Method for detecting overheating of cable termination bushings

Abstract

This application relates to a method for detecting overheating in cable termination sleeves. The method includes: acquiring the target area of ​​the cable termination sleeve in an infrared image, and dividing the target area into multiple sub-regions based on its spatial distribution; acquiring the temperature value of each sub-region, and calculating a first parameter reflecting the temperature difference within the target area based on the temperature value of each sub-region; if the first parameter exceeds a temperature difference threshold, determining a second parameter based on the temperature values ​​of multiple points within the target area; if the second parameter does not meet a preset uniform distribution condition, determining the classification result of the target area based on its corresponding temperature characteristics; if the classification result is a true overheating category, outputting a detection result indicating that the cable termination sleeve has an overheating defect; if the classification result is an artifact category, outputting a detection result indicating that the cable termination sleeve does not have an overheating defect. This method can reduce the false detection rate.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a method for detecting the overheating of cable terminal bushings. Background Technology

[0002] With the development of intelligent operation and maintenance technology for power systems, infrared thermal imaging technology has become the main means of detecting overheating defects in cable terminal bushings of transmission lines due to its advantages such as non-contact temperature measurement and full-field visibility. Traditional methods of detecting overheating defects usually rely on manual visual interpretation of infrared images or on determining overheating based on a single temperature threshold.

[0003] In traditional techniques, the overall outline of the cable terminal bushing is first extracted using an image segmentation model. Then, the highest and lowest global temperatures within the outline area are statistically analyzed, the relative temperature rise is calculated, and compared with a preset threshold to determine whether there is a heating defect.

[0004] However, the traditional method has a high false alarm rate, which means that maintenance personnel need to spend a lot of time manually reviewing the alarm results. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for detecting the heat of cable terminal bushings that can reduce the false detection rate, in order to address the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for detecting the overheating of a cable termination bushing, including:

[0007] The target area of ​​the cable terminal sleeve in the infrared image is obtained, and the target area is divided into multiple sub-regions according to the spatial distribution of the target area;

[0008] The temperature values ​​of each sub-region are obtained separately, and the first parameter reflecting the temperature difference within the target region is calculated based on the temperature values ​​of each sub-region.

[0009] If the first parameter exceeds the temperature difference threshold, the second parameter is determined based on the temperature values ​​at multiple locations within the target area; the second parameter is used to characterize the uniformity of temperature distribution within the target area.

[0010] If the second parameter does not meet the preset uniform distribution condition, the classification result of the target region is determined based on the temperature characteristics of the target region.

[0011] If the classification result is a true heat generation category, the output cable terminal bushing will show a detection result indicating a heat generation defect; if the classification result is an artifact category, the output cable terminal bushing will show a detection result indicating no heat generation defect.

[0012] Secondly, this application also provides a heating detection device for cable termination bushings, comprising:

[0013] The segmentation module is used to obtain the target area of ​​the cable terminal sleeve in the infrared image and divide the target area into multiple sub-regions according to the spatial distribution of the target area;

[0014] The acquisition module is used to acquire the temperature values ​​of each sub-region and calculate the first parameter reflecting the temperature difference within the target region based on the temperature values ​​of each sub-region.

[0015] The first determining module is used to determine the second parameter based on the temperature values ​​of multiple locations within the target area when the first parameter exceeds the temperature difference threshold; the second parameter is used to characterize the uniformity of temperature distribution within the target area.

[0016] The second determining module is used to determine the classification result of the target region based on the temperature characteristics of the target region when the second parameter does not meet the preset uniform distribution condition.

[0017] The output module is used to output the detection result of the cable terminal bushing having a heating defect when the classification result is a true heating category; and to output the detection result of the cable terminal bushing not having a heating defect when the classification result is an artifact category.

[0018] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0019] The target area of ​​the cable terminal sleeve in the infrared image is obtained, and the target area is divided into multiple sub-regions according to the spatial distribution of the target area;

[0020] The temperature values ​​of each sub-region are obtained separately, and the first parameter reflecting the temperature difference within the target region is calculated based on the temperature values ​​of each sub-region.

[0021] If the first parameter exceeds the temperature difference threshold, the second parameter is determined based on the temperature values ​​at multiple locations within the target area; the second parameter is used to characterize the uniformity of temperature distribution within the target area.

[0022] If the second parameter does not meet the preset uniform distribution condition, the classification result of the target region is determined based on the temperature characteristics of the target region.

[0023] If the classification result is a true heat generation category, the output cable terminal bushing will show a detection result indicating a heat generation defect; if the classification result is an artifact category, the output cable terminal bushing will show a detection result indicating no heat generation defect.

[0024] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0025] The target area of ​​the cable terminal sleeve in the infrared image is obtained, and the target area is divided into multiple sub-regions according to the spatial distribution of the target area;

[0026] The temperature values ​​of each sub-region are obtained separately, and the first parameter reflecting the temperature difference within the target region is calculated based on the temperature values ​​of each sub-region.

[0027] If the first parameter exceeds the temperature difference threshold, the second parameter is determined based on the temperature values ​​at multiple locations within the target area; the second parameter is used to characterize the uniformity of temperature distribution within the target area.

[0028] If the second parameter does not meet the preset uniform distribution condition, the classification result of the target region is determined based on the temperature characteristics of the target region.

[0029] If the classification result is a true heat generation category, the output cable terminal bushing will show a detection result indicating a heat generation defect; if the classification result is an artifact category, the output cable terminal bushing will show a detection result indicating no heat generation defect.

[0030] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0031] The target area of ​​the cable terminal sleeve in the infrared image is obtained, and the target area is divided into multiple sub-regions according to the spatial distribution of the target area;

[0032] The temperature values ​​of each sub-region are obtained separately, and the first parameter reflecting the temperature difference within the target region is calculated based on the temperature values ​​of each sub-region.

[0033] If the first parameter exceeds the temperature difference threshold, the second parameter is determined based on the temperature values ​​at multiple locations within the target area; the second parameter is used to characterize the uniformity of temperature distribution within the target area.

[0034] If the second parameter does not meet the preset uniform distribution condition, the classification result of the target region is determined based on the temperature characteristics of the target region.

[0035] If the classification result is a true heat generation category, the output cable terminal bushing will show a detection result indicating a heat generation defect; if the classification result is an artifact category, the output cable terminal bushing will show a detection result indicating no heat generation defect.

[0036] The aforementioned method for detecting overheating in cable terminal bushings divides the target area into multiple sub-regions and calculates a first parameter reflecting the temperature difference within the target area by acquiring the temperature values ​​of each sub-region. This enables refined analysis of the temperature gradient at different locations along the longitudinal direction of the cable terminal bushing, capturing early localized overheating anomalies that are easily masked by traditional global temperature statistics methods. Furthermore, this application only triggers subsequent judgment steps when the first parameter exceeds a temperature difference threshold. The cascaded architecture avoids wasting computational resources on processing normal samples, improving detection efficiency. When the first parameter exceeds the temperature difference threshold, this application further determines a second parameter characterizing the uniformity of temperature distribution based on the temperature values ​​of multiple points within the target area. It utilizes the principle that physical reflection presents a relatively uniform temperature distribution in infrared images, while actual overheating presents a locally concentrated temperature distribution for rapid filtering. Only when the second parameter does not meet the preset uniform distribution condition is the classification model activated to determine the classification result. This dual-stage filtering mechanism, combining uniformity filtering and classification model classification, distinguishes between true overheating and artifact categories in samples with uneven temperature distribution in the second stage, effectively reducing the false alarm rate and the review burden on maintenance personnel. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is an internal structural diagram of a computer device in one embodiment;

[0039] Figure 2 This is a flowchart illustrating a method for detecting the heat generation of a cable terminal bushing in one embodiment.

[0040] Figure 3 This is a flowchart illustrating the steps for determining the target region in one embodiment;

[0041] Figure 4 This is a flowchart illustrating the steps for determining the joint loss in one embodiment;

[0042] Figure 5 This is a flowchart illustrating the step of determining the second parameter in one embodiment;

[0043] Figure 6 This is a flowchart illustrating the steps for determining the classification result in one embodiment;

[0044] Figure 7This is a flowchart illustrating a method for detecting heat generation in a cable terminal bushing, as described in another embodiment.

[0045] Figure 8 This is a schematic diagram of an infrared image of a cable terminal bushing in one embodiment;

[0046] Figure 9 This is a schematic diagram of the infrared temperature measurement display interface of the cable terminal bushing in one embodiment;

[0047] Figure 10 This is a schematic diagram of the extraction and reflection filtering of the center line sub-region in one embodiment;

[0048] Figure 11 This is a structural block diagram of a heating detection device for a cable terminal bushing in one embodiment. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 1 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores relevant data during the heat detection process of cable terminal bushings. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for heat detection of cable terminal bushings.

[0051] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0052] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting the overheating of a cable termination bushing is provided, which can be applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 201 to 205. Wherein:

[0053] Step 201: Obtain the target area of ​​the cable terminal sleeve in the infrared image, and divide the target area into multiple sub-regions according to the spatial distribution of the target area.

[0054] The target region refers to the pixel area occupied by the cable terminal bushing in the infrared image, used to characterize the position and extent of the bushing body in the image space. Subsequent temperature analysis and heat detection are performed based on the temperature data within this region. The infrared image refers to the temperature distribution image of the cable terminal bushing acquired by an infrared thermal imaging device. The grayscale value of each pixel in the image corresponds to the radiation temperature of that point on the surface of the object being measured. The sub-region refers to multiple continuous region segments obtained by dividing the target region along its longitudinal direction, used for zoned temperature analysis at different height positions of the bushing.

[0055] In this embodiment, the server acquires pre-stored or real-time acquired infrared images of cable terminal sleeves and uses a preset region extraction algorithm to extract the target region where the sleeve is located from the infrared images. The server determines the longitudinal extension direction of the target region based on its geometry and divides the target region into multiple sub-regions along the longitudinal direction. The server coordinates the spatial location information of each sub-region with the temperature matrix of the infrared image for subsequent temperature value extraction.

[0056] In another embodiment, after acquiring the infrared image, the server extracts the sheath region using an adaptive threshold segmentation method based on the grayscale distribution characteristics of the sheath region in the infrared image, and uses the extracted region as the target region. The server then divides the target region into multiple sub-regions of unequal lengths along the longitudinal direction based on the height information of the target region.

[0057] In another embodiment, the server acquires an infrared image sequence, extracts the target region for each frame of the infrared image in the sequence and divides it into multiple sub-regions, and performs time-series filtering on the temperature values ​​of the sub-regions at the corresponding positions of each frame of the image to serve as temperature data for subsequent steps.

[0058] Step 202: Obtain the temperature value of each sub-region, and calculate the first parameter reflecting the temperature difference within the target region based on the temperature value of each sub-region.

[0059] Here, the temperature value refers to the radiation temperature data corresponding to the pixel location of each sub-region, extracted from the temperature matrix of the infrared image. The temperature matrix is ​​a two-dimensional array obtained after decoding the infrared image, where each element represents the temperature value corresponding to that pixel. The first parameter is a quantitative indicator calculated based on the temperature values ​​of each sub-region, used to characterize the degree of temperature difference between different locations within the target area.

[0060] In this embodiment, the server extracts the temperature data corresponding to each sub-region from the temperature matrix of the infrared image based on the pixel coordinate range of each sub-region determined in step 201. The server counts the highest temperature value in each sub-region, compares the highest temperature values ​​of each sub-region pairwise, and calculates the maximum difference between the highest temperatures as the first parameter.

[0061] In another embodiment, the server calculates the average temperature within each sub-region, using the maximum difference between the average temperatures of each sub-region as the first parameter.

[0062] In another embodiment, the server calculates the median temperature value in each sub-region, uses the maximum difference between the median temperature values ​​in each sub-region as the first parameter, and calculates the standard deviation of the temperature values ​​in each sub-region. The maximum value of the standard deviation is used as the confidence check item of the first parameter. If the maximum value of the standard deviation exceeds a preset range, a re-detection is triggered.

[0063] Step 203: If the first parameter exceeds the temperature difference threshold, determine the second parameter based on the temperature values ​​of multiple locations within the target area; the second parameter is used to characterize the uniformity of temperature distribution within the target area.

[0064] The second parameter refers to a quantitative index calculated based on temperature values ​​from multiple locations within the target area, used to characterize the uniformity of temperature distribution within the target area. Temperature distribution uniformity refers to the spatial dispersion of temperature values ​​at different locations within the target area. When temperature anomalies are caused by optical reflection, the temperature values ​​at each location within the target area exhibit a relatively uniform distribution; when actual heating exists, the temperature values ​​show local concentration and uneven spatial distribution. Location points refer to spatial locations within the target area selected according to preset rules for temperature sampling.

[0065] In this embodiment, the server selects multiple locations within the target area and acquires the temperature value at each location. Based on the temperature values ​​at each location, it calculates a dispersion index of the temperature values ​​as a second parameter. The server compares the second parameter with a preset uniform distribution condition. If the second parameter satisfies the uniform distribution condition, it determines that the current temperature anomaly is caused by optical reflection; otherwise, it proceeds to the next step.

[0066] In another embodiment, the server selects multiple sampling locations along a preset direction within the target area, obtains the temperature value of each sampling location, calculates the standard deviation or variance of the temperature value of each sampling location as a second parameter, compares the second parameter with a preset standard deviation threshold, and determines that the temperature distribution is uniform if the second parameter is less than or equal to the threshold.

[0067] In another embodiment, the server randomly selects multiple location points within the target area, obtains the temperature value of each location point, calculates the range or interquartile range of each temperature value as a second parameter, and compares the second parameter with a preset range threshold or interquartile range threshold.

[0068] Step 204: If the second parameter does not meet the preset uniform distribution condition, determine the classification result of the target region based on the temperature characteristics corresponding to the target region.

[0069] The classification result refers to the category label output after the target area is identified by the classification model. The category label includes at least two categories: true heat generation and artifact generation. Temperature features refer to numerical features extracted from the temperature data of the target area to characterize the temperature distribution pattern. The classification model refers to a pre-trained machine learning or deep learning model used to map input data to corresponding category labels.

[0070] In this embodiment, after determining that the second parameter does not meet the preset uniform distribution condition, the server extracts multiple temperature statistics from the temperature matrix of the target region as temperature features. The extracted temperature statistics include at least one of the following: mean temperature, temperature variance, maximum temperature, minimum temperature, median temperature, kurtosis, skewness, mean temperature gradient magnitude, and variance of temperature gradient magnitude for all pixels within the target region. The server combines the extracted temperature statistics into a feature vector and inputs this feature vector into a pre-trained classification model. The classification model uses a support vector machine or random forest, and calculates the input feature vector using its internal classification decision function, outputting the probability values ​​of the target region belonging to the true heat category and the probability values ​​of it belonging to the artifact category. The server compares the two probability values ​​and determines the category with the larger probability value as the classification result for the target region.

[0071] In another embodiment, the server extracts temperature features from the temperature matrix of the target region. These features include the temperature difference between the highest and lowest temperatures within the target region, the maximum difference between the average temperatures of each sub-region within the target region, and the proportion of pixels within the target region whose temperature exceeds a preset high-temperature threshold. The server inputs the extracted temperature features into a pre-trained multilayer perceptron classification model. The multilayer perceptron receives feature vectors through its input layer, performs nonlinear transformations on the features through at least one hidden layer, and outputs the classification probability of the target region belonging to either the true heat category or the artifact category through its output layer and a Softmax activation function. The server determines the classification result based on the classification probability.

[0072] In another embodiment, the server extracts temperature gradient features from the temperature matrix of the target region as temperature features. The server calculates the temperature gradient magnitude of each pixel in the target region in the horizontal and vertical directions, and calculates the mean, maximum, and variance of the gradient magnitude. The mean, maximum, and variance of the gradient magnitude are then input into the classification model as temperature features.

[0073] In another embodiment, the server divides the temperature matrix of the target area into multiple grid blocks according to a preset grid size, calculates the average temperature value in each grid block, arranges the average temperature values ​​of all grid blocks in spatial order to form a temperature feature vector, and inputs the temperature feature vector into the classification model.

[0074] Step 205: If the classification result is a true heating category, output the detection result that the cable terminal bushing has a heating defect; if the classification result is an artifact category, output the detection result that the cable terminal bushing does not have a heating defect.

[0075] The test results refer to the conclusions regarding the detection of thermal defects in the cable terminal bushing, indicating whether a genuine thermal defect exists in the target bushing. A thermal defect refers to an abnormal increase in localized temperature within the cable terminal bushing due to factors such as internal insulation deterioration, increased contact resistance, or sealing failure.

[0076] In this embodiment, the server obtains the classification result determined in step 204. If the classification result is a true heat generation category, the server generates a detection result indicating that the cable terminal bushing has a heat generation defect, and outputs the detection result to the terminal device of the maintenance personnel or uploads it to the power grid equipment status monitoring platform. If the classification result is an artifact category, the server generates a detection result indicating that the cable terminal bushing does not have a heat generation defect, and records the detection result in the detection log.

[0077] In another embodiment, if the classification result is a true heat generation category, the server further determines the heat generation severity level based on the temperature data of the target area. The heat generation severity level includes at least one of general defect, major defect, and emergency defect. The server determines the corresponding heat generation severity level based on the temperature range where the highest temperature value is located and / or the temperature rise range where the relative temperature rise value is located, and outputs the heat generation severity level along with the detection result of the presence of a heat generation defect.

[0078] In another embodiment, if the classification result is a true heat category, the server generates an alarm message and pushes the alarm message to the corresponding maintenance personnel via SMS, instant messaging tools, or email to trigger manual review and maintenance scheduling. If the classification result is an artifact category, the server stores the current infrared image and the classification result of the target area in an artifact sample library for incremental training of the subsequent classification model.

[0079] In another embodiment, when outputting the detection results, the server also outputs the classification probability value output by the classification model as the classification confidence level, as well as a temperature distribution visualization map of the target area, wherein the temperature distribution visualization map uses different colors to mark the temperature level of each sub-region.

[0080] In the aforementioned method for detecting the heating of cable terminal bushings, the target area is divided into multiple sub-regions, and the temperature value of each sub-region is obtained to calculate a first parameter reflecting the temperature difference within the target area. This enables refined analysis of the temperature gradient at different locations along the longitudinal direction of the cable terminal bushing, and can capture local early heating anomalies that are easily masked by traditional global temperature statistics methods. Furthermore, this application only triggers subsequent judgment steps when the first parameter exceeds a temperature difference threshold. The cascaded architecture avoids wasting computational resources on processing normal samples, thus improving detection efficiency. When the first parameter exceeds the temperature difference threshold, this application further determines a second parameter characterizing the uniformity of temperature distribution based on the temperature values ​​of multiple points within the target area. It utilizes the principle that physical reflection presents a relatively uniform temperature distribution in infrared images, while actual heating presents a locally concentrated temperature distribution for rapid filtering. Only when the second parameter does not meet the preset uniform distribution condition is the classification model activated to determine the classification result. This dual-stage filtering mechanism, consisting of uniformity filtering and classification model classification, distinguishes between true heating and artifact categories for samples with uneven temperature distribution in the second stage, effectively reducing the false alarm rate and the review burden on maintenance personnel.

[0081] In one exemplary embodiment, such as Figure 3 As shown, the above-mentioned "obtaining the target area of ​​the cable terminal sleeve in the infrared image" includes steps 301 to 303. Wherein:

[0082] Step 301: Perform instance segmentation processing on the infrared image to obtain the initial segmentation contour of the cable terminal sleeve.

[0083] Infrared image refers to the temperature distribution image of the cable terminal sleeve acquired by infrared thermal imaging equipment. The gray value of each pixel in the image corresponds to the radiation temperature of that point on the surface of the object being measured. Instance segmentation refers to the processing technique of detecting each individual instance of the target object in the image and performing precise segmentation of that instance at the pixel level. Its output is a pixel-level segmentation mask corresponding to each instance. Initial segmentation contour refers to the original polygonal contour or binary mask boundary of the cable terminal sleeve region segmented from the infrared image by the instance segmentation model, used to characterize the initial occupancy range of the sleeve in the image.

[0084] In this embodiment, the server inputs the acquired infrared image of the cable terminal sleeve into a pre-trained instance segmentation model. The instance segmentation model performs feature extraction and target recognition on the infrared image, identifies the cable terminal sleeve instance in the image, and predicts the category of the instance pixel by pixel, generating a binary mask corresponding to the instance. The server extracts the contour boundary points of the binary mask to form an initial polygonal contour composed of multiple vertices, which serves as the initial segmentation contour of the cable terminal sleeve.

[0085] In another embodiment, the server inputs the infrared image into the joint model of target detection and instance segmentation. The model first detects the target bounding box of the cable terminal sleeve through the target detection branch, and then performs pixel-level segmentation within the bounding box through the segmentation branch, outputting the instance segmentation mask of the sleeve. The server extracts the mask boundary as the initial segmentation contour.

[0086] In another embodiment, the server inputs an infrared image into a lightweight real-time instance segmentation model. The model uses a single-stage architecture to output an instance segmentation mask while maintaining high inference speed. The server extracts the mask boundaries as the initial segmentation contour.

[0087] Step 302: Perform proportional inward shrinkage on the initial segmented contour to obtain the shrunken contour, and determine the minimum bounding rectangle of the shrunken contour.

[0088] In this context, proportional inward shrinkage refers to moving each edge or the entire polygon of the initial segmented contour inward along its normal direction, thus reducing the overall size of the contour proportionally. The shrunken contour refers to the new polygonal contour obtained after proportional inward shrinkage. The minimum bounding rectangle is the rectangle that can enclose the shrunken contour with the smallest area; the sides of this rectangle are parallel to the image coordinate axes or aligned with the principal axis of the sleeve.

[0089] In this embodiment, the server uses a contour shrinkage algorithm to shrink the initial segmented contour inward proportionally. The vertex coordinates of the initial segmented contour are moved towards the geometric center of the contour according to a preset shrinkage ratio to generate a new shrunken contour. The server determines the bounding rectangle of the shrunken contour, calculates the area of ​​the bounding rectangle, and finds the bounding rectangle with the smallest area by rotating the bounding rectangle and comparing the areas. This bounding rectangle is then used as the minimum bounding rectangle.

[0090] In another embodiment, after the server converts the initial segmentation contour into a binary mask, it performs a morphological erosion operation to erode the mask proportionally to obtain a shrunken mask region. The contour of the shrunken mask is extracted as the shrunken contour, and then the minimum bounding rectangle of the shrunken contour is determined.

[0091] In another embodiment, the server transforms the coordinates of each vertex of the initial segmented contour to a polar coordinate system, reduces the polar radius of each vertex proportionally in the polar coordinate system, and then transforms it back to a rectangular coordinate system to obtain the shrunken contour, and then determines its minimum bounding rectangle.

[0092] Step 303: Extract the central region within the smallest bounding rectangle and use the central region as the target region; wherein, the ratio of the horizontal dimension of the central region to the width of the smallest bounding rectangle is a preset ratio, and the ratio of the vertical dimension of the central region to the height of the smallest bounding rectangle is a preset ratio.

[0093] The central region refers to a rectangular sub-region within the smallest bounding rectangle, centered on its geometric center and with dimensions determined by a certain ratio of the rectangle's width and height. This sub-region serves as the final temperature analysis area. The preset ratio refers to a pre-defined proportion relative to the size of the central region and the smallest bounding rectangle; this ratio is greater than 0 and less than 1. The target region refers to the final determined core area of ​​the casing used for subsequent temperature analysis and heat detection.

[0094] In this embodiment, the server determines the coordinates of the geometric center point of the minimum bounding rectangle, calculates the width and height of the central region according to a preset ratio, determines the boundary range of the central region with the geometric center point as the center and the calculated width and height values, and extracts the central region as the target region within the minimum bounding rectangle.

[0095] In another embodiment, the server obtains a preset ratio value that is dynamically adjusted based on the sleeve size, infrared image resolution, and historical detection experience. The server then uses the geometric center of the smallest bounding rectangle as the center and the dynamic ratio value to determine the boundary range of the central region, and extracts the central region as the target region.

[0096] In another embodiment, the server extends the width of the central region horizontally by half a preset proportion to the left and right sides of the minimum bounding rectangle, with the geometric center as the center, and extends the height of the central region vertically by half a preset proportion to the upper and lower sides, with the geometric center as the center, and then extracts the central region as the target region.

[0097] In an exemplary embodiment, the above-mentioned "performing instance segmentation processing on the infrared image to obtain the initial segmentation contour of the cable terminal sleeve" includes:

[0098] An instance segmentation model is used to segment infrared images to obtain the initial segmentation contour of the cable terminal sleeve. The instance segmentation model is trained as follows: labeled samples are acquired, including the sampled image and its corresponding mask label; the sampled image is input into a visual encoder to extract multi-scale visual features; text prompts indicating the segmentation target are input into a text encoder to extract text features, and the multi-scale visual features and text features are fused to obtain fused features; the fused features are input into a decoder to generate a predicted mask for the segmented target in the sampled image; the joint loss is calculated based on the difference between the predicted mask and the mask label, and the parameters of the trained segmentation model are updated based on the joint loss; the above training steps are iteratively executed until the convergence condition is met, resulting in a trained instance segmentation model.

[0099] The instance segmentation model refers to a deep learning-based neural network model used to detect each target instance in an image and perform pixel-level segmentation. The infrared image refers to the temperature distribution image of the cable terminal sleeve acquired by an infrared thermal imaging device. The initial segmentation contour refers to the contour boundary of the pixel-level mask of the cable terminal sleeve region output by the instance segmentation model. Annotated samples refer to data samples used for model training, including sampled images and their corresponding mask annotations. Sampled images refer to sample images used for model training. Mask annotations refer to binary masks obtained after manually or semi-automatically annotating the cable terminal sleeve region in the sampled images, used as supervision signals during model training. The visual encoder refers to a neural network module used to extract visual features from an image, which progressively extracts multi-scale feature representations of the image from local to global through multi-layer convolutional operations. The text encoder refers to a neural network module used to extract semantic features from text prompts, which converts the input text description into a fixed-length vector representation. Multi-scale visual features refer to the set of feature maps with different spatial resolutions and semantic abstraction levels output by the visual encoder at different network layers. Text features refer to the semantic vectors extracted by the text encoder from the text prompts. Textual cues refer to textual descriptions used to indicate segmentation targets, such as "cable terminal sleeve." Fusion features refer to the joint feature representation obtained by fusing visual and textual features through attention mechanisms or concatenation. A decoder is a neural network module used to progressively recover a predicted mask with the same resolution as the input image from the fused features. A predicted mask is a binary mask output by the instance segmentation model after performing pixel-level predictions of the segmentation target in the sampled image. Joint loss is the total loss function, a weighted combination of multiple loss functions, used to comprehensively measure the difference between the predicted mask and the mask annotation, as well as the alignment between visual and textual features. Convergence criteria are preset standards used to determine whether the model training has reached its optimal state.

[0100] In this embodiment, the server pre-acquires a labeled sample set consisting of multiple infrared images and their corresponding pixel-level mask annotations. The server inputs the sampled images from the labeled sample set into a visual encoder. The visual encoder performs progressive downsampling and feature extraction on the sampled images through a multi-layer convolutional neural network, outputting visual feature maps at different scales to form multi-scale visual features. The server inputs the text prompt "cable terminal sleeve" into a text encoder. The text encoder converts the text prompt into a text feature vector through semantic understanding. The server fuses the multi-scale visual features with the text features through a cross-attention mechanism, i.e., using text features as the query vector and visual features as the key and value vectors, calculating attention weights to obtain fused features. The server inputs the fused features into a decoder. The decoder gradually restores the spatial resolution of the feature map through upsampling and convolution operations, finally outputting a predicted mask with the same resolution as the sampled images. The server calculates the joint loss between the predicted mask and the mask annotations, calculates the gradient of the model parameters based on the joint loss using a backpropagation algorithm, and updates the model parameters using an optimizer. The server repeats the above training steps until the joint loss value no longer decreases significantly or reaches the preset maximum number of iterations, saving the current model parameters as a trained instance segmentation model. During the inference phase, the server inputs the infrared image to be detected into the trained instance segmentation model. The model outputs a pixel-level mask of the cable terminal sleeve. The server extracts the contour boundary points of the mask to form the initial segmentation contour.

[0101] In another embodiment, the labeled samples acquired by the server also include negative sample text descriptions, which are text content unrelated to the segmentation target. When calculating the contrastive loss, the server uses the text features corresponding to the text prompts as positive samples and the text features corresponding to the negative sample text descriptions as negative samples. Through contrastive learning, the server narrows the distance between visual features and positive sample text features, and widens the distance between visual features and negative sample text features.

[0102] In another embodiment, when training the instance segmentation model, the server performs data augmentation on the sampled images in the labeled samples. Data augmentation includes at least one of random rotation, random scaling, random cropping, color jittering, and horizontal flipping. The server then inputs the augmented sampled images into the visual encoder for training.

[0103] In one exemplary embodiment, such as Figure 4 As shown, the above-mentioned "calculating the joint loss based on the difference between the predicted mask and the mask label" includes steps 401 to 403. Wherein:

[0104] Step 401: Calculate the segmentation loss based on the predicted mask and the mask label; the segmentation loss is used to measure the overlap between the predicted mask and the mask label.

[0105] In this context, the prediction mask refers to the binary mask output by the instance segmentation model after performing pixel-level predictions of the segmentation target in the sampled image. The value of each pixel in the mask represents the predicted probability or classification result of that pixel belonging to the segmentation target. The mask annotation refers to the true binary mask obtained after manually or semi-automatically annotating the cable terminal conduit area in the sampled image, used as a supervision signal during model training. The segmentation loss is the loss value calculated based on the difference between the prediction mask and the mask annotation, used to measure the degree of overlap between them; a smaller segmentation loss value indicates that the prediction mask and the mask annotation are closer. Overlap refers to the degree to which the prediction mask and the mask annotation coincide spatially.

[0106] In this embodiment, the server obtains the predicted mask output by the instance segmentation model and the mask annotation corresponding to the sampled image. Based on the predicted and ground truth values ​​of corresponding pixels in the predicted mask and mask annotation, the server calculates the intersection and union areas of the predicted mask and mask annotation, and calculates an overlap index based on the intersection and union areas. The server calculates a segmentation loss value based on the overlap index, and the segmentation loss value decreases as the overlap increases.

[0107] In another embodiment, the server uses a cross-entropy-based loss function to calculate the segmentation loss, that is, to calculate the binary cross-entropy loss between the predicted probability of each pixel in the prediction mask and the true label of that pixel in the mask annotation, and to take the average of the cross-entropy losses of all pixels as the segmentation loss.

[0108] In another embodiment, the server uses a contour-based loss function to calculate the segmentation loss, that is, extracting the predicted contour of the predicted mask and the true contour of the mask annotation, and calculating the contour distance or contour shape difference between the two contours as the segmentation loss.

[0109] Step 402: Calculate the contrast loss based on visual features and text features; the contrast loss is used to narrow the distance between visual features and text features in the feature space.

[0110] In this context, visual features refer to the feature vectors or feature maps extracted by the visual encoder from sampled images to represent image content. Text features refer to the feature vectors extracted by the text encoder from text prompts to represent text semantics. Contrastive loss is a loss value calculated based on the similarity between visual and text features, used to bring matching visual-text feature pairs closer together and push mismatched visual-text feature pairs further apart in the feature space. Feature space refers to a multi-dimensional vector space used to represent visual and text features, in which semantically similar visual features and text features are closer together.

[0111] In this embodiment, the server acquires visual features output by the visual encoder and text features output by the text encoder. The server calculates the similarity between the visual features and the text features, and calculates a contrast loss based on the similarity, so that the distance between the visual features and the corresponding text features in the feature space gradually decreases.

[0112] In another embodiment, the server obtains a training batch containing multiple sampled images. For each sampled image in the training batch, the similarity between the visual features of the sampled image and the text features corresponding to all sampled images in the batch is calculated. The similarity between the visual features and their corresponding text features is taken as positive sample similarity, and the similarity between the visual features and other text features in the same batch is taken as negative sample similarity. The contrast loss is calculated based on the positive sample similarity and all negative sample similarities.

[0113] In another embodiment, the server constructs a dual-tower model containing a visual encoder and a text encoder. During training, mismatched visual-text feature pairs are randomly selected as negative sample pairs. The similarity between visual features and positive sample text features is calculated as positive sample similarity, and the similarity between visual features and negative sample text features is calculated as negative sample similarity. The contrast loss is calculated based on the positive sample similarity and the negative sample similarity.

[0114] Step 403: Determine the joint loss based on the weighted sum of the segmentation loss and the contrast loss.

[0115] The joint loss refers to the total loss function, which is a weighted combination of the segmentation loss and the contrastive loss. It is used to simultaneously optimize the segmentation task and the cross-modal feature alignment task during model training. The weighted sum refers to the result obtained by multiplying the segmentation loss and the contrastive loss by their respective weight coefficients and then adding them together.

[0116] In this embodiment, the server obtains the segmentation loss calculated in step 401 and the contrast loss calculated in step 402. The server obtains the first weight coefficient corresponding to the segmentation loss and the second weight coefficient corresponding to the contrast loss, respectively. The server uses the product of the segmentation loss and the first weight coefficient as a first weighting term, and the product of the contrast loss and the second weight coefficient as a second weighting term. The first weighting term and the second weighting term are added together to obtain the joint loss. The server uses the joint loss to update the parameters of the instance segmentation model.

[0117] In another embodiment, the server dynamically adjusts the first weight coefficient corresponding to the segmentation loss and the second weight coefficient corresponding to the contrastive loss during training. In the early stages of training, the server sets a larger first weight coefficient to prioritize the model's learning of the segmentation task; in the later stages of training, the server gradually increases the second weight coefficient to optimize cross-modal feature alignment while maintaining segmentation accuracy.

[0118] In another embodiment, the server further obtains a third loss term, which is a regularization loss calculated based on the smoothness or regularity of the prediction mask, used to constrain the spatial continuity of the prediction mask. The server then performs a weighted sum of the segmentation loss, contrast loss, and the third loss term to determine the joint loss.

[0119] In one exemplary embodiment, such as Figure 5 As shown, the above-mentioned "determining the second parameter based on the temperature values ​​of multiple locations within the target area" includes steps 501 to 503. Wherein:

[0120] Step 501: Determine the center line of the target area and divide the center line into multiple sub-segments evenly.

[0121] The target region refers to the core area of ​​the sheath extracted from the infrared image for temperature analysis. The centerline is a line extending longitudinally along the target region and passing through its geometric center, used to characterize the sheath's central axis position in the image. Sub-segments are multiple continuous line segments obtained by equally dividing the centerline along its extension direction, with each sub-segment corresponding to a longitudinal position interval within the target region.

[0122] In this embodiment, the server determines the longitudinal extension direction of the target area based on its geometry, and determines a straight line or curve passing through the geometric center of the target area as the centerline along the longitudinal extension direction. The server obtains a preset number of sub-segments, calculates the length of each sub-segment based on the total length of the centerline and the preset number of sub-segments, and sequentially cuts multiple equal-length sub-segments along the extension direction of the centerline from one end.

[0123] In another embodiment, the server determines the longitudinal extension direction by calculating the major axis of the minimum bounding rectangle of the target area, and extends it from the geometric center of the target area to both ends along the major axis to the boundary of the target area to obtain the centerline.

[0124] In another embodiment, the server scans the binary mask of the target region line by line, calculates the mean horizontal coordinate of the pixels belonging to the target region in each line of the mask, and connects the mean horizontal coordinate points of each line in sequence to form a center line.

[0125] Step 502: Obtain the highest temperature value within each sub-segment as the temperature value of each sampling point.

[0126] Here, a sub-segment refers to each segment of the line obtained by dividing the centerline in step 501. A sampling point refers to the location point or temperature summary point within each sub-segment used to obtain temperature values. A temperature value refers to the highest temperature value within the corresponding area of ​​that sub-segment.

[0127] In this embodiment, the server determines the pixel band range corresponding to each sub-line segment in the target area based on the position of each sub-line segment on the center line. The server extracts all temperature values ​​within the pixel band range corresponding to each sub-line segment from the temperature matrix of the infrared image, counts the highest temperature value within each pixel band range, and uses the highest temperature value of each sub-line segment as the temperature value of the sampling point corresponding to that sub-line segment.

[0128] In another embodiment, the server expands each sub-segment to both sides with a preset width along its normal direction to form a rectangular sampling area, extracts all temperature values ​​within the rectangular sampling area, and takes the highest temperature value as the temperature value of the sampling point.

[0129] In another embodiment, the server selects multiple sampling points at equal intervals within each sub-segment, obtains the temperature value corresponding to each sampling point, and takes the average or median value of the temperature values ​​of each sampling point as the temperature value of the sub-segment.

[0130] Step 503: Calculate the second parameter based on the temperature value of each sampling point; wherein, the second parameter is multiple temperature difference values ​​between the temperature value of each sampling point and the highest temperature value on the center line; the preset uniform distribution condition is that the number of temperature difference values ​​that do not exceed the preset temperature difference threshold is greater than or equal to the preset quantity threshold.

[0131] The second parameter refers to a quantitative index calculated based on the temperature values ​​of each sampling point, used to characterize the uniformity of temperature distribution within the target area. The highest temperature value on the center line refers to the maximum value among all temperature values ​​of all sampling points in step 502. The temperature difference value refers to the absolute value of the difference between the temperature value of a single sampling point and the highest temperature value on the center line. The preset uniform distribution condition refers to a set of criteria used to determine whether the temperature distribution within the target area is uniform; when this condition is met, the temperature distribution is considered uniform.

[0132] In this embodiment, the server determines the temperature value with the largest value from the temperature values ​​of each sampling point obtained in step 502 as the highest temperature value on the center line. The server calculates the absolute value of the difference between the temperature value of each sampling point and the highest temperature value on the center line, obtaining multiple temperature difference values ​​as a second parameter. The server obtains a preset temperature difference threshold and counts the number of temperature difference values ​​that do not exceed the preset temperature difference threshold. The server obtains a preset quantity threshold and compares the counted quantity with the preset quantity threshold. If the counted quantity is greater than or equal to the preset quantity threshold, it is determined that the preset uniform distribution condition is met, the temperature distribution in the target area is uniform, and the current temperature anomaly is caused by optical reflection.

[0133] In another embodiment, the server normalizes the temperature values ​​of each sampling point and calculates the difference between each normalized temperature value and the highest normalized temperature value, using the normalized difference as a second parameter.

[0134] In another embodiment, the server uses a sliding window method to perform local uniformity detection on the temperature value on the center line. That is, it sequentially calculates the difference between the temperature value of each sampling point and the temperature value of its adjacent sampling points, counts the number of all adjacent differences that do not exceed a preset difference threshold as a second parameter, and compares this number with the preset number threshold.

[0135] In one exemplary embodiment, such as Figure 6 As shown, the above-mentioned "determining the classification result of the target region based on the temperature characteristics corresponding to the target region" includes steps 601 to 604. Wherein:

[0136] Step 601: Crop the temperature hotspot sub-image from the target area.

[0137] The target region refers to the core area of ​​the casing extracted from the infrared image for temperature analysis. The temperature hotspot sub-image refers to a local image patch containing temperature anomaly areas cropped from the target region. This image patch retains the temperature numerical information from the original temperature matrix and is used as input for subsequent classification models.

[0138] In this embodiment, the server determines the location of the pixel with the highest temperature based on the temperature distribution in the target area. Using this location as the center, a rectangular temperature hotspot sub-image is cropped from the target area according to a preset cropping size. The server uses this temperature hotspot sub-image as input data for subsequent classification models.

[0139] In another embodiment, the server calculates the average and standard deviation of the temperature values ​​of all pixels in the target area, determines a temperature threshold based on the average and standard deviation, identifies the connected regions in the target area where the temperature values ​​are higher than the temperature threshold as hotspot regions, determines the outer bounding box of the hotspot region, and crops a temperature hotspot sub-image from the target area based on the outer bounding box.

[0140] In another embodiment, the server uniformly divides the target area into multiple grid blocks, calculates the average temperature value in each grid block, selects the grid block with the highest average temperature value as the seed block, expands and merges adjacent grid blocks with temperatures higher than a preset threshold around the seed block to form a hotspot area, and cuts out a temperature hotspot sub-map from the target area according to the boundary of the hotspot area.

[0141] Step 602: After normalizing the temperature hotspot submap, the temperature features are obtained.

[0142] Normalization refers to scaling the temperature values ​​in the temperature hotspot sub-image to a preset range to eliminate the impact of temperature range differences between different infrared images on the classification model's judgment.

[0143] In this embodiment, the server obtains the maximum and minimum temperature values ​​of all pixels in the temperature hotspot sub-image, and linearly maps the temperature value of each pixel to a preset numerical range based on the maximum and minimum values ​​to obtain a normalized temperature hotspot sub-image. The server directly uses the normalized temperature hotspot sub-image as image features, or extracts statistical features from the normalized temperature hotspot sub-image as temperature features.

[0144] In another embodiment, the server calculates the mean and standard deviation of the temperature values ​​of all pixels in the temperature hotspot submap, subtracts the mean from the temperature value of each pixel and divides it by the standard deviation, so that the normalized temperature data follows a standard normal distribution.

[0145] Step 603: Classify the temperature features using a pre-trained classification model to obtain the classification probability of the target area belonging to the real heat generation category or the artifact category.

[0146] In this context, the classification model refers to a pre-trained machine learning or deep learning model used to map input temperature features to corresponding category labels. The classification probability refers to the probability value of the target region belonging to each category, typically represented by a value between 0 and 1. The true heating category refers to the category caused by actual heating due to internal defects in the cable termination bushing. The artifact category refers to the category caused by artifacts resulting from non-true heating factors such as optical reflection.

[0147] In this embodiment, the server inputs the temperature features obtained in step 602 into a pre-trained classification model. The classification model performs forward propagation calculations on the input features through its internal network structure, and finally outputs the probability values ​​of the target region belonging to the true heat category and the probability values ​​of it belonging to the artifact category through the classification layer.

[0148] In another embodiment, the classification model is a lightweight convolutional neural network (CNN) model. The server directly inputs the normalized temperature hotspot sub-image as image features into the CNN model. The CNN model automatically extracts the temperature distribution features of the temperature hotspot sub-image through convolutional and pooling layers, and outputs the classification probability through fully connected layers and a softmax activation function.

[0149] Step 604: Determine the classification result based on the classification probability.

[0150] Here, classification probability refers to the probability value of the target region belonging to the true fever category and the probability value of belonging to the artifact category output by the classification model in step 603. Classification result refers to the category label finally determined based on the classification probability, including the true fever category and the artifact category.

[0151] In this embodiment of the application, the server obtains the probability value of the target region belonging to the real fever category and the probability value of belonging to the artifact category output by the classification model in step 603, compares the two probability values, and determines the category with the larger probability value as the classification result.

[0152] In another embodiment, the server obtains the probability value of the target area output by the classification model in step 603 belonging to the real fever category, obtains a preset probability threshold, and compares the classification probability with the preset probability threshold. If the classification probability is greater than or equal to the preset probability threshold, the classification result is determined to be the real fever category; if the classification probability is less than the preset probability threshold, the classification result is determined to be the artifact category.

[0153] In an exemplary embodiment, the above-mentioned "classifying temperature features using a pre-trained classification model to obtain the classification probability that the target region belongs to the true heat category or the artifact category" includes:

[0154] The temperature feature is channel-expanded and processed to obtain a channel-expanded feature map. Spatial features are extracted from the channel-expanded feature map to obtain a spatial feature map. The spatial feature map is then processed sequentially with global average pooling, a first fully connected layer, a first activation function, a second fully connected layer, and a second activation function to obtain channel attention weights. These channel attention weights are then multiplied element-wise with the spatial feature map to obtain a weighted feature map. The weighted feature map is then channel-compressed to obtain a channel-compressed feature map. Finally, a residual connection is established between the temperature feature and the channel-compressed feature map to obtain the classification probability.

[0155] In this context, temperature features refer to the input data received by the classification model, specifically the feature representation of the normalized temperature hotspot submap after initial convolution processing within the model, or directly to the normalized temperature hotspot submap itself. This feature map has three dimensions: height, width, and channels. Channel expansion refers to increasing the number of channels in the feature map through a 1-to-1 convolution, enabling the model to learn richer representations in a higher-dimensional feature space. The channel-expanded feature map has more channels than the temperature features. The first activation function refers to introducing a non-linear activation function, allowing the model to fit complex non-linear mapping relationships. Spatial feature extraction refers to extracting features from the spatial dimension of the feature map through depthwise separable convolution or ordinary convolution, used to capture the spatial distribution patterns between different locations in the feature map. Global average pooling calculates the average pixel value across all locations in the spatial dimension for each channel of the feature map, compressing the spatial information of each channel into a scalar value, resulting in a pooling vector with a length equal to the number of channels. Channel attention weights are one-dimensional weight vectors generated by the compression activation module, with the same number of channels as the feature map. Each element in this weight vector measures the importance of the corresponding channel; a larger value indicates a more significant contribution of that channel to the classification task. Channel element-wise multiplication involves multiplying the channel attention weights by each channel of the feature map one by one. Specifically, the i-th channel attention weight is multiplied by all pixel values ​​of the i-th channel of the feature map, used to weight and label each channel of the feature map. Channel compression restores the number of channels in the weighted feature map to the same number as the temperature feature through a one-to-one convolution. Residual connection adds the temperature feature to the channel-compressed feature map element-wise, allowing gradients to propagate directly through shortcut paths, mitigating the vanishing gradient problem in deep networks and accelerating model convergence. Classification probability refers to the probability value output by the classification model after the above forward propagation calculations, indicating that the target region belongs to each category.

[0156] In this embodiment, the classification model receives a normalized temperature hotspot sub-map as input. This temperature hotspot sub-map is a single-channel or three-channel feature map with a preset height and width. The model first performs channel expansion processing on the temperature features using a 1x1 convolution, expanding the number of channels to a preset expansion factor. After the convolution operation, a first activation function is applied to introduce nonlinearity, resulting in a channel-expanded feature map. The model then performs spatial feature extraction on the channel-expanded feature map using a 3x3 depthwise separable convolution. Each channel undergoes an independent spatial convolution operation to capture spatial distribution features. After the depthwise separable convolution, a first activation function is applied to obtain a feature map with extracted spatial features. The model then performs global average pooling on the feature map with extracted spatial features in the spatial dimension, compressing the spatial information of each channel into a scalar value, resulting in a one-dimensional pooling vector. The length of the pooling vector is equal to the number of channels in the feature map with extracted spatial features. The model sequentially inputs the pooling vector into the first fully connected layer for feature transformation, the second activation function introduces non-linearity, the second fully connected layer transforms the feature dimension back to the same length as the number of channels, and the third activation function maps the output value to between 0 and 1, resulting in channel attention weights with the same length as the number of channels in the feature map. The model multiplies each channel of the extracted spatial feature map by the channel attention weights, i.e., weighting all pixel values ​​of the corresponding channel of the feature map with each component of the channel attention weight, resulting in a weighted feature map. The model then performs channel compression on the weighted feature map using a one-to-one convolution, restoring its channel number to the same as the temperature feature, resulting in a channel-compressed feature map. The model determines whether the current processing step size is one and whether the number of channels in the temperature feature and the output feature map are equal. If both conditions are met, the original temperature feature and the channel-compressed feature map are added element-wise for residual connection; otherwise, no residual connection is performed. The model processes the feature map after residual connection or channel compression through global average pooling and fully connected classification layers, and finally outputs the classification probability of the target region belonging to the real heat category and the classification probability of it belonging to the artifact category.

[0157] In another embodiment, the model employs a multi-branch convolutional structure when performing channel expansion, using one-to-one convolution and three-to-three convolution to process the temperature features in parallel, and then concatenating the outputs of the two branches along the channel dimension to obtain the feature map after channel expansion.

[0158] In another embodiment, after channel compression of the weighted feature map, the model passes through a batch normalization layer and a first activation function to obtain a channel-compressed feature map, and then performs a residual connection between the temperature feature and the channel-compressed feature map.

[0159] In an exemplary embodiment, the training process of the above classification model includes:

[0160] Obtain a training sample set, which includes positive and negative samples. Positive samples are infrared image patches labeled with the true fever category, and negative samples are infrared image patches labeled with the artifact category. Perform data augmentation on the training sample set to obtain an augmented training sample set. Input the augmented training sample set into the classification model to be trained, and the classification model to be trained will output the predicted probability of each training sample belonging to the true fever category. Calculate the prediction loss value based on the difference between the true label and the predicted probability of each training sample, and update the parameters of the classification model to be trained based on the prediction loss value. Iterate through the above training steps until the convergence condition is met to obtain the trained classification model.

[0161] The training sample set refers to the set of sample data used to train the classification model, including positive and negative samples. Positive samples are defined as infrared image patches labeled with the true heat generation category, derived from infrared images of cable terminal sleeves confirmed to have true heat generation defects. Negative samples are defined as infrared image patches labeled with the artifact category, derived from infrared images of cable terminal sleeves confirmed to have optical reflection artifacts. Data augmentation refers to the technique of performing a series of random transformations on the original training samples to generate new training samples, used to expand the diversity and quantity of training samples. The augmented training sample set refers to the sample set generated after the original training samples have undergone data augmentation. The classification model to be trained refers to a classification model that has not yet completed training and whose parameters are in an initial or partially updated state. The predicted probability refers to the probability value of the training sample output by the classification model belonging to the true heat generation category. The true label refers to the true category label of the training sample, used as a supervisory signal for comparison with the predicted probability. The loss value is a numerical value calculated based on the difference between the true label and the predicted probability, used to measure the accuracy of the model's prediction. The convergence condition refers to the preset standard used to determine whether the model training has reached the optimal state.

[0162] In this embodiment, the server acquires a pre-constructed training sample set, which includes multiple positive samples and multiple negative samples. Each training sample is a temperature hotspot sub-image cropped from an infrared image and labeled with a corresponding category label. The server performs data augmentation on the training sample set, including at least one of random rotation, random scaling, random cropping, horizontal flipping, and color jittering, to generate an augmented training sample set. The augmented training sample set contains the original training samples and the new samples generated after data augmentation. The server constructs a classification model to be trained, which employs a lightweight convolutional neural network structure. The server sequentially inputs each training sample from the augmented training sample set into the classification model to be trained. The classification model calculates and outputs the predicted probability of each training sample belonging to the true heat category through forward propagation. The server acquires the true label of each training sample and calculates the loss value based on the difference between the true label and the corresponding predicted probability. The server calculates the gradient of the model parameters based on the loss value using a backpropagation algorithm and updates the parameters of the classification model to be trained using an optimizer based on the gradient. The server repeats the above training steps, using the updated model as the classification model to be trained until the convergence condition is met, and then saves the current model parameters as the trained classification model.

[0163] In another embodiment, the negative samples in the training sample set obtained by the server include at least one of the following: strong light direct illumination artifact samples, water surface reflection artifact samples, and metal flange high brightness artifact samples.

[0164] In another embodiment, when performing data augmentation, the server sequentially performs three augmentation operations on the training samples: illumination jitter, Gaussian noise addition, and affine transformation, generating an augmented training sample set. Illumination jitter involves randomly adjusting the image brightness or contrast to simulate imaging effects under different lighting conditions; Gaussian noise addition simulates sensor noise by superimposing random Gaussian noise into the image; and affine transformation simulates imaging effects at different shooting angles and distances by randomly translating, rotating, or scaling the image.

[0165] In an exemplary embodiment, the above-mentioned "determining the classification result based on the classification probability" includes:

[0166] If the classification probability is greater than or equal to a preset probability threshold, the classification result is determined to be the true fever category; if the classification probability is less than the preset probability threshold, the classification result is determined to be the artifact category.

[0167] The classification probability refers to the probability value output by the classification model that the target area belongs to the true heating category. The value ranges from 0 to 1; a higher probability value indicates a higher confidence level in the classification model's judgment that the target area is experiencing true heating. The preset probability threshold is a pre-set probability threshold used to distinguish between the true heating category and the artifact category. The true heating category refers to the category caused by actual heating due to internal defects in the cable termination bushing. The artifact category refers to the category caused by artifacts due to non-true heating factors such as optical reflection. The classification result refers to the final category label determined based on the comparison between the classification probability and the preset probability threshold.

[0168] In this embodiment, the server obtains the classification probability that the target region output by the classification model belongs to the true fever category. The server obtains a preset probability threshold and compares the classification probability with the preset probability threshold. If the classification probability is greater than or equal to the preset probability threshold, the server determines the classification result as the true fever category; if the classification probability is less than the preset probability threshold, the server determines the classification result as the artifact category.

[0169] In another embodiment, the server also obtains the classification probability of the target region belonging to the artifact category output by the classification model, compares the classification probability of the target region belonging to the real fever category with the classification probability of the target region belonging to the artifact category, and determines the category with the larger probability value as the preliminary classification result. The server obtains the classification probability value corresponding to the preliminary classification result, compares the classification probability value with a preset probability threshold, and if the classification probability value is greater than or equal to the preset probability threshold, the preliminary classification result is output; if the classification probability value is less than the preset probability threshold, the preliminary classification result is marked as pending review and pushed to the manual review queue.

[0170] In another embodiment, the server obtains a dynamically adjusted preset probability threshold, which is determined based on the false positive rate and false negative rate requirements of the current detection scenario. In scenarios with strict false positive rate requirements, the server increases the preset probability threshold; in scenarios with strict false negative rate requirements, the server decreases the preset probability threshold.

[0171] In an exemplary embodiment, the first parameter is the temperature difference between the temperature values ​​of multiple sub-regions. Based on this, the method further includes:

[0172] If the first parameter does not exceed the temperature difference threshold, the cable terminal bushing is determined to be non-heating and the detection process ends; if the second parameter meets the uniform distribution condition, the current temperature anomaly is determined to be caused by optical reflection, and the detection result that the cable terminal bushing does not have a heating defect is output.

[0173] In this embodiment, the server obtains the first parameter calculated in step 202 and compares it with a temperature difference threshold. If the first parameter does not exceed the temperature difference threshold, it indicates that the temperature difference between sub-regions within the target area is not significant, and there is no obvious local heating. The server determines that the cable terminal sleeve is not heating and ends the detection process. If the first parameter exceeds the temperature difference threshold, the server continues to execute subsequent steps. In the subsequent steps, the server obtains the second parameter determined in step 203 and compares it with a preset uniform distribution condition. If the second parameter meets the preset uniform distribution condition, it indicates that the temperature distribution within the target area is uniform, and the currently detected temperature anomaly is caused by optical reflection. The server determines that the current temperature anomaly is an artifact, outputs the detection result that the cable terminal sleeve does not have a heating defect, and ends the detection process. If the second parameter does not meet the preset uniform distribution condition, the server continues to execute subsequent steps and enters the classification model processing.

[0174] In another embodiment, when the server determines that the cable terminal sleeve is not heating up and ends the detection process, it stores the current infrared image, target area location information and judgment result in a normal sample library. The normal sample library is used for incremental training of subsequent classification models or for generating a normal temperature distribution baseline.

[0175] In another embodiment, when the server determines that the current temperature anomaly is caused by optical reflection and outputs a detection result indicating that there is no heating defect, it stores the current infrared image, the target area location information, and the determination result in an artifact sample library. The artifact sample library is used for incremental training of the subsequent classification model to enhance the model's ability to recognize optical reflection artifacts.

[0176] In one embodiment, Figure 7 This is a schematic diagram of the overall process of the heat detection method provided in the embodiments of this application, as shown below. Figure 7 As shown, the method includes: a refined extraction step of the sleeve polygon contour based on the Segment Anything Model 3 (SAM3) to segment the cable terminal sleeve in the infrared image and obtain the initial segmentation contour of the sleeve region; a polygon contour proportional inward compression step to proportionally shrink the initial segmentation contour inward to remove background pixel interference mixed in by the contour edge; obtaining the minimum bounding rectangle of the component contour, and truncating one-fifth of the area at the center of the rectangle to obtain the cable terminal sleeve insulation core region as the target region; dividing the composite insulator into upper, middle and lower parts according to the vertical coordinate of the core region; obtaining the highest temperature, lowest temperature, ambient temperature and global highest temperature in the three parts based on the temperature recognition module of the infrared image; calculating the relative temperature difference based on the temperature information of each part in the component, the formula for calculating the relative temperature difference is: relative temperature difference = .

[0177] In the formula and The temperature of the hot spot, The system uses ambient temperature as the reference temperature; it combines relative temperature difference with the highest temperature of the component to determine the level of heating defects; an infrared component heating reflection classification and filtering model is used to perform secondary verification on suspected heating samples; by comparing reflection, background and foreground, it determines the effective high temperature point of the component to distinguish between real heating and optical reflection artifacts; and outputs the defect detection results.

[0178] In one embodiment, the refined extraction of the casing polygon contour based on SAM3 is achieved by employing both the Common Objects in Context (COCO) annotation format and the native SAM mask format. The annotation file is stored in JSON format and includes image ID, category ID, and polygon vertices. Binary mask matrix Mask quality is verified by pixel accuracy (PA) and intersection over union (IoU).

[0179]

[0180] Required qualified samples , .

[0181] Input image Multi-scale features are extracted using a visual encoder:

[0182]

[0183] in For encoder parameters, The length of the feature sequence. For feature dimensions. Text prompts are processed... Encoded as Fusion is achieved through a cross-attention mechanism:

[0184]

[0185] Semantic enhancement expands the concept space through synonym substitution, improving the model's robustness to expressions such as "cable terminal sleeve". The SAM3 backbone parameters are frozen. A lightweight Adapter module is inserted between the encoder and decoder. The Adapter employs a bottleneck structure:

[0186]

[0187] in This is a scaling factor (usually 0.1), used only for training. With a small number of parameters, the number of parameters is reduced by more than 95%. (Regarding attention weights) Apply low-rank decomposition.

[0188]

[0189] rank (generally Gradient updates only apply to The computational complexity is from Down to Combining segmentation accuracy and concept alignment, we define the joint loss as follows:

[0190]

[0191] in Measuring mask overlap Aligning visual-text features using InfoNCE loss:

[0192]

[0193] Figure 8 This is a schematic diagram of an infrared image of a cable termination bushing in one embodiment, such as... Figure 8 As shown, this infrared image was acquired by the FLIR infrared thermal imaging device. Figure 8 The temperature distribution of the cable terminal bushing is shown in the image. Figure 8 This includes the thermal radiation distribution area of ​​the casing body and the surrounding background area. Figure 8 The display shows temperature measurement parameters, including the distance parameter Dist, the reflection temperature parameter Trefl, and the emissivity parameter ε. The distance parameter represents the measurement distance between the infrared thermal imaging device and the sheath, the reflection temperature parameter represents the set value of the ambient reflection temperature, and the emissivity parameter represents the set value of the emissivity of the sheath surface. Figure 8 The highest and lowest temperature points are marked with their corresponding temperature values. The highest temperature point, 45.7, indicates the location with the highest surface temperature of the casing, and the lowest temperature point, 16.5, indicates the location with the lowest surface temperature of the casing. The difference between the highest and lowest temperatures is used for subsequent calculations of the relative temperature rise. Figure 8 The bright areas on the main body of the middle sleeve may represent the actual heating area or optical reflection artifacts caused by direct sunlight or reflection from the metal surface. They need to be distinguished through subsequent zone temperature difference analysis and classification models.

[0194] Obtaining the heating area of ​​the cable terminal bushing: The first step is to output the outline of the cable terminal bushing component from the instance segmentation model, which has filtered out a large amount of background information compared to the target detection model. However, due to the limitation of infrared image resolution, the boundary between the component and the background is not clear enough. Therefore, in this embodiment, the Python clipping library PyClipper is used to compress the polygonal outline output by the model proportionally inward to further eliminate irrelevant background interference. Figure 8 The green polygons in the diagram represent the model's outline for recognition, while the yellow outlines are scaled-down versions of the original outlines.

[0195] In one embodiment, Figure 9 This is a schematic diagram of an infrared temperature measurement display interface for a cable terminal bushing in one embodiment. The infrared image is acquired by an infrared thermal imaging device (FLIR). Figure 9 As shown, the infrared temperature measurement display interface includes an infrared thermal imaging image display area and a temperature parameter display area. The infrared thermal imaging image display area displays an infrared thermal radiation distribution image of the cable terminal sleeve. Figure 9 The highlighted area on the main body of the sleeve can represent either the actual heating area or optical reflection artifacts caused by direct sunlight or reflection from the metal surface. The temperature parameter display area shows the temperature parameter information for the current frame, including the distance parameter Dist, the reflection temperature parameter Trefl, and the emissivity parameter ε. The distance parameter represents the measurement distance between the infrared thermal imaging device and the sleeve, the reflection temperature parameter represents the set value of the ambient reflection temperature, and the emissivity parameter represents the set value of the emissivity of the sleeve surface. Figure 9 The markings indicate the location of the highest temperature point (45.7°C) and the lowest temperature point (16.5°C) along with their corresponding temperature values. The highest temperature point is located in the main body area of ​​the casing, while the lowest temperature point is located in either the main body area of ​​the casing or the background area.

[0196] Let the set of polygon vertices of the original segmented contour be:

[0197]

[0198] The new polygonal outline is obtained after proportional compression using PyClipper. For the compressed contour Find the smallest bounding rectangle whose coordinate range satisfies:

[0199]

[0200]

[0201] Coordinates of the center point of the minimum bounding rectangle for:

[0202]

[0203] To accurately locate the core heat-generating area, a central 1 / 5 region is extracted from the interior of the smallest bounding rectangle, with its horizontal and vertical extents representing the width and height of the original rectangle, respectively. ,Right now:

[0204]

[0205] The cable termination bushing exhibits a significant temperature difference between its high-voltage and low-voltage ends. Directly comparing the entire temperature matrix of the cable termination bushing can lead the algorithm to flag it as an overheating emergency due to this large temperature difference, significantly impacting line maintenance safety. This project divides the cable termination bushing into three segments (upper, middle, and lower) based on the ordinate of the minimum bounding rectangle. This allows for the acquisition of the core area of ​​the cable termination bushing. .

[0206]

[0207] The EXIF ​​tool is used to decode the newly input infrared image and extract its binary thermal radiation matrix. Based on the heat-generating area mask obtained in the previous step and the shrunken component contour mask, the matrix is ​​used for judgment. Extract the corresponding heat matrix This yields the highest temperature of the heating area and its coordinates, as well as the lowest temperature and its coordinates within the contour area of ​​the shrunken component.

[0208]

[0209] The main factors for determining the heating of cable terminal bushing current in each sub-region are relative temperature rise and absolute temperature. The calculation formula is as follows:

[0210]

[0211] in These are the temperatures of the heating points; This is the temperature at the normal corresponding point.

[0212] To address the issue of misidentification due to overheating of cable terminal bushings caused by reflection, the temperature along the centerline of the cable terminal bushing is extracted based on the core area extracted above. The center line is then evenly divided into 10 sub-regions, denoted as:

[0213]

[0214] The highest temperature of each of the 10 centerline sub-regions was obtained. (in , (Represents the sub-region number).

[0215] Based on the highest temperature of the centerline The highest temperature in each sub-region of the center line is respectively Compare the temperatures and calculate the temperature difference between the two. The calculation formula is:

[0216]

[0217] Define the decision criteria and introduce an indicator function. The function satisfies: when hour, ;when hour, .

[0218] Statistical satisfaction The number of sub-regions, i.e., the cumulative sum of the indicator functions:

[0219]

[0220] If the sum satisfies If the condition is met, it proves that the insulator string is a false positive due to reflection; if the condition is not met, it indicates that the insulator string is suspected of overheating.

[0221] In one embodiment, Figure 10 This is a schematic diagram of the extraction and reflection filtering of the center line sub-region in one embodiment, as shown below. Figure 10 As shown, the schematic diagram includes a centerline sub-region extraction section and a reflection filtering result section. The centerline sub-region extraction section demonstrates the process of determining the centerline from the target area, uniformly dividing the centerline into multiple sub-segments, and then extracting the highest temperature point within each sub-segment. Multiple sampling points are marked on the sleeve centerline, each labeled with a corresponding temperature value, such as 28.5℃, 29.7℃, 30.1℃, 30.7℃, 31.4℃, 29.2℃, 31.5℃, and 32.1℃. These temperature values ​​represent the highest temperature within each sub-segment. The centerline sub-region extraction section also includes a diagram illustrating the extraction of the highest temperature point within each sub-region. By sequentially obtaining the highest temperature value of each sub-segment along the centerline direction, the temperature difference between the temperature value at each sampling point and the highest temperature value on the centerline is calculated to determine whether the temperature distribution meets the uniformity distribution condition. The reflection filtering section shows the judgment result after analyzing the temperature distribution uniformity based on the temperature values ​​of the above sampling points, used to distinguish between real heating and optical reflection artifacts. When the temperature values ​​at each sampling point are relatively uniformly distributed, i.e., the number of sampling points whose temperature difference is less than or equal to a preset temperature difference threshold is greater than or equal to a preset quantity threshold, the current temperature anomaly is determined to be caused by optical reflection. The sample is then marked as a false positive for reflection and discarded. The figure uses a visual comparison and analysis of the uniformity of temperature distribution at each sampling point to illustrate the judgment logic of the reflection filtering.

[0222] When the previous step outputs a suspected overheating issue, a secondary verification is triggered. From the corresponding... Medium clipping temperature hotspot subplot Bilinear interpolation resampling to Perform normalization Input MobileNetV3-Small classification network Its core consists of inverted residual blocks and Squeeze-and-Excitation (SE) attention. The forward propagation of a single inverted residual block is defined as follows:

[0223]

[0224] in , For the Sigmoid function, Element-wise multiplication is performed for each channel. The network is terminated by a fully connected layer and a Softmax layer to output the true probability of heat generation. The model training uses binary cross-entropy loss:

[0225]

[0226] The training set includes pseudo-positive samples such as direct strong light, water surface reflection, and bright metal flanges, as well as real heat-generating negative samples. Generalization is enhanced through lighting jitter, Gaussian noise, and affine transformation. The final decision logic is as follows:

[0227]

[0228] The heating level is determined based on the temperature of the cable terminal bushing and the relative temperature rise, as detailed in Table 1:

[0229] Table 1

[0230]

[0231] In one exemplary embodiment, the method further includes:

[0232] Step 1: Perform instance segmentation processing on the infrared image to obtain the initial segmentation contour of the cable terminal sleeve.

[0233] Step 2: Perform proportional inward shrinkage on the initial segmented contour to obtain the shrunken contour, and determine the minimum bounding rectangle of the shrunken contour.

[0234] Step 3: Extract the central region within the smallest bounding rectangle and use the central region as the target region; wherein, the ratio of the horizontal dimension of the central region to the width of the smallest bounding rectangle is a preset ratio, and the ratio of the vertical dimension of the central region to the height of the smallest bounding rectangle is a preset ratio.

[0235] Step 4: Divide the target area into multiple sub-regions based on the spatial distribution of the target area.

[0236] Step 5: Obtain the temperature value of each sub-region, and calculate the first parameter reflecting the temperature difference within the target region based on the temperature value of each sub-region.

[0237] Step 6: If the first parameter exceeds the temperature difference threshold, determine the center line of the target area and divide the center line into multiple sub-segments evenly.

[0238] Step 7: Obtain the highest temperature value within each sub-segment as the temperature value of each sampling point.

[0239] Step 8: Calculate the second parameter based on the temperature value of each sampling point; wherein, the second parameter is multiple temperature difference values ​​between the temperature value of each sampling point and the highest temperature value on the center line; the preset uniform distribution condition is that the number of temperature difference values ​​that do not exceed the preset temperature difference threshold is greater than or equal to the preset quantity threshold; the second parameter is used to characterize the temperature distribution uniformity within the target area.

[0240] Step 9: If the second parameter does not meet the preset uniform distribution condition, cut out the temperature hotspot sub-map from the target area.

[0241] Step 10: After normalizing the temperature hotspot submap, the temperature features are obtained.

[0242] Step 11: Perform channel expansion and processing on the temperature feature to obtain the channel-expanded feature map.

[0243] Step 12: Extract spatial features from the feature map after channel expansion to obtain the feature map after spatial feature extraction.

[0244] Step 13: The feature map after spatial feature extraction is processed sequentially by global average pooling, the first fully connected layer, the first activation function, the second fully connected layer, and the second activation function to obtain the channel attention weights.

[0245] Step 14: Multiply the channel attention weights with the feature maps after spatial feature extraction by channel element-wise to obtain the weighted feature maps.

[0246] Step 15: Perform channel compression on the weighted feature map to obtain the channel-compressed feature map.

[0247] Step 16: Perform residual connection between the temperature features and the channel-compressed feature map to obtain the classification probability.

[0248] Step 17: If the classification probability is greater than or equal to the preset probability threshold, determine the classification result as the true fever category; if the classification probability is less than the preset probability threshold, determine the classification result as the artifact category.

[0249] Step 18: If the classification result is a true heat generation category, output the detection result that the cable terminal bushing has a heat generation defect; if the classification result is an artifact category, output the detection result that the cable terminal bushing does not have a heat generation defect.

[0250] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0251] Based on the same inventive concept, this application also provides a cable terminal bushing heat detection device for implementing the above-described cable terminal bushing heat detection method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more cable terminal bushing heat detection device embodiments provided below can be found in the limitations of the cable terminal bushing heat detection method described above, and will not be repeated here.

[0252] In one exemplary embodiment, such as Figure 11 As shown, a device for detecting the overheating of a cable termination sleeve is provided, comprising: a division module 701, an acquisition module 702, a first determination module 703, a second determination module 704, and an output module 705, wherein:

[0253] The segmentation module 701 is used to obtain the target area of ​​the cable terminal sleeve in the infrared image and divide the target area into multiple sub-regions according to the spatial distribution of the target area.

[0254] The acquisition module 702 is used to acquire the temperature values ​​of each sub-region and calculate the first parameter reflecting the temperature difference within the target region based on the temperature values ​​of each sub-region.

[0255] The first determining module 703 is used to determine a second parameter based on the temperature values ​​of multiple locations within the target area when the first parameter exceeds the temperature difference threshold; the second parameter is used to characterize the uniformity of temperature distribution within the target area.

[0256] The second determining module 704 is used to determine the classification result of the target region based on the temperature characteristics of the target region when the second parameter does not meet the preset uniform distribution condition.

[0257] The output module 705 is used to output the detection result of the cable terminal bushing having a heating defect when the classification result is a true heating category; and to output the detection result of the cable terminal bushing not having a heating defect when the classification result is an artifact category.

[0258] It should be noted that each module in the above-mentioned cable terminal sleeve heating detection device can execute the above-mentioned method embodiment, and its implementation principle and technical effect are similar, so they will not be described again here.

[0259] Each module in the aforementioned cable terminal sleeve heating detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0260] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0261] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0262] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0263] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0264] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0265] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A heat generation detection method for a cable terminal bush, characterized by, The method includes: The target area of ​​the cable terminal sleeve in the infrared image is obtained, and the target area is divided into multiple sub-regions according to the spatial distribution of the target area; The temperature values ​​of each of the sub-regions are obtained respectively, and a first parameter reflecting the temperature difference within the target region is calculated based on the temperature values ​​of each of the sub-regions. If the first parameter exceeds the temperature difference threshold, a second parameter is determined based on the temperature values ​​of multiple locations within the target area; the second parameter is used to characterize the uniformity of temperature distribution within the target area. If the second parameter does not meet the preset uniform distribution condition, the classification result of the target region is determined according to the temperature characteristics corresponding to the target region. If the classification result is a true heat generation category, the output will show a detection result indicating that the cable terminal bushing has a heat generation defect; if the classification result is an artifact category, the output will show a detection result indicating that the cable terminal bushing does not have a heat generation defect.

2. The method of claim 1, wherein, The acquisition of the target area of ​​the cable terminal sleeve in the infrared image includes: The infrared image is segmented to obtain the initial segmentation contour of the cable terminal sleeve. The initial segmented contour is shrunk inward proportionally to obtain the shrunk contour, and the minimum bounding rectangle of the shrunk contour is determined. A central region is extracted within the minimum bounding rectangle, and this central region is taken as the target region; wherein, the ratio of the horizontal dimension of the central region to the width of the minimum bounding rectangle is a preset ratio, and the ratio of the vertical dimension of the central region to the height of the minimum bounding rectangle is the preset ratio.

3. The method of claim 2, wherein, The step of performing instance segmentation processing on the infrared image to obtain the initial segmentation contour of the cable terminal sleeve includes: The infrared image is segmented using an instance segmentation model to obtain the initial segmentation contour of the cable terminal sleeve. The instance segmentation model is trained in the following way: Obtain labeled samples, wherein the labeled samples include sampled images and corresponding mask labels; The sampled image is input into a visual encoder to extract multi-scale visual features from the sampled image; The text prompts used to indicate the segmentation target are input into the text encoder, the text features of the text prompts are extracted, and the multi-scale visual features are fused with the text features to obtain fused features; The fused features are input into the decoder to generate a predicted mask of the segmented target in the sampled image; The joint loss is calculated based on the difference between the predicted mask and the mask label, and the parameters of the trained segmentation model are updated based on the joint loss. The process iteratively executes the steps of inputting the fused features into the decoder to generate a predicted mask of the segmented target in the sampled image, and the steps of calculating the joint loss based on the difference between the predicted mask and the mask annotation, and updating the parameters of the trained segmentation model based on the joint loss, until the convergence condition is met, and a trained instance segmentation model is obtained.

4. The method of claim 3, wherein, The step of calculating the joint loss based on the difference between the predicted mask and the mask label includes: The segmentation loss is calculated based on the predicted mask and the mask annotation; the segmentation loss is used to measure the overlap between the predicted mask and the mask annotation. A contrast loss is calculated based on the visual features and the text features; the contrast loss is used to narrow the distance between the visual features and the text features in the feature space. The joint loss is determined by a weighted sum of the segmentation loss and the comparison loss.

5. The method of claim 1, wherein, The step of determining the second parameter based on the temperature values ​​of multiple locations within the target area includes: Determine the centerline of the target area, and divide the centerline evenly into multiple sub-segments; The highest temperature value within each sub-segment is obtained as the temperature value of each sampling point. The second parameter is calculated based on the temperature value of each sampling point; wherein, the second parameter is a plurality of temperature difference values ​​between the temperature value of each sampling point and the highest temperature value on the center line; the preset uniform distribution condition is that the number of temperature difference values ​​among the plurality of temperature difference values ​​that do not exceed a preset temperature difference threshold is greater than or equal to a preset quantity threshold.

6. The method of claim 1, wherein, The step of determining the classification result of the target region based on the temperature characteristics corresponding to the target region includes: Extract temperature hotspot sub-maps from the target region; After normalizing the temperature hotspot submap, the temperature feature is obtained; The temperature features are classified using a pre-trained classification model to obtain the classification probability that the target region belongs to the true heat generation category or the artifact category. The classification result is determined based on the classification probability.

7. The method of claim 6, wherein, The step of classifying the temperature features using a pre-trained classification model to obtain the classification probability of the target region belonging to the true heat category or the artifact category includes: The temperature feature is subjected to channel expansion and processing to obtain a channel-expanded feature map; Spatial feature extraction is performed on the feature map after channel expansion to obtain a feature map after spatial feature extraction; The feature map after spatial feature extraction is processed sequentially by global average pooling, a first fully connected layer, a first activation function, a second fully connected layer, and a second activation function to obtain channel attention weights; The channel attention weights are multiplied element-wise with the feature map after spatial feature extraction to obtain a weighted feature map. The weighted feature map is subjected to channel compression to obtain a channel-compressed feature map; The temperature feature is residually concatenated with the channel-compressed feature map to obtain the classification probability.

8. The method of claim 6, wherein, The training process of the classification model includes: Obtain a training sample set, which includes positive samples and negative samples. The positive samples are infrared image patches labeled with the actual heat generation category, and the negative samples are infrared image patches labeled with the artifact category. The training sample set is subjected to data augmentation processing to obtain an augmented training sample set; The enhanced training sample set is input into the classification model to be trained, and the classification model to be trained outputs the predicted probability of each training sample belonging to the real fever category. The prediction loss value is calculated based on the difference between the true label of each training sample and the predicted probability, and the parameters of the classification model to be trained are updated based on the prediction loss value. Repeat the steps of obtaining the training sample set and updating the parameters of the classification model to be trained according to the predicted loss value until the convergence condition is met, and a trained classification model is obtained.

9. The method of claim 6, wherein, Determining the classification result based on the classification probability includes: If the classification probability is greater than or equal to a preset probability threshold, the classification result is determined to be the true fever category; If the classification probability is less than the preset probability threshold, the classification result is determined to be an artifact category.

10. The method according to claim 1, characterized in that, The first parameter is the temperature difference between the temperature values ​​of the plurality of sub-regions; the method further includes: If the first parameter does not exceed the temperature difference threshold, it is determined that the cable terminal bushing is not heating up and the detection process ends. If the second parameter satisfies the uniform distribution condition, it is determined that the current temperature anomaly is caused by optical reflection, and the detection result that the cable terminal sleeve does not have a heating defect is output.

Images