Intelligent inspection optimization method for power equipment

Abstract

The invention discloses an intelligent inspection optimization method for power equipment, and belongs to the technical field of intelligent operation and maintenance of a power system, and the method comprises the steps: firstly obtaining multi-source original data of substation equipment, and fusing the multi-source original data to form fused data; then an alliance chain network is constructed, each transformer substation serves as a node, a plurality of transformer substation nodes, an operation and maintenance center node and a supervision mechanism node jointly form the alliance chain network, and fusion data are stored and shared; federal learning adaptive model optimization is adopted, an edge computing MEC server is deployed at each substation node, and each substation node trains an AI model on the edge computing MEC server based on local data; according to the method, distributed security storage, cross-station credible sharing and distributed model training optimization of the substation inspection data are realized, the model generalization ability is improved while data privacy is protected, the problems of data islands, privacy leakage, response delay and the like in a traditional scheme are solved, and the operation and maintenance efficiency and security of the substation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance of power systems, and particularly to an intelligent inspection optimization method for power equipment. Background Art

[0002] With the gradual expansion of the power grid scale, the number of substations and substation equipment in the power system has increased several times, which has caused a sharp rise in the inspection workload of the power system. The inspection work is an important task to ensure the normal operation of the power system. Once there are omissions in the inspection work, it is very likely to cause substation equipment failures, and in severe cases, it may even threaten the safe and stable operation of the entire power system.

[0003] The traditional power system inspection has the following problems:

[0004] (1) Serious data islands: The data of each substation is stored independently, lacking the ability of cross-site collaborative analysis, and it is difficult to form a global operation and maintenance strategy;

[0005] (2) Insufficient real-time performance: The existing system relies on centralized processing in the cloud, and network latency leads to a lag in abnormal response;

[0006] (3) Weak privacy and security: Inspection data is prone to leakage or being tampered with during transmission and storage, and the device status data involves the safety of the power grid and requires a higher level of protection;

[0007] (4) Poor model generalization ability: The AI model trained in a single substation is difficult to meet the requirements of device status recognition in different environments.

[0008] The prior art such as patent CN114665608B proposes an intelligent perception inspection system and method for substations, which is an inspection scheme based on edge-cloud collaboration. However, its data management still relies on a centralized architecture, with a risk of single-point failure; model training requires uploading raw data, with a high risk of data leakage; and it does not solve the problem of cross-substation knowledge migration. Summary of the Invention

[0009] In order to solve the problems existing in the prior art, the present invention provides an intelligent inspection optimization method for power equipment that combines blockchain technology, a federated learning framework, and multi-modal data fusion, so as to achieve distributed and secure storage and cross-site trusted sharing of substation inspection data, as well as distributed model training optimization, improve the model generalization ability while protecting data privacy, solve problems such as data islands, privacy leakage, and response delay in traditional solutions, and significantly improve the operation and maintenance efficiency and security of substations.

[0010] The technical solution of the present invention is as follows:

[0011] An intelligent inspection optimization method for power equipment, comprising:

[0012] Obtain multi-source original data of substation equipment, and fuse the multi-source original data to form fused data;

[0013] Construct a consortium blockchain network. Each substation serves as a node, and multiple substation nodes, operation and maintenance center nodes, and regulatory agency nodes jointly form a consortium blockchain network to store and share the fused data;

[0014] Federated learning adaptive model optimization. Deploy edge computing MEC servers at each substation node respectively. Each substation node trains an AI model on the edge computing MEC server based on local data, and only uploads the gradient parameters of the AI model to the aggregation server. The aggregation server collects the AI model gradient parameters of each substation node, then uses the FedAvg algorithm to integrate the AI model gradient parameters, updates the global model parameters and distributes them to each substation node, and each substation node synchronously updates the global model parameters.

[0015] Furthermore, the multi-source original data of the substation equipment is obtained through a LiDAR laser radar, an infrared thermal imager, an ultrasonic flaw detector, and a vibration sensor; the LiDAR laser radar is used for equipment surface deformation detection and spatial modeling to generate high-precision three-dimensional point cloud data; the infrared thermal imager is used to capture the equipment temperature distribution; the ultrasonic flaw detector is used to detect hidden defects of the equipment, and the hidden defects include mechanical cracks and transformer insulating oil bubbles; the vibration sensor is used to collect equipment mechanical vibration data;

[0016] The process of fusing the multi-source original data to form fused data includes:

[0017] The LiDAR laser radar provides the three-dimensional geometric information (x, y, z) of the equipment, and the infrared thermal imager provides the two-dimensional temperature distribution (T| (x，y) ) of the same equipment surface. Through coordinate system conversion, the pixel coordinates (x, y) of the infrared image are mapped into the three-dimensional point cloud of the laser radar, and finally a four-dimensional point cloud P fused (x, y, z, T) with temperature attribute (T) is generated;

[0018] The ultrasonic flaw detector outputs the defect position P US of the same equipment. Through coordinate system conversion, the defect position P US is mapped into the three-dimensional point cloud of the laser radar, and finally a four-dimensional point cloud P US with defect position P fused (x, y, z, US) is generated;

[0019] The vibration data output by the vibration sensor is marked in the three-dimensional point cloud of the laser radar.

[0020] Preferably, the generation of the four-dimensional point cloud P with temperature attribute (T)fused (x, y, z, T) process includes:

[0021] Determine the origin of the LiDAR coordinate system through the calibration board, establish a global three-dimensional reference system, set marker points on the device surface, synchronously capture the spatial coordinates and temperature values of the marker points by the LiDAR and the infrared thermal imager, and use the least squares method to fit the coordinate transformation matrix to align the marker point coordinates of the LiDAR and the infrared thermal imager in space:

[0022]

[0023] Where:

[0024] T LiDAR→IR : The transformation matrix from the LiDAR coordinate system to the infrared thermal imager coordinate system;

[0025] Represents finding the transformation matrix T that minimizes the objective function by the least squares method;

[0026] Sum the errors of N marker points;

[0027] P LiDAR，i : The three-dimensional coordinates (x, y, z) of the i-th marker point in the LiDAR coordinate system;

[0028] P IR，i : The two-dimensional coordinates (x, y) of the i-th marker point in the infrared thermal imager coordinate system;

[0029] T·P IR，i : The coordinates after mapping the two-dimensional coordinates in the infrared thermal imager coordinate system to the LiDAR coordinate system through the transformation matrix T;

[0030] ||·|| 2 : The square of the Euclidean distance, used to measure the difference between two coordinates;

[0031] Fuse the LiDAR three-dimensional point cloud and the infrared image of the infrared thermal imager in space, map the temperature attribute (T) to the LiDAR three-dimensional point cloud, and generate a four-dimensional point cloud P with temperature attribute (T) fused (x, y, z, T) fused data:

[0032] P fused (x, y, z, T) = LiDAR(x, y, z) ∪ IR(T| (x，y) )

[0033] Where:

[0034] P fused(x, y, z, T): The fused four-dimensional point cloud fusion data with temperature attributes, including spatial coordinates and temperature values;

[0035] LiDAR(x, y, z): The three-dimensional point cloud coordinates collected by the LiDAR laser radar;

[0036] IR(T| (x，y) ): The temperature values measured by the infrared thermal imager on a two-dimensional plane;

[0037] ∪: Represents the spatial association of data, that is, the temperature value is assigned to the corresponding three-dimensional point cloud through coordinate mapping.

[0038] Preferably, the process of generating the four-dimensional point cloud P US with the defect position P fused (x, y, z, US) is as follows:

[0039] Determine the origin of the LiDAR coordinate system through a calibration board, establish a global three-dimensional reference system, set marking points on the device surface, the LiDAR laser radar captures the spatial coordinates of the marking points, and the ultrasonic flaw detector captures the distance and angle of the marking points relative to the ultrasonic flaw detector probe. Use the least squares method to fit the coordinate transformation matrix to align the marking point coordinates of the ultrasonic flaw detector and the LiDAR laser radar in space:

[0040]

[0041] Where:

[0042] T US→LiDAR : The transformation matrix from the ultrasonic coordinate system to the LiDAR coordinate system;

[0043] Indicates finding the transformation matrix T that minimizes the objective function through the least squares method;

[0044] Sum the errors of N marking points;

[0045] P LiDAR，i : The coordinates of the i-th marking point in the LiDAR coordinate system;

[0046] P US，i : The coordinates of the i-th marking point in the ultrasonic coordinate system;

[0047] T·P US，i : The coordinates after mapping the coordinates in the ultrasonic coordinate system to the LiDAR coordinate system through the transformation matrix T;

[0048] ||·|| 2 : The square of the Euclidean distance, used to measure the difference between two coordinates;

[0049] The defect position P output by the ultrasonic flaw detector US is spatially fused with the 3D point cloud of the lidar, and the defect position P US is mapped into the 3D point cloud of the lidar to generate a four-dimensional point cloud P US with the defect position P fused (x, y, z, US) fusion data:

[0050] P fused (x, y, z, US) = LiDAR(x, y, z) ∪ P US ;

[0051] Where:

[0052] P fused (x, y, z, US): The four-dimensional point cloud fusion data with the defect position after fusion, including spatial coordinates and defect position;

[0053] LiDAR(x, y, z): The 3D point cloud coordinates collected by the LiDAR lidar;

[0054] P US : The defect position output by the ultrasonic flaw detector;

[0055] ∪: Represents the spatial association of data, that is, the defect position is assigned to the corresponding 3D point cloud through coordinate mapping;

[0056] The process of annotating the vibration data output by the vibration sensor in the 3D point cloud of the lidar is as follows:

[0057] The vibration sensor is installed at a known position on the surface of the device, and the lidar coordinates of the installation position of the vibration sensor are recorded; the vibration data output by the vibration sensor is associated with the lidar coordinates of the installation position of the vibration sensor, and the vibration data is directly annotated in the 3D point cloud of the lidar.

[0058] To achieve all-round and multi-dimensional perception of the substation equipment status, the LiDAR includes a fixed LiDAR and a mobile LiDAR, the ultrasonic flaw detector includes a fixed ultrasonic flaw detector and a mobile ultrasonic flaw detector, the mobile LiDAR and the mobile ultrasonic flaw detector are arranged on the autonomous inspection humanoid robot body, the mobile LiDAR is used to supplement the blind area data of the fixed LiDAR, and the mobile ultrasonic flaw detector is used to temporarily detect specific areas; a vision camera and an IMU inertial measurement unit are also arranged on the autonomous inspection humanoid robot body, the vision camera is used to identify equipment appearance defects and environmental perception, and the IMU inertial measurement unit is used for the positioning and motion control of the autonomous inspection humanoid robot; the autonomous inspection humanoid robot communicates with the edge computing MEC server through the 5G network, and the edge computing MEC server processes the data collected by the autonomous inspection humanoid robot in real time and generates a dynamic inspection path planning instruction.

[0059] Further, the process of the edge computing MEC server generating the dynamic inspection path planning instruction is as follows:

[0060] The autonomous inspection humanoid robot uploads the point cloud of the mobile LiDAR, the visual image of the vision camera, and the positioning information of the IMU inertial measurement unit to the edge computing MEC server through the 5G network. The edge computing MEC server runs the Cartographer algorithm, fuses multi-sensor data to generate a high-precision grid map, generates an initial optimal path based on the Dijkstra algorithm, and uses the dynamic window method DWA to calculate the speed and steering instructions in real time:

[0061]

[0062] where v is the linear velocity of the robot; ω is the angular velocity; α, β are weight coefficients;

[0063] The edge computing MEC server issues a control instruction through the 5G URLLC channel. The autonomous inspection humanoid robot adjusts its motion state according to the instruction and feeds back its actual position to the edge computing MEC server to form a closed loop.

[0064] Preferably, the process of the consortium blockchain network storing and sharing the fusion data is as follows:

[0065] Perform a hash operation on the fused data to generate a unique data fingerprint. Package the data fingerprint, timestamp, previous block fingerprint, and substation node signature to generate a new block. The substation node that generates the new block broadcasts the new block to all other nodes in the consortium blockchain network. After receiving the new block, other nodes verify it through the PBFT consensus mechanism. The new block that passes the verification of the PBFT consensus mechanism is added to the consortium blockchain network, and all nodes synchronously update their respective ledgers. All data is encrypted during storage and transmission. Define the rules and permissions for data sharing through smart contracts. When one node requests to access the data of other nodes, it sends a data request to the consortium blockchain network. The smart contract automatically verifies whether the data request node has access rights. If the data request node has access rights, the smart contract decrypts the data and provides it to the data request node. Otherwise, the smart contract rejects the request of the data request node.

[0066] Preferably, the process of optimizing the federated learning adaptive model is as follows:

[0067] (1) Initialization phase

[0068] The aggregation server randomly initializes the global model parameters and distributes the initial parameters to all substation nodes. Each substation node labels the local data and constructs a training set where N i is the number of samples of substation node i;

[0069] (2) Local AI model training

[0070] Substation node i downloads the current global model parameters and uses the local data D i to train the model, with the goal of minimizing the loss function

[0071]

[0072] where η is the learning rate, and the initial value is set by the aggregation server; is the gradient of the loss function;

[0073] (3) Parameter encryption and upload

[0074] Substation node i uses the Paillier encryption algorithm to encrypt the local gradient to generate ciphertext The encrypted gradient parameters are transmitted to the aggregation server through the consortium blockchain network. The aggregation server verifies the identity and data integrity of substation node i through the smart contract;

[0075] (4) Global model aggregation

[0076] Dynamically adjust the aggregation weight according to the data volume of substation node i using the FedAvg algorithm:

[0077]

[0078] Among them, α is a regulation factor used to control the influence intensity of data volume on the weight; K is the total number of substation nodes participating in this round of training;

[0079] The aggregation server decrypts the gradient ciphertext and calculates the weighted average gradient:

[0080]

[0081] Update the global model parameters:

[0082]

[0083] Among them, η global is the global learning rate;

[0084] (5) Model parameter distribution and update

[0085] Prune and quantize the global model, and only distribute the differences in global model parameters to each substation node, and each substation node synchronously updates the global model parameters.

[0086] Furthermore, during the local AI model training in step (2), it also includes gradient clipping for preventing gradient explosion to perform norm constraint on the gradient:

[0087]

[0088] Among them, τ is a preset gradient threshold.

[0089] In order to dynamically adjust the global learning rate and improve the convergence speed and stability, an adaptive optimization strategy is also included between step (4) and step (5), specifically as follows:

[0090] (I) Dynamic adjustment of the global learning rate

[0091] Adjust the global learning rate η according to the model convergence speed global :

[0092]

[0093] Among them, is the validation set loss, calculated by the aggregation server;

[0094] (II) Node elimination mechanism

[0095] If the local training loss of substation node i does not decrease in three consecutive rounds, it is marked as a low-quality node and the substation node i is suspended from participating in aggregation.

[0096] The present invention is beneficial in that:

[0097] (1) Data fusion of multi-source raw data can form a visual interface on the geometric structure of the equipment, superimpose temperature, vibration and defect information, and achieve comprehensive and multi-dimensional perception of the equipment status, providing a basis for subsequent management and analysis.

[0098] (2) Multiple substation nodes, operation and maintenance center nodes, and regulatory agency nodes together form a consortium chain network, which adopts the PBFT consensus mechanism to ensure the network's efficiency and fault tolerance. Each node maintains a complete consortium chain ledger to store the hash value and related metadata of all inspection data. Through the chain structure of the consortium chain, the source and historical operations of any data can be traced to ensure data transparency and traceability, and to ensure that the data cannot be tampered with. Through the automatic execution of data verification, permission management, and model parameter exchange rules by smart contracts, it is ensured that only authorized nodes can access specific data, realizing secure and controllable data sharing, breaking data silos, and thus realizing distributed secure storage, cross-site trusted sharing, and efficient management of substation inspection data, significantly improving data security and the overall performance of the system.

[0099] (3) Each substation node uses local data to train the AI ​​model on the edge computing MEC server, maintains the lightweight AI model, and only uploads the gradient parameters of the AI ​​model. Based on the alliance chain, the node contribution is verified, and the model parameters are dynamically weighted and aggregated. The optimized global model only sends the global model parameter differences. The global model is optimized through distributed model training to avoid cross-station transmission of original data. While protecting data privacy, the model generalization ability is improved, which solves the problems of data islands, privacy leakage and response delay in traditional solutions, and significantly improves the operation and maintenance efficiency and safety of substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 It is a flow chart of the intelligent inspection optimization method for electric power equipment of the present invention;

[0101] Figure 2 It is a flowchart of the federated learning adaptive model optimization of the present invention. DETAILED DESCRIPTION

[0102] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0103] like Figure 1As shown in the figure, this embodiment is an intelligent inspection optimization method for power equipment, including: obtaining multi-source original data of substation equipment, and fusing the multi-source original data to form fusion data; then constructing a consortium blockchain network, where each substation is used as a node, and multiple substation nodes, operation and maintenance center nodes, and regulatory agency nodes jointly form a consortium blockchain network to store and share the fusion data; it also includes federated learning adaptive model optimization. An edge computing MEC server is deployed at each substation node. Each substation node trains an AI model on the edge computing MEC server based on local data, and only uploads the gradient parameters of the AI model to the aggregation server. The aggregation server collects the AI model gradient parameters of each substation node, then uses the FedAvg algorithm to integrate the AI model gradient parameters, updates the global model parameters, and distributes them to each substation node. Each substation node synchronously updates the global model parameters.

[0104] In this embodiment, the multi-source original data of substation equipment is obtained through a LiDAR lidar, an infrared thermal imager, an ultrasonic flaw detector, and a vibration sensor; among them, the LiDAR lidar is used for detecting surface deformation of the equipment and spatial modeling, providing geometric structure information of the equipment, generating high-precision three-dimensional point cloud data, so as to accurately locate the surface deformation of the equipment; the infrared thermal imager is used for capturing the temperature distribution of the equipment, so as to identify local overheating anomalies; the ultrasonic flaw detector is used for detecting hidden defects of the equipment and outputting the defect positions. The hidden defects include mechanical cracks and transformer insulating oil bubbles; the vibration sensor is used for collecting mechanical vibration data of the equipment to reflect the structural health status; the collected original data is all subjected to cleaning, denoising, and formatting processing to ensure the accuracy and consistency of the data.

[0105] In this embodiment, the process of fusing the multi-source original data to form fusion data includes:

[0106] The LiDAR lidar provides the three-dimensional geometric information (x, y, z) of the equipment, and the infrared thermal imager provides the two-dimensional temperature distribution (T| (x，y) ) of the same equipment surface. Through coordinate system transformation, the pixel coordinates (x, y) of the infrared image are mapped into the three-dimensional point cloud of the lidar, and finally a four-dimensional point cloud P with temperature attribute (T) is generated fused (x, y, z, T); the detailed process is as follows:

[0107] Determine the origin of the lidar coordinate system through a calibration board, establish a global three-dimensional reference system, set marker points on the equipment surface, and the LiDAR lidar and the infrared thermal imager synchronously capture the spatial coordinates and temperature values of the marker points. Use the least squares method to fit the coordinate system transformation matrix to align the marker point coordinates of the LiDAR lidar and the infrared thermal imager in space:

[0108]

[0109] Wherein:

[0110] T LiDAR→IR : The transformation matrix from the LiDAR coordinate system to the infrared thermal imager coordinate system;

[0111] Indicates finding the transformation matrix T that minimizes the objective function through the least squares method;

[0112] Sum the errors of N marked points;

[0113] P LiDAR，i : The three-dimensional coordinates (x, y, z) of the i-th marked point in the LiDAR coordinate system;

[0114] P IR，i : The two-dimensional coordinates (x, y) of the i-th marked point in the infrared thermal imager coordinate system;

[0115] T·P IR，i : The coordinates after mapping the two-dimensional coordinates in the infrared thermal imager coordinate system to the LiDAR coordinate system through the transformation matrix T;

[0116] ||·|| 2 : The square of the Euclidean distance, used to measure the difference between two coordinates;

[0117] Perform spatial fusion of the LiDAR three-dimensional point cloud and the infrared image of the infrared thermal imager, map the temperature attribute (T) into the LiDAR three-dimensional point cloud, and generate a four-dimensional point cloud P with temperature attribute (T) fused (x, y, z, T) fusion data:

[0118] P fused (x, y, z, T) = LiDAR(x, y, z) ∪ UR(T| (x，y) )

[0119] Wherein:

[0120] P fused (x, y, z, T): The fused four-dimensional point cloud fusion data with temperature attribute, including spatial coordinates and temperature values;

[0121] LiDAR(x, y, z): The three-dimensional point cloud coordinates collected by the LiDAR;

[0122] IR(T| (x，y) ): The temperature values measured by the infrared thermal imager on the two-dimensional plane;

[0123] ∪: Represents the spatial association of data, that is, assigning temperature values to the corresponding three-dimensional point cloud through coordinate mapping.

[0124] The ultrasonic flaw detector outputs the defect position P of the same device US , through coordinate system conversion, the defect position P US is mapped into the three-dimensional point cloud of the lidar, and finally a four-dimensional point cloud P US with the defect position P fused (x, y, z, US) is generated; the detailed process is as follows:

[0125] Determine the origin of the lidar coordinate system through a calibration board, establish a global three-dimensional reference system, set marker points on the device surface, the LiDAR lidar captures the spatial coordinates of the marker points, and the ultrasonic flaw detector captures the distance and angle of the marker points relative to the ultrasonic flaw detector probe. Use the least squares method to fit the coordinate system conversion matrix to align the marker point coordinates of the ultrasonic flaw detector and the LiDAR lidar in space:

[0126]

[0127] Among them:

[0128] T US→LiDAR : The conversion matrix from the ultrasonic coordinate system to the lidar coordinate system;

[0129] Indicates finding the conversion matrix T that minimizes the objective function through the least squares method;

[0130] Sum the errors of N marker points;

[0131] P LiDAR,i : The coordinates of the i-th marker point in the lidar coordinate system;

[0132] P US，i : The coordinates of the i-th marker point in the ultrasonic coordinate system;

[0133] T·P US，i : The coordinates after mapping the coordinates in the ultrasonic coordinate system to the lidar coordinate system through the conversion matrix T;

[0134] ||·|| 2 : The square of the Euclidean distance, used to measure the difference between two coordinates;

[0135] Fuse the defect position P output by the ultrasonic flaw detector US with the lidar three-dimensional point cloud, map the defect position P US into the lidar three-dimensional point cloud, and generate a four-dimensional point cloud P US with the defect position P fused (x, y, z, US) fusion data:

[0136] Pfused (x, y, z, US) = LiDAR(x, y, z) ∪ P US

[0137] Where:

[0138] P fused (x, y, z, US): The fused four-dimensional point cloud fusion data with defect positions, including spatial coordinates and defect positions;

[0139] LiDAR(x, y, z): The three-dimensional point cloud coordinates collected by the LiDAR lidar;

[0140] P US : The defect positions output by the ultrasonic flaw detector;

[0141] ∪: Represents the spatial association of data, that is, the defect positions are assigned to the corresponding three-dimensional point cloud through coordinate mapping.

[0142] The vibration data output by the vibration sensor is marked in the three-dimensional point cloud of the lidar. The detailed process is as follows:

[0143] Install the vibration sensor at a known position on the surface of the device, and record the lidar coordinates of the installation position of the vibration sensor; associate the vibration data output by the vibration sensor with the lidar coordinates of the installation position of the vibration sensor, and the vibration data is directly marked in the three-dimensional point cloud of the lidar.

[0144] Application scenario examples:

[0145] Example 1: Joint detection of transformer overheating and mechanical deformation

[0146] The LiDAR lidar detects that the surface depression deformation of the transformer is greater than 2 mm, and the infrared thermal imaging shows that the temperature in the corresponding area rises by ΔT ≥ 15 °C. The fusion result triggers an alarm of "local overheating caused by mechanical deformation", and it is recommended to immediately conduct on-site verification.

[0147] Example 2: Identification of internal insulating oil bubbles in a transformer

[0148] The ultrasonic flaw detector detects that the amplitude of the echo signal of the internal insulating oil bubbles in the transformer is greater than the threshold, and the vibration sensor captures abnormal high-frequency vibrations greater than 8 kHz. The fusion result is judged as "deterioration of insulating oil", and it is recommended to immediately take samples for testing.

[0149] To achieve all-round and multi-dimensional perception of the status of substation equipment, the LiDAR includes a fixed LiDAR and a mobile LiDAR, and the ultrasonic flaw detector includes a fixed ultrasonic flaw detector and a mobile ultrasonic flaw detector. The fixed LiDAR and the fixed ultrasonic flaw detector are installed around the substation equipment to monitor static parameters in real time; the mobile LiDAR and the mobile ultrasonic flaw detector are carried on the body of the autonomous inspection humanoid robot. They move with the autonomous inspection humanoid robot to cover different areas. The mobile LiDAR is used to supplement the blind area data of the fixed LiDAR, and the mobile ultrasonic flaw detector is used to temporarily detect specific areas, so as to dynamically cover the inspection blind area; a vision camera and an IMU inertial measurement unit are also set on the body of the autonomous inspection humanoid robot. The vision camera is used to identify equipment appearance defects and environmental perception, and the IMU inertial measurement unit is used for the positioning and motion control of the autonomous inspection humanoid robot; an edge computing MEC server is deployed in the substation. The autonomous inspection humanoid robot communicates with the edge computing MEC server through a 5G network. The edge computing MEC server processes the data collected by the autonomous inspection humanoid robot in real time and generates dynamic inspection path planning instructions. The process of the edge computing MEC server generating dynamic inspection path planning instructions is as follows:

[0150] The autonomous inspection humanoid robot uploads the point cloud of the mobile LiDAR, the visual image of the vision camera, and the positioning information of the IMU inertial measurement unit to the edge computing MEC server through a 5G network. The edge computing MEC server runs the Cartographer algorithm, fuses multi-sensor data to generate a high-precision grid map, generates an initial optimal path based on the Dijkstra algorithm, and uses the dynamic window approach DWA to calculate speed and steering instructions in real time:

[0151]

[0152] where, v is the linear velocity of the robot; ω is the angular velocity; α, β are weight coefficients;

[0153] The edge computing MEC server issues control instructions through the 5G URLLC channel. The autonomous inspection humanoid robot adjusts its motion state according to the instructions and feeds back its actual position to the edge computing MEC server to form a closed loop.

[0154] This embodiment adopts the 5G-MEC edge intelligent architecture, and builds a low-latency, highly reliable local intelligent operation and maintenance platform by integrating 5G communication network and mobile edge computing MEC technology, achieving millisecond-level response, and supporting real-time analysis of substation inspection and dynamic decision-making of robots; the end-to-end E2E communication delay of the 5G communication network is ≤10ms, which meets the real-time control requirements, supports multi-channel high-definition video (4K / 8K), laser point cloud and other large-flow data transmission, and ensures the continuity of key services; edge computing MEC servers are deployed in substations to provide localized computing power (such as NVIDIA T4 GPU), and user-plane functions are sunk to MEC nodes. Compared with the traditional path: terminal → base station → core network → cloud, this embodiment can optimize the path: terminal → base station → MEC, reduce cloud dependence, reduce bandwidth requirements, process data locally, avoid leakage of sensitive information, realize real-time, intelligent and autonomous substation inspection, and provide efficient and reliable technical support for power system operation and maintenance.

[0155] Application scenario examples

[0156] Example 1:

[0157] When a temporary obstacle appears on the inspection path, the mobile LiDAR laser radar detects the obstacle, and the data is transmitted to the edge computing MEC server via the 5G communication network. The edge computing MEC server replans the path and issues new instructions. The autonomous inspection humanoid robot receives the instructions and turns to avoid the obstacle.

[0158] Example 2:

[0159] The fixed infrared thermal imager found that the temperature of the transformer was abnormal and exceeded the threshold (greater than 85°C), triggering the alliance chain alarm. The AI ​​model determined it to be a "cooling system failure". The edge computing MEC server sent instructions to the autonomous inspection humanoid robot. After receiving the instructions, the autonomous inspection humanoid robot inspected the abnormal point, remeasured the temperature through the onboard visual camera, and used a mobile ultrasonic flaw detector to check the defects of the abnormal point and the status of the oil pump. The data was transmitted to the edge computing MEC server via the 5G communication network. After analyzing and confirming the fault, a maintenance work order was generated to guide the operation and maintenance personnel to the designated location.

[0160] The process of storing and sharing fused data in the alliance chain network of this embodiment is as follows:

[0161] Perform a hash operation on the fused data to generate a unique data fingerprint. Package the data fingerprint, timestamp, previous block fingerprint, and substation node signature to generate a new block. The substation node that generates the new block broadcasts the new block to all other nodes in the consortium blockchain network. After receiving the new block, other nodes verify it through the PBFT consensus mechanism to ensure the legality and data integrity of the new block. The new block that passes the verification of the PBFT consensus mechanism is added to the consortium blockchain network, and all nodes synchronously update their respective ledgers. All data is encrypted during storage and transmission to ensure data confidentiality, and the original data is not leaked during data sharing and model training. Define the rules and permissions for data sharing through smart contracts to ensure that only authorized nodes can access specific data. When one node requests to access the data of other nodes, it sends a data request to the consortium blockchain network. The smart contract automatically verifies whether the data request node has access permissions. If the data request node has access permissions, the smart contract decrypts the data and provides it to the data request node. Otherwise, the smart contract rejects the request of the data request node.

[0162] Example: Substation A performs a hash operation on the fused data to generate a data hash value H_A. Substation A packages the data hash value H_A, timestamp T_A, previous block fingerprint P_A, and substation node signature S_A to generate a new block B_A. Substation A broadcasts the new block B_A to all other nodes in the consortium blockchain network. After receiving the new block B_A, other nodes verify it through the PBFT consensus mechanism to ensure the legality and data integrity of the new block B_A. If the verification passes, the new block B_A is added to the consortium blockchain network, and all nodes synchronously update their respective ledgers. When Substation B needs to access the data of Substation A, it sends a data request to the consortium blockchain network. The smart contract automatically verifies the permissions of Substation B. If the verification passes, it decrypts the data and provides it to Substation B. All data operations are recorded on the consortium blockchain. Through the chain structure of the consortium blockchain, the source and historical operations of any data can be traced to ensure data transparency and traceability and ensure that the data cannot be tampered with. Automatically execute data verification, permission management, and model parameter exchange rules through smart contracts to ensure that only authorized nodes can access specific data, realize secure and controllable data sharing, break data islands, and thus achieve distributed secure storage, cross-station trusted sharing, and efficient management of substation inspection data, significantly improving data security and the overall performance of the system.

[0163] As Figure 2 shown, in this embodiment, the process of optimizing the federated learning adaptive model is as follows:

[0164] (1) Initialization phase

[0165] The aggregation server randomly initializes the global model parameters And send the initial parameters to all substation nodes. Each substation node labels the local data to construct a training set Where N i Is the number of samples of substation node i;

[0166] (2) Local AI model training

[0167] Substation node i downloads the current global model parameters Use the local data D i Train the model with the goal of minimizing the loss function

[0168]

[0169] Where η is the learning rate, and the initial value is set by the aggregation server; Is the gradient of the loss function;

[0170] Gradient clipping: To prevent gradient explosion, perform norm constraint on the gradient:

[0171]

[0172] Where τ is the preset gradient threshold, such as 1.0;

[0173] (3) Parameter encryption and upload

[0174] Substation node i uses the Paillier encryption algorithm to encrypt the local gradient Generate ciphertext The encrypted gradient parameters are transmitted to the aggregation server through the consortium blockchain network to ensure transmission security. The aggregation server verifies the identity and data integrity of substation node i through a smart contract and rejects illegal parameters;

[0175] (4) Global model aggregation

[0176] Use the FedAvg algorithm to dynamically adjust the aggregation weights according to the data volume of substation node i:

[0177]

[0178] Where α is the adjustment factor, with a default value of 0.5, used to control the influence intensity of the data volume on the weight; K is the total number of substation nodes participating in this round of training;

[0179] The aggregation server decrypts the gradient ciphertext and calculates the weighted average gradient:

[0180]

[0181] Update the global model parameters:

[0182]

[0183] Among them, η global is the global learning rate;

[0184] (5) Model parameter distribution and update

[0185] Prune and quantize the global model to reduce the occupancy of the transmission bandwidth, and only distribute the differences in the global model parameters to each substation node, further reducing the communication cost, and each substation node synchronously updates the global model parameters.

[0186] In order to dynamically adjust the global learning rate and improve the convergence speed and stability, an adaptive optimization strategy is also included between step (4) and step (5), which is specifically as follows:

[0187] (I) Dynamic adjustment of the global learning rate

[0188] Adjust the global learning rate η according to the model convergence speed global :

[0189]

[0190] Among them, is the validation set loss, calculated by the aggregation server;

[0191] (II) Node elimination mechanism

[0192] If the local training loss of substation node i does not decrease for three consecutive rounds, it is marked as a low-quality node, and substation node i is suspended from participating in the aggregation.

[0193] Application scenario example

[0194] Example 1:

[0195] Substation A discovers surface deformation and local overheating of the transformer through data fusion. The fused data is hashed and chained to generate a trustworthy record, which is authorized and shared to other substations such as Substation B and C. Each substation uses the shared data to train an AI model to identify similar defect features.

[0196] Example 2:

[0197] Substation A is in a coastal high-humidity environment, and its trained AI model is good at identifying insulator contamination defects. Substation B is in an inland dry environment, and its trained AI model is good at identifying mechanical structure cracks and high temperatures. After federated aggregation, the global model can simultaneously identify insulator contamination defects, mechanical structure cracks and high temperatures, thus fusing the data features of multiple sites, adapting to the equipment status identification in different environments, realizing the efficient training and continuous optimization of the substation inspection model, and providing reliable technical support for intelligent operation and maintenance.

[0198] Each substation node uses local data to train an AI model on the edge computing MEC server, maintains a lightweight AI model, and only uploads the gradient parameters of the AI model. Based on the consortium chain to verify node contributions, dynamically weighted aggregation of model parameters, and only the differences in the global model parameters are sent down for the optimized global model; the optimization of the global model is achieved through distributed model training, avoiding cross-station transmission of raw data, enhancing the generalization ability of the model while protecting data privacy, solving problems such as data islands, privacy leakage, and response latency in traditional solutions, and significantly improving the operation and maintenance efficiency and security of substations.

[0199] The above embodiments are only used to illustrate the technical solutions of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form. Any technical solutions obtained by means of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims

1. An intelligent inspection optimization method for power equipment, characterized in that, Including: Obtain multi-source original data of substation equipment, and fuse the multi-source original data to form fused data; Construct a consortium blockchain network. Each substation serves as a node. Multiple substation nodes, operation and maintenance center nodes, and regulatory agency nodes jointly form a consortium blockchain network and store and share the fused data; Federated learning adaptive model optimization. Deploy edge computing MEC servers at each substation node respectively. Each substation node trains an AI model on the edge computing MEC server based on local data, and only uploads the gradient parameters of the AI model to the aggregation server. The aggregation server collects the gradient parameters of the AI model of each substation node, then uses the FedAvg algorithm to integrate the gradient parameters of the AI model, and then updates the global model parameters and distributes them to each substation node. Each substation node synchronously updates the global model parameters.

2. The intelligent inspection optimization method for power equipment according to claim 1, wherein The multi-source original data of the substation equipment is obtained through a LiDAR lidar, an infrared thermal imager, an ultrasonic flaw detector, and a vibration sensor; the LiDAR lidar is used for detecting surface deformation and spatial modeling of the equipment to generate high-precision three-dimensional point cloud data; the infrared thermal imager is used for capturing the temperature distribution of the equipment; the ultrasonic flaw detector is used for detecting hidden defects of the equipment, and the hidden defects include mechanical cracks and transformer insulating oil bubbles; the vibration sensor is used for collecting mechanical vibration data of the equipment; The process of fusing the multi-source original data to form fused data includes: The LiDAR laser radar provides three-dimensional geometric information (x, y, z) of the device, and the infrared thermal imager provides two-dimensional temperature distribution (T| (x，y) ), through coordinate system conversion, the pixel coordinates (x, y) of the infrared image are mapped to the three-dimensional point cloud of the lidar, and finally a four-dimensional point cloud P with temperature attribute (T) is generated. fused (x, y, z, T); The ultrasonic flaw detector outputs the defect position P of the same device US , through coordinate system transformation, the defect position P US is mapped into the three-dimensional point cloud of the lidar, and finally a four-dimensional point cloud P US with the defect position P fused (x, y, z, US) is generated; The vibration data output by the vibration sensor is marked in the lidar three-dimensional point cloud.

3. The intelligent inspection optimization method for power equipment according to claim 2, wherein The process of generating a four-dimensional point cloud P with a temperature attribute (T) fused (x, y, z, T) includes: Determine the origin of the lidar coordinate system through a calibration board, establish a global three-dimensional reference system, set marker points on the equipment surface. The LiDAR lidar and the infrared thermal imager synchronously capture the spatial coordinates and temperature values of the marker points, and use the least squares method to fit the coordinate transformation matrix to align the marker point coordinates of the LiDAR lidar and the infrared thermal imager in space: Where: T LiDAR→IR : The transformation matrix from the lidar coordinate system to the infrared thermal imager coordinate system; It means to find the transformation matrix T that minimizes the objective function through the least squares method; Sum the errors of N marked points; P LIDAR，i : The three-dimensional coordinates (x, y, x) of the i-th marked point in the lidar coordinate system; P IR，i : The two-dimensional coordinates (x, y) of the i-th marked point in the coordinate system of the infrared thermal imager; T·P Ir，i : The coordinates obtained by mapping the two-dimensional coordinates in the infrared thermal imager coordinate system to the lidar coordinate system through the transformation matrix T; ||·|| 2 : The square of the Euclidean distance, used to measure the difference between two coordinates; Spatially fuse the 3D point cloud of the lidar with the infrared image of the infrared thermal imager, map the temperature attribute (T) into the 3D point cloud of the lidar, and generate a four-dimensional point cloud P with the temperature attribute (T). fused (x, y, z, T) fusion data: P fused (x,y,z,T) = LiDAR(x,y,z) ∪ IR(T| (x,y) ) Where: P fused (x, y, z, T): The fused four-dimensional point cloud fusion data with temperature attributes, including spatial coordinates and temperature values; LiDAR(x,y,z): The three-dimensional point cloud coordinates collected by the LiDAR lidar; IR(T| (x,y) ): The temperature value measured by the infrared thermal imager on a two-dimensional plane; ∪: Represents the spatial association of data, that is, assigns the temperature value to the corresponding three-dimensional point cloud through coordinate mapping.

4. The intelligent inspection optimization method for power equipment according to claim 2, wherein, The process of generating the four-dimensional point cloud P US with the defective position P fused (x, y, z, US) is as follows: Determine the origin of the lidar coordinate system through a calibration board, establish a global three-dimensional reference system, set marker points on the equipment surface. The LiDAR lidar captures the spatial coordinates of the marker points, and the ultrasonic flaw detector captures the distance and angle of the marker points relative to the ultrasonic flaw detector probe, and uses the least squares method to fit the coordinate transformation matrix to align the marker point coordinates of the ultrasonic flaw detector and the LiDAR lidar in space: Where: T US→LiDAR : Transformation matrix from the ultrasonic coordinate system to the lidar coordinate system; It means to find the transformation matrix T that minimizes the objective function through the least squares method; Sum the errors of N marked points; P LiDAR，i : The coordinates of the i-th marked point in the lidar coordinate system; P US，i : The coordinates of the i-th marked point in the ultrasonic coordinate system; T·P US，i : The coordinates after mapping the coordinates in the ultrasonic coordinate system to the lidar coordinate system through the transformation matrix T; ||·|| 2 : The square of the Euclidean distance, used to measure the difference between two coordinates; The defect position P output by the ultrasonic flaw detector US is spatially fused with the 3D point cloud of the lidar, and the defect position P US is mapped into the 3D point cloud of the lidar to generate a four-dimensional point cloud P US with the defect position P fused (x, y, z, US) Fusion data: P fused (x, y, z, US) = LiDAR(x, y, z) ∪ P US ; Where: P fused (x, y, z, US): The fused four-dimensional point cloud fusion data with defect positions, including spatial coordinates and defect positions; LiDAR(x,y,z): The three-dimensional point cloud coordinates collected by the LiDAR lidar; P US : The defect location output by the ultrasonic flaw detector; ∪: Represents the spatial association of data, that is, assigns the defect position to the corresponding three-dimensional point cloud through coordinate mapping; The process of marking the vibration data output by the vibration sensor in the lidar three-dimensional point cloud is as follows: Install the vibration sensor at a known position on the surface of the device, and record the LiDAR coordinates of the installation position of the vibration sensor; associate the vibration data output by the vibration sensor with the LiDAR coordinates of the installation position of the vibration sensor, and directly label the vibration data in the LiDAR three-dimensional point cloud.

5. The intelligent inspection optimization method for power equipment according to claim 2, wherein, The LiDAR includes a fixed LiDAR and a mobile LiDAR. The ultrasonic flaw detector includes a fixed ultrasonic flaw detector and a mobile ultrasonic flaw detector. The mobile LiDAR and the mobile ultrasonic flaw detector are arranged on the body of the autonomous inspection humanoid robot. The mobile LiDAR is used to supplement the blind area data of the fixed LiDAR, and the mobile ultrasonic flaw detector is used to temporarily detect specific areas; a vision camera and an IMU inertial measurement unit are also arranged on the body of the autonomous inspection humanoid robot. The vision camera is used to identify device appearance defects and environmental perception, and the IMU inertial measurement unit is used for the positioning and motion control of the autonomous inspection humanoid robot; the autonomous inspection humanoid robot communicates with the edge computing MEC server through the 5G network, and the edge computing MEC server processes the data collected by the autonomous inspection humanoid robot in real time and generates dynamic inspection path planning instructions.

6. The intelligent inspection and optimization method for power equipment according to claim 5, wherein The process by which the edge computing MEC server generates dynamic inspection path planning instructions is as follows: The autonomous inspection humanoid robot uploads the LiDAR point cloud of the mobile LiDAR, the visual image of the vision camera, and the positioning information of the IMU inertial measurement unit to the edge computing MEC server through the 5G network. The edge computing MEC server runs the Cartographer algorithm, fuses multi-sensor data to generate a high-precision grid map, generates an initial optimal path based on the Dijkstra algorithm, and uses the dynamic window method DWA to calculate the speed and steering instructions in real time: where v is the linear velocity of the robot; ω is the angular velocity; α, β are weight coefficients; The edge computing MEC server issues control instructions through the 5G URLLC channel. The autonomous inspection humanoid robot adjusts its motion state according to the instructions and feeds back its actual position to the edge computing MEC server to form a closed loop.

7. The intelligent inspection optimization method for power equipment according to claim 1, wherein, The process by which the consortium blockchain network stores and shares the fusion data is as follows: Perform a hash operation on the fused data to generate a unique data fingerprint. Package the data fingerprint, timestamp, previous block fingerprint, and substation node signature to generate a new block. The substation node that generates the new block broadcasts the new block to all other nodes in the consortium blockchain network. After receiving the new block, other nodes verify it through the PBFT consensus mechanism. The new block that passes the verification of the PBFT consensus mechanism is added to the consortium blockchain network, and all nodes synchronously update their respective ledgers. All data is encrypted during storage and transmission. Define the rules and permissions for data sharing through smart contracts. When one node requests to access the data of other nodes, it sends a data request to the consortium blockchain network. The smart contract automatically verifies whether the data request node has access permissions. If the data request node has access permissions, the smart contract decrypts the data and provides it to the data request node; otherwise, the smart contract rejects the request of the data request node.

8. The intelligent inspection optimization method for power equipment according to claim 1, wherein The process of optimizing the federated learning adaptive model is as follows: (1) Initialization phase The aggregation server randomly initializes the global model parameters and distributes the initial parameters to all substation nodes. Each substation node annotates the local data to construct a training set where N i is the number of samples of substation node i; (2) Local AI model training Substation node i downloads the current global model parameters Uses local data D i Trains the model with the goal of minimizing the loss function Among them, η is the learning rate, and its initial value is set by the aggregation server; is the gradient of the loss function; (3) Parameter encryption and upload The substation node i uses the Paillier encryption algorithm to encrypt the local gradient and generate the ciphertext The encrypted gradient parameters are transmitted to the aggregation server through the consortium blockchain network. The aggregation server verifies the identity and data integrity of the substation node i through a smart contract; (4) Global model aggregation Use the FedAvg algorithm to dynamically adjust the aggregation weight according to the data volume of substation node i: where α is a regulatory factor used to control the influence intensity of data volume on the weight; K is the total number of substation nodes participating in this round of training; The aggregation server decrypts the gradient ciphertext and calculates the weighted average gradient: Update the global model parameters: where η global is the global learning rate; (5) Model parameter distribution and update Prune and quantize the global model, and only send the differences in the global model parameters to each substation node, and each substation node synchronously updates the global model parameters.

9. The intelligent inspection optimization method for power equipment according to claim 8, wherein During the process of local AI model training in step (2), it also includes gradient clipping for preventing gradient explosion to perform norm constraint on the gradient: where τ is a preset gradient threshold.

10. The intelligent inspection optimization method for power equipment according to claim 8, characterized in that An adaptive optimization strategy is also included between step (4) and step (5), specifically as follows: (I) Dynamic adjustment of the global learning rate Adjust the global learning rate η according to the model convergence speed global : Among them, is the validation set loss, calculated by the aggregation server; (II) Node elimination mechanism If the local training loss of substation node i does not decrease for 3 consecutive rounds, it is marked as a low-quality node, and substation node i is suspended from participating in the aggregation.

Images