A method for predicting the film thickness of photovoltaic cells based on improved deep learning

Abstract

The present invention provides a method for predicting the film thickness of a photovoltaic cell based on improved deep learning. The specific steps include: (1) collecting the image information and film thickness value of the photovoltaic cell after coating through a camera device, where the image information includes image data of RGB three-channel grayscale values; (2) converting the RGB image data into HSL image data; (3) determining the correlation R based on the HSL image data and the film thickness value, and determining the image data with the strongest correlation with the actual coating thickness according to the correlation R, and determining the color depth value Y value based on the H value and L value in the HSL image data; (4) performing normalization processing on the image data; (5) forming a color depth value and film thickness prediction model based on a BP neural network; (6) predicting the coating thickness of the photovoltaic cell through the film thickness prediction model and the color depth value Y, determining that the error between the predicted film thickness and the actual film thickness is within the allowable range, with a fast calculation speed and good reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic cells, and particularly relates to a method for predicting the film thickness of a photovoltaic cell based on improved deep learning. Background Art

[0002] Photovoltaic cells are the basic components of solar photovoltaic power generation, and their performance directly affects the efficiency and stability of solar cells. In order to improve the light absorption ability and current output efficiency of the cells, it is often necessary to coat the surface of the cells, and the film thickness of the cell coating will have a serious impact on the power generation efficiency of the cells.

[0003] Currently, the mainstream production process in the market is the TopCon photovoltaic cell production process, which prepares a ultra-thin tunneling oxide layer and a highly doped polycrystalline silicon thin layer on the back of the cell. The two together form a passivated contact structure, which provides good surface passivation for the back of the silicon wafer. The ultra-thin oxide layer allows multi-electrons to tunnel into the polycrystalline silicon layer while blocking the recombination of minority carrier holes. Then, the electrons are laterally transported in the polycrystalline silicon layer and collected by the metal, thus greatly reducing the metal contact recombination current and improving the open circuit voltage and short circuit current of the cell.

[0004] PECVD (Plasma Enhanced Chemical Vapor Deposition) is one of the very important processes in the production coating process of photovoltaic cells. The coating quality is also an important indicator reflecting the efficiency of a company's solar cells. The quality and performance of the cell coating are affected by various factors, mainly including coating materials, coating process parameters, and cell surface characteristics, etc. Among them, the control of coating process parameters is the key to ensuring the quality and performance of the thin film.

[0005] In the patent document with the Chinese patent application number 202310186566.7 and the publication date of June 30, 2023, a method for manufacturing a photovoltaic module is disclosed. It uses ion bombardment on the thin film substrate to clean the surface dirt and water vapor, selects different ion bombardment energies according to different roughnesses in different regions, uses magnetron coating when depositing the Mo film, and uses a moving sputtering target. When the moving target moves to different regions, the height of the sputtering target and the sputtering power of the sputtering target are adjusted according to the corresponding roughness of different regions, and the adjusted parameters are corrected to ensure the uniformity and consistency of the coating, avoid uneven coating due to roughness problems, obtain better coating uniformity and consistency, deposit the absorption layer by evaporation method, adjust the temperature by thickness change, ensure that the temperature is in a better range, and obtain a better film deposition rate, avoid loose pores, thereby improving the light transmittance and enhancing the energy conversion efficiency of the finished product.

[0006] For film thickness prediction, generally, it is only necessary to simply obtain the film thickness information through a camera. However, for this simple method of obtaining film thickness information through a camera, there are also many prediction models in the prior art, such as the BP neural network. The BP neural network has obvious advantages in dealing with such data prediction problems. Its outstanding advantage is its strong non-linear mapping ability and flexible network structure. The number of intermediate layers of the network and the number of neurons in each layer can be set according to the model requirements, and its performance also varies with the structural differences.

[0007] However, the traditional BP neural network generally predicts the output based on the input image data. If a large amount of input image data is input into the BP neural network, this will lead to a large computational pressure during the entire prediction process. And if RGB image data is used as the input, it is only the input average value, which is not easy to reflect the real situation of the entire image data, thus unable to reliably predict the actual film thickness. In addition, the Sigmoid function is generally used in the BP neural network. However, due to the composition of this function, there is a problem of gradient disappearance, that is, gradient saturation. And because there is an exponential term of e in this function, the computational cost of this function is very high, increasing the computational pressure of the model. Summary of the Invention

[0008] The object of the present invention is to propose a method for predicting the film thickness of photovoltaic cells based on improved deep learning, with a fast calculation speed, so that the measured film thickness is more accurate.

[0009] To achieve the above object, the present invention provides a method for predicting the film thickness of a photovoltaic cell based on improved deep learning. The specific steps include:

[0010] (1) Collect the image information and film thickness value of the coated photovoltaic cell through a camera device. The image information includes image data of RGB three-channel gray values.

[0011] (2) Convert the RGB image data into HSL image data.

[0012] (3) Determine the correlation R based on the HSL image data and the film thickness value, and determine the image data with the strongest correlation with the actual coating thickness according to the correlation R. Determine the color depth value Y value according to the H value and L value in the HSL image data.

[0013] (4) Normalize the image data.

[0014] (5) Form a color depth value and film thickness prediction model based on the BP neural network.

[0015] (51) Introduce the Tanh function into the hidden layer of the BP neural network.

[0016] (52) Add a momentum term to the weight relationship formula between the hidden layer and the output layer of the BP neural network.

[0017] (53) Form a film thickness prediction model.

[0018] (6) Predict the coating thickness of the photovoltaic cell by the film thickness prediction model and the color depth value Y, and determine that the error between the predicted film thickness and the actual film thickness is within the allowable range.

[0019] In the above method, the image information of the coated photovoltaic cell is extracted by the imaging device, and the image data is extracted from the image information. The image data is processed, and the obtained RGB image data is converted into HSL image data. The image data with the largest correlation with the film thickness is selected as the basic data of the HSL image data for image processing, which can further reduce the amount of image data and improve the operation speed. At the same time, the color depth value Y representing the image parameters of the photovoltaic cell is generated by the HSL image data. Since the HSL image data is the manifestation of the image in the color space and its three values can be processed separately and independently, the basic image data is protected from external influences, and the correlation between the HSL image data and the actual film thickness is the best. However, the span and gap between the two HSL image data are relatively large. Therefore, the color depth value is formed between the two data H and L, and then the color depth value is used as the input value of the neural network to determine the film thickness, so that the data for determining the film thickness is more reliable. When predicting the image data through the neural network, first, the Tanh function is used as the activation function of the BP neural network to introduce nonlinearity into the model. Since the gradient of the Tanh function is steeper and the convergence speed of the function is faster, the accuracy of film thickness prediction is improved. And a momentum term is introduced into the weight relationship in the BP neural network to accelerate the algorithm convergence and jump out of the local minimum. When solving the minimum value of the network function, the reliability and speed of the solution are further improved.

[0020] Further, step (2) specifically includes that the conversion formula between RGB and HSL is:

[0021] ;

[0022] where .

[0023] The range of the value:

[0024] .

[0025] where H is the hue, S is the saturation, and L is the luminance; R, G, and B respectively represent the gray values on the red, green, and blue channels.

[0026] With the above settings, the HSL image data corresponding to the RGB image data is obtained through the conversion relationship, so that it is convenient to obtain the Y value through the HSL image data subsequently, enabling better prediction of the actual film thickness.

[0027] Furthermore, the calculation formula of the correlation R is as follows:

[0028] .

[0029] Wherein:

[0030] .

[0031] x is the input value of the image data, y is the film thickness value, is the average value of the image data, is the average value of the film thickness, and the value range is: .σ x is the variance of the image data, σ y is the variance of the film thickness, x i is the image data of the i-th image, y i is the film thickness value corresponding to the i-th image data.

[0032] With the above settings, by calculating the correlation, the image data with the best correlation between the image data and the actual film thickness can be confirmed, thereby reducing the data volume of the training samples and alleviating the operation pressure of the system.

[0033] Furthermore, the calculation formula of the color depth value Y is as follows:

[0034] Y = A * H+(1 - A)*L.

[0035] Wherein, A is the color depth value coefficient, and H and L are the HSL image data with the best correlation between the image data and the actual film thickness.

[0036] With the above settings, by determining the color depth parameter to determine the color depth value, the model trained by the color depth value and the actual film thickness avoids the interference of H and L on the neural network, and at the same time reduces the data volume of the training samples and alleviates the operation pressure of the system.

[0037] Furthermore, step (52) specifically includes adding a momentum term between the hidden layer and the output layer of the BP neural network, and its expression is:

[0038] .

[0039] Wherein, Δ is the difference between the weight at time t and the weight at time t - 1; w ij is the weight between node i and node j; is the preset ratio between the input layer and the hidden layer; α is the momentum coefficient, and the value range is: 0 <α <1.

[0040] With the above settings, adding a momentum term can reduce the oscillation trend and improve the stability of the training process, that is, when the parameters are updated, the previous direction trend can be maintained.

[0041] Further, the Tanh function in step (51) is as follows:

[0042] .

[0043] In the formula, x is the input value.

[0044] With the above settings, since the negative power of x is introduced in the Tanh function, the problem of the entire function gradient vanishing and high computational cost can be avoided.

[0045] Further, step (4) includes: .

[0046] Where X norm is the normalized image data, X is the size of the original image data, X max , X min is the maximum and minimum values of the original image data.

[0047] With the above settings, after the data is normalized, the dimensional expression becomes a dimensionless expression, which facilitates the comparison and weighting of two different units and different magnitudes of the color depth value and the film thickness, turns the dimensional data set into a scalar, achieves the effect of simplifying the calculation, and facilitates the calculation of the neural network.

[0048] Further, step (53) includes: using the LM regression algorithm as the training model of the BP neural network to finally obtain the film thickness prediction model.

[0049] With the above settings, selecting the LM algorithm as the training function of the network improves the solution speed and convergence accuracy of the function. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is the flowchart of the operation of the present invention.

[0051] Figure 2 is the grayscale value display diagram of the image data of the present invention.

[0052] Figure 3 is the flowchart of the BP neural network of the present invention.

[0053] Figure 4 is the schematic diagram of the BP neural network architecture of the present invention.

[0054] Figure 5 is the schematic diagram of the predicted average error curve of the present invention.

[0055] Figure 6 is the curve diagram of the actual film thickness, the predicted film thickness of the non-normalized processing model, and the predicted film thickness of the normalized processing model of the present invention.

[0056] Figure 7 Schematic diagram for predicting film thickness of the present invention Detailed implementation manners

[0057] The present invention will be further described in detail below in conjunction with the specific implementation manners

[0058] As Figure 1 shown, the present invention provides a method for predicting the film thickness of a photovoltaic cell based on improved deep learning, and the specific steps include:

[0059] (1) Collect the image information and film thickness value of the photovoltaic cell after coating through an imaging device, and the image information includes image data of RGB three-channel gray scale values;

[0060] (2) Convert the RGB image data into HSL image data;

[0061] (3) Determine the correlation R according to the HSL image data and the film thickness value, and determine the image data with the strongest correlation with the actual coating thickness according to the correlation R, and determine the color depth value Y value according to the H value and L value in the HSL image data; The image data with the strongest correlation can be set as the first 200 groups of image data with the strongest correlation, or can be set as image data of other groups;

[0062] (4) Normalize the image data;

[0063] (5) Form a color depth value and film thickness prediction model based on the BP neural network;

[0064] (51) Introduce the Tanh function into the hidden layer of the BP neural network;

[0065] (52) Add a momentum term to the weight relationship formula between the hidden layer and the output layer of the BP neural network;

[0066] (53) Form a film thickness prediction model;

[0067] (6) Predict the film thickness of the photovoltaic cell through the film thickness prediction model and the color depth value Y, and determine that the error between the predicted film thickness and the actual film thickness is within the allowable range.

[0068] Among them, step (2) specifically includes that the conversion formula between the RGB image data and the HSL image data is:

[0069] (1).

[0070] (2), where M1, M2 and I1 in formula (2) are obtained through formula (1).

[0071] Range of values:

[0072] (3).

[0073] Wherein, H is the hue, S is the saturation, and L is the lightness; R, G, and B respectively represent the color values on the red, green, and blue channels.

[0074] The HSL image data corresponding to the RGB image data is obtained through the conversion relationship, so that the actual film thickness can be predicted better.

[0075] In this embodiment, the original image information of the photovoltaic cell is as shown in the RGB image data and the HSL image data Figure 2 shown, where the image information also includes the outer contour features of the photovoltaic cell, which can be obtained by marking the outer dimensions of the cell on the photovoltaic cell, generating cross-shaped contour information, and then calibrating the acquisition area to obtain the RGB image data of this area. The RGB image data is the average RGB gray value.

[0076] The calculation formula for the correlation coefficient R is:

[0077] (4).

[0078] Where:

[0079] (5).

[0080] x is the input value of the image data, y is the film thickness value, is the average value of the image data, is the average value of the film thickness, and the value range is: . σ x is the variance of the image data, σ y is the variance of the film thickness, x i is the i-th image data, y i is the film thickness value corresponding to the i-th image data.

[0081] In this embodiment, by calculating the correlation coefficient, the image data with the best correlation with the actual film thickness is confirmed, thereby reducing the amount of data of the training samples and reducing the computing pressure on the system.

[0082] In this embodiment, through experiments, the correlation between the actual film thickness and the image data is shown in Table 1 below.

[0083]

[0084] Table 1

[0085] From the data in the table, it can be seen that the correlation between H (hue) and L (lightness) and the actual film thickness is the best. Since the span and gap between the two values are relatively large, it is obviously unwise to use them as inputs for deep learning training simultaneously, and it will generate interfering models.

[0086] However, both H and L are independent values. Therefore, a color depth value Y is designed to construct a functional relationship between the two:

[0087] (6).

[0088] Among them, A is the color depth value coefficient, and H and L are the image data with the best correlation with the actual film thickness in the image data.

[0089] The model obtained by training the color depth value and the actual film thickness avoids the interference of H and L on the neural network, reduces the amount of training samples, and alleviates the computing pressure on the system.

[0090] In this embodiment, as Figure 3 shown, the BP neural network (Back Propagation) is a neural network for backpropagation. It can learn and store a large number of input-output pattern mapping relationships without judging the functional form of this mapping relationship. Its characteristics are: data forward propagation and error backpropagation.

[0091] Each neuron receives input signals from other neurons. Each signal is transmitted through a connection with a weight w, and then the total input value is compared with the threshold b of the neuron, and finally the final output is obtained through an activation function.

[0092] In the BP neural network, a single sample has m inputs and n outputs. There are usually several hidden layers (h) between the input layer (i) and the output layer (o). Usually, the number of nodes in the input layer and the output layer is determined, and the number of hidden layer nodes has a greater impact on the performance of the neural network. According to the empirical formula, the number of hidden layer nodes h can be determined:

[0093] (7).

[0094] Among them, g is an adjustment constant between 1 and 10.

[0095] Now let the weight between node i and node j be w ij , the threshold of node j be b j , and the output value of each node be X j .

[0096] The activation function f is commonly the Sigmoid (logistic) function, also known as the output value S of the S-shaped growth curve. j When used in a classifier, the function has a better effect. In forward propagation, the output value of each node is achieved based on the output values of all nodes in the upper layer, the weights between the current node and all nodes in the previous layer, the threshold of the current node, and the activation function.

[0097] (8).

[0098] (9).

[0099] In the formula, X i is the data of the i-th image, and b j is the threshold of node j.

[0100] By continuously adjusting the weights and thresholds of the network in the reverse direction along the steepest descent of the sum of squared relative errors, according to the gradient descent method, the weights and thresholds between the input layer and the hidden layer are calculated as follows:

[0101] (10).

[0102] (11).

[0103] Among them, E(w,b) is the total error value generated by the weights and thresholds, is the total error value of the weights and thresholds between the k-th hidden layer under the i-th input value, is the first derivative of the error value between the i-th input layer and the k-th hidden layer, is the preset ratio between the input layer and the hidden layer.

[0104] To improve the training speed of the network, the following expression for the weight adjustment vector between the hidden layer and the output layer containing the momentum term is obtained:

[0105] (17).

[0106] In the formula, Δ is the difference between the weight at time t and the weight at time t-1, and 𝛼 is the momentum coefficient.

[0107] It can be seen from the above formula that adding the momentum term means taking a part from the previous weight value adjustment and superimposing it on the current weight adjustment amount. 𝛼 is called the momentum coefficient, and its value range is 0 < 𝛼 < 1. The physical meaning represented by the momentum term is the reflection of the accumulated experience before, which has a damping effect on time t. When the weight surface shows a sudden drop, this momentum term can reduce the oscillation trend and improve the stability of the training process, that is, when the parameters are updated, it can maintain the previous direction trend.

[0108] By adding a momentum term, the oscillation trend can be reduced and the stability of the training process can be improved, that is, when the parameters are updated, the previous direction trend can be maintained.

[0109] When adjusting the weights, the standard BP algorithm only adjusts according to the gradient direction of the error at time t, without considering the gradient direction before time t, which easily causes oscillations and slow convergence in the training process. The error between the actual output and the expected output of the system is reduced by changing the connection weights between neurons, that is, the error correction learning rule. The neural network architecture is as Figure 3 shown.

[0110] In this embodiment, an algorithm is proposed to improve the traditional BP neural network which has many defects: there are many local minima in the network, it is easy to fall into local minima; a large number of training times will reduce the learning efficiency and slow down the convergence speed.

[0111] Aiming at the problems of slow solution speed and convergence speed of the gradient descent method in the traditional BP algorithm, it is decided to normalize the data.

[0112] Normalization is essentially a linear transformation, and linear transformation has many good properties, such as not changing the numerical sorting of the original data, etc. These properties determine that the data will not be "invalidated" after being changed, but can improve the performance of the data.

[0113] When normalizing image data, it is more inclined to map the results to the range of [-1, 1]. The conversion function is as follows:

[0114] (15).

[0115] The above formula scales the image data proportionally. Among them, X norm is the normalized data, X is the size of the original data, X max , X min are the maximum and minimum values of the original data respectively. After data normalization, the dimensional expression becomes a dimensionless expression, which is convenient for comparing and weighting two different units and different magnitudes of indicators, namely the color depth value Y and the film thickness T. It can also turn the dimensional dataset into a scalar, and can play a role in simplifying calculations, making it more convenient for neural network operations.

[0116] Since the relationship between the color depth value of the image data and the actual film thickness is not a linear relationship, it is necessary to introduce non-linearity into the network to learn complex patterns, mainly in the hidden layer of the network. The choice of activation function in the output layer of the neural network depends on the type of problem we want to solve. For regression prediction problems, a linear function can be used.

[0117] Film thickness prediction belongs to the problem of data prediction regression. Therefore, the Tanh function is selected as the activation function of the hidden layer of the BP neural network to introduce non-linearity into the model.

[0118] (17).

[0119] Where x is the input value.

[0120] Although the Tanh function still has problems such as vanishing gradients and high computational costs, compared with the Sigmoid function, the Tanh function is centered around zero, making the optimization process easier, and the gradient of the Tanh function is steeper, so the convergence speed of the function is faster.

[0121] In the improved deep learning, the LM algorithm and the Gauss-Newton method are used as the training functions of the BP neural network. A damping factor is added to the LM algorithm, which can achieve good convergence and stability in different situations; while the Gauss-Newton method approximates the optimal solution by using the first-order and second-order derivative information of the objective function. The specific working principles of the LM algorithm and the Gauss-Newton method are prior arts and will not be elaborated here.

[0122] When the damping parameter is large, it is closer to the gradient descent method, and when it is far from the minimum value, the convergence effect is good; when the damping parameter is small, it is closer to the Gauss-Newton method, and when approaching the target, it can obtain a quadratic convergence speed and can quickly converge to the minimum value.

[0123] In this embodiment, the algorithm is experimentally verified through experimental data, and the specific process is as follows:

[0124] The imaging device includes a camera. The camera is a 20 million-pixel high-speed industrial camera of the area array AOI device. The photovoltaic cell samples are photographed by the camera and the image data is analyzed. The film thickness data corresponding to each cell is measured by an ellipsometer, and a total of 2185 groups of sample data sets are obtained.

[0125] Deep learning training parameters: The original size of the photovoltaic cell image is 1024×1024. The 2185 groups of data are divided into a training set (training), a test set (test), and a validation set (vaildation) according to the ratio of 7:1.5:1.5. Since training cannot be carried out endlessly, certain termination conditions must be designed: First, set the maximum number of iterations, and stop training when the data set is iterated to the specified number of times; Second, calculate the prediction accuracy or error of the training set in the network, and stop training when a certain threshold is reached.

[0126] The specific model training parameters are shown in Table 2 below.

[0127]

[0128] Table 2

[0129] The experiment set the target error (Performance) to 0.0000100001 and the detection point (ValidationValidation checkschecks) to 110000. Its function is to detect the network model during training. When the training error is detected to be less than the target error 110000 times, it proves that the network has converged to the minimum value at this time, ends the training, and saves the model.

[0130] Deep learning training uses the color depth value as the input, and the color depth value coefficient a as the weight of H and L. Its value ranges from 0.1 to 0.99. Since there are significant differences in H and L of the battery wafers produced by different processes or different production lines, the color depth value coefficient should be calculated separately according to different production lines or batches of battery wafers. For the two thousand groups of data in the experiment, the color depth value coefficient is calculated at a gradient of 0.1 with the two thousand groups of H and L in the experiment, and the resulting color depth value is used for deep learning training with the actual film thickness. The obtained model is used for 63 groups of film thickness predictions to obtain the error (unit: nmnm) between the actual film thickness and the predicted film thickness.

[0131] The prediction error of the color depth value training model is shown in Table 3 below.

[0132]

[0133] Table 3

[0134] As Figure 5 shown, for the dataset of this experiment, when the color depth value coefficient is 0.4, the average error of the trained model for film thickness prediction is the smallest, only 1.2370 nm. Subsequent experiments are all carried out on the basis of this color depth value.

[0135] The model obtained by deep learning training of 2185 groups of sample data is used to predict the film thickness of 63 battery wafers on the on-site production line. The PECVD coating of the photovoltaic battery wafers on this production line requires the film thickness to be between 64 - 78 nm. Battery wafers with a film thickness outside this range will be regarded as unqualified.

[0136] Before deep learning training, the data was normalized (NormalizationNormalization). Its improvement to the model is shown in Table 4 below. Compared with before normalization, the number of iterations (epochsepochs) is reduced by 5 times, the convergence speed is improved, the accuracy of the model for film thickness prediction is increased by 7.7%, and the average error is reduced by 0.5795 nm. As Figure 6The figure shows the curve graphs of the actual film thickness 1, the predicted film thickness 2 of the non - normalized processing model, and the predicted film thickness 3 of the normalized processing model, which can more clearly show the differences between the two models. Normalization processing transforms the dimensional dataset into a scalar, achieving the effect of simplifying calculations and further improving the performance of the model.

[0137]

[0138] Table 4

[0139] The improvement of the deep - learning neural network model is mainly in the BP back - propagation algorithm part. Aiming at the problems of the slow convergence speed of the Sigmoid function in the traditional BP algorithm, the difficult optimization process, the low efficiency of the gradient descent method as the training function, and the tendency to fall into local minima easily. Therefore, the Tanh function is selected as the activation function of the network to introduce nonlinearity, the LM algorithm is used as the training function of the network, and a momentum term is added in the weight adjustment step.

[0140] The improvement of the deep - learning neural network model is mainly in the BP back - propagation algorithm part. Aiming at the problems of the slow convergence speed of the Sigmoid function in the traditional BP algorithm, the difficult optimization process, the low efficiency of the gradient descent method as the training function, and the tendency to fall into local minima easily. Therefore, the Tanh function is selected as the activation function of the network to introduce nonlinearity, the LM algorithm is used as the training function of the network. The prediction effect of the final model on the film thickness is shown in Table 5 below. The Tanh activation function makes the optimization process easier and improves the convergence speed of the function. Compared with the model before improvement, the number of iterations is reduced by 8 times; using the LM algorithm as the training function of the network combines the advantages of the gradient descent method and the Gauss - Newton method, can adaptively adjust the convergence speed, improves the stability of the training process, enables the model to complete the regression task more accurately, the accuracy of film - thickness prediction is increased by 129%, and the average error is reduced by 1.558 nm; the added momentum term can reduce the oscillation trend during deep - learning training and improve the stability of the model prediction process, controlling the maximum error of film - thickness prediction within 4 nm. The predicted film - thickness error is as Figure 7 shown. The figure shows the error between the predicted film thickness and the actual film thickness of 63 battery wafers at the production site before and after the improvement of the network model.

[0141]

[0142] Table 5

[0143] Final experimental results: 1. Compared with before normalization, the number of iterations (epochspochs) decreased by 5 times, the convergence speed was improved, the accuracy of the model for film thickness prediction increased by 7.7%, and the average error decreased by 0.5795 nm. Normalization transforms the dimensional dataset into a scalar, achieving the effect of simplifying calculations and improving the performance of the model.

[0144] 2. The Tanh activation function makes the optimization process easier and improves the convergence speed of the function. Compared with the model before improvement, the number of iterations decreased by 8 times.

[0145] 3. Using the LM algorithm as the training function of the network combines the advantages of the gradient descent method and the Gauss-Newton method, can adaptively adjust the convergence speed, improve the stability of the training process, enable the model to complete the regression task more precisely, the accuracy of film thickness prediction increased by 12.9%, and the average error decreased by 1.558 nm.

[0146] 4. The added momentum term can reduce the oscillation trend during deep learning training, improve the stability of the model prediction process, and control the maximum error of film thickness prediction within 4 nm.

[0147] The working principle of the present invention: Extract the image information of the coated photovoltaic cell through the imaging device, extract the image data from the image information, process the image data, convert the obtained RGB image data into HSL image data, select the image data with the greatest correlation with the film thickness as the basic data of the HSL image data for image processing, which can further reduce the amount of image data and improve the running speed. At the same time, generate the color depth value Y representing the image parameters of the photovoltaic cell through the HSL image data. Since the HSL image data is the manifestation of the image in the color space and its three values can be processed separately and independently, the basic image data is protected from external influences, and the correlation between the HSL image data and the actual film thickness is the best. However, the span and gap between the two data of the HSL image data are relatively large, so a color depth value is formed between the two data H and L, and then the color depth value is used as the input value of the neural network to determine the film thickness, so that the data for determining the film thickness is more reliable. When predicting the image data through the neural network, first use the Tanh function as the activation function of the BP neural network to introduce non-linearity into the model. Since the gradient of the Tanh function is steeper and the convergence speed of the function is faster, the accuracy of film thickness prediction is improved, and a momentum term is introduced into the weight relationship in the BP neural network to accelerate the algorithm convergence and jump out of the local minimum. When solving the minimum value of the function in the network, the reliability and speed of the solution are further improved.

Claims

1. A method for predicting the film thickness of a photovoltaic cell based on improved deep learning, characterized in that: The specific steps include: (1) Collect the image information and film thickness value of the photovoltaic cell after coating through an imaging device. The image information includes the image data of the RGB three-channel gray scale value; (2) Convert the RGB image data into HSL image data; (3) Determine the correlation R based on the HSL image data and the film thickness value, and determine the image data with the strongest correlation with the actual coating thickness according to the correlation R. Determine the color depth value Y value according to the H value and L value in the HSL image data; (4) Normalize the image data; (5) Form a color depth value and film thickness prediction model based on the BP neural network; (51) Introduce the Tanh function into the hidden layer of the BP neural network; (52) Add a momentum term to the weight relationship formula between the hidden layer and the output layer of the BP neural network; (53) Form a film thickness prediction model; (6) Predict the coating thickness of the photovoltaic cell through the film thickness prediction model and the color depth value Y, and determine that the error between the predicted film thickness and the actual film thickness is within the allowable range.

2. A method for predicting the film thickness of a photovoltaic cell based on improved deep learning according to claim 1, characterized in that: Step (2) specifically includes that the conversion formula between RGB and HSL is: ； Among them ; Value range: ； Wherein, H is the hue, S is the saturation, and L is the brightness; R, G, and B respectively represent the gray scale values on the red, green, and blue channels.

3. A method for predicting the film thickness of a photovoltaic cell based on improved deep learning according to claim 1, characterized in that: The calculation formula of the correlation R is: ； Where: ； The input value of the x image data, and y is the film thickness value. is the average value of the image data. is the average value of the film thickness, and the value range is: ; σ x is the variance of the image data, and σ y is the variance of the film thickness, x i is the image data of the i-th image, y i is the film thickness value corresponding to the i-th image data.

4. A method for predicting the film thickness of a photovoltaic cell based on improved deep learning according to claim 1, characterized in that: The calculation formula of the color depth value Y is: ； Wherein, A is the color depth value coefficient, and H and L are the HSL image data with the best correlation with the actual film thickness in the image data.

5. A method for predicting the film thickness of a photovoltaic cell based on improved deep learning according to claim 1, characterized in that: The Tanh function in step (51) is: ； In the formula, x is the input value.

6. A method for predicting the film thickness of a photovoltaic cell based on improved deep learning according to claim 1, characterized in that: Step (4) includes: ; where X norm is the normalized image data, X is the size of the original image data, X max , X min is the maximum and minimum values of the original image data.

7. A method for predicting the film thickness of a photovoltaic cell based on improved deep learning according to claim 1, characterized in that: Step (53) includes: Using the LM regression algorithm as the training model of the BP neural network to finally obtain the film thickness prediction model.

Images