CN-121998892-A - Multi-mode feedback iteration-based intelligent power equipment identification and operation management method and system

CN121998892ACN 121998892 ACN121998892 ACN 121998892ACN-121998892-A

Abstract

The application relates to the technical field of power equipment management, in particular to an intelligent power equipment identification and operation and maintenance management method and system based on multi-mode feedback iteration, wherein the method comprises the steps of collecting material codes, text data and multi-angle standard image data of power equipment, and establishing an equipment standard map library; the method comprises the steps of generating an AI model for equipment component identification and comparison based on the standard map library, constructing an AI comparison model library, collecting real-time images of equipment to be identified through a mobile terminal when materials go out of the warehouse or inventory is operated, calling the AI model in the AI comparison model library to identify the real-time images, comparing the identified equipment component with standard information in the standard map library, outputting a comparison result, checking the comparison result, and carrying out real-time warning if the identified component is inconsistent with the standard information or has component missing. The intelligent upgrading of the identification and operation and maintenance of the power equipment is realized, and the operation and maintenance cost is reduced.

Inventors

ZHANG JUN
WANG SHENG
XU XINYUAN
SUN HONGZHI
LI JIANGHUA
CAO GANG
MA YAN
DONG WENJIE
WANG XIAOMIN
Kuang Xuelian

Assignees

国网山东省电力公司物资公司
山东鲁软数字科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251201

Claims (10)

1. The intelligent power equipment identification and operation and maintenance management method based on multi-mode feedback iteration is characterized by being applied to a material warehouse-in and warehouse-out or inventory operation site and comprising the following steps of: S1, collecting material codes, text data and multi-angle standard image data of electric equipment, and establishing an equipment standard map library; S2, collecting real-time images of equipment to be identified through the mobile terminal when the materials go out of the warehouse or inventory the operation, calling an AI model in the AI comparison model library to identify the real-time images, comparing the identified equipment parts with standard information in the standard map library, and outputting a comparison result; s3, checking the comparison result, and if the identified part is inconsistent with the standard information or part loss exists, carrying out real-time warning; And S4, taking the checked and confirmed image and label information in the field identification result as a newly added training sample, periodically triggering incremental training of the AI model to optimize the identification performance of the model in a specific operation scene, and updating the optimized model to the AI comparison model library.
2. The intelligent identification and operation and maintenance management method for the electric power equipment based on the multi-mode feedback iteration of claim 1 is characterized in that in S1, training is performed to generate an AI model for equipment component identification and comparison, specifically, a Mask R-CNN model is used for training, so that target detection and instance segmentation of each component in an electric power equipment image are achieved.
3. The method for intelligent identification and operation and maintenance management of power equipment based on multi-modal feedback iteration of claim 2, wherein the step of training using Mask R-CNN model is as follows: S11a, standard image data of equipment are obtained from the standard map library, pixel-level labeling is carried out on power equipment and key components in the image, and labeling information comprising component types, boundary boxes and instance segmentation masks is generated; S12a, inputting standard image data into a backbone network of an initialized Mask R-CNN model to extract a multi-scale image feature map; S13a, generating a candidate target area through an area proposal network of a Mask R-CNN model by utilizing the image feature map extracted by the backbone network; S14a, after the candidate target area is processed through RoIAlign layers of a Mask R-CNN model, inputting the candidate target area into a target detection and instance segmentation head, and executing the following operations in parallel: Performing multi-classification on each candidate target area, accurately identifying the type of the component, and finely adjusting the position of a boundary frame of the component; Generating a binary mask for each candidate target area to realize pixel-level component instance segmentation; S15a, optimizing model parameters by minimizing a multi-task total loss function in the training process, wherein the loss function is a weighted sum of classification loss, bounding box regression loss and mask segmentation loss; S16a, the trained model is exported to be ONNX-format files and deployed into the AI comparison model library for real-time identification and calling.
4. The method for intelligent recognition and operation and maintenance management of power equipment based on multi-modal feedback iteration of claim 3, wherein when training is performed by using Mask R-CNN model, text data is fused to perform joint feature learning, and specifically comprising the following steps: s11b, material coding and text data of the equipment are converted into text feature vectors through a text embedding layer; s12b, fusing the text feature vector and the image feature map extracted by the backbone network to generate a multi-mode fusion feature map fused with semantic context information; And S13b, executing the training process from S13a to S15a by utilizing the multi-mode fusion feature map, and simultaneously carrying out integrated optimization on parameters of the backbone network, the text embedding layer, the area proposal network, the target detection and the example segmentation head by a back propagation algorithm in the training process, so as to finally generate the equipment identification model which can understand the equipment semantic information and fuses the multi-mode features.
5. The method for intelligent identification and operation and maintenance management of power equipment based on multi-modal feedback iteration of claim 4, wherein the step of S15a is a multi-task total loss function By loss of classification Regression loss of bounding box Loss from mask segmentation The weighted sum is constructed as follows: to classify the loss, Regression loss for bounding box, For the mask segmentation loss to be sufficient, The weight coefficients of the classification loss, the bounding box regression loss and the mask segmentation loss are respectively.
6. The method for intelligent identification and operation and maintenance management of a power device based on multi-modal feedback iteration of claim 5, wherein the step of simultaneously optimizing parameters of the backbone network, the text embedding layer, the regional proposal network, the target detection and instance segmentation head in an integrated manner through a back propagation algorithm in S13b comprises: S13b.1, calculating the gradient of the total loss L relative to the final output layer of the model; S13b.2, counter-propagating the gradient from top to bottom along the model architecture, sequentially passing through the following paths: a pre-measuring head path, wherein the gradient flows through the target detection and example segmentation head, the gradient of the partial parameter is obtained through calculation, and the gradient is continuously and reversely propagated to the input end of the pre-measuring head path; A regional proposal network path, wherein the gradient flows through the regional proposal network, the gradient of the partial parameter is obtained through calculation, and the gradient continues to be reversely propagated to the input end of the regional proposal network path; A multi-mode fusion path, namely, a gradient from a pre-measurement head path and a region proposal network path is converged at the multi-mode fusion feature map and back propagation is continued; Backbone network path, namely, converging gradients and flowing through a backbone network of the Mask R-CNN model, and calculating to obtain gradients of image feature extraction parameters; A text embedding layer path, wherein the converged gradients flow through the text embedding layer at the same time, and the gradients of text feature extraction and mapping parameters are obtained through calculation; And S13b.3, updating parameters of the backbone network, the text embedding layer, the area proposal network, the target detection and the example segmentation head at one time and synchronously according to all gradients calculated in the S13b.2 by utilizing an optimizer algorithm.
7. The method for intelligent identification and operation and maintenance management of a power device based on multi-modal feedback iteration of claim 6, wherein the step of S2 comprises: s21, collecting a real-time image of the power equipment to be identified through a camera at a material warehouse-in or inventory site, and uploading the real-time image to a system server; S22, after receiving the real-time image, the system server calls an AI model corresponding to the target device in an AI comparison model library, performs forward reasoning on the real-time image, and outputs an identification result, wherein the identification result comprises the types, the boundary frame positions and the pixel level division masks of all the power device components identified in the image; S23, according to the business bill or the user selection, a standard component list and a standard image of the corresponding equipment material codes are called from a standard map library; s24, comparing the identification result with the standard component list, and executing the following analysis: Checking whether the identified component category completely covers all the necessary components in the standard component list; Analyzing whether each component exists, whether the position is correct and whether the shape is complete or not based on the boundary box position and the segmentation mask; S25, generating a structured comparison report, wherein the report at least comprises: Parts which are correctly identified and which correspond to the standard list; Components that are present in the standard manifest but not identified in the real-time image; Identifying components that are morphologically abnormal or otherwise inconsistent with the standard manifest; And superposing the identified part boundary frame and the mask on the real-time image, and displaying the part boundary frame and the mask in parallel with the standard image.
8. The method for intelligent identification and operation and maintenance management of a power device based on multi-modal feedback iteration of claim 7, wherein the step of S3 comprises: s31, according to the current operation type and the equipment type, automatically matching and loading corresponding check rules; s32, receiving the comparison result, and automatically triggering one or more of the following checks based on the loaded check rule: The integrity check is that when the missing item list in the comparison result is not empty, the part is judged to be missing; The consistency check is carried out, namely an abnormal item list in the comparison result is compared with tolerance standards in a check rule, and when the degree of abnormality exceeds a preset threshold value, the components are judged to be inconsistent; Checking whether the identified part types and the number accord with the current service regulations or not; s33, executing a corresponding real-time alarm strategy according to the verification result: First-level warning, namely, for the condition needing to be confirmed manually, highlighting and text prompting are carried out on the mobile terminal and an application front-end interface; a second-level alarm, namely, for the confirmed missing or inconsistent parts, sending out an audible alarm when triggering a front-end interface alarm, and automatically suspending the subsequent business operation flow; Three-level alarming, namely for the abnormality affecting the operation safety, automatically generating an alarming work order and pushing the alarming work order to a background management system and appointed responsible persons while triggering a primary alarm and a secondary alarm; S34, whether the alarm is triggered or not, the final result of the identification, the verification conclusion and the processing opinion of the operator are stored in a database as a feedback record.
9. The method for intelligent identification and operation and maintenance management of a power device based on multi-modal feedback iteration of claim 8, wherein the step of S4 comprises: S41, automatically screening feedback records which are confirmed by field verification by the system, taking the records which are judged to be identified correctly as positive samples, directly storing images and marking information into an incremental training sample library, marking and correcting the records which are triggered to be identified incorrectly or lack of alarming, and then taking the records as negative samples and storing the negative samples into the incremental training sample library; S42, automatically triggering a model incremental training task according to a preset period or when the number of new samples in an incremental training sample library reaches a set threshold; S43, when executing the incremental training task, specifically comprising: loading a current online service model from an AI comparison model library as a pre-training weight; mixing the new sample in the incremental training sample library with part of samples in the original training data set to form a data set used in the incremental training; training the model by adopting a learning rate lower than that of initial training so as to finely adjust model parameters; S44, performing performance evaluation on the new model after incremental training by using a fixed verification set; If the comprehensive performance index of the new model is not lower than that of the original model, a new version number is allocated to the new model, and the new model is deployed to a test environment in an AI comparison model library.
10. An intelligent power equipment identification and operation management system based on multi-mode feedback iteration is characterized by being applied to a material warehouse-in and warehouse-out or inventory operation site, and comprising: the standard map and model library construction module is used for collecting material codes, text data and multi-angle standard image data of the power equipment to construct an equipment standard map library, training and generating an AI model for equipment part identification and comparison based on the standard map library, and constructing an AI comparison model library; the on-site operation and real-time comparison module is used for collecting real-time images of equipment to be identified through the mobile terminal when the materials go out of the warehouse or inventory operation; the intelligent checking and alarming module is used for checking the comparison result, and alarming in real time if the identified part is inconsistent with the standard information or the part is missing; And the feedback iteration and model optimization module is used for taking the checked and confirmed image and label information in the field recognition result as a newly added training sample, periodically triggering the incremental training of the AI model to optimize the recognition performance of the model under a specific operation scene, and updating the optimized model to the AI comparison model library.

Description

Multi-mode feedback iteration-based intelligent power equipment identification and operation management method and system Technical Field The application relates to the technical field of power equipment management, in particular to an intelligent power equipment identification and operation and maintenance management method and system based on multi-mode feedback iteration. Background Efficient and accurate identification and operation and maintenance management of power equipment are important basic stones for guaranteeing safe and stable operation of a power grid. In the digital and intelligent transformation process of the current power industry, equipment identification and management work still face a plurality of challenges. Currently, the mainstream solutions rely mainly on the following two technical paths: First, an identification method based on manual inspection and simple image processing. The manual identification is easy to cause missed detection and false detection due to factors such as fatigue, distraction and the like, and the identification accuracy and consistency are difficult to ensure. Although a part of schemes introduce a computer vision technology, traditional feature extraction algorithms (such as SIFT, SURF and the like) are mostly adopted, the feature expression capability is limited, the adaptability to complex scenes such as illumination change, shooting angles, component shielding and the like is poor, the generalization capability is weak, and the requirement of fine management of power equipment is difficult to meet. Second, an automated recognition method based on a deep learning model. Such methods have found widespread use in recent years, but they still have limitations. Firstly, most methods rely on only single image mode data, and text information which is widely existing in a power system and is strongly related to equipment cannot be effectively utilized, so that the model lacks semantic understanding capability, and when facing equipment with similar appearance and different models or poor image quality, the recognition confidence is low and errors are prone to occur. Secondly, the existing model is usually a static model, once training is completed, the performance of the model is fixed, and massive field data generated in daily operation and maintenance cannot be utilized for self iteration and optimization, so that the model is difficult to adapt to the fine changes of different substations and different working environments, and the recognition performance can be gradually degraded with time. Disclosure of Invention In order to solve the problems, the invention provides an intelligent power equipment identification and operation and maintenance management method and system based on multi-mode feedback iteration. In a first aspect, the present invention provides a method for intelligent identification and operation and maintenance management of an electrical device based on multi-mode feedback iteration, which is applied to a material warehouse entry or inventory operation site, and includes the following steps: S1, collecting material codes, text data and multi-angle standard image data of electric equipment, and establishing an equipment standard map library; S2, collecting real-time images of equipment to be identified through the mobile terminal when the materials go out of the warehouse or inventory the operation, calling an AI model in the AI comparison model library to identify the real-time images, comparing the identified equipment parts with standard information in the standard map library, and outputting a comparison result; s3, checking the comparison result, and if the identified part is inconsistent with the standard information or part loss exists, carrying out real-time warning; And S4, taking the checked and confirmed image and label information in the field identification result as a newly added training sample, periodically triggering incremental training of the AI model to optimize the identification performance of the model in a specific operation scene, and updating the optimized model to the AI comparison model library. By constructing a multi-mode standard map library and an AI comparison model library, the method realizes the closed loop flow of real-time image acquisition, AI identification, standard comparison, abnormal alarm, sample return and model iteration on a material input-output or inventory site, and breaks through the bottlenecks of low efficiency, large error and poor adaptability of the traditional manual identification. The method has strong scene adaptability and self-evolution capability, improves the identification accuracy of the power equipment and the operation automation level, reduces the artificial dependence and operation and maintenance cost, and supports the requirements of fine and intelligent management of the equipment in the intelligent power grid background. As a further limitation of the technical scheme