Photovoltaic module fault monitoring system and method

Abstract

The invention relates to the technical field of power supply or power distribution circuit systems, in particular to a photovoltaic module fault monitoring system and method, and the system comprises an acquisition module, a power supply dynamic compensation module, a control optimization module, a power supply module and a digital twin module. The acquisition module acquires multi-source data through the infrared thermal imaging sensor, the electroluminescent detection unit and the current and voltage characteristic curve acquisition module, the power supply dynamic compensation module dynamically adjusts wavelet basis function parameters based on an improved whale optimization algorithm, and the power supply module fuses hot spot distribution, current and voltage abnormity and aging trend data. And driving the dynamic adjustment of the protection threshold. And the digital twin module synchronizes real and virtual system data through transfer learning, rehearses a fault path to generate a maintenance instruction and feeds back the maintenance instruction to an optimization prediction model, so that a monitoring, protection and self-optimization closed-loop system is formed, and the fault diagnosis precision and the system reliability in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of circuit systems for controlling power supply or distribution of equipment, and particularly to a photovoltaic module fault monitoring system and method. Background Art

[0002] Photovoltaic module fault monitoring is a key link for real-time diagnosis of the operating state of a photovoltaic system relying on multi-source sensor fusion and intelligent analysis technologies. Its core mechanism lies in identifying abnormal surface temperature distributions of modules through infrared thermal imaging technology to locate hot spot effects caused by cell breakage, bypass diode failure, or shadow occlusion; analyzing microscopic structural abnormalities such as internal cracks and welding defects in cells by combining electroluminescence detection means; and simultaneously using current-voltage characteristic curve tracking technology to quantify the degree of deviation of module output characteristics from the standard curve, thereby judging systematic faults such as PID effect (Potential Induced Degradation), encapsulation material aging, or poor connection of the junction box. The monitoring system can dynamically analyze the fault evolution law through time-series data comparison and pattern recognition algorithms, provide a scientific basis at the level of failure mechanism for operation and maintenance decisions, effectively reduce energy loss, and extend the service life of the module.

[0003] During the operation of the photovoltaic module fault monitoring system in equipment control power supply and distribution circuit devices, the insulation monitoring device is vulnerable to interference from dynamic fluctuations of leakage current in a humid or salt spray environment, resulting in misjudgment of insulation resistance and affecting the reliability of fault isolation; in addition, the coordinated control of the intelligent circuit breaker and the monitoring system needs to balance rapid fault removal and transient overvoltage suppression, but in complex working conditions, it may lead to a mismatch between the protection action and the fault evolution rate due to communication delay or insufficient algorithm convergence, increasing the system-level cascading risk. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a photovoltaic module fault monitoring system and method to solve the problem of misjudgment of insulation resistance caused by dynamic fluctuations of leakage current in a humid or salt spray environment for the insulation monitoring device, and the problem of system-level cascading risk caused by a mismatch between the protection action and the fault evolution rate due to communication delay or insufficient algorithm convergence in the coordinated control of the intelligent circuit breaker and the monitoring system.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: In a first aspect, a photovoltaic module fault monitoring system provided by the present invention includes: a collection module, a power supply dynamic compensation module, a control optimization module, a power supply module, and a digital twin module; The acquisition module is connected to the power supply dynamic compensation module through a distributed monitoring network, and is used to collect photovoltaic module data. The photovoltaic module data includes infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data. The photovoltaic module data is aligned in time series through a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix and transmit it to the power supply dynamic compensation module; The power supply dynamic compensation module receives the leakage current signal in the dynamic health status matrix, dynamically adjusts the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal through a preset improved whale optimization algorithm, generates a filtered leakage current signal, and combines the preset photovoltaic string voltage level output by the acquisition module to calculate the insulation resistance value, generating a dynamic insulation criterion; when the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and a preset DC circuit breaker is linked to execute a gradient tripping strategy based on the preset photovoltaic string voltage level meter; The control optimization module receives the alarm signal triggered by the power supply dynamic compensation module, and optimizes the parsing priority of the tripping control instruction and the arc extinguishing chamber impedance adjustment instruction generated by the alarm signal based on the communication path weight of the preset time-sensitive network; matches the circuit breaker tripping action time window corresponding to the tripping control instruction with the arc extinguishing chamber impedance change curve through a preset dynamic time warping algorithm, and adjusts the arc extinguishing chamber impedance parameter to suppress arc re-ignition; The power supply module receives the infrared hot spot distribution thermal map and the abnormal analysis result of the current-voltage curve generated by the acquisition module, performs feature-level fusion with the component aging trend prediction data output by the digital twin module, generates a power supply loop insulation deterioration assessment result, and drives the dynamic adjustment of the overcurrent protection threshold of the DC power distribution cabinet; The digital twin module receives the dynamic health status matrix, constructs a digital twin model of the photovoltaic array including temperature field distribution and defect space mapping, synchronizes the insulation deterioration assessment result output by the power supply module to a preset virtual system through a preset transfer learning algorithm, generates a preventive maintenance instruction and feeds it back to the spatio-temporal graph convolutional network model of the power supply module.

[0006] Furthermore, for the photovoltaic module fault monitoring system of the present invention, the digital twin module is further used for: Receiving the dynamic health status matrix generated by the acquisition module, and constructing a digital twin model of the photovoltaic array including temperature field distribution data, defect space mapping diagram, and compensated current and voltage curve data; Synchronizing the temperature field distribution data and the defect space mapping diagram in the dynamic health status matrix to a preset virtual system through a preset transfer learning algorithm. The preset virtual system is used to simulate the fault diffusion path to generate a preventive maintenance instruction; The preventive maintenance instruction is input into the spatio-temporal graph convolutional network model in the power supply module to optimize the confidence level of the component aging trend prediction result generated by the power supply module.

[0007] Further, for the photovoltaic module fault monitoring system of the present invention, the acquisition module includes: An infrared thermal imaging sensor, configured to collect temperature field distribution data of the photovoltaic module, and eliminate environmental noise interference through a preset improved variational mode decomposition algorithm to generate denoised temperature field data; An electroluminescence detection unit, configured to process the electroluminescence image of the photovoltaic module based on a preset adaptive threshold segmentation algorithm, and generate a defect space mapping diagram in combination with the electrical topology structure of the photovoltaic module; A current and voltage characteristic curve acquisition module, configured to compensate for line impedance errors through an adversarial generation network and generate compensated current and voltage characteristic curve data; The denoised temperature field data, the defect space mapping diagram, and the compensated current and voltage characteristic curve data are input into a preset bidirectional long short-term memory network, and multi-source data time series alignment is performed based on the operation timestamp of the photovoltaic module to generate a dynamic health state matrix.

[0008] Further, for the photovoltaic module fault monitoring system of the present invention, the power supply dynamic compensation module includes: An adaptive filtering unit, configured to receive the leakage current signal in the dynamic health state matrix, dynamically adjust the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal, and output a filtered leakage current signal; An environmental parameter and filtering threshold mapping table, configured to correct the filtering bandwidth of the adaptive filtering unit through a preset fuzzy logic controller in combination with temperature and humidity sensor data; A random forest regression model, configured to input the filtered leakage current signal into a pre-trained regression model, and calculate the insulation resistance value and generate a dynamic insulation criterion in combination with the preset photovoltaic string voltage level in the current and voltage characteristic curve data generated by the acquisition module.

[0009] Further, for the photovoltaic module fault monitoring system of the present invention, the control optimization module includes: An edge computing node, configured to optimize the communication path selection strategy of a time-sensitive network using a federated learning framework and generate optimized communication path weights; A deep Q-network algorithm unit, configured to dynamically adjust the parsing weights of the opening control instruction and the arc extinguishing chamber impedance adjustment instruction based on the fault feature map generated from the abnormal hot spot distribution heat map output by the power supply module and the analysis result of the current-voltage curve abnormality; A dynamic time warping algorithm unit, configured to predict the best time window for the circuit breaker opening action according to the optimized communication path weights, and absorb the energy of the inductive load through a pre-charge circuit; A model predictive control unit is used to match the arc impedance change curve corresponding to the time window predicted by the dynamic time warping algorithm unit in the arc extinguishing chamber, suppressing the risk of transient overvoltage and arc re-ignition.

[0010] Furthermore, for the photovoltaic module fault monitoring system of the present invention, the power supply module includes: An improved U-Net network is used to perform pixel-level segmentation on the infrared thermal image generated by the acquisition module, generating an abnormal hot spot distribution heat map; An SVM model enhanced by an attention mechanism is used to analyze the abnormal fluctuation characteristics in the current and voltage characteristic curve data generated by the acquisition module, identifying power attenuation caused by the PID effect; A spatio-temporal graph convolutional network is used to perform spatio-temporal feature fusion on the component aging trend prediction data output by the digital twin module and the abnormal hot spot distribution heat map, modeling the insulation deterioration inflection point and driving the dynamic adjustment of the overcurrent protection threshold of the DC power distribution cabinet.

[0011] Furthermore, for the photovoltaic module fault monitoring system of the present invention, the power supply module further includes: A knowledge distillation unit is used to perform knowledge distillation on the multi-scale feature parameters extracted by the improved U-Net network during the infrared thermal image segmentation process and the classification weights generated by the SVM model enhanced by the attention mechanism during the current-voltage curve analysis.

[0012] Furthermore, for the photovoltaic module fault monitoring system of the present invention, the digital twin module includes: A heterogeneous algorithm federated engine is used to run the deep residual network and the Vision Transformer model in parallel, fusing the conflicting evidence in the abnormal hot spot distribution heat map and the current-voltage abnormal analysis result generated by the power supply module through the Dempster-Shafer evidence theory, and outputting a multi-model joint diagnosis result; An improved A* algorithm unit is used to switch to the local cache data of the acquisition module in case of communication anomaly, and execute a fast protection logic based on the temperature field distribution data and the defect space mapping diagram in the dynamic health status matrix, triggering the inverter derating operation mode.

[0013] Furthermore, for the photovoltaic module fault monitoring system of the present invention, the digital twin module further includes: A transfer learning unit is used to transfer the temperature field distribution data and the defect space mapping diagram in the dynamic health status matrix generated by the acquisition module to a preset virtual system through a domain adaptation algorithm. Based on the transferred temperature field distribution data and the defect space mapping diagram, it simulates the thermal aging and defect diffusion path of the insulating material, generating preventive maintenance instructions for the power supply module.

[0014] In a second aspect, the present invention provides a photovoltaic module fault monitoring method, which is applied to the photovoltaic module fault monitoring system, and comprises the following steps: Collect PV module data, including infrared thermal imaging data, electroluminescent defect data, and current-voltage characteristic curve data. Use a preset multi-scale feature alignment algorithm to perform time-series alignment on the PV module data to generate a dynamic health status matrix. Receive the leakage current signal, dynamically adjust the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal through the preset improved whale optimization algorithm, generate the filtered leakage current signal, and calculate the insulation resistance value in combination with the preset photovoltaic string voltage level to generate the dynamic insulation criterion. When the dynamic insulation criterion is lower than the preset threshold, trigger the alarm signal, and link the preset DC circuit breaker to execute the gradient tripping strategy based on the preset photovoltaic string voltage level meter; Receive an alarm signal, optimize the parsing priority of the opening control instruction and the arc extinguishing chamber impedance adjustment instruction generated by the alarm signal based on the communication path weight of the preset time-sensitive network, match the circuit breaker opening action time window corresponding to the opening control instruction and the arc extinguishing chamber impedance change curve through a preset dynamic time warping algorithm, and adjust the arc extinguishing chamber impedance parameter to suppress the reignition of the arc; Receive infrared hot spot distribution thermogram, current and voltage curve abnormal analysis results and component aging trend prediction data, perform feature-level fusion on the infrared hot spot distribution thermogram, current and voltage curve abnormal analysis results and component aging trend prediction data, and generate power supply circuit insulation degradation assessment results, which are used to drive the dynamic adjustment of the overcurrent protection threshold of the preset DC distribution cabinet; The dynamic health status matrix is ​​received, and a digital twin model of the photovoltaic array including temperature field distribution and defect space mapping is constructed. The insulation degradation assessment result output by the power supply module is synchronized to a preset virtual system through a preset transfer learning algorithm, and preventive maintenance instructions are generated and fed back to the spatiotemporal graph convolutional network model.

[0015] Beneficial effects of the present invention: The present invention dynamically adjusts the parameters of wavelet basis functions by improving the whale optimization algorithm, combines a fuzzy logic controller to correct the filtering bandwidth based on temperature and humidity data, accurately separates the leakage current noise components and real insulation fault characteristics in a humid or salt spray environment, generates a dynamic insulation criterion, and reduces the risk of misjudgment of insulation resistance; optimizes the communication path weights based on the federated learning framework of a preset time-sensitive network, combines the dynamic time warping algorithm to predict the circuit breaker opening time window, and adjusts the impedance characteristics of the arc extinguishing chamber in real time through model predictive control to achieve spatio-temporal matching of protection actions and fault evolution, and suppress the risk of arc re-ignition and system-level cascading; the digital twin module synchronizes multi-dimensional monitoring data and virtual simulation results through transfer learning, uses the Monte Carlo method to pre-determine the fault diffusion path to generate preventive maintenance instructions, and feeds back to the spatio-temporal graph convolutional network to optimize the aging trend prediction, forming a self-evolving monitoring system with closed-loop optimization; the heterogeneous algorithm federated engine fuses the multi-model diagnostic results of infrared hot spot distribution and current-voltage anomalies, combines the improved A* algorithm to construct an online monitoring and offline fault-tolerant dual-mode mechanism, and improves the fault diagnosis accuracy and system reliability under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on the drawings without creative efforts.

[0017] Figure 1 It is a flowchart of a photovoltaic module fault monitoring method provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention will be described in detail below with reference to the drawings. To better understand the objectives of the present invention, the present invention will be further described in detail below.

[0019] In a first aspect, a photovoltaic module fault monitoring system provided by the present invention includes: a collection module, a power supply dynamic compensation module, a control optimization module, a power supply module, and a digital twin module; The acquisition module is connected to the power supply dynamic compensation module through a distributed monitoring network, and is used to collect photovoltaic module data. The photovoltaic module data includes infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data. The photovoltaic module data is aligned in time series through a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix and transmit it to the power supply dynamic compensation module; The acquisition module is connected to the power supply dynamic compensation module through a distributed monitoring network. The infrared thermal imaging sensor obtains the surface temperature field distribution data of the photovoltaic module at a preset sampling frequency. The improved variational mode decomposition algorithm performs mode decomposition on the original temperature data. By adaptively adjusting the penalty factor and the number of modes, and combining the kurtosis-energy joint criterion to screen out the effective temperature feature modes, the denoised temperature gradient distribution data is reconstructed. The reconstructed data is mapped to the spatial coordinates of the radiator layout parameters of the photovoltaic module to generate a temperature field distribution map with physical position relevance, providing basic data support for hotspot effect analysis.

[0020] After the electroluminescence detection unit collects the electroluminescence image of the photovoltaic module, it dynamically adjusts the image segmentation parameters based on the adaptive threshold segmentation algorithm. The algorithm expands the defect boundary contour through the region growing method according to the electroluminescence intensity distribution characteristics, and combines the electrical topological connection relationship of the battery cells of the module to establish a spatial mapping model between the defect region and the circuit nodes. The generated defect spatial mapping diagram annotates the geometric morphology and electrical isolation state of the hidden crack and broken grid defects, and quantitatively evaluates the influence weight of the defects on the output characteristics of the string, providing microstructure data for insulation degradation analysis.

[0021] The current and voltage characteristic curve acquisition module monitors the electrical output characteristics of the photovoltaic module in real time, and constructs a line impedance compensation model through a generative adversarial network. The generator network simulates the current and voltage characteristic curves under ideal line conditions, and the discriminator network compares the feature differences between the measured data and the generated data to dynamically correct the measurement deviation caused by the line impedance. The compensated current and voltage characteristic curve data uses a sliding window mechanism to intercept the key inflection points of the characteristic curve, retaining the characteristic information of the module in transient processes such as shadow occlusion and bypass diode operation, forming a high-precision electrical parameter data set.

[0022] The multi-scale feature alignment algorithm performs temporal alignment processing on the above multi-source data. By calculating the feature point matching path at different sensor sampling rates through dynamic time warping, it aligns the temporal benchmarks of the temperature field distribution, defect spatial mapping, and electrical parameters. The forward and backward recurrent units of the bidirectional long short-term memory network extract feature vectors at different time scales respectively, and establish a spatio-temporal correlation model for temperature gradient changes, defect propagation rates, and current-voltage fluctuations. The feature fusion layer splices the spatio-temporal feature vectors of the multi-source data in dimensions, and combines the operation timestamp of the photovoltaic system to establish a unified temporal benchmark, generating a dynamic health status matrix containing thermo-electric coupling features, defect evolution trends, and component health indices. This matrix is transmitted to the power supply dynamic compensation module through a distributed monitoring network, providing multi-dimensional data input for insulation state assessment and protection strategy generation.

[0023] The power supply dynamic compensation module receives the leakage current signal in the dynamic health status matrix, dynamically adjusts the wavelet packet decomposition layer number and threshold parameters of the wavelet basis function of the leakage current signal through a preset improved whale optimization algorithm, generates a filtered leakage current signal, and calculates the insulation resistance value in combination with the preset photovoltaic string voltage level output by the acquisition module, generating a dynamic insulation criterion; when the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and a gradient tripping strategy based on the preset photovoltaic string voltage level is linked to execute a preset DC circuit breaker; After the power supply dynamic compensation module receives the leakage current signal in the dynamic health status matrix, the improved whale optimization algorithm adaptively adjusts the scale factor and displacement parameters of the wavelet basis function by introducing a dynamic weight factor and a simulated annealing mechanism. The algorithm balances the global search and local development capabilities during the iteration process, dynamically optimizes the wavelet packet decomposition layer number according to the spectral characteristics of the leakage current signal, and separates the high-frequency transient fault components and low-frequency environmental noise. The optimized wavelet basis function parameters change dynamically with the signal, improving the signal decomposition accuracy and outputting a filtered leakage current signal, providing a high signal-to-noise ratio data basis for subsequent insulation state analysis.

[0024] The environmental parameter and filtering threshold mapping table receives the environmental data collected by the temperature and humidity sensors in real time, and constructs an association rule library for temperature-humidity-filtering bandwidth through a fuzzy logic controller. The controller dynamically adjusts the boundary conditions of the membership function according to the environmental humidity change rate and temperature gradient value, and calculates the filtering bandwidth correction coefficient. Under high humidity conditions, the fuzzy rule drives the adaptive filtering unit to expand the passband range and suppress the misjudgment caused by low-frequency leakage current fluctuations; in a dry environment, the bandwidth is contracted to enhance the resolution of high-frequency fault components. The corrected filtering parameters are synchronously updated to the wavelet basis function adjustment module, forming an environment-adaptive signal decomposition mechanism.

[0025] The random forest regression model extracts the time domain statistical characteristics and frequency domain energy distribution characteristics of the filtered leakage current signal, and constructs a multidimensional feature vector in combination with the photovoltaic string voltage level parameters output by the acquisition module. The model dynamically allocates feature weights through the Gini coefficient, prioritizes the nonlinear mapping relationship between the voltage level and the leakage current amplitude, and calculates the equivalent resistance of the insulation resistance. When the insulation resistance value is detected to be lower than the dynamic threshold, the model divides the fault risk area according to the voltage level and generates a graded insulation criterion containing fault location information and severity. The criterion data is input into the protection strategy generation module, triggering the corresponding level of alarm signal and linking the circuit breaker to perform a gradient tripping action, giving priority to isolating the high-risk insulation fault circuit.

[0026] The gradient tripping strategy prioritizes fault circuits based on the voltage level of the photovoltaic strings. High-voltage strings correspond to high-risk fault areas and trigger fast tripping instructions; low-voltage strings use delayed tripping logic. After receiving the graded alarm signal, the DC circuit breaker performs the tripping action according to the preset voltage-time mapping relationship, absorbs the energy of the inductive load through the pre-charging circuit, and reduces the risk of arc reignition during the tripping process. The graded protection mechanism avoids the false operation of a single threshold criterion in a complex environment, while reducing unnecessary power outages of non-fault circuits and maintaining the continuity of system power supply.

[0027] The control optimization module receives the alarm signal triggered by the power supply dynamic compensation module, optimizes the parsing priority of the opening control instruction and the arc extinguishing chamber impedance adjustment instruction generated by the alarm signal based on the communication path weight of the preset time-sensitive network; matches the circuit breaker opening action time window corresponding to the opening control instruction with the arc extinguishing chamber impedance change curve through a preset dynamic time warping algorithm, and adjusts the arc extinguishing chamber impedance parameters to suppress the reignition of the arc; The control optimization module deploys the federated learning framework through edge computing nodes to achieve dynamic optimization of communication path weights under the time-sensitive network architecture. The federated learning framework adopts the distributed edge device local model training and global parameter aggregation mechanism. Each edge node builds a path quality assessment matrix based on the local network topology state and data packet transmission delay characteristics, and uploads it to the global model through encrypted gradient parameters for weighted average fusion to generate optimized communication path weights. The optimized weights are synchronized to the dynamic time warping algorithm unit, and the high-reliability communication link transmission switch control instructions and arc chamber impedance adjustment instructions are preferentially allocated to reduce the transmission delay of key protection instructions.

[0028] The deep Q-network algorithm unit receives the abnormal hot spot distribution heat map output by the power supply module and the analysis results of the abnormal current-voltage curve, and analyzes the temperature gradient of the hot spot area, the current imbalance degree, and the PID effect classification probability parameters in the fault feature map. The algorithm constructs a state space with dimensions of fault type, position coordinates, and risk level, and designs a reward function based on the action response speed and fault suppression effect in combination with the protection action effect data in the historical fault handling case library. Through the dual network structure, the exploration and exploitation strategies are balanced, the parsing weights of overcurrent protection, island detection, and insulation locking instructions are dynamically adjusted, and the protection action sequence matching the current fault evolution rate is preferentially scheduled.

[0029] Based on the optimized communication path weight parameters, the dynamic time warping algorithm unit matches the mechanical response curve of the circuit breaker opening mechanism and the timing of the fault current zero crossing. The algorithm aligns the moving trajectory of the circuit breaker contacts and the arc energy decay curve through the dynamic bending path function, and calculates the opening time window that meets the natural zero crossing requirement of the arc current. The pre-charge circuit generates a buffer capacitor charge and discharge control pulse sequence according to the time window parameters, absorbs the magnetic energy released by the inductive load instantaneously during opening in stages, and reduces the probability of arc re-ignition during the contact separation process.

[0030] The model predictive control unit monitors the arc voltage and ion concentration parameters in the arc extinguishing chamber in real time, and constructs a multi-physical field coupling prediction model including the dynamic change characteristics of the arc impedance. Based on the opening time window parameters output by the dynamic time warping algorithm, the control unit calculates the optimal control quantity sequence of the opening degree of the arc extinguishing chamber nozzle, the current of the magnetic blow coil, and the flow rate of the gas blowing medium by using the rolling horizon optimization strategy. During the execution process, the arc impedance characteristics are sampled at high frequency, the model prediction deviation is dynamically corrected, the excitation current intensity of the magnetic blow coil is adjusted to match the arc movement trajectory, and the rising rate of the transient recovery voltage is suppressed. The arc extinguishing chamber pressure feedback signal is synchronously input to the control model to optimize the flow distribution strategy of the gas blowing medium and block the ionization conditions required for arc re-ignition.

[0031] The power supply module receives the infrared hot spot distribution heat map generated by the acquisition module and the analysis results of the abnormal current-voltage curve, performs feature-level fusion with the component aging trend prediction data output by the digital twin module, generates the insulation degradation assessment result of the power supply circuit, and drives the dynamic adjustment of the overcurrent protection threshold of the DC power distribution cabinet; After the power supply module receives the infrared hot spot distribution thermal map generated by the acquisition module, the improved U-Net network extracts the boundary features of the hot spot area at different resolutions through multi-level convolutional operations of the encoder. The decoder uses a skip connection structure to fuse the shallow texture features and deep semantic information to generate a pixel-level segmentation result. The segmented thermal map is spatially aligned with the string current imbalance parameter obtained by the current and voltage characteristic curve acquisition module, and the temperature gradient value and the corresponding current deviation ratio of the hot spot area are marked to form an abnormal distribution thermal map for quantitatively evaluating the severity of the hot spot, providing a spatial positioning basis for insulation degradation analysis.

[0032] When the SVM model enhanced by the attention mechanism analyzes the current and voltage characteristic curve data, it uses the multi-head attention mechanism to dynamically allocate the feature weights of different sections of the curve, focusing on the section where the open circuit voltage drops abnormally under low irradiance. The model projects the high-dimensional features to the low-dimensional space through the kernel function mapping, combines the soft margin classification strategy to construct the PID effect recognition boundary, and generates a diagnostic report including the power attenuation rate and reversibility evaluation. The diagnostic results are associated with the maximum power point tracking log data of the inverter, and the correlation between the PID effect and the insulation material aging rate is established, providing electrical anomaly feature input for the protection threshold adjustment.

[0033] The spatio-temporal graph convolutional network receives the component aging trend prediction data output by the digital twin module, and constructs a spatio-temporal topology graph with the photovoltaic string as the node and the electrical connection relationship as the edge. The time-axis convolutional layer of the network extracts the temporal decay characteristics of the electrical conductivity and dielectric constant of the insulation material, and the spatial graph convolutional layer captures the aging synergy effect between adjacent strings. The feature fusion layer performs cross-modal stitching on the temperature gradient data of the abnormal hot spot distribution thermal map and the aging trend prediction result, and establishes a non-linear mapping model between the hot spot temperature rise rate and insulation degradation. The insulation degradation inflection point prediction data output by the network is input into the protection threshold adjustment module of the DC power distribution cabinet, and the action delay threshold of the ground fault protection is dynamically corrected according to the environmental temperature and humidity parameters, realizing the adaptive matching of the overcurrent protection threshold and the insulation resistance decay curve.

[0034] The digital twin module receives the dynamic health status matrix, constructs a digital twin model of the photovoltaic array including the temperature field distribution and defect space mapping, synchronizes the insulation degradation evaluation result output by the power supply module to a preset virtual system through a preset transfer learning algorithm, generates a preventive maintenance instruction and feeds it back to the spatio-temporal graph convolutional network model of the power supply module.

[0035] After receiving the dynamic health status matrix, the digital twin module extracts the temperature field distribution data and defect space mapping diagram contained in the matrix. The temperature field distribution data eliminates environmental noise interference through an improved variational mode decomposition algorithm, generating temperature gradient distribution data with spatial coordinate mapping relationships; the defect space mapping diagram combines the electrical topological structure and uses an adaptive threshold segmentation algorithm to label the geometric morphology and electrical isolation status of hidden cracks and broken grid defects. The above data is input into the construction module of the photovoltaic array digital twin model to establish a virtual system topological structure with strings as nodes and electrical connections as edges, synchronously mapping the temperature field and defect distribution characteristics of the real system.

[0036] The transfer learning algorithm reduces the feature distribution difference between the real system and the virtual system through a domain adversarial training strategy. The algorithm introduces a gradient reversal operation in the feature extraction layer, forcing the encoder to generate shared feature representations independent of the data source. The temperature field distribution data of the real system is aligned with the thermal conduction simulation results of the virtual system, combined with the electroluminescence defect characteristics and the microstructure model parameters of the virtual components, to generate a defect evolution map with physical consistency. The virtual system constructs a thermal-electrical coupling simulation model based on the aligned dataset and simulates the co-evolution path of insulation material thermal aging and defect diffusion through a hidden Markov chain.

[0037] The preventive maintenance instruction generation module triggers conditions according to the key parameter thresholds in the fault evolution simulation results and uses a random forest classifier to evaluate the risk levels of each evolution path. The input parameters of the classifier include the predicted insulation resistance value, hot spot area expansion rate, and string current imbalance degree in the virtual system, and the output is a set of maintenance strategies including component replacement priorities, cleaning cycle suggestions, and protection threshold adjustment parameters. The generated maintenance instructions are input into the spatio-temporal graph convolutional network model of the power supply module, and the spatio-temporal correlation of aging trend prediction is optimized by updating the material aging coefficient and environmental stress weight parameters in the network node attributes.

[0038] The spatio-temporal graph convolutional network fuses the historical repair record data in the maintenance instructions in the time dimension to establish a dynamic association model between component performance recovery and subsequent aging rates; in the spatial dimension, it combines the string electrical connection topology to predict the cross-regional impact of local hot spots on the insulation performance of adjacent components. The optimized prediction results are fed back to the parameter correction module of the digital twin model to adjust the degradation equation coefficients of the virtual component material properties, forming a closed-loop optimization mechanism for monitoring data and simulation prediction. The protection threshold adjustment parameter in the maintenance instruction drives the dynamic protection logic of the DC power distribution cabinet, matches the insulation resistance decay curve according to the environmental temperature and humidity conditions, and realizes the adaptive adjustment of the overcurrent protection action threshold.

[0039] The photovoltaic module fault monitoring system provided by the present invention realizes the fault monitoring and protection control of the photovoltaic array through the cooperation of multiple modules. The acquisition module obtains the infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data of the photovoltaic modules through a distributed monitoring network, and preprocesses the multi-source data using a multi-scale feature alignment algorithm. The multi-scale feature alignment algorithm intercepts the time-series segments of the data of each sensor through a sliding window mechanism, and uses dynamic time warping to calculate the feature point matching paths at different sensor sampling rates, aligning the time-series benchmarks of the temperature field distribution, defect spatial mapping, and electrical parameters. The preprocessed data is input into a bidirectional long short-term memory network for cross-modal feature fusion, extracting the spatio-temporal correlation features of the temperature gradient, defect expansion rate, and current-voltage fluctuation amplitude, and generating a dynamic health state matrix containing multi-dimensional health indicators.

[0040] The power supply dynamic compensation module receives the leakage current signal in the dynamic health state matrix and dynamically adjusts the scale factor and displacement parameter of the wavelet basis function using an improved whale optimization algorithm. The improved whale optimization algorithm introduces a simulated annealing mechanism during the iteration process, balances the global search and local development capabilities through a dynamic weight factor, and adaptively optimizes the number of layers of wavelet packet decomposition and the frequency band division accuracy. The optimized leakage current signal is input into a random forest regression model, and combined with the photovoltaic string voltage level data output by the acquisition module, an insulation resistance calculation model based on the non-linear mapping relationship between voltage and leakage current is constructed. The model dynamically allocates feature weights through the Gini coefficient, preferentially associates the leakage current amplitude features at high voltage levels, and generates a dynamic insulation criterion reflecting the change in the conductivity of the insulating material. When the insulation criterion is lower than the preset threshold, the gradient tripping strategy divides the fault loop priority according to the string voltage level, triggering the DC circuit breaker to perform a hierarchical tripping action.

[0041] The control optimization module optimizes the transmission path of the tripping control instruction based on the preset time-sensitive network architecture, and uses a federated learning framework to perform distributed training and global aggregation of the communication path weight parameters among edge computing nodes. The federated learning framework encrypts the local model parameter update process through differential privacy technology, and dynamically adjusts the parsing priority of the protection instruction in combination with the network topology state. The dynamic time warping algorithm receives the optimized communication path weight parameters, matches the mechanical response curve of the circuit breaker with the time sequence of the zero crossing point of the fault current, and predicts the tripping time window that meets the arc energy suppression requirements. The arc chamber impedance adjustment instruction is generated according to the time window parameters, and the model predictive control unit monitors the arc impedance change curve in real time, and adjusts the magnetic blow coil current and the gas blowing medium flow parameters through a rolling horizon optimization strategy to suppress the transient overvoltage and block the arc re-ignition condition.

[0042] The power supply module receives the infrared hot spot distribution thermal map and the analysis results of abnormal current and voltage curves generated by the acquisition module, and uses an improved U-Net network to perform pixel-level segmentation on the thermal image. The network fuses shallow texture features and deep semantic information through a skip connection structure, and combines the string current imbalance parameter to label the temperature gradient and power loss ratio of the hot spot area, generating an abnormal hot spot distribution thermal map. The attention mechanism enhanced SVM model focuses on the local abnormal features of the current-voltage characteristic curve, and extracts the open circuit voltage drop mode caused by the PID effect through kernel function mapping. The spatio-temporal graph convolutional network receives the component aging trend prediction data output by the digital twin module, establishes a spatio-temporal correlation model between the insulation material deterioration and the hot spot temperature rise rate, and drives the overcurrent protection threshold of the DC power distribution cabinet to be dynamically adjusted according to the environmental temperature and humidity parameters.

[0043] The digital twin module constructs a digital twin model of the photovoltaic array, and synchronizes the temperature field distribution, defect space mapping, and electrical parameter data of the real system and the virtual system through a transfer learning algorithm. The transfer learning algorithm adopts a domain adversarial training strategy to reduce the feature distribution difference between the real data and the simulation data, generating a virtual component state data set with physical consistency. The digital twin model simulates the fault evolution path under different environmental stresses based on the Monte Carlo method, and combines the Markov chain model to pre-evolve the synergistic effect of hot spot diffusion and insulation deterioration. The preventive maintenance instruction generation module triggers conditions according to the key parameter thresholds in the simulation results, and outputs a maintenance strategy including the component replacement priority and the protection threshold adjustment parameters. The maintenance instruction is fed back to the spatio-temporal graph convolutional network in the power supply module to optimize the material aging coefficient in the node attributes, forming a closed-loop optimization mechanism for the monitoring data and the prediction model. Specifically, for the photovoltaic component fault monitoring system described in the present invention, the digital twin module is further configured to: Receive the dynamic health state matrix generated by the acquisition module, and construct a digital twin model of the photovoltaic array including temperature field distribution data, defect space mapping diagram, and compensated current and voltage curve data; Synchronize the temperature field distribution data and the defect space mapping diagram in the dynamic health state matrix to a preset virtual system through a preset transfer learning algorithm, and the preset virtual system is used to simulate the fault diffusion path to generate preventive maintenance instructions; The preventive maintenance instruction is input into the spatio-temporal graph convolutional network model in the power supply module to optimize the confidence of the component aging trend prediction result generated by the power supply module.

[0044] The digital twin module receives the dynamic health status matrix generated by the acquisition module, and extracts the temperature field distribution data, defect space mapping diagram, and compensated current and voltage characteristic curve data contained in the matrix. The temperature field distribution data eliminates environmental noise interference through an improved variational mode decomposition algorithm, and generates a temperature gradient distribution diagram with a physical coordinate mapping relationship; the defect space mapping diagram combines the electrical topological structure, and uses an adaptive threshold segmentation algorithm to label the geometric morphology and electrical isolation status of hidden cracks and broken grid defects; the compensated current and voltage characteristic curve data corrects the line impedance error through a generative adversarial network, and retains the transient output characteristics of the components. The above data is input into the construction module of the digital twin model through a timestamp alignment mechanism, and a virtual system topological structure with photovoltaic strings as nodes and electrical connection relationships as edges is established, synchronously mapping the temperature field, defect distribution, and electrical parameter characteristics of the real system.

[0045] The transfer learning algorithm reduces the feature distribution difference between the real system and the virtual system through a domain adversarial training strategy. The algorithm introduces a gradient reversal layer in the feature extraction layer, forcing the feature encoder to generate a shared feature representation independent of the data source. The temperature field distribution data of the real system is aligned with the thermal conduction simulation results of the virtual system, and combined with the electroluminescence defect characteristics and the microscopic structure model parameters of the virtual components, to generate a defect evolution map with physical consistency. The virtual system constructs a Markov chain model based on the aligned dataset to simulate the fault diffusion path under different environmental stresses such as salt fog concentration, temperature, and humidity. The Monte Carlo method generates multiple groups of fault evolution trajectories, and combines the coupling relationship between the temperature rise rate in the hot spot area and the attenuation of the electrical conductivity of the insulating material to predict the cross-region impact of hot spot diffusion on adjacent components.

[0046] The preventive maintenance instruction generation module triggers according to the key parameter threshold trigger conditions in the fault evolution simulation results, and uses a random forest classifier to evaluate the risk level of each evolution path. The input parameters of the classifier include the predicted value of the insulation resistance, the expansion rate of the hot spot area, and the current imbalance degree of the string in the virtual system, and the output is a set of maintenance strategies including the component replacement priority, the cleaning cycle suggestion, and the protection threshold adjustment parameters. The generated preventive maintenance instructions are input into the spatio-temporal graph convolutional network model of the power supply module, and the spatio-temporal correlation of the aging trend prediction is optimized by updating the material aging coefficient and the environmental stress weight parameters in the network node attributes. The spatio-temporal graph convolutional network fuses the historical repair record data in the maintenance instructions in the time dimension to establish a dynamic association model between component performance recovery and subsequent aging rate; in the space dimension, it combines the electrical connection topology of the string to predict the cross-region impact of local hot spots on the insulation performance of adjacent components. The optimized prediction results are fed back to the parameter correction module of the digital twin model to adjust the degradation equation coefficients of the virtual component material properties, forming a closed-loop optimization mechanism for monitoring data and prediction models. Specifically, for the photovoltaic module fault monitoring system described in the present invention, the acquisition module includes: An infrared thermal imaging sensor is used to collect temperature field distribution data of a photovoltaic module, and eliminate environmental noise interference through a preset improved variational mode decomposition algorithm to generate denoised temperature field data; An electroluminescence detection unit is used to process the electroluminescence image of the photovoltaic module based on a preset adaptive threshold segmentation algorithm, and generate a defect space mapping diagram in combination with the electrical topology structure of the photovoltaic module; A current-voltage characteristic curve acquisition module is used to compensate for line impedance errors through an adversarial generation network and generate compensated current-voltage characteristic curve data; The denoised temperature field data, the defect space mapping diagram, and the compensated current-voltage characteristic curve data are input into a preset bidirectional long short-term memory network, and multi-source data time series alignment is performed based on the operation timestamp of the photovoltaic module to generate a dynamic health state matrix.

[0047] The acquisition module obtains the surface temperature field distribution data of the photovoltaic module through an infrared thermal imaging sensor, and performs modal decomposition on the original temperature data using an improved variational mode decomposition algorithm. The improved variational mode decomposition algorithm adaptively adjusts the penalty factor and the number of modes, and combines the kurtosis-energy joint criterion to screen effective temperature feature modes, suppressing the influence of environmental radiation noise on the temperature gradient distribution. The reconstructed denoised temperature field data is subjected to spatial coordinate mapping with the radiator layout parameters of the photovoltaic module to generate a temperature field distribution map with physical position relevance, providing basic data support for hot spot effect analysis.

[0048] After the electroluminescence detection unit collects the electroluminescence image of the photovoltaic module, it dynamically adjusts the image segmentation parameters based on the adaptive threshold segmentation algorithm. The algorithm expands the defect boundary contour through the region growing method according to the electroluminescence intensity distribution characteristics, and combines the electrical topology connection relationship of the battery cells of the module to establish a spatial mapping model between the defect region and the circuit nodes. The generated defect space mapping diagram annotates the geometric morphology and electrical isolation state of the hidden crack and broken grid defects, quantifies and evaluates the influence weight of the defects on the output characteristics of the string, and provides microstructure data for subsequent insulation degradation analysis.

[0049] The current-voltage characteristic curve acquisition module monitors the electrical output characteristics of the photovoltaic module in real time, and constructs a line impedance compensation model through an adversarial generation network. The generator network simulates the current-voltage characteristic curve under ideal line conditions, and the discriminator network compares the feature differences between the measured data and the generated data to dynamically correct the measurement deviation caused by the line impedance. The compensated current-voltage characteristic curve data uses a sliding window mechanism to intercept the key inflection points of the characteristic curve, retaining the characteristic information of the module in transient processes such as shadow occlusion and bypass diode operation, and forming a high-precision electrical parameter dataset.

[0050] The denoised temperature field data, defect space mapping diagram, and compensated current and voltage characteristic curve data are input into a bidirectional long short-term memory network for time series alignment processing. The forward and backward recurrent units of the network extract feature vectors at different time scales respectively, and establish a time series correlation model for temperature field change, defect expansion, and electrical parameter attenuation. The feature fusion layer splices the spatio-temporal feature vectors of multi-source data in dimension, combines the operation timestamp of the photovoltaic system to establish a unified time series benchmark, and generates a dynamic health status matrix containing thermoelectric coupling features, defect evolution trends, and component health indices. This matrix is transmitted to subsequent modules through a distributed monitoring network, providing multi-dimensional data input for insulation status assessment and protection strategy generation. Specifically, for the photovoltaic module fault monitoring system described in the present invention, the power supply dynamic compensation module includes: An adaptive filtering unit, which is used to receive the leakage current signal in the dynamic health status matrix, dynamically adjust the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal, and output the filtered leakage current signal; An environmental parameter and filtering threshold mapping table, which is used to combine the temperature and humidity sensor data and correct the filtering bandwidth of the adaptive filtering unit through a preset fuzzy logic controller; A random forest regression model, which is used to input the filtered leakage current signal into a pre-trained regression model, combine the preset photovoltaic string voltage level in the current and voltage characteristic curve data generated by the acquisition module, calculate the insulation resistance value, and generate a dynamic insulation criterion.

[0051] The power supply dynamic compensation module receives the leakage current signal in the dynamic health status matrix, and performs signal decomposition and feature extraction through the adaptive filtering unit. The adaptive filtering unit uses an improved whale optimization algorithm to dynamically adjust the scale factor and displacement parameter of the wavelet basis function, and optimize the frequency band division accuracy of wavelet packet decomposition. The improved whale optimization algorithm introduces a dynamic weight factor and a simulated annealing mechanism during the iteration process to balance the global search and local development capabilities, and adaptively match the spectral characteristic changes of the leakage current signal. The optimized wavelet basis function parameters are dynamically adjusted with the signal, improving the separation accuracy of high-frequency transient components and low-frequency fluctuation noise, and outputting the filtered leakage current signal.

[0052] The environmental parameter and filter threshold mapping table is combined with the real-time environmental data collected by the temperature and humidity sensor, and the temperature-humidity-filter bandwidth association rule base is constructed through the fuzzy logic controller. The fuzzy logic controller dynamically adjusts the boundary conditions of the membership function according to the environmental humidity change rate and the temperature gradient value, and calculates the filter bandwidth correction coefficient. Under high humidity conditions, the fuzzy rule drives the adaptive filter unit to expand the passband range and suppress the misjudgment caused by low-frequency leakage current fluctuations; under dry conditions, the bandwidth is reduced to enhance the resolution of high-frequency fault components. The corrected filter parameters are synchronously updated to the wavelet basis function adjustment module to form an environmentally adaptive signal decomposition mechanism.

[0053] The random forest regression model receives the filtered leakage current signal, extracts the time domain statistical characteristics and frequency domain energy distribution characteristics of the signal, and constructs a multidimensional feature vector by combining the photovoltaic string voltage level parameters in the current and voltage characteristic curve data generated by the acquisition module. The model dynamically adjusts the feature weight distribution through the Gini coefficient, preferentially associates the nonlinear mapping relationship between the voltage level and the leakage current amplitude, and calculates the equivalent resistance of the insulation resistance. When the insulation resistance value is detected to be lower than the dynamic threshold, the model divides the fault risk area according to the voltage level and generates a graded insulation criterion containing fault location information and severity. The criterion data is input into the protection strategy generation module, triggering the corresponding level of alarm signal and linking the circuit breaker to perform gradient tripping action, giving priority to isolating the high-risk insulation fault circuit. Specifically, in the photovoltaic module fault monitoring system of the present invention, the control optimization module includes: Edge computing nodes are used to optimize the communication path selection strategy of the time-sensitive network using a federated learning framework and generate optimized communication path weights; A deep Q network algorithm unit, used to dynamically adjust the analytical weights of the opening control command and the arc extinguishing chamber impedance adjustment command based on the abnormal hot spot distribution thermal map output by the power supply module and the fault characteristic map generated by the abnormal analysis result of the current and voltage curve; A dynamic time warping algorithm unit, used for predicting the optimal time window of the circuit breaker opening action according to the optimized communication path weight, and absorbing the inductive load energy through the pre-charging circuit; The model prediction control unit is used to match the arc impedance change curve corresponding to the time window predicted by the dynamic time warping algorithm unit in the arc extinguishing chamber to suppress transient overvoltage and arc restrike risks.

[0054] The control optimization module deploys a federated learning framework through edge computing nodes to achieve dynamic optimization of communication path weights under a time-sensitive network architecture. The federated learning framework adopts a mechanism of distributed local model training of edge devices and global parameter aggregation. Each edge node constructs a path quality evaluation matrix according to the local network topology state and the characteristics of packet transmission delay, and uploads encrypted gradient parameters to the global model for weighted average fusion to generate optimized communication path weights. The optimized weights are synchronized to the dynamic time warping algorithm unit, which preferentially allocates high-reliability communication links to transmit the breaker control instruction and the arc extinguishing chamber impedance adjustment instruction, reducing the transmission delay of critical protection instructions.

[0055] The deep Q-network algorithm unit receives the abnormal hot spot distribution heat map output by the power supply module and the analysis results of the abnormal current-voltage curve, and analyzes the temperature gradient of the hot spot area, the current imbalance degree, and the PID effect classification probability parameters in the fault feature map. The algorithm constructs a state space with dimensions of fault type, position coordinates, and risk level, and designs a reward function based on the action response speed and the fault suppression effect in combination with the protection action effect data in the historical fault handling case library. Through a dual-network structure to balance the exploration and exploitation strategies, the parsing weights of overcurrent protection, island detection, and insulation locking instructions are dynamically adjusted, and the protection action sequence matching the current fault evolution rate is preferentially scheduled.

[0056] Based on the optimized communication path weight parameters, the dynamic time warping algorithm unit matches the mechanical response curve of the breaker opening mechanism and the zero-crossing timing of the fault current. The algorithm aligns the breaker contact movement trajectory and the arc energy decay curve through a dynamic bending path function, and calculates the opening time window that meets the natural zero-crossing requirement of the arc current. The pre-charge circuit generates a buffer capacitor charge and discharge control pulse sequence according to the time window parameters, and absorbs the magnetic energy released by the inductive load instantaneously during opening in stages, reducing the probability of arc re-ignition during the contact separation process.

[0057] The model predictive control unit monitors the arc voltage and ion concentration parameters in the arc extinguishing chamber in real time, and constructs a multi-physical field coupling prediction model including the dynamic change characteristics of the arc impedance. Based on the opening time window parameters output by the dynamic time warping algorithm, the control unit calculates the optimal control quantity sequence of the arc extinguishing chamber nozzle opening, the magnetic blow coil current, and the gas-blowing medium flow rate by using a rolling horizon optimization strategy. During the execution process, the arc impedance characteristics are sampled at high frequency to dynamically correct the model prediction deviation, and the excitation current intensity of the magnetic blow coil is adjusted to match the arc movement trajectory, suppressing the rising rate of the transient recovery voltage. The arc extinguishing chamber pressure feedback signal is synchronously input to the control model to optimize the flow distribution strategy of the gas-blowing medium, blocking the ionization conditions required for arc re-ignition. Specifically, for the photovoltaic module fault monitoring system of the present invention, the power supply module includes: An improved U-Net network is used to perform pixel-level segmentation on the infrared thermal images generated by the acquisition module to generate a thermal map of the abnormal hot spot distribution. An SVM model enhanced with an attention mechanism is used to analyze the abnormal fluctuation characteristics in the current and voltage characteristic curve data generated by the acquisition module to identify the power attenuation caused by the PID effect. A spatio-temporal graph convolutional network is used to perform spatio-temporal feature fusion on the component aging trend prediction data output by the digital twin module and the thermal map of the abnormal hot spot distribution, model the insulation degradation inflection point, and drive the dynamic adjustment of the overcurrent protection threshold of the DC power distribution cabinet.

[0058] The power supply module performs pixel-level segmentation on the infrared thermal images generated by the acquisition module through an improved U-Net network. The improved U-Net network uses a skip connection structure to fuse shallow texture features and deep semantic information, and extracts multi-scale hot spot region features through a spatial pyramid pooling module. The segmentation result is spatially aligned with the string current imbalance parameter obtained by the current and voltage characteristic curve acquisition module to generate a thermal map of the abnormal hot spot distribution marked with the hot spot position, temperature gradient, and corresponding current deviation value. The electrical topology connection relationship of the photovoltaic module is superimposed on the thermal map to quantitatively evaluate the influence weight of the hot spot on the output power of the string, providing spatial positioning data for subsequent insulation degradation analysis.

[0059] The SVM model enhanced with an attention mechanism analyzes the abnormal fluctuation characteristics in the current and voltage characteristic curve data, and uses a multi-head attention mechanism to dynamically allocate the feature weights of different sections of the curve. The model focuses on the abnormal drop in the open-circuit voltage under low irradiance, which characterizes the PID effect, in the characteristic curve. Through the kernel function mapping, the non-linearly separable data in the high-dimensional feature space is projected into the low-dimensional space, and a PID effect recognition boundary is constructed in combination with the soft margin classification strategy. The classification result is associated with the maximum power point tracking log data of the inverter to generate a PID effect diagnosis report containing the power attenuation rate, reversibility evaluation, and repair suggestions, providing an electrical anomaly feature basis for the protection threshold adjustment.

[0060] The spatio-temporal graph convolutional network receives the component aging trend prediction data output by the digital twin module and the thermal map of the abnormal hot spot distribution, and constructs a spatio-temporal topology graph with the photovoltaic string as the node and the electrical connection relationship as the edge. The network extracts the temporal decay characteristics of parameters such as the conductivity and dielectric constant of the insulating material through the time-axis convolutional layer, and captures the aging synergy effect between adjacent strings in combination with the spatial graph convolutional layer. The feature fusion layer performs cross-modal splicing on the temperature gradient data in the hot spot area and the aging trend prediction result to establish a non-linear mapping model between the hot spot temperature rise rate and the insulation material degradation. The insulation degradation inflection point prediction data output by the network is input into the overcurrent protection threshold adjustment module of the DC power distribution cabinet, and the action delay parameter of the ground fault protection is dynamically corrected according to the historical decay curve of the insulation resistance, realizing the dynamic matching of the protection threshold with the ambient temperature and humidity conditions.

[0061] Specifically, for the photovoltaic module fault monitoring system of the present invention, the power supply module further includes: A knowledge distillation unit, configured to perform knowledge distillation on the multi-scale feature parameters extracted by the improved U-Net network during the infrared thermal image segmentation process and the classification weights generated by the attention mechanism-enhanced SVM model during the current-voltage curve analysis.

[0062] The knowledge distillation unit in the power supply module extracts the multi-scale feature parameters generated by the improved U-Net network during the infrared thermal image segmentation process, including shallow texture features and deep semantic features. The improved U-Net network captures the boundary information of hot spot regions and the temperature gradient distribution features at different resolutions through multi-level convolutional operations of the encoder. At the same time, the decoder fuses cross-layer feature maps through a skip connection structure to generate a pixel-level segmentation result. The knowledge distillation unit synchronously extracts the classification weight matrix generated by the attention mechanism-enhanced SVM model during the current-voltage curve analysis process, and quantifies the influence intensity of electrical characteristics in different sections on the PID effect classification decision.

[0063] The knowledge distillation unit performs modality alignment on the multi-scale hot spot features and classification weights through a channel attention mechanism to generate a joint feature vector with cross-domain relevance. The channel attention mechanism dynamically adjusts the weight allocation ratio of feature channels according to the correlation between the temperature gradient amplitude of the hot spot region and the abnormal current fluctuation amplitude. The aligned feature vector is input into the feature fusion layer of the knowledge distillation unit, and a gated recurrent unit is used to construct a temporal dependence relationship to capture the dynamic association law between the hot spot temperature rise rate and the abnormal fluctuation of the current characteristic curve.

[0064] The distillation loss function constrains the student network to inherit the feature extraction ability of the teacher network during the feature reconstruction process, so that the lightweight classifier of the student network can directly output the composite fault type diagnosis result based on the joint feature vector. The student network learns the cross-modal mapping relationship between the multi-scale spatial feature representation of the improved U-Net network and the electrical anomaly classification logic of the SVM model by minimizing the weighted loss function of the feature reconstruction error and the classification error. The optimized joint feature vector is input into the node attribute update module of the spatio-temporal graph convolutional network, and maps the hot spot temperature gradient and current anomaly features to the spatio-temporal topology graph of the component aging trend prediction model.

[0065] The spatio-temporal graph convolutional network dynamically adjusts the weight distribution of multi-modal features in node state updates through the graph attention mechanism, and establishes a physical correlation model between the temperature anomaly in the hot spot area and the conductivity decay of the insulating material. The prediction result is fed back to the feature alignment module of the knowledge distillation unit, which drives the student network to optimize the weight distribution strategy in the feature fusion process, forming a closed-loop learning mechanism for cross-modal feature optimization and aging trend prediction. The composite fault diagnosis result output by the knowledge distillation unit is synchronously input into the protection threshold adjustment module of the DC power distribution cabinet, and combined with the current environmental humidity parameter to generate a dynamic overcurrent protection action threshold, forming an adaptive matching mechanism for monitoring data and protection parameters.

[0066] Specifically, for the photovoltaic module fault monitoring system described in the present invention, the digital twin module includes: A heterogeneous algorithm federated engine, which is used to run the deep residual network and the Vision Transformer model in parallel, fuse the conflicting evidence in the abnormal hot spot distribution thermal map and the current-voltage anomaly analysis result generated by the power supply module through the Dempster-Shafer evidence theory, and output a multi-model joint diagnosis result; An improved A* algorithm unit, which is used to switch to the local cache data of the acquisition module in case of communication anomaly, and execute fast protection logic based on the temperature field distribution data and the defect space mapping diagram in the dynamic health status matrix, and trigger the inverter derating operation mode.

[0067] The digital twin module runs the deep residual network and the Vision Transformer model in parallel through the heterogeneous algorithm federated engine to construct a multi-modal fault diagnosis framework. The deep residual network extracts the local texture features and spatial context information of the infrared hot spot distribution thermal map, and retains the detailed features of the hot spot boundary through the residual skip connection; the Vision Transformer model analyzes the temporal correlation characteristics of the current-voltage anomaly data based on the self-attention mechanism, and captures the evolution law of the PID effect and insulation deterioration in the time dimension. The intermediate layer feature maps of the two models are input into the Dempster-Shafer evidence theory fusion module. By calculating the basic probability assignment function of the output results of each model and dynamically adjusting the evidence synthesis rule in combination with the conflict factor, the evidence conflict between the abnormal hot spot distribution and the current-voltage analysis result is eliminated, and a multi-model joint diagnosis result is generated. The fusion result annotates the fault type confidence and the spatial positioning coordinates, providing a multi-dimensional decision-making basis for the protection strategy.

[0068] When the improved A* algorithm unit detects an abnormal communication link, it switches to the local cached data stored in the acquisition module and calls the historical temperature field distribution data and the defect space mapping diagram in the dynamic health status matrix. The algorithm constructs a topological search network with the photovoltaic string health index as nodes and the electrical connection relationship as edges, and dynamically adjusts the heuristic function weight based on the spatial constraint conditions of the fault diffusion path. During the search process, it preferentially matches the historical protection case library under the current ambient temperature and humidity parameters, generates a fault risk coefficient by combining the temperature gradient threshold and the defect expansion rate, and calculates the optimal protection path that meets the minimum energy loss and the lowest fault propagation risk. The generated fast protection instruction set directly drives the inverter to adjust the maximum power point tracking parameters through the edge computing node, triggers the string-level derating operation mode, and reduces the risks of local overheating and insulation breakdown.

[0069] After the communication is restored, the digital twin module uploads the protection instruction log generated locally to the evidence theory fusion module of the heterogeneous algorithm federated engine, and updates the prior parameters of the confidence assignment function of the deep residual network and the Vision Transformer model. At the same time, the local protection execution effect data is input into the parameter correction module of the digital twin model, and the equivalent aging rate equation coefficient of the components in the virtual system is adjusted through the transfer learning algorithm to optimize the boundary conditions of the subsequent fault evolution simulation. The deviation value between the protection instruction execution effect and the digital twin prediction result is input into the online fine-tuning module of the federated engine, driving the parameter update of the feature extraction layer of the deep residual network and the self-attention layer of the Vision Transformer, forming a closed-loop feedback link for improving diagnostic accuracy and optimizing the fault tolerance mechanism.

[0070] Specifically, for the photovoltaic module fault monitoring system described in the present invention, the digital twin module further includes: A transfer learning unit for migrating the temperature field distribution data and the defect space mapping diagram in the dynamic health status matrix generated by the acquisition module to a preset virtual system through a domain adaptation algorithm, simulating the thermal aging of insulating materials and the defect diffusion path based on the migrated temperature field distribution data and the defect space mapping diagram, and generating preventive maintenance instructions for the power supply module.

[0071] The transfer learning unit in the digital twin module receives the dynamic health status matrix generated by the acquisition module, and extracts the temperature field distribution data and the defect space mapping diagram in the matrix. The temperature field distribution data eliminates the interference of environmental noise through an improved variational mode decomposition algorithm, and generates temperature gradient distribution data with a spatial coordinate mapping relationship; the defect space mapping diagram combines the electrical topology of the photovoltaic module, and uses an adaptive threshold segmentation algorithm to label the geometric morphology and electrical isolation status of the hidden crack and broken grid defects. The above data is feature-aligned through the gradient reversal layer in the domain adaptation algorithm, reducing the feature distribution difference between the real system and the virtual system, and generating a virtual component state dataset with physical consistency.

[0072] The transfer learning unit adopts a domain adversarial training strategy, introducing a gradient reversal operation in the feature extraction layer, forcing the encoder to generate a shared feature representation independent of the data source. The temperature field distribution data of the real system is cross-domain aligned with the thermal conduction simulation results of the virtual system, combining the electroluminescence defect characteristics and the microstructure model parameters of the virtual component, and establishing a spatio-temporal mapping relationship of the defect evolution map. The virtual system constructs a thermal-electrical coupling simulation model based on the aligned dataset, and simulates the co-evolution path of the thermal aging of the insulating material and the defect diffusion through a hidden Markov chain. The Monte Carlo method generates multiple sets of fault evolution trajectories, and combines the physical equations of the temperature rise rate in the hot spot area and the decay of the electrical conductivity of the insulating material to predict the cross-regional impact of local overheating on the insulation performance of adjacent components.

[0073] The preventive maintenance instruction generation module triggers according to the key parameter threshold trigger conditions in the fault evolution simulation results, and uses a random forest classifier to evaluate the risk levels of each evolution path. The input parameters of the classifier include the predicted value of the insulation resistance in the virtual system, the expansion rate of the hot spot area, and the string current imbalance degree, and the output is a set of maintenance strategies including the component replacement priority, the cleaning cycle suggestion, and the protection threshold adjustment parameters. The generated maintenance instructions are input into the spatio-temporal graph convolutional network model of the power supply module, and by updating the material aging coefficient and the environmental stress weight parameters in the network node attributes, the spatio-temporal correlation of the aging trend prediction is optimized. The spatio-temporal graph convolutional network fuses the historical repair record data in the maintenance instructions in the time dimension, and establishes a dynamic association model between the component performance recovery and the subsequent aging rate; in the space dimension, it combines the string electrical connection topology to predict the cross-regional impact of local hot spots on the insulation performance of adjacent components.

[0074] The transfer learning unit synchronously updates the parameter mapping relationship between the digital twin model and the spatiotemporal graph convolutional network. When the aging rate of the actual component is monitored to deviate from the predicted value, the online fine-tuning mechanism of the model parameters is triggered. The elastic weight solidification technology is used in the fine-tuning process to retain the existing knowledge while adapting to the new aging pattern characteristics, avoiding the prediction oscillation caused by drastic parameter adjustments of the model. The updated maintenance strategy data is synchronously fed back to the simulation boundary condition setting module of the digital twin module to optimize the physical equation coefficients for subsequent fault evolution path prediction, forming a closed-loop optimization link between the real system monitoring data and the virtual simulation model. The protection threshold adjustment parameters in the maintenance instructions drive the dynamic protection logic of the DC distribution cabinet, match the insulation resistance attenuation curve according to the ambient temperature and humidity conditions, and realize the adaptive adjustment of the overcurrent protection action threshold.

[0075] Second, see Figure 1 The present invention provides a photovoltaic module fault monitoring method, which is applied to the photovoltaic module fault monitoring system, and comprises the following steps: Step S101, collecting photovoltaic module data, the photovoltaic module data including infrared thermal imaging data, electroluminescent defect data and current-voltage characteristic curve data, performing time-series alignment on the photovoltaic module data by using a preset multi-scale feature alignment algorithm, and generating a dynamic health status matrix; Step S102, receiving a leakage current signal, dynamically adjusting the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal by using a preset improved whale optimization algorithm, generating a filtered leakage current signal, and calculating the insulation resistance value in combination with a preset photovoltaic string voltage level, generating a dynamic insulation criterion, triggering an alarm signal when the dynamic insulation criterion is lower than a preset threshold, and linking a preset DC circuit breaker to execute a gradient tripping strategy based on a preset photovoltaic string voltage level meter; Step S103, receiving an alarm signal, optimizing the parsing priority of the opening control instruction and the arc extinguishing chamber impedance adjustment instruction generated by the alarm signal based on the communication path weight of the preset time-sensitive network, matching the circuit breaker opening action time window corresponding to the opening control instruction and the arc extinguishing chamber impedance change curve through a preset dynamic time warping algorithm, and adjusting the arc extinguishing chamber impedance parameter to suppress the reignition of the arc; Step S104, receiving the infrared hot spot distribution thermogram, the current and voltage curve abnormal analysis results, and the component aging trend prediction data, and performing feature-level fusion on the infrared hot spot distribution thermogram, the current and voltage curve abnormal analysis results, and the component aging trend prediction data to generate a power supply circuit insulation degradation assessment result, which is used to drive the dynamic adjustment of the overcurrent protection threshold of the preset DC distribution cabinet; Step S105: Receive the dynamic health status matrix, construct a digital twin model of the photovoltaic array including the temperature field distribution and defect space mapping, synchronize the insulation degradation assessment result output by the power supply module to a preset virtual system through a preset transfer learning algorithm, generate a preventive maintenance instruction, and feedback it to the spatio-temporal graph convolutional network model.

[0076] The photovoltaic module fault monitoring method provided by the present invention realizes the state monitoring and protection optimization of the photovoltaic system through multi-step cooperation. In step S101, the infrared thermal imaging sensor collects the surface temperature field distribution data of the photovoltaic module at a preset sampling frequency, performs modal decomposition on the original temperature data by using an improved variational mode decomposition algorithm, and screens out the effective temperature feature modes through the kurtosis-energy joint criterion to reconstruct the denoised temperature field distribution data. The electroluminescence detection unit combines the electrical topological structure and uses an adaptive threshold segmentation algorithm to process the electroluminescence image to generate a defect space mapping diagram annotating the geometric morphology and electrical isolation state of the hidden crack defect. The current and voltage characteristic curve acquisition module compensates the line impedance error through a generative adversarial network to generate high-precision electrical parameter data. The multi-source data is input into a bidirectional long short-term memory network for time series alignment, and the feature vectors of different time scales are extracted through the forward and backward recurrent units, and the temperature gradient, defect expansion rate, and current and voltage fluctuation characteristics are fused to generate a dynamic health status matrix.

[0077] In step S102, the improved whale optimization algorithm dynamically adjusts the scale factor and displacement parameter of the wavelet basis function, optimizes the frequency band division accuracy of the wavelet packet decomposition by introducing a simulated annealing mechanism, and separates the environmental noise and the true insulation fault characteristics in the leakage current signal. The fuzzy logic controller corrects the filtering bandwidth in combination with the temperature and humidity sensor data to suppress the low-frequency interference under high humidity conditions. The random forest regression model extracts the time-domain statistical characteristics and frequency-domain energy distribution characteristics of the filtered leakage current signal, correlates with the photovoltaic string voltage level parameter output by the acquisition module, constructs a voltage-leakage current non-linear mapping model, and generates a dynamic insulation criterion. When it is detected that the insulation resistance value is lower than the dynamic threshold, the gradient tripping strategy divides the fault circuit priority according to the voltage level and triggers the DC circuit breaker to perform a hierarchical tripping action.

[0078] In step S103, the communication path weight is optimized under the time-sensitive network architecture through the federated learning framework. The edge computing node constructs a path quality evaluation matrix according to the local network topology state and transmission delay characteristics, and the global model aggregates the training results of multiple nodes to generate an optimized weight. The dynamic time warping algorithm matches the mechanical response curve of the circuit breaker and the time sequence of the fault current zero crossing, predicts the opening time window and generates a pre-charge circuit control pulse sequence to absorb the inductive load energy in stages. The model predictive control unit rolls and optimizes the arc chamber nozzle opening and magnetic blow coil current parameters based on the arc impedance change curve to suppress the transient overvoltage and block the arc reignition condition.

[0079] In step S104, the improved U-Net network performs pixel-level segmentation of the infrared thermal image through a jump connection structure, and generates a thermal map of abnormal hot spot distribution in combination with the string current imbalance parameter. The SVM model enhanced by the attention mechanism focuses on the abnormal drop characteristics of the open circuit voltage in the current-voltage characteristic curve, and identifies the power attenuation mode caused by the PID effect. The spatiotemporal graph convolutional network integrates the aging trend data and hot spot distribution characteristics output by the digital twin module, establishes a spatiotemporal correlation model between the hot spot temperature rise rate and the degradation of the insulation material, and drives the DC distribution cabinet to dynamically adjust the overcurrent protection threshold according to the ambient temperature and humidity parameters.

[0080] Step S105 constructs a digital twin model of the photovoltaic array, and synchronizes the temperature field distribution and defect space mapping data of the real system and the virtual system through the transfer learning algorithm. The domain adversarial training strategy reduces the difference in feature distribution between the real and simulated data, and generates a virtual component state data set with physical consistency. The digital twin model simulates the fault diffusion path under different environmental stresses based on the Monte Carlo method, and combines the hidden Markov chain to preview the co-evolution process of thermal aging and defect expansion of insulating materials. The preventive maintenance instruction generation module triggers the threshold conditions according to the simulation results, outputs the component replacement priority and protection parameter adjustment strategy, and feeds back to the spatiotemporal graph convolution network to optimize the aging trend prediction model, forming a closed-loop optimization mechanism for monitoring data and simulation prediction.

[0081] The explanations of the technical feature terms in the technical solution of the present invention are as follows: Infrared thermal imaging sensor: a device used to collect temperature field distribution data on the surface of photovoltaic modules. It obtains temperature information of each area of ​​the module through non-contact thermal radiation detection technology, and combines the improved variational mode decomposition algorithm to eliminate environmental radiation noise and generate denoised temperature gradient distribution data.

[0082] Electroluminescence detection unit: A device that detects internal defects of photovoltaic modules based on the principle of electroluminescence. It applies voltage to stimulate the cells to produce fluorescence, uses an adaptive threshold segmentation algorithm to analyze the light and dark differences in the luminescent image, and combines the electrical topology structure to generate a mapping diagram that marks the geometric shape and electrical isolation status of hidden cracks and broken grid defects.

[0083] Generative Adversarial Network (GAN): A deep learning model for compensating line impedance errors, consisting of a generator and a discriminator. The generator simulates the current and voltage characteristic curves under ideal line conditions, and the discriminator compares the feature differences between the measured data and the generated data, dynamically corrects the measurement deviation, and improves the accuracy of electrical parameter acquisition.

[0084] Improved Whale Optimization Algorithm: An intelligent algorithm for optimizing the parameters of wavelet basis functions. By introducing a dynamic weight factor and a simulated annealing mechanism, it balances the global search and local development capabilities, adaptively adjusts the number of layers of wavelet packet decomposition and the frequency band division accuracy, and separates the high-frequency fault components and low-frequency environmental noise in the leakage current signal.

[0085] Fuzzy Logic Controller: An environment-adaptive control module based on fuzzy rules. By constructing a three-dimensional membership function of temperature-humidity-filtering bandwidth, it dynamically adjusts the filtering bandwidth range to suppress the low-frequency fluctuation interference of the leakage current signal in a high-humidity environment.

[0086] Random Forest Regression Model: A machine learning model for calculating the insulation resistance value. It extracts the time-domain statistical features and frequency-domain energy distribution features of the filtered leakage current signal, correlates with the photovoltaic string voltage level parameters to construct a non-linear mapping relationship, and generates a dynamic criterion reflecting the change of the conductivity of the insulating material.

[0087] Federated Learning Framework: A distributed machine learning architecture deployed in a time-sensitive network. Through local training of edge computing nodes and aggregation of global model parameters, it optimizes the weight allocation of the communication path, reduces the transmission delay of critical protection instructions.

[0088] Dynamic Time Warping Algorithm (DTW): A time series matching algorithm for predicting the opening time window of a circuit breaker. By aligning the mechanical response curve of the circuit breaker and the zero-crossing time sequence of the fault current through a dynamic warping path function, it combines with the pre-charge circuit control pulse sequence to absorb the energy of the inductive load in stages.

[0089] Improved U-Net Network: A convolutional neural network for pixel-level segmentation of infrared thermal images. It adopts a skip connection structure to fuse shallow texture features and deep semantic information, combines with a spatial pyramid pooling module to extract multi-scale hot spot region features, and generates a thermal spot distribution heat map annotating the temperature gradient and current imbalance degree.

[0090] Attention Mechanism Enhanced SVM Model: A classification model that combines Support Vector Machine (SVM) and multi-head attention mechanism. It focuses on the abnormal voltage drop section of the open circuit voltage that characterizes the PID effect in the current-voltage characteristic curve, and identifies the power attenuation mode through kernel function mapping and soft margin classification strategy.

[0091] Spatio-Temporal Graph Convolutional Network (ST-GCN): A graph neural network that fuses spatio-temporal features. It constructs a topological graph with photovoltaic strings as nodes and electrical connections as edges, extracts the temporal decay features of insulation parameters through time-axis convolution, and captures the aging synergy effect of adjacent strings through spatial graph convolution to drive the dynamic adjustment of protection thresholds.

[0092] Heterogeneous algorithm federation engine: It runs the multi-model fusion module of deep residual network and Vision Transformer in parallel, integrates the conflicting evidence of hot spot distribution and current and voltage anomaly analysis through DS evidence theory, and outputs joint diagnosis results.

[0093] Improved A* algorithm: A path search algorithm that builds a fault diffusion risk coefficient model based on the historical temperature field and defect mapping data in the dynamic health state matrix when communication is abnormal, and generates inverter derating operation parameters and string isolation priority instructions.

[0094] Transfer learning unit: A module used to synchronize the data of the real system and the digital twin model, narrow the feature distribution differences through domain adversarial training strategy, simulate the fault evolution path based on the Monte Carlo method, and generate preventive maintenance instructions including component replacement priority and protection threshold adjustment parameters.

[0095] In the specific implementation of the present invention, the fault monitoring and protection control of photovoltaic components are realized through the fusion of multi-source sensor data and the collaboration of intelligent algorithms. The infrared thermal imaging sensor of the acquisition module acquires the surface temperature field distribution data of the component at a sampling rate of 5 frames per second, and uses the improved variational mode decomposition algorithm to perform mode decomposition on the original temperature data. The effective temperature characteristic mode is screened by the kurtosis-energy joint criterion to generate a denoised temperature gradient distribution map; the electroluminescent detection unit collects electroluminescent images at a resolution of 50μm, and uses the adaptive threshold segmentation algorithm in combination with the electrical topology to identify the hidden crack defect area, and generate a defect space mapping map with the defect geometry marked; the current and voltage characteristic curve acquisition module compensates for the line impedance error through the adversarial generation network, controls the compensation error within ±0.5%, and generates high-precision electrical parameter data. Multi-source data is input into the bidirectional long short-term memory network for time series alignment, and a spatiotemporal correlation model of temperature field change, defect expansion and electrical parameter attenuation is established to generate a dynamic health state matrix.

[0096] The power supply dynamic compensation module uses an improved whale optimization algorithm to dynamically adjust the scale factor and displacement parameters of the wavelet basis function. The simulated annealing mechanism is introduced to optimize the frequency band division accuracy of the wavelet packet decomposition. The number of iterations is set to 200 times, and the convergence accuracy is 1e-6. The fuzzy logic controller constructs a three-dimensional membership function based on the temperature and humidity sensor data, dynamically adjusts the filter bandwidth range to 10Hz-1kHz, and suppresses the low-frequency leakage current fluctuation interference under high humidity conditions. The random forest regression model extracts the time domain statistical characteristics and frequency domain energy distribution characteristics of the filtered leakage current signal, and generates a dynamic insulation criterion in combination with the photovoltaic string voltage level parameters. When the insulation resistance value is detected to be lower than 50MΩ, a three-level alarm signal is triggered, and the DC circuit breaker is linked to isolate the fault circuit according to the 0.1s, 0.5s, and 1s gradient trip strategy.

[0097] The control optimization module deploys a federated learning framework under a time-sensitive network architecture. The edge computing nodes update the communication path weight parameters at a period of 30 s. The dynamic time warping algorithm predicts the circuit breaker opening time window, and the time window accuracy is controlled within the range of ±2 ms. The model predictive control unit monitors the arc impedance of the arc extinguishing chamber at a sampling frequency of 1 MHz. The rolling horizon optimization strategy calculates the optimal control sequences of the arc extinguishing chamber nozzle opening and the magnetic blow coil current, and suppresses the transient overvoltage below 120% of the rated voltage. The deep Q-network algorithm constructs a reward function based on the historical fault handling case library, dynamically adjusts the parsing weights of the overcurrent protection and island detection instructions, and shortens the response time to 50 ms to achieve the matching of the protection action and the fault evolution rate.

[0098] The improved U-Net network of the power supply module uses a skip connection structure to perform pixel-level segmentation on infrared thermal images. The hot spot positioning accuracy reaches ±5 mm, and an abnormal hot spot distribution heat map is generated by combining the string current imbalance parameter. The SVM model enhanced by the attention mechanism focuses on the characteristics of the open circuit voltage section of 0.5 V - 0.8 V in the current-voltage characteristic curve, identifies the power attenuation mode caused by the PID effect, and improves the classification accuracy. The spatio-temporal graph convolutional network receives the aging trend data output by the digital twin module, establishes an insulation degradation prediction model with a weekly time granularity, and drives the grounding protection threshold of the DC power distribution cabinet to be dynamically adjusted from the initial 100 MΩ to 60 MΩ to match the ambient temperature and humidity changes.

[0099] The digital twin module synchronizes the temperature field distribution data of the real system and the virtual system every week through the transfer learning algorithm, simulates the fault diffusion path under different salt spray concentrations by the Monte Carlo method, and generates a maintenance instruction set including the component replacement priority. The heterogeneous algorithm federated engine aggregates the diagnostic results of the deep residual network and the Vision Transformer model daily, and fuses the confidence levels of multiple models through the D-S evidence theory, improving the fault type recognition accuracy. The improved A* algorithm calls the health status data of the nearest 24 hours during a communication interruption, searches for the optimal protection path based on the fault diffusion risk coefficient, triggers the inverter to operate at a derated power of 80% of the rated power, and forms an online monitoring and offline fault tolerance dual-mode collaborative mechanism. The preventive maintenance instructions are fed back to the spatio-temporal graph convolutional network to optimize the material aging coefficient in the node attributes, forming a closed-loop optimization link between the monitoring data and the prediction model, and effectively suppressing the system-level cascading risk.

[0100] The present invention solves the above problems through the following technical solutions: Aiming at the problem of misjudgment of insulation resistance caused by dynamic fluctuations of leakage current in humid or salt spray environments for insulation monitoring devices, an improved whale optimization algorithm is used to dynamically adjust the scale factor and displacement parameters of wavelet basis functions, optimizing the frequency band division accuracy of wavelet packet decomposition. The algorithm balances global search and local development capabilities through a simulated annealing mechanism, separating high-frequency transient components and low-frequency environmental noise in the leakage current signal. A fuzzy logic controller dynamically corrects the filtering bandwidth threshold in combination with temperature and humidity sensor data, suppressing low-frequency leakage current fluctuation interference under high humidity conditions, and generating a dynamic insulation criterion reflecting the change in the conductivity of insulating materials. A random forest regression model correlates the voltage level of the photovoltaic string with the characteristics of the filtered leakage current, constructs a non-linear voltage-leakage current mapping relationship, and triggers a gradient tripping strategy to preferentially isolate high-risk fault circuits according to the voltage level, avoiding the misjudgment risk of a single threshold criterion under environmental disturbances.

[0101] Aiming at the problems of communication delay and algorithm convergence in the collaborative control of intelligent circuit breakers and monitoring systems, a federated learning framework is deployed based on a preset time-sensitive network. The communication path weights of multi-source protection instructions are aggregated through edge computing nodes to optimize the data transmission priority. The dynamic time warping algorithm matches the mechanical response curve of the circuit breaker with the timing of the zero-crossing point of the fault current, predicting the opening time window that meets the requirements of arc energy suppression, and combining a pre-charge circuit to absorb the energy of inductive loads in stages. The model predictive control unit adjusts the current of the magnetic blow coil in the arc extinguishing chamber and the flow rate of the gas blowing medium in real time, rollingly optimizing the impedance characteristics of the arc extinguishing chamber, suppressing transient overvoltage and blocking the conditions for arc re-ignition, and achieving the spatio-temporal matching of protection actions and fault evolution.

[0102] Furthermore, a virtual system of the photovoltaic array is constructed through a digital twin module. The transfer learning algorithm synchronizes the temperature field distribution and defect space mapping data of the real system, simulating the thermal aging and hot spot diffusion paths of insulating materials. The spatio-temporal graph convolutional network fuses the hot spot temperature rise rate and aging trend prediction data, driving the DC power distribution cabinet to dynamically adjust the overcurrent protection threshold. When communication is abnormal, the heterogeneous algorithm federated engine calls an improved A* algorithm to generate inverter derating operation parameters based on local cached data, forming an online monitoring and offline fault-tolerant dual-mode collaborative mechanism to reduce system-level cascading risks.

Claims

1. A photovoltaic module fault monitoring system, characterized in that, Including: An acquisition module, a power supply dynamic compensation module, a control optimization module, a power supply module, and a digital twin module; The acquisition module is used to acquire photovoltaic module data. The photovoltaic module data includes infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data. The photovoltaic module data is subjected to time series alignment through a preset multi-scale feature alignment algorithm to generate a dynamic health state matrix; The power supply dynamic compensation module receives the leakage current signal in the dynamic health state matrix, dynamically adjusts the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal through a preset improved whale optimization algorithm, generates a filtered leakage current signal, and combines the preset photovoltaic string voltage level output by the acquisition module to calculate the insulation resistance value and generate a dynamic insulation criterion; when the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and a gradient tripping strategy based on the preset photovoltaic string voltage level is linked to a preset DC circuit breaker; The control optimization module optimizes the parsing priority of the tripping control instruction and the arc extinguishing chamber impedance adjustment instruction generated by the alarm signal based on the communication path weight of the preset time-sensitive network; The arc extinguishing chamber impedance parameter is adjusted by matching the circuit breaker tripping action time window corresponding to the tripping control instruction and the arc extinguishing chamber impedance change curve through a preset dynamic time warping algorithm; The power supply module receives the infrared hot spot distribution heat map and the current-voltage curve anomaly analysis result generated by the acquisition module, performs feature-level fusion with the component aging trend prediction data output by the digital twin module, generates a power supply loop insulation degradation evaluation result, and drives the dynamic adjustment of the overcurrent protection threshold of the DC power distribution cabinet; The digital twin module constructs a digital twin model of the photovoltaic array including temperature field distribution and defect space mapping, synchronizes the insulation degradation evaluation result output by the power supply module to a preset virtual system through a preset transfer learning algorithm, generates a preventive maintenance instruction, and feeds it back to the spatio-temporal graph convolutional network model of the power supply module.

2. The photovoltaic module fault monitoring system according to claim 1, wherein The digital twin module is further used for: Receiving the dynamic health state matrix generated by the acquisition module, and constructing a digital twin model of the photovoltaic array including temperature field distribution data, defect space mapping diagram, and compensated current and voltage curve data; Synchronizing the temperature field distribution data and the defect space mapping diagram in the dynamic health state matrix to a preset virtual system through a preset transfer learning algorithm. The preset virtual system is used to simulate the fault diffusion path to generate a preventive maintenance instruction; The preventive maintenance instruction is input into the spatio-temporal graph convolutional network model in the power supply module to optimize the confidence of the component aging trend prediction result generated by the power supply module.

3. The photovoltaic module fault monitoring system according to claim 1, characterized in that, The acquisition module includes: An infrared thermal imaging sensor, which is used to acquire the temperature field distribution data of the photovoltaic module, and eliminates environmental noise interference through a preset improved variational mode decomposition algorithm to generate denoised temperature field data; An electroluminescence detection unit, which is used to process the electroluminescence image of the photovoltaic module based on a preset adaptive threshold segmentation algorithm, and generate a defect space mapping diagram in combination with the electrical topology structure of the photovoltaic module; The current and voltage characteristic curve acquisition module is used to compensate for the line impedance error through an adversarial generative network and generate the compensated current and voltage characteristic curve data; The denoised temperature field data, defect space mapping diagram, and compensated current and voltage characteristic curve data are input into a preset bidirectional long short-term memory network, and multi-source data time series alignment is performed based on the operation timestamp of the photovoltaic module to generate a dynamic health status matrix.

4. The photovoltaic module fault monitoring system according to claim 1, wherein, The power supply dynamic compensation module includes: An adaptive filtering unit, which is used to receive the leakage current signal in the dynamic health status matrix, dynamically adjust the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal, and output the filtered leakage current signal; An environmental parameter and filtering threshold mapping table, which is used to combine the temperature and humidity sensor data and correct the filtering bandwidth of the adaptive filtering unit through a preset fuzzy logic controller; A random forest regression model, which is used to input the filtered leakage current signal into a pre-trained regression model, combine the preset photovoltaic string voltage level in the current and voltage characteristic curve data generated by the acquisition module, calculate the insulation resistance value, and generate a dynamic insulation criterion.

5. The photovoltaic module fault monitoring system according to claim 1, characterized in that, The control optimization module includes: An edge computing node, which is used to optimize the communication path selection strategy of the time-sensitive network using a federated learning framework and generate the optimized communication path weight; A deep Q-network algorithm unit, which is used to dynamically adjust the parsing weights of the opening control instruction and the arc extinguishing chamber impedance adjustment instruction based on the fault feature map generated from the abnormal hot spot distribution heat map output by the power supply module and the abnormal analysis result of the current-voltage curve; A dynamic time warping algorithm unit, which is used to predict the optimal time window for the circuit breaker opening action according to the optimized communication path weight and absorb the inductive load energy through a pre-charge circuit; A model predictive control unit, which is used to match the arc impedance change curve corresponding to the time window predicted by the dynamic time warping algorithm unit in the arc extinguishing chamber to suppress the transient overvoltage and the risk of arc re-ignition.

6. The photovoltaic module fault monitoring system according to claim 1, characterized in that, The power supply module includes: An improved U-Net network, which is used to perform pixel-level segmentation on the infrared thermal image generated by the acquisition module and generate an abnormal hot spot distribution heat map; An SVM model enhanced by an attention mechanism, which is used to analyze the abnormal fluctuation characteristics in the current and voltage characteristic curve data generated by the acquisition module and identify the power attenuation caused by the PID effect; A spatio-temporal graph convolutional network, which is used to perform spatio-temporal feature fusion on the component aging trend prediction data output by the digital twin module and the abnormal hot spot distribution heat map, model the insulation deterioration inflection point, and drive the dynamic adjustment of the over-current protection threshold of the DC power distribution cabinet.

7. The photovoltaic module fault monitoring system according to claim 6, wherein, The power supply module further includes: A knowledge distillation unit, which is used to perform knowledge distillation on the multi-scale feature parameters extracted by the improved U-Net network during the infrared thermal image segmentation process and the classification weights generated by the SVM model enhanced by the attention mechanism during the current-voltage curve analysis.

8. The photovoltaic module fault monitoring system according to claim 1, characterized in that, The digital twin module includes: A heterogeneous algorithm federation engine is used to run the deep residual network and the Vision Transformer model in parallel, fuse the conflicting evidence in the abnormal hot spot distribution heat map generated by the power supply module and the current and voltage anomaly analysis results through the Dempster-Shafer evidence theory, and output a multi-model joint diagnosis result; The improved A* algorithm unit is used to switch to the local cache data of the acquisition module when communication is abnormal, execute fast protection logic based on the temperature field distribution data and defect space mapping diagram in the dynamic health state matrix, and trigger the inverter derating operation mode.

9. The photovoltaic module fault monitoring system according to claim 8, characterized in that, The digital twin module further includes: A transfer learning unit is used to migrate the temperature field distribution data and defect space mapping diagram in the dynamic health status matrix generated by the acquisition module to a preset virtual system through a domain adaptation algorithm, simulate the thermal aging of the insulation material and the defect diffusion path based on the migrated temperature field distribution data and defect space mapping diagram, and generate preventive maintenance instructions for the power supply module.

10. A photovoltaic module fault monitoring method, applied to the photovoltaic module fault monitoring system according to any one of claims 1 to 9, characterized in that, The following steps are involved: Collect PV module data, including infrared thermal imaging data, electroluminescent defect data, and current-voltage characteristic curve data. Use a preset multi-scale feature alignment algorithm to perform time-series alignment on the PV module data to generate a dynamic health status matrix. Receive the leakage current signal, dynamically adjust the wavelet packet decomposition layer number and threshold parameter of the wavelet basis function of the leakage current signal through the preset improved whale optimization algorithm, generate the filtered leakage current signal, and calculate the insulation resistance value in combination with the preset photovoltaic string voltage level to generate the dynamic insulation criterion. When the dynamic insulation criterion is lower than the preset threshold, trigger the alarm signal, and link the preset DC circuit breaker to execute the gradient tripping strategy based on the preset photovoltaic string voltage level meter; Receive an alarm signal, optimize the parsing priority of the opening control instruction and the arc extinguishing chamber impedance adjustment instruction generated by the alarm signal based on the communication path weight of the preset time-sensitive network, match the circuit breaker opening action time window corresponding to the opening control instruction with the arc extinguishing chamber impedance change curve through a preset dynamic time warping algorithm, and adjust the arc extinguishing chamber impedance parameter; Receive infrared hot spot distribution thermogram, current and voltage curve abnormal analysis results and component aging trend prediction data, perform feature-level fusion on the infrared hot spot distribution thermogram, current and voltage curve abnormal analysis results and component aging trend prediction data, and generate power supply circuit insulation degradation assessment results, which are used to drive the dynamic adjustment of the overcurrent protection threshold of the preset DC distribution cabinet; The dynamic health status matrix is ​​received, and a digital twin model of the photovoltaic array including temperature field distribution and defect space mapping is constructed. The insulation degradation assessment result output by the power supply module is synchronized to a preset virtual system through a preset transfer learning algorithm, and preventive maintenance instructions are generated and fed back to the spatiotemporal graph convolutional network model.

Images