Infrared image-oriented electrical equipment heating defect detection method and system

Abstract

The invention relates to the technical field of image processing, and discloses an infrared image-oriented electrical equipment heating defect detection method and system, and the method comprises the following steps: obtaining an infrared image data set of electrical equipment in a transformer substation scene; using an infrared image data set, based on a deep learning model introducing an improved CBAM attention mechanism module, performing identification detection on the power equipment in the infrared image, and obtaining the type and position of the equipment in a rotary detection frame format; according to the obtained detection frame result, obtaining an equipment main body area and a corresponding temperature as an average temperature of the equipment; taking the temperature of the main body area of the equipment as the temperature of the equipment during normal operation, and performing current-induced heating type defect detection on the equipment in the detection frame; according to the invention, the purpose of rapidly and accurately identifying the heating defect of the power equipment can be achieved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for detecting heating defects in power equipment based on infrared images. Background Technology

[0002] In the monitoring and maintenance of power equipment, the detection of overheating defects is a critical issue. With the expansion of power equipment scale and the increasing complexity of the environment, traditional manual inspection or simple sensor detection methods can no longer meet the requirements of real-time performance and accuracy. Especially in high-voltage power equipment, overheating problems are difficult to detect in a timely manner, leading to equipment damage and safety accidents.

[0003] Existing methods only acquire infrared images, then rely on manual interpretation or traditional image processing algorithms. This approach is insufficient for efficiently identifying defects across a large number of devices, resulting in low accuracy and poor efficiency. Therefore, there is an urgent need for an intelligent, efficient, and accurate infrared image analysis method capable of automatically detecting and locating heating defects in power equipment. This would improve the accuracy and efficiency of equipment maintenance and address the shortcomings of traditional methods. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for detecting thermal defects in power equipment based on infrared images, which can quickly and accurately identify thermal defects in power equipment.

[0005] The first aspect of this invention provides a method for detecting heating defects in power equipment based on infrared images, comprising the following steps: Acquire infrared image datasets of power equipment in a substation scenario; Using an infrared image dataset, a deep learning model with an improved CBAM attention mechanism module is used to identify and detect electrical equipment in infrared images, and the equipment category and location are obtained in the form of a rotated detection box. The improved CBAM attention mechanism module performs batch normalization operation on each channel in the channel attention part of the CBAM attention mechanism. Based on the obtained detection results, the main body area of ​​the device and its corresponding temperature are obtained as the average temperature of the device. The temperature of the main body area of ​​the equipment is taken as the temperature during normal operation of the equipment, and the equipment within the detection frame is subjected to current-induced heating type defect detection. Voltage-induced thermal defect detection is then performed on the three-phase equipment within the detection frame.

[0006] Optionally, deep learning models that incorporate an improved CBAM attention mechanism module include: a backbone network, a neck network, and an output end; The backbone network includes a C2f module, a convolutional module, an improved CBAM attention mechanism module, and a spatial pyramid pooling module. The C2f module and the convolutional module extract and reconstruct features of power equipment in infrared images, gradually obtaining feature representations of power equipment from local to global perspectives to obtain feature maps. The improved CBAM attention mechanism module optimizes the feature maps to obtain optimized feature maps. The spatial pyramid pooling module performs multi-scale pooling operations on the optimized feature maps to obtain high-level feature representations. The neck network performs bidirectional fusion of high-level feature representations from top to bottom and bottom to top to obtain a fused feature map; The output is used to output the precise location and category information of the power equipment in the form of a rotated detection box from the fused feature map.

[0007] Optional, improved CBAM attention mechanisms include channel attention components and spatial attention components; In the channel attention part, batch normalization is performed on each channel. By calculating the mean and variance in the batch and using learnable scaling and translation parameters, the channel information is batch normalized to reflect the importance of each channel, strengthen the influence of important channels on the attention weight value, and weaken the influence of unimportant channels on the attention weight value.

[0008] Optionally, the improved formula for the channel attention part is expressed as follows: , in, Input features F Channel weights, It is the Sigmoid activation function. This represents the output tensor obtained after batch normalization of the input batches. MLP stands for Perceptron with shared parameters. Input features F Global max pooling, This indicates element-wise multiplication; , For the input feature map, For input features Output feature map after applying channel attention weights.

[0009] Optionally, obtaining the main area of ​​the device and its corresponding temperature includes the following steps: Otsu's temperature-weighted method is used to segment the device region of the detection frame and an opening operation is performed to obtain the main area mask of the device. Based on the mask of the main equipment area, read the temperature value of the main equipment area where the mask is 0. The main equipment area is where the mask is 0.

[0010] Optionally, obtaining the device body region mask includes the following steps: First, the grayscale value of each pixel within the detection box is obtained and mapped to the range of 0–255. A temperature weighting mechanism is introduced to assign higher weights to pixels with higher temperatures within the detection box, thereby forming a temperature-weighted grayscale statistical distribution. The Otsu method divides pixels into foreground and background candidate regions by traversing all gray thresholds in the gray-level statistical distribution, and calculates the pixel proportion and gray mean of the two types of regions respectively. Then, it calculates the inter-class variance between the foreground and background candidate regions. When the inter-class variance corresponding to a certain gray threshold reaches the maximum, it is considered that the threshold can most effectively distinguish between the foreground and background regions, and the gray value is determined as the optimal threshold. Pixels with gray values ​​less than the optimal threshold within the detection box are marked as the main body area of ​​the device, and pixels with gray values ​​greater than or equal to the optimal threshold are marked as the background area, thus generating an initial mask image. The initial mask image is then subjected to opening operations for noise reduction to obtain the final device body region mask.

[0011] Optional, optimal threshold We obtain it from the following formula: , in, This represents the probability that a pixel is in the foreground. This represents the probability that a pixel is part of the background. The mean gray level of the foreground. This represents the average grayscale value of the background.

[0012] Optionally, the process of performing current-induced thermal defect detection on the equipment within the detection frame includes the following steps: First, obtain the temperature of the hottest point in the equipment area, i.e., the highest temperature; The relative temperature difference of the heat source is obtained based on the average temperature of the equipment area, the air temperature on the day of shooting, and the temperature of the heat source in the equipment area. Based on the equipment type, heating point temperature, and relative temperature difference of the heating point, and in accordance with power industry standards, determine whether the equipment has a current-induced heating defect.

[0013] Optionally, performing voltage-induced thermal defect detection on the three-phase equipment within the detection frame includes the following steps: Obtain the longer geometric center line of each detection frame, read the temperature distribution in the main body area of ​​the device on the center line, and obtain three temperature center lines. During the alignment process, add a deformation penalty term when calculating the distance between two points to obtain the optimized distance between two points. The improved DTW alignment method is used to align the three temperature midlines. The improved DTW alignment method includes: normalizing the amplitude of the three midline temperature sequences, and aligning the three temperature midlines to the shortest temperature midline length using two alignments. By comparing the aligned temperature centerlines in segments, and according to power industry standards, the temperature differences at different corresponding locations on the equipment are compared to determine whether the equipment has voltage-induced heating defects.

[0014] A second aspect of the present invention provides a power equipment heating defect detection system based on infrared images, comprising: The image acquisition module is used to acquire infrared image datasets of power equipment in substation scenarios. The device identification and localization module is built on a deep learning model that incorporates an improved CBAM attention mechanism module. It is used to identify and detect the input infrared image and output the category and location information of the power equipment in a rotated detection box format. The improved CBAM attention mechanism module performs batch normalization on each channel in its channel attention part. The equipment body area and temperature extraction module is used to extract the corresponding equipment body area and its corresponding temperature according to the rotating detection frame, and use it as the average operating temperature of the equipment. The current-induced heating type defect detection module is used to analyze and judge the current-induced heating type defects of the equipment within the detection frame based on the temperature of the main body area of ​​the equipment as the normal operation benchmark. The voltage-induced heating defect detection module is used to analyze and judge voltage-induced heating defects in objects identified as three-phase equipment within the detection frame.

[0015] The technical solution provided by the embodiments of the present invention has the following advantages compared with the prior art: This invention provides a method and system for detecting thermal defects in power equipment based on infrared images. It abandons the inefficient traditional methods relying on manual interpretation or simple image processing algorithms. By introducing a deep learning model with an improved CBAM attention mechanism, it constructs an end-to-end automated detection process, effectively solving the problems of low accuracy and poor efficiency in efficiently identifying defects in a large number of devices. This significantly improves the intelligence level and detection efficiency of equipment maintenance. The deep learning model with the improved CBAM attention mechanism outputs a rotated detection box format, which can more accurately fit the irregular shapes of various power equipment. In particular, the batch normalization operation introduced in the channel attention part can adaptively strengthen key channel information related to the equipment's thermal characteristics while suppressing irrelevant or redundant background interference. This allows the model to focus more on the main body of the equipment and its key thermal areas, solving the problem of target feature extraction in complex environments and improving the accuracy of identifying and locating small, dense, or partially obscured targets in complex substation scenarios. This invention first accurately identifies the equipment type and location, and then extracts the average temperature of the main body area of ​​the equipment as a normal operation benchmark. Subsequent steps involve detecting current-induced heating defects and voltage-induced heating defects for three-phase equipment, respectively. This differentiated analysis strategy based on equipment type and heating principle makes defect judgment more scientific and evidence-based, reduces false alarm rates, and improves the targeting and reliability of defect diagnosis, helping to more accurately locate the root cause of the fault. This invention boasts a high degree of automation from image input to defect type output, capable of processing large-scale infrared image datasets. This provides a solid technical foundation for the normalized and automated infrared inspection and online monitoring of equipment in power systems, helping to detect latent heating defects in advance and prevent equipment damage and safety accidents. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for detecting thermal defects in power equipment based on infrared images, provided in an embodiment of the present invention; Figure 2 A schematic diagram of the improved CBAM attention mechanism module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall model structure provided in the embodiments of the present invention; Figure 4 The results of device body acquisition and opening operation denoising provided in the embodiments of the present invention; Figure 5 The device current-induced heating detection results provided in the embodiments of the present invention; Figure 6 shows the temperature centerline alignment result provided by an embodiment of the present invention, wherein... Figure 6a This is the original temperature curve. Figure 6b Temperature curve after DTW alignment; Figure 7 shows the detection results of voltage-induced heating type defects in the three-phase equipment provided in the embodiment of the present invention. Figure 7a The output is the detection result. Figure 7b This is the result of midline detection after DTW alignment. Figure 7c This is a comparison chart of the segmented results of the midline. Detailed Implementation

[0017] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0018] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0019] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0020] like Figure 1 As shown, the first embodiment of the present invention provides a method for detecting heating defects in power equipment based on infrared images, including the following steps: S1: Obtain an infrared image dataset of power equipment in a substation scenario. In this embodiment of the invention, an infrared camera is used to photograph the power equipment in the substation scenario to construct an infrared image dataset, specifically including the following steps: S1.1: The power equipment that needs to be detected in the substation is photographed using an infrared camera, either manually or by drone. In this embodiment of the invention, the infrared images used are taken manually, totaling more than three thousand original images. S1.2: Label the acquired image in OBB rotated detection box format. The label content includes the device category and the coordinates of the four vertices of the detection box.

[0021] S2: Using an infrared image dataset, based on a deep learning model with an improved CBAM attention mechanism module, electrical equipment in the infrared images is identified and detected, and the equipment category and location are obtained in the form of a rotating detection box. The improved CBAM attention mechanism module performs batch normalization on each channel in the channel attention part of the CBAM attention mechanism. In this embodiment of the invention, the deep learning model adopts the YOLOv8 deep learning model because the YOLOv8 deep learning model is more suitable for OBB rotating box detection, and has a fast detection speed and high accuracy, making it more suitable for infrared detection scenarios. S3: Based on the obtained detection frame results, obtain the main body area of ​​the equipment and the corresponding temperature, which is used as the average temperature of the equipment for current-induced heating type defect detection. S4: Using the temperature of the main body area of ​​the equipment as the temperature during normal operation, perform current-induced heating type defect detection on the equipment within the detection frame to obtain whether the equipment has current-induced heating type defects and the defect level, such as: no defect, general defect, serious defect or emergency defect. S5: Perform voltage-induced heating defect detection on the three-phase equipment within the detection frame to determine whether the equipment has voltage-induced heating defects and the defect level, such as: no defect, general defect, serious defect or emergency defect. For example, if three current transformers and one surge arrester are detected in an image, then current-induced heating detection is performed on all three current transformers and surge arresters, and then voltage-induced heating detection is performed on the three current transformers.

[0022] Optionally, deep learning models that incorporate an improved CBAM attention mechanism module include: a backbone network, a neck network, and an output head. The backbone network comprises a C2f module, a convolutional module, an improved CBAM attention mechanism module, and a spatial pyramid pooling module. The C2f and convolutional modules extract and reconstruct features from power equipment in infrared images, progressively acquiring feature representations of the power equipment from local to global perspectives, resulting in feature maps. The improved CBAM attention mechanism module optimizes the feature maps to obtain optimized feature maps, and the spatial pyramid pooling module performs multi-scale pooling operations on the optimized feature maps to obtain high-level feature representations. Specifically, the convolutional module performs basic local feature perception and spatial downsampling, while the C2f module enhances feature propagation efficiency through its unique cross-layer connection structure. Rate and gradient flow effectively preserve device edges, textures, and fine-grained features during feature extraction. This reduces computational cost while better integrating shallow details (such as device edges and textures) and deep semantic features (such as the overall shape of the device). This gradually constructs a hierarchical feature representation of power equipment, from local details to global structure, yielding an initial feature map. This feature map is then input into an improved CBAM attention mechanism module, which performs weighted optimization on the channel and spatial dimensions to adaptively highlight important channels and key spatial regions related to the heating characteristics of power equipment and suppress redundant feature information. The attention-optimized features are further input into the Spatial Pyramid Pooling (SPPF) module. SPPF fuses contextual information from different receptive fields through parallel multi-scale pooling operations, thereby improving the network's ability to represent power equipment of different sizes and scales, forming a high-level feature representation containing rich semantic information.

[0023] The neck network performs bidirectional fusion of high-level feature representations from top to bottom and bottom to top, resulting in a fused feature map. Specifically, high-level features are input into the neck network, which fuses and propagates features from different levels through a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN), achieving effective integration of multi-scale feature information. The fused features are then input to the output, which performs target detection of power equipment, outputting the equipment's category and location information in a rotated detection box format. This provides a foundation for subsequent extraction of the main equipment region and analysis of thermal defects.

[0024] The output is used to output the precise location and category information of the power equipment in the form of a rotated detection box from the fused feature map.

[0025] Optional, improved CBAM attention mechanisms include channel attention components and spatial attention components; In the channel attention part, batch normalization is performed on each channel. By calculating the mean and variance in the batch and using learnable scaling and translation parameters, the channel information is batch normalized to reflect the importance of each channel, strengthen the influence of important channels on the attention weight value, and weaken the influence of unimportant channels on the attention weight value.

[0026] Optionally, the improved formula for the channel attention part is expressed as follows: , in, Input features F Channel weights, It is the Sigmoid activation function. This represents the output tensor obtained after batch normalization of the input batches. MLP stands for Perceptron with shared parameters. Input features F Global max pooling, This indicates element-wise multiplication; , For the input feature map, For input features The output feature map after applying channel attention weights represents the channel-dimensional weighted enhancement of the original features.

[0027] Specifically, the improved CBAM attention mechanism used in S2 is as follows: CBAM is a concise and efficient attention unit in feedforward convolutional neural networks. It performs attention inference on the feature maps of intermediate layers of the network and applies the obtained weights to the original features to achieve adaptive feature enhancement and selection. As a lightweight and general-purpose functional component, CBAM can be easily embedded into various convolutional network structures, supports end-to-end training, and has good scalability and adaptability. The CBAM attention mechanism consists of two parts: channel attention and spatial attention. The channel attention part first performs global average pooling and global max pooling on the input feature map to obtain feature descriptions of size 1×1×C. Then, the two types of description vectors are input into a shared multilayer perceptron for processing, and the channel weights are obtained through the sigmoid activation function. And the channel weights are combined with the input features F Element-wise multiplication yields the channel-enhanced feature map. , The result is obtained through channel attention calculation. It is an intermediate feature representation after adaptive weighting in the channel dimension. It is used to highlight important feature channels related to the heating defects of power equipment and suppress redundant channel information that contributes little to defect identification, thereby improving the discriminative ability of features. This is not the final output, but rather the input feature for subsequent feature processing. It is further fed into the spatial attention module for weighted optimization at different spatial locations, thereby forming a high-level feature representation containing rich semantic information for subsequent object detection and defect determination. The specific formula is as follows:

[0028] , , in, This represents the obtained channel weights. Let S denote the Sigmoid activation function, MLP denotes a perceptron with shared parameters, and GAP denotes global max pooling. This indicates element-wise multiplication; Since the importance of information varies across channels, to enhance the influence of channels with important features, batch normalization (BN) is performed on each channel in the channel attention part. This is achieved by calculating the mean of the batch. ,variance Furthermore, learnable scaling and translation parameters are used to batch normalize channel information, thereby reflecting the importance of each channel, strengthening the influence of important channels on attention weight values, and weakening the influence of unimportant channels on attention weight values. The specific formula is as follows: , , , , , in, m This represents the number of samples in the current batch. for, These are the normalized eigenvalues. To prevent small constants from being divided by zero, we take 0.0001. This is a scaling parameter that can be learned during training, initially set to 1, and will change continuously during training. These are the translation parameters that can be learned during training, initially set to 0, and changing as training progresses. For a batch of feature tensors, The output tensor is obtained after batch normalization of the input batch X; Therefore, the improved formula for the channel attention part is expressed as follows: , , In the spatial attention part, the input features are first subjected to global average pooling and global max pooling. Then, the two pooling results are concatenated. The concatenated feature map is then fed into the convolutional layer and passed through the Sigmoid activation function to obtain the spatial weight coefficients. and spatial weight coefficients Input features Multiplication is used to perform spatial attention calculations, and the specific formula is as follows: , , Here, Concat represents the concatenation operation. For features The weight coefficients obtained from spatial attention calculation, The output feature map after spatial attention enhancement; In one embodiment of the present invention, the improved CBAM model structure is as follows: Figure 2 As shown, the overall model structure is as follows Figure 3 As shown; the hyperparameters for the training process are set as follows: the number of epochs is 500, the optimizer is SGD optimizer, the initial learning rate is set to 0.01, and the batch size is set to 16.

[0029] The experimental environment configuration used in the examples is shown in the table below.

[0030] Table 1 Experimental Environment Configuration Table In this embodiment, the optimal weights best.pt obtained during training are used for detection on the test set. The resulting detection accuracy is 95.6%, and the recall is 94.1%.

[0031] Optionally, obtaining the main area of ​​the device and its corresponding temperature includes the following steps: Otsu's temperature-weighted method is used to segment the device region of the detection box and an opening operation is performed to obtain the main device region mask. The device region segmentation uses Otsu's temperature-weighted method to divide the image into foreground and background, represented by mask 0 and 1. The foreground is the device, and the background is the area within the detection box other than the main device, including the environment and other non-target device areas. Based on the mask of the main body area of ​​the equipment, read the temperature value of the main body of the equipment where the mask is 0 (meaning it is the equipment). The main body of the equipment where the mask is 0 is the main body area of ​​the equipment.

[0032] Optionally, obtaining the device body region mask includes the following steps: When obtaining the mask of the main area of ​​the device, the gray value of each pixel in the detection box is first obtained and the gray value is mapped to the range of 0-255. Considering that the heating defects of power equipment in infrared images are usually manifested as local high temperature areas, in order to enhance the influence of high temperature areas in the threshold segmentation process, a temperature weighting mechanism is introduced to give higher weights to pixels with higher temperatures in the detection box, thereby forming a temperature-weighted gray statistical distribution. Based on the temperature-weighted gray-level statistical distribution, the Otsu method is used to perform threshold segmentation of pixels within the detection box for foreground and background. The Otsu method divides pixels into foreground candidate regions and background candidate regions by traversing all gray-level thresholds in the gray-level statistical distribution, and calculates the pixel proportion and gray-level mean of the two types of regions respectively. Then, the inter-class variance between the foreground candidate regions and the background candidate regions is calculated. When the inter-class variance corresponding to a certain gray-level threshold reaches the maximum, it is considered that the threshold can most effectively distinguish the foreground region and the background region, and the gray-level value is determined as the optimal threshold. After obtaining the optimal threshold, pixels with gray values ​​less than the optimal threshold within the detection box are marked as the main body area of ​​the device, and pixels with gray values ​​greater than or equal to the optimal threshold are marked as the background area, thus generating an initial mask image. The initial mask image is then subjected to opening denoising to eliminate isolated noise points, resulting in the final mask for the main device region.

[0033] Optional, optimal threshold We obtain it from the following formula: , in, This represents the probability that a pixel is in the foreground. This represents the probability that a pixel is part of the background. The mean gray level of the foreground. This represents the average grayscale value of the background.

[0034] In S3, the specific steps for obtaining the main area and temperature of the device are as follows: S3.1: Use temperature-weighted Otsu's method to perform device region segmentation on the detection box obtained in step 2 (the device region segmentation uses temperature-weighted Otsu's method to divide the image into foreground and background, represented by masks 0 and 1, with the foreground being the device) and use opening operation to obtain the main region mask of the device. S3.2: Based on the mask of the main body area of ​​the equipment, read the temperature value of the main body of the equipment where the mask is 0 (meaning it is the equipment).

[0035] In S3.1, the specific method for obtaining the main area of ​​the device is as follows: Otsu's method is an adaptive thresholding image segmentation method that, based on the grayscale characteristics of an image, finds the optimal threshold for dividing the grayscale level of each pixel in the image. T Divided into foreground F and background B Two parts; In infrared image scenes, heat defects are typically detected only in foreground devices, and overlap between foreground and background devices is common. Therefore, a weighting mechanism for high-temperature regions is introduced, applying temperature weighting to pixels with higher grayscale values. The formula is as follows: , , , , in, The grayscale value of the pixels within the temperature-weighted detection frame is... The probability, For the detection frame with grayscale values ​​of The number of points, e The grayscale value of the pixels within the detection frame after temperature weighting The probability of.

[0036] The grayscale value within each detection box is divided into 0-255, and the grayscale value in [0, The area within the range is defined as the foreground. , The pixels within the specified range are defined as the background. Based on the aforementioned temperature-weighted pixel grayscale values ​​within the detection frame... The probability that a pixel is in the foreground. The probabilities of a pixel being the background are as follows: , , Otsu's method searches for the optimal threshold that maximizes the inter-class variance by iterating through grayscale values ​​from 0 to 255. This distinguishes between foreground and background. [ ] represents the foreground grayscale value range. For the background grayscale value range, the specific optimal threshold selection method is as follows: definition The mean gray level of the foreground. The average gray level of the background. The overall grayscale average value. The inter-class variances for the foreground and background are as follows: , , , , Then, the optimal threshold is selected. for: , After finding the optimal threshold, if the pixel grayscale value is less than the threshold, the mask of that pixel is assigned to 0; otherwise, the mask is assigned to 1. The obtained mask image is then subjected to opening operation to reduce noise and obtain the final device area; In one embodiment of the present invention, the result of S3 performing device body acquisition and opening operation denoising on the detection frame is as follows: Figure 4 As shown.

[0037] Optionally, the process of performing current-induced thermal defect detection on the equipment within the detection frame includes the following steps: First, obtain the temperature of the heating point in the device area. That is, the highest temperature ; Based on the average temperature of the equipment area Temperature on the day of shooting And the temperature of the heating point in the equipment area, to obtain the relative temperature difference of that heating point. : ; Subsequently, based on the equipment type and heating point temperature Relative temperature difference of heating point According to the power industry standard DL / T664-2016, determine whether the equipment has a current-induced heating defect.

[0038] In one embodiment of the present invention, the result of current-induced heating detection of the device is as follows: Figure 5 As shown, the test results are marked with the maximum temperature point within the test frame and the relative temperature difference at that point, and the defect level is distinguished by the color of the test frame.

[0039] Optionally, performing voltage-induced thermal defect detection on the three-phase equipment within the detection frame includes the following steps: Obtain the longer geometric center line of each detection frame. The OBB detection frame is a rectangle with a long side and a short side. The center line on the short side is the long center line. Read the temperature distribution in the main body area of ​​the device on the center line to obtain three temperature center lines. During the alignment process, add a deformation penalty term when calculating the distance between two points to obtain the optimized distance between two points. The three temperature midlines are aligned using an improved DTW (Dynamic Time Warping) alignment method. This improved DTW alignment method includes: During the alignment process, to reduce the temperature error caused by the different shooting distances of the three devices, the temperature sequences of the three center lines were first... Perform amplitude normalization: , in, This represents the average temperature of the k-th temperature midline. For the first k Standard deviation of the temperature midline; Subsequently, a two-stage alignment process was used to align the three temperature midlines to the shortest temperature midline length. The specific alignment method was as follows: For sequences and sequence Construct a Distance matrix, matrix elements express The points and The Align the points and find a path through a set of grid points in the matrix, representing the sequence. and The matching relationship for each point is as follows, and different paths correspond to different matching relationships; the DTW algorithm is used to find a path containing... Optimal path for two binary arrays This makes the sum of the distances of all matching point pairs on the path. Minimum alignment is achieved, as detailed below: , , , , , , , in, The distance between two points in the sequence; To avoid deformation during alignment due to excessive differences in temperature centerline lengths, a deformation penalty term is added when calculating the distance between two points, resulting in the optimized formula for the distance between the two points as follows: , in, , where is the deformation penalty coefficient.

[0040] By comparing the aligned temperature centerlines in segments, and according to the power industry standard DL / T 664-2016, the temperature difference at different corresponding positions of the equipment is compared to determine whether the equipment has a voltage-induced heating defect.

[0041] In one embodiment of the present invention, the temperature centerline alignment result is shown in Figure 6, and the voltage-induced heating defect detection result of the three-phase equipment is shown in Figure 7.

[0042] A second aspect of the present invention provides a power equipment heating defect detection system based on infrared images, comprising: The image acquisition module is used to acquire infrared image datasets of power equipment in substation scenarios. The device identification and localization module is built on a deep learning model that incorporates an improved CBAM attention mechanism module. It is used to identify and detect the input infrared image and output the category and location information of the power equipment in a rotated detection box format. The improved CBAM attention mechanism module performs batch normalization on each channel in its channel attention part. The equipment body area and temperature extraction module is used to extract the corresponding equipment body area and its corresponding temperature according to the rotating detection frame, and use it as the average operating temperature of the equipment. The current-induced heating type defect detection module is used to analyze and judge the current-induced heating type defects of the equipment within the detection frame based on the temperature of the main body area of ​​the equipment as the normal operation benchmark. The voltage-induced heating defect detection module is used to analyze and judge voltage-induced heating defects in objects identified as three-phase equipment within the detection frame.

[0043] This invention provides a method and system for detecting thermal defects in power equipment based on infrared images. This method and system can improve the efficiency of power equipment maintenance, accurately identify thermal problems in equipment, reduce equipment failure losses, improve operational accuracy, save time and costs, facilitate convenient monitoring, and thus enhance the safety and reliability of the power system. It has significant application value for the intelligent detection and maintenance of power equipment.

[0044] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. An infrared image-oriented power equipment heat defect detection method, characterized by, Includes the following steps: Acquire infrared image datasets of power equipment in a substation scenario; Using the infrared image dataset, based on a deep learning model that incorporates an improved CBAM attention mechanism module, electrical equipment in the infrared images is identified and detected, and the equipment category and location are obtained in a rotated detection box format; wherein, the improved CBAM attention mechanism module performs batch normalization operation on each channel in the channel attention part of the CBAM attention mechanism. Based on the obtained detection results, the main body area of ​​the device and its corresponding temperature are obtained as the average temperature of the device. The temperature of the main body area of ​​the equipment is taken as the temperature during normal operation of the equipment, and the equipment within the detection frame is subjected to current-induced heating type defect detection. Voltage-induced thermal defect detection is then performed on the three-phase equipment within the detection frame.

2. The method for detecting heating defects in power equipment based on infrared images as described in claim 1, characterized in that, The deep learning model that incorporates the improved CBAM attention mechanism module includes: a backbone network, a neck network, and an output end; The backbone network includes a C2f module, a convolutional module, an improved CBAM attention mechanism module, and a spatial pyramid pooling module. The C2f module and the convolutional module extract and reconstruct features of power equipment in infrared images, gradually obtaining feature representations of power equipment from local to global perspectives to obtain feature maps. The improved CBAM attention mechanism module optimizes the feature maps to obtain optimized feature maps. The spatial pyramid pooling module performs multi-scale pooling operations on the optimized feature maps to obtain high-level feature representations. The neck network performs bidirectional fusion of high-level feature representations from top to bottom and bottom to top to obtain a fused feature map; The output terminal is used to output the precise location and category information of the power equipment in the form of a rotating detection box from the fused feature map.

3. The method for detecting heating defects in power equipment based on infrared images as described in claim 1, characterized in that, The improved CBAM attention mechanism includes a channel attention component and a spatial attention component; In the channel attention part, batch normalization is performed on each channel. By calculating the mean and variance in the batch and using learnable scaling and translation parameters, the channel information is batch normalized to reflect the importance of each channel, strengthen the influence of important channels on the attention weight value, and weaken the influence of unimportant channels on the attention weight value.

4. The method for detecting heating defects in power equipment based on infrared images as described in claim 3, characterized in that, The improved channel attention component is expressed as follows: ， in, Input features F Channel weights, It is the Sigmoid activation function. This represents the output tensor obtained after batch normalization of the input batches. MLP stands for Perceptron with shared parameters. Input features F Global max pooling, This indicates element-wise multiplication; ， For the input feature map, For input features Output feature map after applying channel attention weights.

5. The method for detecting heating defects in power equipment based on infrared images as described in claim 1, characterized in that, The process of obtaining the main body area of ​​the device and the corresponding temperature includes the following steps: Otsu's temperature-weighted method is used to segment the device region of the detection frame and an opening operation is performed to obtain the main area mask of the device. Based on the mask of the main body area of ​​the equipment, read the temperature value of the main body of the equipment where the mask is 0. The main body of the equipment where the mask is 0 is the main body area of ​​the equipment.

6. The method for detecting heating defects in power equipment based on infrared images as described in claim 5, characterized in that, The process of obtaining the device main body region mask includes the following steps: First, the grayscale value of each pixel within the detection box is obtained and mapped to the range of 0–255. A temperature weighting mechanism is introduced to assign higher weights to pixels with higher temperatures within the detection box, thereby forming a temperature-weighted grayscale statistical distribution. The Otsu method divides pixels into foreground and background candidate regions by traversing all gray thresholds in the gray-level statistical distribution, and calculates the pixel proportion and gray mean of the two types of regions respectively. Then, it calculates the inter-class variance between the foreground and background candidate regions. When the inter-class variance corresponding to a certain gray threshold reaches the maximum, it is considered that the threshold can most effectively distinguish between the foreground and background regions, and the gray value is determined as the optimal threshold. Pixels with gray values ​​less than the optimal threshold within the detection box are marked as the main body area of ​​the device, and pixels with gray values ​​greater than or equal to the optimal threshold are marked as the background area, thus generating an initial mask image; The initial mask image is then subjected to opening and noise reduction processing to obtain the final device body region mask.

7. The method for detecting heating defects in power equipment based on infrared images as described in claim 6, characterized in that, The optimal threshold We obtain it from the following formula: ， in, This represents the probability that a pixel is in the foreground. This represents the probability that a pixel is part of the background. The mean gray level of the foreground. This represents the average grayscale value of the background.

8. The method for detecting heating defects in power equipment based on infrared images as described in claim 1, characterized in that, The method for performing current-induced thermal defect detection on the equipment within the detection frame includes the following steps: First, obtain the temperature of the hottest point in the equipment area, i.e., the highest temperature; The relative temperature difference of the heat source is obtained based on the average temperature of the equipment area, the air temperature on the day of shooting, and the temperature of the heat source in the equipment area. Based on the equipment type, heating point temperature, and relative temperature difference of the heating point, and in accordance with power industry standards, determine whether the equipment has a current-induced heating defect.

9. The method for detecting heating defects in power equipment based on infrared images as described in claim 1, characterized in that, The process of performing voltage-induced thermal defect detection on the three-phase equipment within the detection frame includes the following steps: Obtain the longer geometric center line of each detection frame, read the temperature distribution in the main body area of ​​the device on the center line, and obtain three temperature center lines. During the alignment process, add a deformation penalty term when calculating the distance between two points to obtain the optimized distance between two points. The improved DTW alignment method is used to align the three temperature midlines. The improved DTW alignment method includes: normalizing the amplitude of the three midline temperature sequences, and aligning the three temperature midlines to the shortest temperature midline length using two alignment steps. By comparing the aligned temperature centerlines in segments, and according to power industry standards, the temperature differences at different corresponding locations on the equipment are compared to determine whether the equipment has voltage-induced heating defects.

10. A power equipment heating defect detection system based on infrared images, characterized in that, include: The image acquisition module is used to acquire infrared image datasets of power equipment in substation scenarios. The device identification and localization module is built based on a deep learning model that incorporates an improved CBAM attention mechanism module. It is used to identify and detect the input infrared image and output the category and location information of the power equipment in a rotated detection box format. The improved CBAM attention mechanism module performs batch normalization operation on each channel in its channel attention part. The main body area and temperature extraction module is used to extract the corresponding main body area of ​​the equipment and its corresponding temperature according to the rotating detection frame, and use it as the average operating temperature of the equipment. The current-induced heating type defect detection module is used to perform current-induced heating type defect analysis and judgment on the equipment within the detection frame based on the temperature of the main body area of ​​the equipment as the normal operation benchmark. The voltage-induced heating defect detection module is used to analyze and judge voltage-induced heating defects in objects identified as three-phase equipment within the detection frame.

Images