Industrial equipment feature extraction method, system, equipment and medium

Abstract

The invention discloses an industrial equipment feature extraction method and system, equipment and a medium, and relates to the technical field of intelligent operation and maintenance of industrial equipment, and the method comprises the steps: obtaining data of a plurality of sensors on the industrial equipment; extracting full-dimensional fusion features of samples in single sensor data to obtain single sensor sample state representation; based on the time sequence data of the full-dimensional fusion features of the samples in all sensor data, fusing multi-sensor information to obtain multi-sensor fusion features; sample information in single-sensor sample state representation is reserved, and feature extraction of different states is carried out on multi-sensor fusion features under the supervision of limited operation and maintenance task labels. According to the method, the effective correlation between the marked samples and the unmarked samples across the sensors is established to realize label transfer, so that the challenge of limited marked samples is overcome, and different equipment state characteristics can be adaptively extracted according to different requirements in industrial fault diagnosis and health state recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance of industrial equipment, and particularly relates to an industrial equipment feature extraction method, system, device and medium. BACKGROUND

[0002] At present, with the increasing complexity and cost of industrial equipment systems, the demand for higher operation reliability, stability and safety is also growing. Data-driven intelligent operation and maintenance technology for industrial equipment has received extensive attention. These technologies extract equipment state features from operation data to achieve intelligent maintenance, and accurate state features can significantly improve the efficiency of industrial equipment operation and maintenance tasks such as fault diagnosis, health state recognition and predictive maintenance. Ultimately, this can reduce equipment downtime and maintenance costs.

[0003] Since industrial equipment systems are usually multi-component systems, their complex operating environment, large-scale structure and other factors cause inaccuracies in the measurement results of a single sensor. In addition, state feature extraction models are susceptible to external factors such as sensor failures and strong noise interference, which can lead to poor performance in operation and maintenance tasks such as fault diagnosis and health state recognition. Therefore, relying solely on the signals of a single sensor cannot meet the needs of efficient operation and maintenance. On the contrary, integrating multi-dimensional and multi-measurement point sensor data is crucial for extracting reliable equipment state features. However, most existing research focuses on single-sensor data, and current state feature extraction models are designed independently for a single operation and maintenance task. Therefore, it is of great significance to develop a model that can extract state features using multi-sensor data. Such a model must also adaptively extract different equipment state features according to the specific requirements of the operation and maintenance task. In addition, the high cost of data labeling, as well as concerns about data privacy and security, have resulted in limited labeled samples for various operation and maintenance tasks. This presents a key challenge: how to extract equipment state features based on multi-sensor data in an optimal manner under limited operation and maintenance task labels.

[0004] Current research on multi-sensor signal fusion methods for industrial equipment can be divided into three types: data-level fusion, feature-level fusion, and decision-level fusion. Data-level fusion uses algorithms such as weighted fusion or least squares to integrate raw multi-sensor data into a unified dataset. However, this approach has some problems, such as high dimensionality of the fused data and low fusion efficiency. Feature-level fusion includes two stages: single-sensor state feature extraction based on algorithms such as signal analysis and neural networks, and multi-sensor state feature fusion using methods such as feature combination and attention mechanisms. The fused features have the characteristics of least interference, low dimensionality, and high robustness. Decision-level fusion uses strategies such as average voting or evidence theory to combine the decision results of multiple sensors. Although decision-level fusion has strong anti-interference ability, it discards a large amount of sensor state information in the process. Existing research shows that feature-level fusion is better than the other two methods. It can achieve high-level shared representation of features, reduce the dimensionality of features compared to data-level fusion, and retain key fault information that is often lost in decision-level fusion.

[0005] For feature-level fusion, various methods have been proposed. In single-sensor state feature extraction, traditional signal analysis methods such as time domain, frequency domain, and wavelet transform largely depend on domain expertise and exhibit limited nonlinear learning capabilities. Deep learning-based methods, including convolutional neural networks and transformers, have gained attention due to their adaptive feature learning and powerful nonlinear modeling capabilities. However, they require a large amount of labeled data to ensure model performance and often fail to capture inter-sample relationships, resulting in partial loss of sensor state features. Graph representation learning (GRL) methods based on graph neural networks (GNNs) have gained increasing attention by constructing sample association graphs (SAGs) to model the interrelationships between samples (including labeled and unlabeled) and their inherent features. These methods naturally address the challenge of limited labeled samples by leveraging graph structures to propagate information between labeled and unlabeled data points. For example, Xie et al. use multi-scale graph convolution to extract multi-scale relationship information between labeled and unlabeled samples and deep nonlinear features from individual samples. Yu et al. establish sample associations by calculating the Euclidean distance between time and frequency domain features, and subsequently develop a graph-weighted reinforcement network to obtain comprehensive feature representations. These efforts have driven research in graph-based single-sensor state feature extraction. However, there are still three major challenges to be addressed:

[0006] 1) Existing methods only use the time domain or frequency domain features of the sample, which cannot comprehensively describe the internal state information of the sample, and ignores the actual impact of noise. This leads to inaccurate features, missing features, and redundant features in some features. 2) Existing methods for constructing SAG fail to comprehensively integrate local and global structure information between samples. In addition, the dependence on empirical parameters such as similarity threshold and number of adjacent nodes leads to the lack of real edge connections in the graph. 3) Existing methods only focus on extracting structure-semantic mixed hierarchical information after constructing non-Euclidean graph data from samples, while ignoring in-depth exploration of multi-perspective insights in the graph (e.g., deep structure and semantic hierarchy). This will result in information loss in the graph data.

[0007] In multi-sensor feature fusion, methods such as cascading or weighted combination are affected by noise, feature redundancy, and static weights that cannot be dynamically adjusted, resulting in the loss of some non-linear information. Attention-based methods (e.g., channel or spatiotemporal mechanisms) face challenges such as high computational complexity, low interpretability, and neglecting sensor coupling. GRL-based methods (e.g., GNNs), i.e., optimizing fusion by explicitly modeling the dependence between sensors, have become a research focus. Wang et al. used the cosine similarity between sensor node data to construct a multi-sensor coupling graph and introduced an attention-aware spatiotemporal graph neural network to achieve multi-sensor information fusion. Cao et al. designed a multi-sensor coupling graph by integrating sensor node attribute features and prior knowledge, and then used a graph attention network (GAT) to effectively fuse multi-sensor information.

[0008] Existing expert cognition-based methods for establishing sensor coupling relationships have subjective bias and cognitive limitations, requiring a large amount of manual work and prior knowledge input. Existing similarity-based modeling sensor coupling methods rely on empirical parameter selection (e.g., neighborhood size, similarity threshold) and superficial data mining (only capturing shallow similarity), resulting in suboptimal coupling relationships. And existing methods for extracting fusion information from multi-sensor graph data rely on simple graph pooling (e.g., average or maximum pooling) to derive sample state features, resulting in the loss of a large amount of sensor information, which leads to the final extracted features not reflecting the state of the industrial equipment. SUMMARY

[0009] The present application aims to solve the problems in the prior art by providing an industrial equipment feature extraction method, system, device and medium.

[0010] The present application specifically provides the following technical solutions, an industrial equipment feature extraction method, comprising: Obtaining a plurality of sensor data on an industrial equipment; The full-dimensional fusion features of samples in single sensor data are extracted, and based on the full-dimensional fusion features, a single sensor sample correlation graph is constructed by extracting the proximity relationship between samples in single sensor data, multi-expert feature expression acquisition and fusion are performed based on the single sensor sample correlation graph, and single sensor sample state representation is obtained. Based on the time series data of the full-dimensional fusion features of samples in all sensor data and the sample correlation graph of single sensor, a multi-sensor spatio-temporal graph data is constructed to generate the coupling relationship between multiple sensors, and a multi-sensor graph data is constructed through the coupling relationship, and hierarchical adaptive graph coarsening is performed on the multi-sensor graph data to fuse multi-sensor information, and multi-sensor fusion features are obtained. The sample state information in the single sensor sample state representation is retained, and adaptive classification extraction is performed on the multi-sensor fusion features under the supervision of limited operation and maintenance task labels to obtain device state features.

[0011] The application provides an industrial equipment feature extraction method and system, comprising: A data acquisition module is configured to acquire multiple sensor data on an industrial equipment. A single sensor processing module is configured to extract full-dimensional fusion features of samples in single sensor data, and based on the full-dimensional fusion features, a single sensor sample correlation graph is constructed by extracting the proximity relationship between samples in single sensor data, multi-expert feature expression acquisition and fusion are performed based on the single sensor sample correlation graph, and single sensor sample state representation is obtained. A multi-sensor processing module is configured to construct a multi-sensor spatio-temporal graph data based on the time series data of the full-dimensional fusion features of samples in all sensor data and the sample correlation graph of single sensor to generate the coupling relationship between multiple sensors, and a multi-sensor graph data is constructed through the coupling relationship, and hierarchical adaptive graph coarsening is performed on the multi-sensor graph data to fuse multi-sensor information, and multi-sensor fusion features are obtained. A feature extraction module is configured to retain sample state information in the single sensor sample state representation, and perform adaptive classification extraction on the multi-sensor fusion features under the supervision of limited operation and maintenance task labels to obtain device state features.

[0012] The application provides a computer device, comprising a memory and a processor, the memory stores a program, and the program is executed by the processor to make the processor execute the steps of the above-mentioned industrial equipment feature extraction method.

[0013] The application provides a storage medium, which stores a computer program, and the computer program realizes steps of an industrial equipment feature extraction method when executed by a processor.

[0014] Compared with the prior art, the application has the following remarkable advantages: The application extracts features of a single sensor in full dimension, constructs a sample correlation graph based on the proximity between samples in the extracted single sensor data, and then performs multi-expert feature expression fusion to obtain single sensor state representation, which provides a data basis for subsequent feature selection of multiple sensors, and based on time series data of all sensor data and the sample correlation graph of the single sensor, a coupling relationship between multiple sensors is generated, and relevant processing is performed based on the coupling relationship to obtain multi-sensor fusion features, the application models the explicit coupling relationship between multiple sensors by using a complete data-driven method, realizes optimized multi-sensor information fusion, and then effectively associates samples across sensors to realize label transfer, thereby overcoming the challenge of limited labeled samples, and different device state features can be adaptively extracted according to different requirements, and the current situation of the industrial equipment can be better reflected. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 A flowchart of the method provided by the application is shown in the figure; Figure 2 A sample NCMAG construction architecture diagram provided by the application is shown in the figure; Figure 3 A single sensor sample multi-expert feature representation architecture diagram provided by the application is shown in the figure; Figure 4 A single sensor sample multi-expert feature representation architecture diagram provided by the application is shown in the figure; Figure 5 An original signal waveform diagram collected by a test bench provided by the application is shown in the figure; Figure 6 A structure configuration diagram of a test bench in a PTFD data set provided by the application is shown in the figure; Figure 7 A result diagram of each stage of a model on a CWRU data set provided by the application is shown in the figure; Figure 8 A result diagram of each stage of a model on a PTFD data set provided by the application is shown in the figure; Figure 9 A t-SNE visualization diagram of single sensor and multi-sensor fusion features on a CWRU data set provided by the application is shown in the figure; Figure 10 A t-SNE visualization diagram of single sensor and multi-sensor fusion features on a PTFD data set provided by the application is shown in the figure; Figure 11 A comparison experiment result graph of the model provided by the present application under two case studies; Figure 12 A confusion matrix graph of the comparison model on the CWRU data set provided by the present application; Figure 13 A confusion matrix t graph of the comparison model on the PTFD data set provided by the present application; Figure 14 A t-SNE visualization graph of the comparison model result on the CWRU data set provided by the present application; Figure 15 A t-SNE visualization graph of the comparison model result on the PTFD data set provided by the present application; Figure 16 A reducer performance degradation test platform graph provided by the present application; Figure 17 A raw signal waveform graph collected by the test bench provided by the present application; Figure 18 A comparison experiment result graph of the model on the GDD data set provided by the present application. DETAILED DESCRIPTION

[0016] The technical solutions of the embodiments of the present application will be described clearly and completely below with reference to the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor should fall within the scope of protection of the present application.

[0017] The present application proposes MSMNEG. The main contributions are summarized as follows:

[0018] 1) To solve the problem of extracting the state features of industrial equipment under limited operation and maintenance task labels, while utilizing multi-sensor fusion technology, the present application proposes a multi-sensor multi-non-Euclidean graph fusion method, which can adaptively extract state features for specific tasks for various operation and maintenance scenarios (fault diagnosis and health state recognition).

[0019] 2) A variety of non-Euclidean graph data are constructed. Among them, single-sensor sample graph data (SSSG) is used to complete the representation of single-sensor state features; multi-sensor spatiotemporal graph data (MSSpG) is used to simulate the real coupling relationship between multiple sensors; and multi-sensor sample graph data (MSSG) is used to fuse information from multiple sensors.

[0020] 3) For single-sensor sample state representation, firstly, robust full-dimensional fusion features of samples are extracted. Then, considering the local and global correlation information between samples, and introducing the idea of link prediction, a neighborhood complementary mixed association graph (NCMAG) construction method is proposed to establish SSSG. Finally, a multi-expert (structure, semantic and hybrid) feature representation and fusion framework is designed based on graph data for sample state representation.

[0021] 4) For multi-sensor sample state feature fusion, a MSSpG is constructed, and based on this, a completely data-driven multi-sensor spatial coupling relationship modeling method is designed. Then, a MSSG is generated, and a hierarchical adaptive graph coarsening method is proposed to optimally integrate multi-sensor information.

[0022] 5) To solve the applicability and effectiveness of multi-sensor fusion features in multiple operation and maintenance tasks, a multi-strategy jointly driven feature optimization method is proposed. These strategies include: a limited operation and maintenance task label driven adaptive fusion feature adjustment strategy and a feature enhancement strategy to retain sample information in single-sensor state representation.

[0023] Among them, the signal full-dimensional feature is: the original waveform of the signal provides a single sample feature representation in the particle dimension (PaD), which is represented by x o , which retains all the details and features of the signal and can accurately reflect the true state of the signal.

[0024] The relative position matrix (RPM) constructed by the signal provides the information representation of the correlation between the parts of a single sample in the Cartesian coordinate system in the local dimension (LoD), which is represented by x m . RPM contains redundant characteristics from the original time series, making it easier to capture similarity information between classes and within classes in the generated image.

[0025] The Gram angle difference field matrix (GADF) constructed by the signal provides the information representation of the correlation between the parts of a single sample in the polar coordinate system in the local dimension, which is represented by x d . GADF uses the Gram matrix to calculate the linear correlation between a set of vectors, and finally obtains the difference between different data points, GADF largely retains the integrity of the signal and its time dependence.

[0026] The time and frequency domain features of the signal provide the overall information representation of a single sample itself in the panoramic dimension (LoD), which is represented by x tfIt is shown that 14 time-domain features and 5 frequency-domain features sensitive to mechanical faults are extracted. Time-domain features can provide information about the changes of signals over time. The frequency-domain features of the signal provide information about the signal in the frequency domain, which helps to analyze the different frequency components contained in the signal.

[0027] A detailed flowchart of the proposed method is shown in Figure 1 It mainly includes three stages: single-sensor sample state representation, multi-sensor sample state feature fusion, and multi-sensor fusion feature optimization. The following describes an industrial equipment feature extraction method in this embodiment, including the following steps:

[0028] Step S1: Obtain multiple sensor data on an industrial equipment.

[0029] Step S2: Extract full-dimensional fusion features of samples in single-sensor data, and based on the full-dimensional fusion features, construct a single-sensor sample correlation graph by extracting the proximity relationship between samples in single-sensor data, obtain and fuse multi-expert feature expression based on the single-sensor sample correlation graph, and obtain single-sensor sample state representation; the full-dimensional fusion features represent the semantic information of various dimensional features.

[0030] For the first i sensor S i , full-dimensional features are extracted from sample signals to obtain comprehensive equipment state information, however, sensor signals often have noise, and manually constructing full-dimensional features will introduce redundancy, thereby reducing model performance. The denoising autoencoder architecture is used to automatically fuse multi-dimensional features, wherein the encoding-decoding process enables the model to learn robust and information-rich features from damaged samples, thereby producing a robust full-dimensional fusion representation. First, the model encoding process is performed, specifically as follows:

[0031] Step S21: Extract full-dimensional fusion features of samples in single-sensor data, specifically: Particle dimensional features, local dimensional features, and panoramic dimensional features are extracted from sample signals of single-sensor data. For each dimensional feature of each sample signal, the feature values of each dimensional feature are randomly replaced according to the binomial distribution, i.e., the feature values are randomly set to 0, and the semantic information of each dimensional feature is extracted through self-attention mechanism, attention mechanism, and convolutional neural network, respectively. For x o and x tf , self-attention mechanism (SA) and multi-layer perceptron (MLP) are used to extract the semantic information of particle dimensional features x p1 and panoramic dimensional features x p2 , respectively; forx m and x d , respectively, use channel attention and spatial attention mechanisms to reconstruct sample information and better describe the local dimensional features of the sample. Multiple convolutional neural network layers are used to extract features from each image, and finally, these features are concatenated in the channel dimension. A convolutional neural network layer is used to extract inter-image feature information, which is then flattened to obtain semantic information of the local dimensional features of the sample x p3 .

[0032] The semantic information of each dimensional feature is concatenated to obtain the full-dimensional fusion feature of the sample, and the full-dimensional fusion feature of the sample is decoded to obtain the final reconstructed full-dimensional fusion feature. After concatenation, SA and MLP are used to finally obtain the robust full-dimensional fusion feature of the sample x r The feature set composed of all samples is X i To ensure X i can overcome the influence of noise and effectively extract semantic information, the X i is decoded. Decoding is the inverse process of encoding, and the reconstructed feature is finally obtained through decoding.

[0033] For the first i sensor S i , an NCMAG is constructed based on its features X i to extract the proximity relationship between samples, as shown in Figure 2 . Specifically, the NCMAG combines S i the local and global relationships between samples, and uses link prediction to optimize the correlation graph.

[0034] Step S22: Based on the full-dimensional fusion feature, the proximity relationship between single sensor data samples is extracted to construct a sample correlation graph for a single sensor, specifically: In constructing the NCMAG, the first step is to capture the local structure of sample correlation. Based on the full-dimensional fusion feature, the Euclidean distance between sample feature pairs (full-dimensional fusion features) is calculated to obtain the K nearest neighbor set of each sample point , where K = 3, and based on the K nearest neighbor algorithm (KNN), the local structure of the nearest neighbor set is captured to obtain the sample local correlation structure matrix , where the calculation of element is as follows:

[0035] ; The similarity threshold method captures the global structure of sample association, which is different from the fixed number of KNN, and can adapt to data of different density and distribution, and help to reveal the global structure of data. Specifically, the cosine similarity between sample features X S is calculated s , and the similar node set of each sample point is obtained based on the similarity threshold s = 0.5, and the sample global association structure matrix is constructed based on the similar node set , where the element is calculated as follows:

[0036] ; The local association structure matrix and the global association structure matrix are fused to obtain the comprehensive association matrix , which maximizes the similarity between nodes, and the calculation method of the matrix element is as follows: ; In order to further optimize the graph structure, the GAE is introduced to reconstruct the comprehensive association matrix to obtain , that is, the comprehensive association matrix is reconstructed by using the graph autoencoder, and the reconstructed association graph A P retains the key topological structure of the original graph, and the link prediction is performed on the reconstructed association graph to complete the missing edges, and the optimized sample association graph of single sensor is obtained. This process effectively utilizes the graph structure learning ability of GAE, thereby inferring the true NCMAG.

[0037] ; ; Wherein, GCN(·) represents the update process of GCN, x r is the node feature, is the sigmoid activation function, is the embedding vector of the node, is the adjacency matrix of the completed positive association graph.

[0038] For any sensor channel S i , the SSSG( G i = ( V , A P,i ,X i ​), wherein V is a set of samples. It is worth noting that due to the consistency of mechanical system monitoring sensors in time, each sensor will obtain a consistent number of samples in time.

[0039] Based on graph data G i , three nonlinear feature extraction experts are trained, as shown in Figure 3 . They are responsible for extracting the topological structure features of the sample correlation graph (structure representation expert), the deep semantic information of the sample itself (semantic representation expert), and the semantic features fused with the topological structure of the sample correlation graph (hybrid representation expert) from the graph data G i . Finally, the multi-expert feature representation of the sample under S i is obtained.

[0040] Step S23: Obtain and fuse the multi-expert feature representation based on the sample correlation graph of a single sensor to obtain the single-sensor sample state representation, specifically: Use the Node2vec algorithm (which is a technique that extends language models and unsupervised learning to graph embeddings) to extract the structural features of the samples from the topological structure of the sample correlation graph (graph data G i ) of a single sensor, capturing the homogeneity between adjacent nodes (i.e., similar features of connected nodes) and the structural equivalence between non-adjacent nodes (i.e., nodes with similar positions or roles in the correlation graph have similar features). The implementation of Node2vec is as follows:

[0041] ; wherein, v and x represent nodes in the graph data G i ; c i represents the i th node in the random walk; is the unnormalized transition probability between nodes v and x ; Z is the normalization constant; w vx is the weight of the directed graph, the value of ; wherein, d tx is the shortest path distance between vertex t and vertex x ; the parameter pAnd q is the degree of regulating breadth-first search and depth-first search, by regulating p And q controls the direction of random walk, so that the generated random walk node sequence can reflect the characteristics of both BFS and DFS (local characteristics and global characteristics of neighbor nodes), improving the effect of node embedding. Further, the generated node sequence is trained by a Skip-gram model to convert the nodes into low-dimensional embedding vectors, and finally an embedding matrix , c is the dimension size of the embedding vector, n is the number of nodes, M contains the structure information of the graph G i

[0042] for the sample feature X i Because the sample full-dimensional fusion feature extraction has preliminarily extracted the nonlinear features of the sample, only the multi-scale one-dimensional convolution network is used to extract the multi-scale semantic features of the sample points from the full-dimensional fusion features. Specifically, different sizes of one-dimensional convolution kernels are used for parallel multiple convolution operations on the sample features to extract features at different levels, forming a more complete feature representation and increasing the expression ability of the features. Then, the feature maps of different scales are spliced in the channel direction, and finally a 1x1 convolution kernel is used to fuse the features in the channel direction. The shape of the convolutional feature map is the same as the original Figure 1 , but contains richer hidden features. In this embodiment, the sizes of the one-dimensional multi-scale convolution kernels are 1x1, 1x3, 1x5 and 1x7 respectively.

[0043] Hybrid representation expert-considering sample information transmission feature extraction: using graph convolution network (GCN) to fuse the correlation features between samples (sample correlation graph) and the full-dimensional fusion of the sample itself, and an optimized multi-receptive field enhancement strategy is proposed to promote the transmission of information between samples. The expert effectively uses the node connectivity and feature propagation mechanism, and can effectively transmit the information of the labeled nodes to the unlabeled nodes, thereby enhancing the feature learning and classification performance of a single sensor sample, and using a multi-receptive field enhancement strategy to aggregate the neighborhood information of different order sample features to generate an enhanced representation, i.e. structure-semantic hybrid features. Specifically, based on the graph data G i , the first-order, second-order and high-order sample feature representations are obtained using the GCN network:

[0044] ; wherein, h i,k is the first k ​Node features obtained after layer processing.

[0045] As GCN aggregates neighborhood information, each layer gradually expands the receptive field to generate first-order, second-order, and higher-order feature representations. Each layer can capture both short-range and long-range sample similarities, and this section integrates multi-order features to jointly model short-range and long-range dependencies. SA dynamically assigns weights to features of different orders to adapt to various operation and maintenance tasks. Specifically, the hybrid representation expert aggregates node representations of all layers into a comprehensive representation . To address the limited nonlinearity of sequential feature concatenation, the expert combines SA and MLP to nonlinearly fuse multi-receptive field features, generating the final enhanced representation h .

[0046] Multi-expert feature adaptive fusion strategy based on sample conversion: In the multi-expert feature fusion, existing methods mostly use linear weighted fusion or feature concatenation for feature fusion. The present invention concatenates different feature representations of the same sample at the same position in different experts to form a new sample, and uses self-attention mechanism SA and multi-layer perceptron MLP for nonlinear deep fusion of multi-expert feature information to obtain single sensor sample state representation. The self-attention mechanism SA captures the correlation between different expert features, and the MLP extracts the nonlinear relationship between expert features, and finally outputs the multi-expert fusion features of the sample z , with feature dimension n × d , where d represents the dimension of the latent space of the final node representation.

[0047] Step S3: Based on the time series data of the full-dimensional fusion features of the samples in all sensor data and the sample association graph of the single sensor, a multi-sensor spatio-temporal graph data is constructed to generate the coupling relationship between the multi-sensors, and the multi-sensor graph data is constructed through the coupling relationship, and the multi-sensor information is fused by hierarchical adaptive graph coarsening to obtain multi-sensor fusion features.

[0048] As Figure 4 shown, the present invention aims to completely infer the coupling relationship between multi-sensors from data, and then realize multi-sensor information transmission and fusion. Drawing on the principles of protein interaction dynamics, a completely data-driven multi-sensor spatial relationship modeling method is proposed. This method first assumes that there is a fully connected coupling graph between sensors, and then learns the real coupling relationship between sensors through unsupervised learning.

[0049] Step S31: Based on the time series data of the full-dimensional fusion features of the samples in all sensor data, a multi-sensor spatiotemporal graph data is constructed and a coupling relationship between multiple sensors is generated, specifically: The node set, edge set, and time series data of node features of the sensor data are defined; wherein the edge set is initialized as a fully connected graph, and its adjacency matrix represents the potential coupling relationship between sensors; that is: The MSSpG data is constructed TG i =( TV , TE , TX , T ), wherein TV ={ tv 1, tv 2,..., tv n} represents a set of n sensor nodes; represents an edge set, and the number of edges is TE |, each edge represents whether there is a coupling relationship between sensors (initialized as fully connected, represented by adjacency matrix TA ); TX represents the time series data of the robust full-dimensional fusion features of each sensor, and the time t is the feature vector of node tv i at time , and the feature set of all n nodes is recorded as . The trajectory of node tv i is represented as , wherein T is the time step, and all trajectory data records are .

[0050] The multi-sensor spatiotemporal graph data is constructed by the node set, edge set, time series data of node features, and time.

[0051] The coupling structure inferrer is responsible for generating the probability matrix of multiple coupling relationships between sensors. Given the observed dynamic trajectory of a node , on the fully connected graph of the multi-sensor spatiotemporal graph data, since the true coupling structure is unknown, a graph neural network is used to predict the potential coupling structure, and based on the observed dynamic trajectory of the node, the pairwise interaction between sensor nodes is predicted TA ij .

[0052] Using coupling graphs generated by a coupled structure inferrer and time series data, the dynamics of the sensor system of industrial equipment are simulated through multiple parallel graph neural networks, the dynamics of the sensor system at the next time step are predicted, and the coupling relationship between sensors is updated by minimizing the prediction error.

[0053] Based on pairwise interactions between sensor nodes, a coupled inferrer is constructed using a graph neural network running on a fully connected graph. The inferrer is modeled as... ,in This represents a GNN operating on a fully connected graph. Given an input trajectory... The inferrer's sensor message passing operations are as follows:

[0054] in, TA ij,k Indicates sensor i and j Between the first k Such coupling relationship; It is a state of aggregation influenced by the coupling relationship of the sensors; Indicates sensor i The state change value at the next time step; f out This represents the output transformation based on a multilayer perceptron (MLP). In the edge-to-node ( e → v In state propagation, aggregate state Current status and the previous hidden state As input to the Gated Recurrent Unit (GRU) network, it generates the hidden state for the next time step; finally, the current state... The hidden states are used to predict the distribution of future states, where It is a sensor i In time t +1 is the predicted average state. It is the fixed variance of state prediction.

[0055] Training of multi-sensor coupling relationship modeling method: The multi-sensor coupling modeling method is essentially a variational autoencoder (VAE) architecture. This model uses gradient descent for stepwise optimization, and the training objective is to minimize the VAE loss, achieved by maximizing the lower bound of evidence, i.e., minimizing the sum of the reconstruction error term and the KL divergence term.

[0056] The reconstruction error term is: ; The KL divergence term is the sum of the entropy and a constant: ; where H denotes the entropy function.

[0057] Multi-sensor information fusion based on hierarchical adaptive graph coarsening: the multi-sensor features of any sample are regarded as a graph, and the fused feature representation is obtained by graph pooling, and all sample graphs share the same model structure and parameters.

[0058] Multi-sensor graph (MSSG) construction under multi-coupling relationship: For a sample j , use the multi-coupling relationship between sensors in the previous section to construct the time-aligned multi-sensor graph data SG j ( SV , TA , SX ), where SV is the number of sensor nodes, TA is the adjacency matrix encoding the coupling relationship (including multi-coupling structure), SX denotes the state feature of all sensors in sample j . All samples share the same SV and SA in their multi-sensor graphs.

[0059] Node information updater with multi-sensor message passing: aims to perform inter-sensor information propagation, reveal hidden correlations in multi-sensor data, and derive the node representation at the current level.

[0060] Step S32: Hierarchical adaptive graph coarsening of multi-sensor graph data, specifically: Before performing the next graph coarsening (passing to the next graph coarsening layer) on the multi-sensor sample graph data SG j , update the node features through message passing, where, due to the presence of multiple initial coupling relationships in the initial input, a multi-space graph neural network (GNN) is used to handle multiple initial coupling relationships during the first layer of message passing, i.e. comprehensively capture the influence between sensors, which is different from subsequent layers. The detailed process is as follows:

[0061] ; ; where, denotes the first kA nonlinear neural network corresponding to the coupling relationship; O ij,att,k represent the trainable attention values between sensor nodes i and j ; O ij,att,k TA ij,k is the attention score of selecting one-hop neighbors of the sensor node. Finally, the first layer graph coarsening is performed based on the updated node features.

[0062] For the first l layer node information passing, the input is the graph coarsening result from the previous layer. Specifically, as follows:

[0063] ; ; where, SA l-1 denotes the adjacency matrix obtained after the graph coarsening process for l −1) times. By performing multiple rounds of node information passing, the high-order neighborhood relationship between sensors can be effectively encoded. After the information passing in the l layer, the final sample feature matrix H ′ is obtained.

[0064] The multi-sensor information fusion module based on graph coarsening: for the graph coarsening of any layer, first, the global graph information is extracted from the multi-sensor sample graph data.

[0065] where, the global graph content matrix for the current graph is generated using the learnable linear transformation parameterized by the weight matrix T ∈ R F×N , which captures the overall graph structure, ensures the stability before and after coarsening, and provides global guidance for the graph coarsening process. The detailed steps are as follows:

[0066] ; where, H ij and T ij denote the H th row, the i th column element of the feature matrix j and the corresponding element of the weight matrix C ∈ R N×N′ is the automatically learned global graph content matrix, where each row C ( i ,⋅)∈ R N′Each column corresponds to a node in the source graph C ( i ,⋅)∈ R N′ is mapped to a cluster node in the target coarsened graph. C The elements in the matrix reflect the interaction between the nodes in the source graph and the clusters in the coarsened graph, bridging the gap and ensuring consistency between them.

[0067] Based on the global graph content matrix, the attention score between the nodes in the previous coarsened graph and the clusters in the target coarsened graph is calculated through a cross-level attention mechanism. That is, after obtaining the global graph content matrix, a cross-level attention mechanism is applied between the nodes in the previous coarsened graph and the clusters in the target coarsened graph. The input is the node representation matrix H ={ h 1, h 2,…, h N} is the number of nodes in the source graph G, and F is the feature dimension of each node. Then, the graph coarsening module generates a new coarsened representation matrix H ′={ h 1′ , h 2′ , …, h N′} where h i′ ∈ R F′ is the number of clusters in the coarsened graph. Each cluster is treated as a new node, and the adjacency matrix SA ∈ R N×N is updated to SA ′∈ R N′×N′ .

[0068] First, the attention score matrix M ∈ R N×N′ is generated, whose elements are defined as follows: ; where σ is a LeakyReLU nonlinear activation function; [·||·] represents the concatenation operation; the feature dimension of C (·, j ) is reduced from R N adjust to R N′ ; aT ∈ R 2N′ is a trainable shared attention parameter. Each element Mij representative node i The importance of the feature to the cluster j .

[0069] Finally, in order to improve the attention score, it is normalized as follows: ; Based on the normalized attention score, the coarsened node feature representation and the updated adjacency matrix are generated.

[0070] Generate coarsened representation matrix H ′∈ R N′×F And update the adjacency matrix A ′∈ R N′×N′ , the specific expression is: ; ; Where the original SA is a sparse matrix, but the updated SA ′ becomes a fully connected graph with different connection strengths between nodes. In order to reduce the complexity of the model, Gumbel-SoftMax is used to soft sample the neighborhood relationship, thereby reducing the edge density of the sampled adjacency matrix SA ′.

[0071] ; Where, obeys the standard Gumbel distribution; τ is a temperature parameter that controls the uniformity of the sampling result, which is set to 0.1 in the present application to ensure that the adjacency matrix distribution is approximately one-hot encoded. This operation minimizes the edge density while maintaining the connectivity of the graph.

[0072] Multi-sensor fusion feature optimization: due to the adaptability of the multi-sensor features extracted by the deep learning model, a multi-strategy joint driving multi-sensor fusion feature optimization method is proposed to ensure the effectiveness of the features.

[0073] Step S4: retain sample information in single-sensor sample state representation, ensure the effectiveness of single-sensor sample state representation, and perform feature extraction of multi-sensor fusion features in different states under limited operation and maintenance task label supervision.

[0074] Feature enhancement strategy for retaining sample information in single-sensor state representation: To enhance the expressive power of feature vectors in single-sensor state representation and achieve effective representation of single-sensor information, the state representation features are constrained to preserve the semantic information of samples and the information about the relationships between samples, thereby promoting effective information transmission between samples.

[0075] The sample information in the single-sensor sample state representation is preserved, including the preservation of semantic information of single-sensor samples and the preservation of inter-sample relationship information of single-sensor data.

[0076] (1) Preservation of semantic information of single sensor samples: For any sensor i To preserve the semantic information of single-sensor samples, constraints are established to maintain node information. This is achieved through asymmetric decoding and reconstruction of the semantic information of the samples. The loss function is defined as follows:

[0077] ; in, Indicates the first i Robust full-dimensional fusion features of each sample Indicates the first i Features of each sample after decoding and reconstruction.

[0078] (2) Preservation of inter-sample relationship information in single-sensor data: For any sensor i The relationship information between samples can be obtained using the adjacency matrix. A p This indicates, specifically, for sensors i Sample representation h Use a reconstruction function Reconstruct the initial connectivity relationships from the sample representation (represented as follows) The reconstructed connections are derived using inner product derivation.

[0079] ; Finally, the loss function is defined as follows: ; in, N Indicates the number of nodes. This represents the set of all nodes.

[0080] An adaptive fusion feature adjustment strategy driven by limited operation and maintenance (O&M) task labels: Under the supervision of a limited set of operation and maintenance task labels, this model can adaptively extract equipment status features. For example, if the label is a fault type, the model will extract fault features; if the label is a healthy state, the model will extract healthy state features. This strategy has three advantages:

[0081] (1) The strategy can adaptively select device state information. Since the initially extracted full-dimensional robust features may exhibit redundancy in different operation and maintenance tasks, the task label guided model adaptively extracts relevant state features for each task. (2) The strategy can supervise the effectiveness of multi-sensor fusion features according to label information, ensuring that the optimal fusion features are extracted. (3) The strategy solves the few-shot label problem. By capturing the relationship between samples, the information of labeled samples is propagated to unlabeled samples, thereby realizing label propagation and model training under limited operation and maintenance task labels.

[0082] Specifically, the multi-sensor fusion feature vector of the sample is input into a multi-layer perception (MLP) classifier to obtain its final state information. Limited labeled samples are used to supervise model training, and the classification loss of these samples is minimized through gradient descent to enhance label propagation. Since it is a multi-classification problem, the loss function to be optimized is as follows:

[0083] ; wherein, N is the number of labeled samples; M is the number of classes in the operation and maintenance task; y ic is an indicator function, and if the real class of sample i is c , then the value is 1; p ic represents the probability of the model classifying i to class c .

[0084] Case study: To verify the multi-sensor information fusion capability of MSMNEG (multi-sensor information fusion method, which is a hypothetical proper noun and can be adjusted according to actual conditions) under limited labeled samples in operation and maintenance (O&M) tasks, a case study was conducted on two operation and maintenance tasks: device fault diagnosis and health state recognition. For the device fault diagnosis task, a case study was conducted on the Case Western Reserve University (CWRU) open dataset and the benchmark dataset collected by the laboratory from the power transmission fault diagnosis (PTFD) platform. For the health state recognition task, experiments were conducted on the gear box degradation dataset (GDD) collected by the laboratory from the reducer platform. To ensure statistical reliability, the experiments were repeated 10 times.

[0085] Comparison methods: Seven state-of-the-art multi-sensor fusion methods were selected for comparative analysis. Method 1 (MDMF) represents a machine learning method, methods 2-3 (DSRNet and M2FN) are traditional deep learning methods, and methods 4-7 (MIFDELN, A-TSGNN, ADFG, and MSMLF) constitute the latest multi-sensor fusion technology based on graph neural networks. The advantages and limitations of these models are comprehensively compared from multiple angles.

[0086] For all comparison algorithms, the recommended values of the method parameters are set and further fine-tuned to achieve the best performance on the three datasets. The hyperparameter values of the model on each dataset in the present invention are shown in Table 1.

[0087] Table 1 Parameter configuration of the model on the three operation and maintenance task datasets Industrial equipment fault diagnosis task: Dataset introduction: (1) Case A: Case study on public dataset: The Case Western Reserve University (CWRU) dataset has been widely used to verify the fault diagnosis method of industrial equipment. The main components of the test bench include a 2-horsepower motor, torque sensor / encoder, and dynamometer. In the experiment, acceleration data from three different sensor locations (drive end, fan end, and base) were selected as multi-sensor input. Single-point bearing faults were introduced using electrical discharge machining (EDM), and the detailed fault type specifications are shown in Table 2 (where "12k" represents a sampling frequency of 12 kHz). The sample waveform is shown in Figure 5 . This dataset contains 237 samples per fault class, with a total of 1422 samples. A 9:1 training-test split ratio was adopted, and Table 3 illustrates the four limited labeled sample conditions for the training set division.

[0088] Table 2 Detailed description of fault types Table 3 Four limited labeled sample conditions (2) Case B: Experimental dataset case study: To further verify the multi-sensor fusion capability of MSMNEG under limited labeled samples of industrial equipment, the research team independently developed a PTFD (Power Transmission Fault Diagnosis) experimental platform, as shown in Figure 6The platform consists of a variable speed drive motor, planetary gearbox, parallel shaft gearbox, bearing radial load unit, and magnetic brake. By electro-erosion machining the inner ring, outer ring, and rolling elements of the bearing, five fault conditions are simulated: normal state (NS), outer ring fault (OF), inner ring fault (IF), ball fault (BF), and compound fault (CF). Piezoelectric IEPE accelerometers are deployed to acquire the vibration signals of the bearing rotation under 2 HP load with a sampling frequency of 10.24 kHz. Six vibration signals are collected, corresponding to three-dimensional spatial vibration components at two orthogonal measurement locations (each location includes X, Y, and Z axes), as shown in Figure 7 The final dataset is divided into five categories, with 512 samples per category, for a total of 2560 samples. The training and test sets are divided in a 9:1 ratio, and Table 4 summarizes the four limited labeled sample conditions for the training set.

[0089] Table 4. Four limited labeled sample conditions Fault diagnosis performance: The performance of the proposed method is verified through an operation and maintenance (O&M) fault diagnosis task. As shown in Table 5, under various limited labeled sample conditions in Case A and Case B, the diagnosis results of MSMNEG exhibit robust performance. Notably, with only 5% labeled training samples, MSMNEG achieves a diagnosis accuracy of 96.21% (standard deviation: 0.22) in Case A and 99.90% (standard deviation: 0.09) in Case B. These results highlight the effectiveness and stability of the method in extreme label scarcity scenarios.

[0090] Table 5. Results of fault diagnosis tasks on CWRU dataset To further demonstrate the effectiveness of the model under limited labeled sample conditions, the results obtained by the model at each stage are visualized in the two case studies. As shown in Figures 7-9 MSMNEG gradually extracts single-sensor sample representations, derives multi-sensor fusion features, establishes sample-to-sample relationships between different sensor channels, and couples relationships between multiple sensors, ultimately overcoming the limitations imposed by the scarcity of labeled samples while achieving efficient fusion of multi-sensor information.

[0091] The t-SNE is used to visualize the single-sensor sample features and multi-sensor fusion features obtained by MSMNEG in Case A and Case B, as shown in Figures 9-11The results show that although single-sensor feature visualization can partially reflect the true fault type of the sample, due to sample noise and limitations of the sensor itself, some samples still have clustering errors. In contrast, multi-sensor fusion features effectively integrate complementary information from various sensor signals, overcome these limitations, and show better visualization results. This confirms that the proposed MSMNEG method successfully integrates multi-sensor information while addressing the challenge of limited labeled samples.

[0092] Comparison of different methods: It can be found that compared with other methods, the proposed MSMNEG method performs outstanding performance under all limited labeled sample conditions, with higher diagnostic accuracy and smaller standard deviation. Further observation Figure 11 It can be found that as the number of labeled samples decreases, the diagnostic accuracy of other methods in Case A and Case B decreases sharply, while the accuracy of the proposed method does not decrease significantly. This shows that insufficient labeled samples can damage the performance of existing models, while the proposed method shows stronger robustness when labeled samples are scarce. This is mainly because MSMNEG extracts the correlation relationship between various sensor samples, establishes the connection between labeled samples and unlabeled samples, and promotes the propagation of label information.

[0093] Table 6 Results of fault diagnosis task on Case A dataset Table 7 Results of fault diagnosis task on Case B dataset To further verify the effectiveness of the model, first use the confusion matrix to visualize the classification results of the fault samples. Figures 12-13 The confusion matrix results of different methods under the limited labeled sample conditions of Case A and Case B are shown. It can be observed that MSMNEG only shows a small number of misclassification nodes, significantly less than other comparative models. This excellent performance is due to two key factors: first, by establishing the association between each sensor labeled sample and unlabeled sample, the method achieves effective propagation of label information. Second, and more importantly, the coupling relationship between sensors is explored in depth from the data, and multi-sensor fusion features are obtained using the proposed graph pooling strategy. This dual approach enables the model to effectively utilize sensor complementarity, ultimately resulting in the optimal device fault feature description.

[0094] Finally, the output results of the fully connected layer are visualized using t-SNE. Figures 14-15The visualization results under the limited labeled sample condition 1 of case A and case B are shown. The results show that MSMNEG can more accurately distinguish different fault types, showing smaller intra-class distance and larger inter-class distance in fault categories. Compared with other methods, it achieves the best visualization performance, which further proves that MSMNEG can more effectively extract and fuse multi-sensor data information. This ultimately verifies the superiority of the proposed method when only limited labeled samples are available.

[0095] Industrial equipment health state recognition task: The performance of the proposed method is verified by the equipment health state recognition task. A gear reducer test platform is constructed, which consists of an alternating current motor, a gear reducer, an electromagnetic powder brake, a coupling and multiple sensors, and a vibration sensor is installed to collect running vibration signals. The platform configuration and sensor arrangement are shown in Figure 16 The experimental device uses MCC USB-1608FS-Plus data acquisition card and CT1050LC vibration accelerometer with a sampling frequency of 5 kHz, and the collected vibration data are shown in Figure 17

[0096] After excluding the initial running-in period data and intermediate shutdown records of the gear reducer, a total of 5000 degradation data are obtained. According to the previous research basis, the health state of the gear reducer is divided into: 653 health state (HS) samples, 2026 mild degradation (MiD) samples, 2134 moderate degradation (MoD) samples and 187 fault state (FS) samples. The experiment adopts a training-test division ratio of 9:1, and table 8 details the four limited labeled sample conditions of the training set division.

[0097] Table 8 Four limited labeled sample conditions Health state recognition performance: Table 9 shows the health state recognition results of MSMNEG, indicating that the proposed method can achieve consistent and excellent performance under all limited labeled sample conditions. It is worth noting that in the case of only 5% training samples, the recognition accuracy of MSMNEG reaches 97.51% (standard deviation is 0.53%), which confirms that the multi-sensor fusion features of the model can effectively capture the health state characteristics of the samples.

[0098] Table 9 Results of health state recognition task on GDD dataset ​To further demonstrate the effectiveness of the model, the proposed method was verified through comparative experiments under four limited labeled sample conditions. As shown in Table 10, the proposed MSMNEG method outperforms other methods in all limited sample scenarios, showing higher state recognition accuracy and lower standard deviation. Figure 18 Further analysis shows that MSMNEG exhibits excellent robustness under different sample conditions, mainly because it effectively captures the relationship between labeled samples and unlabeled samples, while establishing a true multi-sensor coupling relationship. Compared with traditional neural networks such as CNN, the sample-related graph structure of MSMNEG can achieve deeper information mining. In addition, compared with the standard GNN, MSMNEG more effectively considers the multi-sensor coupling relationship, thereby enhancing multi-sensor information fusion. In summary, the MSMNEG method can efficiently integrate multi-sensor information for different operation and maintenance (O&M) tasks, extract sample state features related to the task, and overcome the challenge of limited labeled sample availability.

[0099] Table 10 Health state recognition task results on GDD dataset Analysis shows that compared with other benchmark models, MSMNEG produces significantly fewer misclassified samples. This experimental evidence further confirms the excellent ability of MSMNEG in mining and fusing multi-sensor data, demonstrating its unique advantages in limited labeled sample scenarios.

[0100] This invention proposes a multi-sensor information fusion method based on multiple non-Euclidean graph sets for extracting features from industrial equipment under limited task labels. The proposed MSMNEG framework has three key advantages: First, it is designed for multiple tasks, enabling adaptive extraction of different equipment state features according to different needs in industrial fault diagnosis and health status identification. Second, it establishes effective associations between labeled and unlabeled samples across sensors to achieve label transfer, thus overcoming the challenge of limited labeled samples. Third, it employs a fully data-driven approach to model the explicit coupling relationships between multiple sensors, achieving optimized multi-sensor information fusion. Validation experiments on two fault diagnosis datasets and one health status identification dataset demonstrate that MSMNEG outperforms other multi-sensor fusion methods based on machine learning, traditional deep learning, and graph neural networks (GNNs), especially in multi-task and limited-label scenarios. With 5% labeled samples, MSMNEG achieves diagnostic accuracies of 96.21% and 99.90%, respectively, and a state identification accuracy of 98.58%. Additional ablation studies confirm the effectiveness of the model components. A limitation of current work is that the data-driven explicit coupling modeling is limited to static relationships. Future research should focus on dynamic multi-sensor coupling modeling under time-varying industrial conditions to further improve the performance of this method.

[0101] This invention proposes a system for feature extraction of industrial equipment, comprising: a data acquisition module, a single sensor processing module, a multi-sensor processing module, and a feature extraction module.

[0102] The system comprises the following modules: a data acquisition module for acquiring data from multiple sensors on an industrial device; a single-sensor processing module for extracting full-dimensional fusion features from samples in a single sensor data set, and constructing a sample association graph for the single sensor based on these full-dimensional fusion features by extracting proximity relationships between samples in the single sensor data set; and performing multi-expert feature representation acquisition and fusion based on the single-sensor sample association graph to obtain a single-sensor sample state representation. The full-dimensional fusion feature representation incorporates semantic information from various dimensions. The multi-sensor processing module is used to construct multi-sensor spatiotemporal graph data based on the time-series data of the full-dimensional fusion features of samples from all sensor data sets and the single-sensor sample association graph to generate coupling relationships between multiple sensors. It also constructs multi-sensor graph data through these coupling relationships, performs hierarchical adaptive graph coarsening on the multi-sensor graph data, and fuses multi-sensor information to obtain multi-sensor fusion features. The feature extraction module retains the sample state information in the single-sensor sample state representation and, under limited maintenance task label supervision, performs adaptive classification extraction on the multi-sensor fusion features to obtain equipment state features.

[0103] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of an industrial equipment feature extraction method.

[0104] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0105] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of an industrial equipment feature extraction method.

[0106] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0107] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. An industrial equipment feature extraction method, characterized by, The method comprises the following steps: acquiring a plurality of sensor data on an industrial equipment; extracting a full-dimensional fusion feature of a sample in single sensor data, and based on the full-dimensional fusion feature, constructing a sample correlation graph of a single sensor by extracting a proximity relationship between samples in single sensor data, and based on the sample correlation graph of the single sensor, obtaining and fusing multi-expert feature expression to obtain a single sensor sample state representation; the full-dimensional fusion feature represents semantic information spliced with various dimensional features; based on the time series data of the full-dimensional fusion feature of the sample in all sensor data and the sample correlation graph of the single sensor, constructing a multi-sensor spatio-temporal graph data to generate a coupling relationship between the multi-sensor, and constructing a multi-sensor graph data through the coupling relationship, and performing hierarchical adaptive graph coarsening on the multi-sensor graph data to fuse multi-sensor information and obtain a multi-sensor fusion feature; retaining sample state information in the single sensor sample state representation, and under the supervision of limited operation and maintenance task labels, performing adaptive classification extraction on the multi-sensor fusion feature to obtain an equipment state feature.

2. The industrial equipment feature extraction method of claim 1, wherein, The extraction of the full-dimensional fusion feature of the sample in the single sensor data is specifically: extracting particle dimensional features, local dimensional features and panoramic dimensional features from the sample signals of the single sensor data; for each dimensional feature of each sample signal, randomly replacing the feature values of each dimensional feature according to a binomial distribution, and respectively extracting semantic information of each dimensional feature through a self-attention mechanism, an attention mechanism and a convolutional neural network; splicing the semantic information of each dimensional feature to obtain the full-dimensional fusion feature of the sample, and decoding the full-dimensional fusion feature of the sample to obtain the final reconstructed full-dimensional fusion feature.

3. The industrial equipment feature extraction method of claim 1, wherein, Based on the full-dimensional fusion feature, the proximity relationship between samples in single sensor data is extracted to construct a sample correlation graph of a single sensor, specifically: based on the full-dimensional fusion feature, the Euclidean distance between the full-dimensional fusion features of each sample is calculated to obtain a K nearest neighbor set of each sample point, and based on the K nearest neighbor algorithm, the local structure of the nearest neighbor set is captured to obtain a sample local correlation structure matrix; the cosine similarity between sample feature pairs is calculated, and based on a similarity threshold, a similar node set of each sample point is obtained, and a sample global correlation structure matrix is constructed based on the similar node set; fuse the local correlation structure matrix and the global correlation structure matrix to obtain a comprehensive correlation matrix, and use a graph autoencoder to reconstruct the comprehensive correlation matrix, perform link prediction on the reconstructed correlation graph, and complete the missing edges to obtain an optimized sample correlation graph of a single sensor.

4. The industrial equipment feature extraction method of claim 1, wherein, Based on the sample correlation graph of the single sensor, multi-expert feature expression is obtained and fused to obtain a single sensor sample state representation, specifically: extracting a structure feature of a sample point from a topological structure of the sample correlation graph of the single sensor; extracting a multi-scale semantic feature of a sample point from the full-dimensional fusion feature using a multi-scale one-dimensional convolutional network; fuse the sample correlation graph between samples and the full-dimensional fusion feature of the sample itself using a graph convolutional network, and use a multi-receptive field enhancement strategy to aggregate neighborhood information of different order sample features to generate an enhanced representation, i.e., a structure-semantic hybrid feature; The different feature representations of the same sample at different positions of different experts are spliced, and a self-attention mechanism and a multi-layer perception are used for nonlinear deep fusion to obtain a single-sensor sample state representation.

5. The industrial equipment feature extraction method of claim 1, wherein, The time series data based on the full-dimensional fusion features of the samples in all sensor data and the sample correlation graph of a single sensor are used to construct a multi-sensor spatio-temporal graph data to generate a coupling relationship between multiple sensors, specifically as follows: A node set, an edge set, and time series data of node features of the sensor data are defined, wherein the edge set is initialized as a fully connected graph, and an adjacency matrix thereof represents the potential coupling relationship between sensors; The multi-sensor spatio-temporal graph data is constructed by using the node set, the edge set, and the time series data of node features and time; On the fully connected graph of the multi-sensor spatio-temporal graph data, a graph neural network is used, and based on the observed dynamic trajectories of the nodes, the pairwise interaction between the sensor nodes is predicted; Based on the pairwise interaction between the sensor nodes, a coupling structure inferrer running on the fully connected graph is constructed by using the graph neural network; The coupling graph and the time series data generated by the coupling structure inferrer are used to simulate the sensor system dynamics of the industrial equipment by using multiple parallel graph neural networks, predict the sensor system dynamics at the next time step, and update the coupling relationship between the sensors by minimizing the prediction error.

6. The industrial equipment feature extraction method of claim 1, wherein, The multi-sensor graph data is subjected to hierarchical adaptive graph coarsening, specifically as follows: When the multi-sensor sample graph data is subjected to the next graph coarsening, the node features are updated by message passing, wherein in the first layer of message passing, a multi-space graph neural network is used to process multiple initial coupling relationships; A learnable linear transformation parameterized by a weight matrix is used to generate a global graph content matrix for the current graph, and based on the global graph content matrix, the attention scores between the nodes in the previous coarsened graph and the clusters in the next target coarsened graph are calculated by using a cross-hierarchical attention mechanism; Based on the normalized attention scores, the coarsened node feature representation and the updated adjacency matrix are generated.

7. The industrial equipment feature extraction method of claim 1, wherein, The sample state information in the single-sensor sample state representation is retained, including the retention of single-sensor sample semantic information and the retention of the relationship information between samples in single-sensor data.

8. An industrial equipment feature extraction method system characterized by, It comprises: a data acquisition module for acquiring multiple sensor data on an industrial equipment; a single-sensor processing module for extracting full-dimensional fusion features of samples in single-sensor data, constructing a sample correlation graph of a single sensor by extracting the proximity relationship between samples in single-sensor data based on the full-dimensional fusion features, obtaining and fusing multi-expert feature expressions based on the sample correlation graph of the single sensor to obtain a single-sensor sample state representation; the full-dimensional fusion features represent spliced semantic information of features of various dimensions; a multi-sensor processing module for constructing a coupling relationship between multiple sensors based on time series data of full-dimensional fusion features of samples in all sensor data and a sample correlation graph of a single sensor, constructing a multi-sensor spatio-temporal graph data by the coupling relationship, and performing hierarchical adaptive graph coarsening on the multi-sensor graph data to fuse multi-sensor information and obtain a multi-sensor fusion feature; The feature extraction module is configured to retain sample state information in a single-sensor sample state representation, and to perform adaptive classification extraction on multi-sensor fusion features under limited operation and maintenance task label supervision, to obtain device state features.

9. A computer device, comprising: The industrial equipment feature extraction method comprises the steps of: obtaining a single-sensor sample state representation of an industrial equipment; and performing adaptive classification extraction on multi-sensor fusion features under limited operation and maintenance task label supervision, to obtain device state features.

10. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the steps of the industrial equipment feature extraction method according to any one of claims 1 to 7.

Images