CN-122019218-A - Fault diagnosis method based on semi-supervised random configuration network

CN122019218ACN 122019218 ACN122019218 ACN 122019218ACN-122019218-A

Abstract

The invention relates to the technical field of fault diagnosis, in particular to a fault diagnosis method based on a semi-supervised random configuration network. The method comprises the following steps of S1, collecting vibration signals of key components of industrial equipment, S2, preprocessing data, S3, initializing a semi-supervised random configuration network, S4, configuring weights and biases of random candidate nodes, assembling a candidate node pool, S5, selecting optimal nodes from the candidate node pool, S6, adding the optimal nodes into a network model, S7, initializing an unlabeled sample indication matrix, S8, calculating an output weight matrix, and iteratively optimizing an index matrix and an output weight matrix of the unlabeled sample by adopting a joint optimization method, and S9, applying the method to industrial fault diagnosis, thereby having a good fault identification effect.

Inventors

Yuan Panliang
LI SHAOBO
ZHANG CHENGLONG
LIAO ZIHAO

Assignees

贵州大学

Dates

Publication Date: 20260512
Application Date: 20251127

Claims (6)

1. A fault diagnosis method based on a semi-supervised random configuration network is characterized in that the method carries out rapid and efficient modeling through the semi-supervised random configuration network, firstly, pre-processes collected data, initializes network parameters, then utilizes a supervision mechanism to construct a candidate neural node pool, and selects an optimal candidate node as a hidden layer node; The method comprises the following steps: S1, aiming at key parts of industrial equipment, adopting professional equipment to acquire vibration signals; S2, data preprocessing, namely analyzing the acquired signals and extracting the characteristics, normalizing the data to eliminate the influence of the dimension, and thus improving the training effect of the model; S3, initializing a semi-supervised random configuration network, namely dividing unlabeled samples, labeled samples and test set division, setting Laplacian matrix parameters, constructing a Laplacian matrix, and setting the optimal hidden layer node number Critical tolerance error Number of candidate nodes Setting random parameter range And ; S4, randomly configuring the weight and bias of the nodes, and constructing hidden layer candidate neural nodes; S5, selecting candidate nodes by using inequality constraint conditions; S6, determining an optimal neural node, and adding the optimal neural node into the network model; s7, initializing a label-free sample indication matrix; S8, calculating an output weight matrix, and iteratively optimizing an index matrix and an output weight matrix of the unlabeled sample by adopting a joint optimization method; S9, performing fault diagnosis on the test sample; s3, giving a group of preprocessed data, inputting unlabeled samples, labeled samples and test sets, initializing a semi-supervised random configuration network Input data set Wherein As a feature of the sample, In order to be a sample tag, Wherein For the sample feature to be labeled, In order to mark the number of tags, In the number of unlabeled exemplars, , The label being a label of X L , Each element in (a) is equal to Therefore, it is Initializing Laplace matrix parameters and optimizing hidden layer node number in unknown training process Critical tolerance error Number of candidate nodes Setting up Is selected from the range of The selection range of r is Regularization coefficient And The selection range is as follows: the optimization objective of the semi-supervised random configuration network is as follows: where h is the activation function and where, For the output of the kth hidden node, In order to be a laplace matrix, The calculation mode of (2) is as follows: Wherein the method comprises the steps of Is a set of k fields of x, Average edge length for a local adjacency graph 。
2. The method for diagnosing a failure based on a semi-supervised random access network as recited in claim 1, wherein S4 comprises the steps of In which the output weight is randomly selected Bias and method of making same From non-negative real sequences Selecting r, bringing into an activation function Obtaining Group candidate neural node 。
3. The method for diagnosing faults based on the semi-supervised random configuration network of claim 1, wherein S5 specifically comprises the steps that the semi-supervised random configuration network is a single hidden layer network, a network model is built by adding nodes one by one, and the K neural node newly added to the semi-supervised random configuration network is required to meet the following inequality constraint on the assumption that K-1 hidden layer nodes are built: Wherein the method comprises the steps of Network residuals of model K-1 output node in the q dimension, i.e In column q of (2), in which The constraint is a supervision mechanism of the semi-supervised random configuration network, and is used for ensuring that the weight and the bias of the newly added hidden node enable the semi-supervised random configuration network to have a higher convergence rate, and the nodes meeting the inequality constraint are added into the node candidate pool.
4. The method for diagnosing a fault based on the semi-supervised random access network as recited in claim 1, wherein S6 comprises defining a set of variables: Computing nodes in a candidate pool The neural node whose maximum value is selected as the best node.
5. The method for diagnosing faults based on the semi-supervised random access network as recited in claim 1, wherein S7 comprises initializing the settings ; S8, specifically, calculating an output weight according to the following mode: For unlabeled indication matrix The following is performed for each column of (a): Wherein, the Is the ith row and jth column of the unlabeled matrix, , Is that Is selected from the group consisting of the (j) th column, The sum of each row of (2) is 1 The solution can be performed as follows: Because of Therefore, it is Is a convex function and can be solved by Newton's method The expression is as follows: c is the variable quantity, and C and ft control iteration optimization loop; And S8, repeating until the constraint condition is met.
6. A fault diagnosis system of a semi-supervised random configuration network applying the fault diagnosis method as claimed in any one of claims 1 to 5, comprising: The signal acquisition module is used for data acquisition; The data preprocessing module is used for signal analysis, and characteristic extraction and standardization of data; The initialization module is used for initializing semi-supervised random configuration network parameters; the random configuration module is used for configuring node input weights and biases; the hidden layer node selecting module is used for screening nodes and selecting the optimal node; And the output weight joint optimization module is used for calculating the network output weight and carrying out joint iterative optimization with the unlabeled sample indication matrix.

Description

Fault diagnosis method based on semi-supervised random configuration network Technical Field The invention relates to the technical field of fault diagnosis, in particular to a fault diagnosis method based on a semi-supervised random configuration network. Background Fault diagnosis is an important technology for large-scale industrial maintenance, and with the popularization of automation technology and the wide application of sensor technology, industrial equipment can collect a large amount of operation data, and by using the data, fault diagnosis plays a vital role in system safety and process reliability. At present, a method based on rules, models and statistics is often adopted to diagnose industrial faults, and the method has better performance in application of certain scenes, however, in some complex scenes, all three methods are limited. Wherein, rule-based fault diagnosis methods have difficulty in mining deep correlations between fault categories and collected data through expert experience and knowledge. The method based on the model is difficult to establish a complex physical model, and the method based on statistics is difficult to capture the nonlinear coupling relation between the fault type and the collected data and has limited universal performance, so that the three methods have low fault diagnosis precision in complex industrial scenes. In recent years, deep learning has been paid more attention to the field of intelligent manufacturing, and through a multi-layer neural network, a deep learning method has strong feature learning capability and excellent generalization performance towards high-dimensional nonlinear data, however, the deep learning method has limitations in that a large amount of labeling data is required for training, and in real industrial fault diagnosis, the labeling cost of the data is high, so that the labeling data is less. Furthermore, deep learning methods require a significant amount of hardware resources during training and reasoning, which limits their application in resource-constrained environments. For example, CNN and transform methods require multiple convolution and pooling layers, their network structure is complex and require layer-by-layer solution of gradients by back propagation, which makes training CNNs computationally expensive and computationally long. This computational bottleneck is particularly prominent in industrial application scenarios, exposing the inherent limitations of traditional depth architecture. In order to overcome the problems, the academy is developing a lightweight network structure with fast training and few parameters to apply semi-supervised industrial fault diagnosis. Disclosure of Invention The technical problem solved by the invention is to provide a fault diagnosis method based on a semi-supervised random configuration network, and a lightweight network structure with fast training and few parameters. The invention provides a basic scheme that a fault diagnosis method based on a semi-supervised random configuration network comprises the following steps: The method comprises the following steps: S1, aiming at key parts of industrial equipment, adopting professional equipment to acquire vibration signals; S2, data preprocessing, namely analyzing the acquired signals and extracting the characteristics, normalizing the data to eliminate the influence of the dimension, and thus improving the training effect of the model; S3, initializing a semi-supervised random configuration network, namely dividing unlabeled samples, labeled samples and test set division, setting Laplacian matrix parameters, constructing a Laplacian matrix, and setting the optimal hidden layer node number Critical tolerance errorNumber of candidate nodesSetting random parameter rangeAnd; S4, randomly configuring the weight and bias of the nodes, and constructing hidden layer candidate neural nodes; S5, selecting candidate nodes by using inequality constraint conditions; S6, determining an optimal neural node, and adding the optimal neural node into the network model; s7, initializing a label-free sample indication matrix; S8, calculating an output weight matrix, and iteratively optimizing an index matrix and an output weight matrix of the unlabeled sample by adopting a joint optimization method; s3, giving a group of preprocessed data, inputting unlabeled samples, labeled samples and test sets, initializing a semi-supervised random configuration network Input data setWhereinAs a feature of the sample,In order to be a sample tag,WhereinFor the sample feature to be labeled,In order to mark the number of tags,In the number of unlabeled exemplars,,The label being a label of X L,Each element in (a) is equal toTherefore, it isInitializing Laplace matrix parameters and optimizing hidden layer node number in unknown training processCritical tolerance errorNumber of candidate nodesSetting upIs selected from the range ofThe selection range of r isRegularization