A motor fault diagnosis method based on self-supervised mask contrast learning

Abstract

The application discloses a motor fault diagnosis method based on self-supervised mask contrast learning, and the method comprises the following steps: collecting motor fault vibration signals and constructing a fault diagnosis data set; converting time series signals into time-frequency images; pre-training the images by using a mask autoencoder and extracting features; distinguishing different fault types in a feature space through a contrast learning mechanism; and realizing fault type identification by using a classifier. The application combines mask learning and contrast learning, improves the expression ability of fault features, and significantly improves the accuracy and robustness of motor fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis, and specifically relates to a motor fault diagnosis method that combines mask learning and contrastive learning. Background Technology

[0002] With the continuous advancement and development of technology, the automation level of intelligent equipment is becoming increasingly high, and production processes are becoming more and more complex. However, due to the prolonged high-intensity operation of equipment, various malfunctions are prone to occur, leading to serious consequences. Equipment diagnostic technology has emerged in response to the urgent need for equipment safety and maintenance. Currently, there are still some problems with motor equipment diagnostic technology. First, the various technologies for fault detection and diagnosis of mechanical equipment lack a complete theoretical system. Therefore, in practice, they often only address specific faults that may occur in specific equipment, thus preventing the widespread application of current detection and diagnostic technologies. Second, current mechanical equipment fault detection and diagnosis technologies have poor anti-interference capabilities, and their accuracy needs to be improved.

[0003] Traditional fault diagnosis algorithms require processing large amounts of complex, high-dimensional measurement information, resulting in significant computational overhead and modeling complexity. This makes it impossible to extract all the features of the data and difficult to build accurate and effective mathematical models. With the rapid development of deep learning, its powerful data learning and analysis capabilities have been introduced into fault diagnosis systems, providing a completely new approach to the field.

[0004] Self-supervised learning can directly and effectively characterize massive, multi-source, and high-dimensional data in motor faults. Among them, self-supervised feature representation learning modes based on mask learning and contrastive learning each have their own advantages and disadvantages. In view of these advantages and disadvantages, this invention aims to research and develop a motor fault diagnosis system based on self-supervised features and representations, combining mask learning and contrastive learning to achieve a new breakthrough in deep learning in fault diagnosis technology. Summary of the Invention

[0005] To address the shortcomings of current motor fault diagnosis technologies, this invention proposes a motor fault diagnosis method based on self-supervised comparative learning, which solves the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A motor fault diagnosis method based on self-supervised contrastive learning includes the following steps:

[0008] (1) The collected bearing data, measured motor bearing fault data and test bench motor data from Xichuan University were used as fault data. Time-frequency analysis was performed on the data to obtain its time series. Then, wavelet transform was performed to obtain time-frequency images, i.e., the input dataset in image form.

[0009] (2) Multiple data augmentation methods were applied to the samples and combined to obtain an effective representation of the fault characteristics: cropping, resizing, horizontal flipping, rotation, and erasing methods were selected to perform spatial / geometric transformations on the data; color distortion, Gaussian blur, and Sobel filtering methods were selected to perform shape transformations on the data.

[0010] (3) Preprocess the augmented data, i.e., mask image segmentation, and set the segmentation overlap rate according to the size of the RGB values ​​in the neighborhood. Set a high segmentation overlap rate for areas with large RGB values ​​and a low segmentation overlap rate for areas with small RGB values.

[0011] (4) Input the segmented mask patch into the mask autoencoder for pre-training. The goal of the pre-trained decoder is to restore the missing pixels in the image. L2 loss is selected as the pre-training loss function.

[0012] (5) Divide the obtained image matrix data into training fault data and test fault data;

[0013] (6) Comparative self-supervised learning is adopted for training fault data. The input image and its flipped, cropped and erased samples are used as positive samples, and the distorted samples of the input image and other samples in the training set are used as negative samples for comparative learning. The test data is compared with the positive and negative samples in the feature space to learn the feature representation of the samples.

[0014] (7) The extracted features are integrated and normalized through a fully connected layer, and the obtained probabilities are classified by a softmax classifier. Cross-entropy is selected as the loss function for the classification task.

[0015] (8) Vit-base is used as the fine-tuning model, containing 12 encoder blocks labeled from 0 to 11, with a patch size of 16×16. The number of layers whose weights are allowed to be changed during the model's backward iteration is specified, while the remaining layer structure is frozen, forcing its parameters to be locked, which greatly reduces the number of parameters, thereby realizing the update of parameter weights;

[0016] (9) Verify the test fault data. By adjusting the hyperparameters, when the test accuracy meets the requirements, it can be used to diagnose motor fault data and realize fault diagnosis of self-supervised domain comparative learning.

[0017] Furthermore, in step (1), continuous wavelet transform is performed on the time-series signal to obtain a two-dimensional time-frequency image that can simultaneously reflect time-frequency information, which is used for subsequent feature extraction.

[0018] Furthermore, the expression for the loss function in step (4) is as follows:

[0019]

[0020] Where n represents the number of samples, y i Let y represent the target variable. i p This represents the predicted value.

[0021] Furthermore, the expression for the classification task loss function in step (7) is as follows:

[0022]

[0023] In the formula p(x i ) represents the true probability, q(x) i ) represents the predicted probability. Attached Figure Description

[0024] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided:

[0025] Figure 1 It is a two-dimensional image after wavelet transform of the four categories of labels for 2hp load;

[0026] Figure 2 It is a pre-training process framework;

[0027] Figure 3 This is the overall flowchart of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] The motor fault diagnosis method based on self-supervised mask contrast learning of the present invention includes the following three steps: 1) data acquisition; 2) model training; 3) fault diagnosis.

[0030] (1) Data Acquisition: The dataset used in this section comes from the open-source bearing fault dataset of the Electrical Laboratory at Case Western Reserve University. The bearing experimental platform uses a 2HP electric motor manufactured by an Indian company as the power source. Two types of bearings are installed at both the fan end and the drive end of the motor. Accelerometers are placed near and away from the motor bearings to measure the vibration data during motor operation.

[0031] Download the bearing failure dataset from Case Western Reserve University, which includes three main categories of failures: outer ring failure, inner ring failure, and rolling element failure. Each failure category has three different degrees of bearing damage: 0.007 inch, 0.014 inch, and 0.021 inch. Normal data from the base end is also included, with a total of 10 labels. The experimental platform also tested the bearing vibration data of the motor under different loads: 0 hp (no load), 1 hp, 2 hp, and 3 hp (full load). Finally, the experimental platform sampled vibration data from the accelerometer at the motor drive end at 12 kHz and 48 kHz, and also sampled vibration data from the accelerometer at the motor fan end at 12 kHz. All data used in this experiment came from the 12 kHz frequency sampled drive end vibration data.

[0032] Based on different operating conditions, the data was subdivided into four different datasets: A, B, C, and D, corresponding to 0hp (no load), 1hp, 2hp, and 3hp (full load), respectively. Each dataset includes 10 labels. The folders were named A-0hp, B-1hp, C-2hp, and D-3hp. For each fault type, the signal sequence length under a single operating condition was 120,000, the window length was 1200, and the sliding length was 600 sliding samples. Each sequence could collect 200 sequences, resulting in 800 sequences collected for each fault type across the four operating conditions. A total of 8000 wavelet images were generated. Then, each sequence underwent wavelet transform to generate a time-frequency image, with a size of 256*256 pixels and 3 channels.

[0033] (2) Model Training: The model training flowchart is shown in the figure. First, the dataset is acquired, preprocessed, and converted into a two-dimensional time-frequency graph as a data packet, which is then input into the MAE network for pre-training. The evaluation metric is the loss function. Pre-training is completed on a cloud computing platform. The training time is 24:58:12, and the number of parameters is 93,325,440. The loss function of the pre-training process is exported through TensorBoard visualization. Analyzing the experimental results, it is found that after 200 epochs of training, the loss function gradually decreases. After 120 epochs of training, the loss function changes very slowly, indicating that the loss function has gradually converged, and the model has basically completed learning from the input dataset. Pre-training is complete. The model save frequency (save_ckpt_freq) is set to 50, which means that the pre-trained model is saved once every 50 epochs of training. A total of four weight models are generated in the format pth, which will be used in subsequent fine-tuning experiments.

[0034] (3) Fault Diagnosis: The dataset used is the RTS rotor dynamics teaching experimental platform bearing dataset, which has five classification categories: normal bearing, inner ring fault, outer ring fault, mixed fault, and rolling element fault. Therefore, when performing fault diagnosis on this dataset, the softmax output of the linear head needs to be changed to 5 to adapt to the scenario of this dataset. The preprocessed datasets under the three speed conditions mentioned above are integrated into a dataset under mixed conditions and input into the fine-tuned model for fault diagnosis. Only the block layer closest to the output and the classification head are trained for 100 epochs. The batch size during training is 128. The model optimizer used is Adamw with 7,094,794 parameters. The fault diagnosis capability of the network is observed. The results are then processed using the TensorBoard visualization tool to obtain the fault diagnosis information under this dataset.

Claims

1. A motor fault diagnosis method based on self-supervised mask contrast learning, characterized in that, Comprising the following steps: (1) The bearing data collected by the University of West Florida, the actual motor bearing fault data and the test bench motor data are used as fault data, time-frequency analysis is performed on the data to obtain its time series, and wavelet transform is performed to obtain a time-frequency picture, i.e. an input data set in the form of a picture; (2) A variety of data augmentation methods are combined to obtain effective representation of fault features: spatial / geometric transformation methods such as cropping, resizing, horizontal flipping, rotation, and erasing are selected; shape transformation methods such as color distortion, Gaussian blur, and Sobel filtering are selected; (3) The data after data augmentation is preprocessed, i.e. mask image segmentation, the size of the RGB value in the field is set to segment the overlap rate, the area with high RGB value is set to high segmentation overlap rate, and the area with low RGB value is set to low segmentation overlap rate; (4) The segmented mask image block is input into the mask autoencoder for pre-training, the decoder task of pre-training is to restore the missing pixels of the image; L2 loss is selected as the pre-training loss function; (5) The obtained image matrix data is divided into training fault data and testing fault data; (6) Contrastive self-supervised learning is used on the training fault data, the input picture and its flipped, cropped, and erased samples are used as positive samples, and the distorted samples of the input picture and other samples in the training set are used as negative samples for contrastive learning, and the test data is compared with the positive samples and negative samples in the feature space to learn the feature representation of the samples; (7) The extracted features are passed through a fully connected layer to integrate the features of the training fault data and perform normalization, a softmax classifier is selected to classify the obtained probabilities, and cross-entropy is selected as the loss function of the classification task; (8) Vit-base is used as the fine-tuning model, containing 12 encoder blocks with labels from 0 to 11, patchsize is 16x16, the number of layers allowed to change weights during the reverse iteration process of the model is specified, while the remaining layer structure is frozen, forcing the parameter to be locked, greatly reducing the parameter amount, thereby updating the parameter weights; (9) The test fault data is verified, and when the test accuracy meets the requirements, it can be used for motor fault data diagnosis, realizing self-supervised contrastive learning fault diagnosis.

2. The motor fault diagnosis method based on self-supervised mask contrast learning according to claim 1, characterized in that, In step (1), continuous wavelet transform is performed on the time series signal to obtain a two-dimensional time-frequency image that can reflect time-frequency information at the same time, which is used for subsequent feature extraction.

3. The motor fault diagnosis method based on self-supervised mask contrast learning according to claim 1, characterized in that, The expression of the loss function in step (4) is as follows: where n represents the number of samples, y i represents the target variable, y i p represents the predicted value.

4. The motor fault diagnosis method based on self-supervised mask contrast learning according to claim 1, characterized in that, The expression of the classification task loss function in step (7) is as follows: where p(x i ) represents the true probability and q(x i ) is the predicted probability.

Images