Bearing fault diagnosis system based on deep random neural network integration

Abstract

The invention discloses a bearing fault diagnosis system based on deep random neural network integration in the technical field of bearing detection. The bearing fault diagnosis system comprises the following steps: 1, extracting wavelet packet energy spectrum features; 2, initializing a COOT algorithm, and calculating parameters; 3, constructing a DRNNE model, and calculating a fitness value; 4, the global optimal position and the optimal fitness value of the population are stored; and 5, judging whether the fitness value meets an optimization target or reaches the maximum number of iterations of the algorithm. If yes, determining the optimal parameter, and if not, returning to the step 2 for re-determination; and 6, determining optimal parameters n, p, m and q, and outputting a fault diagnosis result. The diagnosis system is based on a BLS learning framework, multiple times of feature enhancement are carried out on mapping nodes, the generalization performance of a fault diagnosis model is improved in combination with an integrated learning strategy, a COOT algorithm is adopted to optimize DRNNE network structure parameters, it is ensured that the model has high feature learning ability, and meanwhile the generalization ability of the model is further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing detection, and in particular relates to a bearing fault diagnosis system based on deep random neural network integration. Background Art

[0002] Bearings are an important component of rotating machinery and are core components of electromechanical equipment in various industries such as aviation, aerospace, and industrial production. In actual working processes, the vibration signal of bearing faults changes continuously with the load, and the vibration signal of bearing faults is weak and easily covered by strong interference signals. Therefore, extracting fault features from bearing vibration signals under variable loads and strong noise environments, and making timely and effective fault diagnosis can effectively avoid the continued deterioration of faults.

[0003] In recent years, artificial intelligence technologies such as machine learning (ML), deep learning (DL), and large language models (LLM) have developed rapidly, and data-driven bearing fault diagnosis methods have gradually replaced fault diagnosis methods based on mathematical models and fault diagnosis methods based on empirical knowledge. Among them, the technical process of data-driven bearing fault diagnosis methods is divided into signal acquisition, feature extraction, and fault classification.

[0004] Although the deep learning model can effectively improve the accuracy of bearing fault diagnosis, the error back propagation algorithm based on gradient descent seriously affects the computational efficiency of the model and is not suitable for real-time analysis of industrial data. Summary of the invention

[0005] The present invention intends to provide a bearing fault diagnosis system based on deep random neural network integration to solve the problem that the existing bearing fault detection system cannot meet industrial needs.

[0006] The bearing fault diagnosis system based on deep random neural network integration in this scheme includes the following steps:

[0007] Step 1: Data preprocessing - wavelet packet energy spectrum feature extraction;

[0008] Step 2: Initialize the COOT algorithm and calculate the parameters, n: the number of feature mappings, p: each mapping is a p-dimensional feature, m: the number of basic learners, q: the dimension of basic learner feature enhancement;

[0009] Step 3: Build the DRNNE model and use the above parameters to calculate the fitness value.

[0010] The algorithm flow is as follows:

[0011] (1) Feature Mapping Stage

[0012] Calculate the input feature X for feature mapping.

[0013] (2) Feature enhancement stage:

[0014] Perform m feature enhancements on the mapping feature nodes to build a basic learner. The specific process is as follows:

[0015] The mapping feature F is calculated by the following formula n Perform m times of feature enhancement.

[0016] H k =φ(F n W hk +β hk ),k=1

[0017] H k =φ(H k-1 W hk +β hk ),k=2,3,…,m

[0018] (3) Output weight calculation:

[0019] The mapping features and the basic learner enhancement features are combined using the formula, and the output weight β of the DRNNE model is calculated.

[0020] (4) Calculation model output:

[0021] The output result Y of each layer is calculated by the following formula k , and adopt voting ensemble learning method to fuse the results and obtain the final result Y.

[0022] Y k =A k β k

[0023] Step 4: Save the global optimal position and best fitness value of the population.

[0024] Step 5: Determine whether the fitness value meets the optimization goal or reaches the maximum number of iterations of the algorithm. If it does, determine the optimal parameters. If it does not, return to step 2 and determine again.

[0025] Step 6: Determine the optimal parameters n, p, m and q, and output the fault diagnosis results.

[0026] The working principle and beneficial effects of this scheme are as follows: First, the wavelet packet energy spectrum is used to extract high-dimensional bearing fault features; second, in order to improve the generalization performance of BLS, a deep random neural network integration model is designed, and the COOT algorithm is used to select the DRNNE network parameters; finally, the experimental results verify the effectiveness of the DRNNE model. Compared with BLS, DRNNE further improves the accuracy of bearing fault diagnosis;

[0027] Based on the BLS learning framework, the diagnostic system performs multiple feature enhancements on the mapping nodes and combines the integrated learning strategy to improve the generalization performance of the fault diagnosis model. At the same time, the Coot Optimization Algorithm (COOT) is used to optimize the DRNNE network structure parameters to ensure that the model has strong feature learning capabilities and further improve the model generalization ability.

[0028] Furthermore, the wavelet packet energy spectrum feature extraction in step 1 uses the 48Khz signal data collected by the drive end, and for the bearing signal data in various working states, the original signal is decomposed by wavelet packet energy spectrum, and the original signal data is decomposed into multiple different frequency bands, with the frequency resolution of the signal high frequency band and the time resolution of the low frequency band. The wavelet packet energy spectrum can effectively improve the bearing fault characteristics.

[0029] Furthermore, in the step 2, the COOT algorithm determines the positions of the leader and the common individuals in the calculation process, and updates the positions of the leader and the common individuals once for each calculation. The COOT algorithm is beneficial to improving the parameter search efficiency of DRNNE and reducing the number of searches.

[0030] Furthermore, the calculation formula in the algorithm flow (1) in step 3 is:

[0031] F i =φ(XW ei +b ei ),i=1,2,…,n,

[0032] And use F n =[F1,F2,…,F n ],

[0033] Perform feature combination on n results, where φ represents the feature mapping activation function and W ei and b ei are random weights and biases, respectively.

[0034] Furthermore, in the algorithm flow (2) in step 3, φ represents the activation function, W hk and b hk The points represent random weights and biases.

[0035] Furthermore, the formula used in the algorithm flow (3) in step three is

[0036] A k =[F n |H k ]

[0037] and

[0038] Among them, C is the regularization parameter. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is the structure diagram of the width learning system;

[0040] Figure 2 This is a schematic diagram of the structure of the CWRU test bench;

[0041] Figure 3 is the original vibration signal of the bearing;

[0042] Figure 4 is the wavelet packet energy spectrum;

[0043] Figure 5 Deep random neural network ensemble model DRNNE

[0044] Figure 6 It is the bearing fault diagnosis process based on COOT-DRNNE;

[0045] Figure 7 This is a graph showing the experimental comparison results. DETAILED DESCRIPTION

[0046] The following is further described in detail through specific implementation methods:

[0047] Bearing fault diagnosis system based on deep random neural network ensemble:

[0048] 1. Related Theories

[0049] 1.1BLS

[0050] BLS is a randomized neural network model, and its network structure is as follows: Figure 1 As shown in the figure, it evolved from Random Vector Functional Link Neural Networks (RVFL). Different from CNN and other neural network models based on back propagation algorithms, BLS network parameters use randomization to assign node input parameters and calculate output weights through ridge regression method, which speeds up learning efficiency.

[0051] The BLS algorithm process can be simplified as follows:

[0052] Assume X = {x1, x2…, x N} is the input feature data, Y={y1,y2,…,y N} is the data label, and N represents the number of samples.

[0053] 1.1.1 Feature mapping stage:

[0054] The input feature X is mapped using formula (1), where n represents the number of feature mappings, and each mapping is a p-dimensional feature. The n-times mapping results are combined using formula (2).

[0055] F i =φ(XW ei +b ei ), i=1,2,…,n (1)

[0056] F n =[F1,F2,…,F n ] (2)

[0057] Among them, φ represents the feature map activation function, W ei and b ei are random weights and biases, respectively.

[0058] 1.1.2 Feature enhancement stage:

[0059] The mapping feature F is transformed into n Perform feature enhancement, where m represents the number of feature enhancements, q represents the dimension of each feature enhancement, and use formula (4) to perform feature combination.

[0060] H j =φ(Z n W hj +β hj ), j=1,2,…,m (3)

[0061] H m =[H1,H2,…,H m ] (4)

[0062] Among them, φ represents the activation function, W hj and b hj The points represent random weights and biases.

[0063] 1.1.3 Output weight calculation:

[0064] Formula (5) is used to map the features of formula (2) and enhance the features of formula (4), and the output weight β of the BLS model is obtained according to the ridge regression calculation method of formula (6).

[0065] A=[F n |H m ] (5)

[0066] β=(A T A+CI) -1 A T Y (6)

[0067] Among them, C is the regularization parameter.

[0068] 1.1.4 Calculation model output:

[0069] The BLS output result Y is calculated by formula (7).

[0070] Y=Aβ (7)

[0071] 1.2COOT Algorithm

[0072] The Coot Optimization Algorithm (COOT), proposed in 2021, is a new swarm intelligence optimization algorithm inspired by the foraging behavior of coots. The movement of their group on the water surface is determined by food, mainly including random movement, chain movement, movement according to the leader's position, and leader movement.

[0073] The COOT algorithm process can be simplified as follows:

[0074] 1.2.1 Parameter initialization:

[0075] Set the population size (N G ), the number of common individuals (N c ), the number of leaders (N L ), maximum number of iterations (T), search space dimension (D), search range upper bound (ub), and lower bound (lb).

[0076] Initialize the positions of ordinary individuals and leaders according to equations (8) and (9).

[0077] C(a)=lb+rand(1,D)×(ub-lb) (8)

[0078] L(b)=lb+rand(1,D)×(ub-lb) (9)

[0079] Where C(a) is the initial position of the common individual, a=1,2,…,N c . L(b) represents the initial position of the leader, b=1,2,…,N L .

[0080] 1.2.2 Random Move:

[0081] According to formula (10), the position of ordinary individuals is updated, and each part of the exploration area is explored by random movement to obtain the local optimum.

[0082] C(a)=C(a)+A×rand×(QC(a)) (10)

[0083] Among them, Q is calculated by formula (11) to realize the random movement of ordinary individuals, t represents the current iteration number, and T represents the maximum iteration number.

[0084] Q=lb+rand(1,D)×(ub-lb) (11)

[0085]

[0086] 1.2.3 Chain Move:

[0087] In order to realize the chain movement of common individuals, the average position of two common individuals is calculated according to formula (13).

[0088] C(a)=0.5×(C(a)+C(a-1)) (13)

[0089] 1.2.4 Move according to the leader position:

[0090] Ordinary individuals use formula (14) to move their positions according to the leader.

[0091] C(a)=L(k)+2×rand×cos(2Rπ)×(L(k)-C(a)) (14)

[0092] k=1+(a%N L ) (15)

[0093] In the formula, k represents the leader number, which is determined by formula (15), a is the current common individual number, and R is a random value in the range of [-1,1]. Leader movement.

[0094] 1.2.5 Leader Movement:

[0095] Since the population position is determined by the leader, in order to achieve global optimization, equation (16) is applied to update the leader's position.

[0096]

[0097] Among them, G represents the current global optimal position, R is a random value in the range of [-1,1], and r is a random value in the range of [0,1].

[0098] 2. DRNNE Fault Diagnosis Model

[0099] 2.1 Data Preprocessing

[0100] This paper uses the bearing fault dataset of Case Western Reserve University (CWRU) as the experimental dataset. This dataset is one of the most authoritative public datasets in the field of bearing fault diagnosis research. Figure 2 As shown in the figure, the following are shown: fan end bearing 1, motor 2, drive end bearing 3, torque sensor + encoder 4, power meter 5.

[0101] The experimental bench adopts the single-point damage method of artificial electrical discharge machining to manufacture three different fault sizes (0.18mm, 0.36mm, 0.54mm) and a normal state on the outer ring, inner ring and rolling element of the rolling bearing, as shown in Table 1.

[0102] Table 1

[0103]

[0104] The system uses the 48Khz signal data collected by the drive end. For the bearing signal data in various working states, the original signal is decomposed by wavelet packet energy spectrum, and the original signal data is decomposed into multiple different frequency bands, with the frequency resolution of the high-frequency band of the signal and the time resolution of the low-frequency band. This section uses 8 layers of wavelet packet decomposition to extract 256-dimensional characteristic information that reflects the bearing fault. The original signal waveform and wavelet packet energy spectrum are shown in Figure 1. Figure 3 and Figure 4 shown.

[0105] in Figure 3 The horizontal axis is the sampling point / n, and the vertical axis is the amplitude.

[0106] Figure 4 The horizontal axis is the frequency band number, and the vertical axis is the normalized energy amplitude.

[0107] 2.2DRNNE model

[0108] Ensemble learning can effectively improve the generalization performance of learning models. In order to ensure the efficiency of fault diagnosis learning and improve the generalization performance of the model, this paper draws on the BLS feature mapping method and proposes a deep random neural network ensemble model DRNNE. DRNNE can enhance the features of the mapping feature nodes of the node layer m times, build m basic learners for ensemble learning, and thus improve the generalization performance of the fault diagnosis model. The DRNNE model is as follows Figure 5 shown.

[0109] The DRNNE algorithm process is as follows:

[0110] (1) Feature mapping stage:

[0111] DRNNE uses the same feature mapping process as BLS. It maps the input feature X using formula (1) and combines the mapping results using formula (2).

[0112] (2) Feature enhancement stage:

[0113] DRNNE performs m feature enhancements on the mapping feature nodes to build a basic learner. The specific process is as follows:

[0114] Through equations (17) and (18), the mapping feature F n Perform m feature enhancements, where m represents the number of base learners and q represents the feature enhancement dimension of the base learners.

[0115] H k =φ(F n W hk +β hk ),k=1 (17)

[0116] H k =φ(H k-1 W hk +β hk ), k=2,3,…,m (18)

[0117] Among them, φ represents the activation function, W hk and b hk The points represent random weights and biases.

[0118] (3) Output weight calculation:

[0119] The mapping features of formula (2) and the enhanced features of the basic learner of formula (18) are combined using formula (19), and the output weight β of the DRNNE model is obtained according to the ridge regression calculation method of formula (6).

[0120] A k =[F n |H k ] (19)

[0121] β k =(A k T A k +CI) -1 A k T Y (20)

[0122] Among them, C is the regularization parameter.

[0123] (4) Calculation model output:

[0124] The output result Y of each layer is calculated by formula (21): k, and adopt voting ensemble learning method to fuse the results and obtain the final result Y.

[0125] Y k =A k β k (twenty one)

[0126] 2.3 Bearing fault diagnosis process based on COOT-DRNNE

[0127] Compared with traditional BLS, the deep random neural network integration model uses multiple feature enhancement results for integrated learning, which is beneficial to improve the generalization performance of the model. However, DRNNE is similar to BLS. The feature mapping is tuned by sparse autoencoders. Usually, for different training data, it is necessary to perform grid search on the number of basic learners m, the number of enhanced feature nodes q, the number of feature mappings n, and the feature mapping nodes p to achieve the optimal accuracy. However, if the search range is large, the grid search efficiency is low. Therefore, in order to further improve the generalization performance and accuracy of the model, this paper introduces COOT to search for four parameters, and uses the optimal parameters as the DRNNE model parameters for bearing fault diagnosis. The specific algorithm flow is as follows: Figure 6 .

[0128] 3. Experimental Results

[0129] To illustrate the effectiveness of the DRNNE model designed in this paper, the experiment selected BLS, SVM, DBN, and CNN as comparison algorithms, among which the penalty parameter of SVM is 1000 and the kernel parameter is 0.01; DBN uses two hidden layers, the number of hidden layer nodes is 100 and 50 respectively, the learning rate is set to 1, and the number of iterations is 1000; CNN uses the Lenet structure, the learning rate is 1, and the number of iterations is 1000. The network structure parameters of BLS and DRNNE are determined by COOT. The number of COOT populations is set to 50, the number of iterations is 20, the search range of the number of enhanced feature nodes q is 1-200, the search range of the number of feature mappings n and feature mapping nodes p is 1-50, the number of BLS feature enhancements m is 1 (the default setting is 1), and the number of DRNNE basic learners m is 2-10. The parameter search based on COOT only needs 1000 times to determine the optimal network structure.

[0130] Table 2 and Figure 7The test accuracy comparison results of the five algorithms are listed. DRNNE achieves the best results in all four evaluation indicators compared with BLS, DBN, CNN, and SVM, and the average fault diagnosis results are all above 95%. Compared with traditional BLS, DRNNE can further improve the fault diagnosis classification accuracy by using ensemble learning methods. In summary, the bearing fault model based on MBLS proposed in this paper has a high classification accuracy for bearing fault diagnosis, the wavelet packet energy spectrum can effectively improve the bearing fault characteristics, and COOT is conducive to improving the parameter search efficiency of DRNNE and reducing the number of searches.

[0131] Table 2

[0132]

[0133] IV. Conclusion

[0134] In order to improve the accuracy of bearing fault diagnosis, this paper proposes a bearing fault diagnosis model based on DRNNE. Firstly, the wavelet packet energy spectrum is used to extract high-dimensional bearing fault features; secondly, in order to improve the generalization performance of BLS, a deep random neural network integration model is designed, and the COOT algorithm is used to select the DRNNE network parameters; finally, the experimental results verify the effectiveness of the DRNNE model. Compared with BLS, DRNNE further improves the accuracy of bearing fault diagnosis.

[0135] The above is only an embodiment of the present invention, and the common knowledge such as the known specific structure and characteristics in the scheme is not described in detail here. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. Bearing fault diagnosis system based on deep random neural network integration, characterized by: The following steps are involved: Step 1: Data preprocessing - wavelet packet energy spectrum feature extraction; Step 2: Initialize the COOT algorithm and calculate the parameters, n: the number of feature mappings, p: each mapping is a p-dimensional feature, m: the number of basic learners, q: the basic learner feature enhancement dimension; Step 3: Build the DRNNE model and use the above parameters to calculate the fitness value. The algorithm flow is as follows: (1) Feature Mapping Stage Calculate the input feature X for feature mapping. (2) Feature enhancement stage: Perform m feature enhancements on the mapping feature nodes to build a basic learner. The specific process is as follows: The mapping feature F is calculated by the following formula n Perform m times of feature enhancement. H k =φ(F n W hk +b hk ),k=1 H k =φ(H k-1 W hk +b hk ),k=2,3,…,m (3) Output weight calculation: The formula is used to construct the mapping features and the basic learner enhancement features for combination, and the output weight β of the DRNNE model is calculated. (4) Calculation model output: The output result Y of each layer is calculated by the following formula k , and adopt voting ensemble learning method to fuse the results and obtain the final result Y. Y k =A k b k Step 4: Save the global optimal position and best fitness value of the population. Step 5: Determine whether the fitness value meets the optimization goal or reaches the maximum number of iterations of the algorithm. If it does, determine the optimal parameters. If it does not, return to step 2 and determine again. Step 6: Determine the optimal parameters n, p, m and q, and output the fault diagnosis results.

2. The bearing fault diagnosis system based on deep random neural network integration according to claim 1 is characterized in that: The wavelet packet energy spectrum feature extraction in step 1 uses the 48Khz signal data collected by the driving end. For the bearing signal data in various working states, the original signal is decomposed by wavelet packet energy spectrum, and the original signal data is decomposed into multiple different frequency bands with the frequency resolution of the high-frequency band of the signal and the time resolution of the low-frequency band.

3. The bearing fault diagnosis system based on deep random neural network integration according to claim 2 is characterized in that: In the step 2, the COOT algorithm determines the positions of the leader and the common individuals in the calculation process, and updates the positions of the leader and the common individuals once for each calculation.

4. The bearing fault diagnosis system based on deep random neural network integration according to claim 3 is characterized in that: The calculation formula in the algorithm flow (1) in step 3 is: F i =φ(XW ei +b ei ),i=1,2,...,n, And use F n =[F1,F2,…,F n ], Perform feature combination on n results, where φ represents the feature mapping activation function and W ei and b ei are random weights and biases, respectively.

5. The bearing fault diagnosis system based on deep random neural network integration according to claim 4 is characterized in that: In the algorithm flow (2) of step 3, φ represents the activation function, W hk and b hk The points represent random weights and biases.

6. The bearing fault diagnosis system based on deep random neural network integration according to claim 4 is characterized in that: The formula used in the algorithm flow (3) in step 3 is: A k =[F n |H k ] and β k = (A k T A k + CI) -1 A k T Y, Among them, C is the regularization parameter.

Images