Width learning electromechanical equipment fault diagnosis method based on graph regularization

Abstract

The invention relates to the technical field of fault diagnosis, in particular to a width learning electromechanical equipment fault diagnosis method based on graph regularization, and the method comprises the steps: obtaining feature parameters and state data of an actuator at each operation moment; constructing a width learning fault diagnosis model; improving an objective function of the width learning fault diagnosis model to obtain an improved objective function; adding a regularization item into the improved objective function based on a manifold learning method to obtain a final objective function; and inputting the input data at the current moment into the target width learning fault diagnosis model, and outputting state data of the electromechanical equipment. According to the method, the robustness of the width learning fault diagnosis model is enhanced, and the fault diagnosis accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a graph regularization-based width learning electromechanical equipment fault diagnosis method. Background Art

[0002] In the past few decades, industrial equipment has become more diversified and intelligent, and its structure has become more complex. Ensuring its safe and reliable operation is more challenging than ever before. In aerospace systems, electromechanical actuators (EMA) are usually key components and have become an important research topic in intelligent industrial fault diagnosis. In order to improve the diagnostic capabilities of various faults, accurate and efficient data-driven fault diagnosis methods have aroused great research interest among scholars.

[0003] Through the current literature search, it is found that the data-driven method mainly uses sensor data to train the model. This method has the characteristics of strong versatility. With the improvement of computer computing power and the development of machine learning methods, the data-driven electromechanical system fault diagnosis method has developed rapidly. The prior art proposes a multi-channel convolutional neural network (CNN) based on SK to classify the health status of bearings; the prior art uses a regularized deep clustering algorithm based on deep autoencoders to improve the feature extraction process and diagnose gearbox faults; and the prior art "Study on Rolling Bearing Fault Prediction Method Based on Deep Learning and Incremental Update" proposes a rolling bearing fault prediction method based on deep learning and incremental update. However, the above methods are all implemented by relying on high-quality labeled data, but in actual engineering applications, due to the influence of acquisition equipment or human factors, incorrect labels will inevitably be generated, which will seriously affect the accuracy of the fault diagnosis model.

[0004] At the same time, although deep learning technology has been widely used, when the data scale is huge, deep learning usually relies on methods such as increasing the number of network layers to improve network performance, which requires a lot of computing resources and is prone to falling into local optimal solutions, gradient explosion or gradient disappearance. At this time, if the data set label error occurs, it will be more difficult to deal with. In 2018, Chen et al. proposed a new wide learning method in "Broad learning system: An effective and efficient incremental learning system without the need for deep architecture". Wide learning has the following characteristics: (1) It develops the "width" of the network by increasing the number of hidden layer nodes instead of increasing the number of network layers. The number of network layers of wide learning is only three, which can effectively reduce the complexity of the model and can be trained on an ordinary PC. (2) It achieves the same accuracy as deep learning by randomly setting weights, but the training time is greatly shortened. (3) The weights from the hidden layer to the output layer are determined by solving the pseudo-inverse method through ridge regression, avoiding the problems of local optimality, gradient vanishing or explosion in the gradient descent method. (4) When the model training is inaccurate, incremental learning can be quickly implemented to expand and modify the original model without retraining the entire network. Because wide learning has a simple structure and fast training speed, it is easy to optimize and improve, and provides a new method for fault diagnosis in label error scenarios.

[0005] In summary, fault diagnosis of complex electromechanical equipment has important theoretical and practical significance and has attracted the attention of many researchers. However, related methods are particularly dependent on high-quality labeled data. How to perform fault diagnosis in scenarios with incorrect labels becomes a difficult problem.

[0006] Therefore, it is necessary to provide a graph regularized width learning electromechanical equipment fault diagnosis method to solve the above problems. Summary of the invention

[0007] The present invention provides a graph regularized based width learning electromechanical equipment fault diagnosis method to solve the existing fault diagnosis of complex electromechanical equipment, which has important theoretical and practical significance and has attracted the attention of many researchers. However, the related methods are particularly dependent on high-quality labeled data, and there is a problem of how to perform fault diagnosis in scenarios with incorrect labels.

[0008] A graph regularized width learning electromechanical equipment fault diagnosis method of the present invention adopts the following technical solution, including: Acquire the characteristic parameters and state data of the actuator at each operating moment, take the characteristic parameters as input data, take the state data corresponding to the actuator as output data, and construct a data set; wherein the characteristic parameters include: temperature data of the actuator and acceleration data of three axes; the state data include: normal operation, blocking fault, spalling fault and motor fault; A width learning fault diagnosis model is constructed based on the data set; the objective function improvement method based on maximum likelihood estimation is used to improve the objective function of the width learning fault diagnosis model to obtain an improved objective function; Based on the manifold learning method, a regularization term is added to the improved objective function to obtain the final objective function; The objective function of the width learning fault diagnosis model is replaced by the final objective function to obtain the final width learning fault diagnosis model; The final width learning fault diagnosis model is trained using the data set to obtain a trained target width learning fault diagnosis model; The input data at the current moment is input into the target width learning fault diagnosis model, and the status data of the electromechanical equipment is output.

[0009] Preferably, the characteristic parameters are sequentially subjected to data cleaning and data standardization, and the standardized characteristic parameters are used as input data.

[0010] Preferably, the step of performing data cleaning on the characteristic parameters is: Traverse the collected characteristic parameters and mark the missing values ​​and outliers. The outliers are determined by the 6σ criterion. The missing values ​​and outliers are replaced by the neighboring value filling method to complete data cleaning.

[0011] Preferably, the step of performing data standardization on the characteristic parameters is: performing data standardization processing on the data set using a Z-score standardization method.

[0012] Preferably, the objective function of the width learning fault diagnosis model is expressed as:

[0013] In the formula, , represents the error probability density function; A weight matrix representing a node in a width learning fault diagnosis model; represents the constraint coefficient; represents the distribution parameter; represents the errors of independent and identically distributed variables; express; It means to find the second-order F-norm.

[0014] Preferably, the improved objective function is:

[0015] In the formula, represents the weight of the training matrix; A matrix representing the expected output of the width learning fault diagnosis model; represents the input data matrix of the width learning fault diagnosis model; A weight matrix representing a node in a width learning fault diagnosis model; represents the constraint coefficient; It means to find the second-order F-norm.

[0016] Preferably, the expression of the final objective function is:

[0017] In the formula, represents the weight of the training matrix; A matrix representing the expected output of the width learning fault diagnosis model; represents the input data matrix of the width learning fault diagnosis model; A weight matrix representing a node in a width learning fault diagnosis model; represents the first regularization parameter; represents the second regularization parameter; represents a Laplacian graph; Represents an operator.

[0018] The beneficial effects of the present invention are: By analyzing the width learning fault diagnosis model, a graph regularization method is introduced to improve width learning when there are label errors in the fault diagnosis data set. The objective function based on maximum likelihood estimation is used for classification, and the manifold regularization term is integrated to retain the local geometric shape of the data. The erroneous labels in the data are adaptively suppressed by assigning weights. This achieves efficient, fast and reliable fault diagnosis in the scenario of erroneous labeling of the data set, enhances the robustness of the width learning fault diagnosis model, and improves the accuracy of fault diagnosis. It has high practical significance and engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1A flowchart of a method for diagnosing electromechanical equipment faults based on graph regularization and width learning of the present invention; Figure 2 This is a framework diagram of a graph regularized width learning electromechanical equipment fault diagnosis method of the present invention; Figure 3 This is a discussion result diagram of the GBLS enhanced node in an embodiment of the present invention; Figure 4 This is a graph showing the effect of GBLS hyperparameters on the training accuracy in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] An embodiment of a method for fault diagnosis of electromechanical equipment based on graph regularization and width learning of the present invention is as follows: Figure 1 and Figure 2 As shown, including: S1, obtaining characteristic parameters and status data of the actuator at each operating moment; Specifically, characteristic parameters and status data of the actuator at each operating moment are obtained, and the characteristic parameters are used as input data, and the status data corresponding to the actuator is used as output data to construct a data set; wherein the characteristic parameters include: temperature data of the actuator and acceleration data in three axes; the status data include: normal operation, blocking fault, spalling fault and motor fault.

[0023] In this embodiment, the characteristic parameters are sequentially cleaned and standardized, and the standardized characteristic parameters are used as input data, wherein the step of data cleaning of the characteristic parameters is: traversing the collected characteristic parameters, marking the missing values ​​and outliers, wherein the outliers are determined using the 6σ criterion, that is, the number greater than the sum of the mean and 6 times the standard deviation is considered an outlier; the missing values ​​and outliers are replaced by the neighboring value filling method, that is, the average value of the 5 points before and after the missing value or outlier is used to fill and replace, and the data cleaning is completed. The data set is processed using the Z-score standardization method to eliminate the difference between the dimensions, complete the data standardization, and the characteristic parameters that have finally completed the data standardization are used as input data. Finally, the data set selected in this embodiment contains a total of 600 groups of data.

[0024] S2, build a width learning fault diagnosis model; An initial width learning fault diagnosis model is constructed. The model building process is consistent with the model building method in "Broad learning system: An effective and efficient incremental learning system without the need for deep architecture", wherein the feature parameters in the data set are used as the input data of the initial width learning fault diagnosis model, and the state data in the data set is used as the output data of the initial width learning fault diagnosis model. The initial width learning fault diagnosis model is trained to obtain the width learning fault diagnosis model.

[0025] S3, improving the objective function of the width learning fault diagnosis model to obtain an improved objective function; Specifically, the objective function improvement method based on maximum likelihood estimation is used to improve the objective function of the width learning fault diagnosis model to obtain an improved objective function.

[0026] Step 31: Analyze the width learning fault diagnosis model: The training of the width learning fault diagnosis model is essentially to use unconstrained least squares to approximate the projection of the ideal label space, that is: (1) In the formula, A matrix representing the expected output of the width learning fault diagnosis model; represents the input data matrix of the width learning fault diagnosis model; A weight matrix representing a node in a width learning fault diagnosis model; represents the constraint coefficient, It means to find the 2nd order v norm; It means to find the second-order u-norm; m refers to the total number of nodes of the width learning fault diagnosis model; is a first-order power; is a second-order power.

[0027] Using ridge regression to solve, we get: (2) In the formula, is the identity matrix.

[0028] The objective function of the width learning fault diagnosis model is analyzed from the perspective of probability theory. The error between the expected output and the actual output of the width learning fault diagnosis model is for: (3) In the formula, is the approximate residual matrix, Indicates The approximate residual of the row. Let the error The probability density function of , represents the distribution parameter, is the error of independent and identically distributed variables, indicating that there are N characteristic indicators, then The maximum likelihood function is: (4) The solution is obtained by finding the maximum likelihood function or the minimum equivalent objective function, that is: (5) in, , considering the weight The regularization of , then the objective function is: (6) In the formula, , represents the error probability density function; A weight matrix representing a node in a width learning fault diagnosis model; represents the constraint coefficient; represents the distribution parameter; represents the errors of independent and identically distributed variables; express; It means to find the second-order F-norm.

[0029] The key to formula (6) is to solve The distribution of , assuming a Gaussian distribution, is: (7) In the formula, The approximate residual represented by ; represents the random coefficient; It represents the square of vector e and takes the 2-norm; It can be deduced into Solution,Through the above analysis, it can be found that, from the perspective of probability theory, the condition for width learning to be valid is that the error vector obeys the Gaussian distribution, but in the label error scenario, the error distribution will deviate from the Gaussian distribution, resulting in a large error in BLS. Therefore, it is necessary to improve the objective function of the width learning fault diagnosis model.

[0030] Step 22: Improve the objective function of the width learning fault diagnosis model: Pre-Assumptions The distribution of is Gaussian distribution, and the weight is derived The robust MLE of , makes the following assumptions: 1) Symmetry: ; 2) Monotonicity: ; 3) Minimum value: ; make , is the probability distribution function, is the probability density function. Then its first-order Taylor expansion is: (8) In the formula, is the high-order residual, is Taylor expansion, T is the order, since it is 1st order, so T=1, for The first-order Taylor expansion of The first derivative of ; is the initial error between the expected output and the actual output of the width learning fault diagnosis model; it can be approximated as: (9) In the formula, Indicates that and After making the difference, find the Hessian matrix; yes The matrix is ​​diagonal, because the minimum value , exist There is a minimum value when right The derivatives are: (10) in, , Represents the nth row and nth column element of matrix D; express The first derivative of the probability density function of ; express No. elements, because and Same sign, so non-negative, then It can be approximated as: (11) (12) in, Is Determine the scalar value, so The robust version of , that is, the improved objective function is: (13) in, is the weight of the training matrix. In this way, the originally assumed Gaussian distribution Distribution is transformed into the determination of matrix weights in the width learning fault diagnosis model; The calculation of width to determine the matrix weights of the learning fault diagnosis model can reduce the weights of samples with incorrect labels and suppress their impact on the global optimal solution.

[0031] S4, adding a regularization term to the improved objective function based on the manifold learning method to obtain the final objective function; Specifically, the method of using manifold regularization terms uses the intrinsic geometric structure of the data to improve the performance of the learning algorithm. If two samples are close in high-dimensional space, they should also be close in low-dimensional manifold space, and their outputs should also be similar. Construct an adjacency graph of samples to capture the underlying manifold structure of label propagation. The adjacency graph is an undirected graph that can be expressed as: (14) in, Represents the i-th point The vector of corresponding characteristic parameters and the jth point The connection weights between the vectors composed of the corresponding feature parameters, Represents the i-th point The mapping result of represents the jth point The mapping result is and Similar, then and It should be similar in any transformation, Tr is the operator, express - The 2nd-norm of . is defined as the Laplacian graph where is a diagonal matrix with diagonal elements , The mapping should have , then the weight of the adjacency graph is: (15) in, express of nearest neighbor set, combined and have: (16) In the formula, represents the weight of the training matrix; A matrix representing the expected output of the width learning fault diagnosis model; represents the input data matrix of the width learning fault diagnosis model; A weight matrix representing a node in a width learning fault diagnosis model; represents the first regularization parameter; represents the second regularization parameter; represents a Laplacian graph; represents the operator; the introduction of manifold regularization term can make the proposed method more robust.

[0032] In the actual calculation process, Generated by iterative update of error residual, let the error matrix be: (17) in, , each iteration is based on get ,Will Abbreviated as , take the pair The derivative of gets: (18) make , then: (19) So given the error weight , feature transformation and Laplace Under the premise of equal: (20) S5. Input the input data at the current moment into the target width learning fault diagnosis model, and output the status data of the electromechanical equipment.

[0033] The method of this embodiment is verified by combining specific data as follows: Step 1: Verify the robustness of the GBLS fault diagnosis model: In order to verify the advantages of the width learning method based on graph regularization in the label error scenario, a comparative analysis is performed with the width learning method. Represents the label error rate, that is, the percentage of label error data in the training set to the total data, and the hyperparameter , , The grid search method is used to query the optimal solution in the range of 1-10, 1-10, and 2-50 respectively. The error rate of the same label In this case, the two models use the same hyperparameters. For example, in the case of 5% label error, the hyperparameters of the two models are set uniformly. is 6, is 6, is 49. The regularization terms in the width learning method based on graph regularization are is 2, for , the weight function is selected as Huber function, and the results are shown in Table 1: Table 1

[0034] The following conclusions can be drawn from Table 1: 1) In the case of label errors, the width learning method based on graph regularization can maintain a high diagnostic accuracy, which proves that the width learning method based on graph regularization has good robustness. 2) Compared with the width learning method, the method proposed in this invention has a higher fault diagnosis accuracy. That is, when there is no wrong label, the fault diagnosis accuracy of width learning based on graph regularization and width learning both reach the highest, among which GBLS performs better. As the number of errors increases, the advantages of the proposed method become more and more obvious. Taking 20% ​​label error as an example, the proposed method improves the fault diagnosis accuracy by 18.6% compared with width learning, achieving good results. 3) The training time of the model does not increase with This shows that GBLS does not incur additional time cost due to processing wrong labels, and each training time is stable at about 0.16s, which further reflects the advantages of the method proposed in the present invention.

[0035] Step 2: Discuss the impact of hyperparameters on the performance of the width learning fault diagnosis model: The fault diagnosis performance of the width learning method based on graph regularization is analyzed under different hyperparameters. , through the grid search method , , The optimal solutions of the three hyperparameters are searched, and the results are shown in Table 2.

[0036] Table 2

[0037] The following conclusions can be drawn from Table 2: 1) As the number of network nodes increases, the fault diagnosis accuracy of width learning and width learning methods based on graph regularization will be improved to a certain extent. However, as the number of nodes increases, the training time also increases. In practical applications, it is necessary to comprehensively determine the appropriate hyperparameters according to the situation.

[0038] 2) When using the grid search method to discuss the selection of model hyperparameters, it is important to select an appropriate range. Otherwise, the optimal solution may appear at the boundary range. Although the increase in parameters makes the optimal solution continue to increase, the improvement is not obvious. In scenarios where the accuracy of fault diagnosis is not required, discussing the optimal solution of hyperparameters within an appropriate range can reduce the time cost required for training.

[0039] Figure 3 Yes For further discussion, is 0.05, is 10, is 10, The results from 1-50 are as follows, and we can conclude that: As the number of nodes increases, that is, the number of enhanced nodes increases, the fault diagnosis accuracy of the width learning method based on graph regularization and the width learning method gradually increases, and then tends to be stable, and the effect is improved to a certain extent, which further verifies the conclusion drawn above. At the same time, it can be found that the difference in the fault diagnosis accuracy of the two methods is almost unchanged, which shows that it is mainly caused by The value of is determined by , and has no obvious relationship with the value of the enhancement node.

[0040] In addition, the width learning method based on graph regularization has two parameters , ,in Its role is to ensure that the output weights conform to the manifold learning criteria. The function is to smooth the solution of the output weight W. These two parameters can balance the corresponding constraints and play an important role in the performance of the model. The value in is Figure 4 As shown. From this we can conclude that: , The value of has a great influence on the accuracy of model fault diagnosis. , The value is , the accuracy of the model dropped significantly. In comparison, when , The value is , the diagnostic accuracy rate remains at a high level.

[0041] The above experiments show that the graph regularized width learning electromechanical equipment fault diagnosis method of the present invention performs excellently in the NASA actuator dataset fault diagnosis problem, can effectively deal with the dataset mislabeling problem, while retaining the learning speed of the BLS method, and can complete model training in a shorter time on an ordinary PC, and has a higher accuracy rate than BLS.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A graph regularized width learning method for electromechanical equipment fault diagnosis, characterized in that: include: Acquire the characteristic parameters and state data of the actuator at each operating moment, take the characteristic parameters as input data, take the state data corresponding to the actuator as output data, and construct a data set; wherein the characteristic parameters include: temperature data of the actuator and acceleration data of three axes; the state data include: normal operation, blocking fault, spalling fault and motor fault; A width learning fault diagnosis model is constructed based on the data set; the objective function improvement method based on maximum likelihood estimation is used to improve the objective function of the width learning fault diagnosis model to obtain an improved objective function; Based on the manifold learning method, a regularization term is added to the improved objective function to obtain the final objective function; The objective function of the width learning fault diagnosis model is replaced by the final objective function to obtain the final width learning fault diagnosis model; The final width learning fault diagnosis model is trained using the data set to obtain a trained target width learning fault diagnosis model; The input data at the current moment is input into the target width learning fault diagnosis model, and the status data of the electromechanical equipment is output.

2. According to claim 1, a method for fault diagnosis of electromechanical equipment based on graph regularization by width learning is characterized in that: The feature parameters are cleaned and standardized in turn, and the standardized feature parameters are used as input data.

3. The method for fault diagnosis of electromechanical equipment based on incremental fuzzy width learning of streaming data according to claim 1 is characterized in that: The steps for data cleaning of feature parameters are as follows: Traverse the collected characteristic parameters and mark the missing values ​​and outliers. The outliers are determined by the 6σ criterion. The missing values ​​and outliers are replaced by the neighboring value filling method to complete data cleaning.

4. The method for diagnosing electromechanical equipment faults based on incremental fuzzy width learning of streaming data according to claim 1 is characterized in that: The steps for data standardization of feature parameters are: using the Z-score standardization method to perform data standardization on the data set.

5. The method for fault diagnosis of electromechanical equipment based on graph regularization based on width learning according to claim 1, characterized in that: The objective function of the width learning fault diagnosis model is expressed as: In the formula, , represents the error probability density function; represents the weight matrix of the nodes in the width learning fault diagnosis model; represents the constraint coefficient; represents the distribution parameter; represents independent and identically distributed errors; express; It means to find the second-order F-norm.

6. The method for fault diagnosis of electromechanical equipment based on graph regularization width learning according to claim 1, characterized in that: The improved objective function is: In the formula, represents the weight of the training matrix; A matrix representing the expected output of the width learning fault diagnosis model; represents the input data matrix of the width learning fault diagnosis model; represents the weight matrix of the nodes in the width learning fault diagnosis model; represents the constraint coefficient; It means to find the second-order F-norm.

7. The method for fault diagnosis of electromechanical equipment based on graph regularization width learning according to claim 1, characterized in that: The final objective function expression is: In the formula, represents the weight of the training matrix; A matrix representing the expected output of the width learning fault diagnosis model; represents the input data matrix of the width learning fault diagnosis model; represents the weight matrix of the nodes in the width learning fault diagnosis model; represents the first regularization parameter; represents the second regularization parameter; represents a Laplacian graph; Represents an operator.

Images