A method for failure determination of a rolling bearing and related product

Abstract

The application relates to a fault judging method for a rolling bearing and a related product, the method comprising: acquiring vibration signals of the rolling bearing collected by multiple sensors; decomposing and reconstructing the vibration signals according to a maximum overlap discrete wavelet transform to obtain a reconstructed signal; determining a multi-scale scatter entropy corresponding to the reconstructed signal and forming a multi-dimensional state feature vector; and detecting the multi-dimensional state feature vector by using a trained neural network detection model to judge a fault type of the rolling bearing. According to the scheme, the problem that the fault detection of the rolling bearing is difficult and low in accuracy is solved.

Description

Technical Field

[0001] This invention generally relates to the field of fault detection technology. More specifically, this invention relates to a method, apparatus, and computer-readable storage medium for diagnosing faults in rolling bearings. Background Technology

[0002] With the normalization of epidemic prevention and control and the increasingly widespread application of mechanical disinfection equipment such as disinfection robots, the high precision and complexity of these devices lead to a high failure rate and difficulty in fault diagnosis. Rolling bearings, as components supporting key parts such as motor shafts, directly affect the performance of the entire unit. Statistics show that approximately 50% of mechanical equipment failures are caused by rolling bearings. Therefore, in the field of mechanical fault diagnosis, the condition detection and fault feature extraction of rolling bearings have always been a research hotspot and a challenge.

[0003] For rolling bearings, on the one hand, their working conditions are relatively complex and variable, and vibration signal acquisition is easily affected by environmental noise and acquisition and transmission equipment; on the other hand, due to the influence of dynamic loads, nonlinear contact forces and other factors, their vibration signals exhibit obvious nonlinear and non-stationary characteristics, which traditional signal processing methods cannot effectively process and are difficult to accurately extract signal features.

[0004] Therefore, it is urgent to solve the problems of high difficulty and low accuracy in detecting rolling bearing faults. Summary of the Invention

[0005] To address one or more of the aforementioned technical problems, this invention proposes to process and fuse vibration signals detected by multiple sensors, and then use these signals to determine faults, thereby effectively improving the accuracy of rolling bearing fault detection. To this end, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for fault diagnosis of rolling bearings, comprising: acquiring vibration signals of the rolling bearing collected by multiple sensors; decomposing and reconstructing the vibration signals according to the maximum overlap discrete wavelet transform to obtain a reconstructed signal; determining the multi-scale scattering entropy corresponding to the reconstructed signal and forming a multi-dimensional state feature vector; and using a trained neural network detection model to detect the multi-dimensional state feature vector to determine the fault type of the rolling bearing.

[0007] In one embodiment, acquiring the vibration signal of the rolling bearing collected by multiple sensors further includes: performing singular value difference spectrum decomposition on the vibration signal and filtering out noise interference components in the vibration signal to obtain a denoised vibration signal.

[0008] In an embodiment, the singular value difference spectrum decomposition of the vibration signal and filtering of noise interference components in the vibration signal comprises: singular value decomposition of the vibration signal of the rolling bearing collected by the plurality of sensors to obtain singular values corresponding to each vibration signal; singular value difference spectrum of the vibration signal is calculated, and singular value reconstruction is performed to reduce noise.

[0009] In an embodiment, the decomposition and reconstruction of the vibration signal according to the maximum overlap discrete wavelet transform to obtain a reconstructed signal comprises: decomposition of the vibration signal according to the maximum overlap discrete wavelet transform to obtain a plurality of component components; the difference degree between the component components is calculated, and the component components sensitive to the signal characteristics are reconstructed to obtain the reconstructed signal.

[0010] In an embodiment, the calculation of the difference degree between the component components comprises:

[0011]

[0012] In the formula, μ n (i), δ n (i), μ un (i), δ un (i) are the mean and standard deviation of the normal signal and the component component decomposed by the maximum overlap discrete wavelet transform respectively, x i and y i are the normal signal and the component signal respectively, DID is the difference degree, and σ1, σ2 are singular values obtained after singular value decomposition.

[0013] In an embodiment, the determination of the multi-scale scatter entropy corresponding to the reconstructed signal and the formation of the multi-dimensional state feature vector comprises: determination of the fine complex multi-scale scatter entropy of each reconstructed signal under different scale factors to form a state feature vector; feature fusion of the state feature vectors corresponding to each sensor to form a multi-dimensional state feature vector.

[0014] In an embodiment, the detection of the multi-dimensional state feature vector by the trained neural network detection model to determine the fault type of the rolling bearing comprises: dividing the multi-dimensional feature vector into training samples and test samples; training the fuzzy neural network by using the training samples to obtain the trained neural network detection model; and detecting the test samples by using the trained neural network detection model to determine the fault type of the rolling bearing.

[0015] In one embodiment, the training of the fuzzy neural network by using the training samples to obtain the trained neural network detection model comprises: calculating the membership of the input variable for a k-dimensional state feature vector; performing fuzzy calculation on each membership by using a multiplication operator to obtain a fuzzy calculation result; calculating an output value of the fuzzy neural network model according to the fuzzy calculation result, the output value comprising a network expected output value and a network actual output value; calculating a corresponding error value according to the network expected output value and the network actual output value; and correcting the correlation coefficient and the parameter in the fuzzy neural network model according to the output result and the error value to obtain the trained neural network detection model.

[0016] In a second aspect, the present application also provides a fault judging device for rolling bearings, comprising: a processor; and a memory storing computer instructions for fault judging of rolling bearings, which, when executed by the processor, cause the device to perform the fault judging method according to one or more of the preceding embodiments.

[0017] In a third aspect, the present application also provides a computer readable storage medium storing computer readable instructions for fault judging of rolling bearings, which, when executed by one or more processors, implement the fault judging method according to one or more of the preceding embodiments.

[0018] According to the scheme of the present application, the effective extraction of fault feature information in signals can be realized by the maximum overlap discrete wavelet packet transform method, ensuring the accuracy of the fault detection process. Through the multi-sensor information fusion method, comprehensive, accurate and reliable detection of rolling bearings is realized, and through the all-around calculation of multi-sensor signals, good robustness is achieved. Further, due to the influence of factors such as background noise and acquisition system, the signals collected by the vibration sensor often contain a large amount of noise and exhibit nonlinear and non-stationary characteristics. The singular value difference spectrum denoising combined with the maximum overlap discrete wavelet packet transform method can effectively reduce the influence of noise factors on the accuracy of signal feature extraction, highlighting the vibration characteristics of the signal itself. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and other objects, features and advantages of the exemplary embodiments of the present application will be more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the present application are shown by way of example, and wherein like reference numerals refer to like elements throughout. In the drawings:

[0020] Figure 1 is a flowchart schematically showing a fault judging method for rolling bearings according to one embodiment of the present application;

[0021] Figure 2is a flow chart schematically showing a failure determination method for a rolling bearing according to another embodiment of the present application;

[0022] Figure 3 is a schematic diagram schematically showing a failure comprehensive simulation test bench according to an embodiment of the present application;

[0023] Figure 4a is a schematic diagram schematically showing a rolling bearing outer ring fracture failure according to an embodiment of the present application;

[0024] Figure 4b is a schematic diagram schematically showing a rolling bearing inner ring wear failure according to an embodiment of the present application;

[0025] Figure 4c is a schematic diagram schematically showing a rolling bearing retainer fracture according to an embodiment of the present application;

[0026] Figure 5a is a rolling bearing inner ring failure vibration signal obtained by a horizontal position acceleration sensor according to an embodiment of the present application;

[0027] Figure 5b is a rolling bearing inner ring failure vibration signal obtained by a vertical position acceleration sensor according to an embodiment of the present application;

[0028] Figure 6 is a singular value of a horizontal position rolling bearing inner ring vibration signal and a singular value difference spectrum thereof according to an embodiment of the present application;

[0029] Figure 7a is a rolling bearing inner ring failure singular value difference spectrum denoising signal obtained by a horizontal position acceleration sensor according to an embodiment of the present application;

[0030] Figure 7b is a rolling bearing inner ring failure singular value difference spectrum denoising signal obtained by a vertical position acceleration sensor according to an embodiment of the present application;

[0031] Figure 8 is a difference degree of each component obtained by MODWPT decomposition of a horizontal position and a vertical position rolling bearing inner ring vibration signal according to an embodiment of the present application;

[0032] Figure 9a is a horizontal position rolling bearing inner ring failure denoising signal after MODWPT reconstruction according to an embodiment of the present application;

[0033] Figure 9b is a vertical position rolling bearing inner ring failure denoising signal after MODWPT reconstruction according to an embodiment of the present application;

[0034] Figure 10 RCMDE distribution curves schematically showing different state rolling bearing reconstruction signals according to an embodiment of the application;

[0035] Figure 11 is a schematic diagram showing different sensor 10 times comparative test results according to an embodiment of the application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are some but not all of the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0037] Rolling bearings are called "joints of industry" and are widely used in many fields such as mechanical engineering, rail transportation, power electronics, aerospace, etc. The health state of rolling bearings directly affects the safety performance of related equipment. According to relevant statistics, about 50% of the failure accidents of rotating machinery equipment are caused by rolling bearings. The information collection capability of a single sensor is limited, and it is difficult to accurately and comprehensively monitor and diagnose the working conditions of rolling bearings, so the fault diagnosis accuracy is difficult to guarantee. Therefore, how to quickly and accurately extract the fault features of such parts comprehensively and accurately, and then perform effective fault diagnosis has always been the focus and difficulty of the research in the field of rotating machinery fault diagnosis.

[0038] Based on this, the present application improves from the aspects of vibration signal collection, processing and fault discrimination to realize comprehensive and reliable judgment of faults.

[0039] The specific embodiments of the application will be described in detail below with reference to the drawings.

[0040] Figure 1 is a flowchart schematically showing a fault judgment method for rolling bearings according to an embodiment of the application.

[0041] As shown in Figure 1 , at step S101, vibration signals of a rolling bearing collected by multiple sensors are acquired. In some embodiments, the vibration signals can also be subjected to singular value difference spectrum decomposition, and noise interference components in the vibration signals are filtered out to obtain denoised vibration signals. The vibration signals x1(t),..., x n (t) collected by multiple sensors are acquired. Singular value difference spectrum decomposition is performed on the signals x1(t),..., x n (t), and noise interference components in the signals are filtered out to obtain denoised signals x′1(t),..., x′n (t).

[0042] At step S102, the vibration signal is decomposed and reconstructed according to a maximum overlap discrete wavelet transform to obtain a reconstructed signal. The maximum overlap discrete wavelet packet transform (MODWPT) method is used to decompose each denoised signal, and k component components c1(t), …, c k (t) are obtained. The difference degrees of each component component are calculated, the component component sensitive to the signal feature is reconstructed, and the reconstructed signal x''1(t), …, x''k(t) is obtained. n

[0043] At step S103, the multi-scale scatter entropy corresponding to the reconstructed signal is determined, and a multi-dimensional state feature vector is formed. In some embodiments, the refined composite multi-scale scatter entropy (RCMDE) of each purified signal under different scale factors is calculated, and a state feature vector T1(r), …, T n (r) is formed. Then, the state feature vectors obtained by each sensor are fused to form a multi-dimensional state feature vector X = [X 1 ({T i (r)}), …, X n ({T i (r)})], i = 1, 2, …, n.

[0044] At step S104, the multi-dimensional state feature vector is detected by using the trained neural network detection model to judge the fault type of the rolling bearing. The multi-dimensional feature vector is divided into training samples P and test samples Q, the fuzzy neural network is trained by using the training samples P, and the trained fuzzy neural network is used to detect the test samples Q, so as to realize the diagnosis of the fault type of the rolling bearing.

[0045] According to the above scheme, the maximum overlap discrete wavelet packet transform method (MODWPT) is the most advanced nonlinear and non-stationary signal analysis method at present. Compared with traditional signal processing methods, it has good adaptability and frequency invariance, and avoids the problem of large difference between local optimum and global feature in the decomposition process. Compared with modern signal processing methods such as empirical mode decomposition (EMD) and local mean decomposition (LMD), MODWPT has a complete theoretical basis and good anti-noise performance, can effectively avoid the end effect and modal aliasing phenomenon caused by envelope fitting and recursive operation, and can more effectively extract the fault feature information in the non-stationary signal, and has been favored by many researchers.

[0046] ​Multi-sensor information fusion technology is a kind of information processing means that has developed very rapidly in recent years, and has been widely used in many fields. Compared with the information collected by a single sensor, the information obtained by multi-sensor has the advantages of comprehensiveness, accuracy, reliability and strong robustness. However, in the field of fault diagnosis, the multi-sensor fault diagnosis model and algorithm are still in the exploratory stage, and there is a lack of model-level methods with strong reliability and high precision. Therefore, in the present application, signal acquisition, noise elimination, multi-sensor signal fusion calculation and other aspects are processed, so as to realize accurate detection of rolling bearing fault.

[0047] Figure 2 is a flow chart schematically showing a fault determination method for a rolling bearing according to another embodiment of the present application.

[0048] As shown in Figure 2 , at step S201, singular value decomposition is performed on the vibration signals of the rolling bearing collected by the plurality of sensors to obtain singular values corresponding to each vibration signal.

[0049] At step S202, the singular value difference spectrum of the vibration signal is calculated, and singular value reconstruction is performed for noise reduction. In some embodiments, the singular value difference spectrum noise reduction is performed on the rolling bearing vibration signals x1(t),…,x n (t) collected by the plurality of sensors, and the specific process is as follows:

[0050] 1) Let x(N)={x1,x2,…,x N} be the rolling bearing vibration signal collected by one sensor in the plurality of sensors, and an m*n order Hankel matrix H can be obtained through phase space reconstruction:

[0051]

[0052] Wherein, N=m+n-1; D m×n is the subspace of the rolling bearing vibration signal without noise; W m×n is the subspace of the noise signal.

[0053] 2) The matrix H can be obtained by singular value decomposition:

[0054]

[0055] Wherein, U and V are orthogonal matrices, ∑ is a non-negative diagonal matrix, S=diag(σ1,σ2,…,σ r ), σ i is the singular value of the matrix H, σ i =∑(i,i), and r is the rank of the matrix H.

[0056] 3) In order to filter out the noise in the signal, the product of the number of rows m and the number of columns n of the Hankel matrix should be as large as possible. The number of rows m and the number of columns n of the Hankel matrix can be determined according to the parity of the number of signal points N, that is:

[0057]

[0058] where N is the number of signal points.

[0059] 4) Let the singular values obtained by singular value decomposition of the matrix H be σ1, σ2, … σ r , normalize each component, that is:

[0060]

[0061] where e = σ1+ σ2+ … + σ r , e1+ e2+ … + e r = 1.

[0062] 5) Calculate the singular value difference spectrum of the signal:

[0063]

[0064] where e i and e i-1 are the normalized singular values of the i-th and i-1-th components of the matrix H, e max and e min are the maximum value and the minimum value of the normalized singular values of each component in the matrix H.

[0065] Since the singular values of the noise are significantly smaller than the singular values of the signal, k can be determined as the distinguishing point of the signal and the noise. Therefore, after retaining the singular values less than or equal to the point k corresponding to the maximum value in the singular value difference spectrum and setting the rest to 0, the singular values are reconstructed, the signal denoising can be realized.

[0066] At step S203, the vibration signal is decomposed according to the maximum overlap discrete wavelet transform to obtain a plurality of component components. In some embodiments, the singular value denoising signal is decomposed by MODOPT, and the specific process is as follows:

[0067] Let be the scale filter and wavelet filter of MODWPT respectively, then and the scale filter g l and the wavelet filter h l in the discrete wavelet transform (DWT) have the following relationship:

[0068]

[0069]

[0070] where g l is the scaling filter of the discrete wavelet transform, h l is the wavelet filter of the discrete wavelet transform;

[0071] satisfying simultaneously:

[0072]

[0073]

[0074] or

[0075]

[0076] where, are MODOPT scaling filters of length l, l-2n and L-l-1, respectively; are MODOPT wavelet filters of length l and L-l-1;

[0077] At scale j, the MODOPT scaling filter and the wavelet filter are padded with 2 j-1 -1 zeros:

[0078]

[0079]

[0080] Thus, according to the Mallat algorithm, the scaling transform coefficients and the wavelet transform coefficients at scale j are calculated as:

[0081]

[0082]

[0083] where V j,t and W j,t are the scaling transform coefficients and the wavelet transform coefficients of the maximal overlap discrete wavelet packet transform at scale j;

[0084] The decomposition coefficients of MODOPT are denoted by W j,n = {W j,n,t , t = 0,..., N-1}, where j is the decomposition level and n is considered as a frequency index varying with j, then the decomposition coefficients of MODOPT are calculated as:

[0085]

[0086] where if n mod 4 = 0 or 3, then If n mod 4 = 1 or 2, then wherein, are the scale filter and the wavelet filter of MODOPT, respectively.

[0087] At step S204, the difference degree between the component components is calculated, and the component components sensitive to the signal feature are reconstructed to obtain a reconstructed signal. In some embodiments, the difference degree of the component components is calculated, and the components sensitive to the signal feature are reconstructed. The specific calculation formula is:

[0088]

[0089] In the formula, μ n (i), δ n (i), μ un (i), δ un (i) are the mean and standard deviation of the normal signal and the component component decomposed by the maximum overlap discrete wavelet transform, respectively, x i and y i are the normal signal and the component signal, DID is the difference degree, and σ1, σ2 are singular values obtained after singular value decomposition.

[0090] At step S205, the fine composite multi-scale scatter entropy of each reconstructed signal under different scale factors is determined to form a state feature vector. In some embodiments, the multi-scale scatter entropy (RCMDE) of each purified signal under different scale factors is calculated, and the multi-dimensional state feature vector T1(t), …, T n (t) is obtained. The specific process is as follows:

[0091] Let x″(t) = [x1, x2, …, x N ] be a VMD reconstructed signal with a length of N. The initial point is continuously divided into [1, τ] to divide the signal x″(t) into non-overlapping regions with a length of τ, and the average value of each region is calculated to obtain a coarse-grained sequence. That is:

[0092]

[0093] The scatter entropy value of the coarse-grained sequence under different scale factors is calculated:

[0094] 1) The signal x″(t) is mapped to y = [y1, y2, …, y N ] in the range of [0, 1] by using the standard normal distribution function, that is,

[0095]

[0096] In the formula, μ and σ are the expectation and variance of the signal x″(t).

[0097] 2) Further map y to the range [1, c] by a linear transformation algorithm, i.e.

[0098]

[0099] wherein, is the i-th class of time series, and round is the rounding function.

[0100] 3) Perform phase space reconstruction on z c , then the embedding vector is:

[0101]

[0102] wherein, j = 1, 2, …, N-(m-1)d, and m and d are the embedding dimension and time delay, respectively. Each is mapped to a scatter pattern wherein, Since the sequence contains m elements, each element can take any integer of [1, c], therefore the number of all possible scatter patterns is c m .

[0103] 4) Calculate the probability p of each scatter pattern m under c :

[0104]

[0105] wherein, is the number of occurrences of each .

[0106] 5) According to the information entropy theory, the scatter entropy of the time series x can be defined as:

[0107]

[0108] 6) The RCMDE under each scale factor τ is defined as:

[0109]

[0110] 7) Under a specified scale factor τ, the RCMDE value is formed into a state feature vector, which is used to discriminate the signal state.

[0111] At step S206, the state feature vectors corresponding to each sensor are fused to form a multi-dimensional state feature vector. In some embodiments, the state feature vectors obtained by each sensor are fused to form a multi-dimensional state feature vector X = [X 1 ({T i (r)}), …, X​n ({T i (r)})],i=1,2,…,n。

[0112] At step S207, the multi-dimensional feature vector is divided into training samples and test samples.

[0113] At step S208, the fuzzy neural network is trained using the training samples to obtain a trained neural network detection model. In some embodiments, the multi-dimensional feature vector is divided into training samples P and test samples Q, the fuzzy neural network is trained using the training samples P, and the trained fuzzy neural network is used to detect the test samples Q to realize diagnosis of the fault type of the rolling bearing. The calculation steps of the fuzzy neural network are as follows:

[0114] 1) Calculate the membership degree of each input variable r k of a k-dimensional state feature vector T = [r1, r2, ∑, r j k, and the membership function adopts a Gaussian type:

[0115]

[0116] where μ is a Gaussian membership function coefficient, is the membership function of the jth variable in the ith fuzzy subset, are the center and width of the membership function of the jth variable in the ith fuzzy subset, respectively, k is the number of feature vectors, and n is the number of fuzzy subsets.

[0117] 2) Perform fuzzy calculation on each membership degree ∑ fuzzy operator ω i :

[0118]

[0119] 3) Calculate the output value y i of the fuzzy model according to the fuzzy calculation result:

[0120]

[0121] 4) Calculate the error e:

[0122]

[0123] where y d is the network expected output, and y c is the network actual output.

[0124] 5) According to the output result, correct the correlation coefficient and parameters

[0125]

[0126] in, Let be the correlation coefficient of the j-th variable in the i-th fuzzy subset, and α be the adjustment parameter for the actual output value of the fuzzy model. This represents the actual output value of the fuzzy model. Let the center of the membership function of the j-th variable in the i-th fuzzy subset be , Let β be the width of the membership function of the j-th variable in the i-th fuzzy subset, and let β be the adjustment coefficient of the actual output parameter of the fuzzy model. The center of the actual membership function of the fuzzy model. This represents the width of the actual membership function of the fuzzy model.

[0127] In step S209, the trained neural network detection model is used to detect the test sample in order to determine the fault type of the rolling bearing.

[0128] In one embodiment, training the fuzzy neural network using training samples to obtain a trained neural network detection model includes: calculating the membership degree of the input variable for a k-dimensional state feature vector; performing fuzzy calculation on each membership degree using a multiplication operator to obtain fuzzy calculation results; calculating the output value of the fuzzy neural network model based on the fuzzy calculation results, the output value including the network's expected output value and the network's actual output value; calculating the corresponding error value based on the network's expected output value and the network's actual output value; and correcting the correlation coefficient and parameters in the fuzzy neural network model based on the output results and the error value to obtain a trained neural network detection model.

[0129] To verify the effectiveness of the above embodiments, the present invention will combine... Figure 3 The rolling bearing fault comprehensive simulation test bench shown further illustrates the effectiveness of this invention in rolling bearing fault detection. For example... Figure 3 As shown, the experimental platform consists of an AC motor, a motor speed controller, a rotating shaft, support bearings, a hydraulic loading system, and test bearings. The bearings used in the experiment can be, for example, LDK UER204 rolling bearings, and their relevant parameters are shown in Table 1.

[0130] Table 1 Parameters of LDK UER204 Rolling Bearing

[0131]

[0132] During fault simulation, the inner ring fault is wear fault, the outer ring fault is crack fault, and the cage fault is breakage fault. Figure 4a , Figure 4b and Figure 4c The simulation results show three failure modes and locations of the bearing. The following section will explain this in conjunction with the detection process of inner ring wear failure in rolling bearings. Figures 5a to 11The data of the inner ring wear fault are shown in the table.

[0133] In the signal acquisition process, in order to comprehensively monitor the state of the rolling bearing, two PCB 352C33 one-way acceleration sensors are fixed to the horizontal and vertical directions of the test bearing through magnetic seats. A DT9837 portable signal acquisition device is used to collect the acceleration signals, the signal sampling frequency is 25.6 kHz, the sampling interval is 1 min, the sampling time is 1.28 s each time, and the analysis time is 0.5 s.

[0134] Figure 5a and Figure 5b are respectively the rolling bearing vibration signals obtained by the horizontal position and vertical position acceleration sensors under the inner ring fault state in the application examples of the present application. Figure 5a and Figure 5b It can be seen that, since the noise filtering device is not used, a large amount of noise components are contained in the collected vibration signals, so that the internal impact components are submerged in the noise, and these impact components are the key to characterize the signal characteristics, and if they are lost, the accuracy of signal feature extraction will be seriously affected, and the reliability of fault diagnosis will be reduced. Therefore, the singular value difference spectrum is used for noise reduction processing.

[0135] The signal sampling point number N = 12800 is an even number, so the number of Hankel matrix rows is 6400, and 6400 singular values are obtained. Since the maximum value of the singular value difference spectrum is usually in the early position, the first 50 singular values of the horizontal direction inner ring fault signal and the singular value difference spectrum thereof are taken, as shown in Figure 6 The first 50 singular values of the horizontal direction inner ring fault signal and the singular value difference spectrum thereof are shown in

[0136] It can be seen from Figure 6 that the maximum mutation point in the singular value difference spectrum is 2. The first 2 singular values are reserved, and the remaining are set to 0 for reconstruction, and the denoised signal is shown in Figure 7a The vertical direction inner ring fault signal is denoised by this method, and the result is shown in Figure 7b .

[0137] It can be seen from Figure 7a and Figure 7b that most of the noise interference components in the inner ring fault signal are filtered out, and the impact characteristics of the signal are highlighted. However, it can still be seen that there is a certain similarity between the waveforms of the inner ring denoised signals at different positions, in order to ensure the accuracy of fault diagnosis, the above denoised signal characteristics need to be further extracted.

[0138] MODWPT is used to decompose the inner ring fault signal after singular value difference spectrum denoising, and the difference degrees of the component components obtained by decomposition are calculated, and the result is shown in Figure 8 It can be seen from Figure 8It can be seen that the difference of components C1-C3 is larger after the inner ring fault signal is denoised by horizontal and vertical singular value difference spectrum and decomposed by MODWPT, and the reconstructed signal is shown in Figure 9a and Figure 9b .

[0139] It can be seen from Figure 9a and Figure 9b that the high-frequency noise and low-frequency background components in the signal are further filtered out, the impact characteristics in the signal are more obvious, and the similarity of the signal time domain waveform is reduced, which provides a guarantee for the accuracy of subsequent signal feature extraction.

[0140] A group of rolling bearing vibration signals under different states are randomly collected and processed to obtain the reconstructed signals according to the method. The RCMDE values of the reconstructed signals are calculated, and the results are shown in Figure 10 . In the calculation process, the embedding dimension m=2, the category c=5, the delay d=1, and the maximum scale factor τ max =20. It can be seen that there is a good distinction between different faults, indicating that the characteristics of the signal are effectively extracted, which provides a guarantee for subsequent fault diagnosis.

[0141] The RCMDE values of the two MODWPT reconstructed signals of the same fault in the horizontal and vertical directions are fused into a high-dimensional state feature vector. The high-dimensional state feature vector is taken as a sample, and 50 samples are randomly taken for each fault, and a total of 150 samples are taken for 3 types of faults. 80% of them are divided into a training set, and the remaining 20% are divided into a test set, as shown in Table 2.

[0142] Table 2 Division of bearing data set in experimental process

[0143]

[0144] To reduce random errors, the training samples are trained 10 times by fuzzy neural network, and then tested by test samples, and finally the corresponding fault diagnosis results are obtained. The results are shown in Figure 11 . For comparative analysis, the fault diagnosis results based on single direction sensor vibration signal are also given, as shown in Figure 11 . The diagnosis results based on multi-sensor fusion signal are obviously better than those of single sensor, which shows that this method is feasible and effective, and also provides a certain reference for fault diagnosis problem under small sample condition.

[0145] Based on this, the application further provides, in a second aspect, a fault judging device for a rolling bearing, comprising: a processor; and a memory storing computer instructions for fault judging of the rolling bearing, which, when executed by the processor, cause the device to perform the fault judging method according to one or more of the preceding embodiments.

[0146] The application further provides, in a third aspect, a computer readable storage medium storing computer readable instructions for fault judging of a rolling bearing, which, when executed by one or more processors, implement the fault judging method according to one or more of the preceding embodiments.

[0147] Although the present specification has shown and described a number of embodiments of the application, it will be apparent to those skilled in the art that many modifications, variations, and substitutions can be made thereunto in the course of implementation. It should be understood that various alternatives to the embodiments of the application described herein can be employed in practicing the application. It is intended that the following claims define the scope of the application and that methods equivalent to those shown and described herein can be utilized without departing from the spirit and scope of the application.

Claims

1. A failure determination method for a rolling bearing, characterized by, The method comprises the following steps: obtaining vibration signals of a rolling bearing collected by multiple sensors; decomposing and reconstructing the vibration signals according to a maximum overlap discrete wavelet transform to obtain reconstructed signals; determining multiscale scatter entropies corresponding to the reconstructed signals and forming a multidimensional state feature vector; detecting the multidimensional state feature vector by using a trained neural network detection model to determine a fault type of the rolling bearing; the step of decomposing and reconstructing the vibration signals according to the maximum overlap discrete wavelet transform to obtain reconstructed signals comprises the following steps: decomposing the vibration signals according to the maximum overlap discrete wavelet transform to obtain multiple component components; calculating a difference degree between the component components and reconstructing the component components sensitive to signal features to obtain reconstructed signals; the step of calculating the difference degree between the component components comprises the following steps: wherein, μ n i δ n i μ un i δ un i x i y i DID is the difference degree;​​​​​​​​​​ the step of determining the multiscale scatter entropies corresponding to the reconstructed signals and forming the multidimensional state feature vector comprises the following steps: determining fine composite multiscale scatter entropies of each reconstructed signal under different scale factors to form a state feature vector; performing feature fusion on the state feature vectors corresponding to each sensor to form a multidimensional state feature vector.

2. The failure determination method according to claim 1, characterized by, The step of obtaining the vibration signals of the rolling bearing collected by the multiple sensors further comprises the following steps: performing singular value differential spectrum decomposition on the vibration signals and filtering out noise interference components in the vibration signals to obtain denoised vibration signals.

3. The failure determination method according to claim 2, characterized by, The step of performing singular value differential spectrum decomposition on the vibration signals and filtering out noise interference components in the vibration signals comprises the following steps: performing singular value decomposition on the vibration signals of the rolling bearing collected by the multiple sensors to obtain singular values corresponding to each vibration signal; calculating a singular value differential spectrum of the vibration signals and performing singular value reconstruction to perform noise reduction.

4. The failure determination method according to claim 1, characterized by, The step of detecting the multidimensional state feature vector by using the trained neural network detection model to determine the fault type of the rolling bearing comprises the following steps: dividing the multidimensional feature vector into training samples and test samples; training a fuzzy neural network by using the training samples to obtain a trained neural network detection model; detecting the test samples by using the trained neural network detection model to determine the fault type of the rolling bearing.

5. The failure determination method according to claim 4, characterized by, The step of training the fuzzy neural network by using the training samples to obtain the trained neural network detection model comprises the following steps: calculating membership degrees of input variables for a k-dimensional state feature vector; performing fuzzy calculation on the membership degrees by using a continuous multiplication operator to obtain a fuzzy calculation result; calculating output values of the fuzzy neural network model according to the fuzzy calculation result, wherein the output values comprise network expected output values and network actual output values; calculating corresponding error values according to the network expected output values and the network actual output values; correcting relevant coefficients and parameters in the fuzzy neural network model according to the output result and the error values to obtain the trained neural network detection model.

6. A fault diagnosis device for rolling bearings, characterized in that, The device comprises: a processor; and a memory storing computer instructions for fault determination of a rolling bearing, wherein when the computer instructions are executed by the processor, the device performs the fault determination method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, Computer readable instructions for failure determination of a rolling bearing are stored thereon, and the computer readable instructions, when executed by one or more processors, implement the failure determination method according to any one of claims 1-5.

Images