CN-122023824-A - Inspection management and control data analysis system and method based on edge calculation

CN122023824ACN 122023824 ACN122023824 ACN 122023824ACN-122023824-A

Abstract

The invention provides a system and a method for analyzing inspection control data based on edge calculation, and relates to the technical field of inspection control data analysis, wherein the system comprises a sensing layer, a data processing layer and a data processing layer, wherein the sensing layer is used for capturing effective inspection data through an acquisition terminal and uploading the effective inspection data to the edge layer; the system comprises a server layer, an edge layer, a data management module and a data management module, wherein the server layer is used for receiving segmentation point parameters and reasoning instructions of the server layer through an edge terminal, calling a lightweight model to execute shallow reasoning to generate an intermediate feature map, collecting calculation load and network bandwidth, reporting the calculation load and the network bandwidth to the server layer, the server layer is used for receiving data through a collaborative reasoning acceleration module, determining deep reasoning logic, screening abnormal data by combining an intelligent analysis module and pushing the abnormal data to the data management module, and the data management module constructs a marked inspection data set and pushing the marked inspection data set to the model management module to complete model iterative training, so that calculation bottleneck is broken through segmentation point collaborative reasoning, resource utilization rate is improved, abnormal recognition accuracy is optimized through model iterative optimization, and intelligence and reliability of inspection management are improved.

Inventors

LU YONGQING
WEI BENQING
Gong Lingxia
Guo Yunya
WANG YAN
WANG CHENGJUN

Assignees

上海华电奉贤热电有限公司

Dates

Publication Date: 20260512
Application Date: 20260121

Claims (10)

1. An inspection management and control data analysis system based on edge calculation, the system comprising: the sensing layer is composed of a multi-source data acquisition terminal and is used for capturing real-time inspection data through the multi-source data acquisition terminal according to the inspection scene requirement, and uploading effective inspection data to the edge layer after invalid inspection data are removed; The edge layer is formed by an edge terminal, is provided with a lightweight model and is used for receiving unified segmentation point parameters and standardized part reasoning instructions issued by the server layer through the edge terminal, calling the deployed lightweight model to execute shallow layer reasoning on the effective patrol data and generate a structured intermediate feature map, and simultaneously collecting real-time computing load of the edge layer and network bandwidth state of the server layer, and reporting the structured intermediate feature map, the real-time computing load and the network bandwidth state to the server layer together; The server layer is composed of a high-computation-power server and is provided with a patrol management and control platform, and the server layer comprises a collaborative reasoning acceleration module, a computation power scheduling and equipment management module, an intelligent analysis module, a data management module and a model management module, and is used for receiving the structured middle feature map, the real-time computation power load and the network bandwidth state reported by the edge layer through the collaborative reasoning acceleration module, determining deep reasoning logic, combining the intelligent analysis module to identify preliminary abnormal patrol data and generate a standardized patrol analysis report, manually rechecking the standardized patrol analysis report, screening out final abnormal patrol data and pushing the final abnormal patrol data to the data management module, storing the effective patrol data uploaded by the perception layer and the final abnormal patrol data pushed by the intelligent analysis module, pushing the marked patrol data set to the model management module to complete model iterative training, and then deploying the iterative shallow model slice to the edge layer and deploying the iterative deep model slice to the server layer.
2. The inspection management and control data analysis system based on edge calculation according to claim 1, wherein the sensing layer captures real-time inspection data through the multi-source data acquisition terminal according to inspection scene requirements, and uploads effective inspection data to an edge layer after invalid inspection data are removed, the system specifically comprises: determining the requirements of the inspection scene according to the topography of the inspection area and the type of the inspection target, determining the installation position of a camera, the cruising path and the landing point of the unmanned aerial vehicle based on the requirements of the inspection scene, and completing the deployment of the multi-source data acquisition terminal; Capturing the real-time inspection data according to a preset acquisition time interval through the multi-source data acquisition terminal, and synchronously recording acquisition time and longitude and latitude coordinates corresponding to the real-time inspection data; Comparing the definition of the real-time inspection data, the duty ratio of the inspection target with a preset definition threshold and a preset duty ratio threshold through a data filtering unit arranged in the multi-source data acquisition terminal, removing the invalid inspection data, namely the unqualified fuzzy inspection data and the empty frame inspection data without the inspection target, obtaining the filtered effective inspection data, carrying out association packaging on the effective inspection data and the corresponding acquisition time and longitude and latitude information, and uploading the effective inspection data to an edge layer in a corresponding inspection area through a local area network.
3. The inspection management and control data analysis system based on edge calculation according to claim 1, wherein the edge layer receives unified segmentation point parameters and standardized part reasoning instructions issued by a server layer through an edge terminal, invokes a deployed lightweight model to perform shallow layer reasoning on the effective inspection data and generate a structured middle feature map, and simultaneously collects real-time computing load of the edge layer and network bandwidth state of the server layer, and reports the structured middle feature map, the real-time computing load and the network bandwidth state to the server layer together, and the system specifically comprises: Receiving the unified dividing point parameters and the standardized part reasoning instructions issued by a server layer through a wireless communication module built in the edge terminal, and analyzing the reasoning task requirements in the standardized part reasoning instructions; Based on the unified segmentation point parameters and in combination with the reasoning task requirements, determining an operation range of shallow reasoning in the lightweight model, and calling a reasoning module corresponding to the lightweight model; Inputting the effective inspection data uploaded by the perception layer into a lightweight model, performing shallow layer reasoning according to the operation range through a reasoning module, extracting core inspection characteristics in the effective inspection data and generating the structured intermediate characteristic diagram; And acquiring the real-time computing load of the edge layer and the network bandwidth state of the edge layer and the server layer through a state monitoring unit built in the edge terminal, packaging the structured middle feature map, the real-time computing load and the network bandwidth state into a reporting data packet, and sending the reporting data packet to the server layer.
4. The inspection management and data analysis system based on edge calculation of claim 1, wherein the shallow reasoning process of the lightweight model is as follows: Preprocessing the received effective inspection data through an edge terminal, wherein the preprocessing comprises the steps of carrying out size normalization on an effective inspection image and carrying out frame sampling on an effective inspection video frame until the input size of a lightweight model is adapted; based on the unified segmentation point parameters issued by the server layer, activating a shallow layer reasoning unit in the lightweight model, and closing a deep layer reasoning unit; Inputting the preprocessed effective inspection data into a shallow layer reasoning unit, and performing feature extraction operation layer by layer, namely extracting basic edge features through a convolution layer, and compressing feature dimensions of the basic edge features through a pooling layer to generate the structured intermediate feature map; and carrying out integrity check on the generated structured intermediate feature map, and associating the structured intermediate feature map with real-time running state data of the edge terminal after confirming no feature loss.
5. The inspection management and control data analysis system based on edge calculation according to claim 1, wherein the collaborative reasoning acceleration module receives the structured middle feature map, the real-time computing load and the network bandwidth state reported by an edge layer, determines a deep reasoning logic, identifies preliminary abnormal inspection data by combining with an intelligent analysis module and generates a standardized inspection analysis report, performs manual review on the standardized inspection analysis report, screens out final abnormal inspection data and pushes the final abnormal inspection data to a data management module, and specifically comprises: Receiving the structured intermediate feature diagram, the real-time computing load and the network bandwidth state reported by an edge layer through a collaborative reasoning acceleration module, and checking the integrity of the structured intermediate feature diagram and the validity of the real-time computing load and the network bandwidth state; determining operation logic and parameters of deep reasoning by combining the current inspection scene through a collaborative reasoning acceleration module; inputting the verified structured intermediate feature map into a deep inference model through an intelligent analysis module, performing feature fusion and abnormality judgment operation, identifying the preliminary abnormal inspection data and generating the standardized inspection analysis report; And displaying the standardized inspection analysis report through a man-machine interaction interface of the server layer, receiving a manual review result, screening out the final abnormal inspection data, including a final abnormal inspection target, a final abnormal inspection position, a final abnormal inspection type and a final abnormal inspection confidence, and pushing the final abnormal inspection target, the final abnormal inspection position, the final abnormal inspection type and the final abnormal inspection confidence to a data management module.
6. The system for edge-based computing inspection management and data analysis of claim 1, wherein the deep inference logic is determined as follows: Summarizing the real-time computing power load and the network bandwidth state reported by all edge layers through a collaborative reasoning acceleration module, and evaluating the supply and demand conditions of global reasoning resources; Dividing a task sequence of deep reasoning according to the priority of the current inspection task; aiming at the high-priority task, matching the structured middle feature diagram reported by the edge terminal with sufficient computing power and stable bandwidth, and preferentially distributing the deep reasoning resources of the server layer; and determining the operation depth of deep reasoning by combining the feature dimension of the structured intermediate feature map and the preliminary abnormal inspection data, and generating the corresponding deep reasoning logic.
7. The inspection management and control data analysis system based on edge calculation according to claim 1, wherein the determining process of the operation depth of the deep inference is as follows: Acquiring feature dimension information of the structured intermediate feature map and the preliminary abnormal inspection data, performing validity verification, and outputting the verified effective feature dimension information and the effective preliminary abnormal inspection data; acquiring feature complexity in the effective feature dimension information, and determining the initial operation depth of deep reasoning by combining a preset feature complexity-operation level mapping relation; Obtaining preliminary abnormal inspection confidence coefficient in the effective preliminary abnormal inspection data, if the preliminary abnormal inspection confidence coefficient is smaller than a confidence coefficient adjustment threshold value, adding a preset adjustment depth to the initial operation depth to obtain deep inference operation depth, and if the preliminary abnormal inspection confidence coefficient is larger than or equal to the confidence coefficient adjustment threshold value, obtaining the deep inference operation depth as the initial operation depth.
8. The inspection management and control data analysis system based on edge calculation according to claim 1, wherein pushing the marked inspection data set to a model management module to complete model iterative training, deploying an iterated shallow model slice to an edge layer and deploying an iterated deep model slice to a server layer, specifically comprises: receiving the marked inspection data set pushed by the data management module through a model management module, and determining a model iteration training target based on the current unified segmentation point parameter; Invoking a model training framework, performing iterative training on the model by using the marked inspection data set, namely training model parameters by using the marked inspection training data set, adjusting model super parameters by using the marked inspection verification data set, and recording the trained model version information; splitting the iterated model into a shallow model slice adapting to the edge layer and a deep model slice adapting to the server layer according to the unified segmentation point parameters; And issuing the iterated shallow model slices to all edge layers through a remote deployment tool, and issuing the iterated deep model slices to a server.
9. The inspection management and control data analysis system based on edge calculation according to claim 1, further comprising a computing power scheduling and equipment management module of the server layer receiving the real-time computing power load and the network bandwidth state reported by an edge terminal, and displaying the online state and hardware parameters of the edge terminal through a visual interface; based on the real-time computing load of the edge terminal and the network bandwidth state, the processing efficiency of the edge terminal for executing shallow layer reasoning is evaluated, and three load levels of high, medium and low are divided; according to the priority of the inspection task, the high-priority inspection task is distributed to the edge terminal with low load and high bandwidth, and the middle-low-priority inspection task is distributed to the edge terminal with middle load; If the response of the edge terminal is monitored to be overtime, judging that the edge terminal is in a fault state, and migrating the incomplete shallow layer reasoning task of the edge terminal to a standby edge terminal with standard load and bandwidth adaptation in the area.
10. An inspection management and control data analysis method based on edge calculation, which is applied to an inspection management and control data analysis system based on edge calculation as claimed in any one of claims 1 to 9, and is characterized in that the method comprises the following steps: Capturing real-time inspection data through the multi-source data acquisition terminal of the sensing layer according to the inspection scene requirement, and uploading effective inspection data to the edge layer after invalid inspection data are removed; Receiving unified segmentation point parameters and standardized part reasoning instructions issued by a server layer through an edge terminal of the edge layer, calling a deployed lightweight model to execute shallow layer reasoning on the effective inspection data and generate a structured middle feature map, collecting real-time computing load of the edge layer and a network bandwidth state of the server layer, and reporting the structured middle feature map, the real-time computing load and the network bandwidth state to the server layer together; The method comprises the steps of receiving a structured middle feature diagram, the real-time computing power load and the network bandwidth state reported by an edge layer through a collaborative reasoning acceleration module of a server layer, determining a deep reasoning logic, combining an intelligent analysis module to identify preliminary abnormal patrol data and generate a standardized patrol analysis report, manually rechecking the standardized patrol analysis report, screening out final abnormal patrol data and pushing the final abnormal patrol data to a data management module, storing the effective patrol data uploaded by a perception layer and the final abnormal patrol data pushed by the intelligent analysis module by the data management module, constructing a labeling patrol data set after preprocessing, pushing the labeling patrol data set to a model management module to complete model iterative training, and deploying an iterated shallow model slice to the edge layer and a iterated deep model slice to the server layer, wherein the server layer comprises the collaborative reasoning acceleration module, the computing power scheduling and equipment management module, the intelligent analysis module, the data management module and the model management module.

Description

Inspection management and control data analysis system and method based on edge calculation Technical Field The invention relates to the technical field of inspection management and control data analysis, in particular to an edge calculation-based inspection management and control data analysis system and method. Background The traditional inspection management and control system is in a single-terminal mode, a sensing layer collects field images and video data through equipment such as cameras, unmanned aerial vehicles and the like, and an inference server bears an inference task. The mode system has the advantages of simple architecture, low initial software and hardware deployment and debugging cost, short period and simple management flow. However, the system has certain limitation, the calculation power and storage resources of a single reasoning server have physical upper limits, and when the patrol range is enlarged and the number of acquisition devices is increased, the calculation power overload and storage overflow of the server can be caused, the reasoning task queues up and is blocked, and the patrol scene requirement can not be met. Meanwhile, the mode has the cost problem, and in order to cope with the calculation bottleneck, a single server with higher configuration needs to be purchased continuously, so that the hardware purchasing cost is increased exponentially. The system in the mode completely depends on a single reasoning server, when the hardware of the server fails, software crashes or the network is interrupted, the whole system patrol reasoning task is directly stopped, the data uploaded by all acquisition equipment cannot be processed, the key patrol data is easy to lose, and high availability cannot be realized. Therefore, it is necessary to provide a system and a method for analyzing inspection management and control data based on edge calculation to solve the above technical problems. Disclosure of Invention In order to solve the technical problems, the invention provides a patrol management and control data analysis system and method based on edge calculation, which are used for solving the problems that an edge terminal and a server of the existing system are difficult to cooperatively infer, the utilization rate of computing power resources of the edge terminal is low, the fault coping capability is weak, and model iteration lacks pertinence and scene adaptation capability is weak. The invention provides an inspection management and control data analysis system based on edge calculation, which comprises: the sensing layer is composed of a multi-source data acquisition terminal and is used for capturing real-time inspection data through the multi-source data acquisition terminal according to the inspection scene requirement, and uploading effective inspection data to the edge layer after invalid inspection data are removed; The edge layer is formed by an edge terminal, is provided with a lightweight model and is used for receiving unified segmentation point parameters and standardized part reasoning instructions issued by the server layer through the edge terminal, calling the deployed lightweight model to execute shallow layer reasoning on the effective patrol data and generate a structured intermediate feature map, and simultaneously collecting real-time computing load of the edge layer and network bandwidth state of the server layer, and reporting the structured intermediate feature map, the real-time computing load and the network bandwidth state to the server layer together; The server layer is composed of a high-computation-power server and is provided with a patrol management and control platform, and the server layer comprises a collaborative reasoning acceleration module, a computation power scheduling and equipment management module, an intelligent analysis module, a data management module and a model management module, and is used for receiving the structured middle feature map, the real-time computation power load and the network bandwidth state reported by the edge layer through the collaborative reasoning acceleration module, determining deep reasoning logic, combining the intelligent analysis module to identify preliminary abnormal patrol data and generate a standardized patrol analysis report, manually rechecking the standardized patrol analysis report, screening out final abnormal patrol data and pushing the final abnormal patrol data to the data management module, storing the effective patrol data uploaded by the perception layer and the final abnormal patrol data pushed by the intelligent analysis module, pushing the marked patrol data set to the model management module to complete model iterative training, and then deploying the iterative shallow model slice to the edge layer and deploying the iterative deep model slice to the server layer. Preferably, the sensing layer captures real-time inspection data through the multi-source data acquisition terminal ac