An edge-computing-based production line equipment fault detection method

Abstract

This invention discloses a fault detection method for production line equipment based on edge computing, comprising: collecting multimodal sensing data and preprocessing it to obtain standardized multimodal sensing data; extracting product defect parameters and operational anomaly indicators to obtain the node anomaly intensity of each equipment node; selecting abnormal nodes to form an anomaly source node set and constructing a directed graph of the equipment; constructing a fault propagation edge weight dynamic control module to form a weighted directed graph; executing an improved Bellman-Ford algorithm to obtain the minimum risk propagation cost, risk propagation path, and cascading fault risk loop; and determining key risk nodes and the potential fault propagation range. This invention, by constructing a fault propagation edge weight dynamic control module and combining it with an improved Bellman-Ford algorithm, achieves early identification of production line equipment anomalies and accurate prediction and early warning of cascading faults.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method for detecting equipment faults in production lines based on edge computing. Background Technology

[0002] With the deepening of intelligent manufacturing and industrial automation, production lines are gradually shifting from single-machine control to multi-device collaborative control. Equipment operating status monitoring, processing quality inspection, and process execution are all incorporated into the real-time processing system of edge nodes to support continuous coordination of production tasks and production line cycle control. Traditional production line equipment fault detection technologies mainly rely on single types of data, such as vibration, temperature, current, load fluctuations, or single-view visual inspection, and often employ static threshold judgment and rule-based analysis strategies. Limited by data representation capabilities and methodological design ideas, it is difficult to effectively integrate multi-source heterogeneous data and fully characterize the dynamic correlation characteristics between different devices, resulting in the real-time performance and accuracy of equipment anomaly identification failing to meet the production quality control requirements under complex operating conditions. Furthermore, local anomaly detection modes ignore the coupling effect of processing dependencies, material transfer, and execution coordination between devices on the operating status, easily leading to problems such as missed detections, false judgments, and delayed alarms.

[0003] Existing technologies attempt to introduce graph structure analysis or statistical feature learning to characterize the relationships between production line equipment. However, these often employ static or averaged dependency structure settings, failing to introduce differentiated characterizations of equipment process coupling strength, buffer isolation capabilities, and material conduction characteristics. Furthermore, they do not fully consider the dynamic interactions between quality anomalies and operational anomalies. Existing edge-side inference schemes lack proactive inference capabilities in propagation path calculation, risk quantification, and cascading fault identification. They struggle to proactively assess potential propagation risks and cannot meet the practical needs of real-time perception, correlation inference, and propagation range identification of multi-node cascading risks in an edge computing environment.

[0004] Therefore, how to provide a fault detection method for production line equipment based on edge computing 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 a fault detection method for production line equipment based on edge computing. This invention utilizes multimodal sensing data fusion analysis, process dependency graph construction, and risk propagation path inference technology to describe in detail the entire process of identifying anomalies in multiple production line devices, analyzing related propagation, and predicting cascading risks at the edge node side. The method extracts product surface defect parameters and equipment operation anomaly indicators in real time at the edge side, constructs a multi-source fusion node anomaly intensity score, and then generates a weighted directed graph of equipment by combining the production line process sequence, material transfer, and control linkage relationships. A fault propagation edge weight dynamic adjustment module is constructed to perform structured modeling of risk transmission between devices. An improved Bellman-Ford algorithm is used to iteratively infer the minimum risk propagation path and identify cascading fault risk loops at the edge node side, achieving real-time determination of the potential impact range, key risk nodes, and propagation links. This method has advantages such as real-time inference from edge deployment, collaborative judgment of anomalies and quality, and quantifiable analysis of propagation risks. It can effectively improve the operational safety and quality consistency of the production line and provide interpretable risk prediction capabilities for multi-device collaborative monitoring.

[0006] A method for detecting production line equipment faults based on edge computing according to an embodiment of the present invention includes: Collect multimodal sensing data from edge computing nodes of production line equipment, preprocess the multimodal sensing data to obtain standardized multimodal sensing data; Based on standardized multimodal perception data, product defect parameters and operational anomaly indicators are extracted respectively to generate product defect anomaly scores and equipment operational anomaly scores, and then fused according to preset fusion rules to obtain the node anomaly intensity of each equipment node. Based on the strength of node anomalies, anomaly nodes are selected to form a set of anomaly source nodes, and a directed graph of equipment is constructed by combining the process dependencies between production line equipment. A fault propagation edge weight dynamic control module is constructed to generate propagation risk cost edge weights for each directed edge in the device directed graph, forming a weighted directed graph; Based on weighted directed graphs, an improved Bellman-Ford algorithm is used to iteratively calculate path costs, obtain the minimum risk propagation cost from each anomaly source node to the remaining nodes and the corresponding risk propagation path, and identify cascading fault risk loops in the weighted directed graph. Based on the minimum risk propagation cost, risk propagation path, and cascading failure risk loop, key risk nodes and potential failure propagation range are identified.

[0007] Optionally, the multimodal sensing data specifically includes equipment operating status data, product quality surface data, process data, and inter-equipment communication and timing data.

[0008] Optionally, the preprocessing of the multimodal sensing data specifically includes time synchronization, sampling rate alignment, noise suppression, scale standardization, and missing value completion.

[0009] Optionally, the step of fusing the node anomaly strength of each device node according to a preset fusion rule includes: Based on product quality surface data in standardized multimodal perception data, defect candidate regions of product surface are extracted, and the area, perimeter, shape features, edge sharpness and brightness variation of defect candidate regions are quantitatively statistically analyzed to obtain the feature parameter set of defect candidate regions. Based on the set of feature parameters, the product defect anomaly score is calculated according to the area, shape complexity, edge sharpness, brightness difference, and texture anomaly of the defect candidate region. Based on the equipment operation status data in the standardized multimodal sensing data, abnormal indicators are identified, and the vibration change, temperature change, load fluctuation amplitude, current deviation and operation stability of the equipment are measured, weighted and summarized to generate an equipment operation anomaly score. Based on the preset weighted fusion rules, the product defect anomaly score and the equipment operation anomaly score are fused to obtain the node anomaly intensity of each equipment node.

[0010] Optionally, the construction of the directed graph of equipment by combining the process dependencies between production line equipment includes: Consistency detection and amplitude smoothing are performed on the abnormal intensity of nodes, and the results are compared with the abnormal intensity threshold. Device nodes whose abnormal intensity exceeds the abnormal intensity threshold are marked as abnormal nodes. The marked abnormal nodes are identified and their node numbers and corresponding abnormality strengths are recorded to form an abnormal source node data set. Based on the process data and inter-equipment communication and timing data in the standardized multimodal sensing data, the sequential processing relationship, material flow relationship and linkage control relationship between equipment nodes in the production line are determined, and the dependency direction between equipment nodes is obtained. A directed graph of devices is constructed based on the dependency directions between device nodes, and the set of abnormal source nodes is written into the corresponding device node attributes to form a directed graph of devices containing abnormal node identifiers.

[0011] Optionally, forming a weighted directed graph includes: A fault propagation edge weight dynamic control module is constructed, which consists of an adjacency topology mapping unit, a state anomaly scoring quantification unit, an edge weight calculation unit, and a dynamic update control unit. The adjacency topology mapping unit reads the directed graph of the device and obtains the directed edges in the directed graph of the device and the corresponding upstream and downstream device node information. It then combines the standardized multimodal sensing data to construct a composite graph structure corresponding to the directed graph of the device. The state anomaly scoring quantification unit performs node anomaly intensity quantification on the upstream equipment nodes of each directed edge based on the composite graph structure, and performs buffer isolation capability quantification on the downstream equipment nodes of each directed edge. It also introduces a unidirectional influence structure shielding matrix. After structural shielding of directed edges that do not have effective process transmission relationship, it generates the anomaly propagation influence quantity and buffer isolation parameters corresponding to the directed edge. The edge weight calculation unit is based on the abnormal propagation impact quantity and the process impact factor to form the propagation impact enhancement term. The structural adjustment channel is introduced to adjust the propagation impact enhancement term. The propagation impact enhancement term is combined with the buffer isolation parameter to generate the propagation risk cost edge weight corresponding to the directed edge, and written into the directed edge of the corresponding equipment directed graph to form a weighted directed graph. The dynamic update control unit updates the abnormal propagation impact, process impact factor, buffer isolation parameters and structural adjustment channel adjustment parameters in real time based on equipment operation status data, product quality surface data and process data, and regenerates the propagation risk cost edge weights, and dynamically updates the weighted directed graph.

[0012] Optionally, identifying cascading fault risk loops existing in the weighted directed graph includes: Based on the directed graph of the equipment and the set of abnormal source nodes extracted from the directed graph of the equipment, the abnormal source nodes are taken as the set of starting nodes for risk propagation, and the remaining equipment nodes are taken as the set of nodes to be evaluated. Based on the set of starting nodes and the set of nodes to be evaluated, the improved Bellman-Ford algorithm is executed. Virtual source nodes are set up, and virtual directed connections with zero propagation cost are established between the virtual source nodes and each abnormal source node. The propagation cost of all nodes is initialized, and the predecessor node records of all nodes are set to empty. In a weighted directed graph, the propagation cost is iterated round by round. In each round, the propagation cost of each directed edge is compared and updated with the previous node. At the same time, the node number and the propagation round are used as row and column dimensions to record the iterative change value of the propagation cost of each node. A propagation path stability scoring matrix is ​​constructed to measure the path credibility. A path split coding graph is constructed to record the path split information of nodes with the same propagation cost source. A dynamic weighted kernel group for propagation rounds is introduced to dynamically weight the propagation cost increment according to the propagation round, giving higher weight to the propagation cost changes corresponding to earlier propagation rounds, and updating the predecessor node and propagation direction of the node in combination with the propagation path stability scoring matrix and path split coding graph. After the propagation cost converges iteratively, for each node to be evaluated, backtrack from the preceding node to the corresponding abnormal source node, and delete the virtual connections related to the virtual source node in the risk propagation path to obtain the risk propagation path corresponding to the node. After completing the propagation cost iteration, the propagation cost of each node to be evaluated is taken as the minimum risk propagation cost. An additional traversal is performed on all directed edges. When there is a directed edge that can further reduce the propagation cost of downstream nodes, it is determined that there is a self-reinforcing cascading fault risk loop in the directed graph of the device. Based on the predecessor node of the corresponding downstream node, the path containing the corresponding fault risk loop is continuously backtracked.

[0013] Optionally, determining key risk nodes and the potential scope of fault propagation includes: Based on the minimum risk propagation cost of each node and the corresponding risk propagation path, the number of node path connections for each node in all risk propagation paths is counted. By comprehensively processing the minimum risk propagation cost and the number of node path connections, a risk quantification score for each node is obtained. Based on risk quantification scoring and risk threshold, key risk nodes are screened to form a set of key risk nodes, and nodes belonging to the cascaded failure risk loop are incorporated into the set of key risk nodes. Based on the risk propagation path of key risk nodes in the set of key risk nodes, and combined with the minimum risk propagation cost of the nodes, the potential scope of fault propagation is determined.

[0014] The beneficial effects of this invention are: This invention proposes a fault detection method for production line equipment based on edge computing. It utilizes multimodal data fusion, process dependency graph modeling, and risk propagation path inference techniques to identify the correlation between the operating status of multiple equipment on the production line and product quality anomalies, and to predict cascading risks. The method collaboratively collects and standardizes equipment operating status data, product surface quality data, process data, and equipment communication timing data at edge nodes. Combining the material flow patterns, process dependencies, and control linkage logic of the production line, it constructs a topologically oriented directed graph of equipment and builds a dynamic adjustment module for fault propagation edge weights. Through structured modeling of the propagation impact between equipment, buffer isolation capabilities, and the intensity of process effects, it generates propagation risk cost edge weights that reflect the ease of risk transmission, forming a weighted directed graph capable of propagation analysis.

[0015] This invention utilizes an improved Bellman-Ford algorithm to iteratively infer propagation risks at the edge node side. Through methods such as virtual source node construction, dynamic weighting of propagation paths, and confidence assessment of split paths, it obtains the minimum risk propagation cost, potential propagation paths, and cascading failure risk loops in real time. Furthermore, it determines key risk nodes and potential propagation ranges based on risk quantification scoring and path connectivity statistics.

[0016] This invention enables real-time inference of risks at multiple nodes in a production line, collaborative modeling of equipment and product quality anomalies, and interpretable identification of cascading propagation paths. It overcomes the limitations of existing technologies, which are restricted to single-node detection, lack propagation analysis, and cannot assess chain risks in real time at the edge. This invention offers advantages such as rapid response, localized data processing, strong security, outstanding predictive capabilities, and high interpretability of propagation mechanisms, effectively improving the intelligent monitoring capabilities and overall operational reliability of production lines. Attached Figure Description

[0017] 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 a flowchart of a production line equipment fault detection method based on edge computing proposed in this invention; Figure 2 This is a schematic diagram of the fault propagation edge weight dynamic control module structure of a production line equipment fault detection method based on edge computing proposed in this invention. Figure 3 This is a flowchart illustrating the execution of an improved Bellman-Ford algorithm for a production line equipment fault detection method based on edge computing, as proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A method for fault detection of production line equipment based on edge computing, comprising: Collect multimodal sensing data from edge computing nodes of production line equipment, preprocess the multimodal sensing data to obtain standardized multimodal sensing data; Based on standardized multimodal perception data, product defect parameters and operational anomaly indicators are extracted respectively to generate product defect anomaly scores and equipment operational anomaly scores, and then fused according to preset fusion rules to obtain the node anomaly intensity of each equipment node. Based on the strength of node anomalies, anomaly nodes are selected to form a set of anomaly source nodes, and a directed graph of equipment is constructed by combining the process dependencies between production line equipment. A fault propagation edge weight dynamic control module is constructed to generate propagation risk cost edge weights for each directed edge in the device directed graph, forming a weighted directed graph; Based on weighted directed graphs, an improved Bellman-Ford algorithm is used to iteratively calculate path costs, obtain the minimum risk propagation cost from each anomaly source node to the remaining nodes and the corresponding risk propagation path, and identify cascading fault risk loops in the weighted directed graph. Based on the minimum risk propagation cost, risk propagation path, and cascading failure risk loop, key risk nodes and potential failure propagation range are identified.

[0020] In this embodiment, the multimodal sensing data specifically includes equipment operating status data, product quality surface data, process data, and inter-equipment communication and timing data.

[0021] In this embodiment, the preprocessing of multimodal sensing data specifically includes time synchronization, sampling rate alignment, noise suppression, scale standardization, and missing value completion.

[0022] In this embodiment, the step of fusing the nodes according to preset fusion rules to obtain the node anomaly strength of each device node includes: Based on product quality surface data from standardized multimodal sensing data, candidate defect regions are extracted from the product surface. The area, perimeter, shape features, edge sharpness, and brightness variation of these candidate defect regions are then quantified and statistically analyzed to obtain a set of feature parameters for each candidate defect region. Specifically, this set of feature parameters is as follows: The area, shape complexity, edge sharpness, brightness difference, and texture anomaly of the defect candidate region are extracted respectively, normalized, and mapped to a unified numerical range. Defect candidate regions with large area, irregular shape, abrupt edge transition, obvious brightness contrast, and texture pattern deviating from the normal feature distribution correspond to high normalized feature values, thus obtaining the feature parameter set of the defect candidate region. Based on the feature parameter set, a product defect anomaly score is calculated according to the area, shape complexity, edge sharpness, brightness difference, and texture anomaly degree of the defect candidate region. The calculation of the product defect anomaly score is specifically as follows: Based on the feature parameter set and combined with the product process inspection standards, corresponding weight coefficients are set for area, shape complexity, edge sharpness, brightness difference and texture abnormality, and each feature parameter is weighted and fused according to the corresponding weight to generate a comprehensive abnormality score for the corresponding defect candidate area. Based on equipment operating status data from standardized multimodal sensing data, abnormal indicators are identified. Vibration changes, temperature changes, load fluctuations, current deviations, and operational stability are measured and weighted to generate an equipment operating anomaly score. Abnormal indicators refer to abnormal operating state parameters where the deviation between the vibration signals, temperature data, load data, current data, and operating stability parameters of each device and the normal operating state parameters exceeds the deviation threshold. The generated equipment malfunction score is specifically as follows: The abnormal deviation of different operating state parameters is normalized, and all abnormal indicators are uniformly mapped to a comparable numerical range. The greater the deviation, the higher the normalized value. According to the importance of different operating state parameters to equipment safety, corresponding weight coefficients are set for vibration change, temperature change, load fluctuation amplitude, current deviation and operating stability, and the normalized abnormal indicators are weighted and summed to obtain the equipment operation abnormality score. Based on preset weighted fusion rules, product defect anomaly scores and equipment operation anomaly scores are fused to obtain the node anomaly intensity of each equipment node, where: The preset weighted fusion rule refers to a fusion strategy that sets fusion weights for product defect anomaly scores and equipment operation anomaly scores based on the degree of correlation between product surface defects and equipment anomalies and the degree of influence of equipment operation anomalies on product defects, and then combines them according to the weights within a unified numerical range. The node anomaly strength of each device node is obtained as follows: The numerical scales of product defect anomaly scores and equipment operation anomaly scores are unified so that the scores are within the same range. Then, the two scores are weighted according to the fusion weights in the preset weighted fusion rules to obtain the node anomaly intensity of the corresponding equipment node.

[0023] In this embodiment, the construction of a directed graph of equipment based on the process dependencies between production line equipment includes: The node anomaly intensity is subjected to consistency detection and amplitude smoothing, and compared with anomaly intensity threshold. Device nodes whose anomaly intensity exceeds the anomaly intensity threshold are marked as anomalous nodes. The specific steps of the node anomaly intensity consistency detection and amplitude smoothing are as follows: Based on the node anomaly intensity change records in the continuous acquisition period, the consistency of the node anomaly intensity time series is detected. If the difference between the node anomaly intensity value in the continuous acquisition period and the adjacent period exceeds the deviation threshold, the corresponding node anomaly intensity is marked as an inconsistent value and corrected. The corrected node anomaly intensity is then smoothed. The node anomaly intensity in adjacent time slices is then weighted and smoothed to eliminate sharp fluctuations and maintain continuous change, thus obtaining the smoothed node anomaly intensity. The marked abnormal nodes are identified and their node numbers and corresponding abnormality strengths are recorded to form an abnormal source node data set. Based on the process data and inter-equipment communication and timing data in the standardized multimodal sensing data, the sequential processing relationship, material flow relationship and linkage control relationship between equipment nodes in the production line are determined, and the dependency direction between equipment nodes is obtained. A directed graph of devices is constructed based on the dependency directions between device nodes, and the set of abnormal source nodes is written into the corresponding device node attributes to form a directed graph of devices containing abnormal node identifiers.

[0024] In this embodiment, forming a weighted directed graph includes: A fault propagation edge weight dynamic control module is constructed, which consists of an adjacency topology mapping unit, a state anomaly scoring quantification unit, an edge weight calculation unit, and a dynamic update control unit. Specifically, the construction of the fault propagation edge weight dynamic control module involves: The output of the adjacency topology mapping unit is connected to the input of the state anomaly scoring quantization unit. The output of the state anomaly scoring quantization unit is connected to the input of the edge weight calculation unit. The output of the edge weight calculation unit is connected to the input of the dynamic update control unit. The output of the dynamic update control unit is simultaneously connected to the inputs of the adjacency topology mapping unit and the edge weight calculation unit to form a back connection, thus constituting a fault propagation edge weight dynamic control module with a forward calculation link and a back connection update link. The adjacency topology mapping unit reads the directed graph of the devices and obtains the directed edges in the directed graph and the corresponding upstream and downstream device node information. It then combines this information with standardized multimodal sensing data to construct a composite graph structure corresponding to the directed graph of the devices, wherein: An upstream equipment node refers to a device node located at the starting position of a directed edge in a directed equipment graph that provides processing results, material flow, and control signals to the remaining equipment nodes. Downstream equipment nodes refer to equipment nodes located at the end of directed edges in the directed equipment graph, and whose processing results, material flow, and control signal inputs are provided by upstream equipment nodes; The construction of the composite graph structure corresponding to the directed graph of the device is specifically as follows: The process dependency relationship between upstream and downstream equipment nodes in the directed graph of the equipment is used as the topological dimension. The equipment operation status data, product quality surface data and process data corresponding to the directed edges in the standardized multimodal perception data are used as the data dimension. Multidimensional mapping is performed according to the correspondence between the topological dimension and the data dimension to form a composite graph structure with the directed relationship between nodes as the main index and the features of multi-source process and status data attached. The anomaly scoring quantification unit quantifies the node anomaly intensity of upstream equipment nodes for each directed edge based on a composite graph structure, and simultaneously quantifies the buffer isolation capability of downstream equipment nodes for each directed edge. A unidirectional influence structure shielding matrix is ​​introduced. After structural shielding of directed edges without effective process transmission relationships, the unit generates the anomaly propagation influence quantity and buffer isolation parameters corresponding to the directed edges, where: The node anomaly intensity quantification process is performed as follows: Based on the node anomaly intensity corresponding to the upstream equipment node in the composite graph structure, an anomaly transmission amplification coefficient is set according to the influence ratio of the upstream equipment node's process on the subsequent process. The transmission characteristics of the node anomaly intensity are converted to obtain the anomaly propagation influence quantity, which represents the degree of potential propagation influence of the upstream equipment node on the downstream node. The quantification of buffer isolation capacity is performed as follows: Based on the equipment process parameters, work-in-process buffer capacity, bypass switching capability and switching delay parameters corresponding to downstream equipment nodes in the composite graph structure, a buffer suppression coefficient is set according to the contribution of each parameter to suppressing fault propagation. After numerical normalization of different parameters, a weighted calculation is performed based on the buffer suppression coefficient to obtain the quantitative value of buffer isolation capability. The unidirectional influence structure shielding matrix is ​​a structured matrix that expresses the directional constraints of process transmission between equipment in a production line. It takes the unidirectional dependency relationship between equipment nodes in a directed graph as the basis for its construction, and takes the equipment processing sequence, material flow logic and control linkage relationship as the basis for matrix generation. The matrix elements reflect the reachability state that only allows unidirectional process transmission, and are presented as a structural shielding mark to identify non-unidirectional process links, describing the asymmetric influence relationship between equipment processes. The edge weight calculation unit constructs a propagation impact enhancement term based on the anomaly propagation impact quantity and the process impact factor. A structural adjustment channel is introduced to adjust this propagation impact enhancement term. The propagation impact enhancement term is then combined with buffer isolation parameters to generate propagation risk cost edge weights corresponding to directed edges. These weights are then written into the directed edges of the corresponding equipment directed graph, forming a weighted directed graph. Process influence factors refer to a set of structural parameters that characterize the strength of process dependencies between equipment in a production line. They take the degree of tight coupling of processing sequence between equipment, the correlation of work-in-process quality, and the strength of control synergy as the basis, and express the influence intensity between equipment nodes due to the process flow in a structural-level quantitative manner. The structural adjustment channel refers to the parameter adjustment structure that runs parallel to the elements constituting the risk of propagation. It uses the topological positional relationship between equipment, the level of process coupling, and the time synchronization characteristics as adjustment dimensions. It is a channelized parameter structure that uses structural weight factors to finely adjust the propagation impact enhancement term, reflecting the deviation of propagation characteristics caused by differences in topological structure. The process of combining the propagation effect enhancement term with the buffer isolation parameter is as follows: The propagation impact enhancement term and the buffer isolation parameter are normalized and converted respectively. The normalized propagation impact enhancement term is taken as the positive impact term reflecting the downstream propagation of the anomaly, and the normalized buffer isolation parameter is taken as the negative suppression term reflecting the ability to suppress the propagation of the anomaly. Structural weight factors are set for the positive impact term and the negative suppression term respectively. The positive impact term is multiplied by the corresponding structural weight factor and taken as an addition term, and the negative suppression term is multiplied by the corresponding structural weight factor and taken as a subtraction term. The addition term is subtracted from the subtraction term for the two terms to obtain the propagation risk cost edge weight corresponding to the directed edge. The dynamic update control unit updates the abnormal propagation impact, process impact factor, buffer isolation parameters, and structural adjustment channel adjustment parameters in real time based on equipment operation status data, product quality surface data, and process data. It also regenerates the propagation risk cost edge weights and dynamically updates the weighted directed graph. Specifically, refreshing the abnormal propagation impact, process impact factor, buffer isolation parameters, and structural adjustment channel adjustment parameters involves: Based on real-time collected equipment operation status data, product quality surface data, and process data, the latest data fragments associated with the corresponding directed edges are extracted and mapped to the dynamic temporal dimension of the composite graph structure. The data components related to anomaly propagation are unified in terms of dimensions and their structural proportions are converted to form new anomaly propagation impact quantities and process impact factors. The data components related to buffer isolation are dynamically converted into buffer capacity, bypass switching capability, and switching delay parameters to form updated buffer isolation parameters. The adjustment parameters of the structural adjustment channel are weighted according to the ratio of positive influence terms to negative suppression terms under the current time series to obtain new adjustment parameters, thus completing the update process.

[0025] In this embodiment, identifying cascading fault risk loops in a weighted directed graph includes: Based on the directed graph of the equipment and the set of abnormal source nodes extracted from the directed graph of the equipment, the abnormal source nodes are taken as the set of starting nodes for risk propagation, and the remaining equipment nodes are taken as the set of nodes to be evaluated. Based on the starting node set and the node set to be evaluated, the improved Bellman-Ford algorithm is executed. A virtual source node is set, and a virtual directed connection with zero propagation cost is established between the virtual source node and each abnormal source node. The propagation cost of all nodes is initialized, and the predecessor node records of all nodes are set to empty. Specifically: The specific steps for setting up the virtual source node are as follows: Add a new node to the weighted directed graph, use the new node as a virtual source node, generate a directed connection between the virtual source node and each abnormal source node with zero propagation cost, and add the corresponding zero-cost directed connection to the existing set of directed edges to form the expanded propagation topology. The initialization of the propagation cost across all nodes is specifically as follows: Set the propagation cost record of each node in the expanded propagation topology to the unreachable cost constant threshold, and set the propagation cost record of the virtual source node to the zero cost initial value to form the initial state table of propagation cost. In a weighted directed graph, propagation cost is iteratively processed round by round. In each round, the propagation cost of each directed edge is compared and updated with that of the preceding node. Simultaneously, the node number and propagation round number are used as row and column dimensions to record the iterative change value of the propagation cost of each node. A propagation path stability scoring matrix is ​​constructed to quantify path reliability, and a path splitting coding graph is constructed to record path splitting information of nodes with the same propagation cost source. The comparison and update of the propagation cost with the preceding node are performed as follows: The candidate propagation cost is obtained by numerically superimposing the current propagation cost of the upstream node and the propagation risk cost of the corresponding directed edge in the weighted directed graph. When the candidate propagation cost is less than the current propagation cost of the downstream node, the propagation cost record of the downstream node is replaced with the candidate propagation cost, and the corresponding upstream node is recorded as the new predecessor node of the downstream node. The construction of the propagation path stability scoring matrix is ​​specifically as follows: A propagation path stability scoring matrix is ​​constructed using node number as row index and propagation iteration round as column index. The change in propagation cost of a node in consecutive propagation iteration rounds is used as the scoring basis. The change in propagation cost in each round is converted into a stability score according to the corresponding row and column position and written into the matrix propagation path stability scoring matrix. The change in propagation cost is inversely proportional to the stability score. The propagation path stability scoring matrix reflects the stability of the propagation path of each node. The construction path split coding graph is specifically as follows: The nodes are numbered as indices. Records with equivalent propagation costs but originating from different predecessor nodes in the same propagation iteration are uniquely encoded and mapped to the graph structure. The codes are then structurally associated with the corresponding predecessor nodes and stored to form a path splitting encoding graph of the node path splitting records. This graph records the structured graph of multiple source paths that a node may have during propagation. Elements in the path splitting encoding graph mark the different predecessor node sources corresponding to the same propagation cost, expressing the parallel splitting relationship of the path. A dynamically weighted kernel group for propagation rounds is introduced to dynamically weight the propagation cost increment according to the propagation round, assigning higher weights to the propagation cost changes corresponding to earlier propagation rounds. This is combined with the propagation path stability scoring matrix and the path splitting coding graph to update the predecessor nodes and propagation directions of the nodes, where: The propagation cost increment refers to the new amount of propagation risk generated during the propagation cost iteration process by superimposing the current propagation cost of the upstream node and the propagation risk cost edge weight from the node to the downstream node; The propagation round dynamic weighted kernel group refers to a set of weights consisting of multiple weighting coefficients corresponding to the propagation iteration rounds. It uses the propagation round number as an index and arranges the different weighting factors corresponding to the early, middle and late stages of the iteration in sequence to form a serialized weighted structure. It has the structural attribute of presenting a differential numerical distribution according to the propagation round. The update of the predecessor node and propagation direction corresponding to the node is specifically as follows: The propagation cost increment of the candidate predecessor node is introduced into a dynamic weighted kernel group of propagation rounds. The propagation cost increment is multiplied by the weighting coefficient corresponding to the current propagation iteration round to obtain the corresponding weighted propagation cost. Based on the joint comparison of the weighted propagation cost and the stability score of the corresponding node in the propagation path stability score matrix, the candidate preceding node with a weighted propagation cost lower than the cost threshold and a stability score higher than the score threshold is determined as the priority preceding node. When there are multiple candidate preceding nodes with the same priority level, based on the split coding information recorded in the path split coding graph, the candidate preceding nodes with the same priority level are sorted according to the split coding number order, and the one with the earlier sorting is selected as the final preceding node. Write the final predecessor node into the predecessor node record of the current node, and update the propagation direction of the current node to the direction in which the final predecessor node points to the current node; After the propagation cost converges iteratively, for each node to be evaluated, backtracking step by step from the preceding node to the corresponding anomaly source node, and deleting the virtual connections related to the virtual source node in the risk propagation path, the risk propagation path corresponding to the node is obtained, where: Iterative convergence of propagation cost refers to the situation where all directed edges no longer update the propagation cost in the same round of iteration, and the propagation cost of downstream nodes no longer changes due to any candidate propagation cost increment. The iteration state of the corresponding round is regarded as the final convergence state of propagation cost and is determined to be iterative convergence of propagation cost. After completing the propagation cost iteration, the propagation cost of each node to be evaluated is used as the minimum risk propagation cost. An additional traversal is performed on all directed edges. When there is a directed edge that can further reduce the propagation cost of downstream nodes, it is determined that there is a self-reinforcing cascading fault risk loop in the device's directed graph. Based on the predecessor nodes of the corresponding downstream nodes, the path containing the corresponding fault risk loop is continuously backtracked, where: The existence of directed edges that can continue to reduce the propagation cost of downstream nodes means that after the propagation cost iteration is completed, if any directed edge in the weighted directed graph is tested, and the result of adding the existing propagation cost of the corresponding upstream node to the edge propagation risk cost is still less than the current propagation cost record of the downstream node, it indicates that the directed edge has the potential to continuously reduce the propagation cost of the downstream node.

[0026] In this embodiment, determining the key risk nodes and the potential scope of fault propagation includes: Based on the minimum risk propagation cost of each node and the corresponding risk propagation path, the number of node path connections for each node in all risk propagation paths is counted. Specifically, the counting of the number of node path connections for each node in all risk propagation paths involves: In all the risk propagation paths formed by backtracking from each node to be evaluated, each risk propagation path is unfolded into a node sequence composed of multiple nodes in sequence. Each device node appearing in the path sequence is counted one by one. When a node is repeatedly included in different risk propagation paths, the count is accumulated once each time it appears, so as to obtain the number of node path connections of each node in all risk propagation paths. The risk quantification score for each node is obtained by comprehensively processing the minimum risk propagation cost and the number of node path connections. Specifically, this comprehensive processing of the minimum risk propagation cost and the number of node path connections involves: For each device node, the minimum risk propagation cost and the corresponding number of node path connections are normalized. The two indicators are in the same order of magnitude. Within the same order of magnitude, the propagation cost is used as the basic risk metric, and the number of path connections is used as the risk amplification coefficient. The two indicators are multiplied and added in a weighted manner to form a quantitative risk score for the node. A set of key risk nodes is formed by screening key risk nodes based on risk quantification scoring and risk thresholds, and nodes belonging to cascading failure risk loops are incorporated into the set of key risk nodes. The formation of the set of key risk nodes is as follows: The risk quantification scores of all equipment nodes are compared numerically. Nodes with risk quantification scores greater than or equal to the risk threshold are marked as risk-exceeding nodes. All marked risk-exceeding nodes are aggregated according to their corresponding node numbers to form a set of key risk nodes consisting of multiple nodes. Based on the risk propagation paths of key risk nodes in the set of key risk nodes, and combined with the minimum risk propagation cost of each node, the potential fault propagation range is determined. Specifically, determining the potential fault propagation range involves: For each critical risk node in the set of critical risk nodes, the node sequence is expanded according to the corresponding risk propagation path. The downstream nodes following the critical risk node in the node sequence are extracted one by one. Based on the minimum risk propagation cost of the critical risk node, the extracted downstream nodes are marked with risk cost association. The marked downstream nodes in the risk propagation path corresponding to all critical risk nodes are collected and merged according to the node number to form a potential fault propagation range consisting of multiple downstream nodes that may be affected by propagation.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to an aluminum heat sink processing production line at an automotive parts manufacturing base. The production line consists of grinding equipment, edge trimming equipment, drilling equipment, cleaning equipment, and an automatic detection unit. Material transfer occurs between the equipment according to the process sequence. A typical problem that has long existed at the base is that malfunctions in a single piece of equipment can easily lead to a decline in the processing quality of subsequent processes. For example, worn cutting tools in the edge trimming equipment can cause burrs on the edges of the heat sink, further affecting drilling accuracy and leaving residue during the cleaning process. However, traditional on-site monitoring methods can only alarm for a single piece of equipment and cannot predict whether the malfunction will spread to subsequent processes, nor can they identify which equipment will be affected.

[0028] In the application of this invention, the equipment operating status, product surface inspection images, and process parameter communication records are all collected at the edge node. The vibration signal acquisition frequency of the grinding equipment is set to 1.2kHz, the tool temperature rise of the edge trimming equipment is recorded in real time every minute, and the sampling frequency of the automatic detection unit camera is set to 18fps. All signals are standardized at the local node. During operation, the tool temperature of the edge trimming equipment rises from 42.7℃ to 58.3℃, the current fluctuation deviates from the normal average by 17.6%, and at the same time, the edge brightness difference of the product captured by the detection unit increases from an average of 15.2 to 37.9, and the surface texture anomaly score rises from 0.11 to 0.38. After fusing the operating anomaly score and surface defect score in real time at the edge node, this invention calculates the anomaly intensity of the edge trimming equipment as 0.82 and infers its potential impact on drilling and cleaning equipment based on the directed process diagram.

[0029] By constructing a weighted directed graph, the propagation risk cost edge weights are automatically generated. An improved Bellman-Ford algorithm infers that the minimum propagation cost of tool anomalies to the drilling equipment is 0.29, and to the cleaning equipment is 0.33. Simultaneously, a cascading risk loop from the cutting edge to cleaning and then back to drilling is detected. Further, based on risk quantification scoring, the drilling and cleaning equipment are identified as key risk nodes, and it is predicted that the drilling depth deviation may increase from ±0.02mm to ±0.09mm within the next 35 minutes, and the cleaning residue rate may increase from 2.7% to 11.3%. By adjusting the tool, limiting drilling speed, and delaying cleaning strategies in advance, the factory reduced the product defect rate from 5.6% to 1.2% in a seven-day continuous test, verifying the invention's ability to identify cascading effects and reduce anomaly propagation in complex production scenarios.

[0030] Table 1. Comparison of Fault Propagation Prediction and Prevention Effects Based on Edge Computing

[0031] As shown in Table 1, the method of this invention has significant advantages in detecting abnormal nodes, predicting cascading faults, and identifying critical nodes in actual production line scenarios. Looking at the monitoring behavior of typical equipment, such as the edge-cutting equipment (items 1 and 2), although traditional monitoring showed increased tool temperature and current deviation, these did not generate effective anomaly alerts because they did not exceed individual thresholds. However, this invention, by integrating product defect anomaly scores and operational anomaly scores and combining them with process dependencies, identifies these weak anomalies as nodes with high potential propagation risk and predicts their impact on downstream drilling equipment. This allows for early warning 37 minutes in advance, reducing the subsequent burr rate and rework rate from 4.1% to 1.3% and from 6.2% to 1.8%, respectively. This demonstrates that this invention has the ability to identify equipment states that have propagation potential but have not yet constituted obvious anomalies.

[0032] In terms of cascading risk prediction, this invention predictively identifies the cleaning-spraying-drying linked production chain formed by sequences 5, 6, and 7. The residue rate of the cleaning equipment, traditionally considered a localized quality problem, is identified by this invention as the starting point of a cascading risk loop. Furthermore, it predicts that this will lead to increased spraying thickness deviation and poor drying and curing, providing early warning before the event occurs. After intervention, the spraying thickness error decreased from 12.3 μm to 3.4 μm, and the poor drying and curing rate decreased from 3.8% to 0.6%. This demonstrates that this method can not only identify single-point anomalies but also predict cross-process quality deterioration chains from the source, preventing localized problems from amplifying into line-wide defects.

[0033] In the case of final inspection at key nodes, traditional monitoring treats the false positive rate as an equipment performance issue. However, this invention identifies the false positive rate as the endpoint of multi-path propagation through risk path analysis. Subsequently, by tracing and addressing the upstream nodes, the false positive rate is reduced from 4.2% to 1.1%. This method can predict cascading anomalies on average 22 to 38 minutes in advance and effectively reduce the overall line defect rate, fully demonstrating its ability to predict large-scale risk propagation and its cost-reduction and efficiency-enhancing value for actual production lines.

[0034] This invention can effectively solve the problems of traditional methods, such as reliance on single-point threshold detection, lack of propagation prediction, and delayed response to multi-source composite anomalies, and achieve true real-time edge judgment and cross-process anomaly suppression.

[0035] 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. A method for fault detection of production line equipment based on edge computing, characterized in that, include: Collect multimodal sensing data from edge computing nodes of production line equipment, preprocess the multimodal sensing data to obtain standardized multimodal sensing data; Based on standardized multimodal perception data, product defect parameters and operational anomaly indicators are extracted respectively to generate product defect anomaly scores and equipment operational anomaly scores, and then fused according to preset fusion rules to obtain the node anomaly intensity of each equipment node. Based on the strength of node anomalies, anomaly nodes are selected to form a set of anomaly source nodes, and a directed graph of equipment is constructed by combining the process dependencies between production line equipment. A fault propagation edge weight dynamic control module is constructed to generate propagation risk cost edge weights for each directed edge in the device directed graph, forming a weighted directed graph; Based on weighted directed graphs, an improved Bellman-Ford algorithm is used to iteratively calculate path costs, obtain the minimum risk propagation cost from each anomaly source node to the remaining nodes and the corresponding risk propagation path, and identify cascading fault risk loops in the weighted directed graph. Based on the minimum risk propagation cost, risk propagation path, and cascading failure risk loop, key risk nodes and potential failure propagation range are identified.

2. The method for fault detection of production line equipment based on edge computing according to claim 1, characterized in that, The multimodal sensing data specifically includes equipment operating status data, product quality surface data, process data, and inter-equipment communication and timing data.

3. The method for fault detection of production line equipment based on edge computing according to claim 1, characterized in that, The preprocessing of multimodal sensing data specifically includes time synchronization, sampling rate alignment, noise suppression, scale standardization, and missing value completion.

4. The method for fault detection of production line equipment based on edge computing according to claim 1, characterized in that, The node anomaly strength of each device node obtained by fusing according to preset fusion rules includes: Based on product quality surface data in standardized multimodal perception data, defect candidate regions of product surface are extracted, and the area, perimeter, shape features, edge sharpness and brightness variation of defect candidate regions are quantitatively statistically analyzed to obtain the feature parameter set of defect candidate regions. Based on the set of feature parameters, the product defect anomaly score is calculated according to the area, shape complexity, edge sharpness, brightness difference, and texture anomaly of the defect candidate region. Based on equipment operation status data in standardized multimodal sensing data, abnormal indicators are identified, and the vibration change, temperature change, load fluctuation amplitude, current deviation and operation stability of the equipment are measured, weighted and summarized to generate an equipment operation anomaly score. Based on the preset weighted fusion rules, the product defect anomaly score and the equipment operation anomaly score are fused to obtain the node anomaly intensity of each equipment node.

5. The method for fault detection of production line equipment based on edge computing according to claim 1, characterized in that, The construction of the directed graph of equipment by combining the process dependencies between production line equipment includes: Consistency detection and amplitude smoothing are performed on the abnormal intensity of nodes, and the results are compared with the abnormal intensity threshold. Device nodes whose abnormal intensity exceeds the abnormal intensity threshold are marked as abnormal nodes. The marked abnormal nodes are identified and their node numbers and corresponding abnormality strengths are recorded to form an abnormal source node data set. Based on the process data and inter-equipment communication and timing data in the standardized multimodal sensing data, the sequential processing relationship, material flow relationship and linkage control relationship between equipment nodes in the production line are determined, and the dependency direction between equipment nodes is obtained. A directed graph of devices is constructed based on the dependency directions between device nodes, and the set of abnormal source nodes is written into the corresponding device node attributes to form a directed graph of devices containing abnormal node identifiers.

6. The method for fault detection of production line equipment based on edge computing according to claim 1, characterized in that, The formation of the weighted directed graph includes: A fault propagation edge weight dynamic control module is constructed, which consists of an adjacency topology mapping unit, a state anomaly scoring quantification unit, an edge weight calculation unit, and a dynamic update control unit. The adjacency topology mapping unit reads the directed graph of the device and obtains the directed edges in the directed graph of the device and the corresponding upstream and downstream device node information. It then combines the standardized multimodal sensing data to construct a composite graph structure corresponding to the directed graph of the device. The state anomaly scoring quantification unit performs node anomaly intensity quantification on the upstream equipment nodes of each directed edge based on the composite graph structure, and performs buffer isolation capability quantification on the downstream equipment nodes of each directed edge. It also introduces a unidirectional influence structure shielding matrix. After structural shielding of directed edges that do not have effective process transmission relationship, it generates the anomaly propagation influence quantity and buffer isolation parameters corresponding to the directed edge. The edge weight calculation unit is based on the abnormal propagation impact quantity and the process impact factor to form the propagation impact enhancement term. The structural adjustment channel is introduced to adjust the propagation impact enhancement term. The propagation impact enhancement term is combined with the buffer isolation parameter to generate the propagation risk cost edge weight corresponding to the directed edge, and written into the directed edge of the corresponding equipment directed graph to form a weighted directed graph. The dynamic update control unit updates the abnormal propagation impact, process impact factor, buffer isolation parameters and structural adjustment channel adjustment parameters in real time based on equipment operation status data, product quality surface data and process data, and regenerates the propagation risk cost edge weights, and dynamically updates the weighted directed graph.

7. The method for fault detection of production line equipment based on edge computing according to claim 1, characterized in that, The identification of cascading fault risk loops in the weighted directed graph includes: Based on the directed graph of the equipment and the set of abnormal source nodes extracted from the directed graph of the equipment, the abnormal source nodes are taken as the set of starting nodes for risk propagation, and the remaining equipment nodes are taken as the set of nodes to be evaluated. Based on the set of starting nodes and the set of nodes to be evaluated, the improved Bellman-Ford algorithm is executed. Virtual source nodes are set up, and virtual directed connections with zero propagation cost are established between the virtual source nodes and each abnormal source node. The propagation cost of all nodes is initialized, and the predecessor node records of all nodes are set to empty. In a weighted directed graph, the propagation cost is iterated round by round. In each round, the propagation cost of each directed edge is compared and updated with the previous node. At the same time, the node number and the propagation round are used as row and column dimensions to record the iterative change value of the propagation cost of each node. A propagation path stability scoring matrix is ​​constructed to measure the path credibility. A path split coding graph is constructed to record the path split information of nodes with the same propagation cost source. A dynamic weighted kernel group for propagation rounds is introduced to dynamically weight the propagation cost increment according to the propagation round, giving higher weight to the propagation cost changes corresponding to earlier propagation rounds, and updating the predecessor node and propagation direction of the node in combination with the propagation path stability scoring matrix and path split coding graph. After the propagation cost converges iteratively, for each node to be evaluated, backtrack from the preceding node to the corresponding abnormal source node, and delete the virtual connections related to the virtual source node in the risk propagation path to obtain the risk propagation path corresponding to the node. After completing the propagation cost iteration, the propagation cost of each node to be evaluated is taken as the minimum risk propagation cost. An additional traversal is performed on all directed edges. When there is a directed edge that can further reduce the propagation cost of downstream nodes, it is determined that there is a self-reinforcing cascading fault risk loop in the directed graph of the device. Based on the predecessor node of the corresponding downstream node, the path containing the corresponding fault risk loop is continuously backtracked.

8. The method for fault detection of production line equipment based on edge computing according to claim 1, characterized in that, The determination of key risk nodes and the potential scope of fault propagation includes: Based on the minimum risk propagation cost of each node and the corresponding risk propagation path, the number of node path connections for each node in all risk propagation paths is counted. By comprehensively processing the minimum risk propagation cost and the number of node path connections, a risk quantification score for each node is obtained. Based on risk quantification scoring and risk threshold, key risk nodes are screened to form a set of key risk nodes, and nodes belonging to the cascaded failure risk loop are incorporated into the set of key risk nodes. Based on the risk propagation path of key risk nodes in the set of key risk nodes, and combined with the minimum risk propagation cost of the nodes, the potential scope of fault propagation is determined.

Images