An Incremental Fuzzy Width Learning Fault Diagnosis Method for Electromechanical Equipment Based on Streaming Data

Abstract

The present invention relates to an incremental fuzzy width learning-based electromechanical equipment fault diagnosis method for streaming data, including: obtaining characteristic parameters and status data of an actuator at each operating moment; constructing a fault diagnosis model; obtaining a target data set; adding additional nodes to an enhancement layer and obtaining a target fault diagnosis model; and using the target fault diagnosis model to obtain an electromechanical equipment fault diagnosis result. The present invention, through a time-weighted incremental clustering method, introduces a weight decay mechanism to adapt to data changes, dynamically adjusts the positions and quantities of fuzzy width learning clustering center points, enables the model to have an online update function in a real-time data stream scenario, and at the same time improves the diagnostic accuracy rate of the model by incrementally updating the additional nodes.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and particularly relates to an incremental fuzzy width learning electromechanical equipment fault diagnosis method based on streaming data. Background Technique

[0002] With the continuous development of modern technology, complex electromechanical equipment has been widely used in the national defense field. For the intelligent fault diagnosis of complex electromechanical equipment, it has important theoretical significance and application value for ensuring the safety and stability of the system, promptly eliminating faults, and avoiding major casualties and property losses. The fault diagnosis method based on machine learning establishes an intelligent fault diagnosis model according to historical data through machine learning technology, without the need for complex physical and mathematical models, and has become an important part of the research on fault diagnosis technology for complex electromechanical systems.

[0003] In practical engineering applications, how to process continuously collected new samples has become an important research content. Through the current literature search, it is found that using incremental learning for real-time health monitoring and diagnosis model update of complex electromechanical equipment can perform fault diagnosis more efficiently. For example, Dai Jinling et al. combined a type of kernel extreme learning machine and data incremental learning in the "Online Fault Detection Method Based on OCKELM and Incremental Learning" to achieve online fault diagnosis of sequentially arriving sample data; Wang Wenbo et al. proposed a rolling bearing fault prediction method based on deep learning and incremental update in the "Research on Rolling Bearing Fault Prediction Method Based on Deep Learning and Incremental Update", enabling the model to have the ability of parameter self-learning and self-update; Lu Songfeng et al. proposed a hierarchical federated incremental learning algorithm based on the cloud, edge, and end architecture in the "Cloud-Edge-Node Collaborative Incremental Federated Learning Algorithm".

[0004] The prior art proposes an efficient incremental learning system without a deep architecture - width learning. Compared with the "deep" network structure favored by deep learning, width learning is more inclined to the "width" development of the network, which makes it have the advantages of simple structure and fast training speed. At the same time, researchers have proposed fuzzy width learning, which combines the Takagi - Sugeno fuzzy system with BLS, improving the model performance while retaining the advantage of fast calculation speed of width learning. In the field of fault diagnosis, width learning provides a fast and efficient method for fault diagnosis. Jiang S B et al. applied incremental width learning to the fault diagnosis of three - phase induction motors in "An efficient fault diagnostic method for three - phase induction motors based on incremental broad learning and non - negative matrix factorization", and optimized the network using empirical mode decomposition, sample entropy, and non - negative matrix factorization; and in a deep - wide learning framework based on domain knowledge, CNN was combined with BLS. However, the width learning method mainly adopts the batch update method. In actual industrial monitoring, due to the variable operating conditions of equipment, once the traditional fault diagnosis model is trained, it cannot be updated continuously, thus affecting the judgment accuracy of the model.

[0005] Therefore, it is necessary to provide an incremental fuzzy width learning electromechanical equipment fault diagnosis method based on streaming data to solve the above problems. Summary of the Invention

[0006] The present invention provides an incremental fuzzy width learning electromechanical equipment fault diagnosis method based on streaming data, which adopts the "streaming data incremental learning" strategy to construct a fault diagnosis model. At the same time, based on the "fuzzy width learning" strategy, the diagnostic accuracy of the fault diagnosis model is improved to solve the problem that in actual industrial monitoring, due to the variable operating conditions of equipment, once the traditional fault diagnosis model is trained, it cannot be updated continuously, thus affecting the judgment accuracy of the model.

[0007] The incremental fuzzy width learning electromechanical equipment fault diagnosis method of the present invention adopts the following technical solutions, including:

[0008] Obtain the characteristic parameters and status data of the actuator at each operating moment, use the characteristic parameters as input data, and use the corresponding status data of the actuator as output data to construct a data set; wherein, the characteristic parameters include: the temperature data of the actuator and the acceleration data in three axial directions; the status data includes: normal operation, blockage fault, spalling fault, and motor fault;

[0009] Construct a fault diagnosis model, which includes an enhancement layer and a mapping layer. The mapping layer is used to perform fuzzy mapping on the input data and output data in the dataset, and output the defuzzified mapping result; the enhancement layer is used to perform non-linear transformation on the defuzzified mapping result and output the fault diagnosis result.

[0010] Set a clustering cluster for each characteristic parameter, set the maturity of the clustering cluster according to the time decay function, and obtain the clustering cluster of each characteristic parameter at each moment according to the maturity of the clustering cluster, the linear weighted sum of all characteristic parameters in each characteristic parameter of the input data of the dataset, and the weighted sum of squares of all characteristic parameters in each characteristic parameter of the input data of the dataset.

[0011] Obtain the clustering center and clustering radius of the clustering cluster, dynamically update the clustering center point and clustering radius of the clustering cluster at the next moment through the clustering radius and radius threshold of the clustering cluster, obtain the target clustering center point and target clustering radius of the clustering cluster, remove the data that does not belong to the range of the clustering center point and target clustering radius in the clustering cluster, and obtain the target clustering cluster of each characteristic parameter; obtain the target dataset according to the target clustering cluster of each characteristic parameter.

[0012] Add the increasing nodes of the enhancement layer of the fault diagnosis model according to the preset number of enhancement nodes, and obtain the fault diagnosis model after adding the enhancement nodes. Obtain the loss function value of the enhanced fault diagnosis model according to the target dataset. When the loss function value does not meet the preset loss function threshold, then add the preset number of enhancement nodes again until the loss function value of the enhanced fault diagnosis model meets the preset loss function threshold, and then use the enhanced fault diagnosis model at this time as the target fault diagnosis model.

[0013] Input the input data at the current moment into the target fault diagnosis model, and output the fault diagnosis result of the electromechanical equipment.

[0014] Preferably, perform data cleaning and data standardization on the characteristic parameters in sequence, and use the standardized characteristic parameters as the input data.

[0015] Preferably, the steps for data cleaning of the characteristic parameters are as follows:

[0016] Traverse the collected characteristic parameters, mark the missing values and abnormal values, where the determination of abnormal values adopts the 6σ criterion.

[0017] Replace the missing values and abnormal values using the adjacent value filling method to complete the data cleaning.

[0018] Preferably, the steps for data standardization of the characteristic parameters are as follows: Use the Z-score standardization method to perform data standardization processing on the dataset.

[0019] Preferably, the mapping layer includes: a fuzzification mapping layer and a defuzzification mapping layer; the fuzzification mapping layer is used to perform fuzzification mapping on the input data and output data in the dataset, and the defuzzification mapping layer is used to defuzzify the fuzzified output;

[0020] The model expression of the fuzzification mapping layer is:

[0021]

[0022]

[0023]

[0024]

[0025] In the formula, represents the output value of the th characteristic parameter under the kth rule of the ith fuzzy subsystem; represents the normalized weight of the th characteristic parameter under the kth rule of the ith fuzzy subsystem; represents the weighted summation function; represents the th characteristic parameter and the mth parameter; represents the th characteristic parameter and the fuzzy coefficient of the tth parameter under the kth rule of the ith fuzzy subsystem; represents the th characteristic parameter and the tth parameter; represents the output when the th characteristic parameter is input to the ith fuzzy subsystem; represents the output when the th characteristic parameter is input to the ith fuzzy subsystem; represents the output of the ith fuzzy subsystem among fuzzy subsystems; represents the output of the th fuzzy subsystem among fuzzy subsystems;

[0026] The model expression of the defuzzification mapping layer is:

[0027]

[0028] ,......,

[0029]

[0030] In the formula, represents the defuzzification output when the fuzzy mapping layer is fuzzy subsystems; represents the defuzzification output of the i-th fuzzy subsystem among fuzzy subsystems; represents the defuzzification output of the -th fuzzy subsystem; represents the defuzzification output of the i-th fuzzy subsystem corresponding to the -th type of characteristic parameter; represents the defuzzification output of the i-th fuzzy subsystem corresponding to the -th type of characteristic parameter; is the mapping coefficient of the s-th type of characteristic parameter under the k-th rule of the i-th fuzzy subsystem; ; represents the weight of the i-th fuzzy subsystem, .

[0031] Preferably, the expression of the enhancement layer is:

[0032]

[0033]

[0034] In the formula, represents the output when the enhancement layer has m enhancement nodes, represents the output of the -th enhancement node of the enhancement layer; represents the output of the -th enhancement node of the enhancement layer; represents the output when the fuzzy mapping layer is fuzzy subsystems; represents the random weight of the j-th enhancement node of the enhancement layer; represents the random drift of the j-th enhancement node of the enhancement layer; represents the random mapping value of the j-th enhancement node of the enhancement layer.

[0035] Preferably, the expression of the fault diagnosis model is:

[0036] Let , then:

[0037]

[0038] In the formula, Represents the output of the fault diagnosis model; The weights of the fault diagnosis model; Represents the weights of the enhancement layer; Represents the output when the enhancement layer has m enhancement nodes; Represents the operation of Defining the previous value as the symbol The following value; Represents the weights of the mapping layer; Represents the rule output of the fuzzy subsystem in the mapping layer; Represents the weights of the i-th fuzzy subsystem in the mapping layer; Represents the mapping coefficient of the i-th fuzzy subsystem in the mapping layer.

[0039] Preferably, the clustering cluster is represented by a triple where, Represents the vector of the linear weighted sum of all feature parameters of the input data of the data set; Represents the vector of the weighted sum of squares of all feature parameters in the input data of the data set; Represents the maturity; where, the cluster center and cluster radius of the clustering cluster are:

[0040]

[0041]

[0042] In the formula, ; Represents the cluster radius of the clustering cluster; Represents the number of types of feature parameters in the input data; Represents the cluster center of the clustering cluster.

[0043] Preferably, the steps of dynamically updating the cluster center point and cluster radius of the clustering cluster at each moment through the cluster radius of the clustering cluster and the radius threshold are:

[0044] Merge the j-th type of feature parameter at the corresponding moment when the cluster radius of the clustering cluster is less than the radius threshold into the clustering cluster at each moment to obtain a new clustering cluster;

[0045] Obtain the cluster center point and cluster radius of the new clustering cluster.

[0046] Preferably, the expression of the fault diagnosis model after each addition of an enhancement node is:

[0047]

[0048]

[0049]

[0050] In the formula, represents the fault diagnosis model after the first addition of enhancement nodes; represents the network structure of the fault diagnosis model before adding enhancement nodes; represents the network structure of the fault diagnosis model after the first addition of enhancement nodes; represents the target weight of the fault diagnosis model after the first addition of enhancement nodes; represents the weight of the fault diagnosis model before adding enhancement nodes; represents the "pseudo-inverse" of the network structure matrix of the network structure of the fault diagnosis model before enhancement nodes; represents the first intermediate quantity obtained by calculating the "pseudo-inverse" of the network structure matrix of the network structures of the fault diagnosis models before and after enhancement nodes; represents the second intermediate quantity obtained by calculating the "pseudo-inverse" of the network structure matrix of the network structures of the fault diagnosis models before and after enhancement nodes.

[0051] The beneficial effects of the present invention are as follows:

[0052] By introducing a weight decay mechanism based on a time-weighted incremental clustering method when processing real-time data streams to adapt to data changes, dynamically adjusting the positions and quantities of fuzzy width learning clustering center points, enabling the model to have an online update function in real-time data stream scenarios, and at the same time, by incrementally updating the enhancement nodes, improving the diagnostic accuracy rate of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0054] Figure 1 is a flowchart of an incremental fuzzy width learning electromechanical equipment fault diagnosis method based on stream data of the present invention;

[0055] Figure 2 is a flowchart of the time-weighted incremental clustering step in the embodiment of the present invention;

[0056] Figure 3 is a schematic diagram of the comparison of the diagnostic accuracy rate between the method of the present invention and FBLS in the embodiment;

[0057] Figure 4 is a schematic diagram of the comparison of the training time between the method of the present invention and FBLS in the embodiment;

[0058] Figure 5Method parameters of the present invention in the embodiments Schematic diagram of the influence on the diagnostic accuracy rate of RFBLS Specific implementation manners

[0059] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0060] An embodiment of a method for diagnosing faults of electromechanical equipment based on incremental fuzzy width learning of streaming data according to the present invention is as Figure 1 shown and includes:

[0061] S1. Obtain the characteristic parameters and status data of the actuator at each running moment;

[0062] Specifically, obtain the characteristic parameters and status data of the actuator at each running moment, use the characteristic parameters as input data, and use the status data corresponding to the actuator as output data to construct a data set; among them, the characteristic parameters include: temperature data of the actuator and acceleration data in three axial directions; the status data includes: normal operation, blockage fault, spalling fault, and motor fault.

[0063] In this embodiment, the characteristic parameters are sequentially subjected to data cleaning and data standardization, and the standardized characteristic parameters are used as input data. Among them, the steps for data cleaning of the characteristic parameters are: traverse the collected characteristic parameters, mark the missing values and abnormal values. Among them, the determination of abnormal values uses the 6σ criterion, that is, a number greater than the sum of the mean and 6 times the standard deviation is regarded as an abnormal value; the missing values and abnormal values are replaced by the adjacent value filling method, that is, the average value of 5 points before and after the missing value or abnormal value is used for filling and replacement to complete the data cleaning. The Z-score standardization method is used to process the data set to eliminate the differences between dimensions, complete the data standardization, and use the finally completed standardized characteristic parameters as input data.

[0064] S2. Construct a fault diagnosis model;

[0065] Specifically, construct a fault diagnosis model. The fault diagnosis model includes an enhancement layer and a mapping layer. The mapping layer is used to perform fuzzy mapping on the input data and output data in the data set and output a defuzzified mapping result; the enhancement layer is used to perform non-linear transformation on the defuzzified mapping result and output a fault diagnosis result.

[0066] The method for establishing the fault diagnosis model in this embodiment is the same as the model establishment method in "Fuzzy broad learning system: Anovel neuro-fuzzy model for regression and classification". Specifically, the mapping layer includes: a fuzzification mapping layer and a defuzzification mapping layer; the fuzzification mapping layer is used to perform fuzzification mapping on the input data and output data in the dataset, and the defuzzification mapping layer is used to defuzzify the fuzzified output;

[0067] Step 21, establish the fuzzification mapping layer model:

[0068] In the Takagi-Sugeno fuzzy system, let the characteristic parameter , , , assume that there are rules in the th fuzzy subsystem in the fuzzy system, then there are rules, that is …… , , , then there is:

[0069] (1)

[0070] In the formula, represents the output value of the th characteristic parameter under the kth rule of the ith fuzzy subsystem; represents the fuzzy coefficient of the tth parameter in the th characteristic parameter under the kth rule of the ith fuzzy subsystem; represents the tth parameter in the th characteristic parameter; represents the weighted summation function; represents the mth parameter in the th characteristic parameter.

[0071] Then, the expression of the th characteristic parameter corresponding to the th fuzzy subsystem is:

[0072] (2)

[0073] In the formula, represents the output when the th characteristic parameter input corresponds to the i-th fuzzy subsystem; represents the th characteristic parameter under the k-th rule of the i-th fuzzy subsystem; represents the th characteristic parameter under the k-th rule of the i-th fuzzy subsystem; where, , , , here, represents the th characteristic parameter under the k-th rule of the i-th fuzzy subsystem; represents the th characteristic parameter under the k-th rule of the i-th fuzzy subsystem; the th parameter of the th Gaussian distribution with a mean of and a standard deviation of is taken as 1 for ease of calculation, and the mean is directly obtained by the algorithm; the output of the

[0074] th fuzzy subsystem for all input parameters in the dataset is:

[0075] Then, the output when the fuzzy mapping layer is fuzzy subsystems is:

[0076] (4)

[0077] In the formula, represents the output when the th characteristic parameter input corresponds to the i-th fuzzy subsystem; represents the output when the th characteristic parameter input corresponds to the i-th fuzzy subsystem; represents the output of the i-th fuzzy subsystem in fuzzy subsystems; represents the output of the th fuzzy subsystem in fuzzy subsystems;

[0078] Step 22, establish a defuzzification mapping layer:

[0079] To retain the input characteristics, the defuzzified output of the i-th fuzzy subsystem for the s-th characteristic parameter is: is:

[0080] (5)

[0081] where represents the defuzzified output of the i-th fuzzy subsystem corresponding to the th type of characteristic parameter; is the mapping coefficient of the s-th characteristic parameter under the k-th rule of the i-th fuzzy subsystem.

[0082] Then the defuzzified output of the th fuzzy subsystem is Equation (6), and the fuzzy mapping layer is Equation (7):

[0083] (6)

[0084] (7)

[0085] where represents the defuzzified output when the fuzzy mapping layer is fuzzy subsystems; represents the defuzzified output of the i-th fuzzy subsystem among the fuzzy subsystems; represents the defuzzified output of the i-th fuzzy subsystem corresponding to the th type of characteristic parameter; represents the defuzzified output of the i-th fuzzy subsystem corresponding to the th type of characteristic parameter; is the mapping coefficient of the s-th characteristic parameter under the k-th rule of the i-th fuzzy subsystem; ; represents the weight of the i-th fuzzy subsystem, .

[0086] Step 23. Establish the enhancement layer as:

[0087] The outputs of the fuzzy rules generated by each fuzzy subsystem are not immediately aggregated into a single value, but are all sent to the enhancement layer for further non-linear transformation, and the output of the th enhancement layer node is:

[0088] (8)

[0089] Then, the output when the enhancement layer has m enhancement nodes, that is, the enhancement layer is:

[0090] (9)

[0091] In the formula, represents the output when the enhancement layer has m enhancement nodes, represents the output of the m-th enhancement node in the enhancement layer; represents the -th enhancement node in the enhancement layer; represents the output when the fuzzy mapping layer has fuzzy subsystems; represents the random weight of the j-th enhancement node in the enhancement layer; represents the random drift of the j-th enhancement node in the enhancement layer; represents the random mapping value of the j-th enhancement node in the enhancement layer.

[0092] Step 24: Establish a fault diagnosis model:

[0093] Combining the mapping layer and the enhancement layer, the final fault diagnosis model obtained is:

[0094] (10)

[0095] In the formula, is the weight of the enhancement layer, represents the output of the fault diagnosis model;

[0096] Let , then:

[0097] (11)

[0098] In the formula, represents the output of the fault diagnosis model; the weight of the fault diagnosis model; represents the weight of the enhancement layer; represents the output when the enhancement layer has m enhancement nodes; represents defining the value before as the symbol and the value after; represents the weight of the mapping layer; represents the rule output of the fuzzy subsystems in the mapping layer; represents the weight of the i-th fuzzy subsystem in the mapping layer; represents the mapping coefficient of the i-th fuzzy subsystem in the mapping layer.

[0099] Using the pseudo-inverse solution technology of ridge regression, the weight of the fault diagnosis model is obtained as:

[0100] (12)

[0101] Among them, .

[0102] S3. Obtain the target data set;

[0103] Specifically, set a clustering cluster for each feature parameter, set the maturity of the clustering cluster according to the time decay function, and obtain the clustering cluster of each feature parameter at each moment according to the maturity of the clustering cluster, the linear weighted sum of all feature parameters in each feature parameter of the input data of the data set, and the weighted sum of squares of all feature parameters in each feature parameter of the input data of the data set;

[0104] Obtain the clustering center and clustering radius of the clustering cluster, dynamically update the clustering center point and clustering radius of the clustering cluster at the next moment through the clustering radius and radius threshold of the clustering cluster, obtain the target clustering center point and target clustering radius of the clustering cluster, remove the data that does not belong to the range of the clustering center point and target clustering radius in the clustering cluster, and obtain the target clustering cluster of each feature parameter; obtain the target data set according to the target clustering cluster of each feature parameter.

[0105] In this embodiment, an incremental clustering step based on time weighting is proposed. By setting the clustering cluster and dynamically adjusting the clustering center point, it is adapted to the model update in the data stream scenario. Specifically, the incremental clustering step is as follows:

[0106] Set a clustering cluster as , and use the triple to represent that the clustering cluster is . When the time is , the feature parameter in the clustering cluster is , is the feature parameter at the nth moment, where , is the jth feature parameter at the ith moment; is the dth feature parameter at the ith moment, and the time stamp at each moment is , then the expression of the maturity is:

[0107] (13)

[0108] Among them, , represents the maturity threshold for whether a cluster is mature. When is greater than or equal to , it is a mature cluster. When is less than , it is an immature cluster. represents the time stamp at the ith moment; is a weight decay function. As time goes by it decays exponentially. Through this function, the weight of each sample gradually decreases over time, so that the attention to new feature parameters is greater than that to the old feature parameters in the dataset. For the model training method based on incremental learning, since it needs to continuously learn and update the model based on the newly collected data after deployment, it is necessary to make the model pay more attention to new samples (new feature parameters) so as to learn new knowledge. Generally , is a parameter for controlling the decay rate. For this embodiment it takes the value of 0.02.

[0109] (14)

[0110] Among them is the linear weighted sum of feature parameters within the time period from 1 to n, represents the vector of the linear weighted sum of all feature parameters of the input data of the dataset.

[0111] (15)

[0112] Among them is the weighted sum of squares of feature parameters within the time period from 1 to n; represents the vector of the weighted sum of squares of all feature parameters in the input data of the dataset.

[0113] Let the center point of the clustering cluster be c (to be consistent with other clustering algorithms, the position vector of the clustering center of the clustering cluster is uniformly called the clustering center), and the radius be , then there is:

[0114] (16)

[0115] In the formula, ; is the position vector of the clustering center, j = 1, 2, 3, 4;

[0116] (17)

[0117] In the formula, represents the clustering radius of the clustering cluster; represents the number of types of feature parameters in the input data.

[0118] Next, during the update process over time, as long as, , is the maximum threshold radius of a clustering cluster corresponding to each characteristic parameter, then at the

[0119] (18)

[0120] At time, the three decay to:

[0121] (19)

[0122] Here it is assumed that the clustering cluster does not receive new data that meets the conditions (that is, the obtained through new input data) within the time period.

[0123] The updated clustering center and clustering radius of the clustering cluster corresponding to each characteristic parameter are:

[0124] (20)

[0125] (21)

[0126] Remove the data in the clustering cluster that does not belong to the range of the clustering center point and the target clustering radius to obtain the target clustering cluster of each characteristic parameter; obtain the target data set according to the target clustering cluster of each characteristic parameter.

[0127] In this embodiment, the clustering cluster evolves over time, and the weight decay mechanism gives higher attention to newer samples than older samples. Since the weight of each clustering cluster decays exponentially over time, the model can adapt to concept drift. In the subsequent use process, when the clustering cluster is a mature cluster, the mature cluster is used as the actually existing clustering center, and the immature cluster is recorded. When the immature cluster becomes mature, it is used as the clustering center. At the same time, over time, when the weight of the mature cluster decreases and becomes an immature cluster, the clustering center is removed in a timely manner. As Figure 2 shown, let , the maturity threshold , , for the maturity process of the clustering cluster, when , let the input data be , at this time a new cluster is formed. According to the above formulas (13) to (21), can be obtained. At this time is an immature cluster; when , let the new input data be , at this time , , at this time , is still an immature cluster; when , let the new input data be , , , , at this time , becomes a mature cluster. Here, it is defaulted that in all cases, the radius of the mature cluster must be less than the maximum threshold radius . For the process of the clustering cluster from mature to immature, at , let the new input data be , let the radius to the clustering center be greater than , at this time, a new cluster , , is an immature cluster, and at the same time , , , , remains a mature cluster; when , let the new input data be , and it belongs to the immature cluster , at this time , , , is still an immature cluster, and at the same time , , , , changes from a mature cluster to an immature cluster.

[0128] S4. Add the increasing nodes of the enhancement layer and obtain the target fault diagnosis model;

[0129] Since step S2 establishes a fault diagnosis model, and step 4 uses the incremental learning method to update the model established in step 2. Different from the traditional model whose model parameters are fixed once trained and deployed, unable to be changed and unable to learn new data features, this patent enables the model to still perform self-learning and self-update based on new data in an incremental learning manner, which better meets the industrial implementation requirements. Therefore, in this embodiment, the number of enhanced nodes in the enhanced layer of the fault diagnosis model is increased according to a preset quantity, and the fault diagnosis model after adding the enhanced nodes is obtained. The loss function value of the enhanced fault diagnosis model is obtained according to the target data set. When the loss function value does not meet the preset loss function threshold, the preset quantity of enhanced nodes is added again until the loss function value of the enhanced fault diagnosis model meets the preset loss function threshold. Then, the enhanced fault diagnosis model at this time is used as the target fault diagnosis model.

[0130] In this embodiment, it is assumed that 20 enhanced nodes are added each time. The network structure of the fault diagnosis model before adding the enhanced nodes and the weights are , and the network structure after adding the enhanced nodes and the weights are , is the weight of the newly added 20 enhanced nodes. Then, the network structure after adding the enhanced nodes each time is:

[0131] (22)

[0132] Among them, , so the weight after adding the enhanced nodes each time is:

[0133] (23)

[0134] According to the network structure and the weights after adding the enhanced nodes each time, the fault diagnosis model after adding the enhanced nodes each time can be obtained. The fault diagnosis model after adding the enhanced nodes each time is:

[0135]

[0136] In the formula, represents the fault diagnosis model after the first addition of enhanced nodes; represents the network structure of the fault diagnosis model before adding the enhanced nodes; represents the network structure of the fault diagnosis model after the first addition of enhanced nodes; represents the target weight of the fault diagnosis model after the first addition of enhanced nodes; represents the weight of the fault diagnosis model before adding the enhanced nodes; denotes the "pseudo-inverse" of the network structure matrix of the fault diagnosis model before the enhanced node; denotes the first intermediate quantity obtained by taking the "pseudo-inverse" of the network structure matrix of the fault diagnosis model before and after the enhanced node; denotes the second intermediate quantity obtained by taking the "pseudo-inverse" of the network structure matrix of the fault diagnosis model before and after the enhanced node.

[0137] Test the fault diagnosis model after adding an enhanced node each time according to the target data set. According to the loss function value of the fault diagnosis model after adding the enhanced node and the preset loss function threshold (in this embodiment, it is 10 -6 ), if the loss function value is greater than or equal to the loss function threshold, then add an enhanced node again until the loss function value is less than the loss function threshold, and then use the fault diagnosis model after adding the enhanced node this time as the target fault diagnosis model.

[0138] S5. Use the target fault diagnosis model to obtain the fault diagnosis result of the electromechanical equipment;

[0139] Input the input data at the current moment into the target fault diagnosis model to output the fault diagnosis result of the electromechanical equipment.

[0140] The present invention will be described below with reference to the accompanying drawings:

[0141] In this embodiment, taking an aviation actuator as an example, four characteristic parameters, namely, the X-axis acceleration, Y-axis acceleration, Z-axis acceleration, and temperature of the aviation actuator, are selected for experiments, and four conditions, namely, normal operation, blockage fault, spalling fault, and motor fault, are used as state parameters. The selected experimental data set contains a total of 600 groups of original data sets.

[0142] Step 1. Verify the effectiveness of obtaining the target data set in step S3:

[0143] Analyze the performance of RFBLS (real-time data flow model) after introducing time-weighted incremental clustering in step S3, and set the number of nodes in the mapping layer to 14, and the number of nodes added each time in the enhancement layer is 60. Let , , , and the results are as shown in Figure 3 , Figure 4 . Among them, Figure 3 represents the relationship between the number of clustering centers and RFBLS and FBLS, and Figure 4 represents the comparison of the training time of RFBLS and FBLS (retrained model). The following conclusions can be drawn from the figure:

[0144] 1) Under lower time costs, the average diagnosis rate of fault diagnosis in real-time data streams reaches 99%, approaching the correct diagnosis rate of the FBLS fault diagnosis in the cloud. However, compared with FBLS (re-training), RFBLS requires less training time for the model update process, about half of the former, which demonstrates the feasibility of the method proposed by the present invention and its advantages over the FBLS fault diagnosis method in real-time data stream scenarios.

[0145] 2) As the clustering centers change, the correct diagnosis rates of both RFBLS and FBLS first decline, then fluctuate, and finally gradually increase steadily. This is because initially, there is less data, the model fitting accuracy is relatively high (and even overfitting may occur), so the correct diagnosis rate is relatively high. As the data increases, the diagnosis rate decreases. At the same time, the increase in data also leads to an increase in the number of clustering centers. However, when the number of centers increases, the diagnosis rate will increase again, so it shows fluctuations. Finally, when the number of clustering centers is relatively large (the network structure is relatively complex) and the data volume is also relatively large, the impact of a single piece of data on the diagnosis rate is relatively small, so it shows a stable or even slightly increasing trend.

[0146] 3) As the clustering centers change, the diagnosis rate will fluctuate accordingly, and the diagnosis rate of RFBLS is smoother. Although the diagnosis rate of FBLS is higher, it is actually found through comparison that the difference between the two does not exceed 0.2%. Analyzing the reason why the fluctuation of RFBLS is smaller than that of FBLS, because FBLS re-clusters each time it trains, which may lead to a relatively large difference in clustering centers between two adjacent times, so there will also be a relatively large difference in the diagnosis rate. While RFBLS adopts the method of incremental clustering, the change in clustering centers between two adjacent times is relatively small, so it is more stable.

[0147] 4) As the data volume increases, the time costs required by both RFBLS and FBLS are increasing. Analyzing the reason for the increase in the training time required for the model to receive data and complete one update, in FBLS, as more data needs to be processed for each clustering in real-time data streams, the training time increases. In RFBLS, although there is no need to process historical data, as the data increases, the number of clusters will also increase, and the amount of data to be calculated for each model update increases, so the training time will also increase.

[0148] Step 2: Verify the effectiveness of node increment:

[0149] Set the hyperparameters of the fault diagnosis model: the number of network layers is 2, the number of nodes in the mapping layer is 6, the number of nodes added each time in the enhancement layer is 20, and the enhancement nodes increase dynamically by 20 each time, finally reaching 100. The results are shown in Table 1. It can be concluded that as the number of enhancement nodes increases, the model accuracy has been improved to a certain extent. This proves that the enhancement node increment method of RFBLS is effective. When the trained model cannot achieve the ideal performance, the diagnostic accuracy of the model can be improved by adding new enhancement nodes.

[0150] Table 1

[0151]

[0152] This embodiment mainly discusses the impact on the diagnostic accuracy rate, and sets to be 14, to be 60, is the parameter of the weight decay function, and set , , Within the range of , the results are as Figure 5 shown. It can be concluded from Figure 5 that as increases, the number of clustering centers as a whole shows a downward trend. This is mainly because the decay of mature clusters becomes faster, resulting in an increase in the situation of becoming immature clusters. This shows that the number of clustering centers does have an impact on the fault diagnosis accuracy rate. In summary, the incremental fuzzy width learning electromechanical equipment fault diagnosis method based on streaming data of the present invention performs excellently in the fault diagnosis problem of the NASA actuator dataset, can effectively process real-time data streams, while retaining the learning speed of the BLS method, can easily complete model training on an ordinary PC in a short time, and has a high fault diagnosis accuracy rate.

[0153] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An incremental fuzzy width learning electromechanical equipment fault diagnosis method based on streaming data, characterized in that Including: Obtain the characteristic parameters and status data of the actuator at each running moment, use the characteristic parameters as input data, and use the corresponding status data of the actuator as output data to construct a data set; among them, the characteristic parameters include: the temperature data of the actuator and the acceleration data in three axial directions; the status data includes: normal operation, blockage fault, spalling fault, and motor fault; Construct a fault diagnosis model, the fault diagnosis model includes an enhancement layer and a mapping layer, the mapping layer is used to perform fuzzy mapping on the input data and output data in the data set, and output the defuzzified mapping result; the enhancement layer is used to perform non-linear transformation on the defuzzified mapping result and output the fault diagnosis result; Set a clustering cluster for each characteristic parameter, set the maturity of the clustering cluster according to the time decay function, and obtain the clustering cluster of each characteristic parameter at each moment according to the maturity of the clustering cluster, the linear weighted sum of all characteristic parameters in each characteristic parameter of the input data of the data set, and the weighted sum of squares of all characteristic parameters in each characteristic parameter of the input data of the data set; Obtain the clustering center and clustering radius of the clustering cluster, dynamically update the clustering center point and clustering radius of the clustering cluster at the next moment through the clustering radius and radius threshold of the clustering cluster, obtain the target clustering center point and target clustering radius of the clustering cluster, remove the data that does not belong to the range of the clustering center point and target clustering radius in the clustering cluster, and obtain the target clustering cluster of each characteristic parameter; obtain the target data set according to the target clustering cluster of each characteristic parameter; Among them, the steps to obtain the target clustering cluster of each characteristic parameter are as follows: Assume a clustering cluster is , and use the triple to represent that the clustering cluster is . When the time is , the characteristic parameter within the clustering cluster is . is the characteristic parameter at the nth moment, where . is the dth characteristic parameter at the ith moment, and the time stamp for each moment is , that is the time stamp at the nth moment. The expression for maturity is: Among them, , represents a maturity threshold indicating whether a cluster is mature. When it is greater than or equal to , it is a mature cluster. When it is less than , it is an immature cluster. represents the timestamp at the i-th moment; is a weight decay function that decays exponentially over time . Through this function, the weight of each sample gradually decreases over time, that is, the attention to new feature parameters is greater than that to old feature parameters in the dataset. For the model training method based on incremental learning, since it needs to continuously learn and update the model based on the newly collected data after deployment, is a parameter for controlling the decay rate, and its value is 0.

02. Among them, is the linear weighted sum of feature parameters within the time period from 1 to n, a vector representing the linear weighted sum of all feature parameters of the input data of the data set; Among them, is the weighted sum of squares of feature parameters within the time period from 1 to n; is a vector representing the weighted sum of squares of all feature parameters in the input data of the data set; Let the center point of the clustering cluster be c and the radius be , then we have: In the formula, ; is the position vector of the clustering center, where j = 1, 2, 3, 4; In the formula, represents the clustering radius of the clustering cluster; represents the number of types of feature parameters in the input data; During the update process over time, as long as , is the maximum threshold radius of a clustering cluster corresponding to each characteristic parameter, then at time, the j-th parameter of this characteristic parameter is merged into the clustering cluster corresponding to this characteristic parameter. At this time, there is: After time it decays to: Hypothetical cluster During no new data meeting the criteria is received within the time period, that is, the new data meeting the criteria is data; The updated clustering center and clustering radius of the clustering cluster corresponding to each characteristic parameter are: Remove the data that does not belong to the range of the clustering center point and target clustering radius in the clustering cluster, and obtain the target clustering cluster of each characteristic parameter; Increase the number of added nodes in the enhancement layer of the fault diagnosis model according to the preset number of enhancement nodes, and obtain the fault diagnosis model after adding the enhancement nodes. Obtain the loss function value of the enhanced fault diagnosis model according to the target data set. When the loss function value does not meet the preset loss function threshold, then increase the preset number of enhancement nodes again until the loss function value of the enhanced fault diagnosis model meets the preset loss function threshold, and use the enhanced fault diagnosis model at this time as the target fault diagnosis model; Input the input data at the current moment into the target fault diagnosis model, and output the fault diagnosis result of the electromechanical equipment.

2. The incremental fuzzy width learning electromechanical equipment fault diagnosis method based on streaming data according to claim 1, wherein Perform data cleaning and data standardization on the characteristic parameters in sequence, and use the standardized characteristic parameters as input data.

3. The incremental fuzzy width learning electromechanical equipment fault diagnosis method based on streaming data according to claim 1, wherein The steps for data cleaning of the characteristic parameters are: Traverse the collected characteristic parameters, mark the missing values and abnormal values, among which, the determination of abnormal values uses the 6σ criterion; Replace the missing values and abnormal values using the adjacent value filling method to complete the data cleaning.

4. A method for incremental fuzzy width learning electromechanical equipment fault diagnosis based on streaming data according to claim 1, characterized in that, The steps for data standardization of the characteristic parameters are: Use the Z-score standardization method to perform data standardization processing on the data set.

5. A method for diagnosing faults of electromechanical equipment based on incremental fuzzy width learning of streaming data according to claim 1, characterized in that, The mapping layer includes: a fuzzy mapping layer and a defuzzification mapping layer; the fuzzy mapping layer is used to perform fuzzy mapping on the input data and output data in the data set, and the defuzzification mapping layer is used to defuzzify the fuzzy output; The model expression of the fuzzy mapping layer is: wherein, represents the output value of the th characteristic parameter under the kth rule of the ith fuzzy subsystem; represents the normalized weight of the th characteristic parameter under the kth rule of the ith fuzzy subsystem; represents the weighted summation function; represents the th mth parameter among the th characteristic parameters; represents the fuzzy coefficient of the th tth parameter among the th characteristic parameters under the kth rule of the ith fuzzy subsystem; represents the th tth parameter among the th characteristic parameters; represents the output when the th characteristic parameters are input into the ith fuzzy subsystem; represents the output of the th ith fuzzy subsystem among th fuzzy subsystems; represents the output of the th th fuzzy subsystem when the fuzzy mapping layer is The expression of the defuzzification mapping layer model is as follows: ,......, In the formula, represents the defuzzified output when the fuzzy mapping layer is fuzzy subsystems; represents the defuzzified output of the i-th fuzzy subsystem among the fuzzy subsystems; represents the defuzzified output of the -th fuzzy subsystem; represents the defuzzified output of the i-th fuzzy subsystem corresponding to the -th characteristic parameter; is the mapping coefficient of the s-th characteristic parameter under the k-th rule of the i-th fuzzy subsystem; represents the output of the fuzzy subsystem in the mapping layer, ; represents the weight of the i-th fuzzy subsystem, .

6. The incremental fuzzy width learning electromechanical equipment fault diagnosis method based on flow data according to claim 5, characterized in that The expression of the enhancement layer is as follows: Wherein, represents the output when the enhancement layer has m enhancement nodes; represents the th enhancement node output of the enhancement layer; represents the th enhancement node output of the enhancement layer; represents the output when the fuzzy mapping layer has fuzzy subsystems; represents the random weight of the jth enhancement node in the enhancement layer; represents the random drift of the jth enhancement node in the enhancement layer; represents the random mapping value of the jth enhancement node in the enhancement layer.

7. A method for diagnosing faults of electromechanical equipment based on incremental fuzzy width learning of streaming data according to claim 6, characterized in that The expression of the fault diagnosis model is as follows: Let , then: Wherein, represents the output of the fault diagnosis model; represents the weight of the fault diagnosis model; represents the weight of the enhancement layer; represents the output when the enhancement layer has m enhancement nodes; represents defining the value before as the symbol for the value after; represents the weight of the mapping layer; represents the rule output of the fuzzy subsystem of the mapping layer; represents the weight of the i-th fuzzy subsystem of the mapping layer; represents the mapping coefficient of the i-th fuzzy subsystem of the mapping layer.

8. A method for diagnosing faults of electromechanical equipment based on incremental fuzzy width learning of streaming data according to claim 7, characterized in that The expression of the fault diagnosis model after adding enhancement nodes each time is as follows: In the formula, represents the fault diagnosis model after the first addition of enhancement nodes; represents the network structure of the fault diagnosis model before adding enhancement nodes; represents the network structure of the fault diagnosis model after the first addition of enhancement nodes; represents the target weight of the fault diagnosis model after the first addition of enhancement nodes; represents the weight of the fault diagnosis model before adding enhancement nodes; represents the "pseudo-inverse" of the network structure matrix of the network structure of the fault diagnosis model before enhancement nodes; represents the first intermediate quantity obtained by taking the "pseudo-inverse" of the network structure matrix of the network structures of the fault diagnosis models before and after enhancement nodes; represents the second intermediate quantity obtained by taking the "pseudo-inverse" of the network structure matrix of the network structures of the fault diagnosis models before and after enhancement nodes.

Images