Fault Identification Method for Panoramic Infrared Images of Electrical Equipment Based on Expert Knowledge Base

Abstract

The present invention discloses a method for fault identification of panoramic infrared images of electrical equipment based on an expert knowledge base, including: (1) establishing an expert database of infrared images of electrical equipment; (2) acquiring multiple consecutive infrared images of electrical equipment and performing preprocessing on the image data; (3) performing infrared image stitching and outputting a panoramic infrared image of the electrical equipment; (4) using an algorithm for fault identification of infrared images of electrical equipment to identify faults in the panoramic infrared image of the electrical equipment. Through the method for fault identification of panoramic infrared images of electrical equipment based on an expert knowledge base, the present invention realizes the overall control of the operating conditions of electrical equipment, as well as the intelligent judgment of fault types and the analysis of fault causes, ensuring the safe and stable operation of electrical equipment.

Description

Technical Field

[0001] The present invention relates to a method for identifying faults in panoramic infrared images of electrical equipment. Background Art

[0002] With the accelerating pace of power grid transformation, higher requirements are also put forward for the safe and stable operation of electrical equipment in the power grid. As an indispensable important safety tool in power inspection, thermal infrared imaging technology plays an important guiding role in detecting abnormal heating points of equipment and discovering potential equipment faults.

[0003] According to the current situation investigation:

[0004] The current infrared image dataset of electrical equipment faults only labels the location of the fault points and the type of equipment, without further in-depth exploration of the types and causes of equipment faults, resulting in the need for manual discrimination in infrared inspection, which seriously reduces the efficiency of electrical equipment fault diagnosis.

[0005] Due to hardware component limitations, an infrared thermal imager can only capture a certain area. For large electrical equipment such as transformers, it is impossible to completely capture the infrared image of the electrical equipment, resulting in the lack of integrity in subsequent fault analysis.

[0006] The above shows that the existing infrared detection methods for electrical equipment have operational limitations in electrical equipment fault diagnosis. Summary of the Invention

[0007] The purpose of the present invention is to provide a method for identifying faults in panoramic infrared images of electrical equipment based on an expert knowledge base, which can intelligently identify the type of electrical equipment during electrical equipment inspection, guide the staff to discover potential electrical equipment faults, analyze the causes of faults, improve the inspection efficiency, and ensure the safe and stable operation of power equipment.

[0008] The technical solution of the present invention is:

[0009] A method for identifying faults in panoramic infrared images of electrical equipment based on an expert knowledge base includes the following technical contents:

[0010] 1. Based on the existing infrared images of electrical equipment faults, an expert knowledge base is established. Specifically, first, the types of faulty electrical equipment in each infrared image of electrical equipment faults are classified, and the position coordinates of abnormal heating points are given; secondly, the types of faults are judged and treatment opinions are given. The establishment of the expert knowledge base involves various main electrical equipment such as transformers, instrument transformers, wall bushings, circuit breakers, etc., covering the vast majority of inspection scenarios.

[0011] 2. When inspecting large electrical equipment, multiple infrared images need to be taken from the same direction of the equipment to determine the overall operating condition of the equipment. Specifically, for the method of stitching panoramic infrared images of electrical equipment, first, sort the multiple continuously taken infrared images of the same electrical equipment in chronological order, and by default, a maximum of 6 images can be stitched at one time; second, extract the feature points from the temperature overlapping area images of two adjacent numbered infrared images; then, match the feature points, form an image transformation matrix with the image with the smaller number as the reference, stitch the adjacent numbered images, and assign the smallest serial number of the two images before stitching to the stitched image; finally, loop through all the images according to the numbers to generate a panoramic infrared image.

[0012] Through the method for fault identification of panoramic infrared images of electrical equipment based on an expert knowledge base, the present invention realizes the overall control of the operating condition of electrical equipment, as well as the intelligent judgment of fault types and the analysis of fault causes, ensuring the safe and stable operation of electrical equipment. Brief Description of the Drawings

[0013] The present invention will be further described below in conjunction with the drawings and embodiments.

[0014] Figure 1 It is a schematic diagram of the specific structure of the L2-Net network.

[0015] Figure 2 It is a schematic diagram of the specific structure of the Res-DS network.

[0016] Figure 3 It is a diagram of the identification result of using the electrical equipment infrared image fault detection model to identify faults in the panoramic infrared image of electrical equipment.

[0017] Figure 4 It is a schematic diagram of feature point matching.

[0018] Figure 5 It is a schematic diagram of the single stitching result. Detailed Embodiment

[0019] A method for fault identification of panoramic infrared images of electrical equipment based on an expert knowledge base includes the following steps:

[0020] (1) Establish an expert database for infrared images of electrical equipment;

[0021] (2) Obtain multiple consecutive infrared images of electrical equipment and perform image data preprocessing;

[0022] (3) Perform infrared image stitching and output a panoramic infrared image of electrical equipment;

[0023] (4) Use the electrical equipment infrared image fault identification algorithm to identify faults in the panoramic infrared image of electrical equipment.

[0024] The specific method of step (1) is as follows:

[0025] (1) Select infrared images of electrical equipment failures, including transformers, instrument transformers, bushing insulators, circuit breakers, insulators, and lightning arresters, a total of 1000 infrared images, to establish an expert knowledge base;

[0026] (2) For each infrared image of electrical equipment failure, use LabelImage software to frame the position of the fault point and label the category of the faulty equipment;

[0027] (3) Combine the category and location of the faulty equipment, and according to the "Application Specification for Infrared Diagnosis of Energized Equipment" (DL / T 664-2016), conduct thermal image feature analysis and fault feature analysis on each image, and further give fault category judgment and fault handling suggestions;

[0028] (4) Take the fault category judgment, fault handling suggestions, the framed position of the fault point, and the category of the faulty equipment as image labels, and read and save them one by one with the corresponding infrared images of electrical equipment failures to establish an expert database of infrared images of electrical equipment, providing sufficient data support for subsequent fault identification.

[0029] The specific method of step (2) is as follows:

[0030] (1) Take multiple infrared images of the same electrical equipment through an infrared thermal imager, and number the photos in ascending order according to the shooting time;

[0031] (2) Perform bilateral filtering denoising on the infrared images.

[0032] The specific method of step (3) is as follows:

[0033] (1) Extract SIFT feature points from the preprocessed infrared images numbered 1 and 2;

[0034] (2) Generate Res-DS (Residual-Double Scale) feature point descriptors for the feature points in the above two infrared images;

[0035] (3) Match the feature points in the above two infrared images according to the generated feature point descriptors;

[0036] (4) Use the RANSAC method to eliminate the wrong matching point pairs in the image and generate the optimal image transformation matrix model;

[0037] (5) Input the above two infrared images into the image transformation matrix model to generate a spliced infrared image, and number the generated image as 3;

[0038] Repeat the above steps until all infrared images are traversed.

[0039] The specific steps of step (2) in step (III) are as follows:

[0040] a. Introduce the L2-Net network. In the L2-Net network, the feature tower is jointly constructed by a pooling layer and a convolutional operation with a stride of 2. Except for the last convolutional layer, there is a normalization layer after each convolutional layer to keep the weights and bias parameters of the convolutional layer between 0 and 1. The final convolutional operation uses an L2 normalization layer to generate a 128-dimensional feature vector.

[0041] b. Establish a Res-DS (Residual-Double Scale) network model. This network simultaneously uses two L2-Net networks to describe the features of the original infrared image and the 32×32 pixel block around the feature point, adds a residual structure to a single L2-Net network, and adds an information exchange structure between the two L2-Net networks. The Res-DS network not only increases the ability to extract multi-level features of a single-scale image but also fuses the features at both the global and local scales of the image. The specific structure of the Res-DS network is shown in Figure 2 .

[0042] c. Production of the Res-DS network training set; In order to generate more accurate feature point descriptors, a self-made infrared image dataset of electrical equipment is created. This dataset contains 137 groups of infrared images. Each group of infrared images is taken from different angles of the same electrical equipment. Image blocks with a size of 32×32 pixels are extracted from the same region of interest in each image of the same group. Finally, 253 pairs of image blocks and the corresponding original infrared images (matched / unmatched) are selected as the input of the Res-DS network.

[0043] d. Definition of the loss function; In order to increase the distance between the feature descriptors of non-matching pairs and reduce the distance between the feature descriptors of matching pairs, a structural loss function E1 and a geometric loss function E2 are defined; two feature vectors generated from a pair of image blocks and the corresponding original infrared image derive the corresponding cosine similarity matrix S = F1F2 T , calculate the vector L = S - αdiag(S), and obtain the calculation formula for the structural loss function E1 as:

[0044]

[0045] where, l i,j is an element in the vector L. α ∈ (0,1) is used to increase the distance between non-matching pairs and matching pairs; the calculation formula for the geometric loss function E2 is:

[0046] E2 = ||F1,F2||

[0047] where ||·|| is the Euclidean distance. Taking E1 + αE2 as the loss function, α and λ are set to 0.4 and 0.2 respectively;

[0048] e. Train the Res-DS network. Use the gradient descent method to update the network structure parameters for training, with a learning rate of 0.001, a weight of 0.0001, and a learning rate decay rate of 0.9. Perform data augmentation using operations such as random flipping, scaling, 90° rotation, brightness, and contrast adjustment; the matching set size and batch size are 8 and 64 respectively, and the output feature vectors are normalized to a unit norm with a mean of 0. The size of the trained model is 5.9MB. Finally, use this model to generate 128-bit feature point descriptors.

[0049] The specific method of step (iv) is as follows:

[0050] (1) Train the electrical equipment infrared image fault detection model; this model is based on the YOLOV8 deep learning network and uses the electrical equipment infrared image expert database as the training set. The size of the finally generated model is 11.3MB;

[0051] (2) Model testing. Use the electrical equipment infrared image fault detection model to identify faults in the panoramic infrared images of electrical equipment.

Claims

1. A fault identification method for panoramic infrared images of electrical equipment based on an expert knowledge base, characterized in that: It includes the following steps: (1) Establish an expert database for infrared images of electrical equipment; (2) Obtain multiple consecutive infrared images of electrical equipment and perform preprocessing on the image data; (3) Perform infrared image stitching and output a panoramic infrared image of the electrical equipment; (4) Use the fault identification algorithm for infrared images of electrical equipment to identify faults in the panoramic infrared image of the electrical equipment; The specific method of step (3) is: (1) Extract SIFT feature points from the preprocessed infrared images numbered 1 and 2; (2) Generate Res-DS feature point descriptors for the feature points in the above two infrared images; (3) Match the feature points in the above two infrared images according to the generated feature point descriptors; (4) Use the RANSAC method to remove the wrong matching point pairs in the image and generate the optimal image transformation matrix model; (5) Input the above two infrared images into the image transformation matrix model to generate the stitched infrared image, and let the generated image be numbered 3; Loop the above steps until all infrared images are traversed; The specific steps of step (2) in step (3) are: a. Introduce the L2-Net network. In the L2-Net network, the feature tower is jointly constructed by a pooling layer and a convolutional operation with a stride of 2. Except for the last convolutional layer, there is a normalization layer after each convolutional layer to keep the weights and bias parameters of the convolutional layer between 0 and 1. The final convolutional operation uses the L2 normalization layer to generate a 128-dimensional feature vector; b. Establish a Res-DS network model. This network simultaneously uses two L2-Net networks to describe the features of the original infrared image and the 32×32 pixel block around the feature points, adds a residual structure to the individual L2-Net network, and adds an information exchange structure between the two L2-Net networks; c. Production of the Res-DS network training set; in order to generate more accurate feature point descriptors, a self-made infrared image dataset of electrical equipment is made. This dataset contains 137 groups of infrared images. Each group of infrared images is taken from different angles of the same electrical equipment. Image blocks with a size of 32×32 pixels are extracted from the same region of interest in each image of the same group. Finally, 253 pairs of image blocks and the corresponding original infrared images are selected as the input of the Res-DS network; d. Definition of loss function; in order to increase the distance between the feature descriptors of non-matching pairs and reduce the distance between the feature descriptors of matching pairs, the structural loss function E1 and the geometric loss function E2 are defined; two feature vectors generated from a pair of image patches and the corresponding original infrared image Derive the corresponding cosine similarity matrix Calculate the vector L = S - αdiag(S), and the calculation formula for the structural loss function E1 is obtained as follows: where l i,j is an element in vector L; α ∈ (0, 1) is used to increase the distance between non-matching pairs and matching pairs; the calculation formula of the geometric loss function E2 is: E2 = ||F1,F2|| where ||·|| is the Euclidean distance; use E1 + αE2 as the loss function, and α is set to 0.4; e. Train the Res-DS network; use the gradient descent method to update the network structure parameters for training. The learning rate is 0.001, the weight is 0.0001, the learning rate decay rate is 0.9, and data augmentation is performed using operations such as random flipping, scaling, 90° rotation, brightness and contrast adjustment. The matching set size and batch size are 8 and 64 respectively. The output feature vector is normalized to a unit norm with a mean of 0. The size of the trained model is 5.9MB; finally, this model is used to generate 128-bit feature point descriptors.

2. The method for identifying faults in a panoramic infrared image of an electrical equipment based on an expert knowledge base according to claim 1, characterized in that: The specific method of step (1) is as follows: (1) Select 1,000 infrared images of electrical equipment failures, including transformers, instrument transformers, wall bushings, circuit breakers, insulators, and lightning arresters, to establish an expert knowledge base; (2) For each infrared image of electrical equipment failure, use LabelImage software to frame the location of the fault point and label the category of the faulty equipment; (3) Combine the category and location of the faulty equipment, conduct thermal image feature analysis and fault feature analysis on each image, and further give fault category judgments and fault handling suggestions; (4) Take the fault category judgment, fault handling suggestions, the framed location of the fault point, and the category of the faulty equipment as image labels, and read and save them one by one with the corresponding infrared images of electrical equipment failures to establish an expert database of infrared images of electrical equipment, providing sufficient data support for subsequent fault identification.

3. The method for fault identification of panoramic infrared images of electrical equipment based on an expert knowledge base according to claim 1, characterized in that: The specific method of step (2) is as follows: (1) Take multiple infrared images of the same electrical equipment with an infrared thermal imager, and number the photos in ascending order according to the shooting time; (2) Perform bilateral filtering denoising on the infrared images.

4. The method for identifying faults in panoramic infrared images of electrical equipment based on an expert knowledge base according to claim 1, characterized in that: The specific method of step (4) is as follows: (1) Train a fault detection model for infrared images of electrical equipment; this model is based on the YOLOV8 deep learning network and uses the expert database of infrared images of electrical equipment as the training set, and the final size of the generated model is 11.3 MB; (2) Model testing; use the fault detection model for infrared images of electrical equipment to identify faults in panoramic infrared images of electrical equipment.

Images