Incremental learning-oriented neural network fault diagnosis model training method and system

By adopting a semi-supervised learning framework and Tulaplas matrix in armored vehicle fault diagnosis, an incremental learning method for armored vehicle fault diagnosis model is solved, and the problem of waste of resources and insufficient utilization of labelless samples in the existing technology is solved. , realizes efficient fault diagnosis and cost savings.

CN119046739APending Publication Date: 2024-11-29XI AN JIAOTONG UNIV +1
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411286328.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The existing armored vehicle fault diagnosis model needs to be retrained when facing new fault samples, resulting in wasted computing resources and storage space. Traditional supervised learning cannot utilize label-free samples, resulting in high labor and time costs and limited application scope.

Method used

Using the neural network fault diagnosis method for incremental learning, the relationship between labeled samples and unlabeled samples is constructed through the semi-supervised learning framework, the connection between labeled samples and unlabeled samples is established, and the model is expanded. Quantity training, avoid retraining the model, and save computing resources and storage space.

Benefits of technology

It realizes incremental learning of new fault samples based on the old model, saves manpower and time costs, makes full use of labelless samples, improves the efficiency and accuracy of fault diagnosis, and reduces dependence on sample labeling.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046739A_ABST
    Figure CN119046739A_ABST
Patent Text Reader

Abstract

The invention discloses an incremental learning-oriented neural network fault diagnosis model training method and system, and the process of the diagnosis method specifically comprises the steps: carrying out the two times of random mapping of a diagnosis sample, obtaining a feature matrix of the diagnosis sample, selecting an output weight matrix or an increment form of the output weight matrix according to the state of a current fault diagnosis model, and carrying out the training of a fault diagnosis model. Obtaining a prediction label of the diagnosis sample, and realizing fault diagnosis; the method comprises the following steps: introducing a manifold regularization framework, constructing a prediction error expression of labeled samples based on a smoothness hypothesis and a manifold hypothesis, calculating the similarity between every two samples to construct a Laplacian matrix, and establishing a relation between the labeled samples and unlabeled samples; a semi-supervised width learning network fault diagnosis model is constructed, a large amount of manpower and time consumed for accurate state marking of samples are saved, when new samples are collected, incremental learning can be only performed on the new samples to complete updating on the basis of an old fault diagnosis model, and the model does not need to be retrained.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of armored vehicle status monitoring and health management, and in particular to a neural network fault diagnosis model training method and system oriented to incremental learning. Background Art

[0002] At present, the technical level of armored vehicle status monitoring and health management is still at the level of basic functional parameters, mainly conducting online testing of signals such as current, voltage, and temperature and simple threshold judgment. The characterization and judgment of the operation status of armored vehicles are very limited. Abnormal changes in status parameters are often only discovered in the middle and late stages of armored vehicle failures, which can easily cause irreparable and significant losses. Therefore, it is of great significance to accurately and timely discover mechanical equipment failures and take "preventive measures" to ensure operational safety, reduce major accidents, and improve economic benefits.

[0003] Since armored vehicles have many monitoring parameters, high sampling frequency, and a long monitoring process, the "massive" monitoring data poses new challenges to their status monitoring and health management. During the operation of armored vehicles, monitoring parameters are continuously collected as the operation progresses, and the training of existing fault diagnosis models is mostly "once and for all". When new fault samples are collected, the performance of the original model does not meet the requirements, and the model needs to be retrained, which not only consumes computing resources and is inefficient, but also places high demands on sample storage devices.

[0004] In addition, in engineering practice, it is easy to collect a large number of unlabeled samples, and accurate status labeling of all collected samples requires a lot of manpower and time costs. Therefore, the labeled samples obtained are very limited. The traditional supervised learning fault diagnosis model cannot utilize the information of unlabeled samples and can only analyze a small number of expert-labeled samples, which limits its application scope.

[0005] In response to these problems, scholars have conducted relevant research. Sun Ning and others from Hohai University invented an industrial equipment fault diagnosis system based on semi-supervised incremental learning. The semi-supervised labeling module is used to label the unlabeled samples with pseudo labels. The obtained pseudo-label samples will be used to incrementally update the fault diagnosis module. Xiao Zhiqing and others from Fudan University invented a small sample data fault diagnosis method and device based on the transfer incremental learning framework. The benchmark model is trained on the data set of similar faults, and the model weights are incrementally updated with the small sample data set of the same fault. The updated model is used for target fault diagnosis. The above content does not involve the learning of new fault samples. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a neural network fault diagnosis method for incremental learning in response to the shortcomings of the above-mentioned existing fault diagnosis models. New knowledge is incrementally learned from new fault samples based on the original model, and a semi-supervised learning framework is adopted to simultaneously utilize the connection between labeled samples and unlabeled samples to avoid the waste of computing resources and storage space caused by retraining the model, as well as the manpower and time costs of labeling a large number of samples.

[0007] In order to achieve the above object, the present invention adopts the following technical solution: a neural network fault diagnosis model training method for incremental learning, comprising the following steps:

[0008] The initial training samples are randomly mapped twice to obtain the feature matrix;

[0009] The popular regularization framework is introduced. Based on the smoothness hypothesis and the manifold hypothesis, the prediction error of the labeled samples is constructed. The similarity between samples is calculated pairwise to construct the graph Laplacian matrix, and the connection between the labeled samples and the unlabeled samples is established.

[0010] A semi-supervised width learning network fault diagnosis model is constructed. The prediction error of the labeled samples and the graph Laplacian matrix are used to train the fault diagnosis model for the first time and solve the output weight matrix of the model.

[0011] According to whether the incremental learning samples contain labeled samples, the incremental training of the semi-supervised width learning network fault diagnosis model is completed in two categories, that is, the incremental form of the output weight matrix of the model is obtained by updating the original model.

[0012] Furthermore, the initial training samples are randomly mapped twice to obtain the feature matrix including:

[0013] The initial training sample X is transformed into a mapping feature z with n groups of k mapping nodes in each group through randomly generated mapping nodes. i :

[0014]

[0015] In the formula, and Respectively represent the weight matrix and bias of the i-th group of mapping nodes,

[0016] Let the mapping feature z i The merger of n :

[0017] Z n =[z 1 ,z 2 ,...,z n ]

[0018] Mapping feature Z n Through the strengthening nodes, it is further transformed into m strengthening features h j :

[0019]

[0020] In the formula, ζ(·)=1 / (1+e -(·) ), represents the nonlinear activation function; and Respectively represent the weight matrix and bias of the jth reinforcement node, and the reinforcement feature h j The merger of H m :

[0021] H m =[h 1 ,h 2 ,...,h m ]

[0022] Merge all the mapping features and enhanced features to get the feature matrix A:

[0023] A=[Z n |H m ]

[0024] The feature matrix A is transformed into the following form:

[0025] A=[a(x 1 ) T ,...,a(x l ) T ,...,a(x l+u ) T ] T .

[0026] Furthermore, the prediction error of constructing the labeled sample includes:

[0027] Assume that the output weight matrix of the semi-supervised wide neural network is β, then the predicted label of the semi-supervised wide neural network for the labeled sample is for:

[0028]

[0029] In the formula, l is the number of labeled samples in the initial training sample X,

[0030] The network predicts the label of the labeled sample and the true label y of the labeled sample i Do the difference and get the prediction error of the labeled sample:

[0031]

[0032] In the formula, Ci is the penalty coefficient, C i =C 0 / N ti , N ti is the sample x i The number of samples in the corresponding category, C 0 It is a parameter set by humans.

[0033] Furthermore, the similarity between samples is calculated pairwise to construct the graph Laplacian matrix including:

[0034] Minimize the following loss function so that the initial training samples and their labels satisfy the smoothness assumption and manifold assumption:

[0035]

[0036] Where u is the number of unlabeled samples in the initial training sample X; ω ij is the sample x i and x j Pairwise similarity of P(y|x i ) and P(y|x j ) is the sample x i and x j The corresponding conditional probability, for the loss function Approximate and express in matrix form, and get the graph Laplacian matrix L:

[0037] L=DW

[0038] In the formula, σ is the parameter of the Gaussian kernel function, N k (x) represents the first k nearest neighbor samples of sample x.

[0039] Furthermore, the prediction error of the labeled samples and the graph Laplace matrix are used to initially train the fault diagnosis model, and the output weight matrix of the fault diagnosis model is solved, including:

[0040] The objective function for solving the output weight matrix is:

[0041]

[0042] Where L is the graph Laplacian matrix derived from the initial training sample X, is the predicted label of the network's initial training sample, λ is the balance parameter, and the matrix form of the objective function is:

[0043]

[0044] Y is the true label of the initial training sample X; C is a diagonal matrix, C ii =[C i ],

[0045] Finding the gradient of β in the matrix form of the objective function gives:

[0046]

[0047] Furthermore, the incremental form of the output weight matrix of the fault diagnosis model is specifically:

[0048] The subscript t indicates the data state of the sample that the network has learned at time t, the subscript t+1 indicates the data state of the sample that the network has learned at time t+1, and the subscript q indicates the data state of the new sample that needs to be incrementally learned.

[0049] The q new samples are all unlabeled samples, and the output weight matrix at time t+1 is for:

[0050]

[0051] In the formula, I nk+m is the identity matrix of nk+m dimensions; λ is the balance parameter; and At time t, A t With C t The matrix composed of non-zero rows, L t is the graph Laplacian matrix at time t; A q is the feature matrix of q new samples to be incrementally learned; I q is the q-dimensional identity matrix; L q is the graph Laplacian matrix of q new samples to be incrementally learned,

[0052] The q new samples contain q l labeled samples, the output weight matrix at time t+1 for:

[0053]

[0054] In the formula, Yes l The feature matrix corresponding to the labeled samples, I q is the q-dimensional identity matrix, It is C q The matrix consists of non-zero rows.

[0055] The present invention also provides a neural network fault diagnosis model training system for incremental learning, including a feature extraction module, a network training module and an incremental learning module;

[0056] Feature extraction module: transforms the input sample into a series of mapping features through randomly generated mapping nodes, expands the mapping features into enhanced features through random enhancement nodes, merges the mapping features with the enhanced features, and obtains the feature matrix of the input sample;

[0057] Network training module: introduce the popular regularization framework, construct the prediction error expression of the labeled samples based on the smoothness hypothesis and the manifold hypothesis, calculate the similarity between samples pairwise to construct the graph Laplace matrix, and establish the connection between the labeled samples and the unlabeled samples; construct a fault diagnosis model based on the semi-supervised width learning network, use the prediction error of the labeled samples and the graph Laplace matrix to perform initial training on the fault diagnosis model, and solve the output weight matrix of the fault diagnosis model;

[0058] Incremental learning module: According to whether the incremental learning samples contain labeled samples, the incremental training process of the fault diagnosis model based on the semi-supervised width learning network is completed in two categories to obtain the incremental form of the output weight matrix of the model.

[0059] Based on the obtained model, the present invention provides a neural network fault diagnosis method for incremental learning, and the diagnosis process is specifically as follows:

[0060] The characteristic matrix of the diagnosis sample is obtained by two random mappings. According to the state of the current fault diagnosis model, the output weight matrix or its incremental form is selected to obtain the predicted label of the diagnosis sample and realize fault diagnosis.

[0061] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the neural network fault diagnosis model training method for incremental learning described in the present invention.

[0062] At the same time, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, the neural network fault diagnosis model training method for incremental learning according to the present invention can be implemented.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] The present invention provides a neural network fault diagnosis model for incremental learning. By applying a manifold regularization framework to a width learning network, a connection between labeled samples and unlabeled samples is established, and unlabeled samples are fully utilized, thereby saving a large amount of manpower and time costs for accurately labeling the states of samples. In addition, when new samples are collected, a new fault diagnosis model can be updated based on the old fault diagnosis model by incrementally learning only the new samples without retraining the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is the overall flow chart of the present invention.

[0066] Figure 2 Flowchart for initial training of fault diagnosis model.

[0067] Figure 3 It is a schematic diagram of a specific process of an embodiment. DETAILED DESCRIPTION

[0068] 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 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.

[0069] In the description of the present invention, it should be understood that the terms “include” and “comprises” indicate the presence of described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0070] refer to Figure 1 , Figure 2 and Figure 3The training method of the neural network fault diagnosis model for incremental learning provided by the present invention comprises the following steps: obtaining a feature matrix by performing two random mappings on the initial training samples (including labeled samples and unlabeled samples); introducing a popular regularization framework, constructing a prediction error expression of labeled samples based on the smoothness assumption and the manifold assumption, calculating the similarity between samples pairwise to construct a graph Laplace matrix, and establishing a connection between labeled samples and unlabeled samples; constructing a semi-supervised width learning network fault diagnosis model, using the prediction error of labeled samples and the graph Laplace matrix to perform initial training on the fault diagnosis model, and solving the output weight matrix of the model; when incremental training samples (also including labeled samples and unlabeled samples) arrive, the incremental training process of the semi-supervised width learning network fault diagnosis model is completed in two categories according to whether the incremental learning samples contain labeled samples, and the incremental form of the output weight matrix of the model is obtained.

[0071] The initial training samples are randomly mapped twice. The original high-dimensional data is mapped to two lower-dimensional spaces through random projection or transformation methods, which can reduce the dimension of the data while retaining the important information of the original data as much as possible. After two random mappings, two feature matrices can be obtained, corresponding to the results of the two mappings, which will serve as the basis for subsequent analysis; popular regularization is used to enhance the generalization ability of the model. By adding a regularization term related to the popular structure of the data in the objective function, the complexity of the model is constrained to avoid overfitting; the smoothness hypothesis is that the class labels of two samples that are very close to each other in the dense data area are similar. The smoothness hypothesis emphasizes the local smoothness between data points, that is, similar data points should have similar labels; the manifold hypothesis refers to embedding high-dimensional data into a low-dimensional manifold. When two samples are located in a small local neighborhood in the low-dimensional manifold, they have similar class labels. For labeled samples, the loss function can be used to measure the difference between the model prediction value and the actual label, thereby constructing a prediction error expression. In the prediction error expression, the popular regularization term is introduced to reflect the popular structure information of the data point.

[0072] The diagnostic process of the neural network fault diagnosis model for incremental learning is to cut the collected physical data to be diagnosed into samples, obtain the feature matrix through the same two random mappings, superimpose the output weight matrix corresponding to the current model, and obtain the predicted label of the diagnostic sample. The predicted label represents the state of the armored vehicle, and the fault type is determined according to the corresponding relationship between the label and the armored vehicle fault type.

[0073] The self-learning process of the mechanical equipment health management model for incremental learning provided by the present invention includes the following steps:

[0074] Specifically, the process of obtaining the feature matrix of the initial training sample includes:

[0075] The time series vibration signal collected by the vibration acceleration sensor is cut according to the number of data points contained in each sample to obtain the initial training sample X, which can be divided into two parts: labeled samples and their labels Unlabeled samples Among them, l and u are the number of labeled samples and the number of unlabeled samples respectively;

[0076] Assume that the number of groups of mapping nodes is n, each group includes k mapping nodes, and the number of reinforcement nodes is m. First, the initial training sample X is transformed into a series of mapping features z through the mapping nodes. i :

[0077]

[0078] in, and They represent the weight matrix and bias of the i-th group of mapping nodes respectively, and the merging of mapping features is recorded as Z n :

[0079] Z n =[z 1 ,z 2 ,...,z n ]

[0080] Where Z n ∈R (l+u)*nk .

[0081] Then, map feature Z n Through the strengthening nodes, it is further transformed into a series of strengthening features h j :

[0082]

[0083] where ζ(·)=1 / (1+e -(·) ), represents the nonlinear Sigmoid activation function; and They represent the weight matrix and bias of the jth reinforcement node respectively, and the merging of reinforcement features is H m :

[0084] H m =[h 1 ,h 2 ,...,h m ]

[0085] Among them, H m ∈R (l+u)*m , merge all the mapping features and enhanced features to get the feature matrix A:

[0086] A=[Z n |Hm ]

[0087] A can also be expressed as:

[0088] A=[a(x 1 ) T ,...,a(x l ) T ,...,a(x l+u ) T ] T

[0089] where A∈R (l+u)*(nk+m) , l and u are the number of labeled samples and the number of unlabeled samples respectively, and nk+m represents the number of all feature nodes.

[0090] Get the prediction error of the labeled sample, including:

[0091] Assume that the output weight matrix of the semi-supervised width learning network is β, then the predicted label of the semi-supervised width learning network for the labeled sample is for:

[0092]

[0093] From this, the prediction error of the labeled sample can be obtained:

[0094]

[0095] Among them, y i is the true label of the labeled sample, C i is the penalty coefficient. Different penalty coefficients are associated with prediction errors from samples of different categories because, in labeled samples, the sample size of some classes will far exceed that of other classes. Traditional width learning networks tend to better fit classes with more sample sizes during training. For example, in the diagnosis of normal and faulty classes, if the samples contained in the normal class far exceed those in the faulty class, the model will selectively learn the features of normal class samples and ignore the faulty class. Therefore, classes with small sample sizes will be ignored, while classes with large sample sizes will be overfitted, which usually results in poor generalization ability and poor accuracy performance of the width learning network in the test samples. Therefore, it is necessary to perform different degrees of penalty processing on samples of different categories to solve this problem:

[0096] C i =C 0 / N ti

[0097] Where N ti is the sample x i The number of samples in the corresponding category, C 0 is a hyperparameter.

[0098] Specifically, the corresponding Laplacian matrix is ​​constructed according to the sample similarity, including:

[0099] According to the smoothness assumption and the manifold assumption: (1) If two input samples in a high-density region are similar, then the output values ​​corresponding to the two input samples in the high-density region are also similar; (2) Assuming that the samples in the high-dimensional space are embedded in a low-dimensional submanifold, if two points are similar, then the output values ​​corresponding to the two points are also similar. In order to make the input samples meet these two assumptions, the manifold regularization framework needs to minimize the following loss function:

[0100]

[0101] Among them, ω ij is the sample x i and x j The pairwise similarity of P(y|x i ) and P(y|x j ) is the sample x i and x j The corresponding conditional probability. Since the conditional probability is difficult to calculate, the following expression is used to approximate the above loss function:

[0102]

[0103] in and They are respectively the network for sample x i and x j The predicted label of the sample x i and x j The pairwise similarity ω ij It can be calculated by the following formula:

[0104]

[0105] Among them, σ is the parameter of Gaussian kernel function; N k (x) represents the first k nearest neighbor samples of sample x. The pairwise similarity of all samples ω ij The similarity matrix W can be constructed, that is, W = [ω ij ], summing each column of the similarity matrix W can get a diagonal matrix D:

[0106]

[0107] Thus, we can get the graph Laplacian matrix L:

[0108] L=DW

[0109] Where W is the similarity matrix, and D is the diagonal matrix obtained based on the similarity matrix to measure the pairwise similarity of samples. L can be normalized by the following formula:

[0110]

[0111] Therefore, the matrix form of the above formula is:

[0112]

[0113] The objective function is:

[0114]

[0115] Where L∈R (l+u)×(l+u) is the graph Laplacian matrix derived from all input samples, is the network's predicted label for all input samples, and λ is a hyperparameter.

[0116] The matrix form of the objective function is:

[0117]

[0118] is the true label of the input sample. The first l rows are the true labels of the labeled samples, and the rest are 0. C is a (l+u)×(l+u) diagonal matrix. The first l diagonal elements C ii =[C i ] and the rest are 0.

[0119] Find the gradient of β for the matrix form of the objective function,

[0120] When the number of input samples is greater than the sum of the number of mapping nodes and enhancement nodes:

[0121] β * =(I nk+m +A T CA+λA T LA) -1 A T CY

[0122] I nk+m is the identity matrix of nk+m dimensions.

[0123] When the number of input samples is less than the sum of the number of mapping nodes and enhancement nodes:

[0124] β * =A T (I l+u +CAA T +λLAA T ) -1 CY

[0125] In the formula, I l+u is the identity matrix of dimension l+u.

[0126] Specifically, the incremental process includes:

[0127] Assume that at time t, the sample that the model has learned is X t , where the labeled samples are Its sample label is Note X t and The corresponding feature matrices are A t and At time t+1, the new samples that need to be learned can be divided into the following two cases:

[0128] 1. There are q samples, all of which are unlabeled samples.

[0129] Second, q samples, denoted as X q , which contains q l There are labeled samples, denoted as q l The feature matrix and label corresponding to the labeled samples are recorded as and

[0130] Let the feature matrix of the new sample to be learned at time t+1 be A q , so in the first case, it can be obtained that the feature matrix of the new sample at time t+1 and the feature matrix of the initial training sample are The feature matrix of the new sample at time t+1 and the feature matrix of the labeled sample in the initial training sample In the second case,

[0131] In engineering, the number of samples actually participating in network training is usually very large, which is greater than the sum of the number of mapping nodes and enhancement nodes in the width learning network. Therefore, the present invention only analyzes this situation and obtains the following equation:

[0132]

[0133] in It is C t+1 For convenience, we will denote the non-zero rows of t for:

[0134]

[0135] Then, the Laplacian matrix of the graph of the new sample and the initial training sample at time t+1 can be expressed as:

[0136]

[0137] Among them, L t and L q They are the initial training samples X t and the new sample X at time t+1 q The graph Laplacian matrix W. t,q From X t To X q The similarity matrix W q,t From X q To X t The similarity matrix of t,q and D q,t are respectively for S t,q and S q,t The diagonal matrix obtained by summing each column is set as P t+1 for:

[0138]

[0139] Among them B t and E t They are:

[0140]

[0141] Therefore, P t+1 The inverse matrix of is:

[0142]

[0143] in, Get the incremental formula of the network output weight matrix at time t+1:

[0144]

[0145] In the second case, according to the relationship between time t+1 and time t, P t+1 can be written as:

[0146]

[0147] Among them, B t and E t The same as above are:

[0148]

[0149] Therefore, P t+1 The inverse matrix of is:

[0150]

[0151] in,

[0152] Define K t for:

[0153]

[0154] but It can be expressed as:

[0155]

[0156] but It can be expressed as:

[0157]

[0158] Among them, J t for:

[0159]

[0160] At this point, we can get the incremental formula of the network output weight matrix at time t+1 in the second case:

[0161]

[0162] Specifically, the diagnostic process is:

[0163] The sample to be diagnosed is X test The feature matrix A is obtained by the same two random mappings test , superimpose the output weight matrix β corresponding to the current model * (corresponding formula β * =(I nk+m +A T CA+λA T LA) -1 A T CY or

[0164] β * =A T (I l+u +CAA T +λLAA T ) -1 CY) or (Corresponding Or

[0165] The predicted labels of the diagnostic samples can be obtained to achieve fault diagnosis.

[0166] The invention provides a training method for a neural network fault diagnosis model for incremental learning, comprising: obtaining a feature matrix by performing two random mappings on an initial training sample; introducing a popular regularization framework, constructing a prediction error expression of a labeled sample based on a smoothness assumption and a manifold assumption, calculating the similarity between samples in pairs to construct a graph Laplace matrix, and establishing a connection between labeled samples and unlabeled samples; constructing a semi-supervised width learning network fault diagnosis model, performing initial training on the fault diagnosis model using the prediction error of the labeled sample and the graph Laplace matrix, and solving the output weight matrix of the model; when an incremental training sample arrives, completing the incremental training process of the semi-supervised width learning network fault diagnosis model in two categories according to whether the incremental learning sample contains a labeled sample, and obtaining an incremental form of the output weight matrix of the model.

[0167] The samples to be diagnosed are mapped twice to obtain the feature matrix, and the current output weight matrix is ​​superimposed to obtain the predicted label of the diagnosis sample to achieve fault diagnosis.

[0168] On the other hand, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the neural network fault diagnosis model training method for incremental learning described in the present invention can be implemented.

[0169] The computer device may be a laptop computer, a desktop computer or a workstation.

[0170] The present invention can also provide a computer device, including a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the neural network fault diagnosis model training method for incremental learning described in the present invention.

[0171] The processor may be a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), or an off-the-shelf field programmable gate array (FPGA).

[0172] The memory described in the present invention may be an internal storage unit of a notebook computer, a desktop computer or a workstation, such as a memory or a hard disk; or an external storage unit, such as a mobile hard disk or a flash memory card.

[0173] Computer-readable storage media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media may include: read-only memory (ROM), random access memory (RAM), solid-state drive (SSD) or optical disk, etc. Among them, random access memory may include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0174] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A neural network fault diagnosis model training method for incremental learning, It is characterized in that The following steps are involved: The initial training samples are randomly mapped twice to obtain the feature matrix; The popular regularization framework is introduced. Based on the smoothness hypothesis and the manifold hypothesis, the prediction error of the labeled samples is constructed. The similarity between samples is calculated pairwise to construct the graph Laplacian matrix, and the connection between the labeled samples and the unlabeled samples is established. A semi-supervised width learning network fault diagnosis model is constructed. The prediction error of the labeled samples and the graph Laplacian matrix are used to train the fault diagnosis model for the first time and solve the output weight matrix of the model. According to whether the incremental learning samples contain labeled samples, the incremental training of the semi-supervised width learning network fault diagnosis model is completed in two categories, that is, the incremental form of the output weight matrix of the model is obtained by updating the original model.

2. The neural network fault diagnosis model training method for incremental learning according to claim 1, It is characterized in that The initial training samples are randomly mapped twice to obtain the feature matrix including: The initial training sample X is transformed into a mapping feature z with n groups of k mapping nodes in each group through randomly generated mapping nodes. i : In the formula, and Respectively represent the weight matrix and bias of the i-th group of mapping nodes, Let the mapping feature z i The merger of n : WITH n =[of 1 ,With 2 ,...,With n ] Mapping feature Z n Through the strengthening nodes, it is further transformed into m strengthening features h j : In the formula, ζ(·)=1 / (1+e -(·) ), represents the nonlinear activation function; and Respectively represent the weight matrix and bias of the jth reinforcement node, and the reinforcement feature h j The merger of H m : H m =[h 1 ,h 2 ,...,h m ] Merge all the mapping features and enhanced features to get the feature matrix A: A=[Z n |H m ] The feature matrix A is transformed into the following form: A=[a(x 1 ) T ,...,a(x l ) T ,...,a(x l+u ) T ] T 。 3. The neural network fault diagnosis model training method for incremental learning according to claim 1, It is characterized in that The prediction error of constructing labeled samples includes: Assume that the output weight matrix of the semi-supervised wide neural network is β, then the predicted label of the semi-supervised wide neural network for the labeled sample is for: In the formula, l is the number of labeled samples in the initial training sample X, The network predicts the label of the labeled sample and the true label y of the labeled sample i Do the difference and get the prediction error of the labeled sample: In the formula, C i is the penalty coefficient, C i =C 0 / N ti , N ti is the sample x i The number of samples in the corresponding category, C 0 It is a parameter set by humans.

4. The neural network fault diagnosis model training method for incremental learning according to claim 1, It is characterized in that The similarity between two samples is calculated to construct the Laplacian matrix, which includes: Minimize the following loss function so that the initial training samples and their labels satisfy the smoothness assumption and manifold assumption: Where u is the number of unlabeled samples in the initial training sample X; ω ij is the sample x i and x j Pairwise similarity of P(y|x i ) and P(y|x j ) is the sample x i and x j The corresponding conditional probability, for the loss function Approximate and express in matrix form, and get the graph Laplacian matrix L: L=DW is the parameter of the Gaussian kernel function, N k (x) represents the first k nearest neighbor samples of sample x.

5. The neural network fault diagnosis model training method for incremental learning according to claim 1, It is characterized in that The prediction error of the labeled samples and the graph Laplace matrix are used to initially train the fault diagnosis model, and the output weight matrix of the fault diagnosis model is solved, including: The objective function for solving the output weight matrix is: Where L is the graph Laplacian matrix derived from the initial training sample X, is the predicted label of the network's initial training sample, λ is the balance parameter, and the matrix form of the objective function is: Y is the true label of the initial training sample X; C is a diagonal matrix, C ii =[C i ], Finding the gradient of β in the matrix form of the objective function gives:

6. The neural network fault diagnosis model training method for incremental learning according to claim 1, It is characterized in that The incremental form of the output weight matrix of the fault diagnosis model is as follows: The subscript t indicates the data state of the sample that the network has learned at time t, the subscript t+1 indicates the data state of the sample that the network has learned at time t+1, and the subscript q indicates the data state of the new sample that needs to be incrementally learned. The q new samples are all unlabeled samples, and the output weight matrix at time t+1 is for: In the formula, I nk+m is the identity matrix of nk+m dimensions; λ is the balance parameter; and At time t, A t With C t The matrix composed of non-zero rows, L t is the graph Laplacian matrix at time t; A q is the feature matrix of q new samples to be incrementally learned; I q is the q-dimensional identity matrix; L q is the graph Laplacian matrix of q new samples to be incrementally learned, The q new samples contain q l labeled samples, the output weight matrix at time t+1 for: In the formula, Yes l The feature matrix corresponding to the labeled samples, I q is the q-dimensional identity matrix, It is C q The matrix consists of non-zero rows.

7. A neural network fault diagnosis model training system for incremental learning, It is characterized in that include: Feature extraction module: transforms the input sample into a series of mapping features through randomly generated mapping nodes, expands the mapping features into enhanced features through random enhancement nodes, merges the mapping features with the enhanced features, and obtains the feature matrix of the input sample; Network training module: introduce the popular regularization framework, construct the prediction error expression of the labeled samples based on the smoothness hypothesis and the manifold hypothesis, calculate the similarity between samples pairwise to construct the graph Laplace matrix, and establish the connection between the labeled samples and the unlabeled samples; construct a fault diagnosis model based on the semi-supervised width learning network, use the prediction error of the labeled samples and the graph Laplace matrix to perform initial training on the fault diagnosis model, and solve the output weight matrix of the fault diagnosis model; Incremental learning module: According to whether the incremental learning samples contain labeled samples, the incremental training process of the fault diagnosis model based on the semi-supervised width learning network is completed in two categories to obtain the incremental form of the output weight matrix of the model.

8. A neural network fault diagnosis method for incremental learning, It is characterized in that The diagnostic process is as follows: The characteristic matrix of the diagnosis sample is obtained by two random mappings. According to the state of the current fault diagnosis model, the output weight matrix or its incremental form is selected to obtain the predicted label of the diagnosis sample and realize fault diagnosis.

9. A computer device, It is characterized in that It includes a processor and a memory, the memory is used to store a computer executable program, the processor reads the computer executable program from the memory and executes it, and when the processor executes the computer executable program, it can implement the neural network fault diagnosis model training method for incremental learning as described in any one of claims 1 to 6.

10. A computer-readable storage medium, It is characterized in that A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, the method for training a neural network fault diagnosis model for incremental learning according to any one of claims 1 to 6 can be implemented.

Citation Information

Cited By

  • Image classification method and system based on online semi-supervised incremental learning

    CN119863644A

  • Actually measured gravity and magnetic data denoising method based on priori knowledge deep learning

    CN120045926A

  • Feature weighting-based semi-supervised industrial equipment fault image identification method

    CN121259468A