Damage detection method for photovoltaic string board

Abstract

The invention relates to the technical field of photovoltaic equipment fault detection, and provides a damage detection method for a photovoltaic group string board, and the method achieves the precise detection through the three steps: 1, collecting the data of three dimensions, i.e., electrical, physical images, environment and operation, and carrying out the abnormal elimination and smooth optimization, thereby obtaining a detection result; forming a structured data snapshot by using a time-space alignment algorithm of millisecond-level time synchronization and three-level space identification; 2, quantitatively extracting electrical, image and time sequence fault fingerprint features, and realizing fault and cause diagnosis by adopting CNN (image recognition) + LSTM (time sequence capture) + XGBoost (multi-feature classification) fusion AI model; and 3, verifying a diagnosis result by relying on a knowledge base integrating historical data, engineer experience and industry standards and a rule engine, and performing model increment / total training and pilot verification through operation and maintenance feedback data to form closed-loop optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic device fault detection, in particular to a damage detection method for photovoltaic string panels. BACKGROUND

[0002] As the core component of photovoltaic power generation system, the operation state of photovoltaic string directly affects the power generation efficiency and safety of power station, and the accurate detection of its damage condition is the key to the operation and maintenance of power station. However, the existing photovoltaic string damage detection technology has four core technical problems: first, the data collection dimension is single and the coordination is poor, only electrical parameters or physical image data are collected, environmental and operation data are not integrated, and different data sources have time stamp inconsistency and physical coordinate mismatching, which cannot form a coherent and global high-quality data basis, resulting in incomplete fault feature extraction; second, the fault diagnosis capability is limited, relying on a single AI model (such as using only CNN to process images or XGBoost to process electrical data), which is difficult to consider image fine-grained features, time series trend features and electrical quantitative features, and the fault fingerprint is not standardized and quantified, which is easy to confuse similar faults such as transient shadow and continuous hot spot, hidden crack hot spot and blocked hot spot; third, the diagnosis result lacks multi-source verification, only relying on model output, without combining operation and maintenance engineer experience rules and industry standards, which is easy to misjudge the theory and practice; fourth, the system has no self-iteration capability, the model and rules are fixed after deployment, which cannot adapt to new faults in different climate zones or long-term operation of components, and the diagnosis performance decays over time.

[0003] Therefore, a damage detection method for photovoltaic string panels is proposed, which builds an electrical-physical image-environmental operation three-dimensional data collection system, forms a structured data snapshot through a space-time alignment algorithm, quantifies fault features and builds a CNN+LSTM+XGBoost multi-model fusion AI diagnosis system, establishes a knowledge base and rule engine integrating historical data, experience rules and industry standards, and designs an operation and maintenance feedback-data screening-model iteration-pilot test verification closed-loop optimization mechanism to realize the precision, standardization and continuous optimization of fault detection. SUMMARY

[0004] In view of the shortcomings of the prior art, the present application provides a damage detection method for photovoltaic string panels.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a damage detection method for photovoltaic string panels, comprising the following steps: Step 1: Multi-dimensional data acquisition and preprocessing: Collect multi-dimensional data covering electrical parameters, physical images, and environmental operating conditions. Perform preprocessing operations such as abnormal data rejection, time series smoothing, and image optimization. Electrical parameters include string voltage, current, power, and IV curve. Physical images include infrared thermal imaging and visible light imaging. Environmental operating conditions include irradiance, temperature, and actual power generation. Step 2: Data fusion and feature extraction: Use a spatio-temporal alignment algorithm to fuse the preprocessed data to form a structured data snapshot. Extract electrical features, image features, and time series features based on the data snapshot. Electrical features include maximum power point parameters and fill factor. Image features include hot spot temperature difference and surface obstruction characteristics. Time series features include power fluctuation patterns and environmental correlations. Step 3: AI diagnosis and rule correction: Build a diagnostic model with a deep learning + machine learning fusion architecture. Input the extracted features to output fault types and confidence levels. Use a forward reasoning algorithm to match feature parameters and rule conditions based on a structured knowledge base and rule engine. Correct the confidence levels to form the final diagnosis result. Step 4: Closed-loop verification and model optimization: Visualize the diagnosis results and collect feedback data through the operation and maintenance terminal. Perform incremental or full training on the diagnosis model based on effective feedback data to achieve iterative optimization of the model.

[0006] Preferably, the electrical parameter acquisition in step 1 includes: String-level parameter acquisition: Real-time acquisition of voltage, current, and power data through an anti-electromagnetic interference DC side acquisition module. The acquisition module meets national electromagnetic compatibility test standards. IV curve acquisition: Regularly scan and acquire curve data through a string-level tester. The scanning start condition is irradiance fluctuation ≤5% / h and environmental temperature ≥5℃. The scanning data covers the open-circuit voltage to short-circuit current interval. Insulation resistance monitoring: Real-time monitoring of ground insulation resistance through an insulation resistance tester. Set an alarm threshold and preferentially store continuous data before and after the alarm.

[0007] Preferably, the physical image acquisition in step 1 includes: Infrared thermal imaging acquisition: Combination of fixed online thermal imager and unmanned aerial vehicle portable thermal imager for acquisition. The acquisition environment meets the temperature fluctuation ≤2℃ / h and the component surface has no temporary obstructions. The thermal image resolution is ≥640×512 pixels and automatically marks the hot spot parameters. Visible light imaging acquisition: Use a network high-definition camera for acquisition. Avoid backlight scenes during acquisition and ensure that the solar elevation angle is ≥30°. Automatically identify surface features such as component breakage and burn marks and store location information.

[0008] Preferably, the data preprocessing in step 1 specifically includes: Abnormal data rejection: 3σ criterion is used to identify jump data, and environmental threshold method is used to reject image data under bad weather and IV curve under low irradiance condition; Time series data smoothing: sliding average method is used to process the time series parameters of voltage and current, and the sliding window size is 5-10 minutes; Image data optimization: Gamma correction is used to eliminate uneven illumination for visible light images, Wiener filter is used to deblur for blurred images, and non-uniformity correction is used to eliminate noise for infrared thermal images.

[0009] Preferably, the spatio-temporal alignment algorithm in step 2 comprises: Time alignment: millisecond level timestamp format is used, and NTP protocol time synchronization server is used to ensure that the time deviation of each data source is ≤1 second; Spatial alignment: a three-level identification system of inverter-string-module is used, and perspective transformation algorithm is used to map and correlate the image pixel coordinates and module physical coordinates.

[0010] Preferably, the fusion architecture in step 3 comprises: Convolutional neural network module: input infrared thermal image and visible light image fusion features, extract 128-dimensional image feature vector; Recurrent neural network module: input 24-hour time series data, extract 64-dimensional time series feature vector; Gradient boosting tree module: input 12-dimensional electrical feature vector and above-mentioned fusion feature vector, output probability distribution of 12 types of faults.

[0011] Preferably, the knowledge base in step 3 uses production rule storage, each rule contains three elements of condition-conclusion-confidence correction value, and the core rules include: If the infrared image shows a strip-shaped hot area and the string voltage drops by 30%-35%, the conclusion is bypass diode failure and the confidence is corrected; If the maximum power point voltage of IV curve drops by ≥0.2V and there is a hot spot in the infrared image, the conclusion is hot spot failure and the confidence is corrected.

[0012] Preferably, the screening criteria of effective feedback data in step 4 include: The time difference between feedback data and diagnosis data is ≤7 days; Contains verification results with power recovery rate ≥70% after maintenance; With the qualification certificate of operation and maintenance personnel and on-site supporting image.

[0013] Preferably, the visualization display in step 4 includes: Three-dimensional platform display: Based on WebGL technology, a three-dimensional model of the power station is constructed, and red, yellow and blue are used to mark high, medium and low risk failures respectively. The color saturation is positively correlated with the confidence level. Report output: Generate a PDF format report containing fault location diagram, characteristic parameters and maintenance suggestions, support correlation of original data snapshot and image data.

[0014] Compared with the prior art, the application provides a damage detection method for photovoltaic string boards, which has the following beneficial effects: 1. In the scheme, a time-space alignment algorithm of millisecond-level time synchronization + inverter-string-module three-level space identification is proposed by constructing an electrical-physical image-environment operation three-dimensional collection system. Image and physical coordinate mapping is realized through perspective transformation, and finally a structured data snapshot containing four modules is formed. Break the traditional data fragmentation dilemma, provide global and coherent high-quality data support for subsequent accurate diagnosis, and avoid the problem of missed judgment caused by uncoordinated data.

[0015] 2. In the scheme, the feature threshold of IV curve, thermal imaging and time series data is quantified by quantifying the fault fingerprint feature. At the same time, a CNN+LSTM+XGBoost fusion architecture is constructed. CNN identifies image fine-grained features, LSTM captures time series trends, and XGBoost fuses multiple features for classification. Realize multi-feature collaborative diagnosis, effectively distinguish similar types such as transient shadow and continuous failure, and greatly improve the diagnosis accuracy.

[0016] 3. In the scheme, by constructing a structured knowledge base, integrating AI historical data, experienced engineer experience rules and industry standards, designing a forward reasoning rule engine, and dynamically correcting the confidence level by matching data features and rule conditions. For example, for diode faults, combined with multiple conditions such as thermal zone morphology and voltage change, avoid single feature misjudgment, make the diagnosis conclusion more in line with the actual operation and maintenance scene, and reduce the risk of misjudgment.

[0017] 4. In the scheme, by designing the operation and maintenance feedback-data screening-model iteration-pilot verification closed loop: collect maintenance data through mobile APP, select effective data for monthly incremental and quarterly full amount training, and optimize the after in multiple climate zones Pilot verification. This mechanism realizes the self-adaptive adjustment of the system for the first time, which can adapt to different climate zones and new faults, avoid performance degradation, ensure that the diagnosis accuracy continues to improve with the accumulation of operation and maintenance data, and ensure long-term reliable operation. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 A method for detecting damage to photovoltaic string panels includes the following steps: Step 1: Multi-dimensional data collection, preprocessing and fusion This step provides a high-quality data foundation for subsequent fault diagnosis through standardized data collection and precise data fusion. Specifically, it includes two sub-steps: data collection and data preprocessing fusion. 1. Data Collection Data collection must cover three dimensions: electrical data, physical image data, and environmental and operational data. The parameters collected in each dimension must comply with industry testing standards (such as GB / T37408-2019 "Guideline for Fault Detection of Photovoltaic Modules"), as detailed below: (1) Electrical data String-level data acquisition: The inverter's built-in DC-side acquisition module is used to acquire string voltage (accuracy ±0.1V), current (accuracy ±0.05A), and power (accuracy ±1%) in real time, with the acquisition frequency set to 1 minute / time. The acquisition module must have electromagnetic interference immunity and comply with the GB / T17626.2-2006 electromagnetic compatibility test standard.

[0021] Refined electrical data acquisition: IV curve scanning: A string-level IV curve tester (such as the portable tester IVT-1000) is used, with data acquisition frequency once a week. An automatic trigger condition is set to automatically start scanning when the irradiance is stable (irradiance fluctuation <5% / h) and the ambient temperature is >5℃. Each scan lasts 10 minutes. A data filtering algorithm is used to remove abnormal curves caused by transient interference, ultimately retaining 3 sets of valid curves. Curve acquisition must cover the complete range from string open-circuit voltage to short-circuit current, with a voltage sampling interval of 0.1V, and one current data point collected for each interval.

[0022] Insulation resistance monitoring: An insulation resistance tester (measurement range 0.1MΩ-100MΩ, accuracy ±5%) is used to monitor the string-to-ground insulation resistance in real time, with a sampling frequency of 5 minutes / time. An alarm threshold is set, and an audible and visual alarm is triggered when the impedance is lower than 1MΩ. At the same time, continuous data for 30 minutes before and after the alarm is stored first.

[0023] (2) Physical image data High-resolution infrared thermal imaging (IR) collection: Fixed points use online infrared thermal imager (model such as FLIRA700), collect once every 2 hours, covering all groups of the power station; drones carry portable infrared thermal imagers (model such as DJIZenmuseH20N), collect once a week for full coverage to verify fixed point data. The collection environment needs to meet: environmental temperature fluctuation <2℃ / h, no fallen leaves, snow and other temporary obstructions on the surface of the module. Thermal image parameters require: resolution not less than 640x512 pixels, temperature measurement range -20℃-120℃, accuracy ±2℃. After collection, automatically mark the coordinates of hot spots, temperature difference (ΔT) and thermal field distribution form data.

[0024] High-resolution visible light imaging collection: fixed points use network high-definition cameras (resolution not less than 1920x1080 pixels, frame rate 25fps), collect once every 4 hours; drones carry the same high-definition cameras, collect once a week for full coverage. The collection time needs to avoid backlight scenes (solar elevation angle >30°), after collection, automatically identify surface features such as module breakage, burn marks (area >1cm 2 ), bird droppings (diameter >2cm), shadow obstruction (coverage area >5% of module area), and store feature position and size data.

[0025] (3) Environmental and operating data High-precision environmental data collection: Use integrated environmental sensors (model such as Kipp&ZonenCMP11), collection parameters include irradiance (measurement range 0-2000W / ㎡, accuracy ±5W / ㎡, collection frequency 1 minute / time), environmental temperature (-40℃-85℃, accuracy ±0.5℃), wind speed (0-30m / s, accuracy ±0.3m / s), module backboard temperature (-40℃-85℃, accuracy ±0.5℃). The sensor installation position needs to be close to the module array (distance ≤5m) and avoid being blocked by the module or support, the installation height is flush with the module center point.

[0026] Operating state data collection: Real-time acquisition of actual power generation at the time of detection through the power station SCADA system, and calculation of theoretical expected power based on module nameplate parameters and real-time environmental data, using formula (1): Formula (1) Formula (1) character description: is the theoretical power generation (kW), is the rated power of the module (kW), is the real-time irradiance (W / ㎡), is the rated irradiance of the module (take 1000W / ㎡), Power temperature coefficient of module (% / ℃, value range -0.3% to -0.5%, provided by module nameplate), Module backboard temperature (℃), Module rated temperature (25℃).

[0027] Power vs. frequency and electrical data acquisition frequency is consistent (1 minute / time), automatically calculate the deviation value of actual power and theoretical power (P ) and store.

[0028] 2. Data preprocessing and fusion Data preprocessing and fusion is based on industrial-grade data processing server (configuration: CPU Intel Xeon E5-2690v4, memory 32GB, hard disk 1TB SSD) to achieve, the specific process is as follows: (1) Data preprocessing: Abnormal data rejection: 3σ criterion is used to identify and reject jump data (such as current suddenly jumps from 5A to 0A and lasts <10 seconds), and environmental threshold method is used to reject interference data (such as image data under heavy rain / sand weather, IV curve when irradiance <100W / m2); Time series data smoothing: sliding average method is used to process voltage, current, power, impedance and other time series data, and the sliding window size is set to 5 minutes, and the arithmetic mean value is taken within the window; Image data optimization: gray scale correction algorithm (such as Gamma correction, Gamma value 1.2) is used to eliminate the influence of uneven illumination on visible light images, and wiener filter algorithm is used to deblur the blurred images caused by unmanned aerial vehicle shaking; non-uniformity correction algorithm is used for infrared thermal image to eliminate sensor noise.

[0029] (2) Data fusion: The preprocessed multi-dimensional data is associated to form a data snapshot by using space-time alignment algorithm, the specific implementation is as follows: Time alignment: unified millisecond time stamp format (YYYY-MM-DDHH:MM:SS.fff) is used, and time synchronization server (supports NTP protocol) is used to ensure that the time deviation of each data source is <1 second; Space alignment: three-level identification system of inverter number-string number-module number (such as INV-01-STR-03-MOD-15, representing 1st inverter, 3rd string, 15th module) is used, and the pixel coordinates of image data and the physical coordinates of module are mapped through perspective transformation algorithm; Data storage: data snapshot is stored in structured format, containing four core modules and corresponding fields: Basic information module: contains timestamp (format: YYYY-MM-DD HH:MM:SS.fff), spatial identification (uses inverter number - string number - component number three-level identification); Electrical data module: contains real-time acquisition values of voltage, current, power, IV curve data (stored in the form of voltage value-current value pairs, such as [0, 8.5] representing current 8.5A when voltage is 0V), insulation impedance value, and data quality identification (0 represents excellent, 1 represents good, and 2 represents poor); Image data module: contains infrared thermal image storage path, visible light image storage path, hot spot temperature difference (ΔT), component surface stain area (unit: cm 2 ), and data quality identification (0 represents excellent, 1 represents good, and 2 represents poor); Environmental and operational data module: contains real-time irradiance, ambient temperature, wind speed, component backboard temperature, actual power generation, theoretical power generation, power deviation value, and data quality identification (0 represents excellent, 1 represents good, and 2 represents poor).

[0030] Step 2: Feature extraction and AI intelligent diagnosis This step realizes the preliminary diagnosis of fault type and cause by quantifying fault fingerprint features and constructing a fusion AI model, which includes two sub-links: feature engineering and AI diagnosis model. 1. Feature engineering: quantifying fault fingerprint Based on data snapshots, extract three types of fault features: electrical features, image features, and time series features. Each feature has a clear quantitative threshold, as follows: (1) Electrical features extracted from IV curve Shadow / dust feature: under low irradiance (<500W / ㎡), the IV curve current decreases proportionally, and the ratio of current drop to irradiance drop is in the range of 0.9-1.1 (when irradiance drops by 10%, current drops by 9%-11%). Calculate the similarity of curve shape to normal curve by cosine similarity algorithm, which is >90%; Hot spot feature: voltage step appears near the maximum power point (MPP), with voltage drop ≥0.2V or current drop ≥0.5A; fill factor (FF) is calculated by formula (2), which is decreased by ≥5% compared to normal string; Formula (2) Character explanation of formula (2): FF is the fill factor, Pmax is the maximum power of the component (kW), Voc is the open-circuit voltage of the component (V), Isc is the short-circuit current of the component (A); Aging / Degradation Characteristics: Fill factor (FF) decreases by ≥8%, series resistance calculated using IV curve fitting algorithms (such as least squares method). The open-circuit voltage is increased by ≥20% compared to the newer components. The decrease is ≤3%; Diode conduction characteristics: The string voltage suddenly drops by 30% to 35% (compared to the normal string voltage under the same environment), and the voltage fluctuation after the drop is <2% / h, short-circuit current ( (The decrease is ≤5%).

[0031] (2) Image features extracted from thermal imaging Hot spot characteristics: The hot spot is located inside the module (non-edge area, ≥2cm from the module edge), has an irregular shape, a temperature difference ΔT ≥5℃, and an area >2cm². 2 Furthermore, the displacement of the hotspot location in three consecutive data collections was less than 1 cm. Shading characteristics: The shape of the hot zone is consistent with the outline of the shading object (such as leaves in irregular shapes or utility poles in long strips), the temperature difference of the hot zone ΔT = 2℃-4℃, and the position of the hot zone changes with the solar angle, with a displacement of ≥10% of the component width per hour; Characteristics of diode conduction / connection failure: The hot zone is strip-shaped, with a length ≥ 80% of the module length and a width of 10% to 20% of the module width. The temperature difference of the hot zone is ΔT = 3℃-6℃, and the hot zone is distributed along the series connection direction of the string (from the positive terminal to the negative terminal of the module).

[0032] (3) Operational characteristics extracted from time series data Dust characteristics: power loss changes gradually, daily power loss <0.5%, and power recovery rate after cleaning or rain is calculated according to formula (3) ≥80%; Formula (3) Formula (3) character description: For power recovery rate, Power after cleaning (kW) Power (kW) before cleaning. Normal power of the component (kW); Shadow characteristics: Power loss is periodic (e.g., appears daily from 9:00 to 15:00, consistent with the shadow trajectory of the obstruction) or transient (lasting <30 minutes). The correlation between power fluctuation amplitude and irradiance fluctuation amplitude calculated by Pearson correlation coefficient is >0.9. Persistent fault characteristics (such as hot spots, diode faults): power loss duration > 24 hours, and the correlation between power loss value and changes in environmental factors such as irradiation and temperature < 0.3.

[0033] 2. AI Diagnostic Model: Structured Fault Classifier The AI diagnosis model adopts a deep learning + machine learning fusion architecture, developed based on Python3.8 environment and TensorFlow2.5 framework, as follows: (1) Model selection and architecture Basic model architecture: Convolutional Neural Network (CNN): ResNet-18 model is adopted, with input being the fusion features of infrared thermal images and visible light images (size 224x224x3), and through 6 convolutional layers and 4 residual blocks to extract spatial features, outputting an image feature vector of 128 dimensions; Recurrent Neural Network (RNN): LSTM structure is adopted, with input being the time series data of the past 24 hours (including power, voltage, current, environmental parameters, a total of 8 feature dimensions, time step 24), and through 3 layers of LSTM units to extract time features, outputting a time series feature vector of 64 dimensions; Gradient Boosting Tree (XGBoost): input is the 12-dimensional electrical feature vector extracted from the IV curve and the fusion feature vector output by CNN / RNN (a total of 12+128+64=204 dimensions), and through the construction of 100 decision trees to realize fault classification, outputting the probability values of each fault type.

[0034] Model advantages: CNN solves the fine-grained recognition of image features (such as distinguishing between hidden crack hot spots and occluded hot spots), LSTM solves the trend capture of time series features (such as distinguishing between transient shadows and persistent faults), XGBoost improves the classification accuracy after multi-feature fusion, and through testing, the fusion model improves the accuracy by 12% compared with single CNN model and 8% compared with single XGBoost model.

[0035] (2) Training process Training data set construction: Laboratory simulation data: In a constant temperature and constant irradiance laboratory (temperature control range -10℃-60℃, irradiance control range 0-1200W / ㎡), 100 photovoltaic modules of the same type (specification: single crystal silicon 300W, Voc=36V, Isc=8.5A) are selected, and typical faults (including 12 types of faults such as cell hidden crack, bypass diode failure, surface stain, and tree branch occlusion) are artificially created, with 100 groups of data collected for each type of fault, a total of 1200 laboratory data; Field measurement data: Select 10 photovoltaic power stations in different climate zones (covering temperate, tropical, plateau, etc. Climate types), test photovoltaic modules running for 2-10 years, mark fault types through EL test (confirm hidden cracks), artificial visual inspection (confirm shading / burn marks), electrical parameter comparison (confirm attenuation), etc. A total of 5000 sets of field data were collected, including 1500 sets in temperate climate zones, 1200 sets in tropical climate zones, 1000 sets in plateau climate zones, and 1300 sets in other climate zones; Dataset annotation: Cross-annotated by 5 engineers with more than 10 years of photovoltaic operation and maintenance experience, including fault type (12 types), cause (20 types), and characteristic parameter value. The final dataset label accuracy is ≥98%.

[0036] Model training and optimization: Data division: The annotated dataset is divided into training set (4340 groups), validation set (1240 groups), and test set (620 groups) in the ratio of 7:2:1, and the data proportion of each climate zone is kept consistent during division; Hyperparameter optimization: The CNN model uses the Adam optimizer with an initial learning rate of 0.001, which is reduced by 10% every 5 epochs. The LSTM model uses the RMSprop optimizer with a learning rate of 0.0001. The XGBoost model optimizes the hyperparameters through grid search, and the optimal parameters are: learning rate 0.1, tree depth 5, leaf node minimum sample size 3, and decision tree number 100; Training termination condition: The CNN and LSTM models stop training when the validation set loss function (cross-entropy loss, formula (4)) does not decrease for 3 consecutive epochs. The XGBoost model uses five-fold cross-validation and stops training when the validation set accuracy is ≥95%; Formula (4) Formula (4) character description: is the cross-entropy loss value, is the number of samples, is the true label of the ith sample (0-no fault, 1-fault, using one-hot encoding to represent fault type), is the prediction probability of the ith sample.

[0037] (3) Diagnosis process Feature input: After the system receives a new data snapshot, it automatically calls the preprocessing module to complete outlier removal and smoothing, and then generates a 204-dimensional feature vector (12-dimensional electrical features + 128-dimensional image features + 64-dimensional time series features) through feature extraction algorithms; Model inference: Adopting a series inference mode, first, the CNN model outputs the fault probability distribution of image features (12 types of fault probability), and the LSTM model outputs the fault probability distribution of time series features. Then, the two types of probability distribution and the electrical feature vector are input into the XGBoost model, and the final fault probability is calculated through a weighted fusion algorithm (weight coefficients: CNN output 0.3, LSTM output 0.2, electrical features 0.5). Result output: The model outputs the top 3 fault types and causes in terms of confidence, and the output format is as follows: ① hot spot (cause: battery piece hidden crack, confidence 85%); ② local shadow obstruction (cause: tree obstruction, confidence 10%); ③ normal (confidence 5%). Set the confidence threshold, when the highest confidence is <60%, trigger manual review reminder, and mark the corresponding data snapshot as pending review.

[0038] Third step: closed-loop diagnosis and verification This step realizes the continuous improvement of diagnosis accuracy through rule engine verification and operation and maintenance feedback optimization, including two sub-links: diagnosis knowledge base and rule engine, and diagnosis result visualization and feedback. 1. Diagnosis knowledge base and rule engine Knowledge base construction: Based on MySQL database, construct a structured knowledge base, integrating three types of core content: ① historical diagnosis data output by AI model (including fault type, feature parameter, and confidence); ② experience rules of 10 senior photovoltaic operation and maintenance engineers (such as strip hot zone + voltage drop for diode fault); ③ detection criteria in industry standards (GB / T37408-2019 "Photovoltaic Module Fault Detection Guide", GB / T29319-2012 "Photovoltaic Power Generation System Access to Distribution Network Technology Regulations"). Each rule in the knowledge base contains four elements: condition-conclusion-threshold-confidence correction value, stored in production rule representation.

[0039] Core rule example: Rule 1: If the IR image shows a strip-shaped hot zone (hot zone length ≥ 80% of the component length, width 10%-20%) and the string voltage drops 30%-35% (relative to the normal string voltage in the same environment) and the voltage stabilization time > 5 minutes, the conclusion is bypass diode fault, and the confidence correction is 90%; if only the first two conditions are met, the confidence correction is 60%; if only one condition is met, this rule is not triggered. Rule 2: If the IV curve has a voltage step (decrease ≥ 0.2V) near the MPP and the IR image has a hot spot (ΔT ≥ 5℃, area > 2cm 2 ) and the visible light image has no obstruction (stain coverage <5%), the conclusion is true hot spot fault (battery piece hidden crack), and the confidence correction is 92%; if the visible light image has no obstruction, the confidence correction is 70%. Rule 3: If the ratio of power loss to irradiance drop is 0.9-1.1, the IR image temperature distribution is uniform (ΔT < 3°C), and the visible light image surface is dirty (coverage > 10%), the conclusion is uniform dirt, and the confidence is corrected to 88%. If the power recovery rate after cleaning is ≥80%, the confidence is increased to 95%.

[0040] Rule engine operation: Using forward reasoning algorithm, when the AI model outputs the diagnosis result, the rule engine automatically traverses the rules in the knowledge base, matches the feature parameters in the data snapshot with the rule conditions, and if the conditions are met, the confidence is corrected according to the rules. The corrected result is the final diagnosis conclusion.

[0041] Rule update: Rule evaluation and update is performed every quarter. The specific process is as follows: ① Statistic the actual coincidence rate of each rule (the matching degree of rule conclusion and operation feedback result); ② If the coincidence rate < 85%, the rule threshold is adjusted by the engineer team (such as correcting the voltage drop range in rule 1 to 28%-35%); ③ The updated rule is imported into the knowledge base, and the correction effect is verified by 100 test data to ensure that the coincidence rate ≥ 90% before taking effect.

[0042] 2. Visualization and feedback of diagnosis results (1) Visualization display (clear display content and carrier) Operation platform display: Based on WebGL technology, a three-dimensional visualization platform of power station is developed. In the three-dimensional layout diagram of the platform, different colors are used to highlight the fault groups (red: high-risk faults such as hot spots / diode faults, yellow: medium-risk faults such as shadows, blue: low-risk faults such as dirt). The color saturation is positively correlated with the fault confidence (the saturation is the highest when the confidence is 100%). Clicking on the fault group pops up a diagnosis report, which includes: fault type, cause, confidence, affected component number, fault location diagram (marking hot spot / shadow area on the component, superimposed on the visible light image), maintenance suggestion (including operation steps and safety precautions, such as replacing bypass diode of component No. 15, operating in cloudy weather first, and disconnecting the DC switch before operation).

[0043] Data export: Supports exporting PDF format diagnosis report, which includes cover (power station name, detection time, report number), table of contents, fault overview, detailed diagnosis results (including original data snapshot, IV curve diagram, infrared thermal image), feature extraction results, model inference process, maintenance suggestion, and attachments (original data table). It is convenient for auditing and tracing.

[0044] (2) Operation feedback and model optimization Feedback data collection: Develop a mobile operation and maintenance APP (supporting Android 8.0 and above, iOS 13.0 and above system), after the on-site maintenance of the operation and maintenance personnel, fill in the feedback form through the APP, the mandatory items of the form include: actual fault type (drop-down selection, consistent with the diagnosis result type), cause confirmation method (single selection, such as EL test / visual inspection / electrical test), maintenance effect (fill in the quantitative indicators such as power recovery rate), feedback person (need to fill in the engineer qualification certificate number), feedback time; The form attachments need to upload maintenance site photos (at least 2, including fault component close-up, component status after maintenance) as evidence.

[0045] Data screening and labeling: The system automatically screens valid feedback data, screening standards: ① The time difference between feedback data and diagnosis data is <7 days (to ensure that the fault has not changed); ② The maintenance effect is quantifiable (such as power recovery rate >70%); ③ Contains evidence photos and the photos are clear and identifiable; ④ The feedback person has a valid photovoltaic operation and maintenance engineer qualification certificate. The effective data screened is cross-checked and labeled by two engineers, and the labeling results are divided into three categories: correct diagnosis (fault type and cause are matched), partial correct (fault type is matched but cause is wrong), and incorrect diagnosis (fault type is not matched).

[0046] Model optimization: Incremental training: Incremental training is performed once a month, and the effective feedback data (≥100 groups) in the month is added to the training set. The XGBoost / CNN / LSTM model is updated in fine-tuning mode (freeze the bottom layer network parameters, only train the top classifier), and the test set accuracy after training needs to be improved by ≥1% compared with the previous optimization; Full training: Full training is performed once every quarter, and the cumulative effective feedback data (≥300 groups) is combined with the original training set to retrain the model. After training, the test set accuracy is evaluated, and if the accuracy decreases by >3%, the feature quantization threshold is re-optimized (such as adjusting the hot spot ΔT threshold to 4℃); Pilot verification: The optimized model is first deployed in 2-3 pilot power stations (selected in different climate zones), and the diagnosis accuracy is calculated after running for 1 month. If the accuracy is improved by ≥2% and the misjudgment rate is <3%, it is fully promoted in all power stations.

[0047] It should be noted that the above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art can realize that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed by hardware or software depends on the specific application and design constraints of the technical solutions. The obvious changes or modifications derived therefrom are still within the scope of protection of the present application.

Claims

1. A method for damage detection of a photovoltaic string panel, the method comprising: The method comprises the following steps: ​ Step 1: Multi-dimensional data acquisition and preprocessing: Collect multi-dimensional data covering electrical parameters, physical images, and environmental operating conditions. Perform preprocessing operations such as abnormal data elimination, time series smoothing, and image optimization on the data. Electrical parameters include string voltage, current, power, and IV curve. Physical images include infrared thermal imaging and visible light imaging. Environmental operating conditions include irradiance, temperature, and actual power generation. Step 2: Data fusion and feature extraction: Use a space-time alignment algorithm to fuse the preprocessed data to form a structured data snapshot. Extract electrical features, image features, and time series features based on the data snapshot. Electrical features include maximum power point parameters and fill factors. Image features include hot spot temperature difference and surface obstruction characteristics. Time series features include power fluctuation patterns and environmental correlations. Step 3: AI diagnosis and rule correction: Build a diagnostic model with a deep learning + machine learning fusion architecture. Input the extracted features to output fault types and confidence levels. Use a forward reasoning algorithm to match feature parameters and rule conditions based on a structured knowledge base and rule engine. Correct the confidence levels to form the final diagnosis result. Step 4: Closed-loop verification and model optimization: Visualize the diagnosis results and collect feedback data through the operation and maintenance terminal. Perform incremental or full training on the diagnosis model based on effective feedback data to achieve model iteration optimization.

2. A method for damage detection of a photovoltaic string panel according to claim 1, characterized in that: The electrical parameter acquisition in Step 1 includes: String-level parameter acquisition: Real-time acquisition of voltage, current, and power data through an anti-electromagnetic interference DC side acquisition module. The acquisition module meets national electromagnetic compatibility test standards. IV curve acquisition: Regularly scan to obtain curve data through a string-level tester. The scanning start condition is irradiance fluctuation ≤5% / h and environmental temperature ≥5℃. The scanning data covers the open-circuit voltage to short-circuit current interval. Insulation resistance monitoring: Real-time monitoring of ground insulation resistance through an insulation resistance tester. Set an alarm threshold and preferentially store continuous data before and after the alarm.

3. A method for damage detection of a photovoltaic string panel according to claim 1, characterized in that: The physical image acquisition in Step 1 includes: Infrared thermal imaging acquisition: Combination of fixed online thermal imager and unmanned aerial vehicle portable thermal imager for acquisition. The acquisition environment meets the temperature fluctuation ≤2℃ / h and the component surface has no temporary obstructions. The thermal image resolution is ≥640×512 pixels and automatically marks the hot spot parameters. Visible light imaging acquisition: Use a network high-definition camera for acquisition. Avoid backlight scenes during acquisition and ensure that the solar elevation angle is ≥30°. Automatically identify surface features such as component breakage and burn marks and store location information.

4. A method for damage detection of a photovoltaic string panel according to claim 1, characterized in that: The data preprocessing in Step 1 specifically includes: Abnormal data elimination: Use the 3σ criterion to identify jump data and use environmental threshold method to eliminate image data under adverse weather and IV curve under low irradiance conditions. Time series data smoothing: Use the sliding average method to process voltage and current time series parameters with a sliding window size of 5-10 minutes. Image data optimization: Use Gamma correction to eliminate uneven lighting for visible light images, use Wiener filtering to deblur for blurred images, and use non-uniformity correction to eliminate noise for infrared thermal images.

5. A method for damage detection of photovoltaic string boards according to claim 1, characterized in that: The space-time alignment algorithm in Step 2 includes: Time alignment: millisecond timestamp format, ensure time deviation ≤1 second between data sources through NTP protocol time synchronization server; Space alignment: three-level identification system of inverter-string-module, map image pixel coordinates and module physical coordinates through perspective transformation algorithm.

6. A method for damage detection of a photovoltaic string panel according to claim 1, wherein: The fusion architecture in step 3 includes: Convolutional neural network module: input infrared thermal image and visible light image fusion features, extract 128-dimensional image feature vector; Recurrent neural network module: input 24-hour time series data, extract 64-dimensional time series feature vector; Gradient boosting tree module: input 12-dimensional electrical feature vector and the above fusion feature vector, output probability distribution of 12 types of faults.

7. A method for damage detection of photovoltaic string boards according to claim 1, characterized in that: The knowledge base in step 3 uses production rules storage, each rule contains three elements of condition-conclusion-confidence correction value, and the core rules include: If the infrared image shows a strip-shaped hot area and the string voltage drops by 30%-35%, the conclusion is bypass diode failure and the confidence is corrected; If the IV curve maximum power point voltage drops by ≥0.2V and the infrared image exists hot spot, the conclusion is hot spot failure and the confidence is corrected.

8. A method for damage detection of a photovoltaic string panel according to claim 1, wherein: The screening criteria for effective feedback data in step 4 include: Time difference between feedback data and diagnosis data ≤7 days; Include verification results with power recovery rate ≥70% after maintenance; With the qualification certificate of operation and maintenance personnel and on-site supporting image.

9. A method for damage detection of a photovoltaic string panel according to claim 1, wherein: The visualization display in step 4 includes: Three-dimensional platform display: build a three-dimensional model of the power station based on WebGL technology, use red, yellow and blue to mark high, medium and low risk faults, and the color saturation is positively correlated with the confidence; Report output: generate PDF format report containing fault location diagram, characteristic parameters and maintenance suggestions, support association with original data snapshot and image data.

Images