CNN-LSTM time-frequency fusion-based rolling bearing fault diagnosis method and system, computer equipment and storage medium

Abstract

The invention discloses a CNN-LSTM time-frequency fusion-based rolling bearing fault diagnosis method and system, computer equipment and a storage medium. The CNN-LSTM time-frequency fusion-based rolling bearing fault diagnosis method comprises the following steps of 1, collecting a vibration signal in an operation process of a to-be-detected bearing; step 2, inputting the vibration signal into a trained time-frequency fusion CNN-LSTM network model to obtain a fault type of the bearing to be tested; the time-frequency fused CNN-LSTM network model comprises a fast Fourier transform, a convolutional neural network, a long short-term memory network and a full connection layer which are connected in sequence; according to the method, the convolutional neural network and the long-short-term memory network are combined, and the time domain and frequency domain features of the bearing signal are automatically extracted, so that fault feature extraction and bearing fault diagnosis are realized, and accurate, rapid and reliable bearing fault diagnosis is carried out on the rolling bearing.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling bearing fault diagnosis, and particularly to a rolling bearing fault diagnosis method, system, computer device and storage medium based on CNN-LSTM time-frequency fusion. Background Art

[0002] Rolling bearings are one of the most important mechanical components in rotating mechanical equipment. Under difficult working conditions such as high load and high impact, rolling bearings are prone to various faults, which can lead to damage or even shutdown of the entire equipment. The faults generated by rolling bearings not only cause economic losses but also pose safety hazards. Therefore, accurate, rapid and reliable bearing fault diagnosis for rolling bearings is of great importance.

[0003] In recent years, with the rapid development of deep learning, its advantage is that it integrates feature extraction and recognition classification, and can directly extract features from signals and classify fault types. This not only greatly simplifies the diagnosis process but also reduces the difficulty of building a model. Therefore, methods based on deep learning have been widely used in the field of rolling bearing fault detection. However, existing methods only extract features through the time domain or frequency domain, which may lead to insufficient feature extraction and thus affect the final recognition success rate. Summary of the Invention

[0004] Object of the Invention: To solve the problem that existing rolling bearing fault detection methods only extract features through the time domain or frequency domain, resulting in insufficient feature extraction and low recognition success rate, the present invention proposes a rolling bearing fault diagnosis method, system, computer device and storage medium based on CNN-LSTM time-frequency fusion.

[0005] Technical Solution: A rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion includes the following steps:

[0006] Step 1: Collect vibration signals during the operation of the bearing to be measured;

[0007] Step 2: Input the vibration signals into a trained CNN-LSTM network model with time-frequency fusion to obtain the fault type of the bearing to be measured;

[0008] The CNN-LSTM network model with time-frequency fusion includes a fast Fourier transform, a convolutional neural network, a long short-term memory network and a fully connected layer connected in sequence.

[0009] Further, in the CNN-LSTM network model with time-frequency fusion, the following operations are performed:

[0010] Perform a fast Fourier transform on the vibration signals to obtain frequency domain signals from the vibration signals;

[0011] Obtain the time-domain signal from the vibration signal;

[0012] Input the time-domain signal and the frequency-domain signal into their respective convolutional neural networks respectively to extract their respective local features;

[0013] Use a long short-term memory network to capture the global features of the time-domain signal and the global features of the frequency-domain signal from the local features of the time-domain signal and the local features of the frequency-domain signal;

[0014] Fuse and splice the global features of the time-domain signal and the global features of the frequency-domain signal, and perform linear transformation and non-linear activation through a fully connected layer to obtain the probability labels of each category.

[0015] Further, performing a fast Fourier transform on the vibration signal to obtain a frequency-domain signal from the vibration signal includes: using a divide-and-conquer method to perform a fast Fourier transform to obtain the frequency-domain signal, where the divide-and-conquer method is to divide the vibration signal into an odd part and an even part, recursively perform Fourier transforms on the odd part and the even part respectively, and finally merge the transformation results of the odd part and the even part.

[0016] Further, the convolutional neural network includes three stacked CNN modules, and each CNN module is composed of a one-dimensional convolutional layer, a max pooling layer, and a batch normalization layer.

[0017] Further, each one-dimensional convolutional layer uses the same convolutional kernel size, stride, and activation function, but the number of filters used in the three stacked CNN modules gradually increases.

[0018] Further, for the frequency-domain signal, the filter sizes in its convolutional neural network are 16, 32, and 64 respectively; for the time-domain signal, the filter sizes in its convolutional neural network are 32, 64, and 128 respectively.

[0019] The present invention also discloses a rolling bearing fault diagnosis system based on CNN-LSTM time-frequency fusion, including:

[0020] A vibration signal acquisition module for acquiring the vibration signal during the operation of the bearing to be measured;

[0021] A diagnosis module for inputting the vibration signal into the trained time-frequency fusion CNN-LSTM network model to obtain the fault type of the bearing to be measured;

[0022] The time-frequency fusion CNN-LSTM network model includes a fast Fourier transform, a convolutional neural network, a long short-term memory network, and a fully connected layer connected in sequence.

[0023] Further, in the time-frequency fusion CNN-LSTM network model, the following operations are performed:

[0024] Perform a fast Fourier transform on the vibration signal to obtain a frequency-domain signal from the vibration signal;

[0025] Obtain a time-domain signal from the vibration signal;

[0026] Input the time-domain signal and the frequency-domain signal into their respective convolutional neural networks to extract their respective local features;

[0027] Use a long short-term memory network to capture the global features of the time-domain signal and the global features of the frequency-domain signal from the local features of the time-domain signal and the local features of the frequency-domain signal;

[0028] Fuse and splice the global features of the time-domain signal and the global features of the frequency-domain signal, and perform linear transformation and non-linear activation through a fully connected layer to obtain probability labels for each category.

[0029] The present invention discloses a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion.

[0030] The present invention discloses a storage medium storing a rolling bearing fault diagnosis program. When the rolling bearing fault diagnosis program is executed by at least one processor, it implements the steps of a rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion.

[0031] Advantageous effects: Compared with the prior art, the present invention has the following advantages:

[0032] The method of the present invention combines a convolutional neural network and a long short-term memory network to automatically extract the time-domain and frequency-domain features of bearing signals, thereby realizing fault feature extraction and bearing fault diagnosis, and performing accurate, fast, and reliable bearing fault diagnosis on rolling bearings. Description of the Drawings

[0033] Figure 1 It is a flowchart of the rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion proposed by the present invention;

[0034] Figure 2 It is a time-frequency fusion CNN-LSTM network model;

[0035] Figure 3 It is an LSTM network architecture diagram. Detailed Embodiments

[0036] Now, the technical solutions of this embodiment will be further elaborated in combination with the drawings and embodiments.

[0037] As Figure 1 shown, this embodiment proposes a rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion, which mainly includes the following steps:

[0038] Step 1: Collect the vibration signals during the operation of the bearing and preprocess the vibration signals to construct the training set and test set of the vibration signals. The specific operations include:

[0039] Set the sampling parameters, where the sampling parameters include the sampling frequency f and the sampling duration t. Use an accelerometer to sample the acceleration signal received by the bearing and collect the sample data during the operation of the bearing;

[0040] Among them, the sample data is obtained by artificial fault injection, mainly including normal bearings and fault bearings. The fault bearings include 9 different types of fault bearings, which are classified according to different fault positions and different damage diameters. Specifically: sample bearing 0 is a normal bearing, and the distribution of 9 fault bearings is as follows: sample bearing 1 is an inner ring fault bearing with a damage diameter of R1, sample bearing 2 is an inner ring fault bearing with a damage diameter of R2, sample bearing 3 is an inner ring fault bearing with a damage diameter of R3, sample bearing 4 is a rolling element fault bearing with a damage diameter of R1, sample bearing 5 is a rolling element fault bearing with a damage diameter of R2, sample bearing 6 is a rolling element fault bearing with a damage diameter of R3, sample bearing 7 is an outer ring fault bearing with a damage diameter of R1, sample bearing 8 is an outer ring fault bearing with a damage diameter of R2, and sample bearing 9 is an outer ring fault bearing with a damage diameter of R3. Among them, R1 is mild fault data, R2 is moderate fault data, and R3 is severe fault data.

[0041] Output the results of the accelerometer sampling in numerical form, save them in CSV format, place the vibration signal of the normal bearing sample 0 in the first column of the CSV file, place sample bearing 1 in the second column of the CSV file, and so on, for a total of ten columns, to complete the data saving.

[0042] Perform segmentation processing on the obtained CSV file, divide it with a window step of 1024. If each column of sample data has N, then 10*N / 1024 groups of samples are obtained, and the ten types of sample data are divided, and the training set and test set are divided according to a ratio of 7:3.

[0043] Step 2: Construct a time-frequency fusion CNN-LSTM network model, which includes four parts: fast Fourier transform, convolutional neural network, long short-term memory network, and fully connected layer. Use this model for fault diagnosis, and the time-frequency fusion CNN-LSTM network model is as Figure 2 shown. The specific construction process includes:

[0044] S2_1: Perform a fast Fourier transform on the sample to obtain a frequency-domain signal from the one-dimensional time series, specifically including:

[0045] Since the original bearing vibration signal is essentially a one-dimensional time series signal, the obtained original bearing vibration signal can be expressed as: X(n) = [x1, x2,..., x n ∈ R n , where n represents the number of sample data, and x n represents the amplitude of the sample data. Perform a fast Fourier transform on it using the divide-and-conquer method to obtain its frequency-domain signal. The divide-and-conquer method divides the signal into odd and even parts, then recursively performs Fourier transforms on these two groups of data respectively, and finally obtains the final solution by combining the transformation results of the even and odd parts.

[0046] For a sequence x(n) of length N, where N is a power of 2, its fast Fourier transform formula based on the divide-and-conquer method is expressed as:

[0047]

[0048] In the formula, W N = e -j2π / N is the rotation factor, and k = 0, 1, 2,..., N - 1 are the sampling points.

[0049] S2_2: Extract the local features of the sequence from the processed time-domain signal and frequency-domain signal, specifically including:

[0050] Extract the local features of the respective sequences from the obtained time-domain and frequency-domain signals through their respective convolutional neural networks. The convolutional neural network contains three stacked CNN modules, respectively denoted as Conv1, Conv2, and Conv3. Each CNN module consists of a one-dimensional convolutional layer, a max-pooling layer, and a batch normalization layer. The output of the one-dimensional convolutional layer is expressed as:

[0051]

[0052] In the formula, is the one-dimensional input of the l - 1 layer, is the convolutional kernel, b l is the bias, and σ(·) is the activation function.

[0053] Each convolutional layer uses the same convolutional kernel size, stride, and activation function, but as the depth increases, the number of filters gradually increases. The convolutional kernel size is 3, the stride is 1, and the ReLU function is used as the activation function. The filter sizes of the time-domain convolutional neural network are 32, 64, and 128 respectively, and the filter sizes of the frequency-domain convolutional neural network are 16, 32, and 64 respectively. The parameters of the max-pooling layer are all the same, the convolutional kernel size is 2, the stride is 3, and the ReLU function is used as the activation function.

[0054] S2_3: Extract the global features of the sequence from the time-domain signal and frequency-domain signal processed by the CNN, specifically including:

[0055] Use a long short-term memory network to extract the global features of the sequence from the time-domain signal and frequency-domain signal processed by the CNN. The LSTM network architecture is as Figure 3 shown. The LSTM cell updates the memory cell and hidden state through the combined action of the forget gate, input gate, and output gate. The working process of the LSTM cell is as follows: The forget gate determines what information to discard based on the previous hidden state and the current input; the input gate determines which new information to control and add to the memory cell; based on the outputs of the forget gate and input gate, the state of the memory cell is updated; finally, the output gate calculates the final output based on the current state of the memory cell. The formula of the LSTM cell is as follows:

[0056] i t = σ s (W i x t + U i h t-1 + b i );

[0057] f t = σ s (W f x t + U f h t-1 + b f );

[0058] o t = σ s (W o x t + U o h t-1 + b o );

[0059]

[0060] h t = o t * σ tanh (c t );

[0061] In the formula, x and h represent the input and hidden states, t and t - 1 are time steps, i t , f, and o t are the input gate, forget gate, and output gate, and c t are the cell input vector and state vector. σ s and σ tanhThey are the sigmoid and tanh activation functions, the weights W and U in different gates, and b is the bias.

[0062] S2_4: Fuse and splice the time-domain and frequency-domain features captured by the long short-term memory network and output the final result, specifically including:

[0063] Fuse and splice the time-domain and frequency-domain features captured by the feature extraction layer and the multi-head attention layer. The fused features will be input into the fully connected layer. The fully connected layer performs linear transformation and non-linear activation on the integrated features and finally outputs the probability labels of 10 categories. The activation function uses the Softmax function to convert the output values into a probability distribution, making the predicted values of the 10 categories all between 0 and 1, and the sum of all predicted values is 1.

[0064] Step 3: Train the CNN-LSTM model with the training set of vibration signals to obtain the trained model and parameters. The specific operations include:

[0065] Initialize the parameters of the CNN-LSTM model;

[0066] For each training sample, the model performs forward propagation through the input data, calculates the predicted result of the model, compares the predicted result with the true label, and uses the cross-entropy loss function to calculate the gap between the predicted result and the actual value;

[0067] Use the backpropagation algorithm to calculate the gradient of the loss function with respect to each parameter. In backpropagation, the calculation starts from the output layer first, propagates the error layer by layer forward, and then calculates the gradient of each layer through the chain rule. Update the parameters with the Adam optimization algorithm. The goal is to gradually reduce the value of the loss function, that is, to improve the performance of the model by updating the parameters.

[0068] Step 4: Input the test data set into the trained CNN-LSTM model to obtain the fault types of the bearing signals to be measured. The specific operations include:

[0069] Input the test set into the trained CNN-LSTM model, output the probabilities of the samples belonging to different fault categories, take the fault category with the highest probability as the result of the rolling bearing fault diagnosis, and calculate evaluation indicators such as accuracy and F1 score at the same time, and draw a confusion matrix to intuitively display the predicted results of the model on each category.

Claims

1. A rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion, characterized by: The following steps are involved: Step 1: Collect the vibration signal of the bearing to be tested during operation; Step 2: Input the vibration signal into the trained time-frequency fusion CNN-LSTM network model to obtain the fault type of the bearing to be tested; The time-frequency fused CNN-LSTM network model includes a fast Fourier transform, a convolutional neural network, a long short-term memory network and a fully connected layer connected in sequence.

2. The rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion according to claim 1 is characterized in that: In the time-frequency fused CNN-LSTM network model, the following operations are performed: Perform fast Fourier transform on the vibration signal to obtain a frequency domain signal from the vibration signal; Obtaining time domain signals from vibration signals; Input the time domain signal and frequency domain signal into their respective convolutional neural networks to extract their respective local features; A long short-term memory network is used to capture the global features of the time domain signal and the global features of the frequency domain signal from the local features of the time domain signal and the local features of the frequency domain signal; The global features of the time domain signal and the global features of the frequency domain signal are fused and spliced, and linear transformation and nonlinear activation are performed through the fully connected layer to obtain the probability labels of each category.

3. The rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion according to claim 2 is characterized in that: The method of performing fast Fourier transform on the vibration signal to obtain the frequency domain signal from the vibration signal includes: using a divide-and-conquer method to perform fast Fourier transform to obtain the frequency domain signal, the divide-and-conquer method is to divide the vibration signal into an odd part and an even part, recursively perform Fourier transform on the odd part and the even part respectively, and finally merge the transform results of the odd part and the even part.

4. The rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion according to claim 2 is characterized in that: The convolutional neural network includes three stacked CNN modules, each of which consists of a one-dimensional convolution layer, a maximum pooling layer, and a batch normalization layer.

5. The rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion according to claim 4 is characterized in that: Each 1D convolutional layer uses the same kernel size, stride, and activation function, but the number of filters adopted in the three stacked CNN modules gradually increases.

6. The rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion according to claim 5 is characterized in that: For frequency domain signals, the filter sizes in its convolutional neural network are 16, 32, and 64 respectively; for time domain signals, the filter sizes in its convolutional neural network are 32, 64, and 128 respectively.

7. A rolling bearing fault diagnosis system based on CNN-LSTM time-frequency fusion, characterized by: include: A vibration signal acquisition module is used to collect vibration signals during the operation of the bearing to be tested; The diagnosis module is used to input the vibration signal into the trained time-frequency fusion CNN-LSTM network model to obtain the fault type of the bearing to be tested; The time-frequency fused CNN-LSTM network model includes a fast Fourier transform, a convolutional neural network, a long short-term memory network and a fully connected layer connected in sequence.

8. The rolling bearing fault diagnosis system based on CNN-LSTM time-frequency fusion according to claim 7 is characterized in that: In the time-frequency fused CNN-LSTM network model, the following operations are performed: Perform fast Fourier transform on the vibration signal to obtain a frequency domain signal from the vibration signal; Obtaining time domain signals from vibration signals; Input the time domain signal and frequency domain signal into their respective convolutional neural networks to extract their respective local features; A long short-term memory network is used to capture the global features of the time domain signal and the global features of the frequency domain signal from the local features of the time domain signal and the local features of the frequency domain signal; The global features of the time domain signal and the global features of the frequency domain signal are fused and spliced, and linear transformation and nonlinear activation are performed through the fully connected layer to obtain the probability labels of each category.

9. A computer device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion as described in any one of claims 1 to 6 are implemented.

10. A storage medium, characterized in that: The storage medium stores a rolling bearing fault diagnosis program, and when the rolling bearing fault diagnosis program is executed by at least one processor, the steps of a rolling bearing fault diagnosis method based on CNN-LSTM time-frequency fusion as described in any one of claims 1 to 6 are implemented.

Images