Bearing fault diagnosis method and device and computer readable storage medium

Abstract

The invention provides a bearing fault diagnosis method and device and a computer readable storage medium, and the method comprises the steps: carrying out the sampling of original vibration data of a bearing under different working conditions, and dividing the collected original sample data into a training set and a test set; performing CEEMDAN decomposition and wavelet decomposition on the data, and constructing a sample feature set; performing high-dimensional mapping on the sample feature set by using a kernel Fisher discrimination analysis method to obtain a mapped training set and test set; performing iterative training on the DNN model through the training set obtained through mapping, and optimizing model parameters; performing identification accuracy evaluation on the DNN model by using the test set obtained by mapping, and outputting a fault diagnosis model; and acquiring vibration data of the bearing under different working conditions, inputting the vibration data into the fault diagnosis model, and acquiring a diagnosis result of the bearing fault. The problems that in the prior art, the bearing fault type recognition accuracy is not high, and potential faults of the bearing cannot be found in time are solved.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a bearing fault diagnosis method, apparatus and computer-readable storage medium. Background Technology

[0002] Rolling bearings, as key components of rotating machinery such as hydraulic turbines, play a vital role in power generation and other equipment. Stable bearing performance is crucial for maintaining efficient equipment operation and ensuring operational safety. However, in practical applications, as bearing operating time increases, their mechanical properties gradually degrade, often leading to typical failures such as pitting, tooth breakage, and wear. Furthermore, mechanical equipment, especially large-scale equipment, has a high degree of coupling between components, and various failure factors exhibit complex correlations. Therefore, accurate and efficient diagnosis of rolling bearing failure modes is of significant practical value for improving equipment reliability and conducting targeted bearing maintenance.

[0003] Traditional diagnostic methods for rolling bearings primarily rely on single dimensional ratios to determine faults. However, this approach is less effective when dealing with complex, multi-dimensional faults in mechanical structures. With technological advancements, researchers have discovered that when a rolling bearing experiences a localized fault, its vibration signal exhibits nonlinear and non-stationary characteristics, generating periodic pulse impacts. Therefore, by decomposing and extracting features from these signals, fault identification and diagnosis can be effectively achieved. Based on this, methods such as Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), and Adaptive Noise-Complete Ensemble Empirical Mode Decomposition (CEEMDAN) have been proposed in recent years.

[0004] Empirical Mode Decomposition (EMD) decomposes a signal into a series of intrinsic mode components (IMFs), enabling detailed analysis of the signal's informational characteristics, particularly suitable for non-stationary signals. However, EMD may introduce mode aliasing during the decomposition process, causing components of different frequencies to be incorrectly decomposed into the same IMF, thus affecting the accurate extraction of fault features. To address this issue, Ensemble Empirical Mode Decomposition (EEMD) incorporates white noise during the EMD process to reduce mode aliasing and improve decomposition accuracy. However, the reconstructed signal after EEMD decomposition may contain residual noise, which increases reconstruction error and reduces the integrity of the signal decomposition.

[0005] Although the adaptive noise-integrated empirical mode decomposition (CEEMDAN) can further optimize the signal decomposition process and improve the accuracy and robustness of the decomposition by adaptively selecting noise, the problem of feature boundary aliasing still exists after vibration signal decomposition. This results in unclear feature distinction, low accuracy in identifying bearing fault types, and inability to detect potential bearing faults in a timely manner. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a bearing fault diagnosis method, device, and computer-readable storage medium. It solves the problem that after the vibration signal is decomposed, the feature distinction is not obvious, resulting in low accuracy in identifying bearing fault types and the inability to detect potential bearing faults in a timely manner.

[0007] On one hand, embodiments of the present invention provide a bearing fault diagnosis method, which includes the following steps:

[0008] Step S100: Sample the original vibration data of the bearing under different working conditions, and divide the collected original sample data into training set and test set according to a preset ratio;

[0009] Step S200: Perform CEEMDAN decomposition and wavelet decomposition on the original sample data in the training set and test set respectively to construct the corresponding sample feature sets of the training set and test set.

[0010] Step S300: Use the kernel Fisher discriminant analysis method to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set to obtain the mapped training set and test set;

[0011] Step S400: Input the training set obtained through mapping into the DNN model, iteratively train the DNN model, and optimize the model parameters;

[0012] Step S500: Use the test set obtained from the mapping to evaluate the recognition accuracy of the trained DNN model and output the fault diagnosis model.

[0013] Step S600: Obtain vibration data of the bearing under different operating conditions, input the sample features obtained after executing steps S200-S300 into the fault diagnosis model, and obtain the diagnosis results of bearing fault.

[0014] Preferably, in step S200, the original sample data in the training set and the test set are subjected to CEEMDAN decomposition and wavelet decomposition respectively to construct the corresponding sample feature sets of the training set and the test set, which specifically includes the following steps:

[0015] Step S201: Normalize the original sample data in the training set and the test set;

[0016] Step S202: Perform CEEMDAN decomposition on the normalized data to obtain several time-series signals;

[0017] Step S203: Perform wavelet decomposition on each of the time-series signals to obtain approximate and detail signals at different scales;

[0018] Step S204: Calculate the signal energy for the approximate and detail signals at each scale obtained by wavelet decomposition;

[0019] Step S205: Use the calculated energy values ​​as features and integrate them into a feature vector to form the sample feature set of the training set and the test set.

[0020] Preferably, in step S202, the CEEMDAN decomposition of the normalized data to obtain several time-series signals includes:

[0021] Step S2021: Add Gaussian white noise following a normal distribution to the normalized original sample data x(t) to obtain noisy data x. i (t); x i (t)=x(t)+εω i (t)i=1,2,3,…,N,where N represents the number of times white noise is added and ε represents the noise standard deviation;

[0022] Step S2022: Use the CEEMDAN algorithm to process the noisy data x. i (t) is decomposed separately to obtain each noisy data x. i (t) multi-frequency component IMF i,1 (t), calculate the first multi-frequency component IMF1(t) of the original sample data x(t); Where T represents the number of decompositions;

[0023] Step S2023: Calculate the residual r1(t) of the original sample data x(t) based on the first multi-frequency component IMF1(t); r1(t) = x(t) - IMF1(t);

[0024] Step S2024, E i [ω i [(t)] is defined as the i-th multi-frequency component after decomposition, ε j To find the noise standard deviation, we decompose r1(t)+ε1E1[ω1(t)] to obtain the second multi-frequency component IMF2(t);

[0025]

[0026] Where I represents the number of decompositions;

[0027] Step S2025: Calculate the Kth residual r of the original sample data x(t). k (t); r k (t)=r k-1 (t)-IMF k (t), k = 2, 3, ..., K;

[0028] Step S2026: Calculate the (k+1)th multi-frequency component IMF. k+1 (t);

[0029] Step S2027: Repeat steps S2024-S2026 until the remaining multi-frequency components can no longer be decomposed, and the remaining residual balance R(t) satisfies At that time, the final decomposition of the time series signal x(t) is obtained;

[0030]

[0031] Preferably, in step S203, each time-series signal is decomposed using the following formula to obtain approximate and detail signals at different scales:

[0032]

[0033] Among them, A j D represents an approximate signal, used to reflect the overall trend of a signal sequence; j The detail signal is used to reflect random perturbations in different frequency domains of the original signal; k represents the time-shift discretization of the integer node marker, and h k-2n With g k-2n denoted as the transmission coefficients of the orthogonal wavelet, and j represents the decomposition scale.

[0034] Preferably, in step S204, the signal energy is calculated for the approximate and detail signals at each scale obtained by wavelet decomposition using the following formula:

[0035]

[0036] Among them, EA j ED represents the approximate signal energy. j Let A represent the detail signal, j represent the decomposition scale, and A represent the decomposition scale. j D represents an approximate signal. j This indicates a detail signal.

[0037] Preferably, in step S300, the step of using the kernel Fisher discriminant analysis method to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set to obtain the mapped training set and test set includes:

[0038] Step S301: Define the sample feature set as containing D-dimensional features and a total of C categories; where, in the D-dimensional feature samples, the samples belonging to C1 and C2 are respectively x1={x1,x2,…,x n1}、x2={x1,x2,…,x n2}, n1 + n2 = n, where n represents the total number of samples;

[0039] Step S302: Through nonlinear mapping Map each sample feature to a high-dimensional feature space;

[0040] Step S303: In the high-dimensional feature space, calculate the mean vector of each category. Where i represents different categories and i = 1, 2; j = 1, 2, ..., n i ;

[0041] Step S303: Construct the inter-class scatter matrix between the two categories.

[0042] Step S304: Construct the intra-class scatter matrix for each category.

[0043] Step S305: Maximize the objective function J f (ω) Find the optimal projection direction ω opt ;

[0044] Step S306: Map the sample features of the training set and test set to a high-dimensional space, and along the optimal projection direction ω. opt Projection is performed to obtain the mapped training and test sets.

[0045] Preferably, in step S400, the mapped training set is input into the DNN model through the following steps to iteratively train the DNN model and optimize the model parameters:

[0046] Step S401: Initialize network parameters: Set the number of hidden layers L, and determine the number of nodes p in each layer. l (l = 1, 2, ..., L), initialize the weights W between the hidden layer and the output layer. l and offset vector b l (l=2,3,…,L), set the learning rate η and the iteration threshold ε, and select the neuron activation function f(·);

[0047] Step S402: Calculate the output H of the hidden layer and the output layer. l H l =f(z) l )=f(W l Y l-1 +bl ), l=2,3,…,L; where H1 represents the output from the input layer to the first hidden layer, H l This represents the final output of the network.

[0048] Step S403: Calculate the loss function J and the output layer gradient δ L ;

[0049] Where X represents the input data and Y represents the actual label;

[0050]

[0051] Step S404: Backpropagate from the output layer to the hidden layer and calculate the hidden layer gradient δ. l ;

[0052]

[0053] Step S405: Based on the input training set, repeat steps S402-S404 until the preset number of iterations is reached or the error constraint condition is met, and end the training process; otherwise, return to step S402 and continue training.

[0054] Preferably, in step S401, the neuron activation function f(·) includes the ReLU function and the tanh function.

[0055] Secondly, another embodiment of the present invention provides a bearing fault diagnosis device that performs a bearing fault diagnosis method as described in any one of the above technical solutions, characterized in that it includes:

[0056] The sampling module is used to sample the original vibration data of the bearing under different operating conditions, and divide the collected original sample data into training set and test set according to a preset ratio;

[0057] The feature extraction module is used to perform CEEMDAN decomposition and wavelet decomposition on the original sample data in the training set and the test set, respectively, to construct the corresponding sample feature sets of the training set and the test set.

[0058] The mapping module is used to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set using the kernel Fisher discriminant analysis method to obtain the mapped training set and test set;

[0059] The training module is used to input the training set obtained through mapping into the DNN model, perform iterative training on the DNN model, and optimize the model parameters.

[0060] The testing module is used to measure the recognition accuracy of the trained DNN model using the test set obtained by mapping, and to obtain the final fault diagnosis model.

[0061] The results output module is used to acquire vibration data of the bearing under different operating conditions, input the sample features obtained after executing steps S200-S300 into the fault diagnosis model, and output the bearing fault diagnosis results.

[0062] Thirdly, another embodiment of the present invention provides a computer-readable storage medium, characterized in that it stores a processor-executable program, which, when executed by a processor, is used to implement a bearing fault diagnosis method as described in any of the above technical solutions.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] This invention employs the CEEMDAN algorithm to effectively process nonlinear and non-stationary vibration signals, significantly improving the quality of signal decomposition. Combined with kernel Fisher feature extraction technology, it accurately captures the most representative features, thereby greatly improving the accuracy of bearing fault diagnosis. Compared with traditional methods, this invention is more accurate in identifying bearing fault types and significantly reduces the false diagnosis rate. Furthermore, by utilizing CEEMDAN decomposition and kernel Fisher feature extraction technology, this invention not only reduces the feature dimensionality but also effectively removes redundant information, making the feature space more concise. This greatly improves the efficiency of rolling bearing fault diagnosis, providing strong support for timely maintenance. This early warning mechanism significantly reduces equipment downtime caused by faults, significantly lowers maintenance costs, and enhances equipment reliability and stability.

[0065] This invention enables continuous monitoring of bearing status, timely detection of abnormalities and early warning. By promptly identifying potential bearing problems, it prevents further escalation of faults and equipment damage, ensuring normal equipment operation and extending equipment lifespan, thereby bringing long-term economic benefits. Attached Figure Description

[0066] Figure 1 This is a flowchart of the steps of a bearing fault diagnosis method proposed in this invention; Figure 2 This is a schematic diagram of the bearing fault diagnosis process proposed in this invention; Figure 3 This is a schematic diagram of the CEEMDAN decomposition process for data proposed in this invention. Figure 4 This is a network topology diagram of the DNN model proposed in this invention; Figure 5This is a schematic diagram of the frame structure of a bearing fault diagnosis device proposed in this invention. Detailed Implementation

[0071] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0072] like Figure 1 As shown in the figure, an embodiment of the present invention provides a bearing fault diagnosis method, including the following steps:

[0073] Step S100: Sample the original vibration data of the bearing under different working conditions, and divide the collected original sample data into training set and test set according to a preset ratio;

[0074] Step S200: Perform CEEMDAN decomposition and wavelet decomposition on the original sample data in the training set and test set respectively to construct the corresponding sample feature sets of the training set and test set.

[0075] Step S300: Use the kernel Fisher discriminant analysis method to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set to obtain the mapped training set and test set;

[0076] Step S400: Input the training set obtained through mapping into the DNN model, iteratively train the DNN model, and optimize the model parameters;

[0077] Step S500: Use the test set obtained from the mapping to evaluate the recognition accuracy of the trained DNN model and output the fault diagnosis model.

[0078] Step S600: Obtain vibration data of the bearing under different operating conditions, input the sample features obtained after executing steps S200-S300 into the fault diagnosis model, and obtain the diagnosis results of bearing fault.

[0079] In a specific embodiment, refer to Figure 2 , Figure 2This is a schematic diagram of the bearing fault diagnosis process proposed in this invention. In actual production, high-precision equipment, such as high-precision vibration sensors, is used to monitor the operating bearings in real time, collecting vibration signals under different operating conditions (such as normal operation, overload, and insufficient lubrication). The collected bearing states include normal, external crack, internal crack, wear, and missing bearings. The collected raw sample data is divided into an 80% training set and a 20% test set. Then, CEEMDAN decomposition is performed on each sample data in both the training and test sets, decomposing the signal into several intrinsic mode functions (IMFs) to better handle nonlinear and non-stationary signals. Wavelet decomposition is then performed on each IMF to calculate the wavelet energy at each scale, constructing a feature vector. Using kernel Fisher discriminant analysis, the feature vectors of the training and test sets are mapped in high dimension to find the optimal projection direction, achieving feature mapping. This effectively improves the feature resolution in the fault diagnosis stage, removes redundant information, simplifies the feature space, and improves the efficiency of rolling bearing fault diagnosis in mechanical equipment. Finally, a deep neural network (DNN) model is used to train the model and diagnose faults. Based on the diagnosis results, timely maintenance measures are taken, such as lubrication and bearing replacement, to avoid equipment failure.

[0080] Preferably, in step S200, the original sample data in the training set and the test set are subjected to CEEMDAN decomposition and wavelet decomposition respectively to construct the corresponding sample feature sets of the training set and the test set, which specifically includes the following steps:

[0081] Step S201: Normalize the original sample data in the training set and the test set;

[0082] Step S202: Perform CEEMDAN decomposition on the normalized data to obtain several time-series signals;

[0083] Step S203: Perform wavelet decomposition on each of the time-series signals to obtain approximate and detail signals at different scales;

[0084] Step S204: Calculate the signal energy for the approximate and detail signals at each scale obtained by wavelet decomposition;

[0085] Step S205: Use the calculated energy values ​​as features and integrate them into a feature vector to form the sample feature set of the training set and the test set.

[0086] In this embodiment, the original vibration data in the training and test sets are normalized to eliminate the dimensional influence between different signals. Then, the CEEMDAN algorithm is applied to the normalized data to decompose it into several time-series signals, namely Intrinsic Mode Functions (IMFs). Wavelet decomposition is performed on each time-series signal to obtain approximate signals and detail signals at different scales, thereby refining the frequency characteristics of the signal. The signal energy of each approximate signal and detail signal obtained from wavelet decomposition at each scale is calculated to quantify the signal intensity in different frequency ranges. The calculated energy values ​​are used as features and integrated into a feature vector containing energy information of the signal in different frequency bands.

[0087] like Figure 3 As shown, Figure 3 This is a schematic diagram of the CEEMDAN decomposition process proposed in this invention.

[0088] In step S202, the normalized data is decomposed using CEEMDAN to obtain several time-series signals, including:

[0089] Step S2021: Add Gaussian white noise following a normal distribution to the normalized original sample data x(t) to obtain noisy data x. i (t); x i (t)=x(t)+εω i (t)i=1,2,3,…,N,where N represents the number of times white noise is added and ε represents the noise standard deviation;

[0090] Step S2022: Use the CEEMDAN algorithm to process the noisy data x. i (t) is decomposed separately to obtain each noisy data x. i (t) multi-frequency component IMF i,1 (t), calculate the first multi-frequency component IMF1(t) of the original sample data x(t); Where T represents the number of decompositions;

[0091] Step S2023: Calculate the residual r1(t) of the original sample data x(t) based on the first multi-frequency component IMF1(t); r1(t) = x(t) - IMF1(t);

[0092] Step S2024, E i [ω i [(t)] is defined as the i-th multi-frequency component after decomposition, ε j To find the noise standard deviation, we decompose r1(t)+ε1E1[ω1(t)] to obtain the second multi-frequency component IMF2(t);

[0093]

[0094] Where I represents the number of decompositions;

[0095] Step S2025: Calculate the Kth residual r of the original sample data x(t). k (t); r k (t)=r k-1 (t)-IMF k (t), k = 2, 3, ..., K;

[0096] Step S2026: Calculate the (k+1)th multi-frequency component IMF. k+1 (t);

[0097] Step S2027: Repeat steps S2024-S2026 until the remaining multi-frequency components can no longer be decomposed, and the remaining residual balance R(t) satisfies At that time, the final decomposition of the time series signal x(t) is obtained;

[0098]

[0099] Preferably, in step S203, each time-series signal is decomposed using the following formula to obtain approximate and detail signals at different scales:

[0100]

[0101] Among them, A j D represents an approximate signal, used to reflect the overall trend of a signal sequence; j The detail signal is used to reflect random perturbations in different frequency domains of the original signal; k represents the time-shift discretization of the integer node marker, and h k-2n With g k-2n denoted as the transmission coefficients of the orthogonal wavelet, and j represents the decomposition scale.

[0102] In this embodiment, wavelet decomposition of each time series signal is the process of decomposing the time series signal from high resolution to low resolution. The time series signal is decomposed step by step from scale j+1 to scale j, and finally the original time series signal is decomposed into J+1 groups of sequence signals with different characteristics.

[0103] Preferably, in step S204, the signal energy is calculated for the approximate and detail signals at each scale obtained by wavelet decomposition using the following formula:

[0104]

[0105] Among them, EA j ED represents the approximate signal energy. jLet A represent the detail signal, j represent the decomposition scale, and A represent the decomposition scale. j D represents an approximate signal. j This indicates a detail signal.

[0106] Preferably, in step S300, the step of using the kernel Fisher discriminant analysis method to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set to obtain the mapped training set and test set includes:

[0107] Step S301: Define the sample feature set as containing D-dimensional features and a total of C categories; where, in the D-dimensional feature samples, the samples belonging to C1 and C2 are respectively x1={x1,x2,…,x n1}、x2={x1,x2,…,x n2}, n1 + n2 = n, where n represents the total number of samples;

[0108] Step S302: Through nonlinear mapping Map each sample feature to a high-dimensional feature space;

[0109] Step S303: In the high-dimensional feature space, calculate the mean vector of each category. Where i represents different categories, i = 1, 2; j = 1, 2, ..., n i ;

[0110] Step S303: Construct the inter-class scatter matrix between the two categories.

[0111] Step S304: Construct the intra-class scatter matrix for each category.

[0112] Step S305: Maximize the objective function J f (ω) Find the optimal projection direction ω opt ;

[0113] Step S306: Map the sample features of the training set and test set to a high-dimensional space, and along the optimal projection direction ω. opt Projection is performed to obtain the mapped training and test sets.

[0114] The kernel Fisher discriminant analysis method used in this embodiment can significantly improve the accuracy of bearing fault diagnosis. Its principle is to map the original sample features to a high-dimensional feature space through nonlinear mapping, and then perform identification analysis in the new feature space. Specifically, by projecting the sample features onto the optimal projection direction, the inter-class dispersion is maximized and the intra-class dispersion is minimized after projection, thereby effectively distinguishing samples of different categories.

[0115] Preferred, such as Figure 4 As shown, Figure 4 This is a network topology diagram of the DNN model provided by the present invention;

[0116] In step S400, the mapped training set is input into the DNN model through the following steps to iteratively train the DNN model and optimize the model parameters:

[0117] Step S401: Initialize network parameters: Set the number of hidden layers L, and determine the number of nodes p in each layer. l (l = 1, 2, ..., L), initialize the weights W between the hidden layer and the output layer. l and offset vector b l (l=2,3,…,L), set the learning rate η and the iteration threshold ε, and select the neuron activation function f(·);

[0118] Step S402: Calculate the output H of the hidden layer and the output layer. l H l =f(z) l )=f(W l Y l-1 +b l ), l=2,3,…,L; where H1 represents the output from the input layer to the first hidden layer, H l This represents the final output of the network.

[0119] Step S403: Calculate the loss function J and the output layer gradient δ L ;

[0120] Where X represents the input data and Y represents the actual label;

[0121]

[0122] Step S404: Backpropagate from the output layer to the hidden layer and calculate the hidden layer gradient δ. l ;

[0123]

[0124] Step S405: Based on the input training set, repeat steps S402-S404 until the preset number of iterations is reached or the error constraint condition is met, and end the training process; otherwise, return to step S402 and continue training.

[0125] In this embodiment, the DNN network model used has multiple hidden layers compared to the traditional BPNN model, enabling it to learn and represent more complex data features, such as processing and analyzing high-dimensional and complex vibration data, including nonlinear features useful for bearing fault diagnosis. Each hidden layer can capture different levels of abstract features of the input data. Through the connection weights between the input layer, hidden layer, and output layer, these weights are continuously adjusted during training to minimize prediction errors. Node thresholds and their activation functions are used to calculate and pass the outputs between layers, ultimately obtaining the network output value. The DNN model, through deep learning, can adapt to different types of bearing fault data. By adjusting the network structure and parameters, it can optimize the diagnosis of specific types of faults, achieving more accurate bearing fault diagnosis, helping to reduce unnecessary maintenance and replacement, and thus lowering maintenance costs.

[0126] like Figure 5 As shown, this embodiment of the invention also proposes a bearing fault diagnosis device, which performs any of the bearing fault diagnosis methods described above, including:

[0127] The sampling module is used to sample the original vibration data of the bearing under different operating conditions, and divide the collected original sample data into training set and test set according to a preset ratio;

[0128] The feature extraction module is used to perform CEEMDAN decomposition and wavelet decomposition on the original sample data in the training set and the test set, respectively, to construct the corresponding sample feature sets of the training set and the test set.

[0129] The mapping module is used to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set using the kernel Fisher discriminant analysis method to obtain the mapped training set and test set;

[0130] The training module is used to input the training set obtained through mapping into the DNN model, perform iterative training on the DNN model, and optimize the model parameters.

[0131] The testing module is used to measure the recognition accuracy of the trained DNN model using the test set obtained by mapping, and to obtain the final fault diagnosis model.

[0132] The result output module is used to acquire vibration data of the bearing under different operating conditions, input the sample features obtained after executing steps S200-S300 into the fault diagnosis model, and output the diagnosis results of bearing faults.

[0133] The present invention also proposes a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to implement any of the bearing fault diagnosis methods described above.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A bearing fault diagnosis method, characterized in that: The steps include: Step S100, sampling the original vibration data of the bearing under different working conditions, and dividing the collected original sample data into a training set and a test set according to a preset ratio; Step S200, performing CEEMDAN decomposition and wavelet decomposition on the original sample data in the training set and the test set respectively, so as to construct corresponding sample feature sets of the training set and the test set; Step S300, using the kernel Fisher discriminant analysis method to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set to obtain the mapped training set and the test set; Step S400: input the training set obtained by mapping into the DNN model, perform iterative training on the DNN model, and optimize the model parameters; Step S500: Use the test set obtained by mapping to evaluate the recognition accuracy of the trained DNN model and output a fault diagnosis model; Step S600, obtaining vibration data of the bearing under different working conditions, inputting the sample features obtained after executing steps S200-S300 into the fault diagnosis model, and obtaining the diagnosis result of the bearing fault.

2. A bearing fault diagnosis method according to claim 1, characterized in that: In step S200, the original sample data in the training set and the test set are respectively subjected to CEEMDAN decomposition and wavelet decomposition to construct corresponding sample feature sets of the training set and the test set, which specifically includes the following steps: Step S201, normalizing the original sample data in the training set and the test set; Step S202, performing CEEMDAN decomposition on the normalized data to obtain a number of time series signals; Step S203, performing wavelet decomposition on each of the time series signals to obtain approximate and detail signals of different scales; Step S204, calculating the signal energy of the approximate and detail signals of each scale obtained by wavelet decomposition; Step S205: The calculated energy values ​​are used as features and integrated into feature vectors to form sample feature sets of the training set and the test set.

3. A bearing fault diagnosis method according to claim 2, characterized in that: In step S202, the normalized data is decomposed by CEEMDAN to obtain a number of time series signals, including: Step S2021: Add Gaussian white noise that obeys normal distribution to the normalized original sample data x(t) to obtain the noisy data x i (t); x i (t) = x(t) + εω i (t)i=1,2,3,…,N, where N is the number of times white noise is added and ε is the standard deviation of the noise; Step S2022: Use CEEMDAN algorithm to process the noise-added data x i (t) are decomposed respectively to obtain the noise-added data x i Multi-frequency component IMF of (t) i,1 (t), calculate the first multi-frequency component IMF1(t) of the original sample data x(t); Where T represents the number of decompositions; Step S2023, calculating the residual r1(t) of the original sample data x(t) according to the first multi-frequency component IMF1(t); r1(t)=x(t)-IMF1(t); Step S2024: E i [ω i (t)] is defined as the i-th multi-frequency component after decomposition, ε j is the noise standard deviation, r1(t)+ε1E1[ω1(t)] is decomposed to obtain the second multi-frequency component IMF2(t); Where, I represents the number of decompositions; Step S2025: Calculate the Kth residual r of the original sample data x(t) k (t); r k (t) = r k-1 (t)-IMF k (t), k = 2, 3, ..., K; Step S2026: Calculate the k+1th multi-frequency component IMF k+1 (t); Step S2027, repeat steps S2024-S2026 until the remaining multi-frequency components can no longer be decomposed and the remaining residual R(t) satisfies When , the final decomposed time series signal x(t) is obtained; 4. A bearing fault diagnosis method according to claim 2, characterized in that: In step S203, each time series signal is subjected to wavelet decomposition by the following formula to obtain approximate and detail signals of different scales: Among them, A j Represents an approximate signal, which is used to reflect the overall change trend of the signal sequence; D j represents the detail signal, which is used to reflect the random disturbances in different frequency domains in the original signal; k represents the time-shift discreteness of the whole node mark, h k-2n With g k-2n They represent the transfer coefficients of the orthogonal wavelet, and j represents the decomposition scale.

5. A bearing fault diagnosis method according to claim 2, characterized in that: In step S204, the signal energy is calculated for the approximate and detail signals of each scale obtained by wavelet decomposition using the following formula: Among them, EA j Represents the approximate signal energy, ED j represents the detail signal, j represents the decomposition scale, A j represents the approximate signal, D j Indicates detail signal.

6. A bearing fault diagnosis method according to claim 1, characterized in that: In step S300, the kernel Fisher discriminant analysis method is used to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set to obtain the mapped training set and the test set, including: Step S301: define a sample feature set that contains D-dimensional features, with a total of C categories; among the D-dimensional feature samples, the samples belonging to C1 and C2 are respectively x1 = {x1, x2, ..., x n1 }, x2={x1,x2,…,x n2 }, n1+n2=n, n represents the total number of samples; Step S302: nonlinear mapping Map each sample feature to a high-dimensional feature space; Step S303: Calculate the mean vector of each category in the high-dimensional feature space Where i represents different categories, i = 1, 2; j = 1, 2, ..., n i ; Step S303: Construct an inter-class discreteness matrix between two classes Step S304: Construct the intra-class dispersion matrix for each class Step S305: By maximizing the objective function J f (ω) Find the optimal projection direction ω opt ; Step S306: Map the sample features of the training set and the test set to a high-dimensional space and project them along the optimal projection direction ω opt Perform projection to obtain the mapped training set and test set.

7. A bearing fault diagnosis method according to claim 1, characterized in that: In step S400, the training set obtained by mapping is input into the DNN model through the following steps, and the DNN model is iteratively trained to optimize the model parameters: Step S401, initialize network parameters: set the number of hidden layers L, determine the number of nodes in each layer p l (l=1,2,…,L), initialize the weights W between the hidden layer and the output layer l and the offset vector b l (l=2,3,…,L), set the learning rate η and iteration threshold ε, and select the neuron activation function f(·); Step S402: Calculate the output H of the hidden layer and the output layer l ;H l =f(z l )=f(W l Y l-1 +b l ),l=2,3,…,L; where H1 represents the output from the input layer to the first hidden layer, H l Represents the final output of the network; Step S403: Calculate the loss function J and the output layer gradient δ L ; Among them, X represents the input data and Y represents the true label; Step S404: Back propagate from the output layer to the hidden layer to calculate the hidden layer gradient δ l ; Step S405: Repeat steps S402 to S404 according to the input training set until the preset number of iterations is reached or the error constraint condition is met, and the training process ends; otherwise, return to step S402 to continue training.

8. A bearing fault diagnosis method according to claim 7, characterized in that: In step S401, the neuron activation function f(·) includes a ReLU function and a tanh function.

9. A bearing fault diagnosis device, executing a bearing fault diagnosis method according to any one of claims 1 to 8, characterized in that: include: A sampling module is used to sample the original vibration data of the bearing under different working conditions, and divide the collected original sample data into a training set and a test set according to a preset ratio; A feature extraction module, used to perform CEEMDAN decomposition and wavelet decomposition on the original sample data in the training set and the test set, respectively, to construct corresponding sample feature sets of the training set and the test set; A mapping module, used to perform high-dimensional mapping on the corresponding sample feature sets of the training set and the test set using a kernel Fisher discriminant analysis method to obtain a mapped training set and a test set; The training module is used to input the training set obtained through mapping into the DNN model, iteratively train the DNN model, and optimize the model parameters; The testing module is used to measure the recognition accuracy of the trained DNN model using the mapped test set to obtain the final fault diagnosis model; The result output module is used to obtain the vibration data of the bearing under different working conditions, input the sample features obtained after executing steps S200-S300 into the fault diagnosis model, and output the diagnosis result of the bearing fault.

10. A computer-readable storage medium, characterized in that: A program executable by a processor is stored therein, and when the program executable by the processor is executed by the processor, it is used to implement a bearing fault diagnosis method as described in any one of claims 1 to 8.

Images