Photovoltaic cell hot spot early warning method based on machine vision

Abstract

The invention discloses a photovoltaic cell hot spot early warning method based on machine vision. The method comprises the following steps: S1, image data acquisition; s2, image preprocessing; s3, symbol rules are constructed, and a photovoltaic working principle and a hot spot generation mechanism are converted; s4, generating a deep symbol learning model; s5, a population is initialized through a differential evolution algorithm, and candidate hyper-parameter combinations are generated; s6, candidate solution fitness evaluation, detection precision and risk evaluation accuracy; s7, optimizing a candidate solution; and S8, real-time hot spot detection and early warning report generation. According to the invention, the hyper-parameters are optimized by using high-resolution image acquisition, symbol reasoning and deep neural network fusion technologies and a differential evolution algorithm, and efficient and real-time detection and early warning of the hot spots of the photovoltaic cell panel in a complex environment are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of machine vision, and in particular to an early warning method for hot spots of photovoltaic cells based on machine vision. Background Art

[0002] In recent years, with the rapid development of new energy technologies, photovoltaic power generation has been widely used globally. As the core component of photovoltaic power generation systems, the safe and efficient operation of photovoltaic panels is an important prerequisite for ensuring power supply. However, due to the long-term exposure of photovoltaic panels to complex natural environments, affected by various factors such as light intensity, temperature fluctuations, and sand erosion, local temperature anomalies (i.e., hot spot phenomena) often occur. Hot spots not only reduce the energy conversion efficiency of photovoltaic panels but may also cause local material aging, generate microcracks, and even trigger safety accidents such as fires in severe cases. Therefore, how to perform real-time and accurate hot spot monitoring and early warning on photovoltaic panels has become a hot research topic.

[0003] In the prior art, the detection of hot spots on photovoltaic panels mainly relies on high-resolution thermal imaging cameras or visible light cameras to collect image data, and then through traditional image preprocessing methods such as denoising, enhancement, and normalization operations, algorithms such as threshold segmentation, edge detection, and region growing are used to extract hot spot regions. However, these methods have obvious defects. First, due to the complex external lighting conditions and severe noise interference, traditional image processing algorithms often have difficulty accurately distinguishing the subtle differences between real hot spots and the background, resulting in frequent false alarms and missed detections. Second, traditional methods mainly focus on the gray-scale features and edge information of images, lacking a systematic description of the working principle, energy conversion efficiency, and hot spot generation mechanism of photovoltaic panels, making the detection results lack physical significance and reliability. In addition, although some deep learning methods can automatically learn image features, due to relying only on a large amount of data for training, their black-box characteristics and generalization ability are insufficient, and it is also difficult to fully explain the internal mechanism of hot spot generation, thus affecting the detection accuracy.

[0004] To overcome the above deficiencies, in recent years, researchers have attempted to combine symbolic reasoning techniques with deep learning methods. By transforming the working principle of photovoltaic panels and the hot spot generation mechanism into symbolic rules, a symbolic reasoning framework is constructed and then fused with a deep neural network to generate a deep symbolic learning model, thereby achieving efficient identification of hot spots on photovoltaic panels. After collecting image data using machine vision technology, through preprocessing, symbolic rule construction, and symbolic reasoning, key physical parameters such as incident optical power, output current, output voltage, energy conversion efficiency, and temperature difference are organically combined, thus accurately depicting the internal mechanism of hot spot generation. By introducing a symbolic attention module to dynamically allocate the weights of symbolic rules, the deep neural network can not only learn the hidden features in the image during training but also integrate the physical information of the photovoltaic panel, improving the accuracy and robustness of hot spot detection.

[0005] On the other hand, the performance of the deep symbolic learning model highly depends on the reasonable setting of model hyperparameters. Traditional grid search or random search methods are inefficient in high-dimensional parameter spaces and cannot meet the requirements of real-time warning systems for response speed and accuracy. Therefore, a differential evolution algorithm is introduced as a global optimization tool to automatically tune the hyperparameters of the deep symbolic learning model. Through a series of steps such as initializing the population, generating candidate solutions, fitness evaluation, and mutation, crossover, and local search, the candidate solutions are gradually iteratively updated until the optimal hyperparameter combination is obtained, thus significantly improving the performance of the model in hot spot detection and risk assessment.

[0006] In summary, the existing technologies have problems such as insufficient image preprocessing, inaccurate feature extraction, insufficient description of physical mechanisms, and low efficiency of hyperparameter optimization in the detection of hot spots on photovoltaic panels. Based on this, the present invention proposes a method for early warning of hot spots on photovoltaic cells based on machine vision. This method uses high-resolution image acquisition and preprocessing, symbolic rule construction and symbolic reasoning, the fusion of a deep neural network and a symbolic attention module, and a differential evolution algorithm for hyperparameter optimization to form a complete and efficient hot spot detection and warning system. This method can not only accurately reflect the physical state of photovoltaic panels during actual operation but also achieve real-time monitoring and warning, greatly improving the safety and reliability of the operation of photovoltaic panels and providing strong technical support for the stable operation of new energy power generation systems.

[0007] Therefore, how to provide a method for early warning of hot spots on photovoltaic cells based on machine vision is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose an early warning method for hot spots of photovoltaic cells based on machine vision. The present invention makes full use of high-resolution image acquisition, image preprocessing, symbolic reasoning and deep neural network fusion technologies, and introduces a differential evolution algorithm for hyperparameter optimization. Specifically, the present invention first uses a high-resolution thermal imaging camera or a visible light camera to collect image data of a photovoltaic panel in real time, and then preprocesses the collected images, including denoising, enhancement and normalization. Then, by converting the working principle of the photovoltaic panel and the hot spot generation mechanism into symbolic rules, a symbolic reasoning framework is constructed, and combined with a deep neural network to generate a deep symbolic learning model, and a comprehensive analysis of the physical parameters and thermal information contained in the collected images is performed. Further, the present invention initializes the population using a differential evolution algorithm, generates multiple candidate solutions, globally optimizes the hyperparameters of the deep symbolic learning model, and realizes the mutation, crossover and local search of the candidate solutions through fitness evaluation, and finally obtains the optimal hyperparameter combination, and applies the optimized model to real-time hot spot detection, thereby generating an early warning report. This method has the advantages of high detection accuracy, strong real-time performance, good robustness and strong early warning ability, effectively improving the safety and operation efficiency of photovoltaic panels in complex environments.

[0009] An early warning method for hot spots of photovoltaic cells based on machine vision according to an embodiment of the present invention includes the following steps:

[0010] S1. Use a collection device to collect image data of a photovoltaic panel in real time;

[0011] S2. Preprocess the image data, including denoising, enhancement and normalization;

[0012] S3. Convert the working principle of the photovoltaic panel and the hot spot generation mechanism into symbolic rules, and construct a symbolic reasoning framework;

[0013] S4. Combine the symbolic reasoning framework with a deep neural network to generate a deep symbolic learning model;

[0014] S5. Based on the generated deep symbolic learning model, use a differential evolution algorithm to initialize the population and generate multiple candidate solutions, where each candidate solution represents a hyperparameter combination of a deep symbolic learning model;

[0015] S6. Perform fitness evaluation on the candidate solutions, and the evaluation criteria are hot spot detection accuracy and risk assessment accuracy;

[0016] S7. Optimize the candidate solutions through mutation, crossover and local search according to the fitness evaluation results, iteratively update the population until the optimal solution is reached and applied to real-time hot spot detection;

[0017] S8. Perform real-time hot spot detection according to the optimized deep symbolic learning model and generate an early warning report.

[0018] Optionally, S3 specifically includes:

[0019] S31. Collect the working parameter data of the photovoltaic panel. Define the incident light power as P in , the output current as I cell and the output voltage as V cell , calculate the output power P out = I cell ×V cell , calculate the energy conversion efficiency

[0020] S32. Collect the temperature data related to hot spots. Define the temperature of the hot spot area as T hotspot and the ambient temperature as T ambient , calculate the temperature difference ΔT = T hotspot - T ambient ;

[0021] S33. Set the hot spot generation threshold T th , and establish a judgment rule: when ΔT ≥ T th , define the hot spot state H = 1, otherwise H = 0;

[0022] S34. Construct a set of symbolic variables S = {η, I cell , V cell , P out , T hotspot , T ambient , ΔT};

[0023] S35. Establish a set of symbolic rules which includes rule R1: R2: P out = I cell ×V cell , R3: ΔT = T hotspot - T ambient and R4: ΔT ≥ T th ;

[0024] S36. Construct a symbolic reasoning framework whose mapping relationship is defined as

[0025] S37. Use the symbolic reasoning function f to substitute the set of symbolic variables S and the set of symbolic rules into the symbolic reasoning framework to obtain the hot spot detection result D, and its formula is

[0026] Optionally, S3 specifically includes:

[0027] S41. Let the set of symbolic variables S constructed in Claim 2 be used as the input vector X of the deep neural network, and define X = S;

[0028] S42. Construct a symbolic attention module to calculate the symbolic attention vector α, and its formula is:

[0029]

[0030] where softmax is a function that normalizes the result of the linear combination of the inputs into a probability distribution, is the symbolic attention weight matrix, is the bias vector, is the set of symbolic rules, ω r is the weight coefficient corresponding to the rule r, φ r (·) is the feature mapping function under the rule r;

[0031] S43. Combine the symbolic attention vector with the input vector through element-wise multiplication to obtain the integrated input vector Z, and its expression is:

[0032] Z = α ⊙ X;

[0033] where ⊙ represents element-wise multiplication;

[0034] S44. The symbolic inference function is defined as:

[0035]

[0036] where, φ r (h) is the output after mapping the input feature h according to the rule r, ω r is the corresponding rule weight;

[0037] S45. Construct a deep neural network that integrates symbolic inference. Suppose the network has a total of L layers. For the l-th layer, l = 1, 2,..., L, define its output as:

[0038]

[0039] where, h (0) = Z, W (l) and b (l) are the weight matrix and bias vector of the l-th layer respectively, σ(·) is the activation function, λ (l) is the scalar of the symbolic integration factor in the l-th layer;

[0040] S46. Define the final output of the deep neural network as:

[0041]

[0042] where, is the activation function of the output layer;

[0043] S47. Set the deep symbolic learning model M DSL as:

[0044]

[0045] where, are model parameters;

[0046] S48. Define the loss function of the model as:

[0047]

[0048] where, represents the loss function, Y i is the output of the model for the i-th sample, y i is the corresponding actual hot spot state, N is the total number of samples, and β is the regularization coefficient;

[0049] S49. Use the gradient descent method to optimize the parameter θ, and its parameter update formula is:

[0050]

[0051] where, α is the learning rate;

[0052] S410. Apply the trained deep symbolic learning model M DSL to the real-time hot spot detection task.

[0053] Optionally, the S5 specifically includes:

[0054] S51. Set the population size of the differential evolution algorithm as NP and the dimension of the hyperparameters to be optimized in the deep symbolic learning model as d, where NP is the total number of candidate solutions and d represents the number of hyperparameters;

[0055] S52. For each candidate solution H i , i = 1, 2,..., NP, define the candidate solution vector as:

[0056] H i = [θ i,1 , θ i,2 , …, θ i,d ;

[0057] where, θ i,j represents the value of the j-th hyperparameter in the i-th candidate solution;

[0058] S53. Set the lower bound vector L = [L1, L2, …, L d and the upper bound vector U = [U1, U2, …, Ud , where L j and U j represent the lower and upper bounds of the j-th hyperparameter, respectively;

[0059] S54. For each candidate solution H i , generate an initial solution according to the following formula:

[0060] H i = L + rand i ·(U - L);

[0061] where rand i = [r i,1 , r i,2 , …, r i,d is a random vector drawn from the uniform distribution U(0, 1), and the multiplication is element-wise multiplication;

[0062] S55. Construct the initial population P as:

[0063] P = {H i | i = 1, 2, …, NP};

[0064] where each candidate solution H i represents a hyperparameter combination in the deep symbolic learning model M DSL (X; θ), and let θ = H i ;

[0065] S56. Record and output the initial population P as the basis for subsequent fitness evaluation and differential evolution operations.

[0066] Optionally, the S6 specifically includes:

[0067] S61. For each candidate solution H i , i = 1, 2, …, NP, let the corresponding deep symbolic learning model be M DSL (X; H i ), where X is the input vector defined in claim 3;

[0068] S62. In the validation dataset D val , for each candidate solution H i calculate the matching situation between the output Y i = M DSL (X; H i ) of the hot spot detection model and the actual hot spot state y;

[0069] S63. Define the hot spot detection accuracy Acc hotspot (H i ) as:

[0070]

[0071] Among them, N is the total number of verification samples, Y i,j is the candidate solution H i the predicted output for the j-th sample, y j is the actual hot spot state of the j-th sample, δ(a, b) is the indicator function, which takes 1 when a = b, and 0 otherwise;

[0072] S64. Define the risk assessment accuracy Acc risk (H i ) is defined as:

[0073]

[0074] Among them, γ(a, b) is the risk assessment function, which takes 1 when the prediction of the risk state by the candidate solution H i is consistent with the actual risk state, and 0 otherwise;

[0075] S65. Set the overall fitness function F(H i ) as:

[0076] F(H i ) = w1·Acc hotspot (H i ) + w2·Acc risk (H i );

[0077] Among them, w1 and w2 are non-negative weight coefficients, satisfying w1 + w2 = 1;

[0078] S66. For each candidate solution H in the verification dataset i calculate the fitness value F(H i ), and record and output the fitness values of all candidate solutions as the basis for candidate solution evaluation.

[0079] Optionally, the specific steps of S7 include:

[0080] S71. For each candidate solution H i in the population P = {H i |i = 1, 2,..., NP}, according to the mutation operation of the differential evolution algorithm, generate a mutation vector V i,1 , and its formula is: i,2 ,…,θ i,d = [θ i is:

[0081] V i = H r1 + μ·(H r2 - H r3 );

[0082] Among them, Hr1 , H r2 and H r3 are candidate solutions randomly selected from P and different from each other, and μ is a preset mutation factor;

[0083] S72. For each candidate solution H i and its corresponding mutation vector V i perform a crossover operation to generate a trial vector U i = [u i,1 , u i,2 , …, u i,d , and its crossover formula is

[0084]

[0085] where r i,j is a value randomly drawn from the uniform distribution U(0, 1), CR is the crossover probability, and j rand is a random index to ensure that at least one dimension is taken from V i ;

[0086] S73. Calculate the fitness values for each candidate solution H i and the corresponding trial vector U i respectively. Define the fitness function F(H) as:

[0087] F(H) = w1·Acc hotspot (H) + w2·Acc risk (H);

[0088] where Acc hotspot (H) is the hotspot detection accuracy, Acc risk (H) is the risk assessment accuracy, w1, w2 ≥ 0 and satisfy w1 + w2 = 1;

[0089] S74. Compare F(U i ) with F(H i ). If F(U i ) ≥ F(H i ), then update the candidate solution, that is, let H i ←U i ;

[0090] S75. Perform local search optimization on the updated candidate solution H i . Define the local search formula as:

[0091] H′ i = H i + δ·n;

[0092] where H′ i represents the candidate solution H iThe updated candidate solution after local search optimization, where δ is the local search step size, is a random vector drawn from the normal distribution N(0, I d ); if F(H′ i ) ≥ F(H i ) then update the candidate solution H i ←H′ i ;

[0093] S76. Repeat the steps from S71 to S75 until the termination condition T cond is satisfied, where the termination condition is defined as:

[0094] or

[0095] where Θ is a preset threshold, is the variance of the population fitness, and ∈ is a small positive number;

[0096] S77. Select the candidate solution H * with the highest fitness, which satisfies:

[0097]

[0098] where arg max represents the value of the independent variable that makes a certain function reach its maximum, and use H * as the optimal hyperparameter combination for real-time hot spot detection.

[0099] The beneficial effects of the present invention are as follows:

[0100] By organically combining a number of advanced technologies such as high-resolution image acquisition, image preprocessing, symbol inference and deep neural network fusion, and hyperparameter optimization of the differential evolution algorithm, the present invention realizes real-time, accurate detection and early warning of hot spots on photovoltaic panels. In the present invention, first, image data of photovoltaic panels are collected using a high-resolution thermal imaging camera or a visible light camera, and through preprocessing steps such as denoising, enhancement and normalization, it is ensured that the collected images have high quality and stability, providing a reliable basis for subsequent processing. Subsequently, in combination with the working principle of photovoltaic panels and the hot spot generation mechanism, key physical parameters such as incident light power, output current, output voltage, energy conversion efficiency, and the temperature of the hot spot area and the ambient temperature are transformed into symbolic rules, a set of symbolic variables and a set of symbolic rules are constructed, and a symbolic inference framework is used to systematically express these rules, so as to accurately reflect the internal mechanism of hot spot generation on photovoltaic panels.

[0101] On this basis, the present invention effectively integrates the symbolic reasoning framework with the deep neural network to generate a deep symbolic learning model with a symbolic attention module, enabling the system to not only automatically extract implicit features in images but also embed the physical characteristics of photovoltaic panels into the model, greatly improving the accuracy and robustness of hot spot detection. To further enhance the performance of the model in complex environments, the present invention introduces a differential evolution algorithm to globally optimize the hyperparameters of the deep symbolic learning model. By initializing the population, generating candidate hyperparameter combinations, and performing mutation, crossover, and local search optimization on the candidate solutions using fitness evaluation (where the fitness function comprehensively considers the accuracy of hot spot detection and risk assessment accuracy), the present invention can quickly screen out the optimal hyperparameter combination, thus ensuring that the model has high-efficiency real-time response capabilities and excellent detection effects during actual operation.

[0102] In summary, the present invention not only achieves high-precision detection and real-time warning of hot spots on photovoltaic panels but also has achieved remarkable results in reducing the false alarm rate, enhancing system robustness, optimizing model performance, and improving real-time performance. By combining physical rules with deep learning, the present invention effectively makes up for the deficiencies of traditional image processing methods and single deep neural networks in dealing with complex lighting conditions, noise interference, and background clutter, greatly improving the accuracy and stability of detection results. At the same time, the introduction of the differential evolution algorithm for hyperparameter optimization significantly improves the search efficiency of the system in the high-dimensional parameter space, making the overall warning system more suitable for the dynamic monitoring and intelligent maintenance of photovoltaic panels. This technical solution can generate a warning report in a timely manner before the occurrence of hot spots, providing accurate decision-making basis for operation and maintenance personnel, thereby effectively reducing the risks of equipment failures and safety accidents caused by hot spots and enhancing the overall operation efficiency and economic benefits of the photovoltaic power generation system. The technical solution implemented by the present invention not only provides an advanced intelligent monitoring means for the new energy power generation field but also has significant application prospects and promotion values in ensuring the safe and stable operation of photovoltaic panels, extending the service life of equipment, and reducing operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0104] Figure 1 is the overall flowchart of a method for early warning of hot spots on photovoltaic cells based on machine vision proposed by the present invention;

[0105] Figure 2 is the flowchart of fitness evaluation, mutation, crossover, and local search optimization of candidate solutions in step S7 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0106] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only schematically showing the basic structure of the present invention, so they only show the components related to the present invention.

[0107] Reference Figure 1 and Figure 2 , a method for early warning of hot spots in photovoltaic cells based on machine vision, comprising the following steps:

[0108] S1. Use a collection device to collect image data of a photovoltaic panel in real time;

[0109] S2. Preprocess the image data, including denoising, enhancement, and normalization;

[0110] S3. Convert the working principle and hot spot generation mechanism of the photovoltaic panel into symbolic rules to construct a symbolic reasoning framework;

[0111] S4. Combine the symbolic reasoning framework with a deep neural network to generate a deep symbolic learning model;

[0112] S5. Based on the generated deep symbolic learning model, use the differential evolution algorithm to initialize the population and generate multiple candidate solutions, where each candidate solution represents a combination of hyperparameters of a deep symbolic learning model;

[0113] S6. Evaluate the fitness of the candidate solutions, and the evaluation criteria are hot spot detection accuracy and risk assessment accuracy;

[0114] S7. Mutate, cross, and perform local search optimization on the candidate solutions according to the fitness evaluation results, iteratively update the population until the optimal solution is reached and applied to real-time hot spot detection;

[0115] S8. Perform real-time hot spot detection according to the optimized deep symbolic learning model and generate a warning report.

[0116] In this embodiment, S3 specifically includes:

[0117] S31. Collect the working parameter data of the photovoltaic panel, define the incident light power as P in , the output current as I cell and the output voltage as V cell , calculate the output power P out = I cell × V cell , calculate the energy conversion efficiency

[0118] S32. Collect the temperature data related to hot spots, define the temperature of the hot spot area as T hotspot and the ambient temperature as T ambient , calculate the temperature difference ΔT = Thotspot -T ambient ;

[0119] S33. Set the hot spot generation threshold T th , and establish a decision rule: when ΔT≥T th , define the hot spot state H = 1, otherwise H = 0;

[0120] S34. Construct a set of symbolic variables S = {η, I cell , V cell , P out , T hotspot , T ambient , ΔT};

[0121] S35. Establish a set of symbolic rules which includes rule R1: R2: P out = I cell × V cell , R3: ΔT = T hotspot - T ambient and R4: ΔT≥T th ;

[0122] S36. Construct a symbolic reasoning framework whose mapping relationship is defined as

[0123] S37. Use the symbolic reasoning function f to substitute the set of symbolic variables S and the set of symbolic rules into the symbolic reasoning framework to obtain the hot spot detection result D, and its formula is

[0124] In this embodiment, the S3 specifically includes:

[0125] S41. Let the set of symbolic variables S constructed in claim 2 be the input vector X of the deep neural network, and define X = S;

[0126] S42. Construct a symbolic attention module to calculate the symbolic attention vector α, and its formula is:

[0127]

[0128] where softmax is a function that normalizes the linear combination result of the input into a probability distribution, is the symbolic attention weight matrix, is the bias vector, is the set of symbolic rules, ω r is the weight coefficient corresponding to rule r, φ r (·) is the feature mapping function under rule r;

[0129] S43. Combine the symbolic attention vector with the input vector through element-wise multiplication to obtain the integrated input vector Z, and its expression is:

[0130] Z = α ⊙ X;

[0131] where ⊙ represents element-wise multiplication;

[0132] S44. The symbolic reasoning function is defined as:

[0133]

[0134] where φ r (h) is the output after mapping the input feature h according to the rule r, and ω r is the corresponding rule weight;

[0135] S45. Construct a deep neural network that integrates symbolic reasoning. Assume the network has L layers. For the l-th layer, l = 1, 2, …, L, define its output as:

[0136]

[0137] where h (0) = Z, W (l) and b (l) are the weight matrix and bias vector of the l-th layer respectively, σ(·) is the activation function, and λ (l) is the scalar of the symbolic integration factor in the l-th layer;

[0138] S46. Define the final output of the deep neural network as:

[0139]

[0140] where, is the activation function of the output layer;

[0141] S47. Set the deep symbolic learning model M DSL as:

[0142]

[0143] where, are the model parameters;

[0144] S48. Define the loss function of the model as:

[0145]

[0146] where, represents the loss function, Y i is the output of the model for the i-th sample, and yi is the corresponding actual hot spot state, N is the total number of samples, and β is the regularization coefficient;

[0147] S49. Optimize the parameter θ using the gradient descent method, and its parameter update formula is:

[0148]

[0149] where α is the learning rate;

[0150] S410. Apply the trained deep symbolic learning model M DSL to the real-time hot spot detection task.

[0151] In this embodiment, the S5 specifically includes:

[0152] S51. Set the population size of the differential evolution algorithm to NP and the dimension of the hyperparameters to be optimized in the deep symbolic learning model to d, where NP is the total number of candidate solutions and d represents the number of hyperparameters;

[0153] S52. For each candidate solution H i , i = 1, 2,..., NP, define the candidate solution vector as:

[0154] H i = [θ i,1 , θ i,2 ,..., θ i,d ;

[0155] where θ i,j represents the value of the jth hyperparameter in the ith candidate solution;

[0156] S53. Set the lower bound vector L = [L1, L2,..., L d and the upper bound vector U = [U1, U2,..., U d , where L j and U j respectively represent the lower bound and upper bound of the jth hyperparameter;

[0157] S54. For each candidate solution H i , generate the initial solution according to the following formula:

[0158] H i = L + rand i ·(U - L);

[0159] where rand i = [r i,1 , r i,2 ,..., r i,d is a random vector drawn from the uniform distribution U(0, 1), and the multiplication is element-wise multiplication;

[0160] S55. The initial population P is constituted as follows:

[0161] P = {H i | i = 1, 2, …, NP};

[0162] where each candidate solution H i represents a combination of hyperparameters in the deep symbolic learning model M DSL (X; θ), and let θ = H i ;

[0163] S56. Record and output the initial population P as the basis for subsequent fitness evaluation and differential evolution operations.

[0164] In this embodiment, the specific steps of S6 are as follows:

[0165] S61. For each candidate solution H i generated in claim 4, where i = 1, 2, …, NP, let the corresponding deep symbolic learning model be M DSL (X; H i ), where X is the input vector defined in claim 3;

[0166] S62. In the validation dataset D val , calculate the matching situation between the output Y i of the hot spot detection model and the actual hot spot status y for each candidate solution H i = M DSL (X; H i );

[0167] S63. Define the hot spot detection accuracy Acc hotspot (H i ) as:

[0168]

[0169] where N is the total number of validation samples, Y i,j is the predicted output of the candidate solution H i [[ID=6l]]for the j-th sample, y j is the actual hot spot status of the j-th sample, and δ(a, b) is an indicator function that takes 1 when a = b and 0 otherwise;

[0170] S64. Define the risk assessment accuracy Acc risk (H i ) as:

[0171]

[0172] where γ(a, b) is a risk assessment function. When the candidate solution H iWhen the prediction of the risk status is consistent with the actual risk status, take 1; otherwise, take 0.

[0173] S65. Set the overall fitness function F(H i ) as:

[0174] F(H i ) = w1·Acc hotspot (H i ) + w2·Acc risk (H i );

[0175] Wherein, w1 and w2 are non - negative weight coefficients, satisfying w1 + w2 = 1;

[0176] S66. For each candidate solution H in the validation dataset i Calculate the fitness value F(H i ), and record and output the fitness values of all candidate solutions as the basis for candidate solution evaluation.

[0177] In this embodiment, the S7 specifically includes:

[0178] S71. For each candidate solution H i ∣i = 1, 2, …, NP} in the population P = {H i = [θ i,1 , θ i,2 , …, θ i,d , generate a mutant vector V i according to the mutation operation of the differential evolution algorithm. Its formula is:

[0179] V i = H r1 + μ·(H r2 - H r3 );

[0180] Wherein, H r1 , H r2 and H r3 are randomly selected from P and are different from each other, and μ is a preset mutation factor;

[0181] S72. Perform a crossover operation on each candidate solution H i and its corresponding mutant vector V i to generate a trial vector U i = [u i,1 , u i,2 , …, u i,d . Its crossover formula is

[0182]

[0183] Where ri,j is a value randomly drawn from the uniform distribution U(0, 1), CR is the crossover probability, and j rand is a random index to ensure that at least one dimension is taken from V i ;

[0184] S73. For each candidate solution H i and the corresponding trial vector U i calculate the fitness values respectively. Define the fitness function F(H) as:

[0185] F(H) = w1·Acc hotspot (H) + w2·Acc risk (H);

[0186] where Acc hotspot (H) is the hotspot detection accuracy, Acc risk (H) is the risk assessment accuracy, w1, w2 ≥ 0 and satisfy w1 + w2 = 1;

[0187] S74. Compare F(U i ) with F(H i ). If F(U i ) ≥ F(H i ), then update the candidate solution, that is, let H i ←U i ;

[0188] S75. Perform local search optimization on the updated candidate solution H i . Define the local search formula as:

[0189] H′ i = H i + δ·n;

[0190] where H′ i represents the updated candidate solution after local search optimization of the candidate solution H i , δ is the local search step size, is a random vector drawn from the normal distribution N(0, I d ); if F(H′ i ) ≥ F(H i ), then update the candidate solution H i ←H′ i ;

[0191] S76. Repeat the steps of S71 to S75 until the termination condition T cond is satisfied, where the termination condition is defined as:

[0192] or

[0193] where Θ is a preset threshold value, is the variance of the population fitness, and ∈ is a small positive number;

[0194] S77. Select the candidate solution H with the highest fitness * , which satisfies:

[0195]

[0196] where argmax represents the value of the independent variable that makes a certain function reach the maximum value, and use H * as the optimal hyperparameter combination and apply it to real-time hot spot detection.

[0197] Example 1:

[0198] To verify the feasibility of the present invention in implementation, the present invention is applied to a large-scale photovoltaic power plant in East China as an example. The installed capacity of this power plant reaches 200 MW and it has nearly 20,000 photovoltaic panels. Due to being in a high-temperature and high-radiation environment for a long time, local hot spot phenomena often occur in the photovoltaic panels, resulting in a decrease in the overall energy conversion efficiency of the panels and potential safety hazards. Therefore, the present invention uses the early warning method for photovoltaic cell hot spots based on machine vision, deeply improves the problems of low detection accuracy and high false alarm rate of traditional image processing methods in complex environments, and verifies the effectiveness and advancement of the present invention through practical applications.

[0199] In this embodiment, first, a high-resolution thermal imaging camera and a visible light camera are installed on-site at the power plant to collect all-weather and real-time image data of the photovoltaic panels. The collected images are processed by a preprocessing module for denoising, image enhancement, and normalization, making the detailed features of the images clearer and laying a solid foundation for subsequent processing. Then, through a comprehensive analysis of the working principle of the photovoltaic panels and the hot spot generation mechanism, physical parameters such as incident light power, output current, and output voltage, as well as the temperature difference between the hot spot area and the ambient temperature, are transformed into symbolic rules to construct a set of symbolic variables and a set of symbolic rules. Using these symbolic rules, a symbolic reasoning framework is established to match the actual data collected with the theoretical rules, thereby initially determining the hot spot state.

[0200] Subsequently, in this embodiment, a symbolic reasoning framework is combined with a deep neural network to construct a deep symbolic learning model. The model uses a symbolic attention module to dynamically weight symbolic rules, extracts implicit features in the image through a multi-layer deep neural network, and fuses the physical parameter information of the photovoltaic panels to output the detection result. To further improve the model performance, this embodiment introduces a differential evolution algorithm to globally optimize the hyperparameters of the deep symbolic learning model. By initializing the population, generating multiple candidate solutions, and iteratively updating the candidate solutions using a fitness function (considering both the hot spot detection accuracy and the risk assessment accuracy), the optimal hyperparameter combination is finally selected, enabling the model to maintain high detection accuracy and fast response ability in complex environments.

[0201] In the actual application process, in this embodiment, about 20,000 photovoltaic panels at the power plant site were monitored in real time for three consecutive months. The monitoring system runs continuously for 24 hours every day, collects and processes image data in real time, and quickly generates a warning report when a hot spot anomaly is detected. To further verify the superiority of the present invention in actual applications, this embodiment also selects two areas in the same photovoltaic power plant for a comparative test. One area uses traditional image processing methods for hot spot detection, and the other area uses the machine vision-based hot spot early warning method of the present invention.

[0202] To more intuitively demonstrate the beneficial effects of the present invention in actual applications, the following table is a comparative data table of the traditional method and the method of the present invention in the hot spot detection of photovoltaic panels. This table details key indicators such as data collection time, collection location, sample quantity, detection accuracy rate, false alarm rate, average response time, etc., fully proving the significant advantages of the method of the present invention in improving detection accuracy, reducing false alarm rate, accelerating the system response speed, and enhancing system robustness.

[0203] Table 1: Comparative Data Table of Hot Spot Detection Effects of Photovoltaic Panels

[0204]

[0205] The above table shows the comparison of multiple key indicators between the traditional image processing method and the machine vision-based photovoltaic cell hot spot early warning method of the present invention. The data shows that the method of the present invention is significantly superior to the traditional method in terms of detection accuracy rate, which has increased from 81.2% of the traditional method to 96.5%. This means that the system can more accurately identify the hot spot areas on the photovoltaic panels, thereby effectively reducing the missed detection situation. At the same time, the false alarm rate has dropped from 12.5% to 2.3%, indicating that the present invention performs better in reducing false alarms, which helps to reduce unnecessary maintenance interventions and resource waste.

[0206] In addition, in terms of the average response time, the method of the present invention shortens the response time from 8.0 seconds of the traditional method to 2.8 seconds, greatly improving the real-time warning ability of the system. This improvement is particularly crucial for quickly responding to the abnormal state of photovoltaic panels, enabling timely measures to be taken at the initial stage of hot spots to prevent the situation from deteriorating. The energy conversion efficiency degradation rate has decreased from 18.5% to 5.2%, further demonstrating the advantages of the present invention in maintaining the overall performance of photovoltaic panels and effectively reducing the energy loss caused by local hot spot phenomena. In terms of maintenance response time and accident warning success rate, the data also shows obvious improvements. The accident warning success rate has increased from 75.0% to 98.0%, indicating that the present invention is more accurate and timely in warning and can better ensure the safe operation of the power generation system.

[0207] Generally speaking, by introducing the fusion of symbolic reasoning and deep neural network and the automatic optimization of hyperparameters by the differential evolution algorithm, the present invention not only makes breakthroughs in image data processing and feature extraction, but also greatly improves the accuracy, real-time performance and system robustness of hot spot detection in actual operation. The data in the table fully proves the superiority of the present invention compared with traditional image processing methods, providing strong technical support for the safe operation and efficient management of photovoltaic panels.

[0208] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for early warning of hot spots in photovoltaic cells based on machine vision, characterized in that, It includes the following steps: S1. Use a collection device to collect image data of the photovoltaic panel in real time; S2. Preprocess the image data, including denoising, enhancement, and normalization; S3. Convert the working principle and hot spot generation mechanism of the photovoltaic panel into symbolic rules to construct a symbolic reasoning framework; S4. Combine the symbolic reasoning framework with a deep neural network to generate a deep symbolic learning model; S5. Based on the generated deep symbolic learning model, use the differential evolution algorithm to initialize the population and generate multiple candidate solutions, where each candidate solution represents a combination of hyperparameters of the deep symbolic learning model; S6. Conduct fitness evaluation on the candidate solutions, and the evaluation criteria are hot spot detection accuracy and risk assessment accuracy; S7. According to the fitness evaluation results, perform mutation, crossover, and local search optimization on the candidate solutions, iteratively update the population until the optimal solution is reached and applied to real-time hot spot detection; S8. Conduct real-time hot spot detection according to the optimized deep symbolic learning model and generate a warning report.

2. The early warning method for hot spots of a photovoltaic cell based on machine vision according to claim 1, characterized in that, The specific content of S3 includes: S31. Collect the working parameter data of the photovoltaic panel, define the incident light power as P in , the output current as I cell and the output voltage as V cell , calculate the output power P out = I cell × V cell , calculate the energy conversion efficiency S32. Collect temperature data related to hot spots and define the temperature of the hot spot area as T gotspot and the ambient temperature is T ambient , calculate the temperature difference ΔT = T hotspot - T ambient ; S33. Set the hot spot generation threshold T th , and establish a judgment rule: when ΔT≥T th , define the hot spot state H = 1, otherwise H = 0; S34. Construct the set of symbolic variables S = {η, I cell , V cell , P out , T hotspot , T ambient , ΔT}; S35. Establish a symbol rule set which includes rules R2: P out = I cell × V cell , R3: ΔT = T hotspot -T ambient and R4: ΔT ≥ T tg ; S36. Construct a symbolic reasoning framework The mapping relationship is defined as S37. Using the symbolic reasoning function f, substitute the symbolic variable set S and the symbolic rule set into the symbolic reasoning framework to obtain the hot spot detection result D, and its formula is 3. The early warning method for hot spots of photovoltaic cells based on machine vision according to claim 1, characterized in that, The specific content of S3 includes: S41. Let the set of symbolic variables S constructed in claim 2 be the input vector X of the deep neural network, and define X = S; S42. Construct a symbolic attention module to calculate the symbolic attention vector α, and its formula is: Among them, softmax is a function that normalizes the result of the linear combination of inputs into a probability distribution. is the symbolic attention weight matrix. is the bias vector. is the set of symbolic rules, ω r is the weight coefficient corresponding to rule r, φ r (·) is the feature mapping function under rule r. S43. Combine the symbolic attention vector with the input vector through element-wise multiplication to obtain the integrated input vector Z, and its expression is: Z = α ⊙ X; where ⊙ represents element-wise multiplication; S44, Symbolic reasoning function It is defined as: Among them, φ r (h) is the output after mapping the input feature h according to the rule r, and ω r is the corresponding rule weight; S45. Construct a deep neural network integrating symbolic reasoning. Suppose the network has a total of L layers. For the l-th layer, l = 1, 2,..., L, define its output as: where, g (0) = Z, W (l) and b (l) are the weight matrix and bias vector of the l-th layer respectively, σ(·) is the activation function, and λ (l) is the scalar of the symbol integration factor in the l-th layer; S46. Define the final output of the deep neural network as: Among them, is the output layer activation function; S47. Set the depth symbol learning model M DSL It is: Among them, are model parameters; S48. Define the loss function of the model as: Among them, represents the loss function, Y i is the output of the model for the i-th sample, y i is the corresponding actual hot spot state, N is the total number of samples, and β is the regularization coefficient; S49. Use the gradient descent method to optimize the parameter θ, and its parameter update formula is: where α is the learning rate; S410. Apply the trained deep symbolic learning model M DSL to the real-time hot spot detection task.

4. The early warning method for hot spots of photovoltaic cells based on machine vision according to claim 1, characterized in that, The specific content of S5 includes: S51. Set the population size of the differential evolution algorithm as NP and the dimension of the hyperparameters to be optimized in the deep symbolic learning model as d, where NP is the total number of candidate solutions and d represents the number of hyperparameters; S52. For each candidate solution H i = 1, 2,..., NP, define the candidate solution vector as: H i = [θ i，1 , θ i，2 , …, θ i，d ; where, θ i，j represents the value of the j-th hyperparameter in the i-th candidate solution; S53. Set the lower bound vector L = [L1, L2,..., L d and the upper bound vector U = [U1, U2,..., U d , where L j and U j respectively represent the lower bound and the upper bound of the j-th hyperparameter; S54. For each candidate solution H i , generate an initial solution according to the following formula: H i = L + rand i ·(U - L); where rand i = [r i，1 , r i，2 , …, r i，d is a random vector drawn from the uniform distribution U(0, 1), and the multiplication is element-wise multiplication; S55. Construct the initial population P as: P = {H i | i = 1, 2, …, NP}; where each candidate solution H i represents a hyperparameter combination in the deep symbolic learning model M DSL (X; θ), and let θ = H i ; S56. Record and output the initial population P as the basis for subsequent fitness evaluation and differential evolution operations.

5. The early warning method for hot spots of photovoltaic cells based on machine vision according to claim 1, characterized in that, The specific content of S6 includes: S61. For each candidate solution H i , where i = 1, 2,..., NP, let the corresponding deep symbolic learning model be M DSL (X; H i ), where X is the input vector defined in claim 3; S62. In the validation dataset D val for each candidate solution H i calculate the match between the output Y i = M DSL (X; H i ) and the actual hot spot status y; S63. Define the hot spot detection accuracy Acc hotspot (H i ) is defined as: where N is the total number of verification samples, Y i，j is the candidate solution H i is the predicted output for the j-th sample, y j is the actual hot spot state of the j-th sample, δ(a, b) is the indicator function, which takes 1 when a = b and 0 otherwise; S64. Define the risk assessment accuracy Acc risk (H i ) is defined as: Among them, γ(a, b) is a risk assessment function, and when the prediction of the risk state by the candidate solution H i is consistent with the actual risk state, it takes 1, otherwise it takes 0; S65. Set the overall fitness function F(H i ) as follows: F(H i ) = w1·Acc hotspot (H i ) + w2·Acc risk (H i ); where w1 and w2 are non-negative weight coefficients, satisfying w1 + w2 = 1; S66. For each candidate solution H in the validation dataset i calculate the fitness value F(H i ), and record and output the fitness values of all candidate solutions as the basis for candidate solution evaluation.

6. The early warning method for hot spots of photovoltaic cells based on machine vision according to claim 1, characterized in that, The specific content of S7 includes: S71. For population P = {H i Each candidate solution H in |i=1,2,…,NP} i =[θ i，1 ,θ i，2 ,…,θ i，d ], based on the mutation operation of the differential evolution algorithm, generate the mutation vector V i , the formula is: V i = H r1 + μ·(H r2 - H r3 ); Among them, H r1 , H r2 and H r3 are candidate solutions randomly selected from P and different from each other, and μ is a preset mutation factor; S72. For each candidate solution H i and its corresponding mutation vector V i perform a crossover operation to generate a trial vector U i = [u i，1 , u i，2 , …, u i，d , and its crossover formula is: where r i，j is a value randomly drawn from the uniform distribution U(0, 1), CR is the crossover probability, and j rand is a random index to ensure that at least one dimension is taken from V i ; S73. For each candidate solution H i and the corresponding trial vector U i calculate the fitness value respectively, and define the fitness function F(H) as: F(H) = w1·Acc hotspot (H) + w2·Acc risk (H); Among them, Acc hotspot (H) is the hot spot detection accuracy, Acc risk (H) is the risk assessment accuracy, w1, w2 ≥ 0 and satisfy w1 + w2 = 1; S74. Compare F(U i ) with F(H i ). If F(U i ) ≥ F(H i ), then update the candidate solution, i.e., set H i ← U i . S75. Perform local search optimization on the updated candidate solution H i and define the local search formula as follows: H′ i = H i + δ·n; Among them, H′ i represents the updated candidate solution after local search optimization of the candidate solution H i , where δ is the local search step size, and is a random vector drawn from the normal distribution N(0, I d ); if F(H′ i ) ≥ F(H i ), then update the candidate solution H i ← H′ i ; Repeat the steps of S71 to S75 until the termination condition T is satisfied cond , where the termination condition is defined as: where Θ is a preset threshold value, is the variance of the population fitness, and ∈ is a small positive number; S77. Select the candidate solution H with the highest fitness * , which satisfies: where arg max represents the value of the independent variable that maximizes a certain function, and H * is applied as the optimal hyperparameter combination to real-time hot spot detection.

Images