CN-121999272-A - Method and system for detecting infrared abnormal heating point in switch cabinet based on deep learning method and readable storage medium

CN121999272ACN 121999272 ACN121999272 ACN 121999272ACN-121999272-A

Abstract

The invention discloses a detection method, a detection system and a readable storage medium for infrared abnormal heating points in a switch cabinet based on a deep learning method, which comprise the steps of S1, collecting switch cabinet infrared images with abnormal heating points, S2, marking the collected switch cabinet infrared images to form an original data set, S3, dividing the data set into a training set and a testing set, S4, training the training set through a built target detection network model to obtain a first optimized target detection network model, S5, inputting the testing set into the first optimized target detection network model to test to judge whether convergence conditions are met, S6, bringing the weight of the target detection network model obtained by training once into the first optimized target detection network model again to conduct repeated training and verification, S7, extracting weight files of the target detection network model after training and deploying the weight files at edge ends. The invention realizes the rapid real-time detection of the infrared abnormal heating point in the switch cabinet.

Inventors

YUAN JUNFENG
DONG CHUNLEI
JI HUI
WU YONGCHAO
WANG HU
GUO GUANG

Assignees

国网河北省电力有限公司衡水供电分公司
国家电网有限公司
北京中科创益科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251222

Claims (10)

1. The method for detecting the infrared abnormal heating point in the switch cabinet based on the deep learning method is characterized by comprising the following steps of: s1, deploying acquisition equipment around a switch cabinet, and acquiring an infrared image of the switch cabinet with an abnormal heating point; s2, marking the collected infrared images of the switch cabinet to form an original data set; s3, randomly dividing the data set into a training set and a testing set; S4, training the training set through the built target detection network model, finally solving a loss function of the anchor frame formed by clustering and the marked detection frame, and performing error back transmission to optimize the target detection network model to obtain network weight and obtain a first optimized target detection network model; s5, inputting the test set into a first optimized target detection network model for testing so as to judge whether the target detection network model reaches a convergence condition or not; S6, carrying the weight of the target detection network model obtained through training once into the first optimized target detection network model again, and repeating training and verification until the loss function and the average accuracy converge, and stopping training; And S7, extracting the weight file of the trained target detection network model and deploying the weight file at the edge end to obtain an automatic recognition model of the abnormal heating point in the infrared image.
2. The method for detecting infrared abnormal heating points in a switch cabinet based on the deep learning method according to claim 1, wherein in step S4, training the training set for the built target detection network model comprises: The training set sequentially passes through the feature extraction part, the feature fusion part and the detection head of the built target detection network, and the feature images forming the pictures are fused in different scales in the process of feature layer-by-layer extraction.
3. The method for detecting infrared abnormal heating points in a switch cabinet based on the deep learning method according to claim 2, wherein the target detection network model in step S4 is a modified YOLOv model, and the modified YOLOv model includes: Optimization of the feature extraction section, optimization of the activation function, and optimization of the detection head.
4. The method for detecting infrared abnormal heating points in a switchgear cabinet based on the deep learning method according to claim 3, wherein in step S4, the optimization of the feature extraction section includes: The standardized composite convolution layer is modified into an SPD-Conv module.
5. The method for detecting infrared abnormal heating points in a switch cabinet based on the deep learning method according to claim 3, wherein in step S4, optimizing the activation function comprises: the original ReLU activation function is replaced with FReLU activation functions.
6. The method for detecting infrared abnormal heating points in a switch cabinet based on the deep learning method according to claim 3, wherein in step S4, optimizing the detection head comprises: a Dyhead multi-head attention mechanism is introduced in the detection head part.
7. The method for detecting infrared abnormal heating points in a switch cabinet based on a deep learning method according to claim 6, wherein the Dyhead multi-head attention mechanism introduced in the detection head part in the step S4 comprises scale attention, space attention and channel attention.
8. The method for detecting infrared abnormal heating points in the switch cabinet based on the deep learning method according to claim 1, wherein the marking method in the step S2 comprises manual or semi-automatic manual marking.
9. A deep learning method-based detection system for infrared abnormal heating points in a switch cabinet, wherein the system is used for realizing the deep learning method-based detection method for infrared abnormal heating points in the switch cabinet, and the system comprises the following components: The image acquisition module is used for deploying acquisition equipment around the switch cabinet and acquiring infrared images of the switch cabinet with abnormal heating points; The image labeling module is used for labeling the collected infrared images of the switch cabinet to form an original data set; The random division module is used for randomly dividing the data set into a training set and a testing set; The optimization module is used for training the training set through the built target detection network model, finally solving a loss function of the anchor frame formed by clustering and the marked detection frame, and performing error back transmission to optimize the target detection network model to obtain network weight and obtain a first optimized target detection network model; The test module is used for inputting the test set into the first optimized target detection network model for testing so as to judge whether the target detection network model reaches the convergence condition or not; the training and verifying module is used for bringing the weight of the target detection network model obtained by training once into the first optimized target detection network model again, and repeating training and verifying until the loss function and the average accuracy rate converge, and stopping training; the model deployment module is used for extracting the weight file of the trained target detection network model and deploying the weight file at the edge end to obtain an automatic recognition model of the abnormal heating point in the infrared image.
10. A readable storage medium, wherein one or more programs are stored in the readable storage medium, and the one or more programs are executable by one or more processors to implement the method for detecting infrared abnormal heating points in a switch cabinet based on a deep learning method according to any one of claims 1 to 8.

Description

Method and system for detecting infrared abnormal heating point in switch cabinet based on deep learning method and readable storage medium Technical Field The invention relates to the field of high-voltage electrical equipment monitoring, in particular to a detection method and system for infrared abnormal heating points in a switch cabinet based on a deep learning method and a readable storage medium. Background Abnormal temperature rise of current carrying elements (such as bus bar connectors and breaker contacts) in the high-voltage switch cabinet is caused by poor contact, overload and the like, and is a great hidden danger for inducing fusion welding, fire and power failure accidents of equipment. Infrared thermal imaging technology is an effective means to detect such temperature rise, but existing methods that rely on manual interpretation of infrared fluctuation thermal maps have significant drawbacks: the detection efficiency is low, the subjectivity is strong, and the large-scale inspection is difficult to deal with; is easy to be interfered by the environment (structure shielding, background temperature and emissivity setting deviation), so that tiny or hidden heating points are missed to be detected or misjudged; the positioning accuracy is not enough, and the manual work is difficult to correspond the high temperature region in the heat map to specific physical components (such as specific bolts and contact blades) in the cabinet fast and accurately, and accurate maintenance is hindered. The prior art lacks an automatic, intelligent and high-precision switch cabinet infrared abnormal heating point identification method so as to meet the requirements of efficient and reliable equipment state monitoring. Therefore, it is needed to provide a method and a system for detecting infrared abnormal heating points in a switch cabinet based on a deep learning method, and a scheme of a readable storage medium. Disclosure of Invention In order to solve the problems, the technical scheme of the invention provides a detection method, a detection system and a readable storage medium for infrared abnormal heating points in a switch cabinet based on a deep learning method, which can be used for rapidly detecting the occurrence of the abnormal heating points by an infrared probe in switch cabinet equipment. According to an embodiment of a first aspect of the present invention, there is provided a method for detecting an infrared abnormal heating point in a switch cabinet based on a deep learning method, including: s1, deploying acquisition equipment around a switch cabinet, and acquiring an infrared image of the switch cabinet with an abnormal heating point; s2, marking the collected infrared images of the switch cabinet to form an original data set; s3, randomly dividing the data set into a training set and a testing set; S4, training the training set through the built target detection network model, finally solving a loss function of the anchor frame formed by clustering and the marked detection frame, and performing error back transmission to optimize the target detection network model to obtain network weight and obtain a first optimized target detection network model; s5, inputting the test set into a first optimized target detection network model for testing so as to judge whether the target detection network model reaches a convergence condition or not; S6, carrying the weight of the target detection network model obtained through training once into the first optimized target detection network model again, and repeating training and verification until the loss function and the average accuracy converge, and stopping training; And S7, extracting the weight file of the trained target detection network model and deploying the weight file at the edge end to obtain an automatic recognition model of the abnormal heating point in the infrared image. In the above scheme, in step S4, training the training set for the constructed target detection network model includes: The training set sequentially passes through the feature extraction part, the feature fusion part and the detection head of the built target detection network, and the feature images forming the pictures are fused in different scales in the process of feature layer-by-layer extraction. In the above solution, the object detection network model in step S4 is a modified YOLOv model, and the modified YOLOv model includes: Optimization of the feature extraction section, optimization of the activation function, and optimization of the detection head. In the above-described aspect, in step S4, the optimizing of the feature extraction section includes: The standardized composite convolution layer is modified into an SPD-Conv module. In the above scheme, in step S4, the optimizing the activation function includes: the original ReLU activation function is replaced with FReLU activation functions. In the above scheme, in step S4, the optimization of the detect