Bearing fault diagnosis method based on crown porcupine optimization variable mode decomposition

Abstract

The invention discloses a bearing fault diagnosis method based on crown porcupine optimization variable mode decomposition. A bearing vibration signal is collected; processing the bearing vibration signal by using crown porcupine optimization variable mode decomposition to obtain an intrinsic mode function IMFs; converting the IMFs into a two-dimensional wavelet time-frequency graph through continuous wavelet transform, and performing feature fusion through a convolutional neural network to form a new two-dimensional wavelet time-frequency graph; the method comprises the following steps: constructing a feature fusion network which takes a residual network and a Swin Transform network as parallel channels; and inputting the two-dimensional wavelet time-frequency graph into the feature fusion network for learning and fault diagnosis. The bearing vibration signal is subjected to variable mode decomposition and two-dimensional wavelet time-frequency graph processing, and the convolutional neural network and the parallel channel feature fusion network are combined, so that high-accuracy diagnosis of bearing faults under the conditions of limited samples and variable working conditions is realized. The method has high practicability and accuracy, and the reliability of fault diagnosis can be effectively improved.

Description

Technical Field

[0001] The present invention relates to signal processing and deep learning, and in particular to a bearing fault diagnosis method based on deep learning. The method is mainly used for bearing fault diagnosis in the industrial field and is applicable to rolling bearings of various rotating machines. Background Art

[0002] In the operation and maintenance of industrial equipment, vibration monitoring and fault diagnosis are one of the key technologies. Traditional vibration analysis methods, such as time domain analysis, frequency domain analysis, and time-frequency domain analysis, are widely used in equipment status monitoring and fault warning. In recent years, with the development of Industry 4.0 and intelligent manufacturing, the use of data-driven methods for fault prediction and diagnosis has become a research hotspot. These methods rely on advanced data acquisition technology and machine learning algorithms to improve the accuracy and timeliness of fault diagnosis. At present, vibration signal analysis technology mainly includes the following aspects:

[0003] 1. Time domain analysis: Analyze vibration signals through statistical features (such as mean, variance, peak value, etc.). Although the method is simple, its ability to identify complex faults is limited.

[0004] 2. Frequency domain analysis: Use Fourier transform to convert time domain signals into frequency domain signals to identify frequency characteristics. Commonly used techniques include power spectral density analysis, but the processing effect of non-stationary signals is poor.

[0005] 3. Time-frequency domain analysis: By decomposing the signal into time-frequency features through wavelet transform and other technologies, non-stationary signals can be better processed. This method has been widely used in equipment fault diagnosis, but it still faces challenges in feature extraction and information fusion.

[0006] In recent years, the application of deep learning methods in vibration signal analysis has been significantly improved. Deep learning technologies such as convolutional neural networks (CNN) and recurrent neural networks (RNN) have been applied to feature extraction and pattern recognition, effectively improving the accuracy of fault diagnosis. However, these methods still require a large amount of labeled data and computing resources.

[0007] Problems with current technology

[0008] 1. Insufficient feature extraction: Traditional methods have limitations in feature extraction of complex fault modes and have difficulty capturing subtle changes in signals.

[0009] 2. Complex data processing: Although time-frequency domain analysis can process non-stationary signals, it is highly complex in the process of feature extraction and information fusion, resulting in high computational overhead.

[0010] 3. High demand for model training: Deep learning models usually require a large amount of labeled data and computing resources, and data acquisition and model training in practical applications are difficult.

[0011] 4. Poor adaptability: When faced with changes in the equipment working environment or new failure modes, the existing models have poor adaptability and generalization capabilities and need to be continuously adjusted and optimized. Summary of the invention

[0012] With the complexity of industrial equipment and the diversity of operating environments, practical applications often face the challenges of limited sample data and variable operating conditions, which puts higher requirements on the accuracy and reliability of fault diagnosis models.

[0013] The present invention aims to provide a bearing fault diagnosis method based on crown porcupine optimized variable modal decomposition to improve the diagnostic accuracy under limited samples and variable working conditions. The specific implementation steps are as follows:

[0014] 1. Collect vibration signals: Obtain the vibration signal data of the bearing under different working conditions. These vibration signal data reflect the working status of the bearing and possible fault information.

[0015] 2. Use crested porcupine variable mode decomposition to process vibration signals: The collected vibration signals are decomposed into intrinsic mode functions (IMFs) through the crested porcupine optimized variable mode decomposition method. This decomposition method can effectively extract characteristic information from the signal and reduce the impact of noise.

[0016] 3. Generation and feature fusion of two-dimensional wavelet time-frequency graph: The obtained intrinsic mode function is converted into a two-dimensional wavelet time-frequency graph using continuous wavelet transform. Then, the convolutional neural network is used to perform feature fusion on these time-frequency graphs to form a new two-dimensional wavelet time-frequency graph, which helps to extract more discriminative features.

[0017] 4. Construct a feature fusion network: Construct a parallel channel feature fusion network that includes a residual network and a Swin Transformer network. The residual network can effectively capture detailed features, while the Swin Transformer network is good at processing global information. The combination of the two can improve the effect and accuracy of feature fusion.

[0018] 5. Training and application: The processed two-dimensional wavelet time-frequency graph is input into the feature fusion network for training to achieve accurate diagnosis of bearing faults. By learning and optimizing the model, the fault diagnosis accuracy under limited samples and variable working conditions can be further improved.

[0019] Furthermore, in step 1, by changing the motor frequency or load size to obtain vibration signal data under different working conditions, the bearing vibration data under different working conditions in a real industrial environment can be better simulated. The frequency is selected as 19.88 Hz and 29.87 Hz respectively, and the load can be selected as 0 pounds, 50 pounds, 250 pounds, 270 pounds and 300 pounds.

[0020] Furthermore, in step 2, the variable mode decomposition is an adaptive, completely non-recursive method of modal variation and signal processing. This technology has the advantage of being able to determine the number of modal decompositions, and its adaptability is reflected in determining the number of modal decompositions of a given sequence according to actual conditions, and in the subsequent search and solution process, the optimal center frequency and limited bandwidth of each mode can be adaptively matched, and the effective separation of the intrinsic mode component (IMF) and the frequency domain division of the signal can be achieved, thereby obtaining the effective decomposition components of the given signal, and finally obtaining the optimal solution to the variational problem.

[0021] Assuming that the original bearing vibration signal f is decomposed into k components, the decomposition sequence is guaranteed to be a modal component with a limited bandwidth and a center frequency, and the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original bearing vibration signal. The VMD constrained variational model is as follows:

[0022]

[0023] where u k is each mode function, ω k is the modal center frequency. ||·|| demodulates the signal through Gaussian smoothing (i.e., the square root of the L2 norm gradient) to obtain the bandwidth description of each modal function, δ(t) is the Dirac function, and * is the convolution operator. For Hilbert transform, after the bearing vibration signal passes through the Hilbert transform, the positive frequency part is multiplied by -j, that is, the phase is shifted by -π / 2 while keeping the amplitude unchanged, and for the negative frequency component, it is shifted by π / 2, which converts a one-dimensional bearing vibration signal into a bearing vibration signal on a two-dimensional complex plane. The modulus and phase angle of the complex number represent the amplitude and phase of the bearing vibration signal. It is used to adjust the spectrum of each mode so that it is modulated to the corresponding baseband.

[0024] The crested porcupine optimization algorithm simulates four defense strategies of crested porcupines when facing predators, namely: the first defense strategy (visual intimidation), the second defense strategy (sound intimidation), the third defense strategy (smell attack), and the fourth defense strategy (physical attack). The first and second defense strategies of the above four defense strategies are used for global exploration, and the third and fourth defense strategies are used for local development. The algorithm proposes a new strategy called cyclic population reduction technique to simulate the preposition that not all CPs activate their defense mechanisms, but only those CPs that are threatened activate their defense mechanisms. This strategy promotes the convergence rate and population diversity.

[0025] The process of crown porcupine optimization variational mode decomposition is as follows:

[0026] First, the population is initialized and some algorithm parameters are set; then the bearing vibration signal is decomposed using VMD, the fitness of each individual in the population is calculated, and the individual with the highest fitness is calculated as the current global optimal solution. The population begins to search the solution space and updates the position of each individual in the population, and then determines whether the updated position is better, until the number of iterations meets the end condition. Finally, the bearing vibration signal is decomposed by VMD using the optimized parameters to obtain multiple IMFs.

[0027] Furthermore, in step 3, the continuous wavelet transform (CWT) is a method for simultaneously analyzing the bearing vibration signal in the time domain and the frequency domain. It transforms the bearing vibration signal by using wavelet functions of different scales and positions to obtain the local characteristics of the bearing vibration signal. The core idea of ​​CWT is to perform wavelet decomposition on the bearing vibration signal at different scales (frequencies) and positions. To achieve this goal, CWT uses a wavelet function (wavelet), usually called a mother wavelet or a basic wavelet. This wavelet function is a waveform with adjustable scale. The formula of CWT is expressed as:

[0028]

[0029] Where f(t) is the original signal, ψ * is the complex conjugate of the wavelet function, a is the scale parameter, and b is the translation parameter. The original bearing vibration signal f is decomposed by the crown porcupine variable mode to obtain the intrinsic mode function IMF, which is the different bearing vibration signals f(t). After the bearing vibration signal f(t) is transformed by the wavelet transform, the result is the wavelet coefficient C, which is a function of the scale a and position b. The wavelet coefficient C contains the information of the bearing vibration signal at each scale a and position b.

[0030] By constructing convolutional layers, pooling layers, RELU layers, and fully connected layers, a two-dimensional wavelet time-frequency graph feature extraction network is constructed. By changing the convolution kernel size of different convolutional layers, multiple convolution modules are constructed to perform deep feature extraction. Finally, all the extracted feature vectors are fused. The fusion formula is as follows:

[0031]

[0032] Represents the jth feature map extracted from the i-th time-frequency feature image in the convolutional layer. represents the jth channel of the fused time-frequency feature image. fuse represents the element fusion rule. Because downsampling will lose information, the two-dimensional wavelet time-frequency image feature extraction network does not perform any downsampling. The size of the feature map and the input image in the entire two-dimensional wavelet time-frequency image feature extraction network are consistent. Finally, the single channel image is superimposed three times through two layers of convolution to achieve image reconstruction.

[0033] Furthermore, in step 4, the core of the residual network is the residual block, and each residual block usually contains two or three convolutional layers. Batch normalization is often added after each convolutional layer to accelerate training and improve stability. In each residual block, the fused time-frequency feature image not only passes through the convolutional layer, but is also directly added to the output of the convolutional layer. This process is called "jump connection" or "shortcut connection". ReLU (linear rectifier unit) is usually used as an activation function to increase the nonlinearity of the residual network. The Swin Transformer network consists of multiple Swin blocks, each of which includes a window self-attention mechanism and a feedforward neural network. The fused time-frequency feature image is divided into multiple patches and then converted into sequence data through linear embedding. The Swin Transformer network adopts a hierarchical structure and performs self-attention calculations through windows of different sizes. As the depth of the Swin Transformer network increases, the window size gradually increases, thereby increasing the receptive field of the model. The window of each Swin block is mobile, and the extraction of multi-scale information of the network is enhanced by moving the window. The fused time-frequency feature image is cropped to a size of 224*224 and then input into the parallel channel feature fusion network of the residual network and the Swin Transformer network. The local features are extracted using multiple convolutional layers of the residual network and the global features are extracted using the Swin module of the Swin Transformer network. The local feature vector and the global feature vector are connected and input into the fully connected layer of the parallel channel feature fusion network for learning and classification, and finally the fault type is obtained.

[0034] Compared with the prior art, the present invention has the following technical advantages.

[0035] 1. Improve feature extraction accuracy: The crested porcupine variable mode decomposition technology can perform multi-scale and multi-modal decomposition of bearing vibration signals, improve the feature extraction capability of complex fault modes, and can more accurately capture subtle changes in bearing vibration signals compared to traditional methods.

[0036] 2. Reduce computational complexity: Through feature fusion, the features of different modes are effectively integrated, which reduces the need for complex calculation steps in time-frequency domain analysis and improves computational efficiency.

[0037] 3. Reduce data requirements: The method of the present invention can achieve higher diagnostic accuracy with relatively less labeled data by optimizing the feature extraction and fusion process, thereby reducing the dependence on a large amount of labeled data.

[0038] 4. Enhanced adaptability: Combined with advanced feature fusion and machine learning techniques, the diagnostic model of the present invention has stronger adaptability and generalization capabilities, and can effectively respond to changes in equipment working conditions and new failure modes. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Optimizing the VMD algorithm flow chart for crested porcupines

[0040] Figure 2 Optimize the variable mode decomposition iterative curve, decomposition results and Hilbert marginal spectrum for bearing vibration signals using crested porcupine

[0041] picture

[0042] Figure 3 It is a two-dimensional wavelet time-frequency graph feature fusion network diagram

[0043] Figure 4 Parallel channel feature fusion network diagram

[0044] Figure 5 The overall framework of the proposed method, including bearing vibration signal collection, bearing vibration signal preprocessing, parallel channel feature fusion network training and test results DETAILED DESCRIPTION

[0045] 1. Collect bearing vibration signals

[0046] 1. Install sensors on the bearing equipment to collect bearing vibration signals in real time. Select the specifications and installation location of the sensor to ensure the representativeness and accuracy of the bearing vibration signal.

[0047] 2. Preprocess the collected bearing vibration signals, including denoising and normalization, to ensure the quality of the bearing vibration signals.

[0048] 2. Modal Decomposition Processing of Crested Porcupine

[0049] The crown porcupine optimization algorithm is used to perform variable mode decomposition on the preprocessed bearing vibration signal. The specific steps include:

[0050] 1. Initialize the variable mode decomposition parameters and set the number of decomposition levels.

[0051] 2. Through iterative optimization algorithm, the intrinsic mode functions (IMFs) in the bearing vibration signal are extracted. Each IMF represents a different frequency component in the signal.

[0052] 3. Save the IMFs as independent signal groups for subsequent processing.

[0053] 3. Generate two-dimensional wavelet time-frequency diagram

[0054] Apply continuous wavelet transform (CWT) to each IMF and convert it into a two-dimensional wavelet time-frequency diagram. The steps include:

[0055] 1. Select a suitable wavelet function (such as Morlet wavelet) and set the scale parameters.

[0056] 2. Calculate the two-dimensional wavelet time-frequency diagram of IMF to capture the time-frequency characteristics of the bearing vibration signal.

[0057] 3. The generated two-dimensional wavelet time-frequency diagram is used as feature input for subsequent processing.

[0058] 4. Feature Fusion Processing

[0059] Convolutional neural network (CNN) is used to extract and fuse the features of two-dimensional wavelet time-frequency graph. The steps include:

[0060] 1. Design a CNN architecture, including multiple convolutional layers and pooling layers, to extract features from the time-frequency graph.

[0061] 2. The extracted features are fused through the fully connected layer to generate a comprehensive feature map.

[0062] 5. Building a parallel feature fusion network

[0063] Design a feature fusion network that combines the residual network and the Swin Transformer network. The steps include:

[0064] 1. Build a residual network to capture local features and solve the gradient disappearance problem.

[0065] 2. Build a Swin Transformer network to process global information and improve the long-term dependency ability of the model.

[0066] 3. Fuse the outputs of the two and obtain the final feature map through feature cascading or weighted fusion technology.

[0067] Network parameters are shown in the table

[0068] Table 1 Swin Transformer network parameter description

[0069]

[0070] Table 2 ResNet34 network parameter description

[0071]

[0072] 6. Training and Validation

[0073] Use the collected vibration signal data to train the network. The steps include:

[0074] 1. Divide the two-dimensional wavelet time-frequency diagram and label data into training set and validation set.

[0075] 2. Train the network using the cross entropy loss function and an appropriate optimization algorithm such as the Adam optimizer.

[0076] 3. Evaluate model performance through the validation set and adjust hyperparameters to optimize the model.

[0077] VII. Application and troubleshooting

[0078] Apply the trained model to new bearing vibration signal data for real-time fault diagnosis. The steps include:

[0079] 1. Collect new data and preprocess it.

[0080] 2. Perform variable mode decomposition, wavelet transform and feature fusion processing.

[0081] 3. Input into the trained model and output the fault diagnosis results.

[0082] 8. System maintenance and optimization

[0083] 1. Update the model regularly to adapt to new working environments or changes in working conditions.

[0084] 2. Collect data from actual applications to retrain and optimize the model and improve diagnostic accuracy.

Claims

1. A bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition, characterized in that: The method comprises the following steps: Step 1, collecting bearing vibration signals; Step 2, using the crown porcupine optimized variable mode decomposition to process the bearing vibration signal to obtain the intrinsic mode function IMFs; Step 3: convert the intrinsic mode functions IMFs into a two-dimensional wavelet time-frequency diagram through continuous wavelet transform, and perform feature fusion through a convolutional neural network to form a new two-dimensional wavelet time-frequency diagram; Step 4: Construct a feature fusion network with the residual network and Swin Transformer network as parallel channels; Step 5: Input the two-dimensional wavelet time-frequency diagram into the feature fusion network for learning and fault diagnosis.

2. A bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 1, characterized in that: In the step 1, an acceleration sensor is used to collect bearing vibration signals in three directions of x, y, and z of the axle box respectively, and after signal filtering, the bearing vibration signal of one channel is selected as a characteristic signal.

3. The bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 1 is characterized in that: In step 2, a sliding window with a length of 784 is used to intercept the one-dimensional bearing vibration signal, and then the crown porcupine optimized variable mode decomposition algorithm is used to process the bearing vibration signal to remove the noise frequency of the bearing vibration signal. Each intrinsic mode function obtained represents each frequency component in the original bearing vibration signal.

4. The bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 1 is characterized in that: In step 3, continuous wavelet transform is selected to process the intrinsic mode function so that it is converted from the time domain bearing vibration signal into the time-frequency feature image; the continuous wavelet transform continuously transforms the bearing vibration signal in the time domain without discretization processing; so that the continuous wavelet transform is more sensitive to the time domain characteristics of the signal; Wavelet functions of different scales capture the changes of bearing vibration signals in different frequency ranges; continuous wavelet transform can provide spectral information at multiple scales, thereby realizing multi-scale analysis of bearing vibration signals; a convolutional neural network is constructed, and the key features of the time-frequency feature images are extracted using the convolution layer and then fused.

5. The bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 1 is characterized in that: In step 4, a residual network and a Swin Transformer network are constructed respectively, and local features and global features in the wavelet time-frequency feature image are extracted respectively by using multi-layer convolution and multi-head self-attention, and then the local features and global features are fused and learned.

6. A bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 4, characterized in that: The mother wave function of continuous wavelet transform is cmor3-3, and the number of scales is 256.

7. The bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 5 is characterized in that: The optimizer selected for feature fusion network construction is Adam, the learning rate is set to 0.0002, the number of iterations is 50, the batch size is 32, and the loss function is CrossEntropyLoss.

8. The bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 1 is characterized in that: The decomposition process of the crown porcupine optimized variable mode is as follows: First, the population is initialized and some algorithm parameters are set; then, VMD is used to decompose the bearing vibration signal, and the fitness of each individual in the population is calculated, and the individual with the highest fitness is calculated as the current global optimal solution; The population starts to search the solution space and updates the position of each individual in the population, and then determines whether the updated position is better, until the number of iterations meets the end condition; finally, the bearing vibration signal is decomposed by VMD using the optimized parameters to obtain multiple IMFs.

9. The bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 1 is characterized in that: In step 3, the formula of continuous wavelet transform CWT is expressed as: Where f(t) is the original signal, ψ * is the complex conjugate of the wavelet function, a is the scale parameter, and b is the translation parameter; the original bearing vibration signal f is decomposed by the crown porcupine variable mode to obtain the intrinsic mode function IMF, which is the different bearing vibration signals f(t). After the bearing vibration signal f(t) is transformed by the wavelet transform, the result is the wavelet coefficient C, which is a function of the scale parameter a and the translation parameter b; the wavelet coefficient C contains the information of the bearing vibration signal on each scale parameter a and the translation parameter b; By constructing convolutional layers, pooling layers, RELU layers, and fully connected layers, a two-dimensional wavelet time-frequency graph feature extraction network is constructed; by changing the convolution kernel size of different convolutional layers, multiple convolution modules are constructed to perform deep feature extraction; finally, all the extracted feature vectors are fused, and the fusion formula is as follows: Represents the jth feature map extracted from the i-th time-frequency feature image in the convolutional layer; Represents the jth channel of the fused time-frequency feature image; fuse represents the element fusion rule; the size of the feature map and the input image in the entire two-dimensional wavelet time-frequency image feature extraction network are consistent, and finally, the single channel image is superimposed three times through two layers of convolution to achieve image reconstruction.

10. The bearing fault diagnosis method based on crested porcupine optimized variable mode decomposition according to claim 1 is characterized in that: In step 3 and step 4, the core of the residual network is the residual block, each residual block contains two or three convolutional layers; batch normalization is often added after each convolutional layer; in each residual block, the fused time-frequency feature image not only passes through the convolutional layer, but is directly added to the output of the convolutional layer, that is, "jump connection" or "quick connection"; the linear rectifier unit ReLU is used as the activation function to increase the nonlinearity of the residual network; the Swin Transformer network is composed of multiple Swin blocks, each of which includes a window self-attention mechanism and a feedforward neural network; the fused time-frequency feature image is divided into multiple patches, and then converted into sequence data through linear embedding; the Swin Transformer network adopts a hierarchical structure and performs self-attention calculations through windows of different sizes; as the depth of the Swin Transformer network increases, the window size gradually increases and the receptive field increases; the window of each Swin block is mobile, and the extraction of multi-scale information of the network is enhanced by moving the window; the fused time-frequency feature image is cropped to a size of 224*224 and then input into the residual network and Swin In the parallel channel feature fusion network of the Transformer network, multiple convolutional layers of the residual network are used to extract local features and the Swin module of the Swin Transformer network is used to extract global features. The local feature vector and the global feature vector are connected and input into the fully connected layer of the parallel channel feature fusion network for learning and classification, and finally the bearing fault type is obtained.

Images