Photovoltaic equipment fault diagnosis method based on big data and deep learning

Abstract

The present invention relates to the technical field of photovoltaic equipment fault diagnosis, and discloses a photovoltaic equipment fault diagnosis method based on big data and deep learning, including obtaining operation log data, sensor data and maintenance record data of photovoltaic equipment; using a fully connected neural network based on deep pseudo feedback to extract features from the acquired data to obtain a first feature; using an autoencoder neural network based on a photovoltaic equipment condition vector to reduce the feature dimension of the first feature to obtain a second feature; using a support vector machine algorithm based on error correction learning to build a classifier model, using the second feature to train the classifier model; and using the trained classifier model to diagnose photovoltaic equipment faults. The present invention realizes accurate and rapid diagnosis of photovoltaic equipment faults.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic equipment fault diagnosis, and in particular to a photovoltaic equipment fault diagnosis method based on big data and deep learning. Background Art

[0002] With the rapid development of the photovoltaic industry and the widespread deployment of photovoltaic systems around the world, how to ensure the stable operation and efficient performance of photovoltaic equipment has become the focus of the industry. Photovoltaic equipment may be affected by various internal and external factors during operation, resulting in performance degradation or failure. These factors include equipment aging, environmental impact, installation errors, etc. Therefore, accurate and rapid diagnosis and handling of these faults is crucial to improving photovoltaic power generation efficiency and system reliability.

[0003] The Chinese invention patent with application number CN202410291219.5 proposes a method for classifying and identifying five different states: normal operation, open circuit fault, short circuit fault, aging fault, and shadow blocking fault. A fault diagnosis model based on DBN was established. The data set was generated based on the Matlab / Simulink model of the photovoltaic array under different operating conditions. The DBN model was trained and the BA algorithm was used to optimize the number of neurons in each hidden layer to achieve the optimal training result. The test data set was then input into the BA-DBN model for photovoltaic array fault diagnosis, and the accuracy was significantly higher than that of K-nearest neighbors (KNN), backpropagation neural network (BPNN) and the original DBN algorithm, reaching 98%.

[0004] The Chinese invention patent with application number CN202410444547.4 proposes a photovoltaic equipment fault diagnosis method based on real-time calculation of photovoltaic module parameters, including the following steps: obtaining real-time data of photovoltaic modules and preprocessing; using comprehensive analysis method to detect abnormal data on the acquired real-time data of photovoltaic modules; locating abnormal photovoltaic equipment according to the identification results of abnormal real-time data, and analyzing the fault type of the abnormal photovoltaic equipment after positioning; recommending corresponding maintenance plans for abnormal photovoltaic equipment according to the fault type of the abnormal photovoltaic equipment, and automatically generating a fault report. The present invention can more accurately identify abnormal real-time data by using comprehensive analysis method, effectively improve the accuracy of fault detection, and use drones combined with infrared thermal imaging technology and image processing technology to quickly inspect large-area photovoltaic power stations in a short time, realize fault type identification of abnormal photovoltaic equipment, and greatly improve maintenance efficiency.

[0005] The Chinese invention patent with application number CN202410332923.0 proposes a data-driven photovoltaic power generation fault prediction and diagnosis system and method. The method includes pre-dividing the photovoltaic panel into i divided areas, collecting real-time feature data of the photovoltaic panels in the i divided areas per unit time; inputting the real-time feature data into a pre-built fault prediction model to obtain the operating status of the photovoltaic panel after a unit time; the operating status includes the absence of a fault and the presence of a fault, and the presence of a fault includes local blockage damage and temperature abnormality; collecting real-time image data of the corresponding photovoltaic panel according to the temperature abnormality; inputting the collected image data into a pre-built fault diagnosis model to obtain the fault type, and the fault type includes internal abnormality and external abnormality; facilitating timely repair or replacement of the faulty part, reducing the probability of fault occurrence, timely discovering heat spots and improving the efficiency of treating heat spots, so as to ensure the normal operation and power generation efficiency of the photovoltaic panel power generation components.

[0006] The above technical solution has the following problems that still need to be further solved:

[0007] 1. Data expansion in photovoltaic equipment fault diagnosis is usually limited by the amount and quality of original data, and there is a lack of effective methods to increase sample diversity and quantity, which leads to limited model training and insufficient generalization ability.

[0008] 2. In terms of data feature extraction for photovoltaic equipment fault diagnosis, traditional neural networks are prone to the problem of fixed network structure, which makes it impossible to effectively extract features for different data characteristics, affecting the stability and performance of training.

[0009] 3. In terms of data feature dimensionality reduction for photovoltaic equipment fault diagnosis, existing methods often lead to the loss of key information, reducing the explanatory power and accuracy of the model, especially in complex data scenarios.

[0010] 4. In the training of diagnostic classifiers for photovoltaic equipment fault diagnosis, traditional algorithms often have insufficient classification performance when processing complex data with unclear boundaries, resulting in limited accuracy and application scope of the model. Summary of the invention

[0011] In view of the above-mentioned deficiencies in the prior art, the present invention provides a photovoltaic equipment fault diagnosis method based on big data and deep learning.

[0012] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0013] A photovoltaic equipment fault diagnosis method based on big data and deep learning includes the following steps:

[0014] Obtain operation log data, sensor data, and maintenance record data of photovoltaic equipment;

[0015] A fully connected neural network based on deep pseudo feedback is used to extract features from the acquired data to obtain the first feature;

[0016] An autoencoder neural network based on the photovoltaic device condition vector is used to reduce the dimension of the first feature to obtain the second feature;

[0017] A support vector machine algorithm based on error correction learning is used to build a classifier model, and the second feature is used to train the classifier model;

[0018] The trained classifier model is used for photovoltaic equipment fault diagnosis.

[0019] Preferably, the fully connected neural network based on deep pseudo feedback includes:

[0020] 1 input layer, 3 hidden layers and 1 feature extraction output layer;

[0021] The number of hidden layers and the number of neurons are dynamically adjusted according to the change of loss during the training process.

[0022] Preferably, a fully connected neural network based on deep pseudo feedback is used to extract features from the operation log data, sensor data, and maintenance record data to obtain a first feature, including:

[0023] Initializing parameters of a fully connected neural network; the parameters of the fully connected neural network include weights and bias parameters of the neural network;

[0024] The operation log data, sensor data and maintenance record data are input into the fully connected neural network through forward propagation, and the feature vector after feature extraction is output through the feature extraction output layer. At the same time, a pseudo feedback signal is generated according to the error between the output feature and the target feature.

[0025] According to the strength of the pseudo-feedback signal and the current learning efficiency of the model, the network structure is dynamically adjusted, and the adaptive structure of the fully connected neural network is adjusted by increasing or decreasing the number of neurons;

[0026] Use gradient descent to update the weights and biases of the fully connected neural network to minimize the error of the feature extraction output layer;

[0027] Repeat the above steps until the preset stop iteration condition is met.

[0028] Preferably, the network structure is dynamically adjusted according to the strength of the pseudo feedback signal and the current learning efficiency of the model, and the adaptive structure adjustment of the fully connected neural network is performed by increasing or decreasing the number of neurons, specifically:

[0029]

[0030] in, Adjust the rate for the structure; For the Activity measurement of layers; The neural network The number of neurons in the layer; is a symbolic function; is the activity measurement threshold.

[0031] Preferably, an autoencoder neural network based on a photovoltaic device condition vector is used to perform feature dimension reduction on the first feature to obtain a second feature, including:

[0032] Randomly initialize the parameters of the autoencoder neural network according to the data dimension of the first feature and the expected encoding dimension;

[0033] The encoder is used to combine the first feature with the photovoltaic device condition vector and then generate a coded representation through an activation function;

[0034] Decode the encoded representation using a decoder to obtain a reconstructed vector;

[0035] Use the loss function to evaluate the error between the reconstructed vector and the first feature while performing sparsity constraints;

[0036] Parameters are updated according to the momentum factor of the autoencoder neural network and the adaptive learning rate of the autoencoder neural network;

[0037] Repeat the above steps until the preset stop iteration condition is met.

[0038] Preferably, a loss function is used to evaluate the error between the reconstructed vector and the first feature, and a sparsity constraint is performed at the same time, specifically:

[0039]

[0040] in, is the loss function of the autoencoder neural network, The number of training samples input for the current batch, For the The input vector of the encoder of the autoencoder neural network, For the The reconstruction vector output by the decoder; is the sparsity penalty coefficient; Calculates the Kullback-Leibler divergence function; is the number of neurons in the hidden layer of the autoencoder neural network; is the desired sparse activation rate, is the actual sparse activation rate.

[0041] Preferably, a support vector machine algorithm based on error correction learning is used to construct a classifier model, and the second feature is used to train the classifier model, including:

[0042] Initialize the parameters of the support vector machine, set the initial kernel function and penalty parameters, and initialize the bias parameters of the support vector machine;

[0043] According to the prediction results of the current model on the training set, the hardness of the classification boundary is dynamically adjusted by incrementally adjusting the penalty parameter and the width parameter;

[0044] Error correction learning is used to identify misclassified samples in the training data, and additional learning weights are applied to data points near the classification boundary;

[0045] According to the characteristic properties of the current training data, the kernel function is dynamically adjusted or selected to match the training data structure;

[0046] Repeat the above steps until the preset stop iteration condition is met.

[0047] Preferably, according to the prediction results of the current model on the training set, the hardness of the classification boundary is dynamically adjusted by incrementally adjusting the penalty parameter and the width parameter, specifically:

[0048]

[0049]

[0050]

[0051] in, is the penalty parameter increment of the support vector machine; is the loss function of the support vector machine; For the The true label of the data, Support vector machine for The prediction results of the data, is the number of samples in the input batch; is the first adjustment coefficient of the support vector machine, is the second adjustment coefficient of the support vector machine, is the target loss threshold of the support vector machine; For the The width parameter of the kernel function for each iteration, For the The width parameter of the kernel function for each iteration; For the The penalty parameter for the iteration, For the The penalty parameter for the iteration.

[0052] Preferably, obtaining operation log data, sensor data and maintenance record data of the photovoltaic equipment further includes:

[0053] A generative adversarial network based on deep manifold learning is used to augment the acquired data.

[0054] Preferably, a generative adversarial network based on deep manifold learning is used to expand the acquired data, including:

[0055] Initialize the parameters of the generative adversarial network;

[0056] Clustering is performed based on the acquired data, the initial data distribution is determined, and the cluster center is calculated;

[0057] In the alternating training phase of the generator and the discriminator, the generator is used to generate pseudo data corresponding to each cluster center using the current cluster center, and the parameters of the generator are updated;

[0058] The generated pseudo data is sent to the discriminator, and the discriminator is used to classify each sample as true or false, and the parameters of the discriminator are updated;

[0059] In each iteration, the generated pseudo data is re-clustered and the cluster center is dynamically adjusted;

[0060] Repeat the above steps until the preset stop iteration condition is met.

[0061] The present invention has the following beneficial effects:

[0062] 1. The present invention adopts a generative adversarial network based on deep manifold learning, through the game process of the generator and the discriminator, to effectively increase the number of samples and improve the sample quality in order to solve the problems of limited data volume and insufficient sample diversity.

[0063] 2. The present invention adopts a five-layer fully connected neural network based on deep pseudo feedback. To address the problems of gradient vanishing, explosion or local optimal solution that traditional neural networks may encounter during feature extraction, the training stability and performance of the network are improved through adaptive structural adjustment.

[0064] 3. The present invention adopts an autoencoding neural network based on the conditional vector of photovoltaic equipment. To address the problems of information loss in feature dimensionality reduction and insufficient model generalization ability, the present invention enhances feature abstraction ability and generalization performance by utilizing conditional vectors and sparsity constraints.

[0065] 4. The present invention adopts a support vector machine algorithm based on error correction learning. To solve the problem of low classification accuracy caused by unclear classification boundaries, the classification performance of the model for data points in the boundary area is enhanced through error correction learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 The figure is a flow chart of a photovoltaic equipment fault diagnosis method based on big data and deep learning. DETAILED DESCRIPTION

[0067] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0068] like Figure 1 As shown, a photovoltaic equipment fault diagnosis method based on big data and deep learning provided by an embodiment of the present invention includes the following steps S1 to S5:

[0069] S1. Obtaining operation log data, sensor data and maintenance record data of photovoltaic equipment;

[0070] In an optional embodiment of the present invention, the data collected by the present invention is derived from the operation log, sensor data and maintenance record of the photovoltaic equipment. All collected data are stored in a structured JSON format to form a large-scale data set. In one embodiment, the attributes of the data include: Indicates the power output of the photovoltaic panel, Indicates the ambient temperature, Indicates light intensity, Indicates the output voltage, Indicates the output current, Indicates the surface status of photovoltaic panels, Indicates the number of days the equipment has been running. Indicates fault history, Indicates the number of maintenance times, It should be noted that this embodiment is only used to illustrate a data format and type of the present invention. In actual applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0071] Furthermore, the collected data is labeled. The labeling method of the present invention is manual labeling. In one embodiment, the labeled categories include: photovoltaic panel damage, inverter failure, cable connection problem, photovoltaic bracket problem, monitoring system failure, and battery storage system problem.

[0072] It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of effective training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model, while affecting the accuracy of the model. The present invention uses a generative adversarial network algorithm based on deep manifold learning to generate samples, thereby achieving data expansion. The generative adversarial network algorithm based on deep manifold learning includes two parts: a generator and a discriminator. The generator is responsible for generating high-quality pseudo data, and the discriminator is responsible for distinguishing the generated data from the real data. The present invention adopts a deep manifold learning strategy in the generator, so that it can better capture and imitate the distribution of data in high-dimensional space, thereby generating data samples that are more in line with the actual distribution.

[0073] Specifically, the training process of the generative adversarial network algorithm based on deep manifold learning is as follows:

[0074] S101, initialize the parameters of the generative adversarial network algorithm, and set the generator to , the discriminator is , the cluster center is , the number of network layers of the generator is , then the parameter initialization method of the generator and discriminator is expressed as:

[0075]

[0076]

[0077] In the formula, and are the parameters of the generator and the discriminator respectively; Indicates that the parameters are initialized using a normal distribution with a mean of 0 and a variance of the unit matrix; It means that it obeys a specific distribution. represents the identity matrix, is a normal distribution.

[0078] Furthermore, clustering is performed based on the initial data, which is the data obtained after collection and annotation, to determine the preliminary data distribution and calculate the cluster center, which is expressed as:

[0079]

[0080] In the formula, Indicates cluster centers, and the total number of cluster centers is indivual; is the first randomly selected clusters; is the number of clusters, Represents the initial data.

[0081] S102, in the alternating training phase of the generator and the discriminator, the generator uses the current cluster center to try to generate pseudo data corresponding to each cluster center. In this process, the parameter update method of the generator is expressed as:

[0082]

[0083] In the formula, Indicates parameter update operation; represents the L2 norm; is the generator parameter; is the learning rate; represents the gradient of the generator parameters; is the discriminator function; is a generator function; It is Input noise; is the clustering loss weight, Represents a parameter update operation.

[0084] Further, Characterize the distance between the generator output and the cluster center, and the calculation method is expressed as:

[0085]

[0086] In the formula, To generate the characteristic dimension of data; The generator is Dimensional output; It is The cluster center is in Dimensional coordinates.

[0087] Furthermore, the generated data is fed into the discriminator, which classifies each sample as true or false and updates the parameters of the discriminator to more accurately identify false data and real data. In this process, the parameter update method of the discriminator is expressed as:

[0088]

[0089] In the formula, is the discriminator parameter; represents the gradient of the discriminator parameters; is a real data sample; is the input noise of the generator. Preferably, Set to 0.01, Represents a parameter update operation.

[0090] S103. In each iteration process, the generated data is re-clustered and the cluster center is dynamically adjusted. The dynamic adjustment method is expressed as:

[0091]

[0092] In the formula, For the updated Cluster centers; is the updated learning rate of the cluster center; It belongs to Generate a dataset of clusters; For the The cluster centers in the generated dataset of clusters.

[0093] Furthermore, the calculation method of the updated learning rate of the cluster center is expressed as:

[0094]

[0095] In the formula, Represents natural base The exponential function with base , is the standard deviation parameter that controls the update rate. Preferably, Set to 0.01.

[0096] S104, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0097] After the data augmentation model training is completed, the trained data augmentation model is used to increase the number of samples. In one embodiment, assuming that the original collected samples are 800, and the data augmentation model generates 200 samples, the expanded data set contains 1000 samples. By generating adversarial networks to achieve efficient data expansion, the diversity and coverage of training data are improved, and the overfitting phenomenon in model training is effectively alleviated.

[0098] S2. Using a fully connected neural network based on deep pseudo feedback to extract features from the acquired data to obtain the first feature;

[0099] In an optional embodiment of the present invention, the expanded data is input into a feature extraction model to train the feature extraction model. The present invention uses a 5-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient vanishing, gradient explosion, or falling into a local optimal solution may be encountered, affecting the stability of training and the performance of the model. The present invention uses a deep pseudo-feedback neural network algorithm for feature extraction. The deep pseudo-feedback neural network algorithm includes 1 input layer, 3 hidden layers, and 1 feature extraction output layer. The present invention uses an adaptive adjustment mechanism to dynamically adjust the number of hidden layers and the number of neurons according to the loss changes during the training process, so that the neural network can autonomously optimize its structure according to the specific data characteristics and complexity, thereby improving training efficiency and model accuracy.

[0100] Specifically, the training process based on the deep pseudo-feedback neural network algorithm is as follows:

[0101] S201, initializing the parameters of the neural network, including the weight and bias parameters of the neural network. In one embodiment, the initialization method is expressed as:

[0102]

[0103]

[0104] In the formula, Indicates compliance with a specific distribution; The neural network The weight of the layer; The neural network Bias of the layer; and The neural network is Layer and The number of neurons in the layer; Represents a uniform distribution.

[0105] S202, the input data is input to the neural network by forward propagation, and each layer of the neural network calculates its output until it reaches the feature extraction output layer. The feature extraction output layer is not only used to output the feature vector after feature extraction, but also to generate a pseudo feedback signal. The pseudo feedback signal is generated according to the error between the output feature and the target feature. The process of forward propagation is expressed as:

[0106]

[0107] In the formula, For the The output of the layer; For the The output of the layer; is the ReLU activation function.

[0108] Furthermore, the pseudo feedback signal affects the output of the neural network. The calculation of the change in layer output is expressed as:

[0109]

[0110]

[0111] In the formula, The pseudo feedback caused Changes in layer outputs; represents a false feedback signal; is the pseudo feedback signal learning rate; is the output layer weight of the neural network, express The transpose of is the neural network output; is the target output; is the derivative of the activation function; It is The input of the layer; represents the element-wise multiplication operation. Preferably, Set to 0.01.

[0112] S203, dynamically adjust the network structure according to the strength of the pseudo feedback signal and the current learning efficiency of the model, and adjust the adaptive structure of the neural network by increasing or decreasing the number of neurons, which is expressed as:

[0113]

[0114] In the formula, Adjust the rate for the structure; For the The activity measurement of the layer, i.e., the standard deviation of the output, characterizes the dispersion of the output of the layer; The neural network The number of neurons in the layer; is a symbolic function; is the activity measurement threshold, and neurons are added when this value is exceeded. Set to 0.01.

[0115] Further, The calculation method of the layer activity measurement is expressed as:

[0116]

[0117] In the formula, The neural network is The average of the layer outputs; The neural network is Tier The output of a neuron.

[0118] S204, using the gradient descent method to update the weights and biases of the neural network to minimize the error of the output layer. Specifically, the updating method of the weights and biases of the neural network is expressed as:

[0119]

[0120]

[0121] In the formula, is the learning rate of the neural network; is the cross entropy loss function used to supervise neural network training; For the parameters The gradient of For the parameters Preferably, Set to 0.01, Represents a parameter update operation.

[0122] S205, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0123] Based on this, the deep pseudo-feedback based neural network adaptively adjusts its structure, optimizes the network's responsiveness to specific data characteristics, and improves the accuracy of feature extraction and the training efficiency of the model.

[0124] S3, using an autoencoder neural network based on a photovoltaic device condition vector to perform feature dimension reduction on the first feature to obtain a second feature;

[0125] In an optional embodiment of the present invention, the data after feature extraction is input into a feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention uses an autoencoder neural network based on a photovoltaic device conditional vector to perform feature dimensionality reduction. On the basis of a traditional autoencoder neural network structure, a photovoltaic device conditional vector encoding mechanism and sparsity constraints are used to improve the feature abstraction capability and generalization performance of the model.

[0126] Specifically, the training process of the autoencoder neural network algorithm based on the photovoltaic device condition vector is as follows:

[0127] S301, according to the dimension of the input feature extracted data and the expected encoding dimension, the parameters of the autoencoder neural network are randomly initialized, and the weight of the autoencoder neural network is set to , the bias of the autoencoder neural network is , the photovoltaic device condition vector is , the photovoltaic equipment condition vector is predefined by the expert system according to the operation data of the photovoltaic equipment, and the initialization method is expressed as:

[0128]

[0129]

[0130] In the formula, and Represent the dimensions of the encoder input and output of the autoencoder neural network, Represents the generation of standard evenly distributed random numbers.

[0131] S302, assuming that the feature vector obtained by feature extraction is The input vector of the autoencoder neural network is the feature vector obtained by feature extraction and the feature vector of the photovoltaic device condition. The encoder processes the input through the activation function to generate the encoding representation , the calculation method is expressed as:

[0132]

[0133] In the formula, and are the weight and bias of the encoder respectively, represents the vector concatenation operation, is the ReLU activation function, The encoded representation generated for the encoder.

[0134] In one embodiment, the photovoltaic device condition vector includes parameters directly related to the operation of the photovoltaic device, for example, the parameters in the condition vector include:

[0135] a-device temperature parameter, which indicates the temperature of the device during operation, which may affect its performance and failure probability. In this embodiment, the value of this parameter may be 35;

[0136] b-voltage level, indicating the voltage level required for the device to operate. In this embodiment, the value of this parameter may be 380;

[0137] c-load percentage, which indicates the ratio of the current load of the device to the maximum load capacity. In this embodiment, the value of this parameter may be 0.75;

[0138] d-running time, indicating the running time of the device from startup to the current moment. In this embodiment, the value of this parameter may be 400;

[0139] e-maintenance frequency, which indicates the number of times the equipment is maintained within a specified time. In this embodiment, the value of this parameter may be 10.

[0140] S303, the encoded features are sent to the decoder, and the decoder attempts to reconstruct the original input. Let the reconstruction vector output by the decoder be , the calculation method is expressed as:

[0141]

[0142] In the formula, and are the weight and bias of the decoder respectively, It is an activation function that combines linear and nonlinear responses and is used to normalize the output values.

[0143] Furthermore, the activation function combining linear and nonlinear responses can enhance the model's ability to learn the nonlinear characteristics of the input data. for , then the calculation method of the activation function combining linear and nonlinear responses is expressed as:

[0144]

[0145] In the formula, is the nonlinear influencing factor, is a linear influence factor, and both the nonlinear influence factor and the linear influence factor are trainable parameters. is the hyperbolic tangent function.

[0146] S304, use the loss function to evaluate the error between the reconstructed vector and the original input, and perform sparsity constraints at the same time, set for ,Right now, is the input of the encoder of the autoencoder neural network, then the calculation method of the loss function is expressed as:

[0147]

[0148] In the formula, is the loss function of the autoencoder neural network, is the number of training samples input in the current batch, It is The input vector of the encoder of the autoencoder neural network, It is The reconstruction vector output by the decoder; is the sparsity penalty coefficient; It is the Kullback-Leibler divergence calculation function, which is used to measure the difference between the actual activation rate and the target activation rate; is the number of neurons in the hidden layer of the autoencoder neural network, which is equal to the number of neurons in the hidden layer of the encoder and the number of neurons in the hidden layer of the decoder; is the desired sparse activation rate, is the actual sparse activation rate.

[0149] Furthermore, the calculation of Kullback-Leibler divergence can more accurately measure and punish the activity of neurons, and promote the model to learn more useful and sparse feature representations. The calculation method is expressed as:

[0150]

[0151] Where, the expected sparse activation rate It is a preset parameter, which indicates the average activation rate of the hidden layer neurons in the model. It is usually set to a value less than 1. For example, if it is set to 0.05, it means that the vast majority of neurons in the hidden layer are expected to remain inactive under a given input, and only a few neurons are activated, which helps the network learn more important and meaningful features and improve its generalization ability. The actual sparse activation rate It is the actual average activation rate of neurons in the hidden layer during the actual training process. Under the current input batch of training data and network parameters, the average output of each hidden neuron for all input samples (that is, the output after the activation function) is calculated, which is the actual activation rate of the neuron.

[0152] S305. During the iterative training process, let the momentum factor of the autoencoder neural network be , the adaptive learning rate of the autoencoder neural network is , then the parameter update method is expressed as:

[0153]

[0154]

[0155] In the formula, Indicates a parameter update operation. is the learning rate of the autoencoder neural network, and They are the loss functions of the autoencoder neural network. Weight and bias The gradient of is the momentum coefficient of the autoencoder neural network, is the weight change of the encoding neural network in the previous iteration, is the bias change of the encoding neural network in the last iteration. Preferably, Set to 0.01, Set to 0.2, Represents a parameter update operation.

[0156] S306, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0157] Based on this, the autoencoder neural network retains more key feature information through conditional vectors and sparsity constraints during the feature dimensionality reduction process, thereby improving the interpretability of the data and the applicability of the model.

[0158] S4, using a support vector machine algorithm based on error correction learning to build a classifier model, and using the second feature to train the classifier model;

[0159] In an optional embodiment of the present invention, the reduced-dimensional data is input into a classifier to train the classifier model. The present invention adopts a support vector machine algorithm based on error correction learning as the classifier model. On the basis of the support vector machine kernel function, a constraint-driven learning mechanism is adopted to enhance the generalization ability and accuracy of the model by utilizing additional constraints and optimization objectives.

[0160] Specifically, the training process of the support vector machine algorithm based on error correction learning is as follows:

[0161] S501, initialize the parameters of the support vector machine, set the initial kernel function and penalty parameters, and initialize the bias parameters of the support vector machine, expressed as:

[0162]

[0163]

[0164]

[0165] In the formula, is the kernel function of the support vector machine; For the support vector machine The input vector, that is, The feature vector after dimensionality reduction; For the support vector machine The input vector, that is, The feature vector after dimensionality reduction; is the width parameter of the kernel function, is the penalty parameter, is the total number of samples, is the average value of all samples, is the preset minimum penalty threshold, is the initial bias; is the maximum value function.

[0166] S502, automatically adjusting the hardness of the classification boundary according to the performance of the current model on the training set, wherein the training set is a data set obtained from the data set after dimension reduction in whole or in accordance with a preset ratio. When dynamically adjusting the hardness of the classification boundary, the present invention incrementally adjusts the penalty parameter and the width parameter to adapt to the distribution characteristics of the training data, and the adjustment method is expressed as follows:

[0167]

[0168]

[0169]

[0170] In the formula, is the penalty parameter increment of the support vector machine; represents the loss function of the support vector machine, specifically adopting the cross entropy loss function used for supervising the training of the support vector machine; For the The true label of the data, Support vector machine for The prediction results of the data, is the number of samples in the input batch; is the first adjustment coefficient of the support vector machine, is the second adjustment coefficient of the support vector machine, is the target loss threshold of the support vector machine; is the width parameter of the kernel function, For the The width parameter of the kernel function for each iteration, For the The width parameter of the kernel function for each iteration; is the penalty parameter, For the The penalty parameter for the iteration, For the The penalty parameter for the iterations. Preferably, Set to 0.3, Set to 0.9.

[0171] S503, using error correction learning to identify misclassified cases in the training data, especially data points near the classification boundary, and applying additional learning weights to the data points near the classification boundary, forcing the model to focus on optimizing these difficult-to-classify cases. Specifically, the present invention uses error correction items For each sample, an error correction learning weight is added to focus on optimizing the misclassified samples, which can be expressed as:

[0172]

[0173]

[0174] In the formula, For the The error correction items of samples, For the The iteration The error correction items of samples, For the The iteration Error correction items for samples; is the error correction learning rate; is the model weight of the support vector machine, For the The model weights of the support vector machine for the iteration, For the The model weights of the support vector machine of the iteration; is the learning rate of error correction learning, represents the gradient; is the total number of samples. Preferably, Set to 0.03.

[0175] S504. According to the characteristic properties of the current training data, dynamically adjust or select the most suitable kernel function to better match the data structure, improve the adaptability and accuracy of the model, and the adjustment method of the function is expressed as:

[0176]

[0177] In the formula, is the kernel function of the support vector machine, is the sum function of the adjusted support vector machine, that is, The kernel function of the support vector machine of the iteration; represents the L2 norm; is the current dataset after feature dimensionality reduction The structural density function of represents the structure density function, is the dataset after the current feature dimension reduction, that is, The dataset after the feature dimension reduction of the iteration, is the structural density function influencing factor. Preferably, Set to 0.3.

[0178] In one embodiment, the structural density function is calculated based on the mean and standard deviation of the current feature-reduced data set, which provides a measure of the complexity of data distribution. The calculation method is expressed as:

[0179]

[0180] In the formula, is the sample in the dataset after the current feature dimension reduction; and They are respectively the mean and standard deviation of the current dataset after feature dimensionality reduction.

[0181] S505, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 500 times.

[0182] Based on this, the support vector machine algorithm based on error correction learning significantly improves the classification accuracy and generalization ability of the model by focusing on optimizing data points near the classification boundary.

[0183] S5. Use the trained classifier model to perform photovoltaic equipment fault diagnosis.

[0184] In an optional embodiment of the present invention, a trained model is used to diagnose photovoltaic equipment faults. In one embodiment, the collected raw data is input into a trained feature extraction and feature dimension reduction model for feature processing. Further, the processed features are input into a classifier model for classifier training to obtain classification results. In this embodiment, the classification categories include: photovoltaic panel damage, inverter failure, cable connection problem, photovoltaic bracket problem, monitoring system failure, and battery storage system problem.

[0185] Furthermore, according to the classified failures, different failure causes are analyzed, including:

[0186] The causes of photovoltaic panel damage may be due to physical impact, environmental factors (such as high temperature, humidity) or material aging; the causes of inverter failure may be due to overload, overheating, aging or failure of electronic components; the causes of cable connection problems may be aging, corrosion or poor contact of the cables due to long-term exposure to harsh environments; the causes of photovoltaic bracket problems may be problems with the brackets due to weathering, corrosion or improper installation; the causes of monitoring system failures may be due to software problems, hardware failures or communication interruptions; the causes of battery storage system problems may be damage to the battery due to overcharging, over-discharging, high temperature or battery aging.

[0187] Furthermore, different maintenance suggestions are proposed according to different fault causes, including:

[0188] If the photovoltaic panel is damaged, check whether the photovoltaic panel is cracked or damaged, and replace the damaged photovoltaic panel if necessary; if the inverter fails, check the input and output voltage and current of the inverter to confirm whether there is overload or short circuit. If the internal components of the inverter are damaged, professional technicians may be required to repair or replace them; if there is a problem with the cable connection, check whether all cable connections are firm, remove corrosion, and replace damaged cables if necessary; if there is a problem with the photovoltaic bracket, check whether the bracket is stable, repair or replace damaged bracket parts; if the monitoring system fails, check the hardware connection and software settings of the monitoring system to ensure normal data transmission. If necessary, update the software or replace the hardware; if there is a problem with the battery storage system, check the battery charging and discharging status, monitor the battery temperature, and replace the battery if necessary.

[0189] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0190] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1A function specified in one or more boxes.

[0191] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0192] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0193] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A photovoltaic equipment fault diagnosis method based on big data and deep learning, characterized in that: The following steps are involved: Obtaining operation log data, sensor data, and maintenance record data of the photovoltaic device, wherein obtaining the operation log data, sensor data, and maintenance record data of the photovoltaic device further includes: The obtained data is expanded by using a generative adversarial network based on deep manifold learning; the data expansion is performed on the obtained data by using a generative adversarial network based on deep manifold learning, including: initializing the parameters of the generative adversarial network; assuming that the generator is , the discriminator is , the cluster center is , the number of network layers of the generator is , then the parameter initialization method of the generator and discriminator is expressed as: In the formula, are the parameters of the generator and the discriminator respectively; Indicates that the parameters are initialized using a normal distribution with a mean of 0 and a variance of the unit matrix; It means that it obeys a specific distribution. represents the identity matrix, is a normal distribution; Clustering is performed based on the initial data, which is the data obtained after collection and annotation, to determine the preliminary data distribution and calculate the cluster center, which is expressed as: In the formula, Indicates cluster centers, and the total number of cluster centers is indivual; is the first randomly selected clusters; is the number of clusters, Represents the initial data; In the alternating training phase of the generator and the discriminator, the generator uses the current cluster center to try to generate pseudo data corresponding to each cluster center. In this process, the parameter update method of the generator is expressed as: In the formula, Indicates parameter update operation; represents the L2 norm; is the generator parameter; is the learning rate; represents the gradient of the generator parameters; is the discriminator function; is a generator function; It is Input noise; is the clustering loss weight, Indicates parameter update operation; Characterize the distance between the generator output and the cluster center, and the calculation method is expressed as: In the formula, To generate the characteristic dimension of data; The generator is Dimensional output; The generator is The output of the dimension is The cluster center is in Dimensional coordinates; The generated pseudo data is sent to the discriminator, which classifies each sample as true or false and updates the parameters of the discriminator to more accurately identify pseudo data and real data. In this process, the parameter update method of the discriminator is expressed as: In the formula, is the discriminator parameter; represents the gradient of the discriminator parameters; is a real data sample; is the input noise of the generator is the input noise of the generator; In each iteration, the generated data is re-clustered and the cluster center is dynamically adjusted. The dynamic adjustment method is expressed as: In the formula, For the updated Cluster centers; is the updated learning rate of the cluster center; It belongs to Generate a dataset of clusters; For the The cluster centers in the generated data set of clusters; The calculation method of the updated learning rate of the cluster center is expressed as: In the formula, Represents natural base The exponential function with base , is the standard deviation parameter that controls the update rate; Repeat the above steps until the preset stop iteration condition is met; A fully connected neural network based on deep pseudo feedback is used to extract features from the acquired data to obtain the first feature; An autoencoder neural network based on the photovoltaic device condition vector is used to reduce the dimension of the first feature to obtain the second feature; A support vector machine algorithm based on error correction learning is used to build a classifier model, and the second feature is used to train the classifier model; The trained classifier model is used to diagnose photovoltaic equipment faults.

2. A photovoltaic equipment fault diagnosis method based on big data and deep learning according to claim 1, characterized in that: The fully connected neural network based on deep pseudo feedback includes: 1 input layer, 3 hidden layers and 1 feature extraction output layer; The number of hidden layers and the number of neurons are dynamically adjusted according to the change of loss during the training process.

3. According to claim 1, a photovoltaic equipment fault diagnosis method based on big data and deep learning is characterized in that: The fully connected neural network based on deep pseudo feedback is used to extract features from the operation log data, sensor data and maintenance record data to obtain the first feature, including: Initializing parameters of a fully connected neural network; the parameters of the fully connected neural network include weights and bias parameters of the neural network; The operation log data, sensor data and maintenance record data are input into the fully connected neural network through forward propagation, and the feature vector after feature extraction is output through the feature extraction output layer. At the same time, a pseudo feedback signal is generated according to the error between the output feature and the target feature. According to the strength of the pseudo-feedback signal and the current learning efficiency of the model, the network structure is dynamically adjusted, and the adaptive structure of the fully connected neural network is adjusted by increasing or decreasing the number of neurons; Use gradient descent to update the weights and biases of the fully connected neural network to minimize the error of the feature extraction output layer; Repeat the above steps until the preset stop iteration condition is met.

4. A photovoltaic equipment fault diagnosis method based on big data and deep learning according to claim 3, characterized in that: According to the strength of the pseudo feedback signal and the current learning efficiency of the model, the network structure is dynamically adjusted, and the adaptive structure of the fully connected neural network is adjusted by increasing or decreasing the number of neurons. Specifically: in, Adjust the rate for the structure; For the Activity measurement of layers; The neural network The number of neurons in the layer; is a symbolic function; is the activity measurement threshold; Represents a parameter update operation.

5. A photovoltaic equipment fault diagnosis method based on big data and deep learning according to claim 1, characterized in that: The autoencoder neural network based on the photovoltaic device condition vector is used to reduce the dimension of the first feature to obtain the second feature, including: Randomly initialize the parameters of the autoencoder neural network according to the data dimension of the first feature and the expected encoding dimension; The encoder is used to combine the first feature with the photovoltaic device condition vector and then generate a coded representation through an activation function; Decode the encoded representation using a decoder to obtain a reconstructed vector; Use the loss function to evaluate the error between the reconstructed vector and the first feature while performing sparsity constraints; Parameters are updated according to the momentum factor of the autoencoder neural network and the adaptive learning rate of the autoencoder neural network; Repeat the above steps until the preset stop iteration condition is met.

6. A photovoltaic equipment fault diagnosis method based on big data and deep learning according to claim 5, characterized in that: The loss function is used to evaluate the error between the reconstructed vector and the first feature, while performing sparsity constraints, specifically: in, is the loss function of the autoencoder neural network, The number of training samples input for the current batch, For the The input vector of the encoder of the autoencoder neural network, For the The reconstruction vector output by the decoder; is the sparsity penalty coefficient; Calculates the Kullback-Leibler divergence function; is the number of neurons in the hidden layer of the autoencoder neural network; is the desired sparse activation rate, is the actual sparse activation rate.

7. A photovoltaic equipment fault diagnosis method based on big data and deep learning according to claim 1, characterized in that: The classifier model is constructed by using the support vector machine algorithm based on error correction learning, and the classifier model is trained using the second feature, including: Initialize the parameters of the support vector machine, set the initial kernel function and penalty parameters, and initialize the bias parameters of the support vector machine; According to the prediction results of the current model on the training set, the hardness of the classification boundary is dynamically adjusted by incrementally adjusting the penalty parameter and the width parameter; Error correction learning is used to identify misclassified samples in the training data, and additional learning weights are applied to data points near the classification boundary; According to the characteristic properties of the current training data, the kernel function is dynamically adjusted or selected to match the training data structure; Repeat the above steps until the preset stop iteration condition is met.

8. A photovoltaic equipment fault diagnosis method based on big data and deep learning according to claim 7, characterized in that: According to the prediction results of the current model on the training set, the hardness of the classification boundary is dynamically adjusted by incrementally adjusting the penalty parameter and width parameter, specifically: , , in, is the penalty parameter increment of the support vector machine; is the loss function of the support vector machine; For the The true label of the data, Support vector machine for The prediction results of the data, is the number of samples in the input batch; is the first adjustment coefficient of the support vector machine, is the second adjustment coefficient of the support vector machine, is the target loss threshold of the support vector machine; For the The width parameter of the kernel function for each iteration, For the The width parameter of the kernel function for each iteration; For the The penalty parameter for the iteration, For the The penalty parameter for the iteration.

Images