Rotary machinery fault diagnosis method based on improved multi-scale fuzzy entropy

Abstract

The invention discloses a rotating machine fault diagnosis method based on improved multi-scale fuzzy entropy, and belongs to the technical field of rotating machine fault diagnosis, and the method comprises the following steps: S1, EMD decomposition; s2, feature extraction; s3, feature dimension reduction; and S4, performing fault diagnosis. According to the method, the fuzzy entropy with the capability of processing data uncertainty and fuzziness is expanded to multiple time scales through an offset coarse graining method, and data can be used for multiple times, so that each window can capture more subtle changes and trends, mutation among different windows is reduced, and the accuracy of data processing is improved. The local features of the original signal are accurately reflected; a nonlinear dynamic model based on the improved multi-scale fuzzy entropy is established to quantify the time complexity, and a contrast experiment proves that the improved multi-scale fuzzy entropy has better robustness.

Description

Technical Field

[0001] The invention relates to the technical field of rotating machinery fault diagnosis, and in particular to a rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy. Background Art

[0002] Rotating machinery plays a key role in industrial production. If it fails, it will not only lead to production stagnation, but also may cause safety accidents. Therefore, timely and accurate detection and diagnosis of rotating machinery faults has become an important task to ensure equipment safety and improve production efficiency. Traditional fault diagnosis methods usually rely on experience and real physical models, but in complex working environments, they often fail to meet the needs of real-time and high efficiency because they rely on expert experience, lack of real-time performance, and high maintenance costs. For example, some researchers pointed out that the limitations of expert experience in complex fault conditions may cause some subtle faults to be ignored. Some researchers mentioned that fault diagnosis based on physical models often fails under dynamic and nonlinear conditions, resulting in poor diagnostic results. In addition, traditional fault diagnosis methods have high requirements on the quality of input data, and noise and abnormal data may significantly affect the results. At present, with the advancement of sensor technology, researchers can set up fault conditions in the laboratory according to actual production conditions to obtain more complete data resources. This has made data-driven methods gradually emerge, especially in feature extraction methods based on entropy.

[0003] As a tool to measure uncertainty and complexity, entropy is widely used in rotating machinery fault diagnosis. It can effectively extract fault feature information by analyzing the complexity of the signal. Other researchers have proposed a rotating machinery fault diagnosis method based on sample entropy. By performing sample entropy analysis on the fault signal, it can effectively distinguish between normal and faulty states. The results show that this method shows high accuracy in identifying early faults. Some researchers used improved composite multiscale approximate entropy to analyze the fault signal of rolling bearings and found that this entropy is highly sensitive to fault modes, can capture small signal changes, and significantly improves the accuracy of fault classification. Some researchers introduced fuzzy entropy as a feature extraction tool and combined it with support vector machine for fault diagnosis, achieving good results. Studies have shown that fuzzy entropy has strong robustness in processing noisy signals. Some researchers used multiscale entropy method to analyze fault signals, and the results showed that this method significantly improved the sensitivity of fault detection. Some researchers combined multiscale permutation entropy with deep learning, used multiscale permutation entropy to extract signal features, and performed fault classification through convolutional neural network. This method shows high classification accuracy when processing complex faults, especially in noisy environments, which is better than traditional methods.

[0004] However, the current diagnostic method based on entropy algorithm still has some problems that need to be overcome:

[0005] 1) In order to conduct research on multiple time scales, the commonly used coarse-graining method may lose local changes in the signal. When the scale factor is large, the original data length is drastically shortened. This will lead to problems such as the generated features are not continuous and smooth enough, the extracted features are biased, and some signal features are ignored.

[0006] 2) The commonly used multi-scale entropy is to take the average of the sequence entropy values ​​after coarsening with different scale factors as the result, which may lead to overfitting. If there are many similar patterns in the signal, averaging the entropy values ​​of each subsequence may introduce redundant information. This makes the model dependent on these redundant features and cannot capture the subtle changes and trends of the signal well, thus affecting its generalization ability.

[0007] The above problems need to be solved urgently. Therefore, a rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy is proposed. Summary of the invention

[0008] The technical problem to be solved by the present invention is: how to solve the problems existing in the above-mentioned diagnosis method based on entropy algorithm, and provide a rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy.

[0009] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0010] S1: EMD decomposition

[0011] Perform EMD decomposition on the time series samples of the original vibration signal of the rotating machinery and take the first three IMF components;

[0012] S2: Feature extraction

[0013] The improved multi-scale fuzzy entropy algorithm is used to extract the features of the three IMF components respectively to obtain the preliminary feature vectors.

[0014] S3: Feature Dimensionality Reduction

[0015] Use the LDA algorithm to reduce the dimension of the preliminary feature vector to obtain the feature vector after dimension reduction;

[0016] S4: Troubleshooting

[0017] Input the reduced-dimensional feature vector, classify it using SVM, and obtain the fault diagnosis result.

[0018] Furthermore, in step S1, the specific process of EMD decomposition is as follows:

[0019] S11: For the input sample x(t), the spline interpolation method is used to fit the maximum and minimum points of the original vibration signal to form the upper and lower envelopes; the average envelope x0(t) is calculated, and then the original vibration signal is subtracted from the mean to obtain a new signal h1(t):

[0020] h1(t)=x(t)-x0(t);

[0021] S12: Determine whether the new signal h1(t) satisfies the characteristic condition of IMF; if so, h1(t) is the first IMF component, i.e., imf1; if not, continue to loop through step S11 until h1(t) satisfies the characteristic condition of IMF;

[0022] S13: Separate imf1 from the original vibration signal to obtain the separated signal r1(t):

[0023] r1(t)=x(t)-imf1;

[0024] S14: Repeat the above steps until the first three IMF components are obtained.

[0025] Furthermore, in step S12, the characteristic conditions are specifically as follows:

[0026] In the entire data segment, the number of extreme points and the number of zero-crossing points are equal or the difference cannot exceed one;

[0027] At any time, the upper and lower envelopes are locally symmetric with respect to the time axis.

[0028] Furthermore, in step S2, the specific processing process of improving the multi-scale fuzzy entropy algorithm is as follows:

[0029] S21: Input the first three IMF components obtained and set the coarse-grained scaling factor sequence:

[0030] scales=(scale1,scale2,...,scale N ),1<scale i <N

[0031] Among them, scales is the coarse-grained scale;

[0032] S22: Set the offset factor k=1. When the scale factor is scale1, k is incremented from 1 until it is equal to scale1. The time series obtained by coarsening under each different offset factor k is calculated and its fuzzy entropy is calculated. Then the operation is repeated under the next scale factor.

[0033] S23: Combining all fuzzy entropies is the improved multi-scale fuzzy entropy, and outputting the entropy matrix That is, we get the preliminary feature vector.

[0034] Furthermore, in step S22, the calculation process of fuzzy entropy is as follows:

[0035] S221: Given an N-dimensional time series {u(i), 1≤i≤N};

[0036] S222: Construct an m-dimensional vector according to the following formula:

[0037]

[0038] Where i = 1, 2, ..., N-m + 1, It is the m consecutive values ​​of u starting from the i-th one minus the mean value u0(i),

[0039] S223: Definition and The distance between is the maximum value of the difference between the corresponding elements:

[0040]

[0041] Where, i=1,2,...,Nm,i≠j;

[0042] S224: Through the fuzzy function definition and Similarity

[0043]

[0044] Among them, the fuzzy function is an exponential function, n and r are the gradient and width of the blur function boundary respectively;

[0045] S225: Define the fuzzy similarity mean function as follows:

[0046]

[0047] Similarly, let the dimension be m+1, and repeat steps S221 to S224 to obtain:

[0048]

[0049] S226: Calculate fuzzy entropy:

[0050]

[0051] Furthermore, in step S3, the specific process of the LDA algorithm is as follows:

[0052] S31: Calculate the mean vector μ for each category in the preliminary feature vector i and the population mean vector μ;

[0053] S32: Calculate the intra-class scatter matrix S ω , global divergence matrix S t , and get the inter-class scatter matrix S b =S t -S ω ;

[0054] S33: Pair Matrix Perform eigenvalue decomposition and sort the eigenvalues ​​from large to small;

[0055] S34: Take the eigenvectors corresponding to the first p largest eigenvalues, and pass the eigenvalue x′=ω T x will n The dimensional sample is reduced to p dimensions.

[0056] Furthermore, in step S4, the specific processing process of SVM is as follows:

[0057] S41: Select penalty parameter C>0, construct and solve the convex quadratic programming problem:

[0058] satisfy Get the optimal solution

[0059] S42: Calculation Select α * A component of Satisfy the condition 0<α j <C, calculation

[0060] S43: Get the separating hyperplane ω * ·x+b * =0, and then the corresponding classification decision function is obtained:

[0061] f(x)=sign(ω * ·x+b * );

[0062] S44: Classify using the classification decision function to obtain a fault diagnosis result.

[0063] Compared with the prior art, the present invention has the following advantages:

[0064] 1) The fuzzy entropy with the ability to handle data uncertainty and fuzziness is extended to multiple time scales through an offset coarse-graining method; the advantage of this offset coarse-graining is that the data will be used multiple times, so each window can capture more subtle changes and trends; the mutations between different windows are reduced, and the local characteristics of the original signal are accurately reflected.

[0065] 2) In order to avoid the phenomenon of overfitting of classification results caused by redundant features, the present invention proposes an improved multi-scale fuzzy entropy and establishes a nonlinear dynamic model based on the improved multi-scale fuzzy entropy to quantify the time complexity; and through comparative experiments, it is proved that the improved multi-scale fuzzy entropy has better robustness.

[0066] 3) A rotating machinery fault diagnosis method integrating data decomposition, feature extraction, feature dimensionality reduction and fault classification is proposed; this method first uses EMD decomposition to decompose the original data into three layers to obtain three intrinsic modal components; then the signal features extracted by the improved multi-scale fuzzy entropy algorithm are input into the linear judgment analysis for feature dimensionality reduction; finally, the support vector machine algorithm is used as a classifier to realize classification, and the K-fold cross-validation is used to evaluate the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of a rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy in Embodiment 1 of the present invention;

[0068] Figure 2 is a schematic diagram of a framework of a rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy in Embodiment 1 of the present invention;

[0069] Figure 3 This is an example diagram of EMD algorithm decomposition in Embodiment 1 of the present invention;

[0070] Figure 4 is a schematic diagram of improved coarse-graining when the scale factor is 3 in Example 1 of the present invention;

[0071] Figure 5 is a schematic diagram of the process of improving the multi-scale fuzzy entropy algorithm in the first embodiment of the present invention;

[0072] Figure 6 exemplifies nine types of fault samples in Embodiment 2 of the present invention, wherein (a)-(i) correspond to nine types of fault samples respectively;

[0073] Figure 7 This is a comparison chart of 10 folds and average accuracy of different methods in Example 2 of the present invention. DETAILED DESCRIPTION

[0074] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0075] Embodiment 1

[0076] like Figure 1 , Figure 2 As shown, this embodiment provides a technical solution: a rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy, which is a new type of fault diagnosis method for time series, based on the realization of improved multi-scale fuzzy entropy and support vector machine for intelligent diagnosis of rotating machinery system faults, including the following main steps:

[0077] Step 1:

[0078] The time series samples of the original vibration signal of the rotating machinery are decomposed by EMD and the first three IMF components are taken; the IMF components include the oscillation information of the original samples in a specific frequency range, while retaining the nonlinear and non-stationary characteristics of the signal.

[0079] In this step, the first three IMF components are selected for the following reasons:

[0080] 1. The IMF components of EMD decomposition of different signals may be different. In order to ensure the uniformity of subsequent feature dimensions, the first three are selected;

[0081] 2. The first three IMF components can usually effectively cover most of the data features in the signal and are sufficient to provide accurate feature information;

[0082] 3. Selecting the first three IMF components can reduce the amount of calculation and improve the efficiency of the algorithm.

[0083] Step 2

[0084] Secondly, the improved multi-scale fuzzy entropy algorithm is used to extract features of the three IMF components respectively to obtain preliminary feature vectors.

[0085] Step 3

[0086] The LDA algorithm is used to reduce the dimension of the preliminary feature vector to obtain a feature vector reduced to 9 dimensions.

[0087] Step 4

[0088] Input the reduced-dimensional feature vector, classify it using SVM, and obtain the fault diagnosis result.

[0089] Step 5

[0090] Finally, K-fold cross-validation is used to measure the effectiveness of the rotating machinery fault diagnosis method.

[0091] As more specific, in step one, the idea of ​​EMD (empirical mode decomposition) is to decompose the nonlinear and non-stationary signals in the vibration signal into several intrinsic mode functions. Each IMF component reflects the different frequency components in the signal (the result of calculating the improved multi-scale fuzzy entropy for each IMF can reflect the dynamic characteristics on different time scales). This method extracts local features through an iterative process, emphasizes the time-frequency analysis of the signal, and can effectively capture the instantaneous frequency and amplitude changes in the vibration signal. The present invention uses EMD as a data processing method for vibration signals because, compared with other decomposition methods, EMD can be decomposed according to the characteristics of the signal itself. It can effectively extract dynamic features on different time scales, and has time-frequency analysis capabilities to obtain more comprehensive signal features. Specifically, the EMD method can decompose non-stationary signals into a trend term and several intrinsic mode functions with meaningful instantaneous frequencies. It needs to process vibration signal data based on three assumptions: 1) The signal has at least two extreme points, a maximum and a minimum. 2) The characteristic time scale is defined by the time between the two extreme points. 3) If the data lacks extreme points but has deformation points, the extreme points can be obtained by differentiating the data once or several times, and then the decomposition result can be obtained by integration. Figure 3 As shown, different frequency components in the original signal can be separated to better extract features. The following are its main steps:

[0092] Step 1: For the input sample x(t), the spline interpolation method is used to fit the maximum and minimum points of the original vibration signal to form the upper and lower envelopes; the average envelope x0(t) is calculated, and then the original vibration signal is subtracted from the mean to obtain a new signal h1(t):

[0093] h1(t)=x(t)-x0(t)

[0094] Step 2: Determine whether the new signal h1(t) meets the characteristic conditions of IMF; if it does, h1(t) is the first IMF component, i.e., imf1. If it does not, continue to loop through step 1 until h1(t) meets the characteristic conditions of IMF.

[0095] In this step, the characteristic conditions are as follows:

[0096] (1) In the entire data segment, the number of extreme points and the number of zero-crossing points must be equal or differ by no more than one.

[0097] (2) At any time, the average value of the upper envelope formed by the local maximum points and the lower envelope formed by the local minimum points is zero, that is, the upper and lower envelopes are locally symmetrical with respect to the time axis.

[0098] Step 3: Separate imf1 from the original vibration signal and obtain the separated signal r1(t):

[0099] r1(t)=x(t)-imf1

[0100] Step 4: Repeat the above steps until the nth (n=3 in this embodiment) IMF component is obtained. Finally, the result of the empirical mode decomposition is as follows:

[0101]

[0102] Among them, r n (t) is the remainder, which represents the central tendency of the vibration signal.

[0103] As more specific, in step two, by evaluating the fuzziness and uncertainty of the vibration signal, fuzzy entropy can effectively extract fault-related features. Compared with the traditional fuzzy entropy algorithm, the multi-scale fuzzy entropy algorithm extracts multi-level signal features and obtains potential fault modes by calculating entropy values ​​at different time scales. However, as the scale factor of entropy increases, the length of the coarse-grained time series will decrease significantly. This sudden change in sequence length causes the entropy value to fluctuate dramatically, making it difficult to extract key features. In order to solve this problem, the present invention proposes a new time segmentation method, which aims to replace the traditional segmentation method to overcome the defect of large entropy value volatility under high scale factors. For example, when the sequence length is 21 and the scale factor is 3, offset coarse-graining is performed to generate three subsequences of length 19, 10, and 7, respectively, reducing the sudden change in entropy value and calculation deviation. As Figure 4 As shown in the figure, it is a schematic diagram of improved coarse-graining. If the traditional coarse-graining is directly performed, only one subsequence of length 7 will be generated. It discards most of the local information and affects the construction of features. Combining it with the fuzzy entropy algorithm, the improved multi-scale fuzzy entropy algorithm is obtained. The algorithm flow is as follows: Figure 5 As shown, the following are its main steps:

[0104] Step 1: Input the equipment vibration signal sample (in this embodiment, the first three IMF components obtained above), and set the coarse-grained scaling factor sequence:

[0105] scales=(scale1,scale2,...,scale N ),1<scale i <N;

[0106] Wherein, scales is the coarse-grained scale. The present invention relates to multi-scale fuzzy entropy. Therefore, given different coarse-grained scales, time series at different scales are obtained. The subscript number represents the scale.

[0107] Step 2: Set the offset factor k = 1. When the scale factor is scale1, k increases from 1 until it is equal to scale1. The time series obtained by coarsening at each different k is calculated and its fuzzy entropy is calculated. Then the operation is repeated at the next scale factor.

[0108] Step 3: The fuzzy entropy algorithm is:

[0109] 3.1: For a given N-dimensional time series {u(i), 1≤i≤N};

[0110] 3.2: Construct an m-dimensional vector according to the following formula:

[0111]

[0112] Where i = 1, 2, ..., N-m + 1, It is the m consecutive values ​​of u starting from the i-th one minus the mean value u0(i),

[0113] 3.3: Definition and The distance between is the maximum value of the difference between the corresponding elements:

[0114]

[0115] Where, i=1,2,...,Nm,i≠j;

[0116] It should be noted that in step 3.2, for example, 5 vectors are constructed, i is from 1 to 5. In step 3.3, the distance between each two vectors is calculated; i is defined in front and j in the back, and the distance between two vectors is calculated, that is, the distance between [1,2] / [2,3] / [3,4] / [4,5]; the i subscript represents the front vector in [,], so it is up to 4, which is different from step 3.2; and because the distance between different vectors is calculated, i and j are not equal;

[0117] 3.4: Through the fuzzy function definition and Similarity

[0118]

[0119] Among them, the fuzzy function is an exponential function, n and r are the gradient and width of the blur function boundary respectively;

[0120] 3.5: Define the fuzzy similarity mean function (used to calculate the average fuzzy similarity under each embedding dimension in the signal) as follows:

[0121]

[0122] Similarly, let the dimension be m+1 and repeat steps 3.1 to 3.4 to obtain:

[0123]

[0124] 3.6: Calculate fuzzy entropy:

[0125]

[0126] Step 4: Combining all fuzzy entropies is the improved multi-scale fuzzy entropy, and outputting the entropy matrix

[0127] As more specific, in step three, LDA linear discriminant analysis is a supervised learning algorithm. By maximizing the divergence between classes and minimizing the divergence within classes, LDA can effectively reduce high-dimensional data to low dimensions, simplify the feature space, and improve computational efficiency. LDA assumes that all types of sample data are Gaussian distributed, and the covariance matrix is ​​the same and full rank. At the same time, LDA can provide higher classification accuracy when dealing with fault modes with obvious class differences. The following are the main steps of LDA:

[0128] Input data set X = [(x1,y1),(x2,y2),...,(x n ,y n )], where any sample x i is an n-dimensional vector, y i is the category label, and the dimension reduced to is p. The data set in this embodiment is a preliminary feature vector obtained through feature extraction.

[0129] Step 1: Calculate the mean vector μ for each category of the dataset i and the population mean vector μ;

[0130] Step 2: Calculate the intra-class scatter matrix S ω , global divergence matrix S t , and get the inter-class scatter matrix S b =S t -S ω ;

[0131] Step 3: Matrix Perform eigenvalue decomposition and sort the eigenvalues ​​from large to small;

[0132] Step 4: Take the eigenvector corresponding to the first p largest eigenvalues, and use the eigenvalue x′=ω T x will n The dimensional sample is reduced to p dimensions.

[0133] As more specific, in step four, support vector machine (SVM) is a commonly used machine learning algorithm that can process high-dimensional vibration signal data. It provides high classification accuracy by constructing an optimal hyperplane for classification, especially when the number of samples is small. In addition, SVM has good robustness and is insensitive to noise and outliers, making it more reliable in practical applications. Its main idea is to separate data of different fault categories by finding a hyperplane in the feature space, and maximize the distance from the data point closest to the hyperplane to the hyperplane. The following are the main steps of SVM:

[0134] Input vibration signal data set X = [(x1, y1), (x2, y2), ..., (x n ,y n )], where any sample x i is a p-dimensional vector, y i is a category label, i=1, 2, ..., N; the vibration signal data set in this embodiment is a feature vector after dimensionality reduction;

[0135] Step 1: Select the penalty parameter C>0, construct and solve the convex quadratic programming problem:

[0136] satisfy Get the optimal solution

[0137] Step 2: Calculation Select α * A component of Satisfy the condition 0<α j <C, calculation

[0138]

[0139] Step 3: Get the separating hyperplane ω * ·x+b * =0, and then the corresponding classification decision function is obtained:

[0140] f(x)=sign(ω * ·x+b * ).

[0141] Embodiment 2

[0142] In this embodiment, the classification efficiency of the method of the present invention is verified on the Case Western Reserve University data set. In the experiment, we used the seed fault test data of the Case Western Reserve University Bearing Data Center. When the motor load is 2 horsepower and the motor speed is 1750rpm, 9 types of working data, including inner and outer ring faults in different positions, rotor faults, and normal working conditions, totaling 10 categories, are sorted out and made into a data set. Starting from the starting data, a subsequence of 1024 length is taken as a sample, with a step size of 512, and then the next sample is taken. The first 233 samples of each category are taken, and a total of 2330 samples of ten categories are used for experiments. The specific types of faults are shown in Table 1.

[0143] Table 1 Specific conditions of various faults on the platform

[0144] Serial number Fault (Yes / No) Fault location Fault Diameter (inches) Abbreviation Type 1 No Type 2 Yes Inner Circle 0.007 IR7 Type 3 Yes Rolling element 0.007 RE7 Type 4 Yes Outer ring 0.007 OR7 Type 5 Yes Inner Circle 0.014 IR14 Type 6 Yes Rolling element 0.014 RE14 Type 7 Yes Outer ring 0.014 OR14 Type 8 Yes Inner Circle 0.021 IR21 Type 9 Yes Rolling element 0.021 RE21 Type 10 Yes Outer ring 0.021 OR21

[0145] The specific fault types are shown in Table 1. There are 10 types of data in total. Type 1 is the data under normal working conditions, and types 2 to 10 are the data when the motor fails. The main fault locations of the motor are inner ring fault, outer ring fault and rotating body fault. Due to the different fault diameters, it can be subdivided into 0.007 inches, 0.014 inches and 0.021 inches. Figure 6 These are the specific forms of nine fault samples.

[0146] The following compares the diagnostic results of different entropy algorithms including fuzzy entropy, multi-scale fuzzy entropy, fuzzy entropy combination, fuzzy entropy waveform factor and the proposed method. The scale factors of multi-scale fuzzy entropy and improved multi-scale fuzzy entropy are set to scales = [2,3,4], the embedding dimension is 2, and the time delay is 1. The features obtained by the improved multi-scale fuzzy entropy are reduced to 3 dimensions using the LDA linear discriminant analysis method. SVM classifiers are used for classification and K-fold crossover method is used for verification. K is set to 10, that is, the data set is divided into 10 folds, each with 233 samples. By Figure 7 It can be seen that the classification accuracy of the method of the present invention is relatively high in each fold. The experimental results of the comparative method are averaged ten times, and the results of the highest and lowest accuracy rates within ten times are shown in Table 2.

[0147] Table 2 Average, maximum and minimum classification accuracy of different diagnostic methods

[0148]

[0149] As can be seen from Table 2, under the conditions of motor load of 2 horsepower and motor speed of 1750rpm, the accuracy results of the proposed method of the present invention are the highest, indicating the superiority of the diagnostic performance of the proposed method. It can be seen from the results of Method 1 and Method 2 that different EMD decomposition layers will have a significant impact on the diagnostic effect, and as the number of decomposition layers increases, the diagnostic effect is gradually improved. Obviously, when the number of decomposition layers is not enough, the final diagnostic effect of the ordinary feature extraction method is poor. For example, when the 3-layer decomposition is performed, the average classification accuracy of Method 1 and Method 4 does not reach 90%, and cannot be applied in practice. And it can be seen from the diagnostic results of Method 2, Method 3, and Method 4 that when using fuzzy entropy to extract feature details, if the number of EMD decomposition layers is high, overfitting problems will occur. The average accuracy of Method 3 and Method 4 decreased by 0.12%, and the highest and lowest accuracy remained basically unchanged. Therefore, if the entropy feature is further extracted, overfitting problems are prone to occur. Therefore, the present invention proposes a method to improve the classification effect by improving the form of feature extraction method at a low decomposition level, that is, introducing IMFE as a feature extraction method, and then using LDA for dimensionality reduction. The experimental results show that its diagnostic effect is high when the EMD decomposition level is 3, and the classification accuracy is high, which verifies the feasibility of the proposed method.

[0150] In summary, the rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy in the above-mentioned embodiment combines EMD decomposition, improved multi-scale fuzzy entropy algorithm, LDA linear discriminant analysis and SVM classification, and focuses on exploring the feature extraction method; the proposed improved multi-scale fuzzy entropy algorithm has the advantages of few adjustable parameters, strong algorithm robustness, and the ability to effectively distinguish different types of signals; simulation experiments were carried out from the robustness of time length and the recognition performance of different signals, verifying its stability in feature extraction; the diagnosis method was compared with several other different types of rotating machinery diagnosis methods, and the diagnosis results had a greater competitive advantage over traditional diagnosis methods, verifying the good diagnostic ability of this method.

[0151] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy, characterized in that: The following steps are involved: S1: EMD decomposition Perform EMD decomposition on the time series samples of the original vibration signal of the rotating machinery and take the first three IMF components; S2: Feature extraction The improved multi-scale fuzzy entropy algorithm is used to extract the features of the three IMF components respectively to obtain the preliminary feature vectors. S3: Feature Dimensionality Reduction Use the LDA algorithm to reduce the dimension of the preliminary feature vector to obtain the feature vector after dimension reduction; S4: Troubleshooting Input the reduced-dimensional feature vector, classify it using SVM, and obtain the fault diagnosis result.

2. A rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy according to claim 1, characterized in that: In step S1, the specific process of EMD decomposition is as follows: S11: For the input sample x(t), the spline interpolation method is used to fit the maximum and minimum points of the original vibration signal to form the upper and lower envelopes; the average envelope x0(t) is calculated, and then the original vibration signal is subtracted from the mean to obtain a new signal h1(t): h1(t)=x(t)-x0(t); S12: Determine whether the new signal h1(t) satisfies the characteristic condition of IMF; if so, h1(t) is the first IMF component, i.e., imf1; if not, continue to loop through step S11 until h1(t) satisfies the characteristic condition of IMF; S13: Separate imf1 from the original vibration signal to obtain the separated signal r1(t): r1(t)=x(t)-imf1; S14: Repeat the above steps until the first three IMF components are obtained.

3. The rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy according to claim 2 is characterized in that: In step S12, the characteristic conditions are specifically as follows: In the entire data segment, the number of extreme points and the number of zero-crossing points are equal or the difference cannot exceed one; At any time, the upper and lower envelopes are locally symmetric with respect to the time axis.

4. The rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy according to claim 1 is characterized in that: In step S2, the specific processing process of improving the multi-scale fuzzy entropy algorithm is as follows: S21: Input the first three IMF components obtained and set the coarse-grained scaling factor sequence: scales=(scale1,scale2,...,scale N ),1<scale i <N Among them, scales is the coarse-grained scale; S22: Set the offset factor k=1. When the scale factor is scale1, k is incremented from 1 until it is equal to scale1. The time series obtained by coarsening under each different offset factor k is calculated and its fuzzy entropy is calculated. Then the operation is repeated under the next scale factor. S23: Combining all fuzzy entropies is the improved multi-scale fuzzy entropy, and outputting the entropy matrix That is, we get the preliminary feature vector.

5. A rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy according to claim 4, characterized in that: In step S22, the calculation process of fuzzy entropy is as follows: S221: Given an N-dimensional time series {u(i), 1≤i≤N}; S222: Construct an m-dimensional vector according to the following formula: Where i = 1, 2, ..., N-m + 1, It is the m consecutive values ​​of u starting from the i-th one minus the mean value u0(i), S223: Definition and The distance between is the maximum value of the difference between the corresponding elements: Where, i=1,2,...,Nm,i≠j; S224: Through the fuzzy function definition and Similarity Among them, the fuzzy function is an exponential function, n and r are the gradient and width of the blur function boundary respectively; S225: Define the fuzzy similarity mean function as follows: Similarly, let the dimension be m+1, and repeat steps S221 to S224 to obtain: S226: Calculate fuzzy entropy:

6. The rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy according to claim 1 is characterized in that: In step S3, the specific process of the LDA algorithm is as follows: S31: Calculate the mean vector μ for each category in the preliminary feature vector i and the population mean vector μ; S32: Calculate the intra-class scatter matrix S ω , global divergence matrix S t , and get the inter-class scatter matrix S b =S t -S ω ; S33: Pair Matrix Perform eigenvalue decomposition and sort the eigenvalues ​​from large to small; S34: Take the eigenvectors corresponding to the first p largest eigenvalues, and pass the eigenvalue x′=ω T x will n The dimensional sample is reduced to p dimensions.

7. The rotating machinery fault diagnosis method based on improved multi-scale fuzzy entropy according to claim 1 is characterized in that: In step S4, the specific processing process of SVM is as follows: S41: Select penalty parameter C>0, construct and solve the convex quadratic programming problem: satisfy 0<α i <C,i=1,2,...,N, get the optimal solution S42: Calculation Select α * A component of Satisfy the condition 0<α j <C, calculation S43: Get the separating hyperplane ω * ·x+b * =0, and then the corresponding classification decision function is obtained: f(x)=sign(ω * ·x+b * ); S44: Classify using the classification decision function to obtain a fault diagnosis result.

Images