Intelligent station photovoltaic module fault diagnosis system based on unmanned aerial vehicle inspection

Abstract

The invention discloses a smart station photovoltaic module fault diagnosis system based on unmanned aerial vehicle inspection, and relates to the technical field of photovoltaic power generation station operation and maintenance detection and fault diagnosis. Time-varying shadows are distinguished from candidate defects by constructing a multi-frame sequence for the surface area of the same module and calculating pixel-level time sequence dynamics; a dynamic noise mask and a stable feature mask are generated, so that follow-up diagnosis is preferentially based on stable region information, and the probability that dynamic light and shadow are misjudged as defects is reduced; infrared temperature rise features and visible light appearance morphological features are jointly extracted in a stable feature mask, thermal anomaly evidences and appearance anomaly evidences are put into the same diagnosis chain for mutual identification, and the situation that a wrong conclusion is directly triggered due to evidence distortion of a single mode under the strong interference condition is avoided.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant operation and maintenance testing and fault diagnosis technology, and in particular to a smart photovoltaic module fault diagnosis system based on drone inspection. Background Technology

[0002] Currently, automated inspection solutions using drones have emerged in the operation and maintenance of photovoltaic power plants. Drones are equipped with radiation-type infrared thermal imagers and high-resolution visible light cameras to acquire images of the modules, and combine deep learning and other methods to identify and locate thermal and appearance anomalies. Among them, thermal infrared images are helpful in discovering temperature rise characteristics such as hot spots and anomalies related to sub-strings / bypass diodes, while visible light images are more suitable for recording appearance defects such as glass breakage, delamination, shading, and heavy dust / bird droppings. To ensure data quality, existing acquisition processes usually need to constrain flight altitude, overlap, and camera angle to reduce the impact of glare reflected from the module surface on imaging.

[0003] However, when the above-mentioned technology system is applied to water surface / floating photovoltaic power stations, due to the limited accessibility of the array, inspection tends to rely more on aerial methods such as drones. At the same time, the on-site environment presents stronger maintenance characteristics such as humidity, corrosion and bird droppings. In terms of imaging, the water surface is prone to solar flares and sky reflections that change with the wind and waves. Highlight areas and textures change rapidly over time. In addition, the slight undulation of the floating body causes fluctuations in the observation angle, making it easier for irregular bright spots, local overexposure or reflective textures to appear in visible light images, thus causing continuous interference to defect identification based on texture / brightness features.

[0004] To address complex lighting interference, existing studies have proposed strategies such as infrared and visible light multimodal fusion and using multi-frame consistency for screening / deduplication to improve robustness. However, in cases where strong reflections and dynamic light and shadows highly overlap with the physical location of components, false detections and missed detections may still occur, leading to a decrease in the confidence of defect identification and affecting the reliability of diagnostic conclusions. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a smart photovoltaic module fault diagnosis system based on drone inspection to solve the problem of severely reduced confidence in defect identification caused by dynamic noise from reflections and shadows on water surfaces.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a smart photovoltaic module fault diagnosis system for power plants based on unmanned aerial vehicle (UAV) inspection, comprising:

[0009] The drone is equipped with an infrared thermal imager and a visible light camera.

[0010] Communication unit;

[0011] The analysis server receives infrared image data and visible light image data using the communication unit.

[0012] The analysis server is configured to register multiple frames of infrared images and multiple frames of visible light images obtained at consecutive time points on the surface area of ​​the same photovoltaic module and construct an image sequence.

[0013] Pixel-level temporal dynamics indices are calculated based on the image sequence, and dynamic noise masks and stable feature masks are generated.

[0014] Temperature rise features and appearance features are extracted within a stable feature mask; an electrothermal coupling model is run based on the electrical parameters of the electrical branch corresponding to the surface region and the environmental parameters to obtain the theoretical heat distribution, and a consistency score is given for the theoretical heat distribution, the temperature rise features and the appearance features; the fault diagnosis status of the surface region is output based on the consistency score.

[0015] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the image sequence contains at least 3 frames, and the acquisition time interval between two adjacent frames is less than a preset time window, so that dynamic light and shadow interference such as reflections and shadows presents observable time-varying characteristics in the sequence.

[0016] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the pixel-level temporal dynamics index includes at least one of the mean, variance, or weighted combination of the intensity difference between corresponding pixels in adjacent frames after registration.

[0017] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the stable feature mask is obtained by comparing the pixel-level temporal dynamic index with a threshold, and the threshold is adaptively determined based on the statistical distribution of the temporal dynamic index within the surface area of ​​the photovoltaic module.

[0018] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the registration includes: calculating inter-frame geometric transformation based on feature point matching or optical flow estimation, and performing cross-modal alignment between infrared and visible light images.

[0019] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the temperature rise characteristic includes at least one of the following: temperature rise amplitude relative to the surrounding background, temperature rise connected domain area, or temperature gradient.

[0020] The appearance features include at least one of bright spot morphology, edge / crack candidate texture, or stain occlusion candidate texture.

[0021] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the operating electrical parameters include at least one of current and voltage, and the environmental parameters include at least one of ambient temperature, wind speed, or irradiance, and are provided by the site monitoring system or sensor system.

[0022] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the consistency score is obtained by fusing a first verification factor and a second verification factor; wherein the first verification factor characterizes the degree of matching between the temperature rise feature and the theoretical heat distribution, and the second verification factor characterizes the degree of matching between the appearance morphology feature and the pre-stored defect morphology feature library.

[0023] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the analysis server is further configured to: when the consistency score is in the suspected range, issue a reshoot command to the UAV, causing the UAV to change at least one of the shooting angle, exposure, or focal length of the suspected area to obtain supplementary images.

[0024] As a preferred embodiment of the intelligent photovoltaic module fault diagnosis system based on UAV inspection described in this invention, the repeat shooting instruction further includes changing the polarization filter conditions or changing the incident angle range to reduce high brightness interference caused by specular reflection.

[0025] Through the above technical solution, the present invention can achieve at least the following beneficial effects:

[0026] To address the issue of false detection caused by the formation of bright spot textures similar to defects in visible light images due to mirror reflections, reflections, and ripple projections on water surfaces, which change rapidly over time, a multi-frame sequence is constructed for the surface area of ​​the same component, and pixel-level temporal dynamics are calculated. This allows time-varying light and shadow to be distinguished from candidate defects, generating dynamic noise masks and stable feature masks. Subsequent diagnosis prioritizes information from stable regions, reducing the probability of dynamic light and shadow being misjudged as defects.

[0027] To address the issue that dynamic light and shadow highly overlap with the physical location of components, making it difficult to separate them using only a single frame or mode, thus reducing the confidence level of real defects, this paper proposes to jointly extract infrared temperature rise features and visible light appearance morphology features within a stable feature mask. This integrates thermal anomaly evidence and appearance anomaly evidence into the same diagnostic chain for mutual verification, avoiding the distortion of evidence under strong interference conditions by a single mode and the direct triggering of erroneous conclusions.

[0028] To address the issue that existing multimodal fusion methods may still misinterpret instantaneous brightness caused by reflections as anomalies or texture drift caused by reflections as surface defects, this paper introduces an electrothermal coupling model based on electrical and environmental operating conditions to generate theoretical heat distribution predictions. This model is then matched with observed temperature rise characteristics, and physical interpretability is used as a constraint. This suppresses false anomalies that do not conform to the electrothermal mechanism at the scoring level, thereby improving the credibility and interpretability of diagnostic conclusions.

[0029] To address the issues of unstable diagnosis caused by rapid changes in scene lighting and potential overexposure, blurring, or registration failure in single acquisitions, a suspected range and reshoot trigger strategy are set. When the result is uncertain, a reshoot command is automatically issued to adjust the shooting angle, exposure, or focal length. Furthermore, the polarization filter conditions or incident angle range can be changed to reduce high-brightness interference, achieving closed-loop optimization from acquisition to diagnosis and reducing misjudgments caused by fluctuations in the quality of a single acquisition.

[0030] To address the need for rapid location and actionable conclusions in the operation and maintenance of water surface stations, a unified process is developed that combines dynamic noise isolation, stable area feature extraction, physical consistency scoring, and status output. This process not only provides a diagnostic status but also offers categorized conclusions such as thermal anomalies and appearance anomalies, providing a more direct basis for decision-making in subsequent maintenance actions such as re-inspection, cleaning, and replacement. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0032] Figure 1 This is a framework diagram of a smart power station photovoltaic module fault diagnosis system based on drone inspection, as shown in the embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0035] Example 1:

[0036] like Figure 1 As shown, this application proposes a smart power station photovoltaic module fault diagnosis system based on UAV inspection, which is configured with:

[0037] The drone is equipped with an infrared thermal imager and a visible light camera.

[0038] Communication unit;

[0039] The analysis server receives infrared and visible light image data via a communication unit.

[0040] In this invention, the photovoltaic module surface area refers to a target area on the module surface in the image coordinate system, which can be the full area of ​​a single module, a sub-region of the module, or a set of areas composed of several pixels; the image sequence refers to a set of multiple frames of images collected at consecutive time points for the same photovoltaic module surface area, including infrared image frames and visible light image frames at the corresponding time; the dynamic noise mask refers to a binary or multi-value mask that identifies pixels / regions that change significantly over time, used to characterize dynamic light and shadow interference areas such as reflections, shadows, and ripple projections; the stable feature mask refers to a mask that identifies pixels / regions that do not change significantly within a time window; the temperature rise feature refers to a set of features related to temperature anomalies obtained from infrared images; the appearance morphology feature refers to a set of features related to appearance anomalies obtained from visible light images; the consistency score refers to a score or confidence level obtained by fusing the matching degree of theoretical thermal distribution and observed temperature rise features with the matching degree of appearance morphology features; the fault diagnosis status includes at least one or more of the three categories of normal, suspected, and fault, and can be further subdivided into subcategories such as thermal anomaly, shading / stain, and structural appearance.

[0041] The analysis server is configured to register multiple frames of infrared images and multiple frames of visible light images obtained at consecutive time points on the surface area of ​​the same photovoltaic module and construct an image sequence; the determination of the surface area of ​​the same photovoltaic module can be achieved using any of the following methods or a combination thereof:

[0042] (1) Component detection is performed based on geometric features such as the edge of the component board frame, the position of the junction box, and the corner points, and the polygons generated at the four corners of the detected components are used as the target areas;

[0043] (2) Based on the GIS / aerial survey orthophoto of the station component layout, the UAV image is projected onto a unified coordinate system and the target area is determined by mapping the component row and column numbers;

[0044] (3) In the case of manual annotation or initial database construction, the target tracking method is used to map the target region of the previous frame to the next frame through inter-frame geometric transformation, and triggers re-detection when the confidence is insufficient.

[0045] When detection or tracking fails, the system marks the frame as invalid and removes it from the image sequence, or triggers a retake to complete the sequence.

[0046] Pixel-level temporal dynamic indicators are calculated based on image sequences to generate dynamic noise masks and stable feature masks. To avoid misclassifying real defects into dynamic noise regions, the stable feature mask can add consistency constraints: when a region exhibits continuous temperature rise anomalies in an infrared image and has spatial connectivity within a time window, even if there are local reflective changes in visible light, the region can be marked as a candidate stable anomaly region and included in subsequent physical consistency scoring; conversely, when a region is mainly bright in visible light and does not support infrared temperature rise, and its dynamic indicators are significantly higher than the surrounding background, it is preferentially classified into the dynamic noise mask to reduce its impact on fault determination.

[0047] Temperature rise characteristics and appearance morphology characteristics are extracted within a stable feature mask; an electrothermal coupling model is run based on the electrical parameters of the electrical branches corresponding to the surface area and environmental parameters to obtain the theoretical heat distribution, and a consistency score is given for the theoretical heat distribution, temperature rise characteristics, and appearance morphology characteristics; the fault diagnosis status of the surface area is output based on the consistency score; the consistency score can be obtained by fusing a first verification factor and a second verification factor, and at least one judgment threshold or judgment range is set: when the consistency score is lower than the first threshold, normal is output; when the consistency score is in the suspected range, suspected is output and a re-shoot is triggered; when the consistency score is higher than the second threshold, fault is output.

[0048] The diagnostic status can further include fault type labels: when the first check factor is dominant and there is a continuous temperature rise anomaly, output a thermal anomaly type; when the second check factor is dominant and the appearance match is clear, output an appearance anomaly type; when both are supported and the consistency score is the highest, output a composite anomaly or a fault type with higher confidence.

[0049] The correspondence between surface areas and electrical branches can be established through component layout and wiring topology: when the database is built for the first time, a mapping table is established between component row and column numbers and group string numbers; after the UAV acquires the image, the component number is obtained through component identification, the electrical branch number to which it belongs is obtained from the mapping table, and the corresponding electrical parameter data stream is selected accordingly.

[0050] When the mapping table is missing or incomplete, manual annotation of the correspondence between component numbers and electrical branches is allowed, and the annotation results are written into the mapping table for subsequent automated diagnosis.

[0051] The electrothermal coupling model is used to generate the theoretical heat distribution or expected temperature rise of a target area under given operating electrical and environmental parameters. The model can adopt an engineering approximation: based on the balance between the heat generation and convective heat dissipation terms of the component under the current operating conditions, the component temperature field is estimated by combining ambient temperature, wind speed, and irradiance conditions; the theoretical heat distribution can be output as a pixel-level two-dimensional distribution or as a zone-level expected temperature rise (e.g., divided by substring / cell region).

[0052] In this embodiment, the image sequence contains at least 3 frames, and the acquisition time interval between two adjacent frames is less than a preset time window, so that dynamic light and shadow interference such as reflections and shadows can be observed to be time-varying in the sequence.

[0053] To improve the availability of the sequence, the acquisition side can set quality control rules: the flight altitude and ground resolution should meet the requirements for component defect identification; adjacent frames should maintain sufficient overlap to support registration; when overexposure, severe motion blur, or obvious thermal image defocus occurs, the frame should be marked as an invalid frame and discarded; when the number of valid frames is insufficient to form a sequence, a reshoot should be triggered.

[0054] In this embodiment, the pixel-level temporal dynamics index includes at least one of the mean, variance, or weighted combination of the pixel intensity difference between adjacent frames after registration.

[0055] Pixel-level temporal dynamics metrics can be calculated on the registered image sequence, preferably using robust statistics to reduce the impact of occasional noise.

[0056] For each pixel location, calculate the sequence of intensity differences between adjacent frames, and calculate the mean, variance, range, or absolute deviation of the median, etc.

[0057] To reduce exposure variations and overall brightness shifts in cloud shadows, local normalization or background subtraction can be performed on each frame before calculating the difference.

[0058] For abnormal pixels caused by overexposure, saturation, or obvious occlusion, an invalid flag can be set, and an invalid value can be ignored during statistics.

[0059] Optionally, the dynamic index can be spatially smoothed or morphologically processed to obtain a coherent dynamic noise region.

[0060] In one implementation, after registering multiple frames of infrared images with multiple frames of visible light images, the surface area of ​​the same photovoltaic module is denoted as a pixel set. In continuous For each pixel position in a frame of visible light image Calculate pixel-level temporal dynamics metrics to make the highlight jumps caused by specular reflection and the texture drift caused by water reflection exhibit measurable time-varying properties.

[0061] When there are exposure or brightness fluctuations between visible light frames, for the first... Frame visible light intensity By performing robust normalization, we obtain the standardized strength:

[0062] ,

[0063] in, Indicates the first Frame at pixel position Standardization intensity Indicates the first Frame at pixel position The original visible light intensity, Indicates the first Frame in region The stable position of the quantity within. Indicates the first Frame in region Intra-scale robustness;

[0064] The above and The normalization parameters are constructed using the median and the median absolute deviation, ensuring that a small number of bright pixels do not dominate the normalization parameters.

[0065] ,

[0066] in, Indicates the region Inner pixel position The operator for taking the median. This represents the coefficient used to convert the absolute deviation of the median into a scaling factor. Indicates prevention A stable term that is zero;

[0067] After obtaining the normalized intensity, a pixel-level difference sequence is constructed based on the differences between adjacent frames:

[0068] ,

[0069] in, Indicates pixel position In the Frame and the Intensity difference amplitude between frames, Represents the absolute value operator. Represents the number of frames in an image sequence and satisfies , Indicates the frame index;

[0070] Robust statistics are calculated for each pixel based on the difference sequence to characterize the degree of temporal jumps and fluctuations:

[0071] ,

[0072] in, Indicates pixel position The mean difference between adjacent frames, Indicates pixel position The standard deviation of the difference between adjacent frames, Represents the summation operator;

[0073] When occasional spikes exist in the reflective brightness, the median absolute bias is introduced to suppress the impact of these spikes on dynamic performance evaluation.

[0074] ,

[0075] in, Indicates pixel position In time index The median difference in the dimension, Indicates to The operator for taking the median. Indicates pixel position The differential robust volatility;

[0076] To identify texture drift caused by water reflections, the optical flow amplitude of adjacent visible light frames is used as a dynamic complement; for the 1st... Frame to the Frame at pixel position Estimating optical flow vectors The value of its width is:

[0077] ,

[0078] in, Indicates pixel position The optical flow amplitude between two adjacent frames Represents the L2 norm operator, Indicates pixel position The optical flow vector, Indicates pixel position Median optical flow amplitude over time;

[0079] By fusing the difference mean, difference standard deviation, difference robustness fluctuation, and median optical flow amplitude, a pixel-level temporal dynamics index is obtained:

[0080] ,

[0081] in, Indicates pixel position Time-series dynamic indicators, The fusion weights represent the difference mean. The fusion weights represent the standard deviations of the differences. The fusion weights represent the differential robust volatility. The fusion weight represents the median optical flow amplitude;

[0082] To ensure consistent weight values, a normalized weight relationship is adopted:

[0083] ,

[0084] The above inequalities are used to limit the range of values ​​for each fusion weight, and the meaning of each weight is referred to in the aforementioned definition;

[0085] In the mapping from dynamic indicators to masks, single-threshold or double-threshold hysteresis rules are introduced to reduce false detections; quantile adaptive threshold determination is used to ensure that the dynamic labeling ratio has a controllable upper bound within the target area.

[0086] ,

[0087] in, Indicates a low threshold. Indicates a high threshold. Indicates in the region Internal Pick quantile operators, Indicates low quantiles. Indicates high quantiles;

[0088] Under the hysteresis rule, the dynamic noise mask is jointly determined by the high-threshold seed and the low-threshold candidate, reducing dynamic mislabeling caused by isolated noise.

[0089] ,

[0090] in, This indicates a high-threshold seed mask. Indicates a low-threshold candidate mask. Indicates a dynamic noise mask. This is an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise. Indicates in Inner selection and Operators for connected components that have connectivity relations;

[0091] The stable feature mask is extracted from low-dynamic pixels, allowing subsequent temperature rise features and appearance features to avoid reflective and drift areas:

[0092] ,

[0093] in, This represents a stable feature mask; the meanings of the other parameters are as defined above.

[0094] The basis for controlling the false detection rate of quantile thresholds is reflected in: when When taking the range of 0.90 to 0.98, Corresponding area The upper tail portion of the internal dynamic index, the dynamic noise mask is triggered by the upper tail seed and expands within the low threshold candidate, causing the dynamic label ratio to vary. There is an upper bound constraint relationship; as the reflective area increases, The statistical distribution shows an upward tail. and This adaptively improves performance, thereby reducing the likelihood of misclassifying stable regions as dynamic regions.

[0095] Specifically, the dynamic index in the above implementation is based on a multi-frame visible light sequence of the same region after registration. By robustly normalizing the brightness fluctuations within the frame, the impact of exposure changes on temporal difference is reduced. After constructing the difference sequence of adjacent frames at the pixel level, the mean and fluctuation are calculated simultaneously, so that the brightness jumps caused by specular reflection form a statistically distinguishable time-varying feature. To address spike interference, the median absolute deviation is introduced to suppress the influence of a small number of outliers on dynamic estimation. To address the texture displacement caused by water reflection, which is prone to being missed by relying solely on the difference amplitude, the optical flow amplitude is used as a supplementary term, so that texture drift is reflected in the dynamic index. The mapping between the dynamic index and the mask adopts a quantile adaptive threshold, combined with connected component selection with double threshold hysteresis, to reduce dynamic mislabeling caused by isolated noise and isolate unstable regions within the dynamic noise mask, providing more stable pixel support for the subsequent extraction of temperature rise features and appearance features, thereby reducing false detections caused by reflections and mirrors.

[0096] In this embodiment, the stable feature mask is obtained by comparing pixel-level temporal dynamic index with a threshold, and the threshold is adaptively determined based on the statistical distribution of the temporal dynamic index in the surface area of ​​the photovoltaic module.

[0097] The threshold can be adaptively determined in the following manner:

[0098] Quantile strategy: Use the high quantile of the statistical distribution of dynamic index in the target area as the threshold so that only the most unstable pixels enter the dynamic noise mask.

[0099] Dual threshold hysteresis: A high threshold is set to determine entry into the dynamic noise region, and a low threshold is set to determine exit from the dynamic noise region, in order to reduce mask flicker;

[0100] Scene self-learning: Using the distribution of dynamic indicators of multiple component regions within the same voyage as a reference, the global threshold under the current wind / wave / light conditions is obtained, and fine-tuning is allowed for local areas.

[0101] The goal of threshold selection is to prioritize areas with significant time-varying characteristics, such as specular reflection, reflections, and ripple projections, into the dynamic noise mask, while preserving fixed component structures and long-term defect areas in the stable feature mask.

[0102] In this embodiment, registration includes: calculating inter-frame geometric transformation based on feature point matching or optical flow estimation, and performing cross-modal alignment between infrared and visible light images;

[0103] Cross-modal alignment optimization includes two parts: time synchronization and spatial calibration.

[0104] Time synchronization: The acquisition timestamps of the infrared thermal imager and the visible light camera are aligned through unified triggering or software synchronization; when there is a time deviation, interpolation or nearest neighbor matching is used to make the infrared frame and the visible light frame correspond within the same time window;

[0105] Spatial calibration: Intrinsic and relative extrinsic parameter calibrations are performed on infrared and visible light imaging systems to obtain the mapping relationship from infrared coordinates to visible light coordinates. In actual calculations, the infrared image can be resampled to visible light resolution, or both can be mapped to a unified reference coordinate system, and an evaluation index of alignment error is given. When the alignment error exceeds the preset allowable range, the frame pair will not participate in subsequent pixel-level dynamic calculations and feature extraction.

[0106] In this embodiment, the temperature rise feature includes at least one of the temperature rise amplitude relative to the surrounding background, the area of ​​the temperature rise connected domain, or the temperature gradient.

[0107] Appearance features include at least one of bright spot morphology, edge / crack candidate texture, or stain occlusion candidate texture;

[0108] Temperature rise characteristics may include at least: the difference between the regional average temperature and the average temperature of the surrounding reference area, the area and shape parameters of the temperature rise anomaly connected domain, and the temperature gradient or boundary steepness parameters; appearance morphology characteristics may include at least: the morphological parameters of suspected bright spots, the intensity and orientation consistency parameters of edge / linear structures, and texture statistical features used to characterize stain covering or attachments.

[0109] The surrounding reference area can be a non-abnormal area within the same component or a region at the same location of adjacent components.

[0110] In this embodiment, the operating electrical parameters include at least one of current and voltage, and the environmental parameters include at least one of ambient temperature, wind speed or irradiance, and are provided by the site monitoring system or sensor system.

[0111] The operating electrical parameters are preferably obtained from monitoring data of the string, combiner box or inverter side of the same electrical branch as the surface area of ​​the target component, including at least one of current or voltage; the environmental parameters are preferably obtained from the field meteorological station or local sensors, including at least one of ambient temperature, wind speed or irradiance.

[0112] To ensure the correspondence between model calculation and image acquisition, electrical and environmental parameters should have timestamps and be aligned with the acquisition time of the image sequence. When the sampling period is longer than the image acquisition interval, nearest neighbor matching or time window averaging can be used to obtain the parameter values ​​at the corresponding time.

[0113] In this embodiment, the consistency score is obtained by fusing the first verification factor and the second verification factor; wherein the first verification factor represents the degree of matching between the temperature rise characteristics and the theoretical heat distribution, and the second verification factor represents the degree of matching between the appearance morphology characteristics and the pre-stored defect morphology characteristic library.

[0114] The defect morphology feature library contains feature templates or feature distribution parameters for at least several types of typical defects. Defect categories may include: obstructions / attachments, heavy dust accumulation, bird droppings, broken glass, or abnormal borders, etc. The features corresponding to each category can be extracted offline from historical inspection samples and added to the library.

[0115] Matching criteria can be based on similarity comparison or classifier output: when the appearance morphology feature has a similarity to a template of a certain defect category that is higher than a threshold, a corresponding second verification factor is generated; when multiple categories are satisfied at the same time, the highest one is selected by sorting by similarity or confidence, or multiple category candidates and their confidence are output for subsequent fusion.

[0116] In one implementation, the consistency score calculation defines the first verification factor as thermal distribution consistency and the second verification factor as appearance consistency, and then merges the two into an interval. The internal consistency score is used to drive fault diagnosis status output and suspected repeat shooting triggers; specifically:

[0117] When stable feature mask As described above, and with cross-modal alignment completed, robust convergence of multiple frames of infrared temperature within the mask is performed to form an observational thermal map:

[0118] ,

[0119] in, Indicates pixel position The observed temperature convergence value, Indicates time index The operator for taking the median. Indicates the first Frame infrared image at pixel position Temperature value, Indicates the number of infrared frames used for convergence and satisfies , Indicates and Pixel positions in the same coordinate system;

[0120] To reduce the impact of ambient temperature and overall bias on similarity, the observed heatmap and theoretical heat distribution are converted into temperature rise maps relative to the background; the observed background temperature uses a robust baseline within the mask.

[0121] ,

[0122] in, Indicates the baseline of the observed background temperature. Indicates the area Pixel position within, This represents the set of pixels representing the surface region to be diagnosed. Indicates the position of the observed temperature rise map at the pixel location. The value, Represents a stable feature mask;

[0123] The electrothermal coupling model outputs the theoretical heat distribution. Then, the theoretical temperature rise diagram was obtained using the same baseline format:

[0124] ,

[0125] in, Indicates the theoretical background temperature baseline. Indicates the theoretical temperature rise map at the pixel position The value, This indicates the theoretical heat distribution at the pixel location. The temperature value is given; the meanings of the other parameters are as defined above.

[0126] This implementation also includes: defining a first verification factor, and ensuring thermal distribution consistency. :

[0127] Under mask constraints, normalized cross-correlation is used to characterize the overall shape similarity, and the cross-correlation results are mapped to... :

[0128] ,

[0129] in, This represents the normalized cross-correlation value under mask constraints. Indicates in the region Seeking sum within, The interval mapping result representing the cross-correlation similarity. This represents a stable term to prevent the denominator from being zero; the meanings of the other parameters are as defined above.

[0130] When there are significant differences in local structures, structural similarity can be used to supplement the consistency of brightness and contrast.

[0131] ,

[0132] in, Indicates structural similarity. express exist The mean within, express exist The mean within, express exist within variance, express exist within variance, express and exist Covariance within, and A stability constant representing structural similarity;

[0133] When it is necessary to incorporate the morphology of hot spot connected regions into the consistency, the temperature rise map is binarized and the intersection-union ratio is calculated to reflect the morphological consistency of hot spot location and area:

[0134] ,

[0135] in, The binarized threshold representing the theoretical temperature rise. The binarization threshold representing the observed temperature rise. Indicates in the region Internal quantile Operators for taking quantiles, Indicates the quantile sites used for hot spot extraction;

[0136] ,

[0137] in, This represents a binary image of a theoretical hot spot. This represents a binary image of the observed hotspot. This indicates an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. The meanings of the other parameters are as defined above.

[0138] ,

[0139] in, Indicates the similarity of hot spot crossover. This represents a stable term to prevent the denominator from being zero; the meanings of the other parameters are as defined above.

[0140] The first verification factor is obtained by fusing the above multiple thermal consistency measures:

[0141] ,

[0142] in, Indicates the first check factor. The fusion weights represent the cross-correlation similarity. The fusion weights represent the structural similarity. The fusion weights represent the similarity scores of the intersection-union ratio;

[0143] ,

[0144] The above formula is used to limit , , The values ​​of the parameters are constrained, and the meanings of the other parameters are as defined above.

[0145] This implementation method further includes: defining a second verification factor, and ensuring consistency in appearance. :

[0146] The appearance morphology features (bright spot morphology, crack candidate texture, stain occlusion candidate texture, etc.) within the stable feature mask are organized into appearance feature vectors and similarity matching is performed with a pre-stored defect morphology feature library; let the appearance feature vector be... Defect database The class prototype vector is Based on cosine similarity, the first... Class matching degree:

[0147] ,

[0148] in, Indicates the first Cosine similarity, Indicates mapping to cosine similarity, This represents the appearance feature vector of the region to be tested. Indicates the first in the defect library The characteristic prototype vector of a class, Represents the vector dot product. Represents the L2 norm, This represents a stable term to prevent the denominator from being zero. Indicates the defect database index;

[0149] When the defect library performs statistical modeling of the intra-class distribution, Mahalanobis distance is introduced and converted into similarity to suppress mismatches where the directions are similar but the amplitudes are abnormal.

[0150] ,

[0151] in, Indicates the first Mahalanobis-like distance Indicates transpose. Indicates the first in the defect library The covariance matrix of appearance features of a class Denotes its inverse matrix. This represents the similarity obtained from the Mahalanobis distance mapping. This represents the attenuation coefficient of the distance mapping;

[0152] The first result is obtained by fusing the two appearance matching degrees. The overall matching score is calculated, and the maximum value is taken as the second verification factor.

[0153] ,

[0154] in, Indicates the first Overall matching degree of class appearance The fusion weights represent the cosine similarity. Indicates the second check factor. This represents the maximum value operator. This indicates the number of categories in the defect morphology feature library.

[0155] In this implementation, a consistency score is obtained by further fusion. And the suspected interval boundary:

[0156] When the desired consistency score is output in the form of confidence level, logistic regression fusion with interaction terms is used to map the two validation factors to... :

[0157] ,

[0158] in, This represents the fused linear response value. Indicates consistency score, Indicates the bias term. This represents the weight coefficient of the first verification factor. This represents the weighting coefficient of the second check factor. This represents the weight coefficient of the interaction item. Represents an exponential function;

[0159] The suspected interval is divided using two thresholds: a high consistency score indicates a fault, while a low consistency score indicates a non-fault; a reshoot is triggered in the middle interval. The interval boundaries can be determined by the ROC statistics of historical labeled samples. Let the historical sample index be... The sample label is The sample score is For any threshold Define the false positive rate and the true positive rate:

[0160] ,

[0161] ,

[0162] in, Indicates the threshold is The false positive rate at that time Indicates the threshold is True positive rate at that time Indicates the number of historical labeled samples. Indicates the first Consistency score of each sample Indicates the first The label of each sample Indicates a faulty sample. Indicates a non-faulty sample. and This indicates the stable term; the meanings of the remaining parameters are as defined above.

[0163] Given the upper bound of the target false positive rate With target missed detection constraint At that time, the upper and lower bounds are determined according to the set selection method, so that the suspected interval covers statistically uncertain samples:

[0164] ,

[0165] in, Indicates the suspected upper bound of the interval. Indicates the suspected lower bound of the interval. This indicates the upper bound of the false positive rate. This indicates the upper bound of the missed detection constraint. Indicates the threshold is The false negative rate function is defined as the proportion of faulty samples that are judged as non-faulty.

[0166] The false negative rate function is expressed using the same counting method as the above formula:

[0167] ,

[0168] in, Indicates the threshold is The false negative rate at that time This indicates the stable term; the meanings of the remaining parameters are as defined above.

[0169] when When this happens, the area is classified as a suspected zone and a reshooting strategy is triggered. Output fault diagnosis status in real time, when Outputs non-fault diagnosis status at times.

[0170] Specifically, the consistency scoring in the above implementation places thermal distribution matching and appearance matching within the same scoring framework; thermal distribution verification robustly converges the infrared multi-frame temperatures within a stable feature mask, eliminates the overall bias with the background baseline, and then uses cross-correlation and structural similarity to characterize the overall and local consistency of the temperature rise map, while introducing hot spot crossover ratio to reflect the morphological consistency of the location and area of ​​connected regions; appearance verification organizes appearance morphologies such as bright spots, cracks, and occlusions into vectors, calculates similarity with the prototype in the defect library, uses cosine similarity to reflect directional consistency, and uses Mahalanobis distance to constrain the statistical range within the class, and takes the maximum matching degree as the appearance verification factor after the two are fused; in the fusion stage, logistic regression with interaction terms is used to map the two factors into a confidence score, so that the misjudgment when one factor is high but the other factor is low is suppressed; the upper and lower bounds of the suspected interval are selected by the statistical curve of historical labeled samples, and the boundary selection is controlled by the false positive rate and the missed detection constraint, thereby concentrating uncertain samples in the suspected interval to trigger re-shooting.

[0171] Example 2

[0172] Based on Example 1, the analysis server is further configured to: when the consistency score is in the suspected range, issue a reshoot command to the drone, so that the drone changes at least one of the shooting angle, exposure or focal length of the suspected area to obtain supplementary images.

[0173] The conditions for triggering a repeat shot include at least the following: the consistency score is in the suspected range, the dynamic noise mask coverage exceeds the preset ratio, or the registration error exceeds the allowable range, causing the region to be unstable.

[0174] The repeat shot instruction may include: spatial indication information of the target area (component number / area coordinates / relative waypoints), recommended shooting distance and field of view, exposure parameter or focal length setting suggestions, and requirements for the number of repeat shot frames and time window. After the repeat shot is completed, the system will merge the supplementary image with the original sequence, prioritize the use of frames with higher quality and lower dynamic noise coverage for consistency score updates, and output the updated diagnostic status.

[0175] In this embodiment, the repeat shooting instruction also includes changing the polarization filter conditions or changing the incident angle range to reduce the high brightness interference caused by specular reflection.

[0176] When the drone payload supports a polarizing filter, the high brightness of the specular reflection can be reduced by switching the polarization direction or enabling / disabling the filter. When the payload does not support a polarizing filter, the reflection can be reduced by changing the incident angle range: while ensuring resolution, adjust the heading or gimbal pitch angle to make the specular reflection direction of the component deviate from the camera's line of sight, thereby reducing the proportion of the high-brightness saturation area.

[0177] The system can record the dynamic noise mask coverage under different viewing angles / exposure conditions, and use the acquisition conditions with lower coverage as the preferred reshooting strategy.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0179] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A smart photovoltaic module fault diagnosis system for power plants based on unmanned aerial vehicle (UAV) inspection, characterized in that, include: The drone is equipped with an infrared thermal imager and a visible light camera. Communication unit; The analysis server receives infrared image data and visible light image data using the communication unit. The analysis server is configured to register multiple frames of infrared images and multiple frames of visible light images obtained at consecutive time points on the surface area of ​​the same photovoltaic module and construct an image sequence. Pixel-level temporal dynamics indices are calculated based on the image sequence, and dynamic noise masks and stable feature masks are generated. Temperature rise features and appearance features are extracted within a stable feature mask; an electrothermal coupling model is run based on the electrical parameters of the electrical branch corresponding to the surface region and the environmental parameters to obtain the theoretical heat distribution, and a consistency score is given for the theoretical heat distribution, the temperature rise features and the appearance features; the fault diagnosis status of the surface region is output based on the consistency score.

2. The intelligent power station photovoltaic module fault diagnosis system based on UAV inspection according to claim 1, characterized in that, The image sequence contains at least 3 frames, and the acquisition time interval between two adjacent frames is less than a preset time window, so that dynamic light and shadow interference such as reflections and shadows can be observed to be time-varying in the sequence.

3. The intelligent photovoltaic module fault diagnosis system for power stations based on UAV inspection according to claim 1, characterized in that, The pixel-level temporal dynamics index includes at least one of the mean, variance, or weighted combination of the intensity differences between corresponding pixels in adjacent frames after registration.

4. The intelligent power station photovoltaic module fault diagnosis system based on UAV inspection according to claim 1, characterized in that, The stable feature mask is obtained by comparing the pixel-level temporal dynamic index with a threshold, which is adaptively determined based on the statistical distribution of the temporal dynamic index within the surface area of ​​the photovoltaic module.

5. A smart power station photovoltaic module fault diagnosis system based on UAV inspection as described in claim 1 or 4, characterized in that, The registration includes: calculating inter-frame geometric transformation based on feature point matching or optical flow estimation, and performing cross-modal alignment between infrared and visible light images.

6. The intelligent photovoltaic module fault diagnosis system for power stations based on UAV inspection according to claim 1, characterized in that, The temperature rise characteristic includes at least one of the following: the temperature rise amplitude relative to the surrounding background, the area of ​​the temperature rise connected domain, or the temperature gradient. The appearance features include at least one of bright spot morphology, edge / crack candidate texture, or stain occlusion candidate texture.

7. The intelligent power station photovoltaic module fault diagnosis system based on UAV inspection according to claim 1, characterized in that, The operating electrical parameters include at least one of current and voltage, and the environmental parameters include at least one of ambient temperature, wind speed, or irradiance, and are provided by the site monitoring system or sensor system.

8. A smart photovoltaic module fault diagnosis system for power stations based on UAV inspection according to claim 1, characterized in that, The consistency score is obtained by fusing a first verification factor and a second verification factor; wherein the first verification factor represents the degree of matching between the temperature rise feature and the theoretical heat distribution, and the second verification factor represents the degree of matching between the appearance morphology feature and the pre-stored defect morphology feature library.

9. A smart power station photovoltaic module fault diagnosis system based on UAV inspection according to claim 1, characterized in that, The analysis server is also configured to: when the consistency score is in the suspected range, issue a reshoot command to the drone, causing the drone to change at least one of the shooting angle, exposure, or focal length of the suspected area to obtain supplementary images.

10. A smart power station photovoltaic module fault diagnosis system based on UAV inspection according to claim 9, characterized in that, The repeat shooting command also includes changing the polarization filter conditions or changing the incident angle range to reduce the high brightness interference caused by specular reflection.

Images