semi-supervised learning, where high-accuracy deep learning models can be employed to deal with limited data scenarios. In [122], the authors proposed a novel framework for wind turbine fault diagnosis with limited data availability. They propose the use of adversarial training for generalization of the feature representation of the unlabeled dataset. The experimental procedure consists of the normal operating state and five different faulty operating states (bearing pedestal loosening, rolling element fault, inner and outer race fault, and shaft coupling misalignment. The results indicate high performance of the proposed method, with accuracy higher than 90%, even with a small portion of labeled data in the used dataset.

6.4. Deep Learning

Deep learning refers to artificial neural network architectures with many (deep) layers. The main advantage of deep learning is the elimination of preprocessing and feature extraction stages, as it can automatically extract meaningful features from raw data and offer an end-to-end learning framework, as shown in Figure 8.

Moreover, DL demonstrates very good performance in cases where large datasets are available. Due to their growing performance and applications, multiple reviews on deep learning fault diagnosis on generators and rotating machinery have been conducted. In [123], the authors conducted an overview of DL methods in wind turbine fault detection tasks. Their work included supervised and unsupervised applications of deep ANNs for condition monitoring of wind turbines. From their findings, unsupervised approaches are more popular in DL methods due to the limited data availability, the imbalance in datasets, and their overall quality. Since wind turbine fault detection relies more on SCADA data, the authors proposed the extension of their dimensionality in order to improve the performance and the generalization capability of DL approaches. In another wind turbine generator fault diagnosis review paper [26], the authors focus more on the generator system of wind turbines. They conduct an overview of the mechanical and electrical failures of the different parts, along with a failure analysis of them. A taxonomy of the methods is also conducted based on model, signal, knowledge, and hybrid-based approaches. Among the key findings of the authors is the recognition of the need for hybrid approaches, as they demonstrate higher performance than classical methods in most cases. Moreover, the authors propose a qualitative and quantitative assessment of false rate performance. Considering bearings’ fault diagnostics in electric generators, a comprehensive review of DL approaches was conducted in [89]. The performance of the different DL approaches is compared in terms of classification performance. CWRU bearing dataset-based methods are compared for evaluating the highest performance of the different DL frameworks. This work highlights several key points, including data quality issues (noisy data, imbalanced classes, or limited labels), size of data, model complexity, generalization ability of the different models, and the importance of model explainability and interpretation.

Sensor systems for data acquisition and transmission play an important role in DL applications. As reported in Section 4, multiple sensing quantities can be explored for automated fault diagnostics through DL. However, the most used sensors for DL-based applications are vibration sensors, temperature sensors, and current/voltage sensors. Considering DL bearing fault detection, vibration sensors are commonly used due to the high reflection of mechanical characteristics of the system [89]. In wind turbine applications, vibration-, current-, and temperature-sensing DL applications have been reported for condition monitoring of electric generators [90]. Considering data acquisition for DL, vibration and temperature sensing can be achieved by placing accelerometers and temperature sensors on critical spots of electric generator along with wired (e.g., Ethernet) or wireless (e.g., Wi-Fi) transmission. Current sensing does not require access to the site of the electric generator and can be achieved from the power output of the generator. Compared to vibration and temperature, wired transmission does not require cables near the generator area [124]. In contrast to extensive data transmission, edge AI approaches [124] employ DL methods on the site, within embedded systems on the generator site. In this way, only decisions and important data are transmitted, providing real-time decisions and low latency. For example, in [125], vibration signals were used for electric generator fault diagnosis through a Random Forest classifier. The model achieved high performance metrics up to 99% while the model was employed in an edge AI system, integrating Raspberry Pi, offering a cost-efficient approach.

7. Remarks, Limitations, and Future Perspectives

Stator and rotor winding faults are considered the most challenging ones due to rapid evolution that can lead to catastrophic failures. Winding fault diagnosis procedure leans on detecting subtle changes in current/voltage signatures, magnetic flux, and vibration signals, often caused by interturn shorts or insulation breakdown. Methods that combine signal and model-based techniques together with data-driven techniques to localize fault and assess severity in real-time are gaining interest, especially where varying load conditions affect the diagnostic procedure, adding challenging complexity. Generator bearing faults can be detected by monitoring vibration, thermal, or acoustic signals that derive from wear, fatigue, or ongoing lubrication degradation. These monitoring methods, paired with signal-processing and data-driven diagnostics, are increasingly adopted for early warning and real-time maintenance in turbines and large machines. In PM generators, demagnetization faults manifest through magnetic field imbalances, vibrations, and current harmonics distortion. For the diagnosis procedure to be effective, these subtle signatures must be detected via flux monitoring, signal processing, or data-driven algorithms, considering any machine-specific parameter that may conceal fault patterns.

Signal processing methods selection varies depending on the quantity measured and the properties of the signal. The varying operating conditions of real-world electric generators pose challenges that necessitate the use of advanced signal analysis techniques. For this reason, in this review, the various techniques are compared, and recommendations are provided for the selection of the appropriate technique that satisfies the requirements of each application. With the integration of ML in fault diagnosis tasks, it can be concluded that new fault detection methods can lead to automated and credible fault diagnosis schemes, resulting in more reliable generator operation. However, several challenges persist in fault diagnosis regarding multiple sensing quantities, advanced signal processing, and ML methods. With the advancements of sensor technology, optimal sensor selection, but also sensor fusion techniques, need to be further investigated. Moreover, embedded and edge processing is a challenge in applications like wind turbines, where transmission of large volumes of data is limited. Limited data scenarios and different operating conditions across similar electric generators, which reflect real-world cases, present a significant challenge. In this direction, semi-supervised learning and transfer learning should be investigated further. Nevertheless, interpretation and explainability of ML methods are also critical in fault diagnosis tasks for root cause analysis, but also for enhancing trust in ML-based fault diagnosis decisions. In Figure 9, important challenges that determine future directions in fault diagnosis of electric generators are demonstrated.

8. Conclusions

Condition monitoring and fault detection can ensure reliability and stability in electric generators. Especially in large generators and the ones that are critical for power grid operation, their use is essential. The aim of this paper is to provide brief information regarding all the faults that may occur during electric generator operation and the state-of-the-art methods for their monitoring and identification. Fault types are categorized, and the signals required for condition monitoring are presented. In addition, a presentation of traditional and innovative methods for fault detection and diagnosis is provided, including fault diagnosis schemes, advanced signal processing, and machine learning methods.