A photovoltaic module anomaly identification method based on laser detection

Abstract

The present application relates to the technical field of control adjustment system, and more particularly to a photovoltaic module abnormality identification method based on laser detection, which generates a clock signal synchronized with the timing of laser scanning through a hardware synchronous trigger, drives a temperature and humidity sensor, a salt spray sensor and an illumination intensity sensor to collect environmental parameters, and constructs a multi-source data set containing salt crystallization trend and environmental interference level. Based on the temperature and humidity segmented compensation reflectivity baseline, the salt spray mode switching optimizes the dust accumulation determination threshold, and the light correlation sensitivity adjusts the hot spot detection parameters, the model dynamic correction is realized. Matching the fault database triggers the hierarchical alarm mechanism, and combining the decision tree classifier verifies the time continuity to generate a diagnosis report. Through spiral path planning, multi-machine contract network protocol and wind disturbance trajectory correction, a dynamic detection path is constructed, a closed-loop control logic is formed by executing data feedback, and the misjudgment and missed detection problems in high-salt fog, temperature variation and light fluctuation environments are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of control and regulation system technology, and in particular to a method for identifying anomalies in photovoltaic modules based on laser detection. Background Technology

[0002] Laser-based photovoltaic (PV) module inspection refers to the intelligent monitoring of PV modules using laser inspection technology. This technology acquires the optical features of the PV panel surface through high-precision laser scanning equipment and combines this with multispectral analysis algorithms to identify the module's operating status. In the environment of high salt spray and dynamically changing temperature and humidity in salt fields, the laser inspection system captures abnormal features such as cracks, hot spots, and dust accumulation on the PV module surface through non-contact measurement. It then constructs a multi-dimensional fault model using reflected light intensity and wavelength shift information. The system compares the detection data with preset operating parameters and dynamically optimizes the judgment threshold using an environmental variable correction module, achieving accurate location and classification of abnormal features. This effectively improves the operational stability and maintenance response efficiency of PV arrays under complex climatic conditions, providing technical support for the efficient development of clean energy in fish-solar complementary scenarios.

[0003] In the field of photovoltaic (PV) power generation system control and regulation, PV module anomaly identification technology based on laser detection suffers from insufficient adaptability of control strategies to dynamic environments. Existing methods rely on fixed thresholds or single-parameter models for anomaly detection, which cannot effectively adapt to the complex operating conditions of PV arrays caused by salt spray corrosion, temperature gradient changes, and fluctuations in light intensity. When laser scanning equipment acquires multi-dimensional optical features of the PV module surface, existing control and regulation systems struggle to simultaneously analyze the coupling relationship between environmental interference factors and equipment operating parameters, leading to misjudgments or missed detections during anomaly feature extraction. This deficiency directly affects the real-time monitoring accuracy of PV array operation, reduces fault location efficiency, and increases operation and maintenance costs and energy loss risks. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a laser-based method for identifying anomalies in photovoltaic modules. This method solves the problem of misjudgment and missed detection caused by insufficient analysis of the coupling between environmental interference factors and equipment parameters in existing photovoltaic module anomaly identification methods under dynamic salt spray erosion, temperature gradient changes, and light fluctuations.

[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0006] This invention provides a method for anomaly identification of photovoltaic modules based on laser detection, comprising:

[0007] Step 1: Collect dynamic environmental parameters of the area where the photovoltaic module is located. The dynamic environmental parameters include temperature and humidity, salt spray concentration and light intensity. Generate a multi-source dataset containing environmental variables by using timestamps and synchronizing with the laser scanning time sequence. Input the multi-source dataset into the laser emitter control unit and the preset data fusion model.

[0008] Step 2: Dynamically adjust the output wavelength of the laser emitter based on the salt spray concentration data in the multi-source dataset. When the salt spray concentration exceeds the preset threshold range, switch to the near-infrared band. At the same time, adjust the passband range of the tunable filter at the receiver according to the light intensity data in the multi-source dataset to suppress ambient light noise.

[0009] Step 3: Perform multi-dimensional feature extraction on the adjusted reflected laser signal, including time-frequency domain features of reflection intensity based on wavelet transform, wavelength shift features of multispectral decomposition, and three-dimensional spatial positioning features combined with UAV attitude data, and generate multi-dimensional anomaly feature vector by fusing temperature gradient data from multiple source datasets.

[0010] Step 4: Construct a temperature and reflection intensity compensation model based on temperature and humidity data from the multi-source dataset, adjust the dust accumulation judgment threshold range based on salt spray concentration data from the multi-source dataset, and optimize the hot spot detection sensitivity parameters based on light intensity data from the multi-source dataset.

[0011] Step 5: Input the multidimensional abnormal feature vector into the preset fault model database for similarity matching, trigger the hierarchical alarm mechanism based on the matching result, and verify the temporal continuity of the multidimensional abnormal feature vector through the decision tree classifier to eliminate instantaneous environmental interference signals.

[0012] Step 6: Generate an optimized detection path based on the anomaly localization results. The path includes dense scanning area division based on multi-dimensional anomaly feature vectors, multi-UAV collaborative task allocation, and dynamic trajectory correction based on real-time environmental response. The path execution data is then fed back to a preset data fusion model to update the temperature and reflection intensity compensation model parameters.

[0013] Furthermore, in the aforementioned laser-based photovoltaic module anomaly identification method, the generation of a multi-source dataset containing environmental variables includes:

[0014] A synchronous clock signal is generated by using the laser emission pulse signal as a reference clock source through a hardware synchronous trigger.

[0015] The temperature and humidity sensor and the salt spray sensor trigger a hardware interrupt based on the synchronous clock signal to collect temperature and humidity data and salt spray concentration data synchronously with the laser scanning cycle.

[0016] The light intensity sensor integrates the laser echo signal through an integrating circuit to generate illuminance waveform data aligned with the synchronous clock signal.

[0017] The temperature and humidity data, salt spray concentration data, and illuminance waveform data are aligned by timestamp and integrated into a multi-source dataset. The multi-source dataset is then input into the particle deposition rate calculation module. Based on the salt spray concentration data, the amount of salt crystallization deposition per unit time is calculated, and the salt crystallization formation trend and environmental interference level are output.

[0018] Furthermore, in the aforementioned laser-based photovoltaic module anomaly identification method, step 2 includes:

[0019] Based on the historical salt spray concentration data in the multi-source dataset, a sliding window mechanism is used to dynamically update the salt spray concentration threshold boundary, which is used to determine whether laser wavelength switching is triggered.

[0020] When the laser emitter switches to the near-infrared band, the laser pulse width is simultaneously adjusted to the preset penetration optimization range;

[0021] The receiver scattering noise estimator performs inversion calculations on the aerosol particle size distribution based on Mie scattering theory, and compares the backscattering coefficient differences between adjacent scanning points using differential detection technology. When the difference exceeds a set threshold, the corresponding scattering interference signal is eliminated.

[0022] Furthermore, in the aforementioned photovoltaic module anomaly identification method based on laser detection, step 3 includes: decomposing the reflection intensity signal to the fifth level of detail coefficients using the Db4 wavelet basis, setting a sliding energy window in the 200-400Hz frequency band, and marking the energy fluctuation within the window as a crack feature corresponding to a sudden change in reflectivity when the energy fluctuation exceeds a preset threshold.

[0023] By analyzing multispectral reflectance data using the etalon interferometry method, the characteristic peaks corresponding to the bandgap of photovoltaic materials are identified, and a hot spot temperature mapping table is established based on the linear relationship between the characteristic peak wavelength shift and temperature change.

[0024] The pitch and roll angles from the UAV attitude data are input into the coordinate transformation matrix. After correcting the spatial distortion of the laser point cloud data, the second derivative of reflectivity is calculated using the Laplacian operator. When the derivative value exceeds the set range, it is marked as a dust gradient distribution area.

[0025] Furthermore, in the aforementioned photovoltaic module anomaly identification method based on laser detection, step 4 includes: based on the temperature and humidity data in the multi-source dataset, dividing the temperature range into five segments from -20℃ to 80℃, with each segment associated with an independent reflectivity baseline value, the reflectivity baseline value being used to normalize the reflectivity intensity data;

[0026] The dual-range dust accumulation determination mode is activated based on the salt spray concentration data in the multi-source dataset. When the salt spray concentration is ≥5mg / m³, the mode is switched to salt crystallization mode, and morphological filtering is added to distinguish between salt crystallization and regular dust accumulation.

[0027] Based on the light intensity data in the multi-source dataset, the hot spot detection sensitivity is linearly weighted and correlated with the light intensity. When the light intensity is <500W / m², the detection trigger frequency is reduced to 70% of the baseline value.

[0028] Furthermore, in the laser-based photovoltaic module anomaly identification method, step 5 includes: when the matching degree between the multidimensional anomaly feature vector and the preset fault model exceeds 85%, triggering a level one alarm, locking the fault feature evidence chain in the multidimensional anomaly feature vector, and freezing the environmental parameter update operation in the preset data fusion model.

[0029] When the matching degree is between 60% and 85%, a level 2 alarm is triggered, a re-inspection command is generated to drive the drone to re-inspect the abnormal area within 3 minutes, and the coordinate data of the re-inspection area is input into the path planning engine;

[0030] When the salt spray concentration in the multi-source dataset exceeds 5 mg / m³ or the light intensity is below 500 W / m², a level three alarm is triggered, the verification process of the backup sensor is initiated, and the pulse frequency in the laser scanning parameters is adjusted to a safe mode.

[0031] Furthermore, in the aforementioned photovoltaic module anomaly identification method based on laser detection, step 6 includes: generating a spiral expansion scanning path with a pitch increasing by 0.5m centered on the anomaly point based on the anomaly positioning coordinates in the multidimensional anomaly feature vector, and increasing the scanning line spacing from 10cm to 20cm with the number of scans.

[0032] Based on the remaining battery power data and sensor health status scores of the drones in the multi-source dataset, the detection area is divided into multiple sub-regions using an improved contract net protocol, and assigned to multiple drones to perform collaborative tasks.

[0033] When the wind speed in the real-time environmental monitoring data exceeds 8 m / s, a 15° yaw compensation angle is inserted into the flight trajectory, the spacing between hovering acquisition points is increased to 1.5 times the standard density, and the modified trajectory is smoothed using a Bezier curve interpolation algorithm.

[0034] Furthermore, the photovoltaic module anomaly identification method based on laser detection further includes: the master drone generating bidding information including the topological connection relationship of the abnormal area based on the abnormal positioning coordinates, and broadcasting it to the slave drones through a wireless mesh network; the slave drones calculating bidding weights based on remaining power, flight stability error value, and sensor accuracy deviation value, wherein the bidding weights are positively correlated with the remaining power and negatively correlated with stability error and accuracy deviation; based on the bidding weights, the sub-region to be detected is divided into topological association units and environmental parameter similarity units, wherein the topological association units are divided according to the physical connection relationship of the abnormal components, and the environmental parameter similarity units are divided according to the areas with salt spray concentration difference ≤15% and temperature gradient difference ≤5℃.

[0035] Furthermore, the laser-based photovoltaic module anomaly identification method also includes a closed-loop feedback mechanism:

[0036] The closed-loop feedback mechanism includes a forward control chain and a reverse calibration process;

[0037] The forward control chain inputs environmental monitoring data from the multi-source dataset into the device parameter optimization module every 30 seconds to generate an adjustment suggestion table for laser output power and UAV flight speed, and increases the laser output power to 80% of the increase in UAV flight speed according to the suggestion table;

[0038] The reverse calibration process cross-validates the detection data of the same area using three drones, calculates the average deviation of the sensor measurements as the drift compensation, and updates the exponential parameters in the battery degradation coefficient model based on the battery consumption rate (unit: mAh / km) and trajectory positioning error (unit: cm) during path execution.

[0039] Furthermore, in the laser-based photovoltaic module anomaly identification method, step 5 includes: the root node calculates the environmental interference level based on the salt spray concentration, temperature gradient, and light intensity data in the multi-source dataset; when the interference level is ≥4, an anti-interference detection branch is selected; the intermediate node verifies the temporal continuity of the multidimensional anomaly feature vector; when the anomaly signal lasts for ≥2 seconds, it is determined to be a valid anomaly, otherwise it is marked as transient interference; the leaf node matches the multidimensional anomaly feature vector with a standardized fault type coding library to generate a diagnostic report including fault coordinates (X,Y,Z), type coding, and confidence level ≥90%.

[0040] Beneficial effects of this invention;

[0041] This invention effectively solves the problem of misjudgment and missed detection in photovoltaic module anomaly identification through dynamic environmental parameter coupling analysis and multi-dimensional data fusion mechanism. It generates a multi-source dataset by simultaneously collecting temperature, humidity, salt spray concentration, and light intensity. A sliding window threshold update triggers the laser wavelength to switch to the near-infrared band, and ambient light noise is suppressed based on the light intensity-adjusted filter passband. Wavelet transform is used to extract crack time-frequency features, etalon interferometry is used to quantify hot spot temperature, and UAV attitude correction is used to locate dust accumulation gradients. Temperature gradients are fused to construct a multi-dimensional anomaly feature vector. Segmented compensation of reflectivity baseline based on temperature and humidity, optimization of dust accumulation judgment through salt spray mode switching, and dynamic adjustment of light sensitivity improve the robustness of hot spot detection. An alarm mechanism based on matching degree grading is used, combined with a decision tree classifier to verify time continuity and generate a high-confidence diagnostic report. Dynamic detection path optimization is achieved using spiral path planning, contract network protocol collaborative task allocation, and wind disturbance trajectory correction, and the execution data is fed back to the model for iterative updates, forming a closed-loop control logic. The above technical solutions significantly improve the anomaly identification accuracy and system adaptability under complex salt spray, temperature change, and light fluctuation environments. Attached Figure Description

[0042] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0043] Figure 1 This is a flowchart of a photovoltaic module anomaly identification method based on laser detection, provided as an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings. To better understand the objectives of this invention, it will be described in further detail below.

[0045] Please see Figure 1 The present invention provides a method for anomaly identification of photovoltaic modules based on laser detection, comprising:

[0046] Step 1: Collect dynamic environmental parameters of the area where the photovoltaic module is located. The dynamic environmental parameters include temperature and humidity, salt spray concentration and light intensity. Generate a multi-source dataset containing environmental variables by using timestamps and synchronizing with the laser scanning time sequence. Input the multi-source dataset into the laser emitter control unit and the preset data fusion model.

[0047] The technical solution in step 1 of this invention achieves accurate capture of environmental parameters through a multi-sensor synchronous acquisition and data fusion mechanism. The specific implementation process is as follows:

[0048] A hardware synchronization trigger uses the laser emission pulse signal as a reference clock source to generate a clock signal that is strictly synchronized with the laser scanning action. The synchronization trigger employs an FPGA programmable logic device, receiving the rising edge signal of the pulse output from the laser drive circuit as the trigger source to generate a synchronization clock signal with millisecond-level precision, providing a unified time reference for all sensors. The temperature and humidity sensor uses a capacitive sensing element, detecting ambient humidity based on the principle of dielectric constant change, and writes the data to a circular buffer queue via the I2C communication protocol. The salt spray sensor detects the sodium ion concentration in the air through electrochemical electrodes, and the acquired signal is stored in a shared memory area after analog-to-digital conversion. Both types of sensors trigger sampling on the rising edge of the synchronization clock signal, ensuring that the data acquisition time is strictly aligned with the laser scanning cycle.

[0049] The light intensity sensor preprocesses the laser echo signal using an integrating circuit, which consists of an operational amplifier and an RC filter network. This circuit attenuates the AC component in the echo signal and extracts the DC level representing the ambient light intensity. The preprocessed signal is converted into a digital signal by an analog-to-digital converter. The sampling time is controlled by a synchronous clock signal, generating illuminance waveform data aligned with the timestamps of temperature, humidity, and salt spray data. The data from each sensor are aligned using a timestamp matching algorithm, and linear interpolation is used to compensate for micro-time-sequence deviations caused by transmission delays, forming a multi-source dataset with consistent time dimensions.

[0050] The integrated multi-source dataset is input into the particle deposition rate calculation module, which calculates the amount of salt crystallization deposition per unit area based on the time series of salt spray concentration data. The calculation module uses a sliding window to statistically analyze the mean and variance of salt spray concentration, and derives a salt spray crystallization rate model in conjunction with temperature and humidity parameters, outputting a salt crystallization trend index and environmental interference level parameters. These environmental interference level parameters, along with the multi-source dataset, are input into the laser emitter control unit to provide a basis for subsequent laser wavelength switching and power adjustment; simultaneously, they are input into a preset data fusion model to perform spatiotemporal correlation analysis with laser reflection signal characteristics, supporting the dynamic correction of the anomaly detection model.

[0051] The above steps achieve temporal consistency of multi-source data through hardware synchronization triggering, separate the spatial distribution characteristics of environmental parameters through sensor-specific acquisition mechanisms, and finally establish a dynamic evaluation model of environmental interference through data fusion and calculation modules. Each step forms a progressive logical chain of "time synchronization → data acquisition → signal processing → data fusion," providing highly consistent environmental parameter input for subsequent anomaly identification and solving the feature drift problem caused by asynchronous multi-source data in traditional methods.

[0052] Step 2: Dynamically adjust the output wavelength of the laser emitter based on the salt spray concentration data in the multi-source dataset. When the salt spray concentration exceeds the preset threshold range, switch to the near-infrared band. At the same time, adjust the passband range of the tunable filter at the receiver according to the light intensity data in the multi-source dataset to suppress ambient light noise.

[0053] The technical solution in step 2 improves detection accuracy in complex environments by dynamically adjusting laser parameters and noise suppression strategies. The specific implementation process is as follows:

[0054] Based on historical salt spray concentration data from a multi-source dataset, a sliding window mechanism is used to dynamically update the salt spray concentration threshold boundary. The sliding window calculates the mean and variance of the salt spray concentration every 10 minutes, and adjusts the threshold range based on current temperature and humidity parameters. When the real-time salt spray concentration exceeds the upper threshold, the laser emitter control unit receives a switching command and switches the output wavelength from the visible light band to the near-infrared band. The wavelength switching module adjusts the drive current of the laser diode through an electro-optic modulator to match the emission spectral characteristics of the near-infrared band, and simultaneously adjusts the laser pulse width to a preset penetration optimization range, improving the laser's penetration efficiency and signal-to-noise ratio in the salt spray environment.

[0055] The passband range of the tunable filter at the receiver is dynamically adjusted based on the illumination intensity data from the multi-source dataset. Illumination intensity is quantized by a photosensitive sensor and then input into a digital signal processor to calculate the passband width adjustment. Under high illumination intensity, the passband width is compressed to 70% of the standard range, filtering out high-frequency components of ambient stray light; under low illumination intensity, the passband width is expanded to 120% of the standard range, retaining the low-frequency effective components of the reflected signal. The filter adjustment parameters are output to the voltage-controlled filter circuit via a digital-to-analog converter module, achieving real-time dynamic adaptation of the passband.

[0056] The scattering noise suppression module is based on the Mie scattering theory to invert the aerosol particle size distribution model and combines differential detection technology to compare the differences in backscattering coefficients between adjacent scanning points. When the difference in scattering coefficients between adjacent points exceeds a set threshold, it is determined to be environmental scattering noise, and the interference signal is eliminated through a digital filtering algorithm. The differential detection adopts a dual-channel parallel processing architecture to calculate the spatial gradient and temporal fluctuation characteristics of the scattered signal, distinguishing between real reflected signals and random noise. The noise-suppressed signal is input into the feature extraction module to provide high-purity data input for subsequent anomaly identification.

[0057] The above steps form a closed-loop technology chain of "environmental perception → equipment optimization → signal purification" through wavelength switching driven by salt spray concentration, filter adjustment correlated with light intensity, and dynamic suppression of scattering noise. The salt spray threshold model, linked with laser parameters, addresses the penetration challenge in high-salt-spray environments; the linkage between light intensity and filter passband suppresses ambient light interference; and scattering noise modeling and differential detection technology improve signal quality. Each step achieves coordinated control based on environmental parameters from multi-source datasets, ultimately providing high-quality laser reflection signals adapted to complex operating conditions for anomaly feature extraction.

[0058] Step 3: Perform multi-dimensional feature extraction on the adjusted reflected laser signal, including time-frequency domain features of reflection intensity based on wavelet transform, wavelength shift features of multispectral decomposition, and three-dimensional spatial positioning features combined with UAV attitude data, and generate multi-dimensional anomaly feature vector by fusing temperature gradient data from multiple source datasets.

[0059] The technical solution in step 3 extracts the abnormal features on the surface of the photovoltaic module through multi-dimensional signal processing and data fusion methods. The specific implementation process is as follows:

[0060] The reflected laser signal was decomposed using the Db4 wavelet basis at multiple scales, down to the fifth level of detail coefficients to focus on the energy distribution characteristics in the 200-400Hz frequency band. A sliding energy window was set in this frequency band, and the standard deviation of the signal energy within the window was calculated as a fluctuation index. When the fluctuation value exceeded a preset threshold, it was marked as a region of abrupt reflectivity change. This abrupt reflectivity change is related to the change in local light scattering characteristics caused by microcracks on the photovoltaic panel surface, and the initial location of the crack is achieved through the frequency domain energy distribution characteristics.

[0061] Multispectral reflectance data was analyzed using etalon interferometry, and the reflectance signals of different wavelengths were separated using the Fabry-Perot interferometry principle to pinpoint the characteristic peaks corresponding to the bandgap of the photovoltaic material. The wavelength shift of the characteristic peaks showed a linear relationship with the module temperature change. A temperature-wavelength shift mapping table was established through calibration experiments, converting the real-time detected wavelength shift into hotspot temperature values. The mapping table is stored in a database, supporting quantitative assessment and dynamic tracking of the hotspot region temperature.

[0062] The pitch and roll angles from the UAV attitude data are input into a coordinate transformation matrix to calculate the actual coordinates of the laser-scanned point cloud in three-dimensional space. Coordinate transformation eliminates point cloud distortion caused by changes in UAV flight attitude, reconstructing the spatial reflectivity distribution of the photovoltaic panel surface. The Laplacian operator is applied to the corrected reflectivity data to calculate the second-order derivative, extracting the spatial variation characteristics of the reflectivity gradient. When the derivative value exceeds a set range, it is identified as a dust accumulation gradient distribution region. This method enhances the gradual reflectivity variation characteristics caused by dust accumulation through spatial differential operations, improving the sensitivity of dust accumulation detection.

[0063] Temperature gradient data from multiple sources is integrated to spatiotemporally correlate the temperature gradient distribution with features such as cracks, hot spots, and dust accumulation. The temperature gradient data is generated by calculating the difference between an infrared thermal imager and a temperature sensor. Combined with the timestamp information of the anomaly feature vector, a multidimensional anomaly feature vector is constructed, encompassing the time, frequency, spectral, and spatial domains. This vector is then input into a pre-defined fault model database for similarity matching, providing multidimensional input for subsequent graded alarms and fault classification.

[0064] The above steps extract anomalous features from four dimensions: time-frequency analysis, spectral analysis, spatial correction, and temperature correlation. A unified multidimensional feature vector is generated through data fusion. Wavelet decomposition identifies high-frequency abrupt changes corresponding to cracks, etalon interferometry quantifies hot spot temperature, UAV attitude correction and differential calculation locate dust accumulation gradients, and temperature gradient data enhances the environmental correlation of anomaly detection. Through the synergistic effect of signal processing, optical analysis, and spatial modeling techniques, each sub-step forms a technical chain of "signal decomposition → feature analysis → spatial correction → data fusion," ultimately achieving comprehensive capture and feature analysis of anomalies on the photovoltaic module surface.

[0065] Step 4: Construct a temperature and reflection intensity compensation model based on temperature and humidity data from the multi-source dataset, adjust the dust accumulation judgment threshold range based on salt spray concentration data from the multi-source dataset, and optimize the hot spot detection sensitivity parameters based on light intensity data from the multi-source dataset.

[0066] The technical solution in step 4 achieves accurate anomaly detection in complex environments by dynamically adjusting the detection model parameters. The specific implementation process is as follows:

[0067] Based on temperature and humidity data from a multi-source dataset, the temperature range is divided into five segments from -20℃ to 80℃, each spanning 20℃. Each temperature segment is associated with an independent reflectivity baseline value, which is determined by the statistical mean of the reflectivity of the photovoltaic panel surface at different temperatures, calibrated in the laboratory. Real-time reflectivity intensity data is normalized according to the baseline value corresponding to the current temperature segment to eliminate reflectivity drift caused by temperature fluctuations. The normalized reflectivity intensity is then input into the dust accumulation detection module to reduce the interference of temperature on dust accumulation detection.

[0068] The dust accumulation detection threshold range is dynamically adjusted based on salt spray concentration data. A dual-range detection mode is activated when the salt spray concentration is ≥5 mg / m³. In the conventional mode, a single threshold is used to determine the dust accumulation area; in the salt crystallization mode, morphological filtering is added before threshold detection. The morphological filtering uses a combination of opening and closing operations, matching the particle distribution characteristics of salt crystals with circular structural elements to suppress high-frequency noise from salt spray particles while preserving low-frequency, slowly varying signals in the dust accumulation area. The detection threshold in the salt crystallization mode is 15% higher than in the conventional mode, reducing the probability of misidentifying salt spray particles as dust accumulation.

[0069] The sensitivity parameters for hot spot detection are optimized based on illumination intensity data, with the sensitivity exhibiting a linear weighted relationship with illumination intensity. For every 100 W / m² decrease in illumination intensity, the sensitivity parameter is adjusted downwards by 10%. When the illumination intensity is <500 W / m², the detection trigger frequency is reduced to 70% of the baseline value, and the signal-to-noise ratio is improved in low-light environments by extending the signal integration time. The sensitivity adjustment parameters are stored in a configuration file, and the corresponding detection strategy is dynamically loaded based on real-time illumination intensity to balance detection speed and accuracy requirements.

[0070] The above steps eliminate the influence of environmental thermal effects on reflectivity through a temperature compensation model, suppress misjudgments in high-salt-fog environments through salt spray mode switching, and optimize detection stability in low-light scenes through light sensitivity adjustment. Temperature normalized data provides standardized input for dust accumulation judgment, salt crystallization filtering enhances the anti-interference ability of threshold judgment, and dynamic sensitivity adjustment maintains the reliability of hot spot detection. Each correction model is linked in real time based on environmental parameters from multi-source datasets, forming a technical closed loop of "environmental perception → parameter correction → detection optimization," improving the robustness of the anomaly recognition system in dynamic environments.

[0071] Step 5: Input the multidimensional abnormal feature vector into the preset fault model database for similarity matching, trigger the hierarchical alarm mechanism based on the matching result, and verify the temporal continuity of the multidimensional abnormal feature vector through the decision tree classifier to eliminate instantaneous environmental interference signals.

[0072] The technical solution in step 5 achieves accurate classification and interference elimination of abnormal signals through a hierarchical judgment and dynamic verification mechanism. The specific implementation process is as follows:

[0073] Multidimensional anomaly feature vectors are input into a pre-defined fault model database for similarity matching. This database stores feature vector templates and confidence thresholds for fault types such as cracks, hot spots, and dust accumulation. The matching process uses a cosine similarity algorithm to calculate the similarity value between the real-time feature vector and the template. When the similarity exceeds 85%, it is determined to be a high-confidence fault. The database dynamically updates the template library based on historical detection data and laboratory calibration results. New fault type feature vectors are imported after expert verification to improve the model's coverage and adaptability.

[0074] A tiered alarm mechanism is triggered based on the matching degree results. A Level 1 alarm is triggered when the matching degree is >85%, locking key parameters in the abnormal feature vector as the evidence chain and freezing environmental parameter updates to maintain the stability of the current detection model. A Level 2 alarm is triggered when the matching degree is between 60% and 85%, generating a re-inspection command to drive the UAV to perform a spiral path re-inspection of the abnormal area. The re-inspection data is then updated with fault coordinates after secondary verification by a preset data fusion model. A Level 3 alarm is triggered when the salt spray concentration is >5 mg / m³ or the light intensity is <500 W / m², initiating a backup sensor calibration process and adjusting the laser pulse frequency to a safe mode to reduce the impact of environmental interference on the detection results.

[0075] The decision tree classifier verifies the temporal continuity of multidimensional anomaly feature vectors. The root node calculates the environmental interference level based on salt spray concentration, temperature gradient, and light intensity. When the interference level is ≥4, an anti-interference detection branch is activated. Intermediate nodes employ a sliding time window mechanism with a window length of 2 seconds, statistically analyzing the continuous percentage of anomaly feature values ​​within the window. When the percentage of consecutively sampled anomalies is ≥80%, it is considered a valid anomaly; otherwise, it is marked as transient interference. Leaf nodes match the valid anomaly feature vectors with a standardized fault coding library, generating a diagnostic report including three-dimensional coordinates (X, Y, Z), fault type code, and a confidence level ≥90%. This report is transmitted to the maintenance platform via an encrypted protocol to trigger a maintenance response.

[0076] The above steps utilize a three-tiered logical chain—fault template matching, tiered alarm triggering, and time continuity verification—to classify and mitigate interference from abnormal signals. Dynamic database updates enhance model generalization capabilities, the tiered alarm mechanism adapts to different confidence levels, and the decision tree verification layer filters transient noise and solidifies valid anomaly evidence. Each step, based on multi-dimensional data input and dynamic parameter adjustments, forms a closed-loop technology of "feature matching → alarm response → interference filtering," ultimately outputting highly reliable fault diagnosis results to support operational and maintenance decisions.

[0077] Step 6: Generate an optimized detection path based on the anomaly localization results. The path includes dense scanning area division based on multi-dimensional anomaly feature vectors, multi-UAV collaborative task allocation, and dynamic trajectory correction based on real-time environmental response. The path execution data is then fed back to a preset data fusion model to update the temperature and reflection intensity compensation model parameters, forming a closed-loop control logic from data acquisition to diagnostic decision-making.

[0078] This invention provides a method for anomaly identification of photovoltaic modules based on laser detection. The technical solution includes a closed-loop control process encompassing data acquisition, equipment regulation, feature extraction, model correction, alarm decision-making, and path optimization. The specific steps are as follows:

[0079] Step 1 (Data Acquisition and Synchronization): A hardware synchronization trigger uses the laser emission pulse signal as a reference clock source to generate a synchronization clock signal. The temperature and humidity sensor and salt spray sensor trigger hardware interrupts based on this synchronization signal, synchronously acquiring temperature, humidity, and salt spray concentration data with the laser scanning cycle. The illuminance sensor integrates the laser echo signal through an integrating circuit to generate illuminance waveform data aligned with the synchronization clock signal. The above data, aligned by timestamps, is integrated into a multi-source dataset and input to the laser emitter control unit and a preset data fusion model. The multi-source dataset is further input to the particle deposition rate calculation module, which calculates the salt crystallization deposition rate based on the salt spray concentration data, outputting the salt crystallization development trend and environmental interference level, providing basic data for subsequent environmental interference analysis.

[0080] Step 2 (Dynamic Adjustment of Laser and Filter): Based on historical salt fog concentration data from a multi-source dataset, a sliding window mechanism is used to dynamically update the salt fog concentration threshold boundary. This threshold is used to determine whether to trigger laser wavelength switching. When the salt fog concentration exceeds the preset threshold, the laser emitter switches to the near-infrared band and simultaneously adjusts the pulse width to a preset penetration optimization range. The receiver suppresses ambient light noise through a tunable filter, whose passband range is dynamically adjusted according to the illumination intensity data. Simultaneously, the scattering noise estimator inverts the aerosol particle size distribution based on Mie scattering theory and compares the backscattering coefficient differences between adjacent scanning points using differential detection technology. When the difference exceeds a set threshold, scattering interference signals are eliminated to ensure the purity of the reflected signal.

[0081] Step 3 (Multidimensional Feature Extraction and Fusion): The reflection intensity signal is decomposed to the fifth level of detail coefficients using the Db4 wavelet basis. A sliding energy window is set in the 200-400Hz frequency band. When the energy fluctuation within the window exceeds a preset threshold, it is marked as a crack feature corresponding to a sudden change in reflectivity. Multispectral reflectance data is analyzed using the etalon interferometry method to lock the characteristic peaks corresponding to the bandgap width of the photovoltaic material and establish a linear mapping relationship between hot spot temperature and characteristic peak wavelength offset. The pitch and roll angles in the UAV attitude data are used as input coordinate transformation matrices. After correcting the spatial distortion of the laser point cloud data, the second derivative of reflectivity is calculated using the Laplacian operator. When the derivative value exceeds a set range, it is marked as a dust gradient distribution area. After fusing the temperature gradient data, a multidimensional abnormal feature vector is generated, providing multidimensional input for subsequent fault determination.

[0082] Step 4 (Threshold Correction and Sensitivity Optimization): Based on temperature and humidity data from the multi-source dataset, the temperature range is divided into five segments from -20℃ to 80℃. Each segment is associated with an independent reflectivity baseline value for normalizing reflectivity intensity data. A dual-range dust accumulation judgment mode is activated based on salt spray concentration data: when the salt spray concentration is ≥5mg / m³, the mode switches to salt crystallization mode, adding morphological filtering to distinguish between salt crystallization and regular dust accumulation. Hot spot detection sensitivity is linearly weighted and correlated with light intensity; when the light intensity is <500W / m², the detection trigger frequency is reduced to 70% of the baseline value to avoid false triggering under low light conditions.

[0083] Step 5 (Fault Matching and Hierarchical Alarm): Multi-dimensional anomaly feature vectors are input into a preset fault model database for similarity matching. A Level 1 alarm is triggered when the matching degree exceeds 85%, locking the fault feature evidence chain and freezing environmental parameter updates. A Level 2 alarm is triggered when the matching degree is between 60% and 85%, driving the drone to re-inspect the abnormal area within 3 minutes and inputting the coordinate data into the path planning engine. A Level 3 alarm is triggered when the salt spray concentration is >5 mg / m³ or the light intensity is <500 W / m², initiating the backup sensor verification process and adjusting the laser pulse frequency to safe mode. In the decision tree classifier, the root node calculates the environmental interference level based on salt spray concentration, temperature gradient, and light intensity (an anti-interference branch is enabled when the level is ≥4). Intermediate nodes verify the temporal continuity of the abnormal signal (a valid anomaly is determined when the signal lasts ≥2 seconds). Leaf nodes match the standardized fault coding library to generate a diagnostic report.

[0084] Step 6 (Path Optimization and Closed-Loop Feedback): Based on the abnormal positioning coordinates, a spiral expansion scanning path with a pitch increasing by 0.5m is generated, and the scanning line spacing increases from 10cm to 20cm with the number of scans. Multi-UAV collaborative task allocation is achieved through an improved contract network protocol, dividing sub-regions based on remaining battery power and sensor health status scores. Real-time environmental monitoring data triggers dynamic trajectory correction: when wind speed > 8m / s, a 15° yaw compensation angle is inserted, the spacing between hovering acquisition points is increased to 1.5 times the standard density, and a Bezier curve interpolation algorithm is used to smooth the trajectory. Path execution data is fed back to the preset data fusion model to update the temperature-reflection intensity compensation model parameters, forming a closed-loop control logic of data acquisition → diagnosis → path correction.

[0085] The multi-source dataset generated in step 1 provides data input for subsequent steps. Steps 2-S4 improve detection accuracy by dynamically adjusting laser parameters, filters, and threshold models. Step 5 verifies the effectiveness of anomalies by combining fault models and decision trees. Step 6 optimizes the detection path based on anomaly localization and provides feedback to correct model parameters. Each step is tightly coupled with the data flow and control flow, ultimately constructing an environment-adaptive, high-precision anomaly identification and closed-loop control system.

[0086] Specifically, in the aforementioned laser-based photovoltaic module anomaly identification method, the generation of a multi-source dataset containing environmental variables includes:

[0087] A synchronous clock signal is generated by using the laser emission pulse signal as a reference clock source through a hardware synchronous trigger.

[0088] The temperature and humidity sensor and the salt spray sensor trigger a hardware interrupt based on the synchronous clock signal to collect temperature and humidity data and salt spray concentration data synchronously with the laser scanning cycle.

[0089] The light intensity sensor integrates the laser echo signal through an integrating circuit to generate illuminance waveform data aligned with the synchronous clock signal.

[0090] The temperature and humidity data, salt spray concentration data, and illuminance waveform data are aligned by timestamp and integrated into a multi-source dataset. The multi-source dataset is then input into the particle deposition rate calculation module. Based on the salt spray concentration data, the amount of salt crystallization deposition per unit time is calculated, and the salt crystallization formation trend and environmental interference level are output.

[0091] The technical solution for generating multi-source datasets containing environmental variables described in this invention achieves accurate capture of environmental parameters through multi-sensor synchronous acquisition and data fusion mechanisms. The specific implementation process is as follows:

[0092] A hardware synchronization trigger uses the laser emission pulse signal as a reference clock source to generate a clock signal that is strictly synchronized with the laser scanning cycle. This synchronization mechanism uses the pulse signal output from the laser driver circuit to trigger the hardware timing unit, ensuring that the acquisition time of each sensor is aligned with the timing of the laser scanning action, thus eliminating data misalignment problems caused by timing deviations.

[0093] The temperature and humidity sensor and the salt spray sensor trigger a hardware interrupt based on the synchronized clock signal, initiating data acquisition at the start of the laser scanning cycle. The temperature and humidity sensor uses a capacitive sensing element to capture changes in the dielectric constant caused by ambient humidity, while the salt spray sensor detects the concentration of sodium ions in the air through an electrochemical electrode. The data collected by both types of sensors is transmitted to a data buffer queue via a serial communication interface, maintaining a synchronized sampling frequency with the laser scanning cycle to form a time-coherent sequence of environmental parameters.

[0094] The illuminance sensor integrates the laser echo signal using an integrating circuit to generate illuminance waveform data aligned with a synchronous clock signal. The integrating circuit employs an operational amplifier and an RC filter network to smooth transient noise in the laser echo signal and extract the DC component characterizing ambient light intensity. The illuminance waveform data is converted into a digital signal via an analog-to-digital converter, and its sampling time is controlled by the synchronous clock signal, aligning with the timestamps of temperature, humidity, and salt spray data.

[0095] The temperature and humidity data, salt spray concentration data, and illuminance waveform data are aligned by timestamp and integrated into a multi-source dataset. During data alignment, an interpolation algorithm is used to compensate for differences in sampling intervals between different sensors, generating an environmental parameter matrix with a unified time baseline. This multi-source dataset is input into a particle deposition rate calculation module, which calculates the salt crystallization deposition rate per unit area based on the time series of salt spray concentration data. Combined with temperature and humidity parameters, the salt crystallization formation trend is deduced, and a quantitative environmental interference level parameter is output. This environmental interference level parameter provides a dynamic correction basis for subsequent laser parameter adjustments and fault diagnosis.

[0096] The above steps achieve temporal consistency of multi-source data through hardware synchronization triggering, capture the spatial distribution characteristics of environmental parameters through sensor-specific acquisition mechanisms, and finally establish a dynamic assessment model of environmental interference through data fusion and calculation modules. These steps form a progressive logical chain of "time synchronization → data acquisition → signal processing → data fusion → interference assessment," providing reliable environmental parameter inputs for subsequent anomaly identification.

[0097] Specifically, in the aforementioned laser-based photovoltaic module anomaly identification method, step 2 includes:

[0098] Based on the historical salt spray concentration data in the multi-source dataset, a sliding window mechanism is used to dynamically update the salt spray concentration threshold boundary, which is used to determine whether laser wavelength switching is triggered.

[0099] When the laser emitter switches to the near-infrared band, the laser pulse width is simultaneously adjusted to the preset penetration optimization range;

[0100] The receiver scattering noise estimator performs inversion calculations on the aerosol particle size distribution based on Mie scattering theory, and compares the backscattering coefficient differences between adjacent scanning points using differential detection technology. When the difference exceeds a set threshold, the corresponding scattering interference signal is eliminated.

[0101] Step 2 improves detection accuracy in complex environments by dynamically adjusting laser parameters and noise suppression strategies. The specific implementation process is as follows:

[0102] Based on historical salt spray concentration data from a multi-source dataset, a sliding window mechanism is used to dynamically update the salt spray concentration threshold boundary. The sliding window updates the salt spray concentration statistics every 10 minutes, calculates the mean and variance of the salt spray concentration within the window, and dynamically corrects the threshold boundary range by combining environmental temperature and humidity parameters. The threshold boundary serves as the trigger condition for laser wavelength switching; when the real-time salt spray concentration exceeds the upper limit of the threshold, a wavelength switching command is sent to the laser transmitter control unit.

[0103] After receiving a wavelength switching command, the laser emitter switches its output wavelength from the visible light band to the near-infrared band, while simultaneously adjusting the laser pulse width to a preset optimized penetration range. The pulse width adjustment module matches the pulse duration based on the transmission characteristics of the near-infrared band, optimizing the penetration efficiency of laser energy in a salt spray environment. The wavelength switching operation and pulse parameter adjustment are controlled collaboratively by a digital signal processor to reduce signal distortion during mode switching.

[0104] The receiver-side scattering noise estimator performs inversion calculations on the aerosol particle size distribution based on Mie scattering theory, establishing a correlation model between aerosol particle size and scattering intensity. Differential detection technology compares the differences in backscattering coefficients between adjacent scanning points; when the difference in scattering coefficients between adjacent points exceeds a preset threshold, it is identified as a scattering interference signal and filtered out. The differential detection technology employs a dual-channel signal processor to calculate the spatial gradient change of the scattering signal in real time, and combines this with the aerosol particle size distribution model to distinguish between environmental noise and effective reflected signals.

[0105] The above steps control laser wavelength switching through dynamic threshold control of salt spray concentration, improve near-infrared detection efficiency through pulse parameter optimization, and suppress environmental interference through scattering noise modeling and differential detection techniques. Each step achieves coordinated control based on environmental parameters from multi-source datasets, forming a closed-loop technology of "environmental perception → equipment optimization → noise suppression," providing a high signal-to-noise ratio laser reflection signal for subsequent feature extraction.

[0106] Specifically, in the laser-based photovoltaic module anomaly identification method, step 3 includes: decomposing the reflection intensity signal to the fifth level of detail coefficients using the Db4 wavelet basis, setting a sliding energy window in the 200-400Hz frequency band, and marking the energy fluctuation within the window as a crack feature corresponding to a sudden change in reflectivity when the energy fluctuation exceeds a preset threshold.

[0107] By analyzing multispectral reflectance data using the etalon interferometry method, the characteristic peaks corresponding to the bandgap of photovoltaic materials are identified, and a hot spot temperature mapping table is established based on the linear relationship between the characteristic peak wavelength shift and temperature change.

[0108] The pitch and roll angles from the UAV attitude data are input into the coordinate transformation matrix. After correcting the spatial distortion of the laser point cloud data, the second derivative of reflectivity is calculated using the Laplacian operator. When the derivative value exceeds the set range, it is marked as a dust gradient distribution area.

[0109] Step 3 extracts surface anomaly features of the photovoltaic module using multidimensional signal processing and spatial correction methods. The specific implementation process is as follows:

[0110] The reflected intensity signal was decomposed down to the fifth level of detail coefficients using the Db4 wavelet basis, and multi-scale analysis was used to focus on the energy distribution characteristics in the 200-400Hz frequency band. A sliding energy window was set in this frequency band, and the standard deviation of the signal energy within the window was calculated as a fluctuation index. When the fluctuation value exceeded a preset threshold, it was marked as a region of abrupt change in reflectivity. The abrupt change in reflectivity is related to the change in local light scattering characteristics caused by microcracks on the photovoltaic panel surface, and the initial location of the crack is achieved through the frequency domain energy distribution characteristics.

[0111] Multispectral reflectance data was analyzed using etalon interferometry, and the reflectance signals of different wavelengths were separated using the Fabry-Perot interferometry principle to pinpoint the characteristic peaks corresponding to the bandgap of the photovoltaic material. The wavelength shift of the characteristic peaks showed a linear relationship with the module temperature change. A temperature-wavelength shift mapping table was established through calibration experiments, converting the real-time detected wavelength shift into hotspot temperature values. The mapping table is stored in a database, supporting quantitative assessment and dynamic tracking of the hotspot region temperature.

[0112] The pitch and roll angles from the UAV attitude data are input into a coordinate transformation matrix to calculate the actual coordinates of the laser-scanned point cloud in three-dimensional space. Coordinate transformation eliminates point cloud distortion caused by changes in UAV flight attitude, reconstructing the spatial reflectivity distribution of the photovoltaic panel surface. The Laplacian operator is applied to the corrected reflectivity data to calculate the second-order derivative, extracting the spatial variation characteristics of the reflectivity gradient. When the derivative value exceeds a set range, it is identified as a dust accumulation gradient distribution region. This method enhances the gradual reflectivity variation characteristics caused by dust accumulation through spatial differential operations, improving the sensitivity of dust accumulation detection.

[0113] The above steps extract anomalous features from the time-frequency domain, spectral domain, and spatial domain, respectively: wavelet decomposition identifies cracks corresponding to high-frequency abrupt signals; etalon interferometry quantifies hot spot temperature through wavelength shift; and UAV attitude correction and differential operations locate dust accumulation gradients. The multi-dimensional feature extraction results are fused to generate anomaly feature vectors, providing multi-dimensional data support for subsequent fault classification. Through the synergistic effect of signal processing, optical analysis, and spatial modeling techniques, each step achieves comprehensive capture and feature analysis of surface anomalies in photovoltaic modules.

[0114] Specifically, the photovoltaic module anomaly identification method based on laser detection includes step 4: based on the temperature and humidity data in the multi-source dataset, dividing the temperature range into five segments from -20℃ to 80℃, with each segment associated with an independent reflectivity baseline value, which is used to normalize the reflectivity intensity data.

[0115] The dual-range dust accumulation determination mode is activated based on the salt spray concentration data in the multi-source dataset. When the salt spray concentration is ≥5mg / m³, the mode is switched to salt crystallization mode, and morphological filtering is added to distinguish between salt crystallization and regular dust accumulation.

[0116] Based on the light intensity data in the multi-source dataset, the hot spot detection sensitivity is linearly weighted and correlated with the light intensity. When the light intensity is <500W / m², the detection trigger frequency is reduced to 70% of the baseline value.

[0117] Step 4 involves dynamically adjusting the detection threshold and sensitivity based on environmental parameters to improve the environmental adaptability of anomaly detection. The specific implementation process is as follows:

[0118] Based on temperature and humidity data from a multi-source dataset, the temperature range is divided into five segments from -20℃ to 80℃, each spanning 20℃, and associated with independent reflectivity baseline values. These baseline values ​​are determined by statistically averaging the reflectivity of photovoltaic panel surfaces at different temperatures, calibrated in the laboratory. These baseline values ​​are used to normalize the real-time reflectivity data, eliminating reflectivity drift caused by temperature changes. The normalized reflectivity data is then input into the dust accumulation detection module to reduce the impact of temperature interference on dust accumulation detection.

[0119] The dual-range dust accumulation detection mode is activated based on salt spray concentration data from a multi-source dataset: when the salt spray concentration is ≥5mg / m³, the mode switches to salt crystallization mode, adding morphological filtering to the conventional dust accumulation detection algorithm. Morphological filtering employs a combination of opening and closing operations, matching the particle distribution characteristics of salt crystals through structuring elements to suppress high-frequency noise in the salt crystallization region and enhance the distinction between dust accumulation and salt crystals. The detection threshold in salt crystallization mode is 15% higher than in the conventional mode, reducing the probability of misclassifying salt spray particles as dust accumulation.

[0120] Based on illumination intensity data from a multi-source dataset, the hotspot detection sensitivity is linearly weighted and correlated with illumination intensity. For every 100 W / m² decrease in illumination intensity, the detection sensitivity is reduced by 10%. When the illumination intensity is <500 W / m², the detection trigger frequency is reduced to 70% of the baseline value. This adjustment is achieved by dynamically adjusting the signal sampling interval, extending the integration time under low illumination conditions to improve the signal-to-noise ratio, while avoiding false triggers caused by signal fluctuations. The linkage logic between the sensitivity parameter and the trigger frequency is stored in a configuration file, supporting automatic switching of the detection strategy according to ambient illumination conditions.

[0121] The above steps achieve adaptive fusion of multi-dimensional environmental parameters through temperature segmentation compensation, salt spray mode switching, and light sensitivity adjustment. The temperature compensation model provides standardized data input for dust accumulation and hot spot detection, the salt crystallization mode suppresses misjudgments in high salt spray environments, and the dynamic adjustment of light sensitivity optimizes detection stability in low-light scenes. Each correction model is updated in real time based on environmental parameters from multi-source datasets, forming a technical closed loop of "environmental perception → threshold correction → detection optimization," thereby improving the robustness of anomaly identification under complex working conditions.

[0122] Specifically, in the laser detection-based photovoltaic module anomaly identification method, step 5 includes: when the matching degree between the multidimensional anomaly feature vector and the preset fault model exceeds 85%, triggering a level one alarm, locking the fault feature evidence chain in the multidimensional anomaly feature vector, and freezing the environmental parameter update operation in the preset data fusion model.

[0123] When the matching degree is between 60% and 85%, a level 2 alarm is triggered, a re-inspection command is generated to drive the drone to re-inspect the abnormal area within 3 minutes, and the coordinate data of the re-inspection area is input into the path planning engine;

[0124] When the salt spray concentration in the multi-source dataset exceeds 5 mg / m³ or the light intensity is below 500 W / m², a level three alarm is triggered, the verification process of the backup sensor is initiated, and the pulse frequency in the laser scanning parameters is adjusted to a safe mode.

[0125] Step 5 achieves dynamic adaptation of abnormal responses through a tiered alarm mechanism. The specific implementation process is as follows:

[0126] When the matching degree between the multidimensional anomaly feature vector and the preset fault model exceeds 85%, a Level 1 alarm is triggered. At this time, the system locks the feature evidence chain in the multidimensional anomaly feature vector that is strongly correlated with the fault type. The evidence chain includes key parameters such as crack frequency domain energy distribution, hot spot wavelength offset, and dust accumulation gradient value, and stores them in real time to non-volatile memory. At the same time, the environmental parameter update operation in the preset data fusion model is frozen, and the dynamic correction of environmental variables such as temperature, humidity, and salt spray concentration is suspended to maintain the stability of the current detection model and prioritize the handling of high-confidence faults.

[0127] When the matching degree is between 60% and 85%, a level two alarm is triggered. The system generates a re-inspection command, including the coordinates of the abnormal area, a preliminary fault type classification, and a confidence score, and sends it to the UAV control unit via a wireless communication protocol. Within 3 minutes, the UAV performs a re-inspection task on the abnormal area using a spiral expansion scanning path. After the re-inspection data is verified a second time by a preset data fusion model, the updated abnormal coordinates are input into the path planning engine. Based on the re-inspection results, the path planning engine re-divides the dense scanning area and optimizes the coverage and accuracy of subsequent detection trajectories.

[0128] A level-three alarm is triggered when the salt spray concentration in the multi-source dataset exceeds 5 mg / m³ or the light intensity is below 500 W / m². The system initiates the backup sensor calibration process, switches to the redundant sensor channel to collect environmental parameters, and verifies the reliability of the measured values ​​by cross-comparing the primary and backup sensor data. Simultaneously, the pulse frequency in the laser scanning parameters is adjusted to a safe mode, reducing the pulse frequency to 50% of the standard value to minimize the impact of laser scattering noise on signal quality under high salt spray or low light conditions. In safe mode, the system extends the integration time of a single scan to improve the signal-to-noise ratio and prioritizes the use of anti-interference feature extraction algorithms to process reflected signals.

[0129] The above steps achieve hierarchical control of alarm response through confidence level classification and environmental limit violation criteria: Level 1 alarms focus on rapid location and evidence consolidation of high-confidence faults; Level 2 alarms improve the accuracy of judging medium-confidence anomalies through a re-examination mechanism; and Level 3 alarms activate redundant verification and parameter protection strategies to address environmental interference. Each alarm level triggers corresponding closed-loop response logic based on multi-dimensional data input, forming a linkage system of "fault judgment → response strategy → path correction," ultimately achieving a dynamic balance between anomaly identification and system self-protection in complex environments.

[0130] Specifically, in the photovoltaic module anomaly identification method based on laser detection, step 6 includes: generating a spiral expansion scanning path with a pitch increasing by 0.5m centered on the anomaly point based on the anomaly positioning coordinates in the multidimensional anomaly feature vector, and increasing the scanning line spacing from 10cm to 20cm with the number of scans.

[0131] Based on the remaining battery power data and sensor health status scores of the drones in the multi-source dataset, the detection area is divided into multiple sub-regions using an improved contract net protocol, and assigned to multiple drones to perform collaborative tasks.

[0132] When the wind speed in the real-time environmental monitoring data exceeds 8 m / s, a 15° yaw compensation angle is inserted into the flight trajectory, the spacing between hovering acquisition points is increased to 1.5 times the standard density, and the modified trajectory is smoothed using a Bezier curve interpolation algorithm.

[0133] Step 6 achieves efficient execution of anomaly detection and environmental adaptation through dynamic path planning and multi-machine collaborative control. The specific implementation process is as follows:

[0134] Based on the anomaly location coordinates in the multidimensional anomaly feature vector, a spiral expansion scanning path with a pitch increasing by 0.5m is generated centered on the anomaly point. The initial pitch of the spiral path is set to 0.5m, and the spacing between scan lines increases from 10cm to 20cm with the number of scans. The interval between adjacent scan circles is dynamically adjusted using a linear interpolation algorithm. This path design gradually expands the detection range to cover potential risk points around the anomaly area while avoiding resource waste caused by repeated scans. The spiral path generation logic is embedded in the UAV navigation control unit, supporting real-time coordinate transformation and waypoint updates.

[0135] Based on the remaining battery power data and sensor health status scores of drones from multi-source datasets, the detection area is divided into multiple sub-regions using an improved contract network protocol. Drones with remaining battery power above 80% are preferentially assigned to remote sub-regions, while drones with sensor health scores below a threshold only perform low-precision scanning tasks. The improved contract network protocol introduces a dynamic adjustment mechanism for weighting factors. The master drone publishes bidding information based on the sub-region area, environmental interference level, and task priority. Subordinate drones calculate their bidding weights based on their own status and feed them back to the master, ultimately forming a task allocation scheme that balances efficiency and reliability.

[0136] When the wind speed in the real-time environmental monitoring data exceeds 8 m / s, the flight control system inserts a 15° yaw compensation angle between preset waypoints to correct the UAV's heading deviation caused by wind resistance. The spacing between hovering data collection points is increased to 1.5 times the standard density to compensate for positioning errors caused by wind speed by increasing the number of sampling points. The trajectory correction data is input into a Bézier curve interpolation algorithm to generate a smooth transition path using the current waypoint and the target waypoint as control vertices, eliminating trajectory jitter caused by sharp turns. The corrected trajectory data is synchronously updated to other cooperating UAVs to maintain the spatiotemporal consistency of the multi-UAV scanning paths.

[0137] The above steps achieve refined coverage of abnormal areas through spiral path expansion, optimize multi-machine resource allocation through contract network protocol, and improve detection stability under complex weather conditions through environmental response trajectory correction. Each step uses anomaly location data as input and path execution data as feedback to a preset data fusion model to form closed-loop control, ultimately constructing a complete technology chain from anomaly identification to dynamic path optimization.

[0138] Specifically, the laser-based photovoltaic module anomaly identification method further includes: the master drone generating bidding information including the topological connection relationship of the abnormal area based on the abnormal positioning coordinates, and broadcasting it to the slave drones via a wireless mesh network; the slave drones calculating bidding weights based on remaining power, flight stability error values, and sensor accuracy deviation values, wherein the bidding weights are positively correlated with the remaining power and negatively correlated with stability error and accuracy deviation; based on the bidding weights, the sub-region to be detected is divided into topological association units and environmental parameter similarity units, wherein the topological association units are divided according to the physical connection relationship of the abnormal components, and the environmental parameter similarity units are divided according to areas with salt spray concentration differences ≤15% and temperature gradient differences ≤5℃.

[0139] The multi-machine collaborative task allocation technical solution of this invention optimizes detection efficiency through dynamic weight calculation and region partitioning strategies. The specific implementation process is as follows:

[0140] The master drone analyzes the physical connection diagram of the photovoltaic array based on the anomaly location coordinates, generating bidding information including the topology of the anomaly area. The topology connection diagram is constructed using the series and parallel electrical parameters of the photovoltaic modules and their installation location coordinates, identifying the relationship between the anomaly module and its adjacent modules. The bidding information, after data encapsulation, is broadcast to slave drones via a wireless mesh network. The encapsulated content includes the coordinate range of the sub-region, detection accuracy requirements, and task priority parameters, supporting multi-hop relay transmission to expand communication coverage.

[0141] After receiving the bidding information, the subordinate drones calculate the bidding weights based on the remaining battery percentage, flight stability error value, and sensor accuracy deviation value. The remaining battery power is positively weighted linearly, the flight stability error value is calculated using the standard deviation of gyroscope and accelerometer data as a negative weighting factor, and the sensor accuracy deviation value is negatively vectorized based on historical detection error rates. The output of the weight calculation formula is normalized to a 0-1 range and fed back to the master drone via a wireless mesh network, forming a capability evaluation matrix for each drone.

[0142] The master-controlled UAV divides the sub-region to be detected into topologically related units and environmentally similar units based on bidding weights. Topologically related units are defined according to the physical connections of anomalous components, prioritizing the allocation of high-weight UAVs to areas directly electrically connected to anomalous components. Environmentally similar units are divided according to standard clustering based on salt spray concentration differences ≤15% and temperature gradient differences ≤5℃, merging areas with similar salt spray and temperature distributions into the same detection unit, assigning UAVs with low sensor accuracy deviations to perform high-precision scanning. The division results generate a detection waypoint sequence for each UAV through a path planning engine and are synchronously updated to the collaborative control platform.

[0143] The above steps divide the detection area using both topological association and environmental similarity dimensions, and dynamically allocate tasks based on the UAV's status. The master control unit's topology analysis and bidding mechanism achieves targeted coverage of abnormally associated areas, while the slave control unit's weight calculation quantifies differences in equipment capabilities. Finally, a partitioning strategy balances detection efficiency and accuracy. Each step is dynamically adjusted based on real-time data, forming a technical closed loop of "topology modeling → capability assessment → intelligent partitioning," providing a scalable task allocation framework for multi-UAV collaborative detection.

[0144] Specifically, the laser-based photovoltaic module anomaly identification method also includes a closed-loop feedback mechanism:

[0145] The closed-loop feedback mechanism includes a forward control chain and a reverse calibration process;

[0146] The forward control chain inputs environmental monitoring data from the multi-source dataset into the device parameter optimization module every 30 seconds to generate an adjustment suggestion table for laser output power and UAV flight speed, and increases the laser output power to 80% of the increase in UAV flight speed according to the suggestion table;

[0147] The reverse calibration process cross-validates the detection data of the same area using three drones, calculates the average deviation of the sensor measurements as the drift compensation, and updates the exponential parameters in the battery degradation coefficient model based on the battery consumption rate (unit: mAh / km) and trajectory positioning error (unit: cm) during path execution.

[0148] The closed-loop feedback mechanism described in this invention improves system stability and detection accuracy through forward parameter optimization and reverse data calibration. The specific implementation process is as follows:

[0149] The forward control chain inputs environmental monitoring data from the multi-source dataset into the device parameter optimization module every 30 seconds. This module generates an adjustment suggestion table for laser output power and UAV flight speed based on parameters such as salt spray concentration, light intensity, and temperature gradient. The laser output power adjustment strategy is negatively correlated with UAV flight speed; as flight speed increases, the power is increased by 80% of the speed increase to balance signal acquisition quality at high speeds. The adjustment parameters are sent in real-time to the laser drive circuit and UAV ESC module via a digital signal processor, supporting dynamic response to the impact of environmental changes on detection performance.

[0150] The reverse calibration process involves three UAVs simultaneously scanning the same area to collect multispectral reflectance data and laser point cloud coordinates. The three sets of data are input into a cross-validation algorithm to calculate the mean and variance of the reflectance intensity at the same detection point. When the variance exceeds a preset threshold, it is marked as a sensor drift anomaly. The drift compensation is calculated using a sliding window to determine the average deviation of the sensor measurements. This deviation value is input into the calibration database to drive a tunable filter and attitude controller for parameter compensation. Simultaneously, based on the battery consumption rate (mAh / km) and trajectory positioning error (cm) in the path execution data, the exponential parameters in the battery degradation coefficient model are updated. These exponential parameters are used to predict the remaining flight time of the UAV and dynamically adjust the battery weight coefficient in the task allocation strategy.

[0151] The forward control chain and reverse calibration process are interconnected via a data bus, forming a bidirectional feedback loop. Forward optimization parameters influence the detection path and equipment status in real time, while reverse calibration results correct sensor drift and battery model errors. Environmental data, equipment parameters, and calibration values ​​are periodically integrated through a preset data fusion model to generate an updated set of control commands, driving the laser, drone, and sensors into the next detection cycle. This mechanism continuously iterates and optimizes to reduce the impact of environmental interference and equipment degradation on the detection results, maintaining the system's robustness and reliability under all operating conditions.

[0152] In the above steps, the forward control chain focuses on the dynamic adaptation of equipment parameters, while the reverse calibration process emphasizes data reliability and accurate modeling of equipment status. Through data interaction and command collaboration, they construct a closed-loop logic, realizing a complete feedback loop of "environmental perception → parameter adjustment → data verification → model update," ultimately forming an adaptive anomaly identification and equipment management system.

[0153] Specifically, in the laser-based photovoltaic module anomaly identification method, step 5 includes: the root node calculates the environmental interference level based on the salt spray concentration, temperature gradient, and light intensity data in the multi-source dataset; when the interference level is ≥4, an anti-interference detection branch is selected; the intermediate node verifies the temporal continuity of the multidimensional anomaly feature vector; when the anomaly signal lasts for ≥2 seconds, it is determined to be a valid anomaly, otherwise it is marked as transient interference; the leaf node matches the multidimensional anomaly feature vector with a standardized fault type coding library to generate a diagnostic report including fault coordinates (X,Y,Z), type coding, and confidence level ≥90%.

[0154] The technical solution in step 5 achieves accurate identification and classification of abnormal signals through the hierarchical decision logic of the decision tree classifier. The specific implementation process is as follows:

[0155] The root node calculates the environmental interference level based on salt spray concentration, temperature gradient, and light intensity data from a multi-source dataset. Salt spray concentration is obtained using an electrochemical sensor, temperature gradient is calculated from the difference between an infrared thermal imager and a temperature sensor, and light intensity is quantified using a photosensitive element. The interference level uses a weighted scoring model, with salt spray concentration accounting for 50%, and temperature gradient and light intensity each accounting for 25%. The scoring results are divided into levels 1-5. When the interference level is ≥4, an anti-interference detection branch is activated. This branch uses a noise reduction algorithm to increase the wavelet decomposition level of the reflected signal and expands the passband range of the tunable filter to suppress high-frequency noise.

[0156] Intermediate nodes receive detection branch instructions from the root node to verify the temporal continuity of the multidimensional anomaly feature vector. Temporal continuity analysis employs a sliding time window mechanism with a window length of 2 seconds. Within the window, energy values, wavelength offsets, and gradient derivatives of the anomaly feature vector are sampled at 10ms intervals. When the anomaly feature values ​​of consecutive sampling points all exceed a threshold, it is determined to be a valid anomaly; if the cumulative proportion of anomaly signals within the window is less than 80%, it is marked as transient interference. Transient interference data is stored in a buffer queue for subsequent statistical analysis, while valid anomaly data is input to leaf nodes for fault classification.

[0157] Leaf nodes match valid anomaly feature vectors with a standardized fault type coding library. This library stores feature vector templates and confidence thresholds for fault types such as cracks, hot spots, and dust accumulation. The matching process uses a cosine similarity algorithm to calculate the similarity between real-time feature vectors and templates. A diagnostic report is generated when the similarity is ≥90%. The report includes the three-dimensional coordinates (X, Y, Z) of the fault point, the fault type code, and the confidence value. The coordinate data is obtained through a UAV positioning system and a laser point cloud spatial mapping algorithm. The diagnostic report is transmitted to the operation and maintenance management platform via an encrypted protocol, triggering the corresponding maintenance work order dispatch process.

[0158] The above steps construct a technical chain for anomaly identification through hierarchical decision-making logic: the root node assesses the intensity of environmental interference and selects a detection strategy, intermediate nodes filter transient noise interference, and leaf nodes complete high-confidence matching of fault types. The data flow between nodes is "environmental parameters → interference assessment → signal verification → fault classification," forming a closed-loop decision-making process from environmental perception to fault decision-making. The hierarchical design of the decision tree classifier balances anti-interference capability in complex environments with the accuracy of fault classification, ultimately outputting a structured diagnostic report to support operational and maintenance responses.

[0159] In specific embodiments of the present invention, considering complex application scenarios involving high salt spray, temperature gradient changes, and light fluctuations, the following technical solutions are used to achieve accurate identification and optimized control of photovoltaic module anomalies:

[0160] A hardware synchronous trigger is constructed using an FPGA programmable logic device to receive the rising edge signal of the laser drive circuit pulse to generate a synchronous clock source. This source drives the temperature and humidity sensor and the salt spray sensor to trigger a hardware interrupt at the start of the laser scanning cycle, synchronously acquiring temperature, humidity, and salt spray concentration data. An integrator circuit composed of an RC filter network and an operational amplifier is integrated into the light intensity sensor to filter the laser echo signal, extract the DC component of the ambient light intensity, and generate illuminance waveform data aligned with the synchronous clock via an analog-to-digital converter. The multi-source data is integrated using a timestamp alignment algorithm and input into the particle deposition rate calculation module. Based on the salt spray concentration time series, the module calculates the salt crystallization deposition trend and outputs environmental interference level parameters, providing a basis for subsequent laser parameter adjustments.

[0161] The salt spray concentration threshold boundary is dynamically updated using a sliding window mechanism. When the real-time salt spray concentration exceeds the threshold, the laser emitter switches to the near-infrared band. The driving current is adjusted through an electro-optic modulator to match the near-infrared spectral characteristics, and the pulse width is simultaneously optimized to the optimal range for penetration. The receiver's tunable filter dynamically adjusts the passband width based on illumination intensity data: under high illumination, the passband is compressed to filter out high-frequency stray light; under low illumination, the passband is extended to retain low-frequency effective signals. The scattering noise suppression module inverts the aerosol particle size distribution based on Mie scattering theory and combines differential detection technology to compare the differences in backscattering coefficients between adjacent scanning points. When the difference exceeds a threshold, digital filtering is triggered to eliminate interference signals and improve the signal-to-noise ratio of the reflected signal.

[0162] The reflection intensity signal was decomposed into five levels using the Db4 wavelet basis, focusing on energy fluctuation characteristics in the 200-400Hz frequency band. The reflectivity abrupt changes caused by cracks were identified by calculating the standard deviation of the sliding energy window. Multispectral reflectance data was analyzed using etalon interferometry to separate characteristic peaks, and hotspot temperature was quantified based on the linear mapping relationship between wavelength offset and temperature. UAV attitude data was input using a coordinate transformation matrix to correct point cloud distortion, and the second derivative of reflectivity was calculated using the Laplacian operator to extract dust gradient distribution characteristics. Temperature gradient data was fused to construct a multidimensional anomaly feature vector encompassing the time-frequency, spectral, and spatial domains, which was then input into a fault model database for similarity matching.

[0163] The temperature range is divided into five segments and associated with a baseline reflectance value. Reflectance intensity data is normalized to eliminate the influence of temperature drift. When the salt spray concentration is ≥5 mg / m³, the salt crystallization mode is activated. Morphological filtering is used to distinguish between salt crystallization and dust accumulation, and the judgment threshold is increased by 15%. Hot spot detection sensitivity is linearly correlated with light intensity. Under low light conditions, the trigger frequency is reduced to 70% of the baseline value, and the integration time is extended. A matching degree exceeding 85% triggers a level one alarm and freezes environmental parameter updates; a matching degree between 60% and 85% triggers a UAV re-inspection and path update; when salt spray or light intensity exceeds limits, backup sensor verification and a safety mode are activated. The decision tree classifier calculates the environmental interference level through the root node, verifies abnormal signals lasting ≥2 seconds through intermediate nodes, and generates a diagnostic report with a confidence level ≥90% through leaf nodes.

[0164] Based on abnormal coordinates, a spiral scanning path with an increasing pitch of 0.5m is generated, and the scanning line spacing increases from 10cm to 20cm with each scan. An improved contract network protocol divides sub-regions based on remaining battery power and sensor health scores, assigning multiple UAVs for collaborative detection. When wind speed is >8m / s, a 15° yaw compensation angle is inserted, and hovering points are densified; Bézier curves are used for smooth trajectory correction. Path execution data is fed back to the fusion model to update the temperature-reflection intensity compensation model, forming a closed-loop control of data acquisition, diagnosis, and path correction, continuously optimizing the system's detection accuracy and stability in dynamic environments.

[0165] The above implementation method solves the problem of misjudgment and missed detection under salt spray erosion, temperature change and light fluctuation by using environmental parameter coupling analysis, equipment dynamic optimization, multi-dimensional feature fusion and closed-loop feedback mechanism.

[0166] The key technical features involved in the technical solution of this invention are explained below to support those skilled in the art in understanding the implementation details of the solution:

[0167] A timing control module built on an FPGA programmable logic device receives pulse signals from the laser drive circuit as a trigger source and generates a synchronization clock signal with millisecond-level precision. This signal is used to coordinate the data acquisition timing of the temperature and humidity sensor, salt spray sensor, and light sensor, ensuring strict synchronization between multi-source data and laser scanning action, thus solving the timing deviation problem caused by asynchronous operation of multiple sensors in traditional methods.

[0168] Dynamic updating of salt spray concentration threshold boundaries: A sliding window mechanism (window period of 10 minutes) is used to statistically analyze the mean and variance of salt spray concentration, and the threshold range is dynamically corrected in conjunction with real-time temperature and humidity parameters. When the salt spray concentration exceeds the upper limit of the threshold, the laser wavelength is switched to the near-infrared band to optimize the laser's penetration efficiency in the salt spray environment and avoid signal attenuation caused by high salt spray concentration.

[0169] Tunable filter passband dynamic adjustment: The filter passband width is adjusted in real time based on light intensity data: under high light conditions, the passband is compressed to 70% of the standard range to filter out high-frequency noise from ambient stray light; under low light conditions, the passband is expanded to 120% to retain the effective low-frequency components of the reflected signal. This technology achieves adaptive passband adjustment through a voltage-controlled filter circuit, improving the signal-to-noise ratio under different lighting conditions.

[0170] Db4 wavelet basis decomposition and sliding energy window: The reflection intensity signal is decomposed into five levels using the Db4 wavelet basis to extract detail coefficients in the 200-400Hz frequency band. A sliding window is set in this frequency band to calculate the standard deviation of the signal energy. When the fluctuation value exceeds a preset threshold, it is determined to be a sudden change in reflectivity caused by microcracks on the photovoltaic panel surface, thus realizing the frequency domain feature localization of the crack.

[0171] Etennae interferometry characteristic peak locking: Utilizing the Fabry-Perot interferometry principle to separate multispectral reflection signals and identify characteristic peaks corresponding to the bandgap of photovoltaic materials. A linear mapping relationship between characteristic peak wavelength shift and module temperature is established through calibration experiments, converting real-time wavelength shifts into hotspot temperature values ​​to support quantitative assessment of hotspot temperatures.

[0172] The Db4 wavelet basis decomposition to the fifth level detail coefficients involved in the technical solution of this invention are explained below to support those skilled in the art in understanding its specific role in crack detection:

[0173] Hierarchical decomposition and frequency band coverage:

[0174] The first layer of detail coefficients (D1) corresponds to the highest frequency signal components (e.g., 1600-3200Hz), mainly including high-frequency noise in the laser reflection signal (such as sensor circuit noise and environmental electromagnetic interference). This layer of data is usually filtered out or noise-reduced to avoid high-frequency noise interfering with subsequent analysis.

[0175] The second layer of detail coefficients (D2) covers the 800-1600Hz frequency band and reflects short-term fluctuations in the signal. This layer may include localized reflectivity variations caused by minor scratches or dust on the photovoltaic panel surface, but it needs to be combined with higher-level decomposition to distinguish noise from actual cracks.

[0176] The third layer, detail coefficient (D3), corresponds to the 400-800Hz frequency band and captures mid-to-high frequency signal characteristics. This layer can detect periodic reflectivity fluctuations caused by minute defects (such as microcracks) inside the photovoltaic panel material, but further verification of temporal continuity is needed to eliminate transient interference.

[0177] The fourth layer, detail coefficient (D4), covers the 200-400Hz frequency band and characterizes the mid-to-low frequency signal components. This layer is sensitive to the gradual change in reflectivity caused by large cracks on the photovoltaic panel surface, but due to its wide frequency band, it may be mixed with other interferences. Therefore, it needs to be combined with the fifth layer to accurately focus on the target frequency band.

[0178] The fifth level detail factor (D5) focuses on the energy distribution in the 200-400Hz frequency band (analyzed jointly by D4 and D5), specifically designed to detect abrupt changes in reflectivity caused by microcracks on the photovoltaic panel surface. This frequency band is highly correlated with changes in light scattering characteristics caused by cracks (such as optical path distortion and multiple reflections). The standard deviation is calculated using a sliding energy window, and when the fluctuation value exceeds a preset threshold, it is identified as a crack feature.

[0179] Technical considerations for choosing the fifth level of decomposition:

[0180] Frequency band adaptability: The reflectivity abrupt change signal caused by photovoltaic panel cracks has significant energy concentration characteristics in the 200-400Hz frequency band. The fifth layer decomposition can accurately extract this feature, avoiding high-frequency noise and low-frequency background interference.

[0181] Balance of computational efficiency: Decomposing to the fifth level can retain key frequency band information while reducing computational redundancy caused by higher-level decomposition (such as the sixth level), thus meeting real-time requirements.

[0182] Advantages of the Db4 wavelet basis: The Db4 (Daubechies 4th order) wavelet has tight support and fourth-order vanishing moment characteristics, which can effectively capture transient changes in signals (such as the change in reflectivity at the crack edge) while suppressing artifact interference.

[0183] The aforementioned hierarchical strategy achieves accurate extraction of crack features by progressively stripping away noise and focusing on the target frequency band, providing reliable input for the construction of multidimensional anomaly feature vectors.

[0184] UAV attitude correction and differential calculation: The pitch and roll angles of the UAV are input into the coordinate transformation matrix to calculate the actual spatial coordinates of the laser point cloud, eliminating point cloud distortion caused by changes in flight attitude. The Laplacian operator is applied to the corrected reflectivity data to calculate the second-order differential value, extracting the dust gradient distribution characteristics and enhancing the sensitivity of slowly varying signal detection.

[0185] Morphological filtering and salt crystallization mode switching: When the salt spray concentration is ≥5mg / m³, the salt crystallization mode is activated, and morphological filtering combining opening and closing operations is added before dust accumulation judgment. Circular structural elements are used to match the distribution characteristics of salt crystal particles, suppressing high-frequency noise interference, while increasing the dust accumulation judgment threshold by 15% and reducing the probability of misjudging salt spray particles.

[0186] Hierarchical validation of decision tree classifiers:

[0187] Root node: Calculates the environmental interference level based on salt spray concentration (weight 50%), temperature gradient (25%), and light intensity (25%). When the interference level is ≥4, the anti-interference detection branch is enabled to enhance the number of wavelet decomposition layers and the passband width of the filter.

[0188] Intermediate node: A 2-second sliding window is used to verify the time continuity of abnormal signals. When the proportion of abnormal feature values ​​within the window is ≥80%, it is determined to be a valid abnormality.

[0189] Leaf nodes: Match valid abnormal feature vectors with a standardized fault coding library to generate a diagnostic report that includes three-dimensional coordinates, fault type, and confidence level ≥90%.

[0190] Improved Contract Network Protocol Task Allocation: The master UAV generates bidding information based on the topology of abnormal areas and broadcasts it to subordinate UAVs via a wireless mesh network. Subordinate UAVs calculate their bidding weights based on remaining battery power (positive weight), flight stability error (negative weight), and sensor accuracy deviation (negative weight). The master control unit divides the topology-related units (electrically connected areas) and environmental parameter-similar units (salt spray difference ≤15%, temperature gradient difference ≤5℃) according to their weights, and assigns high-weight UAVs to perform high-precision scanning.

[0191] Bézier curve trajectory correction: When the wind speed is >8m / s, a 15° yaw compensation angle is inserted into the flight trajectory to counteract the heading deviation caused by wind resistance. A Bézier curve interpolation algorithm is used to generate a smooth transition path with the current waypoint and the target waypoint as control vertices to eliminate trajectory jitter caused by sharp turns. At the same time, the spacing between hovering data collection points is increased to 1.5 times the standard density to compensate for positioning errors.

[0192] Closed-loop feedback mechanism:

[0193] Forward control chain: Every 30 seconds, environmental data is input into the device parameter optimization module to generate a laser power and drone speed adjustment suggestion table. The laser power is dynamically adjusted according to 80% of the speed increase.

[0194] Reverse calibration process: Cross-validate data from the same area using three drones, calculate the average sensor deviation as drift compensation, and update the battery degradation model parameters based on battery consumption rate (mAh / km) and trajectory error (cm) to optimize task allocation strategy.

[0195] The aforementioned technical features, through dynamic environment adaptation, multi-dimensional data fusion, and closed-loop control logic, solve the problems of abnormal misjudgment under salt spray corrosion, temperature gradient changes, and light fluctuations.

[0196] The Mie scattering theory involved in the technical solution of this invention is explained as follows to support those skilled in the art in understanding its specific application in anomaly detection:

[0197] Definition and Application of Mie Scattering Theory: Mie scattering theory is a mathematical model describing the scattering behavior of uniform spherical particles on electromagnetic waves (such as lasers). It is applicable to scenarios where the particle size is similar to the incident light wavelength (particle size range of approximately 0.1-10 μm). In this invention, the aerosol particles (such as salt crystals and dust) in salt spray environments are mostly within this range, and their scattering characteristics directly affect the quality of the laser reflection signal. Using Mie scattering theory, the particle size distribution and concentration of aerosols can be calculated, providing a theoretical basis for distinguishing between effective reflected signals and scattering noise.

[0198] Specific application of Mie scattering theory in this invention:

[0199] Aerosol particle size inversion calculation: The receiver-side scattering noise estimator, based on the Mie scattering formula, calculates the aerosol particle size distribution by measuring the backscattering coefficient (i.e., the ratio of reflected laser signal intensity to incident intensity) and combining parameters such as laser wavelength (e.g., 1064 nm in the near-infrared band) and the refractive index of salt spray particles. For example, based on the variation characteristics of scattered light intensity with angle, the average particle size and concentration of salt spray particles are derived, providing quantitative input for noise suppression.

[0200] Scattering interference signal elimination: Compare the differences in backscattering coefficients between adjacent scan points using differential detection technology.

[0201] If the difference in scattering coefficients between adjacent points exceeds a set threshold (e.g., a sudden change in scattering caused by a concentrated area of ​​salt spray particles), it is determined to be environmental scattering noise.

[0202] By combining the particle size distribution data obtained from the Mie scattering theory, we can distinguish between random noise (such as scattering from isolated large particles) and systematic interference (such as scattering from a uniform salt spray layer), and use digital filtering algorithms to specifically eliminate interference signals.

[0203] Mie scattering theory and laser parameter linkage optimization: Dynamically adjusting laser parameters based on Mie scattering inversion results.

[0204] When the detected aerosol particle size is small (e.g., <1μm), the laser pulse width is increased to enhance penetration.

[0205] If the aerosol concentration is high, the passband of the synchronously compressed tunable filter is used to suppress high-frequency scattering noise.

[0206] Advantages of Mie scattering theory: By quantitatively analyzing the scattering characteristics of salt spray particles through Mie scattering theory, traditional empirical noise suppression is transformed into precise control based on a physical model, solving the pain points of low signal-to-noise ratio of laser reflection signals and difficulty in extracting abnormal features in high salt spray environments. Combined with differential detection technology, it can effectively distinguish between real component anomalies (such as cracks and hot spots) and environmental scattering interference, improving detection accuracy under complex working conditions.

[0207] The above explanation shows that Mie scattering theory is the core scientific basis for scattering noise modeling and suppression in this invention, and its application runs through the entire process of environmental interference analysis, equipment parameter optimization and signal processing.

[0208] The technical solution of this invention solves the problem of false positives and false negatives in existing photovoltaic module anomaly identification methods under complex environments through dynamic environment adaptation and multi-dimensional data fusion mechanisms. The specific technical means are as follows:

[0209] Multi-source datasets are generated by simultaneously collecting temperature, humidity, salt spray concentration, and light intensity data. Threshold boundaries are dynamically updated based on historical salt spray concentration data, triggering a laser wavelength switch to the near-infrared band to enhance salt spray penetration. The passband range of a tunable filter is adjusted using light intensity data to suppress ambient light noise interference. Simultaneously, aerosol particle size distribution is inverted based on Mie scattering theory, and scattering noise is eliminated using differential detection technology, achieving real-time coupled analysis of environmental interference factors and laser parameters. These techniques, through the coordinated control of environmental perception and equipment optimization, address signal distortion issues in high salt spray and fluctuating light conditions.

[0210] Wavelet transform is used to extract time-frequency domain features of reflection intensity to identify cracks. Multispectral wavelength shift is analyzed using etalon interferometry to quantify hotspot temperature. After correcting spatial distortion using UAV attitude data, reflectivity gradients are calculated to locate dust accumulation distribution. Multidimensional anomaly feature vectors are constructed by fusing temperature gradient data and input into a fault model database for similarity matching. Temperature ranges are divided based on temperature and humidity data and correlated with reflectivity baseline values, dynamically normalizing reflection intensity data. Dust accumulation detection modes are switched according to salt spray concentration, and morphological filtering parameters are optimized. Hotspot detection sensitivity is adjusted based on illumination intensity. Environmental parameter-driven model adaptive correction reduces the interference of temperature drift and salt spray crystallization on anomaly detection.

[0211] A tiered alarm mechanism is triggered based on the matching results: high-confidence faults lock the evidence chain and freeze parameter updates; medium-confidence anomalies drive the UAV to re-inspect and update the path planning; and environmental exceedances trigger sensor verification and safety modes. The decision tree classifier evaluates the interference level through the root node, verifies the time continuity through intermediate nodes, and matches the fault coding library through the leaf nodes, filtering instantaneous interference and generating a diagnostic report. The path optimization module generates a dynamic detection path based on spiral scanning, multi-drone cooperative contract network protocol, and wind disturbance trajectory correction, and executes data feedback to update the temperature-reflection intensity compensation model of the fusion model, forming a closed-loop control logic of "data acquisition → diagnosis → path correction → model iteration".

[0212] The above technical solution achieves accurate anomaly identification under complex working conditions through dynamic coupling analysis of environmental parameters and equipment parameters, multi-dimensional feature fusion, and closed-loop feedback mechanism.

Claims

1. A method for anomaly identification of photovoltaic modules based on laser detection, characterized in that, include: Step 1: Collect dynamic environmental parameters of the area where the photovoltaic module is located. The dynamic environmental parameters include temperature, humidity, salt spray concentration and light intensity. Generate a multi-source dataset containing environmental variables by using timestamps and synchronizing with the laser scanning time sequence. Step 2: Dynamically adjust the output wavelength of the laser emitter based on the salt spray concentration data in the multi-source dataset. When the salt spray concentration exceeds the preset threshold range, switch to the near-infrared band. At the same time, adjust the passband range of the tunable filter at the receiver according to the light intensity data in the multi-source dataset to suppress ambient light noise. Step 3: Perform multi-dimensional feature extraction on the adjusted reflected laser signal, including time-frequency domain features of reflection intensity based on wavelet transform, wavelength shift features of multispectral decomposition, and three-dimensional spatial positioning features combined with UAV attitude data, and generate multi-dimensional anomaly feature vector by fusing temperature gradient data from multiple source datasets. Step 4: Construct a temperature and reflection intensity compensation model based on temperature and humidity data from the multi-source dataset, adjust the dust accumulation judgment threshold range based on salt spray concentration data from the multi-source dataset, and optimize the hot spot detection sensitivity parameters based on light intensity data from the multi-source dataset. Step 5: Input the multidimensional abnormal feature vector into the preset fault model database for similarity matching, trigger the hierarchical alarm mechanism based on the matching result, and verify the temporal continuity of the multidimensional abnormal feature vector through the decision tree classifier to eliminate instantaneous environmental interference signals. Step 6: Generate an optimized detection path based on the anomaly localization results. The path includes dense scanning area division based on multi-dimensional anomaly feature vectors, multi-UAV collaborative task allocation, and dynamic trajectory correction based on real-time environmental response. The path execution data is then fed back to a preset data fusion model to update the temperature and reflection intensity compensation model parameters.

2. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, Step 1 includes: A synchronous clock signal is generated by using the laser emission pulse signal as a reference clock source through a hardware synchronous trigger. The temperature and humidity sensor and the salt spray sensor trigger a hardware interrupt based on the synchronous clock signal to collect temperature and humidity data and salt spray concentration data synchronously with the laser scanning cycle. The light intensity sensor integrates the laser echo signal through an integrating circuit to generate illuminance waveform data aligned with the synchronous clock signal. The temperature and humidity data, salt spray concentration data, and illuminance waveform data are aligned by timestamp and integrated into a multi-source dataset. The multi-source dataset is then input into the particle deposition rate calculation module. Based on the salt spray concentration data, the amount of salt crystallization deposition per unit time is calculated, and the salt crystallization formation trend and environmental interference level are output.

3. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, Step 2 includes: Based on the historical salt spray concentration data in the multi-source dataset, a sliding window mechanism is used to dynamically update the salt spray concentration threshold boundary, which is used to determine whether laser wavelength switching is triggered. When the laser emitter switches to the near-infrared band, the laser pulse width is simultaneously adjusted to the preset penetration optimization range; The receiver scattering noise estimator performs inversion calculations on the aerosol particle size distribution based on Mie scattering theory, and compares the backscattering coefficient differences between adjacent scanning points using differential detection technology. When the difference exceeds a set threshold, the corresponding scattering interference signal is eliminated.

4. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, Step 3 includes: The reflection intensity signal is decomposed to the fifth level of detail coefficients using the Db4 wavelet basis. A sliding energy window is set in the 200-400Hz frequency band. When the energy fluctuation within the window exceeds the preset threshold, it is marked as a crack feature corresponding to the abrupt change in reflectivity. By analyzing multispectral reflectance data using the etalon interferometry method, the characteristic peaks corresponding to the bandgap of photovoltaic materials are identified, and a hot spot temperature mapping table is established based on the linear relationship between the characteristic peak wavelength shift and temperature change. The pitch and roll angles from the UAV attitude data are input into the coordinate transformation matrix. After correcting the spatial distortion of the laser point cloud data, the second derivative of reflectivity is calculated using the Laplacian operator. When the derivative value exceeds the set range, it is marked as a dust gradient distribution area.

5. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, Step 4 includes: Based on the temperature and humidity data in the multi-source dataset, the temperature range is divided into five segments from -20℃ to 80℃. Each segment is associated with an independent reflectivity baseline value, which is used to normalize the reflectivity intensity data. The dual-range dust accumulation determination mode is activated based on the salt spray concentration data in the multi-source dataset. When the salt spray concentration is ≥5mg / m³, the mode is switched to salt crystallization mode, and morphological filtering is added to distinguish between salt crystallization and regular dust accumulation. Based on the light intensity data in the multi-source dataset, the hot spot detection sensitivity is linearly weighted and correlated with the light intensity. When the light intensity is <500W / m², the detection trigger frequency is reduced to 70% of the baseline value.

6. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, Step 5 includes: When the matching degree between the multidimensional abnormal feature vector and the preset fault model exceeds 85%, a level one alarm is triggered, the fault feature evidence chain in the multidimensional abnormal feature vector is locked, and the environmental parameter update operation in the preset data fusion model is frozen. When the matching degree is between 60% and 85%, a level 2 alarm is triggered, a re-inspection command is generated to drive the drone to re-inspect the abnormal area within 3 minutes, and the coordinate data of the re-inspection area is input into the path planning engine; When the salt spray concentration in the multi-source dataset exceeds 5 mg / m³ or the light intensity is below 500 W / m², a level three alarm is triggered, the verification process of the backup sensor is initiated, and the pulse frequency in the laser scanning parameters is adjusted to a safe mode.

7. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, Step 6 includes: Based on the abnormal location coordinates in the multidimensional abnormal feature vector, a spiral expansion scanning path with a pitch increasing by 0.5m is generated with the abnormal point as the center, and the scanning line spacing increases from 10cm to 20cm with the number of scans. Based on the remaining battery power data and sensor health status scores of the drones in the multi-source dataset, the detection area is divided into multiple sub-regions using an improved contract net protocol, and assigned to multiple drones to perform collaborative tasks. When the wind speed in the real-time environmental monitoring data exceeds 8 m / s, a 15° yaw compensation angle is inserted into the flight trajectory, the spacing between hovering acquisition points is increased to 1.5 times the standard density, and the modified trajectory is smoothed using a Bezier curve interpolation algorithm.

8. The photovoltaic module anomaly identification method based on laser detection according to claim 7, characterized in that, Also includes: The master drone generates bidding information, including the topological connection relationship of the abnormal area, based on the abnormal positioning coordinates, and broadcasts it to the slave drones through the wireless mesh network; The subordinate UAV calculates its bidding weight based on the remaining battery power, flight stability error value, and sensor accuracy deviation value. The bidding weight is positively correlated with the remaining battery power and negatively correlated with the stability error and accuracy deviation. Based on the bidding weights, the sub-region to be detected is divided into topologically related units and environmentally similar units. The topologically related units are divided according to the physical connection relationship of the abnormal components, and the environmentally similar units are divided according to the regions with salt spray concentration differences ≤15% and temperature gradient differences ≤5℃.

9. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, It also includes a closed-loop feedback mechanism: The closed-loop feedback mechanism includes a forward control chain and a reverse calibration process; The forward control chain inputs environmental monitoring data from the multi-source dataset into the device parameter optimization module every 30 seconds to generate an adjustment suggestion table for laser output power and UAV flight speed, and increases the laser output power to 80% of the increase in UAV flight speed according to the suggestion table; The reverse calibration process cross-validates the detection data of the same area using three drones, calculates the average deviation of the sensor measurements as the drift compensation, and updates the exponential parameters in the battery degradation coefficient model based on the battery consumption rate and trajectory positioning error during path execution.

10. The photovoltaic module anomaly identification method based on laser detection according to claim 1, characterized in that, Step 5 includes: The root node calculates the environmental interference level based on the salt spray concentration, temperature gradient and light intensity data in the multi-source dataset. When the interference level is ≥4, the anti-interference detection branch is selected. The interference level adopts a weighted scoring model, with salt spray concentration accounting for 50% of the weight, and temperature gradient and light intensity each accounting for 25%. The scoring results are divided into 1-5 levels. The intermediate node verifies the temporal continuity of the multidimensional anomaly feature vector. When the anomaly signal lasts for ≥2 seconds, it is determined to be a valid anomaly; otherwise, it is marked as transient interference. The leaf node matches the multidimensional anomaly feature vector with a standardized fault type coding library to generate a diagnostic report that includes fault coordinates (X,Y,Z), type coding, and confidence level ≥90%.

Images