A rolling bearing acoustic vibration signal fault diagnosis method, system and device

Abstract

The present invention discloses a rolling bearing acoustic and vibration signal fault diagnosis method, system and equipment, and relates to the field of rolling bearing fault diagnosis classification. The method comprises: collecting acoustic signals and vibration signals under the working state of the rolling bearing; extracting acoustic sensitive features and vibration sensitive features respectively; inputting the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputting the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result; on the basis of retaining the conflict information of the acoustic signal diagnosis result and the vibration signal diagnosis result, adopting the conflict proportion redistribution rule to redistribute the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fusion diagnosis result. The present invention can effectively eliminate the interference features in the mixed domain feature set of complex faults of rolling bearings, select acoustic sensitive features and vibration sensitive features, and solve the problem that a single vibration signal or acoustic signal is difficult to fully reflect the real operating state of the rolling bearing.

Description

Technical Field

[0001] The present invention belongs to the technical field of rolling bearing fault diagnosis classification, and in particular relates to a rolling bearing acoustic vibration signal fault diagnosis method, system and equipment. Background Art

[0002] Rolling bearings are key components of mechanical equipment shaft support, and their normal working condition directly affects the operation of the equipment. Slight defects on their surface may cause the entire system to fail, resulting in huge property losses. More than 30% of mechanical equipment failures are caused by rolling bearing failures. Traditional rolling bearing monitoring methods are mostly fixed life replacement or regular safety inspections, which consume manpower and material resources and affect the production process; intelligent maintenance methods use advanced sensors and detection technologies to obtain equipment status information for online detection, predict the remaining life of the equipment based on the established statistical model, continuously update the remaining life of the equipment, and achieve scientific and economical maintenance. Therefore, bearing status detection and online fault diagnosis are of great significance to ensure the safe and economical operation of equipment.

[0003] At present, the classification and condition monitoring of bearing faults are mostly based on vibration signals collected by vibration sensors; the early fault characteristics of rolling bearings are very weak and are affected by the transmission path and signal propagation medium. The weak fault signal characteristics are not only reflected in the vibration signal, but also in the acoustic radiation signal. It is difficult to fully reflect the actual operating status of the rolling bearing by using only a single vibration signal or acoustic signal. It is very meaningful to fuse the diagnosis results of the acoustic and vibration signals for decision making. The feature extraction of acoustic and vibration signals is a key technology for bearing fault diagnosis, which is directly related to the accuracy of diagnosis. A large number of scholars have studied fault diagnosis methods based on time domain features and frequency domain features. Some scholars have also combined the time domain, frequency domain, and time-frequency domain into a hybrid domain feature index set as the original data set for training fault classifiers, trying to fully reflect the fault information. However, the selection of more feature quantities will increase the difficulty of model calculation and cause serious redundancy between features, resulting in reduced accuracy of fault identification and longer diagnosis time. Therefore, a method for evaluating the sensitivity of high-dimensional fault feature indicators in the hybrid domain is needed to eliminate the information redundancy and interference features in the hybrid domain feature set and extract features that are sensitive to acoustic and vibration signals. Summary of the invention

[0004] The purpose of the present invention is to provide a rolling bearing acoustic and vibration signal fault diagnosis method, system and equipment. By performing feature extraction, feature selection, preliminary diagnosis and fusion decision on the acoustic signal and vibration signal of the rolling bearing respectively, the tSNE-ASC feature selection method proposed in the present invention can effectively eliminate the interference features in the mixed domain feature set of complex faults of the rolling bearing, and select acoustic sensitive features and vibration sensitive features; the present invention uses the DSmT acoustic and vibration decision fusion method to effectively reduce the diagnostic uncertainty caused by a single vibration signal or acoustic signal, and has a high diagnostic accuracy under variable load and non-steady speed conditions.

[0005] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:

[0006] The present invention provides a rolling bearing acoustic and vibration signal fault diagnosis method, the method comprising: collecting acoustic signals and vibration signals of the rolling bearing in a working state;

[0007] Respectively extracting the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal;

[0008] Inputting the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputting the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result;

[0009] On the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, the acoustic signal diagnosis result and the vibration signal diagnosis result are reallocated by adopting a conflict proportion reallocation rule to obtain a fusion diagnosis result.

[0010] In an optional embodiment of this solution, the acoustic sensitive feature is selected in the following manner:

[0011] The acoustic sensitive features are selected in the following manner:

[0012] Obtaining acoustic signal samples of the rolling bearing under different fault types and different operating conditions;

[0013] Performing variational mode decomposition on each of the acoustic signal samples to obtain an intrinsic mode function of each of the acoustic signal samples;

[0014] Extracting acoustic signal features from each of the intrinsic mode functions, such as the signal features including time domain features and nonlinear features such as time domain features and nonlinear features;

[0015] Constructing a feature set matrix for each of the acoustic signal time domain features and each of the nonlinear features;

[0016] Reducing the dimension of each of the feature set matrices to two dimensions respectively, to obtain a two-dimensional feature set matrix of each of the feature set matrices;

[0017] Calculating the average silhouette coefficient of each of the two-dimensional feature set matrices, and selecting features that are sensitive to acoustic fault signals as acoustic sensitive features according to the average silhouette coefficient;

[0018] The vibration sensitivity feature is selected in the following manner:

[0019] Obtaining vibration signal samples of the rolling bearing under different fault types and different working conditions;

[0020] Performing variational mode decomposition on each of the vibration signal samples to obtain an intrinsic mode function of each of the vibration signal samples;

[0021] Extracting vibration signal features from each of the intrinsic mode functions, wherein the signal features include time domain features and nonlinear features;

[0022] Constructing a feature set matrix for each of the vibration signal features to obtain the time domain features and nonlinear features of each of the intrinsic mode functions;

[0023] Constructing a feature set matrix for each of the time domain features and each of the nonlinear features;

[0024] Reducing the dimension of each of the feature set matrices to two dimensions respectively, to obtain a two-dimensional feature set matrix of each of the feature set matrices;

[0025] The average silhouette coefficient of each of the two-dimensional feature set matrices is calculated, and features that are sensitive to vibration fault signals are screened out as vibration sensitive features based on the average silhouette coefficient.

[0026] In an optional embodiment of the present scheme, the step of reducing the dimensions of all the feature set matrices to two dimensions to obtain a two-dimensional feature set matrix includes using a t-distributed random neighbor embedding method to reduce the dimensions of all the feature set matrices to two dimensions to obtain a two-dimensional feature set matrix.

[0027] In an optional embodiment of the present scheme, the acoustic sensitive feature includes one or more combinations of approximate entropy, sample entropy, fuzzy entropy, mean value, variance, peak value, kurtosis, root mean square, margin factor, pulse coefficient and shape coefficient.

[0028] In an optional embodiment of the present scheme, the vibration-sensitive feature type includes one or more combinations of Li index, approximate entropy, sample entropy, fuzzy entropy, mean value, variance, peak value, kurtosis, root mean square, crest factor, margin factor, pulse coefficient and shape coefficient.

[0029] In an optional embodiment of the present solution, the method further includes the step of performing fault diagnosis on the acoustic sensitive feature and the vibration sensitive feature to obtain an acoustic signal diagnosis result and a vibration signal diagnosis result respectively, including:

[0030] Obtaining acoustic signal samples and vibration signal samples of the rolling bearing under different fault types and different working conditions;

[0031] Extracting the acoustic sensitive features of each of the acoustic signal samples to construct an acoustic signal data set, and extracting the vibration sensitive features of each of the vibration signal samples to construct a vibration signal data set;

[0032] The acoustic fault classifier and the vibration fault classifier are trained using the acoustic signal data set and the vibration signal data set respectively.

[0033] In an optional embodiment of the present scheme, the fault type includes a combination of one or more of a normal bearing, an outer ring single point fault, an inner ring single point fault, a roller single point fault, an outer ring roller composite fault, an inner ring roller composite fault, an outer ring multi-point fault, an inner ring multi-point fault and a roller multi-point fault.

[0034] In an optional embodiment of the present solution, the step of reallocating the acoustic signal diagnosis result and the vibration signal diagnosis result by using a conflict ratio redistribution rule on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fused diagnosis result includes:

[0035] The acoustic signal diagnosis result and the vibration signal diagnosis result are used as evidence bodies to establish a generalized identification framework;

[0036] According to the focal element in the generalized identification framework, a super power set is obtained by performing intersection and union operations;

[0037] According to the generalized identification framework and the super power set, a credibility distribution function on the generalized identification framework is obtained;

[0038] Based on the DSmT fusion rule, the acoustic signal diagnosis result and the vibration signal diagnosis result are used as generalized confidence allocations of identification framework elements.

[0039] According to the reliability distribution function, a fusion diagnosis result is obtained.

[0040] The present invention also provides a rolling bearing acoustic vibration signal fault diagnosis system, the system comprising:

[0041] A signal acquisition module is used to collect acoustic signals and vibration signals of the rolling bearing in working state;

[0042] A sensitive feature separation module, used to extract the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal respectively;

[0043] A preliminary diagnosis module, used for inputting the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputting the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result;

[0044] The fusion decision module is used to reallocate the acoustic signal diagnosis result and the vibration signal diagnosis result by adopting a conflict proportion redistribution rule on the basis of retaining the conflict information of the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fusion diagnosis result.

[0045] The present invention also provides a rolling bearing acoustic vibration signal fault diagnosis device, the device comprising:

[0046] Testing machines for loading rolling bearings;

[0047] A soundproof enclosure, wrapping the testing machine;

[0048] Sound sensor, used to collect sound signals of rolling bearing in working state;

[0049] Vibration sensor, used to collect vibration signals of rolling bearings in working state;

[0050] A processing unit is used to receive the sound signal and the vibration signal, and perform the following processing:

[0051] Respectively extracting the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal;

[0052] Inputting the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputting the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result;

[0053] On the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, the acoustic signal diagnosis result and the vibration signal diagnosis result are reallocated by adopting a conflict proportion reallocation rule to obtain a fusion diagnosis result.

[0054] The implementation of this solution includes but is not limited to the following technical effects:

[0055] (1) On the basis of retaining the conflict information between the acoustic signal diagnosis results and the vibration signal diagnosis results, the conflict proportion redistribution rule is used to redistribute the acoustic signal diagnosis results and the vibration signal diagnosis results to obtain a fusion diagnosis result. This solves the problem that it is difficult to fully reflect the true operating status of the rolling bearing by using only a single vibration signal or acoustic signal.

[0056] (2) Redundant information and interference features in the acoustic signal and vibration signal are eliminated, and the acoustic sensitive features that are sensitive to the acoustic fault signal and the vibration sensitive features that are sensitive to the vibration fault signal are screened out. This method has better discrimination and recognition, and has higher diagnostic accuracy for fault diagnosis.

[0057] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0059] Figure 1 This is a flow chart of a rolling bearing acoustic and vibration signal fault diagnosis method of this scheme.

[0060] Figure 2 for Figure 1 Step S2 is a schematic diagram of the process of constructing a feature data set of acoustic vibration signals and feature selection.

[0061] Figure 3 for Figure 1 Step S3 is a flowchart of obtaining the acoustic signal diagnosis result and the vibration signal diagnosis result.

[0062] Figure 4 for Figure 1 Step S4 is a schematic diagram of the process of fusion decision-making.

[0063] Figure 5 This is a structural block diagram of a rolling bearing acoustic and vibration signal fault diagnosis system in this scheme.

[0064] Figure 6 This is a structural block diagram of a rolling bearing acoustic and vibration signal fault diagnosis device in this scheme.

[0065] Figure 7 It is a two-dimensional diagram of the entropy of vibration signal samples.

[0066] Figure 8 It is a two-dimensional diagram of the skewness of the vibration signal.

[0067] Fig. 9 A two-dimensional illustration of the entropy of acoustic signal samples.

[0068] Fig.10 2D diagram of acoustic signal skewness.

[0069] Fig.11This is a diagram showing the diagnostic results of the first group of acoustic signals.

[0070] Fig.12 This is a diagram showing the diagnostic results of the first group of vibration signals.

[0071] Fig.13 This is a diagram showing the diagnostic results of the second group of acoustic signals.

[0072] Fig.14 This is a diagram showing the diagnostic results of the second group of vibration signals.

[0073] Fig.15 This is a diagram showing the diagnostic results of the third group of acoustic signals.

[0074] Fig.16 This is a diagram showing the diagnostic results of the third group of vibration signals.

[0075] Fig.17 This is a diagram showing the diagnostic results of the fourth group of acoustic signals.

[0076] Fig.18 This is a diagram showing the diagnostic results of the fourth group of vibration signals.

[0077] Fig.19 This is a diagram showing the diagnosis results of non-sensitive features of acoustic signals.

[0078] Fig. 20 Diagram of the diagnosis results of the non-sensitive characteristics of the vibration signal.

[0079] In the accompanying drawings, the components represented by the reference numerals are listed as follows:

[0080] 11-Signal acquisition module;

[0081] 12-Sensitive feature separation module;

[0082] 13- Preliminary diagnosis module;

[0083] 14- Fusion decision module;

[0084] 21- Testing machine;

[0085] 22-soundproof enclosure;

[0086] 231 - first acoustic sensor, 232 - second acoustic sensor;

[0087] 241- unidirectional acceleration vibration sensor, 242- triaxial acceleration vibration sensor;

[0088] 25-Processing unit. DETAILED DESCRIPTION

[0089] 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.

[0090] See also Figures 1 to 20 It should be noted that the illustrations provided in this embodiment are only used to illustrate the basic concept of the present invention in a schematic manner, and the illustrations only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0091] In order to solve the problem that it is difficult to fully reflect the actual operating status of a rolling bearing by using only a single vibration signal or acoustic signal, the present embodiment discloses a rolling bearing acoustic and vibration signal fault diagnosis method, which collects acoustic signals and vibration signals under the working state of the rolling bearing; extracts acoustic sensitive features in the acoustic signal and vibration sensitive features in the vibration signal respectively according to the tSNE-ASC feature selection method; inputs the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputs the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result; on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, adopts the conflict proportion redistribution rule to redistribute the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fused diagnosis result. Among them, Figure 1 A schematic flow chart of the rolling bearing acoustic vibration signal fault diagnosis method of the present invention is shown.

[0092] Now combine Figure 1 The steps S1 to S4 of the rolling bearing acoustic vibration signal fault diagnosis method of this embodiment are specifically described.

[0093] First, step S1 is performed to collect the acoustic signal and vibration signal of the rolling bearing in the working state, wherein the acoustic signal of the rolling bearing is collected by an acoustic sensor, and the acoustic sensor can be composed of a first acoustic sensor 231 and a second acoustic sensor 232 to form a dual channel. The vibration signal of the rolling bearing is collected by a vibration sensor, and the vibration sensor is composed of a unidirectional acceleration vibration sensor 241 and two three-directional acceleration vibration sensors 242.

[0094] Next, step S2 is performed to extract the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal. Figure 2As shown, step S2 may further include: the acoustic sensitive feature is selected in the following manner (step S21), and the vibration sensitive feature is selected in the following manner (step S22), and finally the acoustic sensitive feature and the vibration sensitive feature are obtained (step S23).

[0095] Step S21 may further include:

[0096] Step S211: Obtain acoustic signal samples of the rolling bearing under different fault types and different working conditions.

[0097] Step S212: Perform variational mode decomposition on each acoustic signal sample to obtain the intrinsic mode function (IMF) of each acoustic signal sample, and determine the decomposition level K of VMD (variational mode decomposition) according to the spectral correlation coefficient method. The specific method is as follows: Calculate the spectral correlation coefficient of each IMF according to formula (1):

[0098] ρ1, ρ2, ..., ρk, where ρi (1<i<k) is its minimum value. If ρi<ε (ε is the set threshold), then k-1 is the minimum decomposition level; otherwise, let k=k+1, and repeat the above cycle until ρi<ε is satisfied to determine the minimum decomposition level k. Where: |Pk| is the modulus of the Fourier transform of the original signal. |Q| is the modulus of the Fourier transform of the kth IMF component.

[0099]

[0100] The VMD decomposition layer number K of this scheme can be 8, that is, each sample is decomposed into 8 IMFs.

[0101] Step S213, extract the acoustic signal features for each intrinsic mode function. The acoustic signal features may include time domain features and nonlinear features. The types of time domain features of each IMF are counted: mean, variance, peak, kurtosis, root mean square, crest factor, margin factor, pulse coefficient, shape coefficient, and skewness. Reconstruct the phase space of each IMF, determine the embedding dimension and time delay according to the mutual information method and the nearest neighbor method, and calculate the type of nonlinear features of each IMF in the reconstructed phase space: maximum Lyapunov index (Lyapunov index), correlation dimension, approximate entropy, fuzzy entropy, and sample entropy.

[0102] Step S214 constructs a feature set matrix for each acoustic signal feature, and each feature set matrix is ​​in the form of (9×100)×8.

[0103] Step S215, reduce the dimension of each feature set matrix to two dimensions respectively, and obtain the two-dimensional feature set matrix of each feature set matrix. Manifold learning method is widely used in fault diagnosis as a basic method of pattern recognition. It effectively realizes data dimension reduction, data visualization, etc. by mapping low-dimensional manifold structure from high-dimensional data. t-distributed Stohastic Neighbor Embedding (tSNE) is a nonlinear, non-parametric popular learning method, and its clustering effect for high-dimensional data is generally better than principal component analysis (Principal Component Analysis, PCA) method. Its basic idea is to map the data points of the original high-dimensional space to the low-dimensional space, while keeping the probability distribution between each other unchanged, and transform the distance measure between high and low dimensions into probability measure, so that the KL divergence between the two distributions is minimized, and the sample distribution of the low-dimensional space is obtained. That is, tSNE reduces the high-dimensional data set to a low dimension. If the data set has a good classification effect in the low dimension, the original high-dimensional data also has a good classification effect. In other words, it is to map the high-order matrix to the t-distribution. Of course, in other applications of this scheme, it can also be mapped to other normal distributions. As long as the two-dimensional expansion of the high-dimensional matrix can be achieved and the probability distribution is kept unchanged and the relative entropy is minimized, it belongs to the protection scope of this scheme. When the t-distribution scheme is used, the basic steps of the tSNE algorithm are as follows:

[0104] Assume that the high-dimensional dataset sequence X = {x 1 , x 2 ,...x N}, N represents the dimension of the original data, then the high-dimensional data point x i and x j The similarity conditional probability distribution between f|i :

[0105]

[0106] Where: i Is x i The variance of the Gaussian centered at is determined by a pre-given perplexity factor and binary search.

[0107] Calculate the value of 2 data points x in high-dimensional space i With x j The joint probability density p ij

[0108]

[0109] In the low-dimensional space, the Gaussian distribution is replaced by the t distribution with a degree of freedom of 1, and the joint probability distribution of the data points in the low-dimensional space is reduced to Y n(where n≤N, n is the dimension of the data after dimensionality reduction):

[0110]

[0111] KL divergence (Kullback-Leibler divergences) is used to measure the distribution similarity between data points before and after dimensionality reduction, and the cost function is obtained:

[0112]

[0113] In order to maximize the similarity of the probability distribution obtained before and after dimensionality reduction, the gradient descent method is used to optimize the KL divergence, and the gradient is as follows:

[0114]

[0115] When using the gradient descent method to optimize the cost function, a relatively large momentum is introduced to accelerate the optimization process. During the gradient search iteration, the current gradient is added to the exponential decay sum of the gradient of the previous iteration. The update rule is:

[0116]

[0117] where Y (t) represents the solution obtained at the tth iteration, the learning rate η, and the momentum term at the tth iteration

[0118] After being processed by the above algorithm, a two-dimensional feature set matrix is ​​obtained after the dimension reduction of all feature set matrices. Step S24, respectively calculate the average silhouette coefficient of all two-dimensional feature set matrices. If the average silhouette coefficient is greater than 0, they are marked as acoustic sensitive features and vibration sensitive features respectively. That is, the t-distributed Stohastic Neighbor Embedding (tSNE) method is combined with the method of calculating the average silhouette coefficient (ASC / Average Silhouette Coefficient), which effectively eliminates the interference features in the mixed domain feature set of complex faults of rolling bearings, effectively reduces the uncertainty of diagnosis, and has a high diagnostic accuracy under variable load and non-steady speed conditions.

[0119] Step S216: Calculate the average silhouette coefficient of each two-dimensional feature set matrix, and select the features that are sensitive to the acoustic fault signal as the acoustic sensitive features according to the average silhouette coefficient. The silhouette coefficient in this scheme is a way to evaluate the clustering effect. i The mean of is called the average silhouette coefficient of the clustering results. It combines the two factors of cohesion and separation. The calculation steps are as follows:

[0120] 1) Calculate the average distance a from sample i to other samples in the same cluster i ,This paper adopts the squared Euclidean distance.

[0121] a i The smaller it is, the more likely sample i should be clustered into this cluster. i It is called the intra-cluster dissimilarity of sample i.

[0122] 2) Calculate the number of samples i to another cluster C j The average distance b of all samples ij , called sample i and cluster C j The dissimilarity of .

[0123] Define the inter-cluster dissimilarity of sample i: b i =min{b i1 , b i2 , ...b ik}, b i The larger it is, the less sample i belongs to other clusters.

[0124] 3) According to the intra-cluster dissimilarity a of sample i i and the inter-cluster dissimilarity b i , define the silhouette coefficient of sample i:

[0125]

[0126] but

[0127]

[0128] S i ∈[-1,1] (10)

[0129] s i If it is close to 1, it means that the clustering of sample i is reasonable; s i If s is close to -1, it means that sample i should be classified into another cluster. i is approximately 0, indicating that sample i is on the boundary of two clusters. i The mean of is called the average silhouette coefficient (ASC) of the clustering result. ASC is a measure of whether the clustering is reasonable and effective.

[0130] When 0<MSC≤1, it means that the clustering is reasonable, and the larger the ASC value, the more reasonable the clustering.

[0131] When -1≤MSC<0, it means that the clustering is unreasonable. The smaller the ASC value, the more unreasonable the clustering is.

[0132] This scheme uses whether the ASC value is greater than 0 as the criterion for feature selection, and finally screens out the acoustic sensitive features.

[0133] The extraction of vibration-sensitive features in step S22 specifically includes:

[0134] Step S221, obtaining vibration signal samples of the rolling bearing under different fault types and different working conditions.

[0135] Step S222: Perform variational mode decomposition on each vibration signal sample to obtain the intrinsic mode function of each vibration signal sample.

[0136] Step S223: Obtain the vibration signal characteristics of each intrinsic mode function. The vibration signal characteristics may be time domain characteristics and nonlinear characteristics.

[0137] Step S224: construct a feature set matrix for each signal feature.

[0138] Step S225 , reduce the dimension of each feature set matrix to two dimensions to obtain a two-dimensional feature set matrix of each feature set matrix.

[0139] Step S226, calculating the average silhouette coefficient of each two-dimensional feature set matrix, and selecting features that are sensitive to vibration fault signals as vibration sensitive features based on the average silhouette coefficient.

[0140] The specific modeling and calculation methods of steps S221 to S226 are the same as those of steps S211 to S216.

[0141] Then, step S3 is performed to input the acoustic sensitive features into the trained acoustic fault classifier to obtain the acoustic signal diagnosis result, and the vibration sensitive features are input into the trained vibration fault classifier to obtain the vibration signal diagnosis result. Figure 3 As shown, it may further include steps S31 to S33.

[0142] Step S31, obtaining acoustic signal samples and vibration signal samples of the rolling bearing under different fault types and different working conditions.

[0143] Step S32: extract the acoustic sensitivity features of each acoustic signal sample to construct an acoustic signal data set, and extract the vibration sensitivity features of each vibration signal sample to construct a vibration signal data set.

[0144] Step S33 uses the acoustic signal data set and the vibration signal data set to train the acoustic fault classifier and the vibration fault classifier respectively.

[0145] Then, step S4 is performed to reallocate the acoustic signal diagnosis result and the vibration signal diagnosis result by using the conflict ratio redistribution rule on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, so as to obtain a fusion diagnosis result. Figure 4 As shown, it may further include steps S41 to S44.

[0146] Step S41: Use the acoustic signal diagnosis results and the vibration signal diagnosis results as evidence bodies to establish a generalized recognition framework. The recognition framework in this solution is a non-empty finite set composed of complete elements.

[0147] θ={θ 1 ,θ 2 , ...θ n}, and it should be noted that in this scheme, the acoustic signal diagnosis result and the vibration signal diagnosis result are used as two evidence bodies to establish the recognition framework Θ.

[0148] Step S42: According to the focal element in the generalized identification framework, a super power set is obtained by performing intersection and union operations. A super power set is generated according to Θ. Super power set D θ It is defined as the set of all elements in the generalized identification framework Θ and the new elements generated by the corresponding combination of ∩ and ∪.

[0149] φ,θ 1 ,θ 2 , ...θ n ∈D θ , if θ 1 ,θ 2 ,∈D θ , then θ 1 ∪θ 2 ,∈D θ .

[0150] Step S43: According to the generalized identification framework and the hyperpower set, a credibility distribution function on the generalized identification framework is obtained. Assume that in the identification framework Θ, there is a set of mappings satisfy:

[0151] m(φ)=0 and Then m(A) is called the credibility distribution function of A in the generalized identification framework Θ.

[0152] Step S44: Based on the DSmT fusion rule and the reliability distribution function, a fusion diagnosis result is obtained. The fusion rule is as follows:

[0153] Assume that m1(·), m2(·), ..., mk(·) represent the generalized credibility distribution functions provided by k different independent information sources, then the fusion rule for k independent information sources is:

[0154]

[0155] As can be seen from formula (11), the DSmT theory adopted in this scheme calculates the corresponding conflict information in the identification framework in principle, making the credibility distribution of each piece of evidence more reasonable. The above content in this scheme can solve the credibility distribution problem of conflict information, but due to the introduction of intersection operation in the fusion decision, the fusion result classification is numerous and the fusion judgment becomes complicated. In order to simplify the fusion decision operation, some singular results need to be redistributed. It can be considered that the generation of conflict information originates from the single focal element with a clear decision in the identification framework, and the size of the conflict is proportional to its own confidence assignment. Therefore, the conflict information after each combination can be redistributed according to the confidence assignment of the single focal element. The PCR5 conflict redistribution formula for two elements is:

[0156]

[0157] Where: Y∈D θ / {φ}From formula (10), it can be seen that the conflict reliability between X and Y consists of two parts: the conflict between X in source 1 and Y in source 2, and the conflict between Y in source 1 and X in source 2. PCR5 distributes these two parts of conflict reliability to the combined reliability of X and Y according to the proportional relationship between the original reliability values ​​of X and Y. In this scheme, source 1 and source 2 represent the acoustic signal and vibration signal, respectively. X and Y represent the acoustic signal diagnosis result and vibration signal diagnosis result generated according to the acoustic signal and vibration signal, respectively. Finally, a fusion diagnosis result is obtained, and the result is generated quickly.

[0158] See also Figure 5 As shown, the present solution also provides a rolling bearing acoustic and vibration signal fault diagnosis system, including a signal acquisition module 11, a sensitive feature separation module 12, a preliminary diagnosis module 13 and a fusion decision module 14. The signal acquisition module 11 is used to collect the acoustic signal and vibration signal of the rolling bearing in the working state. The sensitive feature separation module 12 is used to extract the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal respectively. According to the acoustic sensitive feature type and the vibration sensitive feature type, the acoustic sensitive features sensitive to the acoustic fault signal are separated from the acoustic signal, and the vibration sensitive features sensitive to the vibration fault signal are separated from the vibration signal. The preliminary diagnosis module 13 is used to input the acoustic sensitive features into the trained acoustic fault classifier to obtain the acoustic signal diagnosis result, and input the vibration sensitive features into the trained vibration fault classifier to obtain the vibration signal diagnosis result. The acoustic sensitive features and the vibration sensitive features are used to perform fault diagnosis, and the acoustic signal diagnosis result and the vibration signal diagnosis result are obtained respectively. The fusion decision module 14 is used to reallocate the acoustic signal diagnosis result and the vibration signal diagnosis result by adopting a conflict ratio redistribution rule on the basis of retaining the conflict information of the acoustic signal diagnosis result and the vibration signal diagnosis result, so as to obtain a fusion diagnosis result.

[0159] See also Figure 6As shown, the present solution also provides a rolling bearing acoustic and vibration signal fault diagnosis device, including a test machine 21, a soundproof cover 22, sound sensors 231, 232, vibration sensors 241, 242 and a processing unit 25. The test machine 21 is used to load and test the rolling bearing. The test machine covers the soundproof cover 22, which can be made of a high-density board with sound-absorbing material attached to the inside. The sound sensors 231, 232 can be composed of a first sound sensor 231 and a second sound sensor 232 to form a dual channel, fixed inside the soundproof cover 22, and the sound sensors 231, 232 are used to collect the acoustic signal of the rolling bearing. The vibration signal of the rolling bearing is collected by the vibration sensors 241, 242, and the vibration sensors 241, 242 are composed of a unidirectional acceleration vibration sensor 241 and two three-directional acceleration vibration sensors 242. After receiving the acoustic signal and the vibration signal, the processing unit 25 performs the above-mentioned rolling bearing acoustic and vibration signal fault diagnosis method to obtain a fusion diagnosis result.

[0160] In order to obtain different fault type data of rolling bearings, different damage points of rolling bearings are processed. The inner ring is processed by laser, and the outer ring and roller are processed by wire cutting. According to the different damage locations and damage degrees of rolling bearings, the rolling bearing test data are divided into 9 types of fault types, and category labels of different fault types are constructed. The specific parameter descriptions of the 9 types of faults are listed in Table 1. Other fault types can also be set. For example, I have a single point on the outer ring, and the outer ring multi-point setting is two faults on the outer ring, and the outer ring multi-point setting is three faults on the outer ring. These belong to the category of multi-point faults of the outer ring and are all within the protection scope of this solution.

[0161] Table 1 Rolling bearing fault parameters

[0162]

[0163] This test collects vibration signals and acoustic signals of different fault types, with a sampling frequency of 20480Hz. According to the rolling bearing speed and applied radial load, the acoustic signals and vibration signals of the rolling bearing are collected under the load of 3kn and 4kn, the speed is 2000r / min, 3000r / min, 4000r / min stable speed and 2000~4000~2000r / min lifting speed. The description of each working condition is listed in Table 2. 100 samples of acoustic signals and vibration signals of each fault type under each working condition are collected, and the length of each sample is 1024 sampling points.

[0164] Table 2 Operating parameters

[0165]

[0166] Multiple samples of acoustic and vibration signals under various fault types and working conditions are obtained, and variational modal decomposition is performed on each sample to obtain the intrinsic mode function of each sample. The time domain characteristics and nonlinear characteristics of all intrinsic mode functions are obtained, and the time domain characteristics of each intrinsic mode function are: average value, variance, peak value, kurtosis, root mean square, crest factor, margin factor, pulse coefficient, shape coefficient, and skewness; each intrinsic mode function is reconstructed in phase space, and the embedding dimension and time delay are determined according to the mutual information method and the nearest neighbor method respectively. The nonlinear characteristics of each intrinsic mode function are calculated in the reconstructed phase space: maximum Lyapunov index (Lyapunov index), correlation dimension, approximate entropy, fuzzy entropy and sample entropy. The feature set matrices of all time domain features and nonlinear features are constructed respectively. All feature set matrices are reduced to two dimensions to obtain two-dimensional feature set matrices. The average contour coefficients of all two-dimensional feature set matrices are calculated respectively. If the average contour coefficient is greater than 0, they are marked as acoustic sensitive features and vibration sensitive features respectively.

[0167] Two-dimensional clustering of some feature sets such as Figure 7 , 8 , 9, and 10. It can be seen that the sample entropy of each fault type has good discrimination and can be used as a sensitive feature for fault diagnosis classification. The skewness does not have good discrimination and is used as a non-sensitive feature.

[0168] The average silhouette coefficient (ASC) of each feature set matrix after dimensionality reduction is calculated, as shown in Tables 3 and 4. Features with ASC greater than 0 are sensitive features, and features with ASC less than 0 are non-sensitive features.

[0169] Table 3. Average silhouette coefficient of vibration signal feature set

[0170]

[0171]

[0172] Table 4. Average silhouette coefficient of acoustic signal feature set

[0173]

[0174] The vibration sensitive feature types include: Li index, approximate entropy, sample entropy, fuzzy entropy, mean, variance, peak, kurtosis, root mean square, crest factor, margin factor, pulse coefficient, shape coefficient, a total of 13. The non-sensitive features are correlation dimension and skewness.

[0175] The acoustic sensitive feature types include: approximate entropy, sample entropy, fuzzy entropy, mean value, variance, peak value, kurtosis, root mean square, margin factor, pulse coefficient, shape coefficient, a total of 11. The non-sensitive features are: Li index, correlation dimension, crest factor, skewness. In the process of fault diagnosis, the acoustic sensitive features sensitive to the acoustic fault signal are separated from the acoustic signal according to the acoustic sensitive feature type and the vibration sensitive feature type, and the vibration sensitive features sensitive to the vibration fault signal are separated from the vibration signal for fault diagnosis.

[0176] See also Figures 1 to 4 As shown in , 11 to 18, the acoustic signal dataset and the vibration signal dataset are constructed respectively with the acoustic sensitive features and the vibration sensitive features, and are divided into training set, validation set and test set. The format of the dataset of each working condition of the vibration signal is 900×(8×13), and the format of the dataset of each working condition of the acoustic signal is 900×(8×11), based on the support vector machine fault diagnosis classification.

[0177] In the first group, 90% of the data sets of working conditions 1, 2, 3, and 4 are randomly selected as training sets, and 10% are used as test sets for fault diagnosis using 5-fold cross validation (this paper uses 5-fold cross validation).

[0178] The second group uses the data sets of working conditions 1, 2, 3, and 4 under a load of 3KN as the training set, and the data sets of working conditions 5, 6, 7, and 8 under a load of 4KN as the test set for cross-load fault diagnosis.

[0179] The third group uses the data sets of working conditions 5, 6, and 7 at steady speed as the training set, and the data set of working condition 8 with speed increase or decrease as the test set for fault diagnosis.

[0180] The fourth group uses the data sets of steady speed conditions 5, 6, and 7 as training sets, and the data set of speed-up and speed-down condition 4 across loads as testing sets for fault diagnosis.

[0181] The confusion matrix of the acoustic vibration signal diagnosis result is as follows: Figures 11 to 18 shown.

[0182] like Figure 4 As shown, step S4, on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, convert the primary acoustic and vibration diagnosis result into the generalized confidence distribution of each fault mode to obtain the fused diagnosis result. The diagnosis accuracy of each type of fault after fusion is shown in Table 5.

[0183] Table 5. Fusion decision diagnostic accuracy

[0184]

[0185] The fault diagnosis classification accuracy of the four groups is shown in Table 6. The training set and test set of the first group have the same working conditions, the training set and test set of the second group have different working conditions with different loads, the training set and test set of the third group have steady-state speed and lifting speed respectively, and the training set has steady-state speed. The test set has lifting speed. The diagnostic accuracy of the second, third and fourth groups under different loads, different speeds and different loads and speeds after acoustic-vibration fusion all reached more than 99%, with high diagnostic accuracy. Among them, the diagnostic accuracy of the acoustic signal is more than 90%, and its diagnostic accuracy is slightly lower than that of the vibration signal because the acoustic signal is interfered by environmental noise.

[0186] Table 6. Fault diagnosis accuracy of four groups

[0187]

[0188] See also Figures 1 to 4 , 19, and 20, in order to verify that the tSNE-ASC feature selection method can effectively distinguish sensitive features from non-sensitive features. The non-sensitive feature data sets under working conditions 1, 2, 3, and 4 were used for comparative tests, and fault diagnosis was implemented based on the support vector machine. 90% of the data sets were randomly used as training sets and 10% as test sets. The confusion matrix of the fault diagnosis results of the non-sensitive features of the acoustic vibration signal is shown in Fig.19 and Fig. 20 As shown in Figure 2, the accuracy of acoustic vibration diagnosis is only 81.9% and 62.5%. The comparative experiment proves the effectiveness of the tSNE-ASC feature selection method in this scheme.

[0189] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A rolling bearing acoustic vibration signal fault diagnosis method, characterized in that: The method comprises: Collect the acoustic and vibration signals of rolling bearings in working condition; Respectively extracting the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal; Inputting the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputting the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result; On the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, adopting a conflict proportion redistribution rule to redistribute the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fusion diagnosis result; The step of reallocating the acoustic signal diagnosis result and the vibration signal diagnosis result by using a conflict ratio redistribution rule on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fusion diagnosis result comprises: The acoustic signal diagnosis result and the vibration signal diagnosis result are used as evidence bodies to establish a generalized identification framework; According to the focal element in the generalized identification framework, a super power set is obtained by performing intersection and union operations; According to the generalized identification framework and the super power set, a credibility distribution function on the generalized identification framework is obtained; Based on the DSmT fusion rule and the reliability distribution function, a fusion diagnosis result reliability distribution function is obtained.

2. The method according to claim 1, characterized in that: The acoustic sensitive features are selected in the following manner: Obtaining acoustic signal samples of the rolling bearing under different fault types and different operating conditions; Performing variational mode decomposition on each of the acoustic signal samples to obtain an intrinsic mode function of each of the acoustic signal samples; Extracting acoustic signal features from each of the intrinsic mode functions, wherein the signal features include time domain features and nonlinear features; Constructing a feature set matrix for each of the acoustic signal features; Reducing the dimension of each of the feature set matrices to two dimensions respectively, to obtain a two-dimensional feature set matrix of each of the feature set matrices; Calculating the average silhouette coefficient of each of the two-dimensional feature set matrices, and selecting features that are sensitive to acoustic fault signals as acoustic sensitive features according to the average silhouette coefficient; The vibration sensitivity feature is selected in the following manner: Obtaining vibration signal samples of the rolling bearing under different fault types and different working conditions; Performing variational mode decomposition on each of the vibration signal samples to obtain an intrinsic mode function of each of the vibration signal samples; Extracting vibration signal features from each of the intrinsic mode functions, wherein the signal features include time domain features and nonlinear features; Constructing a feature set matrix for each of the vibration signal features; Reducing the dimension of each of the feature set matrices to two dimensions respectively, to obtain a two-dimensional feature set matrix of each of the feature set matrices; The average silhouette coefficient of each of the two-dimensional feature set matrices is calculated, and features that are sensitive to vibration fault signals are screened out as vibration sensitive features based on the average silhouette coefficient.

3. The method according to claim 2, characterized in that The step of reducing the dimensions of all the feature set matrices to two dimensions to obtain a two-dimensional feature set matrix includes using a t-distributed random neighbor embedding method to reduce the dimensions of all the feature set matrices to two dimensions to obtain a two-dimensional feature set matrix.

4. The method according to claim 2, characterized in that: The acoustic sensitive features include one or more combinations of approximate entropy, sample entropy, fuzzy entropy, mean value, variance, peak value, kurtosis, root mean square, margin factor, pulse coefficient and shape coefficient.

5. The method according to claim 2, characterized in that: The vibration sensitivity feature types include one or more combinations of Li index, approximate entropy, sample entropy, fuzzy entropy, mean value, variance, peak value, kurtosis, root mean square, crest factor, margin factor, pulse coefficient and shape coefficient.

6. The method according to claim 1, characterized in that The method further includes the steps of performing fault diagnosis on the acoustic sensitive feature and the vibration sensitive feature to obtain acoustic signal diagnosis results and vibration signal diagnosis results respectively, including: Obtaining acoustic signal samples and vibration signal samples of the rolling bearing under different fault types and different working conditions; Extracting the acoustic sensitive features of each of the acoustic signal samples to construct an acoustic signal data set, and extracting the vibration sensitive features of each of the vibration signal samples to construct a vibration signal data set; The acoustic fault classifier and the vibration fault classifier are trained using the acoustic signal data set and the vibration signal data set respectively.

7. The method according to any one of claims 2 to 6, characterized in that: The fault types include a combination of one or more of normal bearing, outer ring single point fault, inner ring single point fault, roller single point fault, outer ring roller composite fault, inner ring roller composite fault, outer ring multi-point fault, inner ring multi-point fault and roller multi-point fault.

8. A rolling bearing acoustic vibration signal fault diagnosis system, characterized in that: The system comprises: A signal acquisition module is used to collect acoustic signals and vibration signals of the rolling bearing in working state; A sensitive feature separation module, used to extract the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal respectively; A preliminary diagnosis module, used for inputting the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputting the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result; A fusion decision module, for reallocating the acoustic signal diagnosis result and the vibration signal diagnosis result by using a conflict ratio redistribution rule on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, so as to obtain a fusion diagnosis result; The step of reallocating the acoustic signal diagnosis result and the vibration signal diagnosis result by using a conflict ratio redistribution rule on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fusion diagnosis result comprises: The acoustic signal diagnosis result and the vibration signal diagnosis result are used as evidence bodies to establish a generalized identification framework; According to the focal element in the generalized identification framework, a super power set is obtained by performing intersection and union operations; According to the generalized identification framework and the super power set, a credibility distribution function on the generalized identification framework is obtained; Based on the DSmT fusion rule and the reliability distribution function, a fusion diagnosis result reliability distribution function is obtained.

9. A rolling bearing acoustic vibration signal fault diagnosis device, characterized in that: The device comprises: Testing machines for loading rolling bearings; A soundproof enclosure, wrapping the testing machine; Sound sensor, used to collect sound signals of rolling bearing in working state; Vibration sensor, used to collect vibration signals of rolling bearings in working state; A processing unit is used to receive the sound signal and the vibration signal, and perform the following processing: Respectively extracting the acoustic sensitive features in the acoustic signal and the vibration sensitive features in the vibration signal; Inputting the acoustic sensitive features into a trained acoustic fault classifier to obtain an acoustic signal diagnosis result, and inputting the vibration sensitive features into a trained vibration fault classifier to obtain a vibration signal diagnosis result; On the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result, adopting a conflict proportion redistribution rule to redistribute the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fusion diagnosis result; The step of reallocating the acoustic signal diagnosis result and the vibration signal diagnosis result by using a conflict ratio redistribution rule on the basis of retaining the conflict information between the acoustic signal diagnosis result and the vibration signal diagnosis result to obtain a fusion diagnosis result comprises: The acoustic signal diagnosis result and the vibration signal diagnosis result are used as evidence bodies to establish a generalized identification framework; According to the focal element in the generalized identification framework, a super power set is obtained by performing intersection and union operations; According to the generalized identification framework and the super power set, a credibility distribution function on the generalized identification framework is obtained; Based on the DSmT fusion rule and the reliability distribution function, a fusion diagnosis result reliability distribution function is obtained.

Images