A photovoltaic module failure monitoring system and method

Abstract

The present application relates to the technical field of power supply or power distribution circuit system, and particularly relates to a photovoltaic module fault monitoring system and method, comprising a collection module, a power supply dynamic compensation module, a control optimization module, a power supply module and a digital twin module. The collection module acquires multi-source data through an infrared thermal imaging sensor, an electroluminescence detection unit and a current-voltage characteristic curve collection module, the power supply dynamic compensation module dynamically adjusts wavelet base function parameters based on an improved whale optimization algorithm, the power supply module fuses hot spot distribution, current-voltage abnormality and aging trend data, and drives dynamic adjustment of a protection threshold. The digital twin module synchronizes real and virtual system data through transfer learning, preplays a fault path to generate maintenance instructions and feeds back an optimized prediction model, forms a closed-loop system of monitoring, protection and self-optimization, and improves fault diagnosis accuracy and system reliability in complex environments.

Description

Technical Field

[0001] This invention relates to the technical field of circuit systems for equipment control power supply or distribution, and in particular to a photovoltaic module fault monitoring system and method. Background Technology

[0002] Photovoltaic module fault monitoring is a crucial step in real-time diagnosis of the photovoltaic system's operational status, relying on multi-source sensor fusion and intelligent analysis technologies. Its core mechanism involves using infrared thermal imaging to identify abnormal temperature distributions on the module surface, pinpointing hot spot effects caused by cell damage, bypass diode failure, or shading. Combined with electroluminescence detection, it analyzes microstructural anomalies such as microcracks and welding defects within the cells. Simultaneously, it utilizes current and voltage characteristic curve tracking technology to quantify the degree to which the module's output characteristics deviate from the standard curve, thereby identifying systemic faults such as Potential Induced Degradation (PID), encapsulation material aging, or poor junction box contact. Through time-series data comparison and pattern recognition algorithms, the monitoring system can dynamically analyze fault evolution patterns, providing a scientific basis for operation and maintenance decisions at the failure mechanism level, effectively reducing energy loss and extending the module's service life.

[0003] During the operation of the photovoltaic module fault monitoring system, the insulation monitoring device is susceptible to interference from dynamic fluctuations in leakage current in humid or salt spray environments, which can lead to misjudgment of insulation resistance and affect the reliability of fault isolation. In addition, the coordinated control of the intelligent circuit breaker and the monitoring system needs to take into account both rapid fault clearing and transient overvoltage suppression. However, under complex operating conditions, communication delays or insufficient algorithm convergence may lead to a mismatch between protection actions and fault evolution rates, increasing the risk of system-level cascading. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a photovoltaic module fault monitoring system and method, which solves the problem of misjudgment of insulation resistance caused by dynamic fluctuations in leakage current in humid or salt spray environments, as well as the system-level cascading risks caused by the mismatch between protection actions and fault evolution rates due to communication delays or insufficient algorithm convergence in the collaborative control of intelligent circuit breakers and monitoring systems.

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

[0006] In a first aspect, the present invention provides a photovoltaic module fault monitoring system, comprising: a data acquisition module, a power supply dynamic compensation module, a control optimization module, a power supply module, and a digital twin module;

[0007] The acquisition module is connected to the power supply dynamic compensation module through a distributed monitoring network to collect photovoltaic module data, including infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data. The photovoltaic module data is time-series aligned by a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix and transmit it to the power supply dynamic compensation module.

[0008] The power supply dynamic compensation module receives the leakage current signal from the dynamic health status matrix, dynamically adjusts the wavelet packet decomposition level 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 calculates the insulation resistance value by combining it with the preset photovoltaic string voltage level output by the acquisition module to generate a dynamic insulation criterion; when the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and the preset DC circuit breaker is linked to execute a gradient tripping strategy based on the preset photovoltaic string voltage level.

[0009] The control optimization module receives the alarm signal triggered by the power supply dynamic compensation module, optimizes the parsing priority of the tripping control command and the arc-extinguishing chamber impedance adjustment command generated by the alarm signal based on the communication path weight of the preset time-sensitive network, and matches the circuit breaker tripping action time window and the arc-extinguishing chamber impedance change curve corresponding to the tripping control command with the preset dynamic time warping algorithm, and adjusts the arc-extinguishing chamber impedance parameters to suppress arc reignition.

[0010] The power supply module receives the infrared hot spot distribution thermal map and current and voltage curve anomaly analysis results generated by the acquisition module, performs feature-level fusion with the component aging trend prediction data output by the digital twin module, generates the power supply circuit insulation degradation assessment results, and drives the overcurrent protection threshold of the DC distribution cabinet to be dynamically adjusted.

[0011] The digital twin module receives the dynamic health status matrix, constructs a photovoltaic array digital twin model including temperature field distribution and defect spatial mapping, and synchronizes the insulation degradation assessment results output by the power supply module to the preset virtual system through a preset transfer learning algorithm, generates preventive maintenance instructions and feeds them back to the spatiotemporal graph convolutional network model of the power supply module.

[0012] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the digital twin module is also used for:

[0013] Receive the dynamic health status matrix generated by the acquisition module, and construct a digital twin model of the photovoltaic array including temperature field distribution data, defect space mapping map and compensated current and voltage curve data;

[0014] The temperature field distribution data and defect space mapping map in the dynamic health status matrix are synchronized to a preset virtual system through a preset transfer learning algorithm. The preset virtual system is used to simulate the fault propagation path and generate preventive maintenance instructions.

[0015] The preventive maintenance command is input into the spatiotemporal graph convolutional network model in the power supply module to optimize the confidence level of the component aging trend prediction results generated by the power supply module.

[0016] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the acquisition module includes:

[0017] An infrared thermal imaging sensor is used to collect temperature field distribution data of photovoltaic modules and generate denoised temperature field data by eliminating environmental noise interference through a preset improved variational mode decomposition algorithm.

[0018] An electroluminescence detection unit is used to process the electroluminescence image of a photovoltaic module based on a preset adaptive threshold segmentation algorithm, and generate a defect space mapping map by combining the electrical topology of the photovoltaic module.

[0019] The current and voltage characteristic curve acquisition module is used to compensate for line impedance errors by using an adversarial generation network to generate compensated current and voltage characteristic curve data.

[0020] The denoised temperature field data, defect space mapping map, and compensated current and voltage characteristic curve data are input into a preset bidirectional long short-term memory network. Based on the operating timestamp of the photovoltaic module, multi-source data time sequence alignment is performed to generate a dynamic health status matrix.

[0021] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the power supply dynamic compensation module includes:

[0022] An adaptive filtering unit is used to receive the leakage current signal in the dynamic health status matrix, dynamically adjust the wavelet packet decomposition level and threshold parameter of the wavelet basis function of the leakage current signal, and output the filtered leakage current signal.

[0023] An environmental parameter-to-filter threshold mapping table is used to combine temperature and humidity sensor data and correct the filter bandwidth of the adaptive filter unit through a preset fuzzy logic controller.

[0024] A random forest regression model is used to input the filtered leakage current signal into a pre-trained regression model, and to calculate the insulation resistance value and generate a dynamic insulation criterion by combining the preset photovoltaic string voltage level in the current and voltage characteristic curve data generated by the acquisition module.

[0025] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the control optimization module includes:

[0026] Edge computing nodes are used to optimize the communication path selection strategy of time-sensitive networks using a federated learning framework, and generate optimized communication path weights.

[0027] The deep Q-network algorithm unit is used to dynamically adjust the analytical weights of the tripping control command and the arc-extinguishing chamber impedance adjustment command based on the fault feature map generated by the abnormal hot spot distribution heat map and the abnormal current and voltage curve analysis results output by the power supply module.

[0028] The dynamic time warping algorithm unit is used to predict the optimal time window for the circuit breaker tripping action based on the optimized communication path weight, and to absorb inductive load energy through the pre-charging circuit.

[0029] 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, thereby suppressing the risk of transient overvoltage and arc reignition.

[0030] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the power supply module includes:

[0031] An improved U-Net network is used to perform pixel-level segmentation of the infrared thermal image generated by the acquisition module to generate a thermal map of abnormal hot spot distribution.

[0032] An attention-enhanced SVM model is used to analyze abnormal fluctuation characteristics in the current and voltage characteristic curve data generated by the acquisition module and identify power attenuation caused by PID effect.

[0033] A spatiotemporal graph convolutional network is used to fuse the component aging trend prediction data output by the digital twin module with the abnormal hot spot distribution heat map in terms of spatiotemporal features, model the insulation degradation inflection point, and drive the dynamic adjustment of the overcurrent protection threshold of the DC distribution cabinet.

[0034] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the power supply module further includes:

[0035] The knowledge distillation unit is used to perform knowledge distillation on the multi-scale feature parameters extracted by the improved U-Net network during infrared thermal image segmentation and the classification weights generated by the attention mechanism-enhanced SVM model in current-voltage curve analysis.

[0036] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the digital twin module includes:

[0037] The heterogeneous algorithm federated engine is used to run deep residual networks and Vision Transformer models in parallel. It integrates conflicting evidence from the abnormal hot spot distribution heat map and current and voltage anomaly analysis results generated by the power supply module through Dempster-Shafer evidence theory, and outputs multi-model joint diagnostic results.

[0038] An improved A* algorithm unit is used to switch to the local cached data of the acquisition module when communication is abnormal. Based on the temperature field distribution data and defect space mapping in the dynamic health status matrix, it executes fast protection logic and triggers the inverter derating operation mode.

[0039] Furthermore, in the photovoltaic module fault monitoring system of the present invention, the digital twin module further includes:

[0040] The transfer learning unit is used to transfer the temperature field distribution data and defect space mapping map 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 defect space mapping map, the unit simulates the thermal aging of the insulation material and the defect propagation path, and generates preventive maintenance instructions for the power supply module.

[0041] Secondly, the present invention provides a photovoltaic module fault monitoring method, applied to the aforementioned photovoltaic module fault monitoring system, comprising the following steps:

[0042] Data from photovoltaic modules is collected, including infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data. The photovoltaic module data is time-series aligned using a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix.

[0043] The system receives leakage current signals and dynamically adjusts the wavelet packet decomposition level and threshold parameters of the wavelet basis function of the leakage current signal through a preset improved whale optimization algorithm. It generates a filtered leakage current signal and calculates the insulation resistance value based on the preset photovoltaic string voltage level to generate a dynamic insulation criterion. When the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and the preset DC circuit breaker is linked to execute a gradient tripping strategy based on the preset photovoltaic string voltage level.

[0044] Upon receiving an alarm signal, the parsing priority of the tripping control command and the arc-extinguishing chamber impedance adjustment command generated by the alarm signal is optimized based on the communication path weight of the preset time-sensitive network. The circuit breaker tripping action time window corresponding to the tripping control command is matched with the arc-extinguishing chamber impedance change curve through a preset dynamic time warping algorithm, and the arc-extinguishing chamber impedance parameter is adjusted to suppress arc reignition.

[0045] The system receives infrared hotspot distribution thermal maps, current and voltage curve anomaly analysis results, and component aging trend prediction data. It performs feature-level fusion of the infrared hotspot distribution thermal maps, current and voltage curve anomaly analysis results, and component aging trend prediction data to generate power supply circuit insulation degradation assessment results. The power supply circuit insulation degradation assessment results are used to drive the dynamic adjustment of the overcurrent protection threshold of the preset DC distribution cabinet.

[0046] The system receives the dynamic health status matrix, constructs a digital twin model of the photovoltaic array including temperature field distribution and defect spatial mapping, synchronizes the insulation degradation assessment results output by the power supply module to the preset virtual system through a preset transfer learning algorithm, generates preventive maintenance instructions and feeds them back to the spatiotemporal graph convolutional network model.

[0047] Beneficial effects of this invention;

[0048] This invention dynamically adjusts wavelet basis function parameters by improving the whale optimization algorithm, and combines it with a fuzzy logic controller to correct the filter bandwidth based on temperature and humidity data. This accurately separates leakage current noise components from actual insulation fault characteristics in humid or salt spray environments, generating dynamic insulation criteria and reducing the risk of misjudgment of insulation resistance. It optimizes communication path weights based on a federated learning framework using a preset time-sensitive network, predicts the circuit breaker tripping time window using a dynamic time warping algorithm, and adjusts the arc-extinguishing chamber impedance characteristics in real time through model predictive control. This achieves spatiotemporal matching between protection actions and fault evolution, suppressing arc reignition and system-level cascading risks. The digital twin module synchronizes multi-dimensional monitoring data and virtual simulation results through transfer learning, uses the Monte Carlo method to pre-simulate fault propagation paths to generate preventative maintenance instructions, and feeds them back to the spatiotemporal graph convolutional network to optimize aging trend prediction, forming a closed-loop optimized self-evolving monitoring system. The heterogeneous algorithm federated engine integrates multi-model diagnostic results of infrared hotspot distribution and current and voltage anomalies, and combines an improved A* algorithm to construct a dual-mode mechanism for online monitoring and offline fault tolerance, improving fault diagnosis accuracy and system reliability under complex operating conditions. Attached Figure Description

[0049] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0050] Figure 1 This is a flowchart of a photovoltaic module fault monitoring method provided in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings. To better understand the objectives of this invention, it will be described in further detail below.

[0052] In a first aspect, the present invention provides a photovoltaic module fault monitoring system, comprising: a data acquisition module, a power supply dynamic compensation module, a control optimization module, a power supply module, and a digital twin module;

[0053] The acquisition module is connected to the power supply dynamic compensation module through a distributed monitoring network to collect photovoltaic module data, including infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data. The photovoltaic module data is time-series aligned by a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix and transmit it to the power supply dynamic compensation module.

[0054] The acquisition module is connected to the power supply dynamic compensation module via a distributed monitoring network. An infrared thermal imaging sensor acquires surface temperature field distribution data of the photovoltaic module at a preset sampling frequency. An improved variational mode decomposition algorithm performs mode decomposition on the raw temperature data. By adaptively adjusting the penalty factor and the number of modes, and combining the kurtosis-energy joint criterion, effective temperature feature modes are selected to reconstruct the denoised temperature gradient distribution data. The reconstructed data is then spatially mapped to the heat sink layout parameters of the photovoltaic module to generate a temperature field distribution map with physical location correlation, providing basic data support for hot spot effect analysis.

[0055] After acquiring electroluminescence images of the photovoltaic module, the electroluminescence detection unit dynamically adjusts the image segmentation parameters based on an adaptive threshold segmentation algorithm. The algorithm expands the defect boundary contours using a region growing method based on the electroluminescence intensity distribution characteristics, and establishes a spatial mapping model between the defect region and circuit nodes by combining the electrical topology connections of the module's cells. The generated defect spatial mapping map annotates the geometric morphology and electrical isolation status of hidden cracks and broken grid defects, quantifies the impact weight of defects on string output characteristics, and provides microstructural data for insulation degradation analysis.

[0056] 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 an adversarial generative network. The generator network simulates the current and voltage characteristic curves under ideal line conditions, while the discriminator network compares the characteristic differences between the measured data and the generated data, dynamically correcting the measurement deviation caused by the line impedance. The compensated current and voltage characteristic curve data uses a sliding window mechanism to extract the key inflection points of the characteristic curves, preserving the characteristic information of the module during transient processes such as shading and bypass diode operation, forming a high-precision electrical parameter dataset.

[0057] A multi-scale feature alignment algorithm performs time-series alignment processing on the aforementioned multi-source data. Through dynamic time warping, it calculates feature point matching paths at different sensor sampling rates, aligning the time-series benchmarks for temperature field distribution, defect spatial mapping, and electrical parameters. The forward and backward recurrent units of a bidirectional long short-term memory network extract feature vectors at different time scales, establishing a spatiotemporal correlation model for temperature gradient changes, defect propagation rates, and current and voltage fluctuations. The feature fusion layer concatenates the spatiotemporal feature vectors from the multi-source data dimensionally, combining them with the photovoltaic system's operating timestamps to establish a unified time-series benchmark, generating a dynamic health status matrix that includes thermal-electric coupling characteristics, defect evolution trends, and component health indices. This matrix is ​​transmitted to the power supply dynamic compensation module via a distributed monitoring network, providing multi-dimensional data input for insulation condition assessment and protection strategy generation.

[0058] The power supply dynamic compensation module receives the leakage current signal from the dynamic health status matrix, dynamically adjusts the wavelet packet decomposition level 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 calculates the insulation resistance value by combining it with the preset photovoltaic string voltage level output by the acquisition module to generate a dynamic insulation criterion; when the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and the preset DC circuit breaker is linked to execute a gradient tripping strategy based on the preset photovoltaic string voltage level.

[0059] After receiving the leakage current signal from the dynamic health status matrix, the power supply dynamic compensation module improves the whale optimization algorithm by introducing a dynamic weighting factor and simulated annealing mechanism to adaptively adjust the scaling factor and shift parameters of the wavelet basis function. During the iteration process, the algorithm balances global search and local exploitation capabilities, dynamically optimizing the wavelet packet decomposition level based on the spectral characteristics of the leakage current signal to separate high-frequency transient fault components from low-frequency environmental noise. The optimized wavelet basis function parameters change dynamically with the signal, improving signal decomposition accuracy and outputting a filtered leakage current signal, providing a high signal-to-noise ratio data foundation for subsequent insulation status analysis.

[0060] An environmental parameter-filter threshold mapping table receives environmental data collected in real time from temperature and humidity sensors. A fuzzy logic controller constructs a rule base for the association of temperature, humidity, and filter bandwidth. The controller dynamically adjusts the boundary conditions of the membership function based on the rate of change of ambient humidity and the temperature gradient, calculating the filter bandwidth correction coefficient. Under high humidity conditions, fuzzy rules drive the adaptive filtering unit to expand the passband range and suppress misjudgments caused by low-frequency leakage current fluctuations; in dry environments, the bandwidth is narrowed to enhance the resolution of high-frequency fault components. The corrected filter parameters are synchronously updated to the wavelet basis function adjustment module, forming an environmentally adaptive signal decomposition mechanism.

[0061] A random forest regression model extracts the time-domain statistical features and frequency-domain energy distribution features of the filtered leakage current signal. Combined with the photovoltaic string voltage level parameters output by the acquisition module, a multi-dimensional feature vector is constructed. The model dynamically allocates feature weights using the Gini coefficient, prioritizing the association of the nonlinear mapping relationship between voltage level and leakage current amplitude to calculate the equivalent resistance value 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, prioritizing the isolation of high-risk insulation fault circuits.

[0062] The gradient tripping strategy prioritizes fault circuits based on the voltage levels of the photovoltaic strings. High-voltage strings correspond to high-risk fault areas, triggering rapid tripping commands; low-voltage strings employ delayed tripping logic. Upon receiving the graded alarm signals, the DC circuit breaker executes the tripping action according to a preset voltage-time mapping relationship. A pre-charging circuit absorbs energy from the inductive load, reducing the risk of arc reignition during the tripping process. This graded protection mechanism avoids malfunctions caused by single-threshold criteria in complex environments, while also reducing unnecessary power outages in non-faulty circuits, maintaining system power continuity.

[0063] The control optimization module receives the alarm signal triggered by the power supply dynamic compensation module, optimizes the parsing priority of the tripping control command and the arc-extinguishing chamber impedance adjustment command generated by the alarm signal based on the communication path weight of the preset time-sensitive network, and matches the circuit breaker tripping action time window and the arc-extinguishing chamber impedance change curve corresponding to the tripping control command with the preset dynamic time warping algorithm, and adjusts the arc-extinguishing chamber impedance parameters to suppress arc reignition.

[0064] The control optimization module deploys a federated learning framework on edge computing nodes to dynamically optimize communication path weights in a time-sensitive network architecture. The federated learning framework employs a distributed edge device local model training and global parameter aggregation mechanism. Each edge node constructs a path quality assessment matrix based on its local network topology and data packet transmission latency characteristics. This matrix is ​​then uploaded to the global model via encrypted gradient parameters for weighted averaging and fusion, generating optimized communication path weights. The optimized weights are synchronized to the dynamic time warping algorithm unit, prioritizing the allocation of highly reliable communication links for transmitting tripping control commands and arc-extinguishing chamber impedance adjustment commands, thereby reducing the transmission latency of critical protection commands.

[0065] The deep Q-network algorithm unit receives the abnormal hotspot distribution heatmap and current-voltage curve anomaly analysis results output by the power supply module, and analyzes the temperature gradient, current imbalance, and PID effect classification probability parameters of the hotspot region in the fault feature map. The algorithm constructs a state space with fault type, location coordinates, and risk level as dimensions, and combines protection action effect data from the historical fault handling case library to design a reward function based on action response speed and fault suppression effect. Through a dual network structure balancing exploration and utilization strategy, the parsing weights of overcurrent protection, islanding detection, and insulation blocking commands are dynamically adjusted, prioritizing the scheduling of protection action sequences that match the current fault evolution rate.

[0066] The dynamic time warping algorithm unit matches the mechanical response curve of the circuit breaker's tripping mechanism with the zero-crossing timing of the fault current based on optimized communication path weight parameters. The algorithm aligns the circuit breaker contact trajectory with the arc energy decay curve using a dynamic bending path function, calculating the tripping time window that satisfies the requirement of natural zero-crossing of the arc current. The pre-charging circuit generates a buffer capacitor charging and discharging control pulse sequence based on the time window parameters, absorbing the magnetic energy released by the inductive load during tripping in stages, reducing the probability of arc reignition during contact separation.

[0067] The model predictive control unit monitors the arc voltage and ion concentration parameters within the arc-extinguishing chamber in real time, constructing a multi-physics coupled predictive model that incorporates the dynamic changes in arc impedance. Based on the tripping time window parameters output by the dynamic time warping algorithm, the control unit employs a rolling time-domain optimization strategy to calculate the optimal control sequence for the arc-extinguishing chamber nozzle opening, the magnetic blow-out coil current, and the gas-blowing medium flow rate. During execution, high-frequency sampling of arc impedance characteristics dynamically corrects model prediction biases, adjusts the excitation current intensity of the magnetic blow-out coil to match the arc trajectory, and suppresses the rise rate of the transient recovery voltage. The arc-extinguishing chamber pressure feedback signal is synchronously input into the control model to optimize the gas-blowing medium flow distribution strategy and block the ionization conditions required for arc reignition.

[0068] The power supply module receives the infrared hot spot distribution thermal map and current and voltage curve anomaly analysis results generated by the acquisition module, performs feature-level fusion with the component aging trend prediction data output by the digital twin module, generates the power supply circuit insulation degradation assessment results, and drives the overcurrent protection threshold of the DC distribution cabinet to be dynamically adjusted.

[0069] After receiving the infrared hotspot distribution heatmap generated by the acquisition module, the power supply module uses an improved U-Net network to extract hotspot region boundary features at different resolutions through multi-level convolution operations of the encoder. The decoder employs a skip connection structure to fuse shallow texture features and deep semantic information, generating pixel-level segmentation results. The segmented heatmap is spatially aligned with the string current imbalance parameters obtained by the current and voltage characteristic curve acquisition module. The temperature gradient values ​​and corresponding current deviation ratios of the hotspot regions are labeled to form an abnormal distribution heatmap that quantitatively assesses the severity of hotspots, providing spatial location data for insulation degradation analysis.

[0070] When analyzing current and voltage characteristic curve data, the attention-enhanced SVM model employs a multi-head attention mechanism to dynamically allocate feature weights for different curve segments, focusing on open-circuit voltage drop segments under low irradiance. The model projects high-dimensional features to a low-dimensional space through kernel function mapping and constructs a PID effect identification boundary using a soft-spacing classification strategy, generating a diagnostic report that includes power decay rate and reversibility assessment. The diagnostic results are correlated with the inverter's maximum power point tracking log data to establish a correlation between the PID effect and the aging rate of insulation materials, providing electrical anomaly feature input for protection threshold adjustment.

[0071] The spatiotemporal graph convolutional network receives component aging trend prediction data output from the digital twin module and constructs a spatiotemporal topology graph with photovoltaic strings as nodes and electrical connections as edges. The network's time-axis convolutional layer extracts the temporal decay characteristics of the conductivity and dielectric constant of the insulating material, while the spatial graph convolutional layer captures the synergistic aging effects between adjacent strings. The feature fusion layer performs cross-modal concatenation of the temperature gradient data from the thermal map of abnormal hotspot distribution with the aging trend prediction results, establishing a nonlinear mapping model between the hotspot 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 distribution cabinet. Based on environmental temperature and humidity parameters, the action delay threshold of the ground fault protection is dynamically corrected, achieving adaptive matching between the overcurrent protection threshold and the insulation resistance decay curve.

[0072] The digital twin module receives the dynamic health status matrix, constructs a photovoltaic array digital twin model including temperature field distribution and defect spatial mapping, and synchronizes the insulation degradation assessment results output by the power supply module to the preset virtual system through a preset transfer learning algorithm, generates preventive maintenance instructions and feeds them back to the spatiotemporal graph convolutional network model of the power supply module.

[0073] After receiving the dynamic health status matrix, the digital twin module extracts the temperature field distribution data and defect spatial mapping map contained within the matrix. The temperature field distribution data is processed using an improved variational mode decomposition algorithm to eliminate environmental noise interference, generating temperature gradient distribution data with spatial coordinate mapping relationships. The defect spatial mapping map, combined with the electrical topology, uses an adaptive threshold segmentation algorithm to label the geometric morphology and electrical isolation status of hidden cracks and broken grid defects. This data is input into the construction module of the photovoltaic array digital twin model to establish a virtual system topology with strings as nodes and electrical connections as edges, synchronously mapping the temperature field and defect distribution characteristics of the real system.

[0074] The transfer learning algorithm narrows the feature distribution differences between the real and virtual systems through a domain adversarial training strategy. The algorithm introduces a gradient inversion operation at the feature extraction layer, forcing the encoder to generate shared feature representations independent of the data source. Cross-domain feature alignment is performed between the temperature field distribution data of the real system and the thermal conduction simulation results of the virtual system. This, combined with electroluminescence defect features and the microstructural model parameters of the virtual components, generates a physically consistent defect evolution map. Based on the aligned dataset, the virtual system constructs a thermo-electric coupling simulation model, simulating the co-evolution path of thermal aging and defect diffusion in insulating materials through a hidden Markov chain.

[0075] The preventative maintenance instruction generation module evaluates the risk level of each evolution path using a random forest classifier based on the key parameter threshold triggering conditions in the fault evolution simulation results. The classifier input parameters include the predicted insulation resistance, hot spot area expansion rate, and string current imbalance in the virtual system. The output includes a set of maintenance strategies such as component replacement priorities, cleaning cycle suggestions, and protection threshold adjustment parameters. The generated maintenance instructions are input into the spatiotemporal graph convolutional network model of the power supply module. By updating the material aging coefficient and environmental stress weight parameters in the network node attributes, the spatiotemporal correlation of aging trend prediction is optimized.

[0076] The spatiotemporal graph convolutional network integrates historical repair records from maintenance instructions in the temporal dimension to establish a dynamic correlation model between component performance recovery and subsequent aging rate. In the spatial dimension, it combines string electrical connection topology to predict the cross-regional impact of local hotspots 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 between monitoring data and simulation prediction. The protection threshold adjustment parameters in the maintenance instructions drive the dynamic protection logic of the DC distribution cabinet, matching the insulation resistance decay curve according to environmental temperature and humidity conditions to achieve adaptive adjustment of the overcurrent protection action threshold.

[0077] The photovoltaic module fault monitoring system provided by this invention achieves fault monitoring and protection control of photovoltaic arrays through multi-module collaboration. The acquisition module obtains infrared thermal imaging data, electroluminescence defect data, and current and voltage characteristic curve data of the photovoltaic module through a distributed monitoring network. A multi-scale feature alignment algorithm is used to preprocess the multi-source data. This algorithm uses a sliding window mechanism to extract time-series segments of data from each sensor, and utilizes dynamic time warping to calculate feature point matching paths at different sensor sampling rates, aligning the time-series benchmarks of 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 spatiotemporal correlation features of temperature gradient, defect propagation rate, and current and voltage fluctuation amplitudes to generate a dynamic health status matrix containing multi-dimensional health indicators.

[0078] The power supply dynamic compensation module receives leakage current signals from the dynamic health status matrix and uses an improved whale optimization algorithm to dynamically adjust the scaling factor and displacement parameters of the wavelet basis function. The improved whale optimization algorithm introduces a simulated annealing mechanism during the iteration process, balancing global search and local exploitation capabilities through dynamic weight factors, and adaptively optimizing the number of wavelet packet decomposition levels and frequency band division accuracy. The optimized leakage current signal is input into a random forest regression model, combined with photovoltaic string voltage level data output from the acquisition module, to construct an insulation resistance calculation model based on the voltage-leakage current nonlinear mapping relationship. The model dynamically allocates feature weights through the Gini coefficient, prioritizing the correlation of leakage current amplitude characteristics at high voltage levels, generating a dynamic insulation criterion reflecting changes in the conductivity of the insulation material. When the insulation criterion falls below a preset threshold, a gradient tripping strategy prioritizes fault circuits according to string voltage levels, triggering the DC circuit breaker to perform graded tripping actions.

[0079] The control optimization module optimizes the transmission path of the tripping control command based on a preset time-sensitive network architecture. It employs a federated learning framework to perform distributed training and global aggregation of communication path weight parameters among edge computing nodes. The federated learning framework encrypts the local model parameter update process using differential privacy technology and dynamically adjusts the parsing priority of protection commands based on network topology status. The dynamic time warping algorithm receives the optimized communication path weight parameters, matches the circuit breaker's mechanical response curve with the zero-crossing timing of the fault current, and predicts the tripping time window that meets the arc energy suppression requirements. The arc-extinguishing chamber impedance adjustment command is generated based on the time window parameters. The model prediction control unit monitors the arc impedance change curve in real time and adjusts the magnetic blow-out coil current and air-blowing medium flow parameters through a rolling time-domain optimization strategy to suppress transient overvoltages and block arc reignition conditions.

[0080] The power supply module receives the infrared hotspot distribution heatmap and current-voltage curve anomaly analysis results generated by the acquisition module, and uses an improved U-Net network to perform pixel-level segmentation of the thermal image. The network fuses shallow texture features and deep semantic information through a skip connection structure, and combines string current imbalance parameters to label the temperature gradient and power loss ratio of the hotspot region, generating an abnormal hotspot distribution heatmap. An attention-enhanced SVM model focuses on local anomalies in the current-voltage characteristic curves, and extracts the open-circuit voltage drop pattern caused by the PID effect through kernel function mapping. A spatiotemporal graph convolutional network receives component aging trend prediction data output by the digital twin module, establishes a spatiotemporal correlation model between insulation material degradation and hotspot temperature rise rate, and drives the overcurrent protection threshold of the DC distribution cabinet to dynamically adjust according to ambient temperature and humidity parameters.

[0081] The digital twin module constructs a digital twin model of the photovoltaic array, synchronizing temperature field distribution, defect spatial mapping, and electrical parameter data between the real and virtual systems through a transfer learning algorithm. The transfer learning algorithm employs a domain adversarial training strategy to reduce the difference in feature distribution between real and simulated data, generating a virtual component state dataset with physical consistency. The digital twin model simulates fault evolution paths under different environmental stresses using the Monte Carlo method, and combines a Markov chain model to predict the synergistic effect of hot spot diffusion and insulation degradation. The preventive maintenance instruction generation module outputs a maintenance strategy including component replacement priority and protection threshold adjustment parameters based on key parameter threshold triggering conditions in the simulation results. Maintenance instructions are fed back to a spatiotemporal graph convolutional network to optimize the material aging coefficient in node attributes, forming a closed-loop optimization mechanism between monitoring data and the prediction model.

[0082] Specifically, in the photovoltaic module fault monitoring system of the present invention, the digital twin module is further used for:

[0083] Receive the dynamic health status matrix generated by the acquisition module, and construct a digital twin model of the photovoltaic array including temperature field distribution data, defect space mapping map and compensated current and voltage curve data;

[0084] The temperature field distribution data and defect space mapping map in the dynamic health status matrix are synchronized to a preset virtual system through a preset transfer learning algorithm. The preset virtual system is used to simulate the fault propagation path and generate preventive maintenance instructions.

[0085] The preventive maintenance command is input into the spatiotemporal graph convolutional network model in the power supply module to optimize the confidence level of the component aging trend prediction results generated by the power supply module.

[0086] 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 map, and compensated current and voltage characteristic curve data contained in the matrix. The temperature field distribution data is processed using an improved variational mode decomposition algorithm to eliminate environmental noise interference and generate a temperature gradient distribution map with physical coordinate mapping relationships. The defect space mapping map, combined with the electrical topology, 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 uses an adversarial generative network to correct line impedance errors while preserving the transient output characteristics of the components. This data is input into the digital twin model's construction module through a timestamp alignment mechanism to establish a virtual system topology with photovoltaic strings as nodes and electrical connections as edges, synchronously mapping the temperature field, defect distribution, and electrical parameter characteristics of the real system.

[0087] The transfer learning algorithm narrows the feature distribution differences between the real and virtual systems through a domain adversarial training strategy. A gradient inversion layer is introduced into the feature extraction layer, forcing the feature encoder to generate shared feature representations independent of the data source. Temperature field distribution data from the real system and thermal conduction simulation results from the virtual system are aligned across domains. Combined with electroluminescence defect features and microstructural model parameters of the virtual components, a physically consistent defect evolution map is generated. Based on the aligned dataset, the virtual system constructs a Markov chain model to simulate fault propagation paths under different salt spray concentrations, temperature, humidity, and environmental stresses. The Monte Carlo method generates multiple sets of fault evolution trajectories and, by combining the coupling relationship between the temperature rise rate of the hot spot region and the attenuation of the insulation material's conductivity, predicts the cross-regional impact of hot spot propagation on adjacent components.

[0088] The preventive maintenance instruction generation module evaluates the risk level of each evolution path using a random forest classifier based on the key parameter threshold triggering conditions in the fault evolution simulation results. The classifier input parameters include the predicted insulation resistance, hot spot area expansion rate, and string current imbalance in the virtual system. The output includes a set of maintenance strategies such as component replacement priorities, cleaning cycle suggestions, and protection threshold adjustment parameters. The generated preventive maintenance instructions are input to the spatiotemporal graph convolutional network model of the power supply module. By updating the material aging coefficient and environmental stress weight parameters in the network node attributes, the spatiotemporal correlation of aging trend prediction is optimized. The spatiotemporal graph convolutional network integrates historical repair record data from the maintenance instructions in the time dimension to establish a dynamic correlation model between component performance recovery and subsequent aging rate; 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 between monitoring data and the prediction model.

[0089] Specifically, the photovoltaic module fault monitoring system of the present invention includes, in its acquisition module:

[0090] An infrared thermal imaging sensor is used to collect temperature field distribution data of photovoltaic modules and generate denoised temperature field data by eliminating environmental noise interference through a preset improved variational mode decomposition algorithm.

[0091] An electroluminescence detection unit is used to process the electroluminescence image of a photovoltaic module based on a preset adaptive threshold segmentation algorithm, and generate a defect space mapping map by combining the electrical topology of the photovoltaic module.

[0092] The current and voltage characteristic curve acquisition module is used to compensate for line impedance errors by using an adversarial generation network to generate compensated current and voltage characteristic curve data.

[0093] The denoised temperature field data, defect space mapping map, and compensated current and voltage characteristic curve data are input into a preset bidirectional long short-term memory network. Based on the operating timestamp of the photovoltaic module, multi-source data time sequence alignment is performed to generate a dynamic health status matrix.

[0094] The acquisition module obtains surface temperature field distribution data of photovoltaic modules through infrared thermal imaging sensors, and performs mode decomposition on the raw 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 kurtosis-energy joint criteria to select effective temperature feature modes, suppressing the influence of environmental radiation noise on the temperature gradient distribution. The reconstructed denoised temperature field data is spatially mapped with the heat sink layout parameters of the photovoltaic modules to generate a temperature field distribution map with physical location correlation, providing basic data support for hot spot effect analysis.

[0095] After acquiring electroluminescence images of the photovoltaic module, the electroluminescence detection unit dynamically adjusts the image segmentation parameters based on an adaptive threshold segmentation algorithm. The algorithm expands the defect boundary contours using a region growing method based on the electroluminescence intensity distribution characteristics, and establishes a spatial mapping model between the defect region and circuit nodes by combining the electrical topology connections of the module's cells. The generated defect spatial mapping map annotates the geometric morphology and electrical isolation status of hidden cracks and broken grid defects, quantifies the impact weight of defects on string output characteristics, and provides microstructural data for subsequent insulation degradation analysis.

[0096] 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 an adversarial generative network. The generator network simulates the current and voltage characteristic curves under ideal line conditions, while the discriminator network compares the characteristic differences between the measured data and the generated data, dynamically correcting the measurement deviation caused by the line impedance. The compensated current and voltage characteristic curve data uses a sliding window mechanism to extract the key inflection points of the characteristic curves, preserving the characteristic information of the module during transient processes such as shading and bypass diode operation, forming a high-precision electrical parameter dataset.

[0097] Denoising temperature field data, defect spatial mapping maps, and compensated current and voltage characteristic curves are input into a bidirectional long short-term memory network for time-series alignment processing. The network's forward and backward recurrent units extract feature vectors at different time scales to establish a time-series correlation model of temperature field changes, defect propagation, and electrical parameter decay. The feature fusion layer concatenates the spatiotemporal feature vectors from multiple sources, combining them with the photovoltaic system's operating timestamps to establish a unified time-series benchmark, generating a dynamic health status matrix that includes thermal-electric coupling characteristics, defect evolution trends, and component health indices. This matrix is ​​transmitted to subsequent modules via a distributed monitoring network, providing multi-dimensional data input for insulation condition assessment and protection strategy generation.

[0098] Specifically, in the photovoltaic module fault monitoring system of the present invention, the power supply dynamic compensation module includes:

[0099] An adaptive filtering unit is used to receive the leakage current signal in the dynamic health status matrix, dynamically adjust the wavelet packet decomposition level and threshold parameter of the wavelet basis function of the leakage current signal, and output the filtered leakage current signal.

[0100] An environmental parameter-to-filter threshold mapping table is used to combine temperature and humidity sensor data and correct the filter bandwidth of the adaptive filter unit through a preset fuzzy logic controller.

[0101] A random forest regression model is used to input the filtered leakage current signal into a pre-trained regression model, and to calculate the insulation resistance value and generate a dynamic insulation criterion by combining the preset photovoltaic string voltage level in the current and voltage characteristic curve data generated by the acquisition module.

[0102] The power supply dynamic compensation module receives the leakage current signal from the dynamic health status matrix and performs signal decomposition and feature extraction through an adaptive filtering unit. The adaptive filtering unit uses an improved whale optimization algorithm to dynamically adjust the scaling factor and shift parameters of the wavelet basis function, optimizing the frequency band division accuracy of the wavelet packet decomposition. The improved whale optimization algorithm introduces a dynamic weighting factor and simulated annealing mechanism during the iteration process to balance global search and local exploitation capabilities, adaptively matching the spectral characteristics 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.

[0103] An environmental parameter mapping table, combined with real-time environmental data collected by temperature and humidity sensors, is used to construct a rule base for the association of temperature, humidity, and filter bandwidth through a fuzzy logic controller. The fuzzy logic controller dynamically adjusts the boundary conditions of the membership function based on the rate of change of ambient humidity and the temperature gradient, calculating the filter bandwidth correction coefficient. Under high humidity conditions, the fuzzy rules drive the adaptive filtering unit to expand the passband range and suppress misjudgments caused by low-frequency leakage current fluctuations; in dry environments, the bandwidth is narrowed to enhance the resolution of high-frequency fault components. The corrected filter parameters are synchronously updated to the wavelet basis function adjustment module, forming an environmentally adaptive signal decomposition mechanism.

[0104] The random forest regression model receives the filtered leakage current signal, extracts its time-domain statistical features and frequency-domain energy distribution features, and combines these with the photovoltaic string voltage level parameters from the current and voltage characteristic curve data generated by the acquisition module to construct a multi-dimensional feature vector. The model dynamically adjusts the feature weight allocation through the Gini coefficient, prioritizing the association of the nonlinear mapping relationship between voltage level and leakage current amplitude to calculate the equivalent resistance value 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 coordinating the circuit breaker to perform gradient tripping action, prioritizing the isolation of high-risk insulation fault circuits.

[0105] Specifically, in the photovoltaic module fault monitoring system of the present invention, the control optimization module includes:

[0106] Edge computing nodes are used to optimize the communication path selection strategy of time-sensitive networks using a federated learning framework, and generate optimized communication path weights.

[0107] The deep Q-network algorithm unit is used to dynamically adjust the analytical weights of the tripping control command and the arc-extinguishing chamber impedance adjustment command based on the fault feature map generated by the abnormal hot spot distribution heat map and the abnormal current and voltage curve analysis results output by the power supply module.

[0108] The dynamic time warping algorithm unit is used to predict the optimal time window for the circuit breaker tripping action based on the optimized communication path weight, and to absorb inductive load energy through the pre-charging circuit.

[0109] 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, thereby suppressing the risk of transient overvoltage and arc reignition.

[0110] The control optimization module deploys a federated learning framework on edge computing nodes to dynamically optimize communication path weights in a time-sensitive network architecture. The federated learning framework employs a distributed edge device local model training and global parameter aggregation mechanism. Each edge node constructs a path quality assessment matrix based on its local network topology and data packet transmission latency characteristics. This matrix is ​​then uploaded to the global model via encrypted gradient parameters for weighted averaging and fusion, generating optimized communication path weights. The optimized weights are synchronized to the dynamic time warping algorithm unit, prioritizing the allocation of highly reliable communication links for transmitting tripping control commands and arc-extinguishing chamber impedance adjustment commands, thereby reducing the transmission latency of critical protection commands.

[0111] The deep Q-network algorithm unit receives the abnormal hotspot distribution heatmap and current-voltage curve anomaly analysis results output by the power supply module, and analyzes the temperature gradient, current imbalance, and PID effect classification probability parameters of the hotspot region in the fault feature map. The algorithm constructs a state space with fault type, location coordinates, and risk level as dimensions, and combines protection action effect data from the historical fault handling case library to design a reward function based on action response speed and fault suppression effect. Through a dual network structure balancing exploration and utilization strategy, the parsing weights of overcurrent protection, islanding detection, and insulation blocking commands are dynamically adjusted, prioritizing the scheduling of protection action sequences that match the current fault evolution rate.

[0112] The dynamic time warping algorithm unit matches the mechanical response curve of the circuit breaker's tripping mechanism with the zero-crossing timing of the fault current based on optimized communication path weight parameters. The algorithm aligns the circuit breaker contact trajectory with the arc energy decay curve using a dynamic bending path function, calculating the tripping time window that satisfies the requirement of natural zero-crossing of the arc current. The pre-charging circuit generates a buffer capacitor charging and discharging control pulse sequence based on the time window parameters, absorbing the magnetic energy released by the inductive load during tripping in stages, reducing the probability of arc reignition during contact separation.

[0113] The model predictive control unit monitors the arc voltage and ion concentration parameters within the arc-extinguishing chamber in real time, constructing a multi-physics coupled predictive model that incorporates the dynamic changes in arc impedance. Based on the tripping time window parameters output by the dynamic time warping algorithm, the control unit employs a rolling time-domain optimization strategy to calculate the optimal control sequence for the arc-extinguishing chamber nozzle opening, the magnetic blow-out coil current, and the gas-blowing medium flow rate. During execution, high-frequency sampling of arc impedance characteristics dynamically corrects model prediction biases, adjusts the excitation current intensity of the magnetic blow-out coil to match the arc trajectory, and suppresses the rise rate of the transient recovery voltage. The arc-extinguishing chamber pressure feedback signal is synchronously input into the control model to optimize the gas-blowing medium flow distribution strategy and block the ionization conditions required for arc reignition.

[0114] Specifically, in the photovoltaic module fault monitoring system of the present invention, the power supply module includes:

[0115] An improved U-Net network is used to perform pixel-level segmentation of the infrared thermal image generated by the acquisition module to generate a thermal map of abnormal hot spot distribution.

[0116] An attention-enhanced SVM model is used to analyze abnormal fluctuation characteristics in the current and voltage characteristic curve data generated by the acquisition module and identify power attenuation caused by PID effect.

[0117] A spatiotemporal graph convolutional network is used to fuse the component aging trend prediction data output by the digital twin module with the abnormal hot spot distribution heat map in terms of spatiotemporal features, model the insulation degradation inflection point, and drive the dynamic adjustment of the overcurrent protection threshold of the DC distribution cabinet.

[0118] The power supply module performs pixel-level segmentation of the infrared thermal images generated by the acquisition module using an improved U-Net network. The improved U-Net network employs a skip connection structure to fuse shallow texture features with deep semantic information, and extracts multi-scale hotspot region features through a spatial pyramid pooling module. The segmentation results are spatially aligned with the string current imbalance parameters obtained by the current and voltage characteristic curve acquisition module to generate an abnormal hotspot distribution heatmap labeled with hotspot locations, temperature gradients, and corresponding current deviation values. The electrical topology connections of the photovoltaic modules are overlaid on the heatmap to quantitatively assess the impact weight of hotspots on string output power, providing spatial location data for subsequent insulation degradation analysis.

[0119] An attention-enhanced SVM model analyzes abnormal fluctuations in current and voltage characteristic curves, employing a multi-head attention mechanism to dynamically allocate feature weights across different curve segments. The model focuses on the abnormal open-circuit voltage drop characteristics under low irradiance, which characterize the PID effect, within the characteristic curves. It projects nonlinearly separable data from the high-dimensional feature space to a low-dimensional space using kernel function mapping, and constructs a PID effect identification boundary using a soft-spacing classification strategy. The classification results are correlated with the inverter's maximum power point tracking log data to generate a PID effect diagnostic report containing power decay rate, reversibility assessment, and repair suggestions, providing an electrical anomaly characteristic basis for protection threshold adjustment.

[0120] The spatiotemporal graph convolutional network receives component aging trend prediction data and abnormal hotspot distribution heatmaps output from the digital twin module, constructing a spatiotemporal topology graph with photovoltaic strings as nodes and electrical connections as edges. The network extracts the temporal decay characteristics of parameters such as conductivity and dielectric constant of the insulation material through time-axis convolutional layers, and captures the synergistic aging effects between adjacent strings by combining them with spatial graph convolutional layers. The feature fusion layer performs cross-modal concatenation of temperature gradient data and aging trend prediction results in the hotspot region, establishing a nonlinear mapping model between the hotspot temperature rise rate and 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 distribution cabinet. Based on the historical decay curve of the insulation resistance, the action delay parameters of the ground fault protection are dynamically corrected, achieving dynamic matching between the protection threshold and environmental temperature and humidity conditions.

[0121] Specifically, in the photovoltaic module fault monitoring system of the present invention, the power supply module further includes:

[0122] The knowledge distillation unit is used to perform knowledge distillation on the multi-scale feature parameters extracted by the improved U-Net network during infrared thermal image segmentation and the classification weights generated by the attention mechanism-enhanced SVM model in current-voltage curve analysis.

[0123] The knowledge distillation unit in the power supply module extracts multi-scale feature parameters generated by the improved U-Net network during infrared thermal image segmentation, including shallow texture features and deep semantic features. The improved U-Net network captures hotspot region boundary information and temperature gradient distribution features at different resolutions through multi-level convolutional operations in the encoder, while the decoder fuses cross-layer feature maps through skip connection structures to generate pixel-level segmentation results. Simultaneously, the knowledge distillation unit extracts the classification weight matrix generated by the attention-enhanced SVM model during current-voltage curve analysis, quantifying the influence of electrical characteristics in different segments on the PID effect classification decision.

[0124] The knowledge distillation unit uses a channel attention mechanism to modally align multi-scale hotspot features with classification weights, generating a joint feature vector with cross-domain correlation. The channel attention mechanism dynamically adjusts the weight allocation ratio of feature channels based on the correlation between the temperature gradient amplitude and the current anomaly fluctuation amplitude in the hotspot region. The aligned feature vector is input into the feature fusion layer of the knowledge distillation unit, where a gated recurrent unit is used to construct temporal dependencies, capturing the dynamic correlation between the hotspot temperature rise rate and the anomaly fluctuation of the current characteristic curve.

[0125] The distillation loss function constrains the student network to inherit the feature extraction capabilities of the teacher network during feature reconstruction, enabling the lightweight classifier of the student network to directly output composite fault type diagnostic results based on the joint feature vector. The student network learns and improves the cross-modal mapping relationship between the multi-scale spatial feature representation of the U-Net network and the electrical anomaly classification logic of the SVM model by minimizing the weighted loss function of feature reconstruction error and classification error. The optimized joint feature vector is input into the node attribute update module of the spatiotemporal graph convolutional network, mapping the hotspot temperature gradient and current anomaly features to the spatiotemporal topology graph of the component aging trend prediction model.

[0126] The spatiotemporal graph convolutional network dynamically adjusts the weight allocation of multimodal features in node state updates through a graph attention mechanism, establishing a physical correlation model between temperature anomalies in hot spot regions and the decay of conductivity in insulating materials. The prediction results are fed back to the feature alignment module of the knowledge distillation unit, driving the student network to optimize the weight allocation strategy during feature fusion, forming a closed-loop learning mechanism for cross-modal feature optimization and aging trend prediction. The composite fault diagnosis results output by the knowledge distillation unit are synchronously input into the protection threshold adjustment module of the DC distribution cabinet, generating dynamic overcurrent protection action thresholds based on current environmental humidity parameters, forming an adaptive matching mechanism between monitoring data and protection parameters.

[0127] Specifically, in the photovoltaic module fault monitoring system of the present invention, the digital twin module includes:

[0128] The heterogeneous algorithm federated engine is used to run deep residual networks and Vision Transformer models in parallel. It integrates conflicting evidence from the abnormal hot spot distribution heat map and current and voltage anomaly analysis results generated by the power supply module through Dempster-Shafer evidence theory, and outputs multi-model joint diagnostic results.

[0129] An improved A* algorithm unit is used to switch to the local cached data of the acquisition module when communication is abnormal. Based on the temperature field distribution data and defect space mapping in the dynamic health status matrix, it executes fast protection logic and triggers the inverter derating operation mode.

[0130] The digital twin module utilizes a heterogeneous algorithm federated engine to run a deep residual network and a VisionTransformer model in parallel, constructing a multimodal fault diagnosis framework. The deep residual network extracts local texture features and spatial context information from infrared hotspot distribution heatmaps, preserving detailed hotspot boundary features through residual skip connections. The VisionTransformer model, based on a self-attention mechanism, analyzes the temporal correlation characteristics of abnormal current and voltage data, capturing the evolution of PID effects and insulation degradation over time. The intermediate layer feature maps of the two models are input into the Dempster-Shafer evidence theory fusion module. By calculating the basic probability allocation function of each model's output and dynamically adjusting the evidence synthesis rules based on conflict factors, the module eliminates evidence conflicts between abnormal hotspot distribution and current and voltage analysis results, generating a multi-model joint diagnostic result. The fusion result is labeled with fault type confidence and spatial location coordinates, providing multi-dimensional decision-making basis for protection strategies.

[0131] When the improved A* algorithm unit detects a communication link anomaly, it switches to the locally cached data stored in the acquisition module and calls the historical temperature field distribution data and defect spatial mapping map in the dynamic health status matrix. The algorithm constructs a topology search network with the photovoltaic string health index as nodes and electrical connection relationships as edges, and dynamically adjusts the heuristic function weights based on the spatial constraints of the fault propagation path. During the search process, it prioritizes matching the historical protection case library under the current environmental temperature and humidity parameters, and generates a fault risk coefficient by combining the temperature gradient threshold and the defect propagation rate, calculating the optimal protection path that satisfies 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, triggering the string-level derating operation mode to reduce the risk of local overheating and insulation breakdown.

[0132] After communication is restored, the digital twin module sends the locally generated protection command log back to the evidence theory fusion module of the heterogeneous algorithm federated engine to update the prior parameters of the confidence assignment function of the deep residual network and the Vision Transformer model. Simultaneously, local protection execution performance data is input into the parameter correction module of the digital twin model, which adjusts the coefficients of the equivalent aging rate equation of components in the virtual system through transfer learning algorithms, optimizing the boundary conditions for subsequent fault evolution simulations. The deviation between the protection command execution performance and the digital twin prediction results is input into the online fine-tuning module of the federated engine, driving parameter updates in the feature extraction layer of the deep residual network and the self-attention layer of the Vision Transformer, forming a closed-loop feedback loop that improves diagnostic accuracy and optimizes the fault tolerance mechanism.

[0133] Specifically, the photovoltaic module fault monitoring system of the present invention further includes, in the case of the digital twin module:

[0134] The transfer learning unit is used to transfer the temperature field distribution data and defect space mapping map 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 defect space mapping map, the unit simulates the thermal aging of the insulation material and the defect propagation path, and generates preventive maintenance instructions for the power supply module.

[0135] 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 defect spatial mapping map from the matrix. The temperature field distribution data is processed using an improved variational mode decomposition algorithm to eliminate environmental noise interference, generating temperature gradient distribution data with spatial coordinate mapping. The defect spatial mapping map, combined with the electrical topology of the photovoltaic module, 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 undergoes feature alignment through a gradient inversion layer in the domain adaptation algorithm, reducing the difference in feature distribution between the real and virtual systems, and generating a physically consistent virtual component state dataset.

[0136] The transfer learning unit employs a domain adversarial training strategy, introducing a gradient inversion operation in the feature extraction layer to force the encoder to generate shared feature representations independent of the data source. Temperature field distribution data from the real system is aligned across domains with thermal conduction simulation results from the virtual system. By combining electroluminescence defect features with microstructural model parameters of the virtual components, a spatiotemporal mapping relationship for defect evolution is established. Based on the aligned dataset, the virtual system constructs a thermo-electric coupling simulation model, simulating the co-evolution path of thermal aging of insulating materials and defect propagation through a hidden Markov chain. The Monte Carlo method generates multiple sets of fault evolution trajectories, combining the physical equations of temperature rise rate in hotspot regions and conductivity decay of insulating materials to predict the cross-regional impact of local overheating on the insulation performance of adjacent components.

[0137] The preventative maintenance instruction generation module uses a random forest classifier to assess the risk level of each evolution path based on the key parameter threshold triggering conditions in the fault evolution simulation results. The classifier input parameters include the predicted insulation resistance, hot spot area expansion rate, and string current imbalance in the virtual system. The output includes a set of maintenance strategies such as component replacement priorities, cleaning cycle suggestions, and protection threshold adjustment parameters. The generated maintenance instructions are input into the spatiotemporal graph convolutional network model of the power supply module. By updating the material aging coefficient and environmental stress weight parameters in the network node attributes, the spatiotemporal correlation of aging trend prediction is optimized. The spatiotemporal graph convolutional network integrates historical repair record data from the maintenance instructions in the time dimension to establish a dynamic correlation model between component performance recovery and subsequent aging rate; 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.

[0138] 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 deviates from the predicted value, an online fine-tuning mechanism for the model parameters is triggered. During the fine-tuning process, an elastic weight solidification technique is used to retain existing knowledge while adapting to new aging mode characteristics, avoiding prediction oscillations caused by drastic parameter adjustments. 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 of subsequent fault evolution path prediction, forming a closed-loop optimization link between 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, matching the insulation resistance decay curve according to the environmental temperature and humidity conditions to achieve adaptive adjustment of the overcurrent protection action threshold.

[0139] Secondly, please refer to Figure 1 This invention provides a photovoltaic module fault monitoring method, applied to the aforementioned photovoltaic module fault monitoring system, comprising the following steps:

[0140] Step S101: Collect photovoltaic module data, which includes infrared thermal imaging data, electroluminescence defect data, and current-voltage characteristic curve data. The photovoltaic module data is time-series aligned using a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix.

[0141] Step S102: Receive the leakage current signal, dynamically adjust the wavelet packet decomposition level and threshold parameter of the wavelet basis function of the leakage current signal through a preset improved whale optimization algorithm, generate a filtered leakage current signal, and calculate the insulation resistance value in combination with the preset photovoltaic string voltage level to generate a dynamic insulation criterion. When the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and the preset DC circuit breaker is linked to execute a gradient tripping strategy based on the preset photovoltaic string voltage level.

[0142] Step S103: Receive alarm signal, optimize the parsing priority of the tripping control command and the arc-extinguishing chamber impedance adjustment command generated by the alarm signal based on the communication path weight of the preset time-sensitive network, match the circuit breaker tripping action time window and the arc-extinguishing chamber impedance change curve corresponding to the tripping control command through the preset dynamic time warping algorithm, and adjust the arc-extinguishing chamber impedance parameters to suppress arc reignition.

[0143] Step S104: Receive infrared hot spot distribution thermal map, current and voltage curve anomaly analysis results, and component aging trend prediction data. Perform feature-level fusion on the infrared hot spot distribution thermal map, current and voltage curve anomaly analysis results, and component aging trend prediction data to generate power supply circuit insulation degradation assessment results. The power supply circuit insulation degradation assessment results are used to drive the dynamic adjustment of the overcurrent protection threshold of the preset DC distribution cabinet.

[0144] Step S105: Receive the dynamic health status matrix, construct a photovoltaic array digital twin model including temperature field distribution and defect spatial mapping, synchronize the insulation degradation assessment results output by the power supply module to the preset virtual system through a preset transfer learning algorithm, generate preventive maintenance instructions and feed them back to the spatiotemporal graph convolutional network model.

[0145] The photovoltaic module fault monitoring method provided by this invention achieves state monitoring and protection optimization of the photovoltaic system through multi-step collaborative processes. In step S101, an infrared thermal imaging sensor collects surface temperature field distribution data of the photovoltaic module at a preset sampling frequency. An improved variational mode decomposition algorithm is used to perform mode decomposition on the original temperature data. Valid temperature feature modes are selected using a kurtosis-energy joint criterion to reconstruct the denoised temperature field distribution data. An electroluminescence detection unit, combined with the electrical topology, uses an adaptive threshold segmentation algorithm to process the electroluminescence image, generating a defect space mapping map that annotates the geometric morphology of hidden cracks and the electrical isolation status. A current and voltage characteristic curve acquisition module compensates for line impedance errors through an adversarial generative network, generating high-precision electrical parameter data. Multi-source data is input into a bidirectional long short-term memory network for time-series alignment. Feature vectors at different time scales are extracted through forward and backward recurrent units, and temperature gradient, defect propagation rate, and current and voltage fluctuation features are fused to generate a dynamic health state matrix.

[0146] In step S102, the improved whale optimization algorithm dynamically adjusts the scaling factor and shift parameters of the wavelet basis function. A simulated annealing mechanism is introduced to optimize the frequency band division accuracy of the wavelet packet decomposition, separating environmental noise and actual insulation fault characteristics in the leakage current signal. A fuzzy logic controller, combined with temperature and humidity sensor data, corrects the filtering bandwidth to suppress low-frequency interference under high humidity conditions. A random forest regression model extracts the time-domain statistical features and frequency-domain energy distribution features of the filtered leakage current signal, correlates them with the photovoltaic string voltage level parameters output by the acquisition module, constructs a voltage-leakage current nonlinear mapping model, and generates dynamic insulation criteria. When the detected insulation resistance value is lower than the dynamic threshold, the gradient tripping strategy prioritizes the fault circuits according to the voltage level, triggering the DC circuit breaker to perform graded tripping actions.

[0147] Step S103 optimizes communication path weights within a time-sensitive network architecture using a federated learning framework. Edge computing nodes construct path quality assessment matrices based on local network topology and transmission delay characteristics. The global model aggregates multi-node training results to generate optimized weights. A dynamic time warping algorithm matches the circuit breaker's mechanical response curve with the zero-crossing timing of the fault current, predicts the tripping time window, and generates a pre-charge circuit control pulse sequence to absorb inductive load energy in stages. The model predictive control unit continuously optimizes the arc-extinguishing chamber nozzle opening and magnetic blow-out coil current parameters based on the arc impedance change curve, suppressing transient overvoltages and blocking arc reignition conditions.

[0148] In step S104, the improved U-Net network performs pixel-level segmentation of the infrared thermal image through a skip connection structure, and generates a thermal map of abnormal hot spot distribution by combining string current imbalance parameters. The attention-enhanced SVM model focuses on the abnormal open-circuit voltage drop characteristics in the current-voltage characteristic curves, identifying power attenuation patterns caused by PID effects. The spatiotemporal graph convolutional network fuses the aging trend data output from the digital twin module with the hot spot distribution characteristics to establish a spatiotemporal correlation model between the hot spot temperature rise rate and insulation material degradation, driving the DC distribution cabinet to dynamically adjust the overcurrent protection threshold according to ambient temperature and humidity parameters.

[0149] Step S105 constructs a digital twin model of the photovoltaic array, synchronizing the temperature field distribution and defect spatial mapping data of the real and virtual systems through a transfer learning algorithm. A domain adversarial training strategy reduces the difference in feature distribution between the real and simulated data, generating a virtual component state dataset with physical consistency. The digital twin model simulates fault propagation paths under different environmental stresses using the Monte Carlo method, and combines hidden Markov chains to pre-determine the co-evolution process of thermal aging of insulation materials and defect propagation. The preventive maintenance instruction generation module triggers threshold conditions based on simulation results, outputs component replacement priorities and protection parameter adjustment strategies, and feeds them back to the spatiotemporal graph convolutional network to optimize the aging trend prediction model, forming a closed-loop optimization mechanism between monitoring data and simulation prediction.

[0150] The explanations of the technical feature terms in the technical solution of this invention are as follows:

[0151] 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 an improved variational mode decomposition algorithm to eliminate environmental radiation noise and generate denoised temperature gradient distribution data.

[0152] Electroluminescence detection unit: A device for detecting internal defects in photovoltaic modules based on the principle of electroluminescence. It excites the cells to produce fluorescence by applying voltage, analyzes the brightness difference in the emission image using an adaptive threshold segmentation algorithm, and generates a mapping map of the geometric shape and electrical isolation state of hidden cracks and broken grid defects by combining the electrical topology.

[0153] Generative Adversarial Networks (GANs) are deep learning models used to compensate for line impedance errors, consisting of a generator and a discriminator. The generator simulates the current and voltage characteristic curves under ideal line conditions, while the discriminator compares the feature differences between measured data and generated data, dynamically correcting measurement deviations and improving the accuracy of electrical parameter acquisition.

[0154] Improved Whale Optimization Algorithm: An intelligent algorithm for optimizing wavelet basis function parameters. By introducing dynamic weighting factors and simulated annealing mechanisms, it balances global search and local exploitation capabilities, adaptively adjusts the number of wavelet packet decomposition layers and frequency band division accuracy, and separates high-frequency fault components and low-frequency environmental noise in leakage current signals.

[0155] Fuzzy logic controller: An environment adaptive control module based on fuzzy rules. By constructing a three-dimensional membership function of temperature-humidity-filter bandwidth, it dynamically adjusts the filter bandwidth range to suppress low-frequency fluctuation interference of leakage current signal under high humidity environment.

[0156] Random forest regression model: A machine learning model used to calculate insulation resistance values. It extracts the time-domain statistical features and frequency-domain energy distribution features of the filtered leakage current signal, associates them with the voltage level parameters of the photovoltaic string to construct a nonlinear mapping relationship, and generates a dynamic criterion reflecting the change in the conductivity of the insulation material.

[0157] Federated learning framework: A distributed machine learning architecture deployed in time-sensitive networks. It optimizes communication path weight allocation and reduces the transmission latency of critical protection commands by training locally on edge computing nodes and aggregating global model parameters.

[0158] Dynamic Time Warping (DTW): A timing matching algorithm used to predict the circuit breaker tripping time window. It aligns the circuit breaker's mechanical response curve with the fault current zero-crossing timing using a dynamic bending path function, and combines the pre-charge circuit control pulse sequence to absorb inductive load energy in stages.

[0159] 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, and combines a spatial pyramid pooling module to extract multi-scale hot spot region features to generate a hot spot distribution heatmap labeled with temperature gradient and current imbalance.

[0160] An attention-enhanced SVM model: A classification model combining Support Vector Machine (SVM) and multi-head attention mechanism, focusing on the open-circuit voltage drop segment in the current-voltage characteristic curve that characterizes the PID effect, and identifying power decay patterns through kernel function mapping and soft-interval classification strategy.

[0161] Spatiotemporal Graph Convolutional Network (ST-GCN): A graph neural network that integrates spatiotemporal features. It constructs a topology graph with photovoltaic strings as nodes and electrical connections as edges. It extracts the temporal decay features of insulation parameters through time axis convolution and captures the synergistic aging effect of adjacent strings by combining spatial graph convolution, thereby driving the dynamic adjustment of the protection threshold.

[0162] Heterogeneous Algorithm Federation Engine: A multi-model fusion module that runs deep residual networks and Vision Transformers in parallel, integrates conflicting evidence from hot spot distribution and current and voltage anomaly analysis through DS evidence theory, and outputs joint diagnostic results.

[0163] Improved A* Algorithm: A path search algorithm that, in the event of communication anomalies, constructs a fault propagation risk coefficient model based on historical temperature field and defect mapping data in the dynamic health state matrix, and generates inverter derating operation parameters and string isolation priority instructions.

[0164] Transfer learning unit: A module used to synchronize data between the real system and the digital twin model. It reduces the difference in feature distribution through domain adversarial training strategy, simulates the failure evolution path based on the Monte Carlo method, and generates preventive maintenance instructions that include component replacement priority and protection threshold adjustment parameters.

[0165] In its specific implementation, this invention achieves fault monitoring and protection control of photovoltaic modules through multi-source sensor data fusion and intelligent algorithm collaboration. The infrared thermal imaging sensor in the acquisition module acquires surface temperature field distribution data of the module at a sampling rate of 5 frames per second. An improved variational mode decomposition algorithm is used to perform mode decomposition on the original temperature data. Effective temperature feature modes are selected using a kurtosis-energy joint criterion, generating a denoised temperature gradient distribution map. The electroluminescence detection unit acquires electroluminescence images at a resolution of 50 μm. Combined with the electrical topology, an adaptive threshold segmentation algorithm is used to identify hidden crack defect regions, generating a defect space mapping map with labeled defect geometry. The current and voltage characteristic curve acquisition module compensates for line impedance errors through an adversarial generative network, controlling the compensation error within ±0.5% to generate high-precision electrical parameter data. Multi-source data is input into a bidirectional long short-term memory network for time-series alignment, establishing a spatiotemporal correlation model of temperature field changes, defect propagation, and electrical parameter decay, generating a dynamic health state matrix.

[0166] 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. It optimizes the frequency band division accuracy of wavelet packet decomposition by introducing a simulated annealing mechanism, with an iteration count set to 200 and a convergence accuracy of 1e-6. The fuzzy logic controller constructs a three-dimensional membership function based on temperature and humidity sensor data, dynamically adjusting the filtering bandwidth range to 10Hz-1kHz to suppress low-frequency leakage current fluctuations under high humidity conditions. A random forest regression model extracts the time-domain statistical characteristics and frequency-domain energy distribution characteristics of the filtered leakage current signal, and combines this with photovoltaic string voltage level parameters to generate dynamic insulation criteria. When the detected insulation resistance value is below 50MΩ, a three-level alarm signal is triggered, and the DC circuit breaker is linked to isolate the faulty circuit according to a 0.1s, 0.5s, and 1s gradient tripping strategy.

[0167] The control optimization module deploys a federated learning framework within a time-sensitive network architecture. Edge computing nodes update communication path weight parameters every 30 seconds. A dynamic time warping algorithm predicts the circuit breaker tripping time window, with the time window accuracy controlled within ±2ms. The model predictive control unit monitors the arc impedance of the arc-extinguishing chamber at a 1MHz sampling frequency. A rolling time-domain optimization strategy calculates the optimal control sequence for the arc-extinguishing chamber nozzle opening and the magnetic blow-out coil current, suppressing transient overvoltages to below 120% of the rated voltage. The deep Q-network algorithm constructs a reward function based on a historical fault handling case library, dynamically adjusting the parsing weights of overcurrent protection and islanding detection commands, reducing the response time to 50ms and achieving matching between protection actions and fault evolution rates.

[0168] The improved U-Net network for the power supply module employs a skip connection structure to perform pixel-level segmentation of infrared thermal images, achieving a hotspot positioning accuracy of ±5mm. Combined with string current imbalance parameters, it generates a thermal map of abnormal hotspot distribution. An attention-enhanced SVM model focuses on the 0.5V-0.8V open-circuit voltage range in the current-voltage characteristic curves, identifying power attenuation patterns caused by PID effects and improving classification accuracy. A spatiotemporal graph convolutional network receives aging trend data from the digital twin module, establishing an insulation degradation prediction model with a weekly time granularity. This drives the DC distribution cabinet's grounding protection threshold to dynamically adjust from an initial 100MΩ to 60MΩ, matching changes in ambient temperature and humidity.

[0169] The digital twin module synchronizes temperature field distribution data between the real and virtual systems weekly using transfer learning algorithms. It simulates fault propagation paths under different salt spray concentrations using the Monte Carlo method, generating a maintenance instruction set that includes component replacement priorities. A heterogeneous algorithm federated engine daily aggregates diagnostic results from a deep residual network and a Vision Transformer model, fusing multi-model confidence levels through DS evidence theory, improving fault type identification accuracy. An improved A* algorithm retrieves the health status data from the most recent 24 hours during communication interruptions, searches for the optimal protection path based on the fault propagation risk coefficient, and triggers the inverter to operate at 80% rated power, forming a dual-mode collaborative mechanism of online monitoring and offline fault tolerance. Preventative maintenance instructions are fed back to a spatiotemporal graph convolutional network to optimize the material aging coefficient in node attributes, forming a closed-loop optimization link between monitoring data and the prediction model, effectively suppressing system-level cascading risks.

[0170] The present invention solves the aforementioned problem through the following technical solution:

[0171] To address the misjudgment of insulation resistance caused by dynamic fluctuations in leakage current in humid or salt spray environments, an improved whale optimization algorithm is employed to dynamically adjust the scale factor and displacement parameters of the wavelet basis function, optimizing the frequency band division accuracy of wavelet packet decomposition. The algorithm balances global search and local exploitation capabilities through simulated annealing, separating high-frequency transient components and low-frequency environmental noise in the leakage current signal. A fuzzy logic controller, combined with temperature and humidity sensor data, dynamically corrects the filtering bandwidth threshold, suppressing low-frequency leakage current fluctuation interference under high humidity conditions and generating dynamic insulation criteria reflecting changes in the conductivity of the insulating material. A random forest regression model correlates the photovoltaic string voltage level with the filtered leakage current characteristics, constructing a voltage-leakage current nonlinear mapping relationship. This triggers a gradient tripping strategy, prioritizing the isolation of high-risk fault circuits based on voltage level, avoiding the risk of misjudgment under environmental disturbances using a single threshold criterion.

[0172] To address the communication latency and algorithm convergence issues in the collaborative control of intelligent circuit breakers and monitoring systems, a federated learning framework is deployed based on a pre-defined time-sensitive network. This framework aggregates the communication path weights of multi-source protection commands through edge computing nodes, optimizing data transmission priorities. A dynamic time warping algorithm matches the circuit breaker's mechanical response curve with the zero-crossing timing of the fault current, predicting the opening time window to meet arc energy suppression requirements. This is combined with a pre-charging circuit to absorb inductive load energy in stages. The model predictive control unit adjusts the current of the arc-extinguishing chamber's magnetic blow-out coil and the flow rate of the air-blowing medium in real time, continuously optimizing the arc-extinguishing chamber's impedance characteristics to suppress transient overvoltages and block arc reignition conditions, achieving spatiotemporal matching between protection actions and fault evolution.

[0173] Furthermore, a virtual photovoltaic array system is constructed using a digital twin module. Transfer learning algorithms synchronize the temperature field distribution and defect spatial mapping data of the real system to simulate the thermal aging of insulation materials and the hot spot diffusion path. A spatiotemporal graph convolutional network fuses hot spot temperature rise rate and aging trend prediction data to drive the DC distribution cabinet to dynamically adjust the overcurrent protection threshold. In the event of communication anomalies, the heterogeneous algorithm federated engine invokes an improved A* algorithm to generate inverter derating operation parameters based on locally cached data, forming a dual-mode collaborative mechanism of online monitoring and offline fault tolerance to reduce system-level cascading risks.

Claims

1. A photovoltaic module fault monitoring system, characterized in that, include: The module includes a data acquisition module, a dynamic power supply 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, which includes raw infrared thermal imaging data, electroluminescence defect data, and raw current-voltage characteristic curve data. The photovoltaic module data is time-series aligned using a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix. The power supply dynamic compensation module receives the leakage current signal from the dynamic health status matrix, dynamically adjusts the wavelet packet decomposition level 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 calculates the insulation resistance value by combining it with the preset photovoltaic string voltage level output by the acquisition module to generate a dynamic insulation criterion; when the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and the preset DC circuit breaker is linked to execute a gradient tripping strategy based on the preset photovoltaic string voltage level. The control optimization module optimizes the parsing priority of the tripping control command and the arc-extinguishing chamber impedance adjustment command generated by the alarm signal based on the communication path weight of the preset time-sensitive network. The circuit breaker tripping action time window corresponding to the tripping control command is matched with the arc-extinguishing chamber impedance change curve by a preset dynamic time warping algorithm, and the arc-extinguishing chamber impedance parameter is adjusted. The power supply module receives the infrared hot spot distribution thermal map and current and voltage curve anomaly analysis results generated by the acquisition module, performs feature-level fusion with the component aging trend prediction data output by the digital twin module, generates the power supply circuit insulation degradation assessment results, and drives the overcurrent protection threshold of the DC distribution cabinet to be dynamically adjusted. The digital twin module constructs a digital twin model of the photovoltaic array, including temperature field distribution and defect spatial mapping. Through a preset transfer learning algorithm, the insulation degradation assessment results output by the power supply module are synchronized to a preset virtual system, generating preventive maintenance instructions and feeding them back to the spatiotemporal graph convolutional network model of the power supply module.

2. The photovoltaic module fault monitoring system as described in claim 1, characterized in that, The digital twin module is also used for: Receive the dynamic health status matrix generated by the acquisition module, and construct a digital twin model of the photovoltaic array including temperature field distribution data, defect space mapping map and compensated current and voltage characteristic curve data; The temperature field distribution data and defect space mapping map in the dynamic health status matrix are synchronized to a preset virtual system through a preset transfer learning algorithm. The preset virtual system is used to simulate the fault propagation path and generate preventive maintenance instructions. The preventive maintenance command is input into the spatiotemporal graph convolutional network model in the power supply module to optimize the confidence level of the component aging trend prediction results generated by the power supply module.

3. The photovoltaic module fault monitoring system as described in claim 1, characterized in that, The acquisition module includes: An infrared thermal imaging sensor is used to collect temperature field distribution data of photovoltaic modules and generate denoised temperature field data by eliminating environmental noise interference through a preset improved variational mode decomposition algorithm. An electroluminescence detection unit is used to process the electroluminescence image of a photovoltaic module based on a preset adaptive threshold segmentation algorithm, and generate a defect space mapping map by combining the electrical topology of the photovoltaic module. The current and voltage characteristic curve acquisition module is used to compensate for line impedance errors by using an adversarial generation network to generate compensated current and voltage characteristic curve data. The denoised temperature field data, defect space mapping map, and compensated current and voltage characteristic curve data are input into a preset bidirectional long short-term memory network. Based on the operating timestamp of the photovoltaic module, multi-source data time sequence alignment is performed to generate a dynamic health status matrix.

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

5. The photovoltaic module fault monitoring system as described in claim 1, characterized in that, The control optimization module includes: Edge computing nodes are used to optimize the communication path selection strategy of time-sensitive networks using a federated learning framework, and generate optimized communication path weights. The deep Q-network algorithm unit is used to dynamically adjust the analytical weights of the tripping control command and the arc-extinguishing chamber impedance adjustment command based on the fault feature map generated by the abnormal hot spot distribution heat map and the abnormal current and voltage curve analysis results output by the power supply module. The dynamic time warping algorithm unit is used to predict the optimal time window for the circuit breaker tripping action based on the optimized communication path weight, and to absorb 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, thereby suppressing the risk of transient overvoltage and arc reignition.

6. The photovoltaic module fault monitoring system as described in claim 1, characterized in that, The power supply module includes: An improved U-Net network is used to perform pixel-level segmentation of the infrared thermal image generated by the acquisition module to generate a thermal map of abnormal hot spot distribution. An attention-enhanced SVM model is used to analyze abnormal fluctuation characteristics in the compensated current and voltage characteristic curve data generated by the acquisition module and to identify power attenuation caused by PID effect. A spatiotemporal graph convolutional network is used to fuse the component aging trend prediction data output by the digital twin module with the abnormal hot spot distribution heat map in terms of spatiotemporal features, model the insulation degradation inflection point, and drive the dynamic adjustment of the overcurrent protection threshold of the DC distribution cabinet.

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

8. The photovoltaic module fault monitoring system as described in claim 1, characterized in that, The digital twin module includes: The heterogeneous algorithm federated engine is used to run deep residual networks and Vision Transformer models in parallel. It integrates conflicting evidence from the abnormal hot spot distribution heat map and current and voltage anomaly analysis results generated by the power supply module through Dempster-Shafer evidence theory, and outputs multi-model joint diagnostic results. Improvement A The algorithm unit is used to switch to the local cached data of the acquisition module when communication is abnormal, and to execute fast protection logic based on the temperature field distribution data and defect space mapping in the dynamic health status matrix, triggering the inverter derating operation mode.

9. The photovoltaic module fault monitoring system as described in claim 8, characterized in that, The digital twin module also includes: The transfer learning unit is used to transfer the temperature field distribution data and defect space mapping map 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 defect space mapping map, the unit simulates the thermal aging of the insulation material and the defect propagation path, and generates preventive maintenance instructions for the power supply module.

10. A photovoltaic module fault monitoring method, applied to the photovoltaic module fault monitoring system as described in any one of claims 1 to 9, characterized in that, Includes the following steps: Data from photovoltaic modules is collected, including raw infrared thermal imaging data, electroluminescence defect data, and raw current-voltage characteristic curve data. The photovoltaic module data is time-series aligned using a preset multi-scale feature alignment algorithm to generate a dynamic health status matrix. The system receives leakage current signals and dynamically adjusts the wavelet packet decomposition level and threshold parameters of the wavelet basis function of the leakage current signal through a preset improved whale optimization algorithm. It generates a filtered leakage current signal and calculates the insulation resistance value based on the preset photovoltaic string voltage level to generate a dynamic insulation criterion. When the dynamic insulation criterion is lower than the preset threshold, an alarm signal is triggered, and the preset DC circuit breaker is linked to execute a gradient tripping strategy based on the preset photovoltaic string voltage level. Upon receiving an alarm signal, the parsing priority of the tripping control command and the arc-extinguishing chamber impedance adjustment command generated by the alarm signal is optimized based on the communication path weight of the preset time-sensitive network. The circuit breaker tripping action time window corresponding to the tripping control command is matched with the arc-extinguishing chamber impedance change curve through a preset dynamic time warping algorithm, and the arc-extinguishing chamber impedance parameter is adjusted. The system receives infrared hotspot distribution thermal maps, current and voltage curve anomaly analysis results, and component aging trend prediction data. It performs feature-level fusion of the infrared hotspot distribution thermal maps, current and voltage curve anomaly analysis results, and component aging trend prediction data to generate power supply circuit insulation degradation assessment results. The power supply circuit insulation degradation assessment results are used to drive the dynamic adjustment of the overcurrent protection threshold of the preset DC distribution cabinet. The system receives the dynamic health status matrix, constructs a digital twin model of the photovoltaic array including temperature field distribution and defect spatial mapping, synchronizes the insulation degradation assessment results output by the power supply module to the preset virtual system through a preset transfer learning algorithm, generates preventive maintenance instructions and feeds them back to the spatiotemporal graph convolutional network model.

Images