Bearing small sample diagnosis method based on multi-dimensional data enhancement and feature fusion

Abstract

The invention discloses a bearing small sample diagnosis method based on multi-dimensional data enhancement and feature fusion, and the method comprises the steps: obtaining an original vibration signal and a label of a bearing, and carrying out the preprocessing of the original vibration signal, and obtaining a feature signal; data enhancement is carried out on the original vibration signal and the characteristic signal through a data enhancement model, an expanded sample is obtained, and a loss difference threshold mechanism is set in the data enhancement model; constructing an expansion training set according to the original vibration signal, the tag, the feature signal and the expansion sample; a diagnosis model is constructed, the diagnosis model is trained through an expansion training set, the diagnosis model is used for processing the original vibration signals and the feature signals to obtain feature descriptions of different details, the different detail descriptions are fused and processed in sequence, and a fault diagnosis result is obtained; and obtaining a to-be-detected vibration signal, and processing the to-be-detected vibration signal through the trained diagnosis model to obtain an unbalance diagnosis result of the to-be-detected vibration signal.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and in particular relates to a small sample bearing diagnosis method based on multidimensional data enhancement and feature fusion. Background Art

[0002] In industrial practice, rotating machinery systems operate in a healthy state most of the time, while fault states are relatively short-lived. Therefore, it is usually easy to collect a large amount of vibration data representing normal conditions, but in contrast, the acquisition of fault data is relatively limited. This imbalance between normal and abnormal data poses a considerable challenge to fault diagnosis. In addition, when building an intelligent diagnosis model, it is particularly important to ensure the type and quality of input data, given that the quality and diversity of data have a direct impact on the performance and accuracy of the model.

[0003] At present, many scholars have proposed a series of solutions to improve the accuracy of fault diagnosis. Goodfellow et al. proposed a data synthesis model, Generative Adversarial Network (GAN), which trains the generator and discriminator networks through game theory to generate realistic data samples, but there are problems such as unstable training, difficulty in convergence, and mode collapse. Radford et al. proposed a deep convolutional generative adversarial network (DCGAN) by replacing the ordinary multi-layer neural network with a convolutional neural network based on GAN. Although this convolutional GAN ​​showed stable training effects on many data sets, it promoted the application of CNN in unsupervised scenarios. Li et al. proposed a Wasserstein deep convolutional generative adversarial network (WDCGAN) that improved GAN, generated data for one-dimensional power grid data and used it as the input of the intelligent diagnosis model to achieve small sample anomaly classification. The data generated by the proposed WDCGAN model was similar to the real data and stabilized the training process. Fan et al. used LBP grayscale processing technology to convert one-dimensional vibration signals into grayscale texture images and established original data sets. The original data sets were enhanced by WGAN-GP, and finally CNN was used for accuracy diagnosis. Qin et al. used CWT to convert the original vibration signal into a time-frequency image as the input of the WDCGAN-GP data enhancement model, and used the synthetic data for fault diagnosis and detection.

[0004] Existing technologies have explored the application of data preprocessing and enhancement in the field of fault diagnosis from different perspectives. However, in the current research trend, most diagnostic models are based on single-dimensional single-information input, which can easily lead to the loss of key features and a decrease in diagnostic accuracy. However, multi-source heterogeneous data collected from different sensors used in industrial production cannot be directly used by traditional neural networks, which brings great difficulties to deep learning fault diagnosis based on small samples. Therefore, it is crucial to integrate multimodal information into the diagnostic model. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a small sample bearing diagnosis method based on multi-dimensional data enhancement and feature fusion to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a small sample bearing diagnosis method based on multidimensional data enhancement and feature fusion, comprising:

[0007] Obtaining the original vibration signal and label of the bearing, and preprocessing the original vibration signal to obtain a characteristic signal;

[0008] Performing data enhancement on the original vibration signal and the characteristic signal through a data enhancement model to obtain an expanded sample, wherein the data enhancement model is provided with a loss difference threshold mechanism;

[0009] Construct an extended training set based on the original vibration signal and label, feature signal and extended samples;

[0010] Constructing a diagnostic model, and training the diagnostic model by expanding a training set, wherein the diagnostic model is used to process the original vibration signal and the characteristic signal to obtain multi-faceted feature descriptions, and the multi-faceted feature descriptions are sequentially fused and processed to obtain a fault diagnosis result;

[0011] The vibration signal to be measured is obtained, and the vibration signal to be measured is processed by the trained diagnosis model to obtain the imbalance diagnosis result of the vibration signal to be measured.

[0012] Optionally, the process of preprocessing the original vibration signal includes:

[0013] The original vibration signal is subjected to grayscale processing and continuous wavelet transformation to obtain a grayscale image and a time-frequency image.

[0014] Optionally, the data augmentation model includes a one-dimensional generative adversarial network and a two-dimensional generative adversarial network, wherein the one-dimensional generative adversarial network adopts a one-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty, and the two-dimensional generative adversarial network adopts a two-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty;

[0015] A one-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty is trained through the original vibration signal, and a two-dimensional generative adversarial network is trained respectively through the grayscale image and the time-frequency image. When the loss difference in the data enhancement model training process meets a preset threshold, a trained one-dimensional generative adversarial network and a two-dimensional generative adversarial network are generated, and an expanded sample is generated through the trained one-dimensional generative adversarial network and the two-dimensional generative adversarial network. The expanded sample is used to balance the number of original vibration signals and feature signals.

[0016] Optionally, the diagnostic model adopts a multi-dimensional feature fusion convolutional neural network, wherein the multi-faceted feature descriptions in the multi-dimensional feature fusion convolutional neural network include one-dimensional waveform features, grayscale structured features, and time-frequency features.

[0017] Optionally, after training the diagnostic model by expanding the training set, the following step further includes:

[0018] According to the original vibration signal and label, characteristic signal and expanded sample, an expanded verification set and an expanded test set are constructed. The trained diagnostic model is verified and tested by the expanded verification set and the expanded test set, and the tested diagnostic model is obtained to process the vibration signal to be tested.

[0019] Optionally, the ratio of the expanded training set, the expanded validation set, and the expanded test set is 7:2:1

[0020] Compared with the prior art, the present invention has the following advantages and technical effects:

[0021] The proposed WDCGAN-MDFFCNN bearing imbalance fault diagnosis method not only successfully solves the diagnosis problem under the background of sample imbalance, but also overcomes a series of difficulties such as poor generalization ability of the diagnosis model and low diagnosis accuracy caused by single information input. The experimental results under different working conditions in the CWRU dataset show that:

[0022] In terms of data enhancement, this paper proposes a WDCGAN data enhancement model based on the combination of the advantages of WGAN-GP and DCGAN, and introduces a loss difference threshold mechanism to limit the loss difference between the generator and the discriminator within a certain range, thereby accelerating the convergence speed of the model, improving the quality of generated data, and improving the stability and efficiency of the model, which helps to achieve a faster Nash equilibrium.

[0023] In terms of fault detection, the MDFFCNN diagnostic model is proposed. This model handles fault diagnosis problems under different working conditions through an innovative multi-dimensional feature fusion strategy, effectively integrates feature information from different representation spaces, and generates a comprehensive feature representation to retain multiple features of the original data. It also improves the accuracy and convergence speed of diagnosis, providing an efficient and comprehensive solution for intelligent fault diagnosis of complex mechanical systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0025] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0026] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0027] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0028] In the existing technology, deep learning has been widely used in bearing fault diagnosis, and its intelligent diagnosis method has shown significant advantages. However, in real engineering environments, deep learning fault detection models face a severe challenge: they usually require a large number of data samples for training. However, the actual situation is often that the sample data of normal operation is abundant, while the sample data of fault conditions is very scarce. This sample imbalance phenomenon has caused great trouble to the deep learning fault diagnosis method based on small samples.

[0029] Currently, most diagnostic models rely on single-dimensional, single-information inputs, which is easily constrained by data limitations and is difficult to cover a sufficient sample range, thus affecting the model's generalization ability and diagnostic accuracy. Therefore, diagnostic models that rely solely on single-information inputs have obvious limitations. In order to improve the accuracy and comprehensiveness of diagnosis, it is urgent to integrate multi-dimensional information.

[0030] In summary, the sample imbalance phenomenon in actual industry and the problem that the intelligent diagnosis model is difficult to accurately diagnose due to the feature limitations of single information input, such as Figure 1As shown, the present invention proposes a small sample bearing diagnosis method based on multidimensional data enhancement and feature fusion (WDCGAN-MDCFFCNN). The proposed diagnosis method includes three main stages.

[0031] The first stage: the collected original vibration signal of the bearing is divided and preprocessed, and the grayscale processing technology and continuous wavelet transform are used to convert it into grayscale image and time-frequency image respectively. The original vibration signal is a one-dimensional signal, and the grayscale image and time-frequency image are two-dimensional images. The one-dimensional original vibration signal and the converted two-dimensional image, grayscale image and time-frequency image are divided into training set, verification set and test set according to a certain proportion for training subsequent data enhancement model.

[0032] The second stage: construct the data enhancement model of one-dimensional and two-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty (1DWDCGAN & 2DWDCGAN) respectively, and introduce the loss difference threshold mechanism in the model to make the model converge quickly and reach Nash equilibrium earlier. The one-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty is trained by the original vibration signal, and the two-dimensional generative adversarial network is trained by the grayscale image and the time-frequency image respectively. When the loss difference in the training process of the data enhancement model meets the preset threshold, the trained one-dimensional generative adversarial network and the two-dimensional generative adversarial network are generated, and the expanded samples are generated by the trained one-dimensional generative adversarial network and the two-dimensional generative adversarial network. The expanded samples are used to balance the number of original vibration signals and feature signals. The data enhancement model is used to perform sample synthesis expansion on the original vibration signal, grayscale image and time-frequency image according to different data enhancement ratios, so as to generate more realistic and diverse samples to balance the data set. The model structures and key parameters of the 1DWDCGAN model and the 2DWDCGAN model are shown in Tables 1 and 2, where the different network layer structures are sequentially connected structures from top to bottom, and the generator and discriminator are parallel structures.

[0033] Table 1

[0034]

[0035] Table 2

[0036]

[0037] Specifically, in the 1DWDCGAN model, the generator and the discriminator are connected in parallel, wherein the generator includes a fully-connected layer and four deconvolution layers (deconv) connected in sequence, and the first three deconvolution layers are connected with a normalization layer, and InstanceNorm is used for normalization; wherein the discriminator includes four convolution layers (conv) and a fully-connected layer connected in sequence, and the last three convolution layers are connected with a normalization layer, and InstanceNorm is used for normalization. The specific relevant activation functions, number of channels and convolution kernel size are shown in Table 1 above.

[0038] Specifically, in the 2DWDCGAN model, the generator and the discriminator are connected in parallel, wherein the generator includes five sequentially connected transposed convolutional layers (convTranspose), and the first four transposed convolutions are connected with a normalization layer, and InstanceNorm is used for normalization; wherein the discriminator includes five sequentially connected convolutional layers (conv), and the middle three convolutional layers are connected with a normalization layer, and InstanceNorm is used for normalization. The specific related activation functions, number of channels and convolution kernel size are shown in Table 2 above.

[0039] The loss function of the model used in this article is as follows:

[0040]

[0041] where x is from the real data distribution P r The real data sampled in, λ is the gradient penalty coefficient, P g is the generated data distribution defined by the implicit generative model G, For along P r With P g The random interpolation sampling between them, ||·||2 represents the L2 norm of the gradient.

[0042] The loss difference threshold mechanism is introduced in this model. First, a threshold T is set to monitor the loss difference between the generator and the discriminator, and ΔL is the loss difference. If the absolute value of the loss difference exceeds T, it is considered unbalanced. If ΔL>T, the generator is trained 5 times additionally; conversely, if ΔL<-T, the discriminator is trained 5 times additionally. ΔL is retested after each time until it returns to [-T, T]. After reaching equilibrium, normal training is resumed to make the model converge quickly and reach Nash equilibrium earlier.

[0043] The third stage: construct a diagnostic model of multi-dimensional feature fusion convolutional neural network (MDFFCNN for short). The diagnostic model can use a variety of different information as input, among which the input of the model includes the original vibration signal, grayscale image and time-frequency image mentioned above, extract relevant features from the original vibration signal, grayscale image and time-frequency image, and then combine the features of these information to form a feature description with multiple details. During the training process, the diagnostic model is trained on the enhanced data input, which includes the original and data-enhanced one-dimensional vibration information, grayscale image and time-frequency image. In data processing or actual diagnosis, the diagnostic model extracts important features from the input data and combines them together to form a multi-faceted feature description. This multi-faceted feature description includes the features of the original signal, structural features, and time-frequency features. The above multi-faceted feature descriptions are fused and identified and classified to generate the final diagnostic results.

[0044] The present invention designs a diagnostic model of a multi-dimensional feature fusion convolutional neural network (MDFFCNN for short). The main structure and parameters of the MDFFCNN model are shown in Table 3, where the 1D after each layer in "layer type" indicates one-dimensional convolution, and 2D indicates two-dimensional convolution. At the same time, the last number marked in "layer type" is the layer number, which means the convolution layer.

[0045] Table 3

[0046]

[0047] In this model, different signal types, namely original vibration signals, grayscale images and time-frequency images, are processed through 1DCNN, 2DCNN and 2DCNN connected in parallel. Specifically, 1DCNN includes two convolutional layers connected in sequence, and 2DCNN includes three convolutional layers connected in sequence. After the above 1DCNN, 2DCNN and 2DCNN extract low-order features of the signal, the low-order features are fused through the feature fusion layer (Fusion), and the Softmax function is used in the output layer to output the diagnosis results. The specific activation function, pooling layer, flattening layer and stride parameters of the above convolution are shown in Table 3 above, which will not be repeated here.

[0048] The model can use one-dimensional vibration signals, time-frequency graphs, and grayscale graphs as inputs at the same time, and then use convolutional layers to extract features from the three types of input data and add a maximum pooling layer after each convolutional module to reduce the dimension of the output features. In the last convolutional module, a global average pooling layer is selected to replace the fully connected layer, so as to focus on the global features of the entire feature map. These extracted features are combined through the feature fusion layer to form a feature description with multiple details, where the feature fusion layer performs a splicing operation on different feature maps.

[0049] Finally, the trained diagnostic model is used to process the abnormal samples to be tested for fault diagnosis.

[0050] The data sets under different working conditions of Case Western Reserve University were selected as experimental data sets to verify the effectiveness of this method. First, the vibration signals collected from the drive end of the 1HP and 2HP working conditions were selected as the original input signals. The original signals were preprocessed and converted into two-dimensional structured grayscale images, continuous wavelet transform time-frequency images, and the retained original vibration data. The processed data were sample expanded by 1DWDCGAN and 2DWDCGAN to achieve multidimensional data enhancement. The expanded data set was divided into training set, validation set, and test set in a ratio of 7:2:1 for training and testing of the MDFFCNN diagnostic model, and two sets of single-dimensional and multi-dimensional convolutional neural network diagnostic comparison experiments under different working conditions and different data enhancement ratios were designed. The specific method for diagnosis is shown in Table 4. During the experiment, it was ensured that the 1DCNN and 2DCNN network structures and parameter settings used in methods 1, 2, and 3 were exactly the same as those of method 4 proposed in this patent. In order to avoid random errors, each method was tested ten times. The test results are shown in Tables 5 and 6, where Table 5 shows the 1HP test set accuracy (%) of fault diagnosis using different methods and different data augmentation ratios, and Table 6 shows the 2HP test set accuracy (%) of fault diagnosis using different methods and different data augmentation ratios. The experimental results show that the fault diagnosis accuracy of the proposed diagnosis framework with or without data augmentation is significantly higher than that of the single-dimensional convolutional neural network model, and it has better stability, better robustness, and faster convergence speed.

[0051] Table 4

[0052]

[0053] Table 5

[0054]

[0055] Table 6

[0056]

[0057] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A small sample bearing diagnosis method based on multidimensional data enhancement and feature fusion, characterized in that: include: Obtaining the original vibration signal and label of the bearing, and preprocessing the original vibration signal to obtain a characteristic signal; Performing data enhancement on the original vibration signal and the characteristic signal through a data enhancement model to obtain an expanded sample, wherein the data enhancement model is provided with a loss difference threshold mechanism; Construct an extended training set based on the original vibration signal and label, feature signal and extended samples; Constructing a diagnostic model, and training the diagnostic model by expanding a training set, wherein the diagnostic model is used to process the original vibration signal and the characteristic signal to obtain multi-faceted feature descriptions, and the multi-faceted feature descriptions are sequentially fused and processed to obtain a fault diagnosis result; The vibration signal to be measured is obtained, and the vibration signal to be measured is processed by the trained diagnosis model to obtain the imbalance diagnosis result of the vibration signal to be measured.

2. The method according to claim 1, characterized in that The process of preprocessing the original vibration signal includes: The original vibration signal is subjected to grayscale processing and continuous wavelet transformation to obtain a grayscale image and a time-frequency image.

3. The method according to claim 2, characterized in that The data enhancement model includes a one-dimensional generative adversarial network and a two-dimensional generative adversarial network, wherein the one-dimensional generative adversarial network adopts a one-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty, and the two-dimensional generative adversarial network adopts a two-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty; A one-dimensional Wasserstein deep convolutional generative adversarial network with gradient penalty is trained through the original vibration signal, and a two-dimensional generative adversarial network is trained respectively through the grayscale image and the time-frequency image. When the loss difference in the data enhancement model training process meets a preset threshold, a trained one-dimensional generative adversarial network and a two-dimensional generative adversarial network are generated, and an expanded sample is generated through the trained one-dimensional generative adversarial network and the two-dimensional generative adversarial network. The expanded sample is used to balance the number of original vibration signals and feature signals.

4. The method according to claim 1, characterized in that The diagnostic model adopts a multi-dimensional feature fusion convolutional neural network, wherein the multi-faceted feature descriptions in the multi-dimensional feature fusion convolutional neural network include one-dimensional waveform features, grayscale structured features, and time-frequency features.

5. The method according to claim 1, characterized in that After training the diagnostic model by expanding the training set, the following steps are further included: According to the original vibration signal and label, characteristic signal and expanded sample, an expanded verification set and an expanded test set are constructed. The trained diagnostic model is verified and tested by the expanded verification set and the expanded test set, and the tested diagnostic model is obtained to process the vibration signal to be tested.

6. The method according to claim 5, characterized in that The ratio of the expanded training set, the expanded validation set and the expanded test set is 7:2:1.

Images