Rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network

Abstract

The present invention proposes a rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network, including: collecting vibration signals of rolling bearings under normal operation and different fault states, obtaining data sets of vibration signals under different working conditions; performing fully adaptive noise set empirical mode decomposition on the vibration signals, taking envelope kurtosis as parameter index, obtaining new recombined signals; performing variational mode decomposition on the recombined signals, taking local maximum envelope spectrum kurtosis as parameter index, obtaining the final signal; extracting the time-frequency domain characteristic entropy values ​​of the final signal to form a state characteristic vector matrix of different fault states, and inputting it into the long short-term memory network for training to obtain a training model; collecting the vibration signals of the test bearings, performing the same processing, and then inputting them into the model for fault diagnosis. The present invention effectively solves the problems of poor vibration signal diagnosis effect under strong noise interference and slow operation of large sample size.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical system fault diagnosis and signal processing, and in particular, relates to a rolling bearing fault diagnosis method under strong noise interference based on multi-modal decomposition and long short-term memory network. Background Art

[0002] In the context of modern mass production, rolling bearings are one of the essential components of many mechanical equipment. As key components of rotating machinery such as aerospace engines, high-precision CNC machine tools, and wind power generation equipment, they play a vital role. As the use time of bearings increases, bearing components will experience various forms of damage, such as fatigue spalling, plastic deformation, cracks and fractures, pitting, etc. These conditions make bearings one of the most fragile components in machinery and equipment. Once the bearing is damaged, the performance of the machine will be significantly reduced and the life will be shortened, and even major accidents will occur, causing significant economic losses and endangering personal safety. Therefore, it is of great significance to diagnose the faults of rolling bearings.

[0003] Traditional rolling bearing vibration signal detection is mainly based on time-frequency domain analysis, that is, the vibration signal collected by the sensor is plotted in the time domain and various frequency domain diagrams, the characteristic frequency of the bearing fault under the corresponding conditions is calculated, and the corresponding base frequency and frequency multiplication signals are found in the diagram, and compared with the calculated fault characteristic frequency to determine the fault type. However, under the interference of strong noise, the amplitude of the fault frequency hidden in the noise cannot be separated by simple time-frequency domain analysis. Therefore, for the collected vibration signal, noise reduction processing for various noises of different natures must be performed first, and then the time-frequency domain analysis method is used to determine the fault type. Compared with manual judgment, the currently popular fault diagnosis method based on deep learning has the significant advantages of strong feature extraction ability and high fault identification efficiency. Summary of the invention

[0004] The present invention proposes a rolling bearing fault diagnosis method under strong noise interference based on multi-modal decomposition and long short-term memory network. For vibration signal data of different fault types under strong noise interference, the vibration signal is subjected to fully adaptive noise set empirical mode decomposition, and the intrinsic mode function components obtained by decomposition are reorganized according to the size of envelope kurtosis. The reorganized signal is then subjected to variational mode decomposition, and the local maximum envelope spectrum kurtosis is used as the screening standard to obtain the appropriate intrinsic mode function components, that is, the pure vibration signal with stationary noise and impact noise removed. Then, the singular spectrum entropy, approximate entropy, sample entropy, fuzzy entropy, envelope entropy and multi-scale walk entropy of the vibration signal are extracted to form a state feature vector matrix representing the bearing state. Then, the sample set is trained using the long short-term memory network to achieve accurate identification of different fault states.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] (a) collecting vibration signals of rolling bearings under normal operation and different fault conditions, and classifying the obtained vibration signals to obtain an initial data set of vibration signals of the bearings under normal operation and different fault conditions;

[0007] (b) performing fully adaptive noise ensemble empirical mode decomposition on the vibration signal to obtain multiple intrinsic mode function components, and then using envelope kurtosis as a parameter index to screen out the intrinsic mode function components that are greater than the envelope kurtosis of the original vibration signal, and recombining them into a new vibration signal;

[0008] (c) performing variational mode decomposition on the reorganized vibration signal to obtain multiple new intrinsic mode function components, and then using the local maximum envelope spectrum kurtosis as a parameter index to screen out the intrinsic mode function component with the maximum envelope spectrum kurtosis as the final vibration signal;

[0009] (d) extracting the time-frequency domain characteristic entropy values ​​of the processed vibration signal, namely, singular spectrum entropy, approximate entropy, sample entropy, fuzzy entropy, envelope entropy and multi-scale spread entropy, and generating state characteristic vector matrices of different fault states;

[0010] (e) constructing a rolling bearing long short-term memory network model, and using the state feature vector matrix as the input of the neural network model, and the corresponding fault state label as the expected output of the rolling bearing long short-term memory network, and training the rolling bearing long short-term memory network;

[0011] (f) Collect the vibration signal of the current rolling bearing, and construct its state feature vector matrix using the same method as in steps (b), (c), and (d); and input the matrix into the long short-term memory network model of the rolling bearing trained in step (e) to obtain the fault diagnosis result of the current rolling bearing.

[0012] As a preferred technical solution, in step (a), before collecting the vibration signal data of each fault state, first collect the vibration signal data of the bearing under normal operation, and then collect the bearing fault data according to the three fault states of inner ring, rolling element and outer ring fault in turn; after collecting the normal data and the three types of fault data, pre-process the data, including: preliminary denoising of the vibration signal; retaining four decimal places for the vibration signal;

[0013] As a preferred technical solution, in step (b), a fully adaptive noise set empirical mode decomposition is performed on the vibration signal, and when selecting the best intrinsic mode function from the intrinsic mode function components obtained after the decomposition, the envelope kurtosis of each intrinsic mode function component and the original vibration signal is calculated and judged, and the intrinsic mode function components that meet the requirements are obtained, and they are recombined into a new vibration signal, and the vibration signal removes the interference of stationary noise compared to the original vibration signal;

[0014] As a preferred technical solution, in step (c), the recombined vibration signal is subjected to variational mode decomposition to obtain multiple new intrinsic mode function components, and then the intrinsic mode function component with the local maximum envelope spectrum kurtosis is screened out with the local maximum envelope spectrum kurtosis as the parameter index, and is used as the final vibration signal. Compared with the original signal, the vibration signal removes the interference of stationary noise and impact noise, and is a pure vibration signal;

[0015] As a preferred technical solution, in step (d), the comprehensive time-frequency domain characteristic entropy value of the processed vibration signal is extracted, namely, singular spectrum entropy, approximate entropy, sample entropy, fuzzy entropy, envelope entropy and multi-scale spread entropy, to generate state characteristic vector matrices of different fault states;

[0016] As a preferred technical solution, in step (d), the constructed state feature vector matrix is ​​formed based on the entropy value combination of multiple time-frequency domain features, and is in the form of:

[0017] Among them, SVDE is the singular spectrum entropy value, AE is the approximate entropy value, SE is the sample entropy value, FE is the fuzzy entropy value, EE is the envelope entropy value, and ME is the multi-scale distribution entropy value.

[0018] As a preferred technical solution, in step (e), the long short-term memory network used is an improved recurrent neural network model, which includes an input layer, an output layer and several hidden layers. The hidden layer is composed of memory units, which are responsible for storing state information and transmitting it along the time axis.

[0019] By adopting the above technical solution, the beneficial effects of the present invention are as follows:

[0020] The present invention adopts two modal decomposition methods, namely fully adaptive noise set empirical mode decomposition and variational mode decomposition, to successfully remove stationary noise and impact noise from the vibration signal under strong noise interference, and extract a relatively pure vibration signal;

[0021] The present invention proposes to construct a rolling bearing state feature vector matrix by using singular spectrum entropy, approximate entropy, sample entropy, fuzzy entropy, envelope entropy and multi-scale spread entropy of different fault states. The state feature vector matrix contains information on different fault types of the bearing. By analyzing the state feature vector matrix, the difficulty of extracting fault features can be reduced, and the problem of inaccurate diagnosis when the fault is weak in the vibration signal under strong noise interference can be better solved.

[0022] The present invention greatly reduces the difficulty of data feature extraction by constructing a bearing state feature vector matrix, effectively improving the training efficiency of the long short-term memory network. At the same time, it can effectively solve the problems of missing original data sets and sample imbalance;

[0023] The rolling bearing long short-term memory network constructed by the present invention includes an input layer, an output layer and a plurality of hidden layers, wherein the hidden layers are composed of memory units; the rolling bearing long short-term memory network can learn long-term temporal dependencies, is less affected by gradient vanishing and gradient exploding than ordinary recurrent neural networks, and has better network performance.

[0024] The present invention collects vibration signal data of the normal operation and three fault states of the bearing respectively, extracts the time-frequency domain characteristic entropy value therefrom, and then adopts the long short-term memory network for learning, thereby ensuring that the method can predict the fault state of the bearing. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the embodiments described herein are only used to illustrate and explain the invention and are not used to limit the present invention.

[0026] Figure 1 is a flow chart of a fault diagnosis method constructed according to a preferred embodiment of the present invention;

[0027] Figure 2 is a time domain diagram of a rolling bearing inner race fault collected according to a preferred embodiment of the present invention;

[0028] Figure 3 It is a comparative histogram of the kurtosis of the intrinsic mode function envelope after fully adaptive noise set empirical mode decomposition according to the preferred embodiment of the present invention.

[0029] Figure 4 It is a time domain diagram of a signal reconstructed after fully adaptive noise set empirical mode decomposition according to a preferred embodiment of the present invention.

[0030] Figure 5 It is a comparative histogram of the kurtosis of the envelope spectrum of the intrinsic mode function components after variational mode decomposition according to the preferred embodiment of the present invention.

[0031] Figure 6 It is a time domain diagram of the final signal after variational mode decomposition according to the preferred embodiment of the present invention.

[0032] Figure 7 It is a schematic diagram of a rolling bearing long short-term memory network model constructed according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] The terms of this embodiment are explained, including: Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Variational Mode Decomposition (VMD), Intrinsic Mode Function (IMF), Long Short-Term Memory (LSTM).

[0035] like Figure 1 The flowchart of the rolling bearing fault diagnosis method under strong noise interference based on multi-modal decomposition and long short-term memory network constructed according to the preferred embodiment of the present invention is shown. The method includes the following steps:

[0036] Step 1: The vibration signals of the SKF-6203-2RS deep groove ball bearing under normal operation and different fault conditions are collected at a sampling frequency of 12000Hz. Each vibration signal is preprocessed, including denoising the vibration signal and rounding off the four decimal places after the decimal point of the vibration signal. The time domain diagram of the processed inner ring fault vibration signal is shown in the figure below. Figure 2 shown.

[0037] Step 2: Perform fully adaptive noise ensemble empirical mode decomposition on the vibration signal, calculate the envelope kurtosis of each intrinsic mode function component and the original signal, and compare the envelope kurtosis of each intrinsic mode function component with the original signal. The comparison histogram is as follows: Figure 3As shown in the figure, it can be seen that the envelope kurtosis values ​​of the first, second, third, fifth, sixth, ninth, tenth and fourteenth intrinsic mode function components are higher than the envelope kurtosis value of the original signal. Therefore, the above intrinsic mode function components are reorganized, and the time domain diagram of the reorganized vibration signal is as follows: Figure 4 shown.

[0038] Step 3: Perform variational mode decomposition on the reorganized vibration signal and calculate the envelope spectrum kurtosis of each intrinsic mode function component, such as Figure 5 As shown in the figure, the third eigenmode function component has the largest envelope spectrum kurtosis, so it is selected as the final vibration signal. The time domain diagram of the signal is shown in Figure 6 shown.

[0039] Step 4: Extract the time-frequency domain characteristic entropy values ​​of the final vibration signal after denoising, namely, singular spectrum entropy, approximate entropy, sample entropy, fuzzy entropy, envelope entropy and multi-scale spread entropy, and compose the entropy values ​​into a characteristic state vector matrix.

[0040] Step 5: Construct a rolling bearing long short-term memory network model, such as Figure 7 The state feature vector matrix is ​​used as the input of the rolling bearing long short-term memory network model, and the corresponding fault label is used as the expected output of the rolling bearing long short-term memory network to train the rolling bearing long short-term memory network.

[0041] Step 6: Collect the vibration signal of the current rolling bearing, and construct the state feature vector matrix under normal operation and different fault states through the state feature vector matrix construction method in the above steps 1, 2, 3, and 4. Input it into the long short-term memory network model trained in step 5 for fault diagnosis.

[0042] Furthermore, in step 2, the fully adaptive noise set empirical mode decomposition adds the intrinsic mode components containing auxiliary noise after empirical mode decomposition, rather than directly adding the Gaussian white noise signal to the original signal, and then performs an overall average calculation after obtaining the first-order intrinsic mode components to obtain the final first-order intrinsic mode components, and continues to calculate until a solution is obtained.

[0043] Furthermore, in step 3, in the variational mode decomposition, the intrinsic mode function is defined as a frequency modulation function with bandwidth limitation, and the original signal is decomposed into a specified number of intrinsic mode function components by constructing and solving a constrained variational problem.

[0044] Furthermore, in step 4, the data length used when extracting the entropy value is 100, so 8 time-frequency domain feature entropy values ​​can be extracted for every 100 vibration signals, which can be flexibly modified according to one's own requirements.

[0045] Furthermore, in step 5, the fault label is the expected output of the LSTM model. It is a one-dimensional vector consisting of two elements, 0 and 1, and the number of elements in the vector is the number of categories of the bearing operation status. Take the four categories of bearing operation status as an example, which are: normal operation, inner ring wear, rolling element wear, and outer ring wear. The labels are set as follows: (where the number of 0s in each fault label is 3) The fault label of the normal operation of the bearing is [1,0,0,0], the fault label of the inner ring wear is [0,1,0,0], the fault label caused by rolling element wear is [0,0,1,0], and the fault label caused by outer ring wear is [0,0,0,1].

[0046] Further, in step 6, the long short-term memory network model is shown in Figure 5 , the final output layer outputs a one-dimensional vector, the elements in the vector represent the probability of it being the fault type, and fault diagnosis and fault prediction are performed based on the vector. Taking the three types of bearing fault types mentioned above as an example, the final output one-dimensional vector is [0.05, 0.90, 0.05, 0]. Combined with the fault label in step 6, it can be seen that the vector represents the diagnosis result of this time: the probability of the bearing being in normal operation is 5%, the probability of inner ring wear is 90%, the probability of rolling element failure is 5%, and the probability of outer ring wear is 0%. Therefore, it can be finally judged that the fault type is the inner ring, which is consistent with the actual situation.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network, characterized in that: The following steps are involved: (a) collecting vibration signals of rolling bearings under normal operation and under different fault conditions with different fault degrees, and classifying the obtained vibration signals to obtain an initial data set including vibration signals under normal operation and under different fault conditions; (b) performing fully adaptive noise ensemble empirical mode decomposition on the vibration signal to obtain multiple intrinsic mode function components, using envelope kurtosis as a parameter index, screening out intrinsic mode function components with envelope kurtosis greater than the original vibration signal, and recombining them into a new vibration signal; (c) performing variational mode decomposition on the new vibration signal to obtain multiple new intrinsic mode function components, and then using the local maximum envelope spectrum kurtosis as a parameter index, screening out the intrinsic mode function component with the local maximum envelope spectrum kurtosis and using it as the final vibration signal; (d) extracting the comprehensive time-frequency domain characteristic entropy value of the final vibration signal to generate state characteristic vector matrices of different fault states, wherein the comprehensive time-frequency domain characteristic entropy value includes singular spectrum entropy, approximate entropy, sample entropy, fuzzy entropy, envelope entropy and multi-scale spread entropy; (e) constructing a rolling bearing long short-term memory network model, and using the state feature vector matrix as the input of the neural network model, and the corresponding fault state label as the expected output of the rolling bearing long short-term memory network, and training the rolling bearing long short-term memory network; (f) Collect the vibration signal of the rolling bearing to be tested, and construct its state feature vector matrix using the same method as in steps (b), (c), and (d); and input the matrix into the long short-term memory network model of the rolling bearing trained in step (e) to obtain the fault diagnosis result of the current rolling bearing.

2. The rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network as claimed in claim 1 is characterized in that: In step (a), after collecting the vibration signals of the rolling bearing in normal operation and under different fault conditions, the data is preprocessed, including: preliminary denoising of the vibration signal; retaining four decimal places for the vibration signal.

3. The rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network as claimed in claim 1, characterized in that: In step (b), the vibration signal is subjected to fully adaptive noise set empirical mode decomposition, and the envelope kurtosis value of each intrinsic mode function component and the original vibration signal is calculated and used as a judgment criterion to obtain the intrinsic mode function components that meet the requirements, and recombined them into a new vibration signal. Compared with the original vibration signal, the recombined signal removes the interference of stationary noise.

4. The rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network as claimed in claim 1, characterized in that: In step (c), the reorganized vibration signal is subjected to variational mode decomposition, the envelope spectrum kurtosis value of each intrinsic mode function component is calculated, and the intrinsic mode function component with the largest envelope spectrum kurtosis is selected as the final vibration signal. Compared with the original signal, the vibration signal has removed the interference of stationary noise and impact noise, and is a pure vibration signal.

5. The rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network as claimed in claim 1, characterized in that: In step (d), the constructed state eigenvector matrix is ​​composed of singular spectrum entropy, approximate entropy, sample entropy, fuzzy entropy, envelope entropy and multi-scale spread entropy extracted from the vibration signal.

6. The rolling bearing fault diagnosis method under strong noise interference based on multimodal decomposition and long short-term memory network according to claim 1, characterized in that: In step (e), the long short-term memory network used is an improved recurrent neural network model, which includes an input layer, an output layer and several hidden layers. The hidden layer is composed of memory units, and the memory units are responsible for storing state information and transmitting it along the time axis.

Images