Power inspection model dynamic scheduling collaboration method and system based on hierarchical strategy

Abstract

The invention provides a power inspection model dynamic scheduling cooperation method and system based on a hierarchical strategy, and relates to the technical field of power system inspection, and the method comprises the steps: receiving a power inspection task, collecting multi-modal power inspection data, and carrying out the preprocessing; based on a multi-scale convolution feature segmentation algorithm, key features of the electric power inspection task are extracted; evaluating an electric power inspection task level label corresponding to the electric power inspection task in combination with the multi-dimensional task level evaluation index; a corresponding dynamic model scheduling strategy is selected according to the electric power inspection task level label, and models in the dynamic model scheduling strategy comprise a light-weight small model and a cloud large model; and inputting the key features of the electric power inspection task into a model in the dynamic model scheduling strategy, outputting an electric power inspection identification result and pushing the result to the inspection monitoring platform, so that the dynamic model scheduling strategy can be adopted, dynamic distribution and efficient collaboration of model resources are realized, and the real-time performance, the energy efficiency performance and the intelligent level of the electric power inspection system are improved.

Description

Technical Field

[0001] This invention relates to the field of power system inspection technology, and in particular to a hierarchical strategy-based dynamic scheduling and coordination method and system for power inspection models. Background Technology

[0002] In the process of intelligent transformation of the power industry, the efficiency and accuracy of power inspection are directly related to the reliable power supply capability of the power system.

[0003] In recent years, single-model analysis technology based on deep learning has been widely used in power inspection systems. By automating the analysis and fault identification of multimodal data such as images and videos on the cloud or edge servers, it has significantly replaced the traditional manual inspection mode and effectively improved the automation level and identification accuracy of power inspection operations.

[0004] However, with the diversification of power line inspection scenarios and the increasing sophistication of power line inspection task requirements, existing technologies have gradually revealed significant shortcomings. On the one hand, the inference process of large-scale deep learning models is computationally intensive and energy-intensive, making them unsuitable for long-term deployment on resource-constrained terminal devices and hindering the achievement of long-duration power line inspection operations. On the other hand, fixed model architectures lack flexibility and cannot dynamically adjust the model size according to the complexity and real-time requirements of power line inspection tasks, leading to an imbalance in resource allocation. Furthermore, current systems generally lack hierarchical scheduling and coordination mechanisms for models of different sizes, resulting in wasted resources for simple tasks and insufficient computation for complex tasks, thus reducing the utilization rate of computing resources.

[0005] At the same time, existing technologies also suffer from the problem of uneven data transmission latency and task recognition accuracy. Large-scale deep learning model inference relies on remote data transmission, which cannot meet the requirements of power inspection scenarios with high response speed and low latency transmission. Furthermore, a single model is difficult to take into account the differences in multimodal data features under multiple scenarios, which significantly affects recognition accuracy.

[0006] Therefore, it is necessary to provide a hierarchical strategy-based dynamic scheduling coordination method and system for power inspection models to solve the above-mentioned technical problems. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a hierarchical strategy-based dynamic scheduling and coordination method and system for power inspection models, which solves the problem that existing technologies cannot dynamically allocate and efficiently coordinate model resources, making it difficult to improve the real-time performance, energy efficiency, and accuracy of power inspection.

[0008] This invention provides a hierarchical strategy-based dynamic scheduling and coordination method for power inspection models, the method comprising: Receive power inspection tasks and load corresponding inspection and data collection strategies, collect multimodal power inspection data based on the inspection and data collection strategies and perform preprocessing; Based on the multi-scale convolutional feature segmentation algorithm, key features of power inspection tasks are extracted from the preprocessed multimodal power inspection data; Using multi-dimensional task level evaluation indicators, the power inspection task level label corresponding to the power inspection task is evaluated based on the key characteristics of the power inspection task. Select the corresponding dynamic model scheduling strategy according to the power inspection task level label, wherein the model in the dynamic model scheduling strategy includes lightweight small model and cloud large model; The key features of the power inspection task are input into the model in the dynamic model scheduling strategy, and the power inspection identification results are output and pushed to the inspection monitoring platform.

[0009] Preferably, the step of receiving the power inspection task and loading the corresponding inspection and data acquisition strategy, and collecting multimodal power inspection data based on the inspection and data acquisition strategy and performing preprocessing specifically includes: According to the power inspection task, the pre-stored strategy mapping library is called, and the corresponding inspection and collection strategy is loaded through parameter matching logic. The multimodal power inspection data is collected using either fixed-point collection mode or mobile inspection and collection mode. Signal conditioning technology is used to correct data defects, identify and remove invalid data, and unify the data format based on the data type of the multimodal power inspection data. An adaptive median filtering algorithm is used to dynamically adjust the window size to denoise the power inspection image data; A contrast-limited adaptive histogram equalization algorithm is used to enhance the clarity of the power inspection image data. The multimodal power inspection data is losslessly compressed using encoding compression technology.

[0010] Preferably, the step of extracting key features of the power inspection task from the preprocessed multimodal power inspection data using a multi-scale convolutional feature segmentation algorithm specifically includes: A multi-scale parallel convolutional branch network is constructed, and convolutional kernels of different scales are used to perform layer-by-layer convolutional operations on the preprocessed multimodal power inspection data to extract multi-scale feature information in parallel. A feature pyramid network is employed to fuse the multi-scale feature information through a top-down feature transfer mechanism and lateral connections. A segmentation network with an encoder-decoder structure is used to locate and filter key feature regions in the multi-scale feature information, extract the quantized features corresponding to the key feature regions, and generate the key features of the power inspection task.

[0011] Preferably, the step of using multi-dimensional task level evaluation indicators to evaluate the power inspection task level label corresponding to the power inspection task based on the key characteristics of the power inspection task specifically includes: The key features of the power inspection task are input into the power inspection task analysis unit to calculate the key feature values ​​of the power inspection task and establish a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation index. Based on the mapping relationship, the comprehensive evaluation value of the power inspection task based on the key characteristics of the power inspection task is calculated, and the power inspection task level label is generated.

[0012] Preferably, the key features of the power inspection task are input into the power inspection task analysis unit to calculate the key feature values ​​of the power inspection task and establish a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation index, specifically including: The key features of the power line inspection task are input into the power line inspection task analysis unit via the MQTT protocol. These key features are quantified and subjected to Min-Max standardization to generate key feature values. The corresponding calculation formula is as follows: In the formula, represents the key characteristic value of the power inspection task; x represents the quantitative value of the key characteristic of the power inspection task. The maximum value of the quantified key characteristics of power line inspection tasks; The minimum quantitative value representing the key characteristics of power line inspection tasks; The power inspection task analysis unit classifies the key features of the power inspection task according to the hierarchical rule knowledge base and the lightweight decision tree algorithm, and establishes a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation index, wherein the multi-dimensional task level evaluation index includes task real-time performance, task complexity, and task security level.

[0013] Preferably, the preset power inspection task level threshold, based on the mapping relationship, calculates the comprehensive evaluation value of the key features of the power inspection task, and generates the power inspection task level label, specifically including: Based on the mapping relationship, the key features of the power inspection task corresponding to the real-time performance of the task are determined, the timestamp information of the key features of the power inspection task is identified, and the key feature values ​​of the power inspection task are input into the Linear-SVM regression model to calculate the task real-time performance score. ; Based on the mapping relationship, the key features of the power line inspection task corresponding to the task complexity are determined. A lightweight random forest model is invoked, and each decision tree calculates the importance weight of the key features of the power line inspection task based on its own splitting rules. Based on the key feature values ​​of the power line inspection task and the importance weights, a weighted score for task complexity is calculated. ; Based on the mapping relationship, the key features of the power inspection task corresponding to the task safety level are determined. Based on the distance-weighted K-nearest neighbor algorithm, historical similar tasks are matched in the power inspection task analysis unit, and the task safety level score is calculated by weighting similarity. The corresponding calculation formula is as follows: In the formula, K represents the number of historically similar tasks; This represents the similarity between the power line inspection task and the j-th historical similar task; This represents the security level score of the j-th historically similar task; Indicates the correction factor; The key feature value of the power inspection task is represented by the j-th historically similar task. The Euclidean distance between the key feature value of the power inspection task and the key feature value of the power inspection task of the j-th historical similar task is represented. The CRITIC objective weighting method is used to determine the weights of the task's real-time performance, complexity, and security level. The weighted average is then used to calculate the comprehensive evaluation value of the power inspection task. The corresponding calculation formula is as follows: In the formula, , , These represent the weights of task real-time performance, task complexity, and task security level, respectively. A power inspection task level threshold is preset, and the comprehensive evaluation value of the power inspection task is compared with the power inspection task level threshold. The power inspection task level label is generated by preset task level labeling rules.

[0014] Preferably, the step of selecting a corresponding dynamic model scheduling strategy based on the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models, specifically including: A lightweight convolutional neural network architecture, 8-bit quantization technology, and model distillation technology are adopted to reduce the computational load and resource consumption of model inference and construct the lightweight small model. A multimodal Transformer architecture and cross-modal attention network are adopted to combine the fusion requirements of the power inspection task and construct the cloud-based large model. Based on the intelligent scheduling algorithm, the key characteristics of the power inspection task are perceived in real time according to the power inspection task level label, and the calling order, data transmission priority and resource allocation ratio of the lightweight small model and the cloud large model are dynamically adjusted. If the power inspection task level label is medium or low, the lightweight small model is selected first; if the power inspection task level label is high, the cloud-based large model is selected directly.

[0015] Preferably, the step of inputting the key features of the power inspection task into the model in the dynamic model scheduling strategy, outputting the power inspection identification results, and pushing them to the inspection monitoring platform specifically includes: If the power inspection task level label is medium or low, then the corresponding key features of the power inspection task are read and input into the lightweight small model; if the power inspection task level label is high, then the corresponding key features of the power inspection task are read and input into the cloud-based large model. A confidence threshold is preset. If the confidence level output by the lightweight small model is lower than the confidence threshold, the cloud-based large model is automatically triggered to perform supplementary analysis. If the lightweight small model detects that the key feature of the power inspection task is an abnormal type of power inspection task, the lightweight small model and the cloud-based large model are automatically started for parallel inference. The parallel inference results of the lightweight small model and the cloud-based large model are comprehensively evaluated, and the power inspection identification results are output and pushed to the inspection and monitoring platform.

[0016] Preferably, the step of comprehensively evaluating the parallel inference results of the lightweight small model and the cloud-based large model, outputting the power inspection identification result, and pushing it to the inspection and monitoring platform specifically includes: The confidence scores of the lightweight small model and the cloud large model are normalized and weighted using a feature weighted fusion algorithm. If the confidence score is lower than the confidence score threshold, the key features of the power inspection task corresponding to the confidence score are read, and the key features of the power inspection task are updated probabilistically through a Bayesian confidence correction mechanism to output the power inspection identification result. The power inspection and identification results are pushed to the inspection and monitoring platform through an encrypted transmission channel, and the power inspection and identification results are visualized and stored in a structured manner.

[0017] A hierarchical power inspection model dynamic scheduling and coordination system, the system comprising: The data acquisition and preprocessing module is used to receive power inspection tasks and load the corresponding inspection acquisition strategy, and to collect multimodal power inspection data based on the inspection acquisition strategy and perform preprocessing. The key feature extraction module is used to extract key features of the power inspection task from the preprocessed multimodal power inspection data based on a multi-scale convolutional feature segmentation algorithm. The grade label evaluation module is used to evaluate the power inspection task grade label corresponding to the power inspection task based on the key characteristics of the power inspection task using multi-dimensional task level evaluation indicators. The inspection model scheduling module is used to select the corresponding dynamic model scheduling strategy according to the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models. The inspection result output module is used to input the key features of the power inspection task into the model in the dynamic model scheduling strategy, output the power inspection identification results, and push them to the inspection monitoring platform.

[0018] Compared with existing technologies, the hierarchical strategy-based dynamic scheduling and coordination method and system for power inspection models provided by this invention have the following advantages: This invention can receive power inspection tasks and load corresponding inspection and acquisition strategies, collect multimodal power inspection data based on the inspection and acquisition strategies, and perform preprocessing; extract key features of power inspection tasks from the preprocessed multimodal power inspection data based on a multi-scale convolutional feature segmentation algorithm; evaluate the power inspection task level label corresponding to the power inspection task based on the key features using multi-dimensional task level evaluation indicators; select a corresponding dynamic model scheduling strategy based on the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models; input the key features of the power inspection task into the models in the dynamic model scheduling strategy, output the power inspection identification results, and push them to the inspection and monitoring platform. Thus, a dynamic model scheduling strategy can be adopted to achieve dynamic allocation and efficient collaborative model resources, improving the real-time performance, energy efficiency, and intelligence level of the power inspection system.

[0019] This invention utilizes fixed-point or mobile inspection acquisition modes and multimodal data preprocessing technology to simultaneously collect and optimize multimodal power inspection data, ensuring data integrity and high quality in complex inspection scenarios. Through a multi-scale convolutional feature segmentation algorithm, it accurately locates key feature regions and extracts and fuses key features of power inspection tasks in parallel, ensuring the integrity and accuracy of key features and solving the problems of missed detections and false detections in traditional single-scale feature extraction. Furthermore, this invention employs multi-dimensional task-level evaluation indicators to construct an evaluation system for task real-time performance, task complexity, and task safety level, achieving scientific classification and accurate label generation for power inspection tasks, bypassing traditional manual subjective classification or single-indicator classification. Limitations: This invention dynamically selects lightweight small models and cloud-based large models based on the power inspection task level labels, and can dynamically allocate models according to the type of power inspection task. This invention inputs the key features of the power inspection task into the model in the dynamic model scheduling strategy, outputs the power inspection identification results and pushes them to the inspection monitoring platform. It adopts a pre-set confidence threshold to trigger a supplementary inference mechanism, combined with feature weighted fusion algorithm and Bayesian confidence correction technology, to ensure dynamic allocation and efficient co-model resources, realize the visualization and structured storage of power inspection identification results, effectively solve the problems of wasted computing resources for simple tasks and insufficient identification rate for complex tasks, simplify the circuit inspection calculation link, and improve the accuracy of power inspection identification. Attached Figure Description

[0020] Figure 1 A flowchart of a hierarchical strategy-based dynamic scheduling and coordination method for power inspection models provided in this embodiment of the invention; Figure 2 A system block diagram of a dynamic scheduling and collaborative system for a power inspection model based on a hierarchical strategy, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 The diagram shown is a flowchart of a hierarchical strategy-based dynamic scheduling and coordination method for power inspection models provided in an embodiment of the present invention. Figure 1The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S5 are detailed as follows: S1, Receive power inspection task and load corresponding inspection and collection strategy, collect multimodal power inspection data based on the inspection and collection strategy and perform preprocessing; Power inspection tasks refer to work instructions initiated by the power operation and maintenance system to conduct status checks and fault diagnoses on power facilities. These instructions include, but are not limited to, inspection scenarios, inspection objects, task types, task priorities, and completion deadlines. Examples include routine periodic inspections, abnormal alarm verification inspections, and special fault investigation inspections of substation equipment, transmission lines, distribution room cabinets, and wind turbine components. Different priorities and inspection times are clearly defined during the inspection process. Inspection data collection strategies refer to the equipment configuration schemes, data collection mode rules, and data processing benchmark sets selected from a pre-stored inspection strategy mapping library, based on the inspection scenario, inspection object, task type, task priority, and completion deadline of the power inspection task. For example, based on the inspection scenario and task type, the system automatically configures the combination of data collection equipment and sensor modules, plans the data collection path, adjusts the data collection frequency and accuracy in real time according to task priority, and sets sensor parameters.

[0023] In addition, preprocessing is used to eliminate noise interference in multimodal power inspection data, unify data format, enhance effective features, and improve data quality and subsequent model inference efficiency.

[0024] The process of receiving power line inspection tasks and loading corresponding inspection and data collection strategies, collecting multimodal power line inspection data based on the inspection and data collection strategies, and performing preprocessing specifically includes: According to the power inspection task, the pre-stored strategy mapping library is called, and the corresponding inspection and collection strategy is loaded through parameter matching logic. The multimodal power inspection data is collected using either fixed-point collection mode or mobile inspection and collection mode. Signal conditioning technology is used to correct data defects, identify and remove invalid data, and unify the data format based on the data type of the multimodal power inspection data. An adaptive median filtering algorithm is used to dynamically adjust the window size to denoise the power inspection image data; A contrast-limited adaptive histogram equalization algorithm is used to enhance the clarity of the power inspection image data. The multimodal power inspection data is losslessly compressed using encoding compression technology.

[0025] The process begins by parsing the received power inspection tasks, extracting information such as inspection scene tags, inspection objects, task types, and task priorities. Then, a pre-stored inspection strategy mapping library is invoked, and the corresponding inspection data collection strategy is automatically loaded through multi-dimensional parameter matching logic. If the task target is fixed facilities such as substation equipment or distribution room components, a fixed-point acquisition mode is used. Fixed cameras capture images of key areas such as the equipment's appearance, meters, and connectors at a preset frequency. Infrared detectors collect real-time surface temperature data and generate temperature distribution curves, while a voice module continuously collects ambient sound wave data. If the task target is a mobile inspection scenario such as transmission lines or wind turbines, a mobile acquisition mode is used. A drone flies along a preset cruise path, dynamically collecting multimodal power inspection data along the preset path using visual sensors, an infrared module, and a voice module.

[0026] Furthermore, signal conditioning technology, based on the data type of multimodal power inspection data, focuses on and precisely repairs quality issues arising during the acquisition process due to equipment errors, environmental interference, and transmission fluctuations, addressing different data characteristics and ensuring the integrity, numerical accuracy, and format standardization of various data types. Quantitative indicators are used to traverse the multimodal power inspection data one by one, marking invalid data and deleting it directly from the dataset, recording the reasons for removal. Through format conversion, parameter calibration, structural regularization, and metadata binding, format barriers of multimodal power inspection data are eliminated, achieving data compatibility and interoperability.

[0027] Among them, the adaptive median filtering algorithm is a denoising algorithm that dynamically adjusts the size of the filtering window based on the image noise density. It can effectively remove noise while preserving details such as device edges and cracks. It determines noise density by statistically analyzing the variance of image pixel grayscale values ​​and removes noise through weighted replacement of neighboring pixels, avoiding the excessive blurring problem caused by traditional fixed-window filtering. The contrast-limited adaptive histogram equalization algorithm enhances image brightness and detail by dividing the image into local regions and limiting the histogram contrast range. It divides the image into multiple non-overlapping local regions, equalizes the histogram of each region, and sets a contrast limit threshold to avoid excessive enhancement of local bright areas that leads to detail distortion. This amplifies the grayscale differences of device defects in low-light scenes and suppresses the brightness of overly bright areas in highly reflective scenes, making the overall image brightness and darkness levels more distinct and improving the identification of key defects. Encoding compression technology refers to techniques that reduce data volume through specific algorithms without sacrificing data accuracy. It adopts adaptive compression methods for different modalities of data, reducing data transmission and storage overhead while ensuring data integrity and distortion-free information.

[0028] The above methods enable the dynamic design of strategies for collecting multimodal power inspection data based on power inspection tasks, effectively achieving standardized closed-loop management of the entire process from data collection to preprocessing of multimodal power inspection data.

[0029] S2, Based on the multi-scale convolutional feature segmentation algorithm, extract the key features of the power inspection task from the preprocessed multimodal power inspection data; Among them, the key features of power inspection tasks refer to the set of features that can accurately reflect the operating status, potential fault hazards and operating parameters of power equipment, including but not limited to equipment structural features, numerical attribute features, defect characterization features, and effective feature information covering multimodal data such as images, videos, infrared, and voice.

[0030] The multi-scale convolutional feature segmentation algorithm extracts key features of the power inspection task from the preprocessed multimodal power inspection data, specifically including: A multi-scale parallel convolutional branch network is constructed, and convolutional kernels of different scales are used to perform layer-by-layer convolutional operations on the preprocessed multimodal power inspection data to extract multi-scale feature information in parallel. A feature pyramid network is employed to fuse the multi-scale feature information through a top-down feature transfer mechanism and lateral connections. A segmentation network with an encoder-decoder structure is used to locate and filter key feature regions in the multi-scale feature information, extract the quantized features corresponding to the key feature regions, and generate the key features of the power inspection task.

[0031] Understandably, multi-scale feature information refers to a set of features extracted using convolutional kernels of different sizes, encompassing low-level details, mid-level structure, and high-level semantics. Key feature regions refer to feature regions whose global contextual information is captured by the encoder module and precisely located by the decoder module; these regions contain crucial information for power line inspection tasks. Quantized features refer to features whose physical properties, morphological parameters, and representational information are converted into specific numerical values.

[0032] Specifically, the multi-scale parallel convolutional branch network is a deep learning network containing multiple independent convolutional branches. Each branch embeds a convolutional kernel of a corresponding scale. By computing in parallel and capturing different features simultaneously, it avoids feature omissions caused by single-scale extraction. Small-sized convolutional kernels focus on fine-grained details, while large-sized convolutional kernels capture coarse-grained global features. Subsequently, convolutional operations are performed layer by layer from shallow to deep according to the network hierarchy. Shallow networks focus on extracting basic features such as edges and textures, while deep networks focus on aggregating basic features to form higher-order semantic features. This enables progressive feature enhancement. The feature information output by each branch is temporarily stored according to scale, fully preserving both fine-grained details and coarse-grained global features.

[0033] Then, through a top-down feature propagation mechanism, starting with the coarse-grained feature map output from the deepest layer of the multi-scale parallel convolutional branch network, which contains the richest global semantic information and can accurately represent the overall state of power facilities, an upsampling operation is performed to enlarge its size, making the size of the enlarged feature map completely consistent with the feature map output from the previous level branch. This allows deep global semantic features to be propagated layer by layer downwards, endowing shallow features with semantic associations. Lateral connections, for the feature maps output from the shallow branches at each level, first adjust the channel dimensions to match the number of entity channels with the high-dimensional feature maps propagated from top to bottom, eliminating fusion barriers. Subsequently, the dimension-matched shallow feature maps are aligned element-wise with the propagated high-dimensional feature maps, and a weighted fusion method is used to integrate the two types of features. This preserves the detailed information of shallow features while incorporating the semantic information of deep features. Simultaneously, an attention mechanism is used to strengthen key feature channels and weaken irrelevant interference. The process of upsampling, dimension adjustment, and lateral fusion is repeated until the collaborative processing of features at all levels is completed, generating a fused feature set covering multi-scale feature information.

[0034] Finally, the encoder module gradually reduces the feature map size through downsampling operations, capturing global contextual information of the features and providing a global reference for key feature region localization. The decoder module gradually restores the feature map size through upsampling operations. Simultaneously, by utilizing skip connections between the encoder and decoder, the feature maps output from each layer of the encoder are directly passed to the corresponding layers of the decoder, supplementing the detailed features lost during downsampling and improving localization accuracy. In the key feature region localization and filtering stage, the segmentation network performs pixel-level semantic classification on the fused feature set, outputting different types of region segmentation results. High-confidence regions containing core information are retained through preset filtering conditions, while blurry regions and background interference are removed. Subsequently, a customized transformation method is used for key feature regions of different modalities, converting their physical properties, morphological parameters, etc., into standardized quantized features. Finally, the quantized features of all modalities are organized and classified according to a unified structure to generate standardized key features for power inspection tasks.

[0035] Through the above methods, the key features of power inspection tasks are fully captured, efficiently integrated, and accurately quantified, solving the technical problems of fragmented multimodal data feature extraction, easy loss of fine-grained features, weak feature representation ability, and difficulty in directly supporting model inference in existing technologies.

[0036] S3, using multi-dimensional task level evaluation indicators, evaluate the power inspection task level label corresponding to the power inspection task based on the key characteristics of the power inspection task. The method of using multi-dimensional task-level evaluation indicators to evaluate the power inspection task level label corresponding to the power inspection task based on the key characteristics of the power inspection task specifically includes: The key features of the power inspection task are input into the power inspection task analysis unit to calculate the key feature values ​​of the power inspection task and establish a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation index. Based on the mapping relationship, the comprehensive evaluation value of the power inspection task based on the key characteristics of the power inspection task is calculated, and the power inspection task level label is generated.

[0037] It should be noted that the multi-dimensional task-level evaluation index refers to a set of multi-dimensional evaluation standards covering the important attributes of power inspection tasks. It comprehensively quantifies the key characteristics of power inspection tasks, specifically including three evaluation indicators: task real-time performance, task complexity, and task safety level. These indicators complement each other to ensure the comprehensiveness and accuracy of the evaluation results. The power inspection task analysis unit is a processing module used to evaluate the level of power inspection tasks. The key feature values ​​of a power inspection task refer to a set of unified dimensional values ​​obtained after standardizing, integrating, and quantifying the key features of the power inspection task. These values ​​directly reflect the specific performance of the key features of the power inspection task in the multi-dimensional task-level evaluation index. The comprehensive evaluation value of a power inspection task refers to the comprehensive quantitative result of the power inspection task based on the three evaluation indicators: task real-time performance, task complexity, and task safety level. This value is used to comprehensively and objectively reflect the priority ranking and actual processing difficulty of power inspection tasks. Power inspection task level labels are classification identifiers generated based on the power inspection task level threshold classification rules according to the comprehensive evaluation value of the power inspection task. They are used to clarify the processing priority and appropriate model type of the power inspection task, and are usually divided into three levels: low, medium and high.

[0038] The key features of the power inspection task are input into the power inspection task analysis unit to calculate the key feature values ​​of the power inspection task and establish a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation indicators, specifically including: The key features of the power line inspection task are input into the power line inspection task analysis unit via the MQTT protocol. These key features are quantified and subjected to Min-Max standardization to generate key feature values. The corresponding calculation formula is as follows: In the formula, represents the key characteristic value of the power inspection task; x represents the quantitative value of the key characteristic of the power inspection task. The maximum value of the quantified key characteristics of power line inspection tasks; The minimum quantitative value representing the key characteristics of power line inspection tasks; The power inspection task analysis unit classifies the key features of the power inspection task according to the hierarchical rule knowledge base and the lightweight decision tree algorithm, and establishes a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation index, wherein the multi-dimensional task level evaluation index includes task real-time performance, task complexity, and task security level.

[0039] Among them, the MQTT protocol is a lightweight IoT communication protocol based on a publish-subscribe model, used to transmit remote data between resource-constrained devices and backend processing units. The hierarchical rule knowledge base refers to a structured database storing power inspection business rules and historical evaluation cases, containing templates showing the correspondence between key features of different types of power inspection tasks and multi-dimensional task-level evaluation indicators. The lightweight decision tree algorithm refers to a decision tree model that has been pruned, optimized, and simplified in parameters. By constructing a simple binary tree structure, it learns the complex nonlinear relationship between standardized key features of power inspection tasks and multi-dimensional task-level evaluation indicators, possessing the characteristics of high computational efficiency and low resource consumption.

[0040] Understandably, the key features of the power inspection task are input into the power inspection task analysis unit via the MQTT protocol, and the aforementioned formula is used to convert these key features into specific numerical values. The subsequent Min-Max standardization process maps quantized values ​​of different dimensions and numerical ranges to a uniform 0-1 interval, eliminating evaluation biases caused by dimensional differences. For example, the quantized value of temperature (40℃) and crack length (100mm) are standardized into key feature values ​​for the power inspection task within the 0-1 interval.

[0041] Furthermore, the power inspection task analysis unit first calls the hierarchical rule knowledge base to retrieve the evaluation index categories corresponding to the key features of the power inspection task. If there are multiple evaluation indicators corresponding to a single key feature of the power inspection task, the lightweight decision tree algorithm will combine the weight of the key feature of the power inspection task with the business rules to determine the main evaluation indicators and establish a mapping relationship, generate the corresponding mapping table, and clarify the evaluation dimension corresponding to each key feature of the power inspection task.

[0042] The above methods not only standardize and accurately classify the key characteristics of power inspection tasks, but also establish a clear mapping relationship between the key characteristics of power inspection tasks and multi-dimensional task level evaluation indicators, effectively ensuring the accuracy and reliability of the calculation of the comprehensive evaluation value of power inspection tasks and the generation of power inspection task level labels.

[0043] The preset power inspection task level threshold, based on the mapping relationship, calculates the comprehensive evaluation value of the key features of the power inspection task, and generates the power inspection task level label, specifically including: Based on the mapping relationship, the key features of the power inspection task corresponding to the real-time performance of the task are determined, the timestamp information of the key features of the power inspection task is identified, and the key feature values ​​of the power inspection task are input into the Linear-SVM regression model to calculate the task real-time performance score. ; Based on the mapping relationship, the key features of the power line inspection task corresponding to the task complexity are determined. A lightweight random forest model is invoked, and each decision tree calculates the importance weight of the key features of the power line inspection task based on its own splitting rules. Based on the key feature values ​​of the power line inspection task and the importance weights, a weighted score for task complexity is calculated. ; Based on the mapping relationship, the key features of the power inspection task corresponding to the task safety level are determined. Based on the distance-weighted K-nearest neighbor algorithm, historical similar tasks are matched in the power inspection task analysis unit, and the task safety level score is calculated by weighting similarity. The corresponding calculation formula is as follows: In the formula, K represents the number of historically similar tasks; This represents the similarity between the power line inspection task and the j-th historical similar task; This represents the security level score of the j-th historically similar task; Indicates the correction factor; The key feature value of the power inspection task is represented by the j-th historically similar task. The Euclidean distance between the key feature value of the power inspection task and the key feature value of the power inspection task of the j-th historical similar task is represented. The CRITIC objective weighting method is used to determine the weights of the task's real-time performance, complexity, and security level. The weighted average is then used to calculate the comprehensive evaluation value of the power inspection task. The corresponding calculation formula is as follows: In the formula, , , These represent the weights of task real-time performance, task complexity, and task security level, respectively. A power inspection task level threshold is preset, and the comprehensive evaluation value of the power inspection task is compared with the power inspection task level threshold. The power inspection task level label is generated by preset task level labeling rules.

[0044] Among them, the Linear-SVM regression model learns the linear relationship between key feature values ​​of power inspection tasks and task real-time performance through the support vector machine algorithm, making it suitable for evaluation scenarios with high time sensitivity requirements. The lightweight random forest model is an ensemble learning model optimized through pruning and reducing the number of decision trees. It retains the robustness of multiple decision trees in a random forest while reducing computational complexity by decreasing tree depth and the number of nodes. The splitting rules for each decision tree are designed based on the Gini coefficient or information gain ratio, avoiding the overfitting problem of a single decision tree while maintaining high computational speed. The distance-weighted K-nearest neighbor algorithm is a classification regression algorithm that measures task similarity using Euclidean distance and assigns weights according to similarity. It has strong flexibility and can directly achieve score prediction through historical data matching. The CRITIC objective weighting method is a method of calculating weights based on the dispersion of the data itself and the correlation between indicators, avoiding the bias of subjective weighting and ensuring the objectivity and scientific nature of weight allocation. The task real-time performance score is used to quantify the urgency of power inspection tasks. The task complexity score reflects the resource load required for task execution. The task safety level score is based on risk extrapolation from the safety records of similar historical tasks. The power inspection task level threshold refers to a predefined level classification benchmark based on power inspection industry safety standards, operation and maintenance resource configuration requirements, and historical task processing data, used to define the quantitative range of tasks with different priorities.

[0045] Specifically, firstly, based on the mapping relationship, the key characteristics of power inspection tasks are determined, including the maximum allowable response time limit, real-time data transmission requirements, and emergency event association identifiers. Secondly, the timestamp information of the key characteristics of power inspection tasks is identified, and an association index table between the key characteristics and timestamps is established. The time difference corresponding to the timestamps is converted into a quantified value of the key characteristics of the power inspection tasks, i.e., the key characteristic value of the power inspection tasks. The model fits the correspondence between the characteristic values ​​and the score through internal kernel function calculation, and outputs a normalized task real-time score. Among them, the higher the score, the higher the real-time requirement of the task.

[0046] Based on the mapping relationship, the key features of the power inspection task corresponding to the task complexity are determined, including the number of inspection equipment, work process steps, environmental interference factors, and equipment type complexity. A lightweight random forest model is used, which consists of 10-20 shallow decision trees. Each decision tree divides the key features of the power inspection task into layers according to its own splitting rules. The importance weight is determined by calculating the contribution of the key features of the power inspection task to the decision result. Based on the key feature values ​​of the power inspection task and their corresponding importance weights, a weighted score for task complexity is calculated. .

[0047] Based on the mapping relationship, the key characteristics of the power inspection task corresponding to the task safety level are determined, including equipment operating years, historical failure rate, safety protection level of the work area, and equipment importance level. Based on the distance-weighted K-nearest neighbor algorithm, historical similar tasks are matched in the historical task database of the power inspection task analysis unit, and the task safety level score is calculated by weighting the similarity. .

[0048] Furthermore, key characteristics of power line inspection tasks, including task real-time performance score, task complexity score, and task safety level score, are collected. The standard deviation of the three evaluation indicators is calculated to accurately represent their conflict. A correlation coefficient matrix of the three evaluation indicators is constructed, and the sum of the correlation coefficients between each indicator and other evaluation indicators is calculated to reflect the degree of information overlap. Subsequently, the information carrying capacity of each evaluation indicator is calculated by multiplying the standard deviation by (1 - the sum of correlation coefficients). Finally, the information carrying capacity of each evaluation indicator is divided by the sum of the carrying capacities of the three evaluation indicators to obtain the normalized weights of task real-time performance, task complexity, and task safety level. , , .

[0049] Understandably, the comprehensive evaluation value of the power inspection task is compared one by one with the preset threshold values ​​for each level of power inspection task to determine the corresponding threshold range. Finally, according to the preset task level label rules, when the comprehensive evaluation value of the power inspection task falls into a certain threshold range, the corresponding power inspection task level label for that range is automatically matched.

[0050] The above methods effectively achieve the scientific classification and accurate labeling of power inspection task levels, solving the problems of strong subjectivity, low efficiency, and inconsistent standards in traditional manual assessment.

[0051] S4. Select the corresponding dynamic model scheduling strategy according to the power inspection task level label, wherein the model in the dynamic model scheduling strategy includes lightweight small model and cloud large model. The dynamic model scheduling strategy refers to a system of strategies that, in response to the dynamic changes in power inspection tasks, uses intelligent scheduling algorithms to regulate the calling logic, data transmission rules, and resource allocation schemes of different types of models. This improves resource utilization efficiency and task processing time while ensuring task execution effectiveness. Lightweight small models are lightweight AI models deployed on edge devices or drones for power inspection terminals. Through architecture optimization and technology compression, they reduce computational load, memory consumption, and inference latency, making them suitable for rapid task processing in resource-constrained scenarios. Cloud-based large models are high-performance AI models deployed on near-edge cloud or central cloud nodes, suitable for power inspection tasks with high complexity and high precision requirements.

[0052] The step of selecting a corresponding dynamic model scheduling strategy based on the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models, specifically including: A lightweight convolutional neural network architecture, 8-bit quantization technology, and model distillation technology are adopted to reduce the computational load and resource consumption of model inference and construct the lightweight small model. A multimodal Transformer architecture and cross-modal attention network are adopted to combine the fusion requirements of the power inspection task and construct the cloud-based large model. Based on the intelligent scheduling algorithm, the key characteristics of the power inspection task are perceived in real time according to the power inspection task level label, and the calling order, data transmission priority and resource allocation ratio of the lightweight small model and the cloud large model are dynamically adjusted. If the power inspection task level label is medium or low, the lightweight small model is selected first; if the power inspection task level label is high, the cloud-based large model is selected directly.

[0053] Among them, the intelligent scheduling algorithm refers to the core execution module of the dynamic model scheduling strategy, which integrates three major functions: task feature perception, resource status assessment, and model adaptation decision-making. By parsing the power inspection task level label, it extracts key task features in real time, and simultaneously monitors the remaining resources on the terminal or edge side, the cloud network bandwidth status, and the load of the large model, forming a multi-dimensional decision-making basis.

[0054] Understandably, in the construction of lightweight small models, lightweight convolutional neural network architectures employ designs such as depthwise separable convolutions and bottleneck structures to reduce the number of network parameters and computational load while retaining the effective identification capability of key features in power line inspection. 8-bit quantization technology compresses model parameters from traditional 32-bit floating-point numbers to 8-bit integers, significantly reducing parameter storage and memory access overhead during inference, enabling the model to run quickly on terminal devices with limited computing power. Model distillation technology uses mature, complex models as a benchmark, guiding the training of lightweight models through knowledge transfer.

[0055] Furthermore, during the construction of the large-scale cloud model, a multimodal Transformer architecture is adopted to ensure parallel input and unified encoding of power inspection data such as images, infrared, and sound, capturing deep correlations between different modalities of data, such as the correspondence between abnormal equipment temperature and infrared image features. A cross-modal attention network dynamically allocates attention weights to focus on key feature values ​​of the power inspection task, suppresses redundant interference, and accurately matches the needs of collaborative analysis of multimodal power inspection data.

[0056] Furthermore, the matching selection of power inspection task level labels and model types is based on a precise match between task requirements and model capabilities, avoiding the drawbacks of traditional fixed model calling methods. Lightweight small models reduce the load on cloud computing power and network transmission, improving the efficiency of large-scale execution of daily inspection work; cloud-based large models break through the performance limitations of terminal devices, solving the problems of resource waste, response lag, and insufficient accuracy.

[0057] S5, input the key features of the power inspection task into the model in the dynamic model scheduling strategy, output the power inspection identification results and push them to the inspection monitoring platform.

[0058] The power inspection identification result refers to the final conclusion output after model reasoning and multi-dimensional verification, including the status determination of the power inspection object, the subdivision of anomaly types, confidence level, the matching basis of key features of the power inspection task, and the alarm level. The inspection monitoring platform has functions such as real-time result reception, visualization, hierarchical early warning triggering, historical data storage, and handling process tracking, and supports simultaneous access from multiple terminals such as computers and mobile devices.

[0059] The process of inputting the key features of the power inspection task into the model in the dynamic model scheduling strategy, outputting the power inspection identification results, and pushing them to the inspection monitoring platform specifically includes: If the power inspection task level label is medium or low, then the corresponding key features of the power inspection task are read and input into the lightweight small model; if the power inspection task level label is high, then the corresponding key features of the power inspection task are read and input into the cloud-based large model. A confidence threshold is preset. If the confidence level output by the lightweight small model is lower than the confidence threshold, the cloud-based large model is automatically triggered to perform supplementary analysis. If the lightweight small model detects that the key feature of the power inspection task is an abnormal type of power inspection task, the lightweight small model and the cloud-based large model are automatically started for parallel inference. The parallel inference results of the lightweight small model and the cloud-based large model are comprehensively evaluated, and the power inspection identification results are output and pushed to the inspection and monitoring platform.

[0060] The confidence threshold is a pre-set critical standard used to quantify the reliability of the output results of the lightweight small model. Its value range is [0,1], and it is used to determine whether to trigger the large cloud model to participate in the analysis.

[0061] Specifically, for intermediate or low-level power inspection task level labels, the corresponding key features of the power inspection task can be quickly read through the interface and directly input into a lightweight small model. For high-level power inspection task level labels, multi-modal and high-dimensional key features of the power inspection task can be read.

[0062] Furthermore, the lightweight small model leverages its low latency to achieve rapid inference, automatically comparing the output confidence level with a confidence threshold. If the confidence level is lower than the threshold, it directly sends a supplementary analysis request to the large model in the cloud, and synchronizes the intermediate inference results with the acquired key features of the power inspection task. The large model in the cloud calls data format standardization algorithms and feature mapping alignment algorithms to convert the key features of the power inspection task into a format adapted to its own architecture, aligns the feature mapping matrix of the intermediate inference results with the feature dimensions of the large model in the cloud, and stores the intermediate inference results in a temporary cache in the cloud after feature dimension alignment is completed.

[0063] Furthermore, when the lightweight small model detects that the key features of the power inspection task are of an abnormal type, it completes the feature database comparison through the abnormal type matching algorithm, and then calls the collaborative scheduling algorithm to synchronously send inference start instructions to the lightweight small model and the cloud large model. It synchronizes the key features of the power inspection task and starts resource allocation. Intermediate results such as abnormal area coordinates and feature matching progress are written to the synchronization pool in real time. The lightweight small model can read the feature matching results from the cloud to optimize local accuracy, and the cloud large model can use the abnormal area coordinates of the lightweight small model to focus on the analysis of key points and avoid invalid calculations.

[0064] The process of comprehensively evaluating the parallel inference results of the lightweight small model and the cloud-based large model, outputting the power inspection identification results, and pushing them to the inspection and monitoring platform specifically includes: The confidence scores of the lightweight small model and the cloud large model are normalized and weighted using a feature weighted fusion algorithm. If the confidence score is lower than the confidence score threshold, the key features of the power inspection task corresponding to the confidence score are read, and the key features of the power inspection task are updated probabilistically through a Bayesian confidence correction mechanism to output the power inspection identification result. The power inspection and identification results are pushed to the inspection and monitoring platform through an encrypted transmission channel, and the power inspection and identification results are visualized and stored in a structured manner.

[0065] The process involves using a feature-weighted fusion algorithm to determine dynamic weights, quantifying the core capability differences between the lightweight small model and the cloud-based large model using a model capability assessment algorithm, and combining historical inference data and the analytic hierarchy process (AHP) to determine the weights of the lightweight small model and the cloud-based large model. The output confidence scores of the lightweight small model and the cloud-based large model are extracted, and a normalized weighting formula is used to transform the independent inference results of the lightweight small model and the cloud-based large model into a unified confidence evaluation index. If the normalized weighted confidence score is lower than the confidence score threshold, a feature selection algorithm is used to select the corresponding key features of the power inspection task. First, a prior probability distribution of the key features of the power inspection task is constructed; second, the likelihood probability is calculated using a feature matching algorithm; finally, the prior probability and the likelihood probability are fused according to Bayes' theorem to obtain the corrected confidence score. If the corrected confidence score is still lower than the confidence score threshold, the corresponding key features of the power inspection task are marked as requiring further review.

[0066] It should be noted that the revised confidence level needs to be bound and correlated with the key features of the power inspection task and the power inspection task level label, and analyzed collaboratively to jointly determine whether the power inspection task is normal, abnormal or alarm. The power inspection identification results are pushed to the inspection monitoring platform in real time through an encrypted transmission channel, and the power inspection identification results are visualized and stored in a structured manner, including equipment status map, model call log, alarm level and confidence level, etc. When the inspection monitoring platform detects potential risks, it automatically pushes alarm information and generates an inspection report.

[0067] The above methods effectively solve the problems of traditional single-model output results lacking unified measurement standards, poor traceability, and low computational efficiency, and realize the intelligent and standardized output of power inspection and identification results from data verification to terminal applications.

[0068] like Figure 2 The diagram shown is a system block diagram of a hierarchical strategy power inspection model dynamic scheduling and coordination system provided in an embodiment of the present invention. The system includes: The data acquisition and preprocessing module is used to receive power inspection tasks and load the corresponding inspection acquisition strategy, and to collect multimodal power inspection data based on the inspection acquisition strategy and perform preprocessing. The key feature extraction module is used to extract key features of the power inspection task from the preprocessed multimodal power inspection data based on a multi-scale convolutional feature segmentation algorithm. The grade label evaluation module is used to evaluate the power inspection task grade label corresponding to the power inspection task based on the key characteristics of the power inspection task using multi-dimensional task level evaluation indicators. The inspection model scheduling module is used to select the corresponding dynamic model scheduling strategy according to the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models. The inspection result output module is used to input the key features of the power inspection task into the model in the dynamic model scheduling strategy, output the power inspection identification results, and push them to the inspection monitoring platform.

[0069] Figure 2 The system of the illustrated embodiment can be used to perform corresponding operations. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0070] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the steps of a hierarchical strategy power inspection model dynamic scheduling coordination method as described in any of the above.

[0071] like Figure 3 The diagram shown is a hardware structure schematic of an electronic device according to an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32, and a computer program; wherein... The memory 32 is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0072] Processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0073] Alternatively, the memory 32 can be either standalone or integrated with the processor 31.

[0074] When the memory 32 is a device independent of the processor 31, the device may further include: Bus 33 is used to connect the memory 32 and the processor 31.

[0075] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of a hierarchical strategy power inspection model dynamic scheduling coordination method as described above.

[0076] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0077] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0078] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0079] Through the above embodiments, the present invention can receive power inspection tasks and load corresponding inspection and acquisition strategies, collect multimodal power inspection data based on the inspection and acquisition strategies and perform preprocessing; extract key features of power inspection tasks from the preprocessed multimodal power inspection data based on a multi-scale convolutional feature segmentation algorithm; evaluate the power inspection task level label corresponding to the power inspection task based on the key features of the power inspection task using multi-dimensional task level evaluation indicators; select the corresponding dynamic model scheduling strategy based on the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models; input the key features of the power inspection task into the models in the dynamic model scheduling strategy, output the power inspection identification results and push them to the inspection and monitoring platform, thereby enabling the use of dynamic model scheduling strategies to achieve dynamic allocation and efficient collaborative model resources, improving the real-time performance, energy efficiency and intelligence level of the power inspection system.

[0080] This invention utilizes fixed-point or mobile inspection acquisition modes and multimodal data preprocessing technology to simultaneously collect and optimize multimodal power inspection data, ensuring data integrity and high quality in complex inspection scenarios. Through a multi-scale convolutional feature segmentation algorithm, it accurately locates key feature regions and extracts and fuses key features of power inspection tasks in parallel, ensuring the integrity and accuracy of key features and solving the problems of missed detections and false detections in traditional single-scale feature extraction. Furthermore, this invention employs multi-dimensional task-level evaluation indicators to construct an evaluation system for task real-time performance, task complexity, and task safety level, achieving scientific classification and accurate label generation for power inspection tasks, bypassing traditional manual subjective classification or single-indicator classification. Limitations: This invention dynamically selects lightweight small models and cloud-based large models based on the power inspection task level labels, and can dynamically allocate models according to the type of power inspection task. This invention inputs the key features of the power inspection task into the model in the dynamic model scheduling strategy, outputs the power inspection identification results and pushes them to the inspection monitoring platform. It adopts a pre-set confidence threshold to trigger a supplementary inference mechanism, combined with feature weighted fusion algorithm and Bayesian confidence correction technology, to ensure dynamic allocation and efficient co-model resources, realize the visualization and structured storage of power inspection identification results, effectively solve the problems of wasted computing resources for simple tasks and insufficient identification rate for complex tasks, simplify the circuit inspection calculation link, and improve the accuracy of power inspection identification.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hierarchical strategy-based dynamic scheduling and coordination method for power inspection models, characterized in that, The method includes: Receive power inspection tasks and load corresponding inspection and data collection strategies, collect multimodal power inspection data based on the inspection and data collection strategies and perform preprocessing; Based on the multi-scale convolutional feature segmentation algorithm, key features of power inspection tasks are extracted from the preprocessed multimodal power inspection data; Using multi-dimensional task level evaluation indicators, the power inspection task level label corresponding to the power inspection task is evaluated based on the key characteristics of the power inspection task. Select the corresponding dynamic model scheduling strategy according to the power inspection task level label, wherein the model in the dynamic model scheduling strategy includes lightweight small model and cloud large model; The key features of the power inspection task are input into the model in the dynamic model scheduling strategy, and the power inspection identification results are output and pushed to the inspection monitoring platform.

2. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 1, characterized in that, The process of receiving power line inspection tasks and loading corresponding inspection and data collection strategies, collecting multimodal power line inspection data based on the inspection and data collection strategies, and performing preprocessing specifically includes: According to the power inspection task, the pre-stored strategy mapping library is called, and the corresponding inspection and collection strategy is loaded through parameter matching logic. The multimodal power inspection data is collected using either fixed-point collection mode or mobile inspection and collection mode. Signal conditioning technology is used to correct data defects, identify and remove invalid data, and unify the data format based on the data type of the multimodal power inspection data. An adaptive median filtering algorithm is used to dynamically adjust the window size to denoise the power inspection image data; A contrast-limited adaptive histogram equalization algorithm is used to enhance the clarity of the power inspection image data. The multimodal power inspection data is losslessly compressed using encoding compression technology.

3. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 1, characterized in that, The multi-scale convolutional feature segmentation algorithm extracts key features of the power inspection task from the preprocessed multimodal power inspection data, specifically including: A multi-scale parallel convolutional branch network is constructed, and convolutional kernels of different scales are used to perform layer-by-layer convolutional operations on the preprocessed multimodal power inspection data to extract multi-scale feature information in parallel. A feature pyramid network is employed to fuse the multi-scale feature information through a top-down feature transfer mechanism and lateral connections. A segmentation network with an encoder-decoder structure is used to locate and filter key feature regions in the multi-scale feature information, extract the quantized features corresponding to the key feature regions, and generate the key features of the power inspection task.

4. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 1, characterized in that, The method of using multi-dimensional task-level evaluation indicators to evaluate the power inspection task level label corresponding to the power inspection task based on the key characteristics of the power inspection task specifically includes: The key features of the power inspection task are input into the power inspection task analysis unit to calculate the key feature values ​​of the power inspection task and establish a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation index. Based on the mapping relationship, the comprehensive evaluation value of the power inspection task based on the key characteristics of the power inspection task is calculated, and the power inspection task level label is generated.

5. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 4, characterized in that, The key features of the power inspection task are input into the power inspection task analysis unit to calculate the key feature values ​​of the power inspection task and establish a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation indicators, specifically including: The key features of the power line inspection task are input into the power line inspection task analysis unit via the MQTT protocol. These key features are quantified and subjected to Min-Max standardization to generate key feature values. The corresponding calculation formula is as follows: In the formula, represents the key characteristic value of the power inspection task; x represents the quantitative value of the key characteristic of the power inspection task. The maximum value of the quantified key characteristics of power line inspection tasks; The minimum quantitative value representing the key characteristics of power line inspection tasks; The power inspection task analysis unit classifies the key features of the power inspection task according to the hierarchical rule knowledge base and the lightweight decision tree algorithm, and establishes a mapping relationship between the key features of the power inspection task and the multi-dimensional task level evaluation index, wherein the multi-dimensional task level evaluation index includes task real-time performance, task complexity, and task security level.

6. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 5, characterized in that, The preset power inspection task level threshold, based on the mapping relationship, calculates the comprehensive evaluation value of the key features of the power inspection task, and generates the power inspection task level label, specifically including: Based on the mapping relationship, the key features of the power inspection task corresponding to the real-time performance of the task are determined, the timestamp information of the key features of the power inspection task is identified, and the key feature values ​​of the power inspection task are input into the Linear-SVM regression model to calculate the task real-time performance score. ; Based on the mapping relationship, the key features of the power line inspection task corresponding to the task complexity are determined. A lightweight random forest model is invoked, and each decision tree calculates the importance weight of the key features of the power line inspection task based on its own splitting rules. Based on the key feature values ​​of the power line inspection task and the importance weights, a weighted score for task complexity is calculated. ; Based on the mapping relationship, the key features of the power inspection task corresponding to the task safety level are determined. Based on the distance-weighted K-nearest neighbor algorithm, historical similar tasks are matched in the power inspection task analysis unit, and the task safety level score is calculated by weighting similarity. The corresponding calculation formula is as follows: In the formula, K represents the number of historically similar tasks; This represents the similarity between the power line inspection task and the j-th historical similar task; This represents the security level score of the j-th historically similar task; Indicates the correction factor; The key feature value of the power inspection task is represented by the j-th historically similar task. The Euclidean distance between the key feature value of the power inspection task and the key feature value of the power inspection task of the j-th historical similar task is represented. The CRITIC objective weighting method is used to determine the weights of the task's real-time performance, complexity, and security level. The weighted average is then used to calculate the comprehensive evaluation value of the power inspection task. The corresponding calculation formula is as follows: In the formula, , , These represent the weights of task real-time performance, task complexity, and task security level, respectively. A power inspection task level threshold is preset, and the comprehensive evaluation value of the power inspection task is compared with the power inspection task level threshold. The power inspection task level label is generated by preset task level labeling rules.

7. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 1, characterized in that, The step of selecting a corresponding dynamic model scheduling strategy based on the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models, specifically including: A lightweight convolutional neural network architecture, 8-bit quantization technology, and model distillation technology are adopted to reduce the computational load and resource consumption of model inference and construct the lightweight small model. A multimodal Transformer architecture and cross-modal attention network are adopted to combine the fusion requirements of the power inspection task and construct the cloud-based large model. Based on the intelligent scheduling algorithm, the key characteristics of the power inspection task are perceived in real time according to the power inspection task level label, and the calling order, data transmission priority and resource allocation ratio of the lightweight small model and the cloud large model are dynamically adjusted. If the power inspection task level label is medium or low, the lightweight small model is selected first; if the power inspection task level label is high, the cloud-based large model is selected directly.

8. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 1, characterized in that, The process of inputting the key features of the power inspection task into the model in the dynamic model scheduling strategy, outputting the power inspection identification results, and pushing them to the inspection monitoring platform specifically includes: If the power inspection task level label is medium or low, then the corresponding key features of the power inspection task are read and input into the lightweight small model; if the power inspection task level label is high, then the corresponding key features of the power inspection task are read and input into the cloud-based large model. A confidence threshold is preset. If the confidence level output by the lightweight small model is lower than the confidence threshold, the cloud-based large model is automatically triggered to perform supplementary analysis. If the lightweight small model detects that the key feature of the power inspection task is an abnormal type of power inspection task, the lightweight small model and the cloud-based large model are automatically started for parallel inference. The parallel inference results of the lightweight small model and the cloud-based large model are comprehensively evaluated, and the power inspection identification results are output and pushed to the inspection and monitoring platform.

9. The hierarchical strategy-based dynamic scheduling and coordination method for power inspection models according to claim 8, characterized in that, The process of comprehensively evaluating the parallel inference results of the lightweight small model and the cloud-based large model, outputting the power inspection identification results, and pushing them to the inspection and monitoring platform specifically includes: The confidence scores of the lightweight small model and the cloud large model are normalized and weighted using a feature weighted fusion algorithm. If the confidence score is lower than the confidence score threshold, the key features of the power inspection task corresponding to the confidence score are read, and the key features of the power inspection task are updated probabilistically through a Bayesian confidence correction mechanism to output the power inspection identification result. The power inspection and identification results are pushed to the inspection and monitoring platform through an encrypted transmission channel, and the power inspection and identification results are visualized and stored in a structured manner.

10. A hierarchical strategy-based dynamic scheduling and coordination system for power line inspection models, applied to the method described in any one of claims 1-9, characterized in that, The system includes: The data acquisition and preprocessing module is used to receive power inspection tasks and load the corresponding inspection acquisition strategy, and to collect multimodal power inspection data based on the inspection acquisition strategy and perform preprocessing. The key feature extraction module is used to extract key features of the power inspection task from the preprocessed multimodal power inspection data based on a multi-scale convolutional feature segmentation algorithm. The grade label evaluation module is used to evaluate the power inspection task grade label corresponding to the power inspection task based on the key characteristics of the power inspection task using multi-dimensional task level evaluation indicators. The inspection model scheduling module is used to select the corresponding dynamic model scheduling strategy according to the power inspection task level label, wherein the models in the dynamic model scheduling strategy include lightweight small models and cloud-based large models. The inspection result output module is used to input the key features of the power inspection task into the model in the dynamic model scheduling strategy, output the power inspection identification results, and push them to the inspection monitoring platform.

Images