Automatic detection method and equipment for temperature abnormity of transformer terminal, and medium

Abstract

The invention relates to a transformer terminal temperature anomaly automatic detection method and device and a medium, and the method comprises the steps: inputting a visible light RGB image into a lightweight target detection model, and obtaining an initial position detection frame of a transformer terminal in the visible light RGB image; mapping the initial position detection frame to the infrared image to obtain an initial infrared temperature measurement area; in the initial infrared temperature measurement area and the neighborhood thereof, identifying a high-temperature core area based on pixel temperature distribution; performing spatial consistency verification on the high-temperature core area and the initial position detection frame, optimizing the initial position detection frame according to a verification result, and generating an optimized infrared temperature measurement area; based on the radiation response gain and the environment temperature compensation offset, converting the pixel gray value in the optimized infrared temperature measurement area into a temperature value, and calculating the representative temperature of the optimized infrared temperature measurement area; and obtaining a temperature anomaly diagnosis result according to the representative temperature. Compared with the prior art, the method has the advantages of high efficiency, accuracy, light weight and the like.

Description

Technical Field

[0001] This invention relates to the field of transformer testing, and in particular to an automatic detection method, equipment, and medium for abnormal transformer terminal temperatures. Background Technology

[0002] As a core piece of equipment in a power system, transformers bear high currents and high temperatures at their terminal connections for extended periods, making them one of the most prone to failure areas. Due to issues such as loose connecting bolts, poor conductive contact, and oxidation corrosion, localized overheating can easily occur at the terminals. If not detected in time, this can lead to equipment burnout or even a complete power outage.

[0003] Currently, temperature detection of transformer terminals still mainly relies on manual inspection. Inspectors manually capture images of the terminals using infrared thermal imagers and rely on experience to diagnose faults. This method suffers from numerous problems, including low efficiency, high subjectivity, inaccurate positioning, and unstable temperature extraction. It is particularly prone to missed detections and misjudgments when multiple devices are arranged side-by-side, terminals are obstructed, or the background is complex. Furthermore, existing infrared detection equipment is usually separate from the image analysis system, lacking an integrated, intelligent, and portable operating platform, which is detrimental to real-time on-site detection and handling.

[0004] To improve the accuracy and automation of detection, researchers have recently explored the introduction of deep learning-based object detection algorithms, particularly the YOLO (You Only Look Once) series of models, for infrared image recognition and fault location of power equipment. YOLO, as a single-stage object detection framework, offers advantages such as end-to-end processing, high speed, and flexible structure, and has been widely applied in industrial quality inspection, traffic monitoring, and power line inspection. Meanwhile, lightweight structural design has become a research hotspot for deploying detection algorithms on edge devices such as handheld terminals. Mainstream methods include MobileNet and ShuffleNet, which significantly reduce model computation while maintaining accuracy through strategies such as depthwise separable convolutions, attention pruning, and dynamic structural adjustment. Furthermore, the combination of structural reparameterization and knowledge distillation techniques has further promoted the deployment of high-performance models on mobile devices.

[0005] To further enhance the spatial awareness and robustness of detection models, attention mechanisms and dynamic sensing operators have been widely introduced in recent years. Channel and spatial attention mechanisms can enhance the model's ability to focus on key regions, thereby improving detection accuracy; however, these mechanisms typically come with additional computational overhead and parameter redundancy, which may adversely affect the model's real-time performance and energy efficiency when deployed on handheld devices or at the edge. Dynamic convolutional structures, by adaptively generating convolutional kernels based on input features, effectively improve the model's adaptability when handling complex terminal structures and occlusion situations. However, in practical applications with limited sample sizes or significant variations in input distribution, these methods still suffer from insufficient generalization ability.

[0006] Furthermore, multimodal fusion technology, by establishing a registration relationship between infrared and RGB images, enables joint modeling of spatial structural information and thermal feature information, mitigating to some extent the inaccurate positioning problem caused by factors such as blurred target boundaries and generalized temperature information in infrared images. However, due to redundancy in some areas between infrared and RGB images, and the potential for modal conflicts in complex scenarios such as occlusion and strong reflection, the fused information may be distorted, affecting the stability and accuracy of the detection results. Existing research mainly focuses on implementation on fixed equipment or PC platforms, lacking a terminal-level detection solution that is real-time, lightweight, and integrated for handheld mobile terminals. Therefore, designing a more robust cross-modal fusion strategy to effectively avoid information interference and improve the collaborative characterization ability between heterogeneous modalities, thereby fusing target detection and infrared thermometry capabilities to obtain accurate transformer terminal temperature anomaly detection results, and facilitating lightweight integration, is a technical problem that needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a method, device and medium for detecting abnormal temperature at transformer terminals. By introducing a dynamic detection frame optimization mechanism based on temperature distribution consistency, a closed-loop system is constructed that verifies and optimizes visible light semantic information and infrared temperature information. This design effectively solves the problem of inaccurate temperature measurement caused by terminal obstruction and background heat source interference in complex field environments, thereby achieving accurate and automated detection of abnormal temperature at transformer terminals.

[0008] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, an automatic detection method for abnormal transformer terminal temperature is provided, the specific steps of which include: S1. Synchronously acquire registered infrared and visible light RGB images of the transformer terminal area; S2. Input the visible light RGB image into the lightweight target detection model to obtain the initial position detection box of the transformer terminal in the visible light RGB image; S3. Map the initial position detection box onto the infrared image to obtain an initial infrared temperature measurement area; in the initial infrared temperature measurement area and its neighborhood, identify a high-temperature core area based on pixel temperature distribution; perform spatial consistency verification on the high-temperature core area and the initial position detection box, and optimize the initial position detection box according to the verification result to generate an optimized infrared temperature measurement area. S4. Based on the radiation response gain and ambient temperature compensation offset, convert the pixel grayscale values ​​in the optimized infrared temperature measurement area into temperature values, and calculate the representative temperature of the optimized infrared temperature measurement area. S5. Based on the representative temperature and the criteria for judging heating defects, output the diagnostic results of temperature anomalies at the transformer terminals.

[0009] Furthermore, the neighborhood in S3 is a region formed by extending outward from the initial position detection box as the center, and the boundary is determined by the size ratio of the initial position detection box. The boundary is determined according to the width or height of the initial position detection box.

[0010] Furthermore, the specific steps for identifying the high-temperature core region include: The temperature values ​​of all pixels in the initial infrared temperature measurement area and its neighborhood are input into a clustering algorithm for processing. Pixels with temperatures significantly higher than the background ambient temperature are aggregated into candidate areas. Among the candidate areas, the area closest to the center point of the initial position detection box is selected as the high-temperature core area.

[0011] Furthermore, the rule for determining that the temperature is significantly higher than the background ambient temperature is as follows: calculate the arithmetic mean and standard deviation of the temperature values ​​of all pixels in the initial infrared temperature measurement area and its neighborhood; obtain the significance threshold by adding the product of the sensitivity coefficient and the standard deviation to the arithmetic mean; if the temperature is greater than the significance threshold, it is determined to be a temperature significantly higher than the background ambient temperature.

[0012] Furthermore, the input to the lightweight object detection model in S2 is a registered visible light RGB image. A feature extraction network with CSPDarknet as its backbone performs multi-scale feature encoding on the image, outputting feature maps at different scales. These feature maps are then input to a feature fusion network, which includes a RepGFPN structure and a multi-scale linear attention module. The feature fusion network performs cross-scale feature information fusion and enhancement, outputting three enhanced multi-scale feature maps. Each enhanced multi-scale feature map is input to a decoupling head, which processes classification and regression tasks in parallel through independent convolutional branches. The output is a prediction result generated by the decoupling head, containing the target box location, confidence level, and class probability. After post-processing non-maximum suppression, the initial location detection box is obtained.

[0013] Furthermore, the multi-scale linear attention module in the feature fusion network embeds an ECA channel attention mechanism in the upsampling and downsampling operations of the feature fusion network, and the processing steps include: The intermediate feature map in the input feature fusion network has a channel dimension of D. This intermediate feature map includes feature information consisting of the shape, structure, and contextual relationship of the transformer terminals with surrounding components. Global average pooling is performed on the intermediate feature map to compress the spatial dimension, generating channel description vectors that characterize the importance of different channels in terminal target identification. A one-dimensional convolution kernel is used to convolve the channel description vectors, adaptively capturing the correlation between key channels corresponding to terminal targets of different sizes at different voltage levels. The size of the convolution kernel... k The adaptive expression is: , in, This is an operation to find the nearest odd number; D is the channel dimension; and Preset parameters for convolution; The enhanced multi-scale feature map is obtained by multiplying the generated channel weight coefficients with the original input feature map channel by channel.

[0014] Furthermore, in S4, the pixel grayscale value Convert to temperature value The expression is: , in, This is the infrared temperature measurement area. For radiation response gain, This is the offset for compensating for ambient temperature. Furthermore, the radiation response gain and the ambient temperature compensation offset are obtained through linear fitting, specifically: at least two known temperature points are obtained from the temperature scale on the infrared image, and the gray values ​​of the known temperature points in the infrared image are obtained; a univariate linear regression is performed with the gray values ​​of the known temperature points as independent variables and the corresponding known temperature values ​​as dependent variables; the slope of the fitted line is used as the radiation response gain; and the intercept of the fitted line on the temperature axis is used as the ambient temperature compensation offset.

[0015] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0016] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) The detection process has been automated and intelligent, improving detection efficiency and reliability: This invention introduces a lightweight target detection model to automatically identify terminal positions and realizes automatic temperature extraction and diagnosis based on multimodal image fusion and dynamic optimization mechanism. It transforms the traditional operation mode that relies on human experience into an end-to-end automatic processing flow, overcoming the defects of low efficiency, strong subjectivity and easy misjudgment of manual inspection. This makes the detection process independent of the operator's experience and the results are consistent. It is particularly suitable for large-scale and high-frequency substation inspection scenarios, reducing the dependence on highly skilled personnel and ensuring the objectivity and consistency of inspection results. (2) Improved accuracy and robustness of temperature detection in complex field environments: By using visible light images for high-precision target positioning and using the registration relationship between infrared and visible light images to map the positioning results to infrared images, the problem of inaccurate temperature measurement area caused by blurred target boundaries and missing structural information when relying solely on infrared images is effectively solved. Furthermore, through the dynamic optimization mechanism of the detection frame based on temperature distribution consistency, the temperature measurement area can be automatically corrected when the terminal is partially blocked or there is interference from complex background heat sources, ensuring that the extracted temperature value truly reflects the heating status of the terminal body, thereby improving the accuracy of fault diagnosis, accurately detecting latent heating defects in advance, preventing equipment burnout or power outage accidents caused by terminal overheating, and ensuring the safe and stable operation of the power grid. (3) It has good engineering applicability and system integration, and is easy to deploy and apply on site: The lightweight model design adopted in this invention introduces the ECA attention mechanism, which reduces the computational overhead and enables the whole algorithm to be deployed on edge computing devices such as handheld industrial tablets. Combined with the integrated hardware and software design, it forms a portable and low-power integrated detection terminal, which solves the problems of bulky traditional detection equipment, PC-based processing and inconvenience. It facilitates the on-site operation of inspection personnel, realizes the on-site detection mode of instant diagnosis, and lays the foundation for integration with the background data processing system and the realization of long-term management and intelligent analysis of detection data. Attached Figure Description

[0018] Figure 1 A flowchart for an automatic detection method for abnormal transformer terminal temperature; Figure 2 This is a network structure diagram of a lightweight object detection model; Figure 3 This is a schematic diagram of the decoupling head structure used in the lightweight target detection model; Figure 4 A schematic diagram of the ECA channel attention mechanism; Figure 5 This is a schematic diagram of the criteria for judging the level of heat-induced defects. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] Example 1 While deep learning-based visible light target detection technology has made significant progress in general object recognition, and some researchers have attempted to apply it to power equipment inspection, its application in the specific scenario of transformer terminal temperature detection still faces challenges. First, while visible light models excel at recognizing the shape and structure of terminals, the result is a purely "visual bounding box," unable to directly provide the corresponding temperature data. To achieve temperature measurement, operators still need to manually compare and correlate the detection box with the infrared image, a process that doesn't fundamentally eliminate manual operation and significantly reduces automation. On the other hand, directly using infrared images for terminal target detection results in generally low positioning accuracy due to the inherent low resolution, blurred target edges, and lack of rich texture details in infrared images. The model struggles to accurately distinguish terminal edges from the background, or is prone to misjudging adhesion when different terminals are too close together. This positioning error directly impacts subsequent temperature extraction steps, resulting in temperature values ​​that may not originate from the conductive heating element of the terminal itself, but rather include the background or adjacent components, making the final anomaly diagnosis unreliable.

[0021] This embodiment addresses the problems in existing technologies, such as the separation of infrared temperature measurement and target detection, reliance on manual operation in the detection process, limited accuracy due to subjective judgment, and lack of portability of the detection equipment. Furthermore, in power inspection scenarios involving terminal obstruction, complex backgrounds, and interference from multiple devices, traditional methods struggle to achieve efficient, accurate, and intelligent temperature detection and fault location. This embodiment, however, introduces a target detection algorithm, fuses infrared and RGB image information, performs dynamic optimization, and integrates a lightweight inference module and interactive interface to construct an integrated handheld detection terminal. This terminal is suitable for multiple scenarios including power operation and maintenance, emergency inspection, and equipment quality inspection, demonstrating significant engineering practical value and promising prospects for widespread application.

[0022] like Figure 1 The diagram illustrates an automatic detection method for abnormal transformer terminal temperature according to this embodiment. The specific steps include: S1. Synchronously acquire registered infrared and visible light RGB images of the transformer terminal area; S2. Input the visible light RGB image into the lightweight target detection model to obtain the initial position detection box of the transformer terminal in the visible light RGB image; S3. Map the initial position detection box to the infrared image to obtain the initial infrared temperature measurement area; in the initial infrared temperature measurement area and its neighborhood, identify the high-temperature core area based on the pixel temperature distribution; perform spatial consistency verification on the high-temperature core area and the initial position detection box, and optimize the initial position detection box according to the verification result to generate an optimized infrared temperature measurement area. S4. Based on the radiation response gain and ambient temperature compensation offset, convert the pixel grayscale values ​​in the optimized infrared temperature measurement area into temperature values, and calculate the representative temperature of the optimized infrared temperature measurement area. S5. Based on the representative temperature and the criteria for judging heating defects, output the diagnostic results of abnormal temperature at the transformer terminals.

[0023] In S1, after detection is initiated, the infrared module and the visible light RGB camera simultaneously acquire images of the transformer terminals. The image acquisition module receives both RGB and infrared images; this dual-modal acquisition ensures complementary structural and temperature information, resolving the boundary blurring issue of single infrared images. The acquisition resolution can be set to 1280×720 to balance processing speed and accuracy. (Regarding the infrared image...) With visible light RGB image Feature point sets are extracted from the above: , in, For infrared image feature point set, For the feature point set of visible light RGB image, and The coordinates of feature points in the infrared image and the visible light RGB image with timestamps. and The first in the infrared image The coordinates of the feature points and Visible light RGB image The Middle The coordinates of each feature point. To ensure the temporal synchronization of the two images, a time difference between acquisition is required. Not exceeding the set synchronization threshold Solve timing misalignment issues caused by hardware latency.

[0024] The neighborhood in S3 is a region centered on the initial position detection box, extending outwards by a boundary determined by the size ratio of the initial position detection box. The boundary is determined based on the width or height of the initial position detection box.

[0025] The specific steps for identifying the high-temperature core region include: The temperature values ​​of all pixels in the initial infrared temperature measurement area and its neighborhood are input into a clustering algorithm for processing. Pixels with temperatures significantly higher than the background ambient temperature are aggregated into candidate regions. Among the candidate regions, the region closest to the center point of the initial detection box is selected as the high-temperature core region.

[0026] The rule for determining a temperature that is significantly higher than the background ambient temperature is as follows: Calculate the arithmetic mean and standard deviation of the temperature values ​​of all pixels in the initial infrared temperature measurement area and its neighborhood; obtain the significance threshold by adding the product of the sensitivity coefficient and the standard deviation to the arithmetic mean; if the temperature is greater than the significance threshold, it is determined to be a temperature that is significantly higher than the background ambient temperature.

[0027] When encountering situations where the terminal is partially obscured, has strong surface reflection, or is surrounded by similarly structured interference in the background, the visible light model is prone to problems such as positioning frame misalignment, dimensional inaccuracy, or decreased confidence. In such cases, the temperature value extracted based on the incorrect positioning frame will not accurately reflect the heating status of the terminal body, and may even misjudge the temperature of background heat sources or non-conductive components as the terminal temperature, leading to missed or false alarms. This embodiment, however, constructs a feedback loop that mutually verifies visible light semantic information and infrared temperature information, enabling self-correction in complex scenarios. When the visible light detection frame fails to completely cover the terminal's heating body due to obstruction, reflection, or model errors, the system can automatically identify the high-temperature core area based on the actual temperature distribution in the infrared image and precisely optimize the initial detection frame accordingly. This mechanism ensures that the final image area used for temperature extraction closely matches the actual heating part of the terminal, fundamentally solving the problem of inaccurate temperature measurement data caused by positioning deviations and improving the reliability and robustness of fault diagnosis results.

[0028] like Figure 2 The diagram shows the network structure of the lightweight target detection model in this embodiment. The YOLO series models are lightweighted using the E-ViT structure, achieving efficient target detection on mobile devices. This lightweight target detection model automatically identifies the location of transformer terminals, reducing reliance on human experience. It includes a feature extraction network, a feature fusion network, and a decoupling head, all using CSPDarknet as the backbone for infrared image input. CSPDarknet is an optimized convolutional neural network structure with deep layers and high feature representation capabilities, capable of capturing subtle differences and important patterns in images.

[0029] The feature extraction network incorporates the RepGFPN structure, improving upon the concept of Feature Pyramid Network (FPN) for object detection. It more efficiently fuses multi-scale features, crucial for capturing high-level semantics and low-level spatial details. Within the Fusion Block, structural units consisting of multiple 1x1 convolutional layers repeatedly appear. These convolutional layers not only effectively adjust the channel dimensions of the feature maps but are typically followed by batch normalization (BN) layers and activation functions (Act). This "convolution-normalization-activation" stacking pattern helps stabilize the training process and introduces non-linearity, thereby enhancing the model's expressive power. The Fusion Block also includes a lightweight visual Transformer structure (E-ViT).

[0030] Figure 3 This is a schematic diagram of the decoupling head structure used in a lightweight object detection model. The decoupling head is mainly used to determine feature points in the feature layer, decoupling the classification and regression branches, and then integrating them during prediction. First, a 1×1 convolution is performed to convert the feature map channels to 256 (dimensionless), and then two 3×3 convolutions are performed to achieve classification and regression prediction respectively. Reg is the predicted bounding box position result, where four parameters include the coordinates of the predicted bounding box center point and the width and height (w, h) information; Obj is the confidence result, where one parameter determines the probability that the predicted bounding box contains an object; Cls is the object category determination result, where C parameters determine the type of object contained.

[0031] The multi-scale linear attention module in the feature fusion network embeds an ECA channel attention mechanism during the upsampling and downsampling operations of the feature fusion network, as illustrated in the diagram below. Figure 4 As shown. The processing steps of the ECA channel attention mechanism include: The intermediate feature map in the input feature fusion network has a channel dimension of D. The intermediate feature map includes feature information consisting of the shape, structure, and contextual relationship of the transformer terminals with surrounding components. Global average pooling is performed on the intermediate feature map to compress the spatial dimension, generating channel description vectors that characterize the importance of different channels in terminal target identification. A one-dimensional convolution kernel is used to convolve the channel description vectors, and a cross-channel interaction is performed with the sigmoid function to adaptively capture the key channel correlations corresponding to terminal targets of different sizes at different voltage levels. The size of the convolution kernel is... k The adaptive expression is: , Right now: , in, This is an operation to find the nearest odd number; D is the channel dimension; and Preset parameters for convolution; The enhanced multi-scale feature map is obtained by multiplying the generated channel weight coefficients with the original input feature map channel by channel.

[0032] The data processing of the lightweight object detection model specifically includes: inputting a registered visible light RGB image, performing multi-scale feature encoding on the image through a feature extraction network with CSPDarknet as the backbone, and outputting feature maps of different scales; inputting the feature maps of different scales into a feature fusion network, which includes a RepGFPN structure and a multi-scale linear attention module; the feature fusion network performs cross-scale feature information fusion and enhancement, and outputs three enhanced multi-scale feature maps; inputting each enhanced multi-scale feature map into a decoupling head, which processes classification and regression tasks in parallel through independent convolutional branches; outputting the prediction results generated by the decoupling head, which include the target box position, confidence, and class probability, and obtaining the initial location detection box after post-processing non-maximum suppression.

[0033] Compared to the standard YOLO model, the lightweight object detection model proposed in this embodiment significantly reduces the number of parameters from approximately 35M to approximately 6M, a reduction of over 80%. This results in a corresponding reduction in model size from approximately 140MB to approximately 24MB, greatly facilitating deployment on handheld devices with limited storage and memory resources. In terms of computational resources (FLOPs), this lightweight object detection model also achieves significant optimization, decreasing from 80G to approximately 15G, a reduction of over 81%. This means less computational power is required to process a single image, resulting in lower power consumption and effectively extending the device's battery life.

[0034] In S4, based on the optimized infrared temperature measurement area, the terminal area temperature value is extracted through hotspot clustering and grayscale fitting, accurately obtaining temperature features from the infrared image and avoiding errors from manual extraction. The radiation response gain and ambient temperature compensation offset are obtained through linear fitting, specifically: at least two known temperature points are obtained from the temperature scale built into the infrared image, along with the corresponding grayscale values ​​of these known temperature points in the infrared image; a univariate linear regression is performed with the grayscale values ​​of the known temperature points as the independent variable and their corresponding known temperature values ​​as the dependent variable; the slope of the fitted line is used as the radiation response gain; and the intercept of the fitted line on the temperature axis is used as the ambient temperature compensation offset.

[0035] Pixel grayscale value Convert to temperature value The expression is: , in, This is the infrared temperature measurement area. For radiation response gain, This is the offset for compensating for ambient temperature.

[0036] The criterion for judging overheating defects is based on the temperature characteristic information of the equipment to accurately identify infrared defects in live equipment. The extracted temperature characteristic information mainly includes the maximum temperature of the terminal surface area, the temperature rise between the terminal surface temperature and the ambient temperature reference, and the temperature difference and relative temperature difference between different detection areas of the terminal. The formula for calculating the relative temperature difference is as follows: , in, , This refers to the temperature rise and temperature of the heating point; , This refers to the temperature rise and temperature at the corresponding normal points; The ambient temperature serves as a reference. After extracting temperature information from different regions, the system performs threshold judgments on the temperature characteristic information based on the thermal defect judgment criteria, thereby achieving automatic detection of thermal defects. The thermal defect level judgment criteria are as follows: Figure 5 As shown, the specific diagnostic standards are based on DL / T664--2016 "Specification for Infrared Diagnosis of Live Equipment".

[0037] The backend data processing system, as a crucial foundation of the method of this invention, undertakes tasks such as long-term storage, statistical analysis, fault tracking, and report management of detection data. The backend data processing system is typically deployed on a central server within the enterprise or maintenance unit, achieving real-time data transmission and remote interaction through a wireless communication module between the industrial tablet PC and the system. During data upload, the detection device packages the infrared images, RGB images, and detection results (including target bounding box coordinates, temperature values, diagnostic conclusions, etc.) collected in each inspection task into a standardized data format (such as JSON / XML) and uploads it to the backend server. After receiving the data, the backend system executes the following specific process: Data import and archiving: Uploaded data is parsed, indexed by metadata such as testing time, testing equipment number, and testing personnel ID, and stored uniformly in the historical testing database. The database supports incremental updates, fuzzy search, and breakpoint resume, ensuring data integrity and reliability.

[0038] Historical analysis and trend modeling: The backend system analyzes key detection indicators (such as the maximum temperature rise of a certain terminal, temperature difference change trend, etc.) based on time series models. Potential degradation trends are identified through sliding window averaging and anomaly detection algorithms, providing data support for equipment condition prediction and preventative maintenance.

[0039] Multi-terminal data aggregation and comparative analysis: Supports vertical analysis of data from multiple testing cycles on the same device, as well as horizontal comparison between different devices. The system can bind testing results to substation locations via a GIS map interface, enabling spatial visualization of equipment health status within the area.

[0040] Automatic report generation and export: The system supports one-click generation of standard inspection reports that meet the requirements of power industry inspections. The report content includes inspection time, equipment information, inspection images, inspection result tables, anomaly annotation diagrams, diagnostic conclusions, and handling suggestions, and supports exporting to PDF or Word format for archiving or handover.

[0041] Interface and Integration Capabilities: The backend system provides standard API interfaces, which facilitates integration with the power company's asset management system, SCADA system, data platform, etc., to achieve cross-platform data sharing and collaborative operations.

[0042] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0043] Example 2 This embodiment uses the method provided in Embodiment 1 for simulation and implementation. By comparing it with existing manual infrared thermal imager inspection and fixed PC-based YOLO inspection systems, it demonstrates improved technical performance in terms of accuracy, portability, automation, and inspection efficiency. The experimental results are shown in Table 1. As shown in Table 1, compared with manual inspection and fixed PC-based inspection systems, the handheld inspection device of this invention exhibits significant advantages in inspection accuracy, efficiency, and portability. This invention increases the terminal identification accuracy (mAP) to 92.3% and the temperature extraction accuracy to 94.6%, while reducing the average single-point inspection time to 2.7 seconds, achieving an order-of-magnitude improvement in inspection efficiency. Furthermore, the device's integrated 1.5kg design supports handheld operation and features real-time result display, background data synchronization, and adaptability to complex scenarios, comprehensively overcoming the inherent defects of traditional equipment such as bulkiness, separation, and reliance on manual labor.

[0044] Table 1. Experimental results comparing the method of the present invention with existing technologies. Furthermore, this embodiment also applies the method from Embodiment 1 to different voltage levels, and the results are shown in Table 2. For average accuracy, This represents the average precision. The frame rate is shown in the figure. It is evident that the device described in this invention exhibits excellent generalization performance and stability on transformer equipment at different voltage levels (110kV to 1000kV). Its terminal identification precision and recall consistently remain above 94.2% and 96.8%, respectively, while the average precision (AP) remains stable between 91.8% and 92.6%, demonstrating high robustness of the display model under various operating conditions. Furthermore, the system achieves a processing speed of up to 51.2 frames per second in typical power distribution scenarios (110kV), fully meeting the requirements for real-time on-site processing, proving that this invention has good engineering applicability and promotional value.

[0045] Table 2. Detection results of the method in this embodiment under different voltages. Furthermore, this invention possesses end-to-end fully automated processing capabilities, from image acquisition, target detection, temperature fitting, defect diagnosis to result visualization and report generation. This significantly reduces reliance on human experience and the risk of subjective misjudgment, supporting batch automatic detection and result output for 10 to 15 terminals. The modularly designed lightweight detection model is compatible with various mobile platforms and, through standardized data structures and backend API interfaces, supports seamless integration with existing asset management systems, SCADA systems, and data platforms of power companies, meeting the needs of enterprise-level digital inspection management. Simultaneously, it effectively increases the inspection frequency and fault prevention capabilities of key equipment in substations, enhancing the safety and stability of power grid operation. In pilot applications, this system can help a single substation save 30%–50% of manual inspection time and maintenance costs annually, demonstrating significant potential for cost reduction and efficiency improvement after large-scale deployment. The integrated handheld detection device features a low-power design (power <10W), suitable for continuous use in field scenarios. Combined with its fault prediction and analysis functions, it helps achieve preventative maintenance and energy-saving operation of power equipment.

[0046] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0047] Multiple components in the device are connected to an I / O interface, including: input units such as a keyboard, mouse, etc.; output units such as various types of displays, speakers, etc.; storage units such as disks, optical disks, etc.; and communication units such as network interface cards, modems, wireless transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or the communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method of the present invention described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the method of the present invention by any other suitable means (e.g., by means of firmware).

[0048] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0049] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0050] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0051] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for automatically detecting abnormality of a temperature of a transformer terminal, characterized by, The specific steps include: S1, synchronously collecting a registered infrared image and a visible light RGB image of a terminal area of a transformer; S2, inputting the visible light RGB image into a lightweight target detection model to obtain an initial position detection box of the terminal of the transformer in the visible light RGB image; S3, mapping the initial position detection box to the infrared image to obtain an initial infrared temperature measurement area; identifying a high-temperature core area in the initial infrared temperature measurement area and a neighborhood thereof based on pixel temperature distribution; performing spatial consistency verification on the high-temperature core area and the initial position detection box, optimizing the initial position detection box according to a verification result, and generating an optimized infrared temperature measurement area; S4, converting pixel gray value in the optimized infrared temperature measurement area into a temperature value based on a radiation response gain and an ambient temperature compensation offset, and calculating a representative temperature of the optimized infrared temperature measurement area; and S5, outputting a temperature anomaly diagnosis result of the terminal of the transformer according to the representative temperature and in combination with a heating defect judgment standard.

2. The method of claim 1, wherein the transformer terminal temperature is determined by a temperature sensor. The neighborhood in S3 is a region formed by expanding outward from the center of the initial position detection box by a boundary determined by the size proportion of the initial position detection box, and the boundary is determined according to the width or height of the initial position detection box.

3. The method of claim 1, wherein the transformer terminal temperature is determined by a temperature sensor. The specific steps of identifying the high-temperature core area include: Inputting temperature value data of all pixel points in the initial infrared temperature measurement area and the neighborhood thereof into a clustering algorithm for processing, and aggregating pixel points with temperature significantly higher than the background ambient temperature into a candidate area; and selecting, in the candidate area, an area closest to the center point of the initial position detection box as the high-temperature core area.

4. The method of claim 3, wherein the temperature of the transformer terminal is measured by a temperature sensor. The judgment rule that the temperature is significantly higher than the background ambient temperature is that: calculating an arithmetic mean and a standard deviation of temperature value data of all pixel points in the initial infrared temperature measurement area and the neighborhood thereof; obtaining a significance threshold according to the arithmetic mean plus the product of a sensitivity coefficient and the standard deviation; and determining that the temperature is significantly higher than the background ambient temperature if the temperature is greater than the significance threshold.

5. The method of claim 1, wherein the transformer terminal temperature abnormality is detected automatically. The input of the lightweight target detection model in S2 is the registered visible light RGB image, the image is encoded with multi-scale features through a feature extraction network with CSPDarknet as the backbone, and feature maps of different scales are output; the feature maps of different scales are input into a feature fusion network, the feature fusion network includes a fusion RepGFPN structure and a multi-scale linear attention module; The feature fusion network performs cross-scale feature information fusion and enhancement, and outputs three enhanced multi-scale feature maps; each of the enhanced multi-scale feature maps is input into a decoupling head, the decoupling head processes classification and regression tasks in parallel through independent convolution branches; the prediction result containing the position of the target box, the confidence and the class probability generated by the decoupling head is obtained after non-maximum suppression of post-processing, and the initial position detection box is obtained.

6. The method of claim 5, wherein the step of detecting the abnormal temperature of the transformer terminal is performed by using a temperature sensor. The multi-scale linear attention module in the feature fusion network is that, in the up-sampling and down-sampling operations of the feature fusion network, an ECA channel attention mechanism is embedded, and the processing steps include: An intermediate feature map in the input feature fusion network, a channel dimension of which is D, the intermediate feature map comprising feature information composed of a shape, a structure of a transformer terminal and a context relationship with surrounding components; performing a global average pooling operation on the intermediate feature map to compress a spatial dimension and generate a channel description vector for representing importance of different channels in terminal target identification; performing a convolution operation on the channel description vector using a one-dimensional convolution kernel to adaptively capture inter-key channel correlations corresponding to terminal targets of different sizes under different voltage levels, a size of the convolution kernel being k An adaptive expression of the convolution kernel is: ， wherein, is the nearest odd operation; D is the channel dimension; and is the convolution preset parameter; The generated channel weight coefficients are multiplied with the original input feature map channel by channel to obtain an enhanced multi-scale feature map.

7. The method of claim 1, wherein the transformer terminal temperature anomaly is detected automatically. In the S4, the pixel gray value converted into a temperature value The expression is: ， wherein, is an infrared temperature measurement region, is a radiative response gain, is an ambient temperature compensation offset.

8. The method of claim 7, wherein the transformer terminal temperature is determined by a temperature sensor. The radiation response gain and the ambient temperature compensation offset are obtained through linear fitting, specifically: at least two known temperature points obtained from a temperature scale of the infrared image and corresponding gray values of the known temperature points in the infrared image; performing one-dimensional linear regression fitting with the gray values of the known temperature points as independent variables and the corresponding known temperature values as dependent variables; taking the slope of the obtained straight line as the radiation response gain; and taking the intercept of the obtained straight line on the temperature axis as the ambient temperature compensation offset.

9. An electronic device comprising a memory and a processor, said memory having stored thereon a computer program, characterized in that, The processor implements the method of any one of claims 1-8 when executing the program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1-8.

Images