Multi-source feature adaptive bearing fault diagnosis method and system

Abstract

The application discloses a kind of multi-source feature adaptive bearing fault diagnosis method and system, method includes: obtaining bearing vibration signal, and bearing vibration signal is extracted to feature, and the cooperative feature set is obtained, cooperative feature set includes time domain feature set, frequency domain feature set and wavelet packet energy feature set;According to the feature selection in cooperative feature set in the pre-set improved regularization feature selection method, and the feature subset is obtained;Multi-task loss function is constructed, and neural network is constructed according to multi-task loss function, and neural network is iteratively trained according to feature subset, and the bearing fault diagnosis model is obtained;Real-time bearing vibration signal obtained is input to bearing fault diagnosis model, and the fault diagnosis result corresponding to real-time bearing vibration signal is obtained by bearing fault diagnosis model output. It can effectively reduce equipment maintenance cost, improve production safety and reliability.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis technology, and in particular relates to a multi-source feature adaptive bearing fault diagnosis method and system. Background Technology

[0002] During long-term operation, equipment vibration, temperature, and current signals are often affected by various factors such as environmental noise, changes in operating conditions, and sensor errors. This poses a significant challenge to traditional fault detection methods when dealing with variable operating conditions and small sample data. While traditional signal analysis methods have accumulated rich experience in feature extraction and can reflect the health status of equipment to some extent, they often rely on a single feature extraction method, making it difficult to comprehensively capture the multidimensional information hidden during equipment operation. This limits their ability to identify complex fault modes.

[0003] In recent years, deep learning has made significant progress in the field of intelligent fault diagnosis. Models such as convolutional neural networks, long short-term memory networks, and variational autoencoders can automatically extract features to achieve end-to-end fault detection, avoiding the reliance on manual features in traditional methods. However, most of these models are "black box" structures, lacking physical interpretability and making it difficult to clearly identify the causes of faults. Summary of the Invention

[0004] This invention provides a multi-source feature adaptive bearing fault diagnosis method and system to solve the technical problem that model-based bearing fault diagnosis lacks physical interpretability and makes it difficult to determine the cause of the fault.

[0005] In a first aspect, the present invention provides a multi-source feature adaptive bearing fault diagnosis method, comprising:

[0006] The bearing vibration signal is acquired, and features are extracted from the bearing vibration signal to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set.

[0007] According to the preset improved L1 regularized feature selection method, feature selection is performed in the collaborative feature set to obtain a feature subset;

[0008] A multi-task loss function is constructed, a neural network is constructed based on the multi-task loss function, and the neural network is iteratively trained based on the feature subset to obtain a bearing fault diagnosis model;

[0009] The acquired real-time bearing vibration signal is input into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signal.

[0010] Secondly, the present invention provides a multi-source feature adaptive bearing fault diagnosis system, comprising:

[0011] The extraction module is configured to acquire bearing vibration signals and perform feature extraction on the bearing vibration signals to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set.

[0012] The filtering module is configured to perform feature filtering in the collaborative feature set according to a preset improved L1 regularized feature selection method to obtain a feature subset;

[0013] The module is configured to construct a multi-task loss function, construct a neural network based on the multi-task loss function, and iteratively train the neural network based on the feature subset to obtain a bearing fault diagnosis model.

[0014] The output module is configured to input the acquired real-time bearing vibration signal into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signal.

[0015] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the multi-source feature adaptive bearing fault diagnosis method of any embodiment of the present invention.

[0016] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the multi-source feature adaptive bearing fault diagnosis method according to any embodiment of the present invention.

[0017] This application presents a multi-source feature adaptive bearing fault diagnosis method and system. Employing multiple signal processing methods, including time-domain, frequency-domain, and wavelet packet transform, it achieves comprehensive, multi-scale feature extraction of equipment vibration signals. This overcomes the information loss problem caused by traditional methods relying solely on a single feature domain, improving the robustness and accuracy of fault diagnosis. Furthermore, physical constraints are introduced into the neural network. By jointly training physical index prediction and fault classification tasks, the model's internal feature expressions possess clear physical meaning, enhancing the model's transparency and reliability in engineering applications. This method effectively reduces equipment maintenance costs, improves production safety and reliability, and provides an efficient and transparent fault diagnosis approach for intelligent manufacturing and industrial automation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a multi-source feature adaptive bearing fault diagnosis method provided in an embodiment of the present invention;

[0020] Figure 2 This is a structural diagram of the multi-task physical constraint network constructed in this invention;

[0021] Figure 3 This is a feature importance analysis diagram in a preferred embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram illustrating the verification accuracy under different feature selection numbers in a preferred embodiment of the present invention;

[0023] Figure 5 This is a T-SNE visualization of bearing fault categories before multi-domain feature screening in a preferred embodiment of the present invention;

[0024] Figure 6 This is a T-SNE visualization of bearing fault categories after multi-domain feature filtering in a preferred embodiment of the present invention;

[0025] Figure 7 The diagram shows the accuracy curves and total loss curves for the training set and validation set in a preferred embodiment of the present invention.

[0026] Figure 8 This is a T-SNE visualization of the shared features of the intermediate layer of the test set in a preferred embodiment of the present invention.

[0027] Figure 9 This is a schematic diagram of the test set confusion matrix in a preferred embodiment of the present invention;

[0028] Figure 10 This is a structural block diagram of a multi-source feature adaptive bearing fault diagnosis system provided in an embodiment of the present invention;

[0029] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 1 The flowchart of a multi-source feature adaptive bearing fault diagnosis method of this application is shown.

[0032] like Figure 1 As shown, the multi-source feature adaptive bearing fault diagnosis method specifically includes the following steps:

[0033] Step S101: Obtain the bearing vibration signal and extract features from the bearing vibration signal to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set.

[0034] In this step, statistical analysis is performed on the bearing vibration signal to extract a time-domain feature set, which includes root mean square, variance, kurtosis, skewness, peak factor, and margin factor. Fourier transform is used to convert the bearing vibration signal to the frequency domain to extract a frequency-domain feature set, which includes the spectral centroid and equivalent bandwidth. Wavelet packet analysis is used to extract the energy features of the bearing vibration signal to obtain a wavelet packet energy feature set.

[0035] Specifically, statistical analysis is performed on the bearing vibration signal to extract six time-domain features, including root mean square, variance, kurtosis, skewness, peak factor, and margin factor; let the signal sequence be X = [X1, X2, ..., X...]. N The number of signal samples is N, and x(i) is the i-th sampled value. Let μ1 be the sample mean, and max|x be the absolute value of the mean. i | represents the maximum absolute value of the signal;

[0036] Root mean square:

[0037] variance:

[0038] Kuroshi:

[0039] Skewness:

[0040] Peak factor:

[0041] Margin factor:

[0042] The bearing vibration signal was converted to the frequency domain using Fourier transform, and two frequency domain features were extracted: the spectral centroid and the equivalent bandwidth; the positive frequency component was denoted as... make P k =|X k | 2 , Among them, F s X is the sampling frequency. k f is the complex coefficient of the k-th FFT. k For the k-th frequency, P k For the k-th power spectral component, P tot This is the total power spectrum;

[0043] Spectral centroid:

[0044] Equivalent bandwidth:

[0045] Two energy features of the bearing vibration signal were extracted using wavelet packet analysis. The signal was then decomposed into three levels using the db1 wavelet, yielding eight nodes (frequency bands) in the third level. The coefficients of these nodes are denoted as {w... j，m}, where j = 1, 2, ..., M represents the j-th node, m represents the sample index within that node, and the node energy is . Total energy is If M is the total number of nodes at level 3, then F9 = E tot ,

[0046] Step S102: According to the preset improved L1 regularized feature selection method, feature filtering is performed in the collaborative feature set to obtain a feature subset.

[0047] In this step, a multi-class logistic regression with L1 penalty is solved to obtain a set of sparse weights w, wherein the expression for the multi-class logistic regression is:

[0048]

[0049] In the formula, N is the total number of samples, w T For the model weight vector, The model's prediction score for the true category. Scores for different categories, Let be the weight sub-vector of the k-th class. Let y be the category label of the i-th sample. i The weighted subvector, x iLet y be the feature vector of the i-th sample. i Let be the class label of the i-th sample, α′ be the regularization term strength coefficient, and ||w||1 be the L1 regularization term;

[0050] With |w j |As an importance index of the j-th feature, each feature in the collaborative feature set is sorted from largest to smallest;

[0051] For each candidate feature set, the top p feature subsets are selected, the classifier is retrained, and K-fold cross-validation is performed to calculate the average accuracy. The expression for K-fold cross-validation is as follows:

[0052]

[0053] In the formula, The overall accuracy rate is represented by K, where K is the fold number. Let n be the classification accuracy of the i-th fold. i Let D be the number of samples at the i-th fold. i Let y be the set of samples belonging to the i-th fold. j Let j be the true label of the j-th sample. Let j be the predicted label for the j-th sample. The value is 1 if the condition inside the parentheses is true, and 0 otherwise.

[0054] Select the p-value that maximizes the cross-validation accuracy and the features corresponding to the p-value to obtain a feature subset.

[0055] Step S103: Construct a multi-task loss function, build a neural network based on the multi-task loss function, and iteratively train the neural network based on the feature subset to obtain a bearing fault diagnosis model.

[0056] In this step, for N samples, assuming the model outputs a class probability vector of (where is the probability of the correct class), the sparse cross-entropy classification loss is defined as: In the formula, N is the sample size. The model predicts y for the i-th sample. i The probability value of the category; the physical quantity regression loss uses mean squared error to measure the difference between the predicted physical indicator and the pre-given expected physical label. The model needs to predict the values ​​of these three physical parameters: q = [Rms, Kurtosis, Crest] T Let the shared feature be h, then the prediction of the physical constraint branch is: Among them, W q Let b be the weight matrix. q The bias vector is then used, and the mean squared error is used to measure the difference between the predicted and actual values: In the formula, in the formula, Let q be the predicted physical parameter value of the model for the i-th sample. (i) Let be the true physical parameter value of the i-th sample, Rms be the root mean square, Kurtosis be the kurtosis, Crset be the peak factor, and q be the set of physical parameters that the model needs to predict.

[0057] Physical consistency constraint loss L phys Including damped oscillator constraint loss L ODE and boundary condition loss L BC L ODE The structure is as follows: the signal u(t) satisfies the damped oscillator equation. Where u(t) is the signal at time t. and Let ω and ζ be the first and second derivatives of u(t), respectively, where ω is the natural frequency and ζ is the damping ratio. Since the data is discrete, the central difference method is used to approximate the first derivative. And the second derivative u(t) is and Substituting the above expression into the damped oscillator equation using the discretized approximation yields the residual at each time step. The constraint loss is obtained by averaging the squared residuals at each time step. In the formula, N is the sample size. The model predicts y for the i-th sample. i The probability value of the category;

[0058] Boundary condition loss L BC The structure is as follows: First, a boundary function is constructed as the prior of the signal. Where T′ is the number of sampling points minus 1, the boundary function expression represents the signal transitioning from the initial value to the final value through linear interpolation over the entire time interval, and the loss is calculated as u(t) and u... bj The mean square error between (t) The weighted average of the two yields the physical consistency constraint loss: L phys =αL ODE +βL BC α is the damped oscillator loss coefficient, L ODE For the damped oscillator loss, β is the boundary condition loss coefficient, and L is the loss coefficient. BC For boundary condition loss;

[0059] The expression for the multi-task loss function of the entire model is: L total =λ class L class +λ reg L reg +λ phys L phys L total For multi-task loss function, λclass L represents the weights of the sparse cross-entropy classification loss. class For sparse cross-entropy classification loss, λ reg L represents the weight of the loss in the physical quantity regression task. reg For the physical quantity regression task loss, λ phys L represents the weight of the loss due to physical consistency constraints. phys The loss is a physical consistency constraint; this allows for simultaneous optimization of fault classification accuracy, physical quantity regression accuracy, and system dynamics consistency during training.

[0060] It should be noted that the neural network adopts a hierarchical architecture of input-convolutional multi-branch-residual fusion-multi-head attention mechanism-global convergence-multi-task output: First, it receives 15 steps of single-channel temporal signals. After residual addition within each of two parallel 1-D convolutional branches (64 and 128 channels), the branch features are concatenated in the channel dimension and fed into a deep 256-channel convolutional residual unit to extract multi-scale local patterns. Subsequently, two levels of bidirectional gated recursive units capture contextual dependencies and accumulate temporal information in a residual manner, and then multi-head self-attention is used to strengthen long-range associations. Afterward, global average pooling compresses the temporal features into a 256-dimensional vector, and then passes through two layers containing Dr The fully connected network with dropout enables feature remapping, ultimately outputting 10-dimensional classification probabilities and 3-dimensional continuous physical quantities in parallel, achieving dual tasks of classification and regression. The classification branch uses fully connected layers and dropout layers based on shared features to predict the fault type, and the output uses the Softmax activation function. The physical constraint branch is also based on shared features, outputting predicted values ​​of physical indicators through fully connected layers and dropout layers, using a linear activation function to embed physical prior information into the neural network model. The physical constraint branch is used to regularize the fault diagnosis task, and this branch serves as an auxiliary task to enhance the model's learning of the intrinsic physical characteristics of the signal.

[0061] Step S104: The acquired real-time bearing vibration signal is input into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signal.

[0062] In summary, the method presented in this application employs multiple signal processing techniques, including time-domain, frequency-domain, and wavelet packet transform, to achieve comprehensive, multi-scale feature extraction of equipment vibration signals. This overcomes the information loss problem caused by traditional methods relying solely on a single feature domain, improving the robustness and accuracy of fault diagnosis. Furthermore, physical constraints are introduced into the neural network. By jointly training physical index prediction and fault classification tasks, the internal feature expressions of the model possess clear physical meaning, enhancing the model's transparency and reliability in engineering applications. This method can effectively reduce equipment maintenance costs, improve production safety and reliability, and provide an efficient and transparent fault diagnosis approach for intelligent manufacturing and industrial automation.

[0063] In one specific embodiment, experimental data of rolling bearings were measured. The test bench consisted of a motor, torque sensor / encoder, power meter, and control electronics. A single-point fault was introduced into the test bearing through electrical discharge machining. Accelerometers were installed at the drive end and fan end of the motor housing, respectively, and vibration data were collected at a sampling frequency of 12kHz. This invention used 10 types of data for analysis and experimentation. Normal data and fault diameters of 0.1778mm (including inner ring, outer ring, and rolling elements), 0.3556mm (including inner ring, outer ring, and rolling elements), and 0.5334mm (including inner ring, outer ring, and rolling elements) were selected for analysis at a sampling frequency of 12kHz and an operating condition of 1750rpm. These were labeled with fault tags 0 to 9.

[0064] In the data preprocessing stage, the original continuous signal is first segmented based on a fixed sampling length (784 sampling points) using a sliding window method with a window sliding step size of 150. This ensures that there is partial overlap between adjacent segments, effectively capturing local changes within the signal while maximizing the continuity and integrity of the signal. Furthermore, to ensure the richness and balance of the training data, a fixed number of segmented samples (600 per class) are uniformly extracted from each fault category. Subsequently, all segmented samples and their corresponding labels are randomly shuffled and divided into training, validation, and test sets according to a predetermined ratio (70%:15%:15%), used for model training, model parameter tuning, and model performance evaluation, respectively. To eliminate the influence of different data dimensions and units, the above datasets are further standardized and normalized to ensure that feature values ​​are within a uniform scale, thereby improving the stability, convergence speed, and generalization performance of model training. Finally, to meet the input dimension requirements of deep learning networks, the samples in each dataset were uniformly adjusted to a tensor format with a shape of (sampling length, 1), thus maintaining consistency with the data input requirements of convolutional and recurrent layers in the network.

[0065] Features were extracted from the input raw signal in the time domain, frequency domain, and wavelet transform. The multidimensional features obtained after multi-domain fusion were normalized to ensure comparisons were made under the same scale. Subsequently, L1 regularization-based feature selection was performed, choosing a subset with higher diagnostic contributions. The feature importance analysis is shown in the figure below. Figure 2 As shown, the optimal number of features is then determined through optimization methods to input into the constructed neural network, thereby reducing computational complexity and improving the model's generalization ability. The validation accuracy under different feature selection numbers is shown below. Figure 3 As shown, a feature count of 7 achieves excellent results while reducing dimensionality and improving model training efficiency. Based on the fault category and its corresponding physical characteristics, expected values ​​are pre-set to generate auxiliary physical labels for each sample. These labels guide the deep model to simultaneously focus on fault classification and physical parameter regression during the learning process, improving the model's ability to characterize the physical essence of the signal.

[0066] A bearing fault model based on multi-domain feature fusion and physical constraints is trained, and fault diagnosis is performed using the measured bearing fault signals. The neural network structure diagram constructed in this invention is shown below. Figure 4 As shown, the deep neural network consists of an input layer, a shared feature extraction layer, a global average pooling layer, a classification branch, and a physical constraint branch. The shared feature extraction layer employs multi-scale convolution, residual connections, bidirectional gated recurrent units, and multi-head attention mechanisms to fully capture local and global temporal features in the signal. The global average pooling layer maps features to a fixed dimension before feeding them into the classification and physical constraint branches, respectively, to predict fault categories and regress physical parameters. The feature-filtered data is input into the constructed deep neural network and trained using a joint loss function. The joint loss function is composed of a weighted average of the classification loss and the physical regression loss. In this experiment, the weight of the classification loss is 0.8, the weight of the physical parameter regression loss is 0.2, the weight of the damped oscillator constraint loss is 0.1, and the weight of the boundary condition loss is 0.1. During training, a dynamic learning rate scheduling strategy is used to monitor the validation set loss. That is, when the validation loss does not decrease significantly after 5 consecutive training rounds, the scheduler will multiply the current learning rate by a decay factor (set to 0.5 here) to reduce the learning rate. It is expected that a lower learning rate will help the model escape local optima and achieve more stable convergence. The scheduler sets a minimum learning rate (1e-6 here) to ensure that the learning rate does not decrease to an excessively small value and avoid the training process becoming too slow.

[0067] Design a fault diagnosis experiment.

[0068] The trained model is evaluated and visualized according to the above requirements to obtain fault diagnosis results. The visualization of the original features before feature selection using T-SNE is as follows: Figure 5As shown, however, there is still a lot of aliasing between some categories. The T-SNE visualization of the features after feature filtering using the optimized feature extraction method proposed in this invention is shown below. Figure 6 As shown in the figure, feature filtering removes redundant features and retains the most discriminative features. Points of the same class are more tightly clustered in the T-SNE space, while points of different classes are more spaced apart. This means that in the dimensionality-reduced visualization space, the class boundaries are more distinct, indicating that the filtered features have stronger discriminative power against different classes. The accuracy and total loss curves for the training and validation sets are shown below. Figure 7 As shown, the T-SNE visualization of shared features in the middle layer of the test set is as follows: Figure 8 As shown, the test set confusion matrix is ​​as follows: Figure 9 As shown.

[0069] Please see Figure 10 The diagram shows a structural block diagram of a multi-source feature adaptive bearing fault diagnosis system according to this application.

[0070] like Figure 10 As shown, the multi-source feature adaptive bearing fault diagnosis system 200 includes an extraction module 210, a filtering module 220, a construction module 230, and an output module 240.

[0071] The extraction module 210 is configured to acquire bearing vibration signals and extract features from the bearing vibration signals to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set. The filtering module 220 is configured to filter features in the collaborative feature set according to a preset improved L1 regularized feature selection method to obtain a feature subset. The construction module 230 is configured to construct a multi-task loss function, construct a neural network according to the multi-task loss function, and iteratively train the neural network according to the feature subset to obtain a bearing fault diagnosis model. The output module 240 is configured to input the acquired real-time bearing vibration signals into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signals.

[0072] It should be understood that Figure 10 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 10 The various modules in the document will not be described in detail here.

[0073] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the multi-source feature adaptive bearing fault diagnosis method in any of the above method embodiments.

[0074] In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows:

[0075] The bearing vibration signal is acquired, and features are extracted from the bearing vibration signal to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set.

[0076] According to the preset improved L1 regularized feature selection method, feature selection is performed in the collaborative feature set to obtain a feature subset;

[0077] A multi-task loss function is constructed, a neural network is constructed based on the multi-task loss function, and the neural network is iteratively trained based on the feature subset to obtain a bearing fault diagnosis model;

[0078] The acquired real-time bearing vibration signal is input into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signal.

[0079] Computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the multi-source characteristic adaptive bearing fault diagnosis system, etc. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the multi-source characteristic adaptive bearing fault diagnosis system via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0080] Figure 11 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 11 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 11Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the multi-source feature adaptive bearing fault diagnosis method described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the multi-source feature adaptive bearing fault diagnosis system. The output device 340 may include a display screen or other display device.

[0081] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0082] In one implementation, the above-described electronic device is applied to a multi-source characteristic adaptive bearing fault diagnosis system for a client, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0083] The bearing vibration signal is acquired, and features are extracted from the bearing vibration signal to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set.

[0084] According to the preset improved L1 regularized feature selection method, feature selection is performed in the collaborative feature set to obtain a feature subset;

[0085] A multi-task loss function is constructed, a neural network is constructed based on the multi-task loss function, and the neural network is iteratively trained based on the feature subset to obtain a bearing fault diagnosis model;

[0086] The acquired real-time bearing vibration signal is input into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signal.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-source feature adaptive bearing fault diagnosis method, characterized in that, include: The bearing vibration signal is acquired, and features are extracted from the bearing vibration signal to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set. Based on the preset improvement The regularized feature selection method performs feature filtering on the collaborative feature set to obtain a feature subset; A multi-task loss function is constructed, a neural network is constructed based on the multi-task loss function, and the neural network is iteratively trained based on the feature subset to obtain a bearing fault diagnosis model. The multi-task loss function includes sparse cross-entropy classification loss, physical quantity regression task loss, and physical consistency constraint loss. The expression for the multi-task loss function is: ， In the formula, For multi-task loss function, The weights are the sparse cross-entropy classification loss. For sparse cross-entropy classification loss, The weights for the loss in the physical quantity regression task. For the physical quantity regression task loss, The weights for the loss due to physical consistency constraints. Loss due to physical consistency constraints; The expression for calculating the sparse cross-entropy classification loss is as follows: ， In the formula, For the sample size, The model predicts the i-th sample as follows: The probability value of the category; The expression for calculating the loss of the physical quantity regression task is: ， ， In the formula, Let be the predicted physical parameters of the model for the i-th sample. Let i be the true physical parameter value of the i-th sample. It is the root mean square. For kurtosis, As the peak factor, This is the set of physical parameters that the model needs to predict. The expression for calculating the loss due to the physical consistency constraint is as follows: ， In the formula, The damping oscillator loss coefficient, For the damping oscillator loss, The boundary condition loss coefficient, For boundary condition loss; ， ， In the formula, The duration of the signal. The reconstructed signal output by the physical constraint branch. For ideal signals constrained by boundary conditions. This is the initial value of the vibration signal. This is the end value of the vibration signal; ， ， In the formula, The residuals of the damped oscillator equations, The signal value at time step t+1. The signal value at time step t-1, For the damping ratio, It is the natural frequency; The acquired real-time bearing vibration signal is input into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signal.

2. The multi-source feature adaptive bearing fault diagnosis method according to claim 1, characterized in that, The feature extraction of the bearing vibration signal to obtain the collaborative feature set includes: The bearing vibration signal is statistically analyzed to extract the time-domain feature set, which includes root mean square, variance, kurtosis, skewness, peak factor and margin factor. The bearing vibration signal is converted to the frequency domain using Fourier transform, and the frequency domain feature set is extracted. The frequency domain feature set includes the spectral centroid and the equivalent bandwidth. The energy features of the bearing vibration signal are extracted using wavelet packet analysis to obtain a set of wavelet packet energy features.

3. The multi-source feature adaptive bearing fault diagnosis method according to claim 1, characterized in that, The improvement based on the preset The regularized feature selection method performs feature filtering on the collaborative feature set to obtain a feature subset including: Solve the tape The multi-class logistic regression with penalties yields a set of sparse weights. The expression for the multi-class logistic regression is: ， ， In the formula, For the model weight vector, The model's prediction score for the true category. Scores for different categories, Let be the weight sub-vector of the k-th class. For the first Category labels for each sample The weight sub-vectors, For the first The feature vector of each sample For the first The category label of each sample, The regularization term intensity coefficient, for Regularization term; by As an importance index for the j-th feature, each feature in the collaborative feature set is sorted from largest to smallest; For each candidate feature number, the top-ranked features are selected. Given a subset of features, retrain the classifier and perform K-fold cross-validation to calculate the average accuracy. The expression for K-fold cross-validation is: ， In the formula, For overall accuracy, For folds, Let be the classification accuracy of the i-th fold. Let be the number of samples at the i-th fold. Let be the set of samples belonging to the i-th fold. Let j be the true label of the j-th sample. Let j be the predicted label for the j-th sample. The value is 1 if the condition inside the parentheses is true, and 0 otherwise. Select the one that maximizes the cross-validation accuracy. Value and the stated The feature corresponding to the value is used to obtain a feature subset.

4. A multi-source feature adaptive bearing fault diagnosis system, characterized in that, include: The extraction module is configured to acquire bearing vibration signals and perform feature extraction on the bearing vibration signals to obtain a collaborative feature set, which includes a time-domain feature set, a frequency-domain feature set, and a wavelet packet energy feature set. The filtering module is configured to filter based on preset improvements. The regularized feature selection method performs feature filtering on the collaborative feature set to obtain a feature subset; The module is configured to construct a multi-task loss function, construct a neural network based on the multi-task loss function, and iteratively train the neural network based on the feature subset to obtain a bearing fault diagnosis model. The multi-task loss function includes sparse cross-entropy classification loss, physical quantity regression task loss, and physical consistency constraint loss. The expression for the multi-task loss function is: ， In the formula, For multi-task loss function, The weights are the sparse cross-entropy classification loss. For sparse cross-entropy classification loss, The weights for the loss in the physical quantity regression task. For the physical quantity regression task loss, The weights for the loss due to physical consistency constraints. Loss due to physical consistency constraints; The expression for calculating the sparse cross-entropy classification loss is as follows: ， In the formula, For the sample size, The model predicts the i-th sample as follows: The probability value of the category; The expression for calculating the loss of the physical quantity regression task is: ， ， In the formula, Let be the predicted physical parameters of the model for the i-th sample. Let i be the true physical parameter value of the i-th sample. It is the root mean square. For kurtosis, As the peak factor, This is the set of physical parameters that the model needs to predict. The expression for calculating the loss due to the physical consistency constraint is as follows: ， In the formula, The damping oscillator loss coefficient, For the damping oscillator loss, The boundary condition loss coefficient, For boundary condition loss; ， ， In the formula, The duration of the signal. The reconstructed signal output by the physical constraint branch. For ideal signals constrained by boundary conditions. This is the initial value of the vibration signal. This is the end value of the vibration signal; ， ， In the formula, The residuals of the damped oscillator equations, The signal value at time step t+1. The signal value at time step t-1, For the damping ratio, It is the natural frequency; The output module is configured to input the acquired real-time bearing vibration signal into the bearing fault diagnosis model, and the bearing fault diagnosis model outputs a fault diagnosis result corresponding to the real-time bearing vibration signal.

5. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method according to any one of claims 1 to 3.

Images