Laser scribing planning method, device and equipment and storage medium

Abstract

The invention relates to the technical field of laser scribing, and discloses a laser scribing planning method, device and equipment and a storage medium, and the method comprises the steps: carrying out the kinematics and laser-material interaction dynamics modeling of a dual-light-path laser system, and carrying out the preprocessing of historical scribing data, and obtaining a preprocessing time sequence data set; inputting the preprocessed time sequence data set into a time feature extractor to execute time dynamic feature extraction of the historical lineation trajectory to obtain a time feature vector; calculating spatial interaction features between the double optical paths to obtain spatial feature vectors; inputting the time feature vector and the space feature vector into a mixed density network, and establishing a material response probability model; and based on the material response probability model, performing layered optimization on a lineation path to obtain an optimal control parameter sequence of the dual-optical-path laser head. According to the invention, the global optimization of the double-light-path laser scribing path is realized, the scribing quality and efficiency are balanced, and the scribing positioning precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of laser scribing, and particularly to a laser scribing planning method, device, equipment and storage medium. Background Art

[0002] Laser scribing technology is the core process in the manufacturing of solar cells and has a decisive impact on cell efficiency and product quality. With the development of high-efficiency solar cell technology, the requirements for laser scribing accuracy are constantly increasing. However, most of the existing laser scribing planning methods face the problem of low accuracy. This problem mainly stems from the highly complex and dynamic physical properties between solar cell materials. Different types of solar cell materials have significant differences in the absorption, conduction, and response to laser energy, and these responses have great uncertainties in actual production, making it difficult for traditional deterministic planning methods to cope with the randomness and volatility of material responses.

[0003] Dual-beam laser scribing equipment has important application value in the manufacturing of solar cells due to its high precision, high efficiency, and flexibility, especially for high-efficiency solar cells that require precise division of electrode regions. However, such equipment faces many challenges during use: factors such as the power limitation and movement speed constraint of the laser equipment, the diversity of solar cell materials, and the dynamic changes in the production environment make it particularly difficult to improve production efficiency while ensuring scribing quality. In addition, the dual-beam laser system itself has high complexity, and the collaborative work between the two beams and the precise control of the overall system further increase the difficulty of scribing planning. Summary of the Invention

[0004] The present invention provides a laser scribing planning method, device, equipment and storage medium. The present invention realizes the global optimum of the dual-beam laser scribing path, balances scribing quality and efficiency, and improves scribing positioning accuracy.

[0005] In a first aspect, the present invention provides a laser scribing planning method, and the laser scribing planning method includes: Conduct kinematic and laser-material interaction dynamics modeling on the dual-beam laser system, and preprocess historical scribing data to obtain a preprocessed time series data set; Input the preprocessed time series data set into a time feature extractor to perform time dynamic feature extraction of historical scribing trajectories, and obtain a time feature vector; Construct a graph attention network based on the time feature vector and the surface grid division of the solar cell, and calculate the spatial interaction features between the two beams to obtain a spatial feature vector; Input the time feature vector and the spatial feature vector into a mixture density network to establish a material response probability model; Based on the material response probability model, perform hierarchical optimization of the scribing path to obtain the optimal control parameter sequence of the dual-light-path laser head.

[0006] In a second aspect, the present invention provides a laser scribing planning device, which includes: A preprocessing module, configured to perform kinematic and laser-material interaction dynamics modeling on the dual-light-path laser system, and preprocess historical scribing data to obtain a preprocessed time series data set; A feature extraction module, configured to input the preprocessed time series data set into a time feature extractor to perform time dynamic feature extraction of the historical scribing trajectory, and obtain a time feature vector; A calculation module, configured to construct a graph attention network according to the time feature vector and the surface grid division of the solar cell, and calculate the spatial interaction feature between the dual light paths to obtain a spatial feature vector; An establishment module, configured to input the time feature vector and the spatial feature vector into a mixture density network to establish a material response probability model; A hierarchical optimization module, configured to perform hierarchical optimization of the scribing path based on the material response probability model to obtain the optimal control parameter sequence of the dual-light-path laser head.

[0007] In a third aspect of the present invention, a computer device is provided, including: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor invokes the instructions in the memory to enable the computer device to execute the above-mentioned laser scribing planning method.

[0008] In a fourth aspect of the present invention, a computer-readable storage medium is provided, in which instructions are stored, and when it runs on a computer, it enables the computer to execute the above-mentioned laser scribing planning method.

[0009] In the technical solution provided by the present invention, by combining the Transformer encoder layer and the gated recurrent unit time feature extractor, the long-term dependencies and short-term change features in the historical scribing trajectory can be captured simultaneously, significantly improving the ability to understand the dynamic characteristics of the material response time. Based on the spatial feature extraction method of the graph attention network, the complex interaction relationship between the microstructure on the surface of the solar cell and the double optical paths is effectively modeled, realizing the accurate identification of the key scribing area and improving the scribing positioning accuracy. The material response probability model constructed by the mixture density network breaks through the limitations of the traditional deterministic model and can accurately characterize the multi-modal distribution of the scribing quality. Based on the hierarchical optimization strategy of the material response probability model, the global path planning is carried out by improving the ant colony algorithm, and the local parameter optimization is combined with quadratic programming. Under the condition of meeting the time-varying constraints, the global optimum of the double-optical-path laser scribing path is realized, balancing the scribing quality and efficiency. The real-time monitoring and closed-loop adjustment system realizes the dynamic adjustment of the scribing process. When the material response deviation is detected to exceed the threshold, the control parameters are adjusted online through the model predictive control algorithm and Bayesian update, enabling the system to have the adaptive ability to changes in material properties and fluctuations in the production environment, ensuring the stability and consistency of the scribing quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0011] Figure 1 It is a schematic diagram of the steps of the laser scribing planning method in the embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of the laser scribing planning device in the embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of the computer device in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] An embodiment of the present invention provides a laser scribing planning method, apparatus, device, and storage medium. Terms such as "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention will be described below. Please refer to Figure 1 , an embodiment of the laser scribing planning method in the embodiment of the present invention includes: Step S1: Perform kinematic and laser-material interaction dynamics modeling on the dual-beam laser system, and preprocess the historical scribing data to obtain a preprocessed time series data set; It can be understood that the execution subject of the present invention can be a laser scribing planning device, or a terminal or a server. Specifically, no limitation is made here. The embodiment of the present invention takes the server as the execution subject as an example for illustration.

[0014] Specifically, a kinematic model of the dual-light-path laser system is established to accurately describe the motion characteristics of the laser head during the scribing process. The position, velocity, and acceleration of the laser head in three-dimensional space are defined. The position vector is used to represent the coordinates of the laser head at a certain moment, while the velocity vector and acceleration vector respectively describe the motion state of the laser head. To ensure the stability and accuracy of the laser system, constraints are imposed on the velocity and acceleration to limit the maximum velocity and maximum acceleration of the laser head, ensuring that its motion does not exceed the physical range allowed by the device. Based on these definitions, a complete kinematic model is established, which can describe the trajectory of the laser head during the scribing process. A laser-material interaction dynamics model is established to describe the propagation process of laser energy in solar cell materials. Since the temperature of the material changes with time when irradiated by the laser, the heat conduction theory is used to characterize this process. Different types of solar cell materials have different absorption, conduction, and heat dissipation characteristics of laser energy. Therefore, parameters such as the thermal conductivity, density, and specific heat capacity of the material are considered in the model. At the same time, the interaction between the laser and the material not only depends on the heat conduction effect but also is affected by the material removal efficiency. To accurately describe this process, a material response function is introduced, which is used to characterize the material removal characteristics under different temperature, laser power, and scanning speed conditions. By combining the heat conduction theory and the material response function, a laser-material interaction dynamics model is established. The operating data of the dual-light-path laser system are collected to construct a historical scribing data set. The data collection relies on a high-precision sensor array, including a laser power sensor, an optical path position sensor, a temperature sensor, and a material surface displacement sensor, etc. The change of laser power is obtained through the laser power sensor, and the data is recorded with a high time resolution to ensure that the dynamic change of laser energy input can be accurately reflected. The optical path position sensor is used to record the spatial trajectory of the dual-light-path laser head, including the three-dimensional coordinate information of the two optical paths, ensuring the accuracy of the scribing trajectory. The material surface temperature distribution matrix is collected by a high-precision temperature sensor to reflect the temperature change of the material during the scribing process, thereby optimizing the laser power control strategy. At the same time, to evaluate the scribing quality, the key indicators of the scribing result are collected, including the line width, depth, and edge roughness of the scribing. These data are synchronized through time stamps to form a complete time series data set. The historical scribing data is preprocessed to improve the data quality. Filtering techniques are used to clean the data to remove measurement noise and smooth the data to make it more in line with the actual situation. Multiple filtering methods, such as Kalman filtering and wavelet transform, are used to reduce random interference and improve the stability of the data. The data from different sources are standardized, and the numerical values of each parameter are normalized to eliminate the scale difference between different physical quantities and improve the comparability of the data. Considering the randomness and uncertainty of the material's response during the scribing process, the cleaned data is classified to facilitate subsequent pattern recognition and optimization modeling.For the material response events in the underlined data, clustering analysis or supervised learning methods are used to divide them into different categories, such as normal underlining, partial removal, and excessive removal, etc., to analyze the laser response characteristics of different types of materials. After the above data cleaning and classification processing, a preprocessed time series data set is finally obtained.

[0015] Step S2: Input the preprocessed time series data set into a time feature extractor to perform time dynamic feature extraction of the historical underlining trajectory, and obtain a time feature vector; Specifically, perform a linear transformation on the preprocessed time series dataset to map it to different feature spaces. By constructing a mapping matrix, convert the input data into query, key, and value matrices to adapt to the computational requirements of the self-attention mechanism. The data at each time step is projected into different representation spaces for query, key, and value to ensure that the correlation between time steps can be accurately measured when calculating attention weights. Calculate the attention score matrix based on the query matrix and the key matrix to measure the correlation between different time steps and ensure that the model focuses on key historical information. The calculation of attention scores needs to consider the global information of the input sequence, so first calculate the dot product of the query matrix and the key matrix to obtain the correlation representation between time steps. Since the result of the dot product calculation will become too large as the input scale increases, normalize it with a scaling factor to maintain numerical stability and avoid problems such as gradient explosion or gradient disappearance. The normalized attention score matrix is processed by the softmax function to convert it into a probability distribution, forming the attention weight matrix. Weightedly sum the attention weight matrix and the value matrix and input it into the Transformer encoder layer in the time feature extractor to extract the long-range dependence features of the time series. The Transformer encoder layer contains a multi-head self-attention mechanism and also contains a feed-forward neural network structure to ensure that while focusing on different time steps, it fully learns the complex relationships in the time series data. The introduction of the multi-head self-attention mechanism enables the model to understand the data from multiple perspectives, thereby enhancing the expressive ability of time features. The role of the feed-forward neural network is to process the output of the attention mechanism to form a more abstract representation of time features and improve the expressive ability of the model. After being processed by the Transformer encoder layer, the original time series data is converted into a high-dimensional long-range dependence feature representation, effectively capturing the long-term trends and global information in the scribed trajectory. Input the long-range dependence feature representation into the gated recurrent unit in the time feature extractor to perform local temporal modeling using the characteristics of the recurrent neural network. In the gated recurrent unit, calculate the update gate and the reset gate to control the way information flows. The update gate determines the degree to which the current state retains past information, while the reset gate determines the impact of past information on the current state. By calculating these two gating variables, effectively adjust the weights of historical information, thus avoiding the problem of long-term dependence to a certain extent. Based on the calculation results of the update gate and the reset gate, the model calculates the candidate hidden state to model the information at the current time step. The update gate is used to control the selective memory and forgetting of information, enabling the model to dynamically adjust the information to be retained according to the temporal characteristics of the scribed trajectory, thereby obtaining a more stable local temporal feature representation. Input the local temporal features into the multi-scale temporal convolution module with different convolutional kernel sizes in the time feature extractor to ensure that the model can capture short-term, medium-term, and long-term time patterns simultaneously.The multi-scale temporal convolution module consists of multiple parallel convolutional layers, each with a different convolutional kernel size to provide different receptive fields. Smaller convolutional kernels are used to capture detailed features within a short time range, while larger convolutional kernels extract pattern information over a longer time range. Through this design, the model simultaneously focuses on changes at different time scales, thereby improving the ability to understand the dynamic characteristics of the scribed trajectory. During the actual calculation process, the output of each convolutional layer is processed by a non-linear activation function to enhance the model's expressive power, and weighted fusion is performed through an attention mechanism to ensure that information at different time scales can be reasonably integrated. The data processed by the multi-scale temporal convolution module forms a temporal feature vector.

[0016] Step S3: Construct a graph attention network based on the temporal feature vector and the surface grid division of the solar cell, and calculate the spatial interaction features between the two optical paths to obtain a spatial feature vector; Specifically, the surface of the solar cell is meshed to construct a graph structure and extract the spatial interaction features between the two optical paths. The surface of the solar cell is divided into N×M grid cells, each grid cell serving as a node, and corresponding physical properties and temporal features are assigned to each node, including its position coordinates in two-dimensional or three-dimensional space, material thickness, reflectivity, thermal conductivity, and the temporal feature vector extracted in the previous step, forming a node feature matrix. This matrix not only contains the static physical properties of the material but also integrates the dynamic temporal information during the laser scribing process, enabling subsequent spatial modeling to consider both the inherent properties of the material and the dynamic response during the laser action process. An undirected graph is constructed based on the node feature matrix to describe the spatial relationships between different grid nodes. The Euclidean distance or weighted distance based on material properties between them is calculated based on the spatial distribution of the nodes, and a distance attenuation function is used to construct the adjacency matrix, thereby defining the connectivity between the nodes. The elements of the adjacency matrix decay exponentially according to the physical distance between the nodes to ensure a strong influence between adjacent nodes and a weak connection between distant nodes, thus improving the local feature extraction ability of the model. In this way, the grid nodes on the surface of the solar cell are transformed into a graph structure, enabling subsequent graph neural networks to perform calculations on this topological structure and capture the mutual influence between each grid during the laser scribing process. To more accurately model the interaction relationship between grid nodes, a multi-head attention mechanism is applied to calculate the importance of the nodes and perform weighted learning on the graph structure. During this process, the self-attention score is calculated for each node to measure the correlation strength between the node and its neighbor nodes. The node feature matrix is independently transformed using different attention heads, and the attention coefficients from each node to its neighbor nodes are calculated. To ensure the stability of the calculation, the attention scores are normalized by softmax, making the sum of the attention weights of all neighbor nodes equal to one. Each attention head weights and sums the features of the neighbor nodes according to the calculated attention weights to generate the output of this attention head. Multiple independent attention weights are calculated in parallel through the multi-head attention mechanism, and the outputs of all attention heads are concatenated together at the final stage to obtain a more rich node-level feature representation. Through this step, the features of each node not only contain its own physical properties and temporal information but also integrate the global information from neighbor nodes, thereby improving the model's ability to understand spatial relationships. A dual-optical-path interaction module is designed based on the node-level feature representation, and the material responses at the current positions of the two optical paths are modeled as two subgraphs. Each subgraph respectively contains the local grid nodes around the optical path, and an inter-graph attention mechanism is used to measure the superposition effect of the thermal influence of the two optical paths on the material.During the calculation process, the influence degree of the nodes in optical path 1 on the nodes in optical path 2 is calculated. The specific method is to calculate the inter-graph attention coefficient based on the node features and the physical distance between the two optical paths. This coefficient maps different feature spaces through a learnable multi-layer perceptron, and softmax normalization is used to ensure the stability of the attention distribution. In this way, the influence degree of optical path 1 on optical path 2 is quantified, and a dual-optical-path interaction feature matrix is calculated to describe the synergistic effect between the two optical paths. The dual-optical-path interaction feature matrix not only reflects the spatial position relationship between the two optical paths but also captures the thermal response changes of the material when subjected to the dual-optical-path laser action. The differential pooling algorithm is applied to the dual-optical-path interaction feature matrix to extract the most representative spatial feature vector. During the differential pooling process, a clustering assignment matrix is dynamically generated according to the similarity of node features, and the pooled feature representation is calculated, so that similar nodes are merged, thereby reducing the computational cost and retaining the most important feature information. The differential pooling algorithm dynamically adjusts the importance of different nodes according to the task requirements, enabling the model to adapt to different material properties and optical path layouts, thus improving the generalization ability.

[0017] Step S4: Input the time feature vector and the spatial feature vector into the mixture density network to establish a material response probability model; Specifically, the time feature vector and the spatial feature vector are fused to obtain a spatio-temporal joint feature vector. The time feature vector contains the dynamic evolution pattern during the scribing process, which can reflect key factors such as laser power, motion trajectory, and material temperature change over time. The spatial feature vector captures the interaction relationship between different grid nodes on the surface of the solar cell and quantifies the spatial characteristics under the action of the dual optical path laser. The spatio-temporal joint feature vector is input into a fully connected neural network for hidden feature extraction to construct the intermediate layer representation of the mixture density network. The fully connected neural network consists of multiple hidden layers, and each layer contains a non-linear activation function to enhance the feature expression ability. In this process, the input vector undergoes multiple linear transformations and is mapped to a high-dimensional feature space through the activation function, enabling the model to learn the complex patterns in the spatio-temporal joint features. Through the extraction of hidden features, the network can capture the implicit relationship between the scribing quality and the input features, compress and remove redundant information, and improve the accuracy of subsequent probability distribution modeling. After completing the hidden feature extraction, a Gaussian mixture model parameter set is generated based on the intermediate layer representation for probability distribution modeling of material response. The parameters of the Gaussian mixture model include mixture weights, means, and variances, which are used to define the contributions of each Gaussian component in the probability distribution and the dispersion of data under different modes, respectively. These parameters are generated by the output layer of the neural network, where the mixture weights are normalized through the softmax function to ensure that the sum of all weights is one, and the variance parameters are transformed exponentially to ensure that their values are positive, thereby ensuring the rationality of the probability density function. A conditional probability distribution is constructed based on the Gaussian mixture model parameter set to describe the material response probability under different scribing parameter conditions. The Gaussian mixture model allows the model to represent the scribing quality in a multi-modal manner, enabling it to adapt to different material types and scribing processes. Under this probability representation, each possible value of the scribing quality is modeled by the weighted sum of multiple Gaussian components, thereby enhancing the model's ability to express uncertainty and providing more comprehensive quality prediction information. To ensure that the parameters of the mixture density network can correctly learn the probability distribution characteristics of the data, a negative log-likelihood loss function is defined to measure the gap between the predicted distribution of the model and the real data, and this is used as the optimization objective for network training. The construction method of the negative log-likelihood loss function ensures that while maximizing the likelihood of the observed data, it improves the model's ability to fit multi-modal distributions. To prevent the Gaussian components from degenerating or collapsing during the training process, a component regularization term is introduced to encourage different Gaussian components to learn different patterns in the feature space, thereby ensuring that the model can effectively distinguish the quality change trends under different scribing conditions. The network parameters are optimized through backpropagation, gradually adjusting the model weights to make it show better generalization ability on the training data, and obtaining a preliminarily trained mixture density network. Based on the preliminarily trained mixture density network, a conditional variational autoencoder is constructed to enhance the model's probability modeling ability.The conditional variational autoencoder consists of an encoder and a decoder. The encoder is used to map the input features and the target scribing quality to the latent space and learn the probability distribution of its latent variables. The decoder samples from the latent variables and reconstructs the scribing quality, enabling the model to better adapt to the uncertainty of the data. By imposing regularization constraints on the latent space, the conditional variational autoencoder ensures that the model can generate reasonable scribing quality predictions while avoiding overfitting problems. To improve the reliability of the model, the calibration error metric is calculated to evaluate the deviation between the confidence of the model prediction and the true probability, and it is optimized during the training process to ensure that the probability distribution output by the model can accurately reflect the uncertainty of the scribing quality. Through these optimization steps, the material response probability model is finally obtained.

[0018] Step S5: Based on the material response probability model, perform hierarchical optimization of the scribing path to obtain the optimal control parameter sequence of the dual-light-path laser head.

[0019] Specifically, define the scribing task for the surface of the solar cell and construct a complete set of scribing tasks. Each scribed line segment is represented as a triple containing a starting point, an ending point, and target scribing quality parameters, thereby clarifying the spatial position of each scribed line and the specific requirements for scribing quality. The set of scribing tasks consists of multiple scribed line segments and can comprehensively describe the scribing requirements of the entire solar cell. On this basis, construct a weighted directed graph model, where each node represents the starting point or ending point of a scribed line segment, and the edges in the graph represent possible path connections. Moreover, the weight of each edge comprehensively considers factors such as scribing time, energy consumption, and scribing quality loss to ensure that the scribing scheme achieves a balance between quality and efficiency. Apply the ant colony optimization algorithm to the weighted directed graph model to solve the globally optimal scribing path and ensure the optimal path of the dual-beam laser system in the entire scribing task. The ant colony optimization algorithm performs path optimization by simulating the process of ants searching for food. Each ant determines the next moving direction based on the pheromone concentration and heuristic information of the current path. The state transition probability of each ant is determined by the weighted combination of pheromone and heuristic information, enabling the exploration process to consider the current pheromone distribution and avoid falling into local optimal solutions. During the path search process, the pheromone update rule plays a crucial role. Whenever an ant finds a scribing path, update the pheromone on the path according to the scribing quality evaluation index, so that the paths with better quality receive more pheromone reinforcement, while the pheromone of the paths with poorer quality gradually decays, thereby guiding subsequent ants to preferentially select high-quality paths. Through multiple iterations, the ant colony optimization algorithm can converge to the globally optimal scribing path, ensuring that the scribing task can be planned in the best way. After obtaining the globally optimal scribing path, perform local optimization on each scribed line segment on the path to determine the specific execution parameters. Define a decision variable vector to represent the moving speed of the two-beam laser heads, laser power, and time offset. These variables jointly determine the execution mode of scribing and directly affect scribing quality. To optimize these decision variables, construct an objective function, which is calculated through a material response probability model, ensuring that the optimization process can fully consider the thermal response characteristics of the material and the effect of the laser on the material. To ensure that the optimized solution is physically feasible, introduce equipment constraint conditions, such as the laser power cannot exceed the maximum power allowed by the equipment, the moving speed of the laser head needs to be within the mechanical limits of the equipment, and the time synchronization between the two-beam laser heads also needs to be guaranteed. Under the constraints of these conditions, the optimization problem is transformed into a quadratic programming problem to ensure that the solved parameter combination can achieve the optimal scribing quality while satisfying physical constraints. Since quadratic programming problems are usually difficult to solve directly, the augmented Lagrangian multiplier method is used for solution. The augmented Lagrangian multiplier method solves the optimization problem through iteration by introducing Lagrangian multipliers and adding additional constraint penalty terms to the objective function.In each iteration, the objective function value is calculated based on the current optimization variables, and the scribing quality is evaluated according to the material response probability model. Subsequently, the Lagrange multipliers and constraint penalty terms are adjusted to continuously approach the optimal solution. As the number of iterations increases, the optimization process gradually converges, and the optimal parameter configuration for each scribed segment is obtained, including the moving speed of the dual-beam laser head, the laser power, and the time offset, so as to ensure that the scribing task can be completed with the highest efficiency while ensuring quality. Based on the optimal parameter configuration, samples are taken from the material response probability model to generate multiple possible material response scenarios. Each sampled scenario corresponds to a possible material response situation, and there may be significant differences in material properties between different scenarios, resulting in changes in scribing quality. Based on these scenarios, the performance of different parameter combinations in each scenario is evaluated, and the parameter combination that performs optimally in all scenarios is selected to ensure strong robustness of the scribing process. When finally selecting parameters, multiple factors are comprehensively considered, including the expected value, variance of the scribing quality, and the scribing effect in the worst case, to ensure the stability of the scribing quality even when there are large fluctuations in material response. After this optimization process, the optimal control parameter sequence of the dual-beam laser head is finally obtained, which contains the best execution parameters at each moment during the scribing process.

[0020] Deploy a high-precision real-time monitoring system to ensure that the entire scribing process can be precisely controlled in a dynamic environment. The monitoring system includes a laser power monitor, an optical path position sensor, a high-speed camera, a thermal imager, and a surface profile scanner to achieve a comprehensive perception of the laser scribing process. These sensors collect data in different dimensions. Among them, the laser power monitor is used to detect the real-time change of laser energy to ensure that the power remains within the expected range, while the optical path position sensor is used to track the movement trajectory of the laser head to provide accurate position information. The high-speed camera can capture key moments during the scribing process and record the details of the interaction between the laser and the material. The thermal imager is used to monitor the temperature distribution on the material surface to analyze the influence of heat conduction effects on the scribing quality. At the same time, the surface profile scanner can accurately measure the morphology after scribing to evaluate the depth, width, and edge quality of the scribing. The data of all these sensors are synchronously collected through a high-speed data acquisition card and aggregated into the original monitoring data stream. Process the original monitoring data stream to ensure the accuracy and stability of the data. The Kalman filtering method is used to denoise the data. The measurement noise is removed through a recursive estimation method, and a smooth state estimate value is provided. During the Kalman filtering process, state equations and measurement equations are established to describe the dynamic changes of the system state and the measurement relationship between the sensors and the state. Subsequently, through two steps of prediction update and measurement update, the state estimate value is continuously corrected to improve the accuracy and reliability of the data. After Kalman filtering, relatively stable monitoring data is obtained. Based on the optimal control parameter sequence, a three-layer closed-loop control architecture is constructed to ensure that the laser scribing process can be adjusted and optimized in real time in a complex environment. The closed-loop control architecture includes three levels: reference generation, state estimation, and control execution. Among them, the reference generation layer generates a series of time-discretized reference trajectories and laser power setting values based on the previously optimized optimal control parameter sequence as the target control signal for the entire system; the state estimation layer calculates the real-time state of the system through the fusion of sensor data and makes predictions in combination with the material response model to provide the current system state estimate value; the control execution layer dynamically adjusts the laser power and trajectory based on the state estimation result to ensure that the scribing quality reaches the best state. To improve the accuracy of state estimation, an unscented Kalman filter is used for multi-sensor information fusion in the state estimation layer to make full use of the data of different sensors and perform real-time estimation of the system state and the material response state. Compared with the traditional Kalman filtering method, the unscented Kalman filter has higher accuracy in non-linear system state estimation. By generating a set of sigma points to approximate the probability distribution of the state variables and retaining higher-order non-linear information during the state propagation process, the accuracy of state estimation is improved. During the filtering process, the system state is predicted through the state propagation equation, and the predicted value is corrected in combination with the measurement update equation to obtain a more accurate current system state estimate value.This estimated value is used to evaluate the thermal response of the material and provide real-time feedback information for dynamic adjustment during the control execution. After the state estimation is completed, model predictive control is performed on the current state estimated value of the system based on a three-layer closed-loop control architecture to solve the optimization problem within the rolling time domain, thereby generating an optimal control input sequence. Model predictive control is an optimization-based control method that uses the current state estimated value in each control cycle to predict the future evolution trend of the system and determines the next control input through optimization. When solving the optimization problem, various factors are considered, including the adjustment of laser power, the optimization of the optical path movement trajectory, and the uncertainty of material response, etc. And the optimization objective is to improve the scribing quality while minimizing energy consumption and scribing time as much as possible. Through the rolling optimization method, model predictive control can dynamically adjust the control input at each time step to ensure that the system always maintains the optimal state during actual operation. Calculate the material response deviation index according to the optimal control input sequence to measure the gap between the current system state and the desired scribing quality. When the material response deviation index exceeds the preset threshold, an adaptive adjustment strategy is adopted to ensure that the scribing process can adapt to real-time changes. The recursive least squares method is used to adjust the model parameters in real time to ensure that the model can adapt to changes in material properties at any time. To improve the adaptive ability of the system, the Bayesian update method is used to correct the material response probability model to continuously optimize the model parameters by combining new observation data, making the material response prediction more accurate. At the same time, to improve the safety of the system, multiple safety thresholds are set and the control parameter adjustment is automatically triggered when necessary. Multiple safety levels are set according to indicators such as the real-time temperature of the material, the optical path error, and the scribing quality. When an indicator exceeds the warning threshold, the risk is reduced by adjusting the laser power, reducing the optical path speed, etc., and when an indicator exceeds the danger threshold, an emergency stop mechanism is triggered to prevent the scribing quality from being seriously affected. Through this closed-loop control strategy, it is ensured that the laser scribing system has good adaptability and stability in a complex industrial environment, thus achieving high-precision and high-reliability scribing effects. Through the above optimizations and control adjustments, an adaptive closed-loop control strategy is obtained.

[0021] In this embodiment, by combining a Transformer encoder layer and a gated recurrent unit-based temporal feature extractor, the long-term dependencies and short-term change features in the historical scribing trajectory can be captured simultaneously, significantly improving the understanding ability of the dynamic characteristics of the material response time. Based on the graph attention network-based spatial feature extraction method, the complex interaction relationship between the microstructure on the surface of the solar cell and the dual optical paths is effectively modeled, realizing the accurate identification of the key scribing area and improving the scribing positioning accuracy. The material response probability model constructed by the mixture density network breaks through the limitations of traditional deterministic models and can accurately represent the multimodal distribution of the scribing quality. Based on the hierarchical optimization strategy of the material response probability model, the global path planning is carried out by improving the ant colony algorithm, and the local parameter optimization is combined with quadratic programming. Under the condition of meeting the time-varying constraints, the global optimum of the dual-optical-path laser scribing path is achieved, balancing the scribing quality and efficiency. The real-time monitoring and closed-loop adjustment system realizes the dynamic adjustment of the scribing process. When it is detected that the material response deviation exceeds the threshold, the control parameters are adjusted online through the model predictive control algorithm and Bayesian update, enabling the system to have the adaptive ability to changes in material properties and production environment fluctuations, ensuring the stability and consistency of the scribing quality.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Based on the dual-optical-path laser system, define the three-dimensional position vector, velocity vector, and acceleration vector, and introduce velocity constraints and acceleration constraints to obtain the kinematic model of the dual-optical-path laser system; Based on the heat conduction equation, establish a model for the propagation process of laser energy in the solar cell material, and introduce a material response function to describe the material removal efficiency to obtain the laser-material interaction dynamics model; Based on the kinematic model and the laser-material interaction dynamics model, collect the laser power sequence, optical path position sequence, material surface temperature distribution matrix, and scribing result evaluation index of the dual-optical-path laser system to obtain historical scribing data; Filter the historical scribing data to obtain the cleaned scribing data, and classify the material response events in the cleaned scribing data to obtain the preprocessed time series data set.

[0023] Specifically, establish the kinematic model of the dual-optical-path laser system to describe the motion characteristics of the laser head in three-dimensional space. Define the three-dimensional position vector, velocity vector, and acceleration vector of the laser head to completely characterize its motion state. Let the position vector of the laser head be , where , and respectively represent the positions of the laser head on the three coordinate axes, and the velocity vector represents the instantaneous movement speed of the laser head, while the acceleration vector is used to describe the acceleration state of the laser head in different directions. To ensure that the movement of the laser head conforms to the physical constraints of the device, speed and acceleration constraints are imposed, such that the speed of the laser head does not exceed the maximum allowable speed, and the acceleration does not exceed the maximum allowable acceleration, i.e.: ; ; wherein, represents the maximum operating speed allowed by the laser device, while represents the maximum acceleration limit. These constraints ensure that the laser system does not exceed the physical limits during operation, while reducing mechanical errors caused by high-speed movement and improving the accuracy of scribing. A laser-material interaction dynamics model is constructed to describe the propagation process of laser energy in the solar cell material and analyze the material removal efficiency. Since the laser deposits energy on the material surface and causes local heating, the thermal conduction characteristics of the material play an important role in the scribing process. The heat conduction equation is used to describe the change of the temperature field, i.e.: ; wherein, represents the temperature distribution of the material at position and time , is the thermal conductivity of the material, is the material density, is the specific heat capacity, while represents the thermal energy deposited by the laser on the material surface. Since the solar cell material is usually composed of multiple thin films and the thermal conduction characteristics of each layer are different, the interlayer thermal resistance coefficient is introduced to consider the influence of the interlayer thermal resistance on the temperature distribution. For multi-layer materials, the temperature change needs to satisfy the following boundary conditions: ; ; wherein, and are the thermal conductivities of two layers of materials respectively, while and are the temperatures of adjacent material layers. In this case, the interlayer thermal resistance will affect the heat transfer efficiency and thus the material removal effect. To quantify the material removal efficiency, a material response function is introduced, where represents the local temperature, represents the laser power, represents the scanning speed. The calculation of the material removal efficiency adopts the laser ablation model, and its basic form is: ; wherein, is the material absorption coefficient, is the activation energy for laser-induced material removal, is the gas constant, is the temperature, and is the laser scanning speed. This formula indicates that the material removal rate increases with the increase of laser power, but is simultaneously affected by temperature and scanning speed. After completing the kinematic and dynamic modeling, data acquisition is performed on the dual-beam laser system to record the key parameters during the actual scribing process. To obtain complete historical scribing data, high-precision sensors are used, including a laser power sensor, an optical path position sensor, a temperature sensor, and a surface profile scanner, to collect the key information during the laser scribing process. The laser power sensor is used to record the laser power at each time point, forming a laser power sequence ; secondly, the optical path position sensor records the movement trajectory of the laser head, generating an optical path position sequence ; meanwhile, the temperature sensor is used to measure the temperature distribution on the material surface and form a temperature distribution matrix ; the surface profile scanner is used to measure the geometric shape of the scribing and output scribing quality evaluation indicators such as scribing line width, depth, and edge roughness. These data constitute a complete time series data set. The historical scribing data is filtered to remove noise and improve data quality. On the filtered data, the material response events are classified to better identify the behavior patterns of different materials under the action of the laser, obtaining a preprocessed time series data set.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Perform a linear transformation on the preprocessed time series data set to obtain a query matrix, a key matrix, and a value matrix; Calculate an attention score matrix based on the query matrix and the key matrix, normalize it through a scaling factor, and apply the softmax function to obtain an attention weight matrix; Input the attention weight matrix and the value matrix into the Transformer encoder layer containing a feed-forward neural network in the time feature extractor to obtain a long-range dependence feature representation; Input the long-range dependence feature representation into the gated recurrent unit in the time feature extractor, and control the information flow by calculating the update gate and the reset gate to obtain a gated state; Calculate a candidate hidden state based on the gated state, and control the selective memory and forgetting of information through the update gate to obtain local temporal features; The local temporal features are input into the multi-scale temporal convolution module with different convolutional kernel sizes in the temporal feature extractor, and short-term, medium-term, and long-term temporal patterns are captured simultaneously through parallel 1D convolutional layers to obtain the temporal feature vector.

[0025] Specifically, a linear transformation is performed on the preprocessed time series dataset to extract key features and adapt to subsequent self-attention calculations. Let the input time series dataset be , where represents the system state at the th moment, including key variables such as laser power, optical path position, and material surface temperature. To calculate the attention scores, a linear transformation is performed on the input data to map it to the query matrix, key matrix, and value matrix. Through the weight matrices , , and for projection, that is: ; ; ; where is the query matrix, is the key matrix, is the value matrix, and , , are trainable parameter matrices for feature transformation, respectively. The query matrix is used to measure the focus at the current moment, the key matrix is used to store the content of all time steps, and the value matrix contains the actual temporal features. Calculate the similarity score between the query matrix and the key matrix to determine the information correlation between different time steps. The self-attention mechanism calculates the attention score matrix using the dot product method, that is: ; Since the numerical range of the dot product calculation is large, it is normalized by a scaling factor to maintain numerical stability and avoid the gradient explosion problem. The scaling factor is used for normalization to obtain the normalized attention score matrix: ; where is the dimension of the key matrix. The softmax function is used to convert the normalized attention scores into the attention weight matrix , thus ensuring that the attention distribution of all time steps is normalized to a probability distribution: ; At this time, the attention weight matrix It reflects the correlation between time steps and can dynamically adjust the model's attention to historical information. The attention weight matrix and the value matrix are weighted and summed to obtain the final self-attention output: where, represents the time feature after being processed by the self-attention mechanism. This feature can capture long-range dependencies and provide richer temporal information for subsequent feature extraction. To extract the high-level features of the time series, the self-attention output is fed into the Transformer encoder layer containing a feed-forward neural network. This encoder layer includes residual connections and layer normalization mechanisms to ensure the stability of feature extraction. Specifically, a non-linear mapping is performed through the feed-forward neural network: ; where, and are the trainable weight matrices of the feed-forward neural network, while and are the bias terms. The ReLU activation function is used to introduce non-linearity to enhance the feature representation ability. Through residual connections and layer normalization, the network can be trained more effectively and avoid the problem of gradient vanishing: ; The obtained long-range dependence feature contains both the time features calculated by self-attention and has been further processed by the feed-forward network to ensure the stability of feature representation. The long-range dependence feature representation is input into the gated recurrent unit in the time feature extractor to capture local patterns on short time scales. Taking as the input, the gated recurrent unit calculates the update gate and the reset gate to control the information flow: ; ; where, represents the update gate, which is used to determine how much historical information should be retained at the current moment, while represents the reset gate, which determines the influence degree of past information on the current state. Based on these gated variables, the GRU calculates the candidate hidden state: ; Through the update gate controls the selective memory and forgetting of information to obtain the final local temporal feature: ; At this time, Represents the time features after GRU processing. These features can effectively model short-term dependencies and regulate the information flow through the gating mechanism, thus avoiding the problem of gradient vanishing. To improve the model's ability to express features at different time scales, the local temporal features output by GRU are input into a multi-scale temporal convolution module. This module consists of multiple parallel 1D convolutional layers, and each convolutional layer uses a different kernel size to capture patterns with different time spans. Let the kernel sizes be corresponding to short-term, medium-term, and long-term time features, then the convolution operation is expressed as: ; where is the weight of the th convolutional kernel, and represents the feature maps at different time scales. Through multiple parallel convolutional layers, features within different time ranges are obtained, and finally these features are merged through a concatenation operation to form the final time feature vector: ; The time feature vector contains the long-range dependency features processed by Transformer and the local temporal features calculated by GRU, and at the same time captures the pattern changes at different time scales through the multi-scale temporal convolution module.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Divide the surface of the solar cell into N×M grid cells as the node set, and assign position coordinates, material thickness, reflectivity, thermal conductivity, and the time feature vector to each node to obtain the node feature matrix; Construct an undirected graph based on the node feature matrix, generate the adjacency matrix by calculating the spatial distance between nodes, and obtain the graph structure representation of the grid nodes; Apply the multi-head attention mechanism to the graph structure representation of the grid nodes to obtain the attention weights; Calculate the output of each attention head according to the attention weights, and obtain the node-level feature representation by concatenating the outputs of all attention heads; Design a dual-light path interaction module based on the node-level feature representation, model the material responses at the current positions of the two light paths as two subgraphs, and quantify the superposition effect of the dual-light path thermal influence by calculating the inter-graph attention coefficient to obtain the dual-light path interaction feature matrix; Apply the difference pooling algorithm to the dual-light path interaction feature matrix to obtain the spatial feature vector.

[0027] Specifically, perform grid division on the surface of the solar cell to establish the graph structure representation of the spatial features. Assume the size of the solar cell surface is , divide it into A grid cell, each grid cell represents a node, and the grid set constitutes a node set . To completely describe the characteristics of each grid node, multiple physical attributes are assigned to each node, including position coordinates , material thickness , reflectivity , thermal conductivity and the time feature vector obtained from the time feature extractor , thus forming a node feature matrix : ; Among them, contains time series feature information, such as historical dynamic features of laser power, temperature change, and material response. Through node feature definition, each grid cell not only contains its physical properties but also encodes the time features in the dynamic scribing process. After constructing the node feature matrix, an undirected graph is constructed based on this matrix , where is the node set, is the edge set, describing the connection relationship between adjacent grid cells. To quantify the correlation between nodes, an adjacency matrix is generated by calculating the Euclidean distance between nodes or the weighted distance based on material parameters : ; Among them, represents the connection weight between nodes and , are the spatial positions of the nodes respectively, is a parameter controlling the decay rate of the adjacency weight. Closer nodes have larger connection weights, while farther nodes have weights tending to zero, thus establishing a graph structure representation based on local spatial relationships. The multi-head attention mechanism is applied to the graph structure to extract key spatial features. For each node , the multi-head attention mechanism performs a linear transformation on its features and projects them onto the query matrix , key matrix and value matrix : Among them, is a trainable weight matrix. Calculate the dot product of the query matrix and the key matrix to obtain the attention score: ; Among them, is the dimension of the key matrix. Perform softmax normalization on the attention score to obtain the attention weight: ; Among them, is the neighbor set of node . Combine the attention weights with the value matrix to obtain the output of each attention head: ; Since a single attention head is not sufficient to capture spatial relationships at different scales, independent attention heads are adopted, and the outputs of all attention heads are concatenated to form the final node-level feature representation: ; The features of each node not only contain its own physical information and temporal features, but also incorporate information from its neighboring nodes, thus forming a complete spatial representation. To further model the interaction effect between the two optical paths, a two-optical-path interaction module is designed to quantify the influence of the material response at the current positions of the two optical paths. Extract the nodes where the two optical paths are located and construct two subgraphs and , corresponding to the local regions of the two optical paths respectively. Use the inter-graph attention mechanism to calculate the influence degree of the nodes in optical path 1 on the nodes in optical path 2. Define the inter-graph attention coefficient: ; Among them, and are the node features of optical path 1 and optical path 2 respectively, is the physical distance between the two optical path nodes, represents a multi-layer perceptron for calculating the mutual influence strength between nodes. By summing the attention weights of all nodes, the two-optical-path interaction feature matrix is obtained: ; This matrix describes the thermal influence region formed by the two optical paths on the material surface and the interaction relationship between the two optical paths. To extract key spatial information, apply the differential pooling algorithm to the two-optical-path interaction feature matrix to reduce redundant information and generate the final spatial feature vector. Define the clustering assignment matrix : ; Among them, Determine the clustering relationship by minimizing the similarity loss between nodes. Calculate the pooled feature representation: ; At the same time, update the pooled adjacency matrix: ; The spatial feature vector is composed of the pooled feature matrix It means that it contains the key spatial information after the two - path interaction.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Fuse the time - feature vector and the spatial - feature vector to obtain a spatio - temporal joint feature vector; Input the spatio - temporal joint feature vector into a fully - connected neural network for hidden - feature extraction, obtain the intermediate - layer representation of the mixture - density network, and generate a Gaussian mixture model parameter set based on the intermediate - layer representation; Construct a conditional probability distribution according to the Gaussian mixture model parameter set to obtain a multi - modal distribution representation of the scribing quality; Define a negative - log - likelihood loss function for the multi - modal distribution representation, and at the same time introduce a component regularization term, and optimize the network parameters through backpropagation to obtain a preliminarily trained mixture - density network; Based on the preliminarily trained mixture - density network, construct a conditional variational auto - encoder, enhance the uncertainty modeling ability through the encoder and the decoder, and calculate the calibration error index to obtain a material response probability model.

[0029] Specifically, fuse the time - feature vector and the spatial - feature vector to construct a spatio - temporal joint feature representation. The time - feature vector is obtained by a time - feature extractor and contains the dynamic evolution patterns during the scribing process, such as laser - power change, optical - path trajectory, material temperature and other information, while the spatial - feature vector is extracted by a graph attention network and describes the spatial interaction effect of the two - path laser system on the material surface. Adopt the feature - splicing method to connect the time - feature vector and the spatial - feature vector dimension - by - dimension to obtain a spatio - temporal joint feature vector: ; where represents the feature - splicing operation. Input the spatio - temporal joint feature vector into a fully - connected neural network to extract hidden features and obtain the intermediate - layer representation of the mixture - density network. This fully - connected neural network consists of multiple non - linear transformation units, where each layer contains a weight matrix and a bias term, and is mapped through a non - linear activation function, so that the input features can form a more compact representation in the high - dimensional space. Let the transformation function of the hidden layer be: ; where represents the intermediate - layer representation of the neural network, is the weight matrix, is the bias term, is a non - linear activation function, such as ReLU or ELU. This mapping process can effectively extract the complex relationships in the features, enabling the model to have a stronger expressive ability for material responses under different scribing conditions. Based on the intermediate layer representation, a parameter set of the Gaussian mixture model (GMM) is generated to establish the conditional probability distribution of the material response. The Gaussian mixture model assumes that the probability distribution of the material response is represented as a weighted combination of multiple Gaussian components. Therefore, the GMM is formulated as: ; where, is the number of mixture components, represents the mixing weight of the -th Gaussian component, satisfying and represent the mean and variance of the -th Gaussian component respectively. These parameters are all calculated by the output layer of the neural network. Among them, the mixing weight is normalized by softmax to ensure it is a valid probability distribution, while the variance parameter is transformed by exponentiation to ensure its value is always positive. A negative log - likelihood loss function is defined for the multimodal probability distribution to optimize the parameters of the mixture density network. Since the goal is to make the model maximize the probability of correct data, the loss function is defined as follows: ; where, is the number of training samples, is the true material response value of the -th sample. This loss function enables the model to more accurately characterize the uncertainty of the material response by maximizing the likelihood value of the observed data. To prevent some Gaussian components from collapsing or degenerating, a component regularization term is introduced to encourage sufficient separation between the means of different Gaussian components: ; where, is a hyperparameter used to control the minimum separation between components. This regularization term ensures that the patterns learned by different Gaussian components do not overlap excessively, thereby improving the multimodal expressive ability of the model. Combining the negative log - likelihood loss and the component regularization term, the parameters of the neural network are optimized through the backpropagation algorithm to obtain a preliminarily trained mixture density network. Based on the preliminarily trained mixture density network, a conditional variational auto - encoder is constructed to enhance the probability modeling ability of the model. The conditional variational auto - encoder consists of an encoder and a decoder. The encoder is used to map the input features and the target material response to the latent space and learn the probability distribution of its latent variables, while the decoder samples from the latent variables and reconstructs the target variables. Let the transformation function of the encoder be: ; wherein, represents a latent variable, and are the mean and variance learned by the encoder respectively, represents the parameters of the encoder. The reconstruction process of the decoder is defined as: ; wherein, represents the parameters of the decoder. To optimize the conditional variational autoencoder, maximize the variational lower bound: ; wherein, represents the Kullback-Leibler divergence between the encoder output distribution and the standard normal distribution. This term ensures that the distribution of the latent variable is close to the standard normal distribution, thereby enhancing the generalization ability of the model. By optimizing this objective function, the model's ability to model the uncertainty of material responses is enhanced. To evaluate the prediction confidence of the model, a calibration error metric is calculated to measure whether the output probability of the model can accurately reflect the actual situation. The Expected Calibration Error (ECE) metric is used: ; wherein, is the th confidence bucket, and are the actual accuracy rate and average confidence of this bucket respectively. By optimizing the calibration error, ensure that the prediction confidence of the model is more accurate, and finally obtain a stable material response probability model.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Define the scribing task on the surface of the solar cell, represent each scribing segment as a triple containing the starting point, ending point, and target scribing quality parameter, and construct a scribing task set; Construct a weighted directed graph model based on the scribing task set, apply the ant colony optimization algorithm to the weighted directed graph model, control the exploration of ants through the state transition probability, and adjust the pheromone according to the pheromone update rule to obtain the globally optimal scribing path; For each scribing segment on the globally optimal scribing path, define a decision variable vector to represent the moving speeds, laser powers, and time offsets of the two light paths, calculate the objective function through the material response probability model, and introduce equipment constraint conditions to obtain a quadratic programming problem; Solve the quadratic programming problem using the augmented Lagrangian multiplier method, and through iterative calculation, obtain the optimal parameter configuration for each scribing segment; Based on the optimal parameter configuration, samples are taken from the material response probability model to generate multiple material response scenarios, and the optimal parameter combinations for each material response scenario are selected to obtain the optimal control parameter sequence of the dual-path laser head.

[0031] Specifically, the scribing task on the surface of the solar cell is defined to establish an optimizable path model throughout the scribing process. To precisely describe each scribed line, it is represented as a triple containing the starting point, the ending point, and the target scribing quality parameters, that is, each scribed line segment is defined as: ; where and respectively represent the three-dimensional coordinates of the starting point and the ending point of the th scribed line, and represents the scribing quality parameters, including key indicators such as scribing depth, width, and edge smoothness. The set of all scribed line segments constitutes the scribing task set: ; where is the total number of scribing tasks. This task set defines the scribing layout of the entire solar cell surface. Based on the scribing task set, a weighted directed graph model is established to describe the feasible path connection relationship between different scribing tasks. Define the directed graph , where is the node set, each node corresponding to a scribed line segment, and the edge set represents the connection path between scribed line segments. To quantify the advantages and disadvantages between different scribing paths, weights are assigned to the edges of the graph, indicating the cost from scribed line segment to scribed line segment . This cost comprehensively considers the optical path movement distance, power switching loss, and time adjustment cost: ; where represents the physical distance between two scribed lines, quantifies the cost of the optical path moving from the ending point to the starting point, and represents the degree of change in scribing quality parameters. The weight coefficient controls the influence ratio of different factors. The ant colony optimization algorithm is used to find the global optimal scribing path on the weighted directed graph. Ant colony optimization simulates the exploration process of ants on the path and guides path optimization through the pheromone update mechanism. Define the state transition probability of ant from scribed line segment to scribed line segment as: ; Among them, is the pheromone concentration on the path , controls the importance of the pheromone, controls the influence of the path cost, while is the set of all feasible next segment lines. The ant performs a random walk according to the above probability and updates the pheromone on the path after traversing the entire segment line path: ; Among them, is the pheromone evaporation coefficient, while is the ant on the path the increment of pheromone left. Through multiple iterations, the ant colony gradually tends to find the optimal segment line path, thereby reducing the optical path movement time and power switching loss. After obtaining the globally optimal segment line path, local optimization is performed on each segment line on the path to determine the optimal control parameters of the laser system. Define the decision variable vector as: ; Among them, are the moving speeds of the two optical paths respectively, is the laser power of the two optical paths, is the time offset. To optimize these variables, an objective function is constructed, which is calculated based on the material response probability model and defined as: ; Among them, is the expected value of the scribing quality, while represents the variance of the quality, controls the degree of risk aversion. To ensure the feasibility of the optimization solution, device constraints are imposed, such as maximum speed, maximum power, and synchronization requirements: ; ; This optimization problem belongs to a quadratic programming problem and is solved by the augmented Lagrangian multiplier method. By introducing the Lagrangian multiplier and constructing the augmented Lagrangian function: ; Among them, is the auxiliary variable, is the Lagrangian multiplier, is the penalty coefficient. In the iteration process, and are optimized in turn: ; ; ; After multiple iterations, the method can converge to the optimal parameter configuration for each scribed line segment. Due to the uncertainty of material response, relying solely on a single optimal solution is not sufficient to ensure stable scribing quality. After obtaining the optimal parameter configuration, sampling is performed based on the material response probability model to generate multiple material response scenarios. For each scenario , calculate its optimal parameter combination , and then select the optimal set among all scenarios: ; The obtained optimal control parameter sequence of the dual optical path laser head can ensure that the scribing task still maintains high quality in a complex environment and has a certain degree of robustness to cope with the uncertainties brought by different material characteristics and environmental changes.

[0032] In a specific embodiment, the execution of the laser scribing planning method further includes the following steps: Deploy a real-time monitoring system including a laser power monitor, an optical path position sensor, a high-speed camera, a thermal imager, and a surface profile scanner, and synchronously collect data on the laser scribing process through a high-speed data acquisition card to obtain the original monitoring data stream; Perform Kalman filtering on the original monitoring data stream to obtain the processed monitoring data, and at the same time generate a three-layer closed-loop control architecture including reference generation, state estimation, and control execution based on the optimal control parameter sequence; Use an unscented Kalman filter to perform multi-sensor information fusion on the processed monitoring data, and estimate the system state and material response state in real time through the state propagation equation and the measurement update equation to obtain the current system state estimate value; Perform model predictive control on the current system state estimate value based on the three-layer closed-loop control architecture, solve the optimization problem within the rolling time domain, and obtain the optimal control input sequence; Calculate the material response deviation index according to the optimal control input sequence. When the material response deviation index exceeds the preset threshold, use the recursive least squares method to adjust the model parameters in real time, correct the material response probability model through Bayesian update, and set multiple safety thresholds to automatically trigger the control parameter adjustment to obtain an adaptive closed-loop control strategy.

[0033] Specifically, a high-precision real-time monitoring system is deployed to ensure that the entire scribing process can be precisely controlled in a complex industrial environment. The system consists of a laser power monitor, an optical path position sensor, a high-speed camera, a thermal imager, and a surface profile scanner, and synchronously collects data on the laser scribing process through a high-speed data acquisition card to obtain key process variables. Among them, the laser power monitor is used to monitor the fluctuations in the laser output power in real time to ensure that the power remains within the set range, while the optical path position sensor is used to track the movement trajectory of the laser head and provide accurate position information. The high-speed camera can capture the interaction details between the laser and the material with high time resolution, while the thermal imager is used to measure the temperature distribution on the material surface to evaluate the impact of heat conduction effects on the scribing quality. At the same time, the surface profile scanner is used to measure the morphological characteristics after scribing, such as scribing depth, width, and edge roughness. The data of these sensors are collected at a high frequency and synchronously transmitted to the data processing system to form the original monitoring data stream. The original monitoring data stream is processed by Kalman filtering to improve the stability and reliability of the data. During the Kalman filtering process, a state equation and a measurement equation are established, where the state vector represents the true state of the system at time , including variables such as laser power, optical path position, and material surface temperature, while the measurement vector represents the observed values measured by the sensors. The state equation and the measurement equation are expressed as: ; ; where, is the state transition matrix, which describes the dynamic evolution process of the system state, is the control matrix, which represents the impact of the control input on the system state, and They are process noise and measurement noise respectively, both following Gaussian distribution. Kalman filter recursively estimates the system state through two steps of prediction update and measurement update to maximize the credibility of data. After filtering, smoother and more reliable monitoring data are obtained. A three-layer closed-loop control architecture is constructed based on the optimal control parameter sequence to ensure that the laser scribing process can adapt to dynamic environmental changes and perform real-time optimization. This architecture includes a reference generation layer, a state estimation layer, and a control execution layer. The reference generation layer generates a time-discretized reference trajectory based on the optimal control parameter sequence obtained from previous optimization, such as laser power setting value, optical path movement speed, and scribing time offset, etc.; the state estimation layer calculates the real-time state of the system through sensor data fusion and predicts the future state in combination with the material response probability model to provide the current system state estimation value; the control execution layer adaptively adjusts the laser power, optical path trajectory, and scribing speed based on the state estimation result to ensure that the scribing quality reaches the optimal state. To improve the accuracy of state estimation, an unscented Kalman filter is used for multi-sensor information fusion of the processed monitoring data, and the system state and material response state are estimated in real time through the state propagation equation and the measurement update equation. Unscented Kalman filter can handle the state estimation problem of nonlinear systems more effectively compared with traditional Kalman filter. It uses a set of sigma points to approximate the probability distribution of state variables and propagates these sigma points through nonlinear transformation to calculate the new state estimation value. In the processing of unscented Kalman filter, sigma points are generated: ; where is the state estimation of the previous step, is the state covariance matrix, is the state dimension, is the hyperparameter. Substitute the sigma points into the state propagation equation for time update and correct them in combination with the measurement data to obtain the current state estimation value of the system. After obtaining the current state estimation value of the system, model predictive control is performed based on the three-layer closed-loop control architecture to solve the optimization problem within the rolling time domain and calculate the optimal control input sequence. Model predictive control uses the current state estimation value to predict the system evolution for multiple future time steps and solves the optimization problem at each time step to ensure that the control input can make the system state approach the optimal target. Let the objective function of the optimization problem be: ; where is the current state, is the reference state, is the control input, and are the weight matrices of the state error and the control input respectively. Model predictive control obtains the future The optimal control input sequence for each time step, and during the actual control process, after each time step is executed, the state estimate is updated and rolling optimization is performed to ensure that the control input is always optimal. Calculate the material response deviation index to measure the gap between the current system state and the expected scribing quality. When the material response deviation index exceeds the preset threshold, the recursive least squares method is used to adjust the model parameters in real time, and the material response probability model is corrected through the Bayesian update method. The update formula of the recursive least squares method is as follows: ; where, are the model parameters to be estimated, is the observed value, is the input feature vector, is the gain matrix. Through the update of the recursive least squares method, the model parameters can be quickly adjusted to better adapt to the current environmental changes. At the same time, to improve the safety of the system, multiple safety thresholds are set, and the control parameter adjustment is automatically triggered when necessary. Multiple safety levels are set according to indicators such as material temperature, laser power deviation, and scribing quality. When an indicator exceeds the warning threshold, the laser power is adjusted or the optical path speed is reduced to reduce risks, while when an indicator exceeds the danger threshold, an emergency stop mechanism is triggered to prevent the scribing quality from being seriously affected. Through this closed-loop control strategy, it is ensured that the laser scribing system has good adaptability and stability in a complex environment, thus achieving high-precision and high-reliability scribing effects. An adaptive closed-loop control strategy is obtained.

[0034] This embodiment also includes synergistically allocating and dynamically balancing the dual-path energy based on the optimal control parameters of the dual-path laser head, including: decoupling the space-time of the optimal control parameter sequence of the dual-path laser head, extracting the laser power parameters and position parameters respectively, constructing a four-dimensional dual-path energy-position tensor, and obtaining a dual-path energy-position mapping model; based on the dual-path energy-position mapping model, mathematically modeling the dual-path laser thermal influence region by Fourier series expansion method, calculating the heat flux density distribution at any point (x, y), and calculating the heat accumulation effect through convolution operation to obtain a dual-path thermal field superposition model; discretizing the dual-path thermal field superposition model into an N×N grid, constructing a dual-path thermal influence matrix, and calculating the sensitivity coefficient for each grid cell (i, j) to obtain a dual-path thermal influence sensitivity field; based on the dual-path thermal influence sensitivity field, calculating the power compensation amount to obtain a dual-path power dynamic balance strategy; using deep reinforcement learning to construct a dual-path collaborative control agent, the input state space S includes the current thermal field distribution T(x, y), the positions of the two paths (X1, Y1, X2, Y2) and the powers (P1, P2), the action space A includes the power adjustment amount and the speed adjustment amount, the reward function R is designed as a weighted combination of the scribing quality score and the energy consumption, and the optimal energy allocation strategy is obtained through training by the proximal policy optimization algorithm; based on the optimal energy allocation strategy, designing a dual-path energy balancer, adopting a feedforward-feedback hybrid control architecture, the feedforward control performs open-loop compensation based on the dual-path thermal field superposition model to obtain a dual-path energy real-time regulation mechanism; performing dynamic time-domain analysis on the dual-path energy real-time regulation mechanism, constructing a system transfer function through Laplace transform, calculating the system overshoot, adjustment time and steady-state error, and simultaneously performing Monte Carlo simulation to test the robustness of the system under different material parameter perturbations to obtain the stability evaluation result of the dual-path energy control; based on the stability evaluation result, applying an adaptive fuzzy logic controller to compensate the dual-path energy real-time regulation mechanism, dynamically adjusting the control parameters by establishing a fuzzy rule base "IF {temperature deviation} AND {temperature change rate} THEN {power adjustment amount}", and using Lyapunov stability analysis to ensure the convergence of the closed-loop system, and finally generating a scribing execution plan considering the dual-path thermal coupling effect.

[0035] The laser scribing planning method in the embodiment of the present invention has been described above. Next, the laser scribing planning device in the embodiment of the present invention will be described. Please refer to Figure 2 , an embodiment of the laser scribing planning device in the embodiment of the present invention includes: A preprocessing module, configured to perform kinematic and laser-material interaction dynamics modeling on the dual-path laser system, and preprocess the historical scribing data to obtain a preprocessed time series data set; A feature extraction module, configured to input the preprocessed time series data set into a time feature extractor to perform time dynamic feature extraction of the historical scribing trajectory, and obtain a time feature vector; A calculation module, configured to construct a graph attention network according to the time feature vector and the surface grid division of the solar cell, and calculate the spatial interaction features between the two optical paths to obtain a spatial feature vector; An establishment module, configured to input the time feature vector and the spatial feature vector into a mixture density network to establish a material response probability model; A hierarchical optimization module, configured to perform hierarchical optimization of the scribing path based on the material response probability model to obtain an optimal control parameter sequence for the two-optical-path laser head.

[0036] Through the collaborative cooperation of the above-mentioned various components, by combining a time feature extractor of a Transformer encoder layer and a gated recurrent unit, the long-term dependencies and short-term change features in the historical scribing trajectory can be captured simultaneously, significantly improving the understanding ability of the time dynamic characteristics of the material response. The spatial feature extraction method based on the graph attention network effectively models the microscopic structure on the surface of the solar cell and the complex interaction relationship between the two optical paths, realizes the accurate identification of the key scribing area, and improves the scribing positioning accuracy. The material response probability model constructed by the mixture density network breaks through the limitations of traditional deterministic models and can accurately characterize the multimodal distribution of the scribing quality. The hierarchical optimization strategy based on the material response probability model performs global path planning by improving the ant colony algorithm and combines quadratic programming for local parameter optimization, achieving the global optimum of the two-optical-path laser scribing path under the condition of satisfying time-varying constraints, and balancing the scribing quality and efficiency. The real-time monitoring and closed-loop adjustment system realizes the dynamic adjustment of the scribing process. When it is detected that the material response deviation exceeds the threshold, the control parameters are adjusted online through the model predictive control algorithm and Bayesian update, enabling the system to have the adaptive ability to changes in material characteristics and production environment fluctuations, and ensuring the stability and consistency of the scribing quality.

[0037] Refer to Figure 3 In the embodiment of the present invention, a computer device is further provided. The computer device may be a server, and its internal structure may be as Figure 3As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface, and a database connected via a system bus. Among them, the processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal via a network connection. The computer program, when executed by the processor, implements the above method.

[0038] Those skilled in the art can understand that Figure 3 the structure shown in is only a block diagram of some structures related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0039] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0040] Those of ordinary skill in the art can understand that all or part of the processes in the above embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or an external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0041] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, systems, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0042] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0043] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of the present invention.

Claims

1. A laser scribing planning method, characterized in that, Including: Conduct kinematic and laser-material interaction dynamics modeling on the dual optical path laser system, and preprocess the historical scribing data to obtain a preprocessed time series dataset; Input the preprocessed time series dataset into a time feature extractor to perform time dynamic feature extraction of the historical scribing trajectory, and obtain a time feature vector; Construct a graph attention network based on the time feature vector and the solar cell surface grid division, and calculate the spatial interaction features between the dual optical paths to obtain a spatial feature vector; Input the time feature vector and the spatial feature vector into a mixture density network to establish a material response probability model; Based on the material response probability model, perform hierarchical optimization of the scribing path to obtain an optimal control parameter sequence for the dual optical path laser head.

2. The laser scribing planning method according to claim 1, wherein The conduct kinematic and laser-material interaction dynamics modeling on the dual optical path laser system, and preprocess the historical scribing data to obtain a preprocessed time series dataset, including: Define three-dimensional position vectors, velocity vectors, and acceleration vectors based on the dual optical path laser system, and introduce velocity constraints and acceleration constraints to obtain the kinematic model of the dual optical path laser system; Establish a propagation process model of laser energy in solar cell materials based on the heat conduction equation, and introduce a material response function to describe the material removal efficiency to obtain a laser-material interaction dynamics model; Collect the laser power sequence, optical path position sequence, material surface temperature distribution matrix, and scribing result evaluation index of the dual optical path laser system based on the kinematic model and the laser-material interaction dynamics model to obtain historical scribing data; Perform filtering processing on the historical scribing data to obtain cleaned scribing data, and classify the material response events in the cleaned scribing data to obtain a preprocessed time series dataset.

3. The laser scribing planning method according to claim 1, characterized in that The input the preprocessed time series dataset into a time feature extractor to perform time dynamic feature extraction of the historical scribing trajectory, and obtain a time feature vector, including: Perform a linear transformation on the preprocessed time series dataset to obtain a query matrix, a key matrix, and a value matrix; Calculate an attention score matrix based on the query matrix and the key matrix, normalize it through a scaling factor, and apply the softmax function to obtain an attention weight matrix; Input the attention weight matrix and the value matrix into the Transformer encoder layer containing a feed-forward neural network in the time feature extractor to obtain a long-range dependence feature representation; Input the long-range dependence feature representation into the gated recurrent unit in the time feature extractor, and control the information flow by calculating the update gate and the reset gate to obtain a gated state; Calculate a candidate hidden state based on the gated state, and control the selective memory and forgetting of information through the update gate to obtain local temporal features; Input the local temporal features into a multi-scale temporal convolutional module with different convolutional kernel sizes in the time feature extractor, and simultaneously capture short-term, medium-term, and long-term time patterns through parallel 1D convolutional layers to obtain a time feature vector.

4. The laser scribing planning method according to claim 1, wherein Construct a graph attention network according to the time feature vector and the surface grid division of the solar cell, and calculate the spatial interaction features between the two optical paths to obtain a spatial feature vector, including: Divide the surface of the solar cell into N×M grid cells as the node set, and assign position coordinates, material thickness, reflectivity, thermal conductivity, and the time feature vector to each node to obtain a node feature matrix; Construct an undirected graph based on the node feature matrix, generate an adjacency matrix by calculating the spatial distance between nodes, and obtain the graph structure representation of the grid nodes; Apply the multi-head attention mechanism to the graph structure representation of the grid nodes to obtain attention weights; Calculate the output of each attention head according to the attention weights, and obtain the node-level feature representation by concatenating the outputs of all attention heads; Design a two-optical-path interaction module based on the node-level feature representation, model the material responses at the current positions of the two optical paths as two subgraphs, and quantify the superposition effect of the thermal influence of the two optical paths by calculating the inter-graph attention coefficient to obtain a two-optical-path interaction feature matrix; Apply the differential pooling algorithm to the two-optical-path interaction feature matrix to obtain a spatial feature vector.

5. The laser scribing planning method according to claim 1, characterized in that Input the time feature vector and the spatial feature vector into a mixture density network to establish a material response probability model, including: Fuse the time feature vector and the spatial feature vector to obtain a spatio-temporal joint feature vector; Input the spatio-temporal joint feature vector into a fully-connected neural network for hidden feature extraction to obtain the intermediate layer representation of the mixture density network, and generate a Gaussian mixture model parameter set based on the intermediate layer representation; Construct a conditional probability distribution according to the Gaussian mixture model parameter set to obtain a multi-modal distribution representation of the scribing quality; Define a negative log-likelihood loss function for the multi-modal distribution representation, introduce a component regularization term at the same time, and optimize the network parameters through backpropagation to obtain a preliminarily trained mixture density network; Based on the preliminarily trained mixture density network, construct a conditional variational autoencoder, enhance the uncertainty modeling ability through the encoder and decoder, and calculate the calibration error index to obtain a material response probability model.

6. The laser scribing planning method according to claim 1, wherein Based on the material response probability model, perform hierarchical optimization of the scribing path to obtain the optimal control parameter sequence of the two-optical-path laser head, including: Define the scribing task on the surface of the solar cell, represent each scribing segment as a triple including the starting point, the ending point, and the target scribing quality parameter, and construct a scribing task set; Construct a weighted directed graph model based on the scribing task set, apply the ant colony optimization algorithm to the weighted directed graph model, control the exploration of ants through the state transition probability, and adjust the pheromone according to the pheromone update rule to obtain the globally optimal scribing path; For each scribing segment on the globally optimal scribing path, define a decision variable vector to represent the moving speeds, laser powers, and time offsets of the two optical paths, calculate the objective function through the material response probability model, and introduce device constraint conditions at the same time to obtain a quadratic programming problem; Solve the quadratic programming problem using the augmented Lagrangian multiplier method, and obtain the optimal parameter configuration of each scribing segment through iterative calculation. Based on the optimal parameter configuration, samples are taken from the material response probability model to generate multiple material response scenarios, and the optimal parameter combination for each material response scenario is selected to obtain the optimal control parameter sequence of the dual-path laser head.

7. The laser scribing planning method according to claim 1, wherein The laser scribing planning method further includes: Deploy a real-time monitoring system including a laser power monitor, an optical path position sensor, a high-speed camera, an infrared thermal imager, and a surface profile scanner, and synchronously collect data on the laser scribing process through a high-speed data acquisition card to obtain the original monitoring data stream; Perform Kalman filtering on the original monitoring data stream to obtain the processed monitoring data, and at the same time generate a three-layer closed-loop control architecture including reference generation, state estimation, and control execution based on the optimal control parameter sequence; Use an unscented Kalman filter to perform multi-sensor information fusion on the processed monitoring data, and estimate the system state and material response state in real time through the state propagation equation and the measurement update equation to obtain the current system state estimate; Perform model predictive control on the current system state estimate based on the three-layer closed-loop control architecture, solve the optimization problem within the rolling time domain to obtain the optimal control input sequence; Calculate the material response deviation index according to the optimal control input sequence. When the material response deviation index exceeds the preset threshold, use the recursive least squares method to adjust the model parameters in real time, correct the material response probability model through Bayesian update, and set multiple safety thresholds to automatically trigger control parameter adjustment to obtain an adaptive closed-loop control strategy.

8. A laser scribing planning device, characterized in that, For implementing the laser scribing planning method according to any one of claims 1-7, the laser scribing device includes: A preprocessing module for performing kinematic and laser-material interaction dynamics modeling on the dual-path laser system and preprocessing the historical scribing data to obtain a preprocessed time series data set; A feature extraction module for inputting the preprocessed time series data set into a time feature extractor to perform time dynamic feature extraction of the historical scribing trajectory to obtain a time feature vector; A calculation module for constructing a graph attention network according to the time feature vector and the surface grid division of the solar cell, and calculating the spatial interaction feature between the two optical paths to obtain a spatial feature vector; A building module for inputting the time feature vector and the spatial feature vector into a mixture density network to build a material response probability model; A hierarchical optimization module for performing hierarchical optimization of the scribing path based on the material response probability model to obtain the optimal control parameter sequence of the dual-path laser head.

9. A computer device, characterized in that, It includes a memory and a processor. The memory stores a computer program that can run on the processor, and when the processor executes the computer program, it implements the laser scribing planning method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and when the computer program is run by the processor, the processor is caused to execute the laser scribing planning method according to any one of claims 1 to 7.

Images