Bearing fault artificial intelligence diagnosis method based on physical information embedding

Abstract

The invention discloses a bearing fault artificial intelligence diagnosis method based on physical information embedding. The method comprises the following steps: firstly, mapping a non-stationary vibration signal to a time-frequency domain through a time-frequency analysis method; then calculating a marginal spectrum to integrate time-frequency energy distribution characteristics, realizing energy induction of different frequency components, and effectively revealing potential modulation components and fault frequencies in the signals; on the basis, an amplitude center data value at a position corresponding to the typical fault frequency in the envelope spectrum is extracted as a quantitative evaluation index, so that physically interpretable fault information is explicitly embedded into model input; and finally, realizing automatic classification and recognition of the health state of the bearing by constructing a one-dimensional convolutional neural network which is light in structure and high in feature extraction capability. The diagnosis process provided by the invention gives consideration to the interpretability of signal processing and the nonlinear modeling capability of the depth model, is suitable for the detection task of early weak faults, and has a good application prospect in a complex industrial environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and particularly relates to an artificial intelligence diagnosis method for bearing faults based on the embedding of physical information. Background Art

[0002] In modern mechanical equipment, rolling element bearings, as core support components, are widely used in various rotating machinery and electric motors. Such bearings are typically composed of an inner ring, an outer ring, rolling elements (such as spheres or rollers), and a cage, etc. Their main task is to provide support for rotating components, bear operating loads, and reduce frictional resistance during operation. Although these components have been fully considered for their operating efficiency and stability during the design stage, they are still prone to wear, fatigue cracks, and even structural damage under the continuous influence of high loads, high speeds, and complex environments. If a bearing fails, it may lead to deterioration of equipment performance, increased energy consumption, and further trigger a series of mechanical failures or equipment shutdowns, causing significant economic losses and safety hazards. Therefore, carrying out real-time monitoring of the operating state of rolling bearings and early fault identification has important engineering value and practical significance for ensuring the stability of mechanical systems and extending their service life. Summary of the Invention

[0003] In order to overcome the deficiencies of the prior art, the present invention provides an artificial intelligence diagnosis method for bearing faults based on the embedding of physical information. First, the non-stationary vibration signal is mapped to the time-frequency domain through a time-frequency analysis method; then the marginal spectrum is calculated to integrate the time-frequency energy distribution characteristics, realize the energy induction of different frequency components, and effectively reveal the potential modulation components and fault frequencies in the signal; on this basis, the amplitude center data value at the position corresponding to the typical fault frequency in the envelope spectrum is extracted as a quantitative evaluation index, so as to explicitly embed the physically interpretable fault information into the model input; finally, by constructing a one-dimensional convolutional neural network with a lightweight structure and strong feature extraction ability, the automatic classification and identification of the bearing health state are realized. The diagnostic process proposed by the present invention takes into account both the interpretability of signal processing and the non-linear modeling ability of the deep model, is applicable to the detection task of early weak faults, and has good application prospects in complex industrial environments.

[0004] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0005] Step 1: Signal decomposition based on the extreme value distribution of the marginal spectrum;

[0006] Extract the local components in the signal associated with the fault physical characteristics through a time-frequency analysis method;

[0007] Step 2: Mode screening:

[0008] Screen the key mode that can best represent the fault characteristics from all decomposed modes;

[0009] Step 3: Fault diagnosis:

[0010] The enhanced feature signal is input into a one-dimensional convolutional neural network model to achieve fault classification.

[0011] Preferably, the specific steps of Step 1 are as follows:

[0012] Step 1-1: Perform a short-time Fourier transform on the original vibration signal to obtain a time-frequency spectrum;

[0013] Step 1-2: Calculate the marginal spectrum based on the time-frequency spectrum;

[0014] Step 1-3: Extract the extreme points from the marginal spectrum;

[0015] Step 1-4: Determine the filtering boundary by combining the maximum value and the adjacent minimum value in the extreme points;

[0016] Step 1-5: Construct a time-frequency filter and apply it to the time-frequency spectrum obtained in Step 1-1 to obtain a filtering result;

[0017] Step 1-6: Perform a summation operation on the filtering result in the frequency dimension to obtain a decomposition mode;

[0018] Step 1-7: Repeat Steps 1-4 to 1-6 until all decomposition modes are obtained.

[0019] Preferably, the formula for the short-time Fourier transform is:

[0020]

[0021] where is the experimental signal, represents the window function of the short-time Fourier transform, represents the imaginary symbol, represents the frequency in the Fourier transform, represents the time in the time-domain waveform, represents the result of the short-time Fourier transform, represents the time on the time-frequency plane.

[0022] Preferably, the formula for the marginal spectrum is:

[0023] .

[0024] where, represents the marginal spectrum.

[0025] Preferably, the specific steps of Step 1-5 are as follows:

[0026] The time-frequency filter acts on the short-time Fourier transform result of the original vibration signal to obtain the time-frequency filtering result:

[0027]

[0028] In the formula, represents the time-frequency filter, represents the element-by-element multiplication of the matrix at the corresponding position.

[0029] Preferably, the specific steps of step 1-6 are as follows:

[0030] Perform an addition operation in the frequency dimension, and the calculation formula for the decomposition mode is:

[0031]

[0032] In the formula is the decomposition mode.

[0033] Preferably, the specific step of step 2 is as follows:

[0034] Step 2-1: Calculate the envelope spectrum of each decomposition mode;

[0035] Step 2-2: Extract the amplitude center data value at the target fault frequency as the evaluation index;

[0036] Step 2-3: Select the mode with the maximum evaluation index value as the final output result, that is, the enhanced feature signal.

[0037] Preferably, the specific step of step 2-1 is as follows:

[0038] First, perform the Hilbert transform:

[0039]

[0040] Among them is the pi, is the Hilbert transform;

[0041] Calculate the envelope signal by the Hilbert transform method The formula for is:

[0042]

[0043] In the formula represents calculating the modulus length;

[0044] Finally, obtain the envelope spectrum The formula for is:

[0045]

[0046] In the formula represents the Fourier transform.

[0047] Preferably, the calculation formula for the amplitude center data value at the target fault frequency is:

[0048]

[0049] where is the amplitude corresponding to the fault frequency in each mode envelope spectrum, and μ is the envelope spectrum mean.

[0050] Preferably, the one-dimensional convolutional neural network model in step 3 includes:

[0051] At least one convolutional layer and pooling layer for extracting local features at different scales;

[0052] At least one fully connected layer for feature integration and representation;

[0053] Dropout mechanism to suppress overfitting and improve generalization ability;

[0054] The output layer uses the Softmax function for multi-class fault probability prediction.

[0055] The beneficial effects of the present invention are as follows:

[0056] The present invention combines a physics-driven feature enhancement strategy with a data-driven deep learning model, effectively improving the diagnostic accuracy and stability of the fault diagnosis system under low signal-to-noise ratio, complex background, and multiple fault modes, and has good engineering popularization and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is the flowchart of the method of the present invention;

[0058] Figure 2 is the schematic diagram of the signal with fault information in the embodiment of the present invention;

[0059] Figure 3 is the schematic diagram of the short-time Fourier transform of the fault signal in the embodiment of the present invention;

[0060] Figure 4 is the schematic diagram of the frequency marginal spectrum of the fault signal in the embodiment of the present invention;

[0061] Figure 5 is the schematic diagram of the center data value of 10 modes in the embodiment of the present invention;

[0062] Figure 6 is the schematic diagram of the finally extracted envelope spectrum in the embodiment of the present invention;

[0063] Figure 7It is a schematic diagram of the confusion matrix result of the verification set in the embodiment of the present invention. Detailed implementation manners

[0064] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0065] The present invention aims to provide a bearing fault diagnosis method based on physical information embedding for fault feature enhancement and one-dimensional convolutional neural network, so as to solve the problems in the prior art such as insufficient extraction of fault signal features, low recognition accuracy, and poor model robustness under complex working conditions. The present invention effectively enhances the recognizability of key feature components by introducing physical information closely related to the fault mechanism in the signal processing stage, and combines a convolutional neural network with deep nonlinear modeling ability to realize automatic fault recognition with high accuracy and strong robustness. To achieve the above object, the present invention proposes the following technical solutions:

[0066] (1) Signal decomposition based on the extreme value distribution of the marginal spectrum:

[0067] Extract the local components in the signal that are highly correlated with the fault physical characteristics through time-frequency analysis methods. The specific steps include:

[0068] Perform short-time Fourier transform on the original vibration signal to obtain the time-frequency spectrum;

[0069] Calculate the marginal spectrum based on the time-frequency spectrum;

[0070] Extract the extreme points from the marginal spectrum to determine the central features of the candidate frequency bands;

[0071] Combine the maximum value and the adjacent minimum value to determine the filtering boundary;

[0072] Construct the corresponding time-frequency filter and apply it to the transformation result;

[0073] Perform summation operation in the frequency dimension to form a set of decomposition modes with physical interpretability;

[0074] Repeat the above process until all decomposition modes are obtained.

[0075] (2) Mode screening:

[0076] Screen the key modes that can best represent the fault features from all the decomposition modes, specifically including:

[0077] Calculate the envelope spectrum of each mode;

[0078] Extract the amplitude center data value at the target fault frequency as the evaluation index;

[0079] Select the mode with the maximum evaluation index value as the final output result, that is, the enhanced feature signal.

[0080] (3)Fault diagnosis module:

[0081] The enhanced feature signal is input into a one-dimensional convolutional neural network model for fault classification. The model includes:

[0082] At least one convolutional layer and pooling layer for extracting local features of different scales;

[0083] At least one fully connected layer for feature integration and representation;

[0084] Dropout mechanism to suppress overfitting and improve generalization ability;

[0085] The output layer uses the Softmax function to predict the probabilities of multi-class faults.

[0086] Embodiment:

[0087] Figure 1 The flowchart of the fault feature enhancement and one-dimensional convolutional neural network bearing fault diagnosis method based on physical information embedding is shown. In this embodiment, the method is implemented on the Matlab software platform. Based on the experimental signals, the method is demonstrated and explained.

[0088] Figure 2 Is the experimental signal with faults, Figure 3 Is the short-time Fourier transform of the fault signal, Figure 4 Is the frequency marginal spectrum of the fault signal, Figure 5 Is the central data value of 10 patterns, Figure 6 Is the envelope spectrum finally extracted by pattern screening, Figure 7 Is the confusion matrix result of the validation set of the fault diagnosis module.

[0089] Based on this, in this embodiment, the method includes the following steps:

[0090] (1)Signal decomposition based on the extreme value distribution of the marginal spectrum:

[0091] First step, calculate the short-time Fourier transform of the signal. The formula for the short-time Fourier transform is:

[0092]

[0093] The window size of the short-time Fourier transform is set to 40, and the sampling frequency is set to 25600.

[0094] Second step, calculate the marginal spectrum. The marginal spectrum The formula for is:

[0095]

[0096] Third step, calculate the extreme values of the marginal spectrum;

[0097] Step 4: Calculate the coordinates of the minimum values on both sides of the selected maximum value on the marginal spectrum; determine the filtering boundary;

[0098] Step 5: Construct a time-frequency filter , and the time-frequency filter acts on the short-time Fourier transform to obtain the time-frequency filtering result;

[0099]

[0100] Step 6: Sum in the frequency dimension to obtain the decomposition mode. The calculation formula is:

[0101]

[0102] Repeat Step 4 to Step 6 until all decomposition modes are obtained.

[0103] (2) Mode screening:

[0104] Step 1: Calculate the envelope spectrum of each mode. The formula for the Hilbert transform is:

[0105]

[0106] where . The formula for calculating the envelope signal using the Hilbert transform method is:

[0107]

[0108] Finally, the formula for the envelope spectrum is:

[0109]

[0110] The calculated mode envelope spectrum is as shown in Figure 6 .

[0111] Step 2: Calculate the evaluation parameter of each mode, that is, the central data value of the amplitude at the fault frequency , and the calculation formula is:

[0112]

[0113] In the formula is the amplitude corresponding to the fault frequency in the envelope spectrum of each mode, is the mean value of the envelope spectrum.

[0114] Step 3: Calculate the envelope spectrum of the corresponding mode with the largest central data value as the output result of the mode screening module, that is, the enhanced fault feature, also as shown in Figure 6 .

[0115] (3) Fault diagnosis module, that is, fault diagnosis is implemented based on a one-dimensional convolutional neural network. The confusion matrix of the validation set is as Figure 7 shown. Among them, 0 represents "healthy state", 1 represents "inner raceway fault", and 2 represents "outer raceway fault".

Claims

1. An artificial intelligence diagnosis method for bearing faults based on the embedding of physical information, characterized in that, It includes the following steps: Step 1: Signal decomposition based on the extreme value distribution of the marginal spectrum; Extract the local components associated with the physical characteristics of the fault from the signal through time-frequency analysis methods; Step 2: Mode screening: Screen the key mode that can best characterize the fault characteristics from all decomposition modes; Step 3: Fault diagnosis: The enhanced feature signal is input into a one-dimensional convolutional neural network model to achieve fault classification.

2. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 1, wherein The specific content of the said Step 1 is as follows: Step 1-1: Perform short-time Fourier transform on the original vibration signal to obtain the time-frequency spectrum; Step 1-2: Calculate the marginal spectrum based on the time-frequency spectrum; Step 1-3: Extract the extreme points from the marginal spectrum; Step 1-4: Determine the filtering boundary by combining the maximum value and the adjacent minimum value in the extreme points; Step 1-5: Construct a time-frequency filter and apply it to the time-frequency spectrum obtained in Step 1-1 to obtain the filtering result; Step 1-6: Perform summation operation on the filtering result in the frequency dimension to obtain the decomposition mode; Step 1-7: Repeat Step 1-4 to Step 1-6 until all decomposition modes are obtained.

3. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 2, wherein, The formula for the said short-time Fourier transform is: ； where is the experimental signal, represents the window function of the short-time Fourier transform, represents the imaginary symbol, represents the frequency in the Fourier transform, represents the time in the time-domain waveform, represents the result of the short-time Fourier transform, represents the time on the time-frequency plane.

4. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 3, characterized in that, The formula for the said marginal spectrum is: ； Among them, represents the marginal spectrum.

5. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 4, wherein, The specific content of the said Step 1-5 is as follows: The time-frequency filter acts on the short-time Fourier transform result of the original vibration signal to obtain the time-frequency filtering result: ； In the formula, represents a time-frequency filter, represents the element-by-element multiplication of the corresponding positions of the matrix.

6. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 5, characterized in that The specific content of the said Step 1-6 is as follows: The formula for obtaining the decomposition mode by performing summation operation in the frequency dimension is: ； wherein is the decomposition mode.

7. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 6, wherein The specific content of the said Step 2 is as follows: Step 2-1: Calculate the envelope spectrum of each decomposition mode; Step 2-2: Extract the amplitude center data value at the target fault frequency as the evaluation index; Step 2-3: Select the mode with the maximum evaluation index value as the final output result, that is, the enhanced feature signal.

8. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 7, characterized in that The specific content of the said Step 2-1 is as follows: First perform Hilbert transform: ； wherein is the ratio of a circle's circumference to its diameter; is the Hilbert transform; Calculating the envelope signal by Hilbert transform method The formula is as follows: ； wherein represents calculating the modulus length; Finally, the envelope spectrum is obtained The formula is as follows: ； wherein denotes Fourier transform.

9. The artificial intelligence diagnosis method for bearing faults based on physical information embedding according to claim 8, characterized in that The formula for the amplitude center data value at the target fault frequency is: ； where is the amplitude corresponding to the fault frequency in each mode envelope spectrum, is the mean value of the envelope spectrum.

10. A method for artificial intelligence diagnosis of bearing faults based on physical information embedding according to claim 9, characterized in that, The one-dimensional convolutional neural network model in the said Step 3 includes: At least one convolutional layer and pooling layer for extracting local features of different scales; At least one fully connected layer for feature integration and representation; Dropout mechanism to inhibit overfitting and improve generalization ability; The output layer uses the Softmax function to predict the multi-class fault probability.

Images