CN-122024159-A - Intelligent substation inspection system based on machine vision and infrared thermal imaging

CN122024159ACN 122024159 ACN122024159 ACN 122024159ACN-122024159-A

Abstract

The invention discloses an intelligent substation inspection system based on machine vision and infrared thermal imaging, and belongs to the technical field of intelligent operation and maintenance of power systems. The system comprises a data acquisition module, a historical data preprocessing module, a depth feature extraction module, a multi-mode data fusion module, a model training and reasoning module, a path dynamic planning module and a data safety and compatibility module, wherein the data acquisition module is used for synchronously acquiring real-time machine vision images, infrared thermal imaging data and environmental parameters, and the historical data preprocessing module is used for cleaning, normalizing and labeling historical inspection data, removing abnormal data and redundant information and generating a training data set and a verification data set. The method adopts a time stamp synchronization technology to realize synchronous acquisition and associated storage of machine vision, infrared thermal imaging, positioning and environmental parameters, ensures data time sequence consistency, can dynamically adjust sampling frequency, adapts to different inspection requirements and provides a high-quality data source for subsequent analysis.

Inventors

ZHAN HUAWEI
WANG ZHIQIANG
Zhan Haowen

Assignees

河南师范大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (8)

1. The intelligent substation inspection system based on machine vision and infrared thermal imaging is characterized by comprising a data acquisition module, a historical data preprocessing module, a depth feature extraction module, a multi-mode data fusion module, a model training and reasoning module, a path dynamic planning module and a data safety and compatibility module; The data acquisition module is used for synchronously acquiring real-time machine vision images, infrared thermal imaging data and environmental parameters; The historical data preprocessing module is used for cleaning, normalizing and labeling historical inspection data, removing abnormal data and redundant information and generating a training data set and a verification data set; The depth feature extraction module is used for performing multistage feature extraction on the preprocessed data, wherein the first stage feature extraction is used for extracting basic vision and temperature features, the second stage feature is based on the first stage feature extraction equipment association features, and the third stage feature is based on the second stage feature extraction abstract defect depth features to form a hierarchical feature system; the multi-mode data fusion module is used for performing cross-domain association on the third-level features by adopting a feature level fusion strategy, strengthening the corresponding feature weight of the defects by combining an attention mechanism, inhibiting interference features and generating a fusion type defect feature map; The model training and reasoning module is used for constructing a defect recognition model based on third-level characteristics and the fusion characteristic patterns, completing training, iteration and reasoning, realizing model scene adaptation optimization through an AI self-training platform, outputting defect type, grade and operation and maintenance suggestions, and driving the hardware coordination module to adjust the inspection action; the path dynamic planning module is used for constructing a risk quantification model based on the model reasoning result and the historical defect data, generating a patrol point priority list, fusing real-time obstacle and equipment energy consumption information, and dynamically optimizing a patrol path; and the data security and compatibility module is used for guaranteeing data transmission and storage security by adopting a terminal Bian Yun cooperative encryption mechanism and realizing seamless butt joint with the existing substation management system through the multi-protocol adaptation module.
2. The intelligent substation inspection system based on machine vision and infrared thermal imaging of claim 1, wherein the data acquisition module is used for synchronously acquiring and storing machine vision images, infrared thermal imaging data, positioning information and environmental parameters in an associated mode through a time stamp synchronization technology, and the sampling frequency can be dynamically adjusted according to inspection requirements to ensure data time sequence consistency.
3. The intelligent inspection system of the substation based on machine vision and infrared thermal imaging is characterized in that the historical data preprocessing module performs full-flow processing on historical inspection data, data cleaning is conducted, abnormal data such as temperature measurement drift, image blurring and packet loss transmission are eliminated by adopting an abnormal value detection algorithm, data normalization is conducted, visual image pixel values and infrared temperature data are mapped to the same dimension, dimensional differences are eliminated, data labeling is conducted, information labeling of defect areas, defect types, temperature thresholds and the like is completed based on a semi-automatic labeling tool, data partitioning is conducted, a training set, a verification set and a test set are divided according to a ratio of 7:2:1, meanwhile, a data enhancement strategy is added to expand the data set, and model generalization capability is improved.
4. The intelligent substation inspection system based on machine vision and infrared thermal imaging of claim 1, wherein in the depth feature extraction module, a first-stage feature extraction is performed by processing machine vision data by adopting a lightweight convolutional neural network, extracting basic vision features such as edges, textures, shapes, equipment component contours and the like, and outputting a basic vision feature map; The second-stage feature extraction, namely, based on the basic features extracted in the first stage, constructing a feature association model by adopting an improved residual error network, strengthening the internal relation between the basic features, extracting association features such as position association and appearance defect and corresponding relation of the parts of equipment for the basic features of machine vision, extracting association features such as mapping relation of hot spots and the equipment parts, diffusion trend of temperature abnormality and the like for the basic features of infrared temperature, realizing upgrading of the features from basic description to association analysis, and outputting a second-stage association feature sequence; Third-level feature extraction, namely constructing an abstract feature extraction model based on second-level associated features through a transducer encoder, mining hidden defect essential information in the associated features, extracting abstract appearance defect features for machine vision associated features, extracting abstract temperature abnormal defect features for infrared associated features, and outputting a single-mode three-level depth defect feature sequence to form a vertical feature system of 'basic-association-abstract'.
5. The intelligent substation inspection system based on machine vision and infrared thermal imaging is characterized in that in the multi-mode data fusion module, inter-mode feature correlation is achieved, mutual information entropy of three-level depth features of machine vision and three-level depth features of infrared is calculated, correlation degree of two-level features is quantized, feature components which are strongly related to defect recognition are screened out, redundant features and interference features are removed, an attention mechanism is that a cross-mode attention fusion module is constructed, dynamic weight distribution is conducted on the screened two-level single-mode three-level features, high weight is given to features corresponding to a defect region, low weight is given to background region features, the two-level single-mode depth features are fused into a unified fusion type defect feature map through feature stitching and dimension mapping, and multi-dimensional and deep characterization of equipment defects is achieved.
6. The intelligent substation inspection system based on machine vision and infrared thermal imaging according to claim 1, wherein in the model training and reasoning module, a YOLOv and Transformer mixed architecture is adopted for model construction and training to construct a defect identification model, a three-level vertical extracted single-mode depth feature sequence is used as a basic input, a fusion feature map is used as a core input, a preprocessed training set is utilized for carrying out model training, super parameters such as learning rate, convolution kernel size, attention weight coefficient and the like are dynamically adjusted through a verification set, early-stop strategy is adopted to prevent model overfitting, and the test set is used for verifying model performance and ensuring that model identification accuracy meets the standard; the AI self-training platform supports privately-arranged deployment, provides automatic data labeling, model training, parameter optimization and iterative updating functions, aims at new equipment models and new scene environments, can complete model iterative training by supplementing field data and corresponding three-level characteristic data, combines night enhancement and bad weather adaptation algorithm, optimizes extraction and adaptation capability of the model to three-level characteristics in complex environments, improves model generalization, outputs defect types, grades and operation and maintenance suggestions, namely, the model performs defect reasoning, outputs defect types, positions, confidence and severity grades based on characteristic maps obtained by extracting and fusing three-level characteristics of real-time acquired data, generates corresponding operation and maintenance suggestions, establishes an association mechanism of the defect characteristics and equipment accounts, automatically updates equipment health state files, and provides basis for follow-up inspection optimization, wherein the severity grades comprise slight, general and serious operation and maintenance suggestions comprise real-time early warning, planned overhaul and emergency treatment.
7. The intelligent inspection system of the substation based on machine vision and infrared thermal imaging is characterized in that the path dynamic programming module collects substation environment point cloud data through a laser radar SLAM technology, an initial three-dimensional map is built, map information is dynamically updated in combination with real-time data of fixed monitoring nodes and mobile equipment, response comprises environment changes of temporary obstacles and equipment position adjustment, meanwhile, a risk quantization model is built based on defect levels output by the model, defect risk coefficients corresponding to three-level features and historical defect data, risk values of inspection points are calculated, a point priority list is generated, high-risk points are inspected preferentially, the most comprehensive objects of inspection efficiency maximization, equipment energy consumption minimization and risk coverage are achieved based on an improved greedy algorithm, real-time obstacle information and point priorities are fused, an optimal inspection path is dynamically generated, when the model detects high-risk defects, the path is adjusted in an emergency mode, the nearest mobile equipment is moved to a site to check, and more accurate data are collected for model secondary reasoning.
8. The intelligent substation inspection system based on machine vision and infrared thermal imaging according to claim 1, wherein the data security and compatibility module adopts an end Bian Yun cooperative encryption mechanism, an AES-256 encryption algorithm is adopted in the data transmission process, a partition encryption and access control strategy is adopted in the storage process, special encryption is performed on three-level characteristic data and model parameters, a strict user authority management and security audit mechanism is established, vulnerability scanning and risk assessment are developed regularly to prevent data leakage and tampering, mainstream industrial communication protocols such as OPCUA and MQTT are integrated, seamless docking with an existing substation SCADA system and an equipment management system is realized, real-time synchronization and sharing of original inspection data, three-level characteristic data, defect identification results and model parameters are supported, and large-scale reconstruction of the existing system is not needed.

Description

Intelligent substation inspection system based on machine vision and infrared thermal imaging Technical Field The invention relates to the technical field of intelligent operation and maintenance of power systems, in particular to an intelligent substation inspection system based on machine vision and infrared thermal imaging. Background The substation is used as a core hub of the power system, and the running state of equipment directly determines the safety and stability of the power grid. The traditional substation inspection relies on manual periodic operation, and the problems of high labor intensity, low inspection efficiency, high personal safety risk in a high-pressure environment, high subjectivity of defect identification and the like exist. With the development of intelligent technology, an automatic inspection scheme based on machine vision, infrared thermal imaging and inspection robots gradually replaces manual inspection, but the prior art scheme still has a plurality of defects, particularly has obvious defects in the aspect of feature extraction and model application, and is difficult to meet the inspection requirements of modern substation on high precision, high efficiency and full scene. The core drawbacks of the prior art are mainly manifested in the following aspects: (1) The feature extraction level is shallow and lacks progressive logic, namely the existing system only extracts single-layer basic features, a hierarchical feature system is not constructed, deep defect associated features cannot be mined from basic information, the feature representation capability is weak, and the recognition requirement of complex defects of substation equipment is difficult to adapt. (2) The multi-mode data fusion and feature extraction are disjoint, that is, the existing system mostly adopts a mode of 'fusion before extraction' or 'simple superposition after single-mode extraction', a fusion strategy is designed without combining progressive relation of hierarchical features, and after fusion, the feature redundancy is high, the defect recognition is low, and misjudgment is easily generated due to environmental interference such as illumination, temperature difference and the like. (3) The model training and the characteristic suitability are poor, namely the defect identification model of the existing system is mostly based on single-level characteristic training, the characteristic advantage of depth characteristics is not fully utilized, the preprocessing flow of historical data is imperfect, the model generalization capability is weak, the identification precision is greatly reduced after equipment model or scene is replaced, and closed loop optimization of 'characteristic-model-inspection' is difficult to form. (4) The environment adaptability and the data reliability are insufficient, namely, complex environments such as strong light, overcast and rainy, dense fog, strong electromagnetic field and the like on the site of the transformer substation are easy to cause basic feature extraction distortion, and the existing system lacks an environment self-adaptive correction mechanism combined with hierarchical features, so that the problem of defect missing report and false report is further aggravated. (5) The existing scheme mostly adopts single inspection equipment, inspection coverage blind areas exist, the hardware equipment and the defect identification model lack deep linkage, inspection strategies cannot be dynamically adjusted based on model reasoning results, and inspection efficiency and accuracy are difficult to consider. Disclosure of Invention The invention aims to provide an intelligent substation inspection system based on machine vision and infrared thermal imaging, which can solve the problems of shallow level display, lack of progressive logic, multi-mode data fusion and feature extraction disconnection, poor model training and feature adaptability, insufficient environment adaptability and data reliability and lack of multi-equipment cooperation and model reasoning linkage in the prior art. According to one aspect of the invention, the intelligent substation inspection system based on machine vision and infrared thermal imaging comprises a data acquisition module, a historical data preprocessing module, a depth feature extraction module, a multi-mode data fusion module, a model training and reasoning module, a path dynamic planning module and a data safety and compatibility module; The data acquisition module is used for synchronously acquiring real-time machine vision images, infrared thermal imaging data and environmental parameters; The historical data preprocessing module is used for cleaning, normalizing and labeling historical inspection data, removing abnormal data and redundant information and generating a training data set and a verification data set; The depth feature extraction module is used for performing multistage feature extraction on the preprocessed data, whe