Fault Diagnosis of Power-Shift Systems in Agricultural Continuously Variable Transmissions Using Generative Adversarial Networks

Abstract

The power-shift system employed in agricultural multi-range continuously variable transmissions (CVTs) features a complex structure and control logic, presenting significant challenges to the reliability of agricultural machinery. To enable timely detection of faults, constructing an intelligent fault diagnosis classifier to monitor the system’s health status is essential. Typically, fault samples utilized for classifier development originate from ideal bench tests, characterized by uniform patterns and limited diversity, thereby hindering the algorithm’s generalization capability. This study addresses this issue by proposing a generative adversarial network (GAN) model, integrated with a triple loss function and a novel generator architecture, to augment the fault dataset under laboratory conditions. The generator architecture comprises a variational autoencoder module and an oil pressure point attention mechanism, enabling the generation of diverse and fluctuating virtual samples. Building on this augmented dataset, a fault classifier based on one-dimensional ConvNeXt was developed. Experimental results indicate that the classifier achieves an accuracy of 99.73%. While classifier accuracy decreases with increasing noise levels, the GAN-generated dataset provides more comprehensive training, resulting in an accuracy approximately 3% higher than that achieved using the original dataset.

1. Introduction

To meet diverse speed and load demands, modern agricultural vehicles are commonly fitted with multi-speed transmissions [1]. For example, the Foton Lovol P2654-N tractor offers up to 80 gears. Nevertheless, the proliferation of gears introduces challenges, including intricate gearbox designs, complicated gear selection, and diminished reliability. As a solution, continuously variable transmission (CVT) has emerged as the optimal transmission system for agricultural vehicles. Among CVT variants, hydraulic mechanical transmission (HMT) [2] has been the most advanced and commercially successful agricultural transmission over the past three decades, owing to its gearless operation, user-friendliness, and superior fuel efficiency. HMT integrates the high load-bearing capability of static hydrostatic transmission with the high efficiency of mechanical transmission, thereby achieving synergistic performance advantages [3]. However, HMT systems generally necessitate multiple working ranges to extend their speed spectrum and reduce energy loss. The power-shifting technology employed in these systems operates transparently to the driver but exhibits heightened sensitivity to faults due to its sophisticated control strategy [4]. Faults in the power-shift system, which depends on wet clutches [5], can be challenging to diagnose and repair promptly, resulting in operational delays and financial losses in agriculture. Consequently, there is an urgent requirement for health monitoring and fault diagnosis of HMT’s power-shift system to improve the reliability of agricultural vehicle operations.

The evolution of modern industrial fault diagnosis technology has progressed through three primary stages: traditional manual diagnosis, machine-learning-based diagnosis, and the current mainstream deep-learning-based approach. Traditional manual diagnosis depends on expert knowledge and basic instrumentation, characterized by high subjectivity, lack of standardization, and limited detection capacity. As industrial automation advanced, machine-learning-based diagnosis [6] emerged, which involves employing signal processing techniques to extract manually engineered features from raw data, followed by classification using algorithmic models. A notable example is the neural-network-based leak detection system developed by Wrat et al. [7], which illustrates the full workflow from data acquisition through feature extraction to model development and testing. Nevertheless, the performance of machine learning classifiers is heavily reliant on the quality of feature design and the application of expert knowledge [8]. Deep-learning-based fault diagnosis distinguishes itself through its end-to-end self-learning capability for feature extraction. It enables direct extraction of deep, high-dimensional features from raw signals, thereby substantially reducing manual intervention. The integration of attention mechanisms, exemplified by the Convolutional Block Attention Module (CBAM) [9], further elevates the feature extraction performance of deep learning models, which exhibit enhanced robustness under variable operating conditions and high noise levels. To mitigate the challenge of limited fault samples, generative adversarial networks (GANs) [10] and their variants, such as AttnGAN, are extensively employed to generate high-quality synthetic samples. Vibration source inversion technology also introduces novel methodologies for mechanical fault diagnosis. Furthermore, semi-supervised learning [11] utilizes unlabeled data to improve model generalization, while federated learning [12] addresses cross-institutional data isolation and privacy issues by enabling collaborative learning across multiple data sources. In conclusion, deep learning is emerging as the central driver of intelligent fault diagnosis, marked by autonomous feature learning, strong adaptability to complex working conditions, and superior performance in data-scarce and heterogeneous environments.

Current research on transmission system fault diagnosis primarily centers on gears and bearings. For instance, multi-scale convolutional attention networks enhance sensitivity to localized gear tooth defects [13], whereas deep residual convolutional networks [14] and transfer learning approaches [15] improve bearing diagnostic accuracy under noisy and variable operating conditions. However, fault diagnosis research for agricultural vehicle power-shift systems remains limited, largely due to the lack of fault data resulting from high development costs. Unlike traditional mechanical vibration signals, power-shift system faults typically manifest in hydraulic systems, rendering oil pressure an economical, efficient, and responsive diagnostic indicator. Our team has previously conducted valuable studies using oil pressure data collected from bench tests under ideal conditions; however, models trained on such data exhibit inadequate generalization capability. To address these challenges, this article proposes an innovative fault diagnosis method leveraging generative adversarial networks (GANs). This approach aims to analyze the structure of laboratory-derived oil pressure data and utilize GANs to generate more dynamic and diverse datasets, thereby supporting the development of more robust fault diagnosis systems.

This article is organized as follows: Section 1 provides a concise overview of the research background on fault diagnosis in HMT power-shift systems and identifies the limitations of existing studies. Section 2 details the methodology for sample data acquisition and outlines the architectures of generative adversarial networks and the proposed fault diagnosis algorithms. Section 3 presents the experimental results of the proposed algorithm across two datasets and evaluates its performance metrics. Section 4 concludes the paper with a comprehensive summary of findings.

2. Materials and Methods

2.1. Data Acquisition

Potential issues in the power-shift system include pipeline leaks (F1), pipeline blockages (F2), stuck solenoid valve cores (F3), and stuck clutch pistons (F4), with normal operation designated as F0, as detailed in Table 1. This study employs the mainline pressure of the HMT during gear shifting as sample data for fault diagnosis, which is simulated on a bench using radial-inlet clutches (0–1.03 MPa, 0–178 N·m). In this experiment, solenoid valves controlled clutch switching, and the shifting system’s mainline pressure was recorded at 100 Hz using a USB-6009 data acquisition card (8 AI ports, 2AO ports, and 13 DI/DO ports, National Instruments, Shanghai, China) and a DG2012-A-2 pressure sensor (0–10 V, 0–2 MPa, Senex, Shenzhen, China). Under LabVIEW (v2014) program control, the clutch maintains either an engaged or disengaged state for 3 s prior to switching states. After introducing faults, both clutch operation and data acquisition were fully automated via the program. A total of 30,000 experiments were performed, and 15,000 samples were randomly selected for analysis, ensuring equal sample sizes across F0 to F4 modes. Figure 1 illustrates the hydraulic circuit and mainline pressure curve of the power-shift system.

Table 1. Fault types in power-shift systems.

Label	Fault Description
F0	Normal mode.
F1	A leak at the clutch inlet can impact the clutch’s termination pressure; in severe instances, this may prevent the clutch from transmitting adequate torque. To simulate varying degrees of leakage, a pipeline connected to the oil tank is installed at the clutch entrance, with a throttle valve regulating the hydraulic oil flow within the pipeline.
F2	A blockage in the clutch drive pipeline can impact the movement time of the clutch piston, thus extending clutch engagement time. In severe cases, this may cause power interruption. To simulate varying degrees of blockage, a speed-regulating valve is installed in series along the drive pipeline. By adjusting the valve’s opening, the flow of hydraulic oil is restricted.
F3	A stuck solenoid valve core can lead to loss of control in the shifting system, with its effects on the vehicle varying based on the position of the stuck core. In serious instances, this issue may compromise driving safety or damage the transmission. To simulate such faults, a mechanical device can be used to prevent the valve core from resetting when the solenoid valve is open, keeping the core in the desired position.
F4	Similar to fault F3, a stuck clutch piston can also lead to loss of control in the shifting system, posing serious risks to driving safety or potentially burning out the clutch. When this fault occurs, the piston fails to move. To replicate the clutch pressure present during the fault, the clutch drive pipeline can be closed using a gate valve.

Figure 1. Hydraulic circuit for the power-shift system and corresponding mainline pressure curves during shifting.

2.2. Data Processing

To address the need for rapid fault diagnosis in power-shift systems, stable pressure data is excluded to minimize latency. Experimental results show that clutch pressure attains 90% of its rated value approximately 210 ms after a shift. With a sampling interval of 10 ms, the number of available data points does not exceed 22. Given the generator’s requirement for integer upsampling, the sample length is finalized at 20. Following data extraction, the dataset assumes the following shape:

$$\mathrm{Dataset}.\mathrm{shape} = (\mathrm{15,000}, 20)$$

(1)

To evaluate the overfitting of the fault diagnosis tool, the dataset should be divided into a training set and a testing set using an 8:2 ratio. After this partition, the dimensions of the training and testing sets are as follows:

$$\mathrm{TrainingSet}.\mathrm{shape} = (\mathrm{12,000}, 20)$$

(2)

$$\mathrm{TestSet}.\mathrm{shape}=(3000, 20)$$

(3)

Before inputting the data into the generator, downsampling the dataset is essential. For a dataset with a shape of (15,000, 20), only the indices of its second dimension (0, 4, 7, 19) are retained. As a result, the actual input dataset shape for the generator is:

$$\mathrm{TrainingSet}.\mathrm{shape} = (\mathrm{12,000}, 4)$$

(4)

$$\mathrm{TestSet}.\mathrm{shape}=(3000, 4)$$

(5)

The discriminator must effectively distinguish between fake and real samples, and the input dataset has the following shape:

$$\mathrm{TrainingSet}.\mathrm{shape} = (\mathrm{12,000}, 20)$$

(6)

$$\mathrm{TestSet}.\mathrm{shape}=(3000,20)$$

(7)

To fully utilize data information and assess the model’s generalization ability, we employ K-fold cross-validation for model tuning and testing in the experiment. The TrainingSet is divided into 5 equal folds (S1, S2, S3, S4, and S5), each containing 2400 data samples. In each iteration, 4 folds are used for training, and the remaining 1 fold serves as the validation set.

2.3. Algorithm Design

2.3.1. One-Dimensional ConvNeXt Module

Deep convolutional neural networks (DCNNs) serve as powerful tools for data-driven fault diagnosis. The ConvNeXt [16] model, which builds upon advancements from ResNet [17] and integrates design elements of Transformer architecture, as depicted in Figure 2, achieves performance comparable to state-of-the-art attention-based models. Its robust performance and contemporary architecture render ConvNeXt an ideal choice for processing one-dimensional time series data, such as oil pressure curves, and for developing fault classification systems.

Figure 2. Comparison of ConvNeXt and ResNet residual modules.

ConvNeXt was originally developed for image data. To adapt it effectively to one-dimensional temporal data and improve its performance in fault diagnosis tasks, especially within generative adversarial network (GAN) scenarios, we have implemented targeted modifications and optimizations to the ConvNeXt module. As depicted in Figure 3, the one-dimensional transformation of the ConvNeXt module involves the following adjustments:

Figure 3. Structure of one-dimensional ConvNeXt module.

(1)

Transform all convolutional kernels within the residual module from two-dimensional to one-dimensional.

(2)

To enhance the replication and classification performance of one-dimensional oil pressure curves, the kernel size of the depthwise separable convolutional layer was adjusted from 7 to 3. Consequently, this study employs a 1 × 3 convolution kernel.

(3)

To prevent damage to the generative adversarial network caused by normalization layers, all normalization layers should be removed, and the activation function should be changed from GELU to PReLU.

2.3.2. Discriminator and Fault Classifier

This study utilizes the previously described one-dimensional ConvNeXt module as the foundational component for developing fault classifiers and adversarial neural network discriminators, as depicted in Figure 4. The overarching architecture consists of multiple ConvNeXt stages, linked by downsampling layers, where each stage incorporates several ConvNeXt blocks. The downsampling layer employs a one-dimensional convolutional operation that reduces the number of data features by half while doubling the number of channels. Additionally, the Expand Channel Layer, also a one-dimensional convolutional operation, increases the number of data channels without modifying the count of data features.

Figure 4. Structure of discriminator and generator.

This study employs a fault diagnosis tool and an adversarial network discriminator, both structured with four one-dimensional ConvNeXt stages, where each stage contains two ConvNeXt modules. Contrary to the original ConvNeXt architecture, which uses a 1:1:3:1 hierarchical stacking ratio, extensive testing demonstrated that a 1:1:1:1 hierarchical stacking ratio achieves superior performance in fault classification for one-dimensional oil pressure data.

2.3.3. Generator Model

Generative adversarial networks (GANs) generate samples via adversarial training between the generator (G) and discriminator (D). Nevertheless, conventional GAN architectures exhibit limitations when handling complex time series data, such as oil pressure curves. Consequently, this study reengineered the generator architecture, structuring it into three distinct layers, input layer, computation layer, and output layer, as depicted in Figure 5.

Figure 5. Proposed generator model in this study.

(1) Input Layer. First, equidistant sampling is performed on the original data, followed by restoration to its original length via linear interpolation. The processed data is then split into two pathways; the first pathway is directly transmitted to the computational layer, while the second introduces the core innovation of this study—the variational autoencoder (VAE) [18] mechanism. Through two fully connected layers, the mean and variance of each feature are generated independently. The variance is scaled using random noise drawn from a normal distribution to synthesize new samples. These synthesized samples are further augmented with additional random noise along the channel dimension before being sent to the operational layer. This approach applies random variance scaling to the samples, creating a continuous distribution of input samples. This enhances the diversity of latent spatial inputs and improves the richness of the generated synthetic samples. The variational encoding process is represented as:

$$c_i = m_i+\mathit{exp}({\mathsf{\sigma }}_i)\times z_i$$

(8)

where $z_i~N(0, 1)$, $i$ is the i-th feature of the input data, $c_i$ is the i-th feature value output from the second path’s sample via the variational autoencoder, $c_i$ is the mean of the i-th feature obtained via a fully connected layer, ${\mathsf{\sigma }}_i$ is the variance of the i-th feature obtained via another fully connected layer, and $z_i$ is a random noise value sampled from a standard normal distribution.

(2) Computing Layer. Following the ConvNeXt stage in each computing layer, the two outputs from the input layer are merged and fed into the oil pressure point attention module to compute oil pressure point attention allocation. This process can be represented by the following formula:

$$Q ={ f}_1({\mathit{Input}}_1@W_1)$$

(9)

$$V = f_1({\mathit{Input}}_1@W_2)$$

(10)

$$[\mathit{Weight} = f_2({\mathit{Input}}_2.T@Q)]$$

(11)

$$[\mathit{Value} =V@\mathit{Weight}.T +V]$$

(12)

where $f_1$ is the PReLU activation function, $f_2$ is the softmax activation function, @ is the matrix multiplication operator, ${\mathit{Input}}_1$ and ${\mathit{Input}}_2$ are the outputs from the upper and lower paths of the ConvNeXt stage shown in Figure 5, respectively, and $W_1$ and $W_2$ are two distinct randomly initialized weight matrices.

Each column of matrices $V$, $Q$, and ${\mathit{Input}}_1$ represents the feature vector of a corresponding oil pressure point. For example, the $i$-th column of $V$ corresponds to the $i$-th oil pressure point. The upper input, ${\mathit{Input}}_1$, is processed through two separate fully connected layers to generate the query matrix ($Q$) and the value matrix ($V$). The query matrix undergoes matrix multiplication with ${\mathit{Input}}_2$, resulting in a dependency weight through the softmax function. Subsequently, $V$ is weighted accordingly.

This oil pressure point attention module is inspired by the word-level attention module Fattn from AttnGAN [19]. After adapting the attention mechanism to one-dimensional data, ${\mathit{Input}}_1$ generates $Q$ and $V$ through distinct fully connected layers. The query matrix interacts with ${\mathit{Input}}_2$ to produce a weight matrix, which is subsequently used to compute the final value. Unlike Fattn, which directly generates weights from ${\mathit{Input}}_1$ and ${\mathit{Input}}_2$ and applies them to ${\mathit{Input}}_1$, this design enhances the generator’s expressive capability. The significance of this mechanism lies in enabling each generated oil pressure point to account for the effects of all global pressure points. This provides the neural network with a global receptive field, thereby addressing the local receptive field limitation of the ConvNeXt stage.

(3) Output layer. The output layer consists of a ConvNeXt stage and a fully connected layer. The final generated samples are produced by passing the output from the preceding layer through these components.

2.3.4. Triple Loss Function

In the training of generative adversarial networks, the design of the loss function is critical to ensuring the quality, diversity, and training stability of the generated samples. Traditional GANs often employ binary cross-entropy classification loss, which tends to suffer from mode collapse and unstable training when handling complex data distributions. To address these issues, this study proposes a composite loss function that imposes constraints on the training of both generators and discriminators from multiple perspectives.

(1) Adversarial loss. This study employs RpGAN loss [20] instead of traditional GAN loss. This loss can directly quantify the “authenticity” of generated samples in comparison to real samples, rather than relying on binary classification. By doing so, it preserves distinct decision boundaries around each real sample and effectively mitigates the issue of pattern collapse. Its definition is as follows:

$${ \underset{D}{\mathit{max}}E}_{x~p_{\mathit{data}}(x)z~p_z(z)}[f(D(G(z))-D(x))]$$

(13)

where $f$ is the function operation.

RpGAN introduces a novel loss function that effectively integrates genuine and counterfeit data, allowing for a direct quantification of the authenticity level of forged samples relative to real samples, rather than relying on binary classification loss indirectly. This approach facilitates the maintenance of clear decision boundaries around each authentic sample, thereby reducing the risk of pattern collapse.

To ensure the convergence of the RpGAN loss function, we introduce zero-centered gradient penalties R₁ and R₂ for the discriminator.

$$R_1={\displaystyle \frac{\mathit{\gamma }}{2}}\mathit{E}({\Vert {\nabla }_{\mathit{x}}{\mathit{D}}_{\mathit{\psi }}\Vert }^2)$$

(14)

$$R_2={\displaystyle \frac{\mathit{\gamma }}{2}}\mathit{E}({\Vert {\nabla }_{\mathit{G}}(\mathit{z},{\mathit{x}}^{\prime }){\mathit{D}}_{\mathit{\psi }}\Vert }^2)$$

(15)

Thus, the adversarial losses for the generator and discriminator can be defined as:

$${\mathit{Loss}}_{G_1}={ \mathit{RpGAN}}_G$$

(16)

$${\mathit{Loss}}_{D }={\mathit{RpGAN}}_D+R_1+{ R}_2$$

(17)

$${\mathit{RpGAN}}_G=E(\ln (1+\mathit{exp}(D(G(z,x^{\prime }))-D(x))))$$

(18)

$${\mathit{RpGAN}}_D=E(\ln (1+\mathit{exp}(D(x)-D(G(z,x^{\prime })))))$$

(19)

where $x$ is the real sample, $x^{\prime }$ is the downsampled real sample, and $z$ is the random noise following a standard normal distribution.

(2) L1 loss is defined as the L1 norm between real samples and generated samples, expressed by the following formula:

$${ \mathit{Loss}}_{G_2}= L1\mathit{Loss} = E(|x-x_{\mathit{fake}}|)$$

(20)

$$x_{\mathit{fake}}=G(z,x^{\prime })$$

(21)

where $x_{\mathit{fake}}$ is the forged samples generated by the generator.

Introducing L1 loss to the generator helps address the limitations of adversarial loss when fitting the oil pressure curve. However, this method tends to average out the fake samples. To counteract this issue, we additionally integrate convolutional variance comparison loss.

(3) Convolutional variance loss. Figure 6 outlines the calculation of convolution variance for vectors.

Figure 6. Schematic diagram for calculating the convolution variance of vectors.

The process for calculating the convolution variance of a vector is denoted as g(v), and the formula for the comparison loss associated with convolution variance can be expressed as:

$${\mathit{Loss}}_{G_3}= \mathit{ConvVarLoss} = E({(g(x)-g(x_{\mathit{fake}}))}^2)$$

(22)

The loss is achieved by reducing the variance between each block of the fake sample and the corresponding real sample. This ensures that the fluctuations in the fake sample match those of the real sample, thereby addressing the averaging problem caused by L1 loss. The size of the convolution kernel is determined by the length and volatility of the studied samples. A tentative initial setting involves performing integer division of the sample’s characteristic length by 6. In this study, the selected convolution kernel length is 3.

(4) Total loss. The losses for the generator and discriminator can be defined as:

$${\mathit{LOSS}}_{D }= \mathit{RpGA}{N}_D·R_1 + R_2$$

(23)

$${\mathit{LOSS}}_G=\mathit{RpGA}{N}_G+{\mathit{Loss}}_{G_2}+{ \mathit{Loss}}_{G_3}$$

(24)

Finally, we employ an adversarial training approach to alternately optimize the discriminator and generator by iteratively minimizing ${\mathit{LOSS}}_D$ and ${\mathit{LOSS}}_G$.

2.4. Data Augmentation

The enhanced generator dynamically adjusts the volatility and diversity of the generated samples by modifying the noise intensity of the two inputs. The adjustment mechanism is illustrated in Figure 7. The generator’s input layer comprises adjustment regions “i” and “ii”, where $\alpha$ and $\beta$ serve as the scaling factors for noise in areas “i” and “ii”, respectively. During the training of a generative adversarial network, $\alpha$ is initialized to 1 and $\beta$ to 0. Once the generator is trained, it can generate more volatile and diverse samples by modifying $\alpha$ and $\beta$. In this study, the original training set was expanded by a factor of five, resulting in the following shape:

$$\mathrm{TrainingSet}.\mathrm{shape} = (\mathrm{60,000}, 20)$$

(25)

Figure 7. Schematic diagram for adjusting generator input.

In the training of GANs, no label information is incorporated, and distinct neural networks are trained for each fault category. Consequently, the generated samples avoid introducing label noise or bias.

3. Results and Discussion

3.1. Performance Analysis of Generator

This section examines how the generator’s parameter values influence the morphology of the generated samples. This study defines $\alpha$ within the range [1, 25] and $\beta$ within [0, 1]. When $\alpha$ and $\beta$ exceed these optimal ranges, the data produced by the generator becomes significantly distorted and loses its semantic integrity. Figure 8 illustrates the effect of varying $\alpha$ on the generation of fake samples, with $\beta$ held constant. The research findings indicate that an increase in $\alpha$ leads to greater volatility in the generated samples. Figure 9 illustrates the effect of varying $\beta$ on the generation of fake samples, with $\alpha$ held constant. The research findings indicate that an increase in $\beta$ enhances the diversity of the sample.

Figure 8. Effect of varying .$\alpha$ on generation of fake samples: (a) F0, $\alpha$ = 1; (b) F0, $\alpha$ = 6; (c) F0, $\alpha$ = 12; (d) F0, $\alpha$ = 20; (e) F0, $\alpha$ = 28; (f) F1, $\alpha$ = 1; (g) F1, $\alpha$ = 5; (h) F1, $\alpha$ = 10; (i) F1, $\alpha$ = 16; (j) F1, $\alpha$ = 20; (k) F2, $\alpha$ = 0; (l) F2, $\alpha$ = 1; (m) F2, $\alpha$ = 3; (n) F2, $\alpha$ = 5; (o) F2, $\alpha$ = 10; (p) F3, $\alpha$ = 1; (q) F3, $\alpha$ = 5; (r) F3, $\alpha$ = 10; (s) F3, $\alpha$ = 16; (t) F3, $\alpha$ = 20; (u) F4, $\alpha$ = 1; (v) F4, $\alpha$ = 8; (w) F4, $\alpha$ = 16; (x) F4, $\alpha$ = 20; (y) F4, $\alpha$ = 25.

Figure 9. Effect of varying $\beta$ on generation of fake samples: (a) F0, $\beta$ = 0; (b) F0, $\beta$ = 0.1; (c) F0, $\beta$ = 0.3; (d) F0, $\beta$ = 0.4; (e) F0, $\beta$ = 0.5; (f) F1, $\beta$ = 0; (g) F1, $\beta$ = 0.08; (h) F1, $\beta$ = 0.1; (i) F1, $\beta$ = 0.2; (j) F1, $\beta$ = 0.3; (k) F2, $\beta$ = 0; (l) F2, $\beta$ = 0.03; (m) F2, $\beta$ = 0.1; (n) F2, $\beta$ = 0.2; (o) F2, $\beta$ = 0.3; (p) F3, $\beta$ = 0; (q) F3, $\beta$ = 0.1; (r) F3, $\beta$ = 0.2; (s) F3, $\beta$ = 0.25; (t) F3, $\beta$ = 0.3; (u) F4, $\beta$ = 0; (v) F4, $\beta$ = 0.3; (w) F4, $\beta$ = 0.4; (x) F4, $\beta$ = 0.5; (y) F4, $\beta$ = 0.8.

3.2. K-Fold Cross-Validation

As outlined in Section 2.2, the TrainingSet is randomly partitioned into five subsets, labeled S1 through S5. The experimental protocol involves a series of five trials; in the first trial, S1 serves as the validation set, with the remaining data constituting the training set; in the second trial, S2 acts as the validation set, and the rest forms the training set, and this procedure is iterated until all five subsets have been used as validation sets. For each trial, the current training set undergoes data augmentation, as detailed in Section 2.4. Subsequently, the trained fault classifier is evaluated using the F1-score metric on both the original training set (prior to augmentation) and the corresponding validation set:

$$F1 = 2 \times Precision \times Recall/(Precision + Recall) \times 100\%$$

(26)

The experimental results are presented in Figure 10. Across five experiments, the model achieved an average F1-Score of 99.93% on the training set, with a sample variance of 5.34 × 10⁻³. On the validation set, the average F1-Score was 99.87%, accompanied by a sample variance of 8.06 × 10⁻³. The boxplot distributions of both sets exhibit similarity, and no outliers were observed in any of the five experiments. These findings indicate that the model maintains consistent performance across different subsets and demonstrates robust generalization capability.

Figure 10. Results of K-fold cross-validation: (a) line chart comparison; (b) box plot distribution.

3.3. Evaluation of Classification Results

To assess the noise tolerance of classifiers trained on the enhanced dataset, we trained two identical fault classifiers using both the enhanced and original training sets. Their performance was compared on the original training set and original test set, each overlaid with Gaussian noise of varying standard deviations (SD). Training utilized the Adam optimizer with 400 iterations and a learning rate of 0.009; the training process is depicted in Figure 11. The performance of the two trained models was evaluated on both the training and testing sets after applying Gaussian noise with zero mean and standard deviations of 0, 0.01, 0.02, 0.03, and 0.04. The results are presented in Figure 12.

Figure 11. Classifier training process varies based on different training sets: (a) augmented dataset; (b) original dataset.

Figure 12. Comparison of classification accuracy between two classifiers: (a) training sets; (b) testing sets.

In the noise-free test, both classifiers performed well, achieving accuracy rates of 99.73% and 99.96% for the classifiers trained on enhanced and raw data, respectively, on the test set. This indicates that, even in the absence of attention mechanisms, the classifiers can maintain high accuracy without overfitting. However, after adding Gaussian noise (standard deviation 0.02), the accuracy of the classifier trained on raw data dropped significantly to approximately 95.8%, while the classifier trained on enhanced data maintained an accuracy of 97.26%. When the noise standard deviation was increased to 0.03, the former’s accuracy dropped below 90%, while the latter remains above 92%. These results indicate that classifiers trained on enhanced datasets exhibit excellent noise adaptability.

The confusion matrix presented in Figure 13 further highlights that the performance of the classifier trained on the enhanced dataset outperforms the one trained on the original dataset. When the noise standard deviation is 0.04, the enhanced classifier achieves a diagnostic accuracy of 99% for F4, compared to 96.7% for the original classifier. In addition, the enhanced classifier improved the classification accuracy of F0, F1, F2, and F3 by 7.34%, 2.33%, 19.33%, and −12.67%, respectively. Although the accuracy of F3 has decreased, the overall performance of the enhanced classifier remains superior.

Figure 13. Comparison of confusion matrices for two classifiers on test set with varying levels of added noise: (a) augmented dataset, SD = 0; (b) original dataset, SD = 0; (c) augmented dataset, SD = 0.01; (d) original dataset, SD = 0.01; (e) augmented dataset, SD = 0.02; (f) original dataset, SD = 0.02; (g) augmented dataset, SD = 0.03; (h) original dataset, SD = 0.03; (i) augmented dataset, SD = 0.04; (j) original dataset, SD = 0.04.

3.4. Ablation Experiment

Beginning with only adversarial loss, we sequentially add L1 loss and convolutional variance loss to the generator’s loss function. Three distinct generators were trained, and the experiments outlined in Section 3.2 were replicated for each, yielding three separate fault classifiers. These classifiers were evaluated for classification accuracy across test sets with diverse noise levels. The findings of these ablation studies are presented in Table 2. The results demonstrate that the sequential addition of each loss term leads to a progressive improvement in classifier performance, thus validating the individual contributions of each loss term to the algorithm.

Table 2. Results of ablation experiment.

Standard Deviation of Normal Noise	RpGAN (%)	RpGAN + L1 Loss (%)	RpGAN + L1 Loss + Convolutional Variance Loss (%)
0.00	84.67	96.63	99.73
0.01	84.00	95.00	99.40
0.02	81.67	93.37	97.30
0.03	80.00	90.07	93.00
0.04	74.07	83.33	86.67

3.5. Adaptability Testing

To assess the generalization performance of the algorithm, we used a fault dataset of the power-shift system based on axial-inlet clutches (0–4 MPa, 0–2170 N·m) for testing. The testing platform is illustrated in Figure 14. Due to high testing costs, only 1045 sets of fault data were collected. The laboratory records indicate that the probability of seal ring damage in the system is substantially higher than the probability of solenoid valve spool jamming. Consequently, the fault is classified as F3, while other types of faults are consistent with those in power-shift systems that use radial-inlet clutches.

Figure 14. Test bench for HMT power-shift system utilizing axial-inlet clutches.

Gaussian noise with standard deviations ranging from 0 to 0.04 was introduced to both training and testing sets, and the performance of two classifiers was evaluated. The results are shown in Figure 15. Under noise-free conditions, both models exhibited excellently, with an accuracy of 100% for the original data classifier test set and 100% for the enhanced data classifier. As the noise levels increase, the performance gap between the two gradually becomes significant; when the noise standard deviation is 0.02, the accuracy of the original classifier drops to 96.97% in the training set and 98.89% in the test set, while the enhanced classifier still maintains 97.53% and 100%, respectively. In a high-noise environment of 0.04, the accuracy of the original classifier further reduced to 94.34% and 92.78%, while the enhanced classifier remained at 95.61% and 98.06%. These experimental findings demonstrate that the algorithm exhibits strong anti-interference capability even when applied to diverse datasets.

Figure 15. Comparison of classification accuracy between two classifiers on dataset of axial-inlet clutch faults: (a) training sets; (b) testing sets.

4. Conclusions

Given the significant costs and time investment of inducing faults in real environments, this study aims to develop a robust agricultural vehicle fault diagnosis system using laboratory training datasets. To accomplish this, a reconstructed generative adversarial network is employed to deeply learn the features of the training set, thereby generating more diverse and variable samples. These augmented samples are subsequently used to train a fault classifier. Consequently, a fault diagnosis system with significantly enhanced robustness has been successfully developed. The key contributions of this study are summarized as follows:

(1)

A fault classifier based on a one-dimensional ConvNeXt residual module was designed and implemented. Experimental results show that the classifier achieved a fault diagnosis accuracy of up to 99.73% on the dataset, without incorporating additional attention mechanisms. This significantly outperforms traditional CNN architectures, underscoring the strong potential of deep convolutional neural networks in diagnosing faults in HMT power-shift systems.

(2)

A comprehensive reconstruction of the traditional generative adversarial network architecture has been realized. The oil pressure point attention mechanism, introduced in the computational layer, effectively enhances key feature extraction and global contextual modeling. Integrating an autoencoder structure into the input layer expands the generator’s latent representation space, markedly improving output sample quality. The restructured generator exhibits highly flexible control; by adjusting the noise parameters $\alpha$ (for volatility) and $\beta$ (for diversity), it can generate fault data with varying degrees of fluctuation and pattern features, significantly enhancing the model’s ability to fit complex fault features.

(3)

A new triple loss function for generative adversarial networks has been proposed. Replacing the conventional adversarial loss with the R3GAN loss notably improves the efficiency of the adversarial training process and effectively mitigates mode collapse. Incorporating L1 distance loss in the generator addresses issues related to insufficient fitting capability, while the addition of convolutional variance contrast loss curbs the sample homogenization problem caused by L1 loss, thereby boosting the diversity of generated samples.

(4)

Test results on the fault dataset of the power-shift system with the radial-inlet clutch demonstrate that the GAN-enhanced classifier significantly improves resistance to noise interference. This advantage becomes more pronounced as noise levels increase compared to traditional methods. Comparable outcomes were observed when the GAN enhancement method was applied to the fault dataset of the power-shift system with the axial-inlet clutch. Notably, the two clutches exhibit significant differences in structure, operating pressure, and maximum transmitted torque; however, the proposed methods perform effectively across both, highlighting their strong adaptability and generalization capability.

This study successfully employed enhanced GANs to develop a wet clutch fault classifier. However, several limitations must be acknowledged: first, there is no standardized methodology for determining the generator parameters $\alpha$ and $\beta$, which currently rely on empirical adjustment; second, replicating realistic environmental noise in a laboratory setting presents substantial challenges, necessitating the use of Gaussian noise to assess the algorithm’s anti-interference capability; third, the axial-inlet clutch is not an independent subsystem, so fault data can only be acquired from HMT systems incorporating such clutches, leading to high testing costs and difficulties in obtaining adequate fault data; finally, this research focused solely on common power-shift system faults, whereas agricultural machinery encounters a broader range of fault types in practical operation. These factors collectively create a gap between the current algorithm and full commercial deployment. Future work should prioritize constructing a fault dataset under real-world application conditions during HMT industrialization to further validate and refine the proposed model and establish criteria for parameter determination.