CN-119004186-B - Mechanical equipment associated measuring point migration learning method based on multi-scale convolution network

CN119004186BCN 119004186 BCN119004186 BCN 119004186BCN-119004186-B

Abstract

The invention discloses a mechanical equipment association measuring point migration learning method based on a multi-scale convolution network, which comprises the steps of firstly using source domain data with labels as input of a domain learning network of a source domain to extract association measuring point characteristics, sharing a network structure and weights to the domain learning network of a target domain, then using target domain data without labels as input of the domain learning network of the target domain to extract target measuring point characteristics, using a maximum mean square difference method to carry out inter-domain difference measurement on the association measuring point characteristics and the target measuring point characteristics, constructing a joint loss function by combining classification loss and inter-domain difference loss of the domain learning network of the source domain as an optimization target, using a back propagation algorithm to carry out iterative training, updating target domain learning model parameters and predicting sample types, finally obtaining a target domain learning model with excellent diagnosis effect, and realizing knowledge migration between the association measuring points and the target measuring points of the mechanical equipment. According to the invention, deep features in data are effectively extracted through the multi-scale convolutional neural network, and knowledge migration among different measuring points is realized by utilizing the maximum mean square difference, so that fault diagnosis is performed by utilizing the associated measuring points under the conditions of inconvenient sensor installation, large data distribution difference and lack of label information, and model generalization is effectively improved.

Inventors

LIU CHANG
WANG XI
HE FEIFEI
Ke Haoyang

Assignees

昆明理工大学

Dates

Publication Date: 20260508
Application Date: 20240908

Claims (4)

1. The mechanical equipment associated measuring point migration learning method based on the multi-scale convolution network is characterized by comprising the following steps of: collecting associated measuring point data of equipment fault points as source domain data according to a migration task, and collecting target point data to be diagnosed as target domain data; Carrying out normalization processing on the source domain data and the target domain data, then carrying out data segmentation processing on the normalized data by adopting a sliding window mechanism to obtain samples, determining class labels corresponding to the samples, and finally obtaining a source domain data set and a target domain data set, wherein the class labels comprise normal class labels and different fault class labels; The method comprises the steps of constructing a domain learning network, combining a shared network structure with a training source domain and a target domain, wherein the input of the domain learning network with a label is used for extracting relevant measuring point characteristics, the network structure and the weight are shared to the domain learning network of the target domain, and then the target domain data without the label is used as the input of the domain learning network of the target domain to extract target measuring point characteristics; The domain learning network comprises the steps of firstly, extracting data features by utilizing a multi-scale convolutional neural network, introducing a self-attention module to enhance the characterization capability of the features, then, fusing feature information under different scales, and finally, introducing a maximum mean square difference method to measure the feature difference between a source domain and a target domain in a self-adaptive layer; the multi-scale convolutional neural network comprises two branches, wherein a first branch adopts five layers of overlapped 'convolutional layers + batch normalization layers + ReLU activation functions + max pooling layers', the convolutional kernel size of the first layer in the first branch is larger than the convolutional kernel sizes of other layers in the first branch; The calculation of the maximum mean square difference is as follows: relevant measuring point characteristic set obtained according to network extraction And a target survey point feature set And parameters of a gaussian kernel function And weight The mixed RBF kernel matrix is calculated, and the RBF kernel matrix has the following calculation formula: ; ; ; Wherein, the , , Is a dot product matrix between the features of the sample, , Respectively representing dot product matrices A vector of diagonal elements; next, three hybrid RBF kernel matrices are based on the hybrid RBF kernel function 、、 The maximum mean square difference is calculated, and the maximum mean square difference calculation is shown as follows: ; Wherein, the Is a matrix Is a trace of (1); 、 Respectively represent 、 Is a number of samples of (a).
2. The mechanical equipment association measuring point migration learning method based on the multi-scale convolution network according to claim 1, wherein the sliding window mechanism method is as follows: ; in the formula, The number of samples obtained by the sliding window is indicated, For the sample length s is the window step size.
3. The mechanical equipment association measuring point migration learning method based on the multi-scale convolution network according to claim 1, wherein the convolution kernel sizes adopted by five convolution layers in the first branch are 64, 3 and 3 respectively, and the convolution kernel sizes adopted by four convolution layers in the second branch are 32, 3 and 3 respectively.
4. The mechanical equipment association measuring point migration learning method based on the multi-scale convolution network according to claim 1, wherein the joint loss function is expressed as follows: ; Wherein, the Represents a cross entropy loss function of source domain training, Is the adjustment coefficient of the light source, Representing the inter-domain difference loss obtained from the maximum mean square difference.

Description

Mechanical equipment associated measuring point migration learning method based on multi-scale convolution network Technical Field The invention relates to a mechanical equipment association measuring point migration learning method based on a multi-scale convolution network, and belongs to the field of fault diagnosis of mechanical equipment. Background The continued growth in manufacturing industry has relied on efficient operation of various mechanical devices. However, in the case of long-term, high-load operation of mechanical equipment in an industrial field, the state of the equipment is liable to deteriorate or even malfunction if professional maintenance or maintenance is lacking. This not only affects the life and productivity of the equipment, but also may cause casualties in severe cases. In recent years, with the rapid development of artificial intelligence technology, particularly the rise of deep learning, the artificial intelligence technology has an unprecedented potential in the field of mechanical equipment fault diagnosis. Deep learning can automatically mine hidden features and rules from mass data by constructing a complex neural network model, thereby providing accurate prediction and diagnosis for faults of mechanical equipment. However, the complex layout and operation conditions of mechanical equipment in an industrial field often cause data loss of required positions, and finally, the deep learning model lacks key data, which restricts the application of the deep learning model in the field of fault diagnosis. Therefore, how to effectively use deep migration learning to extract diagnosis knowledge from the mechanical equipment associated measuring points and effectively migrate the diagnosis knowledge to the target measuring points becomes a hot spot and a difficult point of current research. Disclosure of Invention The invention provides a mechanical equipment associated measuring point migration learning method based on a multi-scale convolutional network, which utilizes an improved convolutional neural network to carry out migration learning, successfully extracts equipment fault information and carries out fault characteristic domain adaptation on different measuring points of equipment, realizes migration of diagnosis knowledge between an associated measuring point and a target measuring point, and solves the problems of inconvenient sensor installation and poor knowledge transfer diagnosis capability in mechanical equipment diagnosis. The technical scheme of the invention is as follows: A mechanical equipment association measuring point migration learning method based on a multi-scale convolution network comprises the steps of collecting association measuring point data of equipment fault points according to migration tasks to serve as source domain data, collecting target point data to be diagnosed to serve as target domain data, carrying out normalization processing on the source domain data and the target domain data, carrying out data segmentation processing on the normalized data by adopting a sliding window mechanism to obtain samples, determining class labels corresponding to the samples, finally obtaining a source domain data set and a target domain data set, wherein the class labels comprise normal class labels and different fault class labels, constructing a domain learning network, jointly training the source domain and the target domain through a shared network structure, using the source domain data with the labels as input of the domain learning network of the source domain to extract association measuring point characteristics, sharing the network structure and weights to the domain learning network of the target domain, then using the target domain data without the labels as input of the domain learning network of the target domain to extract target measuring point characteristics, carrying out inter-domain difference measurement on the association measuring point characteristics and the target measuring point characteristics by utilizing a maximum mean square difference method, and combining the inter-domain learning network classification loss and inter-domain difference function of the source domain learning network to achieve optimal iteration training of the target model, and achieve excellent target migration learning effect, and finally achieve object model propagation and excellent object model propagation and channel prediction. The sliding window mechanism method comprises the following steps: In the formula, n represents the number of samples obtained by a sliding window, l is the sample length, and s is the window step size. The domain learning network comprises the steps of firstly extracting data features by utilizing a multi-scale convolutional neural network, introducing a self-attention module to enhance the characterization capability of the features, then fusing feature information under different scales, and finally measuring feature diffe