Floating photovoltaic fault detection method based on irradiance-posture dual normalization and multi-index gating fusion

Abstract

This invention belongs to the field of intelligent operation and maintenance and fault diagnosis technology for photovoltaic power generation. Specifically, it provides a floating photovoltaic fault detection method that integrates irradiance-attitude dual normalization and multi-index gating. This method achieves fault and evidence mapping through scene modeling, data consistency through multi-modal calibration and registration, and full-time symptom capture through three-window synchronous acquisition. It combines irradiance-attitude dual normalization expected power modeling to remove interference and temperature-polarization fusion to generate composite evidence. After extracting multi-dimensional features, it performs two-level detection—edge screening and center ST-GNN deep inference—supplemented by topology and time consistency verification, to achieve accurate identification and graded alarms for module / substring level faults. This invention effectively suppresses surface interference and false positives, improves detection accuracy, adapts to complex water conditions, and automates the entire process, reducing operation and maintenance costs and safety risks, providing core support for intelligent operation and maintenance of floating photovoltaic systems.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance and fault diagnosis technology of photovoltaic power generation. Specifically, it relates to a floating photovoltaic fault detection method that integrates irradiation-attitude dual normalization and multi-index gating. Background Technology

[0002] Floating photovoltaic (FPV) power plants have developed rapidly due to their advantages such as open water and abundant sunshine. However, the special characteristics of the aquatic environment lead to many technical bottlenecks in fault detection, and existing technologies have significant shortcomings: 1. Strong water surface reflection and glare interference: Traditional detection solutions are designed for ground / roof power stations and lack means to suppress water surface reflection and polarized glare, which can easily lead to image saturation, false hot spots, and a significant decrease in detection accuracy.

[0003] 2. The array's attitude fluctuations with the waves are not modeled: The existing solution does not characterize the frequency domain coupling relationship between power and attitude when the photovoltaic array pitches / rolls with the waves, and mistakenly classifies attitude fluctuations as faults, resulting in a high false detection rate.

[0004] 3. Irradiance misjudgment and lack of dynamic correction of water surface albedo: The ordinary array irradiance (POA) transpose model does not consider the influence of water surface specular brightness changes with solar altitude and wave surface roughness, and lacks a mechanism to feed back imaging glare cues to irradiance estimation, resulting in inaccurate expected power baseline.

[0005] 4. Lack of a power model that combines irradiation and attitude: Most online monitoring only performs single-factor corrections for temperature or irradiation, without incorporating attitude / incident angle and incident angle correction (IAM) into the real-time normalization. The residuals contain attitude noise, making them difficult to use as stable evidence of faults.

[0006] 5. The wet-induced / condensation effect is not independently measured and gated: Existing methods lack specific indicators (such as Voc-G hysteresis and decreased polarization) for reversible mismatches such as wet-induced PID and condensation film in the early morning / evening, which are often confused with structural defects.

[0007] 6. Lack or inconsistency in thermal imaging radiation calibration and emissivity / reflection correction processes: Industrial sites often use "empirical emissivity + simplified temperature measurement", and there is a lack of a unified two-point field calibration, reflection temperature and non-uniformity correction (NUC) strategy between different equipment / time periods, resulting in incomparable temperature differences and difficulty in reproducing across days.

[0008] 7. Multimodal registration and time synchronization are not strict: Common implementations are mainly based on image-level or coarse registration, lacking pixel-level extrinsic / homography constraints and cross-device unified time synchronization, which can easily lead to misalignment evidence under FPV micro-motion / jitter.

[0009] 8. Relying solely on temperature difference or single image features, lacking composite evidence: Traditional thresholds / detectors mostly focus on hotspot intensity and area, rarely combining composite patterns such as "edge cold zone geometry," "low polarization degree (DoLP) water film indications," and "thermal-polarization co-occurrence overlap rate," resulting in insufficient type differentiation and positioning accuracy.

[0010] 9. Lack of an interpretable gating index system: Existing AI detection lacks gating and weighting of physical quality such as “wave-induced coherence (CI) / hysteresis (HI) / leakage current humidity sensitivity (LI),” and the model is not stable enough in extreme weather / strong glare / high humidity scenarios.

[0011] 10. Neglecting the consistency constraints of electrical / spatial / structural topology: Based mainly on single-frame or single-node judgment, it lacks topology consistency verification and spatiotemporal persistence logic for the same string / neighborhood / same float, making it difficult to suppress occasional false alarms from group patterns.

[0012] 11. Lack of alarm classification and retest orchestration: Traditional processes often report faults "instantly" and rarely introduce consistency and security upgrade strategies for the same window / cross window / cross day. There is also a lack of automatic retesting and evidence package solidification, resulting in low efficiency of operation and maintenance closed loop. Summary of the Invention

[0013] The technical problem to be solved by the present invention is to provide a floating photovoltaic fault detection method that integrates irradiation-attitude dual normalization and multi-index gating, so as to solve the problems of inaccurate fault detection, high false alarm rate and poor interpretability of the prior art under complex working conditions such as strong water surface reflection, attitude fluctuation and high humidity.

[0014] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a floating photovoltaic fault detection method integrating irradiation-attitude dual normalization and multi-index gating, comprising the following steps: S1. Scene Modeling and Detection Object Conventions: Establish a world coordinate system with the power plant's overall control coordinates as the origin. Register the polygonal boundaries of photovoltaic modules and sub-string partitions, associate the module geometry, installation normal and electrical topology, define the source and semantics of input signals, and clarify the defect types and corresponding observable evidence; S2, Multimodal Calibration, Radiometric Correction and Coordinate-Time Unification: Complete the intrinsic parameter calibration of the visible polarization camera and the long-wave infrared camera LWIR, as well as the cross-modal extrinsic parameter calibration of both, and establish a unified world / array coordinate system and a high-precision time synchronization reference. S3. Pixel Attribution Mapping and Array Topology Modeling: Establish a stable attribution mapping of pixel to component / substring ID and an array-level integrated electrical-geometry-structure topology model, and construct an array topology graph with components / substrings as nodes, including electrical connection edges, spatial proximity edges and structural coupling edges. S4. Three-window acquisition and time synchronization / triggering synchronization standard operating procedure (SOP): The three-window timed acquisition strategy of early morning, noon and evening is adopted. The unified time synchronization of the camera, inertial measurement unit (IMU), electrical parameters and meteorological acquisition equipment is realized through the PTP protocol. The window quality score (DQS) is calculated in real time, and the candidate region of interest (ROI) is automatically scheduled for retesting. S5. Water Surface Environment Suppression Preprocessing and Temperature-Polarization Fusion: Perform imaging preprocessing and cross-modal fusion for water surface scenes to form a temperature-polarization fusion map (TPC). S6. Irradiation-Attitude Dual Normalization Desired Power Modeling and Residual Calculation: Calculate the incident angle, correct it with IAM to obtain the effective irradiation, estimate the battery temperature through Faimin or PVsyst style models, construct the desired power P*(t), and calculate the normalized residual. S7. Construct multimodal features and perform robust normalization: Extract thermal spatial features, polarization texture features, spectral coherence features, electrical and physical features, cross-modal co-occurrence features, and quality and consistency features, and perform MAD normalization. S8. Two-level detection and gating fusion: In the initial edge screening stage, qualified windows are selected through hard gating, and a lightweight classifier is used to output the edge confidence vector of each fault category. The candidate list is generated by fusing with regular fingerprints; in the central deep diagnosis stage, a spatiotemporal graph neural network ST-GNN is constructed, which incorporates electrical series, spatial proximity and structural coupling constraints, outputs the probability and severity of fault categories, and fuses edge and center confidence; S9, Location, Hierarchical Alarm and Consistency / Topology Verification: Fault location is completed based on pixel attribution mapping, spatial confidence weight is calculated through intersection-union ratio (IoU), multi-level alarms are set according to confidence and consistency, time, topology and gating consistency verification is implemented, and evidence package containing core evidence is generated; S10. Evaluation, Deployment and Self-calibration: A three-layer sample set is used for offline evaluation and online canary release. The parameters of the expected power model are updated by robust regression at night / weekly. The ST-GNN model is periodically fine-tuned. A four-dimensional version chain is established to manage the versions of data, features, models and strategies.

[0015] In a preferred embodiment, in step S1, the mapping relationship between the fault defect type and the corresponding observable evidence is as follows: wet-induced PID / condensation mismatch is indicated by the Voc-G hysteresis index HI as observable evidence; water ingress / corrosion in connectors or manifolds is indicated by localized... Co-occurrence of hotspots and polarization decrease and humidity-sensitive admittance fingerprint LI= As evidence, Y represents the low-voltage admittance. Relative humidity; microcracks / solder joint defects are evidenced by stable hot spots and temperature-power consistency residuals; water ingress in the package is evidenced by edge cold band geometry and Gradient along the border is used as evidence; biological attachment / contamination is used as evidence by DoLP and texture statistical changes; attitude occlusion / drift and strong water surface reflection are gating based on power-attitude amplitude squared coherence index CI and polarization glare indication; cable insulation dampness is used as evidence by multi-point micro-hot spots and LI.

[0016] In a preferred embodiment, in step S2, a high-contrast AprilTag or checkerboard planar target is used to perform geometric intrinsic parameter calibration of the visible polarization camera, and the intrinsic parameter matrix of the visible polarization camera is calculated. With distortion parameters The intrinsic parameters of the long-wave infrared (LWIR) camera were calibrated using a hot-cold hybrid target, and the geometric intrinsic parameters of the LWIR camera were calculated. With distortion The cross-modal extrinsic parameters are calculated using a hybrid target combining AprilTag and thermal labeling. First, the target surface in world coordinates is obtained using a visible polarization camera. pose in Then, the coordinates of the thermal marker pixels are extracted from the thermal image, and the rigid body transformation between the thermal / visible coordinate systems is calculated by combining least squares or robust optimization. The homography matrices of the visible polarization camera and the long-wave infrared camera to the array plane are derived respectively. .

[0017] In the preferred embodiment, when calibrating the intrinsic parameters of the long-wave infrared camera, the geometric intrinsic parameters of the long-wave infrared camera (LWIR) are calculated. With distortion Subsequently, a pixel-level non-uniformity correction library was established based on the two-point method. Make pixel readings according to Linear scaling, where, Each pixel Gain correction factor and bias correction factor, This is the raw reading for that pixel. The equivalent temperature reading is corrected for non-uniformity; then, the apparent emissivity of the component cover glass, backplate, and junction box is measured based on material parameters or small sample measurements. The parameters were entered into a parameter table, and the apparent temperature of the reflection was measured using the pleated aluminum foil method. The apparent emissivity of the backplate and junction box and apparent temperature of reflection The parameters are stored in the parameter table as radiation inversion parameters for subsequent temperature inversion corrections, under the near-field radiative transfer approximation. Temperature inversion corrections are performed to obtain comparable surface temperature fields.

[0018] In the preferred embodiment, the steps for calibrating the geometric intrinsic parameters of the visible polarization camera are as follows: 1) To achieve channel gain and bias consistency for polarization cameras under uniform integrating sphere or diffuse target illumination; 2) Assemble a high extinction ratio linear polarizer for fixed step rotation angle sampling, estimate the actual orientation deviation and efficiency matrix of the analyzer microarray, obtain the pixel-level correction coefficients from the Stokes parameter solution, and perform absolute alignment of the polarization angle AoP zero point under a known zero-degree reference. 3) Use a high-contrast AprilTag or checkerboard planar target to perform visible camera geometric intrinsic parameter calibration and calculate the intrinsic parameter matrix of the polarization camera. With distortion parameters .

[0019] In a preferred embodiment, in step S3, the pixel coordinates of each frame in the visible and thermal infrared channels are... respectively or Back projection to world / front coordinate system get and call the R-tree-based spatial index pair Perform point-polygon determination and assign it to a pre-defined polygon. The registered component polygons and their substring partition polygons are assigned a pixel assignment map LUT_ID in the form of a lookup table. The component / substring polygons are geometrically shrunken by 3–5 pixels. When generating LUT_ID, the specular glare mask GLARE_MASK and the thermal image NUC marker pixels are masked.

[0020] In a preferred embodiment, in step S3, an array topology is constructed using components / substrings as the smallest electrical units. Node set For the corresponding component or substring, additional static properties include the installation normal. Effective area Azimuth / tilt, pixel set index and ROI geometric description; additional dynamic attributes updated with the window include window quality score (DQS), cross-modal reprojection error (RMSE_align), glare coverage, temperature-polarization fundamental statistics and power residual statistics; edge set It consists of three subsets, one of which is the electrical connection edge. It records the electrical direction identifier and circuit ID, and secondly, the spatial proximity edge. It also records the Euclidean distance and relative orientation, and the third is the structural coupling edge. The bearing unit identifier and flexible coupling parameters are recorded.

[0021] In the preferred embodiment, the time division criteria for the three windows in step S4 are as follows: the morning window is from sunrise to 15° solar altitude angle, the noon window is the peak solar altitude angle ±2 hours, and the evening window is from 15° solar altitude angle to before sunset.

[0022] In the preferred embodiment, in step S4, the window quality score DQS is a weighted combination of image sharpness / registration RMSE, glare coverage, irradiance stability, IMU attitude variance, and wind / cloud / dew point risk.

[0023] In a preferred embodiment, step S5 involves performing the following imaging preprocessing operation for the water surface scene: 1) Thermal image preprocessing: Perform NUC correction on the raw thermal image, and call up the apparent emissivity of the component cover glass, backplate and junction box recorded in the parameter table in step S2. And combined with the apparent temperature of reflection measured by the pleated aluminum foil method in step S2 Radiation inversion was performed to obtain the quantitatively corrected surface temperature field. In the time domain, exponential smoothing or Kalman filtering is used to suppress thermal noise, and the temperature anomaly map is calculated with the median temperature of the same subarray as a reference. ; 2) Polarization image preprocessing: Calculate the linear polarization degree DoLP and polarization angle AoP based on the pixel-level Stokes solution matrix, and use the Fresnel properties of the mirror to establish a glare detection criterion and generate a glare mask GLARE_MASK; 3) De-jittering and fusion: After wave phase de-jittering and cross-modal registration driven by IMU, a weighted field is constructed to generate a temperature-polarization fusion map (TPC).

[0024] In a preferred embodiment, in step S5, wave phase jitter removal is achieved by bandpass filtering the pitch / roll signal measured by the IMU at 0.2–2 Hz; a reference attitude normal is constructed. Cross-modal registration employs homography plane correction combined with small-amplitude optical flow micro-alignment to ensure that the reprojection error RMSE_align ≤ 0.5 pixels.

[0025] In the preferred embodiment, in step S5, TPC is calculated using a multi-scale guided fusion formula: ; in, For normalization ; For normalization Texture, Water film / pollution index, This represents a feature-level fusion operator that concatenates and / or weights multiple features within the feature space; Temperature weighting; Polarization weights; To prevent the use of tiny positive constants with a denominator of zero; temperature weighting. Follow Monotonically increasing, polarization weight Follow (1) DoLP and the water film / contamination index SI increase monotonically for GLARE_MASK coverage area and local data quality score. Below the preset quality threshold DQS th The pixels, will Set to zero. DQS th Determined through statistical analysis of a healthy sample set.

[0026] In a preferred embodiment, the formula for calculating the incident angle in step S6 is: ; in: The solar vector; The instantaneous normal direction after de-jittering; .

[0027] In a preferred embodiment, step S6, the step of obtaining effective irradiation, is as follows: Direct measurement using array-type irradiation sensors on the irradiated side. If only horizontal / scattering is present, the Perez / Perez-Driesse model is used to obtain the array direct dispersion component by transposing the GHI / DHI / DNI model and then superimposing the water surface reflection term to obtain the following: ; in, DHI is the total horizontal irradiance, DNI is the horizontal diffuse irradiance, and DNI is the direct normal irradiance. The total irradiance of the array surface. The water surface reflection component of the array surface; For the component tilt angle, The equivalent albedo is the water surface albedo; the Perez / Perez-Driesse model is an anisotropic sky irradiance transpose model, used to calculate the direct and scattered components of the array based on GHI / DHI / DNI. ; in, For baseline albedo, The gain coefficient is determined through robust regression of healthy samples; The water surface glimmer index (GLI) is obtained by mapping the area percentage of GLARE_MASK. IAM is used to correct the angle of the direct component: ; in, Cover plate / IAM coefficient; Effective irradiation The calculation formula is: or .

[0028] In a preferred embodiment, the formula for calculating the battery temperature in step S6 is: ; in, For ambient temperature, For wind speed, These are the heat exchange parameters.

[0029] In the preferred embodiment, the formula for calculating the desired power P*(t) in step S6 is as follows: ; in: The system constant loss coefficient; The nominal area of ​​the component; Pollution / water film reduction factor; STC is the temperature coefficient; For baseline efficiency; For effective irradiation; Battery temperature; Low illumination correction factor; Baseline efficiency The calculation formula is: ; in, This is the rated power of STC; The formula for calculating the pollution / water film reduction factor is: ; in, This refers to the sensitivity parameter.

[0030] In the preferred embodiment, the formula for calculating the normalized residual in step S6 is: ; ; in, This represents the measured output power of the component / substring at time t.

[0031] In a preferred embodiment, in step S7, the thermal spatial features include Statistics, area and aspect ratio of hotspot connected regions, minimum distance of hotspots from the edge, bandwidth and length ratio of cold bands along the long side of the component, and the difference between the mean values ​​of the edge band and the center band; polarization texture features include DoLP / AoP gradient, glare coverage; spectral coherence characteristics including attitude coherence index (CI), main peak frequency, modulation depth; electrophysical characteristics including normalized residual. (t) window mean, low-voltage admittance and LI, temperature-power consistency residual and hysteresis index HI; cross-modal co-occurrence characteristics include The Jaccard overlap rate is calculated between the hotspot mask and the low DoLP mask for above-threshold segmentation; quality and consistency features include window quality score (DQS), cross-modal reprojection error (RMSE_align), GLI, and cross-window consistency.

[0032] In the preferred embodiment, in step S8, the triggering condition for hard gating is any of the following conditions: CI exceeds a preset threshold, glare coverage exceeds a threshold, irradiance variation coefficient or dG / dt exceeds a threshold, or dew point risk triggering. The window that triggers hard gating only caches evidence and does not make a judgment.

[0033] In the preferred embodiment, in step S8, the ST-GNN includes a message passing network and a time modeling unit. The loss layer is superimposed with L2 / Huber constraints on electrical string consistency, soft constraints on temperature-power monotonicity, and quality weighting terms. Multi-label binary cross-entropy or focal loss is used to optimize the fault category output, and smoothed L1 loss is used to optimize the severity regression.

[0034] In the preferred embodiment, in step S9, when When the observation threshold is reached and the event occurs within the same window but is not reproduced, an L1 observation event is generated and a retest within the same window is automatically scheduled; when When the suspected threshold is reached and there are two or more instances of exceeding the threshold within the same window, or when the evidence is highly consistent with fingerprint evidence, an L2 suspected event is generated and evidence slices are automatically packaged. When the same component / substring is reproduced in adjacent windows or adjacent days, or when it belongs to a security-related category and the severity reaches the preset security threshold, it is upgraded to L3 confirmed and a work order is triggered.

[0035] The floating photovoltaic fault detection method provided by this invention, which combines irradiation-attitude dual normalization and multi-index gating, has the following beneficial effects: 1. A dual-normalized expected power modeling framework of irradiation and attitude is proposed: the instantaneous incident angle and glass IAM correction, POA irradiation after Perez / Perez-Driesse transpose, and the wind speed-driven Faiman / PVsyst style temperature model are simultaneously incorporated into the same model to form a unified P*(t) baseline for the component output, so that attitude fluctuations and thermal effects are systematically subtracted at the computational level, providing a stable reference for residual criteria.

[0036] 2. Invent a dynamic correction mechanism for water surface glare index (GLI) → albedo: Utilizing the ratio of the area to the specular glare mask generated by polarization imaging, a GLI index that can be updated in real time is constructed and mapped to the equivalent albedo of the water surface. The additive / proportional correction term is directly fed back to the POA transpose, solving the problem of irradiation overestimation or misjudgment caused by Sunglint in FPV scenarios.

[0037] 3. A multimodal preprocessing link combining polarization deglare removal and thermal imaging joint gating is proposed: A specular glare mask GLARE_MASK is constructed on the acquisition side using DoLP / AoP to perform polarization deglare removal and thermal imaging joint gating on the thermal infrared image. Extraction and texture analysis implement pixel-level culling and weight reduction, realizing an evidence cleaning process of "suppressing glare first and then measuring temperature", which significantly reduces false hot spots / false cold zones caused by strong water surface reflection.

[0038] 4. Invent a wave phase de-jittering and attitude normalization method driven by IMU: The wave phase is estimated by using the pitch / roll signal in the 0, 2–2Hz bandpass, and combined with homography correction and small-amplitude optical flow fine adjustment to achieve spatiotemporal alignment of thermal / polarization dual channels for the same reference attitude; and the power-attitude coherence index CI is introduced as a hard gate to distinguish attitude-induced power fluctuations from real faults.

[0039] 5. Propose a hybrid target and fine-tuning mechanism for strict registration of thermal and visible cross-modal data: Design an integrated target combining AprilTag (visible) and heated thermal marker (thermal image). First, calculate the extrinsic parameters and then project the target using planar homography. In the field operation, use the thermal marker to perform sub-pixel affine fine-tuning to ensure that the reprojection RMSE of the two modes is ≤0.5px, supporting the calculation of pixel-level co-occurrence evidence.

[0040] 6. Invented pixel → component / substring ID stable mapping and boundary shrinkage strategy: Based on homography, the pixels are projected onto the array plane and point-polygon assignment is performed. Combined with 3-5px geometric shrinkage and glare mask, an ID lookup table that can be reused across days is generated, which ensures positioning accuracy and suppresses edge and gap artifacts.

[0041] 7. A method for constructing temperature-polarization fusion maps (TPC) is proposed: using... With (1) DoLP / Water Film-Contamination Index (SI) is the dominant factor. Dynamic weight fields are constructed on a pixel-by-pixel basis and multi-scale guided fusion is performed to generate a composite image representation that simultaneously highlights "thermal anomalies" and "water film / contamination textures", providing high-contrast input for subsequent recognition of geometric cold zones and connector anomalies.

[0042] 8. Inventing a dedicated family of FPV features and composite evidence metrics: Simultaneously calculating "hot spot intensity and geometry," "edge cold band width / length," and "edge-to-center temperature difference" at the component / substring granularity. "overlap (with AoP gradient)" "low-DoLP" and the hysteresis index HI (Voc-G loop) and leakage current moisture-sensitive fingerprint LI ( Fingerprint features across modalities and physical quantities constitute an interpretable and gated feature space.

[0043] 9. Propose an engineering SOP for three-window data acquisition, DQS quality scoring, and automatic retesting: data is acquired at different times in the morning, noon, and evening, PPS / PTP time synchronization is unified, and IMU-image cross-correlation is refined to ≤5ms. Window quality scores (DQS) are generated according to wind, cloud, dew point, and glare. When candidate triggers occur, same-window revisit and cross-window verification are automatically scheduled to form a traceable data closed loop.

[0044] 10. Invent a two-level detection architecture and physical prior constraint inference: The edge end uses CI / cloud / glare hard gating and lightweight models to complete high recall initial screening; the center end uses a spatiotemporal graph neural network (ST-GNN) to perform message passing and time modeling on the "component / substring" graph, and outputs multi-label fault probability and severity through prior regularization constraints such as electrical string consistency (KCL / KVL) and temperature-power monotonicity, which significantly reduces false alarms.

[0045] 11. Propose a decision-making mechanism with hierarchical alarm, consistency verification, and topology weighting: The total confidence level is fused with model probability, location IoU, evidence overlap, and quality weights. L1 / L2 / L3 thresholds are set and consistency across windows / windows / days is required to pass. At the same time, electrical / spatial / structural topology consistency is used to weight the confidence level, so as to achieve a robust flow and safe escalation of "observation-suspected-confirmed".

[0046] 12. Inventing gating and learning security strategies: In In self-calibration and central model training, samples with high CI, rapid cloud changes, strong glare, and dew point risk windows are explicitly removed. Robust loss and "golden day" slow updates are introduced to prevent reversible or low-quality effects from being "learned into the model" and to ensure stability and interpretability under long-term drift.

[0047] 13. Provide an interface for automatically generating evidence packages and auditable work orders: Automatically package each L2 / L3 alarm into a single package. Graph, DoLP graph, TPC overlay, Core evidence such as CI / HI / LI curves, registration RMSE, GLARE ratio, and location IoU is integrated with the operation and maintenance work order system to form an auditable and verifiable closed-loop operation and maintenance link.

[0048] 14. Provide the device / system implementation form and edge-center collaboration scheme: At the device level, a coaxially rigidly mounted RadiometricLWIR + polarization camera + IMU + meteorological / electrical parameter integrated node is adopted. The edge end performs preprocessing / feature / gating and network disconnection caching, while the center end performs spatiotemporal inference and work order orchestration. The entire link supports OTA, model grayscale and daily / weekly parameter self-calibration to ensure the feasibility and large-scale deployment of the method. Attached Figure Description

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0051] A method for detecting floating photovoltaic (FPV) faults by integrating irradiance-attitude dual normalization and multi-index gating is proposed, applicable to photovoltaic arrays (FPVs) deployed on floating bodies in water. This method is executed collaboratively by an edge computing unit and a central server, utilizing radiometrically calibrated long-wave infrared (LWIR, 8–14 μm) imaging, on-chip spectrophotometric polarization imaging (DoFP Polarization, outputting Stokes parameters I / Q / U to calculate linear polarization degree DoLP and polarization angle AoP), an inertial measurement unit (IMU), global satellite time synchronization (GNSS) and network precise time synchronization (PTP), and array irradiance / meteorological measurements (POA / GHI). RH, wind speed v, back panel temperature The inputs are cascade electrical parameters (V / I / P), and the data undergoes calibration and spatiotemporal registration, three-window acquisition and synchronization, multi-modal preprocessing for glare and jitter reduction, and a desired power model. The system employs several steps, including constructing a feature engineering approach for FPV (containing coherence index CI, hysteresis index HI, and humidity-sensitive admittance fingerprint LI), constructing two-level detection and gating fusion, and implementing localization and hierarchical alarms. This results in interpretable and traceable judgments for component / substring level faults (such as moisture-induced PID / condensation mismatch, connector water ingress / corrosion, microcracks / solder joints, encapsulation water ingress, biofouling / contamination, attitude occlusion / drift, cable dampness, etc.). The Temperature-Polarization Composite (TPC) serves as image-side composite evidence, the Spatio-Temporal Graph Neural Network (ST-GNN) acts as an inferrer for array-level spatio-temporal and topological consistency constraints, and the Data Quality Score (DQS) is used. (Score) is used for gating and weighting; the above English abbreviations are given in full Chinese upon their first appearance in this technical solution and are used as industry-standard terms. The method of this invention can also be implemented as a device and system: at the device level, it consists of a coaxially rigidly mounted LWIR camera and polarization camera, IMU, timing and meteorological / electrical parameter acquisition terminals, and an edge computing unit; at the system level, it consists of a central data access, feature bin, inference, and work order interface, which communicate via wired / cellular networks and support offline caching and recovery transmission.

[0052] like Figure 1 As shown, the method of the present invention will be described in detail below: S1. Scene Modeling and Detection Object Conventions: Establish a world coordinate system with the power plant's overall control coordinates as the origin. Register the polygonal boundaries of photovoltaic modules and sub-string partitions, associate the module geometry, installation normal and electrical topology, define the source and semantics of input signals, and clarify the defect types and corresponding observable evidence.

[0053] This invention first establishes a set of fault objects and an observable symptom database for the special operating conditions of floating photovoltaic systems. These encompass categories such as moisture-induced PID / condensation mismatch, water ingress or corrosion in connectors / junction boxes, microcracks / solder joint failures, water ingress into the encapsulation, biofouling / contamination, array attitude drift / obstruction, and moisture absorption in cable insulation. For each category, abnormal thermal infrared temperature (hot spots) is specified. Evidence includes edge cold band, polarization degree / polarization angle mode (DoLP / AoP) visible polarization, and electrical side residual / admittance fingerprint. At the same time, quality gating and prior indicators such as attitude coherence index CI, hysteresis index HI, and humidity-sensitive admittance fingerprint LI are given as a unified semantic entry point for subsequent algorithms and criteria.

[0054] At the device level, this invention employs a dual-modal imaging unit consisting of a coaxially rigidly mounted radiometrically calibrated LWIR thermal infrared camera and a polarized visible camera, equipped with an IMU (pitch / roll / heading), GNSS time synchronization and PTP time synchronization, irradiance / weather and backplane temperature sensing, and cascaded electrical parameter acquisition; it sets up thermal / visible dual reference markers (AprilTag and heating thermal marker) and reference micro-batteries; the edge computing unit is responsible for preprocessing, feature extraction and gating, while the central side performs deep diagnostics and maintenance work order interfaces; the cabinet / window has IP66 protection, defogging and lightning protection, and communication supports Ethernet / cellular and network outage caching.

[0055] In a preferred embodiment, this method first performs a unified modeling of the application scenarios, target defects, and observables of floating photovoltaic (FPV) arrays: In terms of static relationships, a world coordinate system W is established with the power plant's overall control coordinates as the origin, and each photovoltaic module and its substrings are partitioned by polygonal boundaries. The system registers components, associates their geometry (long side direction, frame and junction box position), installation normals, and electrical topology (series and bus relationships), and defines the mapping carrier of images and electrical parameters as "component / substring ID". Regarding data and device relationships, it specifies the source and semantics of input signals: Long-wave infrared imaging (LWIR, 8–14 μm) is used to acquire quantitative temperature fields and characterize hot / cold zones; on-chip spectrophotometric polarization imaging (DoFP) outputs Stokes parameters I / Q / U and calculates the degree of linear polarization (DoLP) and angle of polarization (AoP) to characterize specular glare, water film, and contamination textures; the inertial measurement unit (IMU) provides pitch / roll / heading and angular velocity for attitude stabilization and coherence analysis; and time synchronization is achieved by generating second pulses via Global System for Time (GNSS) and using the Precision Time Protocol (PTP). The protocol is distributed to cameras and acquisition terminals. Environmental / meteorological quantities include array or horizontal irradiance (POA / GHI) and ambient temperature. Relative humidity (RH) and wind speed (v), and backsheet temperature can be measured on the module side. Electrical quantities include cascade voltage / current / power (V / I / P) and optional low-voltage segment sampling of the I-V curve; regarding the target and evidence conventions, the defect types and signal fingerprints covered by this method are clearly defined, including wet-induced PID / condensation mismatch (with Voc-G hysteresis index HI as observable evidence), water ingress / corrosion in connectors or junction boxes (with localized... Co-occurrence of hotspots and polarization decrease and humidity-sensitive admittance fingerprint LI= As evidence, Y represents the low-voltage admittance. (Relative humidity), microcracks / poor solder joints (evidence of stable hot spots and temperature-power consistency residuals), water ingress into the package (evidence of edge cold band geometry and) Evidence includes gradient along the border, biofouling / contamination (evidence based on DoLP and texture statistical changes), attitude occlusion / drift and strong water surface reflection (gating based on power-attitude amplitude squared coherence index CI and polarization glare indication), and cable insulation dampness (evidence based on multi-point micro-hot spots and LI). In methodological terminology, the Temperature-Polarimetric Composite (TPC) is defined as a composite image representation after preprocessing and weighted fusion, with the expected power baseline... The normalized residual is the output of the stress power model obtained after dynamic albedo correction driven by irradiation-attitude dual normalization and incident angle correction (IAM) and the Glare Index (GLI). For subsequent criterion fusion, the Data Quality Score (DQS) is used to implement gating weighting for environmental factors such as wind, cloud, dew point risk, and glare. Through the unification of the above scenarios, objects, and symbols, this step establishes a consistent semantic and data interface for the image-geometric-electrical domains, laying the foundation for subsequent steps such as calibration and registration, three-window acquisition and time synchronization, preprocessing, and fusion. Modeling, feature engineering, detection, and hierarchical alerting provide reusable basic conventions and boundary conditions.

[0056] S2. Multimodal calibration, radiometric correction and coordinate-time unification: Complete the intrinsic parameter calibration of the visible polarization camera and the long-wave infrared camera (LWIR) and the cross-modal extrinsic parameter calibration of both, and establish a unified world / array coordinate system and a high-precision time synchronization reference.

[0057] This invention achieves LWIR intrinsic parameter calibration using a hybrid hot-cold target, and visible camera intrinsic and extrinsic parameters using AprilTag. It employs a hybrid target combining AprilTag and thermal markers to calculate cross-modal extrinsic parameters, enabling pixel-level geometric registration between thermal and visible targets. A unified world coordinate system W is established based on GNSS / IMU and integrated with the coordinate systems of the two cameras. Fixed association; using homography to map pixels to the array plane and perform point-polygon association with component / substring vector boundaries to generate a stable "pixel → component / substring ID" lookup table; in terms of time, unified time synchronization is used with PPS / PTP and the time difference is finely calibrated to ≤5ms through IMU-image cross-correlation.

[0058] In a preferred embodiment, after the device is mechanically fixed, the visible polarization camera, the long-wave infrared camera, and the cross-modal extrinsic parameter calibration between them are performed sequentially, and a unified world / array plane coordinate and high-precision timing reference are established: First, for the on-chip beam-of-focal-plane (DoFP), the inter-channel gain and bias are consistent under uniform integrating sphere or diffuse target illumination. Then, a high extinction ratio linear polarizer is assembled for fixed-step rotation angle sampling, the actual azimuth deviation and efficiency matrix of the analyzer microarray are estimated, and the pixel-level correction coefficients of the Stokes parameters (I / Q / U) are obtained. The zero point of the polarization angle AoP (Angle of Polarization) is absolutely aligned under a known zero-degree reference. Then, the visible camera's geometric intrinsic parameter calibration is performed using a high-contrast AprilTag or checkerboard planar target, and the intrinsic parameter matrix is ​​calculated. With distortion parameters For long-wave infrared cameras (LWIR, 8–14 μm), a "hot-cold hybrid target" (high-emissivity matte black stickers and low-emissivity metallic stickers or micro-heated markers are placed at known geometric points) is used to calculate geometric intrinsic parameters under multi-distance and multi-pose sampling. With distortion A pixel-level non-uniformity correction (NUC) library was established based on the two-point method. Make pixel readings according to Linear scaling, where, Each pixel Gain correction factor and bias correction factor, This is the raw reading for that pixel. The equivalent temperature reading is corrected for non-uniformity; subsequently, the apparent emissivity of the component cover glass, backplate, and junction box is measured based on material parameters or small sample measurements. The parameters were written into the parameter table, and the apparent temperature of the reflection was measured using the "wrinkled aluminum foil method". The apparent emissivity of the backplate and junction box and apparent temperature of reflection The parameters are stored in the parameter table as radiation inversion parameters for subsequent temperature inversion corrections, under the near-field radiative transfer approximation. Temperature inversion corrections are performed to obtain a comparable surface temperature field; in terms of cross-modal geometry, a coplanar hybrid target with "visible AprilTag + thermal tag" is set up, and the target surface in world coordinates is first obtained by the visible camera. pose in Then, the coordinates of the thermal marker pixels are extracted from the thermal image, and the rigid body transformation between the thermal / visible coordinate systems is calculated by combining least squares or robust optimization. And derive the homography matrix of the two cameras to the array plane. The acceptance threshold is set at a reprojection mean square error (RMSE) of no more than 0.5 pixels. In the time domain, a GNSS pulse-per-second (PPS) is used to drive the edge host to establish an IEEE-1588 Precision Time Protocol (PTP) master clock and distribute it to the camera and acquisition terminal. Hardware triggering is used to achieve simultaneous exposure of thermal images and polarization frames, and the brightness sequence of the visible channel and the IMU angular velocity are used as the basis for the exposure. Cross-correlation estimation of residual time difference of sequences ,Will Write it into the acquisition driver layer as time offset compensation to ensure multi-source time alignment. The parameters obtained from the above geometric / radiological / polarization and time calibration (including but not limited to) , , , , Stokes solution for the correction matrix, AoP zeros, , , , and The data is uniformly archived as “calib_version (calibration version number) + timestamp (time stamp)” and bound to the device serial number, serving as a fixed input and traceability basis for subsequent pixel mapping, three-window acquisition, temperature-polarization fusion and expected power modeling.

[0059] S3. Pixel Attribution Mapping and Array Topology Modeling: Establish a stable attribution mapping of pixel to component / substring ID and an array-level integrated electrical-geometric-structural topology model, and construct an array topology graph with components / substrings as nodes, including electrical connection edges, spatial proximity edges and structural coupling edges.

[0060] In a preferred embodiment, the homography matrix of the visible / thermal infrared to the frontal plane obtained in step two is used as the basis. , With cross-modal extrinsic parameters This method establishes a stable attribution mapping of "pixel → component / substring ID" and an array-level integrated electro-geo-structural topology model: In terms of static relationships, it integrates the pixel coordinates of each frame in the visible and thermal infrared channels. respectively or Back projection to world / front coordinate system get and call the R-tree-based spatial index pair Perform point-polygon determination and assign it to a pre-defined polygon. The registered component polygons and their substring partition polygons are assigned a pixel attribution map (LUT, Look-Up Table, hereinafter referred to as "LUT_ID") in the form of a lookup table. To avoid boundary and gap errors, it is preferable to perform an equivalent 3-5 pixel geometric indentation on the component / substring polygons (the pixel scale is calculated from the camera's in-camera view distance). When generating LUT_ID, specular glare mask (GLARE_MASK) and thermal image NUC marker pixels are masked so that these pixels are removed in subsequent statistical aggregation. Simultaneously, based on the component geometric template... The system parametrically defines regions of interest (ROIs) such as junction boxes, borders, and busbars, and records the long-side direction vector of the components and the bandwidth parameters of the edge band region to support subsequent direction-sensitive morphological operations and edge-to-center temperature difference measurements. In terms of dynamic relationships, the array topology is constructed with "components / substrings" as the smallest electrical unit. Node set For the corresponding component or substring, additional static properties include the installation normal. Effective area Azimuth / tilt, pixel set index and ROI geometric description, with additional dynamic attributes updated with the window including window quality score (DQS), cross-modal reprojection error (RMSE_align), glare coverage, temperature-polarization basic statistics and power residual statistics, etc.; edge set It consists of three subsets, one of which is the electrical connection edge. (Connected in series with the same type of junction box, or in parallel with the same type of junction box) and record the electrical direction identifier and circuit ID; the second is the spatial proximity edge. (Generated according to K-nearest neighbor or radius threshold) and record Euclidean distance and relative orientation, the third being the structural coupling edge. (Same buoy / same row / same raft) and record the bearing unit identifier and flexible coupling parameters; the node and edge features are synchronously written into the feature warehouse with timestamps, and the mapping and topology version number is saved as "mapping_version" so as to trigger local reconstruction and tracing after camera fine-tuning or array modification; during operation, the edge computing unit aggregates the pixel-level temperature field and polarization field into component / substring-level statistics at the frame level according to LUT_ID and aligns them with the synchronously acquired electrical parameters and writes them into the node attributes, and the center side is based on the topology map The implementation of spatiotemporal and topological consistency inference serves several purposes. First, it ensures that pixel evidence from thermal infrared and polarization images is definitively and reusably mapped to electrical and geometric entities. Second, it guarantees strict comparability of the same component / substring across frames, windows, and days. Third, it lays the foundation for subsequent three-window acquisition control, temperature-polarization fusion (TPC) generation, and expected power... Residual calculation, two-level detection, and hierarchical alarm provide reliable data interfaces and topological constraints.

[0061] S4. Three-Window Acquisition and Time / Trigger Synchronization Standard Operating Procedure (SOP): Adopts a three-window timed acquisition strategy of early morning, noon and evening, and realizes unified time synchronization of camera, inertial measurement unit (IMU), electrical parameters and meteorological acquisition equipment through PTP protocol, calculates window quality score (DQS) in real time, and automatically schedules retesting of candidate regions of interest (ROI).

[0062] This invention employs a three-window acquisition system: "early morning, noon, and evening." Within each window, thermal / polarization frames are synchronously triggered at 1–2 Hz, electrical and meteorological data are acquired at 1–2 Hz, and IMU data is acquired at 100 Hz with downsampling and alignment. Data quality gating for wind, clouds, dew point, and glare is set, and the window quality score (DQS) is calculated. The entire array is covered by multiple preset positions, ensuring >95% component coverage. ROIs that trigger candidates are revisited within the same window and retested across windows to achieve automatic reshooting and quality closure.

[0063] In a preferred embodiment, to balance the observability of the wetted effect with the irradiance angle distribution, this method employs a three-window timing acquisition strategy of "early morning—noon—evening" and a unified time synchronization and hard triggering process: First, the time axis is segmented according to the solar altitude angle. Preferably, the early morning window is defined as from sunrise to approximately 15° of solar altitude angle, the noon window is the peak of the altitude angle ± 2 hours, and the evening window is the time when the altitude angle drops from 15° to before sunset. The edge host executes a preset scanning path and multiple preset position switching within each window to achieve array coverage. In terms of synchronization, the GNSS pulse of second (PPS) drives the edge host to establish the IEEE-1588 Precision Time Protocol (PTP) master clock and distributes a unified time base to the camera and acquisition end. Simultaneously, the host outputs a hardware trigger signal to make the long-wave infrared and polarization camera expose in phase. The IMU continuously samples at ≥100Hz and downsamples and aligns on the host side according to the trigger time. Cascaded electrical parameters (V / I / P) and meteorological parameters (POA or GHI) are synchronized. RH, wind speed v, back panel temperature Data is acquired at 1–2 Hz and aligned to the same timestamp; for data quality and action gating, polarization zero-point fast detection and thermal shutter non-uniformity correction (NUC) are performed at the beginning of each window. A few seconds after the NUC / zero-point completion, the frame is marked as a "dead zone" and not included in the evaluation. During operation, the window quality score (DQS) is calculated in real time, which is composed of image sharpness / registration RMSE, glare coverage (GLARE_MASK ratio), and irradiance stability (…). In this embodiment, the weighted combination of the coefficient of variation, IMU attitude variance, and wind / cloud / dew point risk can be used for calculation: ; in, w 1. w 2.w 3. w 4. w 5 represents the weighting coefficient, obtained through robust regression on a healthy sample set; S sharp The image quality normalization index is obtained by combining image sharpness and registration RMSE. The larger the value, the clearer the image and the smaller the registration error. G glare，ratio This is a normalized value for glare coverage; a higher value indicates more severe glare. S irradiance The normalized value for irradiation stability is derived from the irradiation coefficient of variation and The larger the value obtained through function mapping, the more stable the irradiation. R is the normalized value of the IMU attitude variance; a larger value indicates more severe array attitude jitter. risk This represents the normalized value for wind / cloud / dew point risk levels; a higher value indicates a higher environmental risk. Using this method, when the image is clear, registration is accurate, glare coverage is low, irradiance is stable, and attitude / environmental disturbances are minimal, the window quality score (DQS) achieves a higher value, which is then used for subsequent gating and weighting.

[0064] A hard-gating threshold is set to pause the judgment process and only cache evidence under conditions of strong glare, strong wind, or rapid cloud shadows. Regarding the relationship between coverage and retesting, this method specifies the target component coverage and pixel incident angle distribution for each window, preferably with >95% of components obtaining at least one qualified frame in that window as the standard. Candidate ROIs triggered by edge screening are automatically scheduled for retesting at the end of the same window and adjacent windows to form consistent evidence of "same window - cross window". In terms of data structure and disk storage, a unified "frame recording unit" is generated for each exposure, including frame_id, timestamp_ns, camera-inside and outside homography version numbers (calib_version, mapping_version), NUC / zero point / trigger status, and IMU attitude. With angular velocity, POA / GHI / / RH / v / The electrical parameters V / I / P, preset position index and lens focal length (if applicable), and device health fields are all recorded and sent to the central end via a ring buffer and network disconnection retransmission mechanism. Through the coordination of the above three-window time period division, unified time synchronization and hard trigger synchronization, quality gating and automatic retesting arrangement, this step ensures the comparability of thermal infrared and polarization images, electrical parameters and meteorological quantities in the time and attitude dimensions, providing a high-quality and traceable raw data foundation for the subsequent implementation of water surface environment suppression preprocessing and temperature-polarization fusion (TPC generation) in step five.

[0065] S5. Water Surface Environment Suppression Preprocessing and Temperature-Polarization Fusion: Perform imaging preprocessing and cross-modal fusion for water surface scenes to form a temperature-polarization fused map (TPC).

[0066] This invention performs defect and stripe repair on thermal images, pixel-level two-point field calibration and shutter NUC, emissivity and reflection temperature correction, and edge fidelity denoising; it calculates DoLP / AoP for the polarization channel based on Stokes estimation and generates a specular glare mask GLARE_MASK; it uses an IMU for wave phase stabilization and homography correction, followed by cross-modal sub-pixel registration; and it further applies temperature anomalies... A weighted field is constructed with the polarization / contamination index to form a temperature-polarization fusion map (TPC), and basic statistics (hot spots / cold zones / DoLP / AoP gradient / glare coverage / registration RMSE, etc.) are aggregated and output by component / substring.

[0067] In a preferred embodiment, this method performs imaging preprocessing and cross-modal fusion for water surface scenes based on the synchronized frame and multi-source data obtained in step four to form a temperature-polarization composite (TPC): First, non-uniformity correction (NUC) coefficients are applied to the original thermal image pixel by pixel in the long-wave infrared channel, and bad pixels / stripes are removed. Then, the apparent emissivity is set according to the component material list. And by reflecting apparent temperature (Measured using the pleated aluminum foil method) Radiation inversion was performed to obtain the quantitatively corrected surface temperature field. In the time domain, exponential smoothing or Kalman filtering is used to suppress thermal noise. In the spatial domain, guided filtering (guided by visible brightness or polarization intensity) is used to preserve the details of the junction box and frame. The temperature anomaly map is calculated with the median temperature of the same subarray as a reference. Subsequently, the linear polarization degree DoLP and polarization angle AoP are calculated in the polarization visible channel based on the pixel-level Stokes solution matrix. Sub-pixel affine registration is performed on the four angle frames acquired in time-division to eliminate angular misalignment. Glare detection criteria (high DoLP threshold, brightness quantile, consistency between AoP and water surface normal) are established using the specular Fresnel characteristics to generate a specular glare mask GLARE_MASK, which is used as a hard gate for subsequent statistics and fusion. Then, wave-induced attitude disturbances are filtered out using the pitch / roll measured by the IMU within the 0.2–2Hz bandpass filter to construct a reference attitude normal. The thermal image and polarization frame are respectively subjected to homography-based planar straightening, followed by micro-alignment with a small amount of optical flow to control the cross-modal reprojection error (RMSE_align) to be no greater than 0.5 pixels. After geometric unification, the thermal image is projected to the polarization coordinate system (or vice versa) based on the cross-modal extrinsic parameters obtained in step two, and low-quality pixels and frames are downweighted using DQS (Data Quality Score). To generate the fused representation, a pixel-level weight field constrained by physical meaning is constructed, in which temperature weights are used. Follow Monotonically increasing, polarization weight Follow (1) Increased DoLP and the water film / contamination index SI (measured jointly by ΔDoLP and defogging contrast) affect the data quality score of the GLARE_MASK coverage area and local data. Below the preset quality threshold DQS th The pixels, will Set to zero. DQS th The TPC map was obtained by statistical analysis of a healthy sample set and using multi-scale guided fusion.

[0068] ; in, For normalization , For normalization Texture, This represents a feature-level fusion operator that concatenates and / or weights multiple features within the feature space; Temperature weighting; Polarization weights; To prevent the use of tiny positive constants with a denominator of zero.

[0069] Direction-sensitive morphological operations are performed on the TPC to detect "cold bands" distributed along the long side of the component, and connected component analysis is used to extract geometric quantities such as the area, aspect ratio, and minimum distance to the border of hot spots / cold bands; finally, based on the pixel attribution mapping (LUT_ID) implemented in step three, pixel-level quantities are aggregated to the component / substring granularity to generate basic statistics (such as... The data includes hotspot area, cold band width / length ratio, DoLP mean and AoP gradient, GLARE ratio, RMSE_align, frame-level confidence, etc., and quality flags (GLARE, DEW_RISK, CLOUDY / WINDY). The TPC slices and metadata (NUC / zero-point state, fusion weights, torsion parameters, and version number) are also stored on disk as part of the irradiance-attitude dual-normalized expected power modeling in step six. This is the direct input for residual analysis and the implementation of FPV-specific feature engineering in step seven.

[0070] S6. Irradiation-Attitude Dual Normalization Desired Power Modeling and Residual Calculation: Calculate the incident angle, correct it with IAM to obtain the effective irradiation, estimate the battery temperature using Faiman or PVsyst style models, construct the desired power P*(t), and calculate the normalized residual.

[0071] This invention geometrically calculates the instantaneous incident angle and performs POA transpose (including Perez / Perez-Driesse) and IAM correction on the irradiance. Simultaneously, it designs the injection of the "GLI (derived from the glare mask)" into the water surface albedo. Dynamic correction; thermally, using Faiman or PVsyst style temperature models combined with wind speed to estimate battery temperature; and comprehensively obtaining the desired power. Low-quality samples were removed using cloud / glare / attitude gating, and normalized residuals were calculated. For use in subsequent feature analysis and determination.

[0072] In a preferred embodiment, to isolate the effects of attitude fluctuations, incident angle, and thermal coupling on power generation, this method establishes a baseline for the expected power output at the component / substring level based on the synchronization frame and statistics in step five and calculates the residuals: First, the solar vector is calculated based on the site's latitude and longitude and time synchronization information. The reference attitude normal obtained from step five Or the instantaneous normal after de-jittering Find the angle of incidence and restrictions Direct measurements using array-type irradiation sensors are preferred on the irradiated side. If only horizontal / scattering quantities are present, the Perez / Perez-Driesse model is used to obtain the array direct dispersion components from the transpose of GHI / DHI / DNI and then superimposed with the water surface reflection term. ,in For component tilt angle, Let G be the equivalent albedo of the water surface. To characterize the dynamic effect of sunglint on the albedo, this method introduces the Glare Index (GLI) output from the polarization deglare removal process (implementation step five) and sets... ( For baseline albedo, The gain coefficients are obtained through robust regression on healthy samples. Simultaneously, an incidence angle modifier (IAM) is used to correct the angle of the direct component, for example, using the ASHRAE method. ( (where I am the cover plate / IAM coefficient); thus, the effective irradiation is obtained. The temperature of the solar cells is estimated using a Faiman model that includes a wind speed term or a PVsyst-style model. ( For ambient temperature, For wind speed, (for heat exchange parameters), or when a backplate temperature is available. In the case of obtaining through first-order thermal resistance correction A more precise estimate; in efficiency mapping, let the nominal area of ​​the component be... STC rated power With temperature coefficient Baseline efficiency Introducing low-light correction (Logarithmic or quadratic term) and pollution / water film reduction factor ( To implement the pollution / water film index in step five, (as a sensitivity parameter), and the system constant loss coefficient. Summarizing the cable / inverter / metering deviations, the expected power output of the component / substring is defined as follows: ; in: The system constant loss coefficient; The nominal area of ​​the component; Pollution / water film reduction factor; STC is the temperature coefficient; For baseline efficiency; For effective irradiation; Battery temperature; The low-irradiance correction factor, obtained through calibration on a healthy sample set, is calculated as the normalized ratio of measured power to theoretical power as a function of effective irradiance, after removing faulty and low-quality windows. The relationship between the changes was obtained by robust regression fitting. A monotonically increasing normalized correction function; preferably, it can be in piecewise polynomial or logarithmic form, for example, when season ,when season ,in The threshold irradiance is corrected for low illumination to distinguish between low illumination range and normal irradiance range. The parameters a0, a1, and a2 are obtained by regression from the healthy sample set.

[0073] Baseline efficiency The calculation formula is: ; in, This is the rated power of STC; The formula for calculating the pollution / water film reduction factor is: ; in, This refers to the sensitivity parameter.

[0074] To avoid contaminating the baseline and criteria under strong attitude coupling or low-quality conditions, this method sets gating during the computation and learning phases: the "Power-Attitude Amplitude Squared Coherence Index" (CI, calculated from the MSC of power and pitch / roll sequences, Magnitude-Squared Coherence) exceeding a threshold, glare coverage exceeding a threshold (GLARE_MASK percentage), and rapid cloud changes (…) are all considered. Or the coefficient of variation of the peers exceeds the threshold) and the risk of dew point ( Close to 0 and A window showing a significant decrease, only output For recording purposes only, not for parameter updates or downstream alarms; residuals are defined as... , The measured output power of the component / substring at time t; and the normalized residual is calculated. The method also establishes robust statistics (mean, variance, quantiles) for both peer and cross-peer use for subsequent feature and judgment purposes. To adapt to different sites and seasons, this method employs robust regression (Huber / quantile loss) on the parameter set in the "healthy sample set" (DQS met, no candidate alarms, non-gated window). The low illumination coefficient is updated slowly at night / weekly, and exponential smoothing is used to achieve small step parameter evolution. Physically reasonable ranges and rate of change constraints are set for key parameters. All parameters, gating and statistics are archived with version numbers and timestamps and associated with “mapping_version” in implementation step three and “calib_version” in implementation step two to form a traceable and reproducible expected power baseline and residual sequence, providing stable and physically consistent input for FPV-specific feature engineering in implementation step seven and two-level detection in implementation step eight.

[0075] S7. Construct multimodal features and perform robust normalization: Extract thermal spatial features, polarization texture features, spectral coherence features, electrical and physical features, cross-modal co-occurrence features, and quality and consistency features, and perform MAD normalization.

[0076] This invention constructs a multi-layer feature space, including thermal space features ( Statistics, hotspot geometry, edge cold zone scale and edge-to-mid temperature difference), polarization texture features (DoLP / AoP gradient, glare coverage), spectral coherence characteristics (CI, main peak frequency, modulation depth), electrophysical characteristics (residual statistics, low-voltage admittance and LI, temperature-power consistency residual and HI), cross-modal co-occurrence characteristics (…). It uses features such as low-DoLP overlap rate, as well as quality and consistency features (DQS, registration RMSE, GLI, cross-window consistency), and employs MAD normalization and adaptive thresholding to obtain robust representations.

[0077] In a preferred embodiment, to transform the temperature-polarization fusion map (TPC) obtained in step five and the expected power baseline and residual sequence in step six into interpretable, gated, and reproducible criterion inputs, this method constructs multimodal features and performs robust normalization at the "component / substring × time window" granularity: First, based on the pixel attribution mapping (LUT_ID) and region of interest (ROI) geometry generated in step three, thermal image temperature anomalies of the same component / substring are processed by window (preferably 10–20 min). The thermal-spatial characteristics are calculated by combining the polarization parameter fields DoLP, AoP, and TPC intensity field, including temperature peaks and quantiles, variance and coefficient of variation, area and aspect ratio of hotspot connected regions and their minimum distance from the edge, bandwidth and length ratio of the "cold band" along the long side of the component, and the "edge-center" temperature difference (defined as the difference between the mean values ​​of the edge and center bands). Secondly, polarization-texture features are obtained using the visible polarization channel, calculating the mean, median, and quantile differences of DoLP, and the temperature relative to the "healthy baseline". The gradient norm of the polarization angle AoP (in terms of...) The window mean and variance measure the fragmentation texture of the water film / contamination, and the coverage of the specular glare mask GLARE_MASK is statistically analyzed as a quality and gating input. Furthermore, in the attitude-power coupling dimension, this method defines the attitude coherence index CI using the magnitude-squared coherence (MSC) of the power sequence and the IMU pitch / roll sequence in the 0.2–2 Hz frequency band, and extracts the main peak frequency and modulation depth of this band to characterize the strength and stability of wave-induced disturbances. In the electrical-physical dimension, based on the normalized residuals generated in step six... Calculate its window mean, variance, and upper and lower quantiles. If sampling in the I–V low-pressure range is available, the admittance Y can be estimated equivalently. This method performs robust linear regression of Y and relative humidity RH within a 24–48h sliding window. The regression slope is defined as the humidity-sensitive admittance fingerprint LI. This method is used to characterize water ingress / moisture absorption; simultaneously, a Voc-G hysteresis loop is constructed using samples with the same irradiance level at dawn / dusk, and the hysteresis index HI is defined after normalizing the loop area to the rated open-circuit voltage, used to measure reversible effects such as moisture-induced PID / condensation; furthermore, to characterize the spatial co-occurrence of thermopolarization evidence, this method... The Jaccard overlap rate between the above-threshold segmented hotspot mask and the low DoLP mask is calculated, and its product with the "edge-to-center" temperature difference forms a composite indicator of water ingress in the package. Temperature-power consistency residuals are also introduced. (Under the same POA conditions) Increasing the power output without a proportional decrease is used to indicate microcracks / poor contact. Regarding quality and consistency, window quality score (DQS), cross-modal reprojection error (RMSE_align), glare coverage, and gating trigger count are recorded. Cross-window consistency indices (such as Spearman rank correlation and linear trend slope) are calculated for the feature sequences formed by three windows on the same day to assess the temporal reproducibility of the evidence. To achieve comparability across components / days, this method estimates the median m and median absolute deviation (MAD) of the features for each subarray on the historical statistics of the "healthy sample set" (window quality meets standards, no candidate alarms triggered, and strong gating). Robust normalization is performed, and the normalized dimensionless feature vectors are output in a fixed order of "thermal - polarization - coherence - electrical - co-occurrence - mass" with a version number, so that they can be directly called by the gating fusion and two-level detection in step eight. At the same time, the original quantities are retained for evidence packages and auditable traceability.

[0078] S8. Two-level detection and gating fusion: In the initial edge screening stage, qualified windows are selected through hard gating, and a lightweight classifier is used to output the edge confidence vector of each fault category. The candidate list is generated by fusing with regular fingerprints; in the central deep diagnosis stage, a spatiotemporal graph neural network ST-GNN is constructed, which incorporates electrical series, spatial proximity and structural coupling constraints, outputs the probability and severity of fault categories, and fuses edge and center confidence.

[0079] This invention employs a two-stage detection structure of "edge screening + center-level in-depth diagnosis". On the edge side, hard gating such as CI / cloud / glare and a lightweight model (logistic regression / XGBoost) based on the minimum working feature set are used for candidate screening and retest triggering. On the center side, a spatiotemporal graph neural network (ST-GNN) is used with "components / substrings" as nodes and "electrical / spatial / structural topology" as edges. It combines time series modeling and physical prior regularization (electrical string consistency and temperature-power monotonicity) to output multi-label fault probability and severity, and performs score fusion and calibration with fingerprint evidence (overlap / LI / HI, etc.).

[0080] In a preferred embodiment, to achieve low false alarm and interpretable fault identification under complex surface conditions, this method is based on the normalized feature vector (including thermal, polarization, coherence, electrical, co-occurrence, and mass subset) output in step seven, and the desired power residual. A two-level detection chain, "edge screening - center depth inference," is constructed based on gating indicators (CI / HI / LI, DQS, GLARE coverage, registration RMSE), and cross-source gating fusion is implemented: First, at the edge, for each record of "component / substring × time window," hard gating and weight reduction are performed—when the attitude coherence index CI exceeds the threshold, the glare coverage exceeds the threshold, or rapid cloud changes (irradiance variation coefficient or...) occur... When the threshold is exceeded or the dew point risk is triggered, only cached evidence is used without making a final judgment; in the window where hard gating is not triggered, the minimum working feature set (MFS) is selected and Using first-order statistics as input, a lightweight classifier (such as log-odds regression or gradient boosting tree XGBoost) is employed to output the marginal confidence vector for each fault category. and with rule-based fingerprint triggers (such as " Hot Topics Overlapping, edge cold zone "Gradient along the border" "HI significant" The candidate list and "retest request" are generated by merging "negative residuals in the early morning / evening", etc.; then, at the central end, a graph node is constructed with "components / substrings" as nodes and three types of edges: electrical series connection, spatial proximity, and structural coupling (denoted as respectively). The spatiotemporal diagram of ) is arranged according to the most recent The stacked windows form the node temporal feature tensor. and edge feature tensor The message passing network with edge features (such as GraphSAGE or GatedGCN) and the temporal modeling unit (such as Gated Recurrent Unit (GRU) or Temporal Convolutional Network (TCN)) are fed into a spatio-temporal graph neural network (ST-GNN). To inject physical priors and improve interpretability, (i) string consistency regularization is superimposed on the loss layer: the health score difference of the same string node within the same window is affected by the string consistency regularization. Constraints, (ii) Temperature-power monotonicity constraint: within a non-gated window, A significant increase should be accompanied by (or power) reasonable decrease soft constraints, and (iii) quality weighting: weighting the sample loss with DQS, GLARE and RMSE to reduce the impact of low-quality segments on the model; the training objective uses multi-label binary cross-entropy (BCE) or focal loss for the fault category probability output head, and smooth L1 loss for the severity regression head; under conditions of insufficient labeling or class imbalance, semi-supervised strategies (such as high-confidence pseudo-label self-training and positive-unlabeled PU learning) and consistency regularization (contrast consistency of features of adjacent spatiotemporal nodes) can be combined to improve generalization; in the inference stage, the central end outputs the category probability vector of each node. With severity The calibration probability is obtained through temperature scaling or Platt calibration; ultimately, this step will... and and evidence consistency score (e.g., the degree of matching with fingerprint features, The model parameters (including the Jaccard overlap rate and LI / HI status of low-DoLP) are fused according to weighted rules to form a "candidate event record" for the next step. Its fields include at least: module_id / substr_id, Top-k fault types and calibration probability, severity estimation, gating status and quality weight, evidence summary and retesting suggestion. The model parameters, gating thresholds and fusion weights are archived with version numbers and can be updated periodically by site in small steps with "golden days" health sets to ensure long-term stability and traceability.

[0081] S9, Location, Hierarchical Alarm and Consistency / Topology Verification: Fault location is completed based on pixel attribution mapping, spatial confidence weight is calculated through intersection-union ratio (IoU), multi-level alarms are set according to confidence and consistency, time, topology and gating consistency verification is implemented, and evidence package containing core evidence is generated.

[0082] This invention uses pixel-to-component / substring ID mapping to assign and merge candidate connected components, defines spatial confidence using location IoU and registration RMSE, and fuses model probability and evidence overlap to obtain total confidence. It employs an L1 / L2 / L3 hierarchical alarm strategy and "same window—cross window—cross day" consistency verification, and combines electrical and spatial topology consistency for confidence weighting and false alarm prevention; it automatically generates evidence packages (…). / DoLP / TPC overlay and (CI / HI / LI curves) and arrange peer follow-up visits and cross-window retests to achieve audit friendliness and operation and maintenance closed loop.

[0083] In a preferred embodiment, this method, based on the candidate events output in step eight, combines the pixel attribution mapping (LUT_ID) and array topology map from step three. Complete precise location, hierarchical alarm, and multi-dimensional consistency verification at the component / substring level: First, for the hot spot mask, cold band mask, and low DoLP mask extracted in step five within each time window, connected component annotation is used, and the pixels of each connected component are... or Projected to world / front coordinate system The LUT_ID is called to assign it to a unique component / substring ID; if multiple connected components exist within the same component and the geometric center spacing is less than a preset ratio or is located within the junction box ROI, they are merged into a single event according to the minimum bounding box and topological proximity rules; the positioning reliability is measured by the intersection-over-union (IoU) ratio of the event mask and the corresponding substring polygon and converted into a spatial confidence weight. Simultaneously record the cross-modal reprojection error ( Glare coverage is used for quality downweighting; subsequently, the calibration probability vector at the center end of step eight is used for quality downweighting. confidence level at the edge And combined with evidence consistency scores (such as Jaccard overlap between hotspots and low DoLP, geometric mean of cold zones and edge zones) The conformity of the gradient along the boundary, whether LI / HI is in place) and the quality weights (DQS, (GLARE percentage), and the final confidence level is obtained by weighted fusion or multiplicative fusion. The method includes the Top-k fault types; the severity is calculated and normalized to [0,1] by the feature subset from step seven according to the physically interpretable mapping of the fault class, and used for subsequent action decisions; regarding alarm classification, this method sets a classification threshold system and hysteresis strategy: when When the observation threshold is reached and the event occurs within the same window but is not reproduced, an L1 observation event is generated and a retest within the same window is automatically scheduled; when When the suspected threshold is reached and there are two or more instances of threshold crossing within the same window, or when the evidence is highly consistent with fingerprint evidence, an L2 suspected event is generated and evidence slices are automatically packaged. When the same component / substring reproduces in adjacent windows or adjacent days, or belongs to a security-related category and its severity reaches a preset security threshold, it is upgraded to L3 confirmed and a work order is triggered. To suppress occasional noise and systemic disturbances, this method implements three types of consistency / anti-false alarm checks: the first is time consistency check, which requires the reproduction to be progressively layered from "same window - across window - across day". After NUC / polarization zero point adjustment, a short-time dead zone is set and a delay is set after gate recovery. The second is topology consistency check, in the graph upper electrical series connection side adjacent edge in space Consistency scoring is applied: if adjacent nodes in the same string simultaneously present similar evidence, the confidence level is increased; if only spatially adjacent areas exhibit anomalous patterns and the attitude coherence index (CI) increases, the score is lowered, indicating a tendency towards occlusion / attitude perturbation. For structurally coupled edges... If the same buoy / raft plate forms a strip-shaped cold zone and If the decline is not accompanied by significant hotspots, the focus is on encapsulating water ingress / contamination rather than single-point failures; thirdly, gating consistency verification is used, where candidates triggered by rapid cloud changes, strong glare, and dew point risk windows are only used as retesting objects and not subject to final judgment, and threshold hysteresis is used to prevent alarm jitter; finally, this method generates an "evidence package" and "event record" for each event that reaches L2 / L3, where the evidence package contains at least: Figures, DoLP plots overlaid with TPC, residuals CI / HI / LI curves, positioning IoU, GLARE percentage, DQS and gating flags and their corresponding calib_version, mapping_version and timestamp; event logs should include at least: module_id / substr_id, fault type and calibration probability, Severity, The above records are archived on disk via the edge-center message channel and connected to the operation and maintenance work order interface, forming a closed-loop link from "location - classification - retesting - diagnosis - handling - write-back", which makes fault identification not only spatial accuracy and temporal robustness, but also topological consistency and auditability.

[0084] S10. Evaluation, Deployment and Self-calibration: A three-layer sample set is used for offline evaluation and online canary release. The parameters of the expected power model are updated by robust regression at night / weekly. The ST-GNN model is periodically fine-tuned. A four-dimensional version chain is established to manage the versions of data, features, models and strategies.

[0085] This invention provides an indicator system (mAP / F1, FAR, lead time, location IoU, cross-window consistency, etc.) and achievement thresholds for offline replay evaluation and online grayscale A / B testing. It also specifies edge / center deployment and data persistence standards, time synchronization and NUC / zero-point maintenance, and model and... Parameters are self-calibrated daily / weekly and drift is monitored; the technical solution can be implemented as a method, device (including coaxial dual-modal camera, IMU and edge computing unit) and system (including central inference and work order interface), and stored in a computer-readable medium for execution.

[0086] In a preferred embodiment, to ensure the reproducibility, long-term stability, and auditability of the method, this step provides a closed-loop mechanism from offline verification to online deployment, and from operational monitoring to parameter / model self-calibration: First, the offline evaluation uses three layers of data: a "golden day sample set" (sunny and stable, gating not triggered), a "challenge day sample set" (cloudy / strong wind / low solar altitude), and a "defect sample set" (confirmed by manual inspection or maintenance records). The pipeline is replayed according to the time window in step four to perform playback evaluation, and an ablation test is conducted (removing GLI sequentially → The system employs dynamic correction, CI / HI / LI gating, TPC fusion, and cross-modal correction to verify the impact of each submodule on detection capabilities and false triggers. Secondly, online verification utilizes an A / B testing approach with subarray-level gray-scale deployment. Models / thresholds at the edge and center are differentiated by "model_version" and "policy_version," respectively. The running arrays are grouped according to a preset ratio, and alarm grading pass rates, evidence package verification consistency, and gating trigger structure distribution are compared and observed. Automatic rollback is triggered when monitoring indicators are abnormal, device health is abnormal, or evidence consistency is insufficient. In terms of deployment and maintenance, the edge computing unit is responsible for implementing preprocessing, fusion, and feature calculation in steps five to seven, implementing gating and initial screening in step eight, and evidence slicing buffering. The data is stored, and the LUT_ID, calib_version, mapping_version, and collected metadata are uploaded with the frame. The central side is responsible for implementing step eight, ST-GNN spatiotemporal topology inference, and step nine, hierarchical alarm and evidence package generation, and connects to the work order system through the operation and maintenance interface. The entire lifecycle of data and models is managed using a four-dimensional version chain: "data version - feature version - model version - policy version". All judgment results and evidence packages are archived with timestamps and version numbers for traceability. In terms of self-calibration and continuous learning, the expected power parameter set is... On a healthy sample set, robust regression with small step sizes is performed nightly or weekly (constraining the rate of change and physical upper and lower bounds). ST-GNN uses confirmation labels and high-confidence pseudo-labels for periodic fine-tuning (freezing the lower layers, using robust loss and sample quality weighting), and input / output drift monitoring is set (such as KL divergence of feature distribution, gating trigger rate, and error residual spectrum changes) to trigger read-only mode or parameter backoff. Under the condition of cross-site data compliance, federated optimization (gradient summary calculated locally at the site and sent to the center for aggregation) can be selected to avoid the outflow of raw images and electrical parameters. In terms of operation and maintenance SOP, daily polarization zero point fast detection and thermal imaging NUC, time consistency inspection (image brightness-IMU cross-correlation), lens / window cleaning self-inspection and wind / dew point risk check are required; weekly review should be performed. For pixel mapping (AprilTag or thermal tag reprojection error), the site performs DoLP baseline reset and defogging heating self-check after encountering heavy fog, rain, or extreme weather. In terms of security and compliance, the edge-to-center link adopts TLS transmission, device certificates, and access control lists (ACLs). Evidence packets and decision logs are encrypted and stored, and audit trails are retained. In terms of version release and rollback, atomic OTA and blue-green switching are adopted to ensure that if an anomaly occurs at any stage, it can be restored to the previous stable state with one click according to the version chain. Through the comprehensive implementation of the above evaluation, deployment, and self-calibration mechanisms, the method of this invention ensures that it maintains stable, explainable, and auditable fault identification capabilities under complex working conditions such as strong glare, wave-following attitude, and high humidity during long-term operation, and facilitates large-scale promotion and standardized operation and maintenance.

[0087] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A floating photovoltaic fault detection method integrating irradiation-attitude dual normalization and multi-index gating, characterized in that, Includes the following steps: S1. Scene Modeling and Detection Object Conventions: Establish a world coordinate system with the power plant's overall control coordinates as the origin. Register the polygonal boundaries of photovoltaic modules and sub-string partitions, associate the module geometry, installation normal and electrical topology, define the source and semantics of input signals, and clarify the defect types and corresponding observable evidence; S2, Multimodal Calibration, Radiometric Correction and Coordinate-Time Unification: Complete the intrinsic parameter calibration of the visible polarization camera and the long-wave infrared camera LWIR, as well as the cross-modal extrinsic parameter calibration of both, and establish a unified world / array coordinate system and a high-precision time synchronization reference. S3. Pixel Attribution Mapping and Array Topology Modeling: Establish a stable attribution mapping of pixel to component / substring ID and an array-level integrated electrical-geometry-structure topology model, and construct an array topology graph with components / substrings as nodes, including electrical connection edges, spatial proximity edges and structural coupling edges. S4. Three-window acquisition and time synchronization / triggering synchronization standard operating procedure (SOP): The three-window timed acquisition strategy of early morning, noon and evening is adopted. The unified time synchronization of the camera, inertial measurement unit (IMU), electrical parameters and meteorological acquisition equipment is realized through the PTP protocol. The window quality score (DQS) is calculated in real time, and the candidate region of interest (ROI) is automatically scheduled for retesting. S5. Water Surface Environment Suppression Preprocessing and Temperature-Polarization Fusion: Perform imaging preprocessing and cross-modal fusion for water surface scenes to form a temperature-polarization fusion map (TPC). S6. Irradiation-Attitude Dual Normalization Desired Power Modeling and Residual Calculation: Calculate the incident angle, correct it with IAM to obtain the effective irradiation, estimate the battery temperature through Faimin or PVsyst style models, construct the desired power P*(t), and calculate the normalized residual. S7. Construct multimodal features and perform robust normalization: Extract thermal spatial features, polarization texture features, spectral coherence features, electrical and physical features, cross-modal co-occurrence features, and quality and consistency features, and perform MAD normalization. S8. Two-level detection and gating fusion: In the initial edge screening stage, qualified windows are selected through hard gating, and a lightweight classifier is used to output the edge confidence vector of each fault category. The candidate list is generated by fusing with regular fingerprints; in the central deep diagnosis stage, a spatiotemporal graph neural network ST-GNN is constructed, which incorporates electrical series, spatial proximity and structural coupling constraints, outputs the probability and severity of fault categories, and fuses edge and center confidence; S9, Location, Hierarchical Alarm and Consistency / Topology Verification: Fault location is completed based on pixel attribution mapping, spatial confidence weight is calculated through intersection-union ratio (IoU), multi-level alarms are set according to confidence and consistency, time, topology and gating consistency verification is implemented, and evidence package containing core evidence is generated; S10. Evaluation, Deployment and Self-calibration: A three-layer sample set is used for offline evaluation and online canary release. The parameters of the expected power model are updated by robust regression at night / weekly. The ST-GNN model is periodically fine-tuned. A four-dimensional version chain is established to manage the versions of data, features, models and strategies.

2. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S1, the mapping relationship between the fault defect type and the corresponding observable evidence is as follows: wet-induced PID / condensation mismatch is indicated by the Voc-G hysteresis index HI as observable evidence; water ingress / corrosion in connectors or manifolds is indicated by localized... Co-occurrence of hotspots and polarization decrease and humidity-sensitive admittance fingerprint LI= As evidence, Y represents the low-voltage admittance. Relative humidity; microcracks / solder joint defects are evidenced by stable hot spots and temperature-power consistency residuals; water ingress in the package is evidenced by edge cold band geometry and Gradient along the border is used as evidence; biological attachment / contamination is used as evidence by DoLP and texture statistical changes; attitude occlusion / drift and strong water surface reflection are gating based on power-attitude amplitude squared coherence index CI and polarization glare indication; cable insulation dampness is used as evidence by multi-point micro-hot spots and LI.

3. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S2, a high-contrast AprilTag or checkerboard planar target is used to perform geometric intrinsic parameter calibration of the visible polarization camera, and the intrinsic parameter matrix of the visible polarization camera is calculated. With distortion parameters The intrinsic parameters of the long-wave infrared (LWIR) camera were calibrated using a hot-cold hybrid target, and the geometric intrinsic parameters of the LWIR camera were calculated. With distortion The cross-modal extrinsic parameters are calculated using a hybrid target combining AprilTag and thermal labeling. First, the target surface in world coordinates is obtained using a visible polarization camera. pose in Then, the coordinates of the thermal marker pixels are extracted from the thermal image, and the rigid body transformation between the thermal / visible coordinate systems is calculated by combining least squares or robust optimization. The homography matrices of the visible polarization camera and the long-wave infrared camera to the array plane are derived respectively. .

4. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 3, characterized in that, When calibrating the intrinsic parameters of the long-wave infrared camera, solve for the geometric intrinsic parameters of the LWIR of the long-wave infrared camera. With distortion Subsequently, a pixel-level non-uniformity correction library was established based on the two-point method. Make pixel readings according to Linear scaling, where, Each pixel Gain correction factor and bias correction factor, This is the raw reading for that pixel. The equivalent temperature reading is corrected for non-uniformity; then, the apparent emissivity of the component cover glass, backplate, and junction box is measured based on material parameters or small sample measurements. The parameters were entered into a parameter table, and the apparent temperature of the reflection was measured using the pleated aluminum foil method. The apparent emissivity of the backplate and junction box and apparent temperature of reflection The parameters are stored in the parameter table as radiation inversion parameters for subsequent temperature inversion corrections, under the near-field radiative transfer approximation. Temperature inversion corrections are performed to obtain comparable surface temperature fields.

5. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 3, characterized in that, The steps for calibrating the geometric intrinsic parameters of a visible polarization camera are as follows: 1) To achieve channel gain and bias consistency for polarization cameras under uniform integrating sphere or diffuse target illumination; 2) Assemble a high extinction ratio linear polarizer for fixed step rotation angle sampling, estimate the actual orientation deviation and efficiency matrix of the analyzer microarray, obtain the pixel-level correction coefficients from the Stokes parameter solution, and perform absolute alignment of the polarization angle AoP zero point under a known zero-degree reference. 3) Use a high-contrast AprilTag or checkerboard planar target to perform visible camera geometric intrinsic parameter calibration and calculate the intrinsic parameter matrix of the polarization camera. With distortion parameters .

6. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S3, the pixel coordinates of each frame in the visible and thermal infrared channels are... respectively or Back projection to world / front coordinate system get and call the R-tree-based spatial index pair Perform point-polygon determination and assign it to a pre-defined polygon. The registered component polygons and their substring partition polygons are assigned to pixel-based LUT_IDs in the form of a lookup table. Apply a 3–5 pixel geometric indentation to the component / substring polygons and mask the specular glare mask GLARE_MASK and thermal image NUC marker pixels when generating the LUT_ID.

7. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S3, an array topology is constructed using components / substrings as the smallest electrical units. Node set For the corresponding component or substring, additional static properties include the installation normal. Effective area Azimuth / tilt, pixel set index and ROI geometric description; additional dynamic attributes updated with the window include window quality score (DQS), cross-modal reprojection error (RMSE_align), glare coverage, temperature-polarization fundamental statistics and power residual statistics; edge set It consists of three subsets, one of which is the electrical connection edge. It records the electrical direction identifier and circuit ID, and secondly, the spatial proximity edge. It also records the Euclidean distance and relative orientation, and the third is the structural coupling edge. The bearing unit identifier and flexible coupling parameters are recorded.

8. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S4, the time division criteria for the three windows are as follows: the morning window is from sunrise to 15° solar altitude angle, the noon window is the peak solar altitude angle ±2 hours, and the evening window is from 15° solar altitude angle to before sunset.

9. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S4, the window quality score DQS is a weighted combination of image sharpness / registration RMSE, glare coverage, irradiance stability, IMU attitude variance, and wind / cloud / dew point risk.

10. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S5, the imaging preprocessing operation for the water surface scene is performed as follows: 1) Thermal image preprocessing: Perform NUC correction on the raw thermal image, and call up the apparent emissivity of the component cover glass, backplate and junction box recorded in the parameter table in step S2. And combined with the apparent temperature of reflection measured by the pleated aluminum foil method in step S2 Radiation inversion was performed to obtain the quantitatively corrected surface temperature field. ; In the time domain, exponential smoothing or Kalman filtering is used to suppress thermal noise, and the temperature anomaly map is calculated with the median temperature of the same subarray as a reference. ; 2) Polarization image preprocessing: Calculate the linear polarization degree DoLP and polarization angle AoP based on the pixel-level Stokes solution matrix, and use the Fresnel properties of the mirror to establish a glare detection criterion and generate a glare mask GLARE_MASK; 3) De-jittering and fusion: After wave phase de-jittering and cross-modal registration driven by IMU, a weighted field is constructed to generate a temperature-polarization fusion map (TPC).

11. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 10, characterized in that, In step S5, wave phase de-jittering is achieved by bandpass filtering the pitch / roll signal measured by the IMU at 0.2–2 Hz; a reference attitude normal is constructed. Cross-modal registration employs homography plane correction combined with small-amplitude optical flow micro-alignment to ensure that the reprojection error RMSE_align ≤ 0.5 pixels.

12. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 10, characterized in that, In step S5, TPC is calculated using a multi-scale guided fusion formula: ； in, For normalization ; For normalization Texture, Water film / pollution index, This represents a feature-level fusion operator that concatenates and / or weights multiple features within the feature space; Temperature weighting; Polarization weights; To prevent the use of tiny positive constants with a denominator of zero; temperature weighting. Follow Monotonically increasing, polarization weight Follow (1) DoLP and the water film / contamination index SI increase monotonically for GLARE_MASK coverage area and local data quality score. Below the preset quality threshold DQS th The pixels, will Set to zero. DQS th Determined through statistical analysis of a healthy sample set.

13. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S6, the formula for calculating the incident angle is: ； in: The solar vector; The instantaneous normal direction after de-jittering; .

14. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S6, the step of obtaining effective irradiation is as follows: Direct measurement using array-type irradiation sensors on the irradiated side. If only horizontal / scattering is present, the Perez / Perez-Driesse model is used to obtain the array direct dispersion component by transposing the GHI / DHI / DNI model and then superimposing the water surface reflection term to obtain the following: ； in, DHI represents the total horizontal irradiance, DNI represents the horizontal diffuse irradiance, and DNI represents the direct normal irradiance. The total irradiance of the array surface. The water surface reflection component of the array surface; For the component tilt angle, The equivalent albedo is the water surface albedo; the Perez / Perez-Driesse model is an anisotropic sky irradiance transpose model, used to calculate the direct and scattered components of the array based on GHI / DHI / DNI. ； in, For baseline albedo, The gain coefficient is determined through robust regression of healthy samples; The water surface glimmer index (GLI) is obtained by mapping the area percentage of GLARE_MASK. IAM is used to correct the angle of the direct component: ； in, Cover plate / IAM coefficient; Effective irradiation The calculation formula is: or .

15. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S6, the formula for calculating the battery temperature is: ； in, For ambient temperature, For wind speed, These are the heat exchange parameters.

16. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S6, the formula for calculating the desired power P*(t) is: ； in: The system constant loss coefficient; The nominal area of ​​the component; Pollution / water film reduction factor; STC is the temperature coefficient; For baseline efficiency; For effective irradiation; Battery temperature; Low illumination correction factor; Baseline efficiency The calculation formula is: ； in, This is the rated power of STC; The formula for calculating the pollution / water film reduction factor is: ； in, This refers to the sensitivity parameter.

17. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S6, the formula for calculating the normalized residual is: ； ； in, This represents the measured output power of the component / substring at time t.

18. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S7, the thermal spatial features include Statistics, area and aspect ratio of hotspot connected regions, minimum distance of hotspots from the edge, bandwidth and length ratio of cold bands along the long side of the component, and the difference between the mean values ​​of the edge band and the center band; polarization texture features include DoLP / AoP gradient, glare coverage; spectral coherence characteristics including attitude coherence index (CI), main peak frequency, modulation depth; electrophysical characteristics including normalized residual. (t) window mean, low-voltage admittance and LI, temperature-power consistency residual and hysteresis index HI; Cross-modal co-occurrence features include The Jaccard overlap rate is calculated between the hotspot mask and the low DoLP mask for above-threshold segmentation; quality and consistency features include window quality score (DQS), cross-modal reprojection error (RMSE_align), GLI, and cross-window consistency.

19. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S8, the triggering condition for hard gating is that any of the following conditions are met: CI exceeds the preset threshold, glare coverage exceeds the threshold, irradiance variation coefficient or dG / dt exceeds the threshold, or dew point risk is triggered. The window that triggers hard gating only caches evidence and does not make a judgment.

20. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S8, the ST-GNN includes a message passing network and a time modeling unit. The loss layer is superimposed with L2 / Huber constraints on electrical string consistency, soft constraints on temperature-power monotonicity, and quality weighting terms. Multi-label binary cross-entropy or focal loss is used to optimize the fault category output, and smoothed L1 loss is used to optimize the severity regression.

21. The floating photovoltaic fault detection method based on irradiation-attitude dual normalization and multi-index gating fusion according to claim 1, characterized in that, In step S9, when When the observation threshold is reached and the event occurs within the same window but is not reproduced, an L1 observation event is generated and a retest within the same window is automatically scheduled; when When the suspected threshold is reached and there are two or more instances of exceeding the threshold within the same window, or when the evidence is highly consistent with fingerprint evidence, an L2 suspected event is generated and evidence slices are automatically packaged. When the same component / substring is reproduced in adjacent windows or adjacent days, or when it belongs to a security-related category and the severity reaches the preset security threshold, it is upgraded to L3 confirmed and a work order is triggered.

Images