Intelligent fault diagnosis method for motor based on acoustic signal

Abstract

The application discloses a motor intelligent fault diagnosis method based on acoustic signals, comprising the following steps: collecting acoustic signal data of an induction motor and performing data preprocessing, dividing a single sample under a unified time length, and defining a label by a motor state; calculating information entropy features of each sample under multiple time scales and in other domains, and constructing a multiple-input single-output data set; performing feature selection through a Relief_F algorithm to obtain multiple features most relevant to the output and the correlation weight of each feature; constructing a HANFIS, and according to the correlation weight of the features, circulating input of the high-quality features into the HANFIS, and through training and learning, obtaining a model for motor fault diagnosis; predicting the output from the model, rounding the output result, and obtaining the predicted motor state. The application is suitable for fault detection of an induction motor, and provides a feasible and accurate technical method for motor fault detection based on acoustic signals.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of motor fault diagnosis methods, and particularly relates to a motor intelligent fault diagnosis method based on sound signals. BACKGROUND

[0002] With the development of society and technology, the equipment in the production process is increasingly complex, and the fault diagnosis of the driving equipment in the production process is very critical. If the equipment failure cannot be found and handled in time, there is a huge safety hazard, resulting in economic losses. However, if the fault diagnosis technology can find and alarm in the early stage of failure, it can effectively prevent the occurrence of safety accidents. At present, the motor is the main driving equipment in the production process. Therefore, the fault diagnosis of the motor plays a crucial role in the modern production process.

[0003] At present, in order to realize the fault diagnosis of the motor, the following methods are usually used:

[0004] (1) The diagnosis method based on fault model needs to have a clear understanding of the principle and structure of the fault motor, and a high-precision fault diagnosis model is established through the processing and feature extraction of the signal. However, due to the difficulty in establishing a mathematical model of a modern complex system, and it is difficult to eliminate the noise and redundant information contained in the fault signal in the actual process, it is difficult to extract effective feature information from the complex abnormal fault signal. Therefore, the diagnosis method based on fault model generally has poor effect in the fault diagnosis of complex systems;

[0005] (2) The diagnosis method based on experience knowledge uses the rich experience and deep professional knowledge of experts in the relevant field to establish a rule-based expert system for fault diagnosis of the equipment. However, due to the difficulty in obtaining expert knowledge, it is usually suitable for simple equipment fault diagnosis, and is not suitable for complex system fault diagnosis;

[0006] (3) The diagnosis method based on data driving is based on the collected monitoring signal data, and combines various data mining techniques to obtain the deep feature information hidden therein, so as to judge whether the motor operation is in normal state or fault, so as to achieve the purpose of fault detection and diagnosis.

[0007] Due to the advantages of complete theory, low cost, etc., the current motor fault diagnosis technology based on data driving is mostly carried out by collecting vibration signals, but the fault diagnosis technology based on vibration signals also has some disadvantages: the vibration sensor needs to be rigidly connected to the motor, and it is difficult to install and implement in some complex conditions, such as high temperature, acidic environment, etc.; in addition, some machine learning methods used in the fault diagnosis technology based on data driving have the problem of dimension disaster caused by input increase, such as ANFIS. SUMMARY

[0008] To solve the above problems of the prior art, the application provides an intelligent motor fault diagnosis method based on acoustic signals, which can solve the problem of dimension disaster caused by increasing input, and has the advantage of non-contact diagnosis.

[0009] The purpose of the application is achieved by the following technical scheme: an intelligent motor fault diagnosis method based on acoustic signals, comprising the following steps:

[0010] (1) Collecting acoustic signal data of an induction motor and preprocessing the data, dividing single acoustic signal samples under a unified time length, and defining labels by motor state;

[0011] (2) Calculating information entropy features of each acoustic signal sample under multiple time scales and in other domains, including sample entropy, permutation entropy, fuzzy entropy, wavelet energy spectrum entropy and spectrum entropy under T time scales, a total of 3T+2 features, T being a preset parameter, and constructing a multiple-input single-output data set;

[0012] (3) Feature selection by Relief_F algorithm to obtain multiple acoustic entropy features most relevant to the output and the correlation weight of each feature;

[0013] (4) Constructing a multi-level adaptive neural fuzzy inference system HANFIS (Hierarchical-Adaptive-Network-Based Fuzzy Inference System), selecting a number of features with the highest correlation weight to input HANFIS in a loop, and obtaining a model for motor fault diagnosis through training and learning;

[0014] (5) Predicting the output from the model, rounding the output result to obtain the predicted motor state.

[0015] The application can predict the state of the tested motor without contacting the motor, by extracting motor acoustic signal data, preprocessing and dividing the data, calculating information entropy features under multiple time scales and in other domains, then selecting features by Relief_F algorithm, and finally predicting by the improved HANFIS after training. In experimental verification, the technical method proposed by the application can accurately identify the motor fault type under different working conditions, providing an effective and feasible technical method for motor fault diagnosis based on acoustic signals.

[0016] Further, in step (1), in order to enrich the samples, when slicing and dividing the acoustic signal data of the induction motor, the overlap rate overlap is 50%, obtaining multiple acoustic signal samples with a time length of 1s. In order to facilitate subsequent data processing, different motor states are defined as discrete numbers when defining the output.

[0017] Furthermore, in step (2), the time series of a single acoustic signal sample is G = {g1,…,g N′}, N′ is the number of sampling points in a single acoustic signal sample, and N′ is determined by the sampling frequency of the acoustic sensor that collects the motor acoustic signal; the time scale transformation formula is as follows:

[0018]

[0019] Where τ is the time scale, τ=1,2,…,T,1≤j≤N,N=[N′ / τ]; the transformation sequence under the time scale τ is

[0020] Furthermore, the information entropy calculation in step (2) includes:

[0021] (2.1) The sample entropy calculation method is as follows:

[0022] Set the similarity tolerance r and reconstruction dimension m, and transform the change sequence {x1,…,x N}Reconstruct the phase space and obtain a new vector sequence X with dimension m m (1),…,X m (N-m+1), where vector X m (i) = {x i ,x i+1 ,…,x i+m-1},1≤i≤N-m+1;

[0023] Set two vectors X m (i) With X m (j) The distance d[X m (i),X m (k)] is the absolute value of the maximum difference between the two corresponding parts, that is:

[0024] d[X m (i),X m (j)]=max k＝0,…,m-1 (|x i+k -x j+k |)

[0025] Targeting X m (i), statistics X m (i) With X m (j) The number of j whose distance between them is less than or equal to r and is recorded as V i ;

[0026] Defining intermediate variables

[0027]

[0028] Define the intermediate variable V (m) (r) :

[0029]

[0030] Increase the dimension to m+1 dimension, and perform the same operation for X m+1 (i), count the number of j whose distance between X m+1 (i) and X m+1 (j) is less than or equal to r and record it as U i ;

[0031] Define the intermediate variable

[0032]

[0033] Define the intermediate variable U (m) (r) :

[0034]

[0035] Thus, the obtained V m (r) is the probability of pairing m points of two vector sequences under the similarity tolerance r, and U m (r) is the probability of pairing m+1 points of two vector sequences; then the sample entropy SampEn(m, r) is defined as follows:

[0036]

[0037] When N is a finite value, the sample entropy estimation is:

[0038] SampEn(m, r, N) = -ln[U m (r) / V m (r)]

[0039] The permutation entropy calculation method is as follows:

[0040] Set the reconstruction dimension m and the time delay L, and perform phase space reconstruction on the variation sequence {x1, …, x N} under a certain time scale τ to obtain a new vector sequence X m (1), …, X m (N-m+1) with a dimension of m, where the vector X m (i) = {x i , x i+L , …, x i+(m-1)L}, 1≤i≤N-m+1;

[0041] Perform incremental ordering on each vector X m (i) internally, that is If there are two equal values, then i Sort the subscripts; eventually, an X m (i) is mapped to (j1,j2,…,j m ), there are m! permutations, that is, for each vector X m (i) are mapped to one of the m! permutations;

[0042] Use P1, P2, ..., P K represents the probability distribution of all permutations, where K≤m!;

[0043] The definition of permutation entropy H(m) is as follows:

[0044]

[0045] Finally, normalization is performed, and the formula is:

[0046] 0≤H(m) / ln(m!)≤1

[0047] (2.3) The fuzzy entropy calculation method is as follows:

[0048] Set the reconstruction dimension m, the gradient p and width b of the exponential function boundary; transform the change sequence {x1,…,x N}Reconstruct the phase space and obtain a new vector sequence X with dimension m m (1),…,X m (N-m+1), where X m (i) = {x i ,x i+1 ,…,x i+m-1}-x0(i), 1≤i≤N-m+1, x0(i) represents the mean of m consecutive x, as shown in the following formula:

[0049]

[0050] Set two vectors X m (i) With X m (j) The distance between is the absolute value of the maximum difference between the two corresponding parts, that is:

[0051]

[0052] Define two vectors X using fuzzy functions m (i) With X m (j) Similarity between Right now:

[0053]

[0054] Define the intermediate variable O m :

[0055]

[0056] Increase the dimension to m+1 dimensions, and perform the same operation to obtain the intermediate variable O m+1 :

[0057]

[0058] The definition of FuzzyEn is as follows:

[0059]

[0060] When N is a finite value, the FuzzyEn estimation is:

[0061] FuzzyEn = ln O m - ln O m+1

[0062] The wavelet energy spectrum entropy calculation method is as follows:

[0063] Through wavelet transform, the wavelet energy spectrum of the time series of a single acoustic signal sample on α components is obtained, denoted as E1, E2, …, E α ; E is the sum of the energy of each component; the energy proportion p j of the jth component = E j / E, then the wavelet energy spectrum entropy W EE is defined as:

[0064]

[0065] The spectrum entropy calculation method is as follows:

[0066] Through Fourier transform, the energy spectrum of the time series of a single acoustic signal sample on β frequency components is obtained, denoted as E1, E2, …, E β ; E is the sum of the energy of each frequency component; the energy proportion p j of the jth frequency component = E j / E, then the spectrum entropy SE is defined as:

[0067] SE = -∑ j p j log p j .

[0068] Further, in step (3), feature selection is performed by the Relief_F algorithm to obtain a plurality of sound entropy features most relevant to the output and the relevant weight of each feature, and the selected sound entropy features are arranged in descending order of weight and recorded as F1, F2, …, F n The specific implementation process of the Relief_F algorithm is as follows:

[0069] According to the output result of step (2), a training data set D is constructed, each sample in D is a combination of 3T+2 features and an output label corresponding to a certain sound signal sample; the sample sampling times are e;

[0070] Initially, set all feature weights to 0;

[0071] Randomly select a sample R from D; find k nearest neighbors H j from the same sample set of R j (C), where j = 1, 2, …, k; the weight of each feature I is calculated in turn, and the formula is as follows:

[0072]

[0073] In the formula, class(R) represents the class of R, p(C) is the sample number ratio of class C in all classes; p(class(R)) is the sample number ratio of class(R) in all classes; diff(I, R1, R2) represents the difference between samples R1 and R2 in feature I, as follows:

[0074]

[0075] Repeat the above steps e times to obtain the weight W(I) of each feature I, sort all features according to the weight w(I), and select the top n features to construct a feature set F = F1, F2, …, F n , n≥6.

[0076] SE = -∑ j p j logp j .

[0077] Further, in step (3), feature selection is performed by the Relief_F algorithm to obtain a plurality of sound entropy features most relevant to the output and the relevant weight of each feature, and the selected sound entropy features are arranged in descending order of weight and recorded as F1, F2, …, F n ; The specific implementation process of the Relief_F algorithm is as follows:

[0078] In step (4), HANFIS is applied to fault diagnosis, specifically:

[0079] (4.1) Constructing an adaptive neuro-fuzzy inference system ANFIS (Adaptive-Network-Based Fuzzy Inference System) as a subsystem, which is a five-layer structure;

[0080] For acoustic entropy features F u ,F v Two-input ANFIS, the system function can be described by the following two inference rules:

[0081] Rule 1: If F u is A1 and F v is B1,then f1=p1F u +q1F v +r1;

[0082] Rule 2: If F u is A2 and F v is B2,then f2=p2F u +q2F v +r2;

[0083] Where f1,f2 are the output of each inference rule, A1A2, B1, B2 are the language fuzzy labels in the fuzzy set, p1, p2, q1, q2, r1, r2 are the rule parameters;

[0084] The first layer fuzzifies the input acoustic entropy features through the membership function to get the membership degree in the interval [0,1], the formula is as follows:

[0085] O 1i =μ Ai (F u ),i=1,2

[0086] O 2i =μ Bi (F v ),i=1,2

[0087] Where μ Ai , μ Bi is the membership function, O 1i , O 2i is the membership degree of the fuzzy set, the membership function adopts Gaussian function, and the parameters containing fuzzy information in the Gaussian membership function are determined by model training;

[0088] The second layer calculates the trigger strength w i , that is, the membership degrees of all features are multiplied, the formula is as follows:

[0089] w i =μAi (F u )×μ Bi (F v )

[0090] The third layer calculates the triggering proportion of the i-th rule in the entire rule base That is, the triggering strength w of the i-th rule obtained by the second layer i Normalization is performed to reflect the probability of using each rule in the entire reasoning process, and the formula is as follows:

[0091]

[0092] The fourth layer performs de-fuzzification, that is, the fuzzy output result of each rule is calculated through the linear combination of the input acoustic entropy features, and the formula is as follows:

[0093]

[0094] The fifth layer calculates the output f of the system, that is, the sum of the fuzzy output results of all rules, and the formula is as follows:

[0095]

[0096] Based on the two-input construction process of ANFIS, a three-input structure subsystem is obtained by analogy as a basic unit of cascade;

[0097] (4.2) Constructing a multi-level adaptive neural fuzzy inference system HANFIS;

[0098] The output of the current level h subsystem ANFIS_h and the selected acoustic entropy features are used as the input of the next level h+1 subsystem ANFIS_h+1, and the cascade is continued in this way to form HANFIS;

[0099] Further, in step (4.2), the number of inputs of each ANFIS subsystem is 3 by default; when the number of inputs of ANFIS is too large, it will cause the curse of dimensionality and cannot be calculated; when the number of inputs is too small, the efficiency is low; the number of inputs of the last ANFIS can be flexibly selected to be 2, 3 or 4.

[0100] Further, in step (4.2), the acoustic entropy features are selected in order from the front to the back without repetition according to the feature set F=F1, F2, …, F n in the feature set, until the last feature is input into the last subsystem, so that the model training process is more easily converged.

[0101] Further, in step (4), the feature set F=F1, F2, …, F nThe features of the middle and front [(n+1) / 3] are input into the HANFIS in cycles, and in the deepening process of the HANFIS structure, the upstream feature information forgetting problem occurs, the diagnosis accuracy tends to be saturated or even decreased, so that the information forgetting problem can be solved, and the diagnosis accuracy can be improved.

[0102] Further, in step (5), the model prediction output y is rounded, and the predicted motor state is determined according to the defined motor state label.

[0103]

[0104] In the formula, z is the number of motor states.

[0105] Compared with the prior art, the present application has the following beneficial effects:

[0106] The present application has the advantages of non-contact for fault diagnosis of the motor sound signal data, and has a wider application scenario; inherits the advantages of ANFIS, such as learning mechanism and language reasoning ability of fuzzy system, solves the dimension disaster problem caused by the increase of ANFIS input, and further improves the HANFIS, and alleviates the feature information forgetting problem caused by the increase of system depth. An effective and feasible technical method for motor fault diagnosis based on sound signal is provided. BRIEF DESCRIPTION OF DRAWINGS

[0107] Figure 1 A flowchart of a motor intelligent fault diagnosis method based on sound signals is provided for the embodiment of the present application.

[0108] Figure 2 A simple ANFIS structure diagram is provided for the embodiment of the present application.

[0109] Figure 3 A HANFIS structure diagram is provided for the embodiment of the present application.

[0110] Figure 4 A HANFIS structure diagram of n=8, 4 subsystems is provided for the embodiment of the present application.

[0111] Figure 5 An improved HANFIS structure diagram is provided for the embodiment of the present application.

[0112] Figure 6 A model prediction test set output result precision diagram is provided for the embodiment of the present application. DETAILED DESCRIPTION

[0113] In order to more specifically describe the present application, the technical solutions of the present application are described in detail below in combination with the drawings and specific embodiments.

[0114] The embodiment trains and learns the model of the same type of induction motor in 8 states at 3500r / min, and realizes fault diagnosis.

[0115] As shown in Figure 1 The embodiment provides a motor intelligent fault diagnosis method based on acoustic signals, and the specific process is as follows:

[0116] S01, a sound sensor is arranged near the motor to extract motor acoustic signal data. In the embodiment, the sound sensor is arranged at a distance of 0.6m-1m from the motor. In order to enrich the samples, when the acoustic signal data of the induction motor is divided, the overlap is 50%, and 4800 acoustic signal data samples with a time length of 1s are obtained. In order to facilitate subsequent data processing, when the output is defined, different motor states are defined as discrete numbers. 3800 data are randomly extracted as a training set, and the remaining 1000 data are used as a test set. The definition of each motor state is shown in Table 1:

[0117] Table 1

[0118] Motor condition Label Normal motor 1 Bearing inner ring fault motor 2 Bearing outer ring fault motor 3 Bearing roller fault motor 4 Broken bar fault motor 5 Rotor unbalance motor 6 Rotor misalignment motor 7 Inter-turn short circuit fault motor 8

[0119] S02, feature extraction, calculating sample entropy, permutation entropy, fuzzy entropy, wavelet energy spectrum entropy and frequency spectrum entropy of the training set in 20 time scales, a total of 62 features.

[0120] S03, feature selection is performed through the Relief_F algorithm, and a plurality of acoustic entropy features most related to the output and the correlation weight of each feature are obtained. The weight output by the Relief_F algorithm ranges from-1 to 1, and a larger positive weight is assigned to important features.

[0121] The correlation weight of each feature is shown in Table 2:

[0122] Table 2

[0123]

[0124]

[0125]

[0126]

[0127] According to the feature weight, the first 8 features most related to the output are selected, and the 2nd, 1st, 3rd, 5th, 32nd, 62nd, 11th and 8th features, i.e., 1 time scale permutation entropy, 1 time scale sample entropy, 1 time scale fuzzy entropy, 2 time scale permutation entropy, 11 time scale permutation entropy, spectral entropy, 4 time scale permutation entropy and 3 time scale permutation entropy, are defined as features F1, F2, F3, F4, F5, F6, F7 and F8 in sequence. A multi-input single-output training set is constructed, and the 8 features of the test set are calculated.

[0128] S04, first, an ANFIS is constructed, as shown in Figure 2 The first layer fuzzifies the input acoustic entropy features through a membership function. Specifically, a Gaussian membership function is used to fuzzify the input features to obtain a membership degree in the interval [0, 1].

[0129] The second layer calculates the triggering strength, i.e., the product of the membership degrees of all features.

[0130] The third layer calculates the triggering proportion of each rule in the entire rule base, i.e., the triggering strength of each rule obtained by the second layer is normalized to reflect the probability of using each rule in the entire inference process.

[0131] The fourth layer de-fuzzifies, i.e., calculates the fuzzy output result of each rule through a linear combination of the input acoustic entropy features.

[0132] The fifth layer calculates the output of the system, i.e., the sum of the fuzzy output results of all rules.

[0133] The ANFIS, as a subsystem, takes the output of the current level subsystem and the selected two acoustic entropy features as inputs of the next level subsystem, and cascades in this way to form a HANFIS, as shown in Figure 3 、 4 To avoid forgetting of feature information due to increase in system depth, the top 3 features according to the feature correlation weight are input into the HANFIS in a cycle, as shown in Figure 5 The constructed network is trained on the training set, and the model is obtained after convergence.

[0134] S05, the test set is input into the trained model, and the output is rounded to obtain the predicted motor state in the test set. The model prediction accuracy is 99.4%, as shown in Figure 6 The final motor fault diagnosis model based on acoustic signals is obtained.

[0135] The above embodiments describe the technical solutions and advantages of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the present application. Any modification, supplement and equivalent replacement made within the principle range of the present application shall be included in the protection range of the present application.

Claims

1. A motor intelligent fault diagnosis method based on acoustic signals, characterized in that: The following steps are involved: (1) Collecting and preprocessing the acoustic signal data of the induction motor, dividing the individual acoustic signal samples with a uniform time length, and defining labels based on the motor state; (2) Calculate the information entropy characteristics of each acoustic signal sample at multiple time scales and in other domains, including sample entropy, permutation entropy, fuzzy entropy, and wavelet energy spectrum entropy and spectrum entropy at T time scales, a total of 3T+2 features, and construct a multi-input single-output data set; (3) Feature selection is performed using the Relief_F algorithm to obtain multiple acoustic entropy features that are most relevant to the output and the relevant weights of each feature; (4) Construct a multi-level adaptive neuro-fuzzy inference system HANFIS, select several features with the highest correlation weights and input them into HANFIS in a loop, and obtain a model for motor fault diagnosis through training and learning; (5) The model predicts the output, and the output result is rounded to obtain the predicted motor state.

2. The motor intelligent fault diagnosis method based on acoustic signals according to claim 1 is characterized in that: In step (1), when slicing the acoustic signal data of the induction motor, the overlap rate is 50%, and multiple acoustic signal samples with a time length of 1 second are obtained; when defining the output, different motor states are defined as discrete numbers.

3. The motor intelligent fault diagnosis method based on acoustic signals according to claim 1 is characterized in that: In step (2), the time series of a single acoustic signal sample G = {g1,…,g N′ }, N′ is the number of sampling points in a single acoustic signal sample, and N′ is determined by the sampling frequency of the acoustic sensor; the time scale transformation formula is as follows: Where τ is the time scale, τ=1,2,…,T,1≤j≤N,N=[N ′ / τ]; transformation sequence under time scale τ 4. The motor intelligent fault diagnosis method based on acoustic signals according to claim 3 is characterized in that: The calculation of information entropy in step (2) includes: (2.1) The sample entropy calculation method is as follows: Set the similarity tolerance r and reconstruction dimension m, and transform the change sequence {x1,…,x N }Reconstruct the phase space and obtain a new vector sequence X with dimension m m (1),…,X m (N-m+1), where vector X m (i) = {x i ,x i+1 ,…,x i+-1 },1≤i≤N-m+1; Set two vectors X m (i) With X m (j) The distance d[X m (i),X m (j)] is the absolute value of the maximum difference between the two corresponding parts, that is: d[X m (i),X m (j)]=max k＝0,…,m-1 (|x i+k -x j+k |) Targeting X m (i), statistics X m (i) With X m (j) The number of j whose distance between them is less than or equal to r and is recorded as V i ; Defining intermediate variables Define the intermediate variable V (m) (r): Increase the dimension to m+1 and perform the same operation for X m+1 (i), statistics X m+1 (i) With X m+1 (j) The number of j whose distance between them is less than or equal to r and is recorded as U i ; Defining intermediate variables Define the intermediate variable U (m) (r): Thus, the obtained V m (r) is the probability of pairing m points between two vector sequences under similarity tolerance r, and U m (r) is the probability of pairing m+1 points of two vector sequences; the sample entropy SampEn(m,r) is defined as follows: When N is finite, the sample entropy is estimated as: (2.2) The calculation method of permutation entropy is as follows: Set the reconstruction dimension m and time delay L, and transform the change sequence {x1,…,x N }Reconstruct the phase space and obtain a new vector sequence X with dimension m m (1),…,X m (N-m+1), where vector X m (i) = {x i ,x i+L ,…,x i+(m-1)L },1≤i≤N-m+1; For each vector X m (i) Perform ascending sorting internally, i.e. If there are two equal values, then i Sort the subscripts; eventually, an X m (i) is mapped to (j1,j2,…,j m ), there are m! permutations, that is, for each vector X m (i) are mapped to one of the m! permutations; Use P1, P2, ..., P K represents the probability distribution of all permutations, where K≤m!; The definition of permutation entropy H(m) is as follows: Finally, normalization is performed, and the formula is: 0≤H(m) / ln(m!)≤1 (2.3) The fuzzy entropy calculation method is as follows: Set the reconstruction dimension m, the gradient p and width b of the exponential function boundary; transform the change sequence {x1,…,x N }Reconstruct the phase space and obtain a new vector sequence X with dimension m m (1),…,X m (N-m+1), where X m (i) = {x i ,x i+1 ,…,x i+m-1 }-x0(i), 1≤i≤N-m+1, x0(i) represents the mean of m consecutive x, as shown in the following formula: Set two vectors X m (i) With X m (j) The distance between is the absolute value of the maximum difference between the two corresponding parts, that is: Define two vectors X using fuzzy functions m (i) With X m (j) Similarity between Right now: Define the intermediate variable O m : Increase the dimension to m+1 and perform the same operation to obtain the intermediate variable O m+1 : Fuzzy entropy FuzzyEn is defined as: When N is a finite value, the fuzzy entropy is estimated as: FuzzyEn=lnO m -lnO m+1 (2.4) The calculation method of wavelet energy spectrum entropy is as follows: Through wavelet transform, we can get the wavelet energy spectrum of the time series of a single acoustic signal sample on α components, which are recorded as E1, E2, ..., E α ; E is the sum of the energies of each component; the energy proportion of the jth component p j =E j / E, then the wavelet energy spectrum entropy W EE Defined as: (2.5) The spectrum entropy is calculated as follows: Through Fourier transform, the energy spectrum of the time series of a single sound signal sample at β frequency components is obtained, which is recorded as E1, E2, ..., E β ; E is the sum of the energy of each frequency component; the energy proportion of the jth frequency component p j =E j / E, the spectrum entropy SE is defined as: SE=-∑ j p j logp j 。 5. The motor intelligent fault diagnosis method based on acoustic signals according to claim 1, characterized in that: In step (3), the Relief_F algorithm is used to select features to obtain multiple acoustic entropy features that are most relevant to the output and the relevant weights of each feature. The selected acoustic entropy features are arranged in descending order according to weight and recorded as F1, F2, ..., F n ;The specific implementation process of the Relief_F algorithm is as follows: Construct a training data set D based on the output of step (2), where each sample in D is a combination of 3T+2 features and an output label corresponding to a certain sound signal sample; the number of sample sampling is e; Initially, all feature weights are set to 0; Randomly select a sample R from D; find the k nearest neighbors H from the same sample set of R j , find k nearest neighbors M from different types of sample sets in R j (C), where j = 1, 2, ..., k; calculate the weight of each feature I in turn, the formula is as follows: In the formula, class(R) represents the category of R, p(C) is the proportion of samples of category C in all categories; p(class(R)) is the proportion of samples of class(R) in all categories; diff(I,R1,R2) represents the difference between samples R1 and R2 on feature I, as shown in the following formula: Repeat the above steps e times to obtain the weight W(I) of each feature I, sort all features according to the weight W(I), and select the top n features to construct the feature set F=F1,F2,…,F n ,n≥6.

6. The motor intelligent fault diagnosis method based on acoustic signals according to claim 1, characterized in that: In step (4), HANFIS is applied to fault diagnosis, specifically: (4.1) Construct an adaptive neuro-fuzzy inference system ANFIS as a subsystem, which has a five-layer structure; For the acoustic entropy feature F u ,F v The system function of a two-input ANFIS can be described by the following two inference rules: Rule 1: IfF u isA1andF v isB1,thenf1=p1F u +q1F v +r1; Rule 2: IfF u isA2andF v isB2,thenf2=p2F u +q2F v +r2; Among them, f1, f2 are the outputs of each inference rule, A1, A2, B1, B2 are the linguistic fuzzy labels in the fuzzy set, and p1, p2, q1, q2, r1, r2 are the rule parameters; The first layer fuzzifies the input acoustic entropy features through the membership function to obtain the membership degree in the interval [0,1], as shown in the following formula: The 1i =μ Ai (F u ),i=1.2 The 2i =μ Bi (F v ),i=1.2 Among them, μ Ai ,μ Bi is the membership function, O 1i ,O 2i is the membership of the fuzzy set, the membership function adopts Gaussian function, and the parameters containing fuzzy information in the Gaussian membership function are determined by model training; The second layer calculates the trigger strength w i , that is, multiply the membership of all features, the formula is as follows: w i =μ Ai (F u )×μ Bi (F v ) The third layer calculates the triggering ratio of the i-th rule in the entire rule base That is, the trigger strength w of the i-th rule obtained in the second layer i Normalization is performed to reflect the probability of using each rule in the entire reasoning process. The formula is as follows: The fourth layer performs defuzzification, that is, the fuzzy output of each rule is calculated by the linear combination of the input acoustic entropy features. The formula is as follows: The fifth layer calculates the output f of the system, which is the sum of the fuzzy output results of all rules. The formula is as follows: Based on the two-input construction process of ANFIS, a three-input structural subsystem is obtained by analogy as the basic unit of cascade; (4.2) Construct a multi-level adaptive neuro-fuzzy inference system HANFIS; The output of the current level h subsystem ANFIS_h and the selected acoustic entropy features are used as the input of the next level h+1 subsystem ANFIS_h+1, and so on, to form HANFIS.

7. The motor intelligent fault diagnosis method based on acoustic signals according to claim 6, characterized in that: In step (4.2), the default number of inputs for each ANFIS subsystem is 3; the number of inputs for the last level ANFIS can be flexibly set to 2, 3, or 4.

8. The method for intelligent motor fault diagnosis based on acoustic signals according to claim 6, characterized in that: In step (4.2), the acoustic entropy features selected in step (3) are arranged in descending order according to weight to obtain a feature set F, and the acoustic entropy features are selected from the front to the back without repetition and input into the subsystem according to the feature order in the feature set F until the last feature is input into the last-level subsystem.

9. The motor intelligent fault diagnosis method based on acoustic signals according to claim 6, characterized in that: In step (4), the acoustic entropy features selected in step (3) are arranged in descending order by weight to obtain a feature set F. The number of features in the feature set F is n. The first [(n+1) / 3] features in the feature set F are selected and these features are cyclically input into HANFIS.

10. The motor intelligent fault diagnosis method based on acoustic signals according to claim 1, characterized in that: In step (5), the model prediction output y is rounded and the predicted motor state is determined based on the defined motor state label; Where z is the number of motor states.

Images