Substation equipment thermal fault diagnosis method and system based on infrared image

Abstract

The invention relates to the technical field of substation equipment fault diagnosis, in particular to a substation equipment thermal fault diagnosis method and system based on an infrared image. The invention discloses a substation equipment thermal fault diagnosis method based on an infrared image. The method comprises the following steps: acquiring an equipment temperature distribution image through an infrared thermal imager; carrying out denoising, contrast enhancement and normalization preprocessing on the image; extracting features such as temperature anomaly, temperature gradient and hot spot areas; a YOLOv13 model is adopted to identify the equipment type; inputting the features into a deep learning model for fault classification; and implementing multi-level alarm according to the classification result confidence. According to the method, temperature gradient analysis and a regional dynamic contrast enhancement technology are creatively fused, so that the fault detection precision and the early warning capability are remarkably improved, and intelligent diagnosis and graded early warning of the thermal fault of the substation equipment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of substation equipment fault diagnosis, and in particular to a substation equipment thermal fault diagnosis method and system based on infrared images. BACKGROUND

[0002] As an important part of the power system, substations bear the important task of power transmission and distribution, and the stable operation of equipment is crucial to the safety of the power system. With the increasing operation load of the power system, the risk of substation equipment failure also increases. Traditional substation equipment fault detection methods mainly rely on manual inspection, which is low in efficiency, poor in real-time performance, and easily affected by human factors, resulting in delayed fault detection and even equipment damage or shutdown in severe cases.

[0003] Infrared thermal imaging technology, as a non-contact, fast and efficient detection method, can monitor the temperature distribution of the equipment surface in real time and detect thermal faults caused by abnormal load, aging and other problems during equipment operation. Therefore, equipment fault diagnosis technology based on infrared images has gradually attracted more and more attention and application in substations. Although existing equipment fault diagnosis technology based on infrared images can identify potential fault points through temperature changes, it still has certain limitations, such as insufficient automated identification capability, inability to achieve precise fault diagnosis, and lack of standardized diagnostic methods. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a substation equipment thermal fault diagnosis method based on infrared images to solve the problem of low fault diagnosis accuracy and lack of standardized and automated diagnostic methods in the prior art.

[0005] The technical solution adopted by the present application to solve the technical problem is: a substation equipment thermal fault diagnosis method based on infrared images, comprising the following steps: S1, acquiring an infrared image of the substation equipment by an infrared thermal imager, the infrared image including the overall appearance of the equipment and the temperature distribution of its key parts, wherein the value corresponding to each pixel point represents the temperature of the point; S2, performing noise reduction, contrast enhancement and normalization processing on the acquired infrared image to obtain a preprocessed image; S3, extracting thermal fault features from the preprocessed image, including temperature anomaly detection, temperature gradient analysis, hot spot positioning and contrast enhancement; S4, using a YOLOv13 model to identify the equipment in the preprocessed infrared image to generate a detection result of the equipment; S5. Input the extracted thermal fault features into the deep learning model to perform fault diagnosis and classification, and output the fault type and its corresponding probability. S6. Generate alarm information based on the diagnostic results and provide feedback to maintenance personnel through a multi-level alarm mechanism.

[0006] By using infrared thermal imaging technology to achieve non-contact temperature monitoring, and combining image preprocessing, feature extraction, equipment identification, fault classification and multi-level alarms, a standardized automated diagnostic solution has been formed. By combining traditional infrared detection with deep learning, the problems of low efficiency and high false detection rate of manual inspection have been solved, and the early detection rate and diagnostic accuracy of thermal faults in substation equipment have been significantly improved.

[0007] According to one embodiment of the present invention, the key components of the equipment in step S1 include insulator strings, voltage transformers, current transformers, disconnect switches, circuit breakers, and transformer bushings; the pixels of the original infrared image are represented as follows: .

[0008] The specific target locations for infrared image acquisition (such as insulator strings, instrument transformers, etc.) are clearly defined to ensure data coverage of critical, fault-prone equipment in the substation; the original image pixels are defined. This provides a foundation for subsequent temperature matrix analysis and mathematical modeling, and enhances the operability of the method and the rigor of its technical details.

[0009] According to an embodiment of the present invention, in step S2, the image is denoised by median filtering or mean filtering to normalize the temperature values ​​in the image to a standard temperature range; wherein, the median filtering formula is: ; in, This represents the preprocessed image pixels. This represents the processed image pixels. This indicates the position of the pixel in the processed image. This indicates that it contains the original image pixels. The neighborhood of the region, where median indicates the median filtering operation.

[0010] The core operations of image preprocessing (median filtering and normalization) are described in detail. Image quality is improved and environmental interference is eliminated through denoising and standardization. The introduction of the median filtering formula effectively suppresses impulse noise, while temperature normalization unifies the data scale, which facilitates the stability of subsequent feature extraction and model training.

[0011] According to an embodiment of the present invention, in step S3, the temperature anomaly detection specifically involves first setting the normal temperature range to [T]. min ,T maxNext, determine whether the temperature T(x, y) of a certain region in the image is within the normal temperature range. If T(x, y) > T... max Or T(x, y) < T min If so, then the area is a thermal failure area; Temperature gradient analysis specifically involves calculating the temperature gradient at each pixel location in the image to identify areas of abnormal temperature change on the device surface, which are the hotspots of the fault. The temperature gradient... The calculation formula is: ; Hotspot localization specifically involves using a region growing algorithm to locate hotspot regions S. The formula for calculating hotspot region S is as follows: , Among them, T hotspot G is the temperature threshold for the hotspot. min Set as an empirical threshold or adaptively through machine learning; The contrast enhancement specifically involves performing regional dynamic contrast enhancement on the hotspot region S, resulting in increased image pixel contrast. The calculation formula is: ; in, This represents the average value of pixels in the hotspot region. The standard deviation of pixels in the hot spot region, where γ is the enhancement factor.

[0012] Temperature anomaly detection via threshold [T] min ,T max Quickly locate abnormal areas to directly reflect equipment overload or aging issues; temperature gradient analysis utilizes gradient formulas. Identify regions of sudden temperature changes and capture potential faults that traditional thresholding methods may miss; hotspot localization combines temperature and gradient conditions. This reduces false positives; regional dynamic contrast enhancement only enhances the hot spot area S, avoiding edge distortion caused by global processing, highlighting fault characteristics while retaining background information, and achieving high-precision, low-noise thermal fault location and enhancement.

[0013] According to an embodiment of the present invention, step S4 specifically includes: S41. Obtain infrared images of substation equipment and save them as a dataset; S42. Use the Labelimg tool to label and divide the dataset into training, validation and test sets; S43. Train the YOLOV13 neural network model using the training set, and iteratively update the parameters to generate the trained YOLOV13 neural network model. S44. Input the real-time acquired infrared images of the substation equipment into the trained YOLOV13 neural network model to generate the detection results of the substation equipment.

[0014] The dataset was labeled using the Labelimg tool to ensure the accuracy of the model in identifying targets; parameters were iteratively optimized to improve the generalization ability of equipment identification and adapt to different substation scenarios; the trained model was deployed in the actual monitoring system to achieve automated equipment classification; the efficiency of YOLOv13 ensured real-time performance and was suitable for parallel detection of multiple devices in complex substation environments.

[0015] According to an embodiment of the present invention, step S5 specifically involves using a convolutional neural network to evaluate and diagnose the state of the device based on different fault modes; wherein, the thermal fault feature input is composed of multiple features, which are used to construct a comprehensive feature vector X, expressed as: , Among them, T max For the maximum temperature, T avg The average temperature. For the temperature gradient, A S Where is the area of ​​the hotspot, eccentricity is the regional eccentricity, and elongation is the aspect ratio; The classification formula is: , Where C represents the final classification result, C i For category labels, Given input data X, it belongs to category C. i The posterior probability.

[0016] Integrated feature vector X fused temperature (T) max T avg Temperature gradient Geometric features (hotspot area, eccentricity, etc.) comprehensively characterize the physical and morphological properties of the fault. The classification formula is based on posterior probability. The decision-making process supports the differentiation of multiple fault types; through joint analysis of multi-dimensional features, the classification robustness is significantly improved.

[0017] According to one embodiment of the present invention, the posterior probability The calculation formula is: , in, Category C i The prior probability, It is the marginal likelihood, representing the total probability of observing data X, which is the weighted average of the posterior probabilities of all classes.

[0018] By clarifying the calculation logic of posterior probability through Bayes' theorem, the mathematical basis of model classification is explained, and the likelihood function is discussed. Reflects the correlation between fault characteristics and categories; prior probability By introducing weights from historical fault data, the problem of class imbalance is optimized; the interpretability of the model is enhanced, providing maintenance personnel with a quantitative basis for the probability of faults.

[0019] According to an embodiment of the present invention, the multi-level alarm mechanism in step S6 specifically includes: Let the predicted output P be: The model confidence score Conf is obtained as follows: , Alarm Level is defined as follows: .

[0020] By associating alarm levels (0-3) with model output confidence (Conf), progressive responses are achieved; low confidence levels (Conf < 0.5) do not trigger alarms, reducing the interference of false alarms on operation and maintenance; resource allocation is optimized to prioritize the handling of high-confidence faults (such as level 3 faults that require immediate repair), thereby improving operation and maintenance efficiency.

[0021] A substation equipment thermal fault diagnosis system based on infrared images is also provided, including: The data acquisition module is used to acquire infrared images of substation equipment using an infrared thermal imager; The image processing and analysis module is used for image preprocessing and thermal fault feature extraction; The substation equipment identification module is used to identify equipment types using the YOLOv13 model; The thermal fault diagnosis and prediction module is used for fault classification based on a deep learning model. The fault alarm and feedback module is used to generate multi-level alarm information based on the diagnostic results.

[0022] Infrared thermal imagers provide real-time monitoring to ensure data timeliness; an adaptive temperature gradient enhancement mechanism and a regional dynamic contrast enhancement algorithm are embedded to achieve automated feature extraction; YOLOv13 and deep learning models work together to complete equipment location and fault classification; a hierarchical feedback mechanism directly connects to the operation and maintenance system to form closed-loop management. The modular design facilitates system expansion and maintenance, adapting to the needs of substations of different sizes.

[0023] According to one embodiment of the present invention, the image processing and analysis module includes a temperature gradient adaptive enhancement mechanism and a regional dynamic contrast enhancement algorithm; the fault alarm and feedback module supports a multi-level alarm mechanism based on confidence level.

[0024] Temperature gradient adaptive enhancement mechanism and regional dynamic contrast enhancement algorithm ensure high sensitivity and low false alarm rate in thermal fault detection; multi-level alarm mechanism dynamically responds to fault risks, balancing operation and maintenance costs and safety.

[0025] The beneficial effects of this invention are: (1) Multi-feature fusion improves the comprehensiveness of fault diagnosis. By jointly analyzing and fusing temperature (T) max T avg Temperature gradient Geometric features (hotspot area, eccentricity, etc.) are used to construct a comprehensive feature vector X, overcoming the limitations of traditional methods that rely on a single temperature parameter; after multi-dimensional features are input into a convolutional neural network, the classification model... It can more accurately distinguish between fault types such as overheating, poor contact, and insulation damage; (2) Temperature gradient adaptive enhancement mechanism and regional dynamic contrast enhancement algorithm optimize detection accuracy. The temperature gradient adaptive enhancement mechanism optimizes detection accuracy through dual conditions. Filtering noise avoids misjudging sunlight reflection as a fault, while capturing early slight temperature rises; the regional dynamic contrast enhancement algorithm only dynamically enhances the contrast of hotspot areas S, avoiding edge blurring caused by global processing, making minor faults easier to identify; (3) The intelligent alarm mechanism enables efficient allocation of operation and maintenance resources, based on the confidence level of the model output. The system is divided into four alarm levels. At level 0 (Conf<0.5), low-probability events are ignored to reduce false alarm interference. At level 3 (Conf≥0.9), high-risk faults are responded to immediately. Maintenance personnel can prioritize handling high-confidence alarms to reduce the cost of ineffective inspections. (4) From data acquisition, preprocessing, equipment identification to fault classification, a closed-loop diagnostic process is formed to replace traditional manual inspection, improve efficiency and reduce human error. (5) The modular design of the substation equipment thermal fault diagnosis system based on infrared images ensures scalability, making it easy to adapt to different substation scales, and the algorithm parameters can be dynamically adjusted to adapt to complex field environments. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Figure 1 This is a flowchart of the thermal fault diagnosis method according to Embodiment 1 of the present invention.

[0028] Figure 2 This is a flowchart of image preprocessing in the thermal fault diagnosis method of Embodiment 1 of the present invention.

[0029] Figure 3 This is a flowchart of the thermal fault feature extraction process in the thermal fault diagnosis method of Embodiment 1 of the present invention.

[0030] Figure 4 This is a flowchart of the substation equipment identification process in the thermal fault diagnosis method of Embodiment 1 of the present invention.

[0031] Figure 5 This is a flowchart of the fault diagnosis and classification method in the thermal fault diagnosis method of Embodiment 1 of the present invention.

[0032] Figure 6 This is a flowchart of the fault alarm and feedback process in the thermal fault diagnosis method of Embodiment 1 of the present invention.

[0033] Figure 7 This is a structural block diagram of the thermal fault diagnosis system according to Embodiment 2 of the present invention. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0035] Example 1 like Figure 1 As shown, a method for diagnosing thermal faults in substation equipment based on infrared images includes the following steps: S1. Infrared images of substation equipment are acquired using an infrared thermal imager. The infrared images include the overall appearance of the equipment and the temperature distribution of its key parts. The value corresponding to each pixel represents the temperature of that point. S2. The acquired infrared image is de-noised, contrast-enhanced, and normalized to obtain the pre-processed image; S3. Extract thermal fault features from the preprocessed image, including temperature anomaly detection, temperature gradient analysis, hotspot localization, and contrast enhancement. S4. Use the YOLOv13 model to identify devices in the preprocessed infrared image and generate the detection results of the devices; S5. Input the extracted thermal fault features into the deep learning model to perform fault diagnosis and classification, and output the fault type and its corresponding probability. S6. Generate alarm information based on the diagnostic results and provide feedback to maintenance personnel through a multi-level alarm mechanism.

[0036] In step S1, the infrared imager is installed at the substation site to acquire temperature distribution images of the equipment surface in a remote, non-contact manner. The acquired images should include the overall appearance of the equipment and the temperature conditions of its key components, such as insulator strings, voltage transformers, current transformers, disconnect switches, circuit breakers, and transformer bushings. Assuming the temperature distribution on the equipment surface is represented by a two-dimensional image matrix, each infrared image pixel... The corresponding value represents the temperature at that point, where x and y represent the spatial coordinates of that point.

[0037] like Figure 2 As shown, step S2 preprocesses the acquired infrared image, including denoising, contrast enhancement, and normalization, to eliminate the influence of environmental factors (such as changes in lighting and climate) on image quality and improve image clarity and diagnostic accuracy. The main image preprocessing operations include denoising, contrast enhancement, and normalization. Denoising is performed by removing noise from the image through median filtering and mean filtering. The median filtering formula is: ;in, This represents the preprocessed image pixels. This represents the processed image pixels. This indicates the position of the pixel in the processed image. This indicates that it contains the original image pixels. The neighborhood is defined by median filtering. Normalization normalizes the temperature values ​​in the image to a standard temperature range, facilitating subsequent feature extraction.

[0038] like Figure 3 As shown, in step S3, temperature anomaly detection involves setting a threshold to determine which areas in the image have temperatures exceeding the normal operating range. Assume the set normal temperature range is [T...]. min ,T max If the temperature T(x, y) in a certain region satisfies T(x, y) > T max Or T(x, y) < T min If the temperature gradient is abnormal, the area is considered to have a thermal fault. Temperature gradient analysis determines the anomaly of changes on the device surface by calculating the temperature gradient at each location in the image. The calculation formula is: ; By analyzing temperature gradients, areas with drastic temperature changes can be identified, thereby pinpointing potential hotspots where faults may exist.

[0039] Hotspot localization uses image segmentation techniques and region growing algorithms to locate hotspots in an image that exhibit significant high temperatures. Assuming there is a hotspot region S in the image, the thermal fault characteristics of this region can be represented as: , among which, T hotspot The temperature threshold for hotspots is a common practice in existing technologies that relies solely on absolute temperature, easily missing areas where the temperature rises but does not exceed the threshold. Temperature gradients, however, emphasize the relative temperature rise trend, enabling earlier detection of potential faults. Therefore, this embodiment improves upon this by simultaneously satisfying both temperature and gradient conditions, modifying the calculation formula for the hotspot region S as follows: Among them, G min The threshold is set either empirically or adaptively through machine learning. In this way, not only the temperature itself (T(x,y)) is considered, but also the dynamic feature of the temperature gradient (∇T(x,y)) is introduced. Through gradient analysis, the sensitivity to blurred boundaries and small high-temperature areas is enhanced, which makes it easier to strengthen and classify only the selected hotspot areas S in subsequent image enhancement and classification steps, saving computing resources and focusing on key information.

[0040] Contrast enhancement uses histogram equalization to enhance the contrast of an image. Assuming the image's grayscale value range is [0, 255], the image pixels after contrast enhancement... It can be expressed by the following formula: This method results in low overall contrast of the infrared image when the temperature difference is small, easily masking minor faults. Therefore, this embodiment proposes a local region enhancement strategy, using dynamic range remapping for hotspot areas instead of traditional full-image equalization. The image pixels after contrast enhancement are modified. The calculation formula is: ; in, This represents the average value of pixels in the hotspot region. The standard deviation of pixels in hotspot areas, γ is an enhancement factor, which is adjustable to adapt to different device materials and backgrounds. Local statistics are introduced ( , Standardization is performed on hotspot areas rather than the entire image, avoiding the edge distortion problem caused by global equalization, which is common in image enhancement. Local enhancement is performed only on faulty areas, which can preserve the features of non-faulty areas to the greatest extent and not affect the background, providing a cleaner input for subsequent models.

[0041] like Figure 4 As shown, step S4 specifically involves: acquiring infrared images of substation equipment and saving them as a dataset; using the Labelimg tool to label and divide the dataset into a training set, a validation set, and a test set; training the YOLOv13 neural network model using the training set, iteratively updating the parameters to generate the trained YOLOv13 neural network model; and inputting the real-time acquired infrared images of substation equipment into the trained YOLOv13 neural network model to generate the detection results of the substation equipment (insulator strings, voltage transformers, current transformers, disconnect switches, circuit breakers, and transformer bushings).

[0042] like Figure 5As shown, step S5 uses a convolutional neural network to assess and diagnose the equipment's condition based on different fault modes (overheating, poor contact, insulation failure). Traditional diagnostic methods rely solely on single factors such as temperature or area. This embodiment integrates multiple feature inputs, including temperature anomalies, gradient information, and regional geometry, to construct a higher-dimensional feature vector for classification and diagnosis, improving model robustness and accuracy. The multiple feature inputs construct a comprehensive feature vector X, represented as: , Among them, T max For the maximum temperature, T avg The average temperature. For the temperature gradient, A S Where is the area of ​​the hotspot, eccentricity is the regional eccentricity, and elongation is the aspect ratio; The classification formula is: , Where C represents the final classification result, that is, the category label predicted by the model.

[0043] It is to find all categories C i In the context of conditional probability Largest category C i This is used as the predicted category; that is, the model selects the category with the highest probability as the classification result. i The category labels represent the different categories predicted by the model. For example, C1, C2, and C3 represent the categories of overheating, poor contact, and insulation failure, respectively. Given input data X, it belongs to category C. i The posterior probability indicates whether a sample belongs to class C given the input data X. i The probability of [the number of categories]. That is, for a given input data X, the model calculates the probability of [the number of categories] C for each category. i posterior probability And select the category with the highest probability as the prediction result. This embodiment incorporates spatial shape information. For example, elongated hot spots may indicate busbar overheating, while dot-shaped spots may indicate loose connections. This helps in more refined classification, allowing the subsequent classification model to differentiate between various fault types such as minor, severe, and false alarms based on these diverse inputs.

[0044] A Convolutional Neural Network (CNN) is a deep learning model that extracts image features through multiple convolutional layers, pooling layers, and fully connected layers. Finally, a Softmax layer maps the final features to probabilities for each class. In a CNN, the output of the last fully connected layer represents a score indicating whether a sample belongs to each class; then, the Softmax function transforms these scores into a probability distribution, and the class corresponding to the highest probability is the predicted class. Posterior probability. The posterior probability is usually calculated using Bayes' theorem. The calculation formula is: , in, It is in category C i Below, we observe the likelihood function of data X, which is what the model learns, representing the likelihood function given category C. i Given the condition, the probability of observing data X. It is category C i The prior probability represents prior knowledge that a sample belongs to that category. During training, this is usually estimated by the frequency of the category. Marginal likelihood represents the total probability of observing data X. It is a weighted average of the posterior probabilities of all classes. Since it is a constant, it is usually ignored in actual classification and is therefore the same for all classes.

[0045] like Figure 6 As shown, step S6 involves real-time monitoring of the substation equipment's operating status based on the diagnostic results, and timely feedback of fault information to maintenance personnel via an intelligent alarm system. The alarm information should include the fault type, fault location, and fault severity to facilitate subsequent handling by maintenance personnel.

[0046] In multi-class classification problems, many machine learning models output a probability value for each class. This probability value represents the confidence that the current sample belongs to that class. For fault classification models, these probability values ​​reflect the likelihood of a fault type. Assume the model outputs a class probability distribution: Where P1 and P2 represent the probabilities that the sample belongs to class 1 or 2, respectively, and k is the total number of classes. The confidence level Conf can be directly defined as the maximum probability of that class, expressed as: This represents the confidence level of the model's most likely category.

[0047] Different fault types have varying degrees of severity, and traditional single-threshold alarm mechanisms are prone to false alarms. This embodiment sets up a multi-level alarm strategy based on the probability confidence level output by the model, achieving intelligent hierarchical response. Specifically: Let the predicted output P be: The model confidence score Conf is obtained as follows: , Alarm Level is defined as follows: .

[0048] Among them, 0 corresponds to no alarm, which may be due to model uncertainty or background noise; 1 corresponds to low priority alarm, which may be a reminder that it is not a fault; 2 corresponds to medium priority alarm, which suggests checking; and 3 corresponds to high priority alarm, which requires immediate response.

[0049] Alarm levels are categorized based on the confidence level of the classification model, explicitly incorporating the model's uncertainty into the decision-making process to build a progressive alarm response mechanism, rather than a simple yes / no judgment. Maintenance personnel can then allocate inspection or remote intervention resources according to the level, improving efficiency.

[0050] Traditional thermal fault diagnosis typically relies on a single temperature threshold for judgment, while this embodiment constructs a multi-dimensional feature analysis system based on temperature, gradient, and geometric morphology. Specifically, it first integrates temperature gradient (… By combining temperature (T) and absolute temperature (G) under dual conditions, this approach addresses the insensitivity of traditional methods to progressive temperature rise faults (such as early-stage contact defects). This dynamic feature fusion requires a deep understanding of the physical evolution of thermal faults, and is not simply a combination of factors. Introducing geometric features (eccentricity, aspect ratio, etc.) for auxiliary diagnosis, such as identifying elongated hot spots to distinguish between busbar overheating and point-contact faults, requires a shift away from the conventional thinking that temperature equals fault, and necessitates design considerations based on equipment structural characteristics. The TGAE mechanism (Temperature Gradient Adaptive Enhancement Mechanism) adjusts the gradient threshold G... min Designed to adaptively adjust parameters (optimized through machine learning) rather than using fixed empirical values, this dynamic adjustment requires solving the challenge of distinguishing between the gradient features of environmental noise and real faults. The RDCEA (Regional Dynamic Contrast Enhancement Algorithm) breaks through the limitations of traditional global image enhancement, pioneering a method based on the statistical characteristics of hotspot regions (…). , The local enhancement strategy for YOLOv13 addresses the need for precise matching of the physical locality characteristics of faults, involving complex neighborhood statistical calculations and dynamic range mapping in its technical implementation. A data-driven closed loop is formed from feature extraction to classification and then to alarm. The multi-level alarm mechanism with confidence adjustment quantifies model uncertainty into a basis for operation and maintenance decisions. This design, which precisely binds probability output to actual response levels, requires a deep understanding of the business needs of power operation and maintenance. The transfer application of YOLOv13 to infrared images overcomes the limitations of the model originally designed for visible light images. Its implementation requires special parameter tuning and data augmentation of infrared features.

[0051] Example 2 like Figure 7 As shown, a substation equipment thermal fault diagnosis system based on infrared images includes: The data acquisition module is used to acquire infrared images of substation equipment using an infrared thermal imager; The image processing and analysis module is used for image preprocessing and thermal fault feature extraction; The substation equipment identification module is used to identify equipment types using the YOLOv13 model. The thermal fault diagnosis and prediction module is used for fault classification based on a deep learning model. The fault alarm and feedback module is used to generate multi-level alarm information based on the diagnostic results.

[0052] The image processing and analysis module includes a temperature gradient adaptive enhancement mechanism and a regional dynamic contrast enhancement algorithm; the fault alarm and feedback module supports a multi-level alarm mechanism based on confidence level.

[0053] The TGAE (Temperature Gradient Adaptive Enhancement) and RDCEA (Regional Dynamic Contrast Enhancement) algorithms are deeply integrated into the image processing module (rather than as independent functions), achieving pipeline optimization of preprocessing and feature extraction. A dual-model architecture (YOLOv13+CNN) for device identification and fault diagnosis ensures both real-time performance and improved accuracy through the separation of spatial localization and state analysis. The system supports G... min Online adjustment of key parameters such as γ. Deep integration of the multi-level alarm module with the operation and maintenance system enables an automated process from algorithm output to work order generation. This end-to-end system design goes beyond simple algorithm improvement. Traditional solutions typically treat infrared detection and fault diagnosis as independent components, while this embodiment achieves a performance-coupled improvement in detection and diagnosis through the collaborative design of temperature gradient analysis and local enhancement.

[0054] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for diagnosing thermal faults in substation equipment based on infrared images, characterized in that, Includes the following steps: S1. Infrared images of substation equipment are acquired using an infrared thermal imager. The infrared images include the overall appearance of the equipment and the temperature distribution of its key parts. The value corresponding to each pixel represents the temperature of that point. S2. The acquired infrared image is de-noised, contrast-enhanced, and normalized to obtain the pre-processed image; S3. Extract thermal fault features from the preprocessed image, including temperature anomaly detection, temperature gradient analysis, hotspot localization, and contrast enhancement. S4. Use the YOLOv13 model to identify devices in the preprocessed infrared image and generate the detection results of the devices; S5. Input the extracted thermal fault features into the deep learning model to perform fault diagnosis and classification, and output the fault type and its corresponding probability. S6. Generate alarm information based on the diagnostic results and provide feedback to maintenance personnel through a multi-level alarm mechanism.

2. The method for diagnosing thermal faults in substation equipment based on infrared images according to claim 1, characterized in that: The key components of the equipment in step S1 include insulator strings, voltage transformers, current transformers, disconnect switches, circuit breakers, and transformer bushings; the pixels of the original infrared image are represented as follows: .

3. The method for diagnosing thermal faults in substation equipment based on infrared images according to claim 1, characterized in that: In step S2, the image is denoised using median filtering or mean filtering to normalize the temperature values ​​in the image to a standard temperature range; the median filtering formula is as follows: ； in, This represents the preprocessed image pixels. This represents the processed image pixels. This indicates the position of the pixel in the processed image. This indicates that it contains the original image pixels. The neighborhood of the region, where median indicates the median filtering operation.

4. The method for diagnosing thermal faults in substation equipment based on infrared images according to claim 1, characterized in that: In step S3, the temperature anomaly detection specifically involves first setting the normal temperature range to [T]. min ,T max Next, determine whether the temperature T(x, y) of a certain region in the image is within the normal temperature range. If T(x, y) > T... max Or T(x, y) < T min If so, then the area is a thermal failure area; Temperature gradient analysis specifically involves calculating the temperature gradient at each pixel location in the image to identify areas of abnormal temperature change on the device surface, which are the hotspots of the fault. The temperature gradient... The calculation formula is: ； Hotspot localization specifically involves using a region growing algorithm to locate hotspot regions S. The formula for calculating hotspot region S is as follows: ， Among them, T hotspot G is the temperature threshold for the hotspot. min Set as an empirical threshold or adaptively through machine learning; The contrast enhancement specifically involves performing regional dynamic contrast enhancement on the hotspot region S, resulting in increased image pixel contrast. The calculation formula is: ； in, This represents the average value of pixels in the hotspot region. The standard deviation of pixels in the hot spot region, where γ is the enhancement factor.

5. The method for diagnosing thermal faults in substation equipment based on infrared images according to claim 1, characterized in that, Step S4 is as follows: S41. Obtain infrared images of substation equipment and save them as a dataset; S42. Use the Labelimg tool to label and divide the dataset into training, validation and test sets; S43. Train the YOLOV13 neural network model using the training set, and iteratively update the parameters to generate the trained YOLOV13 neural network model. S44. Input the real-time acquired infrared images of the substation equipment into the trained YOLOV13 neural network model to generate the detection results of the substation equipment.

6. The method for diagnosing thermal faults in substation equipment based on infrared images according to claim 1, characterized in that: Step S5 specifically involves using a convolutional neural network to assess and diagnose the equipment's status based on different fault modes. The thermal fault feature input comprises multiple features, which are combined to form a comprehensive feature vector X, expressed as: ， Among them, T max For the maximum temperature, T avg The average temperature. For the temperature gradient, A S Where is the area of ​​the hotspot, eccentricity is the regional eccentricity, and elongation is the aspect ratio; The classification formula is: , Where C represents the final classification result, C i For category labels, Given input data X, it belongs to category C. i The posterior probability.

7. The method for diagnosing thermal faults in substation equipment based on infrared images according to claim 6, characterized in that: Posterior probability The calculation formula is: ， in, Category C i The prior probability, It is the marginal likelihood, representing the total probability of observing data X, which is the weighted average of the posterior probabilities of all classes.

8. The method for diagnosing thermal faults in substation equipment based on infrared images according to claim 1, characterized in that: The multi-level alarm mechanism in step S6 is as follows: Let the predicted output P be: The model confidence score Conf is obtained as follows: ， Alarm Level is defined as follows: 。 9. A substation equipment thermal fault diagnosis system based on infrared images, characterized in that, include: The data acquisition module is used to acquire infrared images of substation equipment using an infrared thermal imager; The image processing and analysis module is used for image preprocessing and thermal fault feature extraction; The substation equipment identification module is used to identify equipment types using the YOLOv13 model. The thermal fault diagnosis and prediction module is used for fault classification based on a deep learning model. The fault alarm and feedback module is used to generate multi-level alarm information based on the diagnostic results.

10. The substation equipment thermal fault diagnosis system based on infrared images according to claim 9, characterized in that, The image processing and analysis module includes a temperature gradient adaptive enhancement mechanism and a regional dynamic contrast enhancement algorithm; the fault alarm and feedback module supports a multi-level alarm mechanism based on confidence level.

Images