AGENT equipment health assessment system and method based on knowledge graph

Abstract

The present invention relates to the field of intelligent agent evaluation technology, and discloses an AGENT device health evaluation system and method based on a knowledge graph. The method includes the following steps: constructing a resource topology diagram, loading device status and resource capacity data, and providing a structured knowledge foundation for collaborative evaluation; device status includes health and priority; quantifying device resource requirements, generating a device-resource requirement matrix, and clarifying the dependencies of each device on various resources; a maintenance agent predicting device resource requirements, and a resource agent performing deadlock detection based on requirements and resource constraints, outputting a deadlock risk probability and a list of faulty devices. The present invention generates a dynamic scheduling strategy through collaborative optimization of maintenance agents and resource agents, monitors risks in real time, and triggers a rollback mechanism, forming a closed-loop knowledge graph update mechanism, enabling the system to have continuously evolving adaptive capabilities, effectively reducing operation and maintenance costs, and preventing failures caused by system decision conflicts.

Description

Technical Field

[0001] The present invention relates to the field of intelligent agent evaluation, and more specifically, to an agent device health evaluation system and method based on a knowledge graph. Background Art

[0002] With the development of industry and intelligent manufacturing, the health management of complex equipment systems faces challenges such as difficulty in integrating multi-source heterogeneous data, poor real-time evaluation, and weak cross-device collaboration.

[0003] The traditional single-agent (AGENT) model relies on centralized decision-making and is difficult to adapt to the dynamic coupling characteristics of distributed devices, while the static knowledge graph lacks the ability to respond to the evolution of device states in real time; especially when the decision-making layer of the agent model outputs multiple conflicting instructions at the same time, the optimization falls into a local optimum, causing the system to enter an infinite loop, resulting in decision delays or errors. Summary of the Invention

[0004] The present invention provides an AGENT equipment health assessment system and method based on a knowledge graph to solve the technical problems in related technologies of over-reliance on centralized decision-making, inability to analyze conflicts in decisions when making decisions, resulting in decision delays or incorrect decisions.

[0005] The present invention provides an AGENT device health assessment method based on a knowledge graph, comprising the following steps:

[0006] S100, knowledge graph initialization: building a resource topology map, loading device status and resource capacity data, and providing a structured knowledge foundation for collaborative evaluation;

[0007] Device status includes health and priority;

[0008] S200, Resource Demand Modeling: Quantify device resource requirements, generate a device-resource demand matrix, and clarify the dependency of each device on various resources;

[0009] S300, Collaborative Deadlock Prediction: The maintenance agent predicts device resource requirements, and the resource agent performs deadlock detection based on these requirements and resource constraints, outputting a deadlock risk probability and a list of faulty devices.

[0010] S400, Deadlock Avoidance Decision-Making: The maintenance agent optimizes the resource scheduling model, the resource agent quantifies the dynamic deadlock risk, and collaboratively generates the optimal scheduling strategy and maintenance instructions that balance reliability and risk;

[0011] S500, resource coordination execution: allocate resources according to the scheduling sequence, monitor deadlock risks in real time and trigger the rollback mechanism, and finally update the knowledge graph and feedback the execution results.

[0012] Furthermore, in S100, the following steps are specifically included:

[0013] S110, sensor data acquisition: collects real-time sensor data from edge devices, including physical quantities such as vibration, temperature, and current;

[0014] S120, sensor self-test: performs abnormality detection on each sensor channel and identifies failed sensors;

[0015] S130, constructing an anti-interference mask matrix: generating a mask matrix to shield the failed sensor data;

[0016] S140, Anti-interference Feature Extraction: Extract robust features through mask graph convolution;

[0017] S150, feature compression and transmission: dimensionality reduction processing and sending to subsequent intelligent agents.

[0018] Furthermore, in S200, the following steps are specifically included:

[0019] S210, heterogeneous graph construction: The compressed feature matrix output by the scheduling agent is integrated with the resource status to construct a device-resource heterogeneous graph;

[0020] S220, spatial dependency convolution: aggregating neighbor information through graph convolution to capture device-resource coupling relationships;

[0021] S230, temporal evolution modeling: integrating changes in historical state time windows to model dynamic evolution;

[0022] S240, event-triggered momentum update: respond to external events and modify the map in real time.

[0023] Furthermore, in S300, the following steps are specifically included:

[0024] S310, Equipment Health Assessment: Evaluate equipment health status based on a multi-objective optimized graph attention network;

[0025] S320, Dynamic Forecasting of Resource Demand: Combining historical data with current status to predict future resource demand;

[0026] S330, Deadlock Risk Assessment: Construct a resource contention graph to assess the system deadlock risk;

[0027] S340, Dynamic resource adjustment: Resource reallocation strategy based on multi-objective optimization.

[0028] Furthermore, in S330, the calculation formula for quantifying the deadlock risk is as follows:

[0029] Competition graph construction:

[0030] ;

[0031] in, is a collection of resource nodes, is the set of resource competition relationship edges, is the edge weight set, is a competition graph for resources;

[0032] Edge weight calculation:

[0033] ;

[0034] in, Representation device Occupied resources and request resources , is the total number of devices, For resources The available amount, Indicates a resource request and occupy resources The edge weights between

[0035] Deadlock Risk Score:

[0036] ;

[0037] ;

[0038] in, are high risk threshold and low risk threshold, respectively. It is a web page ranking algorithm used to evaluate the importance of nodes. It is the deadlock risk level, which is divided into three levels: High, Medium, and Low. Score the overall deadlock risk of the system;

[0039] Furthermore, in S400, the following steps are specifically included:

[0040] S410, Resource Scheduling Optimization Modeling: Construct an integer programming model to optimize resource scheduling solutions;

[0041] S420, Deadlock Risk Quantification: Calculate the degree of resource contention and quantify the deadlock probability;

[0042] S430, Multi-objective Decision Optimization: Combining reliability and deadlock risk to solve Pareto optimal solutions;

[0043] S440, dynamic scheduling strategy generation: generating resource scheduling sequences and preventive maintenance instructions.

[0044] Furthermore, in S420, the calculation formula for quantifying the deadlock risk is as follows:

[0045] Resource competition:

[0046] ;

[0047] in, is the competitiveness coefficient of resource k, is the fault flag of device i, is an indicator function, which is 1 when the condition is met, otherwise it is 0. For equipment and resources The demand matrix, is the total number of resource types, Index for resource type;

[0048] Deadlock risk factor:

[0049] ;

[0050] in, is the Sigmoid activation function, which maps the output to the [0,1] interval. is the deadlock risk coefficient, close to 1 indicates high risk, and close to 0 indicates low risk.

[0051] Furthermore, in S430, the calculation formula for multi-objective decision optimization is as follows:

[0052] Multi-objective function:

[0053] ;

[0054] ;

[0055] in, is the reliability score of device i, is the deadlock risk coefficient, is the resource scheduling decision vector, Score the health of device i, is the priority of device i, are the weight coefficients of health and priority respectively;

[0056] when When , the constraint adjustment:

[0057] ;

[0058] in, is the deadlock risk threshold, is the resource slack factor, is the predicted resource demand vector;

[0059] Solving optimization:

[0060] ;

[0061] ;

[0062] in, is the total reliability score of the system, is the optimal combination of decision variables.

[0063] Furthermore, in S500, the following steps are specifically included:

[0064] S510, resource scheduling execution: dynamically allocate resources according to the resource scheduling sequence of S440;

[0065] S520, deadlock detection after allocation: real-time detection of deadlock risk after each allocation;

[0066] S530, abnormal rollback mechanism: if deadlock risk is detected, the allocation operation is rolled back immediately;

[0067] S540, execute feedback update: generate resource scheduling report and update knowledge graph.

[0068] The present invention also proposes an AGENT device health assessment system based on a knowledge graph, which is used to perform the steps in the aforementioned AGENT device health assessment method based on a knowledge graph, including:

[0069] Knowledge graph initialization module: Builds a resource topology diagram, describes the physical / logical connection relationship between devices and resources, loads device status data, loads resource capacity data, establishes a structured knowledge base, and supports collaborative decision-making;

[0070] Resource Demand Modeling Module: Quantifies the device's demand for various resources, generates a device-resource demand matrix, clarifies resource dependencies between devices, and labels key resource paths;

[0071] Collaborative deadlock prediction module: The maintenance agent is used to predict the equipment resource demand sequence, and the resource agent is used to detect deadlock risks based on resource constraints, output deadlock risk probability, and generate a list of faulty equipment;

[0072] Deadlock avoidance decision module: This module optimizes the resource scheduling model through maintenance agents. Resource agents quantify dynamic deadlock risks, collaboratively generate optimal scheduling strategies, and output maintenance instruction sequences.

[0073] Resource coordination and execution module: allocates resources to devices according to the scheduling sequence, monitors deadlock risks in real time, triggers rollback mechanisms, updates knowledge graph status, and provides feedback on execution results.

[0074] The beneficial effects of the present invention are:

[0075] The present invention generates a dynamic scheduling strategy that takes into account both security and efficiency through precise modeling of resource requirements and forward-looking prediction of deadlock risks, through the collaborative optimization of maintenance agents and resource agents. During the execution process, risks are monitored in real time and a rollback mechanism is triggered, ultimately forming a closed-loop knowledge graph update mechanism. This enables the system to have continuously evolving adaptive capabilities, effectively reduces operation and maintenance costs, and prevents failures caused by system decision conflicts. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flow chart of the knowledge graph-based AGENT device health assessment method proposed in the present invention;

[0077] Figure 2 This is a module block diagram of the knowledge graph-based AGENT device health assessment system proposed in the present invention;

[0078] Figure 3 It is a multi-agent collaborative sequence diagram based on the health assessment method proposed in this invention.

[0079] In the figure: 101, knowledge graph initialization module; 102, resource demand modeling module; 103, collaborative deadlock prediction module; 104, deadlock avoidance decision module; 105, resource coordination execution module. DETAILED DESCRIPTION

[0080] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. In addition, features described with respect to some examples may also be combined in other examples.

[0081] like Figure 1 As shown in the figure, the knowledge graph-based AGENT device health assessment method includes the following steps:

[0082] S100, knowledge graph initialization: Build a resource topology map, load device status (health, priority) and resource capacity data, and provide a structured knowledge foundation for collaborative evaluation;

[0083] In one embodiment of the present invention, the following steps are specifically included:

[0084] S110, sensor data collection: uses the scheduling agent (SA) to collect real-time sensor data from edge devices, including physical quantities such as vibration, temperature, and current;

[0085] Raw sensor data matrix:

[0086] ;

[0087] in, is the total number of equipment nodes, such as the number of wind turbine gearboxes, is the total number of sensor channels, including different types of sensors such as temperature, vibration, and current. is the actual reading value of the j-th sensor of the i-th node at time t, is a real matrix space with N rows and D columns, represents the sensor data matrix at time t;

[0088] S120, sensor self-test (chi-square test): performs anomaly detection on each sensor channel to identify failed sensors;

[0089] Chi-square statistic calculation:

[0090] ;

[0091] in, is the actual observation value of the i-th sensor in the k-th time window, is the expected value calculated based on historical data, using the moving average method, To detect the time window length, the default setting is 10 time units. To indicate that the detection range covers all D sensor channels, represents the chi-square statistic of the i-th sensor, is the time window index, is the sensor channel index;

[0092] Sensor failure determination:

[0093] ;

[0094] in, is the failure judgment threshold, which is 3.84 at 95% confidence level. is the failure flag of the i-th sensor, 1 means failure, 0 means normal;

[0095] S130, constructing an anti-interference mask matrix: generating a mask matrix to shield the failed sensor data;

[0096] Node-level mask vector:

[0097]

[0098] ;

[0099] in, is the mask identifier of the jth node, which takes a value of 0 or 1. represents the failure state of the i-th sensor on node j, The result is 1 when all sensors are normal, otherwise it is 0. is the total number of sensor channels, Represents an N-dimensional mask vector, containing the sensor working status information of each node;

[0100] Adjacency mask matrix:

[0101] ;

[0102] Where M is the mask matrix, represents the N×N binary matrix space;

[0103] S140, Anti-interference Feature Extraction: Extract robust features through mask graph convolution;

[0104] ;

[0105] ;

[0106] in, To increase the self-connected adjacency matrix, is the original device topology matrix, is the N-dimensional identity matrix, is the Hadamard product operator, which means the corresponding elements of the matrix are multiplied. is the trainable weight matrix, , is the target feature dimension, is the ReLU activation function, which is used to introduce nonlinearity. is the extracted feature matrix with a dimension of N×F, is the degree matrix, which represents the connectivity degree of each node;

[0107] S150, feature compression and transmission: dimensionality reduction processing and sending to subsequent intelligent agents;

[0108] ;

[0109] ;

[0110] in, is the feature dimension after compression, is the maximum pooling operation for dimensionality reduction, F is the original feature dimension, represents the compressed feature matrix, is the feature matrix after extraction;

[0111] S200, resource demand modeling (performed by the Resource Agent (RA)): quantify device resource requirements, generate a device-resource demand matrix, and clarify the dependencies of each device on various resources;

[0112] In one embodiment of the present invention, the following steps are specifically included:

[0113] S210, heterogeneous graph construction: The compressed feature matrix output by SA is integrated with the resource status to construct a device-resource heterogeneous graph;

[0114] Define a node collection:

[0115] ;

[0116] in, It is a device node set, including N device nodes. A collection of resource nodes, including resource nodes (such as spare parts, engineers, etc.), N is the total number of equipment, is the total number of resource types, It is a complete set of nodes, including all device and resource nodes;

[0117] Adjacency matrix block:

[0118] ;

[0119] in, is the topological adjacency matrix between devices, which comes from the system configuration and has a dimension of N×N. is the equipment-resource demand relationship matrix, with dimension N× , is the resource-equipment relationship matrix, equal to , is the relationship matrix between resources, dimension × , is the complete adjacency matrix;

[0120] Initial node embedding:

[0121] ;

[0122] in, The compressed feature matrix output by SA, dimension N× , is the initial embedding matrix of resource nodes, dimension × , randomly initialized, is the feature dimension after compression, is the initial node embedding matrix, with dimension ;

[0123] S220, spatial dependency convolution: aggregating neighbor information through graph convolution to capture device-resource coupling relationships;

[0124] ;

[0125] ;

[0126] in, is the adjacency matrix with self-loops added, I is the identity matrix, is the degree matrix, is the spatial convolution weight matrix, dimension × , is the hidden layer dimension, , ReLU is the rectified linear unit activation function, is the feature matrix after spatial dependent convolution;

[0127] S230, Time Series Evolution Modeling: Fusion of Historical States Changes in time windows to model dynamic evolution;

[0128] ;

[0129] in, for The spatial feature matrix before time, i.e. the historical cache, is the GRU network parameter, including update gate and reset gate weights, is the time window length, GRU is the gated recurrent unit network, is the feature matrix enhanced by time series, dimension × ;

[0130] S240, event-triggered momentum update: responds to external events (such as maintenance work orders, environmental changes), and modifies the map in real time;

[0131] Event code:

[0132]

[0133] Among them, type is the event type code (such as maintenance = 1, fault = 2, etc.), priority is the event priority (1-5 integer), and target is the one-hot encoding vector of the target device ID. are the multi-layer perceptron parameters, is the event vector, whose dimension ;

[0134] Node status correction:

[0135] ;

[0136] in, is the event impact factor. When an emergency occurs, , when normal events, , is an indicator function, which is 1 when i is the target device / resource node, otherwise it is 0. is the feature vector of node i;

[0137] S300, Collaborative Deadlock Prediction (Maintenance Agent (MA) and RA Collaboration): The Maintenance Agent (MA) predicts device resource requirements. The RA combines these requirements with resource constraints to perform deadlock detection and output a deadlock risk probability and a list of faulty devices.

[0138] In one embodiment of the present invention, the following steps are specifically included:

[0139] S310, Equipment Health Assessment (performed by the Diagnostic Agent (DA)): Evaluate equipment health status based on a multi-objective optimized graph attention network;

[0140] Attention weight calculation:

[0141] ;

[0142] in, is the attention weight matrix, is the bias term, For nodes The eigenvector of For vector splicing, is the activation function of the linear rectifier unit with leakage;

[0143] Multi-head attention aggregation:

[0144] ;

[0145] ;

[0146] in, is the number of attention heads (default 8), For the The transformation matrix of the head, For nodes The neighbor set of is the activation function, For the Nodes under the attention head For Node The attention weight of

[0147] Health score:

[0148] ;

[0149] in, is a multi-layer perceptron, For nodes The aggregate feature vector of is the normalized health score;

[0150] Fault judgment threshold:

[0151] ;

[0152] in, is the health threshold (default 0.6), is the rate of change threshold (default 0.1), the amount of change in scoring healthiness;

[0153] S320, Dynamic Forecasting of Resource Demand (RA Execution): Combines historical data with current status to predict future resource demand;

[0154] Demand forecasting model:

[0155] ;

[0156] in, The resource requirements of the faulty device. is the historical resource usage rate, It is a seasonal characteristic. is the GRU network parameter, is a gated recurrent unit network;

[0157] Resource gap calculation:

[0158] ;

[0159] in, is the resource buffer (default 10%), is the current available resources, is the linear rectifier unit activation function;

[0160] S330, Deadlock Risk Assessment (DA→RA): Construct a resource contention graph to assess the system deadlock risk;

[0161] Competition graph construction:

[0162] ;

[0163] in, is a collection of resource nodes, is the set of resource competition relationship edges, is the edge weight set, is a competition graph for resources;

[0164] Edge weight calculation:

[0165] ;

[0166] in, Representation device Occupied resources and request resources , is the total number of devices, For resources The available amount, Indicates a resource request and occupy resources The edge weights between

[0167] Deadlock Risk Score:

[0168] ;

[0169] ;

[0170] in, are high and low risk thresholds (default 0.8, 0.4), It is a web page ranking algorithm used to evaluate the importance of nodes. The deadlock risk level is divided into three levels: High, Medium, and Low, which are used to guide resource allocation decisions. The deadlock risk score of the entire system is in the range of [0,1]. A larger value indicates a higher deadlock risk.

[0171] S340, Resource Dynamic Adjustment (RA Execution): Resource reallocation strategy based on multi-objective optimization;

[0172] Objective function:

[0173] ;

[0174] Constraints:

[0175] ;

[0176] in, is the resource scheduling decision vector, is the weight coefficient, A collection of faulty devices. For equipment Priority, is the equipment-resource requirement matrix;

[0177] Resource release execution:

[0178] ;

[0179] in, For equipment Resource release flag, 1 means release, 0 means keep, is the currently available resource vector, dimension ;

[0180] S400, Deadlock Avoidance Decision (MA and RA Collaboration): MA optimizes the resource scheduling model, RA quantifies the dynamic deadlock risk, and collaboratively generates the optimal scheduling strategy and maintenance instructions that balance reliability and risk;

[0181] In one embodiment of the present invention, the following steps are specifically included:

[0182] S410, Resource Scheduling Optimization Modeling (MA Execution): Build integer programming models to optimize resource scheduling solutions;

[0183] Decision variables:

[0184] ;

[0185] in, is the resource scheduling flag of the i-th device, is the total number of devices;

[0186] Objective function (maximize system reliability):

[0187] ;

[0188] in, The weight coefficients of health and priority respectively control the influence of health and priority (default ), is the health score of device i, ranging from [0,1], The priority of device i, integer value 1-5, is the resource scheduling decision vector, indicating whether to allocate resources to device i;

[0189] Resource constraints:

[0190] ;

[0191] in, is the equipment-resource requirement matrix, with dimensions , is the predicted resource demand vector, dimension , is the currently available resource vector, dimension , is the total number of resource types, Index for resource type;

[0192] S420, Deadlock Risk Quantification (RA Execution): Calculates the degree of resource contention and quantifies the deadlock probability;

[0193] Resource competition:

[0194] ;

[0195] in, is the competitiveness coefficient of resource k, is the fault flag of device i (0 or 1), is an indicator function, which is 1 when the condition is met and 0 otherwise;

[0196] Deadlock risk factor:

[0197] ;

[0198] in, is the Sigmoid activation function, which maps the output to the [0,1] interval. is the deadlock risk coefficient, close to 1 means high risk, close to 0 means low risk, is the competitiveness coefficient of resource k;

[0199] S430, Multi-Objective Decision Optimization (MA and RA Collaboration): Combine reliability and deadlock risk to solve Pareto optimal solutions;

[0200] Multi-objective function:

[0201] ;

[0202] ;

[0203] in, is the reliability score of device i, is the deadlock risk coefficient, is the resource scheduling decision vector, is the priority of device i;

[0204] Constraint adjustment (when hour):

[0205] ;

[0206] in, is the deadlock risk threshold, is the resource slack factor, is the predicted resource demand vector, is the current available amount of resource type k;

[0207] Solving optimization:

[0208] ;

[0209] ;

[0210] in, is the total reliability score of the system, is the optimal combination of decision variables;

[0211] S440, dynamic scheduling strategy generation (MA execution): generating resource scheduling sequences and preventive maintenance instructions;

[0212] Allocation strategy:

[0213] ;

[0214] in, : The set of devices to which resources need to be allocated, : Optimized decision variable value

[0215] Maintenance Instructions:

[0216] ;

[0217] in, is the maintenance instruction type of equipment i, Score the health of device i, is the system deadlock risk factor;

[0218] Resource scheduling sequence:

[0219] ;

[0220] in, is the scheduling weight coefficient (default ), The device sequence for resource scheduling, Returns the sorted index sequence

[0221] S500, Resource Coordination Execution (RA Execution): Allocate resources according to the scheduling sequence, monitor deadlock risks in real time and trigger the rollback mechanism, and finally update the knowledge graph and feedback the execution results;

[0222] S510, resource scheduling execution (RA execution): dynamically allocate resources according to the resource scheduling sequence of S440;

[0223] Traverse the dispatch sequence:

[0224]

[0225] ;

[0226] ;

[0227] in, is the scheduling sequence, The total number of devices to which resources need to be allocated, The index of the device currently being processed;

[0228] Resource scheduling operations:

[0229] ;

[0230] in, For equipment The resource demand vector, is the available resource vector, is the assignment operator;

[0231] Allocation status flag:

[0232] ;

[0233] in, For equipment Resource scheduling status, 1 means allocated, 0 means unallocated;

[0234] S520, deadlock detection after allocation (RA execution): real-time detection of deadlock risk after each allocation;

[0235] Resource contention update:

[0236] ;

[0237] in, For resources The latest competition A collection of devices that need to allocate resources. For equipment Resources demand, For resources The current available amount, is the total number of resource types;

[0238] Deadlock risk reassessment:

[0239] ;

[0240] in, is the updated deadlock risk coefficient, is the Sigmoid activation function, used to normalize the risk value, is an indicator function, which is 1 when the competitiveness is greater than 1, otherwise it is 0

[0241] Risk Assessment:

[0242] ;

[0243] in, Deadlock risk flag, 1 means there is risk, 0 means there is no risk, is the deadlock risk threshold, is the indicator function;

[0244] S530, abnormal rollback mechanism (RA execution): if deadlock risk is detected, the allocation operation is rolled back immediately;

[0245] Rollback conditions:

[0246] ;

[0247] in, It is a deadlock risk flag;

[0248] Resource recovery operation:

[0249] ;

[0250] in, is the available resource vector, For devices The resource demand vector of

[0251] State reset and sequence update:

[0252] ;

[0253] in, For equipment The allocation status mark, To remove a device from the scheduling sequence Operation;

[0254] S540, execution feedback update (RA execution): generate resource scheduling report and update knowledge graph;

[0255] Allocation result statistics:

[0256] ;

[0257] in, is the total amount of allocated resources vector, For equipment The allocation status, For equipment The resource demand vector of

[0258] Remaining resources calculation:

[0259] ;

[0260] in, is the system remaining resource vector, is the currently available resource vector;

[0261] Knowledge Graph Updates:

[0262] ;

[0263] in, is the resource-related knowledge graph subgraph, For the update operation function, The remaining resource data for updating.

[0264] like Figure 2 As shown, based on the steps in the above evaluation method, an AGENT equipment health evaluation system based on knowledge graph is also proposed, which includes the following modules:

[0265] Knowledge Graph Initialization Module 101: Build a resource topology diagram to describe the physical / logical connection between devices and resources, load device status data (health, priority), load resource capacity data (total available resources, type), establish a structured knowledge base, and support collaborative decision-making;

[0266] Resource Demand Modeling Module 102: Quantify the device's demand for various resources, generate a device-resource demand matrix (devices as rows, resources as columns), clarify resource dependencies between devices (mutually exclusive / shared), and mark critical resource paths (highly competitive resources);

[0267] Collaborative deadlock prediction module 103: The maintenance agent (MA) is used to predict the device resource demand sequence, and the resource agent (RA) is used to detect deadlock risks in combination with resource constraints, output the deadlock risk probability (high / medium / low), and generate a list of faulty devices (devices that may be stalled due to resource competition);

[0268] Deadlock avoidance decision module 104: The maintenance agent (MA) optimizes the resource scheduling model (priority scheduling / resource reservation). The resource agent (RA) quantifies the dynamic deadlock risk (real-time risk coefficient), collaboratively generates the optimal scheduling strategy (balancing reliability and risk), and outputs the maintenance instruction sequence (equipment execution order + resource scheduling plan).

[0269] Resource coordination execution module 105: allocates resources to devices according to the scheduling sequence, monitors deadlock risks in real time (dynamically detects resource competition), triggers the rollback mechanism (cancels allocation when the risk exceeds the threshold), updates the knowledge graph status (device health / resource margin), and feedbacks the execution results (success / failure / rollback events).

[0270] like Figure 3 As shown, based on the above method and system, the following example is given: a wind farm maintenance scenario, in which Figure 3 This is the multi-agent collaboration sequence diagram based on the method in this scenario;

[0271] Scenario description: A wind farm needs to perform maintenance on three wind turbines (WT1, WT2, and WT3). The resources involved include: power resources (to drive the maintenance robot), maintenance tool kits (robotic arms, detectors), and spare parts (gears and bearings).

[0272] Division of labor among agents:

[0273] Maintenance Agent (MA): Develops maintenance plans and assesses turbine health;

[0274] Resource Agent (RA): manages resource scheduling and monitors deadlock risks;

[0275] Implementation steps and content:

[0276] Step 1: Knowledge Graph Initialization (RA Execution)

[0277] Build a wind farm knowledge graph, including:

[0278] Equipment nodes: wind turbines (WT1, WT2, WT3), maintenance robots (Robot1, Robot2);

[0279] Resource nodes: electricity, tool kits, spare parts;

[0280] Relationship: WT1 needs a toolkit, Robot1 needs electricity;

[0281] Step 2: Resource Requirements Mapping (MA Execution)

[0282] MA generates a resource requirement matrix based on the maintenance plan, as shown in Table 1:

[0283] Table 1: Resource requirement matrix diagram

[0284]

[0285] Step 3: Collaborative deadlock detection (MA and RA collaboration)

[0286] RA detection resource competition: the toolkit is required by both WT1 and Robot1 (total quantity = 1, required quantity = 2);

[0287] MA calculates the health of the wind turbines: WT1 health is low (needs priority maintenance);

[0288] Output deadlock risk: toolkit competition leads to high risk;

[0289] Step 4: Deadlock avoidance decision (MA and RA collaboration)

[0290] MA optimization allocation: prioritize allocating toolkits to WT1 (low health);

[0291] RA quantitative risk: toolkit competing risk coefficient > threshold;

[0292] Collaborative decision-making: Robot1 delays execution to release toolkit resources;

[0293] Step 5: Resource coordination execution (RA execution)

[0294] RA was assigned by sequence: toolkit → WT1;

[0295] Deadlock monitoring: After allocation, the contention risk is reduced to a safe range;

[0296] Updated map: Remaining toolkits = 0, marking WT1 maintenance completed.

[0297] Based on the above steps, the toolkit competition risk was proactively resolved, WT1 maintenance tasks were completed without interruption, Robot1 execution was delayed to avoid resource waste, and power resources were reallocated to WT2 maintenance. No deadlock occurred throughout the entire process, and maintenance tasks were carried out in an orderly manner according to health priority.

[0298] The above describes the embodiments of the present invention, but the present invention is not limited to the above specific implementation methods. The above specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make many forms under the guidance of the present invention, all of which are protected by the present invention.

Claims

1. The knowledge graph-based AGENT device health assessment method is characterized by: The following steps are involved: S100, knowledge graph initialization: building a resource topology map, loading device status and resource capacity data, and providing a structured knowledge foundation for collaborative evaluation; Device status includes health and priority; S200, Resource Demand Modeling: Quantify device resource requirements, generate a device-resource demand matrix, and clarify the dependency of each device on various resources; S300, Collaborative Deadlock Prediction: The maintenance agent predicts device resource requirements, and the resource agent performs deadlock detection based on these requirements and resource constraints, outputting a deadlock risk probability and a list of faulty devices. In S300, the following steps are specifically included: S310, Equipment Health Assessment: Evaluate equipment health status based on a multi-objective optimized graph attention network; S320, Dynamic Forecasting of Resource Demand: Combining historical data with current status to predict future resource demand; S330, Deadlock Risk Assessment: Construct a resource contention graph to assess the system deadlock risk; In S330, the calculation formula for quantifying the deadlock risk is as follows: Competition graph construction: ； in, is a collection of resource nodes, is the set of resource competition relationship edges, is the edge weight set, is a competition graph for resources; Edge weight calculation: ； in, Representation device Occupied resources and request resources , is the total number of devices, For resources The available amount, Indicates a resource request and occupy resources The edge weights between Deadlock Risk Score: ； ； in, are high risk threshold and low risk threshold, respectively. It is a web page ranking algorithm used to evaluate the importance of nodes. It is the deadlock risk level, which is divided into three levels: High, Medium, and Low. Score the overall deadlock risk of the system; S340, Dynamic Resource Adjustment: Resource reallocation strategy based on multi-objective optimization; S400, Deadlock Avoidance Decision-Making: The maintenance agent optimizes the resource scheduling model, the resource agent quantifies the dynamic deadlock risk, and collaboratively generates the optimal scheduling strategy and maintenance instructions that balance reliability and risk; S500, resource coordination execution: allocate resources according to the scheduling sequence, monitor deadlock risks in real time and trigger the rollback mechanism, and finally update the knowledge graph and feedback the execution results.

2. The AGENT device health assessment method based on knowledge graph according to claim 1 is characterized in that: In S100, the following steps are specifically included: S110, sensor data acquisition: collects real-time sensor data from edge devices, including physical quantities such as vibration, temperature, and current; S120, sensor self-test: performs abnormality detection on each sensor channel and identifies failed sensors; S130, constructing an anti-interference mask matrix: generating a mask matrix to shield the failed sensor data; S140, Anti-interference Feature Extraction: Extract robust features through mask graph convolution; S150, feature compression and transmission: dimensionality reduction processing and sending to subsequent intelligent agents.

3. The AGENT device health assessment method based on knowledge graph according to claim 2 is characterized in that: In S200, the following steps are specifically included: S210, heterogeneous graph construction: The compressed feature matrix output by the scheduling agent is integrated with the resource status to construct a device-resource heterogeneous graph; S220, spatial dependency convolution: aggregating neighbor information through graph convolution to capture device-resource coupling relationships; S230, temporal evolution modeling: integrating changes in historical state time windows to model dynamic evolution; S240, event-triggered momentum update: respond to external events and modify the map in real time.

4. The AGENT device health assessment method based on knowledge graph according to claim 3 is characterized in that: In S400, the following steps are specifically included: S410, Resource Scheduling Optimization Modeling: Construct an integer programming model to optimize resource scheduling solutions; S420, Deadlock Risk Quantification: Calculate the degree of resource contention and quantify the deadlock probability; S430, Multi-objective Decision Optimization: Combining reliability and deadlock risk to solve Pareto optimal solutions; S440, dynamic scheduling strategy generation: generating resource scheduling sequences and preventive maintenance instructions.

5. The AGENT device health assessment method based on knowledge graph according to claim 4 is characterized in that: In S420, the calculation formula for quantifying the deadlock risk is as follows: Resource competition: ； in, is the competitiveness coefficient of resource k, is the fault flag of device i, is an indicator function, which is 1 when the condition is met, otherwise it is 0. For equipment and resources The demand matrix, is the total number of resource types, Index for resource type; Deadlock risk factor: ； in, is the Sigmoid activation function, which maps the output to the [0,1] interval. is the deadlock risk coefficient, close to 1 indicates high risk, and close to 0 indicates low risk.

6. The knowledge graph-based AGENT device health assessment method according to claim 5 is characterized in that: In S430, the calculation formula for multi-objective decision optimization is as follows: Multi-objective function: ； ； in, is the reliability score of device i, is the deadlock risk coefficient, is the resource scheduling decision vector, Score the health of device i, is the priority of device i, are the weight coefficients of health and priority respectively; when When , the constraint adjustment: ； in, is the deadlock risk threshold, is the resource slack factor, is the predicted resource demand vector; Solving optimization: ； ； in, is the total reliability score of the system, is the optimal combination of decision variables.

7. The knowledge graph-based AGENT device health assessment method according to claim 6 is characterized in that: In S500, the following steps are specifically included: S510, resource scheduling execution: dynamically allocate resources according to the resource scheduling sequence of S440; S520, deadlock detection after allocation: real-time detection of deadlock risk after each allocation; S530, abnormal rollback mechanism: if deadlock risk is detected, the allocation operation is rolled back immediately; S540, execute feedback update: generate resource scheduling report and update knowledge graph.

8. The AGENT equipment health assessment system based on knowledge graph is characterized by: The method for performing the steps of the knowledge graph-based AGENT device health assessment method according to any one of claims 1 to 7 comprises: Knowledge graph initialization module: Builds a resource topology diagram, describes the physical / logical connection relationship between devices and resources, loads device status data, loads resource capacity data, establishes a structured knowledge base, and supports collaborative decision-making; Resource Demand Modeling Module: Quantifies the device's demand for various resources, generates a device-resource demand matrix, clarifies resource dependencies between devices, and labels key resource paths; Collaborative deadlock prediction module: The maintenance agent is used to predict the equipment resource demand sequence, and the resource agent is used to detect deadlock risks based on resource constraints, output deadlock risk probability, and generate a list of faulty equipment; Deadlock avoidance decision module: This module optimizes the resource scheduling model through maintenance agents. Resource agents quantify dynamic deadlock risks, collaboratively generate optimal scheduling strategies, and output maintenance instruction sequences. Resource coordination and execution module: allocates resources to devices according to the scheduling sequence, monitors deadlock risks in real time, triggers rollback mechanisms, updates knowledge graph status, and provides feedback on execution results.

Images