A method and system for infrared immune deep learning fault detection in power equipment

Abstract

This invention discloses a method and system for infrared immune deep learning fault detection of power equipment. The system includes edge computing nodes and a cloud server platform. The edge computing nodes are used to standardize, denoise, enhance contrast, and perform pseudo-color mapping on the raw infrared image data to generate standardized infrared images. The cloud server platform uses an immune optimization algorithm to optimize the structural parameters or neuron weights of the deep learning model. It uses a retrained YOLO series, Faster R-CNN, or SSD deep learning model to extract, analyze, and preliminarily determine fault features. It uses a fault determination threshold library combined with a preset fault determination logic for verification and comprehensive judgment, generating a final fault diagnosis report for display and storage. The method is implemented based on the system. The advantages of this invention are: improved accuracy, generalization ability, and robustness of infrared fault detection for power equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault detection technology, and in particular to a method and system for infrared immune deep learning fault detection of power equipment. Background Technology

[0002] Currently, power infrared detection technology exhibits a three-stage coexistence pattern. Basic inspections still rely primarily on manual operation of thermal imagers, accounting for 60% of the market share, with inspection costs exceeding 3,800 yuan per kilometer, but the missed detection rate remains high. 30% of projects utilize drone-based dual-light inspections, but are limited by data transmission delays of 4-6 hours, making real-time early warning difficult. Only 10% of cutting-edge projects deploy AI cloud-edge systems, such as the "Smart Eye" platform used in the State Grid Jiquan Line, which improves the recognition accuracy to 89%. However, environmental interference, such as solar reflection, causes thermal imaging signal-to-noise ratio fluctuations exceeding 8dB, and the 15% difference between the DL / T664-2024 and IEC 62446-3 standards in temperature difference determination further exacerbates misjudgments.

[0003] With the development of science and technology, new methods have emerged in power infrared detection technology. For example, Chinese patent CN120747491A discloses a cable fault infrared image target detection method and system based on an improved YOLOv10. This method solves single-mode defects through RGB-T image fusion, improves positioning accuracy through hierarchical annotation and an improved model, and achieves dynamic risk assessment by combining dynamic bounding boxes and temperature field simulation, significantly improving the accuracy of cable fault detection and risk assessment. However, this method is limited to improving the accuracy of fault detection and risk assessment using the improved YOLOv10 model and does not involve fault detection based on AI cloud-edge systems. Therefore, how to effectively utilize AI cloud-edge systems to design an infrared immune deep learning fault detection system for power equipment, and how to use biological immune optimization algorithms to optimize the structural parameters of the deep learning model to improve accuracy, generalization ability, and robustness, are urgent technical problems that need to be solved. Summary of the Invention

[0004] In view of the problems existing in the background technology, the purpose of this invention is to provide an infrared immune deep learning fault detection method and system for power equipment, so as to solve the problems of accuracy, generalization ability and robustness of existing power equipment fault detection.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0006] On the one hand, the present invention provides an infrared immune deep learning fault detection system for power equipment, including edge computing nodes and a cloud server platform;

[0007] The edge computing nodes are one or more and are deployed in substations or power distribution rooms to collect and preprocess raw infrared image data in real time on site, and send the preprocessing results to the cloud server platform through the 5G communication network.

[0008] The cloud server platform is connected to one or more edge computing nodes via a 5G communication network. It is used to receive preprocessing results uploaded by each edge computing node, optimize parameters using an immune optimization algorithm, extract, analyze and preliminarily determine fault features using a retrained deep learning model, verify and comprehensively determine faults using a fault determination threshold library combined with preset fault determination logic, and generate, display and store the final fault diagnosis report.

[0009] Preferably, the edge computing node includes an infrared thermal imager and an industrial tablet computer;

[0010] The infrared thermal imager is connected to the industrial tablet PC via an RJ45 interface and is used to collect real-time infrared thermal imaging data from the substation or power distribution room and send it to the industrial tablet PC.

[0011] The industrial tablet PC is connected to the cloud server platform via a 5G communication network to receive real-time infrared thermal imaging data sent by the infrared thermal imager and to perform preprocessing and image enhancement.

[0012] Preferably, the cloud server platform integrates a remote service program, including an immune optimization module, a deep learning fault detection module, a fault determination and alarm module, and a human-computer interaction and data management module.

[0013] The immune optimization module is used to simulate the self-non-self recognition, clone selection and memory mechanism of the biological immune system during the training phase of the deep learning model, and to optimize the structural parameters or neuron weights of the deep learning model in order to improve the generalization recognition ability and anti-interference ability of the deep learning model for fault features in infrared images of power equipment.

[0014] The deep learning fault detection module has a built-in deep learning model trained by the immune optimization module, which is used to extract and analyze features from the standardized infrared image and output preliminary detection results including the location, category and confidence level of the fault area.

[0015] The fault determination and alarm module is used to perform result fusion and logical judgment based on the preliminary detection results, combined with the preset fault determination threshold library and equipment operation knowledge graph, to generate a final fault diagnosis report, and to trigger an alarm signal when a fault is diagnosed.

[0016] The human-computer interaction and data management module is communicatively connected to the immune optimization module, the deep learning fault detection module, and the fault determination and alarm module, respectively. It is used to receive user commands, display the standardized infrared image, preliminary detection results and final fault diagnosis report, and store and manage historical diagnostic data, model parameters and alarm records.

[0017] Preferably, the industrial tablet PC is equipped with an application program, including an infrared image acquisition module and an image preprocessing and enhancement module;

[0018] The infrared image acquisition module is used to acquire the original infrared image data of the power equipment to be tested;

[0019] The image preprocessing and enhancement module is used to standardize, denoise, enhance contrast, and perform pseudo-color mapping on the original infrared image data to generate a standardized infrared image.

[0020] Preferably, the image preprocessing and enhancement module includes a temperature information extraction unit, an image registration unit, and a feature enhancement unit;

[0021] The temperature information extraction unit is used to parse a matrix of absolute temperature values ​​or relative temperature differences from the original infrared image data.

[0022] The image registration unit is used to spatially align infrared images of the same device acquired at different time points.

[0023] The feature enhancement unit uses histogram equalization or homomorphic filtering to highlight the difference between the fault area and the background in the standardized infrared image.

[0024] On the other hand, the present invention provides a method for infrared immune deep learning fault detection of power equipment, comprising the following steps:

[0025] S1. Acquire the raw infrared image data of the electrical equipment to be inspected;

[0026] S2. Standardize, denoise, enhance contrast, and perform pseudo-color mapping on the original infrared image data to generate a standardized infrared image;

[0027] S3. Use immune optimization algorithms to optimize the structural parameters or neuron weights of deep learning models;

[0028] S4. Using a deep learning model trained with an immune optimization algorithm, feature extraction and analysis are performed on the standardized infrared image, and preliminary detection results including the location, category, and confidence level of the fault area are output.

[0029] S5. Using a fault determination threshold library combined with preset fault determination logic, the preliminary detection results are verified and comprehensively judged to generate a final fault diagnosis report, and an alarm is triggered when a fault is confirmed.

[0030] S6. Displays standardized infrared images, preliminary detection results, and final fault diagnosis reports; stores historical diagnostic data, model parameters, and alarm records.

[0031] Furthermore, in step S3, the immune optimization algorithm specifically includes the following steps:

[0032] L1. Define the set of data parameters to be optimized for the deep learning model as an antibody population, and define the recognition error or loss function value of the deep learning model on the validation set as the antigen.

[0033] L2. Calculate the affinity between the antibody and the antigen, and evaluate the merits of each parameter set;

[0034] L3. Perform clonal amplification and high-frequency mutation operations on high-affinity antibodies, and inhibit or eliminate low-affinity antibodies;

[0035] L4. Retain the best antibodies generated during the iteration process as memory cells for initializing the next model training or online update;

[0036] L5. Repeat steps L1 to L4 until convergence to the termination condition, obtaining optimal parameters with stronger generalization ability and robustness.

[0037] Furthermore, in step S4, the deep learning model is a multi-task learning model that performs target detection tasks and fault severity classification tasks.

[0038] The multi-task learning model is one or more combinations of the target YOLO series, Faster R-CNN or SSD deep learning model based on immune optimization algorithm;

[0039] The target detection task outputs the bounding box and category label of the fault area. The category label includes wire overheating, joint overheating, insulator deterioration, bushing defects, voltage-induced heating equipment faults, and current-induced heating equipment faults.

[0040] The fault severity classification task outputs a level label representing the severity of the fault, including normal, general defect, severe defect, and critical defect.

[0041] Furthermore, in step S5, the fault determination threshold library is a dynamic threshold library, which can adaptively adjust the temperature threshold, temperature difference threshold and thermal imaging mode used to determine the severity of the fault based on historical and real-time data of ambient temperature and equipment load rate.

[0042] The fault determination logic includes at least one of the following:

[0043] Absolute temperature criterion: Determine whether the absolute temperature of the fault point exceeds the limit value specified by the equipment material or safety standards;

[0044] Relative temperature difference criterion: Determine whether the temperature difference between the fault point and the corresponding normal point exceeds a preset threshold;

[0045] Criteria for comparison with similar devices: Under the same testing conditions, determine whether the temperature difference between corresponding parts of the three phases of the same device exceeds the preset imbalance threshold.

[0046] Thermochromatogram pattern recognition criteria: The matching degree between the thermal distribution pattern of the current infrared image and the typical fault thermochromatograms stored in the knowledge base is calculated.

[0047] Compared with existing technologies, the infrared immune deep learning fault detection method and system for power equipment provided by this invention have the following advantages:

[0048] (1) An immune optimization algorithm was used in the infrared fault detection method for power equipment to optimize the structural parameters or neuron weights of the deep learning model, which has stronger generalization ability and robustness.

[0049] (2) The retrained deep model is used for fault feature extraction, analysis and preliminary judgment, and the fault judgment threshold library is used in combination with the preset fault judgment logic for verification and comprehensive judgment, which improves the accuracy of electrical equipment fault detection.

[0050] (3) An infrared immune deep learning fault detection system for power equipment was designed using edge computing nodes and cloud server platforms. An immune optimization module and a deep learning fault detection module were added to the detection technology, realizing the real-time infrared data acquisition and the visualization and accuracy of electrical equipment fault judgment. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a structural diagram of the infrared immune deep learning fault detection system for power equipment according to an embodiment of the present invention;

[0053] Figure 2 This is a flowchart illustrating the process of the infrared immune deep learning fault detection method for power equipment according to an embodiment of the present invention.

[0054] Figure 3 This is a flowchart illustrating the immune optimization algorithm described in an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1: See Figure 1 Embodiment 1 of the present invention provides an infrared immune deep learning fault detection system for power equipment, comprising edge computing nodes and a cloud server platform, wherein:

[0057] There are one or more edge computing nodes deployed at substations or power distribution rooms to collect and preprocess raw infrared image data in real time on site, and send the preprocessing results to the cloud server platform through the 5G communication network.

[0058] The cloud server platform is connected to one or more edge computing nodes via a 5G communication network to receive preprocessing results uploaded by each edge computing node, optimize parameters using an immune optimization algorithm, extract, analyze and preliminarily determine fault features using a retrained deep learning model, verify and comprehensively determine faults using a fault determination threshold library combined with preset fault determination logic, generate a final fault diagnosis report and display and store it.

[0059] Specifically, edge computing nodes include infrared thermal imagers and industrial tablets, where:

[0060] The infrared thermal imager is connected to the industrial tablet PC via an RJ45 interface to collect real-time infrared thermal imaging data from the substation or power distribution room and send it to the industrial tablet PC.

[0061] The industrial tablet PC connects to the cloud server platform via a 5G communication network to receive real-time infrared thermal imaging data sent by the infrared thermal imager and perform preprocessing and image enhancement.

[0062] Specifically, the cloud server platform integrates remote service programs, including an immune optimization module, a deep learning fault detection module, a fault determination and alarm module, and a human-computer interaction and data management module, among which:

[0063] The immune optimization module is used to simulate the self-non-self recognition, clone selection and memory mechanisms of the biological immune system during the training phase of the deep learning model. It optimizes the structural parameters or neuron weights of the deep learning model to improve the generalization recognition ability and anti-interference ability of the deep learning model for fault features in infrared images of power equipment.

[0064] The deep learning fault detection module has a built-in deep learning model trained by the immune optimization module, which is used to extract and analyze features from standardized infrared images and output preliminary detection results including the location, category and confidence level of the fault area.

[0065] The fault diagnosis and alarm module is used to perform result fusion and logical judgment based on the preliminary detection results, combined with the preset fault diagnosis threshold library and equipment operation knowledge graph, to generate the final fault diagnosis report, and trigger an alarm signal when a fault is diagnosed.

[0066] The human-computer interaction and data management module is connected to the immune optimization module, the deep learning fault detection module, and the fault determination and alarm module, respectively. It is used to receive user commands, display the standardized infrared images, preliminary detection results and final fault diagnosis reports, and store and manage historical diagnostic data, model parameters and alarm records.

[0067] Specifically, the industrial tablet PC is equipped with applications, including an infrared image acquisition module and an image preprocessing and enhancement module, wherein:

[0068] The infrared image acquisition module is used to acquire the raw infrared image data of the power equipment to be inspected;

[0069] The image preprocessing and enhancement module is used to standardize, denoise, enhance contrast, and perform pseudo-color mapping on the raw infrared image data to generate a standardized infrared image.

[0070] Specifically, the image preprocessing and enhancement module includes a temperature information extraction unit, an image registration unit, and a feature enhancement unit, wherein:

[0071] A temperature information extraction unit is used to parse a matrix of absolute temperature values ​​or relative temperature differences from the original infrared image data.

[0072] The image registration unit is used to spatially align infrared images of the same device acquired at different time points;

[0073] The feature enhancement unit uses histogram equalization or homomorphic filtering to highlight the difference between the fault area and the background in the standardized infrared image.

[0074] For example, the temperature matrix of the original infrared image is processed using an 8-bit grayscale normalization formula through histogram equalization, as follows:

[0075]

[0076] in, The cumulative distribution function is... Let be the original temperature in the i-th row and j-th column of the original infrared temperature matrix. The pixels in the vertical direction of the original infrared image. The pixels in the horizontal direction of the original infrared image; It is the minimum value of the cumulative distribution function; The original temperature in the i-th row and j-th column of the original infrared temperature matrix after 8-bit grayscale normalization;

[0077] More specifically, the edge computing nodes consist of a Fluke RSE600 in-line infrared thermal imager and an EDA EM-I22J industrial panel PC. The cloud server platform uses Huawei Cloud and is equipped with an NVIDIA T4 GPU to accelerate deep learning inference.

[0078] The infrared immune deep learning fault detection system for power equipment provided in this invention utilizes edge computing nodes and a cloud server platform. The system standardizes, denoises, enhances contrast, and performs pseudo-color mapping on raw infrared image data using edge computing nodes to generate standardized infrared images. The cloud server platform uses an immune optimization algorithm to optimize the structural parameters or neuron weights of the deep learning model. The retrained deep learning model is then used for fault feature extraction, analysis, and preliminary judgment. A fault judgment threshold library combined with preset fault judgment logic is used for verification and comprehensive judgment, generating a final fault diagnosis report for display and storage. This improves the accuracy, generalization ability, and robustness of infrared fault detection for power equipment. Through design, simulation, and verification, a modular product is formed, enabling rapid portability between different platforms and accelerating the product development process.

[0079] Example 2: See Figure 2 Embodiment 2 of the present invention provides a method for detecting faults in power equipment using infrared immune deep learning, which employs the power equipment infrared immune deep learning fault detection system as described in Embodiment 1, and includes the following steps:

[0080] S1. Acquire the raw infrared image data of the electrical equipment to be inspected;

[0081] S2. Standardize, denoise, enhance contrast, and perform pseudo-color mapping on the original infrared image data to generate a standardized infrared image;

[0082] S3. Use immune optimization algorithms to optimize the structural parameters or neuron weights of deep learning models;

[0083] S4. Using a deep learning model trained with an immune optimization algorithm, feature extraction and analysis are performed on standardized infrared images, and preliminary detection results including the location, category, and confidence level of the fault area are output.

[0084] S5. Using the fault judgment threshold library and the preset fault judgment logic, the preliminary detection results are verified and comprehensively judged to generate the final fault diagnosis report and trigger an alarm when a fault is diagnosed.

[0085] S6. Displays standardized infrared images, preliminary detection results, and final fault diagnosis reports; stores historical diagnostic data, model parameters, and alarm records.

[0086] Specifically, see Figure 3 In step S3, the immune optimization algorithm specifically includes the following steps:

[0087] L1. Define the set of data parameters to be optimized for the deep learning model as an antibody population, and define the recognition error or loss function value of the model on the validation set as the antigen;

[0088] For example, antibody-antigen definition: antibody population Each antibody is a weight vector of the YOLOv7 model. ;

[0089] L2. Calculate the affinity between the antibody and the antigen, and evaluate the merits of each parameter set;

[0090] For example, affinity calculation uses a normalized loss function:

[0091] here, For the normalized loss function, Indicates the original loss. , ;

[0092] L3. Perform clonal amplification and high-frequency mutation operations on high-affinity antibodies, and inhibit or eliminate low-affinity antibodies;

[0093] For example, clonal amplification: High-frequency variants: ,in, ;

[0094] here, The number of cells for clonal expansion. This is the original weight value of the antibody. This represents the weighting value of the antibody after high-frequency mutation operations. To conform to a mean of 0 and a standard deviation of Normally distributed random numbers;

[0095] L4. Retain the best antibodies generated during the iteration process as memory cells for initializing the next model training or online update;

[0096] when This is used to initialize the next model training iteration;

[0097] L5. Repeat the above process until convergence to the termination condition, obtaining the optimal parameters with stronger generalization ability and robustness.

[0098] Specifically, in step S4, the deep learning model is a multi-task learning model that performs object detection and fault severity classification tasks, wherein:

[0099] The multi-task learning model is one or more combinations of the target YOLO series, Faster R-CNN or SSD deep learning model based on immune optimization algorithm;

[0100] The target detection task outputs the bounding box and category label of the fault area. The category label includes wire overheating, joint overheating, insulator deterioration, bushing defects, voltage-induced heating equipment faults, and current-induced heating equipment faults.

[0101] The fault severity classification task outputs a level label representing the severity of the fault, including normal, general defect, severe defect, and critical defect.

[0102] For example, using the immune-optimized YOLOv10, two branches are output:

[0103] Target detection branch: Six types of fault labels are included within the bounding box coordinates and at a confidence threshold of 0.7: conductor overheating, joint overheating, insulator deterioration, bushing defects, voltage-induced heating equipment faults, and current-induced heating equipment faults.

[0104] Severity rating branch: Uses Softmax probability to generate output with 4 levels of classification: normal, general defect, severe defect, and critical defect;

[0105] Specifically, in step S5, the fault judgment threshold library is a dynamic threshold library, which can adaptively adjust the temperature threshold, temperature difference threshold and thermal image mode used to judge the severity of the fault based on historical and real-time data of ambient temperature and equipment load rate.

[0106] For example, the temperature threshold adaptively adjusts with the load rate L: ,here, Temperature threshold This is the temperature threshold under rated load conditions, typically taken as 75℃. This represents the actual load rate. Rated load rate, The adjustment factor is usually set to 0.2.

[0107] Specifically, in step S5, the fault determination logic includes at least one of the following:

[0108] Absolute temperature criterion: Determine whether the absolute temperature of the fault point exceeds the limit value specified by the equipment material or safety standards;

[0109] Relative temperature difference criterion: Determine whether the temperature difference between the fault point and the corresponding normal point exceeds a preset threshold;

[0110] Criteria for comparison with similar devices: Under the same testing conditions, determine whether the temperature difference between corresponding parts of the three phases of the same device exceeds the preset imbalance threshold.

[0111] Thermochromatogram pattern recognition criteria: The matching degree between the thermal distribution pattern of the current infrared image and the typical fault thermochromatograms stored in the knowledge base is calculated.

[0112] For example, a copper conductor with an absolute maximum temperature >110℃ is considered a critical defect, a relative temperature difference >15℃ is considered a serious defect, and a temperature difference between corresponding parts of the three phases of the same equipment >10℃ is considered a general defect.

[0113] Experimental data validated that the generalization ability remained at 89.3%-91.2% mAP@0.5 across three unknown substation datasets. The robustness showed that the detection accuracy decreased by only 2.1% after adding Gaussian noise (SNR=15dB). The real-time performance metrics were ≤80ms for edge preprocessing and ≤120ms for cloud platform inference.

[0114] like Figure 2-3 As shown in the embodiments of the present invention, the infrared immune deep learning fault detection method for power equipment obtains the original infrared image data of the power equipment to be detected, performs standardization, denoising, contrast enhancement, and pseudo-color mapping processing to generate a standardized infrared image, uses an immune optimization algorithm to optimize the structural parameters or neuron weights of the deep learning model, uses the deep learning model trained by the immune optimization algorithm to extract and analyze features from the standardized infrared image, outputs preliminary detection results including the location, category, and confidence level of the fault area, uses a fault judgment threshold library combined with a preset fault judgment logic to verify and comprehensively judge the preliminary detection results, generates a final fault diagnosis report, and triggers an alarm when a fault is diagnosed. This achieves real-time infrared data acquisition and visualization of electrical equipment fault judgment, and improves the accuracy, generalization ability, and robustness of infrared fault detection for power equipment.

[0115] In summary, the infrared immune deep learning fault detection method and system for power equipment provided in this embodiment of the invention adopts a modular design approach and implements the infrared immune deep learning fault detection system for power equipment using edge computing nodes and a cloud server platform. The system utilizes edge computing nodes to standardize, denoise, enhance contrast, and perform pseudo-color mapping processing on the original infrared image data to generate standardized infrared images. The cloud server platform uses an immune optimization algorithm to optimize the structural parameters or neuron weights of the deep learning model. The retrained deep learning model is used to extract, analyze, and preliminarily determine fault features. The fault determination threshold library is combined with preset fault determination logic for verification and comprehensive judgment, generating a final fault diagnosis report for display and storage. This improves the accuracy, generalization ability, and robustness of infrared fault detection for power equipment.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for infrared immune deep learning fault detection in power equipment, characterized in that, Includes the following steps: S1. Acquire the raw infrared image data of the electrical equipment to be inspected; S2. Standardize, denoise, enhance contrast, and perform pseudo-color mapping on the original infrared image data to generate a standardized infrared image; S3. Use immune optimization algorithms to optimize the structural parameters or neuron weights of deep learning models; S4. Using a deep learning model trained with an immune optimization algorithm, feature extraction and analysis are performed on the standardized infrared image, and preliminary detection results including the location, category, and confidence level of the fault area are output. S5. Using a fault determination threshold library combined with preset fault determination logic, the preliminary detection results are verified and comprehensively judged to generate a final fault diagnosis report, and an alarm is triggered when a fault is confirmed. S6. Display the standardized infrared image, preliminary detection results, and final fault diagnosis report, and store historical diagnostic data, model parameters, and alarm records.

2. The infrared immune deep learning fault detection method for power equipment as described in claim 1, characterized in that, In step S3, the immune optimization algorithm specifically includes the following steps: L1. Define the set of data parameters to be optimized for the deep learning model as an antibody population, and define the recognition error or loss function value of the deep learning model on the validation set as the antigen. L2. Calculate the affinity between the antibody and the antigen, and evaluate the merits of each parameter set; L3. Perform clonal amplification and high-frequency mutation operations on high-affinity antibodies, and inhibit or eliminate low-affinity antibodies; L4. Retain the best antibodies generated during the iteration process as memory cells for initializing the next model training or online update; L5. Repeat steps L1 to L4 until convergence to the termination condition, obtaining optimal parameters with stronger generalization ability and robustness.

3. The infrared immune deep learning fault detection method for power equipment as described in claim 1, characterized in that, In step S4, the deep learning model is a multi-task learning model that performs object detection and fault severity classification tasks. The multi-task learning model is one or more combinations of the target YOLO series, Faster R-CNN or SSD deep learning model based on immune optimization algorithm; The target detection task outputs the bounding box and category label of the fault area. The category label includes wire overheating, joint overheating, insulator deterioration, bushing defects, voltage-induced heating equipment faults, and current-induced heating equipment faults. The fault severity classification task outputs a level label representing the severity of the fault, including normal, general defect, severe defect, and critical defect.

4. The infrared immune deep learning fault detection method for power equipment as described in claim 1, characterized in that, In step S5, the fault determination threshold library is a dynamic threshold library, which can adaptively adjust the temperature threshold, temperature difference threshold and thermal image mode used to determine the severity of the fault based on historical and real-time data of ambient temperature and equipment load rate. The fault determination logic includes at least one of the following: Absolute temperature criterion: Determine whether the absolute temperature of the fault point exceeds the limit value specified by the equipment material or safety standards; Relative temperature difference criterion: Determine whether the temperature difference between the fault point and the corresponding normal point exceeds a preset threshold; Criteria for comparison with similar devices: Under the same testing conditions, determine whether the temperature difference between corresponding parts of the three phases of the same device exceeds the preset imbalance threshold. Thermochromatogram pattern recognition criteria: The matching degree between the thermal distribution pattern of the current infrared image and the typical fault thermochromatograms stored in the knowledge base is calculated.

5. A power equipment infrared immune deep learning fault detection system, used to execute the power equipment infrared immune deep learning fault detection method according to any one of claims 1-4, characterized in that, The system includes edge computing nodes and a cloud server platform; The edge computing nodes are one or more and are deployed in substations or power distribution rooms to collect and preprocess raw infrared image data in real time on site, and send the preprocessing results to the cloud server platform through the 5G communication network. The cloud server platform is connected to one or more edge computing nodes via a 5G communication network. It is used to receive preprocessing results uploaded by each edge computing node, optimize parameters using an immune optimization algorithm, extract, analyze and preliminarily determine fault features using a retrained deep learning model, verify and comprehensively determine faults using a fault determination threshold library combined with preset fault determination logic, generate a final fault diagnosis report and display and store it.

6. The infrared immune deep learning fault detection system for power equipment as described in claim 5, characterized in that, The edge computing node includes an infrared thermal imager and an industrial tablet computer. The infrared thermal imager is connected to the industrial tablet computer via an RJ45 interface and is used to collect real-time infrared thermal imaging data from the substation or power distribution room and send it to the industrial tablet computer. The industrial tablet computer is connected to the cloud server platform via a 5G communication network and is used to receive the real-time infrared thermal imaging data sent by the infrared thermal imager and perform preprocessing and image enhancement.

7. The infrared immune deep learning fault detection system for power equipment as described in claim 6, characterized in that, The cloud server platform integrates remote service programs, including an immune optimization module, a deep learning fault detection module, a fault determination and alarm module, and a human-computer interaction and data management module. The immune optimization module is used to simulate the self-non-self recognition, clone selection and memory mechanism of the biological immune system during the training phase of the deep learning model, and to optimize the structural parameters or neuron weights of the deep learning model in order to improve the generalization recognition ability and anti-interference ability of the deep learning model for fault features in infrared images of power equipment. The deep learning fault detection module has a built-in deep learning model trained by the immune optimization module, which is used to extract and analyze features from the standardized infrared image and output preliminary detection results including the location, category and confidence level of the fault area. The fault determination and alarm module is used to perform result fusion and logical judgment based on the preliminary detection results, combined with the preset fault determination threshold library and equipment operation knowledge graph, to generate a final fault diagnosis report, and to trigger an alarm signal when a fault is diagnosed. The human-computer interaction and data management module is communicatively connected to the immune optimization module, the deep learning fault detection module, and the fault determination and alarm module, respectively. It is used to receive user commands, display the standardized infrared image, preliminary detection results and final fault diagnosis report, and store and manage historical diagnostic data, model parameters and alarm records.

8. The infrared immune deep learning fault detection system for power equipment as described in claim 6, characterized in that, The industrial tablet PC is equipped with an application program, including an infrared image acquisition module and an image preprocessing and enhancement module; The infrared image acquisition module is used to acquire the original infrared image data of the power equipment to be tested; The image preprocessing and enhancement module is used to standardize, denoise, enhance contrast, and perform pseudo-color mapping on the original infrared image data to generate a standardized infrared image.

9. The infrared immune deep learning fault detection system for power equipment as described in claim 8, characterized in that, The image preprocessing and enhancement module includes a temperature information extraction unit, an image registration unit, and a feature enhancement unit; The temperature information extraction unit is used to parse a matrix of absolute temperature values ​​or relative temperature differences from the original infrared image data. The image registration unit is used to spatially align infrared images of the same device acquired at different time points. The feature enhancement unit uses histogram equalization or homomorphic filtering to highlight the difference between the fault area and the background in the standardized infrared image.

Images