Bearing fault diagnosis method and system based on characteristic mode decomposition

Abstract

The invention provides a bearing fault diagnosis method and system based on characteristic mode decomposition, and the method comprises the steps: obtaining a decomposition mode of a data sample based on a characteristic mode decomposition technology; training a bearing fault diagnosis model according to the decomposition mode based on a deep neural network; obtaining vibration data to be diagnosed; and based on a transfer learning technology, adjusting pre-training model parameters according to the to-be-diagnosed vibration data, and obtaining a fault diagnosis result corresponding to the to-be-diagnosed vibration data output by the adjusted bearing fault diagnosis model. According to the bearing fault diagnosis method and system based on characteristic mode decomposition, the FMD technology is introduced, the impulse and periodicity of signals are considered at the same time, and the fault period does not need to serve as priori knowledge; main features are extracted by using the convolutional neural network, the calculation speed is accelerated, partial parameters of the pre-training model are adjusted by using the transfer learning technology to adapt to new vibration data, the training time of the new data is shortened, and the model performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method and system based on characteristic mode decomposition. Background Art

[0002] Rolling bearings are one of the most widely used key components in rotating machinery. They play an important role in supporting rotating shafts, reducing friction and bearing loads in various types of mechanical equipment. For example, in automobiles, aircraft, machine tools, wind turbines and other equipment, the performance and reliability of rolling bearings are directly related to the normal operation of the entire equipment. Therefore, the condition monitoring and fault diagnosis of mechanical equipment are of great significance to improving the availability, safety and reliability of mechanical equipment. Related research has received great attention from academia and industry. However, the existing fault diagnosis has significant deficiencies in signal processing.

[0003] Traditional fault diagnosis encounters different challenges in its decomposition methods when dealing with rolling bearing vibration signals. For example, they have difficulty in accurately decomposing fault information because they do not fully consider the characteristics of mechanical faults, namely the impulsiveness and periodicity of the signal. Among the existing methods, the application of variational mode decomposition (VMD) in mechanical fault diagnosis is based on features. That is, the fault component is considered to be a narrowband component, but in addition to the fault component, the measured signal usually contains some interference.

[0004] The performance of filtering-based decomposition methods is highly dependent on the shape and bandwidth of the filter. Among the existing methods, empirical wavelet transform (EWT) and VMD are designed wavelet basis and Wiener filter to decompose the superimposed signal. If the bandwidth is too large, different single components, as well as noise, may be mixed into a single mode. On the contrary, it may lead to more redundant modes and some important fine information loss.

[0005] In view of this, there is an urgent need for a bearing fault diagnosis method and system based on eigenmode decomposition to at least solve the above-mentioned deficiencies. Summary of the invention

[0006] One of the purposes of the present invention is to provide a bearing fault diagnosis method and system based on characteristic mode decomposition. When diagnosing bearing faults, FMD technology is introduced, and the impulsiveness and periodicity of the signal are taken into account at the same time, and there is no need to take the fault period as prior knowledge; convolutional neural network is used to extract the main features to speed up the calculation speed, and transfer learning technology is used to adjust some parameters of the pre-trained model to adapt to new vibration data, thereby reducing the training time of new data and improving model performance.

[0007] An embodiment of the present invention provides a bearing fault diagnosis method based on characteristic mode decomposition, comprising:

[0008] Based on the characteristic mode decomposition technology, the decomposition mode of the data sample is obtained;

[0009] Based on deep neural network, the bearing fault diagnosis model is trained according to the decomposed mode;

[0010] Acquire vibration data to be diagnosed;

[0011] Based on the transfer learning technology, the pre-trained model parameters in the bearing fault diagnosis model are adjusted according to the vibration data to be diagnosed, and the fault diagnosis result corresponding to the vibration data to be diagnosed output by the adjusted bearing fault diagnosis model is obtained.

[0012] Preferably, based on the characteristic mode decomposition technology, the decomposition mode of the data sample is obtained, including:

[0013] Collect data samples of rolling bearings under different fault conditions;

[0014] By initializing a set of uniformly distributed filters covering the entire frequency band, the data samples are decomposed to obtain multiple modes;

[0015] The filter is iteratively updated with CK as the objective function and the fault period is estimated in each filtering process;

[0016] By comparing the correlation coefficients CC of the two modes, the mode with the smallest CK is discarded from the two modes with the largest CC;

[0017] When the termination condition is reached, the decomposition is terminated and the retained mode is obtained as the final decomposition mode;

[0018] The final decomposed modal data is normalized and preprocessed to obtain the decomposed mode, so that the data with different characteristics are in the same numerical range.

[0019] Preferably, collecting data samples of rolling bearings under different fault conditions includes:

[0020] Acquire the collected samples including acceleration signal gs, sampling rate sr, shaft speed rate, load weight l oad and four critical frequencies representing different fault locations as data samples. The four critical frequencies include: outer ball passing frequency BPFO, inner ball passing frequency BPF I, basic train frequency FTF and ball rotation frequency BSF; wherein, the calculation formulas of outer ball passing frequency BPFO and inner ball passing frequency BPF I are as follows:

[0021]

[0022] Where d is the ball diameter, D is the pitch diameter, and the variable f r is the shaft speed, n is the number of rolling elements, and Φ is the bearing contact angle.

[0023] Preferably, based on a deep neural network and according to the decomposition mode, a bearing fault diagnosis model is trained, including:

[0024] Based on deep neural network, target features are extracted according to decomposition mode;

[0025] The target features are divided into training set, validation set and test set by random partitioning method;

[0026] Classify scale maps by fine-tuning the pre-trained SqueezeNet convolutional neural network;

[0027] The validation set is used to monitor the performance of the model and the trained model is evaluated through the test set to obtain a bearing fault diagnosis model that passes the evaluation.

[0028] A bearing fault diagnosis method based on characteristic mode decomposition provided by an embodiment of the present invention also includes:

[0029] When collecting data samples of rolling bearings under different fault conditions, a collection compliance test is performed to determine the data samples that pass the collection compliance test.

[0030] Preferably, collect normative tests, including:

[0031] Based on the bearing fault diagnosis knowledge graph, the diagnosis basis data corresponding to the fault state of the rolling bearing is obtained; the diagnosis basis data includes: a one-to-one corresponding first equipment operation stage and diagnosis basis sub-data;

[0032] Get pre-collected data;

[0033] Parse the pre-collected data to obtain collected sample data corresponding to different collected fault states; the collected sample data includes: one-to-one corresponding second equipment operation phase and collected sample sub-data;

[0034] Compare the diagnostic basis data corresponding to different fault states and the collected sample data corresponding to different collected fault states to determine whether the collected fault states, the second equipment operation stage, and the collected sample sub-data in the pre-collected data are all comprehensive;

[0035] If they are all comprehensive, the collection normativeness test is passed, and the corresponding pre-collected data is used as the data sample;

[0036] If it is not comprehensive, after supplementary collection, the supplementary collected data and the pre-collected data will be used together as data samples.

[0037] Preferably, the supplementary collection includes:

[0038] Obtain a data sample tree construction template; the data sample tree construction template extracts the target fault state as root node data, extracts the target equipment operation stage as first-level sub-node data, and extracts the collection location data as second-level sub-node data according to the data to be constructed;

[0039] Building a template based on the data sample tree and building a standard data sample tree based on the diagnostic data corresponding to different fault states;

[0040] Building a template based on the data sample tree and building a target data sample tree according to the collected sample data corresponding to different collection fault states;

[0041] According to the standard data sample tree and the target data sample tree, determining the missing subtree of the target data sample tree corresponding to the standard data sample tree;

[0042] Get the missing node with the smallest level in each missing subtree;

[0043] Supplementary collection is performed based on the node attributes of the missing nodes.

[0044] Preferably, supplementary collection is performed based on the node attributes of the missing nodes, including:

[0045] Connect to the industrial big data platform and conduct supplementary collection based on the node attributes of the missing nodes.

[0046] Preferably, the industrial big data platform is connected and supplementary collection is performed according to the node attributes of the missing nodes, including:

[0047] Parse the node attributes of the missing nodes and obtain node features; the node features are: the type of missing nodes, the type of upper-layer nodes, and the node data collection conditions;

[0048] According to the missing node types and upper-level node types, information index conditions are constructed;

[0049] Index the industrial big data platform based on information index conditions and obtain the index interface;

[0050] Parse the interface script of the index interface to obtain the conditional description semantics of the interface data;

[0051] Locate target data based on node data collection conditions and condition description semantics;

[0052] Get the capture rules of target data in the interface script;

[0053] According to the capture rules, capture target data;

[0054] When the target data corresponding to all missing nodes are captured, the supplementary collection is completed.

[0055] An embodiment of the present invention provides a bearing fault diagnosis system based on characteristic mode decomposition, comprising:

[0056] The characteristic mode decomposition subsystem is used to obtain the decomposition mode of the data sample based on the characteristic mode decomposition technology;

[0057] The bearing fault diagnosis model training subsystem is used to train the bearing fault diagnosis model based on the deep neural network and the decomposed mode;

[0058] A vibration data acquisition subsystem for diagnosis, used for acquiring vibration data for diagnosis;

[0059] The diagnosis subsystem is used to adjust the pre-trained model parameters in the bearing fault diagnosis model according to the vibration data to be diagnosed based on the transfer learning technology, and obtain the fault diagnosis result corresponding to the vibration data to be diagnosed output by the adjusted bearing fault diagnosis model.

[0060] The beneficial effects of the present invention are:

[0061] The present invention introduces FMD technology when diagnosing bearing faults, taking into account the impulsiveness and periodicity of the signal at the same time, and does not need to take the fault period as prior knowledge; it uses convolutional neural networks to extract main features to speed up the calculation speed, and uses transfer learning technology to adjust some parameters of the pre-trained model to adapt to new vibration data, thereby reducing the training time of new data and improving model performance.

[0062] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in the present application documents.

[0063] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0065] Figure 1 is a schematic diagram of a bearing fault diagnosis method based on characteristic mode decomposition in an embodiment of the present invention;

[0066] Figure 2 Schematic diagram of a bearing fault diagnosis system based on characteristic mode decomposition in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0068] The embodiment of the present invention provides a bearing fault diagnosis method based on characteristic mode decomposition, such as Figure 1 As shown, including:

[0069] Step 1: Based on the characteristic mode decomposition technology, the decomposed mode of the data sample is obtained; wherein the decomposed mode is: a sub-signal decomposed from the data sample, representing a part of the sample characteristics of the data sample;

[0070] Step 2: Based on the deep neural network, according to the decomposition mode, a bearing fault diagnosis model is trained; wherein the bearing fault diagnosis model is: a deep neural network model for identifying and diagnosing bearing faults trained by learning the features in the decomposition mode. During training, the decomposition mode is used as the input of the neural network model, the fault diagnosis results marked in the data sample are used as the output of the neural network model, and the model is trained to a convergence state;

[0071] Step 3: Obtain vibration data to be diagnosed; wherein the vibration data to be diagnosed is: vibration signals to be diagnosed, such as: vibration signals of the outer ring, inner ring, rolling element and cage during the startup phase;

[0072] Step 4: Based on the transfer learning technology, adjust the pre-trained model parameters in the bearing fault diagnosis model according to the vibration data to be diagnosed, and obtain the fault diagnosis result corresponding to the vibration data to be diagnosed output by the adjusted bearing fault diagnosis model; wherein, based on the transfer learning technology, adjust the pre-trained model parameters in the bearing fault diagnosis model according to the vibration data to be diagnosed specifically as follows: based on the transfer learning technology, adjust the weights and biases of the fully connected layer so that the output can adapt to the new vibration data task; fine-tune the filter weights of the convolution layer so that the model can better extract fault-related features from the vibration signal; adjust the stride and window size of the pooling layer, etc., to optimize feature extraction;

[0073] Among them, based on the characteristic mode decomposition technology, the decomposition mode of the data sample is obtained, including:

[0074] Collect data samples of rolling bearings under different fault conditions;

[0075] By initializing a set of uniformly distributed filters covering the entire frequency band, the data samples are decomposed to obtain multiple modes;

[0076] By comparing the correlation coefficients CC of the two modes, the mode with the smallest CK is discarded from the two modes with the largest CC;

[0077] When the termination condition is reached, the decomposition is terminated and the retained mode is obtained as the final decomposition mode;

[0078] The final decomposed modal data is normalized and preprocessed to obtain the decomposed mode, so that the data with different characteristics are in the same numerical range;

[0079] Among them, based on the deep neural network, according to the decomposition mode, the bearing fault diagnosis model is trained, including:

[0080] Based on the deep neural network, the target features are extracted according to the decomposed mode; wherein the target features are: representative features extracted from the decomposed mode, including time domain features and frequency domain features;

[0081] The target features are divided into training set, validation set and test set by random partitioning method;

[0082] Classify scale maps by fine-tuning the pre-trained SqueezeNet convolutional neural network;

[0083] The validation set is used to monitor the performance of the model and the trained model is evaluated through the test set to obtain a bearing fault diagnosis model that passes the evaluation.

[0084] The working principle and beneficial effects of the above technical solution are:

[0085] The specific steps for collecting data samples of rolling bearings under different fault conditions are as follows:

[0086] 1.1 First, select a suitable sensor (such as an accelerometer) to ensure that its frequency response and sensitivity meet the requirements of rolling bearing fault detection; then install the sensor near the rolling bearing to ensure that the data sample (vibration signal) can be accurately collected, and the installation position should be selected in a place that can reflect the fault characteristics, such as the outer ring of the bearing or the supporting structure;

[0087] 1.2 Configure the data acquisition system, including setting the sampling rate, acquisition time and data storage method. The sampling rate is set according to the highest frequency of the vibration signal, usually 2-5 times the signal bandwidth. Determine the duration of data acquisition and perform multiple acquisitions at different stages of equipment operation (such as startup, stabilization, and shutdown) to cover the vibration characteristics under different working conditions;

[0088] 1.3 Introduce different fault states (such as inner ring fault, outer ring fault, rolling element damage, etc.), operate the equipment in each state; record the vibration signal under each fault state, and collect the corresponding operating parameters at the same time;

[0089] 1.4 Constructing data samples: Each data sample should contain the following: acceleration signal gs, sampling rate sr, shaft speed rate, load weight l oad, and four critical frequencies representing different fault locations: outer ball passing frequency (BPFO), inner ball passing frequency (BPF I), basic train frequency (FTF), and ball rotation frequency (BSF). The specific calculation formulas include:

[0090]

[0091] Where d is the ball diameter, D is the pitch diameter, and the variable f r is the shaft speed, NrollingElements is the number of rolling elements, and Φ is the bearing contact angle.

[0092] After obtaining a large number of data samples, these data need to be further processed. In order to verify the importance of filter initialization, a set of uniformly distributed filters covering the entire frequency band are initialized to decompose the data samples. The specific steps to obtain multiple modes are as follows:

[0093] 2.1 Load the original signal input parameters, namely the mode number n and the filter length L, which are used to determine the number of modes decomposed from the signal, and the filter length determines the impulse response length of the FIR filter;

[0094] 2.2 By using the Hanning window to initialize the FIR filter bank and update the filter coefficients using the cutoff frequency, different modes can be adaptively selected at the same time, including:

[0095] The frequency band of the original signal is evenly divided into K segments, where K is set to 5-10;

[0096] The upper cutoff frequency f of these two segments u and the lower cutoff frequency f l , defined as:

[0097] f l = k·f s / 2K

[0098] f u =(k+1)·f s / 2K

[0099] Where K = 0, 1, 2, ..., K-1, ...

[0100] where f s is the sampling frequency of the original signal;

[0101] Use the fir1 function in MATLAB to generate K cutoff frequencies [f l , f u ], and FIR filter with K varying from 0 to K-1;

[0102] The filter length is L, and the "window" is selected as "Hanning";

[0103] Through this initialization, a group of FIR filters that are evenly distributed and cover the entire frequency band can constitute a FIR filter bank.

[0104] By adjusting the filter parameters according to the signal characteristics in subsequent steps, the filter group is made adaptive to the frequency distribution of the signal, such as using objective functions such as correlation kurtosis to evaluate the quality of the filtered signal, adjusting the filter coefficients, and updating the filter coefficients in each iteration to make it more focused on the target frequency components.

[0105] After loading the original signal and initializing the FIR filter group with the Hanning window, the initialized filter group needs to be applied to the original signal. The signal is initially decomposed by convolution operation to obtain multiple initial filtered signals (decomposition modes). These modes correspond to different frequency ranges, retaining the main frequency characteristics of the original signal, and may still contain noise or aliasing information. Therefore, these initial filtered signals are used as input, and the target fault characteristics are further extracted by iteratively updating the filter parameters and optimizing the signal decomposition effect. The filter is iteratively updated with CK as the objective function, and the specific steps of estimating the fault period in each filtering process are as follows:

[0106] 3.1 Calculate the HNR, kurtosis and CK of each signal component:

[0107] Calculate the harmonic noise ratio: for the filtered signal u k Perform spectrum analysis, use fast Fourier transform to calculate the spectrum, extract harmonic components and noise components, and calculate the harmonic power and noise power formula as follows:

[0108]

[0109] Calculate the kurtosis: First calculate the mean and standard deviation of each filtered signal, and calculate the kurtosis according to the formula:

[0110]

[0111] Among them, i ′ is the index of the signal sample, u k [i ′ ] is the kth filter output signal at the i-th ′ The value of the sample point, μ is the signal u k [i ′ ], σ is the mean of the signal u k [i ′ ], N is the number of signal samples;

[0112] Calculate CK: by u k =X*f k Get the filtered signal, * is the convolution operation, f k is the kth filter coefficient. Among them,

[0113]

[0114] Where x(L) is the sample value of the original signal x at time point L;

[0115] The decomposition mode CK can be defined as:

[0116]

[0117] The superscript H is the conjugate transpose operation, and W M As an intermediate variable, used to control the weighted correlation matrix, M is W M Control parameters of

[0118] 3.2 Update filter coefficients:

[0119] The filter needs to extract the target signal features more accurately, that is, the core goal is to maximize the correlation kurtosis CK and periodic correlation of the decomposed signal;

[0120] In order to achieve this goal, the filter coefficients are updated by calculating the CK value of the output signal after each filtering to evaluate the filter performance; and the filter parameters are adjusted according to the changes in the CK value and optimized repeatedly;

[0121] First, maximization with respect to the filter coefficients is equivalent to the eigenvector associated with the largest eigenvalue λ of the following generalized eigenvalue problem:

[0122] R xw f k =R xx f k λ

[0123] During the iteration process, R xw is the cross-correlation matrix between the signal and the filter output, R xx is the signal autocorrelation matrix, the kth filter coefficient f k The value of λ will be updated to keep approaching the set target, i.e. the maximum filtered signal of CK;

[0124] Then, based on the autocorrelation theory, the fault period is estimated from the measured signal, including:

[0125] Rx(τ) represents the autocorrelation function of the signal x(α). The expression of Rx(τ) with respect to the time lag τ can be defined as:

[0126]

[0127] Among them, α is the value of the signal at different time points, and the estimated period of each filtered signal is selected as the autocorrelation spectrum reaching the local maximum R after the zero crossing point. x (τ1) point, estimated period T s =τ1,T s is the period estimated directly from the original signal;

[0128] High CK values ​​signal that both periodicity and impulsivity are strong;

[0129] 3.3 Mode screening:

[0130] First, after calculating the three indicators of the four signal components, by comparing the result graphs, HNR and kurtosis K evaluate the periodicity and impulsivity of the signal respectively, and CK comprehensively considers periodicity and impulsivity and is the most critical evaluation indicator. Therefore, CK is selected as the objective function for fault component extraction;

[0131] Then sort the CK values ​​of all modes from large to small. Modes with higher CK values ​​usually contain more information related to the fault; modes with lower CK values ​​may be noise or frequency components unrelated to the fault. Modes with larger CK values ​​are screened and retained, and modes with weaker impulsiveness or periodicity are discarded.

[0132] Through the above steps, after the initial filter group decomposes the original signal, multiple filtered signals, i.e., multiple modes, are obtained, and the HNR, kurtosis, and related kurtosis CK of each mode are calculated, and the CK value is used as a key indicator to measure the periodicity and impulsivity of the signal for preliminary screening of the mode. By sorting the CK values, the modes with high CK values ​​are retained, which contain more fault-related features, and the modes with low CK values ​​are removed.

[0133] After completing the preliminary screening through the CK value, the redundancy problem between modes has not been solved. For example, similar frequency components may exist in different modes. Therefore, it is necessary to quantify the similarity between modes by calculating the correlation coefficient CC between modes, further analyze the redundant relationship between these modes, and optimize the final mode selection to ensure that the modal decomposition result is more accurate. Therefore, by comparing the correlation coefficient CC of the two modes, the mode with the smallest CK is abandoned from the two modes with the largest CC. The specific steps are as follows:

[0134] 4.1 Calculate the correlation coefficient CC and filter the redundant patterns:

[0135] First, define u p and u q The CC of the two modes is:

[0136]

[0137] in, and u p and u q The mean of the mode, N is the number of signal samples;

[0138] CC(u p ,u q ) has a value range of [-1,1]: when CC approaches 1, the two modes are highly positively correlated, that is, they are highly similar; when CC approaches -1, the two modes are highly negatively correlated; when CC approaches 0, the two modes are unrelated;

[0139] Then generate the correlation matrix, for all patterns {u1,u2,…,u m} Calculate the correlation coefficients pairwise and construct an m×m correlation matrix CC: CC ij =CC(u i ,u j ),1≤ij≤m;

[0140] Then traverse the correlation matrix CC and find the largest off-diagonal element CC ij :CC max =max i≠j CC ij ; This value corresponds to two modes u i and u j , the similarity between the two modalities is the highest, and there may be information redundancy;

[0141] 4.2 Use the estimated period to calculate the CK value of the mode and compare it, and eliminate the mode with smaller CK:

[0142] First get the modal u i and u j The kurtosis CK value of i ) and CK(u j ), using the estimated period T i Calculate CK value, T i is the periodicity estimate for each modal signal after modal decomposition:

[0143] In the modal signal u i The timing of the extraction and period T i The corresponding periodic component, that is, the formula for period-dependent energy is:

[0144] The total energy of the modal signal is defined as:

[0145] Where t represents the discrete time index of the signal (indicating the current time point or sample position); i represents the i-th mode; u i[t] represents the amplitude of the i-th modal signal at time index t, and N is the number of signal samples;

[0146] Calculate CK value: The higher the CK value, the more significant the signal's characteristics in terms of periodicity and impulsivity;

[0147] Then compare CK(u p ) and CK(u q ) size:

[0148] If CK(u p )>CK(u q ), keep u p , remove u q ;

[0149] If CK(u p ) <CK(u q ), keep u q , remove u p ;

[0150] Finally, remove the modality with smaller CK from the candidate modality set and update the modality set {u1,u2,…,u m}, reduce one mode; and record the retained mode and its estimated period T and corresponding CK value for subsequent analysis;

[0151] Through the above steps, the CK value is calculated using the estimated period of the modal signal. Combined with the ratio of period-related energy to total energy, the periodicity and impulsiveness of the mode can be effectively evaluated. For modes with high correlation, the mode with more fault characteristics is selected by comparing the CK values, and redundant modes are eliminated, thereby improving the accuracy and reliability of modal decomposition.

[0152] After gradually eliminating redundant modes by comparing the correlation coefficient CC and the correlation kurtosis CK between the modes, in order to ensure that the remaining modes are independent and highly correlated with the fault characteristics, this process needs to be repeated until the set termination condition is met, such as the correlation coefficient is lower than the threshold or the number of modes reaches the expected value. At this time, the retained mode set will be used as the final decomposition mode, representing the main fault characteristics of the signal for subsequent analysis and diagnosis; when the termination condition is met, the decomposition is terminated and the retained mode is obtained as the final decomposition mode. The specific steps are as follows:

[0153] 5.1 Determine the termination conditions:

[0154] First, the correlation threshold θ is set so that the maximum correlation coefficient CC between the modes max If it is smaller than the set threshold value θ, it means that the correlation between the remaining modes is small enough and there is no obvious redundant mode.

[0155] Then, the number of modes is limited so that the number of retained modes reaches a preset value n, and 2-5 main modes are selected according to the complexity of the fault characteristics;

[0156] When the CK value or signal reconstruction error changes less than the threshold in several consecutive iterations, the decomposition result is stable.

[0157] 5.2 Get the retained mode:

[0158] First, the retained modes are collected, and the final mode set {u1,u2,…u k} is the signal mode retained after multiple rounds of screening and has independence, that is, the correlation between modes is low and the information redundancy is minimal; fault feature correlation, that is, the CK value of the mode is high, with significant periodicity and impulsive characteristics;

[0159] The frequency distribution, time domain characteristics, and energy distribution of each mode are then checked to ensure that these modes have a direct correlation with the equipment operating status or fault characteristics.

[0160] Through the above steps, the decomposition is ended when the termination condition is reached, and the retained modes are obtained, marking the completion of the decomposition process. These modes not only extract the main fault characteristics of the signal, but also eliminate redundancy and noise effects, providing high-quality data support for subsequent diagnosis, classification and prediction.

[0161] After completing the modal decomposition and obtaining the retained mode as the final decomposition mode, normalization preprocessing is performed on these modal data to ensure that the features of different modes are in the same numerical range, so that they are consistent in subsequent analysis and model training. Therefore, the more thoroughly decomposed data information is preprocessed by the normalization method to make the data of different features in the same numerical range. The specific steps are as follows:

[0162] 6.1 Normalization processing:

[0163] First, the more thoroughly decomposed data information is preprocessed by the normalization method, so that the data with different features are in the same numerical range. The steps include:

[0164] The normalization formula is defined as:

[0165]

[0166] Where u is the original data value, u min is the minimum value in the data set, u max is the maximum value in the data set. This formula maps the data to the interval [0,1] so that data with different features are in the same numerical range.

[0167] Then, each retained mode is processed separately, that is, the normalization formula is applied to each mode u one by one to avoid mutual influence between modes; the normalized result is output, and the normalized modal signal y is retained for subsequent analysis (input of deep neural network);

[0168] 6.2 Check the normalized data:

[0169] The normalized data need to be checked to ensure that they are within the expected [0,1] value range;

[0170] Verify whether the normalized data eliminates the scale differences between different modalities to ensure the fairness of subsequent model training and analysis.

[0171] Through normalization, all modal data are ensured to be in the same numerical range, so that the amplitude differences between different features will not affect the subsequent model training and analysis results. The normalized data provides standardized input for subsequent feature extraction, classification model training and fault diagnosis, improving the training effect and accuracy of the model.

[0172] After completing the normalization process of the retained mode, the numerical range of all modal signals has been uniformly normalized, providing a good foundation for subsequent feature extraction. In the next step, the normalized modal signal needs to be analyzed through a deep neural network, and representative features, including time domain features and frequency domain features, are automatically extracted from the data to capture the complex modes and fault features hidden in the signal. Representative features, including time domain features, frequency domain features, etc., are extracted from the data through a deep neural network. The specific steps are as follows:

[0173] 7.1 Input data preparation, building deep neural network structure:

[0174] First, load the normalized modal signal y into the program as the input of the network. Make sure the data format is suitable for the input shape of the deep neural network;

[0175] Then choose a network architecture suitable for extracting time domain features, namely a convolutional neural network;

[0176] The key steps are:

[0177] The shape of the input layer, that is, the input data y, is (batch size ,N1,1), where N1 is the time step, batch size is the number of samples input into the model at one time;

[0178] The convolution layer uses one-dimensional convolution to extract the local features of the time series (signal fluctuations, mutations, etc.). The convolution operation formula is as follows:

[0179]

[0180] Among them, ω[i1] is the convolution kernel, i1 is the index of the current weight in the convolution kernel; k is the convolution kernel size; b is the bias term; j1 is defined as the index in the output feature map z;

[0181] Pooling layer, that is, reducing feature dimensions through maximum pooling or average pooling to retain local important features;

[0182] The fully connected layer summarizes the previously extracted features and maps them into feature vectors;

[0183] 7.2 Model training, extracting time domain features:

[0184] First compile the model and set the training parameters, that is, set the loss function (such as mean square error), optimizer (Adam), learning rate and other parameters;

[0185] Then use all the data for preliminary training, use all pre-processed data, that is, time domain and frequency domain data as input, directly train the model, and calculate the training error. After the training is completed, verify the performance of the model on the same data; then dynamically divide the data set through the code, simply temporarily divide all the data into training set and verification set according to a certain ratio, and quickly realize the performance monitoring of the model;

[0186] The network captures local features such as pulses and mutations of time series signals through the convolution layer, and summarizes high-dimensional time domain features in the fully connected layer, including signal amplitude change trends, waveform mutation characteristics, etc. The features automatically extracted by the network are used as the output of subsequent tasks, that is, the final output of the model can be the signal classification result or the extracted time domain feature vector;

[0187] 7.3 Convert the time domain signal to the frequency domain signal through the function in MATLAB:

[0188] Use the fast Fourier transform (FFT) in MATLAB to convert the time domain signal into a frequency domain signal, that is, spectrum data in complex form. Each complex value contains amplitude and phase information. The specific extracted content is as follows:

[0189] Extract frequency components: According to the sampling frequency f s Sum signal length N, calculate the frequency component, that is, starting from 0Hz, gradually increase until the maximum frequency is f s / 2, which is the Nyquist frequency; and only the frequency value of the positive frequency part is retained;

[0190] Extract amplitude information: It is obtained by calculating the modulus or absolute value of the FFT result, that is, the amplitude corresponding to each frequency reflects the energy of the signal at that frequency; the amplitude is normalized to match the energy distribution of the signal;

[0191] Extract phase information: It is obtained by calculating the angle of the FFT result, that is, the phase angle, which reflects the relative position of different frequency components of the signal in time;

[0192] Finally, the spectrum information is output, and the obtained spectrum information includes: frequency component, amplitude and phase.

[0193] Through these steps, the information in the time domain and frequency domain can be fully utilized by the deep neural network during the training process. The network extracts time domain features and frequency domain features respectively through two branches, and finally summarizes the information in the joint layer to achieve multi-dimensional feature fusion, thereby improving the accuracy of feature extraction and diagnostic ability.

[0194] After obtaining the time domain features and frequency domain features through deep neural networks and MATLAB, a set of high-dimensional feature vectors are generated. The random partitioning method is needed to disrupt the processed data to avoid the adverse effects of the order of data distribution on model training and evaluation, ensure the fairness of the training process and the generalization ability of the model, and divide the preprocessed data into training set, validation set and test set through random partitioning. The specific steps are as follows:

[0195] 8.1 Randomly shuffle the preprocessed data:

[0196] In order to eliminate the possible deviation caused by the original arrangement order of the data, the random partitioning method is used to randomly shuffle the data;

[0197] First, randomly generate an index sequence of length N;

[0198] Then the order of the data is disrupted according to the random index to ensure the randomness of the data distribution;

[0199] 8.2 Divide the data into training set, validation set and test set

[0200] The randomly shuffled data is divided into training set, validation set and test set in a ratio of 7:2:1;

[0201] The training set is used for model parameter learning, that is, forward propagation and back propagation are performed through the data in the training set in the deep neural network to gradually optimize the model parameters and minimize the loss function;

[0202] The validation set is used for model performance monitoring and hyperparameter adjustment. That is, it evaluates the intermediate performance of the model during the training process, helps select the best hyperparameters, and uses early stopping technology to terminate training early when the performance of the validation set no longer improves.

[0203] The test set is used for the final performance evaluation of the model and to measure the generalization ability of the model. That is, after training and validation, an independent evaluation is performed on the test set to understand the performance of the model on unseen data.

[0204] Through the above steps, the reasonable division and use of training sets, validation sets, and test sets can effectively improve the scientific nature of data division, ensure the effectiveness, reliability, and generalization ability of the model, and avoid the imbalance or overfitting of the model in real scenarios.

[0205] After completing the random division of the data, the input scale image data is classified by fine-tuning the pre-trained SqueezeNet convolutional neural network and using its prior knowledge in large-scale image classification tasks; in the fine-tuning stage, the first few layers of the SqueezeNet feature extraction network are frozen, and only the latter layers and the output layer are adjusted to adapt to the fault category of the current task, thereby improving the classification performance and speeding up the training speed. The specific steps for fine-tuning the pre-trained SqueezeNet convolutional neural network to classify the scale map are as follows:

[0206] 9.1 Adjust the output layer:

[0207] Using the SqueezeNet model trained on ImageNet as the base model, in the Squeeze convolution layer, the value of a certain position on the output feature map is defined as:

[0208]

[0209] Among them, Y s,ij It is the output feature map of the Squeeze layer, with a size of n h ×n w ×s,Y s,ij The s in the figure represents the number of channels after compression, and i and j represent the indexes of the height and width of the output feature map (in n h and n w Traverse up); X i,j,k is the input feature map, size is n h ×n w ×n c , n h Represents the height of the input feature map, n w Indicates the width of the input feature map, n c Indicates the number of channels of the input feature map, X i,j,k The i and j in the table represent the spatial positions of the input feature map, respectively, and k represents the kth channel of the input feature map, ranging from 1 to n. c ; W s,1,1,k Where s represents the sth channel of the convolution kernel output feature map, 1,1 represents the size of the convolution kernel, and W s,1,1,k The size is 1×1×n c ;

[0210] In the Expand convolution layer, the output feature map Y eIt is the concatenation of the 1×1 convolution kernel and the 3×3 convolution kernel in the channel dimension, defined as:

[0211]

[0212] Y e =Y e1,ij +Y e2,ij

[0213] Among them, Y e1,ij (i, j represents the index of the height and width corresponding to the spatial position of the output feature map, e1 is the number of channels after the 1×1 convolution) is the output feature map of the 1×1 convolution part, with a size of n h ×n w ×e1; the input of the 1×1 convolution part is the output Y of the Squeeze layer s,i,j,k (s represents the channel index in the input feature map, i, j represent the index of the height and width corresponding to the spatial position of the output feature map, and k is the channel index); the convolution kernel is W e1,1,1,k (e1 indicates that the convolution kernel is used for 1×1 convolution operation, 1,1 indicates that the spatial size of the convolution kernel is 1×1, and k indicates the weight of the convolution kernel on the kth channel of the input feature map), the size is 1×1×s; Y e2,ij (i, j represents the index of the height and width corresponding to the spatial position of the output feature map, e2 is the number of channels after the 3×3 convolution) is the output feature map of the 3×3 convolution part, with a size of n h ×n w ×e2; the input of the 3×3 convolution part is the output Y of the Squeeze layer s,i+m,j+n,k (s represents the channel index of the input feature map, i+m,j+n represents the spatial position offset of the convolution kernel); the convolution kernel is W e2,m,n,k (m, n represent the spatial position offset of the convolution kernel, representing the position of the convolution kernel in space; k represents the weight of the convolution kernel on the kth channel of the input feature map), the size is 3×3×s;

[0214] Y e The size is n h ×n w ×(e1+e2);

[0215] In the pooling layer, the maximum pooling is used to select the maximum value in a pooling window as the output. For a certain area on the input feature map, the output value is defined as:

[0216] P ij =max{X i+m,j+n |0≤m <k;0≤n<k}

[0217] The pooling window is k×k;

[0218] In the fully connected layer, the output is defined as:

[0219] y=f(Wx+b)

[0220] Assume that the input vector is x and its size is n in , W is the weight matrix, its size is n out ×n in ,b is the bias vector, its size is n out , f is the activation function;

[0221] 9.2 Fine-tuning the network:

[0222] It is necessary to further train the network based on the pre-trained model using task-specific data; first, select an optimization strategy, that is, adjust the learning rate. Usually, the learning rate for fine-tuning is smaller than the learning rate for training from scratch. A lower learning rate can be selected to fine-tune the weights of the last few layers.

[0223] Through the above steps, fine-tuning the pre-trained SqueezeNet, using its common features learned on ImageNet, and training the last few layers in combination with the given task data can significantly reduce training time and improve performance.

[0224] After fine-tuning the pre-trained SqueezeNet and completing the training, the next step is to use the validation set to monitor the performance of the model to ensure the effectiveness of the model during the training process and prevent overfitting. By evaluating the performance on the validation set after each training cycle, the model hyperparameters can be adjusted in a timely manner and measures can be taken to avoid overfitting. After the training is completed, the trained model will be finally evaluated on the test set, and the generalization ability and actual performance of the model will be fully understood by calculating performance indicators such as classification accuracy. Therefore, the validation set is used to monitor the performance of the model to prevent overfitting, and the accuracy of the trained model is calculated on the test set to evaluate the model. The specific steps are as follows:

[0225] 10.1 Model prediction, comparison between predicted value and true value:

[0226] First, input the validation set and test set data into the model to obtain the prediction results;

[0227] The predicted value is then compared with the true value. If the predicted value is equal to the true value, the prediction is correct; if the predicted value is not equal to the true value, the prediction is wrong.

[0228] 10.2 Count the number of correct classifications and calculate the accuracy of the model:

[0229] First, count the number of samples predicted correctly in the validation set and the test set;

[0230] Then, the accuracy of the model is calculated by combining the test set and the validation set using the following formula:

[0231]

[0232] TP represents the true positive sample, that is, the sample that is actually positive and predicted to be positive;

[0233] TN represents true negative examples, that is, samples that are actually negative and predicted to be negative;

[0234] FP stands for false positive, which is a sample that is actually a negative class but is predicted to be a positive class;

[0235] FN represents false negative examples, that is, samples that are actually positive but predicted to be negative;

[0236] A high-accuracy model is obtained, and when new vibration data is input into the model, corresponding diagnostic results can be given quickly.

[0237] Through the above steps, the accuracy of the model calculated on the validation set can help monitor model performance and adjust hyperparameters, while the accuracy of the test set is an important indicator for evaluating the generalization ability of the model on unseen data. The calculation method of the validation set and the test set through formulas also provides a scientific basis for optimizing and evaluating model performance.

[0238] After completing the training, verification and testing of the model, the performance and generalization ability of the model have been verified through the accuracy evaluation of the test set, and it has good performance on unknown data. Next, the new vibration data is input into the trained model. The model will use its learned feature extraction and classification capabilities to quickly and accurately determine the fault state of the rolling bearing and output the corresponding diagnostic results, thereby realizing fault detection in actual application scenarios. Therefore, when the new vibration data is input into the model, the model can quickly and accurately determine the fault state of the rolling bearing and give the corresponding diagnostic results. The specific steps are as follows:

[0239] 11.1 Prepare new vibration data (vibration data to be diagnosed), input it into the model and output the diagnosis results:

[0240] First, new vibration signal data is collected in real time from the sensor installed on the rolling bearing to ensure that the sampling frequency and format of the collected data are consistent with the data during model training;

[0241] Then, the new data is processed according to the preprocessing steps of the training data, and the new time domain signal is converted into a specific feature representation, which is passed as input to the saved trained model. The model forward propagates the input data and outputs the prediction result. For the classification task, the model outputs the probability distribution belonging to each category and selects the category with the highest probability as the final diagnosis result.

[0242] The category label output by the final model is the fault status of the rolling bearing. In real-time applications, new vibration signals can be continuously input and fault diagnosis results can be generated in real time.

[0243] 11.2 Deployment Model:

[0244] Integrate the model into the monitoring system, deploy the model to the industrial monitoring platform, and combine it with the real-time data acquisition system to realize automatic fault diagnosis;

[0245] In the deployment environment, ensure that the model runs efficiently, meets the needs of real-time fault detection, and optimizes response speed.

[0246] Through the above steps, this embodiment uses the trained model to pre-process the new vibration data, extract the data of time domain and frequency domain features and input them into the model, use the forward propagation function of the model to generate classification results, and output the fault status, and then combine the real-time monitoring system to deploy the model to achieve fast and accurate rolling bearing fault diagnosis.

[0247] The present invention introduces FMD (Feature Mode Decomposition) technology when diagnosing bearing faults, taking into account the impulsiveness and periodicity of the signal at the same time, and does not need to take the fault period as prior knowledge; it uses convolutional neural networks to extract main features to speed up the calculation speed, and uses transfer learning technology to adjust some parameters of the pre-trained model to adapt to new vibration data, thereby reducing the training time of new data and improving model performance.

[0248] The embodiment of the present invention provides a bearing fault diagnosis method based on characteristic mode decomposition, further comprising:

[0249] When collecting data samples of rolling bearings under different fault conditions, a collection normative test is performed to determine the data samples that have passed the collection normative test. The collection normative test is used to detect whether the data collected at different collection locations in each equipment operation stage for different fault state analysis is comprehensive;

[0250] The working principle and beneficial effects of the above technical solution are:

[0251] The present invention performs collection and standardization detection of data samples, thereby improving the training quality of subsequent models.

[0252] In one embodiment, collecting normative detection includes:

[0253] Based on the bearing fault diagnosis knowledge graph, the diagnostic basis data corresponding to the fault state of the rolling bearing is obtained; the diagnostic basis data includes: the first equipment operation stage and the diagnostic basis sub-data corresponding to each other; wherein, the bearing fault diagnosis knowledge graph is: a structured knowledge base, which contains professional knowledge and experience about rolling bearing fault diagnosis, including information such as fault type, fault cause, fault characteristics, diagnostic method and maintenance strategy; the fault state is: all types of rolling bearing faults obtained according to the bearing fault diagnosis knowledge graph, such as: inner ring fault, outer ring fault and rolling element damage; the equipment operation stage includes: start, stabilize and stop; the sub-data of the equipment operation stage is: the source of the vibration signal for fault diagnosis analysis;

[0254] Acquire pre-collected data; wherein the pre-collected data is: data collected manually in advance for subsequent model training, such as: a vibration signal of the outer ring of the bearing during the startup phase when the outer ring fails;

[0255] Parse the pre-collected data to obtain collected sample data corresponding to different collected fault states; the collected sample data includes: one-to-one corresponding second equipment operation phase and collected sample sub-data;

[0256] Compare the diagnostic basis data corresponding to different fault states and the collected sample data corresponding to different collected fault states to determine whether the collected fault states, the second equipment operation stage, and the collected sample sub-data in the pre-collected data are all comprehensive;

[0257] If they are all comprehensive, the collection normativeness test is passed, and the corresponding pre-collected data is used as the data sample;

[0258] If it is not comprehensive, after supplementary collection, the supplementary collected data and the pre-collected data will be used together as data samples.

[0259] The working principle and beneficial effects of the above technical solution are:

[0260] The present invention introduces a bearing fault diagnosis knowledge graph and obtains diagnostic basis data corresponding to the fault state of the rolling bearing, parses the pre-collected data to obtain collected sample data corresponding to different collected fault states, compares the diagnostic basis data corresponding to different fault states with the collected sample data corresponding to different collected fault states, and determines whether the collected fault state, the second equipment operation stage and the collected sample sub-data in the pre-collected data are all comprehensive. If not, they are supplemented to improve the comprehensiveness of the data sample.

[0261] In one embodiment, the supplemental collection includes:

[0262] Obtain a data sample tree construction template; the data sample tree construction template extracts the target fault state as the root node data, extracts the target equipment operation stage as the first-level sub-node data, and extracts the collection part data as the second-level sub-node data in comparison with the data to be constructed; wherein the data sample tree construction template is used for generating a data sample tree in comparison with the data to be constructed (diagnosis basis data and collection sample data), the data sample tree takes the target fault state as the root node, and its subtrees include different vibration signal extraction parts in different equipment operation stages; the target fault state is the bearing fault type in the data to be constructed; the target equipment operation stage is the equipment operation stage in the data to be constructed;

[0263] Building a template based on the data sample tree and building a standard data sample tree based on the diagnostic data corresponding to different fault states;

[0264] Building a template based on the data sample tree and building a target data sample tree according to the collected sample data corresponding to different collection fault states;

[0265] According to the standard data sample tree and the target data sample tree, determining the missing subtree of the target data sample tree corresponding to the standard data sample tree;

[0266] Get the missing node with the smallest level in each missing subtree; where the missing node contains data information to be supplemented, for example, if the missing node is the root node whose fault state is inner ring fault, then the data information to be supplemented is the data sample of inner ring fault; for example, if the missing node is the startup stage child node under the root node whose fault state is inner ring fault, then the data information to be supplemented is the data sample of the startup stage when the fault state is inner ring fault (which parts of the vibration signal are needed in the startup stage to determine the inner ring fault);

[0267] Supplementary collection is performed based on the node attributes of the missing nodes.

[0268] The working principle and beneficial effects of the above technical solution are:

[0269] The present invention introduces a data sample tree construction template, organizes the data to be constructed (diagnostic basis data and collected sample data) into a tree structure (standard data sample tree and target data sample tree), compares the standard data sample tree and the target data sample tree to obtain the missing subtree, obtains the missing node with the smallest level in each missing subtree and performs supplementary collection according to the node attributes of the missing node. Structuring the data can obtain missing data information faster and improve the supplementation efficiency.

[0270] In one embodiment, supplementary collection is performed based on the node attributes of the missing nodes, including:

[0271] Connect to the industrial big data platform and perform supplementary collection based on the node attributes of the missing nodes. The industrial big data platform is a system that integrates a large amount of industrial data, including sensor data, equipment logs, production process data, etc., to support industrial production, monitoring, optimization, fault diagnosis and other activities;

[0272] The working principle and beneficial effects of the above technical solution are:

[0273] The present invention connects to the industrial big data platform for supplementary collection, and the collected data is more comprehensive.

[0274] In one embodiment, the industrial big data platform is connected and supplementary collection is performed according to the node attributes of the missing nodes, including:

[0275] Parse the node attributes of the missing node and obtain the node features; the node features are: missing node type, upper node type and node data collection conditions; the missing node type is: the node type of the missing node, such as: startup phase, the upper node type is, such as: inner ring fault; the node data collection conditions are: bearing parameters of the rolling bearing, such as: bearing model;

[0276] According to the missing node type and the upper node type, the information index condition is constructed; wherein the information index condition is: the condition of indexing the missing node corresponding to the missing data, for example: the index condition of indexing the data of each collected part in the startup phase when the fault state is an inner circle fault, which is determined according to the index rules preset by the platform, the missing node type and the upper node type, for example: fault state = (inner circle fault) and equipment operation phase = (startup state);

[0277] Indexing is performed on the industrial big data platform based on the information indexing conditions to obtain an indexing interface; wherein the indexing interface is: a front-end interface after indexing on the industrial big data platform based on the information indexing conditions;

[0278] Parse the interface script of the index interface to obtain the conditional description semantics of the interface data; the interface script is: the front-end script; the conditional description semantics is: the data collection conditions of the interface data, such as: the bearing information of the source bearing to be collected;

[0279] According to the node data collection conditions and condition description semantics, the target data is located; wherein the target data is: the interface data corresponding to the condition description semantics whose description conditions are consistent with the node data collection conditions;

[0280] Obtaining the capture rules of the target data in the interface script; wherein the capture rules are obtained according to the access rules preset in the index interface;

[0281] According to the capture rules, capture target data;

[0282] When the target data corresponding to all missing nodes are captured, the supplementary collection is completed.

[0283] The working principle and beneficial effects of the above technical solution are:

[0284] The present invention extracts the missing node type, upper node type and node data collection condition of the missing node, and constructs information index conditions according to the platform index rules, the missing node type and the upper node type; based on the information index conditions, indexing is performed on the industrial big data platform to obtain an index interface, and the conditional description semantics of the interface data is obtained according to the interface script of the index interface, and the target data corresponding to the node data collection condition is located; the target data is captured according to the capture rule of the target data, and when the target data corresponding to all the missing nodes are captured, the supplementary collection is completed, thereby realizing automatic indexing of the big data platform, interface analysis and accurate crawling of required data, which is more intelligent.

[0285] In one embodiment, based on a deep neural network and according to decomposition modes, a bearing fault diagnosis model is trained, including:

[0286] Perform time-frequency analysis on the decomposed modes to obtain the time-frequency distribution diagram;

[0287] Each time-frequency distribution diagram is traversed in turn, and the time-frequency distribution diagram being traversed is used as a side-frequency component analysis diagram; the side-frequency component analysis diagram is: whether there is a time-frequency distribution diagram of frequency components that appear symmetrically on both sides of a certain frequency in the analysis diagram;

[0288] Then, the frequency components in the side frequency component analysis graph are traversed in sequence, and the correlation analysis of the left frequency component and the right frequency component of the frequency component being traversed is performed; wherein the correlation analysis is a process of analyzing the symmetry of the left frequency component and the right frequency component;

[0289] If the result of the correlation analysis is correlation, the corresponding side frequency component analysis diagram is used as the target time-frequency distribution diagram; wherein, correlation means that the symmetry between the left frequency component and the right frequency component meets the preset situation, for example: at least two groups of components appear symmetrically between the left frequency component and the right frequency component;

[0290] Perform envelope demodulation processing according to the target time-frequency distribution diagram and draw the envelope curve;

[0291] Based on the envelope and sideband feature extraction template, the sideband feature is obtained; wherein the sideband feature extraction template is: a template for extracting the additional frequency components appearing near the fault characteristic frequency and its multiples against the envelope spectrum; the sideband is the multiples on both sides of the main fault characteristic frequency, and the sideband feature is: the distribution and interval of the multiples;

[0292] According to the data samples corresponding to the side frequency characteristics and the target time-frequency distribution diagram, the modulation characteristics are obtained; wherein the modulation characteristics include: adjusting the signal type and the modulation signal frequency, for example: the amplitude modulation signal will generate side frequencies on both sides of the fault characteristic frequency, and the frequency modulation signal will generate a series of side frequencies on both sides of the fault characteristic frequency;

[0293] Based on the deep neural network, the bearing fault diagnosis model is trained according to the decomposed mode, the sideband feature code and the modulation signal frequency. When training the bearing fault diagnosis model, the decomposed mode, the sideband feature code and the modulation signal frequency are used as the input of the fault diagnosis neural network model, and the labeled fault diagnosis results in the data sample are used as the output of the fault diagnosis neural network model.

[0294] The working principle and beneficial effects of the above technical solution are:

[0295] Composite faults can cause mutual interference and coupling between vibrations, making it more difficult to extract fault features. For example, the characteristic frequencies of rolling element and inner ring faults will be modulated by the characteristic frequencies of the retainer and the rotation frequency, resulting in a sideband effect. Strong fault features may annihilate weak fault features, making it difficult to identify weak fault features. Therefore, the present invention performs time-frequency analysis on the decomposed mode to obtain a time-frequency distribution diagram, and performs sideband component analysis in the time-frequency distribution diagram, performs envelope demodulation on the target time-frequency distribution diagram with sideband components, determines the sideband features, and further determines the adjustment features based on the sideband features; finally, the decomposed mode, sideband feature code and modulated signal frequency are used as the input of the fault diagnosis neural network model, and the marked fault diagnosis results in the data sample are used as the output of the fault diagnosis neural network model to train the bearing fault diagnosis model, thereby reducing the impact of composite fault interference on the diagnosis accuracy and achieving higher detection accuracy.

[0296] The embodiment of the present invention provides a bearing fault diagnosis system based on characteristic mode decomposition, such as Figure 2 As shown, including:

[0297] The characteristic mode decomposition subsystem 1 is used to obtain the decomposition mode of the data sample based on the characteristic mode decomposition technology;

[0298] A bearing fault diagnosis model training subsystem 2 is used to train a bearing fault diagnosis model based on a deep neural network and decomposed modes;

[0299] The vibration data acquisition subsystem 3 is used to acquire the vibration data to be diagnosed;

[0300] The diagnosis subsystem 4 is used to adjust the pre-trained model parameters in the bearing fault diagnosis model according to the vibration data to be diagnosed based on the transfer learning technology, and obtain the fault diagnosis result corresponding to the vibration data to be diagnosed output by the adjusted bearing fault diagnosis model.

[0301] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A bearing fault diagnosis method based on characteristic mode decomposition, characterized in that: include: Based on the characteristic mode decomposition technology, the decomposition mode of the data sample is obtained; Based on deep neural network, the bearing fault diagnosis model is trained according to the decomposed mode; Acquire vibration data to be diagnosed; Based on the transfer learning technology, the pre-trained model parameters in the bearing fault diagnosis model are adjusted according to the vibration data to be diagnosed, and the fault diagnosis result corresponding to the vibration data to be diagnosed output by the adjusted bearing fault diagnosis model is obtained.

2. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 1, characterized in that: Based on the characteristic mode decomposition technology, the decomposition mode of the data sample is obtained, including: Collect data samples of rolling bearings under different fault conditions; By initializing a set of uniformly distributed filters covering the entire frequency band, the data samples are decomposed to obtain multiple modes; The filter is iteratively updated with CK as the objective function and the fault period is estimated in each filtering process; By comparing the correlation coefficients CC of the two modes, the mode with the smallest CK is discarded from the two modes with the largest CC; When the termination condition is reached, the decomposition is terminated and the retained mode is obtained as the final decomposition mode; The final decomposed modal data is normalized and preprocessed to obtain the decomposed mode, so that the data with different characteristics are in the same numerical range.

3. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 2, characterized in that: Collect data samples of rolling bearings under different fault conditions, including: Acquire the collected samples including acceleration signal gs, sampling rate sr, shaft speed rate, load weight load and four critical frequencies representing different fault locations as data samples. The four critical frequencies include: outer ball passing frequency BPFO, inner ball passing frequency BPFI, basic train frequency FTF and ball rotation frequency BSF. The calculation formulas of outer ball passing frequency BPFO and inner ball passing frequency BPFI are as follows: Where d is the ball diameter, D is the pitch diameter, and the variable f r is the shaft speed, n is the number of rolling elements, and Φ is the bearing contact angle.

4. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 1, characterized in that: Based on the deep neural network and decomposition mode, the bearing fault diagnosis model is trained, including: Based on deep neural network, target features are extracted according to decomposition mode; The target features are divided into training set, validation set and test set by random partitioning method; Classify scale maps by fine-tuning the pre-trained SqueezeNet convolutional neural network; The validation set is used to monitor the performance of the model and the trained model is evaluated through the test set to obtain a bearing fault diagnosis model that passes the evaluation.

5. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 2, characterized in that: Also includes: When collecting data samples of rolling bearings under different fault conditions, a collection compliance test is performed to determine the data samples that pass the collection compliance test.

6. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 5, characterized in that: Collect regulatory testing, including: Based on the bearing fault diagnosis knowledge graph, the diagnosis basis data corresponding to the fault state of the rolling bearing is obtained; the diagnosis basis data includes: a one-to-one corresponding first equipment operation stage and diagnosis basis sub-data; Get pre-collected data; Parse the pre-collected data to obtain collected sample data corresponding to different collected fault states; the collected sample data includes: one-to-one corresponding second equipment operation phase and collected sample sub-data; Compare the diagnostic basis data corresponding to different fault states and the collected sample data corresponding to different collected fault states to determine whether the collected fault states, the second equipment operation stage, and the collected sample sub-data in the pre-collected data are all comprehensive; If they are all comprehensive, the collection normativeness test is passed, and the corresponding pre-collected data is used as the data sample; If it is not comprehensive, after supplementary collection, the supplementary collected data and the pre-collected data will be used together as data samples.

7. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 6, characterized in that: Supplementary collection, including: Obtain a data sample tree construction template; the data sample tree construction template extracts the target fault state as root node data, extracts the target equipment operation stage as first-level sub-node data, and extracts the collection location data as second-level sub-node data according to the data to be constructed; Building a template based on the data sample tree and building a standard data sample tree based on the diagnostic data corresponding to different fault states; Building a template based on the data sample tree and building a target data sample tree according to the collected sample data corresponding to different collection fault states; According to the standard data sample tree and the target data sample tree, determining the missing subtree of the target data sample tree corresponding to the standard data sample tree; Get the missing node with the smallest level in each missing subtree; Supplementary collection is performed based on the node attributes of the missing nodes.

8. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 7, characterized in that: Supplementary collection is performed based on the node attributes of the missing nodes, including: Connect to the industrial big data platform and conduct supplementary collection based on the node attributes of the missing nodes.

9. A bearing fault diagnosis method based on characteristic mode decomposition as claimed in claim 8, characterized in that: Connect to the industrial big data platform and perform supplementary collection based on the node attributes of the missing nodes, including: Parse the node attributes of the missing nodes and obtain node features; the node features are: the type of missing nodes, the type of upper-layer nodes, and the node data collection conditions; According to the missing node types and upper-level node types, information index conditions are constructed; Index the industrial big data platform based on information index conditions and obtain the index interface; Parse the interface script of the index interface to obtain the conditional description semantics of the interface data; Locate target data based on node data collection conditions and condition description semantics; Get the capture rules of target data in the interface script; According to the capture rules, capture target data; When the target data corresponding to all missing nodes are captured, the supplementary collection is completed.

10. A bearing fault diagnosis system based on characteristic mode decomposition, characterized in that: include: The characteristic mode decomposition subsystem is used to obtain the decomposition mode of the data sample based on the characteristic mode decomposition technology; The bearing fault diagnosis model training subsystem is used to train the bearing fault diagnosis model based on the deep neural network and the decomposed mode; A vibration data acquisition subsystem for diagnosis, used for acquiring vibration data for diagnosis; The diagnosis subsystem is used to adjust the pre-trained model parameters in the bearing fault diagnosis model according to the vibration data to be diagnosed based on the transfer learning technology, and obtain the fault diagnosis result corresponding to the vibration data to be diagnosed output by the adjusted bearing fault diagnosis model.

Images