A zero-shot rotating machinery fault diagnosis method based on attribute coding

Abstract

A zero-shot rotating machinery fault diagnosis method based on attribute coding comprises the following steps: step 1: determining the attribute description of the fault scenario; step 2: establishing the attribute coding of the fault scenario; step 3: acquiring multi-source data of the rotating machinery; step 4: forming a multi-domain feature set of the rotating machinery; step 5: constructing a feature subset of the attribute coding; step 6: classifier training; step 7: zero-shot fault diagnosis to realize fault diagnosis of samples of unknown category.

Description

Technical Field

[0001] The present invention relates to a rotating machinery fault diagnosis method, and more particularly to a method for solving the zero-sample problem of a traditional data-driven rotating machinery fault diagnosis method by utilizing an attribute coding method. Background Art

[0002] Rotating machinery occupies an important position in industrial manufacturing production. Monitoring its operating status and accurately and efficiently performing real-time fault diagnosis are of great significance to ensuring its stable operation.

[0003] With the development of machine learning and deep learning, the fault diagnosis of rotating machinery has gradually tended to a data-driven approach, which mainly includes steps such as operating data collection, mechanical state feature extraction, dimensionality reduction and classifier fault identification. The data-driven fault diagnosis method requires a large amount of balanced and comprehensive data to train the model, so as to improve the diagnostic accuracy of the model, and can only diagnose and identify data of the same category involved in the training. However, in actual working conditions, rotating machinery often operates normally for a long time, and the occurrence of a fault is a low-probability event, which causes the collected data to have a sample imbalance problem. The category and number of fault samples cannot meet the requirements of model training, and there is even a problem that the number of samples of a certain fault category is zero, that is, the zero-sample problem of rotating machinery fault diagnosis. Therefore, there is a need in this field for a method that can effectively identify the state of rotating machinery and accurately identify its fault when an unknown fault occurs, that is, when the zero-sample problem occurs. Summary of the Invention

[0004] In order to solve the above-mentioned problems in the field, an embodiment of the present invention provides a zero-sample rotating machinery fault diagnosis method based on attribute coding.

[0005] According to an embodiment of the present invention, a zero-sample rotating machinery fault diagnosis method based on attribute coding is provided, comprising the following steps:

[0006] Step 1: Determine the attribute description of the fault scenario, summarize all potential faults of rotating machinery, and perform disassembly analysis on the fault to determine the fault mode, find the fault cause, define the fault scenario, summarize and organize the relevant information of the above faults to build a fault attribute system. This attribute system includes the unique attributes and common attributes of different faults, and then determine the attribute description of the fault scenario, which provides support for subsequent attribute coding;

[0007] Step 2: Establish attribute coding for fault scenarios. Attribute coding is performed for different faults of rotating machinery, including known and unknown faults. All fault categories are encoded using one-hot encoding to obtain 0-1 codes for different faults.

[0008] Step 3: Acquire multi-source data of the rotating machinery by setting sensors at different locations on the rotating machinery to collect vibration signals of the rotating machinery;

[0009] Step 4: Form a multi-domain feature set of the rotating machinery, extract the time domain features, frequency domain features, energy features and entropy features of the collected vibration signal of the rotating machinery through domain conversion, thereby forming a multi-domain feature set;

[0010] Step 5: Construct the feature subset of attribute coding. Through the recursive feature elimination algorithm based on support vector machine, perform feature selection for each dimension of 0-1 code, select the features with strong distinguishability, eliminate useless features and redundant features, reduce the amount of calculation, and finally form multiple feature subsets, and each dimension of 0-1 code has a corresponding feature subset;

[0011] Step 6: Classifier training: Use the feature subset corresponding to the 0-1 code of each dimension as the input of the classifier, and use the 0-1 code of each dimension instead of the sample label as the output of the classifier for classifier training. Train multiple classifiers, and each dimension of the 0-1 code corresponds to a trained classifier.

[0012] Step 7: Zero-sample fault diagnosis. For the sample data of unknown category faults, according to step 4, the composition of the feature subsets of each dimension of the 0-1 code of the known category fault is used to form the feature subsets of each dimension of the 0-1 code of the unknown category fault sample, and this feature subset is used as the input of the trained classifier. Finally, a string of binary code outputs consisting of 0-1 is obtained, and this output is sequentially compared with the 0-1 code of each category in step 1 based on Euclidean distance, and finally the fault diagnosis of the unknown category sample is realized.

[0013] In an optional embodiment, the zero-sample rotating machinery fault diagnosis method based on attribute coding, in the step one: the fault mode includes the fault location and the fault degree; the fault cause refers to the direct critical factor causing the fault; the fault scenario includes the fault location, the fault degree, the fault impact, and the fault operating condition, that is, the sum of the fault mode and the fault cause; the unique attributes are related attributes that are different from each other for different fault categories, and the common attributes refer to related attributes that all fault categories have.

[0014] In an optional embodiment, in the step 2: the attribute encoding is to encode the various attributes of the rotating machinery, wherein if the attribute is present, the encoding dimension corresponding to the attribute is set to 1, and if the attribute is not present, the encoding dimension corresponding to the attribute is set to 0; the one-hot encoding is to use 0 and 1 to represent the attribute parameters of the rotating machinery, and use N-dimensional binary code to encode the N attributes of the rotating machinery.

[0015] In an optional embodiment, in step four: the time domain features are changes in the vibration signal in the time domain observed through time t, including dimensional and dimensionless features; dimensional time domain features may include maximum value, average value, root mean square value, absolute average amplitude, root mean square amplitude, peak-to-peak value and standard deviation; dimensionless time domain features may include kurtosis, skewness, form factor, peak factor, pulse factor, and margin factor.

[0016] In an optional embodiment, in step four: the frequency domain features include performing fast Fourier transform on the vibration signal and calculating frequency domain features, and the frequency domain features may include center of gravity frequency CF, mean frequency MF, root mean square frequency RMSF, and frequency standard deviation RVF.

[0017] In an optional embodiment, in step 4, the energy feature is marginal spectral energy, which is obtained by the following steps:

[0018] First, the vibration signal is subjected to empirical mode decomposition to obtain the intrinsic mode function (IMF) of the vibration signal from high frequency to low frequency. Then, each IMF is subjected to Hilbert transform to obtain the Hilbert spectrum, where H i (ω,t) is the Hilbert spectrum of each IMF component, ω represents the frequency, and t represents the time corresponding to the frequency

[0019]

[0020] Next, the Hilbert spectrum H of each IMF component is i By integrating (ω, t) in the time domain, the marginal spectrum of each IMF component can be obtained

[0021]

[0022] Then, the marginal spectrum energy can be obtained by squaring the amplitude of the marginal spectrum

[0023]

[0024] Finally, the normalized probability is obtained by the following formula, where p(i) represents the normalized probability, and the marginal spectral energy entropy is obtained by the formula of Shannon entropy, where n represents the number of IMFs

[0025]

[0026] In an optional embodiment, the entropy feature described in step 4 is a feature used to describe the uncertainty and complexity of the vibration signal distribution, including power spectrum Shannon entropy, power spectrum exponential entropy, singular spectrum Shannon entropy, singular spectrum exponential entropy and wavelet energy entropy, which is obtained as follows:

[0027] According to different calculation methods of entropy, it can be divided into Shannon entropy and exponential entropy, where H1 is Shannon entropy and H2 is exponential entropy.

[0028]

[0029]

[0030] Where N represents the total number of probabilities, p i represents probability;

[0031] Among them, the power spectrum is a function used to describe how the power of the vibration signal changes with frequency.

[0032] First, the vibration signal sequence {x t ,t=1,2,…,N} perform fast Fourier transform FFT to obtain X(ω) and calculate the power spectrum density of the vibration signal Where t is the sampling point, ω is the corresponding frequency, and N is the number of sampling points

[0033]

[0034] After that, the power spectrum density is normalized to obtain the probability density function p(ω i )

[0035]

[0036] Then, this probability density function p(ω i ) Substitute the above Shannon entropy and exponential entropy formulas to calculate the power spectrum Shannon entropy and power spectrum exponential entropy respectively;

[0037] The singular spectrum entropy can obtain the local characteristics of the vibration signal in the time domain by performing calculations such as singular value decomposition on the vibration signal.

[0038] First, the vibration signal sequence {x t ,t=1,2,…,N} is used for windowing processing. Windowing processing refers to the process of windowing the vibration signal sequence x with a length of M. t After truncation, the signal length is M, that is, the window size is M, the delay parameter is 1, where x t represents the vibration signal sequence, t represents the sampling point, M is the window size, and N is the vibration signal sequence x t The number of sampling points included is used to obtain the trajectory matrix B of the vibration signal, which divides the vibration signal into (N-M+1)×M dimensions.

[0039]

[0040] Then, the trajectory matrix B is subjected to SVD singular value decomposition

[0041] B=U×S×V'

[0042] Where U and V' represent the orthogonal matrix of singular value decomposition, S is the diagonal matrix, δ i Represents singular values, matrix U is an orthogonal matrix of (N-M+1)×(N-M+1), matrix V' is an orthogonal matrix of M×M, and matrix S is composed of singular values ​​δ i It forms a diagonal matrix of (N-M+1)×M, and in the matrix S, the singular values ​​are arranged on the main diagonal of the matrix S from large to small, that is, δ1≥δ2≥…≥δ M ,

[0043] Then, substitute the singular value into the following formula to find the normalized probability

[0044]

[0045] Substitute this probability into the above formulas of Shannon entropy and exponential entropy to obtain the singular spectrum Shannon entropy and singular spectrum exponential entropy;

[0046] The wavelet packet in wavelet energy entropy is a feature based on the time-frequency domain.

[0047] First, the vibration signal {x t ,t=1,2,…,N} perform n-layer wavelet packet decomposition, and get 2 n nodes, and at the same time obtain the wavelet packet coefficient W of each node in the nth layer i , where i = 1, 2, ..., 2 n ,

[0048] Then, the coefficients are squared to obtain the wavelet packet energy

[0049]

[0050] Then, the wavelet packet energy is normalized to obtain the normalized probability, and the wavelet packet energy entropy is obtained from the normalized probability.

[0051]

[0052] In an optional embodiment, the recursive feature elimination algorithm based on the support vector machine in step 5 is a feature selection method based on the SVM margin maximization principle, and the steps are as follows:

[0053] The weight coefficient of the support vector machine is used as the evaluation criterion. The feature with the lowest score is eliminated in each iteration until the feature set becomes an empty set or a specific number of features is obtained, and the iteration is stopped. The scoring criterion is as follows: represents the score of feature i, c represents the score, Represents the weight coefficient of SVM

[0054]

[0055] In an optional embodiment, the Euclidean distance in step 7 is the distance between two points x(x1, x2, .., x m ) and y(y1,y2,…,y m ), the actual distance between

[0056]

[0057] Among them, dist(X,Y) represents the actual distance between two points in m-dimensional space.

[0058] The zero-sample rolling bearing fault diagnosis method based on attribute coding provided by the embodiments of the present invention includes at least the following advantages. The zero-sample rotating machinery fault diagnosis method based on attribute coding proposed by the embodiments of the present invention can achieve the purpose of fault diagnosis of unknown category samples when only known category sample data is used to train the classifier. All states of the rotating machinery are encoded by attribute coding, and a one-hot encoding method is proposed to increase the dimension of attribute description. A recursive feature elimination algorithm based on support vector machine is used to adaptively select a feature subset for each dimension 0-1 code from the multi-domain feature set, thereby achieving feature dimensionality reduction for each dimension 0-1 code, eliminating useless features and redundant features, and reducing the amount of calculation. The 0-1 code is used to replace the label as the output of the classifier, and the classifier is used to train each dimension 0-1 code, ultimately achieving the output of the unknown category 0-1 code by the classifier trained with known category sample data. And using Euclidean distance as a criterion, fault diagnosis of unknown category samples is achieved. The method provided by the embodiments of the present invention is a guiding zero-sample rotating machinery fault diagnosis method with the characteristics of simple calculation and strong real-time performance. The method provided by the embodiment of the present invention has certain openness and practicality in practical applications, and is applicable to fault diagnosis of rotating machinery such as gearboxes and bearings. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The foregoing features of the present invention may be more readily understood with reference to the accompanying drawings and the following detailed description.

[0060] Figure 1 A flowchart of a zero-sample rolling bearing fault diagnosis method based on attribute coding according to an embodiment of the present invention is shown.

[0061] Figure 2 A flowchart of multi-domain feature extraction in a zero-sample rolling bearing fault diagnosis method based on attribute coding is shown according to one embodiment of the present invention.

[0062] Figure 3A feature selection block diagram in an exemplary embodiment of a zero-sample rolling bearing fault diagnosis method based on attribute coding provided according to an embodiment of the present invention is shown.

[0063] The serial numbers, symbols and codes in the figure are explained as follows

[0064] T: Time domain characteristics of the original vibration signal

[0065] F: Frequency domain characteristics of the original vibration signal

[0066] H: Entropy characteristics of the original vibration signal

[0067] E: Energy characteristics of the original vibration signal

[0068] SVM: Support Vector Machine DETAILED DESCRIPTION

[0069] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0070] It should be noted that, unless otherwise specified, the technical or scientific terms used in the present invention should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0071] Hereinafter, in conjunction with the accompanying drawings, an illustrative embodiment of a zero-sample rotating machinery fault diagnosis method based on attribute coding provided according to an embodiment of the present invention will be described in detail, but the present invention is not limited to the specific embodiment.

[0072] Attribute coding is a coding method based on one-hot encoding. By describing each attribute of a rotating machinery fault, each fault state has an independent 0-1 code. Attribute description refers to the analysis of each rotating machinery fault, selecting representative common characteristics for classification and description. Attribute description allows for the most intuitive summary of each rotating machinery fault state. The zero-shot fault diagnosis method uses the empty set as the intersection of the sample categories of the training set and the test set. The training set is used to train a classifier for each dimension of the 0-1 code, and ultimately, the test set can be diagnosed through distance discrimination, thereby achieving zero-shot rotating machinery fault diagnosis.

[0073] Therefore, by placing different sensors at different locations on a rotating machine, vibration signals of the rotating machine under different conditions are extracted, and these signals are used as data sources for samples of known fault categories. Through domain conversion, features such as time domain, frequency domain, time-frequency domain, entropy, and energy are extracted from the vibration signals of the known rotating machine categories, forming a multi-domain feature set. Based on this, attribute coding is employed to train a model using the feature set of known fault categories for the 0-1 codes of each dimension of the known categories. This allows multiple trained classifiers to perform attribute judgment on sample data of unknown categories, ultimately achieving fault diagnosis for the rotating machine. Based on this, embodiments of the present invention provide a zero-sample rotating machine fault diagnosis method based on attribute coding. The zero-sample rotating machine fault diagnosis method based on attribute coding, provided by embodiments of the present invention, describes the attributes of each type of rotating machine fault, encodes each type of fault using attribute coding, and replaces the sample label with each dimension of the encoded 0-1 code as the final output of each classifier. When sample data of unknown fault categories is used as input to a trained multi-dimensional classifier, 0-1 codes of the unknown category sample data can be obtained, which serve as a diagnostic basis for subsequent rotating machine fault diagnosis.

[0074] Figure 1 A flowchart of a zero-sample rolling bearing fault diagnosis method based on attribute coding according to an embodiment of the present invention is shown. Figure 2 A flowchart of multi-domain feature extraction in a zero-sample rolling bearing fault diagnosis method based on attribute coding is shown according to one embodiment of the present invention.

[0075] refer to Figure 1-2 According to an embodiment of the present invention, a zero-sample rotating machinery fault diagnosis method based on attribute coding is provided, and the method includes the following steps:

[0076] Step 1: Determine the attribute description of the fault scenario, summarize all potential faults of rotating machinery, and perform disassembly analysis on the fault to determine the fault mode, find the fault cause, define the fault scenario, summarize and organize the relevant information of the above faults to build a fault attribute system. This attribute system includes the unique attributes and common attributes of different faults, and then determine the attribute description of the fault scenario, which provides support for subsequent attribute coding;

[0077] Step 2: Establish attribute coding for fault scenarios. Attribute coding is performed for different faults of rotating machinery, including known and unknown faults. All fault categories are encoded using one-hot encoding to obtain 0-1 codes for different faults.

[0078] Step 3: Acquire multi-source data of the rotating machinery by setting sensors at different locations on the rotating machinery to collect vibration signals of the rotating machinery;

[0079] Step 4: Form a multi-domain feature set of the rotating machinery, extract the time domain features, frequency domain features, energy features and entropy features of the collected vibration signal of the rotating machinery through domain conversion, thereby forming a multi-domain feature set;

[0080] Step 5: Construct a feature subset for attribute coding. Through the recursive feature elimination algorithm based on support vector machines, perform feature selection for each dimension of 0-1 code, select highly distinguishable features, eliminate useless features and redundant features, and finally form multiple feature subsets, and each dimension of 0-1 code has a corresponding feature subset;

[0081] Step 6: Classifier training. The feature subset corresponding to each dimension of the 0-1 code is used as the input of the classifier, and each dimension of the 0-1 code replaces the sample label as the output of the classifier for classifier training. Finally, multiple classifiers will be trained, and each dimension of the 0-1 code corresponds to a trained classifier.

[0082] Step 7: Zero-sample fault diagnosis. For the sample data of unknown category faults, according to the composition of the 0-1 code feature subsets of each dimension of the known category faults in step 4, a feature subset of 0-1 codes of each dimension of the unknown category fault sample is formed, and this feature subset is used as the input of the trained classifier. Finally, a string of binary code outputs consisting of 0-1 is obtained, and this output is sequentially compared with the 0-1 codes of each category in step 1 based on Euclidean distance, and finally the fault diagnosis of the unknown category sample is realized.

[0083] Optionally, the fault mode described in step one mainly refers to the external macroscopic manifestation of the fault, such as the fault location, fault degree, etc.; the fault cause mainly refers to the direct key factors causing the fault, such as overload, etc.; the fault scenario mainly refers to the relative or combined situation of various factors within a certain period of time, including the fault location, fault degree, fault impact, fault operating conditions, etc., that is, the sum of the fault mode and the fault cause; the unique attribute refers to the related attributes that are different from each other in different fault categories, and the common attribute refers to the related attributes that all fault categories have. Through the attribute description, we can have the most intuitive understanding of a fault.

[0084] Optionally, the attribute encoding described in step 2 mainly refers to encoding the various attributes of the rotating machinery. For example, if the attribute is present, the encoding dimension corresponding to the attribute is set to 1, and if the attribute is not present, the encoding dimension corresponding to the attribute is set to 0; One-Hot encoding refers to using 0 and 1 to represent some attribute parameters of the rotating machinery, and using N-dimensional binary code to encode the N attributes of the rotating machinery. For example, the location where the rotating machinery fault occurs: [inner ring, outer ring, rolling element], the one-hot encoding of the fault occurring in the inner ring is [1,0,0], the one-hot encoding of the fault occurring in the outer ring is [0,1,0], and the one-hot encoding of the fault occurring in the rolling element is [0,0,1]. Using one-hot encoding, the encoded data obtained by encoding the attributes of the rotating machinery can be sparsely processed. The known category fault refers to the fault category for which data has been collected before model training, and the unknown category fault refers to the fault category for which data has not been collected before model training.

[0085] Optionally, the time-domain features described in step 4 primarily refer to the time-domain changes of the vibration signal observed over time t, and include both dimensional and dimensionless features. Dimensional time-domain features primarily include maximum value, average value, root mean square value, absolute average amplitude, root mean square amplitude, peak-to-peak value, and standard deviation. Dimensionless time-domain features primarily include kurtosis, skewness, form factor, crest factor, impulse factor, and margin factor. The specific calculation formulas are shown in Table 1.

[0086] Table 1 Time domain characteristics

[0087]

[0088]

[0089] Among them, x i Indicates the amplitude of each sampling point, N s Indicates the number of sampling points.

[0090] Optionally, the frequency domain features described in step 4 include performing a fast Fourier transform on the vibration signal and calculating certain frequency domain features. The frequency domain features may include the center of gravity frequency (CF), the mean frequency (MF), the root mean square frequency (RMSF), and the frequency standard deviation (RVF). The specific calculation formulas are shown in Table 2.

[0091] Table 2 Frequency domain characteristics

[0092]

[0093] Where, f j Indicates the sampling frequency, N f represents the number of sampling points, and S(f) represents the power spectrum of the sampling points.

[0094] Optionally, the energy characteristics described in step 4 mainly refer to the marginal spectrum energy. First, perform empirical mode decomposition on the vibration signal to obtain the intrinsic mode function IMF of the vibration signal from high frequency to low frequency. Then, perform Hilbert transform on each IMF to obtain the Hilbert spectrum. i (ω,t) is the Hilbert spectrum of each IMF component, ω represents the frequency, and t represents the time corresponding to the frequency

[0095]

[0096] Then, the Hilbert spectrum H of each IMF component is i By integrating (ω, t) in the time domain, the marginal spectrum of each IMF component can be obtained, where T represents the sampling time of the signal.

[0097]

[0098] Then, the marginal spectrum energy can be obtained by squaring the amplitude of the marginal spectrum

[0099]

[0100] Finally, the normalized probability is obtained by the following formula, where p(i) represents the normalized probability. The marginal spectral energy entropy is obtained by the formula of Shannon entropy, where n represents the number of IMFs.

[0101]

[0102] Among them, the energy of a vibration signal is mainly concentrated in the first few IMFs, so it is only necessary to find the marginal spectrum energy entropy of the first few IMF components of the vibration signal to describe the characteristics of the entire vibration signal.

[0103] Optionally, the entropy feature described in step 4 mainly refers to a feature used to describe the uncertainty and complexity of the vibration signal distribution, and may include power spectrum Shannon entropy, power spectrum exponential entropy, singular spectrum Shannon entropy, singular spectrum exponential entropy, and wavelet energy entropy.

[0104] Among them, according to different ways of calculating entropy, it can be divided into Shannon entropy and exponential entropy, where H1 is Shannon entropy and H2 is exponential entropy

[0105]

[0106]

[0107] Where N represents the total number of probabilities, p i Represents probability.

[0108] Among them, power spectrum refers to a function used to describe the power of vibration signal changes with frequency. First, the vibration signal sequence {x t ,t=1,2,…,N} perform fast Fourier transform FFT to obtain X(ω) and calculate the power spectrum density of the vibration signal Where t is the sampling point, ω is the corresponding frequency, and N is the number of sampling points

[0109]

[0110] After that, the power spectrum density is normalized to obtain the probability density function p(ω i )

[0111]

[0112] Finally, this probability density function is substituted into the above formula to obtain the power spectrum Shannon entropy and power spectrum exponential entropy respectively.

[0113] Among them, the singular spectrum entropy can obtain the local characteristics of the vibration signal in the time domain by performing singular value decomposition and other calculations on the vibration signal. First, the vibration signal sequence {x t ,t=1,2,…,N} is used for windowing processing. Windowing processing refers to the process of windowing the vibration signal sequence x with a length of M. t After truncation, the signal length is M, that is, the window size is M, the delay parameter is 1, where x t represents the vibration signal sequence, t represents the sampling point, M is the window size, and N is the vibration signal sequence x t The number of sampling points included. Thus, the trajectory matrix B of the vibration signal is obtained, which divides the vibration signal into (N-M+1)×M dimensions.

[0114]

[0115] Then, the trajectory matrix B is subjected to SVD singular value decomposition

[0116] B=U×S×V'

[0117] Where U and V' represent the orthogonal matrix of singular value decomposition, S is the diagonal matrix, δ i Represents singular values, matrix U is an orthogonal matrix of (N-M+1)×(N-M+1), matrix V' is an orthogonal matrix of M×M, and matrix S is composed of singular values ​​δ i It forms a diagonal matrix of (N-M+1)×M, and in the matrix S, the singular values ​​are arranged on the main diagonal of the matrix S from large to small, that is, δ1≥δ2≥…≥δ M ;

[0118] Finally, substitute the singular value into the following formula to obtain the normalized probability

[0119]

[0120] And substitute this probability into the formulas of Shannon entropy and exponential entropy above in turn to calculate the singular spectrum Shannon entropy and singular spectrum exponential entropy.

[0121] Among them, the wavelet packet in the wavelet energy entropy is a feature based on the time-frequency domain. First, the vibration signal {x t ,t=1,2,…,N} perform n-layer wavelet packet decomposition, and get 2 n nodes. At the same time, the wavelet packet coefficient W of each node in the nth layer can be obtained i , where i = 1, 2, ..., 2 n Then, the coefficients are squared to obtain the wavelet packet energy

[0122]

[0123] Finally, the wavelet packet energy is normalized to obtain the normalized probability, and the wavelet packet energy entropy is obtained from the normalized probability.

[0124]

[0125] Optionally, in step 5, the recursive feature elimination algorithm based on the support vector machine is a feature selection method based on the SVM margin maximization principle, which uses the weight coefficient of the support vector machine as the evaluation criterion. The feature with the lowest score is eliminated in each iteration until the feature set becomes an empty set or a specific number of features is obtained, and the iteration is stopped. The scoring criterion is as follows, where is the score of feature i, c represents the score, Represents the weight coefficient of SVM

[0126]

[0127] Optionally, the Euclidean distance in step 7 refers to the distance between two points x(x1, x2, .., x m ) and y(y1,y2,…,y m ), the specific calculation formula is as follows

[0128]

[0129] Among them, dist(X,Y) represents the actual distance between two points in m-dimensional space.

[0130] Figure 3 FIG2 shows a feature selection block diagram of an exemplary embodiment of a zero-sample rolling bearing fault diagnosis method based on attribute coding according to an embodiment of the present invention. Figure 3An exemplary embodiment of a zero-sample rolling bearing fault diagnosis method based on attribute coding provided according to an embodiment of the present invention is described in detail.

[0131] In an illustrative example applying an embodiment of the present invention, a rolling bearing, a core component of rotating machinery, was used as an example. Operating data from a rolling bearing with an artificially injected fault was used as experimental data for analysis. The relevant parameters are shown in Table 3. The accelerometer sampling frequency was 12 kHz, and the bearing speed was 1797 rpm. Fault diameters included 0.007 inch, 0.014 inch, and 0.021 inch. Fault types included inner race fault, rolling element fault, and outer race fault.

[0132] Table 3 Test bearing parameters

[0133]

[0134]

[0135] Figure 1 FIG1 shows a flow chart of a zero-sample rolling bearing fault diagnosis method based on attribute coding according to an embodiment of the present invention. Figure 1 A zero-sample rotating machinery fault diagnosis method based on attribute coding according to an embodiment of the present invention is described. The zero-sample rolling bearing fault diagnosis method based on attribute coding according to an embodiment of the present invention may include the following steps.

[0136] Step 1: Attribute description: Summarize all potential rolling bearing faults, analyze and disassemble the faults, determine the fault mode, define the fault scenario, and find the fault cause. Summarize and organize the relevant information of the above faults to build a fault attribute system, including fault conditions, fault location, and fault severity.

[0137] Step 2: Attribute encoding: Attribute encoding is performed for different types of rolling bearing faults, including known and unknown types. All categories are encoded using one-hot encoding to obtain 0-1 codes. The specific encoding is shown in Table 4. Each type of 0-1 code contains 9 dimensions.

[0138] Table 4 Data 0-1 code for each category of rolling bearings

[0139]

[0140] Step 3: Data acquisition: Sensors are placed at different locations on the rotating machinery to collect vibration signals from the rotating machinery at a sampling frequency of 12 kHz. Every 2048 sampling points of each type of sample data are used as a sample data set, and samples of categories 4 and 7 are treated as unknown category samples, that is, these two types of samples are not included in the classification of the classifier, and serve as the final test set;

[0141] Step 4: Form features: Figure 2 As shown in Figure 5, the time domain, frequency domain, energy and entropy features of each known category sample are extracted through domain conversion, thereby forming a multi-domain feature set containing 4 categories and 27 features. The specific feature set is shown in Table 5.

[0142] Table 5 Multi-domain feature set of known categories

[0143]

[0144] Step 5: Feature selection: Using a recursive feature elimination algorithm based on a support vector machine, the threshold for the number of feature subsets is set to 15. A feature selection is performed for each dimension of the 0-1 code, i.e., the feature selection algorithm is performed nine times, ultimately forming nine feature subsets. The feature score rankings for each dimension of the 0-1 code are shown in Table 6. According to the SVM-RFE algorithm's criteria, each round of feature selection for each dimension will eliminate the feature with the lowest score, and the remaining features will undergo a new round of iterative screening until 15 features remain in the feature set. That is, each dimension of the 0-1 code has a corresponding feature subset, and each feature subset contains 15 features. In Table 6, the number 1 represents the feature eliminated due to the lowest score in the first round of screening, the number 12 represents the feature eliminated in the twelfth or final round of screening, and the number 13 represents the features contained in the final feature subsets for each dimension.

[0145] Table 6 Ranking of feature scores of each dimension of 0-1 code

[0146]

[0147]

[0148] Step 6: Classifier training: Select support vector machine as the classifier, and use the feature subset corresponding to each dimension of the 0-1 code as the input of each support vector machine, such as Figure 3 As shown. Each dimension of the 0-1 code replaces the sample label as the output of the support vector machine for classifier training. Finally, 9 support vector machines will be trained, and each dimension of the 0-1 code corresponds to a trained support vector machine.

[0149] Step 7: Fault Diagnosis: For the sample data of unknown categories, a feature subset of the unknown category samples is formed according to Steps 3 and 4. This feature subset is used as the input of the trained classifier, ultimately obtaining a binary code output consisting of a series of 0-1 codes. This output is then sequentially compared with the 0-1 codes of each category in Step 1 using Euclidean distance. The diagnosis results are shown in Table 7. Table 7 also includes the results of two other sets of experiments, randomly selecting two types of samples as unknown fault categories: unknown fault categories 1 and 6, and unknown fault categories 7 and 8. The fault diagnosis accuracy is shown in Table 7.

[0150] Table 7 Fault diagnosis results

[0151]

[0152] Using this method, on-site personnel can diagnose unknown faults and accurately locate the bearing fault location. Based on the diagnostic results, appropriate measures can be taken to promptly maintain or replace the bearing, ensuring the safe and stable operation of the entire equipment.

[0153] This paper proposes a zero-sample rotating machinery fault diagnosis method based on attribute coding. This method requires minimal computation, exhibits strong real-time performance, and can diagnose faults in unknown sample data that was not trained. This model is applicable not only to rolling bearing fault diagnosis in this case study but also to other rotating machinery and related components, such as gearboxes. This invention has excellent scalability and provides valuable insights for others in this field.

[0154] It should be noted that the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment or part of the code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0155] Furthermore, the foregoing describes only some embodiments, which may be changed, modified, added and / or varied without departing from the scope and spirit of the disclosed embodiments, which are illustrative and not restrictive. Furthermore, the embodiments described relate to what are currently considered to be the most practical and preferred embodiments, and it should be understood that the embodiments should not be limited to the disclosed embodiments, but rather are intended to cover different modifications and equivalent arrangements that are included within the spirit and scope of the embodiments. Furthermore, the various embodiments described above may be used in conjunction with other embodiments, such as aspects of one embodiment may be combined with aspects of another embodiment to achieve yet another embodiment. Additionally, each independent feature or component of any given component may constitute another embodiment.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A zero-sample rotating machinery fault diagnosis method based on attribute coding, characterized in that The following steps are involved: Step 1: Determine the attribute description of the fault scenario, summarize all potential faults of rotating machinery, and perform disassembly analysis on the fault to determine the fault mode, find the fault cause, define the fault scenario, summarize and organize the relevant information of the above faults to build a fault attribute system. This attribute system includes the unique attributes and common attributes of different faults, and then determine the attribute description of the fault scenario, which provides support for subsequent attribute coding; Step 2: Establish attribute coding for fault scenarios. Attribute coding is performed for different faults of rotating machinery, including known and unknown faults. All fault categories are encoded using one-hot encoding to obtain 0-1 codes for different faults. Step 3: Acquire multi-source data of the rotating machinery by setting sensors at different locations on the rotating machinery to collect vibration signals of the rotating machinery; Step 4: Form a multi-domain feature set of the rotating machinery, extract the time domain features, frequency domain features, energy features and entropy features of the collected vibration signal of the rotating machinery through domain conversion, thereby forming a multi-domain feature set; Step 5: Construct a feature subset for attribute coding. Through the recursive feature elimination algorithm based on support vector machines, perform feature selection for each dimension of 0-1 code, select highly distinguishable features, eliminate useless features and redundant features, and finally form multiple feature subsets, and each dimension of 0-1 code has a corresponding feature subset; Step 6: Classifier training: Use the feature subset corresponding to the 0-1 code of each dimension as the input of the classifier, and use the 0-1 code of each dimension instead of the sample label as the output of the classifier for classifier training. Train multiple classifiers, and each dimension of the 0-1 code corresponds to a trained classifier. Step 7: Zero-sample fault diagnosis. For the sample data of unknown category faults, according to the composition of the feature subset of 0-1 codes of each dimension for known category faults in step 4, a feature subset of 0-1 codes of each dimension for unknown category fault samples is formed, and this feature subset is used as the input of the trained classifier. Finally, a string of binary code consisting of 0-1 is output, and this output is sequentially compared with the 0-1 codes of each category in step 1 based on Euclidean distance, finally realizing fault diagnosis of unknown category samples.

2. A zero-sample rotating machinery fault diagnosis method based on attribute coding as claimed in claim 1, characterized in that In step one: The failure mode includes the failure location and failure degree; The cause of the failure refers to the direct key factor causing the failure; The fault scenario includes the fault location, fault severity, fault impact, and fault operating conditions, i.e., the sum of the fault mode and fault cause; The unique attributes are related attributes that are different from each other in different fault categories, and the common attributes are related attributes that all fault categories have.

3. A zero-sample rotating machinery fault diagnosis method based on attribute coding as claimed in claim 1, characterized in that In step 2: The attribute coding is to encode the attributes of the rotating machinery, wherein if the attribute is present, the encoding dimension corresponding to the attribute is set to 1, and if the attribute is not present, the encoding dimension corresponding to the attribute is set to 0; One-hot encoding uses 0 and 1 to represent the attribute parameters of rotating machinery and uses N-dimensional binary code to encode N attributes of rotating machinery.

4. A zero-sample rotating machinery fault diagnosis method based on attribute coding as claimed in claim 1, characterized in that In step four: The time domain feature is to observe the change of the vibration signal in the time domain through time t, including dimensioned and dimensionless; Dimensional time domain features include maximum value, average value, RMS value, absolute average amplitude, RMS amplitude, peak-to-peak value, and standard deviation; Dimensionless time domain characteristics include kurtosis, skewness, shape factor, crest factor, impulse factor, and margin factor.

5. A zero-sample rotating machinery fault diagnosis method based on attribute coding as claimed in claim 1, characterized in that In step four: The frequency domain features include performing fast Fourier transform on the vibration signal and calculating frequency domain features, and the frequency domain features include center of gravity frequency CF, mean frequency MF, root mean square frequency RMSF, and frequency standard deviation RVF.

6. A zero-sample rotating machinery fault diagnosis method based on attribute coding as claimed in claim 1, characterized in that In step 4, the energy feature is the marginal spectrum energy, which is obtained as follows: First, the vibration signal is subjected to empirical mode decomposition to obtain the intrinsic mode function (IMF) of the vibration signal from high frequency to low frequency. Then, each IMF is subjected to Hilbert transform to obtain the Hilbert spectrum, where H i (ω,t) is the Hilbert spectrum of each IMF component, ω represents the frequency, and t represents the time corresponding to the frequency Next, the Hilbert spectrum H of each IMF component is i By integrating (ω, t) in the time domain, the marginal spectrum of each IMF component can be obtained Then, the marginal spectrum energy can be obtained by squaring the amplitude of the marginal spectrum Finally, the normalized probability is obtained by the following formula, where p(i) represents the normalized probability, and the marginal spectral energy entropy is obtained by the formula of Shannon entropy, where n represents the number of IMFs 7. A zero-sample rotating machinery fault diagnosis method based on attribute coding as claimed in claim 1, characterized in that The entropy feature described in step 4 is a feature used to describe the uncertainty and complexity of the vibration signal distribution, including power spectrum Shannon entropy, power spectrum exponential entropy, singular spectrum Shannon entropy, singular spectrum exponential entropy and wavelet energy entropy, which are obtained as follows: According to different calculation methods of entropy, it can be divided into Shannon entropy and exponential entropy, where H1 is Shannon entropy and H2 is exponential entropy. Where N represents the total number of probabilities, p i represents probability; Among them, the power spectrum is a function used to describe how the power of the vibration signal changes with frequency. First, the vibration signal sequence {x t ,t=1,2,…,N} perform fast Fourier transform FFT to obtain X(ω) and calculate the power spectrum density of the vibration signal Where t is the sampling point, ω is the corresponding frequency, and N is the number of sampling points After that, the power spectrum density is normalized to obtain the probability density function p(ω i ) Then, this probability density function p(ω i ) Substitute the above Shannon entropy and exponential entropy formulas to calculate the power spectrum Shannon entropy and power spectrum exponential entropy respectively; The singular spectrum entropy is calculated by performing singular value decomposition on the vibration signal to obtain the local characteristics of the vibration signal in the time domain. First, the vibration signal sequence {x t ,t=1,2,…,N} is used for windowing processing. Windowing processing refers to the process of windowing the vibration signal sequence x with a length of M. t After truncation, the signal length is M, that is, the window size is M, the delay parameter is 1, where x t represents the vibration signal sequence, t represents the sampling point, M is the window size, and N is the vibration signal sequence x t The number of sampling points included is used to obtain the trajectory matrix B of the vibration signal, which divides the vibration signal into (N-M+1)×M dimensions. Then, the trajectory matrix B is subjected to SVD singular value decomposition B=U×S×V' Where U and V' represent the orthogonal matrix of singular value decomposition, S is the diagonal matrix, δ i Represents singular values, matrix U is an orthogonal matrix of (N-M+1)×(N-M+1), matrix V' is an orthogonal matrix of M×M, and matrix S is composed of singular values ​​δ i It forms a diagonal matrix of (N-M+1)×M, and in the matrix S, the singular values ​​are arranged on the main diagonal of the matrix S from large to small, that is, δ1≥δ2≥…≥δ M , Then, substitute the singular value into the following formula to find the normalized probability Substitute this probability into the above formulas of Shannon entropy and exponential entropy to obtain the singular spectrum Shannon entropy and singular spectrum exponential entropy; The wavelet packet in wavelet energy entropy is a feature based on the time-frequency domain. First, the vibration signal {x t ,t=1,2,…,N} perform n-layer wavelet packet decomposition, and get 2 n nodes, and at the same time obtain the wavelet packet coefficient W of each node in the nth layer i , where i = 1, 2, ..., 2 n , Then, the coefficients are squared to obtain the wavelet packet energy Then, the wavelet packet energy is normalized to obtain the normalized probability, and the wavelet packet energy entropy is obtained from the normalized probability.

8. The zero-sample rotating machinery fault diagnosis method based on attribute coding according to claim 1 is characterized in that The recursive feature elimination algorithm based on support vector machine in step 5 is a feature selection method based on the principle of SVM margin maximization. The steps are as follows: The weight coefficient of the support vector machine is used as the evaluation criterion. The feature with the lowest score is eliminated in each iteration until the feature set becomes an empty set or a specific number of features is obtained, and the iteration is stopped. The scoring criterion is as follows: represents the score of feature i, c represents the score, Represents the weight coefficient of SVM 9. A zero-sample rotating machinery fault diagnosis method based on attribute coding as claimed in claim 1, characterized in that The Euclidean distance in step 7 is the distance between two points x(x1,x2,..,x m ) and y(y1,y2,…,y m ), the actual distance between Among them, dist(X,Y) represents the actual distance between two points in m-dimensional space.

Images