Multi-level power equipment defect diagnosis method and system based on feature pyramid and physical constraint

Abstract

The application belongs to the technical field of power equipment defect diagnosis, and relates to a multi-level power equipment defect diagnosis method and system based on a feature pyramid and physical constraints, comprising: constructing a power equipment inspection graph; processing power equipment inspection data and the power equipment inspection graph to obtain multi-scale node features of power equipment nodes; based on physical constraints between inspection equipment and power equipment nodes, obtaining current edge features of the power equipment inspection graph; based on the multi-scale node features and the current edge features, constructing a complete power equipment inspection graph, and performing attention enhancement processing on the complete power equipment inspection graph to obtain inspection features of the power equipment nodes; performing multi-layer perception on the inspection features to obtain an abnormal confidence score, and taking power equipment nodes with an abnormal confidence score greater than an abnormal threshold as defective power equipment; and the embodiment improves the robustness of the embodied inspection equipment in identifying abnormalities in a complex high-noise scene.

Description

Technical Field

[0001] This invention relates to the field of power equipment defect diagnosis technology, and specifically discloses a multi-level power equipment defect diagnosis method and system based on feature pyramids and physical constraints. Background Technology

[0002] Insulation degradation and localized overheating in power transmission and transformation equipment are core causes of power grid faults, directly affecting the safe and stable operation of the power grid. Traditional manual inspection methods suffer from low efficiency, subjective diagnostic results, high risks in high-risk environments, and difficulty in achieving high-frequency, full-coverage inspections.

[0003] Existing intelligent inspection solutions based on inspection equipment (such as embodied intelligent robots) mostly rely solely on visual or infrared data for surface defect identification, failing to extract hidden insulation degradation features and exhibiting insufficient accuracy in multimodal data fusion. Furthermore, existing solutions do not consider the multi-physics interactions between the inspection equipment and the power equipment, and the lack of physically constrained modeling cannot provide accurate spatial topology support for defect diagnosis. In addition, existing solutions are weak against environmental noise and motion interference in dynamic inspection scenarios, lacking defect cascading risk projection and proactive safety protection mechanisms, making it difficult to meet the needs of full lifecycle health management for power equipment in complex power grid scenarios.

[0004] In view of this, the present invention provides a multi-level power equipment defect diagnosis method and system based on feature pyramid and physical constraints, which improves the robustness of the embodied inspection equipment in identifying anomalies in complex and high-noise scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-level power equipment defect diagnosis method and system based on feature pyramids and physical constraints, addressing the problem of how to detect defects in power equipment under complex, high-noise conditions. The specific solution is as follows:

[0006] A multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints includes: S1: Constructing a power equipment inspection map based on power equipment inspection data; the power equipment inspection data includes power equipment monitoring data, power equipment sensing data, and inspection status data; S2: Performing feature pyramid processing on the power equipment inspection data and the power equipment inspection map to obtain multi-scale node features of power equipment nodes; the power equipment nodes include equipment master nodes and equipment component nodes; S3: Obtaining the current edge features of the power equipment inspection map based on the physical constraints between the inspection equipment and the power equipment nodes; S4: Constructing a complete power equipment inspection map based on the multi-scale node features and the current edge features, and performing attention enhancement processing on the complete power equipment inspection map to obtain the inspection features of the power equipment nodes; S5: Performing multi-layer perception on the inspection features to obtain anomaly confidence scores, and identifying power equipment nodes with anomaly confidence scores greater than anomaly thresholds as defective power equipment.

[0007] Furthermore, S1 includes: S1.1: constructing a multi-layer power equipment architecture based on the assembly relationship of the main components and auxiliary components of the power equipment; S1.2: constructing a power equipment topology map by taking the main components and auxiliary components as fixed nodes and the electrical connection relationship and auxiliary relationship between components as fixed edges; S1.3: taking the inspection equipment as a mobile node, constructing mobile edges based on the inspection field of the inspection equipment, and integrating the mobile nodes and mobile edges into the power equipment topology map to obtain a power equipment inspection map.

[0008] Furthermore, S2 includes: S2.1: constructing a multimodal feature sequence along the movement trajectory of the inspection equipment; S2.2: processing the target power equipment node based on the multimodal feature sequence and the location of the inspection equipment to obtain multi-scale spatiotemporal features; S2.3: fusing the multi-scale spatiotemporal features and uncertainty variance to obtain multi-scale node features.

[0009] Furthermore, S2.1 includes: S2.1.1: calculating heat loss power based on the surface hot spot temperature and typical thermal resistance parameters of the power equipment nodes; S2.1.2: assessing the aging of the insulating medium within the field of view of the inspection equipment based on the heat loss power; S2.1.3: obtaining the monitored data deviation and prior uncertainty variance based on the power equipment monitoring data and the power equipment operation baseline data; S2.1.4: constructing a multimodal feature sequence based on the aging of the insulating medium, the data deviation, and the uncertainty variance.

[0010] Furthermore, step S2.2 includes: S2.2.1: processing the multimodal feature sequence of the target power equipment node through a long short-term memory network to obtain temporal features; S2.2.2: for the current time, extracting the adjacency matrix of the target power equipment node and the current modal features t of the neighboring nodes based on the power equipment inspection map; S2.2.3: processing the adjacency matrix and the current modal features through a graph convolutional network to obtain spatial features; S2.2.4: concatenating the temporal features and spatial features to obtain spatiotemporal features.

[0011] Further, S3 includes: S3.1: Calculating the electromagnetic attenuation factor based on the location of the inspection equipment, the spatial topology of the power equipment nodes, and the attributes of the inspection equipment; the spatial topology of the power equipment nodes includes the relative coordinate distribution, geometric distance, and assembly connection relationship of each node in three-dimensional physical space; S3.2: Determining the electromagnetic field dynamic weight and thermal field distribution weight of the moving edge based on the distance between the inspection equipment and the power equipment nodes and the electromagnetic attenuation factor; the moving edge refers to the graph edge connected to the inspection equipment; S3.3: Using the electromagnetic field dynamic weight and thermal field distribution weight as the current edge feature of the moving edge.

[0012] Furthermore, S4 includes: S4.1: performing hierarchical convolution based on multi-scale node features and power equipment inspection map to obtain hierarchical convolution results and attention parameters; S4.2: performing noise suppression based on attention parameters, uncertainty variance and environmental noise power spectrum to obtain attention output; S4.3: splicing hierarchical convolution results and attention output to obtain inspection features of power equipment nodes.

[0013] Furthermore, step S4.1 includes: S4.1.1: determining the motion compensation factor of the inspection equipment based on the inspection status data; S4.1.2: performing compensation processing on the multi-scale node features and the motion compensation factor to obtain the compensated enhanced features; S4.1.3: performing graph convolution on the compensated enhanced features and the adjacency matrix of the target power equipment node to obtain the hierarchical convolution result; S4.1.4: performing attention linear projection on the hierarchical convolution result to obtain the attention parameters.

[0014] Furthermore, S4.2 includes: S4.2.1: performing frequency domain analysis on the background information acquired by the inspection equipment to obtain the environmental noise power spectrum; the background information includes environmental audio information and spatial electromagnetic signals; S4.2.2: determining the dynamic noise penalty matrix based on the prior uncertainty variance and the environmental noise power spectrum; S4.2.3: calculating the attention output based on the attention parameters and the dynamic noise penalty matrix.

[0015] This invention also provides a multi-level power equipment defect diagnosis system based on feature pyramids and physical constraints, used to implement the aforementioned multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints. The system includes a graph construction module, an inspection feature acquisition module, and a defect identification module. The graph construction module includes a graph topology construction module, a node feature determination module, an edge feature determination module, and a knowledge graph construction module. The graph topology construction module is used to construct a power equipment inspection graph based on power equipment inspection data. The power equipment inspection data includes power equipment monitoring data, power equipment sensing data, and inspection status data. The node feature determination module is used to perform feature pyramid analysis on the power equipment inspection data and the power equipment inspection graph. The system performs tower processing to obtain multi-scale node features zl of the power equipment nodes; the power equipment nodes include equipment master nodes and equipment component nodes; the edge feature determination module is used to obtain the current edge features of the power equipment inspection map based on the physical constraints between the inspection equipment and the power equipment nodes; the knowledge graph construction module is used to construct a complete power equipment inspection map based on the multi-scale node features and the current edge features; the inspection feature acquisition module is used to perform attention enhancement processing on the complete power equipment inspection map to obtain the inspection features of the power equipment nodes; the defect identification module is used to perform multi-layer perception on the inspection features to obtain anomaly confidence scores, and power equipment nodes with anomaly confidence scores greater than anomaly thresholds are identified as defective power equipment.

[0016] The present invention has the following advantages and beneficial effects: This invention achieves deep fusion of multi-source data, including visual, infrared, ultrasonic, and magnetic field data, through kinematic compensation and dynamic noise assessment. It can effectively suppress motion and environmental dynamic noise, accurately extract hidden insulation degradation features of equipment, and solve the problems of insufficient fusion and weak feature extraction capabilities in existing technologies.

[0017] This invention integrates dynamic observation nodes into twin graphs, constructs electromagnetic and thermal field dual physical edges in real time, quantifies the multi-physics dynamic interaction between inspection equipment and power equipment, realizes real-time linkage between inspection equipment and power equipment, and improves defect diagnosis results.

[0018] This invention introduces attention-enhanced hierarchical cognitive convolution and active redundancy suppression mechanism, combined with prior uncertainty adaptive suppression of dynamic environmental noise and low-quality data interference caused by inspection motion, which greatly improves the robustness of abnormal pattern recognition in complex and noisy scenarios. Attached Figure Description

[0019] Figure 1 This is an exemplary flowchart of a multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints in an embodiment of the present invention; Figure 2This is an exemplary structural diagram of a multi-level power equipment defect diagnosis system based on feature pyramids and physical constraints in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Figure 1 This is an exemplary flowchart of a multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints in an embodiment of the present invention, as shown below. Figure 1 As shown, the methods for diagnosing defects in power equipment include the following: S1: Collect power equipment inspection data through inspection equipment and construct a power equipment inspection map based on the data. Power equipment inspection data includes power equipment monitoring data, power equipment sensing data, and inspection status data. Power equipment monitoring data can be raw data in multiple modalities collected by the inspection equipment in different poses (e.g., lidar point clouds, infrared thermal images, sonar inspection audio, etc.). Power equipment sensing data is data obtained through safe contact with the equipment's grounding casing or using near-field sensing technology. Inspection status data is data obtained through the inspection equipment's odometer information and real-time pose. For example, the inspection equipment can be an embodied intelligent robot equipped with a multi-sensor array, including: a visible light camera (vision), an infrared thermal imager, an ultrasonic sensor, and a magnetic field sensor. The robot is equipped with a six-degree-of-freedom robotic arm for safe contact with the power equipment's grounding casing or down conductor to obtain sensing data (e.g., grounding current, vibration signals, etc.). The robot body also includes an odometer, IMU, and SLAM positioning system for acquiring status data.

[0022] S1.1: Based on the assembly relationships between the main components and auxiliary components of power equipment, a multi-layered power equipment architecture is constructed. For example, for an oil-immersed transformer, the main components include the core and windings, while the auxiliary components include bushings, tap changers, coolers, oil conservators, and gas relays; for a high-voltage circuit breaker, the main components include the arc-extinguishing chamber and operating mechanism, while the auxiliary components include auxiliary contacts, heaters, and buffers. Through the assembly relationships between the main and auxiliary components (such as bolted connections, flange connections, and embedded installations), a multi-level equipment structure tree is formed.

[0023] S1.2: Construct a power equipment topology map by treating main and auxiliary components as fixed nodes, and electrical connections and dependencies between components as fixed edges. The node attributes of the power equipment topology map include equipment type (transformer / circuit breaker / disconnector, etc.), rated voltage, rated capacity, manufacturer, commissioning date, historical defect records, and current operating status; edge attributes include connection type (electrical connection, mechanical connection, signal connection), connection strength (e.g., contact resistance value), physical distance (three-dimensional spatial distance between two components), and dependency identifiers (e.g., "included in", "installed in").

[0024] S1.3: Using the inspection equipment as a mobile node, construct mobile edges based on the inspection field of the inspection equipment, and integrate the mobile nodes and mobile edges into the power equipment topology graph to obtain the power equipment inspection graph.

[0025] S2: Feature pyramid processing is performed on the power equipment inspection data and power equipment inspection maps to obtain multi-scale node features of power equipment nodes. Power equipment nodes include main equipment nodes and component equipment nodes. A main equipment node is an independent and complete power equipment unit in a substation, such as a three-phase oil-immersed transformer, an SF6 circuit breaker, or a set of disconnect switches. Component equipment nodes are the key functional components that constitute the main equipment node, such as the transformer windings, core, bushings, oil conservator, and tap changer; and the circuit breaker's arc-extinguishing chamber, operating mechanism, and auxiliary contacts.

[0026] S2.1: Construct a multimodal feature sequence along the movement trajectory of the inspection equipment.

[0027] S2.1.1: Calculate heat loss power based on surface hotspot temperatures and typical thermal resistance parameters of power equipment nodes. During the inspection equipment's movement along a preset trajectory, an infrared thermal imager acquires infrared thermal images of the power equipment surface at a certain frequency. The system extracts the surface hotspot temperatures from the thermal images. Typical thermal resistance parameters for this equipment model are then queried from the equipment database. Heat loss power is calculated. ; in, The heat loss power of the power equipment node; The surface hotspot temperature of node i of the power equipment; Ambient temperature; These are typical thermal resistance parameters for power equipment nodes.

[0028] S2.1.2: Assess the aging of the insulation medium within the inspection field of view based on heat loss power. Evaluate the degree of aging of the insulation medium according to the historical trend and threshold of heat loss power. For example, if the heat loss power increases by more than 20% compared to the baseline, it is determined that there is a risk of insulation degradation. Here, the baseline refers to the historical average heat loss power of the equipment under healthy conditions.

[0029] S2.1.3: Based on power equipment monitoring data and power equipment operating baseline data, the monitoring data deviation and prior uncertainty variance are obtained. Power equipment operating baseline data are obtained through historical health cycle data of the power equipment. The data deviation between the measured value and the baseline mean is: ; in, This is due to data bias; To take the absolute value; These are the measured values ​​from power equipment monitoring data; Baseline data for the operation of power equipment.

[0030] Mapping the bias to the prior uncertainty variance: ; in, The first adjustable parameter is the variance of the prior uncertainty. α and β are the first and second adjustable parameters, respectively. The first adjustable parameter α is obtained by minimizing the root mean square error between the predicted uncertainty variance and the actual error in the historical data. Specifically, grid search is performed using labeled inspection data (including the actual equipment status) to select the α value that minimizes the root mean square error on the validation set. The second adjustable parameter β is set as the bias term of the baseline data and is calibrated by minimizing the uncertainty variance at the baseline data points. Typically, β is taken as 0.01~0.05. The standard deviation of the baseline data is calculated by measuring the standard deviation of the baseline data during stable operation.

[0031] S2.1.4: Construct a multimodal feature sequence based on the aging status of the insulating medium, data bias, and uncertainty variance. .

[0032] S2.2: Process the target power equipment node based on the multimodal feature sequence and the location of the inspection equipment to obtain multi-scale spatiotemporal features. S2.2.1: Process the multimodal feature sequence of the target power equipment node using a Long Short-Term Memory (LSTM) network to obtain temporal features. S2.2.2: For the current time, based on the power equipment inspection map, extract the adjacency matrix of the target power equipment node and the current modal features of its neighboring nodes. S2.2.3: Spatial features are obtained by processing the adjacency matrix and current modality features through a graph convolutional network. S2.2.4: Temporal and spatial features are concatenated to obtain spatiotemporal features. Based on spatial topology. (Represents the spatial correlation strength between device nodes i and j), combined with the inspection trajectory time window. Multimodal feature sequences within the dataset, multi-scale spatiotemporal features: ; in, Let i be the spatiotemporal characteristics of power equipment node i at the l-th level at the current time t. For the current time interval The multimodal feature sequence; ⊕ represents the concatenation of feature vectors; For each element of the normalized adjacency matrix corresponding to the l-th layer feature pyramid, the strength of the spatial association between device node i and its neighbor node j at that layer is represented. Let represent the current modal features of neighbor node j at the l-th layer scale. LSTM captures temporal dependencies; Graph Convolutional Network (GCN) captures spatial dependencies.

[0033] S2.3: Multi-scale spatiotemporal features and uncertainty variance are fused to obtain multi-scale node features. A variational autoencoder (VAE) is used to sample and fuse features from the posterior distribution. ; in, For node features at scale l, satisfying the distribution ; Let be the mean vector of the features of the l-th scale node output by the variational autoencoder (VAE). Then, a fully connected network is used to integrate the spatiotemporal features. Mapped to the latent space; The standard deviation vector is output by the VAE encoder; The prior uncertainty variance is used to assign higher uncertainty to low-quality data generated by the movement and vibration of inspection equipment, thereby reducing its weight in the fusion features.

[0034] S3: Based on the physical constraints between the inspection equipment and power equipment nodes, the current edge features of the power equipment inspection map are obtained. The physical entity attributes of the inspection equipment (e.g., distance) affect the perception results. The difference in magnetic field radiation intensity between the inspection equipment and other equipment due to changes in spatial location is quantified using an electromagnetic field spatial attenuation model.

[0035] S3.1: Calculate the electromagnetic attenuation factor based on the location of the inspection equipment, the spatial topology of the power equipment nodes, and the attributes of the inspection equipment. The spatial topology of the power equipment nodes includes the relative coordinate distribution, geometric distance, and assembly connection relationship of each node in three-dimensional physical space. The electromagnetic field strength attenuation factor is: ; in, r is the electromagnetic field strength attenuation factor; r is the current position coordinate of the inspection equipment. Coordinates of power equipment node i The distance between them. The electromagnetic field intensity attenuation factor is used to estimate the environmental electromagnetic baseline.

[0036] S3.2: Based on the distance between the inspection equipment and the power equipment node and the electromagnetic attenuation factor, determine the electromagnetic field dynamic weight and thermal field distribution weight of the moving edge; the moving edge refers to the graph edge connected to the inspection equipment. Combining the real-time distance between the inspection equipment and the power equipment node, construct multiple field physical edges: on the one hand, construct the electromagnetic field dynamic weight based on the principle of potential superposition to characterize spatial electromagnetic interference; on the other hand, construct the thermal field distribution weight based on the laws of thermal radiation and thermal convection to characterize heat dissipation along the moving path. Simultaneously, maintain the original topological direct connection and dependency relationships. Based on the principle of potential superposition, determine the electromagnetic field dynamic weight by the electromagnetic field intensity generated by device i at inspection position r: ; in, For electromagnetic field dynamic weights; Let be the azimuth angle of device i relative to the direction of the inspection equipment.

[0037] Determine the weights of the thermal field distribution based on thermal radiation and thermal convection: ; in, ε is the thermal field distribution weight; ε is the emissivity; and σ is the Stefan-Boltzmann constant. Where is the equipment surface area; h is the convective heat transfer coefficient; This refers to the hot spot temperature on the surface of the equipment. λ represents the ambient temperature; λ represents the air thermal diffusion length.

[0038] S3.3: Use the electromagnetic field dynamic weight and thermal field distribution weight as the current edge features of the moving edge.

[0039] S4: Based on multi-scale node features and current edge features, construct a complete power equipment inspection map, and perform attention enhancement processing on the complete power equipment inspection map to obtain the inspection features of power equipment nodes.

[0040] S4.1: Based on multi-scale node features and power equipment inspection maps, hierarchical convolution is performed to obtain hierarchical convolution results and attention parameters.

[0041] S4.1.1: Determine the motion compensation factor based on the inspection status data. This motion compensation factor is calculated from the root mean square of the instantaneous velocity and angular velocity of the inspection equipment within the time window.

[0042] S4.1.2: Compensate the multi-scale node features and motion compensation factor to obtain the compensated enhanced features. The product of the multi-scale node features and the motion compensation factor is used as the compensated enhanced feature.

[0043] S4.1.3: Perform graph convolution on the compensated enhanced features and the adjacency matrix of the target power equipment nodes to obtain the hierarchical convolution result.

[0044] S4.1.4: Perform attention linear projection on the hierarchical convolution results to obtain attention parameters.

[0045] S4.2: Noise suppression is performed based on attention parameters, uncertainty variance, and environmental noise power spectrum to obtain attention output.

[0046] S4.2.1: Perform frequency domain analysis on the background information acquired by the inspection equipment to obtain the environmental noise power spectrum; the background information includes environmental audio information and spatial electromagnetic signals.

[0047] S4.2.2: Based on the prior uncertainty variance and the environmental noise power spectrum, determine the dynamic noise penalty matrix. The dynamic noise penalty matrix, a diagonal matrix, is obtained by integrating over the frequency band corresponding to device i. ; in, This represents the i-th diagonal element of the dynamic noise penalty matrix; Let V be the prior uncertainty variance of power equipment node i; This is the noise fusion coefficient (typical value 0.5); and These represent the upper and lower limits of the characteristic frequency band of power equipment node i, respectively. Environmental noise power spectrum. The results are obtained by performing a Fast Fourier Transform on the blank background signals (such as audio and electromagnetic signals in areas without equipment) collected by the inspection equipment and then taking the square of the modulus.

[0048] S4.2.3: Calculate the attention output based on the attention parameters and the dynamic noise penalty matrix: ; in, For attention output; K is the normalized activation function; Q is the query vector matrix; K is the key vector matrix; V is the scaling factor; V is the value vector matrix; T is the transpose of the matrix; γ is the dynamic noise penalty matrix; γ is a hyperparameter (typical value 0.1~0.5). When a certain device node's When the noise level is high or the ambient noise is strong, Increase the weight of this node in the attention calculation to achieve adaptive redundancy suppression. For example, when the inspection equipment is near a transformer generating strong corona discharge, the ambient noise power spectrum peaks in the 20-50kHz frequency band. After identifying this frequency band, the system increases the weight of the ultrasonic sensor channel that is severely affected by interference. The corresponding terms make the model more dependent on the characteristics of the infrared and magnetic field channels.

[0049] S4.3: Concatenate the layered convolution results and attention output to obtain the inspection features of power equipment nodes.

[0050] S5: Perform multi-layer perception on the inspection features to obtain anomaly confidence scores, and identify power equipment nodes with anomaly confidence scores greater than the anomaly threshold as defective power equipment. Each equipment node i obtains a high-dimensional feature vector. and attention score After concatenating these values, input them into a three-layer multilayer perceptron (MLP) to output anomaly confidence scores: ; in, The anomaly confidence score for power equipment node i. MLP stands for Multilayer Perceptron. For splicing operations; This represents the hierarchical convolution result for node i of the power equipment. This is the attention output for power equipment node i.

[0051] Figure 2 This is an exemplary structural diagram of a multi-level power equipment defect diagnosis system based on feature pyramids and physical constraints, as described in an embodiment of the present invention. Figure 2 As shown, the multi-level power equipment defect diagnosis system based on feature pyramids and physical constraints includes a map construction module, an inspection feature acquisition module, and a defect identification module.

[0052] The graph construction module includes a graph topology construction module, a node feature determination module, an edge feature determination module, and a knowledge graph construction module.

[0053] The topology construction module is used to build power equipment inspection maps based on power equipment inspection data.

[0054] The power equipment inspection data includes power equipment monitoring data, power equipment sensing data, and inspection status data.

[0055] The node feature determination module is used to perform feature pyramid processing on power equipment inspection data and power equipment inspection maps to obtain multi-scale node features zl of power equipment nodes; the power equipment nodes include equipment master nodes and equipment component nodes.

[0056] The edge feature determination module is used to obtain the current edge features of the power equipment inspection map based on the physical constraints between the inspection equipment and the power equipment nodes.

[0057] The knowledge graph construction module is used to build a complete power equipment inspection graph based on multi-scale node features and current edge features.

[0058] The inspection feature acquisition module is used to perform attention enhancement processing on the complete power equipment inspection map to obtain the inspection features of power equipment nodes.

[0059] The defect identification module is used to perform multi-layer perception of inspection features, obtain anomaly confidence scores, and identify power equipment nodes with anomaly confidence scores greater than the anomaly threshold as defective power equipment.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints, characterized in that, include: S1: Construct a power equipment inspection map based on power equipment inspection data; The power equipment inspection data includes power equipment monitoring data, power equipment sensing data, and inspection status data; S2: Perform feature pyramid processing on the power equipment inspection data and power equipment inspection map to obtain multi-scale node features of power equipment nodes; The power equipment node includes a main equipment node and equipment component nodes; S3: Based on the physical constraints between the inspection equipment and the power equipment nodes, obtain the current edge features of the power equipment inspection graph; S4: Based on multi-scale node features and current edge features, construct a complete power equipment inspection map, and perform attention enhancement processing on the complete power equipment inspection map to obtain the inspection features of power equipment nodes; S5: Perform multi-layer perception on the inspection features to obtain anomaly confidence scores, and identify power equipment nodes with anomaly confidence scores greater than the anomaly threshold as defective power equipment.

2. The multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints according to claim 1, characterized in that, S1 includes: S1.1: Based on the assembly relationship of the main components and auxiliary components of power equipment, construct a multi-layer power equipment architecture; S1.2: Construct a power equipment topology map by using main components and auxiliary components as fixed nodes, and electrical connections and relationships between components as fixed edges; S1.3: Using the inspection equipment as a mobile node, construct mobile edges based on the inspection field of the inspection equipment, and integrate the mobile nodes and mobile edges into the power equipment topology graph to obtain the power equipment inspection graph.

3. The multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints according to claim 1, characterized in that, S2 includes: S2.1: Construct a multimodal feature sequence along the movement trajectory of the inspection equipment; S2.2: Based on the multimodal feature sequence and the location of the inspection equipment, the target power equipment node is processed to obtain multi-scale spatiotemporal features; S2.3: The spatiotemporal features and uncertainty variance of multiple scales are fused to obtain multi-scale node features.

4. The multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints according to claim 3, characterized in that, S2.1 includes: S2.1.1: Calculate heat loss power based on the surface hot spot temperature and typical thermal resistance parameters of power equipment nodes; S2.1.2: Based on heat loss power, assess the aging of the insulating medium within the field of view of the inspection equipment; S2.1.3: Based on power equipment monitoring data and power equipment operation baseline data, obtain the monitoring data bias and prior uncertainty variance; S2.1.4: Construct a multimodal feature sequence based on the aging of the insulating medium, data deviation, and uncertainty variance.

5. The multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints according to claim 3, characterized in that, S2.2 includes: S2.2.1: The multimodal feature sequences of the target power equipment nodes are processed by a long short-term memory network to obtain temporal features; S2.2.2: For the current time, based on the power equipment inspection map, extract the adjacency matrix of the target power equipment node and the current modal features t of the neighboring nodes; S2.2.3: Spatial features are obtained by processing the adjacency matrix and current modality features through a graph convolutional network; S2.2.4: Combine the temporal and spatial features to obtain the spatiotemporal features.

6. The multi-level power equipment defect diagnosis method based on feature pyramid and physical constraints according to claim 1, characterized in that, The S3 includes: S3.1: Calculate the electromagnetic attenuation factor based on the location of the inspection equipment, the spatial topology of the power equipment nodes, and the attributes of the inspection equipment; the spatial topology of the power equipment nodes includes the relative coordinate distribution, geometric distance, and assembly connection relationship of each node in three-dimensional physical space; S3.2: Based on the distance between the inspection equipment and the power equipment node and the electromagnetic attenuation factor, determine the electromagnetic field dynamic weight and thermal field distribution weight of the moving edge; the moving edge refers to the graph edge connected to the inspection equipment. S3.3: Use the electromagnetic field dynamic weight and thermal field distribution weight as the current edge features of the moving edge.

7. The multi-level power equipment defect diagnosis method based on feature pyramid and physical constraints according to claim 1, characterized in that, The S4 includes: S4.1: Perform hierarchical convolution based on multi-scale node features and power equipment inspection maps to obtain hierarchical convolution results and attention parameters; S4.2: Noise suppression is performed based on attention parameters, uncertainty variance, and environmental noise power spectrum to obtain attention output; S4.3: Concatenate the layered convolution results and attention output to obtain the inspection features of power equipment nodes.

8. The multi-level power equipment defect diagnosis method based on feature pyramid and physical constraints according to claim 7, characterized in that, S4.1 includes: S4.1.1: Determine the motion compensation factor of the inspection equipment based on the inspection status data; S4.1.2: Compensate the multi-scale node features and motion compensation factors to obtain the compensated enhanced features; S4.1.3: Perform graph convolution on the compensated enhanced features and the adjacency matrix of the target power equipment nodes to obtain the hierarchical convolution result; S4.1.4: Perform attention linear projection on the hierarchical convolution results to obtain attention parameters.

9. The multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints according to claim 7, characterized in that, S4.2 includes: S4.2.1: Perform frequency domain analysis on the background information acquired by the inspection equipment to obtain the environmental noise power spectrum; the background information includes environmental audio information and spatial electromagnetic signals; S4.2.2: Determine the dynamic noise penalty matrix based on the prior uncertainty variance and the environmental noise power spectrum; S4.2.3: Calculate the attention output based on the attention parameters and the dynamic noise penalty matrix.

10. A multi-level power equipment defect diagnosis system based on feature pyramids and physical constraints, characterized in that, The method for implementing the multi-level power equipment defect diagnosis method based on feature pyramids and physical constraints as described in any one of claims 1-9 includes a map construction module, an inspection feature acquisition module, and a defect identification module. The graph construction module includes a graph topology construction module, a node feature determination module, an edge feature determination module, and a knowledge graph construction module. The graph topology construction module is used to construct a power equipment inspection graph based on power equipment inspection data. The power equipment inspection data includes power equipment monitoring data, power equipment sensing data, and inspection status data. The node feature determination module is used to perform feature pyramid processing on power equipment inspection data and power equipment inspection maps to obtain multi-scale node features of power equipment nodes. The power equipment node includes a main equipment node and equipment component nodes; The edge feature determination module is used to obtain the current edge features of the power equipment inspection map based on the physical constraints between the inspection equipment and the power equipment nodes; the knowledge graph construction module is used to construct a complete power equipment inspection map based on multi-scale node features and the current edge features. The inspection feature acquisition module is used to perform attention enhancement processing on the complete power equipment inspection map to obtain the inspection features of power equipment nodes; The defect identification module is used to perform multi-layer perception of inspection features, obtain anomaly confidence scores, and identify power equipment nodes with anomaly confidence scores greater than the anomaly threshold as defective power equipment.

Images