A Review of Stator Insulation State-of-Health Monitoring Methods

Abstract

Tracking the state of the health of electrical insulation in high-power electric machines has always been a topic of great interest due to the high cost of downtime associated with unexpected failures. Over the years, there have been continuous efforts to develop and improve upon methods for testing and categorizing the health and expected lifetime of stator insulation. Methods such as partial discharge, surge, and dissipation factor testing are common examples. With the increasing use of high-specific-power electric machines in new applications such as traction and wind power generation, coupled with the increasing use of wide-bandgap semiconductor device-based inverters, some traditional methods for insulation health tracking may need adjustments or be combined with newer methods to remain accurate and useful. This paper outlines a review of the traditional insulation health tracking methods and newer methods and improvements that have been proposed to address the concerns and shortcomings of traditional methods.

1. Introduction

1.1. Background

Electric machines are the cornerstone of the modern industrialized world and are responsible for a significant portion of energy generation and usage. Their ability to convert between electrical and mechanical energy efficiently means they are widely used in almost every facet of the modern world, from manufacturing, transportation, generation, and beyond. From the very beginning of machine design, the electrical insulation used in machines has been a limiting factor that has been undergoing continuous improvement and refinement to optimize machine size, efficiency, and cost [1,2,3,4]. Despite improvements, 20–25% of failures for electric machines can be traced to insulation issues [5]. Therefore, understanding and improving the ability to measure and predict insulation health continues to be of great importance.

In recent years, electric machines have been increasingly applied with inverters instead of direct line operation. Inverters have been widely adopted in numerous industries due to the advantages they can bring to an operation from reduced mechanical complexity, increased efficiency, easy controllability, and improved power stability [6,7,8]. However, it is known that the use of inverters exacerbates the insulation integrity of motor windings due to additional stress factors such as the high-frequency voltage impulses and fast voltage rise times inherent to the operation of power electronic switches [9,10,11]. Consequently, new insulation materials were used and significant changes were made to electrical insulation in motors to properly withstand the harsh conditions imposed by Si inverters [12].

Recent advances in semiconductor technology have yielded a new generation of semiconductor devices known as WBG devices. These consist primarily of silicon carbide (SiC) and gallium nitride (GaN)-based devices. They offer significant advantages over Si devices in terms of efficiency, size, voltage blocking ability, and operational frequency [13,14]. For these reasons, WBG devices and SiC inverters specifically have been quickly adopted for several applications ranging from electric vehicles to wind power generation [2,14,15,16]. According to [17], the share of SiC-based power devices across industrial and automotive applications is projected to grow significantly, with the SiC device market reaching over USD 4.5 billion by 2027, nearly triple its 2022 value. This growth is driven by the push for higher efficiency, compactness, and thermal performance in motor drives, renewable energy converters, and high-frequency inverters.

However, many of the advantages of WBG devices also directly increase the stress on the machine’s insulation, due to the faster switching and higher slew rates [4,9]. Many studies have endeavored to address the limitations of the use of WBG devices for electric drives. In the paper [18], the authors addressed the issue of uneven high dv/dt stress on the stator windings of induction motors being caused by the fast-switching characteristics of SiC-MOSFETs. They proposed three solutions: using a multilevel inverter topology, implementing a dv/dt filter, and utilizing an integrated motor drive design. Each approach showed significant potential in enhancing the reliability of WBG devices for adjustable speed drives but come at significant additional cost.

As a result, advanced insulation state-of-health (SoH) monitoring techniques are increasingly seen as important components in the preventive maintenance process required to avoid unexpected and costly downtime. This is particularly true for high-power MV and HV machines, as in some industrial applications such as mineral extraction, downtime can cost up to USD 25,000/h [19]. The need for these techniques has been recognized by researchers, and publications in the fields of machine insulation health and fault monitoring have reached exceptional levels over the past 10 years, as shown in Figure 1.

1.2. Contributions and Organization of This Review

This paper seeks to present a comprehensive review of stator insulation health monitoring, including the currently employed methods and standards for insulation testing, developing approaches, limitations, and opportunities for continued improvement. There have been several review papers published on this subject over the years, with early works such as [19,20,21,22], which discussed the contemporary methods in the early 2000s. More recently, ref. [23] presented a review focused on low-frequency methods and [24] many of the online tracking approaches that have recently been gaining popularity. Further work has also been carried out to investigate thermal aging [25] and partial discharge (PD) [26] testing specifically as well as approaches specific to low-voltage machines [2].

The novelty and contribution of this paper is to provide a comprehensive review of both the approaches currently employed in industry and the state of the art for developing SoH tracking methodologies. The techniques were selected based on their demonstrated relevance in industrial standards and operational practice, as well as their emerging potential in academic research. Both offline monitoring approaches such as PD, dissipation factor, and surge testing will be discussed in addition to online methods with the current limitations and avenues for improvement explored. The remainder of this work is organized as follows: In the second section, PD monitoring methods are discussed. Following this, section three is an analysis of offline testing methods such as capacitance tracking, dissipation factor, insulation resistance, and surge testing with their accepted standards and recent improvements. The numerous, new, online tracking methods are discussed in section four, with their similarities and differences to offline methods noted, and the primary limitations of the methods are identified. Lastly, Section 5 provides a comparative evaluation of all the SoH tracking methods. This review paper is needed due to the new stresses placed on insulation health by WBG devices and the relative upsurge of insulation health tracking methods that in some cases deviate greatly from the traditional industrial methods.

2. Partial Discharge Tracking

One of the primary methods for insulation SoH monitoring is tracking the partial discharge (PD) activity, also referred to as a corona discharge. PD has long been used as both an online and offline measurement method as recognized by IEEE 1434 [2,27,28,29,30,31]. PD is a behavior wherein electrical impulses create high electric fields within the insulation, exceeding the critical strength of the insulation. As the voltage between the conductor and system ground rises above the PD inception voltage (PDIV), the high external electric field E_ext creates free electrons within the insulation. PD inception occurs at t_PD-f, when a critical mass of free electrons and ions is reached, and an avalanche occurs between the conductor and the ground, and an internal electric field (E_int) is strengthened. As the voltage is reduced at the end of the pulse, E_ext diminishes until a backwards discharge occurs. This behavior is illustrated in Figure 2 and leads to a localized electrical discharge within the insulation, which causes significant damage to the insulation. Due to this, PD has been studied in some detail with specific attention paid to the inception voltage of PD activity for different test conditions and materials [32,33,34,35,36].

Partial discharge tests are more effective when performed at the end of the machine terminals by energizing one phase at a time with the other two connected to the ground. The device used to record PD during tests is usually an oscilloscope. However, in current times, PD tests are conducted with pulse phase analyzers or phase-resolved PD (PRPD) analyzers. The difference between the PRPD and the scope is that not only does it provide the waveform image, but it records the number, magnitude, and phase of the PD, which is in reference to the AC power supply used. Testing PD for large generators requires a 50/60 Hz resonant or conventional transformer, usually with ratings between 20 and 40 KVA.

PD analysis was primarily developed for use with MV and HV machines [37,38,39,40], but with the advent of WBG inverters, it has also been explored for use with LV machines that have traditionally not experienced the phenomenon [2,27,41]. Additionally, some recent research has found that the high dv/dt of WBG inverters can, among other limitations, disrupt some forms of PD measurement requiring further adjustments for accurate use [11,42].

These issues will need to be resolved with novel innovations within the field such as the use of improved, noise-resistant sensors such as those proposed by Arcones and Wang [11,43] or specialized noise-filtering techniques such as those presented by Cao [44]. Another topic of some interest in improving PD tracking is the use of artificial intelligence (AI) platforms and other advanced signal processing techniques for the interpretation of PD readings [45]. One such method proposed by Wang can reduce the precision error by half in comparison to traditional methods [46]. The method proposed by Wang, shown in Figure 3, consists of an AI model trained with offline PD data that is able to identify the precursors to insulation failure. PD monitoring is a well-studied phenomenon and can warrant a review entirely dedicated to itself, e.g., [26], but its further discussion in this work is limited so as to highlight other new methods in the field of SoH tracking in the following sections.

3. Offline SoH Tracking Methods

Several of the traditional methods for insulation SoH tracking are insulation capacitance, tan-delta or dissipation factor (DF), insulation resistance (IR) or megger testing, and surge testing [19,47,48,49,50]. These methods are considered offline approaches that require the machines under test to be disconnected from their load and often require specialized test equipment. For these reasons, offline tests are impractical for many applications and cannot be conducted with the frequency required for highly accurate SoH tracking. Furthermore, several of these tests are highly focused on the diagnosis of groundwall (GW) insulation, leading to a relative neglect of turn-to-turn (TT) insulation monitoring [50]. Nonetheless, the development of offline SoH tracking methods is under continual improvement. Thus, new methods have been proposed for offline testing using components already present in many inverter drive systems, reducing the need for specialized equipment [51]. Further improvements in the analysis of offline testing data, such as with the use of AI, are also underway, as demonstrated by Nakamura’s work with impulse tests [52].

3.1. Insulation Capacitance Testing

One of the methods for offline SoH testing is the tracking of the insulation’s equivalent capacitance. To use the equivalent insulation capacitance as a metric of insulation SoH, the electrical insulation of the windings is represented as a parallel resistor and capacitor connecting the conductor to the groundwall of the machine in the form shown in Figure 4a [24]. As the insulation deteriorates, the values of the equivalent resistances and capacitances will correspondingly change. This model is used to find both the insulation capacitance and the DF metrics of insulation health. The equivalent capacitance of the insulation can be extracted from the conductor voltage and the capacitive leakage current as follows:

$$I_R\left(\omega \right)=I_{leak}\mathit{sin}\delta$$

(1)

$$I_C\left(\omega \right)=I_{leak}\mathit{cos}\delta$$

(2)

$$\delta =90°-\angle {\displaystyle \frac{V_{ex}}{I_{leak}}}$$

(3)

$$C_{eq}\left(\omega \right)={\displaystyle \frac{I_C}{\omega ·V_{ex}}}$$

(4)

where V_ex is the excitation voltage of the coil and I_leak is the leakage current of the insulation. Traditional offline capacitance testing trends the equivalent capacitance of the groundwall insulation over time, generally using the capacitance at grid frequency.

The capacitance test is more effective on smaller form-wound machines [53]. These tests can detect not only thermal degradation but moisture within the insulation and end-winding contamination [54]. To capture consistent measurements over the years, testers tend to use the same device, a capacitance bridge calibrated for that specific machine. This is because very little changes in capacitance values could cause drifts, and this can potentially lead to inaccurate values. According to [53], there are two types of capacitance testing performed. These include the capacitance map and the winding capacitance test. The winding capacitance test is usually performed directly at the stator winding terminal to avoid cable capacitance, which could distort the readings. The capacitance method involves grounding the stator winding and measuring the capacitance of a plate placed on the end windings of the form-wound coil with respect to the grounded windings. This generates a detailed map which is used for statistical analysis [53]. This test becomes more useful when it is conducted simultaneously with the DF method. This can be achieved using a balanced bridge-type device for the DF method, which measures the R and C components of the coil [53].

It is important to note that in interpreting the amount of degradation, a significant amount of thermal deterioration could only result in a capacitance drop of 1% over the years. The capacitance could reduce by 5% if the windings have been exposed to a lot of moisture, indicating the effect of the ambient stress factor.

3.2. Dissipation Factor (DF) Testing

Similar to the equivalent capacitance, Tanδ or DF testing measures changes in the insulation SoH by trending changes to the equivalent resistance and capacitance of the groundwall insulation. In DF testing, the insulation health is represented as a ratio of the capacitive and resistive leakage currents and can be found as

$$DF=tan\delta ={\displaystyle \frac{\left|I_R\right|}{\left|I_C\right|}}$$

(5)

The process for this testing approach is outlined in IEEE 286 and assumes the insulation is operating as a linear system; thus, for significantly degraded insulation, its accuracy may be reduced [49,55].

All insulation materials have some dielectric strength. Therefore, the DF/power factor (PF) tries to measure these dielectric losses in the insulation over time [53]. These tests, however, are only performed on form-wound machines. It is important to note that DF, PF, and Tanδ are each different methods that are used to test for dielectric stress in the insulation of machines. The Tanδ method, as the name describes, measures the phase angle between the current and the voltage in the stator winding. The reason is that since a dielectric material is capacitive in nature, the phase angle measured should be 90⁰. Changes in the phase angle indicate dielectric loss, which implies a change in the capacitance of the groundwall insulation. Figure 5 shows how the dielectric loss angle is calculated.

The dissipation factor is measured using a balanced bridge-type device. The bridge device has R and C components, which are varied to provide some voltage and loss angle, similar to the Tanδ method. These resistance and capacitance values are used to calculate what the dissipation will be. This method typically has high accuracy, achieving percentage errors of about 0.01% [53]. However, the downside is that the machine needs to be interrupted from its operations in order to test them. This does not allow for continuous tracking of the state of the health of the insulation and thus decreases the accuracy by a large number. Another downside of the DF method is that it is highly sensitive [53]; a long power cable can easily distort the measurements and provide inaccurate readings. The power factor method measures the voltage applied through the stator windings in order to find the current through the coils. The power can be found using both parameters and thus the power factor is expressed in percentage form. The power factor equation is expressed below as

$$Power Factor={\displaystyle \frac{P}{V\times I}}$$

(6)

where P is the active power flow in the windings, and V and I are the measured voltage and current values.

3.3. Insulation Resistance Testing

Insulation resistance (IR) tests, or megger tests, measure the resistance of the insulation under DC excitation, often at high voltages which may exceed the rating of the winding insulation. The IEEE established standards for conducting such tests in the now inactive IEEE 43 [56]. It has been suggested that these standards may need to be revised in regard to their use with strip-on-edge windings [57]. It should be noted that the resistance found by insulation resistance (IR) testing is different from that of the equivalent resistance for capacitance or DF testing. Also, from [58], the reliability of IR tests is dependent on temperature and humidity. For this reason, the test needs to be conducted at the appropriate temperature to ensure accurate results.

An adaptation of IR testing is a polarization index (PI) test, wherein a DC voltage is applied to the test winding and the change in resistance is tracked for set intervals, 1 and 10 min being standard [50,57]. The PI can be calculated as follows:

$$PI={\displaystyle \frac{R_{10}}{R_1}}$$

(7)

where R₁ is the resistance after 1 min and R₁₀ is the resistance after 10 min. Due to the simplicity of both IR testing and polarization index testing, they are often used as preliminary tests to determine if other more intensive testing needs to be performed [50,52]. Insulation resistance can be applied across all machines and windings, unlike the capacitance and the DF methods. [53]. There are four types of currents that flow when DC excitation is applied to the coils to measure insulation. These include the capacitive current, the conduction current, the leakage current, and the absorption current.

The capacitive current is the first to appear when a DC voltage is applied to the windings. This current decays exponentially in a few seconds. As noted in [53], larger stator machines may have capacitance values of up to 1 µF. This current value does not provide enough information on the groundwall insulation and can cause distortion if not allowed to delay before the insulation resistance is measured. The leakage surface current is usually created by conductive moisture or contamination such as oil mixed with dust or salt on the surface of the winding. These currents are typically dominant in strip-on-edge salient pole windings and round rotor windings. An increase in the leakage current directly informs on deterioration in the coils because under ideal conditions, there is no leakage current present in the machine. The presence of conduction current suggests a breakdown in the insulation integrity of the machine windings. It is usually due to physical damage of moisture on the insulation, which causes a steady flow of charges through the insulation. Absorption current is affected by the type of insulation material. It usually is higher at the start of DC excitation and gradually decays as the insulation material stabilizes.

Table 1. Guideline for DC voltage excitation during IR test (adapted from [53]).

Winding Rated Voltage (V)	Insulation Resistance Test Direct Voltage (V)
<100	500
1000–2500	500–1000
2501–5000	1000–2500
5001–12,000	2500–5000
>12,1000	5000–10,000

provides a guideline for DC voltage excitation during IR tests.

3.4. Surge Testing

Surge testing is conducted by applying a high-voltage pulse with a steep rising edge to the machine stator. IEEE standards recommend the use of a surge voltage of 1 kV more than twice the stator’s rated voltage when conducting such tests [59]. While the use of surge testing has been controversial for its theoretical potential to cause damage to the coil under test, it has also been recognized as one of the best ways to identify turn-to-turn insulation deterioration [19,60].

Surge tests can be considered destructive testing [53,60]. The applied high-voltage pulse is used to mimic an actual surge voltage. If the insulation fails when it is exposed to such voltage, it is assumed that the stator would also fail. If the windings survive, it is assumed that the stator is healthy and can operate for a few more years. Turn-to-turn insulation degradation could be detected during surge testing by tracking the change in resonant frequency caused by shorting one turn. This resonant frequency can be expressed as follows:

$$f={\displaystyle \frac{1}{2\pi \sqrt{LC}}}$$

(8)

Fixed-frequency oscillations imply no turn-to-turn fault in the windings. The steep rising edge from the high voltage could also cause a turn fault. A decrease in the inductance of the coils indicates a turn fault induced by the short rise time. Surge test is performed using a surge tester, and the simplified circuit diagram connection is shown in Figure 6. Usually, as a form of quality checking, the surge test is performed before the coils are placed in the slots of the machine. The test can also be conducted on the windings of fully assembled machines. A surge test, from the IEEE standard 522 [59], is recommended to have 3.5 pu voltage impulse magnitude, with a rise time between 0.1 μs and 0.2 μs. The 3.5 pu voltage impulse magnitude is the peak line-to-ground voltage, calculated from the rated voltage. For example, for a rated line-to-line voltage of 2.2 kV, the impulse voltage magnitude applied can be calculated as 3.5 × (2.2/$\sqrt{3}$) × ($\sqrt{2}$) kV, resulting in 6.6 kV peak.

As discussed earlier, there has also been work to further improve the processing of data collected from surge testing with the use of particle swarm optimization to identify the equivalent resistance and capacitance components (RCs), as well as inductance and capacitance (LC) components of the stator winding, which can then be used to track and differentiate damage to the windings in the process shown in Figure 7 [52].

Rotor winding insulation systems are in most cases different from stator windings and are relatively sparsely covered in literature. Rotor insulation systems are typically subjected to either relatively low DC voltage in synchronous rotors or relatively low AC voltage for induction rotors, compared to the stator windings. As a result, their fault condition can be less severe than stator insulation conditions and considered less important. Since this review is mainly focused on stator insulation, the rotor is not covered here. However,

Table 2. Offline rotor winding tests (adapted from [53]).

Name	Description	Performance Difficulty	Effectiveness
Insulation Resistance (IR)	Apply DC voltage for 1 min to measure leakage current	Easy	Only finds contamination and defects
Capacitance	Apply low or high voltage to measure winding capacitance to ground	Moderate	Moderately effective to find thermal or water leak problems
Dissipation Factor (DF)	Apply low or high voltage to measure insulation loss	Moderate	Moderately effective to find thermal or water leak problems
Partial Discharge (PD)	Directly detect PD pulse voltages at rated voltage	Difficult	Finds most problems except end winding vibration; for form-wound only.
Surge Test	Find turn and ground faults by measuring discontinuities in surge impedance	Difficult	Effective if close to dead short circuit

shows a summary of rotor winding tests employing the offline tests, showing their level of difficulty and effectiveness.

4. Online SoH Tracking Methods

Due to the downtime and equipment requirements of offline SoH tracking methods, there is an increasing shift to online analysis methods that do not require downtime of the machine and use integrated sensors for data collection [19,20,61,62]. Online analysis methods have the additional benefit of being able to more continuously track the SoH of a machine instead of exclusively at designated maintenance intervals. Many of the more widely used online SoH tracking methods are adaptations of the traditional offline tests such as insulation resistance, capacitance, and DF tracking [63,64,65,66]. These methods are not new [67], but they have been recently growing in popularity due to advancements in sensors [68] and processing methods.

Many of the online methods involve analysis of the groundwall current that can be found in one of two ways: by tracking the current flow from the stator core to the system ground or by finding the differential current of the machine coil. Both topologies for a single phase are showcased in Figure 8. While either method can be used, the direct measurement of the leakage current might not be feasible for all machines, as the frame must first be isolated from the system ground for the probe to be installed. Additionally, for multiphase machines, this method would capture the sum of all insulation leakage currents. The use of a differential current probe, however, can be problematic due to the small signal magnitudes, meaning such probes can be susceptible to noise [68,69,70]. Another proposal is to use the high-resistance ground resistor present in some inverter topologies for testing, as this has the potential to be more accurate and less complex than the use of current sensors [71].

4.1. Online Tracking of Insulation Impedance

One of the primary online SoH tracking categories includes those methods that approximate the insulation groundwall impedance. This is likely due to the plethora of research that has been carried out on the online calculation of insulation capacitance [54,72,73]. Insulation capacitance has the benefit of being directly linked to the current state of the insulation and thus interpretation of the results is significantly more intuitive than some of the other methods. The overall change in capacitance has been proposed as a measure for remaining useful life (RUL); however, the percent change in capacitance can vary greatly for different applications [72,74,75].

In comparison to offline tests of the insulation equivalent capacitance, which are conducted at grid frequency, online testing has greater flexibility to investigate the capacitance at different frequencies. Tsyokhla proposed the use of a mean of several different frequencies, namely the frequency maxima [75]. Other works have proposed methods to account for phase angle deviations that occur in the measured signals depending on the location of damage within the coil [76]. Another approach proposed by Hu accounts for the spatial distribution of the coaxial lead wires and is able to limit the error of the online calculation method to less than 5% [77]. However, as shown in Figure 9, the deviation between the expected capacitance and the online detection method persists when under higher load conditions, highlighting the need for continued improvement of capacitance calculation methods.

In addition to calculating the capacitive component of the groundwall impedance, other methods of analysis have been proposed, such as using the common mode (CM) and differential mode (DM) impedance to find and differentiate changes in the groundwall and turn-to-turn insulation [78]. Others have suggested the switching tail components of the impedance specifically can be used for better analysis [79]. Other studies have also investigated the possibility of using other transfer functions such as admittance for insulation analysis [80].

4.2. Online Current Analysis Methods

While impedance-based analysis of the insulation SoH is popular, other methods have proposed signature analysis of the current waveforms to reduce the complexity of measurement systems. These methods often assume the insulation is an impedance between the windings and the groundwall, where impedance can be understood as a transfer function whose input and output are the leakage current from the coil to the stator and voltage difference between two nodes. For most inverter-fed machines, it can be assumed that the voltage difference of the insulation system will remain a constant magnitude, i.e., the DC bus voltage. Based on this assumption, any change in the leakage current can be understood as the result of changes to the insulation impedance.

In its simplest form, the insulation health can be extracted from the low-frequency phase and line currents. One such approach proposed by Rayes-Malanche uses fuzzy logic manipulations of these currents to find and compare deviations in the phase current amplitudes [81]. As shown below in Figure 10, this method can clearly delineate between healthy motors and a motor with a minor short circuit in one of the phases. However, more work must be carried out to investigate the efficacy of this method for machines under varying loads.

Several other methods have proposed analysis of the current oscillations in the time domain following switching and used for prediction of RUL [72,74,82,83,84]. These methods rely on a characterization of the properties of the underdamped oscillations, as shown in Figure 11, where the predominant measure of insulation health is the peak oscillation amplitude, A. In addition to this, the oscillation period, T_c, and the oscillation settle time, T_s, have been used to analyze insulation integrity and predict the location of deterioration [83]. While some methods propose the direct analysis of the waveform parameters, others have proposed the use of the parameters to extract the capacitive behavior of the system to perform a capacitance-based analysis [85].

The oscillation amplitude has been found to experience a short rise in magnitude before experiencing an exponential delay until coil failure [86]. The oscillation period and settling time are likewise expected to reduce in magnitude as the insulation degrades [82]. While all three parameters are understood to be useful metrics of insulation health, the amplitude is the most frequently used metric due to its ease of calculation and primary link to the insulation capacitance, although it is also impacted by the insulation resistance. It was shown through Jensen’s work with an accelerated aging test that the predicted RUL of the insulation could be found to within 20 h of the actual insulation life [74].

Other current signature methods rely on a frequency analysis of the high-frequency (HF) current components of either the coil current or the leakage current [84,87,88,89,90]. These methods have several advantages over time domain analysis; namely, time invariance can help reduce the impact of noise. Additionally, many of the frequency components of electric machines are well understood; thus, a frequency domain analysis can be used for interpretation of multiple different machine faults. Typically, the frequency domain behavior is found using a Fourier transform, but other frequency domain transforms, such as the Hilbert transform [91] and wavelet decomposition [92,93], have also been proposed.

Paper [93] proposed a method for online detection and classification of both inter-turn and groundwall faults using Wavelet Particle Decomposition (WPD) to extract features from high-frequency line current transients. Features specifically associated with antiresonance oscillations in the HF current were extracted, allowing them to focus on both insulation faults with no intrusive sensors or measurements in the setup. The signals were reconstructed form the specific wavelet packets (WP1 for GW and WP2 for TT) and were used to calculate the SoH indicators over time to quantify the severity of the insulation.

Another method is to analyze the change in insulation state by monitoring the deviation in the response over a frequency range of interest. As this method lacks any direct comparisons for SoH analysis, an insulation state indicator (ISI) has been proposed, which employs the root mean square deviation (RMSD) between the frequency response of the current to a set baseline [84,94,95,96]. This method can be used for either the coil input current or the leakage current, but when using the coil current, the mean derivative of the current (the red line in Figure 12) must be removed to isolate the HF transient component [88].

An alternative approach proposed by Patel analyzes the transient current component sung wavelet decomposition to investigate changes in the frequency response near the antiresonance frequency [93]. From this, different measures for both the turn-to-turn insulation and the groundwall insulation are extracted. Simulation results, shown in

Table 3. Simulation and experimental results of SoH tracking methods using wavelet decomposition of line current (adapted from [93]).

Degradation Type	Severity (%)	Simulated		Experimental
		GW SoH	TT SoH	GW SoH	TT SoH
Healthy	0	20.56	2.03	25.14	2.03
TT Damage	10	20.51	2.24	24.62	-
	20	20.46	2.64	25.20	2.89
	30	20.42	3.32	-	-
	40	20.37	4.33	-	-
GW Damage	10	23.57	2.25	26.69	-
	20	25.88	2.54	26.95	2.73
	30	27.27	2.90	-	-
	40	29.53	3.34	-	-
Combined Damage	10	23.52	2.61	26.70	-
	20	25.63	3.93	28.69	3.03
	30	27.06	5.98	-	-
	40	29.39	7.81	-	-

, illustrate good sensitivity to changes of the insulation state and initial experimental results are also promising.

It has also been proposed that the coil frequency response could be used to identify other faults such as those within the bearing [96]. In addition to HF analysis, very low frequencies have been proposed for use in fault diagnosis [97]. However, for this, and many other online SoH methods, there is no clear indication of what value is equated with a coil failure; thus, comparison to a failed coil may be needed, limiting their applicability.

4.3. Artificial Intelligence Approaches for SoH Tracking and Diagnostics

Another recent avenue for improvement for SoH tracking methods has been the development of AI models for interpreting SoH measurements [97]. The work carried out by Dayong Zheng in [98] developed an innovative online monitoring method to detect both interturn and groundwall insulation with an error of <6% in damage estimations. Their approach utilized a physics-based broadband impedance model with Principal Component Analysis (PCA) and a decision tree regression machine learning model. The physics-based model generates synthetic data while the PCA extracts and reduces the overall features from the dataset. This filters out noise from the impedance data. The decision tree regression model is trained using the reduced features, and this model learns how combinations of the principal components correspond to different levels of insulation degradation. Its outputs are the estimated degrees of interturn and groundwall aging (ΔCt and ΔCg). Particle swarm optimization was used to ensure that the model closely matches real machine measurements. The low error makes this approach suitable for condition monitoring and preventive maintenance.

Other models using Continuous Wavelet Transforms of the current spectrum have achieved similar results [99]. This paper sets itself apart by proposing a hybrid physics-based and data-driven approach for diagnosing stator insulation. The authors utilized Continuous Wavelet Transform (CWT) to extract features from the high-frequency common-mode data generated from the physics model. These features are used in a proposed convolutional neural network, which is optimized with small kernel sizes and adaptive learning rates to classify groundwall and turn insulation faults. Figure 13 shows the architecture of the proposed CNN in [99]. The model is experimentally validated through a 3 kW permanent magnet synchronous machine with an accuracy of about 94%. This is promising, as it has the potential to improve predictive maintenance and reduce downtime costs. The common mode impedance characteristics of the machine used is characterized in Equation (9) as follows:

$$Z_{cm}={\displaystyle \frac{L\left(C_p+C_{g2}\right)s^2+\frac{L}{R_p}s+1}{\left(C_{g1}+C_{g2}\right)s[L\left(C_p+\frac{C_{g1}{C}_{g2}}{C_{g1}{+C}_{g2}}\right)s^2+\frac{L}{R_p}s+1]}}$$

(9)

Different methods, including Zhang’s use of AI in interpreting impedance frequency response data and Gopinath’s proposed combination of both PD and Tanδ testing with analysis using artificial neural networks, have further improved accuracy [100,101]. In paper [100], a proposed machine learning technique leveraging both an artificial neural network (ANN) and gravitational search algorithm (GSA) was used to assess the insulation failure of HV rotating machines. Features used for training included the leakage current of the machine, capacitance, PD factor, and DF. Results indicated a close relation to the experimental values, with a low percentage error of 1.4% for the ANN-GSA algorithm. The GSA, which is a relatively new metaheuristic algorithm, is based on the law of gravity and mass and has search agents from optimization methods from swarm intelligence. This is used in the optimal selection of hyperparameters in the ANN, improving its performance.

Table 4. Comparison of performance percentage error (adapted from [100]).

Description	ANN	ANN-GA	ANN-GSA
$\tan \delta$	0.173	0.067	0.014

shows performance percentage errors from this proposed work as compared to other ANN algorithms.

With limited training for different failure cases, the accuracy can be limited, but when sufficient training data is provided, the percent error can be as low as 1.4% [100,101]. Another method proposed by Jaen-Cuellar combines the analysis of stator currents with the flux and mechanical behavior, methods also frequently used on their own [102,103], with AI interpretation to identify specific faults in the windings [104]. The proposed method analyzes the three combined SoH measurements to sort the system into a 2D plane, and depending on the space within the plane, the overall state of the insulation can be predicted, as shown in Figure 14.

Yu Zhang’s paper [101] tries to detect faults in hairpin stator windings using impedance data and training a Support Vector Machine (SVM) based on features from the impedance data as summarized in

Table 5. Evaluation results on training SVM based on impedance data (adapted from [101]).

Experiment	Cross Validation	Training	Testing	Test Accuracy
Experiment 1	Cross Validation 1	9 Healthy + 3 Faulty (Faulty: Stator 6)	6 Healthy + 6 Faulty (Faulty: Stator 7 and 8)	95%
	Cross Validation 2	9 Healthy + 3 Faulty (Faulty: Stator 7)	6 Healthy + 6 Faulty (Faulty: Stator 7 and 8)	100%
	Cross Validation 3	9 Healthy + 3 Faulty (Faulty: Stator 8)	6 Healthy + 6 Faulty (Faulty: Stator 6 and 7)	89%
			Overall Experiment accuracy	94.7%
Experiment 2	Cross Validation 1	9 Healthy + 3 Faulty (Faulty: Stator 6)	6 Healthy + 6 Faulty (Faulty: Stator 7 and 8)	100%
	Cross Validation 2	9 Healthy + 3 Faulty (Faulty: Stator 7)	6 Healthy + 6 Faulty (Faulty: Stator 7 and 8)	100%
	Cross Validation 3	9 Healthy + 3 Faulty (Faulty: Stator 8)	6 Healthy + 6 Faulty (Faulty: Stator 6 and 7)	100%
			Overall Experiment accuracy	100%

. The SVM used in this study was C-Support Vector Classification. The impedance data consists of high-frequency data, between 20 kHz and 1 MHz, and low-frequency data, between 1 Hz and 20 kHz. This is taken from 10 stator windings: 5 healthy windings, 3 epoxy faulty windings, and 2 welding faulty windings. The faulty impedance data improves the effectiveness of the SVM classifier. Yu Zhang’s proposed machine learning technique proves to be accurate and cost effective.

From these reviews, we can tell that the primary limitation of using AI/machine learning models for online insulation SoH tracking is their reliance on large datasets to achieve high accuracy. Collecting real-time data is time-consuming and expensive, which is why physics-based models such as the one in [98] are often developed to simulate different scenarios and generate synthetic data that mimics real-time data. Additionally, most commonly used datasets are typically compiled from many different small-scale insulation data. This alone highlights the challenge of obtaining high-quality data for model training.

5. Comparative Evaluation of SoH Methods

Table 6. Summary of offline and online SoH tracking.

Method	Sensitivity	Downtime	Accuracy	Key Limitations	References
Partial Discharge (PD)	High	Low	±15%	Susceptible to noise; might not detect early degradation	[32,33,34,35,36]
Insulation Resistance (IR)	Low	Medium	±25%	Strongly dependent on temperature, humidity, and contamination	[48,56,57,58]
Surge Testing	Moderate	High	±20%	Limited turn-to-turn insulation diagnosis in early stages	[18,53,59,60]
Tan-Delta/Dissipation Factor (DF)	Moderate	Medium	±15%	Requires stable test voltage; surface leakage interference	[49,53,55]
Insulation Capacitance	Moderate	Low	±10%	Temperature-dependent; not reliable for early degradation	[24,53,54]
Impedance-Based Analysis	High	None	±3–6%	Requires accurate modeling	[75,77,78]
Current-Based Analysis	Moderate	None	±5–10%	Requires clean signals; may miss localized faults	[72,74,82,83,84]
Insulation State Indicator (ISI)	High	None	±8%	Heavily model-dependent; requires historic baseline	[84,94,95,96]
Wavelet Particle Decomposition (WPD)	High	None	±6%	Needs high-resolution signal; computationally intensive	[61,92,93]
AI-Based Analysis	Very High	None	±1–6%	Requires large training dataset; performance drops with unseen failure modes; interpretability issues	[46,97,98,99,100,104]

is a summary of both traditional and emerging insulation SoH tracking methods, outlining their diagnostic limitations. Offline methods such as insulation resistance (IR) and surge testing are widely adopted in industry, but as highlighted in Section 3.3 and Section 3.4 of this review, they struggle to detect early-stage degradation and are highly influenced by environmental factors such as temperature and humidity. Their error margins reflect these limitations. For instance, IR measurements are impacted by moisture and thermal fluctuations, and surge tests, while capable of detecting turn-to-turn faults, may risk insulation damage due to their high-voltage application. Additionally, the IR method measures various components like capacitive, conduction, leakage, and absorption currents, but it only becomes meaningful after the capacitive component has decayed, which is typically after 1 min [53]. More offline methods such as dissipation factor (DF) and capacitance testing provide insight into the dielectric strength of the insulation (Section 3.1 and Section 3.2). When properly conducted, these methods are capable of capturing trends in thermal degradation and moisture ingress. DF, in particular, is a reliable indicator when used with stable test conditions and is enhanced when combined with capacitance mapping strategies. However, their medium sensitivity and the need for test interruption limit their usefulness for continuous condition monitoring.

Online SoH methods, by contrast, avoid the downtime and limitations of offline approaches. Impedance-based analysis and current-based signature techniques (Section 4.1 and Section 4.2) use high-resolution data to estimate insulation degradation based on impedance or current deviations under inverter excitation. These approaches have demonstrated greater sensitivity to minor insulation defects, especially under inverter-induced stress conditions typical for WBG devices. For example, the oscillation amplitude and damping behavior of high-frequency switching currents have been proposed as proxies for SoH and are particularly useful in tracking remaining useful life (RUL) under real-time conditions. Wavelet Particle Decomposition (WPD), as reviewed in Section 4.2 and further expanded in [93], isolates transient features from high-frequency leakage or line currents, enabling both groundwall and turn insulation faults to be detected without intrusive sensors. This method maintains a strong balance between accuracy and sensitivity, though it requires high computational power and noise filtering to be effective.

Finally, AI-based SoH techniques show significant potential (Section 4.3). Methods using convolutional neural networks (CNNs), decision tree regression, and hybrid physics-informed models offer exceptional sensitivity and precision as low as 1%. These approaches excel at capturing complex, nonlinear degradation patterns when trained on large datasets. However, as noted, their performance is often limited by dataset availability and interpretability challenges. Despite these challenges, models such as those proposed by Zheng [98] and Zhang [101] have demonstrated accuracies upwards of 94–100% in experimental validation, making them strong candidates for future standardization.

6. Conclusions

Predicting stator insulation failure has and continues to be a topic of great concern due to the high costs of unexpected downtime that can occur when there is a serious fault within the insulation. Historically, there have been numerous approaches for testing the current health of stator insulation that have been standardized by regulatory bodies and other organizations such as IEEE. With the widespread adoption of semiconductor device-based motor drives, these methods were adapted and improved to better reflect the increased stress placed on the insulation. However, the recent development and adoption of WBG devices has once again necessitated improved insulation SoH tracking methods to compensate for the further increased stress placed on the stators.

This paper presented a review of the state-of-the-art for insulation health tracking, starting with the more traditionally established methods such as PD testing and other offline tests. Each of these methods and their corresponding standards were discussed along with their current limitations and recent developments in further improving their usefulness. One common avenue proposed across these methods is the use of AI models to interpret the test results for improved data analysis and noise elimination.

Following this, new online methods for insulation SoH tracking were reviewed, including capacitance tracking and current waveform and spectrum analysis methods. These approaches are appealing because they enable continuous monitoring, reducing unforeseen failure and not requiring scheduled downtime for testing. Numerous variations have been proposed that show promising results in identifying damage at its incipient stages. However, due to their novelty, there are limited tested applications outside laboratory conditions and minimal consensus on standards for the use of such methods. While many of these methods can identify changes to the state of the insulation, further work must be performed to establish damage threshold metrics and enhance their use for predicting the insulation’s RUL. The broader use of AI platforms within the field of insulation SoH monitoring was also discussed, followed by a comparison of both offline and online tracking techniques in terms of sensitivity, downtime, accuracy, and implementation limitations.

Considering the urgent need for accurate prediction of stator insulation health for the continued operation of many industrial processes, the increased stress created by WBG devices, and the numerous benefits proposed by online tracking methods and AI data analysis methodologies discussed in this paper, an increased need for SoH tracking and RUL prediction method development, testing, and standardization is foreseen in academic, industrial, and regulatory laboratories for the continued optimization and use of electric machines necessary for our modern world.