Photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency characteristics

Abstract

The invention relates to a photovoltaic power generation system fault diagnosis method based on a dual-channel CNN and time-frequency characteristics. The method comprises the following steps: S1, collecting and preprocessing historical time sequence data of a photovoltaic power generation system; s2, copying the preprocessed data set to obtain a one-dimensional time sequence data set and a two-dimensional time-frequency graph data set; s3, inputting a dual-channel CNN for feature extraction, enabling a one-dimensional time sequence data set to enter a 1DCNN channel, and enabling a two-dimensional time-frequency graph data set to enter a 2DCNN channel; s4, performing dynamic fusion on the features extracted by the dual-channel CNN through an SE attention module; s5, constructing a classifier based on the SVM, and optimizing two key parameters of a kernel function of the classifier by adopting an adaptive inertia weight particle swarm algorithm to obtain a final SVM classifier; and S6, inputting the fused features into a final SVM classifier, and carrying out fault diagnosis classification. The complex fault diagnosis efficiency and reliability of the photovoltaic power generation system are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation system fault diagnosis technology, and in particular to a photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features. Background Technology

[0002] Photovoltaic power generation is one of the main technologies for utilizing solar energy, converting solar radiation into electrical energy to provide clean and sustainable power. A photovoltaic power generation system consists of multiple components, including photovoltaic arrays, inverters, combiner boxes, connecting lines, and monitoring and control equipment. These components work together to achieve the conversion of solar energy, playing a vital role in the supply of green energy.

[0003] Photovoltaic power generation systems may face various faults during long-term operation, mainly including inverter faults, line faults, and photovoltaic array faults. These faults affect the system's performance and stability. Therefore, fault diagnosis of photovoltaic power generation systems is a crucial link in ensuring reliable system operation. While fault occurrences can be represented by signals, in practical applications, the correspondence between faults and their causes is not one-to-one; a single fault can have multiple causes, making fault feature extraction difficult. Based on this, signal processing techniques, especially modern signal processing theories, methods, and technologies, are used to mine and extract features from the acquired raw data, thereby achieving efficient and accurate fault diagnosis.

[0004] However, photovoltaic system faults under complex operating conditions exhibit multimodal coupling characteristics, such as concurrent faults involving harmonic distortion and power attenuation, and transient high-frequency components in arc faults. This poses a severe challenge to traditional diagnostic methods. Existing technologies mostly employ single-dimensional feature analysis methods. For example, while 1D-CNN time-domain diagnostic models can capture the temporal patterns of current / voltage sequences, they have blind spots in detecting frequency-sensitive faults (such as characteristic harmonics caused by abnormal insulation impedance). While FFT-based frequency-domain analysis methods can identify steady-state spectral features, they exhibit lag in response to time-varying faults such as inverter switching transients and struggle to distinguish the interactive effects of coupled faults. Furthermore, traditional methods employ a single convolutional neural network structure, which, limited by its local receptive field, makes it difficult to simultaneously extract time-domain waveform details and frequency-domain energy distribution features, leading to increased false negative rates for complex faults. Simultaneously, traditional classifiers such as SVM are prone to overfitting when faced with high-dimensional fused features due to fixed kernel functions and parameter sensitivity, limiting their cross-condition generalization ability. Therefore, there is an urgent need to construct an intelligent diagnostic architecture capable of deeply fusing time-frequency features and adapting to dynamic fault modes. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a fault diagnosis method for photovoltaic power generation systems based on dual-channel CNN and time-frequency features. This method overcomes the problem that traditional single CNN methods only focus on time-domain or frequency-domain information when diagnosing photovoltaic system faults, resulting in blind spots for complex coupled faults and thus increasing the missed detection rate of complex faults.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A fault diagnosis method for photovoltaic power generation systems based on dual-channel CNN and time-frequency features includes the following steps:

[0008] S1: Collect and preprocess historical time-series data of the photovoltaic power generation system;

[0009] S2: Copy the preprocessed dataset, perform a short-time Fourier transform on one of the datasets, perform a Fourier transform on the signal within a preset time window, and perform spectral analysis on the entire signal by moving the time window to obtain information about the signal in time and frequency, thus becoming a two-dimensional time-frequency graph dataset, while the other dataset remains a one-dimensional time series dataset.

[0010] S3: Input the one-dimensional time series dataset and the two-dimensional time-frequency graph dataset into a dual-channel CNN for feature extraction. The one-dimensional time series dataset enters the 1DCNN channel, and the two-dimensional time-frequency graph dataset enters the 2DCNN channel.

[0011] S4: Dynamically fuse the features extracted by the dual-channel CNN through the SE attention module;

[0012] S5: A classifier is built based on SVM, and the two key parameters of the kernel function of the classifier are optimized by the adaptive inertial weight particle swarm algorithm to obtain the final SVM classifier.

[0013] S6: Input the fused features into the final SVM classifier for fault diagnosis and classification.

[0014] Further preprocessing is performed as follows: the data is resampled, the signal length is set to q, the sliding window overlap rate is p, a sequence dataset is formed, and corresponding labels are added. Then the data is normalized and mapped to the closed interval [0,1] to facilitate subsequent feature extraction.

[0015] Furthermore, 1DCNN includes one-dimensional convolutional layers and pooling layers, as detailed below:

[0016] One-dimensional convolutional layer: This layer uses convolution and extracts fault features.

[0017] N

[0018] in, Indicates by the first Features extracted by each convolutional kernel; This indicates its corresponding weight; Indicates the bias term; Indicates one-dimensional input; express The number of input data points;

[0019] After the convolution operation, through Activation functions are used to implement nonlinear transformations of convolutional data:

[0020]

[0021] Pooling layer: Employs max pooling to reduce redundant data by lowering dimensionality while retaining useful information. The mathematical formula for max pooling is as follows:

[0022]

[0023] in, Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling output value; Represents the pooling region Central The element at that location.

[0024] Furthermore, 2DCNN includes two-dimensional convolutional layers, pooling layers, and flattening layers, as detailed below:

[0025] The two-dimensional convolutional layer uses convolution to extract fault features. Its mathematical formula is:

[0026] N

[0027] in, Indicates by the first Features extracted by a 2D convolutional kernel; This indicates its corresponding weight; Indicates the bias term; Indicates two-dimensional input; express The number of input data points;

[0028] After the convolution operation, through Activation functions are used to implement nonlinear transformations of convolutional data:

[0029]

[0030] The pooling layer employs max pooling to reduce redundant data by lowering dimensionality while retaining useful information.

[0031]

[0032] in, Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling output value; Represents the pooling region Central Elements at the location;

[0033] The flattening layer flattens the two-dimensional fault features into one-dimensional features, so that the one-dimensional fault features and the two-dimensional fault features can be fused in the fusion layer. The formula is as follows:

[0034]

[0035]

[0036] in, The size extracted for a 2D convolutional neural network is Two-dimensional feature map; For the first Line number The characteristic values ​​of the column; These are the flattened feature vectors.

[0037] Furthermore, the SE attention module is detailed below:

[0038] First, the compression stage compresses the global spatial information of each channel into a scalar, capturing the global feature distribution of the channels. Then, global average pooling is performed on the input feature map to obtain the channel descriptors. :

[0039]

[0040] in, It is the spatial dimension of the feature map. Number of channels; It is the set of real numbers;

[0041] Then, in the activation phase, the nonlinear relationships between channels are learned, channel weights are generated, and then passed through the first fully connected layer. Compressed to a low-dimensional space, using activation:

[0042]

[0043] The original number of channels is restored using a second fully connected layer, and a weight vector is generated using Sigmoid activation.

[0044]

[0045] Finally, feature recalibration is performed, and the weight vector is... The enhanced feature map is obtained by multiplying it channel by channel with the original feature map: .

[0046] Furthermore, the adaptive inertia weighted particle swarm optimization algorithm is used to optimize two key parameters of the SVM kernel function: the penalty factor. (The penalty level for misclassified samples controlled by the classifier) ​​and the RBF kernel function (Determines the local influence range of the kernel function), as follows:

[0047] First, the parameters are encoded and initialized, with the particle position represented as a combination parameter. The particle velocity is initialized to a random small value to avoid the search process getting trapped in a local optimum too early.

[0048] Next, inertia weights are used. The strategy for dynamically adjusting the degree to which particles inherit the current velocity is as follows:

[0049]

[0050] in, The maximum number of iterations, This represents the current iteration number.

[0051] Next, a weighted combination of F1-Score and ROC-AUC from cross-validation was used as the optimization objective:

[0052]

[0053] Randomly generate particle swarms, set (Individual learning factors) (Social learning factor)

[0054] For each particle, train the classification model and calculate the fitness value of cross-validation, update the individual and global optima, and record the historical best for each particle. and global optimal And adjust the speed and position:

[0055]

[0056]

[0057] If the parameter exceeds the preset range, reset it to the boundary.

[0058] Termination condition: Reaching the maximum number of iterations or the fitness value changing less than a threshold for a consecutive number of iterations.

[0059] A photovoltaic power generation system fault diagnosis system based on dual-channel CNN and time-frequency features includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features as described above.

[0060] The present invention has the following beneficial effects:

[0061] This invention enables better diagnosis of complex faults in photovoltaic systems. It solves the problem that traditional methods, which use a single convolutional neural network structure, are limited by their local receptive field and cannot simultaneously extract time-domain waveform details and frequency-domain energy distribution features, leading to a higher rate of missed detections for complex faults. Attached Figure Description

[0062] Figure 1 This is a flowchart of a time-frequency feature fusion method in one embodiment of the present invention;

[0063] Figure 2 This is a flowchart of an improved PSO-SVM classifier in one embodiment of the present invention. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0065] refer to Figure 1-2 In this embodiment, a fault diagnosis method for photovoltaic power generation systems based on dual-channel CNN and time-frequency features is provided, including the following steps:

[0066] S1: Collect historical time-series data of the photovoltaic power generation system, i.e., one-dimensional fault signal current, voltage, power, etc. Resample the data, setting the signal length to 256 and the sliding window overlap rate to 0.75, forming a sequence dataset and adding corresponding labels. Then normalize the data, mapping it to the closed interval [0,1] to facilitate subsequent feature extraction;

[0067] S2: Copy the preprocessed dataset, perform a short-time Fourier transform on one of the datasets, perform a Fourier transform on the signal within a preset time window, and perform spectral analysis on the entire signal by moving the time window to obtain information about the signal in time and frequency, thus becoming a two-dimensional time-frequency graph dataset, while the other dataset remains a one-dimensional time series dataset.

[0068] Preferably, in this embodiment, the short-time Fourier transform has the following specific steps:

[0069] Segmenting the signal: Dividing the input signal into multiple overlapping segments with different time windows;

[0070] Apply a window function to each time window: Window the signal for each time period using a Hanning window.

[0071] Perform Fourier transform: Perform Fourier transform on the windowed signal to obtain the spectral components;

[0072] Overlapping and summing: The spectral components of each time period are superimposed to obtain the overall spectrum.

[0073] S3: Input the one-dimensional time series dataset and the two-dimensional time-frequency graph dataset into a dual-channel CNN for feature extraction. The one-dimensional time series dataset enters the 1DCNN channel, and the two-dimensional time-frequency graph dataset enters the 2DCNN channel.

[0074] S4: Dynamically fuse the features extracted by the dual-channel CNN through the SE attention module;

[0075] S5: A classifier is built based on SVM, and the two key parameters of the kernel function of the classifier are optimized by the adaptive inertial weight particle swarm algorithm to obtain the final SVM classifier.

[0076] S6: Input the fused features into the final SVM classifier for fault diagnosis and classification.

[0077] In this implementation, the 1DCNN includes one-dimensional convolutional layers and pooling layers, as detailed below:

[0078] One-dimensional convolutional layer: This layer uses convolution and extracts fault features.

[0079] N

[0080] in, Indicates by the first Features extracted by each convolutional kernel; This indicates its corresponding weight; Indicates the bias term; Indicates one-dimensional input; express The number of input data points;

[0081] After the convolution operation, through Activation functions are used to implement nonlinear transformations of convolutional data:

[0082]

[0083] Pooling layer: Employs max pooling to reduce redundant data by lowering dimensionality while retaining useful information. The mathematical formula for max pooling is as follows:

[0084]

[0085] in, Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling output value; Represents the pooling region Central The element at that location.

[0086] In this embodiment, the 2DCNN includes two-dimensional convolutional layers, pooling layers, and flattening layers, as detailed below:

[0087] The two-dimensional convolutional layer uses convolution to extract fault features. Its mathematical formula is:

[0088] N

[0089] in, Indicates by the first Features extracted by a 2D convolutional kernel; This indicates its corresponding weight; Indicates the bias term; Indicates two-dimensional input; express The number of input data points;

[0090] After the convolution operation, through Activation functions are used to implement nonlinear transformations of convolutional data:

[0091]

[0092] The pooling layer employs max pooling to reduce redundant data by lowering dimensionality while retaining useful information.

[0093]

[0094] in, Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling output value; Represents the pooling region Central Elements at the location;

[0095] The flattening layer flattens the two-dimensional fault features into one-dimensional features, so that the one-dimensional fault features and the two-dimensional fault features can be fused in the fusion layer. The formula is as follows:

[0096]

[0097]

[0098] in, The size extracted for a 2D convolutional neural network is Two-dimensional feature map; For the first Line number The characteristic values ​​of the column; These are the flattened feature vectors.

[0099] In this embodiment, the SE attention module is specifically as follows:

[0100] First, the compression stage compresses the global spatial information of each channel into a scalar, capturing the global feature distribution of the channels. Then, global average pooling is performed on the input feature map to obtain the channel descriptors. :

[0101]

[0102] in, It is the spatial dimension of the feature map. Number of channels; It is the set of real numbers;

[0103] Then, in the activation phase, the nonlinear relationships between channels are learned, channel weights are generated, and then passed through the first fully connected layer. Compressed to a low-dimensional space, using activation:

[0104]

[0105] The original number of channels is restored using a second fully connected layer, and a weight vector is generated using Sigmoid activation.

[0106]

[0107] Finally, feature recalibration is performed, and the weight vector is... The enhanced feature map is obtained by multiplying it channel by channel with the original feature map: .

[0108] The features from the two parts are dynamically fused using the Squeeze-and-Excitation (SE) attention module. The dynamic fusion weights are then calculated. :

[0109]

[0110] in This represents the mean of the channel weights, reflecting the importance of the characteristics of each branch.

[0111] The fusion formula is:

[0112] In this embodiment, the adaptive inertia weighted particle swarm optimization algorithm is used to optimize two key parameters of the SVM kernel function: the penalty factor. (The penalty level for misclassified samples controlled by the classifier) ​​and the RBF kernel function (Determines the local influence range of the kernel function), as follows:

[0113] First, the parameters are encoded and initialized, with the particle position represented as a combination parameter. The particle velocity is initialized to a random small value to avoid the search process getting trapped in a local optimum too early.

[0114] Next, inertia weights are used. The strategy for dynamically adjusting the degree to which particles inherit the current velocity is as follows:

[0115]

[0116] in, The maximum number of iterations, This represents the current iteration number.

[0117] Next, a weighted combination of F1-Score and ROC-AUC from cross-validation was used as the optimization objective:

[0118]

[0119] Randomly generate particle swarms, set (Individual learning factors) (Social learning factor)

[0120] For each particle, train the classification model and calculate the fitness value of cross-validation, update the individual and global optima, and record the historical best for each particle. and global optimal And adjust the speed and position:

[0121]

[0122]

[0123] If the parameter exceeds the preset range, reset it to the boundary.

[0124] Termination condition: Reaching the maximum number of iterations or the fitness value changing less than a threshold for a consecutive number of iterations.

[0125] A photovoltaic power generation system fault diagnosis system based on dual-channel CNN and time-frequency features includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features as described above.

[0126] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0127] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0130] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A fault diagnosis method for photovoltaic power generation systems based on dual-channel CNN and time-frequency features, characterized in that, Includes the following steps: S1: Collect and preprocess historical time-series data of the photovoltaic power generation system; S2: Copy the preprocessed dataset, perform a short-time Fourier transform on one of the datasets, perform a Fourier transform on the signal within a preset time window, and perform spectral analysis on the entire signal by moving the time window to obtain information about the signal in time and frequency, thus becoming a two-dimensional time-frequency graph dataset, while the other dataset remains a one-dimensional time series dataset. S3: Input the one-dimensional time series dataset and the two-dimensional time-frequency graph dataset into a dual-channel CNN for feature extraction. The one-dimensional time series dataset enters the 1DCNN channel, and the two-dimensional time-frequency graph dataset enters the 2DCNN channel. S4: Dynamically fuse the features extracted by the dual-channel CNN through the SE attention module; S5: A classifier is built based on SVM, and the two key parameters of the kernel function of the classifier are optimized by the adaptive inertial weight particle swarm algorithm to obtain the final SVM classifier. S6: Input the fused features into the final SVM classifier for fault diagnosis and classification.

2. The photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features according to claim 1, characterized in that, The preprocessing is as follows: the data is resampled, the signal length is set to q, the sliding window overlap rate is p, a sequence dataset is formed, and corresponding labels are added. Then the data is normalized and mapped to the closed interval [0,1] to facilitate subsequent feature extraction.

3. The photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features according to claim 1, characterized in that, The 1DCNN includes one-dimensional convolutional layers and pooling layers, as detailed below: One-dimensional convolutional layer: This layer uses convolution and extracts fault features. ,N in, Indicates by the first Features extracted by each convolutional kernel; This indicates its corresponding weight; Indicates the bias term; Indicates one-dimensional input; express The number of input data points; After the convolution operation, through Activation functions are used to implement nonlinear transformations of convolutional data: ； Pooling layer: Employs max pooling to reduce redundant data by lowering dimensionality while retaining useful information. The mathematical formula for max pooling is as follows: ； in, Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling output value; Represents the pooling region Central The element at that location.

4. The photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features according to claim 1, characterized in that, The 2DCNN includes two-dimensional convolutional layers, pooling layers, and flattening layers, as detailed below: The two-dimensional convolutional layer uses convolution to extract fault features. Its mathematical formula is: ,N in, Indicates by the first Features extracted by a 2D convolutional kernel; This indicates its corresponding weight; Indicates the bias term; Indicates two-dimensional input; express The number of input data points; After the convolution operation, through Activation functions are used to implement nonlinear transformations of convolutional data: ； The pooling layer employs max pooling to reduce redundant data by lowering dimensionality while retaining useful information. ； in, Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling output value; Represents the pooling region Central Elements at the location; The flattening layer flattens the two-dimensional fault features into one-dimensional features, so that the one-dimensional fault features and the two-dimensional fault features can be fused in the fusion layer. The formula is as follows: ； ； in, The size extracted for a 2D convolutional neural network is Two-dimensional feature map; For the first Line number The characteristic values ​​of the column; These are the flattened feature vectors.

5. The photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features according to claim 1, characterized in that, The SE attention module is as follows: First, the compression stage compresses the global spatial information of each channel into a scalar, capturing the global feature distribution of the channels. Then, global average pooling is performed on the input feature map to obtain the channel descriptors. : ； in, It is the spatial dimension of the feature map. Number of channels; It is the set of real numbers; Then, in the activation phase, the nonlinear relationships between channels are learned, channel weights are generated, and then passed through the first fully connected layer. Compressed to a low-dimensional space, using activation: ； The original number of channels is restored using a second fully connected layer, and a weight vector is generated using Sigmoid activation. ； Finally, feature recalibration is performed, and the weight vector is... The enhanced feature map is obtained by multiplying it channel by channel with the original feature map: .

6. The photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features according to claim 1, characterized in that, The two key parameters of the kernel function of SVM optimized using the adaptive inertial weighted particle swarm optimization algorithm are: penalty factor. and RBF kernel function The details are as follows: First, the parameters are encoded and initialized, with the particle position represented as a combination parameter. The particle velocity is initialized to a random small value to avoid the search process getting trapped in a local optimum too early. Next, inertia weights are used. The strategy for dynamically adjusting the degree to which particles inherit the current velocity is as follows: ； in, The maximum number of iterations, This represents the current iteration number; Next, a weighted combination of F1-Score and ROC-AUC from cross-validation was used as the optimization objective: ； Randomly generate particle swarms, set , ; For each particle, train the classification model and calculate the fitness value of cross-validation, update the individual and global optima, and record the historical best for each particle. and global optimal And adjust the speed and position: ； ； If the parameter exceeds the preset range, reset it to the boundary. Termination condition: The maximum number of iterations is reached or the fitness value changes less than a threshold for a certain number of consecutive iterations.

7. A fault diagnosis system for photovoltaic power generation systems based on dual-channel CNN and time-frequency features, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the photovoltaic power generation system fault diagnosis method based on dual-channel CNN and time-frequency features as described in any one of claims 1-6.

Images