A method and system for intelligent identification of hot spots in components based on YOLOv12

Abstract

This invention discloses a method and system for intelligent hot spot recognition of photovoltaic components based on YOLOv12, including step one, UAV flight and path planning; step two, data acquisition and image enhancement; step three, T-YOLOv12-PV network architecture and photovoltaic hot spot detection optimization; step four, image registration and intelligent hot spot detection verification; and step five, hot spot geographic information mapping and intelligent operation and maintenance report generation. In step two, the UAV simultaneously triggers an infrared thermal imager and a visible light camera, and the acquired images are enhanced in real time using AI algorithms. In this invention, the T-YOLOv12-PV intelligent detection layer, specifically optimized for photovoltaic scenarios, introduces a temperature prior map-guided attention mechanism and multi-task learning, which can accurately identify and distinguish between real hot spots and occlusion artifacts, significantly improving the detection accuracy and recall rate of small target hot spots in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic technology, specifically relating to a method and system for intelligent identification of hot spots in photovoltaic modules based on YOLOv12. Background Technology

[0002] Traditional hotspot detection primarily relies on manual ground inspections using handheld infrared thermal imagers or scheduled drone patrols based on pre-defined waypoints. Data analysis typically involves visual interpretation of infrared images or preliminary anomaly extraction using simple image processing techniques. Some more advanced methods employ traditional machine learning algorithms to classify extracted image features. These methods often rely on fixed inspection paths that are unsuitable for complex terrain, data processing is highly dependent on expert experience, and the analysis process is independent of data acquisition, resulting in low efficiency and poor consistency.

[0003] Application No. 202410813801.3 discloses an intelligent evolution analysis method and system for operational faults in new energy power plants, relating to the technical field of power plant fault analysis. It generates a set of conditional changes by real-time monitoring of photovoltaic module shading information and cell performance data, and combines this with an infrared thermal imager to identify hot spot distribution areas. This process can capture the dynamic changes of photovoltaic modules in a timely manner, further ensuring the accuracy and real-time nature of the analysis data. After identifying the hot spot distribution areas, it uses real-time monitored glass stress difference data and boundary condition data, combined with a trained mechanical model, to simulate the glass stress distribution caused by the hot spot effect. By obtaining the thermal stress coefficient Rfxs, it can accurately assess the degree of damage to the photovoltaic module glass caused by the hot spot effect. By fitting the fault prediction index Gyzs of the hot spot distribution area, it further evaluates the thermal stress gradient and compares it with a preset difference threshold Q to determine the risk level of glass breakage on the photovoltaic module surface.

[0004] The above comparative solutions require computer simulation and limited monitoring instruments to "predict" hot spots and stress. This results in poor real-time performance, a single data source, and delayed updates, making it unable to dynamically respond to the complex and ever-changing environment in actual inspections, such as sudden stains, bird obstruction, and instantaneous cloud shadows. The fault analysis process is cumbersome and non-automated, requiring manual intervention to perform multiple coefficient calculations and threshold comparisons, making it difficult to apply to rapid surveys and high-frequency inspections of large-scale sites. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a method and system for intelligent hot spot recognition of photovoltaic components based on YOLOv12, including: Step 1, UAV flight and path planning; Step 2, data acquisition and image enhancement; Step 3, T-YOLOv12-PV network architecture and photovoltaic hot spot detection optimization; Step 4, image registration and intelligent hot spot detection verification; Step 5, hot spot geographic information mapping and intelligent operation and maintenance report generation. In Step 2, the UAV simultaneously triggers an infrared thermal imager and a visible light camera, and the acquired images are enhanced in real time using AI algorithms. In this invention, the T-YOLOv12-PV intelligent detection layer, specifically optimized for photovoltaic scenarios, introduces a temperature prior map-guided attention mechanism and multi-task learning, which can accurately identify and distinguish between real hot spots and occlusion artifacts, significantly improving the detection accuracy and recall rate of small target hot spots in complex environments.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Drone flight and path planning; Step 2: Data Acquisition and Image Enhancement; Step 3: Constructing the T-YOLOv12-PV network and optimizing photovoltaic hot spot detection; Step 4: Image registration and intelligent hotspot detection verification; Step 5: Mapping hotspot geographic information and generating intelligent operation and maintenance reports.

[0007] Preferably, in step 2, the drone simultaneously triggers the infrared thermal imager and the visible light camera, and uses AI algorithms to enhance the acquired images in real time.

[0008] Preferably, the T-YOLOv12-PV network is optimized based on the YOLOv12 framework with attention centers, specifically including: multi-channel modification of the input layer, attention enhancement of the backbone network, and optimization of the feature pyramid and detector head; in the input layer, the standard 3-channel RGB input is expanded into a multimodal input. Adopt a lightweight solution or a complete solution; The lightweight solution adopts A total of 3 channels, the complete solution adopts There are 5 channels in total, and the number of channels in the first convolutional layer is adjusted accordingly; IR represents the infrared image. This represents a temperature prior diagram. This represents a grayscale image obtained by converting an RGB color image. The subsequent steps are the same for both the lightweight and complete solutions, as follows: For feature maps in the backbone network C, H, and W represent the number of channels, height, and width of the feature map, respectively; the temperature prior map... downsampling to get ,pass Convolution and the sigmoid function generate attention weights:

[0009] in, This represents a 1×1 convolution. This represents the normalized temperature prior plot after size matching. Weighted enhancement of features:

[0010] in The learnable scaling factor; It represents the Hadamardi (or Hadama) stack.

[0011] Preferably, in step 2, the infrared image is normalized for temperature field, and the average temperature is calculated for the entire infrared image or for the photovoltaic module area. and standard deviation Construct a normalized temperature rise map:

[0012] in Let be the numerical stability constant, and By compressing the temperature to the [0,1] interval through nonlinear mapping, a continuous temperature prior map is generated. ; Indicates the coordinates in the infrared image as The original temperature value of the pixel; At the same time, based on environmentally relevant thresholds:

[0013] in , For adjustment coefficients, Generate high-temperature candidate binary masks as a function of ambient light and temperature. The temperature prior map and high-temperature candidate mask will be provided as additional inputs to the subsequent YOLOv12 network.

[0014] Preferably, in step 3, the enhanced features are fed into the existing attention module of YOLOv12 for fine-grained focusing. In the feature pyramid part, for hotspots, the multi-scale feature fusion path is optimized. The detection head, based on the original bounding box regression and classification head, adds a temperature rise regression head and a risk level classification head. The total loss function... The corresponding adjustment is to multi-task joint optimization:

[0015] in, The weighting coefficients of the bounding box regression loss. The weighting coefficients of the category classification loss. The weighting coefficients of the regression loss due to temperature rise. The weighting coefficients represent the risk level classification losses. , , , These represent the hot spot target bounding box regression loss, hot spot category classification loss, newly added temperature rise regression loss, and newly added risk level classification loss, respectively. Each detector head output includes: bounding box position. Category labels, confidence levels, and predicted temperature rise values and risk score .

[0016] Preferably, in step 4, before detection and analysis, the system performs multimodal data registration and component region mapping, calibrates the intrinsic and extrinsic parameters of the visible light and infrared cameras, and uses the calibrated extrinsic parameter matrix... The infrared image is reprojected onto the visible light coordinate system to obtain the registered infrared image. and the corresponding temperature prior diagram On the registered visible light image, the outer contour of the photovoltaic module is first detected, and the bounding box coordinates and module number of each module are obtained. The bounding box coordinates of the module are:

[0017] Map component regions to and The system establishes an independent analysis area for each component, and performs initial screening of hot spots.

[0018] Preferably, in step 4, the average temperature of the component is obtained by statistically analyzing the pixel values ​​of the infrared image within the component area. and standard deviation The adaptive threshold is calculated using the following formula. :

[0019] in, It is a dynamically adjusted coefficient that reflects the effects of ambient light and temperature fluctuations; After the candidate region selection is completed, multi-dimensional verification is performed, and different algorithms are used to confirm the selected regions, based on the temperature gradient in the infrared image. The rate of temperature change and temperature gradient of the detection area are calculated using the following formula:

[0020] in, and These are the maximum and minimum temperatures within the region, respectively. The diameter of the measurement area.

[0021] Preferably, in step 4, the degree of grid line breakage is measured by calculating a "breakage index" at the breakage location in the image. This can be expressed by the following formula:

[0022] in, It is the length of the broken section of the grid line. It is the total length of the cell grid lines; The system combines historical data and machine learning algorithms to predict the probability of future failures in the area, using temperature fluctuation trends from historical data. The following formula is used to predict the likelihood of future failures in the region:

[0023] in, It is a historical temperature fluctuation trend. It is a machine learning model; The machine learning model is either a random forest or a support vector machine.

[0024] Preferably, in step 5, after hot spot detection is completed, the system maps the detection results with geospatial information, and sets the bounding box of each detected hot spot. Convert the component mapping relationship to the corresponding photovoltaic module number. ,in Represents the array row number. The component column number is used to form a standard identifier. Combined with the pose data collected in real time by the UAV and the pre-built digital elevation model, the pixel coordinates are transformed into three-dimensional geographic coordinates in the global coordinate system of the power station through the coordinate transformation matrix.

[0025] A component hotspot intelligent identification system based on YOLOv12 includes an intelligent flight and data acquisition layer, a data augmentation and preprocessing layer, a T-YOLOv12-PV intelligent detection layer, a multimodal registration verification and mapping layer, and a geographic information fusion and intelligent operation and maintenance layer. The intelligent flight and data acquisition layer is responsible for performing terrain-following flight and synchronous data acquisition covering the photovoltaic array through the dual-light imaging system carried by the UAV, combined with lidar scanning and a dynamic coordinate system. The data augmentation and preprocessing layer performs AI real-time augmentation on the acquired visible light and infrared images, and normalizes the infrared temperature data to generate a temperature prior map and a high temperature candidate mask. The YOLOv12-PV intelligent detection layer receives multimodal input and, through a temperature-guided attention mechanism and a multi-task detection head, enables real-time localization, classification, temperature rise prediction, and risk scoring of hot spots. The multimodal registration verification and mapping layer first registers the visible light and infrared images, and locates the outline of the photovoltaic module in the visible light image. Then, it maps the detection results to the specific module and performs multi-dimensional verification by combining temperature gradient and historical data. The geographic information fusion and intelligent operation and maintenance layer converts the verification results with the module number to the global coordinate system of the power station and generates an intelligent operation and maintenance report that includes geographic coordinates, risk classification, trend prediction and maintenance suggestions.

[0026] The beneficial effects of this invention are as follows: (1) By integrating a three-dimensional geofence, a dynamic coordinate system and a ground-following flight intelligent flight and data acquisition layer, and combining a temperature field normalization and AI-enhanced data augmentation and preprocessing layer, the system ensures complete coverage of the inspection path and high-quality, highly consistent multimodal data acquisition. The T-YOLOv12-PV intelligent detection layer, which is optimized for photovoltaic scenarios, can accurately identify and distinguish between real hot spots and occlusion artifacts by introducing a temperature prior map-guided attention mechanism and multi-task learning, while outputting temperature rise value and risk level, which greatly improves the detection accuracy and recall rate of small target hot spots in complex environments.

[0027] (2) The multimodal registration verification and mapping layer of this invention accurately anchors the algorithm detection results to specific physical components through rigorous image registration and component-level analysis, eliminating positioning ambiguity. Finally, the geographic information fusion and intelligent operation and maintenance layer directly transforms the detection results into a structured report with three-dimensional geographic coordinates, maintenance priorities, and predictive maintenance suggestions through coordinate transformation and trend analysis, effectively supporting the predictive maintenance of photovoltaic power plants, improving operation and maintenance efficiency, and reducing power generation losses. Attached Figure Description

[0028] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart illustrating the data augmentation process of this invention. Figure 3 This is a diagram of the T-YOLOv12-PV network architecture of the present invention; Figure 4 This is a flowchart of the registration and verification process for this invention; Figure 5 This is a simulation diagram illustrating the depth-separable convolution processing principle of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] A component hotspot intelligent identification method based on YOLOv12 includes the following steps: Step 1: Drone flight and path planning. The drone, equipped with a dual-light imaging system, flies over the photovoltaic array and generates a three-dimensional geofence by scanning with a pre-set lidar and combining it with a digital elevation model. A dynamic inspection coordinate system is established based on the solar azimuth angle and the tilt angle of the photovoltaic panels, and a terrain-following flight path covering all components is planned. Step two, data acquisition and image enhancement: During the data acquisition phase, the drone simultaneously triggers the infrared thermal imager and visible light camera, and uses AI algorithms to enhance the acquired images in real time. When lighting conditions are poor or weather conditions change significantly, the system automatically adjusts camera parameters based on environmental perception, optimizing exposure, gain, and contrast to ensure image clarity and detail. Especially in adverse weather or uneven lighting conditions, the AI ​​enhancement algorithm adjusts in real time to reduce environmental interference such as reflections and shadows, ensuring that the acquired images accurately reflect the hot spots and structural features of the photovoltaic modules. Regarding infrared temperature data processing, the system performs temperature field normalization on the infrared images and calculates the average temperature for the entire infrared image or for each photovoltaic module area. and standard deviation Construct a normalized temperature rise map:

[0031] in Let be the numerical stability constant, and By compressing the temperature to the [0,1] interval through nonlinear mapping, a continuous temperature prior map is generated. ; At the same time, the system uses environmentally relevant thresholds:

[0032] in , For adjustment coefficients, Generate high-temperature candidate binary masks as a function of ambient light and temperature. The temperature prior map and high-temperature candidate mask will be provided as additional inputs to the subsequent YOLOv12 network to provide "hot spot probability" prior guidance for the attention mechanism. Through this comprehensive enhancement and preprocessing, the image quality and the availability of temperature information are significantly improved, ensuring the stability and efficiency of data acquisition.

[0033] Step 3: Optimization of T-YOLOv12-PV network architecture and photovoltaic hotspot detection. The T-YOLOv12-PV network is optimized based on the YOLOv12 framework with attention centers. Specifically, this includes: multi-channel modification of the input layer, attention enhancement of the backbone network, and optimization of the feature pyramid and detection head. In the input layer, the standard 3-channel RGB input is expanded to a multimodal input: Lightweight solution A adopts... A total of 3 channels are used in the complete solution B. A total of 5 channels were used, and the number of channels in the first convolutional layer was adjusted accordingly to maintain compatibility with subsequent structures. A temperature-guided attention module was introduced into the backbone network, and TGA processing was inserted before the YOLOv12 R-ELAN or FlashAttention module. For feature maps in the backbone network... Temperature prior map downsampling to get ,pass Convolution and the sigmoid function generate attention weights:

[0034] Its shape is or Weighted enhancement of features:

[0035] in, The temperature-guided attention weight map is a spatial weight matrix derived from temperature information, used to indicate the salience of each location in the feature map relative to the hotspot. This represents the Sigmoid activation function, which compresses the values ​​of the convolution output to the [0,1] interval. To become a standardized weighting coefficient, The size of the one-dimensional convolution kernel is... The convolution operation performs a linear transformation and feature integration on the input temperature prior map along the channel dimension without changing its spatial dimensions. It is the temperature prior map after downsampling, i.e., the original temperature prior map. After spatial downsampling, the feature map is compared with the current feature map of the backbone network. same height and width The later version is used as prior information input into the attention generation module; This represents the output feature map after temperature-guided attention enhancement; it is an enhanced version of the original feature map after incorporating prior temperature information. This represents the input feature map of the current layer in the backbone network, and its shape is... It contains general visual features extracted by the network, symbols This represents element-wise multiplication, also known as the Hadamard product, which is used to precisely apply attention weights to each corresponding spatial location and channel of the feature map. This is a learnable scaling factor, a scalar or channel-number-related vector parameter that is automatically optimized by the network during training. It is used to dynamically adjust the overall influence strength of the temperature attention weights, avoiding over- or under-enhancement issues that may arise from fixed weights. This refers to the temperature-guided attention weight map generated by the first formula, which serves as the modulation signal for feature enhancement. The enhanced features are then fed into YOLOv12's existing attention module for fine-grained focusing. In the feature pyramid section, considering the characteristics of extremely small hotspot targets, the multi-scale feature fusion path is optimized, a high-resolution feature preservation mechanism is added, and the positive and negative sample allocation strategy for small targets is adjusted. In addition to the original bounding box regression and classification heads, the detection head adds a temperature rise regression head and a risk level classification head, and the loss function is adjusted accordingly to multi-task joint optimization.

[0036] While maintaining the real-time detection capabilities of YOLOv12, the entire network significantly improves the accuracy and practicality of hot spot detection through customized optimization for photovoltaic scenarios, supporting the online analysis needs of drone inspections. Furthermore, each detection head output includes: bounding box position. Category labels, confidence levels, and predicted temperature rise values and risk score ,in Using CIoU loss, FocalLoss is used to improve recall for small targets. Using SmoothL1 loss monitoring to improve temperature rise prediction accuracy Based on temperature rise forecasts Symbolic representation, and The meaning is the same. Depending on whether the risk level is represented continuously or discretely, the T-YOLOv12-PV network selects either MSE or cross-entropy loss. While maintaining the real-time detection efficiency of YOLOv12, the entire T-YOLOv12-PV network significantly improves the accuracy and scene adaptability of photovoltaic hot spot detection through temperature prior-guided attention enhancement, small-objective optimization architecture, and multi-task output design.

[0037] Step four, image registration and intelligent hotspot detection verification: Before detection and analysis, the system performs multimodal data registration and component region mapping, and calibrates the intrinsic and extrinsic parameters of the visible light and infrared cameras. The extrinsic parameter matrix obtained from the calibration is then used... The infrared image is reprojected onto the visible light coordinate system to obtain the registered infrared image. and the corresponding temperature prior diagram On the registered visible light image, the outer contour of the photovoltaic module is detected using the lightweight model YOLOv12-Nano or traditional image processing methods to obtain the bounding box coordinates and module number of each module. The bounding box coordinates of the module are as follows:

[0038] Map the component area to and Above, an independent analysis area is established for each component, and the system performs initial screening of hot spots: for each mapped component area, real-time data such as light and temperature are collected through environmental sensors, and adaptive threshold adjustment is performed in combination with AI algorithms; The system obtains the average temperature of the component by statistically analyzing the pixel values ​​of the infrared image within the component area. and standard deviation Based on these statistical characteristics, using the formula:

[0039] Calculate the adaptive threshold for each component, where This is a coefficient dynamically adjusted based on the real-time illumination and temperature fluctuations of the component; this adaptive threshold ensures that the system can accurately screen potential hotspot candidate regions for different components under different environments. After the candidate region screening is completed, the system will perform multi-dimensional verification: Analysis of temperature change rate within the temperature gradient region:

[0040] in, and These are the maximum and minimum temperatures within the region, respectively. The diameter of the measurement area; Assessing component structural integrity using the grid line fracture index:

[0041] in, It is the length of the broken section of the grid line. It is the total length of the cell grid lines; Combine the historical data of this component; Predicting failure probability using machine learning models:

[0042] in, It is a historical temperature fluctuation trend. and These are the average temperature and standard deviation of the current region, respectively. It is a machine learning model; Through this process from overall registration to component-specific analysis, the system can output hotspot detection results with precise component numbers and geographic coordinates.

[0043] Step 5: Hotspot Geographic Information Mapping and Intelligent Operation and Maintenance Report Generation. After hotspot detection is completed, the system accurately maps the detection results with geospatial information, mapping the bounding box of each detected hotspot. Convert the component mapping relationship to the corresponding photovoltaic module number. ,in Represents the array row number. The component column number is used to form a standard identifier. Combined with the pose data collected in real time by the UAV and the pre-built digital elevation model, the pixel coordinates are transformed to three-dimensional geographic coordinates in the global coordinate system of the power station through the coordinate transformation matrix. Based on this precise geographic mapping, the system generates a comprehensive hotspot defect report: the report includes not only the geographic coordinates of the hotspot, but also its temperature rise. Component number It also combines historical inspection data and uses trend analysis models to predict fault evolution trends. The report classifies maintenance priorities (urgent / high / medium / low) based on risk scores and recommends specific maintenance strategies for each hot spot. Finally, it outputs a complete intelligent operation and maintenance report containing a visualized hot spot distribution map, trend prediction curves, and structured maintenance suggestions, supporting predictive maintenance decisions for photovoltaic power plants.

[0044] A component hotspot intelligent identification system based on YOLOv12 includes an intelligent flight and data acquisition layer, a data augmentation and preprocessing layer, a T-YOLOv12-PV intelligent detection layer, a multimodal registration verification and mapping layer, and a geographic information fusion and intelligent operation and maintenance layer. The intelligent flight and data acquisition layer is responsible for performing terrain-following flight and synchronous data acquisition covering the photovoltaic array using a dual-light imaging system mounted on a drone, combined with lidar scanning and a dynamic coordinate system. The data enhancement and preprocessing layer performs real-time AI enhancement on the acquired visible light and infrared images and normalizes the infrared temperature data to generate a temperature prior map and high-temperature candidate masks. The T-YOLOv12-PV intelligent detection layer is the core of this system. It receives multimodal inputs and achieves real-time location, classification, temperature rise prediction, and risk scoring of hot spots through a temperature-guided attention mechanism and a multi-task detection head. The multimodal registration verification and mapping layer first registers the visible light and infrared images and locates the outline of the photovoltaic module in the visible light image. Then, it maps the detection results to specific modules and performs multi-dimensional verification by combining temperature gradients and historical data. The geographic information fusion and intelligent operation and maintenance layer converts the verification results with module numbers to the power plant's global coordinate system and generates an intelligent operation and maintenance report that includes geographic coordinates, risk classification, trend prediction, and maintenance suggestions.

Claims

1. A component hotspot intelligent identification method based on YOLOv12, characterized in that, Includes the following steps: Step 1: Drone flight and path planning; Step 2: Data Acquisition and Image Enhancement; Step 3: Constructing the T-YOLOv12-PV network and optimizing photovoltaic hot spot detection; Step 4: Image registration and intelligent hotspot detection verification; Step 5: Mapping hotspot geographic information and generating intelligent operation and maintenance reports.

2. The component hotspot intelligent identification method based on YOLOv12 according to claim 1, characterized in that, In step 2, the drone simultaneously triggers the infrared thermal imager and the visible light camera, and uses AI algorithms to enhance the acquired images in real time.

3. The component hotspot intelligent identification method based on YOLOv12 according to claim 2, characterized in that, The T-YOLOv12-PV network is optimized based on the YOLOv12 framework with attention centers, specifically including: multi-channel modification of the input layer, attention enhancement of the backbone network, and optimization of the feature pyramid and detector head; in the input layer, the standard 3-channel RGB input is expanded into a multimodal input. Adopt a lightweight solution or a complete solution; The lightweight solution adopts A total of 3 channels, the complete solution adopts There are 5 channels in total, and the number of channels in the first convolutional layer is adjusted accordingly; IR represents the infrared image. This represents a temperature prior diagram. This represents a grayscale image obtained by converting an RGB color image. The subsequent steps are the same for both the lightweight and complete solutions, as follows: For feature maps in the backbone network C, H, and W represent the number of channels, height, and width of the feature map, respectively; the temperature prior map... downsampling to get ,pass Convolution and the sigmoid function generate attention weights: in, This represents a 1×1 convolution. This represents the normalized temperature prior plot after size matching. Weighted enhancement of features: in The learnable scaling factor; It represents the Hadamardi (or Hadama) stack.

4. The component hotspot intelligent identification method based on YOLOv12 according to claim 3, characterized in that, In step 2, the infrared image is normalized for temperature field, and the average temperature is calculated for the entire infrared image or for each photovoltaic module area. and standard deviation Construct a normalized temperature rise map: in Let be the numerical stability constant, and By compressing the temperature to the [0,1] interval through nonlinear mapping, a continuous temperature prior map is generated. ; Indicates the coordinates in the infrared image as The original temperature value of the pixel; At the same time, based on environmentally relevant thresholds: in , For adjustment coefficients, Generate high-temperature candidate binary masks as a function of ambient light and temperature. The temperature prior map and high-temperature candidate mask will be provided as additional inputs to the subsequent YOLOv12 network.

5. The component hotspot intelligent identification method based on YOLOv12 according to claim 4, characterized in that, In step 3, the enhanced features are fed into the existing attention module of YOLOv12 for fine-grained focusing. In the feature pyramid part, the multi-scale feature fusion path is optimized for hotspots. The detection head, based on the original bounding box regression and classification head, adds a temperature rise regression head and a risk level classification head. The total loss function... The corresponding adjustment is to multi-task joint optimization: in, The weighting coefficients of the bounding box regression loss. The weighting coefficients of the category classification loss. The weighting coefficients of the regression loss due to temperature rise. The weighting coefficients represent the risk level classification losses. , , , These represent the hot spot target bounding box regression loss, hot spot category classification loss, newly added temperature rise regression loss, and newly added risk level classification loss, respectively. Each detector head output includes: bounding box position. Category labels, confidence levels, and predicted temperature rise values and risk score .

6. The component hotspot intelligent identification method based on YOLOv12 according to claim 5, characterized in that, In step 4, before detection and analysis, the system performs multimodal data registration and component region mapping, calibrates the intrinsic and extrinsic parameters of the visible light and infrared cameras, and uses the calibrated extrinsic parameter matrix... The infrared image is reprojected onto the visible light coordinate system to obtain the registered infrared image. and the corresponding temperature prior diagram On the registered visible light image, the outer contour of the photovoltaic module is first detected, and the bounding box coordinates and module number of each module are obtained. The bounding box coordinates of the module are: Map component regions to and The system establishes an independent analysis area for each component, and performs initial screening of hot spots.

7. The component hotspot intelligent identification method based on YOLOv12 according to claim 6, characterized in that, In step 4, the average temperature of the component is obtained by statistically analyzing the pixel values ​​of the infrared image within the component area. and standard deviation The adaptive threshold is calculated using the following formula. : in, It is a dynamically adjusted coefficient that reflects the effects of ambient light and temperature fluctuations; After the candidate region selection is completed, multi-dimensional verification is performed, and different algorithms are used to confirm the selected regions, based on the temperature gradient in the infrared image. The rate of temperature change and temperature gradient of the detection area are calculated using the following formula: in, and These are the maximum and minimum temperatures within the region, respectively. The diameter of the measurement area.

8. The component hotspot intelligent identification method based on YOLOv12 according to claim 7, characterized in that, In step 4, the degree of grid line breakage is measured by calculating the "breakage index" at the breakage location in the image. This can be expressed by the following formula: in, It is the length of the broken section of the grid line. It is the total length of the cell grid lines; The system combines historical data and machine learning algorithms to predict the probability of future failures in the area, using temperature fluctuation trends from historical data. The following formula is used to predict the likelihood of future failures in the region: in, It is a historical temperature fluctuation trend. It is a machine learning model; The machine learning model is either a random forest or a support vector machine.

9. The component hotspot intelligent identification method based on YOLOv12 according to claim 8, characterized in that, In step 5, after hot spot detection is completed, the system maps the detection results with geospatial information, and sets the bounding box of each detected hot spot. Convert the component mapping relationship to the corresponding photovoltaic module number. ,in Represents the array row number. The component column number is used to form a standard identifier. Combined with the pose data collected in real time by the UAV and the pre-built digital elevation model, the pixel coordinates are transformed into three-dimensional geographic coordinates in the global coordinate system of the power station through the coordinate transformation matrix.

10. A component hot spot intelligent identification system employing the component hot spot intelligent identification method as described in claim 1, characterized in that, It includes an intelligent flight and data acquisition layer, a data augmentation and preprocessing layer, a T-YOLOv12-PV intelligent detection layer, a multimodal registration verification and mapping layer, and a geographic information fusion and intelligent operation and maintenance layer; The intelligent flight and data acquisition layer is responsible for performing terrain-following flight and synchronous data acquisition covering the photovoltaic array through the dual-light imaging system carried by the UAV, combined with lidar scanning and a dynamic coordinate system. The data augmentation and preprocessing layer performs AI real-time augmentation on the acquired visible light and infrared images, and normalizes the infrared temperature data to generate a temperature prior map and a high temperature candidate mask. The YOLOv12-PV intelligent detection layer receives multimodal input and, through a temperature-guided attention mechanism and a multi-task detection head, enables real-time localization, classification, temperature rise prediction, and risk scoring of hot spots. The multimodal registration verification and mapping layer first registers the visible light and infrared images, and locates the outline of the photovoltaic module in the visible light image. Then, it maps the detection results to the specific module and performs multi-dimensional verification by combining temperature gradient and historical data. The geographic information fusion and intelligent operation and maintenance layer converts the verification results with the module number to the global coordinate system of the power station and generates an intelligent operation and maintenance report that includes geographic coordinates, risk classification, trend prediction and maintenance suggestions.

Images