An electrical equipment anomaly diagnosis method and system based on a multi-agent system

Abstract

The application relates to an electrical equipment abnormality diagnosis multi-agent method based on dynamic task adaptation and federal learning. By introducing a dynamic task adaptation mechanism and a federal learning collaborative diagnosis, an electrical equipment abnormality diagnosis multi-agent system capable of self-adaptation, privacy protection and continuous learning is constructed. The scheme can select the optimal diagnosis strategy for different abnormal characteristics through the dynamic task adaptation mechanism, thereby improving the diagnosis accuracy of composite faults and new abnormalities. Through the federal learning technology, the agents of different power sites can jointly train the diagnosis model without sharing the original data, thereby solving the data island problem and guaranteeing data privacy and security. Meanwhile, the dynamic task adaptation mechanism enables the system to autonomously adjust the diagnosis strategy according to the abnormal characteristics, thereby improving the adaptability of the system to diversified abnormal types.

Description

Technical Field

[0001] This application relates to the field of power plant maintenance and management, and in particular to a method, system, computer device, and computer-readable storage medium for diagnosing electrical equipment anomalies based on a multi-agent system. Background Technology

[0002] With the development of smart grids, a large number of monitoring devices are deployed in power systems, generating massive amounts of multimodal operational data. Existing electrical equipment anomaly diagnosis technologies face numerous challenges:

[0003] First, the diagnostic models have poor adaptability: Most existing diagnostic systems use fixed diagnostic models or rules, which cannot dynamically adjust diagnostic strategies based on the real-time status of the equipment, data characteristics, and fault evolution patterns. When new anomalies or complex faults occur, the diagnostic accuracy drops significantly.

[0004] Second, data silos and privacy risks: Existing diagnostic methods based on centralized data processing require the aggregation of monitoring data from various sites to a central server, which poses a risk of data privacy leakage and is also limited by data transmission restrictions across security zones.

[0005] Third, insufficient collaborative diagnostic capabilities of equipment: Although existing multi-agent-based diagnostic systems have achieved hierarchical diagnosis, the collaborative mechanism between agents is fixed and rigid, lacking the ability to dynamically organize collaborative strategies according to the characteristics of diagnostic tasks, resulting in low diagnostic efficiency.

[0006] Fourth, there are challenges in knowledge sharing and continuous learning: the limited number of fault samples at a single power station makes it difficult to continuously optimize the diagnostic model. Furthermore, traditional centralized learning methods face limitations related to data privacy and transmission bandwidth.

[0007] Therefore, there is an urgent need to design a new system and method with multi-agent collaborative diagnostic capabilities. Summary of the Invention

[0008] This application provides a method, system, computer device, and computer-readable storage medium for diagnosing electrical equipment anomalies based on a multi-agent system, in order to at least solve the problem of poor adaptability of diagnostic models in related technologies.

[0009] In a first aspect, embodiments of this application provide an electrical equipment anomaly diagnosis method based on a multi-agent system, implemented based on a global diagnosis model composed of equipment diagnosis agents, unit collaborative diagnosis agents, and system federated learning agents. The method includes:

[0010] Obtain abnormal data of electrical equipment obtained by the device diagnostic agent based on multimodal sensor data for preliminary detection, and extract multidimensional features from the abnormal data;

[0011] By using the dynamic task adaptation engine within the unit collaborative diagnostic intelligent body, a reinforcement learning strategy is adopted to generate the optimal diagnostic strategy based on the multi-dimensional features, and the propagation path of the abnormal data between electrical devices is predicted by utilizing the device association graph and graph convolutional network model.

[0012] Based on the propagation path, a diagnostic agent alliance is formed, consisting of multiple diagnostic agents; through this alliance, collaborative diagnosis is performed according to the optimal diagnostic strategy, and diagnostic results are generated.

[0013] Specifically, the system uses a federated learning agent to coordinate multiple units to collaboratively diagnose the agent and perform federated learning to update the global diagnostic model.

[0014] In some embodiments, extracting multi-dimensional features from the anomalous data includes:

[0015] Extract the temporal features of the abnormal data, wherein the temporal features include standard deviation, skewness, and kurtosis.

[0016] Extract the frequency domain features of the abnormal data, including the dominant frequency, spectral centroid, and spectral entropy;

[0017] Extract the time-frequency domain features of the abnormal data, wherein the time-frequency domain features are wavelet transform energy distributions;

[0018] In addition, multi-source consistency features are extracted, which are cross-correlation features between multi-modal sensor data;

[0019] The time-domain features, the frequency-domain features, the time-frequency-domain features, and the cross-correlation features are used as the multi-dimensional features.

[0020] In some embodiments, the dynamic task adaptation engine within a unit-cooperative diagnostic intelligent body, employing a reinforcement learning strategy, generates an optimal diagnostic strategy based on the multi-dimensional features, including:

[0021] The multi-dimensional features and device criticality levels are co-encoded to generate states for reinforcement learning;

[0022] Policy optimization is performed using a state-action value function constructed based on the stated state, and the optimal diagnostic policy is output by iteratively updating the state-action value function.

[0023] Specifically, the target value is obtained based on the immediate reward used to evaluate the quality of the diagnostic action and the maximum expected future reward of the next state after discounting factor adjustment. The state-action value function is then corrected using the temporal difference error between the target value and the current value of the state-action value function and a preset learning rate.

[0024] In the process of generating the optimal diagnostic strategy, the merits of the diagnostic actions are evaluated through a multi-objective optimization function, which is constructed based on diagnostic accuracy, diagnostic response time, and diagnostic resource consumption.

[0025] In some embodiments, predicting the propagation path of the anomalous data among electrical devices using device association graphs and graph convolutional network models includes:

[0026] Construct an equipment association graph using the physical connections and functional dependencies between electrical equipment as edges;

[0027] Anomaly propagation prediction is performed using a graph convolutional network model and the device association graph to determine the propagation path of the abnormal data among electrical devices.

[0028] In some embodiments, anomaly propagation prediction is performed using a graph convolutional network model and the device association graph, predicting the propagation path of the anomalous data among electrical devices, including:

[0029] In the input layer of the graph convolutional network model, the initial state features of all electrical devices are defined as baseline data, and in the device association graph, the node features of the detected abnormal electrical devices are modified to abnormal states.

[0030] Through multiple graph convolutional layers of the graph convolutional network model, multi-layer information propagation and feature aggregation are performed based on the data of the input layer. In each layer, feature transformation is performed on the feature matrix of the node in the previous layer, and aggregation operation is performed with the normalized device association graph to propagate abnormal data from the abnormal source node to the same-order neighbor devices and higher-order neighbor devices layer by layer.

[0031] The final state is output through the output layer of the graph convolutional network model, and based on the final state, the probability score of each electrical device being affected by the abnormal data is determined.

[0032] The propagation path is determined by identifying electrical devices in the device association map whose probability score of being affected is higher than a preset threshold.

[0033] In some embodiments, coordinating multiple unit-based collaborative diagnostic agents to perform federated learning to update the global diagnostic model via a system federated learning agent includes:

[0034] The system's federated learning agent distributes global diagnostic model parameters to multiple unit collaborative diagnostic agents.

[0035] The unit collaborative diagnostic agent loads the global diagnostic model parameters into the local model, updates the local model based on local diagnostic case data, and uploads the updated local model to the system federated learning agent.

[0036] The system's federated learning agent aggregates various updated local models to generate an updated global diagnostic model. A differentiated federated aggregation algorithm is used to calculate the client weight of each unit's collaborative diagnostic agent in the global diagnostic model aggregation process based on preset indicators. The updated global diagnostic model is then generated by weighting the updated local models based on the client weights.

[0037] In some embodiments, a differentiated federated aggregation algorithm is used to calculate the client weight of each unit collaborative diagnostic agent in the global diagnostic model aggregation process based on preset indicators, including:

[0038] Determine the first, second, and third preset weighting coefficients for balancing data volume, data quality, and data reliability, respectively;

[0039] The first parameter is obtained based on the ratio of the number of valid diagnostic samples stored locally by the unit collaborative diagnostic agent to the total number of all participating samples, and the first preset weighting coefficient.

[0040] Based on the integrity parameters, consistency parameters, and timeliness parameters of the local data of the unit collaborative diagnostic agent, a data quality score is calculated, and a second parameter is obtained according to the data quality score and the second preset weighting coefficient.

[0041] Based on the historical diagnostic accuracy and response stability of the unit collaborative diagnostic agent, a comprehensive reliability score is calculated. Based on the reliability score and the third preset weighting coefficient, a third parameter is obtained.

[0042] Based on the first, second, and third parameters of the unit collaborative diagnostic agent, its corresponding client weight in the global diagnostic model aggregation process is obtained.

[0043] In some embodiments, the method further includes:

[0044] Before the unit collaborative diagnostic agent uploads the local model update, differential privacy technology is used to add Gaussian noise to the original local model parameters to generate privacy-preserving model parameters for uploading.

[0045] The Gaussian noise has zero mean and is generated based on the identity matrix. The noise scale of the Gaussian noise is determined according to a preset privacy budget and sensitivity. The sensitivity is used to measure the degree of influence of the presence of a single training sample on the local model parameters.

[0046] Secondly, embodiments of this application provide an electrical equipment anomaly diagnosis system based on a multi-agent system, implemented using a global diagnosis model composed of equipment diagnosis agents, unit collaborative diagnosis agents, and system federated learning agents. The system includes:

[0047] The acquisition module is used to acquire electrical equipment abnormality data obtained by the device diagnostic agent based on multimodal sensor data for preliminary detection, and to extract multi-dimensional features from the abnormality data;

[0048] The prediction module is used to diagnose the dynamic task adaptation engine within the intelligent body through unit collaborative diagnosis. It adopts a reinforcement learning strategy to generate the optimal diagnosis strategy based on the multi-dimensional features and uses the device association graph and graph convolutional network model to predict the propagation path of the abnormal data between electrical devices.

[0049] The diagnostic module, based on the propagation path, establishes a diagnostic agent alliance consisting of multiple diagnostic agents; through the diagnostic agent alliance, it performs collaborative diagnosis according to the optimal diagnostic strategy and generates diagnostic results.

[0050] An optimization module is used to coordinate multiple unit collaborative diagnostic agents through system federated learning agents to perform federated learning to update the global diagnostic model.

[0051] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0052] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0053] Compared to related technologies, this application provides a multi-agent method for diagnosing electrical equipment anomalies based on dynamic task adaptation and federated learning. By introducing two core technologies—dynamic task adaptation and federated learning collaborative diagnosis—it constructs an adaptive, privacy-preserving, and continuously learning multi-agent system for diagnosing electrical equipment anomalies. This solution, through the dynamic task adaptation mechanism, can select the optimal diagnostic strategy for different anomaly characteristics, improving the diagnostic accuracy of complex faults and novel anomalies. Through federated learning technology, agents from different power stations can jointly train diagnostic models without sharing raw data, solving the data silo problem and ensuring data privacy and security. Simultaneously, the dynamic task adaptation mechanism enables the system to autonomously adjust its diagnostic strategy based on anomaly characteristics, significantly improving the system's adaptability to diverse anomaly types. Through a dynamic alliance formation mechanism oriented towards anomaly propagation paths, it can predictively organize diagnostic resources, shortening the diagnosis time for complex faults. Attached Figure Description

[0054] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0055] Figure 1 This is a schematic diagram of a system overall architecture according to an embodiment of this application;

[0056] Figure 2 This is a flowchart of an electrical equipment anomaly diagnosis method based on dynamic task adaptation and federated learning according to an embodiment of this application;

[0057] Figure 3 This is a structural block diagram of a multi-agent system for electrical equipment anomaly diagnosis based on dynamic task adaptation and federated learning, according to an embodiment of this application.

[0058] Figure 4 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0060] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0061] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0062] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0063] Existing electrical equipment anomaly diagnosis technologies face problems such as poor adaptability of diagnostic models, data silos and privacy risks, insufficient collaborative diagnosis capabilities, and difficulties in knowledge sharing and continuous learning. In view of this, embodiments of this application provide an electrical equipment anomaly diagnosis method and system based on dynamic task adaptation and federated learning.

[0064] Figure 1 This is a schematic diagram of a system overall architecture according to an embodiment of this application, such as... Figure 1 As shown, the system of the present invention adopts a four-layer closed-loop architecture of perception-diagnosis-decision-learning, and the system mainly includes:

[0065] Data Sensing Layer: This layer consists of a sensor network deployed on various electrical devices (such as vibration sensors, temperature sensors, partial discharge sensors, current and voltage sensors, etc.). This layer is responsible for real-time acquisition of multimodal operating data, such as vibration signals, temperature, partial discharge, current and voltage waveforms, etc.

[0066] Intelligent Agent Collaborative Diagnostic Layer: This layer is the core of the system and consists of three types of core intelligent agents:

[0067] Equipment Diagnostic Agent (DAgent): Each DAgent corresponds one-to-one with critical electrical equipment (such as transformers, switchgear, generators). Each DAgent has lightweight anomaly detection capabilities and is responsible for monitoring the status of its equipment in real time and performing preliminary diagnoses.

[0068] Unit Collaborative Diagnostic Agent (PLAgent): Serving as the "diagnostic coordinator" of the power unit. It aggregates the diagnostic results of all DAgents under its jurisdiction, is responsible for the collaborative analysis of complex faults, diagnostic task planning, and dynamically organizes DAgents to form temporary diagnostic alliances.

[0069] The system's federated learning agent (SCAgent) is deployed in the regional diagnostic center. This agent coordinates the various PLAgents to jointly train and optimize the diagnostic model through a federated learning mechanism (including a federated coordinator, a global optimizer, and a privacy protection module), achieving knowledge sharing and continuous learning.

[0070] Dynamic task adaptation engine: such as Figure 1 As shown internally in PLAgent, this engine is a core module built into PLAgent. It can dynamically generate the optimal diagnostic workflow based on the characteristics of abnormal data (through feature extraction, algorithm selection, strategy generation, and alliance management).

[0071] Federated Learning and Knowledge Base Layer: Located at the bottom of the architecture, this layer provides support for the entire system. It includes a global model pool, a local model library, a diagnostic case library, and a performance evaluation library, supporting knowledge sharing and model updates among agents.

[0072] Figure 2 This is a flowchart of an electrical equipment anomaly diagnosis method based on dynamic task adaptation and federated learning according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps:

[0073] S201, Data Acquisition and Anomaly Detection.

[0074] Specifically, this step is executed by the data sensing layer and the device diagnostic agent (DAgent). The data sensing layer consists of a sensor network deployed on various electrical devices (such as transformers, switchgear, generators, etc.), responsible for real-time acquisition of multimodal operating data. This multimodal operating data includes vibration signals, temperature, partial discharge signals, current and voltage waveforms, etc.

[0075] Each Device Diagnostic Agent (DAgent) is assigned one-to-one with a critical electrical device and deployed on the device itself. Each DAgent has lightweight anomaly detection capabilities and is responsible for monitoring the operating status of its agent device in real time.

[0076] In this embodiment, the lightweight detection model can be a model with low computational overhead, such as a model based on statistical thresholds, isolated forests, or autoencoders, to ensure real-time response capabilities. DAgent uses this model to perform preliminary anomaly detection on the collected multimodal runtime data.

[0077] Step S201 utilizes the DAgent and data awareness layer deployed on the device to achieve real-time acquisition of multimodal data and local lightweight anomaly detection. This design ensures rapid response and on-site processing of anomalies, reduces unnecessary data transmission, and provides timely and accurate input for subsequent task adaptation and collaborative diagnosis.

[0078] S202, Anomaly Feature Extraction and Task Adaptation.

[0079] When the device diagnostic agent (DAgent) detects an anomaly, it first extracts the multi-dimensional features of the anomaly data and reports the features to its unit collaborative diagnostic agent (PLAgent).

[0080] In one embodiment, multi-dimensional feature extraction is performed by DAgent. This feature extraction constitutes an anomaly feature quantification model, including:

[0081] 1) Time-domain characteristics: such as standard deviation, skewness, kurtosis, peak factor, etc. These characteristics are used to describe the amplitude distribution and impulse characteristics of a signal.

[0082] 2) Frequency domain characteristics: such as dominant frequency, spectral centroid, spectral entropy, etc. These characteristics are used to identify rotating component faults or electrical harmonic anomalies.

[0083] 3) Time-frequency domain characteristics: such as wavelet transform energy distribution. Used to capture transient or non-stationary anomalous signals.

[0084] 4) Multi-source consistency characteristics: For example, the cross-correlation of data from multiple sensors (such as vibration signals from different measuring points of the same device). This is used to distinguish between a real fault and sensor failure.

[0085] This feature extraction step provides accurate state input to the dynamic adaptation engine by quantifying the multi-dimensional features of abnormal data, which is the foundation for generating refined diagnostic strategies.

[0086] Furthermore, after receiving features, the PLAgent unit collaborative diagnostic agent uses its built-in dynamic task adaptation engine to dynamically generate customized diagnostic solutions based on the extracted abnormal features. PLAgent acts as a diagnostic coordinator for power units (such as a substation or a power generation unit).

[0087] In one embodiment, the dynamic diagnostic strategy generation algorithm is implemented based on reinforcement learning.

[0088] Specifically, the process includes:

[0089] First, state encoding is performed, which encodes the feature vector extracted from S202 and the equipment criticality level (e.g., the main transformer level is higher than the auxiliary pump) into the state s of reinforcement learning.

[0090] Secondly, a Q-learning strategy is used for updates. In this embodiment, the engine maintains a state-action value function (Q-table or Q-network) to evaluate the expected reward of taking action a (i.e., selecting a certain diagnostic strategy) in state s.

[0091] The Q function updates iteratively according to the following formula:

[0092]

[0093] Where α is the learning rate. Let be the discount factor, r be the immediate reward obtained after executing the diagnostic strategy (e.g., based on the accuracy and speed of the diagnosis), and s' be the next state after the diagnosis. Action This corresponds to a specific diagnostic workflow, such as "calling a high-precision time-frequency analysis algorithm" or "requesting data from the associated device DAgent".

[0094] Finally, policy generation is based on a multi-objective optimization function. When generating policies, the engine aims to balance diagnostic accuracy, response time, and computational resource consumption. The optimization function can be expressed as:

[0095] Simultaneously satisfying constraints, such as response time ResponseTime(x)≤ and resource usage ResourceUsage (x)≤ .

[0096] The strategy generation process utilizes Q-learning, enabling the system to autonomously learn the optimal diagnostic path, balancing accuracy, efficiency, and resources to achieve intelligent and customized diagnostic solutions.

[0097] Step S202 combines multi-dimensional features with reinforcement learning through PLAgent's built-in dynamic task adaptation engine. This overcomes the limitations of fixed diagnostic models, dynamically generating optimal diagnostic solutions based on real-time operating conditions, significantly improving adaptability to complex faults and diagnostic accuracy.

[0098] S203, Dynamic Alliance Formation and Collaborative Diagnosis.

[0099] The unit collaborative diagnostic agent PLAgent performs collaborative diagnosis based on the anomaly type and diagnostic scheme generated by S202.

[0100] In one embodiment, the assembly mechanism is based on prediction of anomaly propagation paths; specifically, the process includes:

[0101] First, construct a device association graph. A pre-constructed electrical device association graph G=(V,E) is created, where nodes V represent electrical devices (such as transformers, switchgear, generators, etc.), and edges E represent the physical or mechanistic connections between devices (such as electrical wiring, control loops, cooling system dependencies, etc.). This graph can be stored in PLAgent or a knowledge base.

[0102] Secondly, Graph Convolutional Networks (GCNs) are used to predict anomaly propagation paths. When an anomaly occurs in a device (such as transformer A), PLAgent uses the GCN model to predict other devices to which the anomaly is most likely to propagate based on the device association graph G.

[0103] Furthermore, the hierarchical propagation formula of GCN can be expressed as:

[0104]

[0105] in, It is the first Feature matrix of layer nodes It is the first The trainable weight matrix of the layer, It is an adjacency matrix with self-connections. yes The degree matrix. It is a normalized adjacency matrix, used for stable computation and to prevent high-degree nodes from dominating information propagation. The diagonal elements... , representing the total number of connections to device i. It is a degree matrix The exponentiation is the square root of each diagonal element and the reciprocal of that root.

[0106] In this embodiment, normalization prevents devices with high degrees from dominating information propagation: if a device connects to many other devices, its features will be over-amplified. Normalization ensures that all devices have similar weights in information propagation. Furthermore, normalization can stabilize numerical calculations: ensuring that feature values ​​are within a reasonable range and avoiding gradient explosion or vanishing gradients.

[0107] Specifically, the prediction workflow is as follows:

[0108] 1) Input layer (l=0): This process uses the initial state features of all devices (i.e., baseline data during normal operation) as the initial feature matrix. When a device (e.g., transformer A) is detected as abnormal by its DAgent, the node characteristics corresponding to transformer A are recorded in... The status has been changed to an abnormal state.

[0109] 2) GCN Layer 1 (l=0→l=1): The GCN model applies the first-order propagation formula. At this point, each device aggregates information from its first-order neighbors (i.e., devices directly connected in the graph). Therefore, the features of devices directly connected to transformer A (e.g., circuit breaker B, cable C) begin to sense and fuse the abnormal state 5 of A.

[0110] 3) GCN Layer 2 (l=1→l=2): The model performs the aggregation operation again. At this time, neighbors B and C have already contained the abnormal information from A, and this information is aggregated again to their respective neighbors (i.e., the second-order neighbors of transformer A, such as load D and protection device E).

[0111] 4) Output layer (l=L): After multiple propagation layers, the model outputs the final state matrix. In this matrix, the final state value of each device node represents its probability score of being affected by the original anomaly.

[0112] Based on GCN's prediction results, PLAgent executes a dynamic alliance formation mechanism:

[0113] PLAgent Acquisition The probability score of all devices affected is calculated and judged according to a preset threshold. 8. All devices with a probability higher than the threshold are identified as highly correlated with this anomaly.

[0114] Subsequently, PLAgent dynamically forms a temporary diagnostic agent alliance based on this predicted path. This alliance precisely includes the DAgents corresponding to all identified high-probability devices. This mechanism shifts the diagnostic approach from isolated single-device analysis to collaborative diagnostics oriented towards fault propagation paths, significantly improving diagnostic efficiency and accuracy.

[0115] Finally, collaborative diagnostics are performed. PLAgent consortium members (DAgent) conduct diagnostic analyses according to the collaborative modes specified in the diagnostic protocol generated in S202. Collaborative modes include, for example, parallel analysis, pipelined analysis, or centralized consultation.

[0116] This collaborative diagnostic step, through flexible modes (parallel, pipeline, centralized consultation), ensures that the dynamic alliance can efficiently execute customized diagnostic solutions and shorten the diagnosis time for complex faults.

[0117] Step S203 uses GCN to predict the anomaly propagation path, achieving a shift from passive response to predictive diagnosis. The dynamically formed intelligent agent alliance replaces the fixed coordination mechanism, enabling diagnostic resources to be precisely focused on the scope of the fault's impact, significantly improving the ability and efficiency in dealing with cascading failures and complex anomalies.

[0118] S204, Federated Model Updates and Continuous Optimization.

[0119] The method of this invention achieves continuous optimization and knowledge sharing of the diagnostic model through a federated learning mechanism. In one embodiment, this mechanism is jointly implemented by PLAgent (as a client) and the system federated learning agent (SCAgent, as a server). SCAgent is deployed in a regional diagnostic center.

[0120] Specifically, the process includes:

[0121] First, local model updates. Each PLAgent periodically (e.g., after completing a successful collaborative diagnosis) updates its local diagnostic model using locally accumulated diagnostic cases (i.e., the anomalies and their characteristics identified in S203). This local model is an algorithm deployed on each PLAgent to perform local diagnostic and collaborative tasks. Specifically, it can refer to the reinforcement learning Q-network used in S202 to dynamically generate the optimal diagnostic solution or the graph convolutional network (GCN) model used in S203 to predict anomaly propagation paths and dynamically form diagnostic alliances.

[0122] PLAgent independently updates these model parameters using locally accumulated diagnostic cases (i.e., anomalies and their characteristics identified in S203), achieving continuous adaptation to specific operating conditions and fault modes of the local power grid through knowledge accumulation.

[0123] Secondly, this embodiment also provides a privacy protection mechanism. Specifically, when PLAgent uploads local model updates (e.g., model parameters), Previously, to protect data privacy, differential privacy technology was used to add Gaussian noise. The processed model parameters... Calculate using the following formula:

[0124]

[0125] in, These are the original local model parameters. These are the parameters used for uploading after adding noise. It is Gaussian noise. The noise mean is 0; S is the sensitivity, which in this embodiment is defined as the upper bound of the difference in model parameters trained on adjacent datasets (differences of only one sample); It's a noise scale, and a privacy budget. Related, The larger the value, the stronger the privacy protection, but the lower the model performance; I is the identity matrix, used to ensure that each dimension of the noise is generated independently, and the covariance matrix is... .

[0126] This mechanism provides quantifiable mathematical privacy guarantees for data learning across power units and security zones. This privacy protection step uses differential privacy technology to add Gaussian noise before model uploading, ensuring that the original data does not leave the local machine during knowledge sharing, thus strictly guaranteeing data security.

[0127] Secondly, differentiated federated aggregation. SCAgent receives privacy-preserving models from multiple PLAgents. Then, aggregation is performed. This embodiment uses an improved differentiated federated aggregation algorithm instead of the standard FedAvg.

[0128] SCAgent first calculates the weight of the k-th client (PLAgent) in the global aggregation. The formula is as follows:

[0129]

[0130] in, : The number of valid diagnostic samples locally owned by the k-th client. The total number of client samples participating in the aggregation, i.e. , The data quality score is calculated based on data integrity, consistency, and timeliness, and normalized to [0,1]. It is a comprehensive score of the accuracy and response stability of the client's historical diagnostics, normalized to [0,1]. If historical data is missing, the initial value is uniformly 0.5. α, β, γ are weighting coefficients that satisfy α+β+γ=1, used to balance the relative importance of data volume, quality and reliability.

[0131] Compared to traditional FedAvg, this weight calculation method additionally considers data quality and diagnostic reliability, preventing low-quality data or unreliable clients from negatively impacting the global model. SCAgent then uses this differentiated weight. Aggregate model updates from all clients to generate a new generation of global model parameters. (The local model parameters uploaded by the k-th client after the (t+1)th round of training):

[0132]

[0133] This differentiated aggregation step takes into account the amount, quality, and reliability of data, enabling the global model to learn from high-quality experiences more effectively, avoid the adverse effects of low-quality data, and improve the effectiveness of model iterative optimization.

[0134] Finally, the model is continuously optimized. SCAgent periodically performs aggregations to generate an improved global model and updates the new global model. The parameters are distributed to each PLAgent. Each PLAgent (or DAgent) updates its local model to the latest global model, thereby achieving continuous improvement in diagnostic capabilities and global knowledge sharing.

[0135] Step S204 addresses the issues of data silos and the difficulty of continuous learning through federated learning and SCAgent coordination. Differential weighted aggregation and differential privacy mechanisms, while ensuring cross-site data security, enable secure and efficient sharing of diagnostic knowledge and continuous evolution of the global model.

[0136] S205, Diagnostic Result Generation and Feedback.

[0137] After the S203 collaborative diagnostics are completed, the collaborative diagnostics team led by PLAgent generates a comprehensive diagnostic result. This result is not just a simple anomaly, but includes a report with detailed information, such as: anomaly type, severity, possible causes of failure (which can be combined with S203 GCN propagation path analysis), and handling recommendations. This comprehensive diagnostic result is ultimately fed back to operations personnel to support their decision-making.

[0138] Step S205, as the final output of the diagnostic process, effectively conveys the agent's analysis results to operations personnel by providing a comprehensive report including type, severity, cause, and recommendations. This ensures that the diagnostic information guides actual operational decisions, completing the closed loop of perception-diagnosis-decision.

[0139] Through the above steps S201 to S205, the method provided in this application embodiment improves the flexibility and accuracy of the diagnostic strategy through a dynamic task adaptation mechanism; improves the diagnostic efficiency of complex faults through a dynamic alliance formation mechanism; and achieves the sharing of diagnostic knowledge and the continuous evolution of the model while ensuring data security by integrating differential privacy-based federated learning.

[0140] On the other hand, embodiments of this application also provide a multi-agent system for diagnosing electrical equipment anomalies based on dynamic task adaptation and federated learning. Figure 3 This is a structural block diagram of a multi-agent system for electrical equipment anomaly diagnosis based on dynamic task adaptation and federated learning, according to an embodiment of this application. Figure 3 As shown, the system adopts a four-layer closed-loop architecture of "perception-diagnosis-decision-learning", including:

[0141] The acquisition module 30 is used to acquire electrical equipment abnormality data obtained by the device diagnostic agent based on multimodal sensor data for preliminary detection, and to extract multi-dimensional features from the abnormality data;

[0142] The prediction module 31 is used to diagnose the dynamic task adaptation engine in the intelligent body through unit collaboration. It generates customized diagnostic strategies based on multi-dimensional features through reinforcement learning strategies, and uses device association graphs and graph convolutional network models to predict the propagation path of abnormal data among electrical devices.

[0143] Diagnostic module 32 is used to dynamically form a diagnostic agent alliance consisting of multiple diagnostic agents based on the propagation path; through the diagnostic agent alliance, collaborative diagnosis is performed according to a customized diagnostic strategy, and diagnostic results are generated.

[0144] The optimization module 33 is used to coordinate multiple units to collaboratively diagnose the agent through the system's federated learning agent, and to perform federated learning to update the global diagnostic model.

[0145] The system improves the diagnostic accuracy of complex and novel anomalies by selecting the optimal diagnostic strategy for different anomaly characteristics. Federated learning technology enables agents at different power stations to jointly train diagnostic models without sharing raw data, solving the data silo problem while ensuring data privacy and security. Simultaneously, a dynamic task adaptation mechanism allows the system to autonomously adjust diagnostic strategies based on anomaly characteristics, significantly enhancing its adaptability to diverse anomaly types. Furthermore, a dynamic alliance formation mechanism oriented towards anomaly propagation paths enables the predictive organization of diagnostic resources, shortening the diagnostic time for complex anomalies.

[0146] In one embodiment, Figure 4 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 4 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 4 As shown, the electronic device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores an operating system, computer programs, and a database. The processor provides computing and control capabilities, the network interface communicates with external terminals via a network, the internal memory provides an environment for the operating system, the computer programs are executed by the processor to implement an electrical equipment anomaly diagnosis method based on a multi-agent system, and the database stores data.

[0147] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0148] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0149] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for diagnosing abnormalities of an electrical device based on a multi-agent system, characterized by, The method is based on a global diagnostic model composed of a device diagnostic agent, a unit collaborative diagnostic agent, and a system federated learning agent. Obtain abnormal data of electrical equipment obtained by the device diagnostic agent based on multimodal sensor data for preliminary detection, and extract multidimensional features from the abnormal data; By using the dynamic task adaptation engine within the unit collaborative diagnostic intelligent body, a reinforcement learning strategy is adopted to generate the optimal diagnostic strategy based on the multi-dimensional features, and the propagation path of the abnormal data between electrical devices is predicted by utilizing the device association graph and graph convolutional network model. Based on the propagation path, a diagnostic agent alliance is formed, consisting of multiple device diagnostic agents; through this alliance, collaborative diagnosis is performed according to the optimal diagnostic strategy, and diagnostic results are generated. Specifically, the system uses a federated learning agent to coordinate multiple units to collaboratively diagnose the agent and perform federated learning to update the global diagnostic model.

2. The method of claim 1, wherein, Extracting multi-dimensional features from the abnormal data includes: Extract the temporal features of the abnormal data, wherein the temporal features include standard deviation, skewness, and kurtosis. Extract the frequency domain features of the abnormal data, including the dominant frequency, spectral centroid, and spectral entropy; Extract the time-frequency domain features of the abnormal data, wherein the time-frequency domain features are wavelet transform energy distributions; In addition, multi-source consistency features are extracted, which are cross-correlation features between multi-modal sensor data; The time-domain features, the frequency-domain features, the time-frequency-domain features, and the cross-correlation features are used as the multi-dimensional features.

3. The method of claim 2, wherein, By employing a dynamic task adaptation engine within a unit-cooperative diagnostic intelligent body and utilizing a reinforcement learning strategy, the optimal diagnostic strategy is generated based on the aforementioned multi-dimensional features, including: The multi-dimensional features and device criticality levels are co-encoded to generate states for reinforcement learning; Policy optimization is performed using a state-action value function constructed based on the stated state, and the optimal diagnostic policy is output by iteratively updating the state-action value function. Specifically, the target value is obtained based on the immediate reward used to evaluate the quality of the diagnostic action and the maximum expected future reward of the next state after discounting factor adjustment. The state-action value function is then corrected using the temporal difference error between the target value and the current value of the state-action value function and a preset learning rate. In the process of generating the optimal diagnostic strategy, the merits of the diagnostic actions are evaluated through a multi-objective optimization function, which is constructed based on diagnostic accuracy, diagnostic response time, and diagnostic resource consumption.

4. The method according to claim 1, characterized in that, Using device association graphs and graph convolutional network models, the propagation path of the abnormal data among electrical devices is predicted as follows: Construct an equipment association graph using the physical connections and functional dependencies between electrical equipment as edges; Anomaly propagation prediction is performed using a graph convolutional network model and the device association graph to determine the propagation path of the abnormal data among electrical devices.

5. The method according to claim 4, characterized in that, Anomaly propagation prediction is performed using a graph convolutional network model and the device association graph. The predicted propagation path of the anomalous data among electrical devices includes: In the input layer of the graph convolutional network model, the initial state features of all electrical devices are defined as baseline data, and in the device association graph, the node features of the detected abnormal electrical devices are modified to abnormal states. Through multiple graph convolutional layers of the graph convolutional network model, multi-layer information propagation and feature aggregation are performed based on the data of the input layer. Each layer performs feature transformation on the feature matrix of the node in the previous layer and performs aggregation operation with the normalized device association graph to propagate abnormal data from the abnormal source node to the same-order neighbor devices and higher-order neighbor devices layer by layer. The final state is output through the output layer of the graph convolutional network model, and based on the final state, the probability score of each electrical device being affected by the abnormal data is determined. The propagation path is determined by identifying electrical devices in the device association map whose probability score of being affected is higher than a preset threshold.

6. The method according to claim 1, characterized in that, Through a system-fed learning agent, coordinating multiple units to collaboratively diagnose the agent, and performing federated learning to update the global diagnostic model, the following is included: The system's federated learning agent distributes global diagnostic model parameters to multiple unit collaborative diagnostic agents. The unit collaborative diagnostic agent loads the global diagnostic model parameters into the local model, updates the local model based on local diagnostic case data, and uploads the updated local model to the system federated learning agent. The system's federated learning agent aggregates various updated local models to generate an updated global diagnostic model. A differentiated federated aggregation algorithm is used to calculate the client weight of each unit's collaborative diagnostic agent in the global diagnostic model aggregation process based on preset indicators. The updated global diagnostic model is then generated by weighting the updated local models based on the client weights.

7. The method according to claim 6, characterized in that, A differentiated federated aggregation algorithm is used to calculate the client weight of each unit's collaborative diagnostic agent in the global diagnostic model aggregation process based on preset indicators, including: Determine the first, second, and third preset weighting coefficients for balancing data volume, data quality, and data reliability, respectively; The first parameter is obtained based on the ratio of the number of valid diagnostic samples stored locally by the unit collaborative diagnostic agent to the total number of all participating samples, and the first preset weighting coefficient. Based on the integrity parameters, consistency parameters, and timeliness parameters of the local data of the unit collaborative diagnostic agent, a data quality score is calculated, and a second parameter is obtained according to the data quality score and the second preset weighting coefficient. A comprehensive reliability score is calculated based on the historical diagnostic accuracy and response stability of the unit collaborative diagnostic agent. A third parameter is obtained based on the comprehensive reliability score and the third preset weighting coefficient. Based on the first, second, and third parameters of the unit collaborative diagnostic agent, its corresponding client weight in the global diagnostic model aggregation process is obtained.

8. The method according to any one of claims 6 and 7, characterized in that, The method further includes: Before the unit collaborative diagnostic agent uploads the updated local model, differential privacy technology is used to add Gaussian noise to the original local model parameters to generate privacy-preserving model parameters for uploading. The Gaussian noise has zero mean and is generated based on the identity matrix. The noise scale of the Gaussian noise is determined according to a preset privacy budget and sensitivity. The sensitivity is used to measure the degree of influence of the presence of a single training sample on the local model parameters.

9. An electrical equipment anomaly diagnosis system based on a multi-agent system, characterized in that, The system is based on a global diagnostic model composed of a device diagnostic agent, a unit collaborative diagnostic agent, and a system federated learning agent. The system includes: The acquisition module is used to acquire electrical equipment abnormality data obtained by the device diagnostic agent based on multimodal sensor data for preliminary detection, and to extract multi-dimensional features from the abnormality data; The prediction module is used to diagnose the dynamic task adaptation engine within the intelligent body through unit collaborative diagnosis. It adopts a reinforcement learning strategy to generate the optimal diagnosis strategy based on the multi-dimensional features and uses the device association graph and graph convolutional network model to predict the propagation path of the abnormal data between electrical devices. The diagnostic module, based on the propagation path, establishes a diagnostic agent alliance consisting of multiple device diagnostic agents; through the diagnostic agent alliance, it performs collaborative diagnosis according to the optimal diagnostic strategy and generates diagnostic results. An optimization module is used to coordinate multiple unit collaborative diagnostic agents through system federated learning agents to perform federated learning to update the global diagnostic model.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 8.

Images