An edge device degradation prediction system based on light graph convolution

Abstract

The application discloses an edge device degradation prediction system based on light graph convolution, comprising: a feature construction module for collecting data and constructing a light feature matrix; a trend analysis module for trend analysis and calculating a degradation trend correlation; a sign relationship construction module for generating a positive sign adjacency matrix and a negative sign adjacency matrix; a sign graph generation module for performing sparse processing to form a sign graph structure; a convolution input generation module for generating a sign convolution input feature representation; a sign convolution propagation module for generating a positive sign convolution result and a negative sign convolution result; a fusion feature generation module for fusion processing to generate a fusion feature representation; a degradation prediction module for generating a degradation prediction result of each device; and a dynamic update module for updating the light feature matrix and the sign graph structure and generating a new degradation prediction result. The application realizes device degradation prediction by using light graph convolution and sign graph update, and has the advantages of structure self-adaptation and prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of equipment health management and prediction technology, and in particular to an edge device degradation prediction system based on lightweight graph convolution. Background Technology

[0002] With the large-scale deployment of edge computing and IoT devices, the demand for real-time health monitoring and degradation prediction in industrial equipment, energy equipment, and complex distributed systems is constantly growing. Existing equipment degradation prediction methods mostly rely on traditional time series analysis or cloud-based deep model inference, typically focusing only on changes in single device characteristics and lacking effective modeling of the relationships between multiple devices, making it difficult to reflect the impact of changes in device operating states. Some studies have attempted to introduce graph structures to represent the connections between devices, but these often use fixed adjacency relationships, failing to express the differences in positive or negative degradation trends between devices.

[0003] Existing degradation prediction methods based on graph neural networks typically require significant computational resources and rely on fixed graph structures for convolutional propagation, making them unsuitable for operating environments with limited computing power at edge nodes and frequent data updates. Furthermore, the lack of differentiated processing for positive and negative correlations during graph convolution results in predictions that fail to effectively reflect the promoting or inhibiting relationships between devices, thus affecting the accuracy of degradation predictions. Existing methods also generally lack dynamic update mechanisms for the graph structure, making it impossible to maintain timely response and adaptability to changing operational metrics.

[0004] Therefore, how to provide an edge device degradation prediction system based on lightweight graph convolution is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an edge device degradation prediction system based on lightweight graph convolution. This invention uses lightweight graph convolution and symbolic graph updating to achieve device degradation prediction, which has the advantages of structural adaptability and accurate prediction.

[0006] An edge device degradation prediction system based on lightweight graph convolution according to an embodiment of the present invention includes: The feature construction module is used to collect operational index data from multiple devices and perform preprocessing to build a lightweight feature matrix; The trend analysis module performs trend analysis on the device feature sequence based on a preset time window and calculates the correlation of degradation trends. The symbol relation construction module generates positive and negative symbol adjacency matrices based on the correlation of degradation trends. The symbol graph generation module is used to perform sparsification processing to form a symbol graph structure; The convolution input generation module is used to apply lightweight graph convolution operators to lightweight feature matrices to generate symbolic convolution input feature representations; The symbolic convolution propagation module performs convolution propagation based on the symbolic graph structure to generate positive and negative symbolic convolution results. The fusion feature generation module is used to fuse the results of positive and negative convolutions to generate fusion feature representations. The degradation prediction module generates degradation prediction results for each device based on the fused feature representation; The dynamic update module is used to update the lightweight feature matrix and symbol graph structure, and generate new degradation prediction results based on the updated symbol graph structure to achieve dynamic prediction and real-time output of device degradation status.

[0007] Optionally, modules can be integrated using the following methods: Operational index data from multiple devices are collected at the edge. Normalization, noise suppression, and feature compression are performed on the collected operational index data to construct a lightweight feature matrix. Perform trend analysis on the feature sequences of each device in the lightweight feature matrix, and calculate the correlation of degradation trends between any two devices. Based on the correlation of degradation trends, determine the sets of positive and negative sign relations respectively, and generate positive and negative sign adjacency matrices; The adjacency matrices of positive and negative symbols are sparsified to form a symbol graph structure; A lightweight graph convolution operator is applied to a lightweight feature matrix to perform a convolution operation, generating a symbolic convolution input feature representation. The input feature representation of the symbolic convolution is propagated by positive and negative symbolic convolutions to generate positive and negative symbolic convolution results. The positive and negative convolution results are fused to obtain the fused feature representation of the device node; Degradation prediction results for each device are generated based on the fusion feature representation of device nodes; When new operational metric data is collected, the lightweight feature matrix and symbolic graph structure are updated, and symbolic convolution propagation is performed to generate new degradation prediction results, thereby realizing dynamic prediction and real-time output of equipment degradation status.

[0008] Optionally, the generation of the degradation trend correlation includes: Within a preset time window, extract time-series feature data corresponding to each device from the lightweight feature matrix, and arrange the feature data in the sampling order to form a device feature sequence. The numerical difference of the equipment feature sequence is calculated according to the order of adjacent sampling points to form a change sequence that reflects the changes in equipment status. Pair any two device change sequences at the same time position within a time window, multiply the change values ​​of the two change sequences at the corresponding time positions, and form a trend product sequence. The trend product sequence is summed within a time window, and the product results at each time position in the trend product sequence are added together to obtain the trend cumulative value. The cumulative trend value is normalized to obtain the degradation trend correlation within the preset range. The degradation trend correlation is used as the quantification result of the degradation trend of any two devices within the time window.

[0009] Optionally, the generation of the positive sign adjacency matrix and the negative sign adjacency matrix includes: The correlation of degradation trend is judged one by one for each device pair. Device pairs with a correlation of degradation trend greater than the first preset threshold are marked as positive sign device pairs, and device pairs with a correlation of degradation trend less than the second preset threshold are marked as negative sign device pairs. Write all positive sign relation device pairs into record locations with the same structure according to the correspondence between devices, forming a positive sign relation set; Write all negative sign relation device pairs into the same structured record positions according to the correspondence between devices to form a negative sign relation set; Construct a positive sign adjacency matrix based on the set of positive sign relations, and construct a negative sign adjacency matrix based on the set of negative sign relations; The positive sign adjacency matrix and the negative sign adjacency matrix are treated as two independent components of the symbolic graph structure.

[0010] Optionally, the generation of the symbolic graph structure includes: The matrix values ​​in the positive sign adjacency matrix are judged according to the third preset threshold. Matrix values ​​that are less than the third preset threshold are set to zero, thus forming a positive sign adjacency matrix that has been thresholded. The matrix values ​​in the negative sign adjacency matrix are judged according to the fourth preset threshold. Matrix values ​​greater than the fourth preset threshold are set to zero, forming a negative sign adjacency matrix after threshold processing. Perform row normalization on the thresholded positive sign adjacency matrix to form a sparse positive sign matrix representation; The thresholded negative sign adjacency matrix is ​​subjected to row normalization to form a sparse negative sign matrix representation; The sparse positive sign matrix representation and the sparse negative sign matrix representation are arranged in the order of the device nodes, and the sparse positive sign matrix representation and the sparse negative sign matrix representation together serve as the symbol graph structure.

[0011] Optionally, the generation of the symbolic convolutional input feature representation includes: Extract the feature values ​​corresponding to each device node from the lightweight feature matrix according to the arrangement order of the device nodes, and use the feature values ​​to form the initial convolution input sequence of the device nodes; Each feature value in the initial convolution input sequence is multiplied by the corresponding convolution weight value in the lightweight graph convolution operator to form a convolution weighted sequence; The convolution weighted sequence is summed to obtain the convolution aggregate value by adding all the weighted values ​​in the convolution weighted sequence; Normalize the convolution aggregate values ​​to form the convolution normalized result; Activation processing is performed on the convolution normalization result, and the convolution normalization result is input into a preset nonlinear activation function to obtain the activated convolution result; The activation convolution result is written into the symbolic convolution input feature representation at the position corresponding to the device node, forming the symbolic convolution input feature representation.

[0012] Optionally, the generation of the positive and negative sign convolution results includes: The positive sign adjacency matrix is ​​read sequentially according to the order of the device nodes. The positive connection values ​​corresponding to each device node are multiplied by the positive connection values ​​and the corresponding feature values ​​in the symbolic convolution input feature representation to form a positive sign convolution weighted sequence. The positive sign convolution weighted sequence is summed to obtain the positive sign convolution aggregate value by adding all the weighted values ​​in the sequence. Normalize the positive sign convolution aggregate value to form the convolution value of the corresponding device node in the positive sign convolution result; Write the convolution value of each device node in the positive sign convolution result to the corresponding position in the positive sign convolution result; The negative sign adjacency matrix is ​​read sequentially according to the order of the device nodes. The negative connection values ​​corresponding to each device node are multiplied by the negative connection values ​​and the corresponding feature values ​​in the symbolic convolution input feature representation to form a negative sign convolution weighted sequence. The negative sign convolution weighted sequence is summed to obtain the negative sign convolution aggregate value by adding all the weighted values ​​in the sequence. Normalize the aggregated values ​​of the negative sign convolution to form the convolution value of the corresponding device node in the negative sign convolution result; Write the convolution value corresponding to each device node in the negative sign convolution result to the corresponding position in the negative sign convolution result.

[0013] Optionally, the generation of the fused feature representation includes: According to the arrangement order of the device nodes, read the forward convolution value corresponding to each device node from the positive sign convolution result, and write the forward convolution value into the corresponding position of the forward sequence; According to the arrangement order of the device nodes, read the negative convolution value corresponding to each device node from the negative sign convolution result, and write the negative convolution value into the corresponding position of the negative sequence; Each positive convolution value in the positive sequence is added to the corresponding negative convolution value in the negative sequence to form a summed value sequence. Normalization is performed on each value in the summation sequence to form a normalized sequence. Activation is performed on each value by inputting each value into a preset nonlinear activation function to generate an activation sequence. Each activation value in the activation sequence is written into the corresponding device node position in the fusion feature representation to form the fusion feature representation.

[0014] Optionally, the generation of the degradation prediction results includes: According to the arrangement order of the device nodes, the fusion feature value corresponding to each device node is read from the fusion feature representation, and the fusion feature value is written into the corresponding position of the fusion feature sequence; Perform numerical transformation processing on each fusion feature value in the fusion feature sequence to generate a transformed value, and write the transformed value into the corresponding position in the transformation sequence; Normalization is performed on each transformed value in the transformed sequence to obtain a normalized numerical sequence. Activation processing is performed on each value in the normalized numerical sequence to generate an activation value, and the activation value is written to the corresponding position in the activation sequence. Perform a status judgment on each activation value in the activation sequence, mark activation values ​​that are greater than the set judgment threshold as degenerate state, and mark activation values ​​that are not greater than the judgment threshold as normal state; The state value corresponding to each device node is written into the corresponding position of the degradation prediction result in the order of the device nodes to obtain the degradation prediction result.

[0015] Optionally, the generation of the new predicted numerical sequence includes: When new operational indicator data is collected, the new operational indicator data is written into a new indicator data sequence according to the order in which the data arrives. Perform numerical processing on the new indicator data sequence, performing numerical calculations on each value in the new indicator data sequence according to a preset processing method to generate a new processed sequence; Write each value in the new sorted sequence into the position of the corresponding device node in the lightweight feature matrix to form the updated lightweight feature matrix; Based on the numerical changes in the updated lightweight feature matrix, a new structure update sequence is generated and written into the corresponding position of the symbolic graph structure to form the updated symbolic graph structure. The updated symbolic graph structure is subjected to symbolic convolutional propagation. The values ​​at each connection position in the updated symbolic graph structure are processed position-by-position with the values ​​at the corresponding positions in the updated lightweight feature matrix, and then aggregated to form a new propagation and aggregation sequence. Numerical transformation processing is performed on the new propagation aggregation sequence. Each value in the new propagation aggregation sequence is numerically calculated according to a preset transformation method to generate a new predicted value sequence. Each predicted value in the new predicted value sequence is written into the new degradation prediction result at the position corresponding to the device node.

[0016] The beneficial effects of this invention are: This invention preprocesses and lightweightly represents device operation index data at the edge, enabling graph structure modeling of device features with low computational cost, laying the foundation for subsequent convolution calculations and correlation analysis. Degradation trend correlations generated based on preset time windows effectively characterize the feature associations between devices over time. Furthermore, by constructing sets of positive and negative sign relationships using dynamic thresholding, the symbolic graph structure can simultaneously express both collaborative and inverse relationship features, thus reflecting the degradation patterns of device groups more realistically than traditional methods relying solely on static adjacency structures. By performing sparsification on the symbolic graph structure, redundant connections are reduced while maintaining key relationships, significantly lowering the computational load in lightweight graph convolution operations and enabling the system to operate stably under the limited computing power of edge devices.

[0017] This invention employs a lightweight graph convolution operator to perform convolution operations on the processed feature matrix, enabling the features of device nodes to be fully aggregated and transformed, providing a suitable input feature representation for symbolic convolution propagation. Based on this, symbolic convolution propagation is performed using both positive and negative symbolic adjacency matrices, enhancing feature information propagation along positive correlations and suppressing it along negative correlations, thus forming a dual-channel convolution result that simultaneously possesses both enhancement and suppression information. By fusing the positive and negative symbolic convolution results, the degradation behavior of device nodes can be characterized in a higher dimension, enabling the fused feature representation to possess correlation, contrast, and difference expression capabilities, significantly improving the accuracy of degradation prediction.

[0018] Furthermore, this invention achieves direct inference of the degradation state of individual devices by performing predictive transformation and state judgment on the fused feature representation, and dynamically links the prediction results with continuously updated operational indicator data. Whenever new operational indicator data is collected, the lightweight feature matrix and symbolic graph structure are updated synchronously, enabling symbolic convolution propagation and prediction generation to perform calculations based on the latest operational state, achieving continuous prediction and real-time output of the degradation state of the entire device group. Through this dynamic update mechanism, this invention is applicable to scenarios with frequently changing equipment operating conditions, freeing degradation prediction from reliance on static models or fixed graph structures, significantly improving the system's adaptability to environmental changes and individual device differences, ultimately achieving high-precision, low-latency degradation prediction results that can operate independently at the edge. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an edge device degradation prediction system based on lightweight graph convolution proposed in this invention; Figure 2 This is a schematic diagram of the symbolic convolution propagation process in an edge device degradation prediction system based on lightweight graph convolution proposed in this invention. Figure 3 This is a schematic diagram of a dynamic update mechanism in an edge device degradation prediction system based on lightweight graph convolution proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 An edge device degradation prediction system based on lightweight graph convolution, comprising: The feature construction module is used to collect operational index data from multiple devices and perform preprocessing to build a lightweight feature matrix; The trend analysis module performs trend analysis on the device feature sequence based on a preset time window and calculates the correlation of degradation trends. The symbol relation construction module generates positive and negative symbol adjacency matrices based on the correlation of degradation trends. The symbol graph generation module is used to perform sparsification processing to form a symbol graph structure; The convolution input generation module is used to apply lightweight graph convolution operators to lightweight feature matrices to generate symbolic convolution input feature representations; The symbolic convolution propagation module performs convolution propagation based on the symbolic graph structure to generate positive and negative symbolic convolution results. The fusion feature generation module is used to fuse the results of positive and negative convolutions to generate fusion feature representations. The degradation prediction module generates degradation prediction results for each device based on the fused feature representation; The dynamic update module is used to update the lightweight feature matrix and symbol graph structure, and generate new degradation prediction results based on the updated symbol graph structure to achieve dynamic prediction and real-time output of device degradation status.

[0022] In this embodiment, the modules are interconnected using the following method: Operational metrics data from multiple devices are collected at the edge. Normalization, noise suppression, and feature compression are performed on the collected operational metrics data to construct a lightweight feature matrix for graph convolution. Perform trend analysis on the feature sequences of each device in the lightweight feature matrix, and calculate the correlation of degradation trends between any two devices. Based on the correlation of degradation trends, determine the sets of positive and negative sign relations respectively, and generate positive and negative sign adjacency matrices; The adjacency matrices of positive and negative symbols are sparsified to form a symbol graph structure; A lightweight graph convolution operator is applied to a lightweight feature matrix to perform a convolution operation, generating a symbolic convolution input feature representation. The input feature representation of the symbolic convolution is propagated by positive and negative symbolic convolutions to generate positive and negative symbolic convolution results. The positive and negative convolution results are fused to obtain the fused feature representation of the device node; Degradation prediction results for each device are generated based on the fusion feature representation of device nodes; When new operational metric data is collected, the lightweight feature matrix and symbolic graph structure are updated, and symbolic convolution propagation is performed to generate new degradation prediction results, thereby realizing dynamic prediction and real-time output of equipment degradation status.

[0023] In this embodiment, the generation of the degradation trend correlation includes: Within a preset time window, extract time-series feature data corresponding to each device from the lightweight feature matrix, and arrange the feature data in the sampling order to form a device feature sequence. The numerical difference of the equipment feature sequence is calculated according to the order of adjacent sampling points. The feature value of the later sampling point is subtracted from the feature value of the previous sampling point to form a change sequence that reflects the change of equipment status. Pair any two device change sequences at the same time position within a time window, multiply the change values ​​of the two change sequences at the corresponding time positions to form a trend product sequence for trend calculation; The trend product sequence is summed within a time window, and the product results at each time position in the trend product sequence are added together to obtain the trend cumulative value. The trend accumulation value is normalized and divided by the sum of the absolute values ​​of all changes in the two change sequences within the time window to obtain the degradation trend correlation degree that falls within the preset range. The degradation trend correlation degree is used as the degradation trend quantification result of any two devices within the time window to characterize the consistency of the characteristic change direction and change magnitude of the two devices within the time window.

[0024] In this embodiment, the generation of the positive sign adjacency matrix and the negative sign adjacency matrix includes: The correlation of degradation trend is judged one by one for each device pair. Device pairs with a correlation of degradation trend greater than the first preset threshold are marked as positive sign device pairs, and device pairs with a correlation of degradation trend less than the second preset threshold are marked as negative sign device pairs. The positive and negative sign relationship is defined as follows: if the correlation of the degradation trends of two devices within a time window is greater than a first threshold, then the sign relationship between the two devices is defined as a positive sign relationship, corresponding to the positive edge in the sign graph; if the correlation of the degradation trends of two devices is less than a second threshold, then it is defined as a negative sign relationship, corresponding to the negative edge in the sign graph. The first threshold and the second threshold are dynamically generated based on the statistical characteristics of the operating indicator data. The statistical characteristics of the operating indicator data include the mean, standard deviation, range, or quantile of the correlation between the degradation trends of any two devices within a preset time window. The first threshold is the value after the mean and standard deviation of the trend correlation are weighted by a preset coefficient, and the second threshold is the value after the mean and standard deviation of the trend correlation are weighted by a preset coefficient. Write all positive sign relation device pairs into record locations with the same structure according to the correspondence between devices, forming a positive sign relation set; Write all negative sign relation device pairs into the same structured record positions according to the correspondence between devices to form a negative sign relation set; Construct a positive sign adjacency matrix based on the positive sign relationship set. Write the matrix value used to identify the positive sign relationship for each device pair in the positive sign relationship set at the corresponding position in the matrix, and write the matrix value used to identify the non-positive sign relationship for each device pair that does not belong to the positive sign relationship set at the corresponding position in the matrix. Construct a negative sign adjacency matrix based on the negative sign relationship set. Write the matrix value used to identify the negative sign relationship for each device pair in the negative sign relationship set at the corresponding position in the matrix. Write the matrix value used to identify the non-negative sign relationship for each device pair that does not belong to the negative sign relationship set at the corresponding position in the matrix. The positive and negative sign adjacency matrices are used as two independent components of the sign graph structure to characterize the positive and negative sign relationships formed between any two devices based on the correlation of degradation trends.

[0025] In this embodiment, the generation of the symbolic graph structure includes: The matrix values ​​in the positive sign adjacency matrix are judged according to the third preset threshold. Matrix values ​​that are less than the third preset threshold are set to zero, thus forming a positive sign adjacency matrix that has been thresholded. The dynamic third threshold is generated based on the statistical characteristics of all matrix values ​​in the positive sign adjacency matrix. The statistical characteristics include the mean and standard deviation of the matrix values. The dynamic third threshold is the value of the mean and standard deviation of the matrix values ​​weighted by a preset coefficient. The matrix values ​​in the negative sign adjacency matrix are judged according to the fourth preset threshold. Matrix values ​​greater than the fourth preset threshold are set to zero, forming a negative sign adjacency matrix after threshold processing. The dynamic fourth threshold is generated based on the statistical characteristics of all matrix values ​​in the negative sign adjacency matrix. The statistical characteristics include the mean and standard deviation of the matrix values. The dynamic fourth threshold is the value of the mean and standard deviation of the matrix values ​​weighted by another preset coefficient. Perform row normalization on the thresholded positive sign adjacency matrix, dividing the non-zero matrix value of each row by the sum of all non-zero matrix values ​​in that row to form a sparse positive sign matrix representation. Perform row normalization on the thresholded negative sign adjacency matrix, dividing the non-zero matrix value of each row by the sum of all non-zero matrix values ​​in that row to form a sparse negative sign matrix representation. The sparse positive sign matrix representation and the sparse negative sign matrix representation are arranged in the order of device nodes to ensure that the two sparse matrices maintain consistency in index position. The sparse positive sign matrix representation and the sparse negative sign matrix representation together serve as a symbol graph structure to represent the positive and negative sign connection relationships after sparsification.

[0026] In this embodiment, the generation of the symbolic convolutional input feature representation includes: Extract the feature values ​​corresponding to each device node from the lightweight feature matrix according to the arrangement order of the device nodes, and use the feature values ​​to form the initial convolution input sequence of the device nodes; Each feature value in the initial convolution input sequence is multiplied by the corresponding convolution weight value in the lightweight graph convolution operator to form a convolution weighted sequence; The convolution weighted sequence is summed to obtain the convolution aggregate value by adding all the weighted values ​​in the convolution weighted sequence; Normalize the convolution aggregate value by dividing the convolution aggregate value by the total number of all feature values ​​in the initial convolution input sequence to form the convolution normalization result. Activation processing is performed on the convolution normalization result, and the convolution normalization result is input into a preset nonlinear activation function to obtain the activated convolution result; The activation convolution result is written into the position corresponding to the device node in the symbolic convolution input feature representation, so that the symbolic convolution input feature representation is composed of the activation convolution values ​​corresponding to all device nodes, thus forming the symbolic convolution input feature representation.

[0027] In this embodiment, the generation of the positive sign convolution result and the negative sign convolution result includes: The positive sign adjacency matrix is ​​read sequentially according to the order of the device nodes. The positive connection values ​​corresponding to each device node are multiplied by the positive connection values ​​and the corresponding feature values ​​in the symbolic convolution input feature representation to form a positive sign convolution weighted sequence. The positive sign convolution weighted sequence is summed to obtain the positive sign convolution aggregate value by adding all the weighted values ​​in the sequence. Normalize the positive sign convolution aggregate value by dividing the positive sign convolution aggregate value by the number of all non-zero connection values ​​of the corresponding device node in the positive sign adjacency matrix to form the convolution value of the corresponding device node in the positive sign convolution result. Write the convolution value of each device node in the positive sign convolution result to the corresponding position in the positive sign convolution result, so that the positive sign convolution result is consistent with the order of the device nodes; The negative sign adjacency matrix is ​​read sequentially according to the order of the device nodes. The negative connection values ​​corresponding to each device node are multiplied by the negative connection values ​​and the corresponding feature values ​​in the symbolic convolution input feature representation to form a negative sign convolution weighted sequence. The negative sign convolution weighted sequence is summed to obtain the negative sign convolution aggregate value by adding all the weighted values ​​in the sequence. Normalize the negative sign convolution aggregate value by dividing the negative sign convolution aggregate value by the number of all non-zero connection values ​​of the corresponding device node in the negative sign adjacency matrix to form the convolution value of the corresponding device node in the negative sign convolution result. Write the convolution value corresponding to each device node in the negative sign convolution result to the corresponding position in the negative sign convolution result, so that the negative sign convolution result is consistent with the device node order.

[0028] In this embodiment, the generation of the fusion feature representation includes: According to the arrangement order of the device nodes, read the forward convolution value corresponding to each device node from the positive sign convolution result, and write the forward convolution value into the corresponding position of the forward sequence; According to the arrangement order of the device nodes, read the negative convolution value corresponding to each device node from the negative sign convolution result, and write the negative convolution value into the corresponding position of the negative sequence; Each positive convolution value in the positive sequence is added to the corresponding negative convolution value in the negative sequence to form a summed value sequence. Normalization is performed on each value in the summed value sequence by dividing the value by a fixed normalization factor to form a normalized sequence. Activation processing is performed on each value in the normalized sequence, and each value is input into a preset nonlinear activation function to generate an activation sequence. Each activation value in the activation sequence is written into the position of the corresponding device node in the fusion feature representation to form the fusion feature representation, so that each position in the fusion feature representation corresponds one-to-one with the device node.

[0029] In this embodiment, the generation of the degradation prediction result includes: According to the arrangement order of the device nodes, the fusion feature value corresponding to each device node is read from the fusion feature representation, and the fusion feature value is written into the corresponding position of the fusion feature sequence; Numerical transformation processing is performed on each fusion feature value in the fusion feature sequence. Each fusion feature value is multiplied by a preset transformation coefficient to generate a transformation value, and the transformation value is written into the corresponding position in the transformation sequence. Normalization is performed on each transformed value in the transformation sequence by dividing each transformed value by a fixed normalization factor to obtain a normalized value sequence. Activation processing is performed on each value in the normalized numerical sequence. Each value is input into a preset activation function to generate an activation value, and the activation value is written to the corresponding position in the activation sequence. Perform a status judgment on each activation value in the activation sequence, mark activation values ​​that are greater than the set judgment threshold as degenerate state, and mark activation values ​​that are not greater than the judgment threshold as normal state; The state value corresponding to each device node is written into the corresponding position of the degradation prediction result in the order of the device nodes, so that the degradation prediction result is composed of the state values ​​corresponding to all device nodes, thus obtaining the degradation prediction result.

[0030] In this embodiment, the generation of the new predicted numerical sequence includes: When new operational indicator data is collected, the operational indicator data is data that can reflect the operating status of the equipment. The new operational indicator data is written into a new indicator data sequence according to the order of data arrival, so that each position of the new indicator data sequence corresponds one-to-one with the corresponding equipment node. The operational indicator data is collected and continuously updated in real time by the edge side, forming a feature sequence that changes over time, and serves as input for trend analysis and symbol graph updates; Perform numerical processing on the new indicator data sequence, performing numerical calculations on each value in the new indicator data sequence according to a preset processing method to generate a new processed sequence; Write each value in the new sorted sequence into the position of the corresponding device node in the lightweight feature matrix to form the updated lightweight feature matrix; Based on the numerical changes in the updated lightweight feature matrix, a new structure update sequence is generated and written into the corresponding position of the symbolic graph structure to form the updated symbolic graph structure. The updated symbolic graph structure is subjected to symbolic convolutional propagation. The values ​​at each connection position in the updated symbolic graph structure are processed position-by-position with the values ​​at the corresponding positions in the updated lightweight feature matrix, and then aggregated to form a new propagation and aggregation sequence. Numerical transformation processing is performed on the new propagation aggregation sequence. The numerical transformation processing is used to transform the fusion feature or propagation aggregation sequence into a predicted value that can be used for degradation state determination in a linear mapping manner. Each value in the new propagation aggregation sequence is numerically calculated according to a preset transformation method to generate a new predicted value sequence. Each predicted value in the new predicted value sequence is written into the position corresponding to the device node in the new degradation prediction result. The new degradation prediction result is composed of the predicted values ​​corresponding to all device nodes, so that the degradation prediction result can continuously output new prediction results as the operation index data is continuously updated.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a smart manufacturing park in a coastal region to predict the degradation of the operating status of various types of equipment, including motors, conveyor structures, and actuators, on multiple production lines. This park operates in a high-humidity, high-salt environment, making equipment prone to varying degrees of wear and performance degradation. Furthermore, the large fluctuations in production tasks cause various equipment to exhibit dynamically changing operating modes over time. In such a complex environment, traditional degradation prediction methods relying on fixed graph structures or single correlation analysis often struggle to track real-time changes in the relationships between equipment, leading to decreased model prediction accuracy and failing to meet the practical requirements of low latency and high stability for edge devices. This invention provides a solution to these problems.

[0032] During use, edge nodes continuously collect operational data from various devices, including vibration, rotational speed, load, temperature rise, and motion change information. This data undergoes normalization, noise suppression, and feature compression via the preprocessing mechanism of this invention, transforming it into a lightweight feature matrix. Since these preprocessing operations are performed directly at the edge, the amount of data is effectively reduced, improving transmission and processing efficiency and laying the foundation for subsequent graph structure construction and convolutional propagation.

[0033] The trend analysis mechanism of this invention is based on a sliding time window. By calculating the characteristic differences between adjacent sampling points and aggregating the trend products of the change sequences of different devices, a degradation trend correlation degree that reflects the consistency of changes among devices is obtained. In a real production environment, some devices may exhibit a significant positive correlation due to sharing the load, while other devices may exhibit a negative correlation due to the regulating effect. When the resistance of the upstream motor increases, the load of the transmission device also increases, and the two show a high positive correlation. However, when certain actuators start up, they cause the speed of other devices to decrease. This phenomenon is represented as a negative relationship in this invention. The trend correlation degree is the core quantitative indicator used to characterize such relationships.

[0034] After obtaining the correlation of degradation trends, this invention constructs a set of positive and negative sign relationships using dynamic threshold determination, and then generates positive and negative sign adjacency matrices. Traditional methods often rely on fixed thresholds, which are not easily adapted to fluctuations in operating conditions. In contrast, this invention generates dynamic thresholds by utilizing the statistical characteristics of correlation within a time window, ensuring that the sign relationship determination remains stable across different load stages and does not shift due to increases or decreases in the overall level of change.

[0035] After constructing the symbolic adjacency matrix, this invention performs sparsification on the symbolic graph, removing weakly weighted connections to make the remaining graph structure more prominent in terms of key relationships, while reducing the computational burden of convolutional propagation. In industrial park scenarios with a large number of devices, without sparsification, the graph structure would be extremely dense, making it difficult to operate on edge nodes. This invention, through the dual effects of dynamic thresholding and sparsification, enables the symbolic graph to adapt to changing operating conditions while maintaining computational lightweightness.

[0036] This invention then uses a lightweight graph convolution operator to generate a symbolic convolution input feature representation on the lightweight feature matrix. Compared to traditional convolution structures, the lightweight graph convolution operator is more concise, has a lower parameter scale, and can execute quickly under the limited computing power of edge devices. Subsequently, the symbolic convolution input feature representation is fed into positive and negative symbolic convolution propagation paths, respectively. Positive symbolic convolution propagation is used to aggregate positive relation features from the positive symbolic adjacency matrix, strengthening the features of devices with a consistent degradation trend; negative symbolic convolution propagation is used to aggregate negative relation features from the negative symbolic adjacency matrix, thereby suppressing or highlighting important information between devices with adversarial relationships. The final positive and negative convolution features are fused to form a fused feature representation, whose comprehensive expressive power is significantly better than single-channel features.

[0037] In the prediction phase of this invention, the fused feature representation is converted into the degradation state of each device. During operation in the park, the system repeatedly predicted the degradation trends of certain devices in advance, including insufficient lubrication due to long-term frictional heat in the bearings of a certain conveyor device, and operational deviations caused by aging seals in a certain pneumatic actuator. These predictions were verified during subsequent inspections by maintenance personnel, proving that this invention possesses early warning capabilities and can effectively reduce the risk of equipment downtime.

[0038] More importantly, this invention possesses continuous update capabilities. When the production line status changes or equipment parameters fluctuate significantly, the system automatically updates the lightweight feature matrix and symbolic graph structure upon receiving new operational index data, and re-executes symbolic convolutional propagation and prediction output based on the updated graph structure. The entire process requires no manual intervention or model reconstruction, maintaining predictive capability in real time according to actual operating conditions. This characteristic is particularly evident during multiple production task switches; regardless of day shifts, night shifts, or different production line combinations, the system maintains the stability and consistency of prediction results.

[0039] Through the above implementation process, this invention demonstrates excellent stability, adaptability, and predictive accuracy in edge environments. It effectively solves the problems of traditional methods, such as their inability to reflect dynamic changes in relationships between devices, high computational burden, and unsuitability for edge deployment. This makes device operation status management more intelligent, real-time, and forward-looking, meeting the urgent needs of modern industrial scenarios for device health management.

[0040] Table 1. Performance Comparison between Lightweight Graph Convolution-Based Edge Device Degradation Prediction System and Traditional Methods

[0041] Experimental results show that the system of this invention has a significant advantage in prediction accuracy. The prediction accuracy of the traditional threshold correlation method is 82%, and that of the traditional static graph convolution method is 87%, while the prediction accuracy of this invention is improved to 92%, representing improvements of 10% and 5%, respectively. This indicates that this invention, through techniques such as trend product sequence calculation, positive and negative sign relationship partitioning, and dual-path sign convolution propagation, enables a finer-grained expression of degradation correlations between devices, thereby achieving more accurate degradation prediction in complex operating environments.

[0042] In terms of real-time performance, this invention also demonstrates outstanding performance. The processing latency of traditional thresholding methods is 220 milliseconds, and that of traditional static graph convolution methods is 180 milliseconds, while the latency of this invention is only 95 milliseconds, representing a 57% and 47% improvement in processing speed, respectively. This improvement is mainly due to the combination of lightweight graph convolution operators and sparse symbolic graph structures, making inference computation more compact; at the same time, the dynamic update mechanism avoids the redundant overhead caused by the frequent reconstruction of the model or recalculation of the full graph structure required by traditional methods, shortening the response time at the edge.

[0043] In terms of edge resource utilization, this invention exhibits significant lightweight characteristics. Traditional thresholding methods have a CPU utilization rate of 68%, and traditional static graph convolution has a rate of 55%, while this invention only has a rate of 34%. CPU resource load is reduced by 50% and 38%, respectively. Because this invention uses a lightweight feature matrix, a sparse adjacency structure, and simplified convolution operations, it can run continuously and stably on edge computing nodes with limited computing power, without interfering with the control tasks of field devices, resulting in higher operational reliability.

[0044] Regarding model update capabilities, this invention establishes a symbolic graph structure that automatically updates as operational metric data changes. Traditional thresholding methods require manual resetting of threshold rules, and traditional static graph convolution methods, due to their fixed adjacency matrix structure, struggle to reflect the dynamics of device relationships. In contrast, this invention, through real-time trend analysis, dynamic threshold generation, and adaptive symbolic graph updates, enables the model to automatically update the graph structure and convolution propagation process upon the arrival of new data, thereby continuously outputting the latest prediction results and achieving continuity in degradation identification.

[0045] This invention excels in anomaly detection lead time. Traditional thresholding methods provide a lead time of 0.5 hours, and traditional static graph convolution methods provide 1.2 hours, while this invention can output degradation trend warnings 3.4 hours in advance. This is more than 6.8 times earlier than traditional thresholding methods and approximately 2.8 times earlier than static graph convolution methods. This is because this invention uses a positive and negative sign relationship to express the heterogeneity of device degradation and strengthens weak signals through trend multiplication, enabling the system to capture potential signs of device state changes much earlier.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An edge device degradation prediction system based on lightweight graph convolution, characterized in that, include: The feature construction module is used to collect operational index data from multiple devices and perform preprocessing to build a lightweight feature matrix; The trend analysis module performs trend analysis on the device feature sequence based on a preset time window and calculates the correlation of degradation trends. The symbol relation construction module generates positive and negative symbol adjacency matrices based on the correlation of degradation trends. The symbol graph generation module is used to perform sparsification processing to form a symbol graph structure; The convolution input generation module is used to apply lightweight graph convolution operators to lightweight feature matrices to generate symbolic convolution input feature representations; The symbolic convolution propagation module performs convolution propagation based on the symbolic graph structure to generate positive and negative symbolic convolution results. The fusion feature generation module is used to fuse the results of positive and negative convolutions to generate fusion feature representations. The degradation prediction module generates degradation prediction results for each device based on the fused feature representation; The dynamic update module is used to update the lightweight feature matrix and symbol graph structure, and generate new degradation prediction results based on the updated symbol graph structure to achieve dynamic prediction and real-time output of device degradation status. The generation of the degradation trend correlation includes: Within a preset time window, extract time-series feature data corresponding to each device from the lightweight feature matrix, and arrange the feature data in the sampling order to form a device feature sequence. The numerical difference of the equipment feature sequence is calculated according to the order of adjacent sampling points to form a change sequence that reflects the changes in equipment status. Pair any two device change sequences at the same time position within a time window, multiply the change values ​​of the two change sequences at the corresponding time positions, and form a trend product sequence. The trend product sequence is summed within a time window, and the product results at each time position in the trend product sequence are added together to obtain the trend cumulative value. The cumulative trend value is normalized to obtain the degradation trend correlation within the preset range. The degradation trend correlation is used as the quantification result of the degradation trend of any two devices within the time window. The generation of the positive sign adjacency matrix and the negative sign adjacency matrix includes: The correlation of degradation trend is judged one by one for each device pair. Device pairs with a correlation of degradation trend greater than the first preset threshold are marked as positive sign device pairs, and device pairs with a correlation of degradation trend less than the second preset threshold are marked as negative sign device pairs. Write all positive sign relation device pairs into record locations with the same structure according to the correspondence between devices, forming a positive sign relation set; Write all negative sign relation device pairs into the same structured record positions according to the correspondence between devices to form a negative sign relation set; Construct a positive sign adjacency matrix based on the set of positive sign relations, and construct a negative sign adjacency matrix based on the set of negative sign relations; The positive sign adjacency matrix and the negative sign adjacency matrix are treated as two independent components of the symbol graph structure; The generation of the positive and negative sign convolution results includes: The positive sign adjacency matrix is ​​read sequentially according to the order of the device nodes. The positive connection values ​​corresponding to each device node are multiplied by the positive connection values ​​and the corresponding feature values ​​in the symbolic convolution input feature representation to form a positive sign convolution weighted sequence. The positive sign convolution weighted sequence is summed to obtain the positive sign convolution aggregate value by adding all the weighted values ​​in the sequence. Normalize the positive sign convolution aggregate value to form the convolution value of the corresponding device node in the positive sign convolution result; Write the convolution value of each device node in the positive sign convolution result to the corresponding position in the positive sign convolution result; The negative sign adjacency matrix is ​​read sequentially according to the order of the device nodes. The negative connection values ​​corresponding to each device node are multiplied by the negative connection values ​​and the corresponding feature values ​​in the symbolic convolution input feature representation to form a negative sign convolution weighted sequence. The negative sign convolution weighted sequence is summed to obtain the negative sign convolution aggregate value by adding all the weighted values ​​in the sequence. Normalize the aggregated values ​​of the negative sign convolution to form the convolution value of the corresponding device node in the negative sign convolution result; Write the convolution value corresponding to each device node in the negative sign convolution result to the corresponding position in the negative sign convolution result.

2. The edge device degradation prediction system based on lightweight graph convolution according to claim 1, characterized in that, The modules are connected in the following way: Operational index data from multiple devices are collected at the edge. Normalization, noise suppression, and feature compression are performed on the collected operational index data to construct a lightweight feature matrix. Perform trend analysis on the feature sequences of each device in the lightweight feature matrix, and calculate the correlation of degradation trends between any two devices. Based on the correlation of degradation trends, determine the sets of positive and negative sign relations respectively, and generate positive and negative sign adjacency matrices; The adjacency matrices of positive and negative symbols are sparsified to form a symbol graph structure; A lightweight graph convolution operator is applied to a lightweight feature matrix to perform a convolution operation, generating a symbolic convolution input feature representation. The input feature representation of the symbolic convolution is propagated by positive and negative symbolic convolutions to generate positive and negative symbolic convolution results. The positive and negative convolution results are fused to obtain the fused feature representation of the device node; Degradation prediction results for each device are generated based on the fusion feature representation of device nodes; When new operational metric data is collected, the lightweight feature matrix and symbolic graph structure are updated, and symbolic convolution propagation is performed to generate new degradation prediction results, thereby realizing dynamic prediction and real-time output of equipment degradation status.

3. The edge device degradation prediction system based on lightweight graph convolution according to claim 2, characterized in that, The generation of the symbolic graph structure includes: The matrix values ​​in the positive sign adjacency matrix are judged according to the third preset threshold. Matrix values ​​that are less than the third preset threshold are set to zero, thus forming a positive sign adjacency matrix that has been thresholded. The matrix values ​​in the negative sign adjacency matrix are judged according to the fourth preset threshold. Matrix values ​​greater than the fourth preset threshold are set to zero, forming a negative sign adjacency matrix after threshold processing. Perform row normalization on the thresholded positive sign adjacency matrix to form a sparse positive sign matrix representation; The thresholded negative sign adjacency matrix is ​​subjected to row normalization to form a sparse negative sign matrix representation; The sparse positive sign matrix representation and the sparse negative sign matrix representation are arranged in the order of the device nodes, and the sparse positive sign matrix representation and the sparse negative sign matrix representation together serve as the symbol graph structure.

4. The edge device degradation prediction system based on lightweight graph convolution according to claim 2, characterized in that, The generation of the symbolic convolution input feature representation includes: Extract the feature values ​​corresponding to each device node from the lightweight feature matrix according to the arrangement order of the device nodes, and use the feature values ​​to form the initial convolution input sequence of the device nodes; Each feature value in the initial convolution input sequence is multiplied by the corresponding convolution weight value in the lightweight graph convolution operator to form a convolution weighted sequence; The convolution weighted sequence is summed to obtain the convolution aggregate value by adding all the weighted values ​​in the convolution weighted sequence; Normalize the convolution aggregate values ​​to form a convolution normalization result. Activate the convolution normalization result and input the convolution normalization result into a preset nonlinear activation function to obtain the activated convolution result. The activation convolution result is written into the symbolic convolution input feature representation at the position corresponding to the device node, forming the symbolic convolution input feature representation.

5. The edge device degradation prediction system based on lightweight graph convolution according to claim 2, characterized in that, The generation of the fusion feature representation includes: According to the arrangement order of the device nodes, read the forward convolution value corresponding to each device node from the positive sign convolution result, and write the forward convolution value into the corresponding position of the forward sequence; According to the arrangement order of the device nodes, read the negative convolution value corresponding to each device node from the negative sign convolution result, and write the negative convolution value into the corresponding position of the negative sequence; Each positive convolution value in the positive sequence is added to the corresponding negative convolution value in the negative sequence to form a summed value sequence. Normalization is performed on each value in the summation sequence to form a normalized sequence. Activation is performed on each value by inputting each value into a preset nonlinear activation function to generate an activation sequence. Each activation value in the activation sequence is written into the corresponding device node position in the fusion feature representation to form the fusion feature representation.

6. The edge device degradation prediction system based on lightweight graph convolution according to claim 2, characterized in that, The generation of the degradation prediction results includes: According to the arrangement order of the device nodes, the fusion feature value corresponding to each device node is read from the fusion feature representation, and the fusion feature value is written into the corresponding position of the fusion feature sequence; Perform numerical transformation processing on each fusion feature value in the fusion feature sequence to generate a transformed value, and write the transformed value into the corresponding position in the transformation sequence; Normalization is performed on each transformed value in the transformed sequence to obtain a normalized numerical sequence. Activation processing is performed on each value in the normalized numerical sequence to generate an activation value, and the activation value is written to the corresponding position in the activation sequence. Perform a status judgment on each activation value in the activation sequence, mark activation values ​​that are greater than the set judgment threshold as degenerate state, and mark activation values ​​that are not greater than the judgment threshold as normal state; The state value corresponding to each device node is written into the corresponding position of the degradation prediction result in the order of the device nodes to obtain the degradation prediction result.

7. The edge device degradation prediction system based on lightweight graph convolution according to claim 2, characterized in that, The generation of the new degradation prediction results includes: When new operational indicator data is collected, the new operational indicator data is written into a new indicator data sequence according to the order in which the data arrives. Perform numerical processing on the new indicator data sequence, performing numerical calculations on each value in the new indicator data sequence according to a preset processing method to generate a new processed sequence; Write each value in the new sorted sequence into the position of the corresponding device node in the lightweight feature matrix to form the updated lightweight feature matrix; Based on the numerical changes in the updated lightweight feature matrix, a new structure update sequence is generated and written into the corresponding position of the symbolic graph structure to form the updated symbolic graph structure. The updated symbolic graph structure is subjected to symbolic convolutional propagation. The values ​​at each connection position in the updated symbolic graph structure are processed position-by-position with the values ​​at the corresponding positions in the updated lightweight feature matrix, and then aggregated to form a new propagation and aggregation sequence. Numerical transformation processing is performed on the new propagation aggregation sequence. Each value in the new propagation aggregation sequence is numerically calculated according to a preset transformation method. The numerically calculated values ​​are then arranged according to the correspondence between device nodes to generate new degradation prediction results.

Images