A Fault Diagnosis Method for Photovoltaic Panels Based on Multi-Source Data Fusion

Abstract

The present invention relates to the technical field of photovoltaic panel fault diagnosis, and discloses a photovoltaic panel fault diagnosis method based on multi-source data fusion. Data collection and annotation are carried out on the operation monitoring system of the photovoltaic panel and on-site detection instruments, and then data augmentation is performed. The feature extraction model, feature dimensionality reduction model, and classifier model are trained in sequence. Finally, the collected data is input into the model for diagnosis. This method collects and annotates the operation data of the photovoltaic panel, performs data augmentation, enhances the sample quality through disentanglement processing, uses a neural network based on adaptive fluctuation optimization for feature extraction, combines an autoencoder neural network of manifold mapping learning to achieve feature dimensionality reduction, and strengthens the classification accuracy through a high-order neural network. The present invention not only significantly improves the accuracy and generalization ability of fault diagnosis, but also optimizes the stability and efficiency of model training, effectively deals with complex data patterns, and ensures the stable operation of the photovoltaic panel in a changing environment.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic panel fault diagnosis, and specifically to a photovoltaic panel fault diagnosis method based on multi-source data fusion. Background Art

[0002] With the rapid development of renewable energy, the stability and efficiency of photovoltaic power generation systems have become important research topics. As the core component of a photovoltaic power generation system, the reliability of the performance of a photovoltaic panel directly affects the operation effect of the entire power generation system. However, during operation, a photovoltaic panel is easily affected by various factors, such as environmental factors and material aging, which may lead to a decrease in efficiency or a failure. Therefore, accurately and quickly diagnosing the operating status and potential faults of a photovoltaic panel is of great significance for ensuring the stable operation of the system and improving power generation efficiency.

[0003] The Chinese invention patent with the application number CN202011063756.2 proposes a method for diagnosing DC-side faults in a household photovoltaic system, which relates to the field of power grid operation and maintenance. Most of the DC-side fault diagnoses in household photovoltaic systems are based on experimental conditions, with too high requirements for the dimension of data. Blindly pursuing accuracy ignores the applicable scenarios. The present invention first finds the three normal users closest to the faulty user, calculates the photo-generated current value according to the single-diode model formula of the photovoltaic module, then calculates the power loss time series of the faulty user within a single day, and then the least squares method can be used to fit the power loss. Combining with the output characteristics of the photovoltaic array, the fault diagnosis result can be obtained according to the fitting result. This technical solution only relies on the readings of the smart meters installed by users and the installation information to complete the analysis and fault diagnosis of the household photovoltaic status in a region, effectively reducing the losses caused by photovoltaic faults that users cannot detect.

[0004] The Chinese invention patent with the application number CN202410580406.5 proposes a method for detecting faults in photovoltaic modules, belonging to the field of photovoltaic power generation, including: establishing a photovoltaic module database; performing data preprocessing on the photovoltaic module database, and the data preprocessing includes normalization processing, denoising by the singular value algorithm, phase space reconstruction, and time series reconstruction; configuring an evaluation function for the photovoltaic module to be monitored; optimizing the parameters of the evaluation function with the preprocessed photovoltaic module database, and completing the configuration of the evaluation fitness function according to the optimization result; generating a first fault monitoring result and a second fault monitoring result, performing two-way trusted authentication, and reporting a fault anomaly according to the two-way trusted authentication result. This application solves the technical problem that the existing fault detection of photovoltaic modules relies on the operating parameters of photovoltaic modules and lacks trusted authentication, resulting in low accuracy of fault detection. It achieves the technical effect of combining data analysis with image recognition through a two-way trusted authentication mechanism to improve the accuracy of fault detection.

[0005] Based on the existing technology, the following deficiencies exist: Conventional data augmentation methods may not effectively handle the dependencies between data features, resulting in low-quality generated samples and affecting the generalization ability of the model. In the neural network feature extraction methods in the existing technology, there may be problems of oscillation and local minimum in the parameter update process, reducing the training efficiency and the stability of the model. Traditional feature dimensionality reduction methods may not retain sufficient key information during the dimensionality reduction process, affecting the accuracy and reliability of the model in practical applications. Existing classifier models may face problems of gradient explosion or training oscillation during the weight update process, limiting the training speed and the final classification accuracy. Summary of the Invention

[0006] (1) Technical problems to be solved

[0007] In view of the deficiencies of the existing technology, the present invention provides a photovoltaic panel fault diagnosis method based on multi-source data fusion, which has the advantages of accurate photovoltaic panel fault diagnosis, etc., and solves the above technical problems.

[0008] (2) Technical solutions

[0009] To achieve the above object, the present invention provides the following technical solutions: A photovoltaic panel fault diagnosis method based on multi-source data fusion, comprising the following steps:

[0010] S1. Data collection and annotation: Collect data from the operation monitoring system and on-site detection instruments of the photovoltaic panel, and store the collected data in accordance with the structured data standard. The data includes voltage value a 1 、current value a 2 、surface temperature a 3 、light intensity a 4 、ambient humidity a 5 、wind speed a 6 、dust deposition amount a 7 、service life a 8 、maintenance history record a 9 、error code a 10 , and manually annotate the collected data;

[0011] S2. Data augmentation: Generate samples based on the SMOTE algorithm based on the non-locality of quantum information;

[0012] S3. Input the augmented data into the feature extraction model for training of the feature extraction model. The feature extraction model uses a neural network feature extraction algorithm based on adaptive fluctuation optimization to extract features;

[0013] S4. Input the data after feature extraction into the feature dimensionality reduction model for training of the feature dimensionality reduction model. The feature dimensionality reduction model uses an autoencoder neural network based on manifold mapping learning for feature dimensionality reduction;

[0014] S5. Input the data after dimensionality reduction into the classifier model for training the classifier model. The classifier model uses a high-order neural network based on gradient accumulation;

[0015] S6. Photovoltaic panel data fault diagnosis: Input the photovoltaic panel data to be detected into the feature extraction model for feature extraction, perform data dimensionality reduction through the data dimensionality reduction model, and judge the fault type through the classifier model.

[0016] As a preferred technical solution of the present invention, the step of generating samples by the SMOTE algorithm based on the non-locality of quantum information in step S2 includes the following steps:

[0017] S2.1. Initialize the algorithm parameters, including the initial superposition angle θ of the quantum state b , entanglement strength ∈ b and the initialization of the adaptively adjusted learning rate λ b . The specific expressions are as follows:

[0018]

[0019] ∈ b = 0.1

[0020] λ b = 0.01

[0021] Among them, θ b represents the superposition state angle of the quantum bit, ∈ b represents the entanglement degree between quantum bits, and λ b represents the learning rate;

[0022] S2.2. Use a non-linear transformation to perform disentanglement processing on the features in the original dataset. The specific expression is as follows:

[0023] zb = f b (x; Θ b ) = tanh(Θ b ·x)

[0024] Among them, x is the input original feature vector, zb represents the disentangled feature vector, Θ b is the weight matrix of the neural network, tanh(*) is the hyperbolic tangent function, and f b (x; Θ b ) represents the non-linear transformation function;

[0025] S2.3. By simulating the behavior of quantum superposition and entanglement, generate new sample points y in the data space. The generation method of the new sample is expressed as:

[0026] y = x + γb ·(σ b ·(zb - x))

[0027] In the formula, y represents the newly generated sample, γ b represents a parameter that controls the degree of deviation of the new sample from the original sample, σ b is a random variable adjusted based on the entanglement degree of the quantum state, x is the input original feature vector, and zb represents the disentangled feature vector;

[0028] S2.4. Automatically adjust the algorithm parameters θ b , ∈ b and λ b after each iteration to optimize the subsequent sample generation process, which is expressed as:

[0029]

[0030] In the formula, Δθ b is the adjustment amount of the initial superposition state angle of the qubit, Δ∈ b is the adjustment amount of the entanglement strength, is the updated initial superposition state angle of the qubit, is the updated entanglement strength, λ b is the adjustment learning rate, and Δθ b and Δ∈ b are expressed as follows:

[0031]

[0032] Among them, η cv is the loss learning rate parameter, and are the partial derivatives of the loss function L with respect to θ b and ∈ b respectively, Δθ b is the adjustment amount of the initial superposition state angle of the qubit, Δ∈ b is the adjustment amount of the entanglement strength;

[0033] S2.5. Repeat the above steps iteratively until the preset maximum number of iterations is reached.

[0034] As a preferred technical solution of the present invention, the random variable σ b adjusted based on the entanglement degree of the quantum state in step S2.3 b and the superposition angle θ b is dynamically adjusted, and the specific expression is as follows:

[0035] σ b = cos(θ b )·∈ b2

[0036] Among them, cos(θ b ) represents the influence of the superposition state on the sample generation position, ∈ b 2 represents the weight of the entanglement degree in the generated sample, and σ b is a random variable adjusted based on the quantum state entanglement degree.

[0037] As a preferred technical solution of the present invention, the process of training based on the neural network feature extraction algorithm optimized by adaptive fluctuation in step S3 is as follows:

[0038] S3.1. Define the neural network structure. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, and the second hidden layer includes 50 neurons. The activation function is the ReLU activation function. At the same time, set the parameters of the neural network, including the initial parameters θ 0 , the position of the wave source S i , the intensity a i and the initial phase φ i . The initial parameters θ 0 of the neural network include the weights w gr and the biases b gr ;

[0039] S3.2. Calculate the loss L GR through forward propagation. Use the neural network parameters θ to perform forward propagation on the input X GR , calculate the output Y and compare it with the target value T. The expression for calculating the loss is:

[0040] Y = frs(X GR ; θ)

[0041] L GR = L GRoss(Y,T)

[0042] Among them, frs(*) represents the network structure, and the output layer of this network structure uses a preset Softmax_gr classifier to obtain the prediction label. L GRoss(Y,T) represents the cross-entropy loss;

[0043] S3.3. Calculate the gradient of each parameter θ according to the loss function which can be expressed as:

[0044]

[0045] Among them, represents the gradient of the loss function L GR with respect to the parameter θ, and respectively represent the loss function L GR for the weights w of the neural network gr and the biases b of the neural network gr partial derivatives;

[0046] S3.4. Generate wave sources at the corresponding parameter positions according to the direction and magnitude of the gradient. At the (t + 1)-th iteration, the intensity of each wave source S i is updated to The calculation method can be expressed as:

[0047]

[0048] In the formula, represents the loss function L GR the gradient of the loss function L with respect to the i-th parameter, α ck is the learning rate, represents the gradient of the i-th parameter, is the attenuation coefficient, represents the intensity of each wave source S at the t-th iteration i of, is the amplification coefficient, L GR represents the loss function;

[0049] S3.5. The fluctuations of all wave sources are superimposed in the parameter space. The update amount of each parameter is determined by the superimposed result of the fluctuations of all wave sources at that point. The specific expression is as follows:

[0050]

[0051] where k is the wave number, is the neural network parameter at the (t + 1)-th iteration, represents the neural network parameter at the t-th iteration, ω is the angular frequency, representing the propagation speed and direction of the wave, Re() is the ReLU activation function, L gr represents the cross-entropy loss function, represents that the loss function is L gr the minimum fluctuation threshold in the historical iteration times of the i-th parameter when the loss function is L, | | is the absolute value symbol, ng represents the total number of wave sources, respectively represent the intensity of the j-th wave source at the (t + 1)-th iteration, is that the loss function is L gr the fluctuation threshold of the i-th parameter when, and its expression is as follows:

[0052]

[0053] where τ 0 is the initial fluctuation threshold, is for L grThe layer sensitivity index of the layer, κ pq is the adjustment coefficient, e represents the natural constant, where The expression of

[0054]

[0055] where, represents the L2 norm, represents the loss function L GR The gradient of L with respect to the k-th parameter, L GRc represents the total number of layers of the network; represents the loss function L GR with respect to the parameter of the L-th GRc layer; represents the parameter of the L-th GRc layer; represents the loss function L GR The gradient of L with respect to the parameter of the k-th layer; represents the parameter of the k-th layer;

[0056] S3.6. Dynamically adjust the attenuation coefficient β and the amplification coefficient γ of each wave source, and the specific expressions are as follows:

[0057]

[0058] In the formula, are the attenuation coefficient and the amplification coefficient of the previous iteration respectively, are the attenuation coefficient adjustment factor and the amplification coefficient adjustment factor respectively, represent the attenuation coefficient and the amplification coefficient respectively, ΔL GR is the change in loss, e represents the natural constant;

[0059] S3.7. Determine whether the preset maximum number of iterations is reached. If not, jump to step S3.2. If so, stop the iteration and complete the training.

[0060] As a preferred technical solution of the present invention, the training of the autoencoder neural network algorithm based on manifold mapping learning in step S4 includes the following steps:

[0061] S4.1. Initialize the weights w cr and biases b cr parameters of the autoencoder, and the specific expressions are as follows:

[0062]

[0063] where, represents the weight of the i-th layer, σ ch represents the standard deviation, n ”in " is the number of nodes in the input layer;

[0064] S4.2. During the encoding process, the data x cr is mapped to a low-dimensional feature space through an encoder; during the decoding process, the original data is reconstructed through a decoder, and the specific expression is as follows:

[0065] The data x cr is mapped to the low-dimensional feature space zr:

[0066]

[0067] The decoder reconstructs the original input x cr ^:

[0068]

[0069] where Sig( ) is the activation function; and are the weights and biases of the encoder respectively; and are the weights and biases of the decoder;

[0070] S4.3. The Riemannian factor is used to implement manifold learning, and the expression for the Riemannian factor R that constrains the encoded features is as follows:

[0071]

[0072] where μ cy represents the preset manifold center, σ R is a parameter that controls the compactness of the manifold, exp represents the exponential function with the natural constant as the base, and zr represents the low-dimensional feature space;

[0073] S4.4. Construct the loss function Lu, and the specific expression is as follows:

[0074] Lu = ||x cr - x cr ^|| 2 + λ cm R

[0075] where λ cm is the regularization coefficient, x cr represents the data before encoding, x cr ^ represents the data after decoding, and R represents the Riemannian factor;

[0076] S4.5. Adjust the network parameters through the backpropagation algorithm;

[0077] S4.6. Determine whether the preset maximum number of iterations is reached. If not, jump to step S4.2; if so, stop the iteration and complete the training.

[0078] As a preferred technical solution of the present invention, the training of the high-order neural network algorithm based on gradient accumulation in step S5 specifically includes the following steps:

[0079] S5.1. Initialize the parameters of the high-order neural network: Initialize the high-order neural network parameters W q and the bias b q ;

[0080] S5.2. Use the data after feature dimensionality reduction as the input feature vector and process it through the network layer. The specific expression is as follows:

[0081] z q = Sig(W q ·x q +b q )

[0082] where x q is the input feature vector, z q is the output of this layer, W q represents the high-order neural network parameters, and Sig() is the Sigmoid activation function;

[0083] S5.3. Use the loss function Lg to evaluate the error between the output of the classifier model and the actual label. The specific expression is as follows:

[0084]

[0085] where y true,i is the one-hot encoding of the true label, y pred,i is the probability output predicted by the classifier model, log(*) represents the logarithmic function, represents the summation of N data;

[0086] S5.4. Calculate the gradients of the high-order neural network parameters and the bias according to the loss function L and The specific expressions are as follows:

[0087]

[0088] where is the gradient of the loss function Lg with respect to the output; and are the gradients of the output with respect to the weight and the bias;

[0089] S5.5. Accumulate the calculated gradients until the preset training cycle is reached. The cumulative gradients of the high-order neural network parameters and the bias and are expressed as follows:

[0090]

[0091] Among them, represents the cumulative gradient for each iteration represents the cumulative gradient for each iteration

[0092] S5.6. After the iteration in step S5.5 is completed, perform clipping processing on the accumulated gradient to limit the maximum value of the gradient. The specific expression is as follows:

[0093]

[0094] Among them, clip(*, -τ q , τ q ) represents the operation of clipping the gradient to the range [-τ q , τ q . τ q represents the gradient clipping threshold;

[0095] S5.7. Determine whether the preset maximum number of iterations is reached. If not, jump to step S5.3. If so, stop the iteration and complete the training.

[0096] Compared with the prior art, the present invention provides a photovoltaic panel fault diagnosis method based on multi-source data fusion, which has the following beneficial effects:

[0097] 1. By using quantum information technology to disentangle data features, the present invention generates high-quality training samples, significantly improving the generalization ability of the model and the processing ability for complex data patterns. The adaptive fluctuation optimization algorithm ensures the smoothness and self-adaptability of network parameter updates, significantly improving the stability and training efficiency of the model.

[0098] 2. Through effective encoding and decoding processes, the present invention ensures the retention of important features, enhancing the accuracy and reliability of the model in practical applications. The gradient accumulation strategy of the high-order neural network significantly reduces the oscillation during the training process, improving the convergence speed and classification accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 is a schematic flowchart of the present invention;

[0100] Figure 2 is a schematic flowchart of the training process of the feature extraction model of the present invention;

[0101] Figure 3 is a schematic flowchart of the training process of the feature dimensionality reduction model of the present invention;

[0102] Figure 4 is a schematic flowchart of the training and application process of the present invention. Specific Embodiments

[0103] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0104] Please refer to Figure 1 - Figure 4 , a photovoltaic panel fault diagnosis method based on multi-source data fusion, comprising the following steps:

[0105] S1. Data collection and annotation to form an original data set, collected from the operation monitoring system of the photovoltaic panel and on-site detection instruments, and the collected data is stored in accordance with the structured data standard. The data includes voltage value a 1 , current value a 2 , surface temperature a 3 , light intensity a 4 , environmental humidity a 5 , wind speed a 6 , dust deposition amount a 7 , service life a 8 , maintenance history record a 9 , error code a 10 , and the collected data is manually annotated, and the annotated categories include "normal operation", "efficiency decline", "local fault", and "overall fault";

[0106] S2. Data augmentation, generating samples based on the SMOTE algorithm of quantum information non-locality;

[0107] The step of generating samples based on the SMOTE algorithm of quantum information non-locality in step S2 includes the following steps:

[0108] S2.1. Initialize the algorithm parameters, including the initial superposition angle θ b of the quantum state, entanglement strength ∈ b and the initialization of the adaptively adjusted learning rate λ b , and the specific expressions are as follows:

[0109]

[0110] ∈ b = 0.1

[0111] λ b = 0.01

[0112] Among them, θ bRepresents the superposition state angle of qubits, ∈ b Represents the entanglement degree between qubits, λ b Represents the learning rate;

[0113] S2.2. Use non - linear transformation to perform disentanglement processing on the features in the original dataset. The specific expression is as follows:

[0114] zb = f b (x; Θ b ) = tanh(Θ b ·x)

[0115] Among them, x is the input original feature vector, zb represents the disentangled feature vector, Θ b is the weight matrix of the neural network, tanh(*) is the hyperbolic tangent function, and f b (x; Θ b ) represents the non - linear transformation function;

[0116] S2.3. By simulating the behaviors of quantum superposition and entanglement, generate new sample points y in the data space. The generation method of the new sample is expressed as:

[0117] y = x + γ b ·(σ b ·(z - x))

[0118] In the formula, y represents the generated new sample, γ b represents the parameter that controls the degree of deviation of the new sample from the original sample, σ b is a random variable adjusted based on the entanglement degree of the quantum state, x is the input original feature vector, zb represents the disentangled feature vector, and γ b is set to 3;

[0119] The random variable σ adjusted based on the entanglement degree of the quantum state b is dynamically adjusted according to the entanglement degree ∈ b and the superposition angle θ b The specific expression is as follows:

[0120] σ b = cos(θ b )·∈ b 2

[0121] Among them, cos(θ b ) represents the influence of the superposition state on the generation position of the sample, ∈ b 2 represents the weight of the entanglement degree in the generated sample, and σ b is a random variable adjusted based on the entanglement degree of the quantum state;

[0122] S2.4. Automatically adjust the algorithm parameters θ after each iteration b , ∈ b and λ b , to optimize the subsequent sample generation process, expressed as:

[0123]

[0124] In the formula, Δθ b is the adjustment amount of the initial superposition state angle of the qubit, Δ∈ b is the adjustment amount of the entanglement strength, is the initial superposition state angle of the updated qubit, is the updated entanglement strength, λ b is the adjustment learning rate, λ b is set to 0.01, and Δθ b and Δ∈ b The expressions are as follows:

[0125]

[0126] Among them, η cv is the loss learning rate parameter, and are the partial derivatives of the loss function L with respect to θ b and ∈ b respectively, Δθ b is the adjustment amount of the initial superposition state angle of the qubit, η cv is set to 0.03, Δ∈ b is the adjustment amount of the entanglement strength;

[0127] S2.5. Repeat the above steps iteratively until the preset maximum number of iterations is reached. The maximum number of iterations is set to 1000 times;

[0128] S3. Input the augmented data into the feature extraction model for training the feature extraction model. The feature extraction model uses a neural network feature extraction algorithm based on adaptive fluctuation optimization to extract features. Input the augmented data into the feature extraction model for training the feature extraction model. The present invention adopts a neural network feature extraction algorithm based on adaptive fluctuation optimization, inspired by the fluctuation phenomenon. Fluctuations exhibit the characteristics of periodic changes and energy transfer, and can transmit and transform energy forms in different media. The adaptive fluctuation optimization algorithm iteratively updates the weights and biases in the network by simulating the propagation and energy distribution characteristics of fluctuations. Each parameter update is not only affected by the current gradient but also by the cumulative influence of the fluctuations of previous parameter updates, similar to the fluctuation influence caused by the superposition of wave sources, making the parameter update process smoother and more adaptive, effectively avoiding the oscillation and local minimum problems in the traditional gradient descent method;

[0129] S3.1. Define the neural network structure. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, and the second hidden layer includes 50 neurons. The activation function is the ReLU activation function. At the same time, set the parameters of the neural network, including the initial parameters θ of the neural network 0 , the wave source S i position intensity a i and initial phase φ i , the initial parameters θ of the neural network 0 include the weights w of the neural network gr and biases b gr ;

[0130] The initialization can be expressed as:

[0131]

[0132] In the formula, is usually selected as the initial position of the parameter, a i and φ i randomly set their initial values; S i is the i-th wave source, which consists of the position intensity a i and initial phase φ i ;

[0133] S3.2. Calculate the loss L through forward propagation GR . Use the neural network parameters θ to perform forward propagation on the input X GR , calculate the output Y and compare it with the target value T. The expression for calculating the loss is:

[0134] Y = frs(X GR ; θ)

[0135] L GR = L GRoss(Y,T)

[0136] where frs(*) represents the network structure, and the output layer of this network structure uses a preset Softmax_gr classifier to obtain the predicted label, and L GRoss(Y,T) represents the cross-entropy loss;

[0137] S3.3. Calculate the gradient of each parameter θ according to the loss function can be expressed as:

[0138]

[0139] where, represents the gradient of the loss function L GR with respect to the parameter θ, and respectively represent the loss function L GR for the weights w of the neural network gr and the biases b of the neural network gr partial derivatives;

[0140] The chain rule is used to calculate the gradient. Suppose the network output Y passes through the Softmax_gr function. Then the gradient calculation method for the weights w gr can be expressed as:

[0141]

[0142] where y k is the output of Softmax, and z k is the linear output before the Softmax layer.

[0143] The calculation method of

[0144]

[0145] can be expressed as:

[0146]

[0147] The calculation method of

[0148]

[0149] S3.4. Generate wave sources at the corresponding parameter positions according to the direction and magnitude of the gradient. At the (t + 1)-th iteration, the intensity of each wave source S i is updated to The calculation method can be expressed as:

[0150]

[0151] where represents the gradient of the loss function L GR with respect to the i-th parameter, α ck is the learning rate, represents the gradient of the i-th parameter, is the attenuation coefficient, represents the intensity of each wave source S i at the t-th iteration, is the amplification coefficient, L GR represents the loss function, α ck is set to 0.01;

[0152] S3.5. The fluctuations of all wave sources are superimposed in the parameter space, and the update amount of each parameter is determined by the superimposed result of the fluctuations of all wave sources at that point. The specific expression is as follows:

[0153]

[0154] where k is the wave number, is the neural network parameter at the (t + 1)-th iteration, represents the neural network parameter at the t-th iteration, ω is the angular frequency, represents the propagation speed and direction of the wave, Re() is the ReLU activation function, represents that the loss function is L gr is the minimum fluctuation threshold in the historical iteration times of the i-th parameter when the loss function is L gr L is the cross-entropy loss function, | | is the absolute value symbol, and ng represents the total number of wave sources, respectively represent the intensities of the j-th wave source at the (t + 1)-th iteration, is that the loss function is L gr is the fluctuation threshold of the i-th parameter when the loss function is L, and its expression is as follows:

[0155]

[0156] where τ 0 is the initial fluctuation threshold, is the layer sensitivity index of the L gr -th layer, κ pq is the adjustment coefficient, e represents the natural constant, τ 0 is set to 0.1, κ pq is set to 3, where the expression of

[0157]

[0158] where ||*|| represents the L2 norm, represents the gradient of the loss function L GR with respect to the k-th parameter, L GRc represents the total number of layers of the network; represents the gradient of the loss function L GR with respect to the parameters of the L GRc -th layer; represents the parameters of the L GRc -th layer; represents the gradient of the loss function L GR with respect to the parameters of the k-th layer; represents the parameters of the k-th layer;

[0159] S3.6. Dynamically adjust the attenuation coefficient β and the amplification coefficient γ of each wave source. The specific expression is as follows:

[0160]

[0161] In the formula, are respectively the attenuation coefficient and the amplification coefficient of the previous iteration, are respectively the attenuation coefficient adjustment factor and the amplification coefficient adjustment factor, respectively represent the attenuation coefficient and the amplification coefficient, ΔL GR is the change in loss, e represents the natural constant, and the parameter and are respectively set to 0.1 and 0.05;

[0162] S3.7. Determine whether the preset maximum number of iterations is reached. If not, jump to step S3.2. If so, stop the iteration and complete the training;

[0163] S4. Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. The feature dimensionality reduction model uses an autoencoder neural network based on manifold mapping learning for feature dimensionality reduction. Input the data after feature extraction into the feature dimensionality reduction model for training the feature dimensionality reduction model. The present invention uses an autoencoder neural network based on manifold mapping learning for feature dimensionality reduction. The autoencoder neural network not only learns the internal structure of the data during the encoding stage, but also tries to reconstruct the input through the decoding process to ensure that the key information can still be retained in the reduced-dimensional features. The model structure of the autoencoder neural network consists of 1 encoder and 1 decoder. The encoder is responsible for mapping the high-dimensional input data to a low-dimensional feature space, and the decoder reconstructs the original input data from this low-dimensional space. The present invention uses the Riemannian factor based on manifold learning to optimize the representation of the feature space. By using the geometric properties on the Riemannian manifold, the local structure of the data can be better understood, and more accurate data reconstruction can be achieved;

[0164] The training of the autoencoder neural network algorithm based on manifold mapping learning in step S4 includes the following steps:

[0165] S4.1. Initialize the weights w cr and the biases b cr of the autoencoder using a Gaussian distribution. The specific expressions are as follows:

[0166]

[0167] Among them, represents the weight of the i-th layer, and σ ch represents the standard deviation. Preferably, n ”in " is the number of nodes in the input layer;

[0168] S4.2. During the encoding process, the data xcr After passing through the encoder, it is mapped to a low-dimensional feature space; during the decoding process, the original data is reconstructed by the decoder, and the specific expressions are as follows:

[0169] Data x cr is mapped to the low-dimensional feature space zr:

[0170]

[0171] The decoder reconstructs the original input x cr ^:

[0172]

[0173] where Sig() is the activation function; and are the weights and biases of the encoder respectively; and are the weights and biases of the decoder;

[0174] Weights are preset by means of principal component analysis, and the calculation method can be expressed as:

[0175]

[0176] In the formula, V k is the matrix composed of the first k eigenvectors obtained from the covariance matrix , and ∑ k is the corresponding diagonal matrix of eigenvalues;

[0177] S4.3. Implement manifold learning using the Riemannian factor. The expression for the Riemannian factor R that constrains the encoded features is as follows:

[0178]

[0179] where μ cy represents the preset manifold center, σ R is the parameter that controls the compactness of the manifold, exp represents the exponential function with the natural constant as the base, and zr represents the low-dimensional feature space;

[0180] S4.4. Construct the loss function Lu, and the specific expression is as follows:

[0181] Lu = ||x cr - x cr ^|| 2 + λ cm R

[0182] where λ cm is the regularization coefficient, R represents the Riemannian factor, and the reconstruction error ||x cr - xcr ^|| 2 The calculation method can be expressed as:

[0183]

[0184] In the formula, z i is the value of z in the i-th dimension, and μ i is the value of the preset manifold center in the i-th dimension;

[0185] S4.5. Adjust the network parameters through the backpropagation algorithm;

[0186] Taking the weight as an example, the calculation method of its gradient can be expressed as:

[0187]

[0188] Furthermore, the gradient is calculated according to the error backpropagation method and can be expressed as:

[0189]

[0190] And, Through the transpose representation of the decoder weight it can be expressed as:

[0191]

[0192] Furthermore, since z is and x cr the result obtained after a linear combination and passing through the activation function, then the calculation method can be expressed as:

[0193]

[0194] In the formula, Sig'( ) is the derivative of the Sigmoid activation function, and diag(vs) represents a diagonal matrix with the vector vs as the diagonal elements, is the transpose of x cr ;

[0195] Furthermore, use the gradient descent method to update the parameters, and the update method can be expressed as:

[0196]

[0197] In the formula, α rv is the learning rate, t represents the number of iterations, is the updated weight parameter, is the weight parameter before update. Preferably, α rv is set to 0.05;

[0198] S4.6. Determine whether the preset maximum number of iterations is reached. If not, jump to step S4.2; if so, stop the iteration and complete the training. The maximum number of iterations is set to 1000 times.

[0199] S5. Input the data after dimensionality reduction into the classifier model for training. The classifier model uses a high-order neural network based on gradient accumulation.

[0200] The specific steps of the training of the high-order neural network algorithm based on gradient accumulation in step S5 are as follows:

[0201] S5.1. Initialize the parameters of the high-order neural network: Initialize the high-order neural network parameters W q and the bias b q ;

[0202] S5.2. Process the data after feature dimensionality reduction as the input feature vector through the network layer. The specific expression is as follows:

[0203] z q = Sig(W q ·x q +b q )

[0204] where x q is the input feature vector, z q is the output of this layer, W q represents the high-order neural network parameters, and Sig() is the Sigmoid activation function;

[0205] S5.3. Use the loss function Lg to evaluate the error between the output of the classifier model and the actual label. The specific expression is as follows:

[0206]

[0207] where y true,i is the one-hot encoding of the true label, y pred,i is the probability output predicted by the classifier model, log(*) represents the logarithmic function, represents the sum over N data;

[0208] y predi is calculated as:

[0209]

[0210] In the formula, z q,i is the value of the i-th node in the output layer of the neural network;

[0211] S5.4. Calculate the gradients of the high - order neural network parameters and biases according to the loss function \(L_g\). and The specific expressions are as follows:

[0212]

[0213] where, is the gradient of the loss function \(L_g\) with respect to the output; and are the gradients of the output with respect to the weights and biases;

[0214] Furthermore, and are calculated as:

[0215]

[0216] Furthermore, the derivative of the Sigmoid function \(Sig'()\) is calculated as:

[0217] \(Sig'(z q ) = Sig(z q )\cdot(1 - Sig(z q ))

[0218] S5.5. Accumulate the calculated gradients until a preset training cycle is reached. The preset training cycle is 20 iterations. The cumulative gradients of the high - order neural network parameters and biases and are expressed as follows:

[0219]

[0220] where, represents the cumulative gradient for each iteration represents the cumulative gradient for each iteration

[0221] S5.6. After the iteration in step S5.5 is completed, perform clipping on the accumulated gradients to limit the maximum value of the gradients. The specific expression is as follows:

[0222]

[0223] where \(clip(*,-\tau q ,\tau q )\) represents the operation of clipping the gradient to the range \([-\tau q ,\tau q \), and \(\tau q \) represents the gradient clipping threshold;

[0224] Update the parameters using the accumulated gradients, expressed as:

[0225]

[0226] In the formula, ← represents the update symbol, and τ q is the gradient clipping threshold; λ q is the learning rate. Preferably, τ q is set to 0.1, and λ q is set to 0.01.

[0227] In one embodiment, the implementation manner of gradient clipping is expressed as:

[0228]

[0229] In the formula, g represents the gradient to be clipped, and τ q is the gradient clipping threshold

[0230] S5.7. Determine whether the preset maximum number of iterations is reached. If not, jump to step S5.3. If so, stop the iteration and complete the training;

[0231] S6. Photovoltaic panel data fault diagnosis: Input the photovoltaic panel data to be detected into the feature extraction model for feature extraction, perform data dimensionality reduction through the data dimensionality reduction model, and judge the fault type through the classifier model. The fault categories include "normal operation", "efficiency degradation", "local fault", and "overall fault".

[0232] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic panel fault diagnosis method based on multi-source data fusion, characterized in that: The following steps are involved: S1. Data collection and annotation constitute the original data set, which is collected from the operation monitoring system of the photovoltaic panel and the on-site detection instrument. The collected data is stored in a structured data standard. The data includes voltage value a1, current value a2, surface temperature a3, light intensity a4, ambient humidity a5, wind speed a6, dust deposition a7, service life a8, maintenance history a9, error code a1, etc. 10 , and manually annotate the collected data; S2, data expansion, sample generation based on the SMOTE algorithm of quantum information non-locality, including the following steps: S2.

1. Initialize algorithm parameters, including the initial superposition angle θ of the quantum state b , entanglement strength ∈ b and adaptively adjusted learning rate λ b The initialization of is as follows: ∈ b ＝0.1 l b =0.01 Among them, θ b represents the superposition state angle of the quantum bit, ∈ b represents the degree of entanglement between quantum bits, λ b represents the learning rate; S2.2, use nonlinear transformation to disentangle the features in the original data set collected in step S1, the specific expression is as follows: zb=f b (x;Θ b )=tanh(Θ b ·x) Where x is the original input eigenvector, zb is the eigenvector after disentanglement, Θ b is the weight matrix of the neural network, tanh() is the hyperbolic tangent function, f b (x; Θ b ) represents a nonlinear transformation function; S2.

3. By simulating the behavior of quantum superposition and entanglement, a new sample point y in the data space is generated. The generation method of the new sample is expressed as: y=x+γ b ·(s b ·(zb-x)) In the formula, y represents the generated new sample, γ b represents the parameter that controls the degree to which the new sample deviates from the original sample, σ b It is a random variable adjusted based on the degree of entanglement of the quantum state; S2.

4. Automatically adjust algorithm parameters θ after each iteration b ,∈ b and λ b , to optimize the subsequent sample generation process, expressed as: In the formula, Δθ b is the adjustment amount of the initial superposition state angle of the quantum bit, Δ∈ b is the adjustment amount of entanglement strength, is the initial superposition state angle of the updated quantum bit, is the updated entanglement strength, λ b To adjust the learning rate, and Δθ b and Δ∈ b The expression is as follows: Among them, η cv is the loss learning rate parameter, and They are the loss function L with respect to θ b and ∈ b The partial derivative of b is the adjustment amount of the initial superposition state angle of the quantum bit, Δ∈ b is the adjustment amount of entanglement strength; S2.5, repeat the above steps until the preset maximum number of iterations is reached; S3, inputting the expanded data into a feature extraction model to train the feature extraction model, wherein the feature extraction model extracts features using a neural network feature extraction algorithm based on adaptive fluctuation optimization; S4, inputting the feature-extracted data into a feature dimension reduction model to train the feature dimension reduction model, wherein the feature dimension reduction model uses an autoencoder neural network based on manifold mapping learning to perform feature dimension reduction; S5, inputting the dimension-reduced data into a classifier model to train the classifier model, wherein the classifier model adopts a high-order neural network based on gradient accumulation; S6. Photovoltaic panel data fault diagnosis: The photovoltaic panel data to be tested is input into the feature extraction model for feature extraction, the data dimension reduction is performed through the data dimension reduction model, and the fault type is judged through the classifier model.

2. A photovoltaic panel fault diagnosis method based on multi-source data fusion according to claim 1, characterized in that: The random variable σ in step S2.3 is adjusted based on the quantum state entanglement degree. b According to the entanglement degree of the quantum state ∈ b and the superposition angle θ b Dynamic adjustment, the specific expression is as follows: s b =cos(θ b )·∈ b 2 Among them, cos(θ b ) represents the influence of the superposition state on the sample generation position, ∈ b 2 represents the weight of entanglement in generating samples, σ b It is a random variable adjusted based on the degree of entanglement of the quantum state.

3. The photovoltaic panel fault diagnosis method based on multi-source data fusion according to claim 1 is characterized in that: The process of training the neural network feature extraction algorithm based on adaptive fluctuation optimization in step S3 is as follows: S3.

1. Define the neural network structure. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, and the second hidden layer includes 50 neurons. The activation function is the ReLU activation function. At the same time, set the parameters of the neural network, including the initial parameters of the neural network θ0, the wave source S i Location Strength i and the initial phase φ i , the initial parameters of the neural network θ0 include the weights w of the neural network gr and bias b gr ; S3.

2. Calculate the loss L through forward propagation GR , using the neural network parameters θ on the input X GR Perform forward propagation on the , calculate the output Y and compare it with the target value T, and the expression for calculating the loss is: Y=frs(X GR ;θ) L GR =L GRoss (Y,T) Among them, frs(*) represents the network structure, and the output layer of the network structure uses the preset Softmax_gr classifier to obtain the predicted label, L GRoss(Y,T) represents the cross entropy loss; S3.

3. Based on the loss function L GR Calculate the gradient of each parameter θ The expression is: in, Denotes the loss function L GR The gradient with respect to the parameter θ is and Respectively represent the loss function L GR The weight w of the neural network gr and the bias b of the neural network gr The partial derivative of S3.

4. Generate a wave source at the corresponding parameter position according to the direction and size of the gradient. At the t+1th iteration, each wave source S i The strength is updated to The calculation method is expressed as: In the formula, Denotes the loss function L GR The gradient of the i-th parameter, α ck is the learning rate, represents the gradient of the i-th parameter, is the attenuation coefficient, Indicates that each wave source S at the tth iteration i The strength of is the amplification factor, L GR represents the loss function; S3.

5. The fluctuations of all wave sources are superimposed in the parameter space. The update amount of each parameter is determined by the superposition result of the fluctuations of all wave sources at this point. The specific expression is as follows: Where k is the wave number, is the neural network parameter of the t+1th iteration, represents the neural network parameters of the tth iteration, ω is the angular frequency, which represents the propagation speed and direction of the wave, Re() is the ReLU activation function, and L gr represents the cross entropy loss function, Denote the loss function as L gr The minimum fluctuation threshold value in the historical iteration number of the i-th parameter at time , || is the absolute value symbol, ng represents the total number of wave sources, They represent the intensity of the jth wave source at the t+1th iteration, The loss function is L gr The fluctuation threshold of the i-th parameter at time , its expression is as follows: Among them, τ0 is the initial fluctuation threshold, For L gr Layer sensitivity index of the layer, κ pq is the adjustment coefficient, e represents the natural constant, where The expression is as follows: Among them, ||*|| represents the L2 norm, Denotes the loss function L GR The gradient of the kth parameter, L GRc Indicates the total number of layers in the network; Denotes the loss function L GR About L GRc Gradients of layer parameters; Indicates L GRc Gradients of layer parameters; Denotes the loss function L GR Gradients about the parameters of the kth layer; Represents the gradient of the k-th layer parameter; S3.

6. Dynamically adjust the attenuation coefficient β and amplification coefficient γ of each wave source. The specific expressions are as follows: In the formula, are the attenuation coefficient and amplification coefficient of the previous iteration respectively, They are the attenuation coefficient adjustment factor and the amplification coefficient adjustment factor, Respectively represent the attenuation coefficient and the amplification coefficient, ΔL GR is the change in loss, e represents a natural constant; S3.

7. Determine whether the preset maximum number of iterations has been reached. If not, jump to step S3.

2. If reached, stop iteration and complete training.

4. The photovoltaic panel fault diagnosis method based on multi-source data fusion according to claim 1 is characterized in that: The training of the autoencoder neural network algorithm based on manifold mapping learning in step S4 comprises the following steps: S4.

1. Initialize the weight w of the autoencoder using Gaussian distribution cr and bias b cr Parameters, the specific expressions are as follows: in, represents the weight of the i-th layer, σ ch represents the standard deviation, n ”in ” is the number of nodes in the input layer; S4.

2. During the encoding process, the data x cr After the encoder, it is mapped to a low-dimensional feature space; during the decoding process, the original data is reconstructed by the decoder. The specific expression is as follows: Data x cr is mapped to the low-dimensional feature space zr: The decoder reconstructs the original input x cr ^: Among them, Sig() is the activation function; and are the weights and biases of the encoder respectively; and are the weights and biases of the decoder; S4.3, Riemann factors are used to implement manifold learning. The expression of the features of the Riemann factor R constraint encoding is as follows: Among them, μ cy represents the preset manifold center, σ R It is a parameter that controls the compactness of the manifold, exp represents an exponential function with a natural constant as the base, and zr represents a low-dimensional feature space; S4.

4. Construct the loss function Lu, the specific expression is as follows: Lu=||x cr -x cr ^|| 2 +λ cm R Among them, λ cm is the regularization coefficient, x cr Represents the data before encoding, x cr ^ represents the decoded data, R represents the Riemann factor, ||x cr -x cr ^|| 2 represents the reconstruction error; S4.5, adjust network parameters through back propagation algorithm; S4.

6. Determine whether the preset maximum number of iterations has been reached. If not, jump to step S4.

2. If reached, stop iteration and complete training.

5. The photovoltaic panel fault diagnosis method based on multi-source data fusion according to claim 1 is characterized in that: The training of the high-order neural network algorithm based on gradient accumulation in step S5 specifically includes the following steps: S5.

1. Initialize the parameters of the high-order neural network: Use random initialization to initialize the high-order neural network parameters W q and bias b q ; S5.

2. The data after feature dimension reduction is used as the input feature vector and processed through the network layer. The specific expression is as follows: z q =Sig(W q ·x q +b q ) Among them, x q is the input feature vector, z q is the output of this layer, W q Represents the parameters of high-order neural network, Sig() is the Sigmoid activation function; S5.

3. Use the loss function Lg to evaluate the error between the classifier model output and the actual label. The specific expression is as follows: Among them, y true,i is the one-hot encoding of the true label, y pred,i is the probability output predicted by the classifier model, log(*) represents the logarithmic function, Indicates the sum of N data; S5.

4. Calculate the gradient of high-order neural network parameters and biases based on the loss function Lg and The specific expression is as follows: in, is the gradient of the loss function Lg with respect to the output; and is the gradient of the output with respect to the weights and biases; S5.

5. Accumulate the calculated gradients until the preset training cycle is reached, and the accumulated gradients of the high-order neural network parameters and biases and The expression is as follows: in, Indicates that the gradient is accumulated at each iteration Indicates that the gradient is accumulated at each iteration S5.6, after the iteration of step S5.5 is completed, the accumulated gradient is clipped to limit the maximum value of the gradient. The specific expression is as follows: Among them, clip(*,-τ q ,τ q ) means clipping the gradient to [-τ q ,τ q ] range, τ q Represents the gradient clipping threshold; S5.

7. Determine whether the preset maximum number of iterations has been reached. If not, jump to step S5.

3. If reached, stop iteration and complete training.

Images