An enterprise-level intelligent point patrol inspection recognition analysis system based on a graph neural network

Abstract

The application discloses an enterprise-level intelligent point patrol inspection recognition analysis system based on a graph neural network, comprising: a point patrol inspection standard management module, which is used for establishing a point patrol inspection standard library and a unified coding system, and generating a point patrol inspection route; a patrol inspection task generation and execution module, which is used for generating a periodic point patrol inspection task set; a data uploading and analysis module, which is used for uploading patrol inspection data to a server through an enterprise network to form a trend sequence; a graph structure construction module, which is used for constructing a patrol inspection network graph to form a patrol inspection feature set; a graph neural network analysis module, which is used for calculating a device health score and an abnormal confidence through an improved MixHop model; an abnormality recognition and work order management module, which is used for triggering an early warning notification and updating an evaluation index according to a preset rule; and a data synchronization and report generation module, which is used for generating a statistical report and an audit log, so that the standardized, networked and intelligent management of enterprise-level point patrol inspection is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of enterprise equipment and industrial information management, and in particular to an enterprise-level intelligent point inspection and identification analysis system based on graph neural networks. Background Technology

[0002] As the number and complexity of production equipment in enterprises continue to increase, equipment inspection management has gradually become a crucial link in ensuring safe production and healthy equipment operation. Traditional inspection methods rely heavily on manual operation, primarily through manual data entry, paper records, or simple mobile terminal data collection. These methods suffer from problems such as poor data accuracy, information delays, difficulties in statistical analysis, and weak fault prediction capabilities, making it difficult to meet the needs of modern enterprises for intelligent equipment operation and maintenance and status awareness. In existing technologies, some enterprises have introduced IoT-based inspection systems, using RFID radio frequency identification or QR code scanning to achieve equipment identification and information collection. However, the functions of such systems are mainly concentrated in the "data collection" and "task execution" stages, lacking the ability to intelligently analyze inspection data and perform deep correlation modeling. Equipment inspection data is often stored in independent tables or logs, failing to form an effective network of relationships between equipment or identify potential dependencies between different inspection points, components, and processes, resulting in low accuracy and efficiency in anomaly identification.

[0003] Existing inspection systems often rely on fixed thresholds or simple statistical methods for anomaly detection. For example, an alarm is triggered if a device parameter exceeds a set threshold. This approach ignores the complex interrelationships between devices and the dynamic nature of changing operating conditions, making it prone to false alarms or missed alarms. Furthermore, traditional systems struggle to perform global analysis based on the temporal sequence, spatial distribution, and process dependence of inspection data, and cannot dynamically assess the health status of equipment. As a result, maintenance personnel must rely on experience to determine equipment anomalies, leading to low levels of system intelligence and autonomous analytical capabilities. Summary of the Invention

[0004] One objective of this invention is to propose an enterprise-level intelligent point inspection identification and analysis system based on graph neural networks. This invention fully utilizes multiple technologies such as radio frequency identification positioning, time series analysis, and data interaction with enterprise information systems. By establishing a standardized inspection coding system, constructing an inspection network diagram, and introducing an improved MixHop model, it achieves intelligent identification, health status assessment, and anomaly classification management of enterprise equipment inspection data.

[0005] An enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to an embodiment of the present invention includes:

[0006] The point inspection standard management module is used to establish a point inspection standard library and a unified coding system. It classifies and organizes all equipment, components and point inspection items of the enterprise, assigns unique codes and sets over-limit thresholds, and generates point inspection routes.

[0007] The inspection task generation and execution module is used to configure RFID tags and register location information at each inspection point, generate a set of periodic inspection tasks based on the inspection route and send them to the handheld terminal. The handheld terminal completes RFID tag verification, collects inspection data and records timestamps, device codes and task numbers.

[0008] The data upload and analysis module is used to encapsulate the inspection data in a structured way, generate structured inspection data entries, upload them to the server through the enterprise network, and aggregate them into a trend sequence based on timestamps, equipment codes and task numbers.

[0009] The graph structure construction module is used to construct an inspection network graph based on trend sequences and equipment, inspection points, process flow and spatial adjacency relationships, generate node sets, edge sets and node attribute vectors, and form an inspection feature set;

[0010] The graph neural network analysis module is used to input the inspection feature set into the improved MixHop model, perform multi-order neighborhood diffusion coding, cross-order attention fusion and task discrimination and calibration, output node representation and edge representation, and calculate equipment health score and anomaly confidence.

[0011] The anomaly identification and work order management module is used to identify and classify anomalies based on anomaly confidence level and over-limit threshold, generate defect entries and automatically create defect work orders, and trigger early warning notifications and performance indicator updates according to preset rules.

[0012] The data synchronization and report generation module is used to synchronize equipment health scores, anomaly confidence levels, and defect work order records to the equipment ledger. It achieves data exchange through the enterprise management system interface, updates equipment technical parameters, maintenance history, and runtime, and generates statistical reports and audit logs.

[0013] Optionally, modules can be integrated using the following methods:

[0014] Establish a standard library and unified coding system for point inspection, assign unique codes to equipment, components and point inspection items and set over-limit thresholds to obtain point inspection routes;

[0015] RFID tags are configured at each inspection point and their location information is registered. Periodic inspection task sets are generated based on the inspection route and sent to handheld terminals.

[0016] The point inspection personnel use handheld terminals to read RFID tags to complete the on-site verification, collect point inspection data, and record the timestamp, equipment code and task number;

[0017] Structured inspection data entries are generated based on inspection data, uploaded to the server via the enterprise network, and aggregated into a trend sequence according to timestamp, equipment code, and task number;

[0018] Based on the trend sequence, an inspection network diagram is constructed on the server side according to equipment, inspection points, process flow and spatial adjacency, forming an inspection feature set;

[0019] The inspection feature set is input into the improved MixHop model to generate node representations and edge representations, and the equipment health score and anomaly confidence are calculated.

[0020] Based on the anomaly confidence level and the limit threshold, anomaly identification and classification are performed, defect entries are generated and defect work orders are automatically created, and early warning notifications and performance indicator updates are triggered according to preset rules.

[0021] The equipment health score, anomaly confidence level and defect work order are synchronized to the equipment ledger. Data exchange is carried out through the enterprise management system interface to update equipment technical parameters, maintenance history and running time, and generate statistical reports and audit logs.

[0022] Optionally, the establishment of the point inspection standard library and unified coding system is achieved by classifying and organizing all equipment, components and inspection items of the enterprise to form a unified data standard.

[0023] Optionally, the generation and distribution of the periodic point inspection task set specifically includes:

[0024] Configure an RFID tag for each inspection point, read the unique identification code of the RFID tag, and bind the unique identification code with the corresponding inspection point number, equipment code, and inspection item number;

[0025] Record the location information of each inspection point and form a location information set;

[0026] Based on the set of inspection routes and location information, all inspection points are sorted according to spatial distance and process dependence to generate an inspection sequence table;

[0027] Based on the inspection sequence relationship table, the route with the shortest total travel distance is selected from the set of feasible routes to obtain the optimal inspection route;

[0028] A set of periodic inspection tasks is generated based on the optimal inspection route and preset inspection cycle parameters, and the set of periodic inspection tasks is distributed to handheld terminals through the enterprise network.

[0029] Optionally, the collection of the point inspection data specifically includes:

[0030] The inspection personnel read the unique identification code of the radio frequency identification tag at the inspection point using a handheld terminal, and match and verify the unique identification code with the task number of the current inspection point to complete the on-site verification.

[0031] After the on-site verification is completed, the handheld terminal loads the device code and inspection item number bound to the inspection point and generates a data item sequence list;

[0032] The inspection data is collected in the order of the data items. The inspection data includes a set of measurement data, a set of meter reading data, and a set of observation data. The handheld terminal records the timestamp, equipment code, and task number.

[0033] Optionally, the formation of the trend sequence specifically includes:

[0034] The handheld terminal encapsulates the collected point inspection data and corresponding task information to generate structured inspection data entries. The structured inspection data entries include point inspection task number, equipment code, timestamp and inspection result parameters.

[0035] The structured inspection data items are uploaded to the server through the enterprise network. The server performs integrity verification on the received data and obtains the verified valid data.

[0036] The server writes the verified valid data into the inspection database to form standardized storage records, and automatically judges the data status according to the threshold of exceeding the limit. The results of exceeding the limit are marked as abnormal data and a preliminary alarm record is generated.

[0037] Based on the time series data of equipment codes and project numbers, the data is aggregated and sorted by timestamp to generate a trend series.

[0038] Optionally, the formation of the inspection feature set specifically includes:

[0039] A node set is obtained based on the equipment code and the inspection point number. The node set includes equipment nodes and inspection point nodes. An edge set is obtained based on the process flow dependency, spatial adjacency and historical collaboration. The edge set refers to the association between any two nodes.

[0040] For each node in the node set, a node attribute vector set is established. The node attribute vector set is formed by concatenating the statistical characteristics of the trend sequence corresponding to the node and the static attribute parameters of the node in a fixed order.

[0041] The weighted adjacency matrix is ​​calculated based on the edge set. The weighted adjacency matrix is ​​obtained by combining the process dependency relationship, spatial distance relationship and historical cooperation relationship in the edge set into three indicators and then weighting them.

[0042] The node set, edge set, node attribute vector set, and weighted adjacency matrix are concatenated to form the inspection network graph;

[0043] Based on the inspection network diagram, using the equipment code and the inspection item number as keys, the corresponding trend sequence is retrieved and aligned with the vectors in the node attribute vector set within a continuous time window to generate an inspection feature set.

[0044] Optionally, the calculation of the device health score and anomaly confidence level specifically includes:

[0045] The inspection feature set is input into the improved MixHop model, which includes a multi-order neighborhood diffusion coding module, a cross-order attention fusion module, and a task discrimination and calibration module. The multi-order neighborhood diffusion coding module refers to introducing an adaptive diffusion time parameter to diffuse the intermediate representations of nodes at each order and outputting a set of intermediate representations of nodes. The cross-order attention fusion module refers to calculating node representations and edge representations based on the set of intermediate representations of nodes. The task discrimination and calibration module refers to calculating the device health score based on the node representations and calculating the anomaly confidence based on the edge representations.

[0046] In the multi-order neighborhood diffusion coding module, each node in the inspection feature set is used to generate a node intermediate representation divided by order. An order-adaptive diffusion time parameter is introduced to diffuse the node intermediate representations of each order, and the node intermediate representation set is output.

[0047] In the cross-order attention fusion module, based on the set of intermediate node representations, the importance weights of each order are calculated and weighted fusion is performed to obtain the node representation. The edge representation is constructed based on the node representations at both ends of each edge and the edge attributes. The edge representation is formed by splicing the element-wise interaction results of the node representations at both ends and the edge attributes in a fixed order.

[0048] In the task discrimination and calibration module, the device health score is calculated based on the node representation, and the anomaly confidence score is calculated based on the edge representation. The device health score is obtained by multiplying the node representation with the weight matrix and adding a bias term for linear mapping, followed by Sigmoid normalization. The anomaly confidence score is obtained by multiplying the edge representation with the weight matrix and adding a bias term for linear mapping, followed by Sigmoid normalization.

[0049] Optionally, the triggering of the early warning notification and the updating of the assessment indicators specifically include:

[0050] Based on the device health score and the anomaly confidence level, each indicator is judged by exceeding the threshold, and samples with anomaly confidence levels greater than the threshold are marked as abnormal nodes.

[0051] Based on the distribution of abnormal nodes, perform anomaly identification and classification to obtain the anomaly level and generate the corresponding set of defect entries;

[0052] Based on the equipment code, inspection point number, and abnormality level of each defect item in the defect item set, a defect work order is automatically created, and the work order number, creation time, and responsible department information are recorded.

[0053] Submit the defect work order to the task center, where it will sequentially execute the acceptance, processing, pre-acceptance and acceptance process, and update the status of the defect work order in real time.

[0054] Based on the equipment running time and the preset lubrication standard, the lubrication cycle deviation parameter is calculated, and when the deviation exceeds the limit, the corresponding lubrication task is automatically generated.

[0055] During the execution of defect work orders and lubrication tasks, the system triggers early warning notifications according to preset rules, pushes the anomaly level and equipment code to the relevant responsible positions, and automatically summarizes and updates the assessment indicators based on the completion rate of defect work orders, response time and accuracy of anomaly handling.

[0056] Optionally, the generation of the statistical reports and audit logs specifically includes:

[0057] Based on equipment health scores, anomaly confidence levels, and defect work orders, an equipment operation dataset is compiled and synchronized to the equipment ledger database;

[0058] During the synchronization process, data exchange is performed through the enterprise management system interface, and the equipment technical parameters, maintenance history, running time and corresponding fields in the management system database are aligned and overwritten and updated.

[0059] After the update is completed, the system generates a structured index table based on the timestamp and device code;

[0060] Perform integrity and consistency checks on the equipment ledger database and obtain the check results;

[0061] Based on the structured index table and verification results, statistical reports are generated and audit logs are output. The statistical reports include the distribution of equipment health scores, the trend of abnormal confidence levels and the closure rate of defect work orders. The audit logs include the data synchronization time, the operator and the verification results.

[0062] The beneficial effects of this invention are:

[0063] This invention introduces an improved MixHop model into an enterprise-level equipment inspection system, constructing a fully intelligent system encompassing data acquisition, feature modeling, intelligent identification, and closed-loop feedback. This significantly improves the automation and analytical accuracy of inspection work. Compared to traditional systems relying on manual inspection, threshold judgment, and single-point data analysis, this invention achieves breakthroughs in three aspects: structuring, networking, and intelligence of inspection data. By establishing a unified point inspection standard library and coding system, standardized management of different equipment, components, and inspection projects within the enterprise is achieved. Data is traceable and correlated throughout its entire lifecycle, providing a solid data foundation for subsequent graph structure analysis.

[0064] In the data modeling and analysis phase, this invention constructs an inspection network graph based on equipment, inspection points, process flows, and spatial adjacency relationships, transforming previously isolated inspection data into a topologically structured, interconnected dataset. Furthermore, by utilizing an improved MixHop model and employing multi-order neighborhood diffusion and cross-order attention fusion mechanisms, it effectively captures the implicit relationship features and anomaly propagation paths between equipment and inspection points, enabling the model to maintain stable discrimination capabilities even in complex multi-equipment, multi-operating-condition environments. This model can calculate equipment health scores and anomaly confidence levels at the global level, achieving accurate quantitative assessment of equipment operating status and overcoming the limitations of traditional methods in learning multi-dimensional interconnected features.

[0065] The system design of this invention also achieves deep integration of the inspection process with enterprise management processes. The system can automatically perform anomaly identification and classification based on anomaly confidence level and exceeding threshold, generating defect entries and defect work orders, and linking with the task center to achieve a closed-loop flow of acceptance, handling, and verification, significantly reducing the risk of manual intervention and missed reports. Simultaneously, it automatically generates lubrication tasks based on equipment runtime and lubrication standards, realizing intelligent execution of preventative maintenance strategies. Through multi-channel early warning and assessment mechanisms, the system can push risk information to responsible positions in real time and automatically update assessment indicators based on task completion rate and handling accuracy, enabling enterprises to have data-driven evaluation and performance tracking capabilities in operation and maintenance management.

[0066] In terms of data interaction and decision support, this invention, through interface integration with enterprise management systems, achieves automatic synchronous updates of equipment health scores, anomaly confidence levels, and work order closed-loop status. The system can generate statistical reports and audit logs in real time, reflecting equipment operating trends and management performance, providing visual analysis support for decision-makers. Overall, this invention not only improves the utilization rate and analysis depth of inspection data but also establishes an integrated management mechanism covering "data collection—intelligent identification—anomaly handling—result feedback."

[0067] Therefore, this invention achieves a comprehensive upgrade of the enterprise point inspection system from "manual recording" to "intelligent sensing," from "independent data" to "correlation analysis," and from "manual response" to "intelligent closed loop." The system boasts advantages such as high recognition accuracy, strong operational stability, a complete management closed loop, and good data traceability. It can be widely applied to intelligent equipment operation and maintenance scenarios in industries such as metallurgy, power, petrochemicals, and manufacturing, and has significant promotional value and practical application significance. Attached Figure Description

[0068] 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:

[0069] Figure 1 This is an overall flowchart of an enterprise-level intelligent point inspection and identification analysis system based on graph neural networks proposed in this invention;

[0070] Figure 2 This is a schematic diagram of the module structure of an improved MixHop model for an enterprise-level intelligent point inspection and identification analysis system based on graph neural networks proposed in this invention. Detailed Implementation

[0071] 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.

[0072] refer to Figure 1-2 An enterprise-level intelligent point inspection and identification analysis system based on graph neural networks includes:

[0073] The point inspection standard management module is used to establish a point inspection standard library and a unified coding system. It classifies and organizes all equipment, components and point inspection items of the enterprise, assigns unique codes and sets over-limit thresholds, and generates point inspection routes.

[0074] The inspection task generation and execution module is used to configure RFID tags and register location information at each inspection point, generate a set of periodic inspection tasks based on the inspection route and send them to the handheld terminal. The handheld terminal completes RFID tag verification, collects inspection data and records timestamps, device codes and task numbers.

[0075] The data upload and analysis module is used to encapsulate the inspection data in a structured way, generate structured inspection data entries, upload them to the server through the enterprise network, and aggregate them into a trend sequence based on timestamps, equipment codes and task numbers.

[0076] The graph structure construction module is used to construct an inspection network graph based on trend sequences and equipment, inspection points, process flow and spatial adjacency relationships, generate node sets, edge sets and node attribute vectors, and form an inspection feature set;

[0077] The graph neural network analysis module is used to input the inspection feature set into the improved MixHop model, perform multi-order neighborhood diffusion coding, cross-order attention fusion and task discrimination and calibration, output node representation and edge representation, and calculate equipment health score and anomaly confidence.

[0078] The anomaly identification and work order management module is used to identify and classify anomalies based on anomaly confidence level and over-limit threshold, generate defect entries and automatically create defect work orders, and trigger early warning notifications and performance indicator updates according to preset rules.

[0079] The data synchronization and report generation module is used to synchronize equipment health scores, anomaly confidence levels, and defect work order records to the equipment ledger. It achieves data exchange through the enterprise management system interface, updates equipment technical parameters, maintenance history, and runtime, and generates statistical reports and audit logs.

[0080] In this embodiment, the modules are interconnected using the following method:

[0081] Establish a standard library and unified coding system for point inspection, assign unique codes to equipment, components and point inspection items and set over-limit thresholds to obtain point inspection routes;

[0082] RFID tags are configured at each inspection point and their location information is registered. Periodic inspection task sets are generated based on the inspection route and sent to handheld terminals.

[0083] The point inspection personnel use handheld terminals to read RFID tags to complete the on-site verification, collect point inspection data, and record the timestamp, equipment code and task number;

[0084] Structured inspection data entries are generated based on inspection data, uploaded to the server via the enterprise network, and aggregated into a trend sequence according to timestamp, equipment code, and task number;

[0085] Based on the trend sequence, an inspection network diagram is constructed on the server side according to equipment, inspection points, process flow and spatial adjacency, forming an inspection feature set;

[0086] The inspection feature set is input into the improved MixHop model to generate node representations and edge representations, and the equipment health score and anomaly confidence are calculated.

[0087] Based on the anomaly confidence level and the limit threshold, anomaly identification and classification are performed, defect entries are generated and defect work orders are automatically created, and early warning notifications and performance indicator updates are triggered according to preset rules.

[0088] The equipment health score, anomaly confidence level and defect work order are synchronized to the equipment ledger. Data exchange is carried out through the enterprise management system interface to update equipment technical parameters, maintenance history and running time, and generate statistical reports and audit logs.

[0089] In this embodiment, the establishment of the point inspection standard library and unified coding system is achieved by classifying and organizing all equipment, components and inspection items of the enterprise to form a unified data standard.

[0090] In this embodiment, the generation and distribution of the periodic point inspection task set specifically includes:

[0091] Configure an RFID tag for each inspection point, read the unique identification code of the RFID tag, and bind the unique identification code with the corresponding inspection point number, equipment code, and inspection item number;

[0092] Record the location information of each inspection point and form a location information set;

[0093] Based on the set of inspection routes and location information, all inspection points are sorted according to spatial distance and process dependence to generate an inspection sequence table;

[0094] Based on the inspection sequence relationship table, the route with the shortest total travel distance is selected from the set of feasible routes to obtain the optimal inspection route;

[0095] A set of periodic inspection tasks is generated based on the optimal inspection route and preset inspection cycle parameters, and the set of periodic inspection tasks is distributed to handheld terminals through the enterprise network.

[0096] In this embodiment, the collection of the point inspection data specifically includes:

[0097] The inspection personnel read the unique identification code of the radio frequency identification tag at the inspection point using a handheld terminal, and match and verify the unique identification code with the task number of the current inspection point to complete the on-site verification.

[0098] After the on-site verification is completed, the handheld terminal loads the device code and inspection item number bound to the inspection point and generates a data item sequence list;

[0099] The inspection data is collected in the order of the data items. The inspection data includes a set of measurement data, a set of meter reading data, and a set of observation data. The handheld terminal records the timestamp, equipment code, and task number.

[0100] In this embodiment, the formation of the trend sequence specifically includes:

[0101] The handheld terminal encapsulates the collected point inspection data and corresponding task information to generate structured inspection data entries. The structured inspection data entries include point inspection task number, equipment code, timestamp and inspection result parameters.

[0102] The structured inspection data items are uploaded to the server through the enterprise network. The server performs integrity verification on the received data and obtains the verified valid data.

[0103] The server writes the verified valid data into the inspection database to form standardized storage records, and automatically judges the data status according to the threshold of exceeding the limit. The results of exceeding the limit are marked as abnormal data and a preliminary alarm record is generated.

[0104] Based on the time series data of equipment codes and project numbers, the data is aggregated and sorted by timestamp to generate a trend series.

[0105] In this embodiment, the formation of the inspection feature set specifically includes:

[0106] A node set is obtained based on the equipment code and the inspection point number. The node set includes equipment nodes and inspection point nodes. An edge set is obtained based on the process flow dependency, spatial adjacency and historical collaboration. The edge set refers to the association between any two nodes.

[0107] For each node in the node set, a node attribute vector set is established. The node attribute vector set is formed by concatenating the statistical characteristics of the trend sequence corresponding to the node and the static attribute parameters of the node in a fixed order.

[0108] The weighted adjacency matrix is ​​calculated based on the edge set. The weighted adjacency matrix is ​​obtained by combining the process dependency relationship, spatial distance relationship and historical cooperation relationship in the edge set into three indicators and then weighting them.

[0109] The node set, edge set, node attribute vector set, and weighted adjacency matrix are concatenated to form the inspection network graph;

[0110] Based on the inspection network diagram, using the equipment code and the inspection item number as keys, the corresponding trend sequence is retrieved and aligned with the vectors in the node attribute vector set within a continuous time window to generate an inspection feature set.

[0111] In this embodiment, the calculation of the device health score and anomaly confidence level specifically includes:

[0112] The inspection feature set is input into the improved MixHop model, which includes a multi-order neighborhood diffusion coding module, a cross-order attention fusion module, and a task discrimination and calibration module. The multi-order neighborhood diffusion coding module refers to introducing an adaptive diffusion time parameter to diffuse the intermediate representations of nodes at each order and outputting a set of intermediate representations of nodes. The cross-order attention fusion module refers to calculating node representations and edge representations based on the set of intermediate representations of nodes. The task discrimination and calibration module refers to calculating the device health score based on the node representations and calculating the anomaly confidence based on the edge representations.

[0113] In the multi-order neighborhood diffusion coding module, each node in the inspection feature set is used to generate a node intermediate representation divided by order. An order-adaptive diffusion time parameter is introduced to diffuse the node intermediate representations of each order, and the node intermediate representation set is output.

[0114] In the cross-order attention fusion module, based on the set of intermediate node representations, the importance weights of each order are calculated and weighted fusion is performed to obtain the node representation. The edge representation is constructed based on the node representations at both ends of each edge and the edge attributes. The edge representation is formed by splicing the element-wise interaction results of the node representations at both ends and the edge attributes in a fixed order.

[0115] In the task discrimination and calibration module, the device health score is calculated based on the node representation, and the anomaly confidence score is calculated based on the edge representation. The device health score is obtained by multiplying the node representation with the weight matrix and adding a bias term for linear mapping, followed by Sigmoid normalization. The anomaly confidence score is obtained by multiplying the edge representation with the weight matrix and adding a bias term for linear mapping, followed by Sigmoid normalization.

[0116] In this embodiment, the triggering of the early warning notification and the updating of the assessment indicators specifically include:

[0117] Based on the device health score and the anomaly confidence level, each indicator is judged by exceeding the threshold, and samples with anomaly confidence levels greater than the threshold are marked as abnormal nodes.

[0118] Based on the distribution of abnormal nodes, perform anomaly identification and classification to obtain the anomaly level and generate the corresponding set of defect entries;

[0119] Based on the equipment code, inspection point number, and abnormality level of each defect item in the defect item set, a defect work order is automatically created, and the work order number, creation time, and responsible department information are recorded.

[0120] Submit the defect work order to the task center, where it will sequentially execute the acceptance, processing, pre-acceptance and acceptance process, and update the status of the defect work order in real time.

[0121] Based on the equipment running time and the preset lubrication standard, the lubrication cycle deviation parameter is calculated, and when the deviation exceeds the limit, the corresponding lubrication task is automatically generated.

[0122] During the execution of defect work orders and lubrication tasks, the system triggers early warning notifications according to preset rules, pushes the anomaly level and equipment code to the relevant responsible positions, and automatically summarizes and updates the assessment indicators based on the completion rate of defect work orders, response time and accuracy of anomaly handling.

[0123] In this embodiment, the generation of the statistical reports and audit logs specifically includes:

[0124] Based on equipment health scores, anomaly confidence levels, and defect work orders, an equipment operation dataset is compiled and synchronized to the equipment ledger database;

[0125] During the synchronization process, data exchange is performed through the enterprise management system interface, and the equipment technical parameters, maintenance history, running time and corresponding fields in the management system database are aligned and overwritten and updated.

[0126] After the update is completed, the system generates a structured index table based on the timestamp and device code;

[0127] Perform integrity and consistency checks on the equipment ledger database and obtain the check results;

[0128] Based on the structured index table and verification results, statistical reports are generated and audit logs are output. The statistical reports include the distribution of equipment health scores, the trend of abnormal confidence levels and the closure rate of defect work orders. The audit logs include the data synchronization time, the operator and the verification results.

[0129] Example 1:

[0130] This embodiment uses the continuous casting workshop of a large steel company as an application scenario. The workshop has 67 key production equipment, including hydraulic pump stations, cooling fans, casting machine drive shafts, and high-temperature circulating water pumps. The company's original point inspection system was a traditional handheld terminal inspection management software. Data was mainly entered manually. The system could only issue alarms for parameters exceeding limits based on fixed thresholds, and could not achieve correlation judgment and trend prediction under complex working conditions. Moreover, manual review and work order processing were time-consuming.

[0131] To enhance the intelligence level of inspection management, the company adopted the enterprise-level intelligent point inspection identification and analysis system based on graph neural networks proposed in this invention. The system is deployed on the company's private cloud server and integrates with the company's existing MES and EAM systems. After deployment, all devices and inspection points are uniquely identified and bound via RFID tags. The system automatically generates inspection routes and distributes them to the handheld terminals of inspection personnel. During the inspection process, inspection personnel only need to scan the RFID tag to verify their presence. The terminal automatically collects inspection data, including oil temperature, vibration, current, pressure, and speed, and uploads it to the server in real time.

[0132] The system encapsulates the inspection data in a structured manner on the server side, aggregating it by equipment code, project number, and timestamp to form time-series trend data. Subsequently, the server constructs an inspection network graph based on equipment topology, process flow dependencies, and spatial location relationships, mapping equipment, components, and inspection points as nodes, and pipeline connections, process couplings, and historical collaboration relationships as edges. Finally, it integrates trend features and static attribute features to form an inspection feature set.

[0133] During the analysis phase, the system invokes the improved MixHop model, achieving multi-layer feature propagation through a multi-order neighborhood diffusion mechanism and dynamically fusing information from different levels using a cross-order attention mechanism to generate node and edge representations, thereby calculating the health score and anomaly confidence for each device. The model training phase uses three months of historical inspection data as the training set, containing nearly 80,000 valid records. After optimization, the model's anomaly detection accuracy reaches 97.6%, with a false positive rate controlled below 2.1%, significantly higher than the accuracy of the enterprise's original anomaly detection based on fixed thresholds.

[0134] In actual operation, the system can automatically identify potential anomalies in the equipment's operating trends. For example, the outlet pressure of the cooling water pump did not exceed the limit for two consecutive shifts, but the improved MixHop model detected abnormal fluctuations in its node features in the high-order neighborhood diffusion. The model output anomaly confidence level was 0.9. The system automatically generated a defect work order and notified the maintenance team. Subsequent inspection revealed that the water pump coupling clearance was too large, and maintenance was carried out in advance, avoiding subsequent equipment downtime.

[0135] Meanwhile, the system achieves closed-loop inspection management based on anomaly identification. Defect work orders are automatically submitted to the task center, where the system assigns responsible departments and processing deadlines. After repairs are completed, the system automatically records the repair time, responsible person, and repair method, and generates a closed-loop record. Weekly, the system automatically summarizes the health score distribution, anomaly trend curve, and work order closure rate, generates statistical reports, and pushes them to the equipment management department, forming a complete closed loop of inspection data.

[0136] Table 1. Performance Evaluation of Intelligent Point Inspection and Identification Systems

[0137] Evaluation Project Original system This invention system Increase Average delay (minutes) for uploading inspection data 144 6 ↓95.8% Anomaly detection accuracy 81.4% 97.6% +16.2% False alarm rate 7.8% 2.1% ↓73.1% Average anomaly early detection time (hours) 3.2 22.4 +600% Number of unplanned outages (quarterly) 19 times 13 times ↓31.6% Average work order closed-loop time (hours) 42 19 ↓54.8% Maintenance labor costs (monthly) 186,000 yuan 157,000 yuan ↓15.6% System stable operating time (days) 78 90 (continuous) +15.4% Average equipment health score (0~1) 0.72 0.89 +23.6% Data traceability coverage 68% 100% +47.1%

[0138] As shown in Table 1, the system of this invention achieves significant performance improvements throughout the entire inspection process. First, the average upload latency of inspection data has been reduced from 144 minutes in the original system to 6 minutes. This change stems from the structured data encapsulation and the automatic synchronization mechanism of the enterprise network, ensuring the real-time nature of the inspection data. Second, the anomaly recognition model based on graph neural networks demonstrates outstanding performance in multi-dimensional feature fusion, increasing the anomaly recognition accuracy to 97.6% and reducing the false alarm rate to 2.1%, effectively avoiding human error and delayed alarms.

[0139] Regarding anomaly early warning capabilities, the system of this invention achieves early identification of abnormal trends by modeling the graph structure of relationships between devices. It can predict potential anomalies on average 22.4 hours in advance, a six-fold improvement over traditional methods. The number of unplanned equipment downtimes decreased from 19 per quarter to 13, indicating that the system not only improved prediction accuracy but also significantly enhanced equipment operational reliability. Simultaneously, the average work order closure time was shortened from 42 hours to 19 hours, thanks to the system's automatic generation and scheduling mechanism, which greatly improved operational response efficiency.

[0140] From an economic perspective, maintenance labor costs decreased by approximately 15.6%, primarily due to the automation of inspection task allocation, anomaly detection, and work order management, reducing manual review and repetitive inspections. The average equipment health score increased from 0.72 to 0.89, reflecting a significant improvement in overall equipment condition. The system operated stably for up to 90 days without failure, demonstrating high reliability and sustainability. Data traceability coverage reached 100%, ensuring transparency and traceability of inspection data, anomaly records, and work order flow throughout the entire process, providing data support for enterprise quality management and auditing.

[0141] In summary, the system of this invention, by introducing a graph neural network model and a unified coding system, has achieved significant application results in intelligent analysis, reliable early warning, and management closed loop. It has successfully solved the problems of data fragmentation, delayed identification, and decentralized management in existing inspection systems, and has high practicality and promotion value.

Claims

1. An enterprise-level intelligent point inspection and identification analysis system based on graph neural networks, characterized in that, include: The point inspection standard management module is used to establish a point inspection standard library and a unified coding system. It classifies and organizes all equipment, components and point inspection items of the enterprise, assigns unique codes and sets over-limit thresholds, and generates point inspection routes. The inspection task generation and execution module is used to configure RFID tags and register location information at each inspection point, generate a set of periodic inspection tasks based on the inspection route and send them to the handheld terminal. The handheld terminal completes RFID tag verification, collects inspection data and records timestamps, device codes and task numbers. The data upload and analysis module is used to encapsulate the inspection data in a structured way, generate structured inspection data entries, upload them to the server through the enterprise network, and aggregate them into a trend sequence based on timestamps, equipment codes and task numbers. The graph structure construction module is used to construct an inspection network graph based on trend sequences and equipment, inspection points, process flow and spatial adjacency relationships, generate node sets, edge sets and node attribute vectors, and form an inspection feature set; The graph neural network analysis module is used to input the inspection feature set into the improved MixHop model, perform multi-order neighborhood diffusion coding, cross-order attention fusion and task discrimination and calibration, output node representation and edge representation, and calculate equipment health score and anomaly confidence. The anomaly identification and work order management module is used to identify and classify anomalies based on anomaly confidence level and over-limit threshold, generate defect entries and automatically create defect work orders, and trigger early warning notifications and performance indicator updates according to preset rules. The data synchronization and report generation module is used to synchronize equipment health scores, anomaly confidence levels and defect work order records to the equipment ledger, realize data exchange through the enterprise management system interface, update equipment technical parameters, maintenance history and running time, and generate statistical reports and audit logs. The formation of the inspection feature set specifically includes: A node set is obtained based on the equipment code and the inspection point number. The node set includes equipment nodes and inspection point nodes. An edge set is obtained based on the process flow dependency, spatial adjacency and historical collaboration. The edge set refers to the association between any two nodes. For each node in the node set, a node attribute vector set is established. The node attribute vector set is formed by concatenating the statistical characteristics of the trend sequence corresponding to the node and the static attribute parameters of the node in a fixed order. The weighted adjacency matrix is ​​calculated based on the edge set. The weighted adjacency matrix is ​​obtained by combining the process dependency relationship, spatial distance relationship and historical cooperation relationship in the edge set into three indicators and then weighting them. The node set, edge set, node attribute vector set, and weighted adjacency matrix are concatenated to form the inspection network graph; Based on the inspection network diagram, using the equipment code and the inspection item number as keys, the corresponding trend sequence is retrieved and aligned with the vectors in the node attribute vector set within a continuous time window to generate an inspection feature set. The calculation of the device health score and anomaly confidence level specifically includes: The inspection feature set is input into the improved MixHop model, which includes a multi-order neighborhood diffusion coding module, a cross-order attention fusion module, and a task discrimination and calibration module. The multi-order neighborhood diffusion coding module refers to introducing an adaptive diffusion time parameter to diffuse the intermediate representations of nodes at each order and outputting a set of intermediate representations of nodes. The cross-order attention fusion module refers to calculating node representations and edge representations based on the set of intermediate representations of nodes. The task discrimination and calibration module refers to calculating the device health score based on the node representations and calculating the anomaly confidence based on the edge representations. In the multi-order neighborhood diffusion coding module, each node in the inspection feature set is used to generate a node intermediate representation divided by order. An order-adaptive diffusion time parameter is introduced to diffuse the node intermediate representations of each order, and the node intermediate representation set is output. In the cross-order attention fusion module, based on the set of intermediate node representations, the importance weights of each order are calculated and weighted fusion is performed to obtain the node representation. The edge representation is constructed based on the node representations at both ends of each edge and the edge attributes. The edge representation is formed by splicing the element-wise interaction results of the node representations at both ends and the edge attributes in a fixed order. In the task discrimination and calibration module, the device health score is calculated based on the node representation, and the anomaly confidence score is calculated based on the edge representation. The device health score is obtained by multiplying the node representation with the weight matrix and adding a bias term for linear mapping, followed by Sigmoid normalization. The anomaly confidence score is obtained by multiplying the edge representation with the weight matrix and adding a bias term for linear mapping, followed by Sigmoid normalization.

2. The enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to claim 1, characterized in that, The modules are connected in the following way: Establish a standard library and unified coding system for point inspection, assign unique codes to equipment, components and point inspection items and set over-limit thresholds to obtain point inspection routes; RFID tags are configured at each inspection point and their location information is registered. Periodic inspection task sets are generated based on the inspection route and sent to handheld terminals. The point inspection personnel use handheld terminals to read RFID tags to complete the on-site verification, collect point inspection data, and record the timestamp, equipment code and task number; Structured inspection data entries are generated based on inspection data, uploaded to the server via the enterprise network, and aggregated into a trend sequence according to timestamp, equipment code, and task number; Based on the trend sequence, an inspection network diagram is constructed on the server side according to equipment, inspection points, process flow and spatial adjacency, forming an inspection feature set; The inspection feature set is input into the improved MixHop model to generate node representations and edge representations, and the equipment health score and anomaly confidence are calculated. Based on the anomaly confidence level and the limit threshold, anomaly identification and classification are performed, defect entries are generated and defect work orders are automatically created, and early warning notifications and performance indicator updates are triggered according to preset rules. The equipment health score, anomaly confidence level and defect work order are synchronized to the equipment ledger. Data exchange is carried out through the enterprise management system interface to update equipment technical parameters, maintenance history and running time, and generate statistical reports and audit logs.

3. The enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to claim 2, characterized in that, The establishment of the inspection standard library and unified coding system is achieved by classifying and organizing all equipment, components and inspection items of the enterprise to form a unified data standard.

4. The enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to claim 2, characterized in that, The generation and distribution of the periodic point inspection task set specifically includes: Configure an RFID tag for each inspection point, read the unique identification code of the RFID tag, and bind the unique identification code with the corresponding inspection point number, equipment code, and inspection item number; Record the location information of each inspection point and form a location information set; Based on the set of inspection routes and location information, all inspection points are sorted according to spatial distance and process dependence to generate an inspection sequence table; Based on the inspection sequence relationship table, the route with the shortest total travel distance is selected from the set of feasible routes to obtain the optimal inspection route; A set of periodic inspection tasks is generated based on the optimal inspection route and preset inspection cycle parameters, and the set of periodic inspection tasks is distributed to handheld terminals through the enterprise network.

5. The enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to claim 2, characterized in that, The collection of the point inspection data specifically includes: The inspection personnel read the unique identification code of the radio frequency identification tag at the inspection point using a handheld terminal, and match and verify the unique identification code with the task number of the current inspection point to complete the on-site verification. After the on-site verification is completed, the handheld terminal loads the device code and inspection item number bound to the inspection point and generates a data item sequence list; The inspection data is collected in the order of the data items. The inspection data includes a set of measurement data, a set of meter reading data, and a set of observation data. The handheld terminal records the timestamp, equipment code, and task number.

6. The enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to claim 2, characterized in that, The formation of the trend sequence specifically includes: The handheld terminal encapsulates the collected point inspection data and corresponding task information to generate structured inspection data entries. The structured inspection data entries include point inspection task number, equipment code, timestamp and inspection result parameters. The structured inspection data items are uploaded to the server through the enterprise network. The server performs integrity verification on the received data and obtains the verified valid data. The server writes the verified valid data into the inspection database to form standardized storage records, and automatically judges the data status according to the threshold of exceeding the limit. The results of exceeding the limit are marked as abnormal data and a preliminary alarm record is generated. Based on the time series data of equipment codes and project numbers, the data is aggregated and sorted by timestamp to generate a trend series.

7. The enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to claim 2, characterized in that, The triggering of the early warning notification and the updating of the assessment indicators specifically include: Based on the device health score and the anomaly confidence level, each indicator is judged by exceeding the threshold, and samples with anomaly confidence levels greater than the threshold are marked as abnormal nodes. Based on the distribution of abnormal nodes, perform anomaly identification and classification to obtain the anomaly level and generate the corresponding set of defect entries; Based on the equipment code, inspection point number, and abnormality level of each defect item in the defect item set, a defect work order is automatically created, and the work order number, creation time, and responsible department information are recorded. Submit the defect work order to the task center, where it will sequentially execute the acceptance, processing, pre-acceptance and acceptance process, and update the status of the defect work order in real time. Based on the equipment running time and the preset lubrication standard, the lubrication cycle deviation parameter is calculated, and when the deviation exceeds the limit, the corresponding lubrication task is automatically generated. During the execution of defect work orders and lubrication tasks, the system triggers early warning notifications according to preset rules, pushes the anomaly level and equipment code to the relevant responsible positions, and automatically summarizes and updates the assessment indicators based on the completion rate of defect work orders, response time and accuracy of anomaly handling.

8. The enterprise-level intelligent point inspection and identification analysis system based on graph neural networks according to claim 2, characterized in that, The generation of the statistical reports and audit logs specifically includes: Based on equipment health scores, anomaly confidence levels, and defect work orders, an equipment operation dataset is compiled and synchronized to the equipment ledger database; During the synchronization process, data exchange is performed through the enterprise management system interface, and the equipment technical parameters, maintenance history, running time and corresponding fields in the management system database are aligned and overwritten and updated. After the update is completed, the system generates a structured index table based on the timestamp and device code; Perform integrity and consistency checks on the equipment ledger database and obtain the check results; Based on the structured index table and verification results, statistical reports are generated and audit logs are output. The statistical reports include the distribution of equipment health scores, the trend of abnormal confidence levels and the closure rate of defect work orders. The audit logs include the data synchronization time, the operator and the verification results.

Images