Intelligent sensing method and system based on graph structure construction, and apparatus and medium

By optimizing the graph structure through a bi-branch graph structure and an uneven edge weighting strategy, the feature extraction and representation capabilities of graph convolution are improved, solving the problem of insufficient utilization of semantic differences between nodes in graph neural networks, and achieving more efficient multi-view signal classification and perception.

WO2026232498A1PCT designated stage Publication Date: 2026-11-12SHENZHEN TECH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/093821
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-05-09
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

In existing technologies, graph neural networks cannot fully utilize the semantic or structural differences between nodes when constructing graph structures, resulting in poor data perception performance.

Method used

A bi-branch graph structure is adopted, and time-frequency graph and time-domain graph are constructed through continuous wavelet transform. The graph structure is optimized by using k-dimensional trees, and non-uniform edge weighting strategies are introduced, such as 0-1 weighting, negative distance exponential weighting, cosine similarity weighting, and Gaussian kernel weighting, to improve the feature extraction effect of graph convolution.

Benefits of technology

It improves the feature extraction effect and expressive power of graph convolution, and enhances the classification and perception accuracy of multi-view signals, making it suitable for equipment operation status perception in intelligent manufacturing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025093821_12112026_PF_FP_ABST
    Figure CN2025093821_12112026_PF_FP_ABST
Patent Text Reader

Abstract

An intelligent sensing method and system based on graph structure construction, and an apparatus and a medium. The method comprises: acquiring a one-dimensional raw vibration signal; performing continuous wavelet transform on the raw vibration signal, in order to obtain a two-dimensional time-frequency image, and using a k-dimensional tree to establish a time-domain graph structure for the raw vibration signal; performing feature extraction on the two-dimensional time-frequency image, in order to obtain a first feature vector, and performing feature extraction on the time-domain graph structure, in order to obtain a second feature vector; fusing the first feature vector and the second feature vector, in order to obtain a fused vector; and performing classification on the fused vector, in order to obtain a determined category for the raw vibration signal. The present invention can solve the technical problem in the prior art of the data sensing effect requiring further improvement.
Need to check novelty before this filing date? Find Prior Art

Description

Intelligent sensing methods, systems, devices, and media based on graph structures. Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to an intelligent perception method, system, device, and medium based on graph structure construction. Background Technology

[0002] In actual manufacturing systems, the vibration signals collected by equipment possess rich temporal and frequency characteristics. Although a single physical signal is single-mode, it can be transformed using different mathematical methods to obtain multiple feature perspectives (such as the original time domain, frequency domain, and time-frequency plots). Joint modeling of these signals helps to more comprehensively characterize the equipment's operating state. However, these heterogeneous representations differ in dimensionality, structure, and spatial distribution, making it difficult for traditional feature extraction methods to take into account the inherent relationships between various features.

[0003] To address the aforementioned issues, Graph Neural Networks (GNNs) have become an important tool for complex data perception in recent years due to their advantages in processing non-Euclidean structured data. However, in practical applications, the quality of graph construction directly determines the feature extraction performance of GCNs (Graph Convolutional Networks). Most existing methods employ fixed adjacency or simple distance rules, failing to fully utilize the semantic or structural differences between nodes, thus the effectiveness of data perception needs improvement. Summary of the Invention

[0004] The main objective of this invention is to provide an intelligent sensing method, system, device, and medium based on graph structure construction, aiming to solve the technical problem that the effect of data sensing in the prior art needs to be improved.

[0005] To achieve the above objectives, the first aspect of the present invention provides an intelligent sensing method for constructing a graph structure, comprising: acquiring a one-dimensional original vibration signal; performing continuous wavelet transform on the original vibration signal to obtain a two-dimensional time-frequency image; and establishing a time-domain graph structure for the original vibration signal using a k-dimensional tree; extracting features from the two-dimensional time-frequency image to obtain a first feature vector, and extracting features from the time-domain graph structure to obtain a second feature vector; fusing the first feature vector and the second feature vector to obtain a fused vector; and classifying the fused vector to obtain a classification of the original vibration signal.

[0006] Further, the step of extracting features from the two-dimensional time-frequency image to obtain a first feature vector includes: treating the two-dimensional time-frequency image as a grid graph, treating each pixel in the grid graph as a node in the graph, establishing edge connections between adjacent pixels, and constructing an undirected graph G = (V, E), where V is the set of pixels, E is the set of connecting edges, and the feature of the node is X. i =(R i G i B i ), representing the RGB three-channel pixel values, and the edge weight w between nodes. ij =||X i -X j ||; Perform a 2-layer ChebNet convolution on the undirected graph to obtain the first feature vector.

[0007] Furthermore, the step of establishing a time-domain graph structure for the original vibration signal using a k-dimensional tree includes: treating each sampling point of the graph structure as a node in the graph, and establishing edge links between the node and its five nearest neighbors to obtain the time-domain graph structure.

[0008] Further, the step of extracting features from the temporal graph structure to obtain a second feature vector includes: performing a 2-layer GraphConv convolution on the temporal graph structure to extract temporal structural features and obtain a second feature vector.

[0009] Furthermore, the step of extracting features from the temporal graph structure to obtain a second feature vector further includes: during the process of extracting temporal structure features by performing a 2-layer GraphConv convolution on the temporal graph structure, applying non-uniform edge weighting to the convolution result to dynamically adjust the edge weights based on the feature differences and distances between different nodes. The non-uniform edge weighting includes at least one of the following: 0-1 weighting, negative distance exponential weighting, cosine similarity weighting, hot kernel weighting, and Gaussian kernel weighting.

[0010] Furthermore, after applying non-uniform edge weighting to the convolution result, during the process of extracting temporal structural features from the temporal graph structure using a 2-layer GraphConv convolution, the convolution calculation formula includes:

[0011] in, The adjacency matrix is ​​calculated based on a weighted strategy. W is the degree matrix corresponding to the adjacency matrix. (l) It is a learnable weight parameter matrix.

[0012] Furthermore, both the first feature vector and the second feature vector are 64-dimensional graph feature vectors; the fused vector is a 128-dimensional feature vector; and when classifying the fused vector, a fully connected layer is used for classification.

[0013] A second aspect of the present invention also provides an intelligent sensing system based on a graph structure, comprising: a signal acquisition module for acquiring a one-dimensional original vibration signal; a signal processing module for performing continuous wavelet transform on the original vibration signal to obtain a two-dimensional time-frequency image; and establishing a time-domain graph structure for the original vibration signal using a k-dimensional tree; a feature extraction module for extracting features from the two-dimensional time-frequency image to obtain a first feature vector, and extracting features from the time-domain graph structure to obtain a second feature vector; a feature fusion module for fusing the first feature vector and the second feature vector to obtain a fused vector; and a feature classification module for classifying the fused vector to obtain a classification of the original vibration signal.

[0014] A third aspect of the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any one of the above-described intelligent sensing methods based on graph structure construction.

[0015] A fourth aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the intelligent sensing method based on graph structure construction described in any one of the above.

[0016] The intelligent sensing method, system, device, and medium based on graph structure construction provided by this invention have the following advantages: by constructing a bi-branch graph structure, graph models are constructed for the original time-domain signal and its time-frequency image after continuous wavelet transform, and K-dimensional trees are used to optimize the modeling of the graph structure, thereby improving the feature extraction effect and expressive ability of graph convolution, and ultimately improving the effect of feature fusion to complete the classification and sensing task. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 is a flowchart of the intelligent perception method based on graph structure construction provided in an embodiment of this application;

[0019] Figure 2 is a model architecture diagram of the intelligent perception method based on graph structure construction provided in the embodiments of this application;

[0020] Figure 3 is a graph showing the training loss versus training accuracy of the intelligent perception method based on graph structure construction provided in the embodiments of this application.

[0021] Figure 4 is a classification confusion matrix diagram of the intelligent perception method based on graph structure construction provided in the embodiments of this application;

[0022] Figure 5 shows the feature distribution map of the intelligent perception method based on graph structure construction provided in the embodiments of this application, which is visualized by dimensionality reduction through t-distribution random neighborhood embedding after 0-1 weighting.

[0023] Figure 6 shows the feature distribution map of the graph-based intelligent perception method provided in this application embodiment, after Euclidean distance weighting and dimensionality reduction visualization through t-distribution random neighborhood embedding;

[0024] Figure 7 shows the feature distribution map of the intelligent perception method based on graph structure construction provided in the embodiments of this application, which is visualized by dimensionality reduction through t-distribution random neighborhood embedding after distance negative exponential weighting.

[0025] Figure 8 shows the feature distribution map of the intelligent perception method based on graph structure construction provided in the embodiments of this application, which is visualized by dimensionality reduction through t-distribution random neighborhood embedding after cosine similarity weighting.

[0026] Figure 9 is a feature distribution map visualized by t-distribution random neighborhood embedding after hot kernel weighting of the intelligent perception method based on graph structure construction provided in the embodiments of this application;

[0027] Figure 10 is a feature distribution map visualized by dimensionality reduction through t-distribution random neighborhood embedding after Gaussian weighting of the intelligent perception method based on graph structure construction provided in the embodiments of this application;

[0028] Figure 11 is a framework diagram of an intelligent sensing system based on graph structure provided in an embodiment of this application;

[0029] Figure 12 is a schematic block diagram of the electronic device provided in the embodiment of this application. Detailed Implementation

[0030] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0031] In actual manufacturing systems, the vibration signals collected by equipment possess rich temporal and frequency characteristics. Although a single physical signal is single-mode, it can be transformed using different mathematical methods to obtain multiple feature perspectives (such as the original time domain, frequency domain, and time-frequency plots). Joint modeling of these signals helps to more comprehensively characterize the equipment's operating state. However, these heterogeneous representations differ in dimensionality, structure, and spatial distribution, making it difficult for traditional feature extraction methods to take into account the inherent relationships between various features.

[0032] In recent years, Graph Neural Networks (GNNs) have become an important tool for understanding complex data due to their advantages in processing non-Euclidean structured data. However, in practical applications, the quality of graph construction directly determines the feature extraction performance of GNNs. Most existing methods use fixed adjacency or simple distance rules, which cannot fully utilize the semantic or structural differences between nodes.

[0033] To address the aforementioned issues, embodiments of the present invention provide an intelligent sensing method, system, device, and medium based on graph structure construction, which can improve the data sensing effect.

[0034] Please refer to Figure 1, which illustrates an intelligent perception method based on graph structure construction, including:

[0035] S101. Obtain the one-dimensional original vibration signal;

[0036] S102. Perform continuous wavelet transform on the original vibration signal to obtain a two-dimensional time-frequency image; and use a k-dimensional tree to establish a time-domain graph structure for the original vibration signal.

[0037] S103. Perform feature extraction on the two-dimensional time-frequency image to obtain the first feature vector, and perform feature extraction on the time-domain image structure to obtain the second feature vector;

[0038] S104. Merge the first feature vector and the second feature vector to obtain the fused vector;

[0039] S105. Classify the fused vector to obtain the classification of the original vibration signal.

[0040] In step S101, the original vibration signal is the vibration signal collected by the equipment in the actual manufacturing system.

[0041] In steps S102-S104, the time-domain graph structure is a time-series signal. To fully extract deep features from the modal information, this embodiment adopts a dual-branch graph structure processing method: one branch processes the two-dimensional time-frequency image generated by continuous wavelet transform, and the other branch processes the original one-dimensional vibration signal. Both branches use graph neural networks to perform graph convolution to extract features, obtaining a first feature vector and a second feature vector. Both the first and second feature vectors are 64-dimensional graph feature vectors; the fused vector is a 128-dimensional feature vector.

[0042] Therefore, the intelligent perception method based on graph structure construction provided in this embodiment constructs a bi-branch graph structure to build graph models for the original time-domain signal and its time-frequency image after continuous wavelet transform, and uses a K-dimensional tree to optimize the graph structure model, thereby improving the feature extraction effect and expressive power of graph convolution, and ultimately improving the effect of feature fusion to complete the classification perception task.

[0043] In some embodiments, the step of extracting features from the two-dimensional time-frequency image to obtain a first feature vector includes: treating the two-dimensional time-frequency image as a mesh graph, treating each pixel in the mesh graph as a node in the graph, establishing edge connections between adjacent pixels, and constructing an undirected graph G = (V, E), where V is the set of pixels, E is the set of connecting edges, and the feature of each node is X. i =(R i G i B i ), representing the RGB three-channel pixel values, and the edge weight w between nodes. ij =||X i -X j ||; Perform a 2-layer ChebNet convolution on the undirected graph to obtain the first feature vector.

[0044] In this embodiment, the original vibration signal is transformed into a two-dimensional time-frequency image through continuous wavelet transform. The image can be viewed as a regular grid diagram, where each pixel is a node in the graph, and edges are established between adjacent pixels to construct an undirected graph G = (V, E), where V is the set of pixels and E is the set of connecting edges. Node features X i =(R i G i B i ), representing RGB three-channel pixel values, with edge weights between nodes being w. ij =||X i -X j ||, defined based on the similarity between pixels. This branch goes through 2 layers of ChebNet convolution, with the feature dimension increasing from 3 to 128 to 64.

[0045] In some embodiments, the step of using a k-dimensional tree to establish a time-domain graph structure for the original vibration signal includes: treating each sampling point of the graph structure as a node in the graph, and establishing edge links between the node and its five nearest neighbors to obtain the time-domain graph structure.

[0046] In some embodiments, the step of extracting features from the temporal graph structure to obtain a second feature vector includes: performing a 2-layer GraphConv convolution on the temporal graph structure to extract temporal structural features and obtain a second feature vector.

[0047] In this embodiment, a graph structure is constructed using a KD-Tree (k-dimensional tree) for the original one-dimensional vibration signal. Each sampling point is considered a node in the graph, and edges are established to connect it to its five nearest neighbors. The constructed graph structure is input into the improved GCN model, and temporal structural features are extracted through two layers of GraphConv convolutions, with the convolutional layer configuration being 1024→256→64.

[0048] In some embodiments, the step of extracting features from the temporal graph structure to obtain a second feature vector further includes: during the process of extracting temporal structure features by performing a 2-layer GraphConv convolution on the temporal graph structure, applying non-uniform edge weighting to the convolution result to dynamically adjust the edge weights based on feature differences and distances between different nodes, wherein the non-uniform edge weighting includes at least one of: 0-1 weighting, negative distance exponential weighting, cosine similarity weighting, hot kernel weighting, and Gaussian kernel weighting.

[0049] To improve graph convolution performance, this embodiment introduces a non-uniform edge weighting strategy, enabling GCN to dynamically adjust edge weights based on feature differences and distances between nodes, thereby enhancing its ability to model heterogeneous graph structures. Common edge weighting methods are as follows:

[0050] (1) 0-1 weighted: Only considers whether the edge exists. It is simple to calculate but ignores the distance.

[0051] (2) Euclidean distance weighted: Similarity is represented by the geometric distance between nodes, and the expression is:

[0052] (3) Distance-negative exponential weighting: Weights decay for distant nodes, expressed as follows:

[0053] (4) Cosine similarity weighting: This method considers node feature similarity and is suitable for high-dimensional data. The expression is:

[0054] (5) Thermonuclear weighting: Smooths distance variations and reduces noise interference; the expression is:

[0055] (6) Gaussian kernel weighting: Considering the Gaussian decay characteristics of distance, this improves modeling capabilities. The expression is:

[0056] In some embodiments, after applying non-uniform edge weighting to the convolution result, during the process of extracting temporal structural features by performing a 2-layer GraphConv convolution on the temporal graph structure, the convolution calculation formula includes:

[0057] in, The adjacency matrix is ​​calculated based on a weighted strategy. W is the degree matrix corresponding to the adjacency matrix. (l) It is a learnable weight parameter matrix.

[0058] The 64-dimensional feature vectors output from the two branches are concatenated to form a 128-dimensional fused feature. This fused feature is then input into a fully connected layer to ultimately achieve classification and discrimination for multi-view intelligent perception tasks.

[0059] Therefore, this invention introduces a KD-Tree nearest neighbor structure and various edge weight design strategies (such as Euclidean distance weighting, cosine similarity, and hot kernel functions) to construct a more expressive graph structure, thereby improving the modeling performance of graph convolution under heterogeneous feature representation. Although this method does not change the core computational framework of graph convolution, it achieves an overall improvement in perception effect through graph modeling optimization, making it suitable for multi-view modeling and intelligent recognition scenarios for signals from the same source.

[0060] Furthermore, the method provided in this embodiment achieves a higher accuracy than traditional feature extraction methods in multiple typical working condition recognition experiments, demonstrating good engineering applicability and suitability for the perception and analysis of equipment operating status in intelligent manufacturing.

[0061] To verify the above claims, this application also conducted practical verification, as detailed below:

[0062] (1) Dataset partitioning: The dataset is divided into training set, validation set and test set according to a ratio of 7:2:1;

[0063] (2) GCN model training: Train two GCN branch models respectively to learn multi-view features;

[0064] (3) Feature fusion and classification: The two branches of features are concatenated and input into the classifier to complete multi-class discrimination;

[0065] (4) Model evaluation: Validate using the test set and output the recognition accuracy, confusion matrix and weighted average performance index.

[0066] The model is shown in Figure 2.

[0067] To verify the effectiveness of the graph-based intelligent sensing method provided in this invention in multi-condition bearing fault diagnosis, a publicly available bearing vibration dataset was selected for functional verification.

[0068] The dataset used comes from the Electrical Engineering Laboratory at Case Western Reserve University, and includes four typical data subsets: normal state, drive-end bearing failure, fan-end bearing failure, and operating data at different speeds. This invention selects "12k Drive End Bearing Fault Data" as the research object and sets three different speed conditions (1797 rpm, 1772 rpm, and 1750 rpm, denoted as conditions A, B, and C). Each condition includes four bearing states: inner race failure, outer race failure, rolling element failure, and normal state. Different failure degrees (fault diameters of 0 mils, 7 mils, 14 mils, and 21 mils) are assigned to each fault state, forming multiple fault level sample sets, as shown in Table 1.

[0069] Table 1 Description of bearing failure dataset

[0070] The original vibration signal was copied into two branches and input into the GCN diagnostic model according to the above method. The features were then fused. The number of training samples, validation samples and test samples were 1631, 466 and 233 respectively. The model training loss, training accuracy, validation loss and validation accuracy curves are shown in Figure 3.

[0071] As shown in the figure, after 100 epochs of training, the training loss is close to 0, and the training accuracy reaches nearly 100%, with no underfitting. After training, to evaluate the fault classification ability of the GCN model, the test set was input into the trained model. The confusion matrix between the classification results and the classification accuracy is shown in Figure 4, where C1-C10 represent ball 1-3 / inner 1-3 / health / outer 1-3. Based on the accuracy: Precision = TP / (TP+FP)

[0072] Where TP represents the number of samples correctly predicted by the model, and FP represents the number of samples incorrectly predicted. Recall can be calculated as: Recall = TP / (TP + FN)

[0073] FP represents the rows of samples with incorrect predictions. The weighted average of each category is calculated as: F1-score = 2 * Precision / Recall

[0074] The classification accuracy in the figure can be calculated to be 95.5%. As can be seen from the figure, the diagnostic model makes basically correct predictions for the 10 types of fault criteria.

[0075] Different weighting methods were used to assess the model's classification accuracy. Each method was performed three times, divided into groups A, B, and C. The three groups were then averaged to obtain the final classification accuracy, as shown in Table 2. The weighting methods included 0-1 weighting, Euclidean distance, negative distance exponent, cosine similarity, and Gaussian kernel function. In one-dimensional bearing data, each one-dimensional signal represents a node, where a and b represent two different nodes. Different edge weighting methods were used to measure the relationship between the two nodes. The table shows that compared to the original edge structure Euclidean distance, 0-1 weighting, Gaussian kernel function, heat kernel, and negative distance exponent weighting negatively impacted the diagnostic results. Cosine similarity improved the diagnostic accuracy. This simplifies the graph construction to some extent, eliminates the need for advanced prior knowledge, reduces the influence of human experience, and achieves more efficient end-to-end fault diagnosis.

[0076] Table 2 Accuracy of Fault Classification with Different Weights

[0077] The results show that the graph structure modeled using Euclidean distance outperforms other methods in classification accuracy, achieving an average accuracy of 94.41%. Furthermore, Figures 5-10 illustrate the feature distribution extracted by the model under different weighting methods through T-SNE dimensionality reduction visualization. Euclidean distance and cosine similarity enhance class discriminability and increase feature clustering, further demonstrating that the improved graph structure significantly improves classification performance.

[0078] In summary, this embodiment verifies the robustness and effectiveness of the graph structure construction method described in this invention under multi-condition fault data scenarios, and is particularly suitable for multi-physical domain signal modeling and intelligent diagnosis tasks based on graph structures.

[0079] Please refer to Figure 11. This application embodiment also provides an intelligent sensing system based on graph structure construction, including: a signal acquisition module 1, used to acquire a one-dimensional original vibration signal; a signal processing module 2, used to perform continuous wavelet transform on the original vibration signal to obtain a two-dimensional time-frequency image; and to establish a time-domain graph structure for the original vibration signal using a k-dimensional tree; a feature extraction module 3, used to extract features from the two-dimensional time-frequency image to obtain a first feature vector, and to extract features from the time-domain graph structure to obtain a second feature vector; a feature fusion module 4, used to fuse the first feature vector and the second feature vector to obtain a fused vector; and a feature classification module 5, used to classify the fused vector to obtain a classification judgment of the original vibration signal.

[0080] The intelligent perception system based on graph structure provided in this embodiment constructs a bi-branch graph structure to build graph models for the original time-domain signal and its time-frequency image after continuous wavelet transform, and uses a K-dimensional tree to optimize the modeling of the graph structure, thereby improving the feature extraction effect and expressive power of graph convolution, and ultimately improving the effect of feature fusion to complete the classification perception task.

[0081] In some embodiments, the feature extraction module 3 includes an undirected graph construction unit, configured to treat the two-dimensional time-frequency image as a mesh graph, treat each pixel in the mesh graph as a node in the graph, establish edge connections between adjacent pixels, and construct an undirected graph G = (V, E), where V is the set of pixels, E is the set of connecting edges, and the feature of the node is X. i =(R i G i B i ), representing the RGB three-channel pixel values, and the edge weight w between nodes. ij =||X i -X j ||;The first convolutional unit is used to perform two layers of ChebNet convolution on the undirected graph to obtain the first feature vector.

[0082] In some embodiments, the signal processing module 2 is specifically used to treat each sampling point of the graph structure as a node in the graph, and to establish edge links between the node and its five nearest neighbors to obtain a time-domain graph structure.

[0083] In some embodiments, the feature extraction module 3 further includes a second convolution unit, used to perform a 2-layer GraphConv convolution on the temporal graph structure to extract temporal structural features and obtain a second feature vector.

[0084] In some embodiments, the feature extraction module 3 further includes: a non-uniform edge weighting unit, used to perform non-uniform edge weighting on the convolution result during the process of the second convolution unit performing 2-layer GraphConv convolution to extract temporal structural features of the temporal graph structure, so as to dynamically adjust the edge weights based on the feature differences and distances between different nodes, wherein the non-uniform edge weighting includes at least one of: 0-1 weighting, negative distance exponential weighting, cosine similarity weighting, hot kernel weighting, and Gaussian kernel weighting.

[0085] In some embodiments, after applying non-uniform edge weighting to the convolution result, during the process of extracting temporal structural features by performing a 2-layer GraphConv convolution on the temporal graph structure, the convolution calculation formula includes:

[0086] in, The adjacency matrix is ​​calculated based on a weighted strategy. W is the degree matrix corresponding to the adjacency matrix. (l) It is a learnable weight parameter matrix.

[0087] In some embodiments, the first feature vector and the second feature vector are both 64-dimensional graph feature vectors; the fused vector is a 128-dimensional feature vector; and when classifying the fused vector, a fully connected layer is used for classification.

[0088] This application embodiment also provides an electronic device, as shown in FIG12. The electronic device includes: a memory 601, a processor 602, and a computer program stored in the memory 601 and executable on the processor 602. When the processor 602 executes the computer program, it implements the intelligent sensing method based on graph structure construction described above.

[0089] Furthermore, the electronic device also includes at least one input device 603 and at least one output device 604.

[0090] The aforementioned memory 601, processor 602, input device 603, and output device 604 are connected via bus 605.

[0091] The input device 603 can specifically be a camera, touch panel, physical buttons, or mouse, etc. The output device 604 can specifically be a display screen.

[0092] The memory 601 can be a high-speed random access memory (RAM) or a non-volatile memory, such as a disk storage device. The memory 601 is used to store a set of executable program code, and the processor 602 is coupled to the memory 601.

[0093] Furthermore, this application embodiment also provides a computer-readable storage medium, which may be disposed in the electronic device of the above embodiments, and the computer-readable storage medium may be the aforementioned memory 601. The computer-readable storage medium stores a computer program, which, when executed by the processor 602, implements the graph-based intelligent sensing method described in the foregoing embodiments.

[0094] Furthermore, the storage medium of the computer can also be a USB flash drive, a portable hard drive, a read-only memory (ROM), RAM, a magnetic disk, or an optical disk, or any other medium that can store program code.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0096] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0097] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0098] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0099] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0100] The above is a description of an intelligent sensing method, system, electronic device, and storage medium based on graph structure construction provided by the present invention. For those skilled in the art, based on the ideas of the embodiments of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An intelligent sensing method based on graph structure construction, characterized in that, include: Obtain the original one-dimensional vibration signal; The original vibration signal is subjected to continuous wavelet transform to obtain a two-dimensional time-frequency image; and a k-dimensional tree is used to establish a time-domain graph structure for the original vibration signal. Feature extraction is performed on the two-dimensional time-frequency image to obtain a first feature vector, and feature extraction is performed on the time-domain graph structure to obtain a second feature vector; The first feature vector and the second feature vector are fused to obtain a fused vector; The fused vector is classified to obtain the classification of the original vibration signal.

2. The intelligent sensing method based on graph structure construction according to claim 1, characterized in that, The step of extracting features from the two-dimensional time-frequency image to obtain a first feature vector includes: The two-dimensional time-frequency image is used as a mesh graph, and each pixel in the mesh graph is treated as a node in the graph. Edges are established between adjacent pixels to construct an undirected graph G = (V, E), where V is the set of pixels, E is the set of connecting edges, and the node is characterized by X. i =(R i G i B i ), representing the RGB three-channel pixel values, and the edge weight w between nodes. ij =||X i -X j ||; The undirected graph is subjected to two layers of ChebNet convolution to obtain the first feature vector.

3. The intelligent sensing method based on graph structure construction according to claim 1, characterized in that, The process of establishing a time-domain graph structure for the original vibration signal using a k-dimensional tree includes: Each sampling point of the graph structure is treated as a node in the graph, and an edge link is established between the node and its five nearest neighbors to obtain the time-domain graph structure.

4. The intelligent sensing method based on graph structure construction according to claim 1, characterized in that, The step of extracting features from the time-domain graph structure to obtain a second feature vector includes: The temporal graph structure is subjected to a 2-layer GraphConv convolution to extract temporal structural features, resulting in a second feature vector.

5. The intelligent sensing method based on graph structure construction according to claim 4, characterized in that, The step of extracting features from the time-domain graph structure to obtain the second feature vector further includes: In the process of extracting temporal structural features by performing a 2-layer GraphConv convolution on the temporal graph structure, the convolution result is weighted unevenly to dynamically adjust the edge weights based on the feature differences and distances between different nodes. The uneven edge weighting includes at least one of the following: 0-1 weighting, negative distance exponential weighting, cosine similarity weighting, hot kernel weighting, and Gaussian kernel weighting.

6. The intelligent sensing method based on graph structure construction according to claim 5, characterized in that, After applying non-uniform edge weighting to the convolution result, during the process of extracting temporal structural features from the temporal graph structure using a 2-layer GraphConv convolution, the convolution calculation formula includes: in, The adjacency matrix is ​​calculated based on a weighted strategy. W is the degree matrix corresponding to the adjacency matrix. (l) It is a learnable weight parameter matrix.

7. The intelligent sensing method based on graph structure construction according to claim 1, characterized in that, Both the first feature vector and the second feature vector are 64-dimensional graph feature vectors; The fusion vector is a 128-dimensional feature vector; When classifying the fused vector, a fully connected layer is used for classification.

8. An intelligent sensing system based on graph structure, characterized in that, include: The signal acquisition module is used to acquire one-dimensional raw vibration signals; The signal processing module is used to perform continuous wavelet transform on the original vibration signal to obtain a two-dimensional time-frequency image; and to use a k-dimensional tree to establish a time-domain graph structure for the original vibration signal. The feature extraction module is used to extract features from the two-dimensional time-frequency image to obtain a first feature vector, and to extract features from the time-domain graph structure to obtain a second feature vector; The feature fusion module is used to fuse the first feature vector and the second feature vector to obtain a fused vector; The feature classification module is used to classify the fused vector to obtain the classification of the original vibration signal.

9. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 7.