Current heating equipment detection and analysis method based on artificial intelligence
By employing an AI-based method for detecting and analyzing current-induced heating equipment, and utilizing drones to collect images, perform image processing, and detect targets, the problem of distance error and high false negative rate in the detection of thermal defects in power equipment has been solved. This enables accurate identification and timely detection of equipment defects, ensuring power grid safety.
Patent Information
- Application Number
- CN202511261577.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-05
- Publication Date
- 2025-12-19
AI Technical Summary
Existing thermal defect detection methods for power equipment suffer from uncontrollable distance attenuation errors and high rates of missed detection for minor defects. Traditional infrared thermal imaging technology cannot effectively compensate for distance differences and cannot identify the correlation between equipment status, resulting in low measured temperature values and high missed detection rates.
An AI-based current-induced heating equipment detection and analysis method is adopted. Visible light and thermal images are collected by UAVs, and SSD target detection network is used to identify equipment type and partition. Temperature compensation is calculated, and calibrated thermal images are generated through image registration and thermal image segmentation. Combined with multi-scale feature extraction and anomaly detection model, accurate identification of equipment defects is achieved.
It improves the accuracy and efficiency of temperature identification in thermal imaging of power equipment, enabling timely detection of minor thermal defects in equipment and ensuring the safe and stable operation of the power grid.
Smart Images
Figure CN121170637A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to an artificial intelligence-based method for detecting and analyzing current-induced heating equipment. Background Technology
[0002] Currently, thermal defect detection in power equipment mainly relies on infrared thermal imaging technology, but it faces the following technical bottlenecks:
[0003] 1. Uncontrollable distance attenuation error: The temperature value measured by the thermal imager decreases significantly with the increase of the detection distance (empirical attenuation coefficient 0.5-1℃ / meter). Especially when the same image contains multiple devices (such as the dense equipment area of a substation), it is impossible to independently compensate for the equipment zones at different distances, resulting in a systematically low temperature value measured by distant devices (maximum error of ±5℃).
[0004] 2. High rate of missed detection for minor defects: Traditional threshold methods (such as fixed temperature alarms) ignore the correlation between equipment status and thermal pattern characteristics, resulting in a missed detection rate of over 40% for local hot spots with too small a video screen area. Summary of the Invention
[0005] To address the aforementioned problems in the prior art, this invention provides an artificial intelligence-based method for detecting and analyzing current-induced heating devices.
[0006] An artificial intelligence-based method for detecting and analyzing current-induced heating devices includes the following steps:
[0007] Step 1: The drone collects visible light and thermal images of the power equipment and performs preprocessing.
[0008] Step 2: Use the constructed SSD target detection network to identify targets in the visible light image and determine the device type and partition.
[0009] Step 3: Calculate the monocular distance based on the device pixel width for the visible light image, and calculate the temperature compensation according to the distance by device partition;
[0010] Step 4: Segment the thermal image according to the device partition and image registration, and generate a calibrated thermal image based on temperature compensation;
[0011] Step 5: Use the thermal imaging image to compare with the thermal imaging image model to detect temperature anomalies and read the temperature data;
[0012] Step 6: Determine the type of equipment defect based on the anomaly detection results.
[0013] Furthermore, the preprocessing specifically includes
[0014] Radiation correction: Perform black level correction, radiation calibration, and atmospheric transmittance compensation on thermal images;
[0015] Geometric correction: Eliminates lens distortion and drone attitude shift, ensuring pixel-level alignment between visible light and thermal images;
[0016] Image enhancement and noise suppression: The Retinex algorithm is used to optimize shadow areas in the visible light image; nonlocal mean filtering is used to remove noise in the thermal image.
[0017] Image registration: The visible light image and the thermal image are time-aligned according to the timestamp. The aligned visible light and infrared images are then input into the generative adversarial network to generate a texture-temperature correlation feature map.
[0018] Furthermore, the construction of the SSD target detection network is specifically as follows:
[0019] Creating a sample dataset: Convert the labeled data of power equipment into VOC or COCO format, adapt it to the SSD training framework, and refine the labeling to the component level;
[0020] SSD Network Construction and Initialization: A pre-trained VGG16 network is used as the feature extraction backbone network. The weights of the convolutional layers are retained for transfer learning. Additional convolutional layers are added to generate multi-scale feature maps, each corresponding to a different receptive field, to adapt to differences in device size.
[0021] Multi-scale feature extraction and target detection: After the backbone network extracts basic features, it outputs multi-level feature maps through additional convolutional layers to predict targets at different scales; then, the SE module is embedded in the SSD to improve the feature weights of key components;
[0022] Loss calculation and model training: The loss function adopts the multi-task loss function; and hard example mining is used to backpropagate only to the negative samples with the highest loss.
[0023] Training strategy optimization: RMSProp was used as the optimizer with an initial learning rate of 0.001; the shallow weights of the backbone network were frozen, and the deep layers and newly added convolutional layers were fine-tuned to accelerate convergence and prevent overfitting.
[0024] Post-processing and equipment partitioning: Non-maximum suppression is used to filter predicted boxes by category, and the highest confidence results are retained to output the equipment type and bounding box; and semantic segmentation is used to assist partitioning, dividing key temperature zones according to equipment structure.
[0025] Furthermore, the distance calculation formula in step 3 is as follows: Local distances are calculated for different zones, and a distance-temperature attenuation mapping table is established to compensate for temperature.
[0026] Furthermore, step 4 specifically involves:
[0027] The thermal image is segmented based on device partitioning and image registration, and a calibrated segmented thermal image is generated based on temperature compensation.
[0028] Step 401: Accurately map the device partition coordinates identified by the SSD network in Step 2 onto the thermal image;
[0029] Step 402: For each pixel within the partition mask, apply compensation coefficients to generate a temperature calibration matrix;
[0030] Step 403: The compensated temperature calibration matrix is superimposed with the grayscale value of the original thermal image to generate a calibrated thermal image corresponding to the physical temperature value and the pseudo-color.
[0031] Furthermore, step 5, which involves detecting temperature anomalies by comparing thermal imaging images with the model, specifically includes the following steps:
[0032] Step 501: Thermodynamic features are extracted using a multi-scale feature pyramid.
[0033] Step 502, Static benchmark comparison: including dynamic benchmark library construction and residual map generation;
[0034] The dynamic benchmark library is constructed by using historical normal thermal images of the same type of equipment under the same operating conditions as a benchmark template, and extracting the average temperature T of the corresponding zone in the template. base Maximum allowable temperature rise ΔT allow and heat distribution patterns;
[0035] The residual heatmap (ResidualMap) is generated using the following formula: ResidualMap = |T| current -T base |·Mask ROI Among them, Mask ROI For equipment partitioning mask; T current The actual temperature measurement value of the device or area at the current moment or under the current state;
[0036] Step 503, Locate the abnormal region: Use dual threshold segmentation to locate the abnormal region. First threshold: mark the region where ΔT > ΔT in the residual heatmap. allow For the region, a secondary threshold is applied: morphological opening operations are used to eliminate noise while retaining connected regions with a minimum area of 5×5 pixels.
[0037] Furthermore, step 6 specifically involves matching the abnormal detection results with the constructed temperature defect analysis library to determine possible abnormal situations that may occur in the equipment.
[0038] The beneficial effects of this invention are as follows: This invention improves the accuracy and efficiency of temperature identification in thermal imaging images of equipment during power grid inspections. It can promptly detect minor thermal defects and anomalies in equipment, making the analysis and judgment of thermal defects more scientific and accurate, and enabling the formulation of more scientific planned defect handling. This prevents ordinary thermal defects from developing into critical defects and ensures the safe and stable operation of the power grid. Attached Figure Description
[0039] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0041] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] Example 1: Refer to Figure 1 This invention provides an artificial intelligence-based method for detecting and analyzing current-induced heating devices, comprising the following steps:
[0043] Step 1: The drone collects visible light and thermal images of the power equipment and performs preprocessing.
[0044] Step 2: Use the constructed SSD target detection network to identify targets in the visible light image and determine the device type and partition.
[0045] Step 3: Calculate the monocular distance based on the device pixel width for the visible light image, and calculate the temperature compensation according to the distance by device partition;
[0046] Step 4: Segment the thermal image according to the device partition and image registration, and generate a calibrated thermal image based on temperature compensation;
[0047] Step 5: Use the thermal imaging image to compare with the thermal imaging image model to detect temperature anomalies and read the temperature data;
[0048] Step 6: Determine the type of equipment defect based on the anomaly detection results.
[0049] Preprocessing specifically includes
[0050] Radiation correction: Perform black level correction (eliminate sensor dark current), radiometric calibration (convert DN value to radiance) on thermal images, and atmospheric transmittance compensation (correct attenuation error based on ambient temperature and humidity).
[0051] Geometric correction: Eliminates lens distortion (radial / tangential distortion model) and UAV attitude shift (IMU data-assisted image registration) to ensure pixel-level alignment between visible light and thermal images;
[0052] Image enhancement and noise suppression: The visible light image uses the Retinex algorithm to optimize shadow areas; the thermal image uses nonlocal mean filtering (NLM) to remove noise.
[0053] Image registration: The visible light image and the thermal image are time-aligned according to the timestamp. The aligned visible light and infrared images are then input into the generative adversarial network to generate a texture-temperature correlation feature map.
[0054] The construction of the SSD object detection network is as follows:
[0055] Create a sample dataset: Convert the power equipment annotation data (such as bounding box coordinates and category labels) into VOC or COCO format, adapt it to the SSD training framework, and the annotation needs to be refined to the component level (such as transformer bushings, surge arrester bodies, and cable joints).
[0056] SSD Network Construction and Initialization: A pre-trained VGG16 network is used as the feature extraction backbone network. The weights of the convolutional layers are retained for transfer learning. Additional convolutional layers are added to generate multi-scale feature maps, each corresponding to a different receptive field, to adapt to differences in device size.
[0057] Multi-scale feature extraction and target detection: After the backbone network extracts basic features, it outputs multi-level feature maps (decreasing in size) through additional convolutional layers to predict targets at different scales (shallow feature maps (high resolution) detect small parts (such as bolts and hardware); deep feature maps (low resolution) detect large equipment (such as transformer bodies and towers)); and performs position prediction on each feature map (class probability (Softmax output: background / equipment type); bounding box offset (4 parameters: center point coordinates, width and height adjustment)); then embeds the SE (Squeeze-and-Excitation) module into SSD to enhance the feature weights of key components;
[0058] Loss calculation and model training:
[0059] The loss function adopted is the multi-task loss function, where total loss = classification loss (CrossEntropy) + localization loss (SmoothL1), and the formula is:
[0060]
[0061] Where N is the number of positive samples, α is the weight balancing parameter (taken as 1 in this embodiment), and L cnof For classification loss; L loc For classification loss;
[0062] Hard Negative Mining is used to backpropagate only to the negative samples (background false positives) with the highest loss, thus solving the problem of imbalanced positive and negative samples.
[0063] Training strategy optimization: RMSProp is used as the optimizer with an initial learning rate of 0.001 and a decay of 94% every 2 epochs; the shallow weights of the backbone network are frozen, and the deep layers and newly added convolutional layers are fine-tuned to accelerate convergence and prevent overfitting.
[0064] Post-processing and equipment partitioning: Non-maximum suppression (NMS) is used to filter predicted bounding boxes by category, retaining the highest confidence results to output equipment type and bounding boxes; and semantic segmentation is used to assist partitioning, dividing key temperature zones according to equipment structure (based on the bounding boxes output by SSD, the equipment area is cropped and input into the U-Net network for pixel-level semantic segmentation (e.g., dividing transformer oil conservator, heat sink, high-voltage bushing)).
[0065] Partitioning rules:
[0066] Structurally regular equipment (such as insulator strings) is divided into sections according to physical connections.
[0067] Irregular equipment (such as cable joints) is divided into sections according to thermal imaging temperature measurement requirements.
[0068] The distance calculation formula in step 3 is: Local distances are calculated for different zones, and a distance-temperature attenuation mapping table is established to compensate for temperature.
[0069] Step 4 specifically involves:
[0070] The thermal image is segmented based on device partitioning and image registration, and a calibrated segmented thermal image is generated based on temperature compensation.
[0071] Step 401: Accurately map the device partition coordinates identified by the SSD network in Step 2 onto the thermal image;
[0072] Step 402: For each pixel within the partition mask, apply compensation coefficients to generate a temperature calibration matrix;
[0073] Step 403: The compensated temperature calibration matrix is superimposed with the grayscale value of the original thermal image to generate a calibrated thermal image corresponding to the physical temperature value and the pseudo-color.
[0074] Step 5, which involves detecting temperature anomalies by comparing thermal images with the model, specifically includes the following steps:
[0075] Step 501: Thermodynamic features are extracted using a multi-scale feature pyramid.
[0076] Multi-scale feature pyramid:
[0077] Using an improved U-NetwithMS-CAM (multi-scale channel attention module):
[0078] Encoder: Extracts local texture (3×3 convolution kernel) and global temperature distribution (dilated convolution) through 4 levels of downsampling;
[0079] MS-CAM: An attention mechanism is embedded at the jump connection to weightedly fuse shallow details (such as bolt hotspots) with deep semantics (such as overload band temperature rise);
[0080] Output: Generate a feature map with a resolution of 256×256, retaining small targets at the 5cm level.
[0081] Step 502, Static benchmark comparison: including dynamic benchmark library construction and residual map generation;
[0082] The dynamic benchmark library is constructed by using historical normal thermal images of the same type of equipment under the same operating conditions as a benchmark template, and extracting the average temperature T of the corresponding zone in the template. base Maximum allowable temperature rise ΔT allow and heat distribution patterns (e.g., uniform / gradual);
[0083] The residual heatmap (ResidualMap) is generated using the following formula: ResidualMap = |T| current -T base |·Mask ROI Among them, Mask ROI For equipment partitioning mask; T current The actual temperature measurement value of the device or area at the current moment or under the current state;
[0084] Step 503, Locate the abnormal region: Use dual threshold segmentation to locate the abnormal region. First threshold: mark the region where ΔT > ΔT in the residual heatmap. allow The area
[0085] Secondary thresholding: Morphological opening operations are used to eliminate noise, retaining connected regions with a minimum area of 5×5 pixels. The code is as follows:
[0086] contours=find_contours(high_heat)
[0087] valid_anomalies=filter(lambda c:ellipticity(c)>0.7and area(c)>5px,contours)
[0088] return mask_from_contours(valid_anomalies)
[0089] Step 6 specifically involves matching the anomaly detection results with the constructed temperature defect analysis library to determine possible abnormalities in the equipment.
[0090] Defect Analysis Library Construction
[0091] Library structure: Contains three types of defect rule bases:
[0092] National Value Database: Based on the DL / T664-2016 standard, the absolute temperature thresholds for critical / serious / general defects are defined (e.g., sleeve joints >110℃ are critical).
[0093] Relative temperature difference chamber: A temperature difference of >80% between similar equipment is considered a critical defect (formula: T% = x100%).
[0094] Pattern library: Thermal characteristics of typical faults (such as local overheating of cables appearing as "star-shaped" and flashover of insulators appearing as "band-shaped temperature rise").
[0095] In the description of embodiments of the present invention, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise stated, "a plurality of" means two or more.
[0096] In the description of embodiments of the present invention, specific features, structures, materials or characteristics may be combined in any suitable manner in one or more embodiments or examples.
[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for detecting and analyzing current-induced heating devices based on artificial intelligence, characterized in that, Includes the following steps: Step 1: The drone collects visible light and thermal images of the power equipment and performs preprocessing. Step 2: Use the constructed SSD target detection network to identify targets in the visible light image and determine the device type and partition. Step 3: Calculate the monocular distance based on the device pixel width for the visible light image, and calculate the temperature compensation according to the distance by device partition; Step 4: Segment the thermal image according to the device partition and image registration, and generate a calibrated thermal image based on temperature compensation; Step 5: Use the thermal imaging image to compare with the thermal imaging image model to detect temperature anomalies and read the temperature data; Step 6: Determine the type of equipment defect based on the anomaly detection results.
2. The method for detecting and analyzing current-induced heating devices based on artificial intelligence according to claim 1, characterized in that, The preprocessing specifically includes Radiation correction: Perform black level correction, radiation calibration, and atmospheric transmittance compensation on thermal images; Geometric correction: Eliminates lens distortion and drone attitude shift, ensuring pixel-level alignment between visible light and thermal images; Image enhancement and noise suppression: The Retinex algorithm is used to optimize shadow areas in the visible light image; The thermal image uses nonlocal mean filtering to remove noise; Image registration: The visible light image and the thermal image are time-aligned according to the timestamp. The aligned visible light and infrared images are then input into the generative adversarial network to generate a texture-temperature correlation feature map.
3. The method for detecting and analyzing current-induced heating devices based on artificial intelligence according to claim 1, characterized in that, The construction of the SSD target detection network is as follows: Creating a sample dataset: Convert the labeled data of power equipment into VOC or COCO format, adapt it to the SSD training framework, and refine the labeling to the component level; SSD Network Construction and Initialization: A pre-trained VGG16 network is used as the feature extraction backbone network. The weights of the convolutional layers are retained for transfer learning. Additional convolutional layers are added to generate multi-scale feature maps, each corresponding to a different receptive field, to adapt to differences in device size. Multi-scale feature extraction and target detection: After the backbone network extracts basic features, it outputs multi-level feature maps through additional convolutional layers to predict targets at different scales; then, the SE module is embedded in the SSD to improve the feature weights of key components; Loss calculation and model training: The loss function adopts the multi-task loss function; and hard example mining is used to backpropagate only to the negative samples with the highest loss. Training strategy optimization: RMSProp was used as the optimizer with an initial learning rate of 0.001; the shallow weights of the backbone network were frozen, and the deep layers and newly added convolutional layers were fine-tuned to accelerate convergence and prevent overfitting. Post-processing and equipment partitioning: Non-maximum suppression is used to filter predicted boxes by category, and the highest confidence results are retained to output the equipment type and bounding box; and semantic segmentation is used to assist partitioning, dividing key temperature zones according to equipment structure.
4. The method for detecting and analyzing current-induced heating devices based on artificial intelligence according to claim 1, characterized in that, The distance calculation formula in step 3 is as follows: Local distances are calculated for different zones, and a distance-temperature attenuation mapping table is established to compensate for temperature.
5. The method for detecting and analyzing current-induced heating devices based on artificial intelligence according to claim 1, characterized in that, Step 4 specifically involves: The thermal image is segmented based on device partitioning and image registration, and a calibrated segmented thermal image is generated based on temperature compensation. Step 401: Accurately map the device partition coordinates identified by the SSD network in Step 2 onto the thermal image; Step 402: For each pixel within the partition mask, apply compensation coefficients to generate a temperature calibration matrix; Step 403: The compensated temperature calibration matrix is superimposed with the grayscale value of the original thermal image to generate a calibrated thermal image corresponding to the physical temperature value and the pseudo-color.
6. The method for detecting and analyzing current-induced heating devices based on artificial intelligence according to claim 1, characterized in that, Step 5, which involves detecting temperature anomalies by comparing thermal imaging images with the model, specifically includes the following steps: Step 501: Thermodynamic features are extracted using a multi-scale feature pyramid. Step 502, Static benchmark comparison: including dynamic benchmark library construction and residual map generation; The dynamic benchmark library is constructed by using historical normal thermal images of the same type of equipment under the same operating conditions as a benchmark template, and extracting the average temperature T of the corresponding zone in the template. base Maximum allowable temperature rise ΔT allow and heat distribution patterns; The residual heatmap (ResidualMap) is generated using the following formula: ResidualMap = |T| current -T base |·Mask ROI Mask ROI For equipment partitioning mask; T current The actual temperature measurement value of the device or area at the current moment or under the current state; Step 503, Locate the abnormal region: Use dual threshold segmentation to locate the abnormal region. First threshold: mark the region where ΔT > ΔT in the residual heatmap. allow For the region, a secondary threshold is applied: morphological opening operations are used to eliminate noise while retaining connected regions with a minimum area of 5×5 pixels.
7. The method for detecting and analyzing current-induced heating devices based on artificial intelligence according to claim 1, characterized in that, Step 6 specifically involves matching the abnormal detection results with the constructed temperature defect analysis library to determine possible abnormal situations that may occur in the equipment.