Motor current fault diagnosis method based on de-noising diffusion probability model

Abstract

The invention discloses a motor current fault diagnosis method based on a de-noising diffusion probability model, and belongs to the technical field of mechanical equipment state monitoring and fault diagnosis, and the method comprises the steps: obtaining an original motor current signal sample, carrying out the wavelet transformation, obtaining a time-frequency grayscale image, and obtaining a time-frequency grayscale image; dividing the sample into a training sample used for diffusion model training and a test sample of a fault diagnosis model; constructing a diffusion model DDPM, carrying out training by adopting the training sample, and generating a pseudo sample based on the trained diffusion model; constructing a fault diagnosis model MAF-Cnet, and training the MAF-Cnet based on the training sample and the pseudo sample to obtain the trained MAF-Cnet; and inputting a test sample into the trained MAF-Cnet for diagnosis to obtain a motor current fault diagnosis result. The method solves the core problem that the generalization ability of the diagnosis model is insufficient due to scarcity of motor fault samples, improves the diagnosis accuracy, and is wider in application scene.

Description

Technical Field

[0001] This application relates to the field of mechanical equipment condition monitoring and fault diagnosis technology, and in particular to a method for diagnosing motor current faults based on a noise reduction diffusion probability model. Background Technology

[0002] While deep learning-based fault diagnosis methods reduce the reliance on manual feature engineering, their performance is highly dependent on a large amount of high-quality labeled data. In real-world industrial scenarios, fault data for specific equipment, especially severe fault data, is often difficult to obtain in large quantities, resulting in a scarcity of training samples. This makes the aforementioned deep models prone to overfitting and a decline in generalization ability.

[0003] To alleviate the problem of insufficient training data, data augmentation techniques are widely used, such as random oversampling, synthetic minority class oversampling, and their adaptive variants. However, traditional oversampling methods, which generate new samples through interpolation to balance the dataset, produce samples with insufficient diversity and may introduce noise, thus offering limited improvement to the performance of diagnostic models.

[0004] In recent years, generative adversarial networks (GANs) and their variants have been used for data augmentation. For example, GANs using auxiliary classifiers can generate more diverse fault samples. However, such generative models generally suffer from problems such as unstable training processes, difficulty in convergence, and pattern collapse, resulting in unsatisfactory quality and diversity of generated samples.

[0005] Therefore, there is an urgent need for a data augmentation method that can stably generate high-quality and diverse fault samples to improve the accuracy and robustness of fault diagnosis models under small sample conditions. Summary of the Invention

[0006] To address the aforementioned shortcomings in the existing technology, this application provides a motor current fault diagnosis method based on a denoising diffusion probability model, which solves the problems of low quality and limited types of existing detection data, as well as low detection accuracy of existing motor fault diagnosis models.

[0007] To achieve the aforementioned objectives, the technical solution adopted in this application is as follows: This application provides a motor current fault diagnosis method based on a denoising diffusion probability model, including: S1: Obtain the original motor current signal sample, perform wavelet transform on the original motor current signal sample to obtain the time-frequency grayscale image, and divide the time-frequency grayscale image into training samples for diffusion model training and test samples for fault diagnosis model. S2: Construct a diffusion model DDPM and train it using training samples. Generate pseudo samples based on the trained diffusion model DDPM. S3: Construct the fault diagnosis model MAF-Cnet, and train the fault diagnosis model MAF-Cnet based on training samples and pseudo samples to obtain the trained fault diagnosis model MAF-Cnet; S4: Input the test samples into the trained fault diagnosis model MAF-Cnet for diagnosis to obtain the motor current fault diagnosis results.

[0008] Further, S2 includes: S201: Randomly sample from the probability distribution of the standard Gaussian distribution to obtain the initial noise image; S202: The trained diffusion model DDPM is used to estimate the residual noise in the initial noisy image through the U-Net network, and the initial noisy image is reversed based on the estimated residual noise until the preset number of denoising steps are completed or the image features meet the preset sample quality requirements to obtain the fault sample. S203: Perform quality inspection on the fault samples, screen out the fault samples that meet the quality requirements, and obtain pseudo samples.

[0009] Furthermore, the fault diagnosis model MAF-Cnet includes: a 3×3 convolutional layer, a ReLU activation function, a multi-scale module, a max pooling layer, a CBAM attention module, a fully connected layer with a ReLU activation hidden layer, and an output layer. The multi-scale module includes four heterogeneous convolutional paths, namely, a dimensionality compression path, a concentrated scale capture path, a large scale capture path, and a multi-sensory vision fusion path.

[0010] Further, the training of the fault diagnosis model MAF-Cnet based on training samples and pseudo-samples to obtain the trained fault diagnosis model MAF-Cnet includes: A1: Unify the image specifications of training samples and pseudo samples, extract primary features from the unified training samples and pseudo samples through a 3×3 convolutional layer, and perform non-linear transformation through the ReLU activation function to obtain low-level features and obtain the primary feature map. A2: The primary feature map is processed through a multi-scale module, and four heterogeneous convolutional paths are executed in parallel to capture multi-granularity features; A3: The primary feature map and multi-granularity features are concatenated, and the concatenated features are compressed using a max pooling layer to obtain a compressed multi-scale feature map. A4: The compressed multi-scale feature map is calibrated using the CBAM attention module to obtain the calibrated feature map; The calibration process includes: channel attention generates weights through global pooling and fully connected layers to filter important channels; spatial attention generates weights through channel pooling and convolutional layers to locate key regions; the weights are multiplied by the feature map in sequence to output the calibrated feature map. A5: The calibrated feature map is subjected to secondary feature extraction through a 3×3 convolutional layer, and then the feature is compressed through a max pooling layer to obtain a high-level feature map. A6: The high-level feature map is flattened into a one-dimensional vector, and a fully connected layer with ReLU activation hidden layer is used for fault diagnosis. The output layer is then used to map the fault probability through softmax and output the maximum fault probability. A7: Calculate the loss for the maximum failure probability and the actual failure using cross-entropy loss, and update the parameters of the fault diagnosis model MAF-Cnet using backpropagation with the Adam optimizer to obtain the trained fault diagnosis model MAF-Cnet.

[0011] Further, S4 includes: S401: Based on the image specifications of the training samples and pseudo samples, standardize the image specifications of the test samples and perform preprocessing. S402: Input the preprocessed test samples into the trained fault diagnosis model MAF-Cnet for diagnosis, output the maximum fault probability, and obtain the preliminary diagnosis results; S403: Record preliminary diagnostic results and calculate overall accuracy and accuracy for various fault types.

[0012] The beneficial effects of this application are: This application provides a motor current fault diagnosis method based on a denoised diffusion probability model. It generates and acquires abnormal motor data by constructing a diffusion model (DDPM), and then processes and analyzes the data using a fault diagnosis model (MAF-Cnet) to obtain motor current fault diagnosis results under small sample conditions. This approach allows for the use of smaller samples, considers more comprehensive factors, and yields more scientifically weighted and higher-precision motor anomaly diagnosis results. It can be applied to various future motor fault detection scenarios, thus having a wider range of applications. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0014] Figure 1This is a flowchart illustrating a motor current fault diagnosis method based on a denoising diffusion probability model, provided in an embodiment of this application.

[0015] Figure 2 This is a schematic diagram of a fault diagnosis result based on data augmentation, provided as an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0017] Example 1: This application provides a motor current fault diagnosis method based on a denoising diffusion probability model. This method can be found in [reference needed]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a motor current fault diagnosis method based on a denoising diffusion probability model provided in this application, including: S1: Obtain a two-dimensional time series image of the current signal, and divide it into a dataset for training the diffusion model and a test set required for the diagnostic model by wavelet transforming the grayscale.

[0018] In one embodiment of this application, the processor acquires motor current data and obtains time-frequency grayscale images through continuous wavelet transform. These images are then divided into training samples for a diffusion model and test samples for a diagnostic module. The diffusion model is trained using the obtained training sample dataset. After training, samples are generated through the diffusion model's denoising process. Pseudo-samples and some real samples are mixed for training the diagnostic model. Through sample processing, more refined factors can be analyzed, resulting in more accurate evaluation results; image classification enables both sample training and diagnosis.

[0019] S2: Construct the diffusion model DDPM to augment the data, generate high-quality and diverse fault samples, and select fault samples to obtain enhanced pseudo samples.

[0020] Specifically, it includes: S201: Initial noise sampling: Random sampling is performed from a standard Gaussian probability distribution to obtain the initial noise image. ,in This is the preset maximum number of diffusion steps.

[0021] S202: Reverse Denoising and Sample Generation: Call the trained denoising diffusion model DDPM to generate samples from the initial noisy image. The reverse process (i.e., the reverse denoising process) is executed, and the reverse process is executed at most once. The model performs a step-by-step denoising operation. In each step, the model estimates the residual noise in the current noisy image using a U-Net network and denoises the image based on the residual noise. This denoising operation is repeated until a preset number of denoising steps are completed or the image features meet a preset sample quality requirement, resulting in the final fault sample. .

[0022] S203: Sample Screening and Output: This involves processing the generated fault samples. Quality checks are performed (e.g., screening by indicators such as image clarity and feature similarity), and samples that meet the quality requirements are output to expand the faulty samples and obtain pseudo samples.

[0023] S3: Design an end-to-end multi-scale attention fusion convolutional network diagnostic model (MAF-CNet) for motor current fault diagnosis, and train the fault diagnosis model MAF-Cnet using training samples and pseudo samples to obtain the trained MAF-Cnet.

[0024] In one embodiment of this application, the fault diagnosis model MAF-Cnet includes: a 3×3 convolutional layer, a ReLU activation function, a multi-scale module, a max pooling layer, a CBAM attention module, a fully connected layer with a ReLU activation hidden layer, and an output layer. The multi-scale module includes four heterogeneous convolutional paths, namely, a compressed dimensionality path, a captured concentrated scale path, a captured large scale path, and a fused multi-sensory vision path.

[0025] The training process includes: S301: Integrate two types of datasets into a training set: one is the original grayscale time-frequency image dataset of motor current faults used during diffusion model training, i.e., training samples; the other is the enhanced fault sample dataset generated by the S2 diffusion model, i.e., pseudo samples; and unify the image size to 64×64 pixels to ensure input standardization.

[0026] S302: The primary features are extracted using the 3×3 convolutional layer (zero-padding strategy) of the fault diagnosis model MAF-Cnet, and non-linear transformation is performed through the ReLU activation function to obtain low-level features such as edges and textures, generating a primary feature map.

[0027] S303: The primary feature map input multi-scale module passes through four heterogeneous convolutional paths in parallel (1×1 compression dimension, 3×3 capture medium-scale, 5×5 capture large-scale, 1×1+3×3+5×5 fusion multiple receptive fields); after the output feature map channels are concatenated, the dimension is compressed by the max pooling layer to retain key features, and the compressed multi-scale feature map is output.

[0028] S304: The compressed multi-scale feature map is input into the CBAM attention module for calibration: Channel attention generates weights through global pooling and fully connected layers to filter important channels; Spatial attention generates weights through channel pooling and convolutional layers to locate key regions; The weights are multiplied by the feature map in sequence to output the calibrated feature map.

[0029] S305: The calibrated feature map is extracted twice by 1-2 layers of 3×3 convolution (batch normalization after each convolution) to enhance the high-level semantic features related to the fault; then it is downsampled by the max pooling layer to compress the dimension and output the refined high-level feature map.

[0030] S306: The high-level feature map is flattened and converted into a one-dimensional vector. The input is a fully connected layer with 1-2 ReLU activated hidden layers. The output layer is mapped to five types of fault probabilities through softmax, and the one with the highest probability is the diagnosis result.

[0031] S307: The cross-entropy loss is used to calculate the loss of the maximum failure probability and the true failure. The Adam optimizer is selected, and all parameters are updated by backpropagation. The number of iterations and the learning rate decay strategy are set until the accuracy of the validation set (accounting for 10%-20% of the training set) is stable and the loss converges, resulting in a high-quality diagnostic model MAF-Cnet.

[0032] S4: Input the test samples into the trained MAF-Cnet for diagnosis to obtain motor current fault diagnosis results. The formulas for calculating the accuracy of various fault types are as follows: .

[0033] In the formula, Indicates the true category is Fault accuracy at that time Indicates the true category is The predicted category is The number of samples, Indicates all that do not belong to In the sample, the prediction does not belong to the category. The number of samples.

[0034] In one embodiment of this application, the test sample includes raw motor current fault data or normal grayscale time-frequency images that were not used in training, avoiding overlap with the training set. The images are standardized to 64×64 pixels according to the image specifications of training samples and pseudo-samples to match the model input specifications. The test samples undergo preprocessing consistent with the training set (e.g., grayscale normalization to [0,1]) to eliminate data distribution differences, remove blurry or feature-deficient invalid samples, and retain valid samples. The trained MAF-CNet model is loaded: the network configuration and convergence parameters (including convolutional layer, CBAM, and fully connected layer parameters) are read, and the inference environment is initialized (e.g., specifying GPU / CPU) to ensure model inference capability and parameter integrity. The preprocessed test samples are batch-input into the loaded MAF-CNet. The model automatically executes the diagnostic process, outputs five types of fault probabilities, determines the diagnostic category (including normal) based on the maximum probability, and generates preliminary results. The model's diagnostic performance is validated based on the result record table, calculating core performance indicators: overall accuracy (number of correct answers / total number of samples) and accuracy for each type of fault.

[0035] Example 2: To verify the accuracy and stability of this invention, a method for accuracy testing based on a motor dataset from the University of São Paulo laboratory in Brazil is proposed. The main tests consist of the following steps: Step 1: Data presentation, including: Step 101: For the data obtained from the experiment, the experimental setup is located within the campus of the Engineering College of São Paulo (EESC) at the University of São Paulo (USP), Brazil, specifically in the Laboratory of Intelligent Automation of Processes and Systems (LAIPS) and the Laboratory of Intelligent Control of Motors (LACIME).

[0036] The three-phase induction motor is the WEG manufacturer's W22 standard line model, 1 hp, 220V / 380V, 3.02A / 1.75A, 4 poles, 60Hz, rated torque of 4.1 Nm, and rated speed of 1715 rpm. The rotor is a squirrel-cage type, consisting of 34 bars. It is driven via a control panel that allows selection of drive type (star or delta) and power supply type (direct grid voltage or via a three-phase inverter). The rotary torque wrench used in this application is a Transtec model MT-103, with a maximum speed of 2000 rpm. Its main function is to allow visualization of the torque present on the shaft, which will vary to simulate various operating conditions of the induction motor. The resistance torque is manually adjusted by changing the field winding voltage of the DC generator. To reduce the amplitude of the grid voltage, a 1800W single-phase voltage change is performed using a Variac, and a single-phase rectifier is used to power the field winding, converting the AC voltage to DC voltage.

[0037] The vibration sensor used is a Vibrocontrol uniaxial accelerometer, model PU 2001, with a sensitivity of 10 mV / mm / s and a frequency range of 5 to 2000 Hz. It can provide an integrated acceleration signal over time, i.e., provide a measurement of vibration velocity. A total of five accelerometers are used simultaneously, located on the non-drive side of the motor, the drive side, the housing, the motor axis, and the support table. These monitoring points allow for the measurement of axial, tangential, and radial velocities.

[0038] Current was measured using an AC current probe with a capacity up to 50A RMS and an output voltage of 10mV / A, corresponding to the Yokogawa model 96033. Voltage was measured directly at the MIT terminal using an oscilloscope voltage tip from the manufacturer, Yokogawa.

[0039] Step 102: To simulate a broken bar fault in the squirrel-cage rotor of a three-phase induction motor, it is necessary to drill a hole in the rotor. The drilling is performed using a bench drill with a 6mm diameter drill bit to ensure that the hole diameter exceeds the width of the rotor bar and that the tip is located at half the longitudinal length of the rotor.

[0040] First, a rotor without holes, i.e., a healthy rotor, was tested, and then the rotors were replaced sequentially to obtain a database of monitored variables. Experiments were conducted using the aforementioned experimental setup to build the database. Direct start-up tests were performed on both healthy and defective motors, balancing the three-phase supply voltage and 60Hz frequency.

[0041] To prepare a reliable database for testing, loads of 0.5 Nm, 1.0 Nm, 1.5 Nm, 2.0 Nm, 2.5 Nm, 3.0 Nm, 3.5 Nm, and 4.0 Nm were applied to the shaft of the three-phase induction motor. Ten repetitions were performed for each load condition on the motor shaft. Using a data acquisition system, the following variables were collected for each experiment under each load: voltages at phases A, B, and C; currents at phases A, B, and C; and mechanical vibration velocities in the tangential direction of the housing, the tangential direction of the base, the axial direction of the drive side, the radial direction of the drive side, and the radial direction of the non-drive side.

[0042] This experimental procedure is used to detect and diagnose faults in healthy motors and rotor motors containing 1, 2, 3, and 4 adjacent broken bars. The database is organized into a Matlab application structure, where the "struct_rs_R1" structure presents experimental data for defect-free induction motors, "struct_r1b_R1" corresponds to a rotor with one broken bar, "struct_r2b_R1" corresponds to a rotor with two broken bars, "struct_r3b_R1" corresponds to a rotor with three broken bars, and "Struct_r4b_R1" corresponds to a rotor with four broken bars.

[0043] Step 2: Experimental testing, including: Step 201: Generate Quality Comparison: Experimentally compare the generation algorithms. Use the generation models GAN, WGAN-GP, DCGAN, and WGAN to generate 100 data points, and then compare them with the original data.

[0044] The experiment selected two evaluation metrics—Moving Distance (EMD) and Structural Similarity Index (SSIM)—to rigorously evaluate the quality of the generated samples by comparing them with real samples. The definitions and functions of each metric are as follows: (1) EMD measures the similarity of the feature distributions between two samples, and is suitable for measuring the similarity of two samples from a global distribution perspective; the smaller the EMD, the closer the feature distributions of the two samples are. (2) SSIM measures the similarity of the data structure between two samples, and is suitable for measuring data similarity from a more detailed local structure comparison perspective; the closer the SSIM is to 1, the more similar the two samples are considered to be.

[0045] Table 1. Comparison of EMD and Structural Similarity Index among the various models.

[0046] In the quantitative evaluation of image generation quality, this application selected two key indicators, EMD and SSIM, to systematically compare the performance of Deep Diffusion Probabilistic Model (DDPM), Wasserstein Generative Adversarial Network (WGAN), Deep Convolutional Generative Adversarial Network (DCGAN), and Gradient Penalized Wasserstein Generative Adversarial Network (WGAN-GP) from two dimensions: global feature distribution and local structural consistency. The analysis results show that DDPM significantly outperforms the other methods in both indicators. Its EMD value is the lowest (0.0006), indicating that the feature distribution of its generated samples is closest to that of the real samples; simultaneously, its SSIM value is the highest (0.5090), reflecting the highest similarity of the generated images in terms of local structural details. In contrast, the EMD values ​​of WGAN, DCGAN, and WGAN-GP are 0.0017, 0.002, and 0.0018, respectively, all higher than DDPM, indicating a larger global distribution bias in the samples generated by these models. Their SSIM values ​​are 0.2706, 0.3114, and 0.3214, respectively, significantly lower than DDPM, further confirming their limitations in reconstructing fine structures. In summary, DDPM demonstrates a clear advantage in the overall quality and structural fidelity of the generated images, while adversarial training-based methods (especially WGAN-GP), although slightly better than standard WGAN and DCGAN, still do not reach the performance level of diffusion models.

[0047] Step 202: Experimental comparison of generation algorithms: 600 data points were generated using GAN, WGAN-GP, DCGAN and WGAN, and mixed with the real data used for training the generation model to form a new dataset, which was used to train the diagnostic model and obtain the classification accuracy.

[0048] Table 2 Comparison of Data Generated by Each Model

[0049] like Figure 2 As shown, in the fault diagnosis experiment based on data augmentation, the performance of different generation algorithms under the same diagnostic model was further evaluated. Through comparison of diagnostic accuracy, it was found that the data augmentation strategy based on DDPM significantly outperformed other adversarial generation methods. Specifically, when trained with 300 generated samples, DDPM achieved a diagnostic accuracy of 0.702, while DCGAN, WGAN, and WGAN-GP only reached 0.475, 0.507, and 0.546, respectively. When the number of generated samples increased to 600, the diagnostic performance of all methods improved, with DDPM showing the most outstanding performance, achieving an accuracy of 0.96, far exceeding WGAN-GP's 0.837, WGAN's 0.776, and DCGAN's 0.706. This indicates that the high-quality images generated by DDPM can more effectively improve the generalization ability of the diagnostic model. Furthermore, the confusion matrix results demonstrate that the algorithm proposed in this application exhibits good recall and precision across most categories. The values ​​of many main diagonal units exceed 0.94, and some even reach 0.98, reflecting the high reliability and stability of the overall diagnostic results, further verifying the practicality and superiority of the method of the present invention in fault diagnosis tasks.

[0050] This application addresses the core problem of insufficient generalization ability of diagnostic models due to the scarcity of motor fault samples in industrial practice. It proposes a fault diagnosis method combining data augmentation based on a denoised diffusion probability model and a CBAM attention mechanism diagnostic model. Firstly, in terms of data generation quality, DDPM outperforms traditional generative models such as WGAN and DCGAN on both EMD and SSIM evaluation metrics, meaning it generates higher-quality, more effective images that significantly improve diagnostic accuracy, with a significantly higher accuracy than other generation methods. In comparisons of different diagnostic models, the CBAM diagnostic model combined with DDPM data augmentation proposed in this application demonstrates the best performance. Therefore, this method allows for the use of smaller samples, consideration of more comprehensive factors, and the acquisition of more scientifically weighted and higher-precision motor anomaly diagnostic results. It can be applied to various future motor fault detection scenarios, thus having a wider range of applications.

[0051] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.

Claims

1. A method for diagnosing motor current faults based on a denoised diffusion probability model, characterized in that, include: S1: Obtain the original motor current signal sample, perform wavelet transform on the original motor current signal sample to obtain the time-frequency grayscale image, and divide the time-frequency grayscale image into training samples for diffusion model training and test samples for fault diagnosis model. S2: Construct a diffusion model DDPM and train it using training samples. Generate pseudo samples based on the trained diffusion model DDPM. S3: Construct the fault diagnosis model MAF-Cnet, and train the fault diagnosis model MAF-Cnet based on training samples and pseudo samples to obtain the trained fault diagnosis model MAF-Cnet; S4: Input the test samples into the trained fault diagnosis model MAF-Cnet for diagnosis to obtain the motor current fault diagnosis results.

2. The motor current fault diagnosis method based on a denoising diffusion probability model according to claim 1, characterized in that, The S2 includes: S201: Randomly sample from the probability distribution of the standard Gaussian distribution to obtain the initial noise image; S202: The trained diffusion model DDPM is used to estimate the residual noise in the initial noisy image through the U-Net network, and the initial noisy image is reversed based on the estimated residual noise until the preset number of denoising steps are completed or the image features meet the preset sample quality requirements to obtain the fault sample. S203: Perform quality inspection on the fault samples, screen out the fault samples that meet the quality requirements, and obtain pseudo samples.

3. The motor current fault diagnosis method based on a denoised diffusion probability model according to claim 1, characterized in that, The fault diagnosis model MAF-Cnet includes: a 3×3 convolutional layer, a ReLU activation function, a multi-scale module, a max pooling layer, a CBAM attention module, a fully connected layer with a ReLU activation hidden layer, and an output layer. The multi-scale module includes four heterogeneous convolutional paths: a compressed dimensionality path, a captured concentrated scale path, a captured large scale path, and a fused multi-sensory field of view path.

4. The motor current fault diagnosis method based on a denoising diffusion probability model according to claim 3, characterized in that, The process of training the fault diagnosis model MAF-Cnet based on training samples and pseudo-samples to obtain the trained fault diagnosis model MAF-Cnet includes: A1: Unify the image specifications of training samples and pseudo samples, extract primary features from the unified training samples and pseudo samples through a 3×3 convolutional layer, and perform non-linear transformation through the ReLU activation function to obtain low-level features and obtain the primary feature map. A2: The primary feature map is processed through a multi-scale module, and four heterogeneous convolutional paths are executed in parallel to capture multi-granularity features; A3: The primary feature map and multi-granularity features are concatenated, and the concatenated features are compressed using a max pooling layer to obtain a compressed multi-scale feature map. A4: The compressed multi-scale feature map is calibrated using the CBAM attention module to obtain the calibrated feature map; The calibration process includes: channel attention generates weights through global pooling and fully connected layers to filter important channels; spatial attention generates weights through channel pooling and convolutional layers to locate key regions; the weights are multiplied by the feature map in sequence to output the calibrated feature map. A5: The calibrated feature map is subjected to secondary feature extraction through a 3×3 convolutional layer, and then the feature is compressed through a max pooling layer to obtain a high-level feature map. A6: The high-level feature map is flattened into a one-dimensional vector, and a fully connected layer with ReLU activation hidden layer is used for fault diagnosis. The output layer is then used to map the fault probability through softmax and output the maximum fault probability. A7: Calculate the loss for the maximum failure probability and the actual failure using cross-entropy loss, and update the parameters of the fault diagnosis model MAF-Cnet using backpropagation with the Adam optimizer to obtain the trained fault diagnosis model MAF-Cnet.

5. The motor current fault diagnosis method based on a denoising diffusion probability model according to claim 4, characterized in that, The S4 includes: S401: Based on the image specifications of the training samples and pseudo samples, standardize the image specifications of the test samples and perform preprocessing. S402: Input the preprocessed test samples into the trained fault diagnosis model MAF-Cnet for diagnosis, output the maximum fault probability, and obtain the preliminary diagnosis results; S403: Record preliminary diagnostic results and calculate overall accuracy and accuracy for various fault types.

Images