Multi-modal data driven intelligent fault diagnosis method and system for photovoltaic module

Abstract

The invention relates to the technical field of photovoltaic module fault diagnosis and management, and discloses a multi-modal data-driven photovoltaic module intelligent fault diagnosis method and system, and the method comprises the steps: collecting the electrical parameters, environment variables and physical characteristic data of a photovoltaic module in real time, carrying out the preprocessing, and outputting the processed standardized data; constructing a multi-type database, storing the processed standardized data, extracting a plurality of fault feature vectors in combination with a mixed fault detection model of a physical mechanism model and a data-driven algorithm, comparing the fault feature vectors with corresponding fault feature vector thresholds, and judging fault types; based on the fault feature vectors and the component parameters, fault inversion is carried out through a COMSOL multi-physics field simulation model, and a fault position is determined; and the fault types and the fault positions are collected, and a diagnosis result is output in a visual mode. Reliable technical support is provided for intelligent operation and maintenance of the photovoltaic power station, and the accuracy and timeliness of fault diagnosis are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic module fault diagnosis and management technology, and more specifically to a multimodal data-driven intelligent fault diagnosis method and system for photovoltaic modules. Background Technology

[0002] With the continuous expansion of photovoltaic (PV) installed capacity, the health management technology of PV modules has received significant attention. Annual power generation losses due to PV modules can reach 5-30%, with module failures accounting for over 60% of these losses. Therefore, research on PV module fault diagnosis has become a crucial element for the high-quality operation of PV power generation systems.

[0003] Traditional methods for detecting photovoltaic module faults mainly include artificial intelligence detection, infrared image detection, and... IV Curve analysis and other methods exist. While infrared image analysis utilizes temperature differences to detect hot spot faults, it is susceptible to environmental interference, leading to misjudgments. Traditional image processing suffers from difficulty in eliminating interference. Although attempts have been made to fuse visible light and infrared images and improve stitching algorithms, they still cannot comprehensively cover fault type identification, and the efficiency of some algorithms needs improvement, with limited adaptability to complex fault scenarios. IV Although the curve method has been improved by introducing fill factor and aging index, there are still errors in accurately judging the degree of aging failure. When faced with complex failures, multi-parameter fitting and classification are prone to confusion and are difficult to distinguish accurately.

[0004] Therefore, this invention provides a multimodal data-driven intelligent fault diagnosis method and system for photovoltaic modules to solve the problems of incomplete fault detection, inaccurate location, and difficulty in detecting complex faults in existing photovoltaic module fault detection methods. Summary of the Invention

[0005] In view of this, the present invention provides a multimodal data-driven intelligent fault diagnosis method and system for photovoltaic modules. By constructing a multi-type fault database and integrating electrical parameters, environmental variables, and physical characteristic data, a hybrid fault detection model based on physical mechanisms and data-driven approaches is built. Utilizing COMSOL electro-thermal-mechanical multiphysics coupling simulation technology, a knowledge base covering seven typical fault characteristics—short circuit, open circuit, hot spot, shading, dust accumulation, cracks, and aging—is established, enabling intelligent diagnosis and fault location of abnormal states in photovoltaic modules. This overcomes the limitations of single methods, improves fault location accuracy, solves the problem of ambiguous location in traditional methods, meets the needs of refined operation and maintenance of photovoltaic power plants, and improves the completeness and accuracy of fault detection. In the future, this system can be integrated with a drone inspection system, using drones equipped with thermal imaging and EL imaging equipment to collect physical characteristic data of photovoltaic modules and transmit it back to the system in real time for analysis and diagnosis. This is suitable for environments with complex terrain and where manual inspection is difficult.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A multimodal data-driven intelligent fault diagnosis method for photovoltaic modules includes: Real-time acquisition of electrical parameters, environmental variables, and physical characteristics of photovoltaic modules, followed by preprocessing and output of standardized data. A multi-type database is constructed and the processed standardized data is stored. A hybrid fault detection model combining physical mechanism model and data-driven algorithm is used to extract multiple fault feature vectors. The fault feature vectors are then compared with the corresponding fault feature vector thresholds to determine the fault type. Based on the fault feature vector and component parameters, and through the COMSOL multiphysics simulation model, the fault location is determined by fault inversion. The fault types and fault locations are categorized, and the diagnostic results are output in a visual manner.

[0007] Preferably, the electrical parameters include open-circuit voltage, short-circuit current, maximum power, and fill factor; the environmental variables include irradiance and temperature; and the physical characteristic data include thermal images and electroluminescence maps.

[0008] Preferably, the pretreatment includes: Kirchhoff electrical verification, noise filtering, and data cleaning were performed on the electrical parameters. Irradiance normalization and temperature compensation are applied to environmental variables to correct for the impact of the environment on electrical parameters; Noise filtering, image registration, cell segmentation, and region of interest extraction are performed on the physical feature data.

[0009] Preferably, a hybrid fault detection model combining physical mechanism models and data-driven algorithms is used to extract feature vectors for various faults. These feature vectors are then compared with corresponding fault feature vector thresholds to determine the fault type, including: Based on Kirchhoff's laws, electrical parameters are analyzed to determine whether they are normal, and electrical characteristic vectors are output, including open-circuit voltage deviation rate and short-circuit current deviation rate. Based on digital image processing technology, thermal images and electroluminescence images are analyzed to determine whether there are fault characteristics, and image feature vectors are output, including the proportion of dark areas, irradiance attenuation rate and heat flux density ratio. By analyzing electrical data through exponential decay fitting, relevant feature parameters are extracted, and the data-driven model feature vector is output, including power decay rate, decay residual ratio, power fluctuation and power reversibility. The fault feature vector is obtained by fusing electrical feature vectors, image feature vectors, and data-driven model feature vectors. The fault feature vector is compared with the preset thresholds for seven types of fault feature vectors, and the fault type is determined by combining Bayesian network probabilistic inference.

[0010] Preferably, the fault types include short circuit, open circuit, hot spot, shadow, dust accumulation, crack, and aging.

[0011] Preferably, the vector threshold for short-circuit faults is:

[0012] in, The vector threshold for short-circuit faults. The open-circuit voltage deviation rate threshold. The deviation rate threshold for the short-circuit current. The heat flux density ratio threshold, The power attenuation rate threshold; The vector threshold for open circuit faults is:

[0013] in, The vector threshold for open circuit faults. e 2 represents the attenuation residual ratio threshold; The vector threshold for hot spot faults is:

[0014] in, The vector threshold for hot spot faults; The vector threshold for shadow faults is:

[0015] in, The vector threshold for shadow faults, The threshold for the proportion of dark areas. The threshold for irradiance attenuation rate. The power fluctuation threshold; The vector threshold for dust accumulation faults is:

[0016] in, The vector threshold for dust accumulation faults. The power reversibility threshold; The vector threshold for crack failure is:

[0017] in, The vector threshold for crack failure; The vector threshold for aging faults is:

[0018] in, The vector threshold for aging faults.

[0019] Preferably, the fault location is determined by fault inversion based on fault feature vectors and component parameters, using a COMSOL multiphysics simulation model, including: A COMSOL multiphysics simulation model is established based on a three-dimensional hierarchical structure of components. Processed standardized data is loaded as boundary conditions, and the fault region in the simulation is initialized and constrained based on the fault feature vector. The multiphysics coupling equations are solved using the COMSOL multiphysics simulation model, and the simulation results are calculated and output, including the current distribution output by the electric field control equation, the temperature distribution output by the thermal field control equation, and the stress distribution output by the force field control equation. The simulation results are compared with the measured physical characteristic data. If they do not match, the model parameters are iteratively optimized until the simulation results match the measured data, thereby determining the spatial coordinates of the fault.

[0020] Preferably, in the multiphysics coupling equations, the electric field control equation is:

[0021] in, For spatial gradient, For electrical conductivity, V 'This represents the potential distribution; Thermodynamic control equations:

[0022] in, C is the material density. P For specific heat capacity, Thermal conductivity, For temperature gradient, For heat flux density, J For current density, E Electric field strength; Force field governing equations:

[0023] in, Let F be the stress tensor and F be gravity.

[0024] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a multimodal data-driven intelligent fault diagnosis method and system for photovoltaic modules, which has the following effects: (1) A complete closed loop was constructed, from data acquisition, processing, fault database construction to diagnostic decision output. The core objective is to accurately identify and locate seven common photovoltaic module faults (short circuit, open circuit, hot spot, shading, dust accumulation, crack, and aging) to provide decision support for efficient operation and maintenance.

[0025] (2) The fault knowledge base mainly includes fault detection and fault location. The fault detection stage innovatively adopts a hybrid fault detection model that combines physical models and data-driven models. The physical model quickly identifies structural faults (such as short circuits and open circuits) and fault types with obvious anomalies in thermal images and EL maps; the data-driven model can learn complex and nonlinear fault types from a large amount of historical data and identify subtle change features (such as early aging and performance degradation) that are difficult for the physical model to describe. The fault location stage innovatively integrates COMSOL multiphysics simulation for precise location. By accurately modeling the photovoltaic module, the electro-thermal-mechanical multiphysics field coupling inversion of the photovoltaic module is realized, which can accurately locate the physical location of the fault from the physical level, providing clear guidance for subsequent maintenance and greatly saving troubleshooting time and costs. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1 A flowchart of a multimodal data-driven intelligent fault diagnosis method for photovoltaic modules provided by the present invention; Figure 2 A flowchart of the steps of a photovoltaic module data preprocessing method provided by the present invention; Figure 3 A selection block diagram for seven fault feature vector threshold parameters of a photovoltaic module provided by the present invention; Figure 4 A flowchart illustrating the steps of a hybrid fault diagnosis model for photovoltaic modules provided by this invention; Figure 5 A block diagram of digital image processing for a photovoltaic module provided by the present invention; Figure 6 This invention provides a content framework diagram for photovoltaic module fault location. Figure 7 A flowchart illustrating the relationship between the output of diagnostic results for photovoltaic modules is provided by this invention. Figure 8 The intelligent diagnostic decision output relationship diagram provided by this invention; Figure 9 A block diagram of a multimodal data-driven intelligent fault diagnosis system for photovoltaic modules provided by the present invention. Detailed Implementation

[0028] 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.

[0029] This invention discloses a multimodal data-driven intelligent fault diagnosis method for photovoltaic modules, such as... Figure 1 As shown, it includes: Real-time acquisition of electrical parameters, environmental variables, and physical characteristics of photovoltaic modules, followed by preprocessing and output of standardized data. A multi-type database is constructed and the processed standardized data is stored. A hybrid fault detection model combining physical mechanism model and data-driven algorithm is used to extract multiple fault feature vectors. The fault feature vectors are then compared with the corresponding fault feature vector thresholds to determine the fault type. Based on the fault feature vector and component parameters, and through the COMSOL multiphysics simulation model, the fault location is determined by fault inversion. The fault types and fault locations are categorized, and the diagnostic results are output in a visual manner.

[0030] In this embodiment of the invention, sensors and testing instruments are used to monitor and collect electrical parameters, environmental variables, and physical characteristic data of photovoltaic modules in real time, and the collected data is preprocessed to provide comprehensive and basic data information for subsequent diagnosis, ensuring the accuracy and usability of the data and preparing data for the efficient operation of the photovoltaic module fault diagnosis system.

[0031] (1) Data acquisition section Four-wire system IV The scanner collects electrical parameters, including open-circuit voltage. Short-circuit current Maximum power MPP Fill factor FF Electrical parameters, etc.

[0032] Physical characteristic data includes thermal images and electroluminescence images; specifically, thermal images and electroluminescence images are acquired by an infrared thermal imager and an EL tester.

[0033] Environmental variables include irradiance, temperature, and humidity. Specifically, irradiance and temperature / humidity variables are collected by irradiance sensors and temperature / humidity sensors.

[0034] (2) Data preprocessing section, such as Figure 2 As shown, the acquired real-time data undergoes electrical verification, Savitzky-Golay noise filtering, data cleaning, and environmental compensation. Specifically: The operating electrical parameters are subjected to Kirchhoff electrical verification, noise filtering, and data cleaning to fill in missing data and delete values ​​that are outside the reasonable range.

[0035] Irradiance normalization and temperature compensation are applied to environmental variables to correct for their impact on electrical parameters. Specifically, irradiance normalization and temperature compensation are applied to environmental data to correct for the effects of irradiance and temperature on electrical parameters. Irradiance normalization corrects the measured current and voltage to standard test conditions. To determine whether a circuit conforms to Kirchhoff's laws, its mathematical description is as follows:

[0036] in, The current value after compensation; This is the measured current; Standard test irradiance; G This is the measured irradiance; The current temperature coefficient; Standard test temperature; T This is the measured temperature.

[0037]

[0038] in, The voltage value after compensation; This is the measured voltage; Voltage temperature coefficient; This is the voltage irradiation compensation coefficient; This is the natural logarithm of the irradiance ratio, used to quantify the nonlinear effect of irradiance differences on voltage.

[0039] The physical feature data, including thermal images and electroluminescence images, undergoes noise filtering, image registration, cell segmentation, and region of interest extraction. The specific processing steps are as follows: The EL image was denoised using a non-local means (NLM) algorithm, and grayscale normalization was performed on the EL image. The thermal image was smoothed using Gaussian filtering and calibrated according to the material emissivity.

[0040] The mathematical description of grayscale normalization is as follows:

[0041] Where M is the original pixel value. L avg_normal The average brightness of a normal solar cell. L target This is the reference brightness value.

[0042] The emissivity calibration formula is:

[0043] in, gray denoted as grayscale value, μ as material emissivity, and k and c as calibration coefficients.

[0044] In addition to the preprocessing operations mentioned above, image registration, cell segmentation, and region of interest (ROI) extraction operations are also required for the EL image and thermal image.

[0045] In this embodiment of the invention, the fault knowledge base layer consists of a multi-functional database, fault detection, and fault location, enabling the identification of fault types and the precise location of faults.

[0046] The multi-functional database section uses PostgreSQL, providing a powerful, flexible, high-performance, and cost-effective integrated data platform for building a photovoltaic module fault database.

[0047] (1) Data classification and storage Time-series data stores pre-processed electrical and environmental data in real time, dynamically recording data changes. Relational data stores basic component profiles and information on past fault types, times, and handling measures. Image data stores pre-processed thermal images, EL images, and... IV Image information such as curves. Fault feature vector data, storing threshold values ​​for seven types of fault feature vectors.

[0048] (2) Definition of threshold for seven types of fault feature vectors In this implementation plan, the seven fault feature vector thresholds are mainly used to diagnose seven types of faults in photovoltaic modules, namely short circuit, open circuit, hot spot, shadow, dust accumulation, crack, and aging. By combining the locking of essential parameters by physical model and the capture of dynamic behavior by data-driven approach, and by adding Bayesian network to quantify uncertainty, a multi-dimensional fault feature library is constructed.

[0049] For each type of threshold setting, a threshold optimization method based on the Receiver Operating Characteristic Curve (ROC) is employed. Threshold types include interval or single-value outputs, and the setting process is based on the core physical mechanism analysis and verification statistics for each type of fault. The threshold setting follows these procedures: Step 1: Selection of feature parameter types based on core physical properties. For example... Figure 3 As shown. Finally, based on the failure mechanisms of different faults, the most discriminative feature subsets are selected to construct a highly discriminative feature set for each type of fault.

[0050] Step 2: Determining the precise threshold boundary through statistical optimization based on ROC curves. First, based on the unique physical characteristics of each fault, the approximate range of variation for characteristic parameters is determined. Then, secondary classification statistical analysis of a large amount of measured data is needed to optimize the threshold boundary. This includes collecting a large amount of component data with clearly defined fault labels and using single-feature ROC analysis to determine the characteristic parameters (such as the heat flux ratio in short-circuit faults) for each fault type. v 5) Iterate through the threshold U, calculating the TPR (True Positive Rate) and FPR (False Positive Rate). Finally, plot FPR on the x-axis and TPR on the y-axis, combining this with U to form an ROC curve. Optimal threshold selection criteria: Find the optimal threshold U that maximizes the difference between TPR and FPR. This point represents the threshold with the strongest statistical discriminative power.

[0051] For each characteristic parameter of each type of fault, the above two steps are required to obtain its corresponding critical threshold U. A .

[0052] 1) Short-circuit fault characteristic parameters and thresholds When a photovoltaic module experiences an internal short circuit, a low-resistance path forms at the fault point, leading to an abnormal increase in current, a sharp drop in module output voltage and maximum power, and the generation of localized high temperatures at the short circuit point. Based on these significant electrical and physical characteristics, and through statistical optimization of the ROC curve, a strongly correlated diagnostic feature parameter vector and a judgment threshold are defined for short-circuit faults. The threshold expression can be described as:

[0053] The mathematical description of each characteristic parameter is as follows:

[0054]

[0055]

[0056]

[0057] This is the vector threshold for short-circuit faults. The open-circuit voltage deviation rate threshold. The deviation rate threshold for the short-circuit current. The heat flux density ratio threshold; This is the power attenuation rate threshold.

[0058] 2) Circuit breaker characteristic parameters and thresholds When a photovoltaic module experiences an internal open circuit, the fault point blocks the current path, causing the loop current to drop sharply to near zero, and the output power to essentially return to zero. On an infrared thermogram, the open-circuit area, due to the lack of current flow, exhibits a low-temperature dark spot, lower than the normal area. This leads to the definition of a diagnostic feature parameter vector and a judgment threshold strongly correlated with open-circuit faults. The threshold expression can be described as:

[0059] Among them, feature parameters The mathematical description is as follows:

[0060] in, The vector threshold for open circuit faults; This is the threshold for the attenuation residual ratio.

[0061] 3) Aging Fault Characteristic Parameters and Thresholds Photovoltaic module aging is a global, gradual performance degradation process. Its core characteristic is the continuous and irreversible decline in photoelectric conversion efficiency caused by material degradation and cell performance decline, manifested as a year-on-year decrease in maximum output power. Based on this unique and slow global degradation mode, a strongly correlated diagnostic feature parameter vector and judgment threshold are defined for aging faults. The threshold expression can be described as:

[0062] in, The vector threshold for aging faults.

[0063] 4) Shadow Fault Characteristic Parameters and Thresholds The core characteristics of shading faults in photovoltaic modules are uneven illumination and dynamic fluctuations. Shading leads to a decrease in the maximum output power of the module. Moving shadows (such as clouds) cause frequent fluctuations in output power. Local shading presents as shape-dependent dark areas in EL images and may lead to a significant decrease in overall short-circuit current. Based on these characteristics, a strongly correlated diagnostic feature parameter vector and a judgment threshold are defined for shading faults. The threshold expression can be described as:

[0064] The mathematical description of each characteristic parameter is as follows:

[0065]

[0066]

[0067] The vector threshold for shadow faults; The threshold for the proportion of dark areas in an EL image; The threshold for irradiance attenuation rate; This is the power fluctuation threshold.

[0068] 5) Characteristic parameters and thresholds of dust accumulation faults Dust accumulation in photovoltaic modules is a progressive, global, and reversible performance degradation. Dust covering the glass surface leads to decreased light transmittance and reduced effective irradiance, causing a continuous and stable decline in the module's maximum output power. With uniform dust accumulation, performance degradation exhibits global uniformity, typically not forming localized abnormal areas on infrared thermograms or EL images. Based on these characteristics, a strongly correlated diagnostic feature parameter vector and a judgment threshold are defined for dust accumulation faults. The threshold expression can be described as:

[0069] Among them, feature parameters The mathematical description is as follows:

[0070] The vector threshold for dust accumulation faults; This is the power reversibility threshold.

[0071] 6) Hot spot fault characteristic parameters and threshold The core characteristic of hot spot faults is that a localized solar cell transforms from a power-generating unit into an energy-consuming load, leading to abnormally high local temperatures and nonlinear power loss. Faulty solar cells exhibit significant high-temperature hotspots on infrared thermograms. Based on this unique characteristic of localized high temperature and nonlinear power loss, a strongly correlated diagnostic feature parameter vector and a judgment threshold are defined for hot spot faults. The threshold expression can be described as:

[0072] in, This is the vector threshold for hot spot faults.

[0073] 7) Crack Fault Characteristic Parameters and Thresholds Crack failure is a common latent defect in photovoltaic modules, primarily caused by mechanical stress. Its core characteristics manifest as thin black lines and a mesh-like pattern in EL images, and a slight temperature rise in infrared images. Cracks lead to a continuous and stable decrease in module output power. Based on the unique EL dark morphology, micro-thermal effect, and stable power attenuation characteristics, a strongly correlated diagnostic feature parameter vector and judgment threshold are defined for crack failure. The threshold expression can be described as:

[0074] in, The vector threshold for crack faults.

[0075] The above seven types of fault thresholds are stored separately in the fault feature vector data of the database.

[0076] In this implementation scheme, the fault category is determined through the fault detection loop of the fault knowledge base. The determination process is as follows: Figure 4 As shown; in the fault detection part, a hybrid fault detection model is used to extract feature parameter vectors that are strongly correlated with faults, and then compared with the thresholds of seven types of fault feature vectors to obtain the fault type.

[0077] (1) Physical model feature extraction The physical model uses Kirchhoff's laws to analyze electrical parameters such as current and voltage of photovoltaic modules at the circuit level, identify electrical anomalies caused by faults such as short circuits and open circuits, and output electrical characteristic parameters, including open circuit voltage deviation rate and short circuit current deviation rate. Based on digital image processing technology, it analyzes thermal images and electroluminescent images to determine whether there are fault characteristics and outputs image characteristic parameters, including dark area ratio, irradiance attenuation rate and heat flux density ratio. 1) Extraction of electrical anomaly features Kirchhoff's current law focuses on the conservation of node currents, while Kirchhoff's voltage law focuses on loop voltage relationships. Each cell in a healthy module follows a single-diode equivalent circuit model, mathematically described as follows:

[0078] in, It is a photoelectric current source; This is the diode saturation current; It is a series resistor; These are parallel resistors; These are terms related to thermal voltage.

[0079] If the current circuit does not conform to the original theoretical model, it is preliminarily judged that a circuit fault has occurred. Then, the measured open-circuit voltage is selected. V oc,act ) and measured short-circuit current ( I sc,act These are two fundamental characteristic parameters in photovoltaic module fault diagnosis. Among them, I sc,act Mainly affected by photocurrent source ( The effects of light energy capture and current conduction paths are reflected in the component's ability to capture light energy and the smoothness of the current conduction path, and are therefore often used to correlate faults such as shadows, dirt, and open circuits. V oc,act This mainly depends on the integrity of the solar cells and is a key indicator for diagnosing voltage-type faults such as short circuits in solar cells. V oc,actand I sc,act The main purpose of correcting to standard test conditions (STC) is to eliminate interference from environmental factors, achieve standardized performance comparison, and accurately diagnose faults.

[0080] The mathematical description of correcting the actual measured open-circuit voltage to standard test conditions is as follows:

[0081] in, This is the open-circuit voltage after environmental compensation; Measured open-circuit voltage.

[0082] The mathematical description of the actual measured short-circuit current corrected to standard test conditions is as follows:

[0083] in, This refers to the short-circuit current after environmental compensation. This is the measured short-circuit current.

[0084] When a fault occurs, changes in the circuit structure will cause the current and voltage to deviate from normal operating conditions. The open-circuit voltage deviation rate is calculated and output. ) and short-circuit current deviation rate ( ):

[0085]

[0086] in, This is the open-circuit voltage of the component when it is operating normally under standard test conditions. This is the short-circuit current when the component is operating normally under standard test conditions.

[0087] Feature parameters extracted from electrical anomaly processing , Output.

[0088] 2) Feature extraction in digital image processing The physical model part involves digital processing of images and data calculations, such as... Figure 5 As shown, the percentage of dark areas in the EL filter was obtained respectively. ), Irradiance attenuation rate ( ), heat flux density ratio ( Three feature vectors are used to constrain the identification of photovoltaic module fault types and output image feature vectors.

[0089] Step 1: Image Feature Extraction EL images show the near-infrared light emitted by a photovoltaic module after it is powered on. Normal areas are brightly lit, while defective areas (such as cracks) appear dark because current cannot pass through them.

[0090] Let the EL image be a grayscale matrix. Each pixel value M I ( i , j The dark area is defined as follows: ∈ [0, 255] (0 represents all black, 255 represents all white).

[0091] in, It is an adaptive threshold (determined by the Otsu algorithm). The dark area proportion is calculated as follows:

[0092] Where m is the number of rows in the grayscale matrix and n is the number of columns in the grayscale matrix.

[0093] When the surface irradiance of the solar module is lower than the ambient irradiance, the actual light intensity received by the solar module surface will be lower than under ideal environmental conditions. Irradiance attenuation rate. The mathematical description is as follows:

[0094] in, The effective irradiance of the component surface. This refers to environmental irradiance.

[0095] The surface irradiance of the component cannot be directly obtained from the thermal image. The ambient irradiance can be obtained by fitting the data using the linear least squares method. Average gray value of EL image The linear relationship between them:

[0096] in, denoted as EL image average grayscale value; a is the slope of the fitted straight line, and b is the intercept. This equation is established under the ideal condition of a clean and unobstructed photovoltaic module surface. When dust or shading exists on the module surface, its EL image is acquired and the average grayscale value is calculated. .Will Substituting this value into the calibration model above, the calculated value is the effective irradiance of the component surface. .

[0097]

[0098] Finally, the calculated effective surface irradiance At the same time as the ambient irradiance measured Substituting into the attenuation rate formula, we can obtain the result. .

[0099] The heat flux density ratio indicates the non-uniformity of heat distribution, particularly the degree of heat concentration in the fault area. The mathematical description of heat flux density is:

[0100] in, q ( i,j ) is located at coordinates ( i,j The energy radiated outward per unit time and per unit area by a pixel; (Boltzmann constant); ε is the emissivity of the material (photovoltaic glass ≈ 0.85, silicon wafer ≈ 0.7); T ( i,j ) is located at coordinates ( i,j The component surface temperature measured at ( ); The ambient temperature.

[0101] Average heat flux density

[0102] Maximum heat flux density

[0103] Heat flux density ratio .

[0104] Where N = m × n, is the total number of pixels.

[0105] Step 2: Output feature parameters Feature parameters extracted from data image processing and data computation Output.

[0106] (2) Data-driven model feature extraction Data-driven algorithms include exponential decay fitting and Bayesian networks. Exponential decay fitting targets performance degradation faults (aging, long-term hot spots), fitting time-series electrical data to extract feature vectors such as decay rate. Bayesian networks utilize historical data to construct probabilistic associations between "faults and data features," inferring the probability of occurrence and feature correlations of various faults under multi-source data.

[0107] 1) Extraction of performance degradation feature parameters The data-driven algorithm uses exponential decay fitting to target performance degradation faults (aging, long-term hot spots) and extracts feature vectors of the data-driven model, including power decay rate, decay residual ratio, power fluctuation, and power reversibility feature parameters.

[0108] Exponential decay models are commonly used to describe performance degradation failures, such as aging and long-term hot spots. Their mathematical description is as follows:

[0109] in, P ( t )for t Performance parameters at time; A is the initial amplitude; B is the decay rate; B is the asymptotic value of the steady-state power. t For time; This model can capture the trend of equipment performance changes over time. Using the Levenberg-Marquardt (LM) algorithm, the optimal parameters are obtained by fitting historical power data using a nonlinear least squares method. This allows the model to fit the predicted power value. Compared with actual observed power data points The sum of squared residuals (SSR) between them is minimized.

[0110] The power attenuation rate is obtained directly from the fitting results. . The larger the value, the faster the component's performance degrades. The power attenuation rate is obtained through fitting. :

[0111] The attenuation residual ratio reflects the degree of deviation between the actual attenuation and the ideal attenuation model. The mathematical description of the attenuation residual ratio is as follows: Calculate the residuals for all data points:

[0112] in, For the first f The deviation between the model's predicted value and the ideal model value at each time point.

[0113] Calculate its root mean square RMS :

[0114] in, Q This is the total number of time steps used to calculate the residual sequence.

[0115] With initial rated power For example:

[0116] Power fluctuation indicates the degree to which a system is affected by random disturbances. The mathematical description of power fluctuation is: Calculate the above residual sequence Standard deviation (STD):

[0117] in, This represents the average residual.

[0118] Power invertibility indicates the system's self-recovery capability. After fitting, the residual sequence is analyzed. Behavior within a specific period. If the residual... If a regular periodic change is observed, it proves that reversible decay exists; if the residual... If the deviation is persistent and unidirectional (actual power consistently falls below the model prediction and does not recover), it indicates irreversible degradation. Power reversibility The mathematical description can be expressed as:

[0119] By analyzing various types of data through a physical mechanism model, corresponding feature parameters are extracted, and the feature vector of the physical model is output. After fitting historical power data, the output data drives the algorithm to fit the feature vector of the exponential decay fitting part. The feature vectors from the physical mechanism model and the data-driven model are fused into a single vector to obtain the comprehensive feature vector extracted by the hybrid fault detection model. :

[0120] 2) Bayesian network output fault correlation The data-driven algorithm utilizes a Bayesian network, supported by the probability relationships between "faults" and "features" in a historical fault dataset. By inputting the historical fault dataset and multi-source feature data from the current system state, it learns the weight relationships between faults and features based on historical data, performs comparative corrections on real-time data, and outputs the target fault probability. l Characteristic parameters help to accurately identify fault types.

[0121] (3) Priority-based fault type discrimination output A hierarchical decision tree approach is used to perform a sequential comparison of fault vectors, starting with the easiest and progressing to the most difficult, and from the outside in. The fault type determination output flowchart is as follows: Figure 7 As shown: First, distinguish external reversible faults (first determine if there is a shadow fault, then determine if there is a dust accumulation fault); then determine serious faults (i.e., hot spot faults), which are high-risk faults and require priority alarm; next, distinguish electrical faults (first determine if there is a short circuit fault, then determine if there is an open circuit fault), as electrical faults have obvious symptoms and are easily distinguished by electrical parameters; finally, diagnose internal damage and aging (first determine if there is a crack fault, then determine if there is an aging fault). The following uses a shadow fault as an example to illustrate the specific comparison process: First, determine if it's due to shadow occlusion, then extract... In and Threshold in the matrix The comparison is performed, and a Bayesian network is used to learn the weight relationship between faults and features based on historical data, outputting the target fault probability. l If all thresholds are met and the probability of outputting a shadow fault is high, then it is determined to be a shadow fault. Output the result and the process ends; otherwise, proceed to the next fault, the dust accumulation fault identification process.

[0122] In this embodiment, when a system fault is detected, fault inversion is triggered for fault location. The fault location flowchart is as follows: Figure 6 As shown.

[0123] (1) COMSOL Multiphysics Simulation Modeling COMSOL multiphysics simulation is used to perform accurate geometric modeling and boundary condition loading of photovoltaic modules.

[0124] 1) Geometric modeling A three-dimensional layered structure for photovoltaic modules is constructed, consisting of a glass layer, an EVA encapsulation layer, a cell layer, an EVA encapsulation layer, and a TPT backsheet layer, from top to bottom.

[0125] 2) Boundary condition loading An accurate electro-thermal-mechanical coupling model of photovoltaic modules is established by inputting boundary conditions.

[0126] (2) Solving the multiphysics coupling model The multiphysics coupling equations are solved using a COMSOL multiphysics simulation model, outputting the simulated current density distribution, temperature distribution, and stress distribution. IV Simulation results such as curves.

[0127] The governing equations for the electric field are:

[0128] The mathematical description of the output current density is as follows:

[0129] in, For spatial gradient, For electrical conductivity, V 'For potential distribution, J The current density is calculated using the electric field control equation. J The distribution is used to analyze current distribution, such as current anomalies or decreases caused by hot spots or ash accumulation.

[0130] Thermodynamic control equations:

[0131] in, C is the material density. P For specific heat capacity, Thermal conductivity, For temperature gradient, For heat flux density, E The electric field strength is used to output the temperature distribution through the thermal field control equation, which is then used to analyze the component surface temperature and match it with thermal images. Examples include matching high-temperature regions in thermal images of hot spot faults, localized temperature rise phenomena in crack faults, and analyzing temperature non-uniformity caused by dust accumulation or obstruction.

[0132] Force field governing equations (static equilibrium equations):

[0133] The stress tensor is mathematically described as follows:

[0134] It is the total strain, composed of the coupling of mechanical and thermal strain. Its mathematical description is:

[0135] Mechanical strain and thermal strain The mathematical description is as follows:

[0136]

[0137] Where F is gravity and D is the elasticity matrix (which depends on the material's Young's modulus, Poisson's ratio, etc.). It is the initial strain. u It is a displacement vector. The coefficient of thermal expansion is The reference temperature is under stress-free conditions. It is an identity matrix. The stress distribution is output through the force field control equations to analyze the mechanical state of components, and is an important indicator for diagnosing physical damage (such as cracks and microcracks).

[0138] (3) Fault inversion logic The simulation results are compared with the measured physical characteristic data (thermal images, EL maps). If they match, the fault location is output; if they do not match, the fault parameters, location, severity, and other data are adjusted until a match is found. For hot spot faults, the following conditions must be met simultaneously: the high-temperature area of ​​the thermal field matches the thermal image; there is an abnormal current branch in the electric field caused by local resistance increase; and there is a region of concentrated thermal strain in the force field. When a crack fault occurs, the stress concentration point coincides with the location of the dark EL area; the electric current at the crack point returns to zero; and the thermal field exhibits a significant local temperature rise. Dust accumulation and shadow faults depend on the negative correlation between the decrease in electric field EL dark area current density and temperature distribution.

[0139] (4) Parameter optimization iteration When the simulation results (current density distribution, temperature distribution, etc.) do not match the measured data (thermal images, EL images, etc.), iterative optimization is performed. The local resistance value corresponding to the dark area of ​​the EL is corrected so that the simulated current density distribution map matches the EL grayscale distribution map; the material contact thermal resistance is optimized so that the simulated temperature field profile matches the morphology of the high-temperature area in the thermal image.

[0140] By iterating through parameters, the simulation results are made to approximate the measured data, and the spatial coordinates of the fault are finally determined.

[0141] The intelligent diagnostic decision output layer in this implementation plan includes two parts: fault diagnosis information collection and diagnostic information visualization. The intelligent diagnostic decision output relationship diagram is as follows: Figure 8 As shown.

[0142] (1) Fault diagnosis information collection section The fault type, along with fault-related location, time, and environmental data, are collected into the database.

[0143] (2) Visualization of diagnostic information Output using data visualization technology IV The system provides readable charts such as curves, and automatically integrates fault type, location, environmental data, and direct database connection to generate HTML format reports. It uses the real-time communication protocol (MQTT) to push fault warnings to user terminals in real time, and uses a front-end pop-up component (Vue dynamic rendering) to display core risk indicators and corresponding handling suggestions.

[0144] This invention provides a multimodal data-driven intelligent fault diagnosis system for photovoltaic modules, such as... Figure 9 As shown, it includes: Multimodal data acquisition and preprocessing module: used to acquire electrical parameters, environmental variables and physical characteristics of photovoltaic modules in real time, and preprocess and output standardized data after processing; Fault Knowledge Base Module: Used to build multi-type databases and store processed standardized data. It combines a hybrid fault detection model with physical mechanism models and data-driven algorithms to extract various fault feature vectors and compares the fault feature vectors with the corresponding fault feature vector thresholds to determine the fault type. Based on the fault feature vector and component parameters, and through the COMSOL multiphysics simulation model, the fault location is determined by fault inversion. Intelligent diagnostic decision output module: used to collect fault types and fault locations, and output diagnostic results in a visual manner.

[0145] The specific implementation process of the system of the present invention is the same as the method described above, and will not be repeated here. Please refer to the method section.

Claims

1. A multi-modal data-driven intelligent fault diagnosis method for photovoltaic modules, characterized in that, The method comprises the following steps: Real-time acquisition of electrical parameters, environmental variables and physical characteristic data of the photovoltaic module, and preprocessing to output processed standardized data; Construction of a multi-type database to store the processed standardized data, combined with a hybrid fault detection model of physical mechanism model and data-driven algorithm, extraction of multiple fault feature vectors, comparison of the fault feature vectors with corresponding fault feature vector thresholds, and determination of fault types; Determination of fault location based on fault feature vectors and module parameters, and fault inversion through a COMSOL multi-physics simulation model; Collection of fault types and fault locations, and output of the diagnosis results in a visual manner.

2. The multi-modal data-driven photovoltaic module intelligent fault diagnosis method according to claim 1, characterized in that, The electrical parameters include open-circuit voltage, short-circuit current, maximum power and fill factor; the environmental variables include irradiance and temperature; and the physical characteristic data include thermal images and electroluminescence images.

3. The multi-modal data-driven photovoltaic module intelligent fault diagnostic method of claim 1, wherein, The preprocessing includes: Kirchhoff electrical verification and noise filtering, data cleaning of the electrical parameters; Irradiance normalization and temperature compensation of the environmental variables to correct the influence of the environment on the electrical parameters; Noise filtering, image registration, cell segmentation and region of interest extraction of the physical characteristic data.

4. The multi-modal data-driven photovoltaic module intelligent fault diagnostic method of claim 2, wherein, The hybrid fault detection model of physical mechanism model and data-driven algorithm extracts multiple fault feature vectors, compares the fault feature vectors with corresponding fault feature vector thresholds, and determines fault types, including: Analysis of electrical parameters based on Kirchhoff's law to determine whether the electrical parameters are normal, and output of electrical feature vectors, including open-circuit voltage deviation rate and short-circuit current deviation rate; Analysis of thermal images and electroluminescence images based on digital image processing technology to determine whether there are fault features, and output of image feature vectors, including dark area proportion, irradiance attenuation rate and heat flux density ratio; Exponential decay fitting analysis of electrical data to extract corresponding feature parameters, and output of data-driven model feature vectors, including power attenuation rate, residual error ratio, power fluctuation degree and power reversibility; Fusion of electrical feature vectors, image feature vectors and data-driven model feature vectors to obtain fault feature vectors; Comparison of the fault feature vectors with preset seven-type fault feature vector thresholds, combined with Bayesian network probability reasoning, to determine fault types.

5. The multi-modal data-driven photovoltaic module intelligent fault diagnostic method of claim 4, wherein, The fault types include short circuit, open circuit, hot spot, shadow, dust accumulation, crack and aging.

6. The multi-modal data-driven photovoltaic module intelligent fault diagnosis method according to claim 5, wherein the vector threshold of the short circuit fault is: The vector threshold of the open circuit fault is: wherein, a vector threshold for short circuit faults, a deviation rate threshold for open circuit voltage, a deviation rate threshold for short circuit current, a heat flow density ratio threshold, a power decay rate threshold; The vector threshold of the hot spot fault is: wherein, a vector threshold for open circuit faults, a decay residual ratio threshold; The vector threshold of the shadow fault is: wherein, a vector threshold for hot spot failure; The vector threshold of the dust accumulation fault is: wherein, a vector threshold for shadow failure, a dark area ratio threshold, an irradiance decay rate threshold, a power fluctuation degree threshold; The vector threshold of the crack fault is: wherein, a vector threshold for a fouling fault, a power reversibility threshold; The vector threshold of the aging fault is: wherein, is a vector threshold for a crack failure; Determination of fault location based on fault feature vectors and module parameters, and fault inversion through a COMSOL multi-physics simulation model, including: wherein, is the vector threshold for aging failures.

7. The multi-modal data-driven photovoltaic module intelligent fault diagnostic method of claim 4, wherein, Establishment of a COMSOL multi-physics simulation model based on the three-dimensional layered structure of the module, loading of the processed standardized data as boundary conditions, initialization and constraint of the fault area in the simulation based on the fault feature vectors; ​ The multi-physics field coupling equations are solved by a COMSOL multi-physics simulation model, and simulation results are calculated and output, including current distribution output by an electric field control equation, temperature distribution output by a thermal field control equation, and stress distribution output by a force field control equation; The simulation results are compared with the measured physical characteristic data, if not matched, the model parameters are iteratively optimized until the simulation results match the measured data, so as to determine the spatial coordinates of the fault.

8. The multi-modal data-driven photovoltaic module intelligent fault diagnostic method of claim 7, wherein, In the multi-physics field coupling equations, the electric field control equation is: wherein is the spatial gradient, is the electrical conductivity, V is the electric potential distribution; The thermal field control equation is: wherein, is the density of the material, C P is the specific heat capacity, is the thermal conductivity, is the temperature gradient, is the heat flux density, J is the current density, E is the electric field strength; The force field control equation is: wherein is the stress tensor, F is the gravitational force.

9. A multi-modal data driven photovoltaic module intelligent fault diagnostic system, characterized in that, It includes: A multi-modal data acquisition and preprocessing module is used to acquire and preprocess the electrical parameters, environmental variables and physical characteristic data of the photovoltaic module in real time, and output the processed standardized data; A fault knowledge base module is used to build a multi-type database and store the processed standardized data, combine a hybrid fault detection model of a physical mechanism model and a data-driven algorithm, extract a plurality of fault feature vectors, and compare the fault feature vectors with corresponding fault feature vector thresholds to determine the fault type; Based on the fault feature vectors and the component parameters, a fault inversion is performed through a COMSOL multi-physics simulation model to determine the fault location; An intelligent diagnostic decision output module is used to collect the fault type and fault location, and output the diagnostic results in a visual manner.

Images