A method and system for power inspection task offloading state encoding

Abstract

The application relates to the technical field of electric digital data processing, and discloses a power inspection task unloading state coding method and system, which comprises the following steps: acquiring multi-modal original observation data collected by an inspection terminal in a power inspection task, and identifying key components and state information of a target power equipment based on an image analysis model; performing defect analysis on the key components to generate an evaluation result containing defect types and quantitative severity; fusing task calculation attributes, communication environment parameters and equipment resource states to construct a multi-dimensional state vector; performing normalization and vectorization processing on the dimensional features of the multi-dimensional state vector to output a state space representation for task unloading decision-making of a reinforcement learning intelligent agent. The application solves the problems of disconnection of general state coding, inaccurate decision-making and untimely response in the prior art, and achieves the purposes of precise decision-making, targeted response and energy efficiency improvement.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing technology, and in particular to a method and system for encoding the unloading status of power inspection tasks. Background Technology

[0002] With the development of smart grids, power equipment inspection is gradually evolving from manual to automated and intelligent methods. Traditional power equipment inspection mainly relies on standardized procedures and the experience of operation and maintenance personnel, which has limitations such as low efficiency, high error rate, and high skill requirements for personnel. Mobile terminals such as drones and inspection robots have been widely used in transmission lines, substations, and other scenarios, undertaking various inspection tasks such as insulator damage identification and conductor joint temperature monitoring.

[0003] In edge computing architectures, inspection terminals often face limited computing resources, requiring task offloading mechanisms to transfer some computing tasks to edge servers or the cloud. Existing technologies still have room for improvement in inspection efficiency and the accuracy of intelligent model analysis. While there are methods for detecting abnormal overheating of power equipment based on infrared video data from power line inspections, they lack effective integration with task offloading decisions.

[0004] Existing task offloading methods are mostly based on general reinforcement learning frameworks, whose state spaces typically only contain general parameters such as network latency, bandwidth, equipment power consumption, and task size, lacking modeling of the specific characteristics of power inspection tasks. These methods suffer from insufficient specificity in power scenarios, making it difficult to dynamically optimize offloading strategies based on equipment status. For example, for severely damaged insulator areas, local processing should be prioritized to ensure real-time response; while for joints with slight temperature anomalies, a certain delay can be tolerated, and offloading can be performed at the edge for in-depth analysis. Furthermore, the current state representation does not effectively integrate semantic information about power equipment defects, such as component type, defect location, and severity level, leading to a disconnect between task offloading decisions and actual maintenance needs, impacting overall inspection efficiency and safety. Equipment reasoning capability evaluation should consider factors such as equipment defect classification, comparison of power information differences, power fault defect reasoning, and equipment relationship topology; however, existing systems have failed to effectively integrate these factors into task scheduling decisions.

[0005] For example, Chinese Patent CN120892795A discloses a simulation test optimization method, system, storage medium, and electronic device for power inspection equipment, providing the following technical solution: acquiring multi-source data and constructing a digital twin model of the power inspection equipment based on the multi-source data; acquiring the inspection task parameters of the power inspection equipment and calculating and outputting the optimal power inspection scheme; deploying at least two digital twin models of the power inspection equipment in a virtual power inspection scenario and conducting virtual scenario testing to form virtual power inspection test data; conducting actual power inspection scenario testing on the power inspection equipment, and optimizing the power inspection equipment based on the virtual power inspection scenario test data and the actual power inspection scenario test data to form an optimized power inspection equipment scheme, thereby improving simulation test efficiency. However, the aforementioned power inspection equipment simulation test optimization method, system, storage medium, and electronic equipment cannot achieve intelligent task offloading decisions based on power equipment characteristics and defect severity. It lacks fine-grained coding of equipment semantics, defect types, and levels, resulting in a disconnect between simulation testing and actual inspection business needs. Its optimization is limited to model building while neglecting the accuracy of task scheduling. Summary of the Invention

[0006] This invention solves the problems of disconnected general status coding, inaccurate decision-making, and untimely response in the prior art. It proposes a method and system for unloading status coding in power inspection tasks, which achieves the goals of accurate decision-making, targeted response, and improved energy efficiency.

[0007] The purpose of this invention is to provide a task offloading status coding method and system based on the characteristics of power equipment in power inspection scenarios. By constructing a multi-dimensional status coding scheme that integrates the identification results of key components of power equipment, defect type and severity assessment, the reinforcement learning agent can accurately perceive the characteristics of inspection tasks, improve the pertinence and decision-making efficiency of task offloading strategies, reduce the waste of human resources, and increase the frequency and efficiency of intelligent operation and maintenance of power system inspection.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for encoding the unloading status of power inspection tasks includes: Acquire multimodal raw observation data collected by inspection terminals during power line inspection tasks, and identify key components and status information of target power equipment based on image analysis models; Defect analysis is performed on the key components to generate assessment results that include defect types and quantified severity. Then, by integrating task computation attributes, communication environment parameters, and device resource status, a multi-dimensional state vector is constructed. The features of each dimension of the multi-dimensional state vector are normalized and vectorized to output a state space representation for the reinforcement learning agent to make task offloading decisions.

[0009] By integrating multi-source data and intelligent analysis, reinforcement learning agents can accurately perceive the characteristics of power inspection tasks, significantly improve the targeting and decision-making efficiency of task offloading strategies, reduce human resource waste, and increase the frequency and overall efficiency of intelligent operation and maintenance of power system inspections, thus making up for the shortcomings of traditional methods.

[0010] A system for encoding the unloading status of power inspection tasks includes: a data acquisition module, an identification and analysis module, and a decision execution module; the data acquisition module includes a data acquisition unit and a key component identification unit, the key component identification unit is connected to the defect assessment unit in the identification and analysis module, the decision execution module is equipped with a reinforcement learning decision unit and a task scheduling execution unit, and the identification and analysis module is connected to the reinforcement learning decision unit through its internally set status encoding engine.

[0011] The system architecture features a modular design, enabling seamless integration of data acquisition, analysis, and decision-making. It supports real-time status updates and online optimization, improving the overall efficiency and scalability of the task unloading system.

[0012] Preferably, the image analysis model is a deep learning-based visual recognition model that extracts key component features of power equipment from visible light images and infrared thermal imaging data, and outputs recognition results containing component category semantic labels and spatial location information; the component category semantic labels are encoded through a predefined power equipment component dictionary.

[0013] By leveraging deep learning technology to improve the accuracy of key component identification and semantic understanding capabilities, ensuring that component type information is standardized and encoded, and enhancing the professionalism of state representation, the system can optimize decisions for different power equipment characteristics, thereby improving the business adaptability of inspection tasks.

[0014] Preferably, the defect analysis specifically includes: generating an evaluation result containing defect type coding and quantified severity score by integrating an image segmentation algorithm and a temperature analysis algorithm; the defect type coding is mapped according to a predefined semantic hierarchy of defects, converting the defect categories described in the text into discrete numerical values ​​that retain the order of severity.

[0015] By fusion of algorithms to preserve the semantic hierarchy of defects, state coding can reflect the relative severity of different defect types, improve the understanding of the impact of reinforcement learning agents on business, thereby achieving more accurate offloading decisions and reducing the false negative rate of high-risk tasks.

[0016] Preferably, the quantitative severity score is calculated by integrating visual damage features, temperature anomaly features, and equipment operation history features through a lightweight neural network model or expert system. The score is a numerical value that is continuously distributed in the range of 0 to 1 to characterize the overall severity level of the defect.

[0017] By integrating information from multiple sources to classify defects, the scoring accurately reflects the overall severity level of defects, enhances the reliability of state vectors, helps agents distinguish between urgent and non-urgent tasks, optimizes resource allocation, and improves inspection safety and efficiency.

[0018] Preferably, the multi-dimensional state vector consists of at least a device semantic feature dimension, a defect state dimension, a task attribute dimension, and an environmental state dimension; the device semantic feature dimension represents the type attribute of the key component of the power equipment targeted by the current task; the defect state dimension represents the existence, type, and severity of the identified defects on the component; the task attribute dimension represents the computational storage and real-time requirements of the inspection task itself; and the environmental state dimension represents the communication conditions and computing resource availability during task execution.

[0019] Multidimensional state vectors comprehensively cover business and technical parameters, constructing a refined state representation, enabling the decision-making system to dynamically adjust strategies based on equipment specificity, improving the targeting of task offloading, and thus enhancing overall inspection efficiency.

[0020] Preferably, the multi-dimensional state vector also includes a historical decision dimension, which represents the decision-making experience of similar tasks in recent times. The state vector also includes a dynamic update mechanism, which refreshes the state vector in real time according to the scene during the inspection. The triggering conditions include, but are not limited to, data collection triggering, defect change triggering, environmental change triggering, and periodic triggering.

[0021] By introducing historical experience and dynamic update mechanisms, the timeliness and adaptability of the state vector are enhanced, ensuring that the decision-making system can respond quickly to environmental changes, reduce the task failure rate, optimize resource utilization, and improve the stability of the system in complex scenarios.

[0022] Preferably, the normalization and vectorization processing specifically includes: employing specific mathematical transformation methods for dimensional features with different physical meanings and dimensions; for task computation, performing linear normalization after logarithmic transformation to compress its numerical range; for communication signal strength, normalizing it using a linear mapping function based on its typical physical range; for categorical features, converting them into numerical features using one-hot encoding or numerical mapping encoding; and mapping all dimensional features to a unified numerical range after processing, forming a standardized state space representation that can be directly processed by reinforcement learning agents.

[0023] Ensuring that features across all dimensions are comparable within a unified range simplifies input processing for reinforcement learning models, improves decision-making speed and accuracy, and reduces computational complexity, making the system more suitable for deployment on resource-constrained edge devices.

[0024] Preferably, the multimodal raw observation data includes visible light images, infrared thermal imaging data, and device data, wherein the device data includes device location information and sensor status.

[0025] Preferably, the data acquisition unit is connected to the key component identification unit, the defect evaluation unit is connected to the state coding engine, and the reinforcement learning decision unit is connected to the task scheduling and execution unit; the state coding engine fuses the identification and evaluation results with other contextual information to generate a standardized state vector; the reinforcement learning decision unit receives the state vector and outputs the task unloading target node.

[0026] The reinforcement learning decision unit quickly outputs the unloading target based on the standardized state vector, optimizes task scheduling, reduces latency, and improves the system's response speed and reliability in power inspection scenarios.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0028] 1. This invention significantly improves the professionalism of task offloading decisions by constructing a multi-dimensional state coding scheme that integrates the identification results of key components of power equipment with defect assessment. Traditional methods often rely on general parameters such as network latency or task size, ignoring the specific needs of power operations, resulting in a serious disconnect between state coding and actual operation and maintenance. This scheme combines equipment semantic features with defect severity scoring, enabling the reinforcement learning agent to accurately perceive the characteristics of inspection tasks. This ensures that the offloading strategy closely aligns with the actual needs of the power scenario, effectively reducing the waste of human resources and significantly improving inspection frequency and efficiency.

[0029] 2. This invention, by introducing defect semantic information and severity scoring, enables the intelligent agent to deeply understand the business impact of different defects, thereby making more accurate decisions. The system can prioritize local processing for high-risk tasks to ensure real-time performance, while selecting an offloading path that tolerates delays for minor anomalies. This semantic-based decision-making mechanism overcomes the limitations of traditional methods that rely solely on technical parameters, improving the targeted nature of the response.

[0030] 3. This invention significantly optimizes resource utilization efficiency through refined state modeling, thereby comprehensively improving overall inspection efficiency. The system is integrated into edge computing devices, supporting dynamic updates and online learning. This integrated application enables drone or robot platforms to adjust strategies in real time, effectively reducing the rate of missed detections of critical defects and response latency. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of a method for encoding the unloading status of power inspection tasks according to the present invention.

[0032] Figure 2 This is a system block diagram of one embodiment of a system for encoding the unloading status of power inspection tasks according to the present invention. Detailed Implementation

[0033] See Figures 1-2 As shown, a method for power line inspection task unloading status coding includes: Acquire multimodal raw observation data collected by inspection terminals during power line inspection tasks, and identify key components and status information of target power equipment based on image analysis models; Defect analysis is performed on the key components to generate assessment results that include defect types and quantified severity. Then, by integrating task computation attributes, communication environment parameters, and device resource status, a multi-dimensional state vector is constructed. The features of each dimension of the multi-dimensional state vector are normalized and vectorized to output a state space representation for the reinforcement learning agent to make task offloading decisions.

[0034] A system for encoding the unloading status of power inspection tasks includes: a data acquisition module, an identification and analysis module, and a decision execution module; the data acquisition module includes a data acquisition unit and a key component identification unit, the key component identification unit is connected to the defect assessment unit in the identification and analysis module, the decision execution module is equipped with a reinforcement learning decision unit and a task scheduling execution unit, and the identification and analysis module is connected to the reinforcement learning decision unit through its internally set status encoding engine.

[0035] like Figure 1 In one embodiment shown, Figure 1 This is an overall flowchart of a method for encoding the unloading status of a power inspection task according to the present invention. First, the present invention initiates the process by acquiring multimodal raw observation data collected by the inspection terminal during the power inspection task, including visible light images, infrared thermal imaging data, and equipment location information and sensor status. Then, based on a pre-trained deep learning visual recognition model, key component features of the power equipment are extracted from this data, and a recognition result containing semantic labels for component categories and spatial location information is output. The component categories are encoded using a predefined power equipment component dictionary.

[0036] Next, defect analysis is performed on the identified key components. By integrating image segmentation and temperature analysis algorithms, an evaluation result is generated that includes defect type coding and a quantified severity score. The defect type coding maps according to a predefined semantic hierarchy of defects, converting the textual description of defect categories into discrete numerical values ​​that retain the order of severity. Then, the quantified severity score is calculated by fusing visual impairment features, temperature anomaly features, and equipment operating history features through a lightweight neural network model or expert system. This score is a continuous value in the range of 0 to 1 to characterize the overall severity level of the defect.

[0037] Then, by integrating task calculation attributes, communication environment parameters, and device resource status, a multi-dimensional state vector is constructed. This vector consists of at least the device semantic feature dimension, defect status dimension, task attribute dimension, and environment status dimension. Optionally, it also includes a historical decision dimension to represent decision-making experience of similar tasks in recent times, and supports a dynamic update mechanism to refresh the state vector in real time based on data collection, defect changes, environmental changes, or periodic triggering conditions.

[0038] Finally, the features of each dimension of the multi-dimensional state vector are normalized and vectorized: for features with different physical meanings and dimensions, specific mathematical transformation methods are adopted. For example, logarithmic transformation is used followed by linear normalization to compress the numerical range of the task computation, and linear mapping is performed on the communication signal strength according to a typical range. Categorical features are converted to numerical types using one-hot encoding or numerical mapping. After all feature processing, they are mapped to a unified numerical range, forming a standardized state space representation, which can be directly used by the reinforcement learning agent for task offloading decisions. The entire process ensures the accuracy and real-time performance of state encoding, improving the intelligence level of power line inspection.

[0039] In another embodiment, the present invention provides the following technical solution: A task offloading status encoding method based on power equipment characteristics in a power inspection scenario includes the following steps: 1. Obtain raw observation data for power line inspection tasks, including but not limited to visible light images, infrared thermal imaging data, equipment location information, and sensor status.

[0040] 2. Analyze the observed data using a pre-trained deep learning model to identify key components of the target power equipment and output component category labels and spatial location information.

[0041] 3. Based on the identification results, perform defect detection and severity assessment on key components, and generate defect type codes and level scores.

[0042] 4. Combining task computation requirements, communication environment parameters, and device resource status, construct a multi-dimensional state vector, which shall include at least the following dimensions: Equipment semantic feature dimension: includes the key component type code corresponding to the current task (such as "insulator", "conductor joint", "surge arrester" etc.); Defect status dimension: includes defect presence flag, defect type code (such as "crack", "flashover mark", "oxidation corrosion") and severity score (0 to 1 continuous value or graded label) output by expert system or model; Task attribute dimensions: including the computational load (CPU cycles), data size, and processing latency requirements; Environmental status dimensions include current communication link bandwidth, signal strength, edge node load rate, and remaining terminal battery power. Historical decision dimension (optional): includes the uninstallation choices and execution results (success / failure, delay exceeding limits, etc.) of the most recent N similar tasks.

[0043] 5. Normalize and vectorize the features of each dimension mentioned above to form a unified reinforcement learning state space representation for the agent to make decisions on task unloading actions.

[0044] The specific processing methods for each dimension of features are as follows: Equipment semantic feature encoding: One-hot encoding technology is used to represent the key component types of power equipment, and the encoding order is [transformer, circuit breaker, surge arrester, ...].

[0045] Defect presence flag processing: The defect detection result is binarized into 0-1 flag bits, where 1 indicates that a defect was detected and 0 indicates that no defect was detected.

[0046] Defect type semantic encoding: Defect types such as "cracks", "flashover marks", and "oxidation corrosion" are converted into continuous values ​​through predefined mapping relationships. The specific mapping rules are as follows: "Defect-free" is defined as 0; "Corrosion" is defined as 1; "Flashover trace" is defined as 2; A "crack" is defined as 3; A “through crack” is defined as 4; "Severe damage" is defined as 5; "Oxidative corrosion" is defined as 6; "Mechanical wear" is defined as 7.

[0047] This mapping preserves the relative severity relationships between different types of defects, enabling the reinforcement learning state space to reflect the semantic hierarchy of defects in power equipment.

[0048] Defect severity quantification: The severity of defects is directly represented by a continuous value of 0 to 1. This value is calculated by an expert system or a lightweight neural network model by fusing multi-source information such as visual feature saliency and temperature anomaly amplitude.

[0049] Task attribute dimension processing: Computational Feature Processing: Addressing the large computational demands (typically 10) in power image analysis tasks. 6 ~10 12 (CPU cycles), using a combination of logarithmic transformation and linear normalization, the original computational load is mapped to the standardized interval [0,1]. The specific transformation is: C norm equal to log 10 Divide C by 6 and then subtract 1.

[0050] Data volume feature processing: The data volume of images, videos and other data generated by the inspection task is linearly normalized to map the original data size to the range of [0,1].

[0051] Latency requirement processing: The latency constraints of task processing are linearly normalized and converted into standardized latency requirement indicators.

[0052] Environmental state dimension processing: Communication link bandwidth processing: The real-time measured communication bandwidth is linearly normalized and converted into a standardized index within the range of [0,1].

[0053] Wireless signal strength processing: In view of the large fluctuations in signal strength at the power field (usually -100~-50dBm), a dedicated linear mapping function is designed to convert the signal strength into a continuous value in the range of [0,1].

[0054] Edge node load rate processing: The real-time load status of edge computing nodes is standardized and directly converted into load indicators in the range of [0,1].

[0055] Terminal remaining power processing: Normalize the battery power of the inspection terminal (such as drones and inspection robots) and convert it into a continuous value in the range of [0,1].

[0056] Historical decision-making dimension processing: Historical Unloading Selection Encoding: The unloading selection (local, edge, cloud) of historical tasks is encoded as [0.0, 0.5, 1.0] and normalized.

[0057] Historical execution results are represented as follows: the task execution result is binary converted into a success / failure flag, which is directly converted into a 0-1 value.

[0058] Historical task delay status representation: The task execution timeout status is binarized and converted into a 0-1 value.

[0059] The severity score of defects is calculated by integrating multi-source information such as visual feature saliency, temperature anomaly range, and component operating years, and then using a weighted fusion or lightweight neural network model to achieve equipment defect classification and fault reasoning.

[0060] The neural network model has approximately 1200 parameters, and its structure is as follows: 1. Input layer Visual feature saliency (3D): [Crack length percentage, surface damage area, color anomaly]; Temperature anomaly magnitude (3D): [temperature rise magnitude, percentage of anomaly areas, concentration of hotspots]; Component operation information (3D): [Years of operation, number of historical defects, maintenance record score].

[0061] 2. Feature Selection Layer Device type embedding vector: Encode the device type (“insulator”, “conductor_joint”, “arrester”) as a 3-dimensional one-hot vector; Gating mechanism: Feature weight vectors are generated through a small MLP to enable feature selection for device perception.

[0062] 3. Multi-branch processing layer Visual branch: 3→8→4 (fully connected layer, ReLU activation); Temperature branch: 3→8→4 (fully connected layer, ReLU activated); Running information branch: 3→6→3 (fully connected layer, ReLU activated).

[0063] 4. Fusion Layer The output is obtained by concatenating the three branches (4+4+3=11 dimensions); 11→8→5 (fully connected layer, ReLU activated); Introduce defect type embedding: map defect types to 5-dimensional embedding vectors and add them to the fused features.

[0064] 5. Nonlinear reinforcement layer 5→3→1 (fully connected layer, Softplus activated); Output range control: sigmoid (output) ensures a range of 0 to 1.

[0065] 6. Output layer: Output defect severity score (0~1).

[0066] The training and inference process of a neural network is as follows: The training settings are as follows: Dataset Construction: The training data comes from the historical inspection data of three 500kV substations and twelve 220kV transmission lines; Total sample size: 8,642 valid samples (including a positive to negative sample ratio of 3:7); Tag source: Jointly assessed by three senior power system engineers according to the DL / T 664-2016 standard; Data augmentation: The sample size was increased to 25,926 by means of rotation, brightness adjustment, and Gaussian noise injection.

[0067] The loss function is specifically: L equals α multiplied by MSE(y) pred y true ), plus β multiplied by RankLoss, plus γ multiplied by Reg; where α=0.7 is the mean squared error weight, β=0.2 (to ensure the defect level order of "severe damage > through crack > crack"), γ=0.1 is the L2 regularization coefficient, and λ=10 -4 .

[0068] Optimizer and training parameters: Optimizer: AdamW (learning rate equal to 3 × 10) -4 , β1=0.9, β2=0.999, weight_decay=10 -4 ); Batch size: 64 (limited by edge device memory); Training rounds: 150 rounds, early stop mechanism (patience=20); Learning rate scheduling: Cosine annealing strategy, minimum learning rate = 10 -6 ; Hardware platform: NVIDIA Tesla T4 GPU × 2 (training time: 2.3 hours).

[0069] The inference performance is specifically as follows: Reasoning speed: Real-world performance on typical inspection terminal hardware (Qualcomm QCS610 SoC, 2.2GHz CPU, Adreno 612 GPU): Mean inference latency: 18.7 ± 2.3 ms; 99th percentile delay: <25 ms; It meets the real-time requirements of power inspection (<50 ms).

[0070] Energy consumption performance: Energy consumption per inference cycle: 3.2 ± 0.4 mJ; In continuous operation mode (5 inferences per second): Power consumption: 16.8 mW; Impact on drone / robot battery life: <0.05% / minute; Compared to traditional methods (such as SVM), energy consumption is reduced by 67%.

[0071] The deployment and optimization are as follows: Model quantization: Deployed using 8-bit integer quantization (INT8); The model size was compressed from 4.7KB to 1.2KB; Accuracy loss: <0.5% RMSE.

[0072] Edge device adaptation: Supports both TensorFlow Lite and ONNX Runtime engines; Memory usage: Static memory <15KB, dynamic memory <50KB.

[0073] The integration method with the power equipment inspection system is as follows: / / Example API call DefectSeverityScore calculate_defect_severity( const VisualFeatures* visual, const ThermalFeatures* thermal, const OperationalInfo* operational, const DeviceType device_type ).

[0074] The status coding scheme supports dynamic updates, refreshing the status vector in real time during the inspection process as new data is collected and analyzed, ensuring the timeliness of decision-making.

[0075] The conditions for triggering a state update are: Data acquisition trigger: Updated after each frame of image analysis is completed (change rate > 15%); Defect change trigger: Update immediately when the defect score changes by more than 0.1 or a new defect is discovered; Environmental change trigger: Update when communication quality changes >10dBm or battery level drops to critical thresholds (20%, 10%); Periodic trigger: When there are no other triggers, it updates automatically every 500ms (reduced to 100ms in high-risk situations).

[0076] Compared to traditional solutions, the performance indicators of this invention are as follows: index This invention Traditional methods Average update latency 32.5ms 85.7ms High-risk update frequency 10Hz 3Hz Energy consumption per update 5.7mJ 17.5mJ Resource usage 23.5% CPU 40%+ CPU

[0077] like Figure 2 In one embodiment shown, Figure 2 This is a system block diagram of a power inspection task unloading status coding system according to one embodiment of the present invention. The power inspection task unloading status coding system includes: a data acquisition module, an identification and analysis module, and a decision execution module. The data acquisition module includes a data acquisition unit and a key component identification unit. The key component identification unit is connected to a defect assessment unit within the identification and analysis module. The decision execution module includes a reinforcement learning decision unit and a task scheduling execution unit. The identification and analysis module is connected to the reinforcement learning decision unit through an internally configured status coding engine. The data acquisition unit is connected to the key component identification unit, the defect assessment unit is connected to the status coding engine, and the reinforcement learning decision unit is connected to the task scheduling execution unit. The status coding engine fuses the identification and assessment results with other contextual information to generate a standardized status vector. The reinforcement learning decision unit receives the status vector and outputs the task unloading target node. The power inspection task unloading status coding system is connected to a power inspection terminal.

[0078] In another embodiment, the present invention adopts the following technical solution: A power line inspection task offloading system, comprising: Data acquisition module: used to acquire multimodal sensing data from the inspection terminal; Key component identification module: Implements equipment component detection and classification based on convolutional neural networks or visual Transformer models; Defect assessment module: integrates image segmentation and temperature analysis algorithms, and outputs defect type and severity level; State coding engine: responsible for fusing recognition and evaluation results with other contextual information to generate standardized state vectors; Reinforcement learning decision-making module: Receives state vectors and outputs task unloading target nodes (local, edge, cloud). Task scheduling and execution module: Completes task distribution and resource coordination based on the decision results.

[0079] The system is deployed in drones, inspection robots, or edge gateway devices and supports online learning and policy updates.

[0080] The advantages of this invention are: 1. Highly professional: The identification results of key components of power equipment and the assessment of the severity of defects are integrated into the status representation of task unloading, which solves the problem of the disconnect between traditional general status coding and power business needs; 2. More accurate decision-making: By introducing defect semantic information, reinforcement learning agents can understand the business impact of different defect types, thereby prioritizing local processing or low-latency offloading paths for high-risk tasks; 3. More targeted response: For example, when a "through crack" is detected in an insulator and the severity score is > 0.8, the system automatically tends to process it locally to reduce transmission delay and ensure the timeliness of emergency alarms; 4. Good scalability: The status coding structure is modular, supporting the addition of new equipment types and defect categories, and is suitable for different voltage levels and power grid scenarios; 5. Improve overall inspection efficiency: Through refined condition modeling, optimize resource utilization efficiency, reduce the rate of missed detection of key defects and response delay, and make up for the shortcomings of traditional power transmission and transformation equipment inspection and maintenance methods.

[0081] Example 1: Detection of abnormal temperature at transformer joints in substations During a routine inspection of a 500kV substation, an inspection robot was checking the bushing joint area of ​​the main transformer. As a critical connection point, even slight temperature anomalies in the transformer joints could indicate a serious fault.

[0082] Data Acquisition: The inspection robot is equipped with a visible light camera and an infrared thermal imager to collect multimodal data of the transformer joint area. The visible light image resolution is 1920×1080, the infrared thermal image resolution is 320×240, and the temperature measurement range is -20℃ to 150℃.

[0083] Key component identification: The key component identification module uses an improved YOLOv5 model to identify the target area containing the "transformer bushing joint" component, with spatial coordinates (x1, y1, x2, y2) = (420, 310, 580, 450).

[0084] Defect assessment: Visual analysis revealed slight oxidation marks on the joint surface, with crack length accounting for 3.2% and surface damage area of ​​1.8 cm². 2 ; Infrared data showed a local temperature rise of 18.5℃, with the abnormal area accounting for 35% and the hotspot concentration at 0.7. The connector has been in operation for 12 years, and historical records show that it has undergone two minor repairs, with a maintenance record score of 75. Defect assessment module output: Defect type code "oxidation thermal", severity score 0.65.

[0085] Status Code: Device semantic features: [1, 0, 0] ("Transformer" class, encoding order is [Transformer, Circuit Breaker, Surge Arrester]); Defect status: [1, 6, 0.65] (Defect exists, type code 6 corresponds to "oxidation corrosion", score 0.65); Task attributes: computation time = 8.5e8 cycles, data size = 3.2MB, maximum allowable latency = 5s; Environmental conditions: Bandwidth = 12Mbps, Signal strength = -68dBm, Edge node load = 45%, Battery level = 85%; Historical decisions: [Edge, Success], [Cloud, Success], [Local, Success]; After normalization, a 15-dimensional state vector is formed to input the reinforcement learning decision module.

[0086] Task offloading decision: After evaluation by the reinforcement learning model (based on the PPO algorithm), it was determined that although the defect exists, its severity is moderate (score 0.65 < 0.7), and the current communication conditions are good. Therefore, it was decided to offload the task to the edge node for in-depth analysis, while retaining local processing capabilities to handle more urgent tasks.

[0087] in conclusion: 1. In this embodiment, the system accurately identifies the combined defects of oxidation corrosion and temperature anomaly in the transformer joint, and reasonably selects the edge unloading strategy based on the severity score (0.65), thus avoiding unnecessary occupation of limited local computing resources.

[0088] 2. Compared with traditional offloading decision-making methods based solely on task size and network conditions, this system improves edge computing resource utilization by 31% and reduces average task response time by 27% while ensuring detection quality.

[0089] 3. By integrating equipment-specific status codes, the system can distinguish between defects that "require immediate handling" and those that "can be analyzed later." For such moderately severe joint problems, edge unloading is selected instead of local processing, allowing the inspection robot to continue performing subsequent tasks, improving overall inspection efficiency by 22%.

[0090] 4. Experimental data show that this method reduces the missed detection rate of transformer joint defects from 8.7% to 5.2% compared to the traditional method, and particularly improves the ability to identify early temperature anomalies.

[0091] Example 2: Corrosion Detection of Lightning Arresters for Transmission Lines During the inspection of 110kV overhead transmission lines, drones were used to inspect composite insulator surge arresters. Surge arrester corrosion can lead to protection failure during lightning strikes, making it a key monitoring target.

[0092] Data acquisition: The drone hovered 2.5 meters away from the lightning arrester and acquired high-resolution visible light images (4096×2160) and near-infrared images, while also recording the GPS location and ambient temperature and humidity.

[0093] Key component identification: The visual Transformer model identified the "lightning arrester" component, accurately located its metal end and composite jacket area, and found a significant discolored area on the metal end.

[0094] Defect assessment: Visual characteristics: Corrosion area accounts for 28.5%, edge irregularity is 0.63, and color anomalousness is 0.82; Temperature data: No significant temperature rise (only 2.3℃), abnormal area accounts for 5%, hotspot concentration is 0.21; Operational information: This surge arrester has been in operation for 8 years with no history of defects and a maintenance record score of 90. Defect assessment module output: Defect type code "corrosion", severity score 0.42.

[0095] Status Code: Equipment semantic features: [0, 0, 1] ("Surge arrester" class, encoding order is [transformer, circuit breaker, surge arrester]); Defect status: [1, 1, 0.42] (Defect exists, type code 1 corresponds to "corrosion", score 0.42); Task attributes: computation time = 2.1e9 cycles, data size = 8.7MB (including high-resolution images), maximum allowable latency = 10s; Environmental conditions: Bandwidth = 5Mbps (weak signal in mountainous areas), signal strength = -85dBm, edge node load = 80%, battery level = 45%; Historical decisions: [Cloud, Success], [Edge, Failure], [Local, Success]; The state coding engine normalizes each dimension, and performs weighted processing, especially for weak signal environments in mountainous areas.

[0096] Task unloading decision: The reinforcement learning decision module analyzes the state vector, taking into account: The severity of the defect is low (0.42 < 0.5). Current communication conditions are poor (bandwidth is only 5Mbps); Battery level is low (45%). High load on edge nodes (80%) Historical decision-making data shows that edge node processing success rates are low in this environment; The model decides to process the task in segments: basic corrosion identification is completed locally, while detailed analysis tasks are offloaded to the cloud using a segmented transmission strategy.

[0097] in conclusion: 1. This embodiment verifies how the system can make intelligent unloading decisions based on the specific state of the equipment in mountainous environments with limited communication conditions. For low-severity surge arrester corrosion problems, the system selects a hybrid "local + cloud" processing strategy to avoid transmitting large amounts of high-definition image data in weak signal environments.

[0098] 2. Compared with traditional methods, this strategy reduces task completion time by 35% and data transmission volume by 52%, effectively extending the working time of drones in low-battery conditions.

[0099] 3. By introducing equipment semantic features (such as surge arresters) and defect type features, the system can identify the characteristic that corrosion defects usually do not require emergency response, thereby optimizing resource allocation. In actual testing, this method reduced the false alarm rate of surge arrester defects by 19% while ensuring a critical defect detection rate of 98.5%.

[0100] 4. This embodiment particularly demonstrates the importance of the historical decision dimension: the system learns the pattern of edge node processing failure in similar environments, actively avoids high-load edge nodes, and improves the task success rate from 76% in the traditional method to 94%.

[0101] Example 3: Early Warning of Mechanical Faults in Substation Circuit Breakers During the spring maintenance of the 220kV substation, inspection robots conducted detailed inspections on the operating mechanisms of SF6 circuit breakers, focusing on monitoring early signs of wear on mechanical components.

[0102] Data acquisition: The robot uses a high frame rate camera (120fps) to capture video of the circuit breaker's opening and closing process, while simultaneously acquiring vibration sensor data and acoustic signals to form a multimodal dataset.

[0103] Key component identification: Based on the improved Mask R-CNN model, the key moving components of the circuit breaker operating mechanism are accurately segmented, including the linkage, crank arm and contact system.

[0104] Defect assessment: Visual characteristics: motion trajectory deviation 5.2mm (exceeding the standard by 2.1mm), component wear area ratio 12.7%, motion incoordination 0.68; Vibration data: Abnormal vibration frequency components accounted for 38%, amplitude exceeded the standard by 1.8 times, and vibration energy concentration was 0.75; Operational information: This circuit breaker has been in operation for 15 years, and historical records show 3 mechanical adjustments. The maintenance record score is 65 points. Defect assessment module output: Defect type code "mechanical_wear", severity score 0.78.

[0105] Status Code: Device semantic features: [0, 1, 0] ("Circuit breaker" class, encoding order is [Transformer, Circuit breaker, Surge arrester]); Defect status: [1, 7, 0.78] (Defect exists, type code 7 corresponds to "mechanical wear", score 0.78); Task attributes: computation time = 3.5e9 cycles (including video analysis), data size = 15.2MB, maximum allowable latency = 2s (due to mechanical failure early warning). Environmental conditions: Bandwidth = 15Mbps, Signal strength = -62dBm, Edge node load = 30%, Battery level = 60%.

[0106] Historical decisions: [Local, Success], [Local, Success], [Edge, Timeout].

[0107] Task offloading decision: After evaluating the reinforcement learning decision module, the following was considered: The defect severity is high (0.78 > 0.75 threshold). Mechanical fault early warning has extremely high real-time requirements (time delay requires a normalized value of 0.92); Local computing resources are sufficient (60% battery, no other high-priority tasks); Historical data shows that local processing has a high success rate; The model decisively chose the "local processing" strategy, and the inspection robot immediately activated its onboard AI chip for real-time analysis.

[0108] in conclusion: 1. This embodiment demonstrates the system's rapid response capability to high-risk mechanical faults. When a circuit breaker mechanical wear severity score of 0.78 is detected, the system prioritizes local processing, ensuring analysis is completed within 2 seconds, a 43% reduction in average response time compared to traditional methods (3.5 seconds).

[0109] 2. By incorporating the characteristics of equipment type (circuit breaker) and defect type (mechanical wear) into the status coding, the system understands that such defects require high real-time processing and chooses to process them locally even if the current network conditions are good (bandwidth 15Mbps) to avoid any potential delays.

[0110] 3. In practical applications, this method successfully issued early warnings for three potential circuit breaker failures, preventing possible cascading power grid faults. Comparative tests show that for mechanical defects with a severity level >0.75, this system can trigger the local processing strategy 100%, while the traditional method only has an accuracy rate of 68%.

[0111] 4. This embodiment verifies the core advantage of the present invention: encoding power expertise into the state space enables task offloading decisions to be based not only on technical parameters but also on the actual business needs of power operation and maintenance. In the circuit breaker mechanical fault early warning scenario, the system reduces the missed detection rate of critical defects from 12.3% in the traditional method to 3.7%, while optimizing the resource allocation of non-critical tasks, resulting in an overall improvement in computing resource utilization of 37%.

Claims

1. A method for encoding the unloading status of power inspection tasks, characterized in that, include: Acquire multimodal raw observation data collected by inspection terminals during power line inspection tasks, and identify key components and status information of target power equipment based on image analysis models; Defect analysis is performed on the key components to generate assessment results that include defect types and quantified severity. Then, by integrating task computation attributes, communication environment parameters, and device resource status, a multi-dimensional state vector is constructed. The features of each dimension of the multi-dimensional state vector are normalized and vectorized to output a state space representation for the reinforcement learning agent to make task offloading decisions.

2. The method for encoding the unloading status of power inspection tasks according to claim 1, characterized in that, The image analysis model is a deep learning-based visual recognition model that extracts key component features of power equipment from visible light images and infrared thermal imaging data, and outputs recognition results containing component category semantic labels and spatial location information; the component category semantic labels are encoded through a predefined power equipment component dictionary.

3. A method for encoding the unloading status of power inspection tasks according to claim 1 or 2, characterized in that, The defect analysis specifically includes: generating an evaluation result containing defect type coding and quantified severity score by integrating image segmentation algorithm and temperature analysis algorithm; the defect type coding is mapped according to a predefined semantic hierarchy of defects, converting the defect categories described in the text into discrete numerical values ​​that retain the order of severity.

4. The method for encoding the unloading status of power inspection tasks according to claim 3, characterized in that, The quantitative severity score is calculated by integrating visual damage features, temperature anomaly features, and equipment operation history features through a lightweight neural network model or expert system. The score is a numerical value that is continuously distributed in the range of 0 to 1 to characterize the overall severity level of the defect.

5. The method for encoding the unloading status of power inspection tasks according to claim 4, characterized in that, The multi-dimensional state vector is composed of at least the device semantic feature dimension, the defect state dimension, the task attribute dimension, and the environment state dimension; the device semantic feature dimension represents the type attribute of the key component of the power equipment targeted by the current task; the defect state dimension represents the existence, type, and severity of the identified defects on the component; the task attribute dimension represents the computation, storage, and real-time requirements of the inspection task itself. The environmental state dimension represents the communication conditions and computing resource availability during task execution.

6. The method for encoding the unloading status of power inspection tasks according to claim 5, characterized in that, The multi-dimensional state vector also includes a historical decision dimension, which represents the decision-making experience of similar tasks in recent times. The state vector also includes a dynamic update mechanism, which refreshes the state vector in real time according to the scene during the inspection. The triggering conditions include, but are not limited to, data collection triggering, defect change triggering, environmental change triggering, and periodic triggering.

7. The method for encoding the unloading status of power inspection tasks according to claim 6, characterized in that, The normalization and vectorization processing specifically includes: employing specific mathematical transformation methods for dimensional features with different physical meanings and dimensions; for task computation, performing linear normalization after logarithmic transformation to compress its numerical range; for communication signal strength, normalizing it using a linear mapping function based on its typical physical range; for categorical features, converting them into numerical features using one-hot encoding or numerical mapping encoding; and mapping all dimensional features to a unified numerical range after processing, forming a standardized state space representation that can be directly processed by reinforcement learning agents.

8. A method for encoding the unloading status of power inspection tasks according to claim 1 or 7, characterized in that, The multimodal raw observation data includes visible light images, infrared thermal imaging data, and device data, including device location information and sensor status.

9. A system for encoding the unloading status of power inspection tasks, employing the method for encoding the unloading status of power inspection tasks as described in any one of claims 1-8, characterized in that, include: The system includes a data acquisition module, an identification and analysis module, and a decision execution module. The data acquisition module comprises a data acquisition unit and a key component identification unit. The key component identification unit is connected to the defect assessment unit within the identification and analysis module. The decision execution module includes a reinforcement learning decision unit and a task scheduling and execution unit. The identification and analysis module is connected to the reinforcement learning decision unit through its internal state coding engine.

10. A system for encoding the unloading status of power inspection tasks according to claim 9, characterized in that, The data acquisition unit is connected to the key component identification unit, the defect evaluation unit is connected to the state coding engine, and the reinforcement learning decision unit is connected to the task scheduling and execution unit. The state coding engine fuses the identification and evaluation results with other contextual information to generate a standardized state vector. The reinforcement learning decision unit receives the state vector and outputs the task unloading target node.

Images