Optimised Neural Network Model for Wind Turbine DFIG Converter Fault Diagnosis

Abstract

This research introduces an enhanced fault detection approach, variational mode decomposition (VMD), for identifying open-circuit IGBT faults in the grid-side converter (GSC) of a doubly fed induction generator (DFIG) wind turbine system. VMD has many advantages over other decomposition methods, notably for non-stationary signals and noise. VMD’s robustness stems from its ability to decompose a signal into intrinsic mode functions (IMFs) with well-defined centre frequencies and bandwidths. The proposed methodology integrates VMD with a hybrid convolutional neural network–long short-term memory (CNN-LSTM) architecture to efficiently extract and learn distinctive temporal and spectral properties from three-phase current sources. Ten operational scenarios with a wind speed range of 5–16 m/s were simulated using a comprehensive MATLAB/Simulink version R2022b model, including one healthy condition and nine unique IGBT failure conditions. The obtained current signals were decomposed via VMD to extract essential frequency components, which were normalised and utilised as input sequences for deep learning models. A comparative comparison of CNN-LSTM and CNN-only classifiers revealed that the CNN-LSTM model attained the greatest classification accuracy of 88.00%, exhibiting enhanced precision and resilience in noisy and dynamic environments. These findings emphasise the efficiency of integrating advanced signal decomposition with deep sequential learning for real-time, high-precision fault identification in wind turbine power electronic converters.

1. Introduction

Technology underwent significant improvement throughout the Industrial Revolution over the past decades, which emphasised the creation and expansion of renewable energy [1]. The improved technical abilities of machines have replaced laborious manual activities, promoting human advancement [2]. The acknowledgement of improved processing powers and progress in big data has revolutionised the use of artificial intelligence (AI) technologies and impacted Industry 4.0 [3], including intelligent systems like renewable energy systems (RES) [4].

The International Energy Agency (IEA) forecasts that renewable energy sources (RESs) will produce over 50% of worldwide electricity by 2025, surpassing coal as the primary source. Moreover, the renewable energy capacity is projected to reach 10,800 GW by 2040 [5,6]. Wind energy, as a form of renewable energy, serves as a principal method for reducing carbon emissions in the energy sector, with its collective contribution to electrical energy production projected to increase from 10% in 2021 to 40% by 2030 and 70% by 2050, as illustrated in Figure 1 [6]. Wind energy encounters a challenge due to the erratic nature of wind. Hence, the wind turbines do not consistently function at peak capacity [7]. By integrating technical, ecological, and private interests while delivering advantages for all, AI can achieve what is often considered unfeasible in wind-turbine-based renewable energy [8].

Grid-integrated doubly fed induction generators (DFIGs) driven by wind turbines are progressively dominating the renewable energy sector due to their enhanced durability, reduced maintenance expenses, and streamlined control compared to traditional generators [9,10]. The usage of power electronic converters has grown widespread in contemporary technologies [11]. As shown in Figure 2 [12], the DFIG configuration has two power electronic converters: the rotor side converter (RSC) and the grid side converter (GSC), linked by a shared DC-link [12,13,14,15,16]. A 2009 industry review of the reliability of power electronics converters indicates that the power converter is one of the least reliable components in a renewable energy system operating under extreme environmental conditions [11,17].

According to studies [18,19], power electronic frequency converters account for approximately 13% of total failures and are responsible for 18.4% of the downtime across various types of monitored wind turbines. An examination in [11] confirms that semiconductor devices and capacitors are the most sensitive components of a power converter [18]. Wind speed fluctuations can produce substantial variations in the SMPS’s input or output currents, triggering serious electromagnetic interference (EMI) problems and necessitating complicated and expensive filters [20]. High switching frequencies, dual (bipolar) power supply needs, and complex feedback circuitry make switched-mode power amplifier design difficult [21].

Power converters break down like other systems. Such failures require costly maintenance, especially if the power electronic converter equipment is remote or offshore [22]. For this purpose, semiconductor switching ripples and packaging, notably bond wire degradation and die-attach solder fatigue, are the primary failure modes. The semiconductor element’s average junction temperatures and fluctuating cycles cause the failure modes [23]. Several studies on these topics have lately been published, among others [24,25]. The pattern balances the system’s availability and size. Still, current research does not establish relationships between these failure modes, and measurement systems do not incorporate them into the design process. It is more appropriate to characterise the methodologies employed in these articles as availability evaluation approaches. In contrast, optimal power converter system design tasks [26,27,28] do not include system availability.

This work presents a hybrid diagnostic model utilising CNN and LSTM for the detection and classification of IGBT open-circuit faults in the grid side converter of a DFIG wind turbine. The model utilises CNN layers to autonomously extract distinctive spatial features from three-phase current waveforms, followed by LSTM layers that represent the temporal progression of these features for precise classification. The suggested strategy, in contrast to conventional signal processing approaches, eliminates the need for manual feature extraction or threshold adjustment, hence facilitating real-time implementation and scalable deployment.

The authors trained and validated the model using synthetic data generated from a detailed MATLAB/Simulink model of a DFIG wind energy system, simulating ten different operational states, one representing healthy operation and nine representing distinct open-circuit fault scenarios. Simulation results in Section 5.3, show that the CNN-LSTM model outperforms standalone CNN, LSTM, and multi-layer perceptron (MLP) models in terms of accuracy, robustness, and diagnostic granularity.

These findings demonstrate the viability of employing deep learning diagnostics for predictive maintenance and fault tolerance in contemporary wind turbine systems, facilitating the development of more intelligent and resilient renewable energy infrastructure.

1.1. Research Gap and Motivation

The increasing integration of variable renewable energy (VRE) sources, especially wind energy, into power grids has presented novel technical issues in maintaining the stability, reliability, and fault tolerance of power electronic converters. DFIG-based wind turbines are among the most extensively utilised topologies in grid-connected wind energy systems, owing to their cost-effectiveness and controllability. The dynamic and stochastic characteristics of wind create considerable fluctuation in system performance, especially during fault conditions and varying wind speeds.

While numerous fault detection strategies for DFIG-based systems, including model-based estimation, signal processing methods, and traditional machine learning, demonstrate efficiency in controlled environments, they frequently exhibit insufficient flexibility to non-stationary operational conditions. These encompass abrupt wind fluctuations and thermal stress, which are critical for practical wind farm applications. Moreover, current methodologies generally lack model transparency, contextual interpretability, and good diagnostic accuracy across varying operating settings [29].

Previous research conducted by the authors [30,31,32] has investigated the use of AI to enhance data precision, system transparency, and grid-integrated control methodologies in DFIG systems. Nonetheless, a significant study gap persists: the inadequate capacity of conventional and certain AI-based approaches to effectively manage the fault categorisation of power converter systems amid variable wind and heat conditions. Recent research on artificial neural networks (ANNs) indicates potential improvements when accounting for system harmonics, junction temperature effects, and power electronics switching; however, their implementation under variable wind conditions has not been thoroughly confirmed.

Thus, there is a pressing need for an AI-driven, lifecycle-aware diagnostic and optimisation framework that enhances real-time adaptation and prediction under varying wind conditions while also improving the dependability, transparency, and performance of DFIG-based wind energy systems.

This study will address this gap by incorporating wind speed fluctuations into the ANN training process and improving feature extraction through deep learning-inspired fusion methods. Among these, deep learning architectures such as convolutional neural networks (CNNs) and long short-term memory (LSTM) networks have shown great promise in condition monitoring applications. CNNs are particularly effective in capturing local patterns and distortions in electrical signals through spatial convolutional filters. LSTMs, on the other hand, are designed to capture long-term temporal dependencies and sequential characteristics of time-series data, making them suitable for modelling electrical waveforms over time.

1.2. Research Contribution and Future Work

The authors summarise the paper’s contribution in the following statements:

This paper primarily includes model-based methods, signal processing approaches, and machine learning models, such as pattern recognition and classification algorithms, to enhance the efficiency and sustainability of power converter components and configurations. The contributions of this paper are threefold:

• Development of a high-fidelity Simulink model for fault injection in GSCs of DFIG wind turbines.
• Designing a CNN-LSTM-based time-series classification model tailored for multivariate current signals.
• Performance benchmarking against other neural architectures with accuracy and confusion matrix evaluations.

These results establish the feasibility of using deep learning-based diagnostics for predictive maintenance and fault tolerance in modern wind turbine systems, paving the way for more intelligent, more resilient renewable energy infrastructure.

The authors presented the manuscript in distinct sections. Section 1 encompasses the introduction, explaining the gap analysis, challenges, and the authors’ contributions to the proposed study. Section 2 includes a comprehensive review analysis. Section 3 delineates the modelling of wind-driven turbines and the detailed description of DFIG. Section 4 provides details regarding the suggested methodology under investigation. Section 5 includes details of the comparison model simulations and their analysis. Section 6 contains the conclusion and recommendations for future research.

2. Article Reviews

The global shift toward sustainable energy has significantly accelerated the deployment of wind energy systems. Wind energy has emerged as a sustainable solution to meet global electricity demands. Among the various technologies used in wind power generation, the doubly fed induction generator (DFIG)-based wind turbine has become a dominant configuration due to its ability to operate over a wide range of wind speeds, its cost-effective partial-scale converter, and its capability to provide reactive power support to the grid.

The researchers in [12] established that the DFIG topology includes two power electronic converters, the rotor side converter (RSC) and the grid side converter (GSC), connected through a common DC-link. While the RSC controls the active and reactive power between the rotor and grid, the GSC is responsible for maintaining the stability of the DC-link voltage and ensuring compliance with grid codes by regulating power quality at the point of common coupling.

The research review [29] demonstrated that the GSC’s exposure to persistent switching (voltage ramping rate > 1 V/s) [33], thermal stress (e.g., junction temperature >150 °C) [11], and voltage transients using the feedback of the stator current reference (FSCR) strategy [34] render it very vulnerable to faults, especially open-circuit failures in insulated gate bipolar transistor (IGBT) devices. These errors frequently go unnoticed in initial stages, yet can result in waveform distortion, diminished power quality, and irreversible damage if not swiftly rectified. Further, the researchers established that the conventional fault detection techniques depend on established criteria, which may exhibit suboptimal performance in fluctuating wind conditions. This research seeks to utilise an advanced artificial neural network model’s pattern recognition ability to create an intelligent fault detection system that dynamically adjusts to wind fluctuations and improves classification performance using feature fusion approaches.

Traditional defect detection techniques, such as signal thresholding with fixed limits [35,36], predominantly depend on hardware redundancy, thresholding, and rule-based logic [37]. These methods often exhibit limited adaptability, a lack of pattern recognition, and high sensitivity to operational conditions, frequently triggering false alarms during transient states or fluctuating wind speeds. The proposed CNN-LSTM hybrid model addresses these limitations by combining convolutional layers for spatial feature extraction with LSTM units for modelling long-term temporal dependencies in time-series data, as demonstrated in recent deep learning applications for fault diagnosis [38]. Unlike fixed-threshold methods, CNN-LSTM architectures can identify subtle, non-linear variations in three-phase current signals across variable wind conditions and diverse fault scenarios, enabling real-time, accurate, and robust fault classification [39].

Furthermore, the non-linear and dynamic behaviour of wind energy systems significantly complicates fault identification using conventional fixed-rule systems [40]. Recent studies have progressively investigated data-driven diagnostic methodologies utilising machine learning (ML) and deep learning (DL) models capable of discerning intricate fault patterns from raw signal data [41,42].

The study [43] examines stator flux dynamics during voltage dips, representing transient rotor-induced electromotive force as the primary driver of rotor overcurrent. A stator current-based rotor voltage compensation technique is proposed, along with a low-voltage ride-through (LVRT) protection scheme that incorporates crowbar and chopper circuits to mitigate converter limits during significant voltage drops.

The research work [44] aims to enhance the reliability of DFIG-based wind energy conversion systems (DFIG-WECSs) by introducing an improved fault ride-through (FRT) method. The study presents a dual-mode operation in which the system functions in wind speed estimation-based maximum power point tracking (WSE-MPPT) mode under normal conditions to optimise power extraction. It transitions to an enhanced crowbar-based rotor active impedance fault ride-through (FRT) mode during fault conditions. It mitigates post-fault inrush currents and transient oscillations with an active impedance-based crowbar management method while also guaranteeing optimal power extraction during regular operation without necessitating supplementary expensive sensors.

In [44], the authors assessed the performance of the system under three-phase symmetrical faults with significant voltage dips, illustrating how the suggested technique enhances system stability. A comparative study with conventional methods showed that the suggested strategy surpasses typical crowbar protection strategies by mitigating rotor and stator current surges and enhancing transient recovery.

The research [43] examines methods to enhance the low-voltage ride-through (LVRT) capabilities of doubly fed induction generator (DFIG)-based wind turbines under significant grid voltage disturbances. The research examines stator flux behaviour during voltage dips and determines that transient rotor-induced EMF is the primary factor contributing to rotor overcurrent.

The authors suggest an enhanced control technique utilising stator current-based rotor voltage compensation to mitigate overcurrent and improve LVRT performance. Furthermore, they present a hardware protection mechanism that integrates crowbar and chopper circuits to regulate high transient currents. The proposed strategy is validated via MATLAB/Simulink simulations, showcasing its capacity to uphold grid stability and diminish converter stress.

A key contribution of [43] is the dynamic study of stator flux, which identifies transient rotor-induced EMF as a primary factor in rotor overcurrent during voltage dips. Additionally, rotor voltage adjustment based on stator current is achieved by stator current feedback to regulate rotor voltage, hence enhancing system responsiveness to failures. In [43], the crowbar and chopper protection scheme mitigates high rotor currents and stabilises the DC bus voltage. MATLAB/Simulink validation illustrates the efficacy of the proposed control techniques in mitigating overcurrent, voltage variations, and torque oscillations.

The researchers in [45] investigated an innovative control method to improve the low-voltage ride-through (LVRT) capacity of wind turbines utilising doubly fed induction generators (DFIG). The primary hurdles in DFIG-based systems encompass overshooting, stability concerns, and increased response time under fault conditions. Conventional hysteresis-based fault detection techniques frequently result in erroneous activation of protective measures, such as crowbar deployment. This paper presents fuzzy logic control (FLC) for fault detection, thereby improving the precision of voltage-dip detection and facilitating a more rapid reaction to grid failures.

The work proposed the Salp swarm optimisation algorithm (SSOA) for controller tuning to optimise the gains of the proportional–integral (PI) controller and the values of DC-link capacitance, thereby enhancing dynamic response and transient stability. The fault ride-through (FRT) performance improved, with active power overshoot decreasing from 10.12 × 10⁶ to 3.78 × 10⁶ and reactive power overshoot decreasing from 15.01 × 10⁶ to 6.10 × 10⁶. In [45], MATLAB/Simulink simulations exhibited enhanced power quality, decreased oscillations, and expedited transient recovery relative to conventional PSO-based optimisation methods.

The study [46] suggested that indirect vector control (IVC) incorporate a super-twisting controller to rectify instantaneous power discrepancies in DFIGs by directly calculating the necessary rotor control voltage, hence improving transient performance. A multilevel fuzzy-modified space vector modulation guarantees a consistent switching frequency, hence facilitating AC filter design. This approach lowered harmonic distortion by 18.02% and 16.22% relative to traditional IVC.

The researchers in [47] introduced a sophisticated approach for fault identification and diagnostics in DFIG-based wind energy systems. The methodology employs an independent component analysis-based correlation coefficient (ICA-CC) for fault detection and an enhanced multihead cross attention with BiLSTM network (EMCABN) for fault categorisation. Furthermore, it utilises a multi-strategy enhanced orchard algorithm (MSEOA) to optimise fault control by managing active/reactive power changes, rotor current harmonics, and oscillations in DC-link voltage. The proposed CNN-LSTM model attained an impressive 88% accuracy in defect identification, proficiently addressing 10 distinct IGBT fault states within a comprehensive and varied dataset. Conversely, the methodology detailed in [47] achieved an accuracy of 98% using MATLAB simulations; however, this outcome was predicated on a more restricted situation, encompassing only two-stage failures and a limited dataset. Although [47] exhibits good performance in a controlled environment, the CNN-LSTM model’s exceptional accuracy under more intricate and varied fault circumstances underscores its better generality and practical applicability for real-world wind turbine scenarios.

The work derives fault-associated characteristics from wind turbine operational data. It identifies irregularities in wind velocity, generator temperature, and voltage variations. Fault classification is performed via EMCABN (BiLSTM combined with attention mechanism) and BiLSTM with multihead cross-attention to categorise defects. It has ensured exceptional accuracy in differentiating among electrical, mechanical, and grid-related issues. It has enhanced system performance by mitigating power oscillations and voltage disturbances. It has executed fault mitigation measures, including pitch angle correction and converter protection. The researcher validated MATLAB/Simulink in [47] and compared EMCABN-MSEOA with typical neural networks, resulting in an impressive accuracy of 98%. It exhibited proficient fault identification, classification, and mitigation in DFIG-based wind turbines.

The research [48] introduced an innovative method for improving fault detection and identification (FDI) in DFIG-based wind turbines. This technology combines neural network predictive control (NNPC) with superconducting magnetic energy storage (SMES) to enhance low-voltage ride-through (LVRT) performance during grid disturbances. The work utilised ANNs with model predictive control (MPC) to forecast wind speed variations and enhance DFIG regulation. This model facilitated an improved real-time fault identification and classification, hence reducing response time to grid problems. Superconducting magnetic energy storage (SMES) is employed to stabilise the DC-link voltage and alleviate transient oscillations in power output. The ANN-based model in [48] detects and categorises stator and rotor problems by analysing voltage amplitude, phase changes, and frequency variations. The suggested technique enhances precision in problem categorisation while minimising false alarms.

The authors in [49] introduced a fault-tolerant control (FTC) technique for doubly fed induction generator (DFIG)-based wind turbines. The research aimed to improve system reliability and sustainability through the application of a state-feedback controller in conjunction with proportional–integral (PI) and sliding-mode controllers (SMC) to regulate active and reactive power transfer between the stator and the grid. This research presents a proportional–integral observer (PIO) for fault estimation and detection, which juxtaposes fault estimation signals against established thresholds. In [49], the authors executed the fault-tolerant control (FTC) by state-feedback control to manage stator and rotor currents, hence assuring stable regulation of active and reactive power. It evaluated the performance of PI, SMC, and state-feedback controllers and accomplished fault detection by juxtaposing estimated fault signals against threshold levels. The efficiency of fault-tolerant control (FTC) in both fault-free and faulty settings was assessed using MATLAB/Simulink, revealing that the state-feedback controller with FTC surpasses proportional–integral (PI) and sliding mode control (SMC) controllers, especially in the presence of non-constant actuator faults. The suggested method improves the stability of wind energy systems by addressing actuator faults before their effect on power regulation, ensuring the DFIG-based system stays functional under diverse fault conditions.

The research [35] presents a fault detection and isolation (FDI) technique for current sensors in wind turbine converters. It specifically addresses deficiencies in the DFIG system, which is essential for wind energy conversion. The study utilised an extended Kalman filter (EKF) organised within a generalised observer scheme (GOS) to identify defects in the rotor-side converter (RSC) and the grid-side converter (GSC). The methodology in [35] utilises the extended Kalman filter (EKF) to estimate current values and juxtaposes them with real sensor readings to identify anomalies. It transforms the defects from the αβ reference frame to the abc frame for accurate faulty localisation with the Clarke transformation. The methodology employs statistical tests (Page–Hinkley method) to distinguish between genuine defects and measurement noise, averting false alarms. The approach is evaluated in an experimental configuration, showcasing its efficacy in identifying and isolating singular and multiple sensor problems.

The research [50] investigated the enhancement of transient stability in DFIG-based wind turbines with deep learning algorithms. It presents a passive fault current limiter (FCL) using an optimised resistance calculation technique to stabilise voltage and current under fault conditions. The study highlights the significance of deep learning models in predicting optimal resistance values for the FCL, thereby maintaining grid stability during both symmetrical and asymmetrical faults.

Transient stability in DFIG wind turbines has improved by introducing a resistance-based fault current limiter (FCL), which mitigates transient oscillations without active controllers while maintaining voltage stability within ±10% of the reference level during failures. It employs recurrent neural networks (RNNs) and long short-term memory (LSTM) networks for fault classification and fault current limiter (FCL) optimisation, utilising real-time fault data to adjust resistance dynamically.

The work assessed the system’s performance under two-phase-to-ground, phase-to-phase, and single-phase-to-ground faults using MATLAB/Simulink version R2022b software. It illustrates that deep-learning-optimised fault current limiters (FCLs) surpass traditional fault ride-through (FRT) systems. A comparative performance analysis was conducted utilising benchmarks of deep-learning-controlled FCL in contrast to conventional crowbar-based protection and series dynamic braking resistors (SDBR). The findings include enhanced voltage regulation, reduced current surges, and superior transient response.

The research [51] presented a data-driven methodology for the detection of problems in power converters within wind turbines. This approach employs ensemble empirical mode decomposition (EEMD) for fault detection, enhancing fault feature extraction and improving classification accuracy through intrinsic mode function (IMF) analysis and permutation entropy (PE) computation. The study [51] analysed three-phase voltage data from the power converter to obtain intrinsic mode functions (IMFs). It employed PE analysis to quantify signal complexity and improve fault discrimination. The suggested technique was evaluated on 22 distinct converter fault scenarios, with an accuracy of 98.3%, hence exhibiting significant resistance to noise and fluctuating operating circumstances. It enhanced fault diagnosis under non-linear and noisy conditions. The PE-based classifier enhanced dependability under varying wind speeds. The researchers compared traditional approaches, FFT, wavelet transformations, and conventional neural network classifiers, which were surpassed. The cited research presents a reliable and computationally efficient diagnostic method for wind turbine converters.

The work [52] introduced a comprehensive fault diagnosis method for DFIG-based wind turbines, proficient in identifying numerous IGBT open-circuit defects solely through rotor current measurements and references. This method avoids the necessity for supplementary sensors. The proposed approach in [52] incorporates chi-square statistics to analyse rotor current deviations. It facilitated fault detection under various operating situations and delivered precise diagnostics without the necessity for fixed thresholds. It can identify and isolate 21 distinct combinations of single and double IGBT open-circuit faults in both sub-synchronous and super-synchronous modes. The distinctions between IGBT failures and rotor current sensor errors within a single fundamental period are established. Experimental validation was conducted on a 15 kW laboratory apparatus, confirming its efficacy in both operational modes.

The authors in [53] introduced a sophisticated failure diagnosis method for wind turbines (WTs) utilising a deep learning model known as Time2Vec-long short-term memory (T2V-LSTM). The research utilised supervisory control and data acquisition (SCADA) data to forecast failures 10 to 210 min before their occurrence, reducing downtime and enhancing maintenance efficiency. Reference [53] utilised SCADA data for predictive maintenance and derived 33 critical parameters from it, including wind speed, rotor speed, electrical power, and temperature. It forecasts eight distinct categories of faults utilising deep learning methodologies. The T2V-LSTM model was developed, incorporating Time2Vec (T2V) encoding to improve the representation of temporal characteristics. Study [53] attained a fault prediction accuracy of 84.97%, surpassing traditional LSTM models. A comparison with traditional machine learning models was conducted. This work [53] yielded elevated recall scores, minimising false negatives and enhancing early defect discovery. Based on MATLAB/Simulink, validation was completed, and test performances were conducted on various wind turbine components, including pitch systems, gearboxes, power conditioning systems (PCS), and hydraulic units. It guarantees that the model precisely identifies defects across various operational contexts.

CNN-LSTM is frequently regarded as superior to T2V-LSTM scenarios, such as fault detection in wind energy systems or other time-series prediction and classification tasks [54]. This is primarily due to its ability to integrate spatial and temporal features through the combination of convolutional and recurrent layers [55], as well as its enhanced capacity to generalise across non-stationary and noisy data.

The research scholars in [56] investigated an innovative artificial neural network predictive control (NNPC) approach to improve the reliability of grid-connected DFIG wind turbines under three-phase fault scenarios. The study examined factors such as torque and power ripples, DC bus variations, and voltage distortions that impair power quality in wind energy systems.

Ref. [56] employed neural networks to forecast system behaviour and optimise control operations for failure mitigation while also minimising torque and power oscillations during fault circumstances. It improves grid stability and power quality while also diminishing reactive power disturbances and averting excessive power swings during faults. The validation of NNPC using MATLAB/Simulink and its comparison with conventional PI and model predictive control (MPC) was conducted. It exhibited exceptional efficacy in diminishing torque ripples and improving transient response.

Ref. [56] presented an active fault-tolerant control (AFTC) system for wind turbines, which combines backstepping active disturbance rejection control (BADRC) with an adaptive neuro-fuzzy inference system (ANFIS) for fault identification and isolation. The research examined generator actuator malfunctions, guaranteeing that the system sustains ideal performance under fault scenarios.

It utilised BADRC to counteract system disturbances and sustain power extraction despite generator actuator defects, hence improving system resilience and efficiency by alleviating the impacts of actuator failures. It employed ANFIS for fault detection, integrating fuzzy logic and artificial neural networks to categorise issues. Reference [56] may obtain 95.10% accuracy in identifying actuator failures, with a recall rate of 100% (no false negatives). This study employed PCA to extract pertinent features for the neuro-fuzzy fault detection model, which has improved fault classification by diminishing noise and enhancing computing efficiency. The study validated the MATLAB simulations and assessed the performance of the AFTC under normal, defective, and fault-compensated conditions. The proposed modelling of BADRC and ANFIS significantly enhanced power extraction, voltage stability, and transient response.

The paper [57] offers an extensive analysis of fault diagnostics and fault-tolerant control (FTC) methodologies for wind turbines. It examined diverse model-based, data-driven, and signal-based fault detection methodologies, emphasising both passive and active fault-tolerant control strategies. The research underscores the necessity of early detection to reduce maintenance expenses and enhance system reliability, especially in offshore wind farms. The overview of wind turbine failures and their effects is shown in [57], highlighting prevalent issues, including problems in the generator, gearbox, yaw system, rotor blade, and power electronics. It has underscored the significance of proactive defect diagnosis in averting expensive repairs.

The review in [57] examined model-based methods such as Kalman filters, observers, and set-membership techniques for residual-based fault detection, as well as data-driven strategies, including ANNs, deep learning, fuzzy logic, and machine learning models for fault classification. Additionally, it discusses signal-based approaches that analyse vibration, thermal, and electrical signals for fault feature extraction. Ref. [57] presents a sophisticated fault detection technique for DFIG-based wind turbines, concentrating on inter-turn short circuit (ITSC) issues within stator windings. It presents a multi-tiered fusion methodology that amalgamates empirical mode decomposition (EMD), least squares wavelet support vector machine (LS-WSVM), and genetic algorithm (GA) optimisation in conjunction with Dempster–Shafer evidential reasoning (DSER) for decision fusion.

It employed feature-level fusion (FLF) through EMD and decision-level fusion (DLF) via DSER, which improved the robustness of fault classification by integrating multi-sensor data. EMD was utilised to extract features from electrical signals (current, power) for the detection of ITSC problems. Employing a distance assessment method results in the minimisation of superfluous features.

The implementation of GA-optimised LS-WSVM for fault classification in [58] enhanced both accuracy and computational efficiency. The experimental validation using MATLAB/Simulink showed that the suggested GA-LS-WSVM model attained a fault classification accuracy of 98.27%, surpassing traditional approaches.

The study [59] examined the application of supervisory control and data acquisition (SCADA) data alongside machine learning techniques for the detection and prediction of wind turbine (WT) problems. The research assesses the efficiency of k-nearest neighbours (kNN) in conjunction with a bagging regressor, extreme gradient boosting (XGBoost), and artificial neural networks (ANNs) for fault classification, specifically emphasising gearbox oil sump temperature monitoring as a preliminary failure signal.

Ref. [59] utilised SCADA data to monitor temperature, rotor speed, power production, and gearbox condition. The kNN algorithm, when combined with a bagging regressor, attained optimal performance through PCA feature selection and grid search optimisation. XGBoost and ANN models demonstrated significant accuracy, although they necessitate distinct tuning methodologies.

Principal component analysis (PCA) and Pearson correlation coefficient (PCC) were employed in [59] to optimise input features. Grid search, random search, and Bayesian optimisation were compared for fine-tuning machine learning models. The proposed kNN model identified flaws a minimum of four weeks prior. It proposed integrating kNN with support vector machines (SVMs) to improve speed and minimise false alarms.

The application of EMD on a collection of signals combined with Gaussian white noise in [60] alleviated mode mixing and enhanced resilience. Variations in signal amplitudes and energy distribution were recorded, ensuring resilience against fluctuations in operating conditions and noise variations. This study presented an uncomplicated, low-complexity method that is readily implementable across various wind-power-producing systems. Reference [60] employed MATLAB/Simulink simulations to model and assess the doubly fed wind power system. The fault diagnosis method was evaluated on a back-to-back converter model, using a sampling frequency of 10 kHz and a sample size of 10,000. The methodology was corroborated by a 1.5 MW doubly fed wind power system simulation model, indicating a diagnostic precision of 99.57% across diverse wind speeds and above 70% under various noise conditions.

The research work [61] introduced an innovative method for identifying open-circuit problems in the insulated gate bipolar transistor (IGBT) modules of wind power converters. The suggested approach incorporates variational mode decomposition (VMD) for signal preprocessing, trend feature analysis for feature extraction, and a deep belief network (DBN) for classification. The emphasis was on grid-side converters (GSC) of permanent magnet synchronous generator (PMSG) wind turbines. It undertook single- and double-IGBT open-circuit problems, which significantly impact power quality and grid stability. It employed variational mode decomposition (VMD) for feature extraction. It analysed three-phase current signals under both normal and defective conditions. It derived trend feature vectors to differentiate fault patterns. A deep belief network (DBN) was employed in [61] to classify various IGBT failure states, demonstrating superior accuracy relative to conventional techniques such as wavelet transform and backpropagation neural networks (BPNN). Ref. [61] validated the DBN-based classifier to exceed traditional fault detection techniques in both accuracy and robustness.

The analysis of the examined literature discloses multiple gaps that the present study intends to rectify. A tabulated form of the reviewed studies is shown in

Table 1. Summary of reviewed DFIG power system fault detection methods.

Reference	Method and Model	Objective	Key Result/Contribution
[10]	DFIG converter topology	Explain DFIG structure and roles of RSC and GSC	Operational framework of converters
[11,29,33]	Review and thermal stress analysis	Identify GSC vulnerabilities	High temp > 150 °C leads to IGBT failure
[34,35,36,37]	Traditional signal thresholding	Fault detection	Limited adaptability under wind variations
[38,39,40,41]	CNN-LSTM	Real-time non-linear pattern detection	High accuracy fault classification
[42]	Crowbar-based WSE-MPPT and FRT	Enhance FRT under symmetrical faults	Stable recovery under faults
[43]	Stator current rotor voltage comp.	Improve LVRT performance	Reduced rotor overcurrent
[44]	Fuzzy logic and SSOA	Optimise PI and enhance response	Reduced active/reactive power overshoot
[43]	ICA-CC + EMCABN + MSEOA	Intelligent fault classification	98% accuracy, grid stability
[45]	NNPC + SMES	LVRT and transient oscillation reduction	Improved detection and voltage stability
[46]	State feedback + SMC	FTC with PI observer	Improved power regulation
[47]	RNN + LSTM + FCL	Deep learning FCL for stability	Superior transient stability
[48]	EEMD + IMF + PE	Data-driven voltage fault diagnosis	98.3% accuracy under noise
[49]	Chi-square current-based	Detect IGBT open-circuit faults	Precise diagnostics, 21 fault types
[50]	T2V-LSTM	SCADA-based fault prediction	84.97% accuracy up to 210 min ahead
[51]	NNPC + BADRC + ANFIS	Fault-tolerant grid control	95.1% actuator fault accuracy
[52]	EMD + LS-WSVM + GA	ITSC detection	98.27% accuracy
[53,54]	SCADA + ML (kNN, ANN, XGBoost)	Gearbox/power failure prediction	Early fault detection
[55]	EMD + noise reduction	Noise-resilient fault classification	99.57% accuracy
[56]	VMD + trend feature + DBN	IGBT open fault classification	Superior DBN-based detection

.

Although a wide array of existing methodologies spans signal processing, fuzzy logic, and sophisticated hybrid deep learning models, most strategies are hindered by either inadequate generalisation in varying wind circumstances or excessive dependence on fixed-threshold diagnostics and restricted fault categories. Although CNN-LSTM and T2V-LSTM have exhibited enhanced accuracy, contemporary models frequently experience:

• Insufficient resilience to noisy or non-stationary inputs ([48,50]).
• Inadequate adaptability to real-time dynamic situations ([34,47]).
• Restricted examination of multi-stage or intricate IGBT fault combinations ([49,56]).
• Reliance on specialised hardware or sensor configurations ([42,43,45]).

3. Modelling and Controlling of DFIG

3.1. Power Control

Wind-turbine-driven DFIG can produce and deliver electrical power at a stable frequency. Figure 3 [62] illustrates the fundamental theory of the contact between the stator and rotor magnetomotive forces (MMF) in a functioning induction machine. The currents in the stator windings produce a magnetic field that rotates at grid frequency, inducing a magnetic field in the rotor windings. The slip frequency will rotate the induced rotor MMF at the following value:

$${\mathsf{\omega }}_{\mathrm{s}\mathrm{l}\mathrm{i}\mathrm{p}}={\mathsf{\omega }}_{\mathrm{m}\mathrm{m}\mathrm{f}}^{\mathrm{r}\mathrm{o}\mathrm{t}\mathrm{o}\mathrm{r}}={\mathsf{\omega }}_{\mathrm{m}\mathrm{m}\mathrm{f}}^{\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{o}\mathrm{r}}-{\mathsf{\omega }}_{\mathrm{r}\mathrm{o}\mathrm{t}\mathrm{o}\mathrm{r}}$$

(1)

The rotor speed does not reach the stator side magnetic field strength. Here, ${\omega }_{slip}$ denotes the slip frequency, associated with the rotor side current and voltage frequency; ${\omega }_{mmf}^{stator}$ represents the stator frequency corresponding to the grid frequency in radians per second; and ${\omega }_{rotor}$ signifies the rotor side rotating frequency in radians per second, which corresponds to the product of the mechanical frequency and the number of magnetic pole pairs.

Figure 4 [63] illustrates the circuit diagram of the DFIGs within the d-q reference frame. One can presume the electrical modelling of the DFIG within the d-q reference frame. The equations for the voltage, flux, and power of the rotor-side and grid-side converters in DFIG-based wind energy systems can be articulated in the d-q frame of reference [64].

$$V_{ds}=R_s{I}_{ds}+\dot{{\phi }_{ds}}-{\omega }_s{\phi }_{qs}$$

(2)

$$V_{qs}=R_s{I}_{qs}+\dot{{\phi }_{qs}}-{\omega }_s{\phi }_{ds}$$

(3)

$$V_{dr}=R_r{I}_{dr}+\dot{{\phi }_{dr}}-{\omega }_r{\phi }_{qr}$$

(4)

$$V_{qr}=R_r{I}_{qr}+\dot{{\phi }_{qr}}-{\omega }_r{\phi }_{dr}$$

(5)

$${\phi }_{ds}=L_s{I}_{ds}+L_m{I}_{dr}$$

(6)

$${\phi }_{qs}=L_s{I}_{qs}+L_m{I}_{qr}$$

(7)

$${\phi }_{dr}=L_r{I}_{dr}+L_m{I}_{ds}$$

(8)

$${\phi }_{qr}=L_r{I}_{qr}+L_m{I}_{qs}$$

(9)

$R_s$ and $R_r$ represent the resistances on the stator and rotor sides, respectively, whereas $L_s$ and $L_r$ denote the self-inductance coefficients of the stator and rotor windings. $L_m$ represents the mutual coupling coefficient between the stator and rotor sides. $V_{ds}$, $V_{qs}$, $I_{ds}$, $I_{qs}$, $V_{dr}$, $V_{qr}$, $I_{dr}$, and $I_{qr}$ represent the voltage and current components on the stator and rotor sides within the d-q Park reference frame [65]. The per-unit electromagnetic torque equation in the d-q Park reference frame is posited by [64]

$$T_e={\phi }_{ds}{I}_{qs}-{\phi }_{qs}{I}_{ds}={\phi }_{qr}{I}_{dr}-{\phi }_{dr}{I}_{qr}=L_m\left(I_{qs}{I}_{dr}-I_{ds}{I}_{qr}\right)$$

(10)

Taking into account solely the resistances on the stator and rotor sides, the reactive and active stator powers of the DFIG are [64]

$$P_s=\left({\displaystyle \frac{3}{2}}\right)\left(V_{ds}{I}_{ds}+V_{qs}{I}_{qs}\right)$$

(11)

$$Q_s=\left({\displaystyle \frac{3}{2}}\right)\left(V_{qs}{I}_{ds}-V_{ds}{I}_{qs}\right)$$

(12)

The active and reactive rotor powers of the DFIG are

$$P_r=\left({\displaystyle \frac{3}{2}}\right)\left(V_{dr}{I}_{dr}+V_{qr}{I}_{qr}\right)$$

(13)

$$Q_r=\left({\displaystyle \frac{3}{2}}\right)\left(V_{qr}{I}_{dr}-V_{dr}{I}_{qr}\right)$$

(14)

Rewriting the system Equations (11)–(14) to account for rotating frames [62]:

$$P_T=P_s+P_r=\left({\displaystyle \frac{3}{2}}\right)\left(\dot{V_{qr}}{\dot{I}}_{qr}+\dot{V_{dr}}{\dot{I}}_{dr}+V_{ds}{I}_{ds}+V_{qs}{I}_{qs}\right)$$

(15)

$$Q_T=Q_s+Q_r=\left({\displaystyle \frac{3}{2}}\right)\left(\dot{V_{qr}}{\dot{I}}_{qr}-\dot{V_{dr}}{\dot{I}}_{dr}+V_{ds}{I}_{ds}-V_{qs}{I}_{qs}\right)$$

(16)

The reactive power on the stator side and the torque, which are the primary objectives of the rotor-side converter control system, are expressed in the following manner [66]. $I_{qs}$ and $I_{qr}$ denote the q component of the stator and rotor side currents, respectively, while $I_{ds}$ and $I_{dr}$ represent the d component of the stator and rotor side currents. $V_{qs}$ and $V_{ds}$ signify the q and d components of the stator side voltage, with p indicating the number of pole pairs of the generator. The flux connections on the stator and rotor sides, as well as the electromagnetic torque, are represented using d-q elements in the synchronous reference frame [67]:

$${\mathsf{\Psi }}_s=L_s{I}_s+L_m{I}_r$$

(17)

$${\mathsf{\Psi }}_r=L_m{I}_s+L_r{I}_r$$

(18)

$${\mathrm{T}}_m={\displaystyle \frac{3}{2}}p{\displaystyle \frac{L_m}{L_s}} \left({\mathsf{\Psi }}_{qs}{I}_{dr}-{\mathsf{\Psi }}_{ds}{I}_{qr}\right)$$

(19)

3.2. DFIG Rotor Side Control

The electrical circuit of the DFIG’s RSC is responsible for the unimpeded regulation of active and reactive powers to optimise the available power. In the conventional electrical grid, the stator flux φ_s remains constant. Moreover, the stator resistance value is negligible for moderately and highly rated power DFIGs; hence, it may be disregarded [32,68]. The Equations (2) and (3) regarding the stator voltages and fluxes are somewhat more comprehensible [69]:

$$V_{ds}=0$$

$$V_{qs}=V_s={\omega }_s{\phi }_{ds}$$

(20)

$${\phi }_{ds}={\phi }_{ds}=L_s{i}_{ds}+L_m{i}_{dr}$$

(21)

$${\phi }_{ds}=0=L_r{i}_{qs}+L_m{i}_{dr}$$

(22)

The stator, magnetising, and rotor inductances are stated as L_s, L_m, and L_r, respectively. The corresponding active and reactive powers are [69]

$${\mathrm{P}}_s={\displaystyle \frac{3}{2}}\left(V_{ds}{I}_{ds}+V_{qs}{I}_{qs}\right)=-{\displaystyle \frac{3}{2}}{\displaystyle \frac{L_m}{L_s}} \left(V_s{I}_{qr}\right)$$

(23)

$${\mathrm{Q}}_s={\displaystyle \frac{3}{2}}\left(V_{qs}{I}_{ds}+V_{ds}{I}_{qs}\right)={\displaystyle \frac{3}{2}}\left({\displaystyle \frac{V_s^2}{{\omega }_s{L}_s}}-{\displaystyle \frac{L_m}{L_s}}{V}_s{i}_{dr}\right)$$

(24)

Equations (23) and (24) demonstrate that the active and reactive powers on the stator side are independent. The components of the stator side power are directly proportional to the direct and quadrature rotor currents. Power systems generally utilise PI controllers to regulate the d-axis and q-axis components of rotor current and control. The reference current functions as the input for the PI current controller, which assesses the reference voltage and regulates the current to a fixed value to sustain a steady stator flux. The control system’s transfer function is specified as

$$\dot{G}\left(p\right)={\displaystyle \frac{I_{rp}^{*}\left(p\right)}{V_{rq}^{*}\left(p\right)}}={\displaystyle \frac{I_{rd}^{*}}{V\left(p\right)}}={\displaystyle \frac{1}{R_r+L_r\sigma p}}$$

(25)

The closed-loop transfer function:

$$\dot{H}\left(p\right)={\displaystyle \frac{I_{rq}\left(p\right)}{I_{rq}^{*}\left(p\right)}}={\displaystyle \frac{I_{rd}\left(p\right)}{I_{rd}^{*}\left(p\right)}}={\displaystyle \frac{\left(K_p^{″}p+K_i^{\prime }\right)\left(L_r\sigma \right)}{S^{″}\left(p\right)}}$$

(26)

S’(p) as the characteristic polynomial is given by

$$\dot{S}\left(p\right)=p^2+\left({\displaystyle \frac{R_r+K_p^{″}}{L_r\sigma }}\right)p+{\displaystyle \frac{K_i^{″}}{L_r\sigma }}$$

(27)

The PI controller’s gains for the rotor dynamics are

$$K_i^{″}=2L_r\sigma {\mu }^2$$

(28)

$$K_p^{″}=2L_r\sigma \mu -R_r$$

(29)

3.3. DFIG Grid Side Control

The GSC system can be mathematically represented in the MATLAB simulation model. The GSC is executed as a three-phase voltage source inverter utilising a space vector pulse-width modulation (PWM) strategy [12]. It comprises six insulated gate bipolar transistors (IGBTs), with each pair constituting one of the three output phases. The DC-Link, consisting of shunt capacitors, guarantees voltage stability throughout the converter. The PWM approach converts the DC-Link voltage into sinusoidal AC waveforms, which are then filtered and delivered to the grid through a transformer [12].

In this study, the LCL filter is incorporated into the grid side converter (GSC) model to reflect realistic operational conditions of grid-connected DFIG systems. The filter is composed of two inductors (one on the converter side and one on the grid side) and a capacitor connected between them, designed to attenuate high-frequency switching harmonics and improve power quality at the point of common coupling (PCC).

The design parameters of the LCL filter follow standard tuning criteria based on grid impedance and switching frequency to ensure stability and effective harmonic suppression. The inclusion of the LCL filter in the Simulink model impacts the current waveforms by introducing a phase shift and attenuating high-frequency noise components. These effects are critical in determining the shape and spectral content of the three-phase current signals that serve as input for fault detection. Therefore, modelling the LCL filter ensures that the current data used in diagnosis accurately represents real-world filtered outputs, making the proposed method more applicable to practical systems.

Figure 5 [12] depicts the topological configuration of the GSC, with T1–T6 denoting the IGBT switches. The rectifier’s DC voltage is transformed into a regulated AC waveform synchronised with the grid. The filtering phase reduces harmonic content and ensures compliance with grid standards.

The regulation of the GSC must achieve two objectives: (1) managing the DC-link voltage, and (2) regulating output phase currents. The active power output of the GSC is directly proportional to wind speed, as described by the following equation:

$$P_{GSC}={\displaystyle \frac{1}{2}}\rho A{C}_p{kv}^3$$

(30)

In this equation, ρ represents air density, A denotes the swept area of the turbine blades, C_p signifies the power coefficient, k is a conversion constant, and v indicates wind speed. This relationship demonstrates that phase current magnitudes fluctuate nonlinearly with wind speed, hindering fault identification. Equations (23) and (24) establish that the stator side power components are directly proportional to the direct and quadrature rotor currents. In vector-oriented control, the d-axis grid voltage is equated to its magnitude, but the q-axis voltage remains null. Therefore, the grid power expression derived from Equations (23) and (24) is as follows:

$$P_s={\displaystyle \frac{3}{2}}{V}_{ds}{I}_{ds}$$

(31)

$$Q_s=-{\displaystyle \frac{3}{2}}{V}_{ds}{I}_{ds}$$

(32)

3.4. Modeling IGBTs Fault in GSC

This study implemented several open-circuit faults in the GSC model by detaching one or more IGBTs to replicate and identify converter malfunctions. Each IGBT switch may fail in an open configuration, impacting the current waveform related to that phase. A total of ten operational scenarios were established: one healthy condition and nine separate open-circuit failure conditions. Fault states encompass single-phase and multiple-phase faults, which disrupt phase current symmetry and spectrum, rendering them appropriate for diagnostic classification using time-series analysis and machine learning. Thus, nine open-circuit failure states and one normal functioning state exist for the six IGBT variants of the GSC.

Table 2. Different types of IGBT states in GSC.

State Types	T1	T2	T3	T4	T5	T6
State 1 (Normal operation)	1	1	1	1	1	1
State 2 (Single open-circuit)	0	1	1	1	1	1
State 3 (Single open-circuit)	1	0	1	1	1	1
State 4 (Single open-circuit)	1	1	0	1	1	1
State 5 (Single open-circuit)	1	1	1	0	1	1
State 6 (Single open-circuit)	1	1	1	1	0	1
State 7 (Single open-circuit)	1	1	1	1	1	0
State 8 (Double open-circuit in phase a)	0	1	1	0	1	1
State 9 (Double open-circuit in phase b)	1	1	0	1	1	0
State 10 (Double open-circuit in the phase c)	1	0	1	1	0	1

illustrates the open-circuit failure status of IGBT modules in the GSC. T1, T2, T3, T4, T5, and T6 denote the respective IGBT modules seen in Figure 5. The values of Ti (i = 1, 2,…, 6) represent the states of IGBT modules. When the value of Ti equals 1, it indicates that Ti is functioning normally at that moment. When the value of Ti is zero, it shows that Ti is experiencing an open-circuit fault at that moment. Simulation outputs comprise three-phase current data (Ia, Ib, Ic) gathered over 2001 samples for each condition. The current signals function as the principal diagnostic indicators in the fault detection framework.

4. Methodology

This section outlines the methodology employed in the study, which integrated model-based, signal-based, and data-driven fault diagnostic approaches, utilising real wind speed data sourced from the NASA POWER database. While some commercial wind turbines are designed to operate at wind speeds of up to 25 m/s, many systems initiate cut-out mechanisms between 20 m/s and 25 m/s to prevent excessive mechanical stress and ensure operational safety.

Table 3. Statistical summary of the wind data.

Statistics	Value
Mean wind speed	8.0 m/s
Standard deviation	2.0 m/s
Minimum wind speed	2.5 m/s
Maximum wind speed	15–18 m/s
Typical range (68%)	6.0–10.0 m/s
Distribution type	Positively skewed, non-normal
Sampling interval	Daily to hourly

presents a wind speed range (mean: 8.0 m/s; max: 15–18 m/s) representative of typical and near-maximum safe operating conditions for DFIG-based turbines. This choice ensures the reliability and realism of the simulated fault scenarios while avoiding the inclusion of extreme conditions that would trigger system shutdown in actual wind turbine operations. Including this operational range enhances the generalisability of the proposed diagnostic model to practical field applications.

4.1. Variational Mode Decomposition (VMD)

Variational mode decomposition (VMD) is an adaptive signal processing method that disaggregates a non-stationary signal into a predetermined number of sub-signals, referred to as modes, each exhibiting distinct sparsity characteristics in the frequency domain. In contrast to empirical mode decomposition (EMD), variational mode decomposition (VMD) is established as a well-defined variational issue, hence improving its resilience and convergence characteristics [70].

VMD aims to break down an input signal f(t) into K modes u_k(t), each characterised by a corresponding central frequency ω_k. The decomposition seeks to minimise the aggregate bandwidths of the modes, subject to the reconstruction of the original signal. The confined variational problem is articulated as [70]

$$min\left\{{\sum }_{k=1}^K{‖{\partial }_t\left[\left({\delta }_t+{\displaystyle \frac{j}{\pi t}}\right)\mathrm{\ast }{u}_k\left(t\right)\right]e^{-j{\omega }_{k}t}‖}_2^2\right\}subject{\sum }_{k=1}^K{u}_k\left(t\right)=f\left(t\right)$$

(33)

This formulation guarantees that each mode is concentrated around its central frequency ω_k, and the aggregation of all modes reconstructs the original signal. The Lagrange multipliers approach and the alternating direction approach of multipliers (ADMM) are utilised to address this issue, resulting in an iterative procedure that converges to the modes u_k(t) and their associated central frequencies ω_k.

This paper used VMD to extract significant temporal and frequency information from the current data. VMD disaggregates a signal into a specified number of intrinsic mode functions (IMFs), with each function representing a unique spectral component. The three-phase currents were decomposed into K = 7 intrinsic mode functions (IMFs). The authors chose all IMFs from each step (a total of six components) to enhance efficiency and preserve the most critical properties for subsequent processing. The resultant IMF matrix for each sample was measured [2001 × 21].

Implications of LCL Filter on Feature Extraction

As the GSC is typically connected to the grid via an LCL filter, in the simulation environment of the study, the LCL filter was incorporated within the GSC model. The presence of the LCL filter significantly influences the nature of the current signals used for fault classification. Since the LCL filter suppresses high-frequency switching harmonics, the current signals at the GSC output are smoother and exhibit attenuated transients compared to raw inverter outputs.

Consequently, when variational mode decomposition (VMD) is applied, the extracted intrinsic mode functions (IMFs) reflect the spectral components present after filtering. This leads to the isolation of more robust and stable frequency bands that are better suited for pattern recognition and classification using deep learning models. The CNN-LSTM architecture is thus trained on filtered, realistic signals, improving the model’s ability to generalise to real turbine systems that incorporate LCL filters. This modelling choice enhances both diagnostic accuracy and deployment readiness for field applications.

In our simulation environment, the LCL filter was modelled as part of the converter system. The CNN-LSTM model was trained on current signals post-filter, meaning the diagnostic model accounts for the waveform characteristics influenced by the filter. This ensures that the extracted features, through variational mode decomposition (VMD), reflect realistic post-filter behaviour.

4.2. Data Normalisation and Augmentation

Each IMF component underwent z-score normalisation to maintain uniformity within the dataset and enhance model convergence during training. Furthermore, data augmentation was implemented by introducing low-level Gaussian noise to the original signals, producing numerous training examples for each failure state, and improving model generalisation. Each operational class was expanded to incorporate ten samples, resulting in 100 training samples (10 classes × 10 samples/class).

4.3. Deep Learning Models

Two distinct deep learning architectures were employed to assess categorisation performance:

4.3.1. CNN-LSTM Hybrid Model

The CNN-LSTM model is a hybrid deep learning architecture that integrates CNNs with long short-term memory (LSTM) networks to utilise spatial feature extraction and temporal sequence modelling. This combination is highly effective for time-series classification, signal processing, fault detection, and activity recognition tasks, particularly when input data demonstrates both local patterns and sequential dependencies [71].

CNNs are engineered to acquire spatial hierarchies of features autonomously and adaptively from input data. They are proficient at extracting local and distinctive patterns using convolutional and pooling layers. In time-series applications, CNNs may extract localised features such as abrupt spikes, edges, or periodic segments from one-dimensional or two-dimensional representations of signals.

LSTM is a subtype of recurrent neural networks (RNNs), adept at modelling temporal and sequential data. LSTM mitigates the vanishing gradient issue inherent in ordinary RNNs with the implementation of gating mechanisms: the input gate, forget gate, and output gate [72].

The CNN-LSTM model is designed with CNN layers first to extract high-level features from raw input data (such as time series or spectrograms), followed by LSTM layers that capture temporal correlations among the derived features. Ultimately, dense layers and a softmax classifier are employed to execute classification. This model initiates with one-dimensional convolutional layers to extract local temporal patterns from IMFs, followed by an LSTM layer that captures long-term temporal dependencies. The architecture comprises a sequence of input layer (21), two convolutional layers utilising ReLU and max pooling, a bilstm Layer (100), and fully connected layers followed by softmax layers for classification, as illustrated in Figure 6 [72].

4.3.2. CNN Only Model

The authors constructed a purely convolutional architecture for comparison. A global average pooling 1D layer was employed to remove the temporal dimension before classification, resolving the discrepancy between input and output lengths in sequence-to-label tasks.

4.4. Classification and Evaluation

The models were trained to categorise the input sequences into one of the ten operational states. The performance of each model was assessed by classification accuracy and confusion matrices. The CNN-LSTM model attained maximum accuracy, illustrating its proficiency in extracting spatial (through CNN) and temporal (via LSTM) characteristics from VMD-processed current data. The fault diagnosis flow for DFIG-based wind turbine utilising CNN-LSTM is illustrated in Figure 7, which is detailed below:

The process commences with initiating the diagnostic workflow for the DFIG-based wind turbine system.

4.4.1. Simulation of DFIG-Based Wind Turbine for Specified Fault Conditions

A comprehensive simulation model of the DFIG wind turbine was built. Multiple operational and fault scenarios were delineated, encompassing normal conditions and nine distinct fault states associated with the GSC. These encompass faults such as line-to-line, line-to-ground, and open-switch failures.

4.4.2. Collection of Three-Phase Current Data from GSC

For each defined failure state in the simulation, three-phase current signals are recorded at the GSC output. The time-series signals function as the principal raw input for future feature extraction and classification tasks.

4.4.3. Feature Extraction by Variational Mode Decomposition (VMD)

The acquired current signals undergo VMD, an advanced signal processing method that disaggregates a signal into many intrinsic mode functions. The objective is to derive pertinent frequency and time-domain features that encapsulate the fault characteristics more concisely and in a distinguishable manner.

4.4.4. Training of CNN-LSTM Neural Network

The features derived from the VMD process are utilised to train a hybrid deep learning model that integrates CNN with LSTM networks. CNN layers extract spatial features, whereas LSTM layers learn temporal dependencies in the signal.

4.4.5. Classification Outcomes

Following training, the CNN-LSTM model categorises fresh data samples into one of ten established states (normal plus nine fault conditions). The evaluation of categorisation performance employs variables such as accuracy, confusion matrix, precision, and recall.

4.4.6. End

The diagnostic procedure culminates in model assessment and validation. The finalised trained model can be utilised for real-time problem detection and classification in DFIG-based wind energy systems.

5. Simulation Results

The authors executed and compared three methodologies to assess the efficiency of the proposed CNN-LSTM model for fault diagnosis in the grid-side converter of a DFIG-based wind turbine: a CNN-LSTM hybrid model, an independent LSTM network, and a multilayer perceptron (MLP). Each model was trained with three-phase current signals (Ia, Ib, Ic) obtained from 10 distinct operational states, encompassing normal operation and nine specific IGBT open-circuit faults.

5.1. Signal Exchange and Fault Simulation

This study utilised a simulation of a DFIG-based wind turbine, as depicted in Figure 8, sourced from the Simscape Electrical library in MATLAB/Simulink. The system is engineered to simulate authentic operational conditions and evaluate fault classification efficiency within realistic wind fluctuations. The GSC model was revised in the designated MATLAB simulation. The system is designed to utilise a conventional three-phase PWM inverter configuration, featuring six IGBT switches that regulate power transfer to the grid. Nine unique IGBT open-circuit fault scenarios were implemented in the GSC to simulate fault conditions, from a healthy working state, yielding ten classes. Three-phase current indications (Ia, Ib, Ic) were recognised for each state under diverse wind circumstances.

Each signal was sampled over 2001 time steps, documenting dynamic behaviour under faulty and healthy conditions. Real wind speed data from the NASA POWER database for the year 2021 was utilised to guarantee a representative spectrum of operational conditions. Table 3 indicates the statistical summary of the wind data.

A selection of 2001 hourly data, spanning hours 3000 to 5000 (of the June, July, and August period), was made to provide adequate variety in wind profiles, as illustrated in Figure 9. The observations were obtained from a particular geographic site in Germany characterised by significant wind activity, hence ensuring dynamic behaviour in the wind turbine system. The simulation incorporates the wind speed signal as an input to the wind turbine model, enabling the generator to react to authentic variations. The focus is on the three-phase current signal at the GSC under ten distinct conditions: one normal condition and nine fault scenarios, encompassing open circuit faults and converter switching problems.

Figure 10 and Figure 11 illustrate the three-phase currents during state 2 and state 9 operations, respectively. All three-phase current samples across 10 states were analysed using VMD at seven levels. Figure 12 and Figure 13 illustrate the waveforms of the three-phase current and the mode coefficient series for the first and fifth levels under failure state No. 10 (double open-circuit in phase c). The blue, red, and yellow curves represent the A, B, and C phase currents in Figure 10, Figure 11, Figure 12 and Figure 13 respectively.

The DFIG’s working principle can be investigated using the d-q model and the well-known theory of rotating fields of transformations from two-to-three and three-to-two axes.

The rotating magnetic field speed N_stator going through the stator windings is controlled by the rotor’s rotational speed N_rotor and the AC currents fed into the rotor windings f_rotor. DFIGs have two working modes; in (i) mode N_r > N_s, the slip is negative, and the generator is super-synchronous, delivering power to the grid from stators and rotor windings. For synchronous generator mode, the stator and rotor windings are positive, and (ii) mode N_r < N_s, slip is then positive [72,73]. The studies [63,74,75] explain the control mechanisms and mathematical models used by the DFIG in detail.

5.2. Neural Network Training

Following aggregation of current signals across all circumstances, this work utilised VMD for feature extraction. The features were subsequently input into the CNN-LSTM neural network, which was trained to categorise the system state into one of 10 predefined groups. All models were trained to utilise the Adam optimiser with an initial learning rate of 0.001, a batch size of 10, and a maximum of 100 epochs. The training and testing datasets were divided using an 80/20 ratio.

5.3. Accuracy Comparison

The classification outcomes, illustrated in the confusion matrix in Figure 14 demonstrate the precision with which the CNN-LSTM model recognised each circumstance. Most classes were accurately classified, with almost all forecasts corresponding appropriately to their actual labels. The confusion matrix determines the model’s efficiency in separating 10 distinct classes representing various fault or operational states in the DFIG-based wind turbine system.

The diagonal elements indicate accurately identified cases, whereas the off-diagonal entries signify misclassifications. Except for class 4, all classes attained accurate classifications without any misclassifications. The results suggest that although the CNN-LSTM model excels in certain classes, further feature enhancement, such as deeper network layers, sophisticated feature selection, or signal fusion techniques, may be necessary to enhance class differentiation, especially for those displaying overlapping frequency or time-domain characteristics.

Figure 15 illustrates the confusion matrix of the CNN model. This matrix shows the classification performance across all ten operational states. CNN exhibited low confusion; however, several errors were misclassified due to overlapping signal characteristics, especially under rapidly fluctuating wind conditions, such as classes 6, 7, and 9. Furthermore, class 4 was completely misclassified.

The models’ classification performance was assessed using accuracy, with the F1-score (the harmonic means of precision and recall) also considered. The CNN-LSTM model attained superior accuracy compared to the independent LSTM, CNN, and MLP. The findings are encapsulated in

Table 4. Model accuracy comparison.

Model	Accuracy (%)	Recall (%)	F1 (%)	Precision (%)
CNN-LSTM	88	90	85.71	84
LSTM	86	87	83.50	80
CNN	84	83	82.33	73.17
MLP	81	79	77.80	76

.

Primary component analysis (PCA) was utilised to condense the high-dimensional feature space into two primary components for visual examination across the 10 classes. The convolutional layers are proficient at extracting spatial characteristics from the input data, whereas the LSTM layer identifies the temporal dependencies essential for differentiating fault patterns across time. The independent CNN, however efficient, is deficient in spatial filtering skills offered by CNN-LSTM. The results confirm the suggested CNN-LSTM architecture as a dependable and precise method for automated fault identification in real-time wind turbine DFIG converter systems.

6. Discussion and Comparison

This study presents a hybrid deep learning model combining CNN and LSTM for diagnosing grid-side converter (GSC) defects in DFIG-based wind turbines. The model was evaluated using simulated three-phase current data over 10 operational circumstances, comprising nine fault states and one healthy state, and was compared to standalone CNN, LSTM, and MLP architectures. The utilisation of authentic wind speed data from the NASA POWER database, along with a comprehensive MATLAB/Simulink simulation environment, guaranteed realistic and rigorous testing circumstances.

The CNN-LSTM model surpassed all competitors, attaining an accuracy of 88%, a recall of 90%, and an F1-score of 85.71%. Its exceptional performance arises from its capacity to collect both spatial variables (through CNN) and temporal features (via LSTM), rendering it highly appropriate for the dynamic, non-linear characteristics of wind energy systems. The CNN and MLP models demonstrated misdiagnosis in overlapping fault states, especially under variable wind conditions.

This study enhances existing literature by improving shortcomings in previous models, including those by [41], which employed CNNs in static conditions, and [38,76], which omitted real-world wind variability and VMD-based preprocessing. In contrast to [76], whose T2V-LSTM model focused on bearing faults, this study aimed to detect defects specific to GSC.

The integration of variational mode decomposition (VMD) markedly increased feature extraction by breaking down current signals into intrinsic mode functions, while z-score normalisation and data augmentation strengthened model generalisation. PCA further validated the model’s discriminative capability among fault categories.

7. Conclusions

This research introduced a comprehensive data-driven approach for identifying and categorising IGBT open-circuit faults in the GSC of a DFIG-based wind turbine through a hybrid deep learning framework. The proposed methodology integrates the variational mode decomposition (VMD) for time-frequency signal analysis with a convolutional neural network–long short-term memory (CNN-LSTM) framework to leverage spatial and temporal attributes of three-phase current signals.

A comprehensive MATLAB/Simulink model of the DFIG system was employed to simulate 10 operational scenarios comprising one normal condition and nine fault states. The three-phase current signals underwent VMD processing to identify predominant intrinsic mode functions (IMFs), which were subsequently normalised and utilised as input features for the deep learning models. The CNN-LSTM network exhibited enhanced diagnostic performance relative to the CNN, LSTM, and MLP models, through greater accuracy and superior fault distinguishability.

The findings validate that combining VMD with deep neural networks markedly improves the model’s capacity to identify non-stationary and non-linear patterns related to converter defects. Furthermore, the application of CNN-LSTM facilitates the system’s ability to comprehend sequential relationships in the present waveforms, rendering it very effective for real-time defect detection under fluctuating wind situations.

Future research will investigate real-time implementation on embedded systems, extended fault coverage, including short circuit, and sensor malfunctions, and integrating empirical wind farm data to assess model generalisation. Integrating adaptive wind speed profiles and hybrid ensemble learning methodologies may improve diagnostic efficiency.

The results emphasise the necessity of efficient and reliable systems for sustainable wind energy production. In the future, diverse learning approaches will be integrated to enhance model performance efficiently, including exciting research in digitalisation. According to shifting research trends, developing a digital twin system (DTS) should improve communication between demand-side management, energy forecasting, reliability, monitoring, data visualisation, and virtual machines.