Rotary machinery intelligent fault diagnosis method and system based on signals

Abstract

The invention discloses a signal-based rotating machinery intelligent fault diagnosis method and system, and the method comprises the steps: 1, obtaining vibration signals of a bearing in different states, and carrying out the labeling of the vibration signals, and forming an original data set; 2, performing feature extraction on the original data set to obtain a feature data set; 3, processing the feature data set by adopting a feature selection evolutionary algorithm, reducing the dimensionality of the feature data set, and obtaining a key fault feature combination; and 4, identifying and classifying the to-be-identified fault according to the key feature combination by using a K-nearest neighbor classification technology, and determining a fault category. According to the method, more accurate and comprehensive feature information can be provided for fault diagnosis through effective preprocessing, feature enhancement and multi-angle analysis, the classification precision of fault modes is improved, and a foundation is laid for a subsequent feature selection part.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a signal-based intelligent fault diagnosis method and system for rotating machinery. Background Art

[0002] Rotating machinery is widely used in power plants, manufacturing, petrochemical, transportation and other fields. Generally speaking, once these equipments fail, it will not only affect the normal production process of the enterprise, but also cause serious accidents with machine damage and death, which will cause serious economic losses. Therefore, the fault diagnosis technology of rotating machinery has become an important research direction in the field of industrial maintenance. Among them, traditional knowledge-based fault diagnosis methods mostly diagnose the system through manually defined rules, but when the system becomes complex, a certain diagnostic accuracy cannot be guaranteed. Therefore, intelligent fault diagnosis methods using artificial intelligence technologies such as machine learning and deep learning have gradually become the main trend in the industrial and engineering fields. Intelligent fault diagnosis is a reasoning process that simulates human thinking. By effectively acquiring, transmitting and processing diagnostic information, it can simulate human experts and make intelligent judgments and decisions on the operating status and faults of the monitored object with flexible diagnostic strategies. At present, signal-based intelligent fault diagnosis methods have received widespread attention in the field of rotating machinery due to their wide applicability and strong anti-interference ability.

[0003] Signal-based methods have three main steps: feature extraction, feature selection, and fault classification. Among them, extracting features from vibration signals is a crucial step in the fault diagnosis process. Vibration signals usually reflect the operating status of mechanical systems and can provide important information about the health of equipment. Useful fault feature information is often drowned out by heavy background noise. Therefore, advanced signal processing technology must be mastered to effectively extract features. Effective feature extraction technology plays a key role in intelligent fault diagnosis. Effective feature extraction technology is not only the basis for building high-performance models, but also the key to achieving data-driven decision-making. With the continuous increase in the amount and complexity of data in fault diagnosis, it will be particularly important to study and apply advanced feature extraction technology. This will provide strong support for intelligent analysis and fault diagnosis in the field of rotating machinery.

[0004] In the process of signal processing, vibration signals usually have complex nonlinear characteristics and may change over time (i.e., non-stationary), which makes it difficult for traditional signal analysis methods to effectively extract useful information. In addition, the collected vibration signals often contain various forms of noise, such as environmental noise, electronic interference, etc. These noises will mask the effective signals and reduce the accuracy of the analysis. Therefore, it is very important to use efficient signal processing tools. In addition, traditional methods often rely on the analysis of a single field (time domain or frequency domain), which makes it difficult to fully reflect the operating status of the system and may not be able to effectively cope with complex signal analysis tasks. A variety of signal processing methods are needed to ensure that the extracted features are representative and effective. Therefore, how to use signal processing tools to conduct a comprehensive analysis of vibration signals from multiple angles to obtain effective fault feature information has become a fault diagnosis problem that needs to be solved urgently. Summary of the invention

[0005] In view of the deficiencies in the prior art, an object of the present invention is to provide a signal-based intelligent fault diagnosis method and system for rotating machinery, which can achieve fault diagnosis with higher classification accuracy.

[0006] The technical solution adopted by the present invention is: a signal-based intelligent fault diagnosis method for rotating machinery, comprising the following steps:

[0007] Step 1: Obtain vibration signals of the bearing under different conditions, label the vibration signals, and form an original data set;

[0008] Step 2, extracting features from the original data set to obtain a feature data set;

[0009] Step 3: Use a feature selection evolutionary algorithm to process the feature data set, reduce the dimension of the feature data set, and obtain a key fault feature combination;

[0010] Step 4: Use K-nearest neighbor classification technology to identify and classify the faults to be identified based on the key feature combination to determine the fault category.

[0011] Further, the step 2 comprises the specific steps of:

[0012] Step 2.1, using variational mode decomposition to decompose the vibration signal and divide all components into high-frequency components and low-frequency components;

[0013] Step 2.2, weighted reconstruction is performed according to the correlation coefficient between the component and the original vibration signal;

[0014] Step 2.3, extracting the time domain, frequency domain and time-frequency domain features of the high-frequency and low-frequency reconstructed signals to form a fault diagnosis feature set;

[0015] In step 2.4, the time domain features, frequency domain features, wavelet packet energy entropy and multi-scale permutation entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal together constitute a feature data set with a total of 56 dimensions; each column of the feature data set represents a feature, each row represents a sample, and corresponds to a class label.

[0016] Further, the specific steps of step 2.1 are:

[0017] Step 2.1.1, use variational mode decomposition to decompose the vibration signal, where the penalty factor of variational mode decomposition is set to 2000, the number of components is set to 10, and u k = {u1, u2, ..., u k}(k=1, 2, ..., 10) are the signal components;

[0018] Step 2.1.2, add the first i components and record them as Q(i), that is

[0019] Q(i)=u1+u2+...+u i (i=1, 2, ..., 10);

[0020] Calculate the mean of Q(i) and perform a T test to see whether its mean is significantly different from 0. The statistic of the T test is:

[0021]

[0022] In the formula, is the mean of Q(i), σ i is the standard deviation of Q(i), n is the length of Q(i);

[0023] Step 2.1.3, when i = m, Q(i) is significantly different from 0, then the high-frequency component combination and the low-frequency component combination are u h = {u1, u2, ..., u m} and u l = {u m+1 ,u m+2 , ..., u k}.

[0024] Further, the specific steps of step 2.2 are:

[0025] Step 2.2.1, calculate the correlation coefficient ρ between each component in the high-frequency combination and the low-frequency combination and the original signal x(t) according to formula (2) and formula (3): h ={ρ1, ρ2, ..., ρ m} and ρ l ={ρ m+1 , ρ m+2 , ..., ρ k}:

[0026]

[0027] In the formula, u h and u h are high frequency components and low frequency components respectively, x(t) is the original signal cov(u h , x(t)) is u h and the covariance of x(t), and δ x(t) u h and the standard deviation of x(t);

[0028] Step 2.2.2, the proportion of its correlation coefficient in the entire combination is used as the weight for reconstruction; the weight of the high-frequency combination γ h ={γ1,γ2,...,γ m} and the weight of the low-frequency combination γ l ={γ m+1 , γ m+2 , ..., γ k} are calculated according to formula (4) and formula (5) respectively:

[0029]

[0030] Step 2.2.3, during reconstruction, different weights are assigned to each component according to its importance; the high-frequency and low-frequency combinations are reconstructed using formula (6) and formula (7) respectively:

[0031] x' h (t) = γ1×u1+γ2×u2+...+γ m ×u m (6)

[0032] x′ l (t) = γ m+1 ×u m+1 +γ m+2 ×u m+2 +...+γ k ×u k (7).

[0033] Further, the specific steps of step 2.3 are:

[0034] Step 2.3.1, extracting 10-dimensional time domain features of the high-frequency reconstructed signal and the low-frequency reconstructed signal, including mean, standard deviation, skewness, kurtosis, peak, root mean square, shape factor, impulse factor, margin factor and peak factor;

[0035] Step 2.3.2, extracting the 4-dimensional frequency domain features of the high-frequency reconstructed signal and the low-frequency reconstructed signal, including the average frequency, center frequency, root mean square frequency and frequency standard deviation;

[0036] Step 2.3.3, extract the wavelet packet energy entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal as features; decompose the signal into three layers of wavelet packets, and use "db4" as the wavelet basis function. The number of nodes in the last layer is 2 3 =8, so the dimension of energy entropy is 8; the specific steps are:

[0037] First, calculate the energy value of each decomposed component; then calculate the proportion of the energy of each component to the total energy to obtain the probability density function P(x); finally, substitute the probability density function into formula (8) to obtain the energy entropy:

[0038]

[0039] Step 2.3.4, respectively extract the multi-scale permutation entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal as features: First, according to formula (9), the time series x(i), i = 1, 2, ... N, is multi-scale coarse-grained to obtain the coarse-grained sequence y j ;

[0040]

[0041] Then, the permutation entropy value of the rough physical and chemical sequence under each scale factor τ (τ = 1, 2, ..., 6) is calculated, for a total of 6 dimensions.

[0042] The present invention also provides an intelligent fault diagnosis system, comprising:

[0043] A data acquisition module is configured to acquire vibration signals of bearings in different states, label the vibration signals, and form an original data set;

[0044] The feature extraction module is configured to extract features from the original data set to obtain a feature data set;

[0045] A feature selection module is configured to process a feature data set using a feature selection evolutionary algorithm to obtain a key feature combination for fault diagnosis;

[0046] The fault diagnosis module is configured to use the K nearest neighbor classification method to identify and classify the fault to be identified according to the key feature combination to determine the fault category.

[0047] The beneficial effects of the present invention are:

[0048] 1. The present invention can provide more accurate and comprehensive feature information for fault diagnosis through effective preprocessing, feature enhancement and multi-angle analysis, which not only improves the classification accuracy of fault modes, but also lays the foundation for the subsequent feature selection part.

[0049] 2. The present invention uses variational mode decomposition to process vibration signals and divides components according to frequency, which helps to reduce the impact of noise and improve the analyzability of signals. Introducing correlation coefficients in the reconstruction process can further optimize the reconstruction effect, highlight the characteristic information related to the fault, and thus improve the accuracy of fault classification.

[0050] 3. The invention extracts features from multiple angles to help understand the characteristics of the signal more deeply and reflect the fault mode more comprehensively. A more comprehensive feature set can make the classification model more discerning.

[0051] 4. The vibration signals based on different experimental platforms can effectively extract features, which not only has high classification accuracy but also greatly reduces the cost of fault diagnosis.

[0052] 5. Compared with several existing feature extraction methods, the fault diagnosis method of the present invention has higher classification accuracy and can be widely used in intelligent fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0054] The present invention is further described below in conjunction with the accompanying drawings. It should be noted that the following detailed descriptions are exemplary and are intended to provide further description of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those generally understood by those of ordinary skill in the art to which the present invention belongs. It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0055] like Figure 1 As shown, the present invention discloses a signal-based intelligent fault diagnosis method for rotating machinery, comprising the following steps:

[0056] Step 1, obtain the vibration signals of the bearing in different states, label the vibration signals, and form the original data set; specifically: the vibration signal of the rotating machinery in the present invention comes from a fault diagnosis test of a centrifugal fan system using a Mitsubishi SB-JR induction motor. The motor is a 3.7KW three-phase induction motor with a voltage of 220V, a motor pole number of 4, a rated speed of 1800rpm, a rotor supported by two bearings, and a defective bearing in the entire experimental device. It is worth noting that the defective bearing is located on the output shaft of the motor. The speed varies between 400rpm / min and 800rpm / min, and the signal is measured by changing the optional input voltage of the motor. The load (i.e., the torque applied to the rotating shaft) can also be changed by using a tool to manually randomly rub the synchronization band between the motor and the rotating shaft when measuring the data. In the experiment, an accelerometer with a bandwidth of 5Hz to 60kHz and an output of 10mV / g was used to measure the vertical vibration signals in normal, outer ring fault, inner ring fault and rolling lift fault states.

[0057] In the experiment, rolling bearings were used, and the defects were outer ring fault, inner ring fault and rolling element fault. All faults were artificially made by wire cutting. The test bearings were of two types, N205 and NU205. Among them, N205 bearings were used for normal, outer ring fault and rolling element fault; NU205 bearings with detachable outer ring were used for inner ring fault. When collecting signals, the sampling frequency was 50kHz and the speed was 800rpm / min. The bearings were tested in four different states: normal, outer ring fault, inner ring fault and rolling element fault. After sliding window processing, a total of 1020 samples with a length of 20,000 were obtained as the original data set x(t). Each row of the original data set corresponds to a class label, that is, a fault state, and each column is the signal data collected under the fault state.

[0058] Step 2: Extract features from the original data set. The specific steps to obtain the feature data set are:

[0059] Step 2.1, use variational mode decomposition to decompose the vibration signal and divide all components into high-frequency components and low-frequency components. The specific steps are:

[0060] Step 2.1.1, use variational mode decomposition to decompose the vibration signal, where the penalty factor of variational mode decomposition is set to 2000, the number of components is set to 10, and u k = {u1, u2, ..., u k}(k=1, 2, ..., 10) are the signal components;

[0061] Step 2.1.2, add the first i components and record them as Q(i), that is, Q(i) = u1+u2+...+ui (i=1, 2, ..., 10);

[0062] Then calculate the mean of Q(i) and perform a T test to see whether its mean is significantly different from 0; the statistic of the t test is:

[0063]

[0064] In the formula, is the mean of Q(i), σ i is the standard deviation of Q(i), n is the length of Q(i);

[0065] Step 2.1.3, when Q(i) is significantly different from 0, it proves that the first i components are high-frequency components and the remaining ones are low-frequency components; for example, when i=m, Q(i) is significantly different from 0, then the high-frequency component combination and the low-frequency component combination are u respectively. h = {u1, u2, ..., u m} and u l = {u m+1 ,u m+2 , ..., u k}.

[0066] Step 2.2, weighted reconstruction is performed according to the correlation coefficient between the component and the original vibration signal. The specific steps are:

[0067] Step 2.2.1, calculate the correlation coefficient ρ between each component in the high-frequency combination and the low-frequency combination and the original signal x(t) according to formula (2) and formula (3): h ={ρ1, ρ2, ..., ρ m} and ρ l ={ρ m+1 , ρ m+2 , ..., ρ k}:

[0068]

[0069] In the formula, u h and u h are high frequency components and low frequency components respectively, x(t) is the original signal cov(u h , x(t)) is u h and the covariance of x(t), and δ x(t) u h and the standard deviation of x(t);

[0070] Step 2.2.2, the proportion of its correlation coefficient in the entire combination is used as the weight for reconstruction; the weight of the high-frequency combination γ h ={γ1,γ2,...,γm} and the weight of the low-frequency combination γ l ={γ m+1 , γ m+2 , ..., γ k} are calculated according to formula (4) and formula (5) respectively:

[0071]

[0072] Step 2.2.3, during reconstruction, different weights are assigned to each component according to its importance; the high-frequency and low-frequency combinations are reconstructed using formula (6) and formula (7) respectively:

[0073] x′ h (t) = γ1×u1+γ2×u2+...+γ m ×u m (6)

[0074] x′ l (y) = γ m+1 ×u m+1 +γ m+2 ×u m+2 +...+γ k ×u k (7).

[0075] Step 2.3, extract the time domain, frequency domain and time-frequency domain features of the high-frequency and low-frequency reconstructed signals to form a fault diagnosis feature set. The specific steps are:

[0076] In step 2.3.1, 10-dimensional time domain features of high-frequency reconstructed signals and low-frequency reconstructed signals are extracted respectively, as shown in Table 1. Where x is the signal to be analyzed, and N is the length of each signal.

[0077] Table 1 Time domain characteristics

[0078]

[0079] In step 2.3.2, the 4-dimensional frequency domain features of the high-frequency reconstructed signal and the low-frequency reconstructed signal are extracted respectively, as shown in Table 2. Where f(i) and s(i) represent the spectrum amplitude and the corresponding frequency respectively, and K is the number of spectral lines.

[0080] Table 2 Frequency domain characteristics

[0081]

[0082]

[0083] Step 2.3.3, extract the wavelet packet energy entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal as features; decompose the signal into three layers of wavelet packets, and use "db4" as the wavelet basis function. The number of nodes in the last layer is 23 =8, so the dimension of energy entropy is 8. The specific steps are:

[0084] First, calculate the energy value of each decomposed component; then calculate the proportion of the energy of each component to the total energy to obtain the probability density function P(x); finally, substitute the probability density function into formula (8) to obtain the energy entropy:

[0085]

[0086] Step 2.3.4, respectively extract the multi-scale permutation entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal as features: First, according to formula (9), the time series x(i), i = 1, 2, ... N, is multi-scale coarse-grained to obtain the coarse-grained sequence y j ;

[0087]

[0088] Then, the permutation entropy value of the rough physical and chemical sequence under each scale factor τ (τ = 1, 2, ..., 6) is calculated, for a total of 6 dimensions.

[0089] In step 2.4, the time domain features, frequency domain features, wavelet packet energy entropy and multi-scale permutation entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal together constitute a feature data set with a total of 56 dimensions; each column of the feature data set represents a feature, each row represents a sample, and corresponds to a class label.

[0090] Step 3: Use the feature selection evolutionary algorithm to process the feature data set, reduce the dimension of the feature data set, and obtain the key fault feature combination.

[0091] Step 4: Use K-nearest neighbor classification technology to identify and classify the faults to be identified based on the key feature combination to determine the fault category.

[0092] In step 4, when the fault occurs again, detect which fault has occurred in the experimental platform, extract the fault feature from the signal according to the feature extraction method, and select the key fault feature of the experimental platform. Then, based on the key fault feature data, use the well-known K nearest neighbor classification technology to obtain the classification label corresponding to the fault feature from the public data set corresponding to the experimental platform, where the distance value K in the K nearest neighbor classification technology is 5, and the classification label obtained at this time is the fault state of the experimental platform.

[0093] The present invention discloses a signal-based intelligent fault diagnosis method for rotating machinery, which provides more accurate and comprehensive feature information for fault diagnosis through effective preprocessing, feature enhancement and multi-angle analysis. This not only improves the classification accuracy of fault modes, but also lays the foundation for the subsequent feature selection part.

[0094] In order to verify the performance of the rotating machinery intelligent fault diagnosis method of the present invention, the rotating machinery intelligent fault diagnosis method of the present invention is first used to perform bone fault diagnosis on the JNU data set, and compared with several other feature extraction methods. The features extracted by different methods are classified using the K nearest neighbor classification technology, and the obtained classification accuracy is used as an indicator for comparison. The specific comparison data is shown in Table 3.

[0095] Table 3 Comparison of classification accuracy of the present invention and other feature extraction methods

[0096] Dataset DWTVFD EEMDBLP MOFD TFWPD UNFD Method of the present invention JNU 96.74％ 87.68％ 94.53％ 91.22％ 94.31％ 98.10％

[0097] In Table 3, DWTVFD is "Single-and Multi-Fault Diagnosis Using Machine Learning for Variable Frequency Drive-Fed Induction Motors", which uses machine learning to perform single-fault and multi-fault diagnosis on variable frequency drive-fed induction motors;

[0098] EEMDBLP is "Euclidean distance based feature ranking and subset selection for bearing fault diagnosis", which is based on Euclidean distance feature ranking and subset selection for bearing fault diagnosis;

[0099] MOFD is "Fault diagnosis of planetary gearbox using multi-criteria feature selection and heterogeneous ensemble learning classification", which is planetary gearbox fault diagnosis based on multi-criteria feature selection and heterogeneous ensemble learning classification;

[0100] TFWPD is "Vibration-based fault diagnosis for railway point machines using multi-domain features, ensemble feature selection and SVM", which uses multi-domain features, ensemble feature selection and SVM to perform vibration-based fault diagnosis for railway point machines.

[0101] UNFD stands for "A modified active learning intelligent fault diagnosis method for rolling bearings with unbalanced samples", an improved active learning intelligent fault diagnosis method for rolling bearings with unbalanced samples.

[0102] From the comparison results, it can be seen that the fault diagnosis method proposed in the present invention can obtain higher classification accuracy, which shows that the fault diagnosis method of the present invention not only effectively separates the high-frequency and low-frequency fault features in the signal, but also enhances the expression of fault information through correlation coefficient weighted reconstruction, thereby significantly improving the accuracy and reliability of fault diagnosis.

[0103] There is still some irrelevant or redundant information in the data set obtained after feature extraction. Therefore, it is necessary to select the most representative features to improve the generalization ability of the classification model and make it more accurate in fault diagnosis. In order to further verify that the features extracted by the method in this paper have greater potential, feature selection is performed on the feature sets obtained by different feature extraction methods. Table 4 shows the average classification accuracy of the features extracted by different feature extraction methods after multiple feature selection methods. Among them, the Wilcoxon rank sum test with a significance level of 0.05 is used to evaluate the performance of the method in 30 independent runs. In the following results, "+", "-" and "=" indicate that the proposed method has obvious advantages, disadvantages and similarities compared with the compared method.

[0104] Table 4 Classification accuracy obtained by using different feature selection methods for the present invention and other feature extraction methods

[0105]

[0106] In Table 4, HFS-CP is “A Fast Hybrid Feature Selection Based on Correlation-Guided Clustering and Particle Swarm Optimization for High-Dimensional Data”, a fast hybrid feature selection for high-dimensional data based on correlation-guided clustering and particle swarm optimization;

[0107] VSCCPSO stands for "Variable-Size Cooperative Coevolutionary Particle Swarm Optimization for Feature Selection on High-Dimensional Data", variable-size cooperative coevolutionary particle swarm optimization for high-dimensional data feature selection;

[0108] CGGA stands for "A correlation guided genetic algorithm and its application to feature selection", a correlation guided genetic algorithm and its application to feature selection;

[0109] MEL stands for "MEL: Efficient Multi-Task Evolutionary Learning for High-Dimensional Feature Selection", efficient multi-task evolutionary learning based on high-dimensional feature selection;

[0110] LAGAM stands for "LAGAM: A Length-Adaptive Genetic Algorithm With Markov Blanketfor High-Dimensional Feature Selection in Classification", a length-adaptive genetic algorithm for high-dimensional feature selection and classification based on Markov blanket.

[0111] From the experimental results, it can be found that the classification accuracy of the features extracted by the present invention exceeds that of all other feature extraction methods after the feature selection algorithm. This phenomenon shows that the method in this paper combines with the evolutionary algorithm to generate high-quality feature combinations, which can effectively optimize the feature space, so that important information is retained and irrelevant or redundant information is eliminated; it also reflects the effectiveness of the evolutionary algorithm as a feature selection, indicating that it plays a positive role in the evaluation and selection process of features.

[0112] The present invention also provides an intelligent fault diagnosis system for processing the above method, comprising:

[0113] A data acquisition module is configured to acquire vibration signals of bearings in different states, label the vibration signals, and form an original data set;

[0114] The feature extraction module is configured to extract features from the original data set to obtain a feature data set;

[0115] A feature selection module is configured to process a feature data set using a feature selection evolutionary algorithm to obtain a key feature combination for fault diagnosis;

[0116] The fault diagnosis module is configured to use the K nearest neighbor classification method to identify and classify the fault to be identified according to the key feature combination to determine the fault category.

[0117] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0118] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A signal-based intelligent fault diagnosis method for rotating machinery, characterized in that: The following steps are involved: Step 1: Obtain vibration signals of the bearing under different conditions, label the vibration signals, and form an original data set; Step 2, extracting features from the original data set to obtain a feature data set; Step 3: Use a feature selection evolutionary algorithm to process the feature data set, reduce the dimension of the feature data set, and obtain a key fault feature combination; Step 4: Use K-nearest neighbor classification technology to identify and classify the faults to be identified based on the key feature combination to determine the fault category.

2. The signal-based intelligent fault diagnosis method for rotating machinery according to claim 1, characterized in that: The vibration signals include vibration signals of four bearing states: normal, inner ring fault, outer ring fault and rolling element fault.

3. The signal-based intelligent fault diagnosis method for rotating machinery according to claim 1, characterized in that: The step 2 includes the following specific steps: Step 2.1, using variational mode decomposition to decompose the vibration signal and divide all components into high-frequency components and low-frequency components; Step 2.2, weighted reconstruction is performed according to the correlation coefficient between the component and the original vibration signal; Step 2.3, extracting the time domain, frequency domain and time-frequency domain features of the high-frequency and low-frequency reconstructed signals to form a fault diagnosis feature set; In step 2.4, the time domain features, frequency domain features, wavelet packet energy entropy and multi-scale permutation entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal together constitute a feature data set with a total of 56 dimensions; each column of the feature data set represents a feature, each row represents a sample, and corresponds to a class label.

4. The signal-based intelligent fault diagnosis method for rotating machinery according to claim 2, characterized in that: The specific steps of step 2.1 are: Step 2.1.1, use variational mode decomposition to decompose the vibration signal, where the penalty factor of variational mode decomposition is set to 2000, the number of components is set to 10, and u k ={u1,u2,...,u k }(k=1,2,...,10) are the signal components; Step 2.1.2, add the first i components and record them as Q(i), that is Q(i)=u1+u2+...+u i (i=1,2,...10); Calculate the mean of Q(i) and perform a T test to see whether its mean is significantly different from 0. The statistic of the T test is: In the formula, is the mean of Q(i), σ i is the standard deviation of Q(i), n is the length of Q(i); Step 2.1.3, when i = m, Q(i) is significantly different from 0, then the high-frequency component combination and the low-frequency component combination are u h ={u1,u2,...,u m } and u l = {u m+1 ,u m+2 ,...,u k }.

5. The signal-based intelligent fault diagnosis method for rotating machinery according to claim 2, characterized in that: The specific steps of step 2.2 are: Step 2.2.1, calculate the correlation coefficient ρ between each component in the high-frequency combination and the low-frequency combination and the original signal x(t) according to formula (2) and formula (3): h ={ρ1,ρ2,...,ρ m } and ρ l ={ρ m+1 ,ρ m+2 ,...,ρ k }: In the formula, u h and u h are high frequency components and low frequency components respectively, x(t) is the original signal cov(u h ,x(t)) is u h and the covariance of x(t), and δ x(t) u h and the standard deviation of x(t); Step 2.2.2, the proportion of its correlation coefficient in the entire combination is used as the weight for reconstruction; the weight of the high-frequency combination γ h ={γ1,γ2,...,γ m } and the weight of the low-frequency combination γ l ={γ m+1 ,γ m+2 ,...,γ k } are calculated according to formula (4) and formula (5) respectively: Step 2.2.3, during reconstruction, different weights are assigned to each component according to its importance; the high-frequency and low-frequency combinations are reconstructed using formula (6) and formula (7) respectively: x' h (t)=γ1×u1+γ2×u2+...+γ m ×u m (6) x′ l (t)=γ m+1 ×u m+1 +g m+2 ×u m+2 +...+c k ×u k (7)。 6. The signal-based intelligent fault diagnosis method for rotating machinery according to claim 2, characterized in that: The specific steps of step 2.3 are: Step 2.3.1, extracting 10-dimensional time domain features of the high-frequency reconstructed signal and the low-frequency reconstructed signal, including mean, standard deviation, skewness, kurtosis, peak, root mean square, shape factor, impulse factor, margin factor and peak factor; Step 2.3.2, extracting the 4-dimensional frequency domain features of the high-frequency reconstructed signal and the low-frequency reconstructed signal, including the average frequency, center frequency, root mean square frequency and frequency standard deviation; Step 2.3.3, extract the wavelet packet energy entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal as features; decompose the signal into three layers of wavelet packets, and use "db4" as the wavelet basis function. The number of nodes in the last layer is 2. 3 =8, so the dimension of energy entropy is 8; the specific steps are: First, calculate the energy value of each decomposed component; then calculate the proportion of the energy of each component to the total energy to obtain the probability density function P(x); finally, substitute the probability density function into formula (8) to obtain the energy entropy: Step 2.3.4, extract the multi-scale permutation entropy of the high-frequency reconstructed signal and the low-frequency reconstructed signal as features: First, according to formula (9), the time series x(i), i = 1, 2, ... N is multi-scale coarse-grained to obtain the coarse-grained sequence y j ; Then, the permutation entropy of the rough physical and chemical sequence is calculated under each scale factor τ (τ = 1, 2, ..., 6), for a total of 6 dimensions.

7. An intelligent fault diagnosis system, characterized in that: include: A data acquisition module is configured to acquire vibration signals of bearings in different states, label the vibration signals, and form an original data set; The feature extraction module is configured to extract features from the original data set to obtain a feature data set; A feature selection module is configured to process a feature data set using a feature selection evolutionary algorithm to obtain a key feature combination for fault diagnosis; The fault diagnosis module is configured to use the K nearest neighbor classification method to identify and classify the fault to be identified according to the key feature combination to determine the fault category.

Images