An infrared temperature field and visible light feature fusion photovoltaic hot spot recognition method, system and device

Abstract

An infrared temperature field and visible light feature fusion photovoltaic hot spot recognition method, system and device. The method comprises: synchronously collecting a visible light image and an infrared temperature matrix; performing preliminary screening through comparison of the maximum value of the infrared temperature and a preset threshold value to filter out risk-free components; for components with temperature exceeding the limit, a target detection model is used to extract a reflection area in the visible light image and its confidence; through a homography transformation, the infrared abnormal points are mapped to the visible light plane to judge spatial coincidence; for suspicious targets with spatial coincidence, a non-symmetric double-confidence weighted decision function is used to calculate a comprehensive score, wherein the infrared temperature confidence weight is higher than the visible light reflection confidence, and the score is compared with a decision threshold value to output a real hot spot or a false hot spot determination result. The present application provides a cross-modal continuous probability fusion recognition method for an edge computing platform, which can distinguish between mirror reflection false hot spots and real hot spots, and is especially suitable for track or wheel type photovoltaic inspection devices equipped with a dual light vision acquisition module.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic module defect detection technology, specifically to a photovoltaic hot spot identification method, system, and device that integrates infrared temperature field and visible light characteristics. Background Technology

[0002] With the widespread adoption of photovoltaic (PV) power generation, hot spot detection of PV modules has become a crucial aspect of ensuring the safe operation of power plants. The hot spot effect refers to the phenomenon where certain cells in a PV module become a load and generate heat due to factors such as shading, microcracks, or performance degradation. In severe cases, this can lead to partial burnout of the module or even a fire. Therefore, efficient and accurate hot spot detection methods are essential for the safe operation and maintenance of PV power plants.

[0003] In the existing technology, there are several solutions for hot spot identification that fuse infrared and visible light information to address this problem.

[0004] The first type of solution involves dual-mode detection using both infrared and visible light. Patent CN113076816B discloses a method for identifying hot spots on solar photovoltaic modules based on infrared and visible light images. It employs a dual-model architecture combining an improved Mask R-CNN detection model and a ResNet causal analysis model, processing infrared and visible light independently before superimposing them for display. This method operates with the two modes running independently, lacking cross-modal decision-level fusion and failing to address the differentiation mechanism for specular reflection-induced false hot spots. Patent CN117173601B discloses a method for identifying hot spots on photovoltaic power station arrays. After image registration using SIFT+RANSAC, a fused image is obtained by subtracting the grayscale difference between infrared and visible light; a difference greater than a threshold is considered a hot spot. This method is essentially a simple arithmetic operation and cannot distinguish between "high grayscale visible light + high infrared temperature caused by sunlight reflection" and "pure infrared high temperature of a real hot spot." Jiang Lin et al. (Acta Energiae Solaris Sinica, 2022, 43(1): 393-397) proposed a photovoltaic array hot spot detection method based on the fusion of visible light and infrared thermal images. They used the Otsu method of OTU to perform threshold segmentation on both infrared and visible light to obtain binary images, and then used Boolean logic subtraction to directly remove the overlapping parts of the visible light segmentation region and the infrared segmentation region. This method employs a "hard rejection" strategy, unconditionally excluding any region that is determined to be reflective by the visible light OTU, leading to the missed detection of real hot spots that may exist simultaneously within the reflective region.

[0005] The second type of approach is a confidence-based photovoltaic detection method. Patent CN120876478A discloses an optimization method for photovoltaic panel hotspot detection based on a dual-model dynamic architecture, which uses confidence to drive adaptive selection of image resolution. This method only utilizes infrared single-mode information and does not use visible light reflectivity features to remove false hotspots; the confidence level is used for resolution selection rather than cross-modal decision fusion.

[0006] The third type of approach is a pure infrared image analysis method. Patent CN115115634B discloses a photovoltaic array hotspot detection method based on infrared images, which identifies hotspot regions through threshold segmentation, morphological operations, and local neighborhood maximum difference. This type of method only uses infrared images and cannot eliminate false hotspots caused by specular reflection. Patent CN122153741A discloses a photovoltaic power station infrared image feature extraction and anomaly diagnosis method, which judges hotspot risk based on temperature distribution change sequences. This method only utilizes infrared temperature matrix data and does not involve cross-modal image registration and fusion.

[0007] The fourth type of approach is a feature-level fusion method based on Transformer. Patent application CN202410965181 discloses a photovoltaic panel hotspot detection method based on dual-spectrum fusion, which uses ResNet50 to extract features separately and then performs cross-attention fusion via a Transformer encoder-decoder. This method requires significant computational power, is unsuitable for edge computing platforms, and the loss weighting is only applied to hyperparameters during the training phase, not to decision-level confidence weighting during the inference phase.

[0008] Based on the above existing technologies, the following core defects exist: (1) The cross-modal decision-making logic is crude - the existing fusion schemes generally adopt binary logic (Boolean subtraction, gray-scale difference subtraction, threshold comparison), that is, spatial overlap means the corresponding area is excluded, ignoring the possibility that real hot spots may exist in the reflective area at the same time, resulting in missed detection; (2) There is a lack of confidence weighting mechanism in the reasoning stage - there is no decision-making scheme that uses "infrared temperature confidence" and "visible light reflectivity confidence" to perform cross-modal continuous probability weighting; (3) The computing strategy is not adapted to the edge - most schemes adopt computationally intensive architectures such as Transformer and Mask R-CNN, without considering the real-time reasoning requirements of the edge platform; (4) There is no cascading filtering mechanism - no pre-screening stage is set up, which causes unnecessary computing power consumption in large-area inspection scenarios. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, the present invention provides a photovoltaic hot spot identification method, system and device that fuses infrared temperature field and visible light features, which mainly solves the problems of missed detection and unnecessary computing power consumption in existing schemes that fuse infrared and visible light information.

[0010] The technical solution of the present invention is as follows: A method for identifying photovoltaic hot spots by fusing infrared temperature field and visible light features includes the following steps: Step 1, Synchronous Acquisition Trigger: Acquire visible light images and infrared temperature matrices at the same time; when the main control unit issues the shooting command, it simultaneously encapsulates the system timestamp; the edge computing processing module binds the visible light images, infrared temperature matrices, and original temperature matrices acquired within the same command cycle to this timestamp, serving as the temporal consistency basis for subsequent cross-modal comparisons.

[0011] Step 2, Initial screening of the infrared temperature matrix: Obtain the maximum temperature T from the infrared temperature matrix. max If Tmax is lower than the preset temperature threshold Tth, then it is determined that there is no risk of hot spots and the identification of the current component ends; if Tmax reaches or exceeds Tth, then proceed to step 3; specifically: as the first-level filter of the system, the global maximum temperature T is directly retrieved from the infrared temperature matrix. max , and the preset temperature threshold T th Compare them. If T max Below T th If the current component is determined to have no hot spot risk, skip steps 3 to 5 and proceed directly to the next inspection position. If T max Reaching or exceeding T th This triggers subsequent advanced decision-making processes. Its function is to filter out the vast majority of normal components with extremely low computational cost, thereby significantly reducing the system's average power consumption and inference latency.

[0012] Step 3, Visible light reflectance feature extraction: Input the visible light image into the trained target detection model, and output the bounding box of at least one reflective region and its classification confidence score C. R C R ∈[0,1]; Specifically: The synchronously acquired visible light images are fed into the YOLO-V8 model, which is quantized by the neural network inference framework and deployed in the neural network acceleration unit of the edge computing processing module, and the bounding box coordinates and classification confidence score C of each reflective region are output. R The YOLO-V8 model introduces false hotspot negative samples such as uniform backlighting and metal frame reflections during the training phase, enabling it to distinguish specular reflection highlights from surface foreign objects and defects.

[0013] Step 4, Spatial Consistency Verification Based on Homography Transformation: Using a pre-calibrated homography transformation matrix H, the coordinates (x, y) of the abnormal high-temperature point in the infrared temperature matrix are mapped to the visible light image plane to obtain the mapped point (x', y'). It is then determined whether the mapped point falls within the bounding box of any of the reflective areas. If it does not, it is determined to be a real hot spot and an alarm is triggered; if it does, the abnormal high-temperature point is marked as a suspicious target, and the process proceeds to Step 5. Specifically, using a pre-calibrated homography transformation matrix H (a 3×3 matrix obtained through spatial calibration between the infrared and visible light sensors), the coordinates (x, y) of the abnormal high-temperature point in the infrared temperature matrix are mapped to the visible light image plane to obtain the corresponding coordinates (x', y'). It is then determined whether the mapped point falls within any of the reflective detection boxes output in Step 3. If it does not, the temperature anomaly is determined to be unrelated to specular reflection, and a real hot spot conclusion is directly output and an alarm is triggered; if it does, it is marked as "suspicious," and the process proceeds to Step 5 for quantitative decision-making. The key logic of this step is that spatial overlap is only considered "questionable" rather than directly rejected, because the end-side YOLO-V8 model itself may misidentify (such as labeling non-reflective areas as reflective). Directly excluding overlapping areas would lead to the missed detection of real hot spots. Another feature of this step is that it does not interpolate the temperature data—it only maps a single anomalous coordinate point to the visible light plane for inspection, avoiding the spurious data introduced by interpolation and augmentation of low-resolution temperature data.

[0014] Step 5, Double-confidence weighted decision: For the questionable target, according to the formula... Calculate its comprehensive decision score, where C T C represents the infrared temperature confidence level, characterizing the degree of temperature anomaly. T ∈[0,1];C R C represents the confidence level for visible light reflectance. R ∈[0,1]; α is the first weight, corresponding to the infrared temperature confidence level C. T β is the second weight, corresponding to the visible light reflectance confidence level C. R ; α and β satisfy the constraints α>β and α+β=1; The calculated S is compared with the preset decision threshold Sth: if S≥Sth, it is determined to be a real hot spot and an alarm is triggered; if S<Sth, it is determined to be a false hot spot caused by specular reflection and the alarm is suppressed, where Sth is the preset decision threshold, 0<Sth<1; Dual confidence weighted decision (core algorithm): For the suspicious targets marked in step 4, a quantitative decision function is introduced to conduct weighted competition between infrared temperature evidence and visible light reflection evidence in a continuous probability space, and the final classification is output based on the comparison result of the comprehensive score and the preset decision threshold.

[0015] The core formula for the dual-confidence weighted decision is:

[0016] Among them, C T C represents the infrared temperature confidence level, characterizing the degree of temperature anomaly. T ∈[0,1], based on T max Deviation from preset temperature threshold T th The degree of temperature calculation - the higher the temperature, the higher the C T The closer a value is to 1, the stronger the physical evidence for the hot spot; C R The visible light reflectance confidence score is directly given by the classification probability score output by the YOLO-V8 model, C. R ∈[0,1], the higher the score, the more significant the visual feature of specular reflection in that area; α is the first weight, corresponding to the infrared temperature confidence level C. T β is the second weight, corresponding to the visible light reflectance confidence level C. R ; α and β satisfy the constraints α>β and α+β=1, reflecting the priority of infrared temperature as direct physical evidence in decision-making.

[0017] The decision rule is: if S≥S th If S < S, it is determined to be a real hotspot and an alarm is triggered; th The hotspot was identified as a false hotspot (caused by specular reflection), and the alarm was suppressed. Among them, S... th The preset decision threshold is 0 < S th <1.

[0018] The design principle of this formula lies in transforming the decision relationship between infrared temperature confidence and visible light reflectance confidence from the traditional "binary mutual exclusion" to a linear weighted competition in a continuous probability space. Asymmetric weights (α>β) ensure that direct physical evidence, infrared temperature, has a higher say in the decision-making process—even if the visible light model determines a region as reflective with a high confidence level, as long as the physical temperature of that region is high enough, the overall score S may still exceed the decision threshold, and the system will still tend to classify it as a real hot spot, thus prioritizing the detection rate and avoiding missed detections.

[0019] Preferably, the YOLO-V8 model, after being quantized by the INT8 neural network inference framework, is deployed in the neural network acceleration unit of the edge computing processing module, and the inference frame rate can reach about 15 FPS under the condition of an input resolution of about 640×640.

[0020] Preferably, the initial temperature screening threshold T th The parameters are set according to the normal operating temperature range of the photovoltaic module, for example, 60℃; the first weight α is set to 0.55, the second weight β is set to 0.45; and the comprehensive decision threshold S is set. th Set it to 0.7.

[0021] Variations of the method include: the weighting function in step 5 is not limited to linear weighting; other fusion functions that satisfy the asymmetric constraint that "infrared temperature weight is higher than visible light reflectance weight" can also be used, such as weighted geometric mean, weighted harmonic mean, etc. Temperature confidence C T The calculation method is not limited to a specific function form, as long as it satisfies the monotonically increasing relationship that "the higher the temperature, the closer the confidence level is to 1".

[0022] Step 6, Alarm Output: Execute alarm triggering or suppression operations based on the judgment result of Step 4 or Step 5.

[0023] The target detection model is a YOLO-V8 model, which is deployed in the neural network acceleration unit of the edge computing processing module after being quantized by the neural network inference framework INT8.

[0024] The preset temperature threshold Tth is 60℃; the first weight α=0.55, the second weight β=0.45; and the preset decision threshold Sth=0.7.

[0025] The infrared temperature confidence level C T Calculation using piecewise linear mapping: The formula is as follows

[0026] And truncate to the interval [0,1], where T danger This is a preset hazardous temperature threshold.

[0027] The infrared temperature confidence level C T The Sigmoid function is used for calculation, and the formula is as follows:

[0028] Where k is a positive constant.

[0029] In step 5, the weighting function uses a weighted geometric mean or a weighted harmonic mean instead of a linear weighting form.

[0030] The abnormally high temperature point is the pixel point corresponding to the maximum temperature Tmax in the infrared temperature matrix.

[0031] A photovoltaic hot spot identification system that fuses infrared temperature field and visible light features includes: Dual-light vision acquisition module, used to simultaneously acquire visible light images and infrared temperature matrix; An edge computing processing module, including a neural network acceleration unit, is configured to execute the above-described method; The alarm output module is used to perform alarm triggering or suppression operations based on the determination result of the edge computing processing module.

[0032] A photovoltaic hot spot identification device that fuses infrared temperature field and visible light features includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The device is characterized in that the processor implements the above-described method when executing the program.

[0033] The beneficial effects of this invention are: (1) Missed detection protection mechanism: The processing strategy for spatial overlap is changed from the "direct exclusion" of the prior art to "quantitative decision after doubt", which protects the real hot spots that may exist in the reflective area from being mistakenly eliminated. This is the essential difference between this solution and the existing fusion solutions based on logical subtraction (such as Jiang Lin et al. 2022, CN117173601B); (2) Decision-making mechanism prioritizing physical evidence: An asymmetric dual-confidence weighting function is introduced, giving infrared temperature (direct physical evidence of hot spots) a higher weight than visible light reflectivity in the decision-making process. Even if the YOLO-V8 model determines a region as reflective with a higher confidence level, the system will still tend to classify it as a real hot spot as long as the physical temperature is high enough—this is crucial in photovoltaic power plant operation and maintenance scenarios where "avoiding missed detections" is the primary safety objective; (3) Real-time inference at the edge: By filtering out most normal frames through pre-temperature screening, and only triggering subsequent complex visual inference and spatial transformation calculations for frames with excessive temperature, the average power consumption and inference latency of the system are significantly reduced in the large-area photovoltaic array inspection scenario, so that the solution can be deployed on edge computing processing modules with limited computing resources; (4) Cross-modal continuous decision-making: The decision relationship between infrared and visible light is transformed from the existing "binary mutual exclusion" to "continuous probability competition". The decision result is no longer a simple "yes / no" but a smooth transition based on quantitative scoring. The weights and thresholds can be adjusted according to the actual application scenario, which is highly flexible. (5) No need to interpolate temperature data: The spatial mapping process only maps a single abnormal coordinate point to the visible light plane, without interpolating and expanding the low-resolution temperature matrix, thus avoiding the interference of pseudo-temperature data introduced by interpolation on subsequent judgment. Attached Figure Description

[0034] Figure 1 This is a system block diagram of the method of the present invention.

[0035] Figure 2 This is a schematic diagram of the system flow of the method of the present invention. Detailed Implementation

[0036] The invention will be further described below with reference to the accompanying drawings, showing that the method of the invention is deployed on a track-mounted photovoltaic inspection robot equipped with a dual-light vision acquisition module (visible light camera and infrared thermal imager). The robot inspects the surface of the photovoltaic module row by row. At each preset acquisition position, the main control unit issues a synchronous shooting command, and the vision processing board executes the six-level cascaded inference process of the method of the invention to determine online whether there are real hot spots on the current module.

[0037] The edge computing processing module employs an embedded system-on-chip (SoC) with an integrated neural network acceleration unit (NPU). The visible light camera and infrared thermal imager are fixedly mounted on the inspection device with a fixed relative pose; their relative pose is obtained through offline calibration to obtain the homography transformation matrix H. A polarizer can be optionally added to the image acquisition device to suppress some specular reflection interference.

[0038] like Figure 1 and Figure 2 As shown, the specific detection process in this embodiment is as follows: S1, Synchronous Acquisition Trigger: When the inspection robot reaches the preset acquisition position, the main control unit sends a synchronous shooting command to the dual-light vision acquisition module. The system timestamp is simultaneously encapsulated when the command is issued; the edge computing processing module binds the visible light image, infrared temperature matrix, and original temperature matrix acquired within the same command cycle to this timestamp, serving as the temporal consistency basis for subsequent cross-modal comparisons.

[0039] S2, Infrared Temperature Matrix Initial Screening: The edge computing processing module directly retrieves the global maximum temperature value Tmax from the infrared temperature matrix. Tmax is compared with a preset initial temperature screening threshold Tth (set to 60℃ in this embodiment). If Tmax < 60℃, the current component is determined to have no hot spot risk, and steps 3 to 5 are skipped, with the robot continuing to the next inspection position. If Tmax ≥ 60℃, it indicates that the current component has an abnormally high temperature area, triggering subsequent advanced judgment processes. This step filters out the vast majority of normal components with extremely low computational cost (only requiring traversing the infrared temperature matrix once to obtain the maximum value and compare). In large-area photovoltaic array inspection scenarios, normal components account for the vast majority; therefore, pre-screening can significantly reduce the system's average power consumption and inference latency.

[0040] S3, Visible Light Reflection Feature Extraction: The visible light image synchronously acquired in step 1 is fed into a Yolo-V8 model deployed on the NPU after INT8 quantization using the neural network inference framework. The model detects specular reflection highlight areas and metallic frame reflection areas in the visible light image, outputting the bounding box coordinates and corresponding classification confidence score CR for each reflection area. CR∈[0,1], with higher scores indicating more significant specular reflection visual features. In this embodiment, in addition to regular reflection samples, the Yolo-V8 model introduces false hotspot negative samples such as uniform backlighting and metallic frame reflection during the training phase, enabling it to distinguish specular reflection highlights from surface foreign objects / defects. After INT8 quantization, the model infers on the NPU, achieving an inference frame rate of approximately 15 FPS under an input resolution of 640×640, meeting the requirements for online real-time detection.

[0041] S4. Spatial consistency verification based on homography transformation: Using the 3×3 homography transformation matrix H obtained from offline calibration, the pixel coordinates (x, y) of the abnormal high temperature point corresponding to T_max in the infrared temperature matrix are mapped to the visible light image plane to obtain the corresponding pixel coordinates (x', y'). The mapping is achieved through homogeneous coordinate transformation:

[0042] Determine whether the mapped point (x', y') falls within any of the reflection detection boxes output in step 3. The determination method is as follows: for each reflection detection box, check whether the upper left corner coordinates (x1, y1) and the lower right corner coordinates (x2, y2) satisfy x1≤x'≤x2 and y1≤y'≤y2. If the mapped point does not fall within any reflection detection box, it indicates that there is no spatial correspondence between the abnormal high temperature point and the specular reflection area in the visible light image. The probability that this abnormal high temperature is caused by a real hot spot is extremely high, and it is directly determined as a real hot spot and an alarm is triggered. If the mapped point falls within a reflection detection box, it indicates that the abnormal temperature point and the reflection area spatially overlap. At this time, it is not directly determined as a false hot spot, but is marked as "doubtful" and proceeds to step 5 for quantitative decision-making.

[0043] The key technical features of this step are: (1) The spatial mapping process only maps a single abnormal coordinate point to the visible light plane for inspection, without interpolating and expanding the low-resolution infrared temperature matrix, thus avoiding the interference of pseudo-temperature data introduced by interpolation on subsequent judgment; (2) Spatial overlap is only regarded as doubtful rather than directly rejected - the YOLO-V8 model itself may misidentify, such as marking non-reflective areas as reflective, and if overlapping areas are directly excluded, it will lead to the missed detection of real hot spots. This strategy of "doubtful rather than directly rejected" is the essential difference between this method and existing fusion schemes based on logical subtraction (such as Jiang Lin et al. 2022, CN117173601B).

[0044] S5, double confidence weighted decision: For the targets marked as questionable in step 4, calculate the comprehensive decision score S.

[0045] First, calculate the infrared temperature confidence level C. T C T Based on T max Deviation from preset temperature threshold T th The degree of confidence is calculated to satisfy a monotonically increasing relationship where "the higher the temperature, the closer the confidence level is to 1". This embodiment uses a piecewise linear mapping method:

[0046] And truncate to the interval [0,1], where T danger This is a hazardous temperature threshold (e.g., set to 80°C). This mapping makes C... T In T max =T th Take 0.5 at T max ≥T danger A value of 1 is used to confirm a hot spot with extremely high confidence.

[0047] Visible light reflectance confidence level C R The classification probability score is given directly from the output of the YOLO-V8 model in step 3.

[0048] Taking α=0.55 and β=0.45, calculate the comprehensive decision score:

[0049] Compare S with the preset decision threshold S th =0.7 for comparison. If S≥0.7, it is determined to be a real hotspot, and an alarm is triggered. If S<0.7, it is determined to be a false hotspot caused by specular reflection, and the alarm is suppressed.

[0050] The asymmetric weighting design of the formula (α=0.55>β=0.45) ensures that direct physical evidence, such as infrared temperature, has a greater influence in decision-making. A typical example illustrates this: if a certain region T... max =75℃, calculated C T ≈0.875; the Yolo-V8 model achieved a high confidence level C R =0.9 indicates it is a reflective area. At this point, S = 0.55 × 0.875 + 0.45 × 0.9 ≈ 0.886, which is greater than 0.7, so the system still classifies it as a real hot spot. Even if the visible light model is "convinced" that the area is reflective, the system still tends to issue an alarm due to the sufficiently high physical temperature. This is a reasonable decision-making approach under the safety objective of "avoiding missed detections." Conversely, if T... max Only slightly above the threshold (e.g., T) max =62℃, C T≈0.55), and the visible light model determines it as reflective with extremely high confidence (C). R =0.95), then S = 0.55 × 0.55 + 0.45 × 0.95 ≈ 0.73, which is still higher than 0.7, but close to the decision boundary. If T max =61℃ and C R If α = 0.95, then S ≈ 0.70, which is exactly on the boundary. At this point, α and S can be fine-tuned according to the conservative or aggressive requirements of the actual application scenario. th To achieve different decision-making tendencies.

[0051] S6, alarm output, receives the judgment result from the upstream unit, performs alarm triggering or suppression operation, and sends the hot spot location coordinates and classification conclusion back to the main control system.

[0052] The above parameters were used to verify the results on a test sample set containing both real hot spots and specular reflection spurious hot spots. The test sample set included: (a) pure real hot spot samples (the infrared high-temperature region and the visible light reflective region do not overlap); (b) pure specular reflection spurious hot spot samples (the infrared high-temperature region and the visible light reflective region overlap, and the high temperature is caused by pure reflection); and (c) composite samples where real hot spots coexist within the reflective region (the infrared high-temperature region and the visible light reflective region overlap, but the high temperature is indeed caused by real hot spots). Experimental results show that compared to the single infrared temperature threshold determination method (using only T...),... max ≥T th (For hotspot identification), the false alarm rate of this method is reduced from about 54.7% to about 4.2%, and the performance is improved by about 92.3%. For composite samples (c), existing fusion schemes based on Boolean subtraction, such as the method of Jiang Lin et al. 2022, directly exclude overlapping regions, which inevitably leads to the missed detection of real hotspots in this scenario; while this method correctly detects real hotspots through asymmetric weighted decision-making.

[0053] Regarding end-to-end inference latency, in large-scale inspection scenarios, only about 5% to 15% of components trigger T. max ≥T th Under these conditions, most components skip the subsequent complex inference steps after the initial screening in step 2. For components that trigger the entire process, the model inference in step 3 takes approximately 67ms (15FPS), and the computational load of steps 4 and 5 is minimal (single-point coordinate transformation and one linear weighting). The end-to-end latency meets the requirements for online real-time detection.

[0054] Variations The weighting function in step 5 is not limited to the linear weighting form. Other fusion functions that satisfy the asymmetric constraint that "the infrared temperature weight is higher than the visible light reflection weight" can also be used, such as weighted geometric mean, weighted harmonic mean, etc.

[0055] Temperature confidence level C TThe calculation method is not limited to piecewise linear mapping; a Sigmoid function can also be used.

[0056] The requirement is simply that the confidence level increases with increasing temperature, approaching 1. The initial temperature screening stage is not limited to searching for global maximum values; it can also search for local maximum values ​​or temperature gradients, as long as it can quickly determine the presence of temperature anomalies with low computational cost.

[0057] The embodiments described with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention. The embodiments should not be considered as limiting the invention, but any improvements made based on the spirit of the invention should be within the scope of protection of the invention.

Claims

1. A method for identifying photovoltaic hot spots by fusing infrared temperature field and visible light features, characterized in that: Includes the following steps, Step 1, Synchronous Acquisition Trigger: Acquire visible light images and infrared temperature matrices at the same time; Step 2, Initial screening of the infrared temperature matrix: Obtain the maximum temperature T from the infrared temperature matrix. max If Tmax is lower than the preset temperature threshold Tth, it is determined that there is no risk of hot spots and the identification of the current component ends; if Tmax reaches or exceeds Tth, proceed to step 3. Step 3, Visible light reflectance feature extraction: Input the visible light image into the trained target detection model, and output the bounding box of at least one reflective region and its classification confidence score C. R C R ∈[0,1]; Step 4, Spatial consistency verification based on homography transformation: Using the pre-calibrated homography transformation matrix H, the coordinates (x, y) of the abnormal high temperature point in the infrared temperature matrix are mapped to the visible light image plane to obtain the mapped point (x', y'). It is determined whether the mapped point falls within the bounding box of any of the reflective areas: if it does not fall within, it is determined to be a real hot spot and an alarm is triggered; if it falls within, the abnormal high temperature point is marked as a suspicious target, and proceed to step 5. Step 5, Double-confidence weighted decision: For the questionable target, according to the formula... Calculate its comprehensive decision score, where C T C represents the infrared temperature confidence level, characterizing the degree of temperature anomaly. T ∈[0,1];C R C represents the confidence level for visible light reflectance. R ∈[0,1]; α is the first weight, corresponding to the infrared temperature confidence level C. T β is the second weight, corresponding to the visible light reflectance confidence level C. R ; α and β satisfy the constraints α>β and α+β=1; Compare the calculated S with the preset decision threshold Sth: if S≥Sth, it is determined to be a real hot spot and an alarm is triggered; if S<Sth, it is determined to be a false hot spot caused by specular reflection and the alarm is suppressed, where Sth is the preset decision threshold, 0<Sth<1. Step 6, Alarm Output: Execute alarm triggering or suppression operations based on the judgment result of Step 4 or Step 5.

2. The photovoltaic hot spot identification method based on the fusion of infrared temperature field and visible light features according to claim 1, characterized in that: The target detection model is a YOLO-V8 model, which is deployed in the neural network acceleration unit of the edge computing processing module after being quantized by the neural network inference framework INT8.

3. The photovoltaic hot spot identification method based on the fusion of infrared temperature field and visible light features according to claim 1, characterized in that: The preset temperature threshold Tth is 60℃; the first weight α=0.55, the second weight β=0.45; and the preset decision threshold Sth=0.

7.

4. The photovoltaic hot spot identification method based on the fusion of infrared temperature field and visible light features according to claim 1, characterized in that: The infrared temperature confidence level C T Calculation using piecewise linear mapping: The formula is as follows And truncate to the interval [0,1], where T danger This is a preset hazardous temperature threshold.

5. The photovoltaic hot spot identification method based on the fusion of infrared temperature field and visible light features according to claim 1, characterized in that: The infrared temperature confidence level C T The Sigmoid function is used for calculation, and the formula is as follows: Where k is a positive constant.

6. The photovoltaic hot spot identification method based on the fusion of infrared temperature field and visible light features according to claim 1, characterized in that: In step 5, the weighting function uses a weighted geometric mean or a weighted harmonic mean instead of a linear weighting form.

7. The photovoltaic hot spot identification method based on the fusion of infrared temperature field and visible light features according to claim 1, characterized in that: The abnormally high temperature point is the pixel point corresponding to the maximum temperature Tmax in the infrared temperature matrix.

8. A photovoltaic hot spot identification system that fuses infrared temperature field and visible light characteristics, characterized in that, include: Dual-light vision acquisition module, used to simultaneously acquire visible light images and infrared temperature matrix; An edge computing processing module, including a neural network acceleration unit, is configured to perform the method of any one of claims 1 to 7; The alarm output module is used to perform alarm triggering or suppression operations based on the determination result of the edge computing processing module.

9. A photovoltaic hot spot identification device that fuses infrared temperature field and visible light features, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 1 to 7.

Images