Bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network

Abstract

The present invention discloses a bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network, the method comprising the following steps: step 1: obtaining and inputting a one-dimensional vibration signal of a bearing through an acceleration sensor; step 2: labeling each data of a data set with a fault category, and randomly dividing it into a training sample set and a test sample set; step 3: increasing the number and diversity of training samples through a VSE method; step 4: constructing TFIF‑Net, and initializing weights and biases; step 5: adjusting the weights and biases of TFIF‑Net through supervised training until a specified number of iterations is reached to complete the training of the diagnosis model; step 6: inputting a training sample set to obtain a classification result. The method proposed by the present invention not only has practical application value in the field of small sample bearing fault diagnosis, but also provides new research ideas for other classification and diagnosis tasks facing small sample limitations.

Description

Technical Field

[0001] The invention relates to a rolling bearing fault diagnosis method, and in particular to a bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network. Background Art

[0002] Bearings are key components of rotating machinery and equipment, and they bear the heavy responsibility of supporting the rotating body. However, due to the complex and harsh working environment, bearings often suffer from faults such as wear, spalling and fatigue. These faults often lead to abnormal operation of the equipment, and in severe cases even cause equipment damage and production accidents. Therefore, effective diagnosis of bearing faults is of great significance to improving the reliability and safety of rotating machinery and equipment.

[0003] Deep learning, an advanced technical means, is gradually becoming the mainstream method in the field of fault diagnosis due to its excellent processing speed, high-precision detection ability and the advantage of not relying on expert knowledge. However, most deep learning methods rely heavily on the number and diversity of training samples. For example, models such as deep belief networks, convolutional neural networks and recurrent neural networks all require a large number of training samples to achieve optimal performance. When the number of training samples is insufficient or unevenly distributed, the performance of deep learning models in fault diagnosis will be severely limited. However, in actual industrial scenarios, due to the high complexity and coordination of rotating machinery and equipment, it is an extremely difficult task to obtain sufficient bearing fault samples. Therefore, how to accurately realize bearing fault diagnosis under small sample conditions is still a challenging problem. At present, common small sample fault diagnosis methods are mainly divided into two categories: diagnosis methods based on data enhancement and diagnosis methods based on small sample learning.

[0004] The core challenge of data augmentation methods is how to create a series of new data with rich diversity while keeping the key feature information of the original data unchanged. Further, data augmentation methods can be divided into three categories: data augmentation methods based on resampling, data augmentation methods based on image transformation, and data augmentation methods based on generative adversarial networks. Resampling methods solve the problem of limited and unbalanced training samples by oversampling minority class samples and undersampling majority class samples. However, such methods can only use the information of existing data and cannot accurately map the distribution of original data. Image transformation methods, such as adding Gaussian white noise, masking noise, signal conversion, amplitude shifting, and time stretching, can effectively increase the number and diversity of sample data, thereby improving the robustness of diagnostic models. However, such methods will inevitably introduce additional noise and bias when processing data. These additional factors may cause excessive interference and distortion to the original data, thereby affecting the accuracy and reliability of the results. Generative adversarial network methods, due to their unique adversarial learning mechanism, can generate simulated data that is highly similar to the distribution of original data. For example, advanced models such as gradient normalized fully-attention Wasserstein generative adversarial networks, structurally similar generative adversarial networks, and autoencoder-based Fourier transform-like methods are able to generate simulated data whose distribution is very close to the real data. However, in the case of severely limited fault data, GAN may suffer from the problem of pattern loss, that is, the generator will gradually stop learning and always generate the same sample data.

[0005] Small sample learning is a new method that has emerged in the field of fault diagnosis in recent years. Its core lies in using a sample number far less than that required by traditional deep learning, but it can achieve results comparable to big data deep learning. In this study, Yang et al. proposed a small sample bearing fault diagnosis method based on convolutional neural networks, gated recurrent units, and attention mechanisms. The method showed excellent performance in identifying and locating key discriminant information in the input data. However, the effectiveness of this method in diagnosing small sample bearing faults has not been fully verified. Su et al. took a different approach and proposed a small sample bearing fault diagnosis method based on data reconstruction hierarchical recursive meta-learning. This method uses a cyclic meta-learning strategy and a one-time learning method for training, which significantly improves the ability to cope with limited data bearing fault diagnosis tasks under different working conditions. However, this method is still limited by small sample learning, and when processing large-scale data, the computational complexity is high, which undoubtedly increases the huge demand for computing resources and time. In addition, Zhang et al. proposed a method for diagnosing bearing faults in small samples, which uses dual-path convolution with attention mechanisms and bidirectional gated recurrent units. The method shows excellent generalization ability and robustness, and can accurately diagnose bearing faults even in complex environments. However, this method still faces challenges in extracting basic feature information when dealing with extremely small or extremely unbalanced sample data.

[0006] Although researchers have conducted in-depth research on small-sample bearing fault diagnosis methods and made some progress, they still face many challenges. Limited original samples cannot support the existing data augmentation methods to generate high-quality new samples. In addition, the complexity of small-sample learning methods leads to slow and inefficient training optimization processes. However, since data augmentation and small-sample learning are still of great value in scenarios where there is a lack of a large amount of labeled data, it is crucial to propose a new data augmentation method. This method should be able to generate new data with a wide variety while retaining the main feature information of the original data. In addition to the need to improve the data augmentation method, the shortcomings of the existing methods are also reflected in the lack of an efficient diagnostic model to match it. Therefore, it is necessary to develop a bearing fault diagnosis model that matches the new data augmentation method. This model should be able to complement the data augmentation method and have a strong feature extraction capability to extract sufficient feature information from the training samples. Summary of the invention

[0007] Due to the high time cost and operational difficulty of obtaining bearing fault data, the number and diversity of training samples are insufficient, which limits the effect of deep learning models in practical applications. The present invention provides a bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network. This method can increase the number and diversity of training samples without changing the main characteristic information of the original vibration signal. By constructing an intelligent fault diagnosis model, the key fault characteristic information of the input data is deeply mined, thereby realizing intelligent diagnosis of bearing faults.

[0008] The objective of the present invention is achieved through the following technical solutions:

[0009] A bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network includes the following steps:

[0010] Step 1: Obtain and input the one-dimensional vibration signal of the bearing through the acceleration sensor;

[0011] Step 2: Label each data set with a fault category label and randomly divide it into a training sample set and a test sample set;

[0012] Step 3: Increase the number and diversity of training samples through the VSE method. The specific steps are as follows:

[0013] Step 31: Decompose the vibration signal of the bearing into a set of intrinsic mode functions using a variational mode decomposition algorithm, each intrinsic mode function corresponding to a frequency component of the original vibration signal;

[0014] Step 32: using a linear extension method based on an autoregressive model, each eigenmode function is extended individually to generate a set of extended eigenmode functions that retain the original characteristic information;

[0015] Step 33: Reconstruct a more complex extended vibration signal by using the extended intrinsic mode function;

[0016] Step 4: Build TFIF-Net and initialize weights and biases. The specific steps for building TFIF-Net are as follows:

[0017] Step 41: Construct a feature extraction module: The feature extraction module is used to mine key fault feature information from time domain and frequency domain data, including two sets of parallel dilated convolution structures, SKNet structure, convolution structure and average pooling structure, where:

[0018] The dilated convolution structure is used as a feature extraction layer to mine long-range and deep fault feature information hidden in time domain and frequency domain data;

[0019] The SKNet structure is used as a feature enhancement layer to receive and enhance the long-distance and short-time features contained in the time domain and frequency domain data output by the feature extraction layer;

[0020] The convolution structure and the pooling structure form a feature selection layer, which is used to extract short-distance fault features from time domain and frequency domain features, while reducing redundant features;

[0021] The dilated convolution structure is composed of a wide dilated convolution structure, a BN structure and a ReLU activation structure;

[0022] The output of the dilated convolution structure is expressed as:

[0023]

[0024] In the formula, C is the size of the convolution kernel, and b i denote the weight and bias of the ith channel, respectively, x(j+h·c) denotes the (j+h·c)th local region of the input data, and h is the hole rate;

[0025] The output of the BN structure is expressed as:

[0026]

[0027] In the formula, γ i is the scaling factor, μ (i,B) and σ (i,B) Represent the mean and variance of each batch, ∈ i To maintain a stable constant, β i is bias;

[0028] The output of the ReLU activation structure is expressed as:

[0029] R i (j) = max{0,B i (j)}

[0030] The SKNet structure includes segmentation, fusion and selection operations. First, the segmentation operation uses convolution kernels of different sizes to create multiple paths. Then, the fusion operation integrates the information from these paths by assigning different attention to each channel, thereby forming a global representation of the selective weights. Finally, the selection operation uses different weights to summarize information from different spatial scales.

[0031] The output of the convolutional structure is expressed as:

[0032] T i =δ[β(V i * )]

[0033] In the formula, δ(·) indicates that the matrix (·) is processed using the ReLU activation structure, β(·) indicates that the matrix (·) is processed using the BN structure, and V i * Indicates the use of convolution structure for the matrix V i to process;

[0034] The output of the pooling structure is expressed as:

[0035] F i =δ[ρ(T i )]

[0036] In the formula, ρ(·) means using the pooling structure to process the matrix (·);

[0037] Step 42: Construct a feature fusion module: The feature fusion module is used to receive and fuse key fault features extracted from time domain and frequency domain data, and is composed of a weighted concatenation layer and a feature fusion layer, wherein:

[0038] The output of the weighted concatenation layer is expressed as:

[0039]

[0040] In the formula, α and β represent the weight factors in the time domain and frequency domain respectively, F i t and F i f Respectively represent the time domain and frequency domain features output by the feature extraction module;

[0041] The feature fusion layer is composed of a convolution structure, a pooling structure and a convolution structure. First, the convolution structure with small kernel learning and abstraction capabilities is used to adaptively fuse the weighted spliced ​​feature information; then, the pooling structure is used to reduce the dimension of the fused features and reduce the complexity of the model while retaining the key feature information; finally, the convolution structure is used to further fuse the time-frequency features;

[0042] Step 43: Construct a classification module: The classification module selects a global average pooling structure. The global average pooling structure first calculates the average value of all pixels in each channel feature matrix, and then generates a feature vector with a dimension equal to the number of categories. The feature vector is directly input into the Softmax classification structure to generate the classification result, where:

[0043] The output of the global average pooling structure is expressed as:

[0044]

[0045] In the formula, Z i Represents the feature matrix output by the feature fusion module;

[0046] The calculation formula of the classification result is as follows:

[0047]

[0048] Step 5: Adjust the weights and biases of TFIF-Net through supervised training until the specified number of iterations is reached to complete the training of the diagnostic model;

[0049] Step 6: Input the training sample set and obtain the classification results.

[0050] Compared with the prior art, the present invention has the following advantages:

[0051] 1. The present invention proposes a bearing fault diagnosis method based on vibration signal delay and time-frequency information fusion network, which includes two stages: (1) preprocessing stage; (2) diagnosis stage. In the preprocessing stage, a method for expanding vibration signals is proposed to increase the number and diversity of training samples. In the diagnosis stage, a time-frequency information fusion network is proposed to learn the basic fault features hidden in the training samples.

[0052] 2. The performance of the proposed method is evaluated on CWRU and laboratory bearing datasets. The results show that the VSE method effectively increases the number and diversity of training samples without changing the main feature information of the original signal. TFIF-Net can make full use of the fault feature information hidden in the time domain and frequency domain data to accurately realize bearing fault diagnosis.

[0053] 3. Compared with the existing methods, the proposed method shows superior performance. The proposed method not only has practical application value in the field of small sample bearing fault diagnosis, but also provides new research ideas for other classification and diagnosis tasks facing small sample limitations. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is the VSE flow chart.

[0055] Figure 2 This is the structure of TFIF-Net. “WC” is the abbreviation of weighted cascade layer.

[0056] Figure 3 Method for the fault diagnosis process.

[0057] Figure 4 is the time domain waveform of CWRU data, where O and E represent original data and extended data, respectively.

[0058] Figure 5 Visualization of feature classification results using T-SNE for the CWRU dataset.

[0059] Figure 6 Ablation experiment results on the CWRU dataset.

[0060] Figure 7 is the time domain waveform of the laboratory data, O and E represent the original data and extended data respectively.

[0061] Figure 8 Visualize feature classification results using T-SNE for the lab dataset.

[0062] Fig. 9 Ablation experiment results on the self-collected dataset. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0064] The present invention provides a bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network. In the preprocessing stage, a vibration signal extension method is proposed to increase the number and diversity of training samples. First, the original vibration signal is decomposed into a set of intrinsic mode functions using a variational mode decomposition algorithm to ensure that each frequency component can receive appropriate attention. Then, these intrinsic mode functions are extended using a linear extension method based on an autoregressive model to obtain a set of extended intrinsic mode functions that retain the original feature information. Finally, the extended vibration signal is reconstructed using these extended intrinsic mode functions. The extended vibration signal is not only a continuation of the original vibration signal, but also complements it to jointly construct a high-quality training sample set. In the diagnosis stage, a time-frequency information fusion network matching the vibration signal extension method is proposed to learn the key fault features hidden in the training samples. The network takes the time domain and frequency domain data of the training samples as input, and accurately realizes bearing fault diagnosis through a customized feature extraction module, feature fusion module and classification module. Among them, the feature extraction module extracts the long-distance fault features contained in the input data by expanding the convolution structure, and then introduces the selective kernel network structure to perform multi-scale enhancement on these features. Finally, a custom feature selection structure is used to extract the short-distance fault features contained in these enhanced features and remove redundant features. The feature fusion module selectively connects the feature matrices output by the feature extraction module by applying time-frequency weighting factors of different proportions, and then uses a custom feature fusion structure to achieve deep fusion of cross-domain features. Figure 3 As shown, the specific steps include:

[0065] Step 1: Obtain and input the one-dimensional vibration signal of the bearing through the acceleration sensor.

[0066] Step 2: Label each data set with a fault category and randomly divide it into a training sample set and a test sample set.

[0067] Step 3: Increase the number and diversity of training samples through the VSE method. Figure 1 As shown, the specific steps of the VSE method are as follows:

[0068] (1) The vibration signal of the bearing is decomposed into a set of intrinsic mode functions using the variational mode decomposition (VMD) algorithm, each of which corresponds to a frequency component of the original vibration signal. The decomposition operation ensures that each frequency component receives proper attention and processing in subsequent processing. In particular, for low-energy frequency components, this processing method can effectively prevent them from being masked by high-energy frequency components, thereby ensuring the integrity and accuracy of the vibration signal. Among them, VMD is a signal decomposition method based on variational Bayesian inference, which can decompose non-stationary and nonlinear signals. When applying VMD, the default value of the penalty factor is 2000. The bearing vibration signal f(x) can be expressed as {μ after VMD decomposition as {μ k}, and its corresponding center frequency can be expressed as {ω k}.

[0069] (2) Using the linear extension method based on the autoregressive model, each eigenmode function is extended individually to generate a set of extended eigenmode functions that retain the original characteristic information. The extended eigenmode function can be expressed as:

[0070]

[0071] In the formula, a i represents the prediction coefficient, p represents the model order, μ k Represents the frequency components of the vibration signal f(x) after decomposition.

[0072] (3) Using these extended intrinsic mode functions, more complex extended vibration signals can be reconstructed. This not only ensures that the basic characteristic information of the original vibration signal is retained, but also introduces new elements, significantly increasing the number and diversity of training samples. These extended vibration signals not only continue the original vibration signals, but also complement each other to form a high-quality training sample set. The extended vibration signal can be expressed as:

[0073]

[0074] In the formula, μ' i Represents μ k prediction signal.

[0075] Step 4: Construct TFIF-Net and initialize weights and biases. TFIF-Net includes feature extraction module, feature fusion module and classification module. Figure 2 As shown, the specific steps to build TFIF-Net are as follows:

[0076] (1) Construct a feature extraction module. The feature extraction module consists of multiple structural layers, mainly including two sets of parallel dilated convolution structures, SKNet structure, convolution structure and average pooling structure. These structures work together to mine key fault feature information from time domain and frequency domain data. The parallel dilated convolution structure is used to mine long-distance and deep fault feature information hidden in time domain and frequency domain data respectively. The essence of the dilated convolution structure is to map long-distance feature information into a larger receptive field. This not only improves the network's noise resistance, but also enables the network to better understand and process complex input data. This receptive field is usually composed of a wide dilated convolution structure, a batch normalization structure (BN) and a rectified linear unit (ReLU) activation structure. In TFIF-Net, this set of parallel receptive fields is called a feature extraction layer. Among them, the BN structure was proposed by Ioffe et al. It can reduce the internal covariance bias and effectively shorten the training time of the model. At present, the BN structure has become a standard tool in the field of deep learning. In TFIF-Net, the BN structure is located between the convolution structure and the activation structure. This layout is designed to ensure accurate data transmission and efficient processing. In addition, ReLU can also be regarded as a soft denoising technique for filtering negative features. Based on this feature, ReLU is used as an activation function in practical applications to reduce the impact of noise. The output of the dilated convolution structure can be expressed as:

[0077]

[0078] In the formula, C is the size of the convolution kernel, and b i They represent the weight and bias of the i-th channel, x(j+h·c) represents the (j+h·c)-th local area of ​​the input data, and h is the hole rate. The output of the BN structure can be expressed as:

[0079]

[0080] In the formula, γ i is the scaling factor, μ (i,B) and σ (i,B) Represent the mean and variance of each batch, ∈ i To maintain a stable constant, β i is the bias. The output of the ReLU activation structure can be expressed as:

[0081] R i (j) = max{0,B i (j)} (5)

[0082] Inspired by the functions of the human visual system, SKNet can use the attention mechanism to adaptively adjust the receptive field and achieve multi-scale feature enhancement. Therefore, the parallel SKNet structure is used to receive and enhance the long-distance and short-time features contained in the time domain and frequency domain data output by the feature extraction layer respectively. In TFIF-Net, this set of parallel SKNet structures is called the feature enhancement layer. In this process, SKNet is mainly divided into three steps: segmentation, fusion, and selection. First, the segmentation operation uses convolution kernels of different sizes to create multiple paths. Then, the fusion operation integrates the information from these paths by assigning different attention to each channel, thereby forming a global representation of the selective weights. Finally, the selection operation uses different weights to summarize information from different spatial scales.

[0083] Short-distance and long-distance patterns in sequence data coexist at multiple distance scales. Capturing multi-scale information of input data can improve the performance of the network. Feature extraction and enhancement layers usually extract and enhance fault features of long-distance patterns, but lack the mining of short-distance fault feature information. In addition, after deep processing of feature extraction and feature enhancement layers, the time domain and frequency domain feature matrices inevitably contain a large number of non-essential features. Feature extraction and feature enhancement layers usually extract and enhance fault features in long-distance patterns, but lack the mining of short-distance fault feature information. In addition, after deep processing of feature extraction and feature enhancement layers, the time domain and frequency domain feature matrices inevitably contain a large number of non-critical features. Therefore, a feature selection layer consisting of a convolutional structure and a pooling structure is defined to extract short-distance fault features from time domain and frequency domain features while reducing redundant features. The convolutional structure using a small convolution kernel plays a vital role in effectively capturing short-distance local details in time domain and frequency domain features and extracting valuable key features from them. The output of the convolutional structure can be expressed as:

[0084] T i =δ[β(V i * )] (6)

[0085] In the formula, δ(·) indicates that the matrix (·) is processed using the ReLU activation structure, β(·) indicates that the matrix (·) is processed using the BN structure, and V i * Indicates the use of convolution structure for the matrix V iThe pooling structure further filters and compresses these features, eliminating redundant information and ensuring that the model only focuses on important features. The output of the pooling structure can be expressed as:

[0086] F i =δ[ρ(T i )] (7)

[0087] Where ρ(·) represents the processing of the matrix (·) using a pooling structure.

[0088] (2) Constructing a feature fusion module: The feature extraction module can extract multi-scale key fault features from different information domains and understand them from different levels. At this stage, an effective cross-domain feature splicing and fusion mechanism is needed to integrate these key features. The traditional splicing layer realizes the horizontal fusion of features by combining features from multiple information domains. This fusion makes full use of the complementary information from different information domains, enabling the model to understand the data more deeply and comprehensively. However, each part of the time series does not contribute the same to the final output. In other words, the contribution of fault features to the final output is not the same in time domain and frequency domain data. Therefore, a feature fusion module consisting of a weighted splicing layer and a feature fusion layer is proposed to receive and fuse the key fault features extracted from time domain and frequency domain data.

[0089] The weighted splicing layer improves the utilization efficiency of cross-domain features by selectively splicing key fault features in the time domain and frequency domain using time-frequency weighting factors of different proportions. This splicing can be selectively spliced ​​according to the importance of cross-domain features, ensuring that the model can more accurately integrate feature information from different information domains. The output of the weighted splicing layer can be expressed as:

[0090]

[0091] In the formula, α and β represent the weight factors in the time domain and frequency domain respectively, F i t and F i f They represent the time domain and frequency domain features output by the feature extraction module respectively.

[0092] The features from multiple information domains are weighted and spliced. Although the integration of cross-domain features is preliminarily achieved, the integration of long-distance up and down information is lacking. In order to deeply fuse the long-distance information of cross-domain features, a feature fusion layer consisting of convolution structure, pooling structure and convolution structure is constructed. First, the weighted spliced ​​feature information is adaptively fused using the convolution structure with small kernel learning and abstraction capabilities. Among them, the small kernel can focus on the learning of local features and will not overfit the global data. Then, the pooling structure is used to reduce the dimension of the fused features and reduce the complexity of the model while retaining the key feature information. Finally, the small kernel convolution structure is used to further fuse the time-frequency features, thereby effectively improving the feature extraction capability and the fault recognition performance of the model.

[0093] (3) Constructing the classification module: In the classification module, a global average pooling structure is selected instead of the traditional fully connected structure. The global average pooling structure first calculates the average value of all pixels in the feature matrix of each channel, and then generates a feature vector with a dimension equal to the number of categories. This feature vector can be directly input into the Softmax structure for classification. The introduction of the global average pooling structure enables the model to enhance the generalization ability while retaining key features. In addition, the global average pooling structure ensures that the output of each layer corresponds to a specific area or feature. This helps to understand how the model extracts key features from the input data and realizes fault diagnosis. The output of the global average pooling structure can be expressed as:

[0094]

[0095] In the formula, Z i Represents the feature matrix output by the feature fusion module. For the bearing fault diagnosis task, the Softmax classification structure is used in the feature classification module to generate classification results. The calculation formulas for these results are as follows:

[0096]

[0097] (4) Optimize the parameters of TFIF-Net, and the final parameters are shown in Table 1. It is worth noting that the size of input data and Softmax is not fixed, but can be adjusted according to the needs of specific cases and application scenarios.

[0098] Table 1

[0099]

[0100]

[0101] Step 5: Adjust the weights and biases of TFIF-Net through supervised training until the specified number of iterations is reached to complete the training of the diagnostic model.

[0102] Step 6: Input the training sample set and obtain the classification results.

[0103] Embodiment 1:

[0104] The data acquisition platform of CWRU includes motor, torque sensor, dynamometer, fan end bearing (SKF-6023), drive end bearing (SKF-6025) and control electronics. The experimental data set comes from the vibration data of the drive end bearing and contains ten different types of working conditions, including health, inner ring failure, ball failure and outer ring failure. These single point faults are caused by electrical discharge machining technology. The fault diameters are 0.007 inches, 0.014 inches and 0.021 inches respectively. The sampling frequency of the torque sensor is 12kHz. The motor load is 1 horsepower and the speed is about 1772 revolutions per minute. This embodiment conducts three sets of experiments with different numbers of training samples, namely Experiment A, Experiment B and Experiment C, and the details are shown in Table 2. The training set includes 10 different working condition samples, each containing 1024 data points.

[0105] Table 2

[0106]

[0107] Figure 4 The time domain waveforms of the original and extended samples in the CWRU dataset are shown. Compared with the original samples, the extended samples show obvious "burr" noise, resulting in a rougher surface and lack of smoothness. This feature enhances the diversity of the extended samples, thereby improving the overall quality of the training sample set. Although there are some differences in the time domain waveforms of the original and extended samples, their envelopes are very similar. This shows that the extended samples successfully retain the main feature information of the original samples.

[0108] Figure 5 The feature classification results of the test set in Experiment A, Experiment B, and Experiment C are shown. Figure 5 (a) and Figure 5 As can be seen in (b), the sample features of different labels are non-discrete and there is no obvious classification error. Figure 5 As shown in (c), two samples originally marked as condition type 10 were misclassified as condition type 8. This error involves misclassification of different sizes of the same type of fault, which has little impact on the test results in practical applications. Despite this error, the accuracy of bearing fault diagnosis remains at 99.6%.

[0109] This embodiment conducted ten independent repeated tests on Experiment A, Experiment B, and Experiment C. The results show that the test accuracy in the same experiment is always high and the fluctuation range is small. Specifically, the average accuracy is 100%, 100%, and 99.3%, respectively, and the standard deviation is 0.00000, 0.00000, and 0.00475, respectively. Importantly, no misclassification between different faults was observed in multiple tests in different experiments. This shows that the method proposed in the present invention can accurately diagnose bearing faults and has strong generalization ability and robustness under small sample conditions.

[0110] The above results show that the proposed method can achieve accurate classification and diagnosis of bearing faults in small samples with low loss and high accuracy. In addition, the method also shows strong generalization ability and robustness in the CWRU public dataset. Even if there are only five original samples for each working condition, the proposed method can still classify and diagnose bearing faults with an accuracy of 99.3%.

[0111] In order to further verify the influence of each structure in the method proposed in the present invention on the final result, this embodiment conducts an ablation experiment on the CWRU bearing dataset. Specifically, this embodiment studies the influence of removing certain specific structures on the result. Among them, "NoExtraction" means removing the feature extraction layer in the proposed method. "NoEnhancement" means removing the feature enhancement layer in the proposed method. "NoSelection" means removing the feature selection layer in the proposed method. "NoConcatenation" means deleting the weighted concatenation layer in the proposed method. "NoFusion" means deleting the feature fusion layer from the proposed method. "ProposedMethod" refers to the entire proposed method.

[0112] Figure 6The results of the ablation experiment are shown. It can be clearly seen from the figure that as the number of original training samples decreases, the contribution of each module to the final result gradually increases. After removing the feature extraction layer, the diagnostic accuracy of Experiment A, Experiment B, and Experiment C decreased by 0.8, 0.8, and 0.7 percentage points, respectively. This shows that the performance of TFIF-Net can be significantly improved by adding an extended wide convolution structure. This structure can effectively retain the key fault information in the input data and capture long-distance fault feature information. After removing the feature enhancement layer, the diagnostic accuracy of Experiment B and Experiment C decreased by 0.1 and 0.3 percentage points, respectively. This shows that the implementation of the SKNet structure allows multi-scale enhancement of fault features, thereby improving the performance of TFIF-Net. After removing the feature selection layer, the diagnostic accuracy of Experiment B and Experiment C decreased by 0.3 and 5.1 percentage points, respectively. This shows that the feature selection layer, including convolution and pooling structures, can eliminate redundant features while retaining useful features. After removing the weighted splicing layer, the diagnostic accuracy of Experiment B and Experiment C decreased by 0.1 and 2.3 percentage points, respectively. This shows that when the number of training samples is small, the key fault feature information in the time domain and frequency domain data contributes differently to the final output. Targeted use of multi-domain feature information can effectively improve the performance of fault diagnosis. After removing the feature fusion layer, the diagnostic accuracy of Experiment B and Experiment C decreased by 0.4 and 2.2 percentage points, respectively. This shows that when the number of training samples is limited, the feature fusion layer proposed in the present invention can deeply fuse cross-domain feature long-range information, thereby improving the performance of TFIF-Net. The above results show that the method proposed in the present invention performs well in the task of bearing fault diagnosis, and each structure is indispensable.

[0113] Embodiment 2:

[0114] The laboratory data acquisition platform includes a drive motor (QABP-90L), an accelerometer, a bearing, a motor controller, a tachometer, a load, and a data acquisition system (not shown). The experimental data set comes from the vibration data of the test bearing, including ten different working conditions, including health, inner ring failure, ball failure, and outer ring failure. The speeds of the test bearings are 1200, 1800, and 2400 revolutions per minute (r / min), respectively. The sampling frequency of the accelerometer is 10kHz. The load of the test bearing is 5.25N. This embodiment conducts three groups of experiments with different numbers of training samples, namely Experiment D, Experiment E, and Experiment F, and the detailed information is shown in Table 3. Each sample contains 1024 data points.

[0115] Table 3

[0116]

[0117] Figure 7The time domain waveforms of the original samples and the extended samples in the laboratory dataset are shown. It can be clearly seen from the figure that compared with the original samples, there is obvious "burr" noise in the extended samples, resulting in a rougher and less smooth surface. The above results show that the VSE method can generate high-quality extended samples, thereby expanding the training sample set with high quality.

[0118] Figure 8 The feature classification results of the test set in Experiment D, Experiment E, and Experiment F are shown. Figure 8 In (a), it can be clearly seen that the sample features of different labels are non-discrete and there is no obvious classification error. However, it can be seen from 8(b) that there are individual misclassifications of samples labeled 6 in the test set of Experiment E. Two working condition samples that were originally labeled 6 were mistakenly classified as working condition types labeled 5. This error belongs to the misclassification between different sizes of the same type of fault, and has little impact on the detection results in practical applications. Despite such errors, the accuracy of bearing fault diagnosis still reached 99.8%. Figure 8 As can be seen from (c), there are individual misclassifications of samples with labels 3, 6, and 7 in the test set of experiment F. One working condition sample with label 3 is mistakenly classified as a working condition type with label 6, two working condition samples with label 6 are mistakenly classified as a working condition type with label 7, and one working condition sample with label 7 is mistakenly classified as a working condition type with label 6. Among them, the classification errors of working condition samples with label 6 and working condition samples with label 7 belong to errors between different sizes of the same type of faults, which have a low impact on the detection results in practical applications. The classification error of the working condition sample with label 3 belongs to errors between the same size of different faults. Despite this, there is still no misclassification between healthy and faulty states, which has a low impact on the detection results in practical applications, and when the number of original samples is 2, the accuracy rate reaches 99.2%.

[0119] The above results show that in the laboratory data set, the proposed method can achieve small sample bearing fault classification and diagnosis with low loss and high accuracy, and has good generalization and robustness. Even when the original sample number is 2, the proposed method can still achieve bearing fault classification and diagnosis with an accuracy of 99.1%.

[0120] In order to further verify the positive influence of each structure in the method proposed in the present invention on the final result, this embodiment performs an ablation experiment on a laboratory bearing data set. Fig. 9 shows the results of the ablation experiment. It can be seen from the figure that as the number of original training samples decreases, the contribution of each structure to the final result gradually increases.

[0121] When the feature extraction layer is removed, the diagnostic accuracy of Experiment D, Experiment E, and Experiment F decreases by 0.1, 1.4, and 6.4 percentage points, respectively. This shows that the extended convolution structure can effectively retain the key information of the input data and capture the long-distance related fault features, thereby enhancing the performance of TFIF-Net. When the feature enhancement layer is removed, the diagnostic accuracy of Experiment E and Experiment F decreases by 0.7 and 0.5 percentage points, respectively. This shows that the introduction of the SKNet structure to perform multi-scale enhancement of the short-term features output by the feature extraction layer can improve the performance of TFIF-Net, and the improvement effect is better when the number of original training samples for each type of working condition is low. When the feature selection layer is removed, the diagnostic accuracy of Experiment D, Experiment E, and Experiment F decreases by 2.2, 1.7, and 1.2 percentage points, respectively. This shows that the feature selection layer proposed in the present invention plays a vital role in removing redundant features and retaining useful features. When the weighted splicing layer is removed, the diagnostic accuracy of Experiment D, Experiment E, and Experiment F decreases by 0.6, 1.6, and 0.4 percentage points, respectively. This shows that the key fault feature information in the time domain and frequency domain data contributes differently to the final output, and the performance of fault diagnosis can be effectively improved by targeted use of multi-domain feature information. When the feature fusion layer is removed, the diagnostic accuracy of Experiment D, Experiment E, and Experiment F decreases by 1.0, 0.5, and 1.6 percentage points, respectively. This shows that the feature fusion layer proposed in the present invention can deeply fuse cross-domain feature long-distance information, thereby improving the performance of TFIF-Net. The above results further show that the method proposed in the present invention performs well in the task of bearing fault diagnosis, and each structure is indispensable.

Claims

1. A bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network, characterized in that The method comprises the following steps: Step 1: Obtain and input the one-dimensional vibration signal of the bearing through the acceleration sensor; Step 2: Label each data set with a fault category label and randomly divide it into a training sample set and a test sample set; Step 3: Increase the number and diversity of training samples through the VSE method. The specific steps are as follows: Step 31: Decompose the vibration signal of the bearing into a set of intrinsic mode functions using a variational mode decomposition algorithm, each intrinsic mode function corresponding to a frequency component of the original vibration signal; Step 32: using a linear extension method based on an autoregressive model, each eigenmode function is extended individually to generate a set of extended eigenmode functions that retain the original characteristic information; Step 33: Reconstruct a more complex extended vibration signal by using the extended intrinsic mode function; Step 4: Build TFIF-Net and initialize weights and biases. The specific steps to build TFIF-Net are as follows: Step 41: Construct a feature extraction module: The feature extraction module is used to mine key fault feature information from time domain and frequency domain data, including two sets of parallel dilated convolution structures, SKNet structure, convolution structure and average pooling structure, where: The dilated convolution structure is used as a feature extraction layer to mine long-range and deep fault feature information hidden in time domain and frequency domain data; The SKNet structure is used as a feature enhancement layer to receive and enhance the long-distance and short-time features contained in the time domain and frequency domain data output by the feature extraction layer; The convolution structure and the average pooling structure form a feature selection layer, which is used to extract short-distance fault features from time domain and frequency domain features while reducing redundant features; Step 43: Construct a classification module: The classification module selects a global average pooling structure. The global average pooling structure first calculates the average value of all pixels in each channel feature matrix, and then generates a feature vector with a dimension equal to the number of categories. The feature vector is directly input into the Softmax classification structure to generate a classification result. Step 42: Construct a feature fusion module: The feature fusion module is used to receive and fuse key fault features extracted from time domain and frequency domain data, and is composed of a weighted concatenation layer and a feature fusion layer, wherein: The output of the weighted concatenation layer is expressed as: In the formula, α and β represent the weight factors in the time domain and frequency domain respectively, F i t and F i f Respectively represent the time domain and frequency domain features output by the feature extraction module; The feature fusion layer is composed of a convolution structure, a pooling structure and a convolution structure. First, the convolution structure with small kernel learning and abstraction capabilities is used to adaptively fuse the weighted spliced ​​feature information; then, the pooling structure is used to reduce the dimension of the fused features and reduce the complexity of the model while retaining the key feature information; finally, the convolution structure is used to further fuse the time-frequency features; Step 5: Adjust the weights and biases of TFIF-Net through supervised training until the specified number of iterations is reached to complete the training of the diagnostic model; Step 6: Input the training sample set and obtain the classification results.

2. The bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network according to claim 1 is characterized in that The extended intrinsic mode function is expressed as: In the formula, a i represents the prediction coefficient, p represents the model order, μ k Represents the frequency components of the vibration signal f(x) after decomposition; The extended vibration signal is expressed as: In the formula, μ i ' represents μ k prediction signal.

3. The bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network according to claim 1 is characterized in that The dilated convolution structure is composed of a wide dilated convolution structure, a BN structure and a ReLU activation structure, where: The output of the dilated convolution structure is expressed as: In the formula, C is the size of the convolution kernel, and b i denote the weight and bias of the ith channel, respectively, x(j+h·c) denotes the (j+h·c)th local region of the input data, and h is the hole rate; The output of the BN structure is expressed as: In the formula, γ i is the scaling factor, μ (i,B) and σ (i,B) Represent the mean and variance of each batch, ∈ i To maintain a stable constant, β i is bias; The output of the ReLU activation structure is expressed as: R i (j)=max{0,B i (j)}.

4. The bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network according to claim 1 is characterized in that The SKNet structure includes segmentation operations, fusion operations, and selection operations. First, the segmentation operation uses convolution kernels of different sizes to create multiple paths. Then, the fusion operation integrates the information from these paths by assigning different attention to each channel, thereby forming a global representation of selective weights. Finally, the selection operation uses different weights to summarize information from different spatial scales.

5. The bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network according to claim 1 is characterized in that The output of the convolutional structure in the feature extraction module is expressed as: T i =δ[β(V i * )] In the formula, δ(·) indicates that the matrix (·) is processed using the ReLU activation structure, β(·) indicates that the matrix (·) is processed using the BN structure, and V i * Indicates the use of convolution structure for the matrix V i to process; The output of the average pooling structure is expressed as: F i =δ[ρ(T i )] Where ρ(·) represents the processing of the matrix (·) using a pooling structure.

6. The bearing fault diagnosis method based on vibration signal extension and time-frequency information fusion network according to claim 1 is characterized in that The output of the global average pooling structure is expressed as: In the formula, Z i Represents the feature matrix output by the feature fusion module; The calculation formula of the classification result is as follows:

Images