Enhanced Fault Diagnosis of Drive-Fed Induction Motors Using a Multi-Scale Wide-Kernel CNN

Abstract

Induction motor (IM) drives are widely used in industrial applications, particularly within the renewable energy sector, owing to their fast dynamic response and robust performance. Reliable condition monitoring is essential to ensure uninterrupted operation, minimize unexpected downtime, and avoid associated financial losses. Although numerous studies have introduced advanced fault detection techniques for IMs, early fault identification remains a significant challenge, especially in systems powered by electronic drives. To address the limitations of manual feature extraction, deep learning methods, particularly conventional convolutional neural networks (CNNs), have emerged as promising tools for automated fault diagnosis. However, enhancing their capability to capture a broader spectrum of spatial features can further improve detection accuracy. This study presents a novel fault detection framework based on a multi-wide-kernel convolutional neural network (MWK-CNN) tailored for drive-fed induction motors. By integrating convolutional kernels of varying widths, the proposed architecture effectively captures both fine-grained details and large-scale patterns in the input signals, thereby enhancing its ability to distinguish between normal and faulty operating states. Electrical signals acquired from drive-fed IMs under diverse operating conditions were used to train and evaluate the MWK-CNN. Experimental results demonstrate that the proposed model exhibits heightened sensitivity to subtle fault signatures, leading to superior diagnostic accuracy and outperforming existing state-of-the-art approaches for fault detection in drive-fed IM systems.

1. Introduction

The induction motor (IM) is a fundamental component across many industries [1] due to its robust performance, high efficiency, and favorable power density. IMs are widely deployed in applications ranging from electric vehicles and aerospace drive systems to household appliances [2,3,4]. Nevertheless, components within IM drive systems are vulnerable to failure from elevated electro-thermal and mechanical stresses; such failures undermine system reliability and can cause costly unplanned downtime. Consequently, fault diagnosis (FD) in converter-fed IMs is essential for ensuring the dependable operation of variable speed drive (VSD) systems, and it has therefore attracted substantial research attention, particularly in the contexts of renewable energy systems and smart grids [5,6,7].

In inverter-driven IMs, stator windings are exposed to combined electromagnetic and mechanical stresses [8] that can precipitate insulation degradation and eventual breakdown if not detected early [9]. Undetected faults may produce irreversible damage to the motor and the wider drive system [10]. Typical manifestations include coil-to-coil, phase-to-phase, and phase-to-ground faults, among others [11].

The stringent performance and reliability demands of modern applications, such as electric vehicles and high-speed rail systems, have elevated fault detection in electric drive systems to a critical research priority [12]. Historically, FD methods for IMs have fallen into two broad categories: model-based and signal-based approaches [13]. Model-based techniques use mathematical models to reproduce both normal and faulty behavior, enabling theoretically grounded anomaly detection and precise residual analysis [14]. In contrast, signal-based methods, including vibration analysis, temperature monitoring, and motor current signature analysis (MCSA), are commonly used to identify bearing defects, misalignment, rotor imbalance, and other mechanical issues [15]. While effective for many applications, signal-based methods are often sensitive to noise, interference, and signal distortion, factors that can degrade diagnostic reliability in practice [16].

Infrared thermography [17] is a widely used technique for detecting abnormal thermal signatures. Similarly, MCSA methods, including Park’s vector [18], extended Park’s vector [19], and negative-sequence impedance [20], have proven effective in diagnosing drive faults in line-fed machines. However, these approaches face significant challenges when applied to IM drives, particularly in detecting early-stage interturn faults and distinguishing them from harmonics generated by the drive system. Despite extensive research on IM fault detection, diagnosing incipient faults and accurately assessing their severity in drive-fed systems remains a complex task. Since fault development is generally gradual [13], precise fault severity estimation is essential for timely maintenance planning and repair prioritization [21].

Drive-fed IMs are now integral to modern industrial operations. As line-fed motors are increasingly replaced by voltage-source inverter (VSI)-fed systems, ensuring operational reliability and enabling early fault detection have become critical for minimizing unplanned downtime and mitigating financial losses [22]. Timely identification of early-stage faults also supports preventive maintenance, reduces the risk of fault escalation, and safeguards the overall drive system [23].

Inverter-fed motors experience elevated electrical stress on stator windings due to the high harmonic content of the supply voltage, which increases thermal loading [24]. Consequently, effective condition monitoring systems must be capable of detecting faults regardless of the type of power supply. Fault diagnosis in converter-fed motors is particularly challenging because of electromagnetic interference (EMI), high switching frequencies, and rapid voltage transients inherent to inverter operation [25]. These factors complicate the distinction between fault-induced anomalies and normal operating patterns, underscoring the need for advanced diagnostic techniques. Despite the importance of this topic, it remains relatively underexplored, especially under variable speed and torque conditions.

Artificial intelligence (AI)-based approaches have shown strong potential for modeling complex systems without requiring extensive prior knowledge or explicit parameter definitions [26]. Such methods can automatically extract input features, making them particularly suitable for high-dimensional and nonlinear systems. However, conventional machine learning (ML) algorithms often struggle to process large datasets and typically rely on manual feature extraction, limiting their adaptability. In contrast, deep learning (DL), a subfield of ML, has recently achieved significant success in FD applications [27,28,29,30]. Hierarchical DL architectures are increasingly incorporated into condition monitoring frameworks to improve both diagnostic accuracy and automation [14,31,32].

This study explores the use of such architectures for the early detection of interturn faults in power-electronics-fed drive systems. Specifically, we propose a novel hybrid framework that combines ML and DL for two-level decision making, fault classification, and severity assessment.

The proposed method is a multi-wide-kernel convolutional neural network (MWK-CNN) designed for fault classification and severity estimation in IM drives. Unlike conventional CNNs, which typically employ small kernels (e.g., 3 × 3), the MWK-CNN uses larger convolutional filters to capture a broader range of contextual features in the input data [33]. Wide kernels offer several advantages, including an increased receptive field, enhanced shape bias, reduced computational complexity, and improved overall model performance. Despite these advantages, wide-kernel CNNs have seen limited adoption in FD applications. For instance, large-kernel CNNs have been applied to one-dimensional vibration signals, and wide-kernel deep convolutional autoencoders have been developed for diagnosing faults in rotating machinery using vibration data [33,34].

Building on this foundational work, we present a multi-size wide-kernel CNN architecture specifically optimized for improving diagnostic accuracy and robustness in electric drive systems. The main contributions of this study are:

• Development of a novel MWK-CNN architecture for FD in IM drives, incorporating multiple kernel sizes across convolutional layers.
• Effective feature extraction from input signals through wide-kernel convolutions, enabling accurate fault detection.
• Enhanced diagnostic accuracy through integration of diverse kernel sizes, outperforming conventional small-kernel CNNs.
• Robustness and generalizability, demonstrated through evaluation on an IM drive dataset collected under a variety of operating conditions.

The remainder of this paper is structured as follows. Section 2 presents an overview of electric drive systems and common fault types. Section 3 reviews the fundamentals of CNNs. Section 4 describes the proposed methodology in detail. Section 5 provides a comparative analysis and in-depth discussion of results. Finally, Section 6 summarizes the key findings and concludes the study.

2. Electrical Drive and Its Faults

Field-oriented (vector) control is a widely adopted strategy for regulating induction-motor drives, particularly in VSD systems [35]. Among its variants, speed-sensorless vector control has gained particular prominence in IM drive applications [36], offering high efficiency and precise speed regulation without relying on external sensors such as encoders. Additional benefits include rapid dynamic response, reduced system cost, enhanced reliability, and a more compact overall design [37].

Electrical-drive faults can take many forms, including overcurrent, overvoltage, communication errors, and failures within the drive’s power electronics or the connected motor [21]. Such faults commonly originate from power-supply disturbances, overload conditions, or malfunctions in the drive’s internal components or communication interfaces [13].

Common types of electrical-drive faults include:

• Overcurrent: Excessive current flow—typically caused by overloads, short circuits, or wiring faults—that can trip protective devices and damage power-stage components.
• Overvoltage: Elevated DC-bus or phase voltages resulting from supply transients, regenerative events, or converter malfunctions.
• Communication faults: Loss or corruption of communication between the drive and its controller or feedback devices (e.g., PLCs, encoders), often due to cable defects, configuration errors, or interface failures.
• Power-electronics faults: Open-circuit or short-circuit failures in switching devices (e.g., IGBTs/MOSFETs), gate-driver faults, or DC-link component failures that degrade converter performance.
• Motor faults: Electrical faults in the motor windings (open or short circuits, including interturn faults), insulation degradation, and phase imbalances that reduce torque capability and increase heating.
• Mechanical faults: Bearing defects, rotor eccentricity, misalignment, and shaft-related issues—typically revealed through vibration signatures and changes in electrical parameters.

These fault classes often interact (for example, a power-electronics fault can induce thermal stress that accelerates insulation failure), which complicates diagnosis and motivates integrated electrical–mechanical monitoring strategies.

3. Overview of the CNN

The effectiveness of a CNN for fault detection depends critically on both its architecture and the characteristics of the input data. Typical CNNs for time-series (one-dimensional) signals comprise convolutional layers (CLs), pooling layers (PLs), and fully connected layers (FCLs). In one-dimensional CLs, learnable filters (kernels) slide along the time axis with a specified stride and padding. At each position a filter computes the dot product between its weights and the local input segment, producing an activation that indicates how well the filter’s pattern matches the local data. Stacking multiple filters produces one-dimensional feature maps that capture local temporal patterns at different receptive fields.

Pooling layers follow CLs to reduce the spatial length of feature maps, thereby lowering computational cost, improving scalability, and reducing the risk of overfitting. For one-dimensional data, the most commonly used pooling operations are max pooling (selecting the maximum value in a sliding window) and average pooling (computing the window mean); both reduce dimensionality and can help limit overfitting. When FCLs are used for classification, a flattening layer converts the final set of feature maps (number of filters × feature length) into a one-dimensional vector that feeds the FCLs, which then learn higher-level combinations of features relevant to the target labels. Below is a concise summary of key CNN components and standard parameterizations used in fault-detection models.

• Normalization: Training deep networks can be significantly slowed by variations in input distributions across layers. Batch normalization (BN) addresses this issue by reducing internal covariate shift, thereby accelerating convergence [32]. Typically, the activation function is applied after the BN layer. During batch normalization, each input x is standardized by subtracting the mean and dividing by the standard deviation of the training data for each dimension k, ensuring stable input distributions for every layer:
  $$x^k={\displaystyle \frac{\left(x^k-E[x^k]\right)}{\sqrt{\left(Var[x^k]\right)}}}.$$

  (1)

  However, normalization can distort the layer’s representation, particularly in the presence of nonlinear transformations. To mitigate this, BN introduces learnable scale and shift parameters for each activation, enabling the network to preserve its expressive capacity:
  $$y^k=p^k{x}^k+q^k,$$

  (2)

  where p^k and q^k are trainable parameters that rescale and shift the normalized activations (x(k)). These parameters are learned jointly with the model weights during training.
• Activation Function: The ReLU activation function introduces nonlinearity after each convolutional block. It is widely favored for its computational efficiency and simplicity. A variant of ReLU with added Gaussian noise is defined as:
  $$Relu(x)=max\{0,x\}+N(o,\sigma (x)),$$

  (3)

  where N(o, σ(x)) is Gaussian noise with mean 0 and variance of σ(x), and x is the output of the convolution operation.
• Fully Connected Layer: The FCL precedes the final classification stage and integrates features extracted by the preceding convolutional and pooling layers. It operates analogously to a traditional multilayer perceptron. Mathematically, it can be expressed as:
  $$FL=f\left(w_l\ast x_l+b_l\right),$$

  (4)

  where x_l is the flattened input vector from the last pooling layer, w_l is the weight vector, b_l is the bias term, and f denotes the activation function. The dot product w_l ∗ x_l combines the learned weights with the input features to produce the final feature representation.
• Optimization: The Adam optimizer is a widely adopted stochastic optimization algorithm for training neural networks. It employs an adaptive learning rate mechanism that updates each parameter individually, improving training efficiency and convergence speed. Adam is particularly well-suited for models with a large number of parameters, as it combines low memory usage with high computational efficiency. It also performs robustly on noisy and non-stationary datasets, which are common in real-world applications [33]. The update process is described as:
  $$g_t={\nabla }^{\theta }{f}_t\left({\theta }_{t-1}\right),$$

  (5)

  $$m_t^{\theta }={\lambda }_1{V}_{t-1}^{\theta }+\left(g_t^2-{\lambda }_1{g}_t^2\right),$$

  (6)

  $$V_t^{\theta }={\lambda }_2{V}_{t-1}^{\theta }+(g_t^2-{\lambda }_2{g}_t^2),$$

  (7)

  $${\widehat{m}}^{\theta }={\displaystyle \frac{m_t^{\theta }}{1-{\lambda }_t^1}},$$

  (8)

  $${\widehat{v}}^{\theta }={\displaystyle \frac{v_t^{\theta }}{1-{\lambda }_t^2}},$$

  (9)

  $${\theta }_t={\theta }_{t-1}-\alpha {\displaystyle \frac{{\widehat{m}}^{\theta }}{\sqrt{{\widehat{v}}^{\theta }}+\epsilon }},$$

  (10)

  where g_t is the gradient of the stochastic objective function f(·) parameterized by θ_t, λ₁ and λ₂ are decay rates, ${\widehat{m}}^{\theta }$ and ${\widehat{v}}^{\theta }$ are bias-corrected estimates of the first and second moments. Parameters θ_t are updated using a corrected momentum approach, with a fixed learning rate α = 0.001 and a small constant ϵ = 10⁻⁹ for numerical stability.
• Loss Function: The categorical cross-entropy (CEL) loss function is commonly employed for classification tasks, particularly in multiclass problems. It measures the divergence between the true label distribution and the predicted probability distribution output by the model:
  $$L\left(y^{true},y^{pred}\right)=-sum(y^{true}\ast \mathrm{l}\mathrm{o}\mathrm{g}(y^{pred}))$$

  (11)

  where y^true and y^pred denote the actual and predicted probability distributions, respectively. CEL is both differentiable and convex, making it suitable for optimization with gradient-based algorithms.

4. Proposed Fault Diagnosis Method

Early fault diagnosis in electrical drives (EDFD) is essential to avoid disruptions in industrial operations. Electrical drive faults may manifest as overcurrent, overvoltage, communication errors, or failures in the drive’s power electronics or connected motor. Such faults can result from power supply disturbances, overload conditions, or issues within the drive’s communication systems or internal components. If undetected, these faults can lead to complete machinery shutdowns and significant production losses. The primary objective of this work is to develop an MWK-CNN for effective EDFD. Electrical signals are collected under various fault conditions and operational settings for subsequent analysis.

CNN models have been widely applied to FD and have demonstrated strong performance in EDFD. However, conventional CNN architectures typically employ small convolutional kernels for feature extraction, limiting their ability to accurately identify and differentiate between various electrical faults. To address this limitation, the proposed MWK-CNN incorporates wide convolution kernels, which expand the receptive field and allow the model to capture both high- and low-frequency features in vibration signals.

This enhanced architecture improves the network’s capacity to learn temporal characteristics inherent in the signals, which are often time-dependent and periodic. As a result, the proposed model achieves more effective feature extraction and higher diagnostic accuracy. A key structural innovation is the use of large convolution kernels in the initial layer, which plays a crucial role in preserving temporal features and enhancing diagnostic performance. The overall workflow of the proposed method is illustrated in Figure 1.

The first layer of the MWK-CNN employs multiple convolution kernels of different sizes, as shown in Figure 2. This multi-kernel design enables the capture of features across multiple frequency bands. The primary motivation for using wide kernels is to enhance the extraction of short-term signal characteristics. Compared to standard CNNs, the wide-kernel architecture offers advantages for EDFD, including improved global feature extraction, reduced overfitting, and more effective learning of relevant patterns.

One of the main strengths of CNNs is their ability to automatically identify informative diagnostic indicators while filtering out irrelevant data. The complete architecture of the MWK-CNN is depicted in Figure 2. The first convolutional layer uses three wide kernels with sizes 150, 75, and 50, respectively, each followed by max-pooling layers with pool sizes of 15, 10, and 5.

Because the first layer uses multiple kernel sizes, a concatenation layer merges their outputs before passing the data to subsequent layers. FCLs follow this concatenation stage. Both convolutional and fully connected layers use the ReLU activation function, while the final output layer applies the SoftMax function for classification. The Adam optimizer is used to train and fine-tune the MWK-CNN model.

5. Test-Setup, Results, and Discussion

5.1. Test-Setup

A comprehensive experimental study was performed using a laboratory prototype and a data-acquisition (DAQ) system equipped with three Hall-effect (LEM) current sensors and three voltage sensors. The test rig comprises a 5.5 kW induction motor mechanically coupled to an AC drive module with an integrated intelligent power element. Controlled experiments on a three-phase induction-motor drive system produced a dataset of 40,040 records under a variety of operating and fault conditions. The data acquisition process was conducted in a controlled laboratory setup, using Hall-effect current sensors, voltage sensors, a torque sensor, and an encoder for speed measurement. These instruments interfaced with a high-precision DAQ, enabling the capture of high-frequency time-series signals under both normal and fault conditions. The DAQ captured high-frequency time-series signals, including stator voltages, phase currents, torque, and rotor speed (measured with an encoder). A torque sensor was used to control and log the applied load during testing. The experimental block diagram is shown in Figure 3, and the motor specification is given in

Table 1. Motor Parameter Specification.

Motor Parameters
Power Rating	5.5 kW
Voltage Rating	415 V
No. of Poles	4
Rated Frequency	50 Hz
Stator Resistance	0.7767 Ω
Rotor Resistance	0.703 Ω
Stator Impedance	0.10773 H
Rotor Impedance	0.10773 H
Mutual Impedance	0.10322 H
J	0.22 kg·m2

. Signals were recorded across multiple load cases and fault scenarios: phase-to-ground short circuit (PTGF), phase-to-phase short circuit (PTPF), overloading (OLF), overvoltage (OVF), undervoltage (UVF), and normal operation (NOM). From these measurements, a labeled dataset was constructed and partitioned into six classes: OLF (0), OVF (1), PTGF (2), PTPF (3), UVF (4), and NOM (5), and then split into training, validation, and test subsets for model development and evaluation. Preprocessing steps included integrity checks and removal of missing values. The cleaned time series were segmented into fixed-length samples before being supplied to the MWK-CNN. Model development and evaluation were performed on a workstation running Microsoft Windows 11, equipped with an Intel i7-8750H processor and 16 GB of RAM. The MWK-CNN was trained using one-dimensional drive-signal samples corresponding to the six predefined operating conditions and validated on held-out data. The speed and current signal under normal and faulty conditions are shown in Figure 4, Figure 5, Figure 6 and Figure 7.

5.2. Feature Description and Extraction

The feature extraction process was performed using signals directly acquired from motor-mounted sensors and the drive controller under diverse operational scenarios, including both normal operation and multiple fault conditions (

Table 2. Experimental input scenarios and abbreviations.

Input and Output Scenario		Abbreviation
Phase Current (Amp)		Ia, Ib, Ic
Line-to-Line Voltage (V)		Vab
Torque (N·m)		Tn
Rated Speed (rad/s)		Speed
Rated Torque (N·m)		Torque
Fault Indicator	Normal	Nom (0)
	Overloading Fault	OLF (1)
	Over-Voltage Fault	OVF (2)
	Phase-to-Ground Fault	PTGF (3)
	Phase-to-Phase Fault	PTPF (4)
	Under Voltage Fault	UVF (5)

and

Table 3. Experimental scenarios, fault indicators, and dataset sizes.

Input Scenario	Fault Indicator	Values
Normal (Unfaulted)	Nom (0)	13,013
Overloading Fault	OLF (1)	7007
Over Voltage Fault	OVF (2)	7007
Phase-to-Ground Fault	PTGF (3)	5005
Phase-to-Phase Fault	PTPF (4)	5005
Under Voltage Fault	UVF (5)	3003

). For time-series modeling, the raw signals were first normalized and then segmented into fixed-length windows, allowing the capture of temporal dynamics. These segmented windows served as inputs to the deep learning framework, enabling it to learn both instantaneous characteristics and time-dependent patterns associated with fault evolution.

To maximize discriminative capability, the segmented signals were processed through convolutional and/or recurrent layers, which automatically learned higher-level feature representations. The use of a deep architecture minimized the need for manual feature engineering, allowing the model to extract informative and fault-relevant patterns directly from the raw inputs. The correlation heat map between the input variables and the target fault classes for the proposed method is presented in Figure 8.

5.3. Results

To evaluate the performance of the proposed MWK-CNN model, the electrical drive dataset was divided into three subsets: training, validation, and testing. The MWK-CNN architecture offers several advantages, including improved feature learning, reduced overfitting, and effective global feature extraction. The model was trained over multiple iterations, with performance assessed at each stage. Fine-tuning was applied to further enhance accuracy. The model’s fault detection (FD) performance was quantitatively assessed using several metrics: confusion matrix (CM), accuracy (Acc), precision (Pre), recall (Rec), and F1-score. The confusion matrix provides a structured summary of classification results, while the metrics quantify the model’s ability to correctly identify fault classes. The evaluation metrics are defined as follows:

$$Acc={\displaystyle \frac{TP+FP}{TP+TN+FP+FN}}$$

(12)

$$Pre={\displaystyle \frac{TP}{TP+FP}}$$

(13)

$$Rec={\displaystyle \frac{TP}{TP+FN}}$$

(14)

$$F1-Score={\displaystyle \frac{2 \ast Pre \ast Rec}{Pre+Rec}}$$

(15)

where TP denotes true positives (correctly identified positive cases), TN denotes true negatives (correctly identified negative cases), FP denotes false positives (incorrectly identified as positive), and FN denotes false negatives (incorrectly identified as negative).

The MWK-CNN model converged after approximately 50 training epochs. Figure 9 presents the training and validation accuracy curves, Figure 10 shows the loss curves, and Figure 11 depicts the training and validation error curves. These plots demonstrate that convergence was achieved within 50 epochs. The model attained an average classification accuracy exceeding 99%, reflecting its robust diagnostic capability. The confusion matrix in Figure 12 confirms that all fault conditions, NOM (0), OLF (1), OVF (2), PTGF (3), PTPF (4), and UVF (5), were correctly identified, yielding an average classification accuracy of 99.38%. Detailed per-class performance metrics, including precision, recall, and F1-score, are summarized in

Table 4. Performance metrics for the proposed MWK-CNN model.

State	Precision	Recall	F1-Score
0	0.89	0.96	0.92
1	1.0	1.0	1.0
2	1.0	0.99	1.0
3	0.94	0.91	0.93
4	0.94	0.85	0.90
5	1.0	0.99	0.99

.

5.4. Comparison with State-of-the-Art Models

This study compares the proposed MWK-CNN model with established deep learning approaches, 1D-CNN, long short-term memory (LSTM), and artificial neural network (ANN), for fault detection in IM drive systems. After training each model, predictions on the testing dataset were evaluated using three statistical metrics: mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²), assessed during both training and validation phases. The results are summarized in

Table 5. Performance comparison of the proposed model and benchmark models.

Model	RMSE	MAE	R2
ANN	0.2675	0.0715	0.7885
LSTM	0.3342	0.1117	0.8883
CNN	0.1802	0.0325	0.9575
Proposed	0.079	0.0062	0.9938

.

In fault detection performance, the LSTM model (RMSE: 0.3342, MAE: 0.1117, R²: 0.8883) underperformed compared to the CNN model (RMSE: 0.1802, MAE: 0.0325, R²: 0.9575). This observation is consistent with findings in [38], which note that while LSTM models excel in capturing long-term temporal dependencies in time-series data, they can be less effective than CNNs in extracting localized features essential for fault classification.

The proposed MWK-CNN achieved the best performance across all models, recording an RMSE of 0.079, MAE of 0.0062, and R² of 0.9938 for drive fault detection during testing and validation. These results are illustrated in Figure 9, Figure 10 and Figure 11, with further visual confirmation provided by spider plots in Figure 13a–c, comparing MAE, RMSE, and R² across all models.

The MWK-CNN yielded the lowest prediction errors and highest R² value, indicating superior diagnostic precision. In contrast, the ANN model exhibited the highest error rates and lowest R² value, making it the least effective among the compared methods.

The proposed MSWK-CNN also consistently outperformed all baseline models in terms of accuracy and F1-score, as shown in

Table 6. Performance comparison of the proposed model and benchmark models.

Model	Accuracy (%)	Precision	Recall	F1-Score
ANN	85.4	0.856	0.847	0.851
Traditional CNN	90.1	0.902	0.897	0.899
LSTM	91.7	0.919	0.913	0.916
ResNet-1D	93.2	0.933	0.931	0.932
TCN	93.5	0.935	0.934	0.934
MSWK-CNN (Proposed)	0.9938	0.96	0.96	0.963

. While ResNet-1D and TCN demonstrated strong performance, the extensive multi-scale wide kernel design of MSWK-CNN enabled more effective capture of both prolonged and transient fault characteristics. Moreover, it maintained competitive inference speed, making it suitable for real-time industrial applications. These results highlight the effectiveness of wide-core feature extraction, particularly for complex scenarios involving combined electrical and mechanical faults across varying time scales.

Wide kernel CNNs generally outperform conventional CNNs and other deep learning models in both image recognition and signal processing tasks. Larger convolutional kernels capture broader contextual and spatial features with fewer layers, enabling more efficient identification of large-scale patterns and dependencies. This larger receptive field is especially beneficial when local features alone are insufficient for accurate classification.

Compared with highly deep networks, wide kernel CNNs can be more computationally efficient and easier to train, as they avoid excessive depth while still modeling complex patterns and correlations. To validate the impact of kernel size selection, an ablation study was conducted by training model variants with different kernel configurations, as shown in

Table 7. Comparative study of the proposed method using different kernel sizes.

Configuration	Kernel Sizes Used	Test Accuracy (%)
Single-scale (large)	[150]	91.7
Single-scale (medium)	[75]	89.5
Single-scale (small)	[50]	87.9
Dual-scale	[75, 150]	93.1
Dual-scale	[50, 75]	91.0
Multi-scale (proposed)	[50, 75, 150]	0.9938

.

The results indicate that the multi-scale kernel configuration surpasses both single- and dual-scale approaches, confirming its ability to capture the diverse temporal dynamics of fault signatures. By combining large and small kernels, the model simultaneously extracts global patterns and fine-grained details, leading to stronger generalization and improved fault discrimination.

5.5. Robustness Testing Under Noisy Conditions

To assess the generalization capability of the proposed MSWK-CNN model under conditions representative of real-world industrial environments, additional experiments were performed by introducing artificial noise into the raw test signals. Two primary types of noise were injected:

• Gaussian White Noise: Simulates broadband electromagnetic interference from switching equipment and nearby devices. Added with signal-to-noise ratios (SNR) of 30 dB, 20 dB, and 10 dB.
• Pulse Interference (Impulse Noise): Models transient spikes caused by inverter switching and sensor malfunctions. Simulated by inserting short-duration, high-amplitude pulses at random intervals.

The comparative performance of the proposed model under varying noise conditions is presented in

Table 8. Performance comparison of the proposed model under different Gaussian noise levels and pulse noise intensities.

Noise Type	SNR or Intensity	Accuracy (%)	F1-Score
No Noise	—	0.9938	0.963
Gaussian Noise	30 dB	96.2	0.961
Gaussian Noise	20 dB	94.8	0.947
Gaussian Noise	10 dB	91.7	0.923
Pulse Noise (Moderate)	1% pulse rate	94.1	0.94
Pulse Noise (Severe)	3% pulse rate	91.9	0.914

.

The results indicate that while noise introduction inevitably reduces diagnostic performance, the MSWK-CNN maintains high accuracy and F1-score even under severe noise conditions. This robustness suggests strong potential for deployment in noisy industrial settings where sensor signals are frequently degraded by interference.

6. Conclusions

This study addresses the critical challenge of accurate drive fault detection by proposing a multi-scale wide-kernel convolutional neural network (MSWK-CNN) capable of directly processing raw drive signals, eliminating the need for manual feature extraction. The proposed model was benchmarked against established deep learning architectures, including 1D-CNN, LSTM, and ANN, using a purpose-built data acquisition system that collected drive signal data under both normal and faulty operating conditions. Fault types considered included PTGF, PTPF, OLF, OVF, UVF, and NOM. The expansive kernel architecture enhances the model’s ability to capture both local and global signal characteristics, thereby improving its capacity to distinguish between healthy and faulty conditions. The MSK-CNN, trained and validated on a comprehensive dataset collected from diverse operating scenarios, demonstrates heightened sensitivity to subtle fault signatures and delivers superior diagnostic performance. Moreover, the proposed approach shows strong potential for future applications in detecting additional drive-related issues, such as inverter pulse failures and gate-level defects. Future research should focus on improving the interpretability of wide-kernel CNN models, as a deeper understanding of their decision-making processes is essential for fostering user confidence and ensuring safe, reliable deployment in critical industrial environments.