A deep learning-based GAN-LSTM framework for friction failure detection and acoustic emission signal enhancement in machine engines

Wasserstein generative adversarial network with gradient penalty (WGAN-GP)

WGAN-GP is an advanced variant of GAN designed to improve training stability and output quality. Unlike traditional GANs that rely on the Jensen Shannon divergence, WGAN-GP optimizes the Wasserstein distance, which provides smoother gradients and avoids mode collapse. To further stabilize training, a gradient penalty is introduced to enforce the Lipschitz constraint without resorting to weight clipping. As a result, WGAN-GP can generate high-fidelity signals while retaining the essential structural properties of the raw AE signal, making it highly effective for denoising tasks in fault diagnosis.

Long short-term memory (LSTM)

LSTM is a specialized type of recurrent neural network (RNN) capable of learning long-term dependencies in sequential data. It overcomes the vanishing gradient problem of standard RNNs by incorporating a memory cell and gating mechanisms, input gate, forget gate, and output gate, that regulate how information is stored, updated, and retrieved. This selective memory control allows LSTMs to effectively capture temporal dependencies in AE signals, ensuring that subtle fault-related patterns over time are preserved. By integrating LSTM with the denoised output of WGAN-GP, the proposed method achieves robust feature extraction and reliable fault classification.

Data collection

In this study, high-quality AE signals were collected from rotating machinery under varying operational conditions to ensure a diverse and representative dataset for fault diagnosis. By capturing signals across different rotational speeds, load levels, and fault types, the dataset provides a rich foundation for training the proposed model.

The Engine Acoustic Emissions Dataset (EAE-I) includes vibration and AE signal measurements of the rotating equipment, predominantly bearings, at a range of operating speed levels and loading conditions. It is usually used to implement fault detection, condition monitoring, and predictive maintenance of industrial systems. It includes signal acquisition of normal bearings and faulty bearings with a range of types of faults, such as inner race faults, outer race faults, and ball defects. It is made up of a total of 10,000 samples with a 50 kHz sample frequency, with the signal duration of 2 to 10 s. The information captured includes both time-domain raw AE waves and frequency-domain formats such as the FFT transformation of the signal.

The Rolling Element Bearing Sound Dataset (REB-II) is a 15,000 AE signal database obtained from a range of operating conditions of the rolling element bearings, which makes the database a rich database to diagnose faults and monitor the status of a machine. It includes normal operating conditions and faulty operating conditions of bearings at a range of rotational velocities and loads to provide a complete analysis of the patterns of the faults. It includes raw time-domain signal samples, frequency-domain analysis, and significant statistical measures like RMS value, kurtosis, and spectral entropy that support discriminating between normal operating conditions and faulty operating conditions. Types of faults are inner race faults, outer race faults, and ball faults, with unique acoustic properties for each of them. Next is the ToyADMOS Dataset (TAD-III), a database of 20,000 AE signals recorded from small industrial equipment with the objective of detecting abnormal noises for machinery faults or wear on the machine component. It encompasses typical operating conditions and malfunctioning operating conditions of various types of equipment, such as motors, gears, and rotating elements, to offer a wide variety of applications to perform research on fault detection. It comprises raw time-series acoustic signal samples, frequency-domain transformations, and calculated measures of statistics such as RMS value, spectral entropy, and peak-to-peak value that aid to distinguish between normal operational conditions and abnormal operational conditions.

The visualization of Fig. 3 and Table 2 presents a detailed overview of the AE (AE) datasets applied within this work, with information about their signal properties and structural variability. The graph representation displays the AE signal of the EAE-I, REB-II, and TAD-III datasets with varying amplitudes, frequency, and levels of noise between the various conditions of the machine. The green, red, and yellow curves indicate the unique patterns of the AE signal with the various datasets representing the fault-related abnormalities differently. To this end, the table offers the major attributes of the datasets with signal types, operating speed levels, conditions of the load, RMS levels, and types of faults to put the datasets into perspective with respect to training the models. The EAE-I offers the time-domain signal with a structured format, while the REB-II includes raw and the FFT-transformation of the signal to provide a time-frequency mix of analysis. The TAD-III is a sequence of the AE signal with a focus on time-series relationships.

Data preprocessing

Preprocessing plays a critical role in the GAN-LSTM framework for AE signal enhancement and fault diagnosis, ensuring that the input data is clean, structured, and suitable for deep learning models (Tong, Zhang & Xie, 2024; Hang et al., 2024). Raw AE signals often suffer from high noise levels, varying signal amplitudes, and non-stationary characteristics, which can negatively impact model performance (He et al., 2024). To overcome these challenges, Algorithm 1 shows the preprocessing workflow that employs wavelet transform-based denoising, z-score normalization, and short-time Fourier transform (STFT) as the core preprocessing techniques, ensuring optimal signal representation for the GAN and LSTM components.

Wavelet transform-based denoising is chosen due to its superior ability to preserve both time and frequency-domain features while removing high-frequency noise. Given an AE signal $x\left(t\right)$, the continuous wavelet transform (CWT) is defined as:

(1) $$W_x\left(a,b\right)={\int }_{-\mathrm{\infty }}^{+\mathrm{\infty }}x\left(t\right){\mathrm{\psi }}^{\ast }\left({\displaystyle \frac{t-b}{a}}\right)dt$$where $a$ and $b$ represent the scale and translation parameters, and ψ(t) is the mother wavelet function. By applying the thresholding techniques within the wavelet domain, the unwanted parts of the noise are eliminated while the most significant fault characteristics of the AE signal are preserved. The quality of the AE signal is enhanced before being supplied to the GAN to enhance the signal. Z-score normalization is applied to maintain the AE signal at a standard scale to avoid the issue of deep learning training due to gradient instability. Given an AE signal vector $X=\left\{x_1,x_2,...,x_n\right\}$, the normalized signal is computed as:

(2) $${\mathrm{X}}^{\mathrm{\prime }}={\displaystyle \frac{X-\mu }{\sigma }}$$where μ is the mean, and σ is the standard deviation of the signal. This conversion guarantees that every input feature is normalized to have zero mean and unit variance, which enables the GAN to produce high-fidelity synthetic AE signals and facilitates the LSTM to effectively learn temporal dependencies without being affected by large magnitude changes. The short-time Fourier transform (STFT) is used to transform non-stationary AE signals into a time-frequency representation, which is especially useful for the analysis of transient events in AE signals. The STFT is mathematically expressed as:

(3) $$\mathrm{X}\left(\tau ,\mathrm{\omega }\right)={\int }_{-\mathrm{\infty }}^{+\mathrm{\infty }}\mathrm{x}\left(\mathrm{t}\right)\mathrm{w}\left(\mathrm{t}-\tau \right)e^{-\mathrm{j}\omega \mathrm{t}}dt$$where w (t − τ) is the window function applied to localize the frequency content at different time intervals. This transformation enables the model to distinguish between different fault types based on their spectral characteristics, significantly improving fault classification accuracy. This preprocessing collectively enhances the proposed model’s ability to learn meaningful fault patterns.

GAN-based AE signal enhancement

Enhancing the AE signal is a central aspect of this work due to the raw AE signal being affected by excessive noise levels, time-varying amplitudes, and non-stationarity that pose challenges to exact fault diagnosis. Traditional denoising techniques such as Fourier filtering and wavelet thresholding are inefficient at preserving the subtle patterns of the fault while removing unnecessary noise. To avoid the aforementioned shortcomings of the traditional techniques, this work uses a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) as the core model to improve the quality of the AE signal. Compared to standard GAN models that are vulnerable to mode collapse and training instability, the smooth convergence of the WGAN-GP can generate high-fidelity signals by optimizing the Wasserstein distance instead of the standard Jensen-Shannon (JS) measure of divergence. The central argument of the usage of the WGAN-GP is that the latter can produce high-quality denoising of the AE signal with the retention of key structural properties of the original fault signal.

The fundamental principle behind WGAN-GP is to approximate the Earth Mover’s Distance (EMD) between real AE signals and generated signals, leading to a more stable training process. The Wasserstein distance between two probability distributions $Pr$ (real AE signals) and $Pg$ (generated AE signals) is defined as:

(4) $$W\left(Pr,Pg\right){=}_{\gamma \in \prod \left(Pr,Pg\right)}^{inf}{\mathbb{E}}_{\left(x,y\right)}\sim \gamma \left[\parallel \mathrm{x}-\mathrm{y}\parallel \right]$$where $\prod \left(Pr,Pg\right)$ represents the set of all possible joint distributions with marginals $Pr$ and $Pg$. Instead of using a discriminator with a sigmoid activation function, as in conventional GANs, WGAN-GP employs a critic network that outputs a real-valued score, encouraging a smooth function space for better gradient updates. To enforce Lipschitz continuity, WGAN-GP introduces a gradient penalty term, computed as:

(5) $$E_{x^{\mathrm{\prime }}\sim P_x^{\mathrm{\prime }}}\left[{\left(\parallel {\mathrm{\nabla }}_{x^{\mathrm{\prime }}}D\left(x^{\mathrm{\prime }}\right){\parallel }_2-1\right)}^2\right]$$where $x^{\prime }$ is a randomly interpolated sample between real and generated AE signals, and $D\left(x^{\prime }\right)$ is the critic’s score.

This gradient penalty ensures that the discriminator behaves as a 1-Lipschitz function, preventing exploding gradients and mode collapse, common issues in traditional GANs.

In this model, WGAN-GP is designed to take raw AE signals from datasets (EAE-I, REB-II, and TAD-III) and learn to generate high-fidelity, denoised versions that retain critical fault characteristics. The generator, parameterized by G(z), transforms a latent noise vector $z\sim P_z$ into a synthetic AE signal:

(6) $$x^{\mathrm{\prime }}=G\left(z\right),z\sim P_z$$while the critic network $D\left(x\right)$ assigns scores to both real and generated AE signals, optimizing the Wasserstein loss function:

(7) $$LD=E_{x\sim P_r}\left[D\left(x\right)\right]-E_{x^{\mathrm{\prime }}\sim P_y}\left[D\left(x^{\mathrm{\prime }}\right)\right]+\lambda E_{x^{\mathrm{\prime }}\sim P_x^{\mathrm{\prime }}}[{\left(\parallel {\mathrm{\nabla }}_{x^{\mathrm{\prime }}}D\left(x^{\mathrm{\prime }}\right){\parallel }_2-1\right)}^2$$where λ is the gradient penalty coefficient. The generator is optimized using:

(8) $$LG=-E_{x^{\mathrm{\prime }}\sim P_y}\left[D\left(x^{\mathrm{\prime }}\right)\right].$$

Ensuring that generated AE signals receive higher critic scores, leading to better quality enhancement. In the context of this research, WGAN-GP plays a pivotal role in removing unwanted noise while preserving essential fault information, thereby improving the quality of AE signals before they are fed into the LSTM-based fault classification model. By learning a smooth transformation from noisy to clean AE signals, WGAN-GP enhances model robustness, ensuring superior fault detection performance across different operating conditions. The complete working flow of the WGAN-GP for signal enhancement has been presented in Algorithm 2.

Figure 4, is the visualization that clearly shows the effect of WGAN-GP-based AE signal enhancement, depicting the conversion from a noisy raw AE signal to a denoised high-fidelity AE signal. In the initial frame, the red waveform is the AE signal prior to processing, with high levels of noise and unpredictable fluctuations masking important fault-related features. These distortions have a substantial impact on the reliability of fault classification models, causing misinterpretation of machine conditions. However, the second frame shows the green waveform, indicating the AE signal after WGAN-GP enhancement. This iteration shows smoother patterns with less noise and maintains important fault characteristics inherent in the structure of the signal. By using Wasserstein loss with gradient penalty, WGAN-GP successfully learns a mapping from noisy to clean AE signals in a stable, robust, and high-quality signal reconstruction manner. This improvement is especially useful for fault classification since it enables the model to concentrate on significant sequential patterns without being affected by random noise. The WGAN-GP-generated denoised AE signals ultimately enhance fault detection precision and generalization, making the method very efficient for practical application in predictive maintenance and structural health monitoring.

LSTM based fault classification

Algorithm 3 represents the fault classification that the proper identification of various types of faults from AE signals facilitates predictive maintenance and the early detection of faults. AE signals are inherently time-dependent and sequential in nature, with traditional machine learning models like support vector machines (SVMs) or convolutional neural networks (CNNs) being ineffective at handling their long-term relationships. To overcome this limitation, long short-term memory (LSTM) networks are used due to their capacity to learn about the temporal relationships and learn from the long-range patterns of the AE signal (Umar et al., 2024). In contrast to traditional recurrent neural networks (RNNs), which are prone to the vanishing gradient issue, the introduction of a gating mechanism by the LSTM enables the network to preserve key characteristics over a large number of time steps, thereby providing a more accurate fault classification.

Mathematically, an LSTM unit consists of three primary gates: the forget gate, the input gate, and the output gate, which controls the flow of information within the network. Given an input sequence $x_t$, the forget gate determines which past information to retain or discard using a sigmoid activation function:

(9) $$f_t=\sigma \left(W_f{h}_{t-1}+U_f{x}_t+b_f\right)$$where $W_f$ and $U_f$ are weight matrices, $h_{t-1}$ is the hidden state from the previous time step, and $b_f$ is the bias term. The input gate then decides how much new information should be stored in the cell state:

(10) $$i_t=\sigma \left(W_i{h}_{t-1}+U_i{x}_t+b_i\right)$$

(11) $$C_t^{\mathrm{\prime }}=tanh\left(W_c{h}_{t-1}+U_c{x}_t+b_c\right)$$where $C_t^{\mathrm{\prime }}$ represents candidate memory values. The cell state is then updated as:

(12) $$C_t=f_t\odot C_{t-1}+i_t\odot C_t^{\mathrm{\prime }}.$$Ensuring the retention of critical fault-related features. Finally, the output gate determines the next hidden state, which is used for classification:

(13) $$o_t=\sigma \left(W_o{h}_{t-1}+U_o{x}_t+b_o\right)$$

(14) $$h_t=o_t\odot tanh\left(C_t\right)$$where $h_t$ captures the essential fault information from the AE signal. The LSTM-based classifier in this model takes the enhanced AE signals from WGAN-GP and learns the sequential relationships between different signal patterns, distinguishing between normal, inner race faults, outer race faults, and ball defects with high accuracy. The network is trained using a categorical cross-entropy loss function:

(15) $$\mathrm{L}=-\sum _{i=1}^N{y}_{i}log(y_i^{\mathrm{\prime }})$$where $y_i$ is the actual fault label and $y_i^{\mathrm{\prime }}$ is the predicted probability. By leveraging time-series dependencies in AE signals, the LSTM model effectively classifies different fault conditions, ensuring higher generalization across varying machine conditions.

Figure 5 is the final classified output of the LSTM-based fault classification model that displays the potential of the model to discriminate between normal and faulty AE signals with time. Each of the graph lines represents the predicted probability of a specific fault type Normal, Inner Race Fault, Outer Race Fault, and Ball Defect at 20 time steps. The peak value of the “Normal” at the majority of time steps reflects the proper classification of the fault-free conditions, while the change of the fault classes reflects the response of the model to the change of the properties of the AE signal. The smooth transitions between the classes and the well-separated classes indicate the effectiveness of LSTM to learn the long-term dependency of the AE signal to provide robust classification. The visualization provides strong empirical evidence that the proposed GAN-LSTM approach is successfully enhancing and classifying the AE signal with a very high degree of accuracy to serve as a robust solution to the challenges of predictive maintenance and real-time fault detection.

Finally, the LSTM classifier and the GAN-based signal enhancement of the AE signal are merged into a complete end-to-end pipeline to supply smooth signal processing with strong fault detection capabilities. The raw signal of the AE is cleaned of noise by the WGAN-GP model, which processes the raw signal to generate high-quality, denoised signals with the significant patterns of the faults remaining intact. These cleaned signals are then supplied to the LSTM classifier that scans their sequence relationships to find the fault labels with a very high degree of accuracy. To improve the performance of the model to the optimal levels possible, the hyperparameters of the model are optimized, that is, the learning rate, the number of LSTM layers, the batch size, and the Wasserstein loss weight factor. The training is stabilized by the usage of techniques such as adaptive learning rate schedule, dropout regularization, and gradient clipping to avert the problem of the model getting trapped into the mode of the training data while learning to avert the problem of the model getting trapped into the mode of the training.

System requirements

The implementation of the proposed AE fault classification model requires a high-performance computing environment to handle large-scale AE signal processing, deep learning training, and real-time inference. Given the computational complexity of WGAN-GP for AE signal enhancement and LSTM for sequential fault classification, a dedicated GPU-enabled system is essential for efficient execution. System configuration comprises an NVIDIA GPU (minimum RTX 3090 or above) with a minimum of 24 GB VRAM for accelerated tensor calculation and deep learning optimizations. A minimum multi-core processor (Intel i9 or AMD Ryzen 9) with a clock speed of at least 3.5 GHz assures efficient preprocessing of AE signals, such as wavelet denoising and STFT transforms. The model requires at least 32 GB RAM to process large AE datasets (EAE-I, REB-II, and TAD-III) and avoid memory bottlenecks in batch training. Fast NVMe SSD storage (at least 1TB) is required to hold processed AE signals, trained models, and real-time inference results. The software environment consists of Python (3.8+), TensorFlow/PyTorch for deep learning, SciPy and NumPy for signal processing, and Matplotlib for result visualization. Moreover, CUDA and cuDNN libraries are necessary for GPU acceleration.

Data distribution for reliable model evaluation

To ensure a robust and unbiased evaluation of the proposed framework, the datasets EAE-I, REB-II, and TAD-III are strategically divided into training, validation, and testing sets in a 70:15:15 ratio as presented in Table 3.

The training set (70%) is applied to refine the model parameters to improve the ability of the GAN to generate quality AE signals, while the 15% validation set is utilized to fine-tune hyperparameters to avoid overfitting. The 15% testing set is then kept separate to measure the real-world generalization of the model. Having the sets distributed with a specific format means that the model is trained on a broad array of AE signal conditions, validated with unknown samples to enhance flexibility, and strictly tested to measure the effectiveness of the model to classify faults with varying conditions of the machine. Having the datasets distributed in the same format all the time will provide a proper comparison while increasing the robustness of the experimental results.

Baseline models

The following baseline models have been used for the evaluation of the proposed research work.

• Umar et al. (2024): this study presents a fault diagnosis technique for milling machines utilizing acoustic emission signals and a hybrid deep learning model optimized with a genetic algorithm. The approach achieves a fault classification accuracy of 92.6%, significantly outperforming traditional methods.

• Chen et al. (2024): this article introduces a deep LSTM-based fault detection method for railway vehicle suspensions. The method employs a goodness-of-fit criterion to quantify deviations between baseline models and newly monitored vibrations, demonstrating superior performance over traditional approaches.

• Fu, Wei & Yang (2024): this article builds a fault diagnosis model based on a parallel network of long short-term memory (LSTM) units and convolutional neural networks (CNN) to capture temporal and spatial features from vibration signals of bearings. The method improves feature extraction quality and robustness of the model.

• Sychev & Batako (2024): used wavelet transform to process audible acoustic emission signals, enabling precise identification of the start and end of frequency changes caused by surface roughness interaction.

Result

Experimental assessment of the developed AE Fault Classification Model was performed using three benchmark datasets, i.e., EAE-I, REB-II, and TAD-III. The experiments validate that the developed model greatly exceeds baseline approaches in terms of classification accuracy, fault detection precision, and robustness. Signal enhancement using WGAN-GP increases the quality of acoustic emission signals, resulting in improved feature preservation and better classification performance.

The outcome in Table 4 indicates that the mean square error (MSE) assessment was used to measure the quality improvement in the AE signal before and after enhancement using WGAN-GP-based. The outcomes indicate the MSE of raw AE signals was 0.025 for EAE-I, 0.032 for REB-II, and 0.028 for TAD-III, reflecting the high amount of noise present in the raw recordings. Following the WGAN-GP application for signal improvement, the MSE values decreased to 0.008, 0.011, and 0.009, respectively, which is an average of 67% improvement in all datasets. This substantial reduction confirms that the WGAN-GP model effectively removes unwanted noise while preserving critical fault-related features, improving fault classification accuracy in later stages. The performance improvement was consistent across all datasets, indicating the generalizability and robustness of the proposed GAN-based signal enhancement method.

From the other experiment, the value of the receiver operating characteristic (ROC) curve in Fig. 6 shows the performance of fault classification by the GAN-LSTM-based model applied and trained on the three data sets-EAE-I, REB-II, and TAD-III. The high ROC-area under the curve (AUC) scores signal that the model provides near-perfect fault detection through great true positive rate (TPR) and minimal false positive rate (FPR). Also noteworthy is the fact that this trend continues with an AUC value of 0.970-97.0% in the REB-II dataset, further proof of the strength of the model in coping with real-world AE signal variability. The EAE-I and TAD-III dataset is still classified strongly, with AUC values of 0.972 and 0.981, respectively, confirming the model’s capability of effective generalization over different fault scenarios. The clear separation from the random classifier baseline accentuates the superior discriminative power of the GAN-LSTM model in favor of a high reliability in the industrial fault diagnosis.

Figure 7 is the overall comparison of the proposed approach with the baselines (Umar et al., 2024; Chen et al., 2024; and Yenealem, 2025) on the datasets of EAE-I, REB-II, and TAD-III. The green bar is the proposed approach, while the red, yellow, and blue are the approaches of the baselines of the CNN (Umar et al., 2024), LSTM (Chen et al., 2024), and another LSTM (Yenealem, 2025), respectively. It is obvious that the proposed approach is the best with the highest accuracy on all the datasets, with the best performance in the classification of the faults. In specific terms: For the case of the dataset of EAE-I, the proposed approach is 94.4%, outperforming Umar et al. (2024) (88.6%), Chen et al. (2024) (89.5%), and Yenealem (2025) (90.1%). Similarly, for the case of the dataset of REB-II, the proposed approach is 97.0%, a much higher value compared to Umar et al. (2024) (90.2%), Chen et al. (2024) (91.8%), and Yenealem (2025) (93.0%). For the case of the dataset of TAD-III, the proposed approach is 96.2%, outperforming Umar et al. (2024) (92.3%), Chen et al. (2024) (93.0%), and Yenealem (2025) (94.2%). The substantial performance difference between the proposed approach and the baselines reflects the strength of the signal improvement of WGAN-GP that increases the signal quality of AE and retains the key features of the signal. The LSTM-based classification architecture also retains the sequence relationships of the AE signal to provide greater classification trustworthiness and fault localization accuracy.

The results demonstrate that the GAN-LSTM model is extremely generalized to other datasets with a considerable minimization of the classification error and the misdiagnosis of the AE-based fault detection systems. The performance improvement of REB-II (97%) and TAD-III (96.2%) also substantiates the strength of the model to other patterns of faults, levels of noises, and conditions of the machine. The analysis confirms that the proposed GAN-LSTM architecture is a superior solution to the conventional CNN and LSTM-based solutions to real-time industrial fault monitoring and predictive maintenance applications.

The proposed method was evaluated against the approach by Sychev & Batako (2024) across three datasets: EAE-I, EAE-II, and EAE-III. For EAE-I, Sychev & Batako (2024) method achieved a false alarm rate of 8.12% and a missed detection rate of 6.05%, whereas the proposed method reduced these values to 3.24% and 2.18%, respectively. On EAE-II, Sychev & Batako (2024) approach recorded a false alarm rate of 7.56% and a missed detection rate of 5.42%, while the proposed method achieved lower rates of 2.97% and 2.04%. Similarly, for EAE-III, Sychev & Batako (2024) method produced a false alarm rate of 6.87% and a missed detection rate of 4.95%, compared to the proposed method’s improved results of 2.63% and 1.92%. These findings clearly demonstrate the superior performance of the proposed method in reducing both false alarms and missed detections across all datasets.

The illustration in Fig. 8, comparably, visualizes the precision, recall, and F1-score of the presented GAN-LSTM model in comparison with baseline approaches (Umar et al., 2024; Chen et al., 2024 and Yenealem, 2025) over two datasets: EAE-I and REB-II. The findings validate the evident excellence of the introduced model, as it outperforms the others by consistently exhibiting better classification performance on all three primary metrics. On the EAE-I dataset, the introduced model decidedly performs better than baselines, illustrating its strength to learn intricate AE signal pattern properties and suppress misclassification mistakes. Likewise, for the REB-II dataset, the GAN-LSTM approach demonstrates even greater performance, signifying the capability of its resistance to capture and distinguish fault states from diverse AE signals. The greater recall values reflect that the suggested model identifies more real faults without omitting significant fault instances, and the high F1-score reflects a good balance between precision and recall. The clear discrimination between the suggested model and the baseline approaches verifies that GAN-based signal enhancement and LSTM-driven sequential learning have a significant benefit over traditional CNN and LSTM classifiers. These findings justify the efficacy and validity of the suggested method as a very dependable solution for AE-based fault detection and predictive maintenance.

Ablation study analysis

The ablation study presented in Table 5 comprehensively evaluates the contribution of each key component in the proposed framework, thereby validating its robustness and effectiveness. The baseline LSTM and CNN models, without any GAN-based enhancement, achieve accuracies of 89.5% and 88.6%, respectively, highlighting their limited ability to cope with AE signal noise. Introducing WGAN-GP without preprocessing raises the accuracy to 92.1%, confirming that GAN-based signal enhancement substantially improves fault classification. Incorporating STFT feature extraction further boosts accuracy to 94.0%, demonstrating the significance of frequency-domain features in capturing essential AE signal characteristics. Finally, the complete proposed configuration (LSTM + WGAN-GP + Full Preprocessing) achieves the highest performance, with 97.0% accuracy, precision of 0.98, recall of 0.97, and F1-score of 0.975. These results clearly indicate that each added component contributes meaningfully, and the integration of GAN-based enhancement with advanced preprocessing yields a robust and highly effective fault classification framework.