Implementation method of multilayer fusion dynamic optimization expert model architecture for photovoltaic production line

Abstract

The invention discloses a multilayer fusion dynamic optimization expert model architecture for a photovoltaic production line and an implementation method. The implementation method comprises the following steps: 1, generating a high-dimensional feature vector by using a convolutional neural network algorithm; 2, defining intelligent optimization modules designed for different key processes of the photovoltaic production line as expert models, wherein each model focuses on a dynamic optimization task of a specific process; 3, the welding optimization expert model uses deep reinforcement learning to optimize the combination of welding temperature, pressure and current; the lamination optimization expert model outputs lamination temperature and time based on an evolutionary optimization predictive control algorithm; the glue filling and curing optimization expert model is combined with Bayesian optimization and fuzzy logic control to adjust the glue amount, the glue filling speed and the environment temperature and humidity; and 4, taking the parameter result obtained by the optimization as the input of the next optimization to form a continuous iterative optimization process, so that the process parameters are continuously and effectively close to the optimal solution, and the assembly quality and the resource utilization efficiency of the production line are improved.

Description

Technical Field

[0001] The present invention discloses an implementation method for a multi-layer fusion dynamic optimization expert model architecture for a photovoltaic production line, belonging to the field of intelligent manufacturing of photovoltaic modules. Background Art

[0002] The production process of photovoltaic modules covers multiple complex technological processes such as cell welding, lamination, encapsulation gluing and curing. These processes are directly related to the power generation efficiency and long-term stability of the modules. However, the current production line still faces significant challenges in these key processes. In terms of the welding process, subtle changes in temperature, pressure and current may lead to welding defects such as false soldering, over-soldering or open circuit. The existing fixed threshold control method lacks the ability of intelligent optimization and is difficult to adapt to the differences between different batches of materials. In the lamination process, the application of fixed parameters makes it difficult for the equipment to adjust in a timely manner when facing the differences in raw material batches and environmental fluctuations. At the same time, it is difficult to find an ideal balance between the response speed of the lamination equipment and parameter adjustment in terms of product quality and production efficiency, which restricts the stability and reliability of the overall process. In addition, the control of the gluing and curing processes depends on manual experience adjustment and is easily affected by environmental changes, resulting in inconsistent glue curing states. The existing system lacks a real-time adaptive adjustment mechanism, which affects the consistency of module encapsulation. At present, the optimization of these processes mainly relies on rule control and empirical parameter setting, such as fixed parameter control. This method lacks flexibility and cannot cope with the fluctuations of raw materials and environmental factors. Although offline optimization can optimize parameters through historical data analysis, the update cycle is long and it is difficult to respond to the dynamic changes of the production line in real time. The adoption of a single control strategy lacks the ability to globally optimize and coordinate a multi-process, multi-variable system and cannot achieve efficient coordination and adaptive control among various processes. Summary of the Invention

[0003] In order to overcome the above deficiencies in the prior art, the present invention provides an expert model for multi-layer fusion and dynamic optimization for a photovoltaic cell production line. This model can intelligently provide process suggestions in real time, effectively respond to changes in raw materials and the environment, and achieve optimized control of the global process. The technical solution of the present invention is as follows:

[0004] A multi-layer fusion dynamic optimization expert model architecture and implementation method for a photovoltaic production line, comprising the following steps:

[0005] Collect data by sensors and control systems deployed at the welding workstation, lamination workstation, and gluing and curing workstation, and extract and fuse the multi-source process data of the photovoltaic production line through a convolutional neural network algorithm to generate a unified high-dimensional feature vector, which is used as the input for the subsequent optimization module;

[0006] Define 3 intelligent optimization modules for the processes of the photovoltaic production line. Each module serves as an independent expert model, respectively focusing on the dynamic optimization tasks of different processes;

[0007] Welding optimization expert model: Adopt the deep reinforcement learning method to optimize the combined parameters of welding temperature, pressure and current;

[0008] Lamination optimization expert model: Construct a predictive control algorithm based on evolutionary optimization to dynamically adjust the lamination temperature, pressure and time, and improve the quality of the photovoltaic lamination process;

[0009] Glue filling and curing optimization expert model: Determine the glue volume and glue filling speed parameters through the Bayesian optimization algorithm, and combine the fuzzy logic controller to adaptively adjust the ambient temperature and humidity to achieve the joint optimization of glue layer uniformity and curing efficiency;

[0010] According to the newly obtained real-time data and the results of the previous round of optimization, use the iterative optimization idea to adjust the process parameters, and use the optimized parameters as the input for the next round of optimization to ensure that the process parameters are continuously improved after each optimization, so that the production process continuously and effectively approaches the global optimal solution, thereby improving product quality, reducing production energy consumption and increasing production efficiency.

[0011] Furthermore, the specific process of data acquisition and preprocessing is as follows:

[0012] At the welding workstation, install temperature sensors near the welding head to monitor the change of welding temperature T w ; Deploy pressure sensors at the welding fixture to measure the pressure P w in real time during the welding process; Integrate current sensors at the output end of the welding power supply to detect the stability of welding current I w ; Obtain the time t w from the welding control system;

[0013] The data during the welding process is time-sequential, so it is necessary to combine the data at multiple time points to form a multi-dimensional time-sequential feature matrix X weld :

[0014]

[0015] In the matrix represents the welding temperature at the i-th sampling, represents the welding pressure at the i-th sampling, represents the welding current at the i-th sampling, represents the welding time at the i-th sampling, and n represents the number of samplings;

[0016] A one-dimensional convolutional network is used to extract key features from time-series data, which is particularly suitable for processing time-series data. By sliding the convolutional kernel, it can capture local time-series features and reflect the dynamic change law of welding parameters over time. The one-dimensional convolutional network is used to extract time-series features from time-series data:

[0017] F weld = Conv1D(X weld , W weld ) + b weld

[0018] Where X weld represents the welding data matrix, W weld , b weld represent the convolutional kernel weight and bias respectively, and F weld represents the welding feature, which is used as the input for the welding expert model;

[0019] In the lamination workstation, temperature sensors are arranged inside the laminator to achieve real-time temperature monitoring; sensors are arranged inside the lamination cavity to monitor pressure, avoiding component damage or uneven lamination caused by excessive local stress; the lamination duration is recorded from the lamination control system to achieve high-precision time control and analysis; in order to improve data accuracy and ensure the time consistency of each sensor's data, data filtering technology and timestamp synchronization mechanism are introduced;

[0020] Data is collected at a certain time step, and the structure of the time-series data can be represented as the following matrix. The data for each time step includes temperature T n , pressure P n and lamination duration t n :

[0021]

[0022] Before data modeling, the data must be properly preprocessed to improve the performance and training speed of the model. If there are missing sensor data, linear interpolation is used to fill in the missing values:

[0023]

[0024] Where y represents the interpolation at the missing position, and y1 and y2 represent the data before and after the missing data respectively;

[0025] In order to make different features have similar scales, data normalization operations are required to scale the data to the range of [0, 1]. After normalization, the scale differences between features are eliminated, which helps to improve the training effect of subsequent models:

[0026]

[0027] where X is the original data, X max , X min represent the minimum and maximum values of this feature respectively, and X norm is the data after normalization;

[0028] To capture the local dynamic features of time series data, local features of each time period are extracted by window sliding. The frequency domain features are calculated within the window to maintain trend information and reflect periodic fluctuations. Therefore, a time window of length W is used to extract local features from the time series data in order to capture the patterns in time series. Through window sliding, the following feature set X window is obtained:

[0029]

[0030] For the normalized data within each time window, its mean and variance are calculated to reflect its overall state and parameter volatility:

[0031]

[0032] where x t represents the data value at the t-th moment within the time window, W represents the size of the time window, μ represents the mean of the data within this time window, and σ 2 represents the variance of the data within this time window;

[0033] At the glue filling and curing workstation, a flow sensor is installed in the glue filling conveying pipeline to measure the actual amount of glue The curing equipment control system monitors and records the glue filling time of each component and the curing time Meanwhile, temperature and humidity sensors are installed inside the curing chamber to monitor the ambient temperature and humidity

[0034] The original glue filling and curing data matrix X glue where each row represents the process parameters of one sampling:

[0035]

[0036] In the matrix represents the amount of glue for the i-th sampling, represents the glue filling speed for the i-th sampling, represents the curing time for the i-th sampling, represents the ambient temperature for the i-th sampling, represents the ambient humidity for the i-th sampling, and n represents the number of samplings;

[0037] After data acquisition, all sensor data needs to be standardized to ensure that data with different units and magnitudes can be combined into the same model. Standardization is performed using the mean and standard deviation to ensure that the mean of each data is 0 and the standard deviation is 1:

[0038]

[0039] In the formula, F glue represents the standardized process data, and μ glue , σ glue represent the mean and standard deviation of the data respectively.

[0040] Furthermore, the specific process of the welding optimization expert model is as follows:

[0041] The state vector of the welding optimization expert model represents the parameters of the current welding process; in deep reinforcement learning, the model adjusts the combination of welding temperature, pressure, or current according to the current state:

[0042] A = [ΔT w , ΔP w , ΔI w

[0043] where ΔT w represents the increment or decrement of temperature; ΔP w represents the increment or decrement of pressure, and ΔI w represents the increment or decrement of current;

[0044] The goal of welding optimization is to make the welding quality meet the standard, improve production efficiency, and reduce energy consumption, that is, to maximize welding quality, minimize energy consumption, and minimize welding time. Therefore, the reward function R can be designed as a weighted sum of multiple goals:

[0045] R = w1·Strength - w2·EnergyCost

[0046] where Strength represents the mechanical strength of the solder joint, which is obtained by the ultrasonic detection probe installed at the welding workstation. The welding energy consumption EnergyCost is proportional to the welding current I w , welding voltage V w , and time t w . The weights w1, w2, w3 of each index are adjusted according to production requirements;

[0047] According to the deep Q - network (DQN) reinforcement learning algorithm, evaluate the Q - value of each action based on the current state Q(S, A; θ) = NeuralNet(S, A; θ), and dynamically adjust the process parameters such as welding temperature, pressure, and current by optimizing the policy, where θ represents the weight parameters of the neural network; ​

[0048] During the training phase, by continuously inputting real-time data of the welding process, the system will generate a large amount of sample data, which will be used to train the model. After each training, the model will update its strategy through Q learning and adjust the welding process parameters until the desired optimization goal is achieved;

[0049] DQN updates parameters by comparing the current Q value with the target Q value, which can be calculated using the following formula:

[0050]

[0051] Where Q target represents the ideal Q value, R represents the reward value directly fed back by the environment after the current state performs an action, the discount factor γ controls the weight of future rewards, S' represents the next state to be transferred to after the action is performed, A' represents the optional actions in the next state, and the parameters θ' of the target network are updated independently from the parameters of the main network for stable training;

[0052] The algorithm will continuously try to more accurately predict the cumulative rewards under different states and actions. The DQN algorithm uses the target network to calculate the target Q value. Its parameters are updated lazily, usually after the main network parameters are updated for a certain number of steps, the parameters of the main network are copied to the target network.

[0053] By minimizing the following loss function Indicates the current estimated Q value QS(,A;θ) and the target Q value Q target The difference between them reduces the instability of network training caused by frequent changes in the target Q value during training:

[0054]

[0055] The parameters of the neural network are constantly updated during the training process to optimize the network's prediction of the Q value. Back propagation is an algorithm used to train neural networks. It calculates the loss function based on the difference between the network's output and the target value. Therefore, the model parameters θ are updated through the back propagation algorithm to gradually reduce the error of the Q value, so that the strategy is gradually optimized:

[0056]

[0057] The learning rate η is used to control the update step size. Represents the gradient of the loss function with respect to the parameters.

[0058] Furthermore, the specific process of the lamination optimization expert model is as follows:

[0059] Each individual X represents a set of feasible lamination process parameters:

[0060] X=[Ti ,P i ,t i ]

[0061] Where T i ,P i ,t i represent lamination temperature, pressure and time respectively;

[0062] The optimization goal is to reduce the bubble rate, improve the bonding strength, and reduce energy consumption. The output of the optimal parameter conditions is: 1. When the fitness function converges, that is, the change is less than 1%, 2. The maximum number of iterations is reached, that is, 1000 times; define the comprehensive fitness function:

[0063] F(X)=-(w1×Q bubble +w2×Q delam +w3×E)

[0064] Where Q bubble Indicates the bubble ratio of the final component. The optical sensor is used to non-destructively detect the bubbles inside the component. The delamination rate Q delam It is used to measure the bonding strength. An ultrasonic detector is used to detect internal delamination defects through high-frequency sound wave reflection. E represents the total electrical energy consumed during the lamination process. The weight coefficients w1, w2, and w3 are used to adjust the importance of the optimization target.

[0065] From the current population P = [X1, X2, ···, XN], randomly select a subset, in which the fitness of each individual is calculated, and the individual X with the highest fitness is selected. best As the winner of the tournament, and replicate it to the next generation:

[0066]

[0067] In the crossover mutation operation, new individuals are generated by exchanging part of the genes of the two parent individuals, thereby increasing the diversity of the population and recombining excellent genes; through Gaussian mutation, fine search can be performed locally, and small changes in parameters can be adjusted to further optimize the objective function. The introduction of randomness helps to increase the diversity of the population and the global search ability, and avoid premature convergence of the algorithm; use Gaussian noise for mutation:

[0068] x i '=x i +e,e~N(0,σ 2 )

[0069] where x i represents the genetic parameter value of the current individual, e represents the random disturbance term that obeys the Gaussian distribution, and x i ' represents the new gene value after Gaussian mutation;

[0070] Using the dynamic model of the system, predict the state changes in the future for a period of time at each sampling moment, and then adjust the current control input through an online optimization control strategy, so that the system compensates for disturbances in advance and maintains the stability of process parameters. By minimizing the weighted sum of squares of the quality index error and the change in control input, MPC uses state-space equations to describe the dynamic changes of the lamination process:

[0071] X k+1 = AX k + BU k

[0072] Y k = CX k

[0073] Among them, the state variable X k represents the current temperature, pressure, and lamination duration, X k+1 represents the predicted state variable at the next moment, Y k represents the output variables at the current moment, the bubble ratio and the delamination rate. The control input U k represents the adjustment amplitude of the temperature, pressure, and lamination duration in the previous step. A, B, and C represent the state transition matrix, input matrix, and output matrix respectively;

[0074] Through quadratic programming solution, calculate the optimal control input and adjust the process parameters to make the actual value as close as possible to the target value. The optimization objective:

[0075]

[0076] Among them, N represents the number of terms in the summation; due to the different characteristics of different materials, in order to make the control strategy better adapted, it is necessary to adaptively adjust the weights w1 and w2 of MPC to adapt to different production environments, adapt to different batches of materials faster, and improve production stability:

[0077]

[0078] Among them, the learning rate α is used to control the optimization speed, is the gradient calculated through historical data, and t represents the number of rounds of parameter updates.

[0079] Furthermore, the specific process of the glue filling and curing optimization expert model is as follows:

[0080] Construct an objective loss function, and by minimizing the difference between the actual process and the target process, optimize the parameters of the glue volume, glue filling speed, and curing time, so that the photovoltaic module reaches the best state during the glue filling and curing process, thereby improving the process quality of the production line:

[0081]

[0082] Among them They are the target glue volume, glue filling speed, and curing time respectively. It represents the process parameters of the i-th sampling. w1, w2, and w3 are the optimization weights of different parameters.

[0083] Gaussian process regression does not require a prior assumption of the specific form of the objective function. Instead, it makes predictions by modeling the Gaussian distribution of the latent function. By modeling the existing data, Gaussian process regression can estimate the distribution of the optimization objective function, thereby predicting the output in the unknown region. The modeled objective function is:

[0084]

[0085] Among them, μ(x) represents the mean function of the objective function, and the covariance function k(x, x') is used to measure the correlation between samples. This model can describe the influence of each parameter on the objective function and provide a reference for parameter adjustment.

[0086] Based on a large amount of production data, by calculating the expected improvement effects under different parameter combinations, comprehensively evaluating multiple possible parameter combinations, directly pointing to the optimal parameter selection, the optimal glue filling parameters are selected by maximizing the expected improvement value. Calculate the expected improvement criterion:

[0087]

[0088] Among them, f * is the current optimal value, and the optimal glue filling parameters are selected by maximizing EI(x).

[0089] According to the real-time uncertain temperature and humidity data and the preset control rules, the control strategy is dynamically adjusted. By using fuzzy sets and fuzzy rules to simulate the human decision-making process, the flexible adjustment of temperature and humidity parameters is realized. The fuzzy logic control system will make adaptive adjustments according to the actual production situation and feedback information. The temperature and humidity parameters will be input into the inference system for calculation:

[0090]

[0091] Among them, w i is the membership degree weight of the fuzzy rule, T e and H e represent the optimized temperature and humidity values respectively. During the optimization process, the Bayesian optimization method will search for the optimal solution according to the current estimate of the objective function, while the fuzzy logic control is used to ensure that the adjustment range of each parameter is within reasonable process conditions. The combination of Bayesian optimization and fuzzy logic control can quickly find the optimal solution within a large range and ensure the stability of the system under rapidly changing production conditions.

[0092] Furthermore, the optimized parameters are used as the input for the next round of optimization. The specific iterative optimization process is as follows:

[0093] During the production process of photovoltaic modules, continuous optimization of process parameters is the key to ensuring high-quality and low-energy consumption production. To achieve a more suitable production effect, the architecture continuously performs iterative optimization of process parameters. The result of each optimization is used as the input for the next optimization to ensure that each round of optimization is better than the previous one. When the data shows a specific trend after multiple consecutive process cycles, the parameters of the production line equipment are dynamically adjusted to optimize the production process and improve product quality;

[0094] The goal of the welding process is to ensure welding quality, improve production efficiency, and reduce energy consumption. The optimization goal is to minimize the objective function. Define the welding objective function J weld :

[0095] J weld = w1·(Q weld - Q target ) 2 + w2·E weld

[0096] where Q weld represents the welding strength, Q target represents the ideal welding strength, E weld represents the energy consumption during the welding process, and w1, w2 represent the weights of different objectives;

[0097] The main objectives of the lamination process are to increase the peel strength, reduce the bubble rate, and optimize the lamination time and temperature. The goal is to reduce the bubble rate and ensure that the peel strength reaches the predetermined target. Design the lamination objective function as J laminate :

[0098]

[0099] where Q peel represents the peel strength, Q target represents the ideal peel strength, B bubble represents the bubble rate, and w1, w2 represent the weights of different objectives;

[0100] The goal of the potting and curing process is to optimize the potting amount, potting speed, curing temperature, and humidity, while ensuring the uniformity of the glue layer, curing quality, and efficiency, minimizing the standard deviation of the glue layer thickness, and ensuring that the curing strength reaches the target. Design the potting and curing objective function as:

[0101]

[0102] where a thickness represents the standard deviation of the glue layer thickness, S curing represents the curing strength, S targetIt represents the ideal peeling strength, and w1 and w2 represent the weights of different objectives;

[0103] The process parameters after each optimization will be used as the input for the next optimization. These optimization results reflect the effects of the previous round of process adjustments, denoted as J. old , The optimized parameters are calculated through the objective function and further optimized based on new real-time data. The new objective value is denoted as j. new , As the number of iterations increases, the value of the objective function will gradually converge. Set a stopping condition:

[0104] μ = j new -j old ≤ ∈

[0105] After each round of iterative optimization, record the minimum value of μ. When the change in the objective function is less than or equal to the threshold ∈, terminate the optimization process; for different thresholds in the architecture, take ∈ weld = 0.01, ∈ laminate = 0.02, ∈ dispensing = 0.015.

[0106] Beneficial effects

[0107] The present invention conducts global optimization on the welding, lamination, potting, and curing processes in the production of photovoltaic modules based on an intelligent optimization expert model. Compared with the prior art, it has the following significant advantages:

[0108] 1. Adopting a multi-layer expert model architecture to achieve process adaptive optimization: The present invention uses a multi-agent expert model to modularize the key processes such as welding, lamination, potting, and curing in the production of photovoltaic modules. Each process is optimized by an independent expert model to ensure the optimization accuracy.

[0109] 2. Multivariate data fusion and cross-process comprehensive decision-making: By deploying sensors at the welding, lamination, potting, and curing workstations, data from different processes are collected, and then a convolutional neural network is used to extract and fuse features from multi-source data to generate high-dimensional feature vectors. This not only captures information from a single process but also reflects the overall state of the entire production line in subsequent analysis.

[0110] 3. End-to-end closed-loop optimization to improve the intelligence level of the photovoltaic module production line: The present invention adopts a real-time feedback mechanism to form a complete closed-loop of data collection - expert model optimization - iterative approximation of the optimal value, forming a self-learning, self-optimizing, and self-adaptive intelligent production system, improving the quality of photovoltaic modules and production efficiency. Description of the drawings

[0111] Figure 1 Intelligent optimization system architecture diagram of the expert model for the photovoltaic module production line;

[0112] Figure 2 Flow chart of online dynamic adjustment and adaptive optimization

[0113] Figure 3 Flow chart of production line data stream and feedback control Specific implementation manners

[0114] Next, the technical solutions in the embodiments of the present invention will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on these embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0115] As Figure 1 shown, a multi-layer fusion dynamic optimization expert model architecture and implementation method for a photovoltaic production line includes the following steps:

[0116] Collect data by sensors and control systems deployed at the soldering workstation, lamination workstation, and potting and curing workstation, and extract and fuse the multi-source process data of the photovoltaic production line through a convolutional neural network algorithm to generate a unified high-dimensional feature vector as the input for the subsequent optimization module;

[0117] Define 3 intelligent optimization modules for the photovoltaic production line processes. Each module serves as an independent expert model and focuses on the dynamic optimization tasks of different processes respectively;

[0118] Soldering optimization expert model: Adopt a deep reinforcement learning method to optimize the combined parameters of soldering temperature, pressure, and current;

[0119] Lamination optimization expert model: Construct a predictive control algorithm based on evolutionary optimization to dynamically adjust the lamination temperature, pressure, and time to improve the quality of the photovoltaic lamination process;

[0120] Potting and curing optimization expert model: Determine the glue volume and potting speed parameters through a Bayesian optimization algorithm, and adaptively adjust the environmental temperature and humidity in combination with a fuzzy logic controller to achieve the joint optimization of glue layer uniformity and curing efficiency;

[0121] According to the newly obtained real-time data and the previous round of optimization results, use the iterative optimization idea to adjust the process parameters, and use the optimized parameters as the input for the next round of optimization to ensure that the process parameters are continuously improved after each optimization, so that the production process continuously and effectively approaches the global optimal solution, thereby improving product quality, reducing production energy consumption, and increasing production efficiency.

[0122] At the welding workstation, a temperature sensor is installed near the welding head to monitor the change in welding temperature; a pressure sensor is deployed at the welding fixture to measure the pressure during welding in real time; a current sensor is integrated at the output end of the welding power supply to detect the stability of the welding current.

[0123]

[0124] Calculate the mean μ weld and the standard deviation σ weld Perform normalization:

[0125]

[0126] Use a convolutional kernel to extract the time series:

[0127]

[0128] The convolution calculation results in:

[0129]

[0130] In the reinforcement learning optimization, the learning rate η is 0.01, and γ represents the discount factor of 0.95. According to the current state, after optimization, adjust the welding temperature, pressure, and current to:

[0131] ΔT w =20,ΔP w =2,ΔI w =3

[0132] At the lamination workstation, temperature sensors are arranged inside the laminator to achieve real-time temperature monitoring; sensors are arranged inside the lamination cavity to monitor the pressure, avoiding component damage or uneven lamination caused by excessive local stress; record the lamination duration from the lamination control system to achieve high-precision time control and analysis; in order to improve data accuracy and ensure the time consistency of sensor data, introduce data filtering technology to remove noise and a timestamp synchronization mechanism.

[0133] During the lamination process, 9 groups of original data are collected as shown below:

[0134]

[0135] After normalization:

[0136]

[0137] Calculate the average value and the standard deviation:

[0138] μ T =149.28,σ T =1.70,μ P =0.47

[0139] σ P = 0.00085, μ t = 428, σ t = 4.44

[0140] The window size is selected as 3, and the weight coefficients in the fitness function are w1 = 0.4, w2 = 0.4, w3 = 0.2. After optimization, the results are obtained as follows:

[0141] T L = 149.8, P L = 0.47, t L = 425

[0142] In the glue filling and curing workstation, the flow sensor is installed in the glue filling conveying pipeline to measure the actual glue volume and the glue filling speed; the curing equipment control system monitors and records the curing time of each component; meanwhile, temperature and humidity sensors are installed inside the curing chamber to monitor the environmental temperature and humidity.

[0143] The original data matrix collected during a production process:

[0144]

[0145] To ensure the scale consistency between different data dimensions, the mean and standard deviation of each parameter are calculated:

[0146]

[0147] Calculation results:

[0148]

[0149] The weights of each index in the target loss function:

[0150] L(V g , v g , t e ) = 0.5·|V g - 140| + 0.3·|v g - 14| + 0.2·|t c - 220|

[0151] After Bayesian optimization and fuzzy logic adjustment, the optimization results are as follows:

[0152]

[0153] The optimized parameters are calculated through the target function and feedback adjustment for the welding process:

[0154] Q weld = 4.8, E weld = 1.0

[0155] In the lamination process, potting and curing processes:

[0156] Q peel = 4.9, B bubble = 0.01, a thickness = 0.3, S curing = 4.7

[0157] Substitute into the objective function for calculation:

[0158] J weld_old = 0.42, J laminate_old = 0.36, J dispensing_old = 0.32

[0159] The process parameters after each optimization will be used as the input for the next optimization and further optimized based on new real-time data. These optimization results reflect the effects of the previous round of process adjustments; (660, 22, 13, 0.3) is used as the input for the welding optimization expert model, (149.8, 0.47, 425) is used as the input for the lamination optimization expert model, and (140, 14, 242, 27, 0.55) is used as the input for the potting and curing optimization expert model to continuously iterate and optimize various process parameters;

[0160] After optimization, adjust the welding temperature, pressure, and current to (647, 27, 14), the lamination temperature, pressure, and time to (151, 0.47, 430), and the glue amount, potting speed, curing time, ambient temperature, and ambient humidity to (138, 15, 240, 24, 0.52). From the objective function:

[0161] J weld_new = 0.45, J laminate_new = 0.38, J dispensing_new = 0.35

[0162] Among them, J laminate_new - J laminate_old ≤ ∈ laminate , the optimized parameters of the lamination optimization expert model have tended to the optimal state and meet the termination conditions. The lamination temperature, pressure, and time can be reset when the production of the current batch stops, and they also serve as the initial values for subsequent production of the same raw materials.

[0163] Finally, it should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the said element.

[0164] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An implementation method of a multi-layer fusion dynamic optimization expert model architecture for a photovoltaic production line, characterized in that Including: Optimization models are separately configured for the welding, lamination, potting, and curing processes of the photovoltaic production line; among them: Data is collected by sensors and control systems deployed at the welding workstation, lamination workstation, potting and curing workstation, and the convolutional neural network algorithm is used to extract and fuse the multi-source process data of the photovoltaic production line to generate a unified high-dimensional feature vector as the input for the subsequent optimization module; Three intelligent optimization modules for the photovoltaic production line process are defined, and each module serves as an independent expert model, respectively focusing on the dynamic optimization tasks of different processes; Welding optimization expert model: The deep reinforcement learning method is used to optimize the combined parameters of welding temperature, pressure, and current; Lamination optimization expert model: A predictive control algorithm based on evolutionary optimization is constructed to dynamically adjust the lamination temperature, pressure, and time to improve the quality of the photovoltaic lamination process; Potting and curing optimization expert model: The Bayesian optimization algorithm is used to determine the glue volume and potting speed parameters, and the fuzzy logic controller is combined to adaptively adjust the environmental temperature and humidity to achieve the joint optimization of glue layer uniformity and curing efficiency; According to the collected real-time data and the results of the previous round of optimization, the iterative optimization idea is used to adjust the process parameters, and the optimized parameters are continued to be used as the input for the next round of optimization to ensure that the process parameters are continuously improved after each optimization, so that the production process continuously and effectively approaches the global optimal solution, thereby improving product quality, reducing production energy consumption, and increasing production efficiency.

2. The method according to claim 1, wherein The data collection and preprocessing process is as follows: At the welding workstation, a temperature sensor is installed near the welding head to monitor the change of welding temperature T w ; A pressure sensor is deployed at the welding fixture to measure the pressure P during the welding process in real time w ; A current sensor is integrated at the output end of the welding power supply to detect the stability of the welding current I w ; The welding control system obtains the time t w ; The data during the welding process is time-series, so it is necessary to combine the data at multiple time points to form a multi-dimensional time-series feature matrix X weld : In the matrix represents the welding temperature of the i-th sampling, represents the welding pressure of the i-th sampling, represents the welding current of the i-th sampling, represents the welding time of the i-th sampling, and n represents the number of samplings; One-dimensional convolutional network is used to extract key features from time series data, which is especially suitable for processing time series data. The local time series features can be captured through the sliding convolutional kernel, reflecting the dynamic change law of welding parameters over time. The one-dimensional convolutional network is used to extract time series features from time series data: F weld = Conv1D(X weld , W weld ) + b weld where X weld represents the welding data matrix, W weld , b weld represent the convolution kernel weights and biases respectively, and F weld represents the welding feature for input to the welding expert model; At the lamination workstation, temperature sensors are arranged inside the laminator to achieve real-time temperature monitoring; sensors are arranged inside the lamination cavity to monitor the pressure to avoid component breakage or uneven lamination caused by excessive local stress; the lamination duration is recorded from the lamination control system to achieve high-precision time control and analysis; in order to improve data accuracy and ensure the time consistency of each sensor data, data filtering technology is introduced to remove noise and timestamp synchronization mechanism; Collect data at a certain time step. The structure of the time series data can be represented as the following matrix. The data for each time step includes the temperature T n represents the temperature and pressure P collected at the nth time n represents the pressure and lamination duration t collected at the nth time n represents the duration collected at the nth time: Before data modeling, the data must be properly preprocessed to improve the performance and training speed of the model. If some sensor data is missing, linear interpolation is used to fill in the missing values: Where y represents the interpolation at the missing position, and y1 and y2 represent the data before and after the missing data respectively; In order to make different features have similar scales, data normalization operation is required to scale the data to the range of [0,1]. After standardization, the scale differences between various features are eliminated, which helps to improve the training effect of the subsequent model: where X is the original data, X max , X min represent the minimum and maximum values of this feature respectively, and X norm is the data after normalization; To capture the local dynamic features of time series data, local features for each time period are extracted by window sliding. Frequency domain features are calculated within the window to maintain trend information and reflect periodic fluctuations. Therefore, a time window of length W is used to extract local features from the time series data in order to capture the patterns in time series. The following feature set X is obtained by window sliding window : For the normalized data within each time window, calculate its average value and variance to reflect its overall state and parameter volatility: where x t represents the data value at the t-th moment within the time window, W represents the size of the time window, μ represents the mean of the data within the time window, and σ 2 represents the variance of the data within the time window; At the glue filling and curing workstation, a flow sensor is installed in the glue filling conveyor pipeline to measure the actual amount of glue The curing equipment control system monitors and records the glue filling time of each component and the curing time Meanwhile, temperature and humidity sensors are installed inside the curing chamber to monitor the ambient temperature and humidity Original potting and curing data matrix X glue Each row represents the process parameters of each sampling: In the matrix represents the amount of glue in the i-th sampling, represents the glue filling speed in the i-th sampling, represents the curing time in the i-th sampling, represents the ambient temperature in the i-th sampling, represents the ambient humidity in the i-th sampling, and n represents the number of samplings; After data collection, all sensor data needs to be standardized to ensure that data of different units and magnitudes can be merged into the same model. The mean and standard deviation are used for standardization to ensure that the mean of each data is 0 and the standard deviation is 1: where F glue represents the standardized process data, μ glue , σ glue represent the mean and standard deviation of the data, respectively.

3. The method according to claim 1, characterized in that The welding optimization expert model method is: State vector of the welding optimization expert model Represents the parameters of the current welding process; in deep reinforcement learning, A represents the action taken, and the model adjusts the combination of welding temperature, pressure, or current based on the current state: A = [ΔT w , ΔP w , ΔI w ​ where ΔT w represents an increase or decrease in temperature; ΔP w represents an increase or decrease in pressure, and ΔI w represents an increase or decrease in current; The goal of welding optimization is to make the welding quality meet the standard, improve production efficiency and reduce energy consumption, that is, to maximize welding quality, minimize energy consumption and minimize welding time. To this end, the reward function R can be designed as the weighted sum of multiple objectives: R=w1·Strength-w2·EnergyCost Among them, Strength represents the mechanical strength of the solder joint, which is obtained by the ultrasonic detection probe installed at the welding workstation. The welding energy consumption EnergyCost is proportional to the welding current I w , the welding voltage V w , and the time t w . The weights w1, w2, w3 of each index are adjusted according to production requirements; According to the reinforcement learning algorithm DQN, the Q value of each action is evaluated based on the current state, Q(,A;θ) = NeuralNet(S,A;θ), and the process parameters such as welding temperature, pressure, and current are dynamically adjusted through optimization strategies, where θ represents the weight parameter of the neural network; During the training phase, the system generates a large amount of sample data by continuously inputting real-time data of the welding process. These data are used to train the model. After each training, the model updates its strategy through Q learning and adjusts the welding process parameters until the desired optimization goal is achieved. DQN updates parameters by comparing the current Q value with the target Q value, which can be calculated using the following formula: where Q target represents the ideal Q value, R represents the reward value directly feedback by the environment after executing the action in the current state, the discount factor γ controls the weight of future rewards, S' represents the next state transferred to after executing the action, A' represents the actions available in the next state, and the parameters θ' of the target network are updated independently of the parameters of the main network for stable training; The algorithm will continuously try to more accurately predict the cumulative rewards under different states and actions. The DQN algorithm uses the target network to calculate the target Q value. Its parameters are updated lazily, usually after the main network parameters are updated for a certain number of steps, the parameters of the main network are copied to the target network. By minimizing the following loss function represents the difference between the currently estimated Q-value Q(S, A; θ) and the target Q-value Q target to reduce the instability problem of network training caused by the frequent change of the target Q-value during the training process: The parameters of the neural network are constantly updated during the training process to optimize the network's prediction of the Q value. Back propagation is an algorithm used to train neural networks. It calculates the loss function based on the difference between the network's output and the target value. Therefore, the model parameters θ are updated through the back propagation algorithm to gradually reduce the error of the Q value, so that the strategy is gradually optimized: The learning rate η is used to control the update step size, which represents the gradient of the loss function with respect to the parameters.

4. The method according to claim 1, wherein The laminate optimization expert model approach is: Each individual X represents a set of feasible lamination process parameters: X = [T i , P i , t i ​ where T i , P i , t i represent the lamination temperature, pressure, and time, respectively; The optimization goal is to reduce the bubble rate, improve the bonding strength, and reduce energy consumption. The output optimal parameter conditions are: condition 1, when the fitness function converges, that is, the change is less than 1%, condition 2, when the maximum number of iterations is reached, that is, 1000 times; Define the comprehensive fitness function: F(X = -(w1×Q bubble + w2×Q delam + w3×E) Among which Q bubble represents the proportion of bubbles in the final component. The bubbles inside the component are non-destructively detected using an optical sensor, and the delamination rate Q delam is used to measure the bonding strength. The internal delamination defects are detected using an ultrasonic detector through the reflection of high-frequency sound waves. E represents the total electrical energy consumed during the lamination process. The weight coefficients w1, w2, and w3 are used to adjust the importance of the optimization objectives; Randomly select a subset from the current population P = [X1, X2, ···, XN]. In this subset, calculate the fitness of each individual and select the individual X with the highest fitness best as the winner of the tournament and copy it to the next generation: In the crossover mutation operation, new individuals are generated by exchanging part of the genes of the two parent individuals, thereby increasing the diversity of the population and recombining excellent genes; through Gaussian mutation, fine search can be performed locally, and small changes in parameters can be adjusted to further optimize the objective function. The introduction of randomness helps to increase the diversity of the population and the global search ability, and avoid premature convergence of the algorithm; use Gaussian noise for mutation: x i ' = x i + e, e ~ N(0, σ 2 ) where x i represents the gene parameter value of the current individual, e represents a random perturbation term subject to a Gaussian distribution, and x i ' represents the new gene value after Gaussian mutation; Using the dynamic model of the system, predict the state changes within a certain period in the future at each sampling moment, and then adjust the current control input through an online optimization control strategy, so that the system compensates for disturbances in advance and maintains the stability of process parameters. By minimizing the weighted sum of squares of the quality index error and the change amount of the control input, MPC uses the state space equation to describe the dynamic changes of the lamination process: X k+1 = AX k + BU k Y k = CX k Among them, the state variable X k represents the current temperature, pressure, and lamination duration. X k+1 represents the predicted state variable at the next moment. Y k represents the output variable bubble ratio and delamination rate at the current moment. The control input U k represents the adjustment amplitude of the temperature, pressure, and lamination duration in the previous step. A, B, and C respectively represent the state transition matrix, input matrix, and output matrix; Through quadratic programming solution, calculate the optimal control input, adjust the process parameters, make the actual value as close as possible to the target value, and the optimization objective is: where N represents the number of terms for summation; due to the different characteristics of different materials, in order to make the control strategy better adapted, it is necessary to adaptively adjust the weights w1 and w2 of MPC to adapt to different production environments, adapt to different batches of materials faster, and improve production stability: where the learning rate α is used to control the optimization speed, is the gradient calculated from historical data, and t represents the round of parameter update.

5. The method according to claim 1, wherein The method of the glue filling and curing optimization expert model is: Construct an objective loss function, optimize the parameters of the glue volume, glue filling speed and curing time by minimizing the difference between the actual process and the target process, so that the photovoltaic module reaches the best state during the glue filling and curing process, thereby improving the process quality of the production line: Among them are the target glue volume, glue filling speed, and curing time respectively represents the process parameters of the i-th sampling, and w1, w2, and w3 are the optimization weights of different parameters Gaussian process regression does not need to pre-assume the specific form of the objective function, but makes predictions by modeling the Gaussian distribution of the latent function; Adopting Gaussian process regression can estimate the distribution of the optimization objective function by modeling the existing data, so as to predict the output of the unknown region, and the modeling objective function is: where μ(x) represents the mean function of the objective function, and the covariance function k(x, x') is used to measure the correlation between samples. This model can express the influence of each parameter on the objective function and provide a reference for parameter adjustment; Based on a large amount of production data, by calculating the expected improvement effects under different parameter combinations, comprehensively evaluate multiple possible parameter combinations, and try to select the optimal parameter combination. By maximizing the expected improvement value, the optimal glue filling parameters are selected, and the expected improvement criterion is calculated by the following formula: where f * is the current optimal value, and the optimal encapsulation parameters are selected by maximizing EI(x); According to the real-time uncertain temperature and humidity data and the preset control rules, dynamically adjust the control strategy, simulate the human decision-making process through fuzzy sets and fuzzy rules, and realize the flexible adjustment of the temperature and humidity parameters. The fuzzy logic control system will make adaptive adjustments according to the actual production situation and feedback information, and the temperature and humidity parameters will be input into the inference system for calculation: where w i is the membership degree weight of the fuzzy rule, and T e , H e represent the optimized temperature and humidity values respectively; During the optimization process, the Bayesian optimization method will search for the optimal solution according to the current estimate of the objective function, while the fuzzy logic control is used to ensure that the adjustment range of each parameter is under reasonable process conditions; The combination of Bayesian optimization and fuzzy logic control can quickly find the optimal solution within a large range and ensure that the system remains stable under rapidly changing production conditions.

6. The method according to claim 1, wherein The specific process of taking the optimized parameters as the input for the next round of optimization and iterative optimization is: During the production of photovoltaic modules, the continuous optimization of process parameters is the key to ensuring high-quality and low-energy consumption production. To achieve a more suitable production effect, the architecture continuously performs iterative optimization of process parameters. The result of each optimization will be used as the input for the next optimization to ensure that each round of optimization is better than the previous one. When the data shows a specific trend after multiple consecutive process cycles, the parameters of the production line equipment are dynamically adjusted to optimize the production process and improve product quality; The goal of the welding process is to ensure welding quality, while improving production efficiency, reducing energy consumption. The optimization goal is to minimize the objective function, and the welding objective function J is defined weld : J weld = w1·(Q weld - Q target ) 2 + w2·E weld Among them, Q weld represents the welding strength, and Q target represents the ideal welding strength. E weld represents the energy consumption during the welding process, and w1 and w2 represent the weights of different objectives; The main objectives of the lamination process are to increase the peel strength, reduce the bubble rate, and optimize the lamination time and temperature. The goal is to reduce the bubble rate and ensure that the peel strength reaches the predetermined target. The lamination objective function is designed as J laminate : Among them, Q peel represents the peel strength, and Q target represents the ideal peel strength. B bubble represents the bubble rate, and w1, w2 represent the weights of different targets; The goal of the potting and curing process is to optimize the glue volume, potting speed, curing temperature and humidity, while ensuring the uniformity of the glue layer, curing quality and efficiency, minimizing the thickness standard deviation of the glue layer, and ensuring that the curing strength reaches the target. The potting and curing objective function is designed as: where a thickness represents the standard deviation of the adhesive layer thickness, S curing represents the curing strength, S target represents the ideal peel strength, and w1, w2 represent the weights for different targets; The process parameters after each optimization will be used as the input for the next optimization. These optimization results reflect the effects of the previous round of process adjustments and are denoted as J old , the optimized parameters are calculated through the objective function and further optimized based on new real-time data. The new objective value is denoted as J new , as the number of iterations increases, the value of the objective function will gradually converge. A stopping condition is set, where μ represents the difference between the objective functions of two consecutive rounds: μ = j new -j old ≤ ∈ After each round of iterative optimization, the minimum μ value is recorded. When the change in the objective function is less than or equal to the threshold ∈, the optimization process is terminated; For different thresholds in the architecture, the threshold of the welding model ∈ weld = 0.01, the threshold of the lamination model ∈ laminate = 0.02, the threshold of the potting and curing model ∈ dispensing = 0.015.

Images