Rotating machine fault diagnosis method and system

Abstract

The invention provides a rotating machine fault diagnosis method and system, and the method comprises the steps: collecting monitoring signals of a rotating machine in different fault states in a certain time period through a plurality of different types of sensors, and converting the monitoring signals into time-frequency images through a constant Q conversion method; performing graying, cutting and normalization processing in sequence to obtain a plurality of normalized pixel matrixes, and stacking each pixel matrix along the channel dimension of the tensor to obtain a plurality of third-order tensor samples in each fault state; then obtaining an intuitionistic fuzzy score of each three-order tensor sample by adopting an exponential membership function and an exponential non-membership function, extracting a low-rank feature tensor sample by utilizing a tensor robust principal component analysis method so as to construct a high-dimensional feature matrix, and finally establishing a nonlinear intuitionistic fuzzy twin support tensor machine model to obtain a solution vector; and a decision function is constructed to obtain a prediction label of each test sample in the test set, and fault diagnosis of the rotating machinery is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical diagnosis, and in particular to a rotating machinery fault diagnosis method and system. Background Art

[0002] As a key equipment in the field of modern intelligent manufacturing, rotating machinery has been widely used in the fields of energy, transportation, and national defense. Due to the complexity and high coupling of its system structure, once a component fails in a rotating machinery, it may affect the normal operation of the entire equipment. Therefore, in order to ensure the long-term safe, efficient and reliable operation of rotating machinery and equipment, it is of great engineering significance to carry out health status monitoring and intelligent diagnosis of rotating machinery and its key components. Fault diagnosis mainly includes three stages: sensor signal acquisition, feature extraction and selection, and fault classification. Faulty machinery will produce periodic shocks during operation, so that the collected monitoring signals often show strong non-stationary characteristics. Therefore, it is necessary to extract fault-related features from the original non-stationary signal for fault diagnosis.

[0003] As a commonly used feature extraction method, the time-frequency analysis method can reveal the non-stationary characteristics of the signal in the time domain and frequency domain, and help to identify the changing laws of different frequency components. As a commonly used time-frequency analysis method, the constant Q transform can achieve higher time resolution at higher frequencies and higher frequency resolution at lower frequencies, better capture the transient characteristic information of the signal, and help to achieve high-performance fault diagnosis. After feature extraction, a diagnostic method is needed to distinguish faults. Intelligent fault diagnosis methods can automatically identify and diagnose fault categories and have received widespread attention, such as convolutional neural networks, graph convolutional networks, deep belief networks, transfer learning, fuzzy inference systems based on adaptive networks, etc. In fact, fault data is difficult to collect, but most methods rely on sufficient data, so there are certain limitations for small samples. Support vector machines are based on the principle of structural risk minimization and the idea of ​​interval maximization. They can not only handle small sample problems, but also have strong generalization performance and have been widely used in the field of fault diagnosis.

[0004] However, the information provided by a single sensor signal is limited and cannot fully reflect the actual operating status of the rotating machinery. Multiple sensors arranged in different positions and directions collect multiple signals, which can provide richer equipment operating status information and effectively improve the accuracy and reliability of diagnosis. In order to combine multiple sensors for fault diagnosis, it is necessary to use a time-frequency analysis method to convert each sensor signal into a corresponding time-frequency image, thereby obtaining multiple time-frequency images. However, the support vector machine method usually pulls multiple time-frequency image pixel matrices into a long vector as the input of the model. This vectorization operation not only destroys the potential correlation information between different sensor signals, but also generates high-dimensional feature vectors, resulting in increased computational complexity, further triggering overfitting problems, and resulting in decreased diagnostic performance. In addition, the collected multi-sensor signals contain a lot of noise, which also brings challenges to rotating machinery fault diagnosis. Summary of the invention

[0005] In order to solve the problems that the current rotating machinery fault diagnosis process cannot effectively extract the time-varying features of multiple sensors and is easily affected by noise, resulting in low diagnostic accuracy and computational efficiency, the present invention provides a rotating machinery fault diagnosis method, which uses a constant Q transformation method to convert the monitoring signals collected by multiple sensors into a time-frequency diagram and performs grayscale, clipping, and normalization processing, and then stacks and reconstructs them into third-order tensor samples along the channel dimension of the tensor, and designs a non-linear non-membership function to construct an intuitive fuzzy set to obtain the intuitive fuzzy score of each tensor sample, and adopts a tensor robust principal component analysis method and establishes a nonlinear intuitive fuzzy twin support tensor machine model, and realizes the fault diagnosis of rotating machinery by solving two relatively small quadratic programming problems, which effectively improves the performance and computational efficiency of rotating machinery fault diagnosis. The present invention also relates to a rotating machinery fault diagnosis system.

[0006] The technical solution of the present invention is as follows:

[0007] A method for diagnosing a rotating machinery fault, characterized in that it comprises the following steps:

[0008] Monitoring signal collection step: using multiple sensors of different types to collect monitoring signals of rotating machinery under different fault states within a certain period of time;

[0009] Steps for time-frequency graph conversion and third-order tensor sample acquisition: divide the monitoring signals collected by all sensors at a certain number of time points in each fault state during the time period into one sample to obtain multiple samples under each fault state; then use the constant Q transformation method to convert each monitoring signal in each sample into a time-frequency graph, and grayscale each time-frequency graph to obtain multiple grayscale graphs; then crop each grayscale graph to the same size, and perform pixel normalization on each grayscale graph of the same size to obtain multiple normalized pixel matrices; then stack each pixel matrix along the channel dimension of the tensor to obtain multiple third-order tensor samples under each fault state;

[0010] Intuitive fuzzy score calculation steps: all third-order tensor samples in each fault state are divided into training samples and test samples according to a preset ratio, and the training samples under all fault states are combined into a training set consisting of positive samples and negative samples, and the test samples under all fault states are combined into a test set, and then the exponential membership function and the exponential non-membership function are used to calculate the membership and non-membership of each training sample in the training set, and the intuitive fuzzy score of each positive sample and negative sample is calculated according to the membership and non-membership, and the intuitive fuzzy score vector of the positive sample is established based on the intuitive fuzzy scores of all positive samples, and the intuitive fuzzy score vector of the negative sample is established based on the intuitive fuzzy scores of all negative samples;

[0011] The steps of extracting low-rank feature tensor samples and constructing high-dimensional feature matrices are as follows: using the tensor robust principal component analysis method to analyze each positive sample and each negative sample in the training set, respectively, extracting multiple positive low-rank feature tensor samples and multiple negative low-rank feature tensor samples, respectively, calculating the total low-rank feature tensor samples according to the positive low-rank feature tensor samples and the negative low-rank feature tensor samples, constructing a high-dimensional feature matrix of positive samples based on multiple positive low-rank feature tensor samples and the total low-rank feature tensor samples using a Gaussian kernel function, and constructing a high-dimensional feature matrix of negative samples based on multiple negative low-rank feature tensor samples and the total low-rank feature tensor samples using a Gaussian kernel function;

[0012] Model establishment and solution vector calculation steps: establish a first nonlinear intuitionistic fuzzy twin support tensor machine model based on the high-dimensional feature matrix of positive samples and the intuitive fuzzy score vector of negative samples, and establish a second nonlinear intuitionistic fuzzy twin support tensor machine model based on the high-dimensional feature matrix of negative samples and the intuitive fuzzy score vector of positive samples, then obtain a first dual optimization model based on the first nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtain a second dual optimization model based on the second nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtain a first solution vector based on the first dual optimization model, and obtain a second solution vector based on the second dual optimization model;

[0013] Prediction label calculation and fault diagnosis steps: Use the tensor robust principal component analysis method to analyze each test sample in the test set, extract multiple test low-rank feature tensor samples, and calculate the high-dimensional feature vector based on each test low-rank feature tensor sample using the Gaussian kernel function. Then, build a decision function based on the high-dimensional feature vector, the first solution vector and the second solution vector. According to the decision function, the prediction label of each test sample in the test set is obtained, thereby realizing the fault diagnosis of rotating machinery.

[0014] Preferably, in the prediction label calculation and fault diagnosis steps, an evaluation index is also used to evaluate the prediction effect of the prediction label to evaluate the prediction effect.

[0015] Preferably, the evaluation indicators include accuracy, precision, recall and F-score.

[0016] Preferably, the different types of sensors include an ICP accelerometer and an incremental encoder, and the monitoring signal includes vibration data and shaft speed data.

[0017] Preferably, the different fault states include an inner ring fault and an outer ring fault of the rotating machinery.

[0018] A rotating machinery fault diagnosis system, characterized by comprising a monitoring signal acquisition module, a time-frequency diagram conversion and third-order tensor sample acquisition module, an intuitive fuzzy score calculation module, a low-rank feature tensor sample extraction and high-dimensional feature matrix construction module, a model building and solution vector calculation module, and a prediction label calculation and fault diagnosis module, which are connected in sequence.

[0019] The monitoring signal acquisition module collects monitoring signals of the rotating machinery under different fault states within a certain time period through multiple different types of sensors;

[0020] The time-frequency graph conversion and third-order tensor sample acquisition module divides the monitoring signals collected by all sensors in each fault state at a certain number of time points within the time period into one sample to obtain multiple samples under each fault state; then uses the constant Q transformation method to convert each monitoring signal in each sample into a time-frequency graph, and grayscales each time-frequency graph to obtain multiple grayscale graphs; then crops each grayscale graph to the same size, and performs pixel normalization on each grayscale graph of the same size to obtain multiple normalized pixel matrices; then stacks each pixel matrix along the channel dimension of the tensor to obtain multiple third-order tensor samples under each fault state;

[0021] The intuitive fuzzy score calculation module divides all third-order tensor samples under each fault state into training samples and test samples according to a preset ratio, and combines the training samples under all fault states into a training set consisting of positive samples and negative samples, and combines the test samples under all fault states into a test set, and then uses an exponential membership function and an exponential non-membership function to respectively calculate the membership and non-membership of each training sample in the training set, and calculates the intuitive fuzzy score of each positive sample and negative sample according to the membership and non-membership, and establishes an intuitive fuzzy score vector of positive samples based on the intuitive fuzzy scores of all positive samples, and establishes an intuitive fuzzy score vector of negative samples based on the intuitive fuzzy scores of all negative samples;

[0022] The low-rank feature tensor sample extraction and high-dimensional feature matrix construction module adopts a tensor robust principal component analysis method to analyze the positive samples and negative samples in the training set respectively, and extracts a plurality of positive low-rank feature tensor samples and a plurality of negative low-rank feature tensor samples respectively, calculates a total low-rank feature tensor sample according to the positive low-rank feature tensor samples and the negative low-rank feature tensor samples, constructs a high-dimensional feature matrix of positive samples based on the plurality of positive low-rank feature tensor samples and the total low-rank feature tensor samples and adopts a Gaussian kernel function, and constructs a high-dimensional feature matrix of negative samples based on the plurality of negative low-rank feature tensor samples and the total low-rank feature tensor samples and adopts a Gaussian kernel function;

[0023] The model establishment and solution vector calculation module establishes a first nonlinear intuitionistic fuzzy twin support tensor machine model according to the high-dimensional feature matrix of positive samples and the intuitive fuzzy score vector of negative samples, and establishes a second nonlinear intuitionistic fuzzy twin support tensor machine model according to the high-dimensional feature matrix of negative samples and the intuitive fuzzy score vector of positive samples, then obtains a first dual optimization model based on the first nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtains a second dual optimization model based on the second nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtains a first solution vector according to the first dual optimization model, and obtains a second solution vector according to the second dual optimization model;

[0024] The prediction label calculation and fault diagnosis module uses a tensor robust principal component analysis method to analyze each test sample in the test set respectively, extracts multiple test low-rank feature tensor samples, calculates a high-dimensional feature vector based on each test low-rank feature tensor sample and uses a Gaussian kernel function, and then constructs a decision function based on the high-dimensional feature vector, the first solution vector and the second solution vector. According to the decision function, the prediction label of each test sample in the test set is obtained, thereby realizing fault diagnosis of rotating machinery.

[0025] Preferably, in the prediction label calculation and fault diagnosis module, an evaluation index is also used to evaluate the prediction effect of the prediction label to evaluate the prediction effect.

[0026] Preferably, the evaluation indicators include accuracy, precision, recall and F-score.

[0027] Preferably, the different types of sensors include an ICP accelerometer and an incremental encoder, and the monitoring signal includes vibration data and shaft speed data.

[0028] Preferably, the different fault states include an inner ring fault and an outer ring fault of the rotating machinery.

[0029] The beneficial effects of the present invention are:

[0030] The present invention provides a method for diagnosing faults of rotating machinery. First, monitoring signals of rotating machinery under different fault states within a certain time period are collected by using multiple sensors of different types, and the monitoring signals are converted into time-frequency images by using a constant Q transformation method. Grayscale, cropping and normalization processing are performed in sequence to obtain multiple normalized pixel matrices. Each pixel matrix is ​​then stacked along the channel dimension of the tensor to obtain multiple third-order tensor samples under each fault state, which can extract multi-sensor time-varying features and fully mine time domain feature information, frequency domain feature information and space domain coupling structure information. Then, all third-order tensor samples under each fault state are divided into training samples and test samples according to a preset ratio, and the training samples under all fault states are combined into a training set consisting of positive samples and negative samples. An exponential membership function and an exponential non-membership function in a nonlinear form are designed to calculate the intuitive fuzzy score of each positive sample and negative sample, thereby highlighting the contribution of different samples, effectively alleviating the influence of noise and outliers on the optimal hyperplane, and improving the robustness of the model. Then, the tensor robust principal component analysis method is used to accurately extract noise-free low-rank feature tensor samples from the original noisy third-order tensor samples. On the one hand, it avoids the coupling information between different monitoring signals from being destroyed by the vectorization process, and on the other hand, it avoids this information from being submerged by noise, effectively alleviating the impact of noise on the model and significantly improving the accuracy of fault diagnosis. Finally, a nonlinear intuitive fuzzy twin support tensor machine model is established. By converting the model into solving two relatively small quadratic programming problems, the predicted label of each test sample in the test set is finally obtained, thereby realizing the fault diagnosis of rotating machinery, effectively improving the performance of rotating machinery fault diagnosis and the computational efficiency of the model.

[0031] The rotating machinery fault diagnosis method provided by the present invention is a rotating machinery fault diagnosis method based on a nonlinear intuitive fuzzy twin support tensor machine model, which solves the problem that the existing diagnostic methods cannot effectively extract the time-varying features of multiple sensors and are easily affected by noise, resulting in low diagnostic accuracy and computational efficiency. Compared with the existing methods, the method of the present invention can extract the time-varying features of multiple sensors and highlight the contribution of different samples, effectively alleviate the influence of noise and outliers on the optimal hyperplane, and improve the robustness of the model. In addition, by introducing tensor robust principal component analysis, a low-rank noise-free feature tensor is accurately extracted from the noise tensor sample and used as an approximation of the original tensor. On the one hand, the coupling information between different monitoring signals is avoided from being destroyed by the vectorization process, and on the other hand, this information is avoided from being submerged by noise, and finally the accuracy of fault diagnosis is significantly improved. And by defining two relatively small quadratic programming problems, the computational efficiency of the model is improved.

[0032] The present invention also relates to a rotating machinery fault diagnosis system, which corresponds to the above-mentioned rotating machinery fault diagnosis method and can be understood as a system for implementing the above-mentioned rotating machinery fault diagnosis method, including a monitoring signal acquisition module, a time-frequency diagram conversion and third-order tensor sample acquisition module, an intuitive fuzzy score calculation module, a low-rank feature tensor sample extraction and high-dimensional feature matrix construction module, a model establishment and solution vector calculation module, and a prediction label calculation and fault diagnosis module connected in sequence. The modules work together to convert multi-sensor monitoring signals into time-frequency images by using a constant Q transformation method, and grayscale, crop, and normalize them, and then transform them along the tensor. The channel dimensions are stacked and reconstructed into third-order tensor samples, which fully exploits the feature information in the time domain and frequency domain and the coupling structure information in the space domain; secondly, a nonlinear non-membership function is designed to construct an intuitive fuzzy set, and the intuitive fuzzy score of each tensor sample is obtained to highlight the contribution of different samples; then the tensor robust principal component analysis method is used to accurately extract noise-free low-rank feature tensor samples from the original noisy tensor samples, which alleviates the impact of noise on the model; finally, a nonlinear intuitionistic fuzzy twin support tensor machine model is established, which effectively improves the performance and computational efficiency of rotating machinery fault diagnosis by solving two relatively small quadratic programming problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the rotating machinery fault diagnosis method of the present invention.

[0034] Figure 2 This is a schematic diagram of the first sensor monitoring signal and the second sensor monitoring signal of fault 1.

[0035] Figure 3 It is the time-frequency grayscale image corresponding to the monitoring signals of the first and second sensors of fault 1.

[0036] Figure 4 This is a schematic diagram of the first sensor monitoring signal and the second sensor monitoring signal of fault 2.

[0037] Figure 5 It is the time-frequency grayscale image corresponding to the monitoring signals of the first and second sensors of fault 2. DETAILED DESCRIPTION

[0038] The present invention will be described below in conjunction with the accompanying drawings.

[0039] The present invention relates to a method for diagnosing rotating machinery faults, which can also be called a method for diagnosing rotating machinery faults based on a nonlinear intuitive fuzzy twin support tensor machine. The method combines rich multi-sensor signals for diagnosis, uses constant Q transform to convert multi-sensor signals into time-frequency images, and performs cropping, normalization and grayscale processing, and then stacks and reconstructs them into tensor samples along the forward slices. The constructed tensor samples fully mine the feature information in the time domain and frequency domain and the coupling structure information in the space domain; secondly, a nonlinear non-membership function is designed to construct an intuitive fuzzy set, and by assigning different intuitive fuzzy scores to each tensor sample, the contribution of different samples is highlighted, thereby improving the robustness of the model. Then, tensor robust principal component analysis is used to accurately extract low-rank and noise-free feature tensors from the original noisy tensor samples, thereby alleviating the impact of noise on the model and achieving a significant improvement in the accuracy of fault diagnosis; finally, a nonlinear intuitive fuzzy twin support tensor machine model is established, and by solving two relatively small quadratic programming problems, the performance and computational efficiency of rotating machinery fault diagnosis are effectively improved. The flowchart of the method is shown as follows. Figure 1 As shown, the following steps are included in sequence:

[0040] 1. Signal collection step: Use multiple different types of sensors to collect monitoring signals of rotating machinery under different fault states within a certain period of time. Among them, different types of sensors may include ICP accelerometers and incremental encoders, etc. Monitoring signals may include vibration data and shaft speed data, etc. Different fault states may include inner ring faults and outer ring faults of rotating machinery.

[0041] 2. Time-frequency graph conversion and third-order tensor sample acquisition steps: divide the monitoring signals collected by all sensors at a certain number of time points in each fault state within the time period into one sample to obtain multiple samples under each fault state; then use the constant Q transformation method to convert each monitoring signal in each sample into a time-frequency graph, and grayscale each time-frequency graph to obtain multiple grayscale graphs; then crop each grayscale graph to the same size, and perform pixel normalization on each grayscale graph of the same size to obtain multiple normalized pixel matrices; then stack each pixel matrix along the channel dimension of the tensor to obtain multiple third-order tensor samples under each fault state. Specifically,

[0042] 1) The monitoring signals collected by all sensors at every L time points in each fault state during the time period are divided into one sample, and multiple samples under each fault state are obtained. Each sample is expressed as: Where K is the number of sensors and L is the time window (i.e., L time points).

[0043] 2) Each column in each sample X (i.e., each monitoring signal) is converted into a time-frequency diagram using the constant Q transformation method, and each time-frequency diagram is gray-scaled to obtain K gray-scale diagrams, and then each gray-scale diagram is cropped to the same size, i.e., cropped to I×J size, and each gray-scale diagram of the same size is pixel normalized, i.e., the pixels are normalized to [0,1], and K normalized pixel matrices are obtained.

[0044] 3) Each pixel matrix is ​​stacked along the channel dimension of the tensor (also known as the forward slicing direction of the tensor) to obtain a third-order tensor sample of size I×J×K under each fault state. Among them, the first order corresponds to the time domain information of the multi-sensor monitoring signal, the second order corresponds to the frequency domain information of the multi-sensor monitoring signal, and the third order corresponds to the spatial domain related information of the multi-sensor monitoring signal.

[0045] 3. Intuitive fuzzy score calculation steps: divide all third-order tensor samples under each fault state into training samples and test samples according to a preset ratio, and combine the training samples under all fault states into a training set consisting of positive samples and negative samples, and combine the test samples under all fault states into a test set, and then use the exponential membership function and exponential non-membership function to calculate the membership and non-membership of each training sample in the training set, respectively, calculate the intuitive fuzzy score of each positive sample and negative sample according to the membership and non-membership, establish the positive sample intuitive fuzzy score vector based on the intuitive fuzzy scores of all positive samples, and establish the negative sample intuitive fuzzy score vector based on the intuitive fuzzy scores of all negative samples.

[0046] Specifically, since the present invention is aimed at a binary classification problem, two faults are randomly selected from all fault states (for example, including four faults: fault 1, fault 2, fault 3, and fault 4) as binary classification experimental objects. Assume that the fault states to be diagnosed are fault 1 and fault 2, and a total of m third-order tensor samples are selected from the two faults. Training is performed, where all training samples of fault 1 represents positive samples (fault 1, label is +1), all training samples of fault 2 represents the negative class sample (fault 2, label is -1), so the training set D can be obtained:

[0047]

[0048] Among them, m represents the number of third-order tensor samples in the training set, y i Represents a third-order tensor sample The corresponding label.

[0049] It should be noted that, among the two faults arbitrarily selected above, if all the training samples of one of the faults are positive samples, then all the training samples of the other fault are negative samples.

[0050] Then, the exponential membership function and exponential non-membership function are used to calculate the membership and non-membership of each training sample (i.e., the third-order tensor sample) in the training set. For the i-th training sample Its exponential membership function μ i and exponential non-membership function v i The calculation formulas are:

[0051]

[0052] Among them, h>0 is an adjustable parameter to avoid the membership degree being 0; r + 、r - Respectively represent the positive and negative class radius; s + Indicates the maximum distance between the positive sample and the negative sample center; s - Indicates the maximum distance between the negative sample and the positive sample center; M + 、M - Respectively represent the positive and negative class centers; they can be calculated by (4-6):

[0053] r + =max i＝1,...,l ||X i -M + ||,r - =max i＝l+1,...,m ||X i -M- || (4)

[0054] s + =max i＝1,...,l ||X i -M - ||,s - =max i＝l+1,...,m ||X i -M + || (5)

[0055]

[0056] Among them, l represents the number of positive samples, and ml represents the number of negative samples.

[0057] Then construct the intuitionistic fuzzy set: (i=1,…,m), and finally the i-th training sample is obtained according to the obtained membership and non-membership through formula (7) The intuitionistic fuzzy score s i :

[0058]

[0059] According to the above formula (7), the intuitive fuzzy scores of each positive sample and each negative sample in the training set can be obtained, and then the intuitive fuzzy score vector of positive samples s1 = [s1, ..., s l ] T , and based on the intuitive fuzzy scores of all negative samples, establish the negative sample intuitive fuzzy score vector s2 = [s l+1 ,...,s m ] T .

[0060] 4. Low-rank feature tensor sample extraction and kernel function calculation steps: Use the tensor robust principal component analysis method to analyze the positive samples and negative samples in the training set respectively, and extract multiple positive low-rank feature tensor samples and multiple negative low-rank feature tensor samples respectively, calculate the total low-rank feature tensor samples according to the positive low-rank feature tensor samples and the negative low-rank feature tensor samples, and construct a high-dimensional feature matrix of positive samples based on the positive low-rank feature tensor samples and the total low-rank feature tensor samples using the Gaussian kernel function, and construct a high-dimensional feature matrix of negative samples based on the negative low-rank feature tensor samples and the total low-rank feature tensor samples using the Gaussian kernel function.

[0061] Specifically, the tensor robust principal component analysis method is used to analyze each positive sample and each negative sample in the training set (i.e. all training samples) Analyze and accurately extract the total low-rank feature tensor samples without noise The process is mainly solved by solving the following convex optimization problem:

[0062]

[0063] in, Represents low-rank feature components; represents the sparse noise component; Represents low-rank feature components The nuclear norm of i Indicates parameters; Represents the sparse noise component 1-norm of ;

[0064] The solution to the above problem is obtained by the Alternating Direction Method of Multipliers And the multiple positive low-rank feature tensor samples extracted from the positive samples (i.e., the training samples of fault 1) are recorded as The multiple negative low-rank feature tensor samples extracted from the negative samples (i.e., the training samples of fault 2) are recorded as

[0065] Then based on the positive low-rank feature tensor samples, negative low-rank feature tensor samples and total low-rank feature tensor samples, that is, as well as And use the Gaussian kernel function, that is: Construct high-dimensional feature matrices of positive samples and negative samples respectively.

[0066] By using the kernel trick, we define and The Gaussian kernel function between And through Will Mapped to the m-dimensional space, a high-dimensional feature matrix of positive samples is constructed based on multiple positive low-rank feature tensor samples and total low-rank feature tensor samples using a Gaussian kernel function, as shown in the following formula:

[0067]

[0068] Based on multiple negative low-rank feature tensor samples and total low-rank feature tensor samples and using Gaussian kernel function, a high-dimensional feature matrix of negative samples is constructed, as shown in the following formula:

[0069]

[0070] 5. Model establishment and solution vector calculation steps: Establish a first nonlinear intuitionistic fuzzy twin support tensor machine model based on the high-dimensional feature matrix of positive samples and the intuitive fuzzy score vector of negative samples, and establish a second nonlinear intuitionistic fuzzy twin support tensor machine model based on the high-dimensional feature matrix of negative samples and the intuitive fuzzy score vector of positive samples, then obtain a first dual optimization model based on the first nonlinear intuitionistic fuzzy twin support tensor machine model and use the Lagrangian function, obtain a second dual optimization model based on the second nonlinear intuitionistic fuzzy twin support tensor machine model and use the Lagrangian function, obtain a first solution vector based on the first dual optimization model, and obtain a second solution vector based on the second dual optimization model.

[0071] Specifically, this step can be understood as the original problem of establishing a nonlinear intuitionistic fuzzy twin support tensor machine model and obtaining its dual problem. By minimizing the sum of the squares of the distances from one type of sample to its corresponding hyperplane and ensuring that the distance from the other type of sample to the hyperplane is at least 1, the original optimization problems of the nonlinear intuitionistic fuzzy twin support tensor machine model are:

[0072]

[0073] in, c1, ξ1, and b1 represent the weight vector, penalty factor, slack variable, and bias of the first hyperplane, respectively; c2, ξ2 and b2 represent the weight vector, penalty factor, slack variable and bias of the second hyperplane respectively; e1 and e2 are column vectors whose elements are all 1; s1=[s1,...,s l ] T ,s2=[s l+1 ,...,s m ] T Represent the intuitionistic fuzzy score vectors of positive samples and negative samples respectively.

[0074] In order to obtain the optimal solution of the original optimization problem (11), it is necessary to transform the original optimization problem into its dual problem. To this end, the Lagrangian function of the original optimization problem (11) is constructed as follows:

[0075]

[0076] Here, α and β represent Lagrange multipliers.

[0077] By taking partial derivatives of the variables and setting them to 0, we get the Karush Kuhn Tucker (KKT) condition:

[0078]

[0079] So we further get:

[0080]

[0081] make Formula (14) can be rewritten as:

[0082] U T Up+G T α=0,ie p=-(U T U) -1 G T α (16)

[0083] Finally, the first dual optimization model of the nonlinear intuitionistic fuzzy twin support tensor machine model (11) is obtained:

[0084]

[0085] Where α represents the Lagrange multiplier. Correspondingly, the first solution vector p can be obtained:

[0086]

[0087] Similarly, the second dual optimization model of the nonlinear intuitionistic fuzzy twin support tensor machine model (12) can be obtained:

[0088]

[0089] Where μ represents the Lagrange multiplier. Correspondingly, the second solution vector q can be obtained:

[0090]

[0091] 6. Prediction label calculation and fault diagnosis steps: Use the tensor robust principal component analysis method to analyze each test sample in the test set, extract multiple test low-rank feature tensor samples, calculate the high-dimensional feature vector based on each test low-rank feature tensor sample and use the Gaussian kernel function, and then construct a decision function based on the first solution vector and the second solution vector of the high-dimensional feature vector. According to the decision function, the prediction label of each third-order tensor sample in the test set is obtained, thereby realizing the fault diagnosis of rotating machinery.

[0092] Specifically, for a given j-th test sample First, the tensor robust principal component analysis method is used, that is, the tensor robust principal component analysis is performed on it using formula (8) to extract multiple test low-rank feature tensor samples Then according to each test low-rank feature tensor sample And use Gaussian kernel function to calculate high-dimensional feature vector Where K(·) represents the Gaussian kernel function. Then, a decision function is constructed based on the high-dimensional feature vector, the first solution vector, and the second solution vector, as shown in the following formula:

[0093]

[0094] Finally, the decision rule of the above decision function is used to predict the test sample The label is obtained, and the above process is repeated for each tensor sample in the test set to obtain the predicted label of each test sample in the test set, thereby realizing the fault diagnosis of rotating machinery.

[0095] Example:

[0096] In this example, the bearing vibration data collected by the University of Ottawa under time-varying speed conditions is used as experimental data. Each sampling data set contains two channels: sensor 1 is the vibration data collected by the ICP accelerometer placed on the experimental bearing seat, and sensor 2 is the shaft speed data collected by the incremental encoder. The sampling frequency of the signal is 20000hz, and the sampling time is 10s. The effectiveness of the present invention is verified by a binary classification experiment of inner race fault IC-1 (fault 1, labeled as 1) and outer race fault OC-1 (fault 2, labeled as -1). The data of the first 10240 sampling time points of fault 1 and fault 2 are selected, and the time window L is set to 1024. In order to verify the anti-noise performance of the present invention, Gaussian white noise with a signal-to-noise ratio (SNR) value equal to 1 is added to each sensor monitoring signal of fault 1 and fault 2 respectively.

[0097] The first sensor monitoring signal and the second sensor monitoring signal of fault 1 are sampled and 100 samples of fault 1 are obtained, where the multi-sensor monitoring signal of each sample is expressed as Use the constant Q transform method to convert each column in X (that is, each monitoring signal) into a time-frequency graph, grayscale each time-frequency graph to obtain multiple grayscale images, and crop them to 64×64 size. Perform pixel normalization on each grayscale image of the same size, that is, normalize the pixels to [0,1], and obtain multiple normalized pixel matrices. For example, for the first sample of fault 1 (the first 1024 time points), Figure 2 (a) and Figure 2 (b) shows the first and second sensor monitoring signals. The time-frequency grayscale images corresponding to the first and second sensor monitoring signals are shown in Figure 2. Figure 3 (a) and Figure 3 (b) shows that the two sensor monitoring signals are converted into two grayscale images of size 64×64, and the pixel matrices corresponding to these grayscale images are stacked along the forward slicing direction (which can also be understood as along the channel dimension of the tensor), and the sample X is reconstructed into a third-order tensor sample of size 64×64×2. A total of 100 tensor samples of fault 1 are obtained, 80 samples are randomly selected for training, and the remaining 20 samples are used for testing.

[0098] Repeat the above process for fault 2. For example, for the first sample of fault 2 (the first 1024 time points), Figure 4 (a) and Figure 4 (b) shows the first and second sensor monitoring signals. The time-frequency grayscale images corresponding to the first and second sensor monitoring signals are shown in Figure 2. Figure 5 (a) and Figure 5 As shown in (b), 100 tensor samples of fault 2 are finally obtained, 80 samples are randomly selected for training, and the remaining 20 samples are used for testing.

[0099] The 80 training samples of fault 1 and the 80 training samples of fault 2 are combined into a training set The membership and non-membership of each training sample are calculated by formula (2) and formula (3), and the intuitive fuzzy score of the training sample is obtained by formula (7):

[0100] For all training samples in the training set We perform robust principal component analysis on the tensor, that is, we accurately extract low-rank feature tensor samples without noise by solving the optimization problem (8) and obtain In this embodiment, a Gaussian kernel function is used, namely: Using formula (9) and formula (10), we can get the high-dimensional feature matrix of positive samples: And the high-dimensional feature matrix of negative samples Combined with the intuitive fuzzy score obtained in step 4 Substitute them into the nonlinear intuitionistic fuzzy twin support tensor machine optimization model (11) and (12). Among them, the parameter c1 = c2, and c1, c2, q are all in [2 -7 ,2 -6 ,2 -5 ,2 -4 ,2 -3 ,2 -2 ,2 -1 ,2 0 ,2 1 ,2 2 ,2 3 ,2 4 ,2 5 ,2 6 ,2 7 ] range, and use the grid search method to determine the optimal parameter value. The solution vector of model (11) is obtained by formula (18): The solution vector of model (12) is obtained by formula (19):

[0101] For the 40 test samples in the test set Perform robust principal component analysis on tensors to obtain multiple test low-rank feature tensor samples For each j=1,...,40, a high-dimensional feature vector is calculated based on each test low-rank feature tensor sample and using the Gaussian kernel function Then build a decision function based on the first solution vector and the second solution vector of the high-dimensional feature vector The prediction label of each third-order tensor sample in the test set is obtained by the decision function. In order to comprehensively evaluate the diagnostic performance of the invented method, the following accuracy, precision, recall and F-score are used as evaluation indicators to evaluate the prediction effect of the prediction label, where:

[0102]

[0103] In the above formula, TP represents the number of positive predictions that are actually positive, FN represents the number of negative predictions that are actually positive, FP represents the number of positive predictions that are actually negative, and TN represents the number of negative predictions that are actually negative.

[0104] In addition, the above experiment was repeated 10 times, and the Accuracy, Precision, Recall and F-score of the 10 experiments and their corresponding optimal parameter values ​​are shown in Table 1, and the confusion matrix is ​​shown in Table 2.

[0105] Table 1 Accuracy and optimal parameter values ​​of 10 repeated experiments

[0106] Accuracy Precision Recall F-score <![CDATA[c1 value]]> Q-value 1 90.00％ 90.00％ 90.00％ 90.00％ 2 8 2 95.00％ 90.91％ 100.0％ 95.24％ 2 8 3 97.50％ 100.0％ 95.00％ 97.44％ 2 8 4 97.50％ 100.0％ 95.00％ 97.44％ 2 8 5 92.50％ 90.48％ 95.00％ 92.68％ 2 8 6 97.50％ 100％ 95.00％ 97.44％ 2 8 7 97.50％ 95.24％ 100.0％ 97.56％ 2 8 8 95.00％ 100.0％ 90.00％ 94.74％ 2 8 9 97.50％ 100.0％ 95.00％ 97.44％ 2 8 10 100.0％ 100.0％ 100.0％ 100.0％ 2 8 average value 96.00％ 96.66％ 95.50％ 96.00％ —— —— Standard Deviation 0.0278 0.0430 0.0350 0.0275 —— ——

[0107] The average accuracy, average precision, average recall, and average F-score of the 10 repeated experiments in Table 1 are 96.00%, 96.66%, 95.50%, and 96.00%, respectively, indicating that the present invention can accurately identify the inner race fault (IC-1) and the outer race fault (OC-1). And a small standard deviation is obtained, which is 0.0278, 0.0430, 0.0350, and 0.0275, respectively, indicating that the present invention has good stability.

[0108] Table 2 Confusion matrix of 10 repeated experiments

[0109]

[0110] It can be seen from Table 2 that the method of the present invention achieved good prediction results for both positive samples (inner ring fault samples) and negative samples (outer ring fault samples) in 10 repeated experiments, further verifying that the present invention can more accurately diagnose this type of rotating machinery faults.

[0111] The present invention also relates to a rotating machinery fault diagnosis system, which corresponds to the above-mentioned rotating machinery fault diagnosis method and can be understood as a system for implementing the above-mentioned method. The system includes a monitoring signal acquisition module, a time-frequency diagram conversion and third-order tensor sample acquisition module, an intuitive fuzzy score calculation module, a low-rank feature tensor sample extraction and high-dimensional feature matrix construction module, a model establishment and solution vector calculation module, and a prediction label calculation and fault diagnosis module, which are connected in sequence. Specifically,

[0112] The monitoring signal acquisition module collects monitoring signals of the rotating machinery under different fault states within a certain time period through multiple different types of sensors;

[0113] The time-frequency graph conversion and third-order tensor sample acquisition module divides the monitoring signals collected by all sensors in each fault state at a certain number of time points within the time period into one sample to obtain multiple samples under each fault state; then uses the constant Q transformation method to convert each monitoring signal in each sample into a time-frequency graph, and grayscales each time-frequency graph to obtain multiple grayscale graphs; then crops each grayscale graph to the same size, and performs pixel normalization on each grayscale graph of the same size to obtain multiple normalized pixel matrices; then stacks each pixel matrix along the channel dimension of the tensor to obtain multiple third-order tensor samples under each fault state;

[0114] The intuitive fuzzy score calculation module divides all third-order tensor samples under each fault state into training samples and test samples according to a preset ratio, and combines the training samples under all fault states into a training set consisting of positive samples and negative samples, and combines the test samples under all fault states into a test set, and then uses an exponential membership function and an exponential non-membership function to respectively calculate the membership and non-membership of each training sample in the training set, and calculates the intuitive fuzzy score of each positive sample and negative sample according to the membership and non-membership, and establishes an intuitive fuzzy score vector of positive samples based on the intuitive fuzzy scores of all positive samples, and establishes an intuitive fuzzy score vector of negative samples based on the intuitive fuzzy scores of all negative samples;

[0115] The low-rank feature tensor sample extraction and high-dimensional feature matrix construction module adopts a tensor robust principal component analysis method to analyze each positive sample and each negative sample in the training set respectively, and extracts a plurality of positive low-rank feature tensor samples and a plurality of negative low-rank feature tensor samples respectively, calculates a total low-rank feature tensor sample according to the positive low-rank feature tensor samples and the negative low-rank feature tensor samples, constructs a high-dimensional feature matrix of positive samples based on the plurality of positive low-rank feature tensor samples and the total low-rank feature tensor samples and adopts a Gaussian kernel function, and constructs a high-dimensional feature matrix of negative samples based on the plurality of negative low-rank feature tensor samples and the total low-rank feature tensor samples and adopts a Gaussian kernel function;

[0116] The model establishment and solution vector calculation module establishes a first nonlinear intuitionistic fuzzy twin support tensor machine model according to the high-dimensional feature matrix of positive samples and the intuitive fuzzy score vector of negative samples, and establishes a second nonlinear intuitionistic fuzzy twin support tensor machine model according to the high-dimensional feature matrix of negative samples and the intuitive fuzzy score vector of positive samples, then obtains a first dual optimization model based on the first nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtains a second dual optimization model based on the second nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtains a first solution vector according to the first dual optimization model, and obtains a second solution vector according to the second dual optimization model;

[0117] The prediction label calculation and fault diagnosis module uses a tensor robust principal component analysis method to analyze each test sample in the test set respectively, extracts multiple test low-rank feature tensor samples, calculates a high-dimensional feature vector based on each test low-rank feature tensor sample and uses a Gaussian kernel function, and then constructs a decision function based on the high-dimensional feature vector, the first solution vector and the second solution vector. According to the decision function, the prediction label of each test sample in the test set is obtained, thereby realizing fault diagnosis of rotating machinery.

[0118] Preferably, in the prediction label calculation and fault diagnosis module, an evaluation index is also used to evaluate the prediction effect of the prediction label to evaluate the prediction effect.

[0119] Preferably, the evaluation indicators include accuracy, precision, recall and F-score.

[0120] Preferably, the different types of sensors include an ICP accelerometer and an incremental encoder, and the monitoring signals include vibration data and shaft speed data.

[0121] Preferably, the different fault states include an inner ring fault and an outer ring fault of the rotating machinery.

[0122] The present invention provides an objective and scientific method and system for rotating machinery fault diagnosis. On the one hand, it combines rich multi-sensor signals for diagnosis. By using the constant Q transform method, the multi-sensor signals are converted into time-frequency images and grayed, cropped, and normalized, and then stacked and reconstructed into third-order tensor samples along the channel dimension of the tensor, the time domain, frequency domain feature information and spatial domain feature information of the signal are fully mined; and by assigning different intuitive fuzzy scores to each tensor sample, the contribution of different samples is highlighted, and the robustness of the model is improved. On the other hand, the tensor robust principal component analysis method is used to accurately extract low-rank noise-free feature tensors from noise tensor samples, which alleviates the interference of noise and achieves a significant improvement in the accuracy of fault diagnosis. Finally, a nonlinear intuitive fuzzy twin support tensor machine model is established and the model is converted into solving two relatively small quadratic programming problems, which improves the computational efficiency of the model.

[0123] It should be noted that the above-described specific implementations can enable those skilled in the art to more fully understand the invention, but do not limit the invention in any way. Therefore, although this specification has described the invention in detail with reference to the drawings and embodiments, those skilled in the art should understand that the invention can still be modified or replaced by equivalents. In short, all technical solutions and improvements that do not deviate from the spirit and scope of the invention should be included in the protection scope of the patent for the invention.

Claims

1. A method for diagnosing a rotating machinery fault, characterized in that: The following steps are involved: Monitoring signal collection step: using multiple sensors of different types to collect monitoring signals of rotating machinery under different fault states within a certain period of time; Steps for time-frequency graph conversion and third-order tensor sample acquisition: divide the monitoring signals collected by all sensors at a certain number of time points in each fault state into one sample, and obtain multiple samples under each fault state; then use the constant Q transformation method to convert each monitoring signal in each sample into a time-frequency graph, and grayscale each time-frequency graph to obtain multiple grayscale graphs; then crop each grayscale graph to the same size, and perform pixel normalization on each grayscale graph of the same size to obtain multiple normalized pixel matrices; Each pixel matrix is ​​then stacked along the channel dimension of the tensor to obtain multiple third-order tensor samples under each fault state; Intuitive fuzzy score calculation steps: all third-order tensor samples in each fault state are divided into training samples and test samples according to a preset ratio, and the training samples under all fault states are combined into a training set consisting of positive samples and negative samples, and the test samples under all fault states are combined into a test set, and then the exponential membership function and the exponential non-membership function are used to calculate the membership and non-membership of each training sample in the training set, and the intuitive fuzzy score of each positive sample and negative sample is calculated according to the membership and non-membership, and the intuitive fuzzy score vector of the positive sample is established based on the intuitive fuzzy scores of all positive samples, and the intuitive fuzzy score vector of the negative sample is established based on the intuitive fuzzy scores of all negative samples; The steps of extracting low-rank feature tensor samples and constructing high-dimensional feature matrices are as follows: using the tensor robust principal component analysis method to analyze each positive sample and each negative sample in the training set, respectively, extracting multiple positive low-rank feature tensor samples and multiple negative low-rank feature tensor samples, respectively, calculating the total low-rank feature tensor samples according to the positive low-rank feature tensor samples and the negative low-rank feature tensor samples, constructing a high-dimensional feature matrix of positive samples based on multiple positive low-rank feature tensor samples and the total low-rank feature tensor samples using a Gaussian kernel function, and constructing a high-dimensional feature matrix of negative samples based on multiple negative low-rank feature tensor samples and the total low-rank feature tensor samples using a Gaussian kernel function; Model establishment and solution vector calculation steps: establish a first nonlinear intuitionistic fuzzy twin support tensor machine model based on the high-dimensional feature matrix of positive samples and the intuitive fuzzy score vector of negative samples, and establish a second nonlinear intuitionistic fuzzy twin support tensor machine model based on the high-dimensional feature matrix of negative samples and the intuitive fuzzy score vector of positive samples, then obtain a first dual optimization model based on the first nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtain a second dual optimization model based on the second nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtain a first solution vector based on the first dual optimization model, and obtain a second solution vector based on the second dual optimization model; Prediction label calculation and fault diagnosis steps: Use the tensor robust principal component analysis method to analyze each test sample in the test set, extract multiple test low-rank feature tensor samples, and calculate the high-dimensional feature vector based on each test low-rank feature tensor sample using the Gaussian kernel function. Then, build a decision function based on the high-dimensional feature vector, the first solution vector and the second solution vector. According to the decision function, the prediction label of each test sample in the test set is obtained, thereby realizing the fault diagnosis of rotating machinery.

2. The rotating machinery fault diagnosis method according to claim 1, characterized in that: In the prediction label calculation and fault diagnosis steps, an evaluation index is also used to evaluate the prediction effect of the prediction label to evaluate the prediction effect.

3. The rotating machinery fault diagnosis method according to claim 2, characterized in that: The evaluation indicators include accuracy, precision, recall and F-score.

4. The rotating machinery fault diagnosis method according to claim 1, characterized in that: The different types of sensors include ICP accelerometers and incremental encoders, and the monitoring signals include vibration data and shaft speed data.

5. The rotating machinery fault diagnosis method according to claim 1, characterized in that: The different fault states include inner ring fault and outer ring fault of the rotating machinery.

6. A rotating machinery fault diagnosis system, characterized in that: It includes a monitoring signal acquisition module, a time-frequency diagram conversion and third-order tensor sample acquisition module, an intuitive fuzzy score calculation module, a low-rank feature tensor sample extraction and high-dimensional feature matrix construction module, a model building and solution vector calculation module, and a prediction label calculation and fault diagnosis module. The monitoring signal acquisition module collects monitoring signals of the rotating machinery under different fault states within a certain time period through multiple different types of sensors; The time-frequency graph conversion and third-order tensor sample acquisition module divides the monitoring signals collected by all sensors at a certain number of time points in the time period under each fault state into one sample, and obtains multiple samples under each fault state; then uses the constant Q transformation method to convert each monitoring signal in each sample into a time-frequency graph, and grayscales each time-frequency graph to obtain multiple grayscale graphs; then each grayscale graph is cropped to the same size, and pixel normalization is performed on each grayscale graph of the same size to obtain multiple normalized pixel matrices; Each pixel matrix is ​​then stacked along the channel dimension of the tensor to obtain multiple third-order tensor samples under each fault state; The intuitive fuzzy score calculation module divides all third-order tensor samples under each fault state into training samples and test samples according to a preset ratio, and combines the training samples under all fault states into a training set consisting of positive samples and negative samples, and combines the test samples under all fault states into a test set, and then uses an exponential membership function and an exponential non-membership function to respectively calculate the membership and non-membership of each training sample in the training set, and calculates the intuitive fuzzy score of each positive sample and negative sample according to the membership and non-membership, and establishes an intuitive fuzzy score vector of positive samples based on the intuitive fuzzy scores of all positive samples, and establishes an intuitive fuzzy score vector of negative samples based on the intuitive fuzzy scores of all negative samples; The low-rank feature tensor sample extraction and high-dimensional feature matrix construction module adopts a tensor robust principal component analysis method to analyze the positive samples and negative samples in the training set respectively, and extracts a plurality of positive low-rank feature tensor samples and a plurality of negative low-rank feature tensor samples respectively, calculates a total low-rank feature tensor sample according to the positive low-rank feature tensor samples and the negative low-rank feature tensor samples, constructs a high-dimensional feature matrix of positive samples based on the plurality of positive low-rank feature tensor samples and the total low-rank feature tensor samples and adopts a Gaussian kernel function, and constructs a high-dimensional feature matrix of negative samples based on the plurality of negative low-rank feature tensor samples and the total low-rank feature tensor samples and adopts a Gaussian kernel function; The model establishment and solution vector calculation module establishes a first nonlinear intuitionistic fuzzy twin support tensor machine model according to the high-dimensional feature matrix of positive samples and the intuitive fuzzy score vector of negative samples, and establishes a second nonlinear intuitionistic fuzzy twin support tensor machine model according to the high-dimensional feature matrix of negative samples and the intuitive fuzzy score vector of positive samples, then obtains a first dual optimization model based on the first nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtains a second dual optimization model based on the second nonlinear intuitionistic fuzzy twin support tensor machine model and using the Lagrangian function, obtains a first solution vector according to the first dual optimization model, and obtains a second solution vector according to the second dual optimization model; The prediction label calculation and fault diagnosis module uses a tensor robust principal component analysis method to analyze each test sample in the test set respectively, extracts multiple test low-rank feature tensor samples, calculates a high-dimensional feature vector based on each test low-rank feature tensor sample and uses a Gaussian kernel function, and then constructs a decision function based on the high-dimensional feature vector, the first solution vector and the second solution vector. According to the decision function, the prediction label of each test sample in the test set is obtained, thereby realizing fault diagnosis of rotating machinery.

7. The rotating machinery fault diagnosis system according to claim 6, characterized in that: In the prediction label calculation and fault diagnosis module, an evaluation index is also used to evaluate the prediction effect of the prediction label to evaluate the prediction effect.

8. The rotating machinery fault diagnosis system according to claim 7, characterized in that: The evaluation indicators include accuracy, precision, recall and F-score.

9. The rotating machinery fault diagnosis system according to claim 6, characterized in that: The different types of sensors include ICP accelerometers and incremental encoders, and the monitoring signals include vibration data and shaft speed data.

10. The rotating machinery fault diagnosis system according to claim 6, characterized in that: The different fault states include inner ring fault and outer ring fault of the rotating machinery.

Images