Photovoltaic module fault diagnosis method based on machine learning

Abstract

The invention relates to the technical field of intelligent photovoltaic power generation operation and maintenance, in particular to a photovoltaic module fault diagnosis method based on machine learning, and the method comprises the following steps: S1, collecting historical multi-dimensional operation data of a photovoltaic module; s2, constructing a tensor decomposition model comprising a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix and a core factor tensor; s3, obtaining a reference core factor tensor and a reference physical quantity factor matrix in a normal state by using historical data of the photovoltaic module in the normal operation state; and S4, performing weighted fusion on the reconstruction error and the factor matrix deviation to obtain a local sensitivity metric value. According to the method, the three-order tensor is constructed, the multi-kernel tensor decomposition technology is introduced, smooth evolution of the time factor is forcibly constrained through the Gaussian kernel function, and the technical problem that environment interference and real faults are difficult to distinguish in the complex dynamic environment through a traditional method is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent photovoltaic power generation operation and maintenance technology, and in particular to a photovoltaic module fault diagnosis method based on machine learning. Background Technology

[0002] With the widespread application of photovoltaic power generation technology, the operation and maintenance management of photovoltaic power plants is crucial for ensuring power generation efficiency and system safety. In the field of photovoltaic module fault diagnosis, timely and accurate identification of internal electrical faults is a core technical requirement. Existing fault diagnosis methods mainly rely on data acquisition and analysis technologies, using sensors deployed in photovoltaic arrays to acquire operational data such as voltage, current, temperature, and irradiance, and then interpreting the data using preset rules or machine learning algorithms.

[0003] However, existing photovoltaic (PV) fault diagnosis technologies have significant limitations in data processing and feature extraction. Traditional methods typically involve simply concatenating data from PV modules at different times, physical dimensions, and sampling points into a one-dimensional vector or two-dimensional matrix for processing. This dimensionality reduction or planarization destroys the original high-order spatial structure of the data, resulting in the loss of multidimensional topological features inherent in the PV data. Specifically, the operating state of a PV module is determined by the evolution trend of the time dimension, the coupling relationship of the physical quantity dimension, and the distribution characteristics of the sampling space dimension. Existing technologies often separate these closely related dimensions, analyzing the time series of a single physical quantity independently, or, although using multiple physical quantities as inputs, failing to effectively model the inherent dynamic tensor structure between them.

[0004] This lack of structural information directly leads to insufficient robustness of the system in complex dynamic environments. In actual operation, photovoltaic modules are highly susceptible to interference from environmental factors such as cloud cover and shadow movement. For example, when clouds pass by, solar irradiance fluctuates drastically, causing a simultaneous and significant drop in the output current of the photovoltaic modules. Because existing technologies struggle to effectively distinguish between parameter changes caused by environmental factors and parameter anomalies caused by internal module faults at the feature level, they often rely solely on the magnitude of a single parameter value for judgment. When the current value drops to a low level due to environmental shading, it is highly likely to trigger threshold-based alarm mechanisms or be misclassified by machine learning models as an open circuit or abnormally low current fault. This high false alarm rate, resulting from the inability to decouple environmental interference, not only increases the workload of maintenance personnel in ineffective inspections but also reduces the reliability and applicability of the fault diagnosis system. Summary of the Invention

[0005] To overcome the above shortcomings, this invention provides a machine learning-based photovoltaic module fault diagnosis method, aiming to solve the problem of high false alarm rate in existing technologies under dynamic weather conditions.

[0006] In a first aspect, the present invention provides the following technical solution: a photovoltaic module fault diagnosis method based on machine learning, comprising the following steps: S1. Collect historical multidimensional operating data of photovoltaic modules, and organize the operating data according to time windows to construct a third-order tensor containing time window dimension, physical quantity dimension and sampling point dimension; S2. Construct a tensor decomposition model that includes a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix, and a core factor tensor. By introducing a multi-kernel function to constrain the time factor matrix and the physical quantity factor matrix, and by minimizing the objective function that includes a reconstruction error term and a regularization constraint term, the core factor tensor and the low-rank factor matrices of each dimension are obtained. S3. Using historical data of photovoltaic modules under normal operating conditions, construct a tensor according to the method described in step S1, and decompose it according to the method described in step S2 to obtain the reference core factor tensor and reference physical quantity factor matrix under normal conditions. S4. Calculate the reconstruction error of the tensor to be diagnosed and the factor matrix deviation between the physical quantity factor matrix decomposed from the tensor and the reference physical quantity factor matrix. Perform weighted fusion on the reconstruction error and the factor matrix deviation to obtain the local sensitivity metric. S5. Compare the local sensitivity metric value with a preset fault threshold. If the local sensitivity metric value is greater than or equal to the fault threshold, it is determined that the photovoltaic module has a fault.

[0007] Preferably, in step S1, the dimensions of the third-order tensor are defined as follows: The time window dimension corresponds to multiple continuously collected time segments, which are used to characterize the evolution of the photovoltaic module's operating status over time; The physical quantities correspond to the DC voltage, DC current, irradiance, and backsheet temperature of the photovoltaic module, and are used to characterize the coupling relationship of multiple physical fields. The sampling point dimension corresponds to the number of data points collected within each time window, and is used to characterize the instantaneous fluctuation characteristics within a single time window.

[0008] Preferably, the approximate decomposition process of the third-order tensor by the tensor decomposition model in step S2 includes: The third-order tensor is represented as the sum of multiple tensors of rank one; Each rank-one tensor is formed by the outer product of the column vectors of the time factor matrix, the column vectors of the physical quantity factor matrix, and the column vectors of the sampling point factor matrix, and is weighted by the corresponding elements of the core factor tensor.

[0009] Preferably, the objective function in step S2 is composed of: Reconstruction error term: used to characterize the Euclidean distance between the original third-order tensor and the approximate tensor reconstructed using the core factor tensor and each factor matrix; Regularization constraint terms: include time constraint terms for the time factor matrix and physical quantity constraint terms for the physical quantity factor matrix; the time constraint terms are calculated based on the trace of the product of the time factor matrix and the inverse of the time kernel matrix; the physical quantity constraint terms are calculated based on the trace of the product of the physical quantity factor matrix and the inverse of the physical kernel matrix.

[0010] Preferably, the selection rule for the multi-core function is as follows: The time kernel matrix is ​​constructed using a Gaussian kernel function to constrain the smooth evolution characteristics of the time factor matrix in the time dimension and prevent non-fault-related time mutation interference. The physical quantity kernel matrix is ​​constructed using a polynomial kernel function to constrain the nonlinear coupling characteristics of the physical quantity factor matrix between different physical parameters, so as to extract the correlation features between physical quantities.

[0011] Preferably, the local sensitivity metric value is calculated in step S4 as follows: The reconstruction error and the deviation of the factor matrix are multiplied by their respective weighting coefficients and then summed; the weighting coefficients are used to balance the influence of the overall reconstruction accuracy and the deviation of local physical quantity characteristics on the fault decision.

[0012] Preferably, the reconstruction error is calculated by: calculating the Frobenius norm between the current third-order tensor to be diagnosed and the tensor reconstructed using the low-rank factor matrices of each dimension and the core factor tensor obtained after decomposing the tensor to be diagnosed.

[0013] Preferably, the factor matrix deviation is calculated as follows: the distance between the physical quantity factor matrix obtained by the current tensor decomposition of the diagnostic tensor and the reference physical quantity factor matrix is ​​calculated; the distance adopts a distance metric based on local weighting to quantify the degree to which the current physical quantity coupling relationship deviates from the normal state reference benchmark.

[0014] Preferably, the fault determination step in step S5 further includes: when it is determined that there is a fault in the photovoltaic module, comparing the magnitude of the deviation values ​​corresponding to each physical quantity dimension in the physical quantity factor matrix, identifying the physical quantity dimension that causes the largest deviation in the factor matrix as the fault source, and outputting the fault source as the fault tracing result.

[0015] Secondly, the present invention provides the following technical solution: a photovoltaic module fault diagnosis system based on machine learning, the system comprising: The multidimensional data construction module is used to collect historical multidimensional operating data of photovoltaic modules and organize the operating data according to time windows to construct a third-order tensor containing time window dimension, physical quantity dimension and sampling point dimension; The feature extraction module is used to construct a tensor decomposition model that includes a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix, and a core factor tensor. By introducing a multi-kernel function to constrain the time factor matrix and the physical quantity factor matrix, the core factor tensor and the low-rank factor matrix of each dimension are obtained by minimizing the objective function that includes the reconstruction error term and the regularization constraint term. The reference model storage module is used to store the reference core factor tensor and reference physical quantity factor matrix of the photovoltaic module under normal operating conditions. An anomaly detection module is used to calculate the reconstruction error of the tensor to be diagnosed and the factor matrix deviation between the physical quantity factor matrix decomposed from the tensor and the reference physical quantity factor matrix based on the reference model, and to perform weighted fusion of the reconstruction error and the factor matrix deviation to obtain a local sensitivity metric. The fault diagnosis module is used to compare the local sensitivity metric value with a preset fault threshold. If the local sensitivity metric value is greater than or equal to the fault threshold, the photovoltaic module is determined to have a fault.

[0016] The present invention has the following beneficial effects: 1. In this invention, a third-order tensor is constructed and a multi-kernel tensor decomposition technique is introduced. Gaussian kernel functions are used to force the smooth evolution of the time factor, while polynomial kernel functions are used to constrain the nonlinear coupling of physical quantity factors. This design enables the model to effectively distinguish between normal environmental fluctuations caused by cloud cover or sudden changes in irradiance and essential state changes caused by degradation of the internal electrical performance of components. This solves the technical problem of traditional methods being unable to distinguish between environmental disturbances and real faults in complex dynamic environments.

[0017] 2. This invention employs an unsupervised detection mechanism based on local sensitivity measurement. It calculates the local distance deviation between the physical quantity factor matrix of the current data and the normal reference model, and then weights and fuses this deviation with the global reconstruction error. This method eliminates the need to collect large amounts of fault label data for model training; it only requires establishing a benchmark based on historical data from normal operation. This enables keen detection of potential faults, overcoming the limitation of newly built power plants or specific components lacking fault samples, which prevents the deployment of diagnostic systems.

[0018] 3. In this invention, a direct mapping relationship between fault determination and specific physical parameters is established using the low-rank physical quantity factor matrix obtained by tensor decomposition. When the system determines a fault, by comparing the magnitude of the factor deviations in each physical quantity dimension, the specific physical quantity causing the anomaly can be directly located. Compared with the black-box model that only outputs binary decision results, this provides maintenance personnel with a clear direction for fault investigation and reduces the time cost of on-site maintenance. Attached Figure Description

[0019] Figure 1 This is a flowchart of a photovoltaic module fault diagnosis method based on machine learning proposed in this invention. Figure 2 This is a framework diagram of a photovoltaic module fault diagnosis system based on machine learning proposed in this invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 In a first embodiment of the present invention, the present invention provides a photovoltaic module fault diagnosis method based on machine learning, such as... Figure 1 As shown, it includes the following steps: S1. Collect historical multidimensional operation data of photovoltaic modules, and organize the operation data according to time windows to construct a third-order tensor containing time window dimension, physical quantity dimension and sampling point dimension; Furthermore, in step S1, the dimensions of the third-order tensor are defined as follows: The time window dimension corresponds to multiple time segments collected continuously, and is used to characterize the evolution of the photovoltaic module's operating status over time; The physical quantity dimensions correspond to the DC voltage, DC current, irradiance, and backsheet temperature of the photovoltaic module, and are used to characterize the coupling relationship of multiple physical fields. The sampling point dimension corresponds to the number of data points collected within each time window, and is used to characterize the instantaneous fluctuation characteristics within a single time window.

[0022] Specifically, this embodiment of the invention provides a photovoltaic module fault diagnosis method based on machine learning. When the method is executed, the first step S1 is to collect historical multidimensional operating data of the photovoltaic module and organize the operating data according to the time window to construct a third-order tensor containing the time window dimension, physical quantity dimension and sampling point dimension.

[0023] In the specific implementation of this step, the system first connects to the sensor network at the photovoltaic array site via a hardware interface. The sensor network includes a DC voltage sensor connected in parallel to the output of the photovoltaic module, a Hall current sensor connected in series in the module circuit, a total radiation meter mounted on the module support plane, and a thermocouple or resistance temperature sensor attached to the center of the module backsheet. Each sensor synchronously acquires analog signals according to a preset sampling frequency, converts them into digital signals via an analog-to-digital converter, and transmits them to the data processing terminal via an industrial fieldbus. The raw data received by the data processing terminal includes time-series sequences of four physical quantities: DC voltage, DC current, solar irradiance, and module backsheet temperature.

[0024] Because different physical quantities have significantly different dimensions and numerical ranges, directly constructing tensors can lead to instability in numerical calculations. Therefore, the system performs Z-score normalization on the historical data sequences of each type of physical quantity.

[0025] After data preprocessing, the system segments the continuous time-series data according to a preset time window length. The time window length is set based on the thermal inertia of the photovoltaic modules and the characteristics of environmental changes, for example, 15 minutes per window. The system allocates a three-dimensional array space in memory to construct a third-order tensor, denoted as . The third-order tensor It contains three orthogonal dimensions: The first dimension is the time window dimension, denoted as... This dimension corresponds to multiple time segment sequences collected consecutively, index. The value range is 1 to , indicating the first A time window. This dimension is used to characterize the long-term evolution of the photovoltaic module's operating status over days, weeks, or months, reflecting the slow-changing characteristics of module performance.

[0026] The second dimension is the physical quantity dimension, denoted as... This dimension corresponds to the four key operating parameters of photovoltaic modules: DC voltage, DC current, irradiance, and module backsheet temperature. (Index) The value range is 1 to , indicating the first This is a physical quantity. This dimension is used to characterize the multiphysics coupling relationship between the internal electrical parameters of the photovoltaic system and the external environmental parameters.

[0027] The third dimension is the sampling point dimension, denoted as... This dimension corresponds to the number of data points collected within each time window. If the time window length is 15 minutes and the sampling frequency is 1Hz, then... Equals 900. Index The value range is 1 to , indicating the first time within the time window Each sampling point. This dimension is used to characterize the instantaneous fluctuation characteristics within a single time window, reflecting the dynamic response of the component over a short period of time.

[0028] The third-order tensor structure constructed using the above method yields significant technical advantages compared to the traditional two-dimensional matrix storage method. First, this structure preserves the multidimensional topological characteristics of the data, avoiding the temporal structure disruption caused by forcibly flattening data from different time windows.

[0029] Secondly, by explicitly defining the physical quantity dimension, the irradiance and temperature of environmental factors are placed in the same tensor slice as the voltage and current of electrical output. This forces the subsequent calculation model to consider the nonlinear mapping relationship between them when extracting features, thus providing a data basis for distinguishing between environmental interference and actual faults.

[0030] Finally, the introduction of sampling point dimension preserves high-frequency sampling information, ensuring that transient anomalies occurring in a short period of time are not smoothly lost due to averaging operations, thus improving the ability to capture weak fault characteristics.

[0031] S2. Construct a tensor decomposition model that includes a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix, and a core factor tensor. By introducing a multi-kernel function to constrain the time factor matrix and the physical quantity factor matrix, and by minimizing the objective function that includes reconstruction error terms and regularization constraint terms, the core factor tensor and the low-rank factor matrices of each dimension are obtained. Furthermore, the objective function in step S2 is composed of: Reconstruction error term: used to characterize the Euclidean distance between the original third-order tensor and the approximate tensor reconstructed using the core factor tensor and each factor matrix; Regularization constraints include time constraints on the time factor matrix and physical quantity constraints on the physical quantity factor matrix. The time constraint is calculated based on the trace of the product of the time factor matrix and the inverse of the time kernel matrix; the physical quantity constraint is calculated based on the trace of the product of the physical quantity factor matrix and the inverse of the physical kernel matrix. The selection rule for the multi-kernel function is as follows: The time kernel matrix is ​​constructed using a Gaussian kernel function to constrain the smooth evolution characteristics of the time factor matrix in the time dimension and prevent non-fault-related time mutations from interfering with the time. The physical quantity kernel matrix is ​​constructed using a polynomial kernel function to constrain the nonlinear coupling characteristics of the physical quantity factor matrix between different physical parameters, so as to extract the correlation features between physical quantities.

[0032] Specifically, after constructing the third-order tensor, the system then executes step S2: constructing a tensor decomposition model containing a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix, and a core factor tensor; constraining the time factor matrix and the physical quantity factor matrix by introducing a multi-kernel function; and solving for the core factor tensor and the low-rank factor matrices of each dimension by minimizing the objective function containing the reconstruction error term and the regularization constraint term.

[0033] In implementing this step, the system uses the Tucker decomposition model as its basic architecture. The system initializes three factor matrices and one core factor tensor in computational memory. The time factor matrix is ​​denoted as... The number of rows corresponds to the time window dimension. The number of columns corresponds to the rank of the time dimension. The physical quantity factor matrix is ​​denoted as The number of rows corresponds to the dimension of the physical quantity. The number of columns corresponds to the rank of the physical quantity dimension. The sampling point factor matrix is ​​denoted as The number of rows corresponds to the dimension of the sampling points. The number of columns corresponds to the rank of the dimension of the sampling points. The core factor tensor is denoted as .

[0034] rank These are preset hyperparameters, usually set to integers less than the corresponding original dimension values, in order to achieve dimensionality reduction and feature compression of the data.

[0035] This tensor decomposition model aims to reconstruct the original third-order tensor through the modular multiplication of the core factor tensor with each factor matrix. An approximate estimate of the reconstructed tensor. The mathematical expression is as follows:

[0036]

[0037] Among them, symbols Representing the modulus of tensors and matrices The product, where n takes the values ​​1, 2, or 3.

[0038] Specifically, Representation matrix With tensor In the product operation on the first dimension, to eliminate environmental noise interference and accurately decouple physical quantity relationships during feature extraction, the system introduces a multi-kernel learning mechanism during the decomposition process. The system first constructs a time kernel matrix. and physical quantity kernel matrix .

[0039] For time kernel matrix The system is constructed using a Gaussian kernel function, which has smoothing properties and can constrain the time factor matrix. The evolutionary trend along the time axis ensures its continuity and smoothness. The [number]th [unit] in the time kernel matrix... Line number Column elements The calculation formula is:

[0040] in, and They represent the first The and the first The time index or feature vector of a time window The bandwidth parameter of the local frequency kernel controls the degree of smoothing. Through this kernel function, the model forces time-adjacent factors to have similar values, thereby suppressing non-fault-related local frequency temporal abrupt changes caused by cloud cover or short-term shading.

[0041] For physical quantity kernel matrix The system is constructed using a polynomial kernel. The polynomial kernel maps low-dimensional physical quantity features to a high-dimensional space, thereby capturing the nonlinear coupling relationships between different physical quantities. The th element in the physical quantity kernel matrix... Line number Column elements The calculation formula is:

[0042] in, and They represent the first The and the first The feature vector of a physical quantity in historical data, with the superscript T indicating the transpose operation of the vector. The degree is the order of the polynomial, for example, 2 or 3. This is a constant term. Through this kernel function, the model can extract the complex nonlinear correlation characteristics between DC voltage, current, irradiance, and temperature.

[0043] Based on the above definition, the system constructs the objective function. The objective function consists of a reconstruction error term and a regularization constraint term. The system solves for the model parameters by minimizing this objective function. The quantization expression is as follows:

[0044] The definitions of each term in the above formula are as follows: The table shows the reconstruction error term, and the original tensor is calculated using the Frobenius norm. With Reconstructed Tensor The Euclidean distance between them is used to ensure the model fitting accuracy.

[0045] The time constraint term is represented by the following formula: The spectral norm of the time factor is constrained by the inverse of the time kernel matrix.

[0046] This represents the physical quantity constraint term, and the calculation formula is: The spectral norm of physical quantity factors is constrained by the inverse matrix of the physical kernel matrix.

[0047] Here are the corresponding regularization weight coefficients, where Let represent the objective function of the tensor decomposition model, which measures the weighted sum of model reconstruction error and physical constraint penalty terms. The optimization objective of the algorithm is to make minimize.

[0048] The system uses alternating least squares or gradient descent to iteratively solve the objective function. In each iteration, other variables are fixed, and only one variable is updated until the objective function is solved. The convergence of the core factor tensor ultimately outputs the core factor tensor. and low-rank factors .

[0049] The aforementioned tensor decomposition technique based on multi-kernel constraints yields significant technical benefits. First, through high-dimensional tensor decomposition, the model effectively extracts the latent high-order features of photovoltaic data, preserving the original structural information of the data. Second, by applying smoothing constraints to the time factor using a Gaussian kernel function, the model can automatically filter out data noise caused by fluctuations in environmental irradiance, preventing normal weather changes from being misjudged as module failures. Finally, by applying nonlinear constraints to the physical quantity factors using a polynomial kernel function, the model can learn the complex mapping relationship between voltage, current, temperature, and irradiance under normal operating conditions. This allows for the accurate identification of internal electrical faults within the module by detecting disruptions to this mapping relationship.

[0050] S3. Using historical data of photovoltaic modules under normal operating conditions, construct tensors according to the method in step S1, and decompose them according to the method in step S2 to obtain the reference core factor tensor and reference physical quantity factor matrix under normal conditions. Furthermore, after establishing the tensor decomposition model and solution algorithm, the system then executes step S3: using historical data of photovoltaic modules under normal operating conditions, a tensor is constructed according to the method in step S1, and decomposed according to the method in step S2 to obtain the reference core factor tensor and reference physical quantity factor matrix under normal conditions.

[0051] In implementing this step, the system first performs a data filtering operation from the historical database. Based on the photovoltaic power plant's operation and maintenance logs and meteorological records, the system selects a continuous time series data segment marked as operating normally under clear weather conditions as a reference dataset. This filtering process automatically removes periods containing equipment fault alarms, downtime maintenance records, and extreme weather events to ensure that the selected data accurately represents the standard operating characteristics of the photovoltaic modules under design conditions.

[0052] For the selected reference dataset, the system strictly executes the data preprocessing workflow defined in step S1, including Z-score normalization and time window partitioning, to construct a data structure with dimensions of [missing information]. The normal state third-order tensor is denoted as Subsequently, the system uses the multi-core constrained tensor decomposition model constructed in step S2 to train the tensor offline. Through iterative calculation, the normal-state third-order tensor is decomposed into a combination of the core tensor and factor matrices of each dimension. Its simplified mathematical reconstruction relation is expressed as:

[0053] In the above relation, The tensor representing the normal state of the input; The reference core factor tensor obtained from the solution represents the energy distribution weights among the features of each dimension under normal conditions. , , The reference factor matrices represent the time dimension, physical quantity dimension, and sampling point dimension, respectively.

[0054] After the decomposition converges, the system performs key parameter extraction and storage operations. From all the obtained decomposition parameters, the system specifically extracts the reference core factor tensor. and reference physical quantity factor matrix These two parameters are defined as the system's fault detection baseline model and are permanently stored in the controller's non-volatile memory or a dedicated configuration table in the database for subsequent online monitoring. Other parameters, such as the time factor matrix, only contain historical time information and do not have reference value for subsequent diagnosis; therefore, the system may choose not to store them to save space.

[0055] The aforementioned reference model construction scheme yields significant technical benefits. Firstly, this step establishes a quantitative and objective health benchmark. This is achieved by storing only the reference core factor tensors. and reference physical quantity factor matrix Instead of storing massive amounts of raw historical data, this significantly reduces the storage overhead and data query latency of the online monitoring system, enabling the algorithm to run efficiently on edge computing devices. Secondly, it references the physical quantity factor matrix. The mathematical model solidifies the multi-physics coupling laws of photovoltaic modules under fault-free conditions, namely the normal projection relationship of voltage, current, temperature, and irradiance in the characteristic space. This allows subsequent fault diagnosis to no longer rely on simple numerical threshold judgments, but rather on the characteristic deviation based on physical laws, thereby effectively reducing the false alarm rate caused by fluctuations in environmental irradiance and significantly improving the system's accuracy in identifying early, minor faults.

[0056] S4. Calculate the reconstruction error of the tensor to be diagnosed and the factor matrix deviation between the physical quantity factor matrix decomposed from the tensor and the reference physical quantity factor matrix. Weight the reconstruction error and the factor matrix deviation to obtain the local sensitivity measure. Furthermore, the local sensitivity metric in step S4 is calculated as follows: The reconstruction error and factor matrix deviation are multiplied by their respective weighting coefficients and then summed. These weighting coefficients balance the influence of overall reconstruction accuracy and local physical quantity characteristic deviations on fault diagnosis. The reconstruction error is calculated by determining the Frobenius norm between the current third-order tensor to be diagnosed and the tensor reconstructed using the low-rank factor matrices of each dimension and the core factor tensor obtained after decomposing the tensor. The factor matrix deviation is calculated by determining the distance between the physical quantity factor matrix obtained from the decomposition of the current tensor to be diagnosed and the reference physical quantity factor matrix. A locally weighted distance metric is used to quantify the degree to which the current physical quantity coupling relationship deviates from the normal state reference baseline.

[0057] Specifically, after the reference model is constructed, the system enters the real-time monitoring stage and executes step S4: calculate the reconstruction error of the tensor to be diagnosed and the factor matrix deviation between the physical quantity factor matrix decomposed from the tensor and the reference physical quantity factor matrix. The reconstruction error and the factor matrix deviation are weighted and fused to obtain the local sensitivity metric.

[0058] In the specific process of implementing this step, the system first acquires the real-time running data within the current time window, and constructs the third-order tensor to be diagnosed according to the method in step S1, denoted as... Next, the system uses the tensor decomposition algorithm from step S2 to decompose the tensor to be diagnosed. This is to refactor the change in the current state relative to the normal baseline.

[0059] In this decomposition, the system uses the reference core factor tensor stored in step S3 as a fixed constraint, and only iteratively updates the physical quantity factor matrix and other dimension matrices to obtain the physical quantity factor matrix in the current state, denoted as . Simultaneously, the system reconstructs an approximate tensor using the parameters obtained from the decomposition, denoted as... .

[0060] Subsequently, the system calculates the reconstruction error, denoted as . This indicator characterizes whether the current data conforms to the expected low-rank structure in its overall distribution, reflecting the macroscopic consistency of the data. It is calculated by solving for the norm of the difference between the original tensor to be diagnosed and the reconstructed tensor. To simplify the calculation and unify the dimensions, the formula is simplified as follows:

[0061] In the above formula, the symbol This represents the Frobenius norm operation, which calculates the Euclidean distance between all elements in the tensor difference. This value increases significantly when a severe fault in the photovoltaic module causes overall distortion of the signal waveform.

[0062] Simultaneously, the system calculates the bias of the factor matrix, denoted as... This metric quantifies the degree of deviation of the coupling relationship between current physical quantities from a normal reference base. The system reads the reference physical quantity factor matrix. And calculate its relationship with the current physical quantity factor matrix. The distance between them. To improve sensitivity to anomalies in specific physical parameters, the system uses a locally weighted approach for calculation, expressed by the formula:

[0063] In the above formula, This represents a preset local weighting vector or matrix. Through this weighting term, the system can assign different weights to deviations in different dimensions of physical quantities such as voltage, current, or temperature, thereby quantifying the degree to which the current physical quantity coupling relationship deviates from the normal state.

[0064] Finally, the system performs a linear weighted fusion of the two indicators to generate the final local sensitivity metric, denoted as . This calculation process aims to balance the impact of overall reconstruction accuracy and the deviation of local physical quantity characteristics on fault determination. The calculation formula is as follows:

[0065] in, and These are the normalized weighting coefficients, used to adjust the proportion of the two sub-indicators in the final measurement.

[0066] The above-mentioned dual-index fusion calculation scheme can produce significant technical effects. Relying solely on reconstruction error often makes it difficult to distinguish between high-frequency noise in the sensor itself and actual component failures, while relying solely on factor bias may result in a sluggish response to sudden system-wide disturbances.

[0067] This step effectively combines information from two dimensions—data distribution consistency and physical coupling consistency—by introducing a local sensitivity metric. This fusion mechanism enables the system to maintain its ability to perceive overall signal waveform distortion while also being able to distinguish minor disruptions to the nonlinear coupling relationships between voltage, current, and temperature. This significantly improves the robustness of fault diagnosis and ensures diagnostic accuracy in complex operating environments.

[0068] S5. Compare the local sensitivity metric value with the preset fault threshold. If the local sensitivity metric value is greater than or equal to the fault threshold, the photovoltaic module is determined to be faulty.

[0069] Furthermore, the fault determination step in step S5 also includes: when it is determined that there is a fault in the photovoltaic module, comparing the magnitude of the deviation values ​​corresponding to each physical quantity dimension in the physical quantity factor matrix, identifying the physical quantity dimension that causes the largest deviation in the factor matrix as the fault source, and outputting the fault source as the fault tracing result.

[0070] Specifically, after calculating the local sensitivity metric, the system executes step S5: comparing the local sensitivity metric with a preset fault threshold. If the local sensitivity metric is greater than or equal to the fault threshold, the photovoltaic module is determined to be faulty.

[0071] In the specific process of implementing this step, the system first retrieves the preset fault threshold from the configuration storage area, denoted as... The threshold is typically set based on the statistical analysis results of the normal state dataset in step S3. For example, the system calculates the probability distribution of the local sensitivity metric values ​​for all samples under normal conditions, and selects the 99th quantile or the mean plus three standard deviations of this distribution as the judgment boundary. The system then uses the local sensitivity metric values ​​calculated in the current time window... With threshold Perform numerical comparisons.

[0072] if The system determines that the current photovoltaic module is operating normally, does not trigger an alarm, and continues data collection for the next time window. If The system determines that there is an abnormality or fault in the photovoltaic module, and then triggers the fault response mechanism, generates a fault alarm signal, and starts the subsequent fault tracing process.

[0073] Furthermore, when a fault is detected in the photovoltaic module, the system automatically executes a fault source identification procedure. The core logic of this procedure is to locate the source of the anomaly by utilizing the differences in the row vector characteristics of the physical quantity factor matrix. The system compares the current physical quantity factor matrix row by row. With reference physical quantity factor matrix Since each row of the matrix corresponds to a specific physical quantity, such as the first row corresponding to DC voltage, the second row corresponding to DC current, etc., the system calculates the difference magnitude between the vectors in each row. For the first row... Each physical quantity, its deviation contribution value The simplified formula for calculation is as follows:

[0074] in, and Let p represent the row vectors of the current physical quantity factor matrix and the reference physical quantity factor matrix, respectively. The norm operation of vectors is used to calculate the Euclidean distance between two vectors, thereby quantifying the magnitude of the difference. The system iterates through all physical quantity dimensions, calculates a set of bias contribution values, and finds the one with the largest value, denoted as . The system is based on The corresponding row index p is used to query the preset physical quantity mapping table, determine the name of the physical parameter that causes the largest deviation, and mark the physical parameter as the main source of failure.

[0075] Finally, the system outputs a diagnostic report through the human-computer interaction interface, which includes the time of the fault occurrence, the fault type determination, and suggestions for the fault source.

[0076] The aforementioned fault diagnosis and tracing technology achieves significant technical benefits. First, the threshold comparison method automates and enables real-time fault diagnosis, allowing for rapid response to abnormal situations without manual intervention and effectively preventing fault escalation. Second, the tracing mechanism based on factor matrix row bias endows the diagnostic model with strong interpretability. Traditional black-box models often only indicate the presence of a fault, but not the location of the problem. This solution can accurately pinpoint whether the fault lies in the voltage loop, current loop, or temperature sensor.

[0077] For example, if the system identifies the deviation in the DC current dimension as contributing the most, maintenance personnel can prioritize checking the combiner box fuses or component connectors; if the backplane temperature deviation is the largest, hot spot effects can be prioritized. This precise fault location capability significantly shortens maintenance troubleshooting time and reduces the maintenance cost of photovoltaic power plants.

[0078] Example 2: To more intuitively illustrate how this invention solves specific problems in the prior art, a detailed description is provided below using a typical dynamic weather application scenario.

[0079] In the actual operating environment of photovoltaic power plants, cloud cover often causes short-term and drastic fluctuations in ambient irradiance. For example, at a certain moment, due to cloud cover, solar irradiance may rapidly drop from 800 watts per square meter to 200 watts per square meter within a short period. Due to the physical laws of the photovoltaic effect, the output current of the photovoltaic modules will simultaneously decrease from 8 amperes to 2 amperes. In this scenario, if a traditional monitoring method based on a single fixed threshold is used, the system may generate false alarms by outputting fault judgments of abnormal current or open circuits because the collected real-time current value is less than the threshold.

[0080] To address the aforementioned problems, this invention provides a photovoltaic module fault diagnosis system based on machine learning, the architecture of which is as follows: Figure 2 As shown. The system includes: The multidimensional data construction module is used to collect historical multidimensional operating data of photovoltaic modules and organize the operating data according to time windows to construct a third-order tensor containing time window dimension, physical quantity dimension and sampling point dimension; The feature extraction module is used to construct a tensor decomposition model that includes a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix, and a core factor tensor. By introducing a multi-kernel function to constrain the time factor matrix and the physical quantity factor matrix, the core factor tensor and the low-rank factor matrices of each dimension are obtained by minimizing the objective function that includes the reconstruction error term and the regularization constraint term. The reference model storage module is used to store the reference core factor tensor and reference physical quantity factor matrix of the photovoltaic module under normal operating conditions. The anomaly detection module is used to calculate the reconstruction error of the tensor to be diagnosed and the factor matrix deviation between the physical quantity factor matrix decomposed from the tensor and the reference physical quantity factor matrix based on the reference model. The reconstruction error and the factor matrix deviation are weighted and fused to obtain the local sensitivity metric. The fault diagnosis module is used to compare the local sensitivity metric value with the preset fault threshold. If the local sensitivity metric value is greater than or equal to the fault threshold, the photovoltaic module is determined to be faulty.

[0081] Specifically, the data acquisition module is equipped with a sensor group and a signal conversion unit connected to the photovoltaic module site. The sensor group specifically includes a DC voltage sensor connected in parallel to the module output terminal, a Hall current sensor connected in series in the circuit, a total radiation meter installed on the bracket, and a thermocouple temperature sensor attached to the module back panel.

[0082] This module is used to synchronously acquire field analog signals at a preset frequency and convert the analog signals into a digital signal sequence containing four-dimensional features: DC voltage, DC current, irradiance, and backplane temperature. The data preprocessing module is communicatively connected to the data acquisition module. It receives the raw digital signal sequence and performs Z-score normalization on the raw data based on stored historical statistical parameters. Then, it constructs a third-order tensor from the normalized data sequence according to a preset time window length.

[0083] The tensor calculation module is connected to the data preprocessing module, and it integrates or connects to a non-volatile storage unit. This storage unit pre-stores a reference physical quantity factor matrix and a reference core factor tensor trained based on historical normal operation data. The tensor calculation module is configured to receive the third-order tensor to be diagnosed, and use the aforementioned stored reference physical quantity factor matrix as a fixed constraint to perform restricted tensor decomposition operations on the tensor to be diagnosed, thereby outputting the reconstructed tensor and the physical quantity factor matrix in the current state.

[0084] The fault diagnosis module is connected to the tensor calculation module to calculate the reconstruction error between the tensor to be diagnosed and the reconstructed tensor, as well as the factor matrix deviation between the current physical quantity factor matrix and the reference physical quantity factor matrix. The two indicators are weighted and fused according to preset weights to obtain the local sensitivity metric. Finally, the metric is compared with the preset fault threshold to output the fault judgment result.

[0085] When processing data using the system in this embodiment, the data acquisition module first collects voltage, current, irradiance, and temperature data for the specified time period. Subsequently, the data preprocessing module standardizes the collected data. Because both irradiance and current values ​​decrease significantly, their standardized values ​​exhibit markedly negative characteristics, resulting in the constructed diagnostic tensor displaying synchronously low values ​​in both the current and irradiance dimensions.

[0086] Next, the tensor calculation module calls the stored reference physical quantity factor matrix to decompose the tensor to be diagnosed. Since the reference physical quantity factor matrix is ​​obtained by learning from normal historical data, its internal eigenvectors have solidified the physical coupling law that irradiance and output current are positively correlated.

[0087] In the current cloud-covered scenario, the synchronous decrease in current and irradiance in the tensor to be diagnosed is consistent with the physical laws described in the reference physical quantity factor matrix. Therefore, when performing fitting operations, the tensor calculation module can calculate a set of coefficients with extremely high matching degree, making the approximate tensor reconstructed based on the reference parameters highly consistent with the original tensor to be diagnosed.

[0088] Finally, the fault diagnosis module calculates various indicators. Since the data characteristics conform to the physical laws defined by the reference model, the difference between the tensor to be diagnosed and the reconstructed tensor is minimal, meaning the reconstruction error is at a low level. Simultaneously, the distance between the currently decomposed physical quantity relationships and the reference benchmark does not increase despite the overall decrease in numerical amplitude, indicating that the factor matrix bias is also at a low level.

[0089] The local sensitivity metric calculated by the fault diagnosis module based on these two low-level indicators will be less than the preset fault threshold. Therefore, the system will ultimately output a judgment that the system is operating normally and no alarm has been triggered.

[0090] As can be seen from the above process, the system in this embodiment utilizes a third-order tensor structure that includes physical quantity dimensions to preserve the multidimensional topological characteristics of the data. During the calculation process, the system does not treat current data as an isolated variable, but rather places it and irradiance data within the same tensor structure for correlation analysis. By introducing a reference physical quantity factor matrix as a decomposition constraint, the system forces the computational model to extract features based on the nonlinear coupling relationships between various physical quantities.

[0091] When changes in environmental factors can reasonably explain changes in electrical parameters, the system generates extremely low reconstruction errors and factor biases, thereby achieving decoupling of environmental interference factors and actual component failures at the algorithm level and effectively solving the false alarm problem caused by traditional threshold methods.

[0092] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A photovoltaic module fault diagnosis method based on machine learning, characterized in that, Includes the following steps: S1. Collect historical multidimensional operating data of photovoltaic modules, and organize the historical multidimensional operating data according to time windows to construct a third-order tensor containing time window dimension, physical quantity dimension and sampling point dimension; S2. Construct a tensor decomposition model that includes a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix, and a core factor tensor. By introducing a multi-kernel function to constrain the time factor matrix and the physical quantity factor matrix, and by minimizing the objective function that includes a reconstruction error term and a regularization constraint term, the core factor tensor and the low-rank factor matrices of each dimension are obtained. S3. Using historical data of photovoltaic modules under normal operating conditions, construct tensors according to the steps, and decompose them according to the method described in step S2 to obtain the reference core factor tensor and reference physical quantity factor matrix under normal conditions. S4. Calculate the reconstruction error of the current tensor to be diagnosed and the factor matrix deviation between the physical quantity factor matrix obtained by decomposing the tensor to be diagnosed based on the tensor decomposition model and the reference physical quantity factor matrix. Then, perform weighted fusion on the reconstruction error and the factor matrix deviation to obtain the local sensitivity metric. S5. Compare the local sensitivity metric value with a preset fault threshold. If the local sensitivity metric value is greater than or equal to the fault threshold, it is determined that the photovoltaic module has a fault.

2. The photovoltaic module fault diagnosis method based on machine learning according to claim 1, characterized in that, The dimensions of the third-order tensor are defined as follows: The time window dimension corresponds to multiple continuously collected time segments, which are used to characterize the evolution of the photovoltaic module's operating status over time; The physical quantities correspond to the DC voltage, DC current, irradiance, and backsheet temperature of the photovoltaic module, and are used to characterize the coupling relationship of multiple physical fields. The sampling point dimension corresponds to the number of data points collected within each time window, and is used to characterize the instantaneous fluctuation characteristics within a single time window.

3. The photovoltaic module fault diagnosis method based on machine learning according to claim 1, characterized in that, The decomposition process of the third-order tensor by the tensor decomposition model includes: The third-order tensor is represented as the sum of multiple tensors of rank one; Each rank-one tensor is formed by the outer product of the column vectors of the time factor matrix, the column vectors of the physical quantity factor matrix, and the column vectors of the sampling point factor matrix, and is weighted by the corresponding elements of the core factor tensor.

4. The photovoltaic module fault diagnosis method based on machine learning according to claim 3, characterized in that, The objective function in step S2 is composed of: Reconstruction error term: used to characterize the Euclidean distance between the original third-order tensor and the tensor reconstructed using the core factor tensor and each factor matrix; Regularization constraint terms: include time constraint terms for the time factor matrix and physical quantity constraint terms for the physical quantity factor matrix; the time constraint terms are calculated based on the trace of the product of the time factor matrix and the inverse of the time kernel matrix; the physical quantity constraint terms are calculated based on the trace of the product of the physical quantity factor matrix and the inverse of the physical kernel matrix.

5. The photovoltaic module fault diagnosis method based on machine learning according to claim 4, characterized in that, The selection rule for the multi-core function is as follows: The time kernel matrix is ​​constructed using a Gaussian kernel function to constrain the smooth evolution characteristics of the time factor matrix in the time dimension and prevent non-fault-related time mutation interference. The physical quantity kernel matrix is ​​constructed using a polynomial kernel function to constrain the nonlinear coupling characteristics of the physical quantity factor matrix between different physical parameters, so as to extract the correlation features between physical quantities.

6. The photovoltaic module fault diagnosis method based on machine learning according to claim 1, characterized in that, The calculation method for the local sensitivity metric value in step S4 is as follows: The reconstruction error and the deviation of the factor matrix are multiplied by their respective weighting coefficients and then summed; the weighting coefficients are used to balance the influence of the overall reconstruction accuracy and the deviation of local physical quantity characteristics on the fault decision.

7. The photovoltaic module fault diagnosis method based on machine learning according to claim 6, characterized in that, The reconstruction error is calculated as follows: the Frobenius norm between the current third-order tensor to be diagnosed and the tensor reconstructed using the low-rank factor matrices of each dimension and the core factor tensor obtained after decomposing the tensor to be diagnosed.

8. The photovoltaic module fault diagnosis method based on machine learning according to claim 6, characterized in that, The factor matrix deviation is calculated as follows: the distance between the physical quantity factor matrix obtained by the current tensor decomposition of the diagnostic tensor and the reference physical quantity factor matrix is ​​calculated; the distance adopts a distance metric based on local weighting to quantify the degree to which the current physical quantity coupling relationship deviates from the normal state reference benchmark.

9. The photovoltaic module fault diagnosis method based on machine learning according to claim 1, characterized in that, The fault determination step in step S5 further includes: when it is determined that there is a fault in the photovoltaic module, comparing the magnitude of the deviation values ​​corresponding to each physical quantity dimension in the physical quantity factor matrix, identifying the physical quantity dimension that causes the largest deviation in the factor matrix as the fault source, and outputting the fault source as the fault tracing result.

10. A photovoltaic module fault diagnosis system based on machine learning, characterized in that, A machine learning-based photovoltaic module fault diagnosis method according to any one of claims 1-9, the system comprising: The multidimensional data construction module is used to collect historical multidimensional operating data of photovoltaic modules and organize the operating data according to time windows to construct a third-order tensor containing time window dimension, physical quantity dimension and sampling point dimension; The feature extraction module is used to construct a tensor decomposition model that includes a time factor matrix, a physical quantity factor matrix, a sampling point factor matrix, and a core factor tensor. By introducing a multi-kernel function to constrain the time factor matrix and the physical quantity factor matrix, the core factor tensor and the low-rank factor matrix of each dimension are obtained by minimizing the objective function that includes the reconstruction error term and the regularization constraint term. The reference model storage module is used to store the reference core factor tensor and reference physical quantity factor matrix of the photovoltaic module under normal operating conditions. An anomaly detection module is used to calculate the reconstruction error of the tensor to be diagnosed and the factor matrix deviation between the physical quantity factor matrix decomposed from the tensor and the reference physical quantity factor matrix based on the reference model, and to perform weighted fusion of the reconstruction error and the factor matrix deviation to obtain a local sensitivity metric. The fault diagnosis module is used to compare the local sensitivity metric value with a preset fault threshold. If the local sensitivity metric value is greater than or equal to the fault threshold, the photovoltaic module is determined to have a fault.

Images