Laser detection method, device and equipment of laser scribing machine and storage medium

Abstract

The invention relates to a laser detection method, device and equipment of a laser scribing machine and a storage medium. The method comprises the following steps: carrying out laser scanning on the surface of a solar cell, collecting a reflected light intensity signal and carrying out adaptive domain transformation processing to obtain a target characteristic matrix; training a material characteristic predictor based on the target characteristic matrix, and predicting material characteristic parameters on the lineation path to obtain a material characteristic prediction value; a state space model of the dual-optical-path laser system is constructed according to the material characteristic predicted value, a Lyapunov control function is established by solving a Lyapunov equation, and constraint conditions for ensuring the stability of the system are obtained; and the constraint condition and the material characteristic prediction value are substituted into a model prediction control optimization problem, a quadratic programming problem is solved through an interior point method, and an optimal control sequence is obtained. According to the invention, accurate prediction of the material characteristics on the scribing path is realized, scribing abnormity caused by sudden change of the material characteristics is reduced, and the anti-interference capability and the control precision of the system are enhanced.

Description

Technical Field

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

[0002] Laser scribing has been widely used in the manufacturing process of solar cells and is a key process for achieving electrode patterning and electrical isolation. With the diversified development of solar cell materials and structures, traditional laser scribing technologies are facing severe challenges. In actual production, there are significant differences in the surface material density and surface conditions of solar cells, which directly affect the propagation and analysis of laser signals, resulting in problems such as insufficient scribing accuracy, poor consistency, and low efficiency. Especially when the laser crosses between different material interfaces, the changes in reflectivity and scattering characteristics will generate signal perturbations, and these perturbations are extremely similar to the signal changes caused by micro-defects on the surface of solar cells, making it complex and difficult to accurately identify surface features.

[0003] Existing laser scribing technologies mostly adopt single-light-path systems and lack the ability to perceive and adapt to changes in material properties in real time. Such systems usually rely on fixed laser parameter settings and cannot dynamically adjust according to the material properties of different regions, and are prone to over-cutting or under-cutting phenomena when encountering material inhomogeneity. At the same time, traditional path planning control algorithms have insufficient robustness to material changes and are difficult to meet the accuracy requirements under large material changes. Especially in the field of high-efficiency photovoltaic manufacturing, even a small change in material properties may lead to a significant difference in energy absorption, which has a significant impact on the scribing quality, and the existing technologies lack effective sensing and response mechanisms. Summary of the Invention

[0004] The main purpose of the present invention is to provide a laser detection method, device, equipment and storage medium for a laser scribing machine. The present invention realizes the accurate prediction of material properties on the scribing path, reduces scribing abnormalities caused by sudden changes in material properties, and enhances the anti-interference ability and control accuracy of the system.

[0005] To achieve the above purpose, the present invention provides a laser detection method for a laser scribing machine, including the following steps: Perform laser scanning on the surface of the solar cell, collect the reflected light intensity signal, and perform adaptive domain transformation processing on the reflected light intensity signal to obtain a target feature matrix; Train a material property predictor with a hybrid architecture of a bidirectional long short-term memory network and a temporal convolutional network based on the target feature matrix, and predict the material property parameters on the scribing path to obtain a material property prediction value; Construct a state - space model of the double - path laser system based on the predicted values of the material properties, establish a control Lyapunov function by solving the Lyapunov equation, and obtain the constraint conditions for ensuring the stability of the system; Substitute the constraint conditions and the predicted values of the material properties into the model predictive control optimization problem, and solve the quadratic programming problem by the interior - point method to obtain the optimal control sequence.

[0006] The present invention also provides a laser detection device for a laser scribing machine, including: An acquisition module, configured to perform laser scanning on the surface of a solar cell, acquire the reflected light intensity signal, and perform adaptive domain transformation processing on the reflected light intensity signal to obtain a target feature matrix; A training module, configured to train a material property predictor with a hybrid architecture of a bidirectional long - short - term memory network and a temporal convolutional network based on the target feature matrix, and predict the material property parameters on the scribing path to obtain the predicted values of the material properties; A construction module, configured to construct a state - space model of the double - path laser system based on the predicted values of the material properties, establish a control Lyapunov function by solving the Lyapunov equation, and obtain the constraint conditions for ensuring the stability of the system; A solution module, configured to substitute the constraint conditions and the predicted values of the material properties into the model predictive control optimization problem, and solve the quadratic programming problem by the interior - point method to obtain the optimal control sequence.

[0007] The present invention also provides a computer device, including a memory and a processor, where a computer program is stored in the memory, and when the processor executes the computer program, the steps of the method described in any one of the above are implemented.

[0008] The present invention also provides a computer - readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in any one of the above are implemented.

[0009] In summary, the technical solution provided by the present invention constructs a dual - optical - path laser system including a detection optical path and a scribing optical path, and achieves a co - axis design with a position deviation of no more than 5μm through precise optical calibration. The detection optical path is made to operate ahead of the scribing optical path, forming a closed - loop control process of "detection - prediction - scribing". It can sense in advance before the material characteristics change, greatly improving the system's response ability and adaptability to material changes. The adaptive domain transformation processing is implemented using a deep neural network, mapping the signal feature matrices under different material conditions to a unified feature space, eliminating signal interference caused by material differences, while retaining surface feature information, enabling the scribing system to maintain consistent detection performance under various material conditions. Combining the hybrid architecture of bidirectional long - short - term memory network and temporal convolutional network, it realizes accurate prediction of material characteristics on the scribing path, reducing scribing abnormalities caused by sudden changes in material characteristics. By constructing a state - space model considering material characteristics and solving the Lyapunov equation, a control Lyapunov function and constraint conditions for ensuring system stability are established, providing theoretical stability guarantee for the scribing process under complex material conditions, effectively preventing the occurrence of divergence and instability phenomena in the control system. The laser power and scanning speed are adaptively adjusted according to the predicted values of material characteristics, ensuring consistent scribing quality under different material conditions, and at the same time, real - time optimization is achieved through model predictive control solved by the interior - point method, meeting the requirements of high - speed scribing. The spatial synchronization and temporal synchronization of the detection optical path and the scribing optical path are realized, and real - time update of model parameters and noise filtering of system states are achieved through recursive least - squares method and Kalman filter, further enhancing the system's anti - interference ability and control accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a schematic diagram of the steps of the laser detection method of the laser scribing machine in an embodiment of the present invention; Figure 2 is a block diagram of the structure of the laser detection device of the laser scribing machine in an embodiment of the present invention; Figure 3 is a schematic block diagram of the structure of a computer device in an embodiment of the present invention.

[0011] The realization, functional characteristics, and advantages of the object of the present invention will be further described in conjunction with the embodiments with reference to the drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] In order to make the object, technical solution, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0013] Referring to Figure 1 , this embodiment provides a laser detection method for a laser scribing machine, including the following steps: S1. Laser scan the surface of the solar cell, collect the reflected light intensity signal, and perform adaptive domain transformation processing on the reflected light intensity signal to obtain the target feature matrix; Among them, a dual - optical - path laser system including a detection optical path and a scribing optical path is constructed. The detection optical path is used to scan the surface of the solar cell, and the scribing optical path is used to perform the actual laser scribing operation. The detection optical path consists of a laser with a wavelength of 1064 nm, a focusing lens group, a beam splitter, and a photodetector. During the scanning process, the laser beam of the detection optical path irradiates the surface of the solar cell. Different material surfaces have different reflection abilities to the laser, and the intensity of the reflected light signal collected by the photodetector also varies. In order to ensure that the detection optical path and the scribing optical path are aligned with the same target position, high - precision optical calibration is adopted to keep them strictly co - axially aligned in space, and the maximum position deviation is controlled within the micron level, so as to ensure that the scanning data of the detection optical path can accurately reflect the processing area of the scribing optical path. The collected reflected light intensity signal is subjected to analog - to - digital conversion in order to be converted into a digital signal for subsequent calculations. The analog - to - digital conversion is completed by a high - speed data acquisition card, and the sampling frequency of this data acquisition card is set to 50 MHz to ensure the accuracy and integrity of the signal conversion. The collected original signal is filtered to extract effective signal information. The filtering is performed using a band - pass filter, and the bandwidth range is set between 100 Hz and 5 kHz to effectively remove the high - frequency noise and low - frequency interference of the system, and the filtered signal is more stable. The filtered reflected light intensity signal is divided by the reflected light intensity of the reference substrate for normalization processing to obtain a normalized reflectivity signal, thereby eliminating the systematic errors caused by changes in the measurement environment or different reflection abilities of the material surface, and improving the stability and comparability of the signal. The normalized reflectivity signal is subjected to time - domain analysis to extract key characteristic parameters. During the time - domain analysis process, the amplitude, rise time, fall time, and duration of the signal are calculated. Among them, the signal amplitude reflects the reflection ability of the material to the laser at the detection point, the rise time and fall time respectively describe the time when the signal changes from the initial state to the peak and from the peak back to the initial state, and the duration characterizes the stability of the signal, which helps to analyze the uniformity of the material surface and identify the defective areas encountered during the scribing process. In order to obtain the frequency - domain characteristics of the signal, a fast Fourier transform is performed on the normalized reflectivity signal to convert the time - domain signal into frequency components, and the amplitude and phase information of the main frequency components are extracted. The time - domain characteristics and frequency - domain characteristics are combined to form a signal characteristic matrix. This matrix contains multiple sampling points, and each point contains a set of characteristic values, including the four parameters in the time - domain and the information of the main frequency components in the frequency - domain. The signal characteristic matrix is input into a deep neural network for adaptive domain transformation processing to eliminate the influence of different material types on the signal characteristics, so that the signals under different material conditions have a more stable distribution. The adaptive domain transformation is realized through a deep neural network. A feature space mapping function is constructed to map the original signal characteristics to a unified feature space, reducing the interference of changes in material properties on the signal characteristics.The deep neural network adopts a four-layer fully connected structure, with different numbers of neurons in each layer. The previous layers use non-linear activation functions for feature extraction, and the last layer maps the output features to a fixed range through normalization to ensure the consistency of features under different material conditions. To ensure that the neural network can accurately learn the feature distributions under different material conditions, the training dataset contains a variety of different material samples, and each sample contains a feature matrix, a material category label, and a surface feature category label. During the training process, an adversarial learning strategy is adopted to optimize the network weights, making the material classifier unable to distinguish different materials, while the surface feature classifier can accurately identify the micro-defects and structural features on the material surface. The feature mapping function obtained through training acts on the signal feature matrix to convert it into a unified feature representation, that is, the target feature matrix.

[0014] Input the signal feature matrix into the first fully connected layer of the deep neural network. This layer performs a linear transformation by calculating the product of the input signal feature matrix and the first-layer weight matrix and adding a bias vector to obtain the output features of the first layer. The first layer is responsible for initially extracting the main features in the signal and converting the input data into a form suitable for processing by subsequent layers, enabling it to have stronger expressive power to eliminate the interference of material properties. Input the output features of the first layer into the second fully connected layer of the deep neural network. This layer performs calculations of the weight matrix and the bias vector and conducts a non-linear transformation on the input features to enhance the network's learning ability for complex signal patterns. The neural network can gradually adjust the distribution of features so that signals under different material conditions have similar representation methods in the transformed feature space, thereby reducing the differences caused by material properties. Apply batch normalization to the output features of the second layer to reduce the distribution deviation of data in different batches and improve the stability of network training. Calculate the mean and variance of the output features of this layer, and then standardize each sample so that its mean is close to zero and its variance is close to one. The normalized features can effectively accelerate the convergence speed of the network and alleviate the problems of vanishing gradients or exploding gradients that occur during the training of deep networks, enabling the network to have better generalization ability under different material conditions. Input the normalized features into the third fully connected layer of the deep neural network. This layer performs matrix multiplication and bias addition on the input features to further adjust the feature distribution and enhance the distinctiveness of the features. The role of this step is to optimize the linear separability of the features so that samples with the same surface features are closer in the new feature space, while samples with different surface features are more dispersed. Input the output features of the third layer into the fourth fully connected layer of the deep neural network. A linear transformation is performed in this layer so that the dimension of the final feature vector becomes 128. During this process, the deep neural network further compresses and optimizes the feature representation, making the output features more compact while retaining the most discriminative information. To prevent the deep neural network from overfitting during training, apply random masking and random zeroing to the feature vector with a dimension of 128. By randomly setting the outputs of some neurons to zero during training, reduce the network's dependence on specific features, thereby enhancing the generalization ability of the model. The random masking is implemented by masking some values in the feature vector with a certain probability in each training iteration, so that the network does not always rely on fixed feature components in different training batches, thereby improving its adaptability under different material surface conditions. Map the feature vector after random zeroing to the target feature space through a linear projection matrix. This mapping process converts it into a d-dimensional target feature matrix by calculating the product of the feature vector and the projection matrix.

[0015] S2. Based on the target feature matrix, train a material property predictor with a hybrid architecture of a bidirectional long short-term memory network and a temporal convolutional network, and predict the material property parameters on the scribing path to obtain the material property prediction values;Specifically, a material property predictor is constructed. The predictor adopts a hybrid architecture of bidirectional long short-term memory network and temporal convolutional network, and organizes the training sample set based on the temporal relationship of the target feature matrix to ensure that the model fully learns the dynamic changes of material properties along the scribing path. Since the target feature matrix contains the normalized reflectivity signal and its time-frequency features obtained from the laser detection process, in order to enable the material property predictor to accurately predict the material property parameters along the scribing path, the training sample set is constructed in chronological order. The target feature matrix within the historical time window is used as the input, and the future material property parameters are used as the supervision signal for training. The material property predictor is trained based on the training sample set, and a multi-task learning strategy is adopted to optimize the network so that it can simultaneously predict multiple material property parameters. The multi-task learning strategy enables different tasks to complement each other's information by sharing the first few layers of the feature extraction part of the network, thereby improving the overall prediction accuracy. During the training process, the input sequence of target feature matrices passes through the temporal convolutional network, which performs one-dimensional convolutional operations in the time dimension through a sliding window to capture the change patterns of features over time. The temporal convolutional network adopts a multi-layer residual block structure and expands the receptive field through dilated convolution, enabling the network to focus on the material property changes over a longer time range. After being processed by the temporal convolutional network, the extracted time series features are then input into the bidirectional long short-term memory network to further capture the dynamic evolution of features over time. The design of the bidirectional long short-term memory network can consider both historical information and future information, enabling the prediction model to not only utilize past material properties but also learn potential temporal features through bidirectional dependencies, improving the prediction accuracy. After training is completed, the material property predictor is used to predict the material property parameters along the scribing path. For prediction, the historical target feature matrix sequence is input into the trained predictor, and a dual-path feature representation is obtained. The dual-path feature representation refers to the results obtained by the model from two different feature extraction paths of the temporal convolutional network and the bidirectional long short-term memory network respectively. The temporal convolutional network extracts local temporal dependencies, while the bidirectional long short-term memory network focuses on global temporal dependency information. Since these two feature representations are complementary in information expression, after obtaining the dual-path feature representation, a feature fusion operation is performed to merge the two different time dependency patterns and obtain a fused feature representation. The feature fusion operation adopts the method of weighted feature concatenation. By calculating the weighted sum of the output features of the temporal convolutional network and the bidirectional long short-term memory network, the final fused feature can take into account both short-term local patterns and long-term global trends. Dimensionality reduction processing is performed on the fused feature to reduce redundant information, improve computational efficiency, and enhance the generalization ability of the model. The fused feature representation is input into the multi-task output module, which includes multiple independent regression sub-networks. Each regression sub-network is used to predict specific material property parameters, including the reflectivity, scattering coefficient, thermal conductivity, heat capacity, and material density of the material.Each regression sub-network adopts a fully connected layer structure and is transformed through a non-linear activation function to ensure that the prediction results can be accurately mapped to the actual range of material properties. During the training process, the multi-task output module assigns different loss weights according to the importance of different tasks and optimizes the network parameters through a weighted mean square error loss function, enabling the model to take into account the differences between different material property parameters while ensuring the overall prediction accuracy. The trained material property predictor can predict the changes in material properties on the future path based on the input sequence of historical target feature matrices and obtain the predicted values of material properties.

[0016] S3. Construct the state space model of the double optical path laser system according to the predicted values of material properties, establish the control Lyapunov function by solving the Lyapunov equation, and obtain the constraint conditions to ensure the stability of the system; It should be noted that the state vector of the dual - optical - path laser system is constructed. This state vector contains the key motion information of the detection optical path and the scribing optical path, including position, velocity, and acceleration, which describes the dynamic characteristics of the system. At the same time, the control input vector is defined. This vector includes the drive signals of the two optical paths and is used to control the precise execution of laser scanning and scribing. The system output vector is defined as the actual scribing position to ensure the accuracy and stability of the scribing path. Through the construction of this series of variables, a basic variable set for state - space representation is formed. Based on the basic variable set of state - space representation and the predicted values of material properties, a non - linear state - transfer function is constructed to obtain the non - linear system dynamic equation. Since during the laser scribing process, material properties affect the absorption of laser energy, heat dissipation, and the efficiency of material removal, the dynamic behavior of the system not only depends on the control input and the current state but is also affected by changes in material properties. To accurately describe this dynamic relationship, the predicted values of material properties are incorporated into the state - transfer function, enabling the system to adapt to the dynamic changes of the material surface and thus ensuring the consistency of scribing quality. At this time, the non - linear system dynamic equation not only reflects the basic physical process of laser scanning and scribing but also takes into account the influence of material properties on motion and energy distribution. The non - linear system dynamic equation is linearly approximated by a first - order Taylor expansion at the current operating point to achieve local linearization and calculate the Jacobian matrix for constructing a time - varying linear model. The core of the linearization process lies in solving the linear approximation of the system under small - range perturbations, making subsequent optimization calculations more efficient and feasible. By calculating the partial derivatives of the state variables with respect to the system dynamic equation, the Jacobian matrix of the state equation is obtained, and combined with the predicted values of material properties, the linearized model of the system is updated at each time step to ensure that the system model can accurately reflect the current operating state. Based on the obtained time - varying linear model, an appropriate positive - definite weight matrix is selected, and a positive - definite matrix is obtained by solving the discrete - time Lyapunov equation to construct a quadratic - form - based control Lyapunov function. The solution of the Lyapunov equation guarantees the stability of the system, and by optimizing the weight matrix, the state of the system can gradually converge to the desired trajectory over time. The construction of the control Lyapunov function is based on the quadratic - form, ensuring that the energy function of the system has a decreasing property, that is, at any initial state, as time goes by, the system state always converges to the stable point without divergence or unstable oscillation. On this basis, the Lyapunov stability constraint is designed according to the control Lyapunov function and transformed into an explicit constraint form regarding the control input to obtain the control constraint that can ensure the stability of the system. The key to this process is to ensure that the rate of change of the Lyapunov function is always negative, enabling the system to operate stably. By embedding the Lyapunov stability constraint into the optimization problem, it is ensured that the control input can not only meet the basic system dynamic constraints but also guarantee that the system remains stable during execution and adapts to changes under different material conditions.

[0017] S4. Substitute the constraint conditions and the predicted values of material properties into the model predictive control optimization problem, and solve the quadratic programming problem by the interior point method to obtain the optimal control sequence.

[0018] Specifically, the surface of the solar cell is divided into m×n grid cells, and a corresponding material property vector is assigned to each grid cell to establish a material distribution model. Since there are differences in material properties in different regions of the solar cell surface, according to the requirements of the cell scribing design, a set of target paths is determined to ensure that the scribing process adapts to the physical properties of different material regions, thereby improving the scribing accuracy and consistency. Through the grid-based modeling method, a high-resolution property distribution is formed on the entire surface of the solar cell to obtain a grid-based cell model. An energy density model is established based on the grid-based cell model and the predicted material property values to describe the physical process of the interaction between the laser and the material. The energy density model is used to calculate the laser energy received by the material per unit area during the scribing process, which is affected by the laser power, the scanning speed, and the material absorptivity. Based on the energy density model, an adaptive adjustment function for the laser power and an adaptive adjustment function for the scanning speed are designed, enabling the system to dynamically adjust the laser parameters in different material regions, thereby ensuring the consistency of the scribing quality. The adaptive adjustment function of the laser power is adjusted according to the changes in the reflectivity and thermal conductivity of the material to compensate for the energy distribution differences caused by different material absorptivities, while the adaptive adjustment function of the scanning speed is optimized based on the thermal diffusion characteristics of the material to ensure the stability of energy transmission during the scribing process. A model predictive control optimization objective function is constructed to achieve the optimal control of the laser power and the scanning speed during the scribing process. This optimization objective function considers three aspects: minimizing the deviation between the scribing trajectory and the target path to ensure the scribing accuracy; optimizing the laser energy distribution to make it uniform throughout the path range, thereby reducing the inconsistency of the scribing quality; minimizing the change rate of the system control input to reduce the impact on the actuator and improve the stability of the system. On the basis of constructing the optimization objective function, combined with the constraint conditions, it is ensured that the physical feasibility is satisfied during the optimization process. The laser power needs to be limited within the maximum and minimum power ranges to prevent damage to the material caused by excessive power or incomplete scribing caused by too low power, while the range of the scanning speed needs to meet the physical constraints of the actuator to ensure the feasibility of the movement. The change rate of the control input also needs to be limited to avoid instability caused by sudden changes. In the final optimization problem, the adaptive adjustment function of the laser power and the adaptive adjustment function of the scanning speed are used as decision variables and solved in the form of quadratic programming. To efficiently solve the above quadratic programming problem of power and speed, the interior point method is used for optimization calculation. The interior point method converts the constrained optimization problem into an unconstrained optimization problem by introducing a barrier function, thereby performing smooth optimization within the feasible region. By constructing a barrier function, the constraint conditions of the laser power and the scanning speed are converted into a part of the internal optimization objective, and the influence of the barrier term is gradually weakened during the optimization process, so that the optimization solution finally approaches the true optimal solution. On this basis, the Newton method is used for iterative solution to quickly find the optimal control parameters.The application of Newton's method can accelerate the convergence process and obtain a high-precision solution with fewer iterations, enabling the entire optimization process to be completed in a shorter time and meeting the real-time control requirements of the laser scribing machine. By solving this optimization problem, an optimal control sequence containing laser power and scanning speed parameters is obtained and applied to the scribing process to ensure uniform, stable, and efficient scribing effects in different material regions, thereby improving the quality and production efficiency of solar cell manufacturing.

[0019] Construct a dual - optical - path collaborative control architecture including a path actuator, a feedback corrector, and a dynamic adjuster, decompose the optimal control sequence to obtain the detection optical path control signal and the scribing optical path control signal. The path actuator analyzes the optimal control sequence and reasonably distributes the control signals according to the physical constraints of the system and the dynamic characteristics of the actuator, enabling the detection optical path and the scribing optical path to move according to the established plan. After completing the decomposition of the control signals, calculate the system processing delay and calculate the leading distance of the detection optical path relative to the scribing optical path based on this delay. Since the detection optical path needs to scan the surface of the solar cell in advance and obtain material characteristic information, while the scribing optical path needs to scribe based on the data provided by the detection optical path, it is ensured that the detection optical path is spatially leading while being strictly synchronized with the scribing optical path in time. Drive the detection optical path to run along the scribing path through the detection optical path control signal and maintain the dynamic leading distance from the scribing optical path, enabling it to collect the required surface information in advance, forming a dual - optical - path operation mode that is leading in space but synchronized in time. The realization of this mode depends on precise time control and synchronization mechanisms to ensure that the data of the detection optical path can be transmitted and applied to the control decision before the scribing optical path executes, thereby effectively improving the response speed and scribing accuracy of the system. After establishing the dual - optical - path operation mode, use the detection optical path to collect the real - time reflection signal on the surface of the solar cell and perform fast feature extraction on this signal to update the material characteristic prediction model in real time. The photodetector of the detection optical path scans the surface of the solar cell at a high frequency and records the reflection signals at different positions. After pre - processing this signal, extract the key time - domain and frequency - domain features, including the amplitude, rise time, fall time, and duration of the signal, etc. After obtaining these features, update the model parameters of the material characteristic predictor online based on the recursive least - squares method, enabling it to adapt to the changes in material characteristics in different regions and correct the prediction error. The application of the recursive least - squares method enables the model to perform incremental updates each time new observation data is input, without the need to retrain the entire neural network, thereby improving the computational efficiency and ensuring that the material characteristic predictor can always reflect the current material state. Compare the actual material characteristics collected by the detection optical path with the predicted material characteristics of the target characteristic predictor and calculate the error vector characterizing the prediction accuracy. The error vector is used to measure the deviation between the prediction model and the actual observation data and provides a basis for the dynamic adjustment of subsequent control parameters. Since the material characteristics may shift due to changes in the manufacturing process or interference from environmental factors, continuously monitor the prediction error and adopt corresponding compensation strategies to maintain the stability and control accuracy of the system. After obtaining the error vector, dynamically adjust the weight matrix in the control Lyapunov function based on this error, and at the same time finely adjust the laser power and scanning speed of the scribing optical path to optimize the control parameters and reduce the scribing error.The weight adjustment of the control Lyapunov function is based on the magnitude and change trend of the error to ensure that the system can more accurately track the target path while meeting the stability constraints. The laser power and scanning speed of the scribing optical path also need to be appropriately adjusted according to the prediction error to compensate for the influence of material property changes on energy absorption and heat conduction. For example, if the deviation between the actual material properties and the predicted values is large, the laser power is increased or the scanning speed is decreased to ensure that the scribing depth and width remain consistent. When the error is small, the amplitude of the power adjustment is reduced to maintain the stable operation of the system. After optimizing the control parameters, clock synchronization is performed on the detection optical path and the scribing optical path according to the new control parameters to ensure the coordinated operation of the entire system. At the same time, to improve the robustness and anti-interference ability of the system, a Kalman filter is introduced for system state estimation, enabling the control system to still maintain high detection accuracy and control stability in the presence of noise and external interference. The Kalman filter predicts the future state based on the current observation data and historical states and compensates for the system noise in real time, thereby improving the accuracy of signal processing and optimizing the feature extraction results of the detection optical path. Through the above collaborative control strategy, the laser scribing machine can achieve high-precision scribing operations in a complex material environment, while ensuring the strict synchronization of the detection optical path and the scribing optical path, thereby improving the scribing quality and enhancing the stability and reliability of the solar cell manufacturing process.

[0020] In one example, a laser scan is performed on the surface of a solar cell to collect the reflected light intensity signal, and an adaptive domain transformation process is performed on the reflected light intensity signal to obtain a target feature matrix, including: A dual-optical-path laser system including a detection optical path and a scribing optical path is constructed, and the detection optical path is used to scan the surface of the solar cell. The reflected light intensity signal is collected by a photodetector in the detection optical path; The reflected light intensity signal is subjected to analog-to-digital conversion to obtain a digitized reflected light intensity signal, and the digitized reflected light intensity signal is filtered to obtain a filtered reflected light intensity signal; The filtered reflected light intensity signal is divided by the reflected light intensity of the reference substrate for normalization processing to obtain a normalized reflectance signal; Time-domain analysis is performed on the normalized reflectance signal to obtain time-domain characteristic parameters including signal amplitude, rise time, fall time, and duration. At the same time, a fast Fourier transform is performed on the normalized reflectance signal to obtain frequency-domain characteristics, and the time-domain characteristic parameters and frequency-domain characteristics are combined into a signal characteristic matrix; The signal characteristic matrix is input into a deep neural network for adaptive domain transformation processing to eliminate signal differences under different material conditions and obtain a target feature matrix.

[0021] In this example, a high-precision optical system is built. The system includes a detection optical path for scanning the surface of a solar cell and a scribing optical path for performing laser scribing. The detection optical path uses a laser with a wavelength of 1064 nm. Its laser beam irradiates the surface of the solar cell after passing through a focusing lens group, and the reflected signal is transmitted to a photodetector through a beam splitter. Since different materials have different reflection abilities to laser, the photodetector can obtain the reflected light intensity signal related to the material properties. At the same time, in order to ensure that the detection optical path and the scribing optical path act on the same position, a high-precision optical calibration method is adopted to strictly align the two in space and control the position deviation within the micron level, so as to ensure that the scanning result of the detection optical path can accurately reflect the processing area of the scribing optical path. The reflected light intensity signal is subjected to analog-to-digital conversion to convert the analog signal into a digital signal for subsequent calculation and processing. The analog-to-digital conversion is completed by a high-speed data acquisition card, and the sampling frequency is set to 50 MHz to ensure high-precision conversion of the reflected light intensity signal. After the analog-to-digital conversion is completed, the obtained digital reflected light intensity signal contains noise and environmental interference, so filtering processing is performed to remove irrelevant information and improve the stability of the signal. The filtering uses a band-pass filter, and its bandwidth range is set between 100 Hz and 5 kHz, so as to effectively remove high-frequency noise and low-frequency interference and make the signal more stable. The filtered signal is normalized to eliminate the signal intensity difference caused by changes in the material surface conditions or the detection environment. The ratio operation is performed between the filtered reflected light intensity signal and the reflected light intensity of the reference substrate to obtain a normalized reflectivity signal. Let be the normalized reflectivity signal, where is the coordinate of the scanning position, and its calculation method is expressed as:

[0022] where, is the reflected light intensity at a certain coordinate point on the surface of the solar cell, The intensity of the reflected light with the reference substrate. Through normalization, the measurement errors caused by different optical properties of the substrates are eliminated, making the signals on the surfaces of different materials have higher comparability and stability. Perform time-domain analysis on the standardized reflectivity signal to extract its key characteristic parameters. During the time-domain analysis, calculate the amplitude, rise time, fall time, and duration of the signal. Among them, the signal amplitude reflects the reflection ability of the material to the laser, while the rise time and fall time respectively describe the time required for the signal to reach the peak from the initial state and to return from the peak to the initial state. The duration characterizes the stability of the signal and helps to identify surface defects or material inhomogeneities of the solar cell. At the same time, in order to obtain more comprehensive signal characteristics, perform a fast Fourier transform on the standardized reflectivity signal to extract its frequency-domain information. The fast Fourier transform can convert the time-domain signal into frequency components and extract the amplitude and phase information of the main frequency components. By combining the time-domain characteristics and frequency-domain characteristics, a signal characteristic matrix is formed, where each sampling point contains a set of characteristic vectors composed of time-domain parameters and frequency-domain parameters. Input the signal characteristic matrix into a deep neural network for adaptive domain transformation processing to eliminate the signal differences under different material conditions, so that the detection system can maintain a consistent feature extraction ability in various material environments. The adaptive domain transformation constructs a nonlinear mapping function through a deep neural network, projects the original signal characteristic matrix into a unified feature space, makes the distribution of signals under different material conditions closer in this space, and thus reduces the signal differences caused by changes in material characteristics. The neural network adopts a multi-layer fully connected structure and introduces a nonlinear activation function in the hidden layer to enhance its learning ability. Let the input signal characteristic matrix be , and the target characteristic matrix after being mapped by the neural network is , then the transformation function is expressed as:

[0023] where, is the nonlinear transformation function of the neural network, is the trainable weight matrix of the neural network. By optimizing the parameters, the input signals under different materials are made to have a consistent distribution in the transformed feature space , thereby improving the robustness of the detection system.

[0024] In an example, input the signal characteristic matrix into a deep neural network for adaptive domain transformation processing to eliminate the signal differences under different material conditions, and obtain the target characteristic matrix, including: Input the signal characteristic matrix into the first fully connected layer of the deep neural network for linear transformation to obtain the output characteristics of the first layer; The first-layer output features are input into the second fully connected layer of the deep neural network for weight matrix and bias vector calculations to obtain the second-layer output features; Batch normalization is applied to the second-layer output features to obtain the normalized features; The normalized features are input into the third fully connected layer of the deep neural network to perform matrix multiplication and bias addition to obtain the third-layer output features; The third-layer output features are input into the fourth fully connected layer of the deep neural network to perform a linear transformation to obtain a feature vector with a dimension of 128; Random masking and random zeroing are applied to the feature vector with a dimension of 128 to prevent overfitting, and then the feature vector with a dimension of 128 is mapped to a d-dimensional space through a linear projection matrix to obtain the target feature matrix.

[0025] In this example, a deep neural network model suitable for material property learning is constructed, and the structure of the input data is defined. Assume that the input signal feature matrix is , which contains time-domain features and frequency-domain features collected from the surface of a solar cell, where each sample corresponds to the complete signal feature vector of a detection point. The dimension of this input matrix is , where is the number of sampling points is the feature dimension, and each sample contains various features from the time domain and the frequency domain, such as signal amplitude, rise time, fall time, duration, and the amplitude and phase information of the main frequency component. This signal feature matrix undergoes a linear transformation through the first fully connected layer to extract preliminary features. Assume that the weight matrix of the first layer is , and the bias vector is , then the first-layer output features are calculated as follows:

[0026] where, is a non-linear activation function, such as ReLU or LeakyReLU, whose purpose is to enhance the non-linear expression ability of the features. The dimension is , where is the number of neurons in the first layer, The dimension is , ensuring that each neuron has an independent bias term. The role of this layer is to extract high-level features from the original signal features while reducing the data dimension, enabling the neural network to learn the key patterns of the input data. The first-layer output features Passed as input to the second fully-connected layer, where more in-depth feature transformation is performed. Let the weight matrix of the second layer be , and the bias vector be . Then the output features of the second layer are calculated as follows:

[0027] where has a dimension of , represents the number of neurons in the second layer, and has a dimension of . This step can enhance the discriminative ability of the features, enabling the material features of similar classes to aggregate in the high-dimensional space, while the material features of different classes maintain a large interval, thereby enhancing the classification and prediction ability of the network. To optimize the training stability of the model and accelerate the convergence speed, batch normalization is performed on the output features of the second layer. The calculation formula for batch normalization is as follows:

[0028] where and represent the mean and standard deviation of the mini-batch samples respectively, and are learnable scaling and offset parameters. The role of batch normalization is to keep the data distribution stable, thereby alleviating the problems of gradient vanishing or gradient explosion, and at the same time enhancing the generalization ability of the model. Taking the normalized features as input, they are passed to the third fully-connected layer, which performs matrix multiplication and bias addition, and the calculation method is as follows:

[0029] where has a dimension of which is the number of neurons in the third layer, has a dimension of . The role of this layer is to further refine the features, enabling the network to form a more compact and effective representation in the high-dimensional space and improving the discriminative ability of the model. The output features of the third layer are passed into the fourth fully-connected layer and a linear transformation is performed, and the calculation method is as follows:

[0030] where has a dimension of to ensure that the output feature dimension is fixed at 128, has a dimension of . The network obtains a high-level feature vector of dimension 128, which contains the most discriminative information extracted from the original signal. To prevent the model from overfitting during training, the 128-dimensional feature vector is randomly masked and randomly set to zero, that is, the outputs of some neurons are set to zero with a certain probability. Suppose the probability of setting to zero is , then the mathematical representation of this operation is as follows:

[0031] where is a random mask matrix with the same shape as , and each element is independently sampled from the Bernoulli distribution , and © represents element-wise multiplication. The purpose of this operation is to force the network to learn more robust features during training and avoid the model's over-reliance on certain specific neurons. The 128-dimensional feature vector after random masking is mapped to the -dimensional target feature space through a linear projection matrix. Let the projection matrix be , then the target feature matrix is calculated as follows:

[0032] where has a dimension of , ensuring that the output target feature matrix meets the expected dimension. The role of this mapping is to convert high-dimensional features into a more compact representation suitable for subsequent processing, so that the features of different materials maintain the optimal discrimination ability in this space.

[0033] In one example, a material property predictor with a hybrid architecture of a bidirectional long short-term memory network and a temporal convolutional network is trained based on the target feature matrix, and the material property parameters on the scribing path are predicted to obtain material property prediction values, including: Construct a material property predictor, which includes a bidirectional long short-term memory network and a temporal convolutional network, and organize the target feature matrix into a training sample set according to the temporal relationship; Based on the training sample set, apply a multi-task learning strategy to train the material property predictor to obtain a trained material property predictor; Input the historical target feature matrix sequence on the scribing path to be predicted into the trained material property predictor to obtain a dual-path feature representation; Perform a feature fusion operation on the dual-path feature representation to obtain a fused feature representation; Input the fused feature representation into the multi-task output module, and calculate the material reflectivity, scattering coefficient, thermal conductivity, heat capacity, and material density respectively to obtain the material property prediction values.

[0034] In this example, a model architecture that combines a bidirectional long short-term memory network and a temporal convolutional network is designed, and the target feature matrix is organized into a training sample set according to the temporal relationship so that the network can learn the change patterns of material properties in the time dimension. The target feature matrix contains the standardized feature data collected from the surface of the solar cell, including time-domain features and frequency-domain features. Each sampling point corresponds to a time step. To train the material property predictor, these features are arranged in a time series to form a training sample set, where each sample consists of the target feature matrix of multiple consecutive time steps as the input sequence, and the material property parameters corresponding to the time steps, such as material reflectivity, scattering coefficient, thermal conductivity, heat capacity, and material density, are used as the supervision signals for model training. After constructing the training sample set, a multi-task learning strategy is applied to train the material property predictor to simultaneously optimize the prediction accuracy of multiple material property parameters. The core idea of multi-task learning is to share the first few layers of the feature extraction part of the model, so that different tasks can complement each other's information and improve the overall prediction effect. To achieve this goal, the input target feature matrix sequence is passed to the temporal convolutional network, which extracts the time series features through one-dimensional convolutional operations and uses dilated convolutions to expand the receptive field, enabling the network to focus on the material property changes in a longer time range. Let the output of the temporal convolutional network be , and its calculation method is:

[0035] where, is the weight of the convolutional kernel of the th layer, is the target feature matrix at time step , the symbol * represents the one-dimensional convolutional operation, is the size of the convolutional kernel. The role of the temporal convolutional network is to extract local temporal patterns from the input sequence, enabling the accurate modeling of the material change trend within a short time. The output of the temporal convolutional network is passed to the bidirectional long short-term memory network, which uses recurrent units in both the forward and backward directions to capture the long-term dependencies of material properties over time, thereby improving the prediction accuracy. Considering both past and future information makes the model's prediction of material properties more global and stable. By combining the local feature extraction ability of the temporal convolutional network and the long-term dependency modeling ability of the bidirectional long short-term memory network, the material property predictor can effectively learn the change trends of different material properties along the scribing path. After the model training is completed, the historical target feature matrix sequence on the scribing path to be predicted is input into the trained material property predictor to generate a dual-path feature representation. Among them, one path of features comes from the temporal convolutional network and is used to capture the material change patterns within a short time range, and the other path of features comes from the bidirectional long short-term memory network and is used to learn the trend information within a long time range. Since these two feature expression methods have their own advantages, a feature fusion operation is performed to combine them into a unified fused feature representation. The fused feature representation is input into the multi-task output module to calculate material property parameters such as material reflectivity, scattering coefficient, thermal conductivity, heat capacity, and material density respectively. The multi-task output module contains five parallel regression sub-networks, each network is responsible for predicting a material property parameter, and the fused features are transformed through a fully connected layer to generate the final prediction value. The loss function of each regression sub-network is the weighted mean squared error, and its optimization goal is to minimize the prediction errors of all material property parameters while ensuring a reasonable trade-off relationship between different tasks. The trained material property predictor can accurately predict the changes in material properties on the future scribing path based on the input historical target feature matrix sequence.

[0036] Among them, before predicting the material property parameters on the scribing path based on the unified feature representation to train the material property predictor of the hybrid architecture of the bidirectional long short-term memory network and the temporal convolutional network and obtaining the material property prediction values, and constructing the state space model of the double optical path laser system according to the material property prediction values, the following steps are further included: applying anomaly detection processing based on the local outlier factor algorithm to the material property prediction values, calculating the local density ratio of each prediction point, and obtaining the material property anomaly distribution map; for the potential anomaly points in the material property anomaly distribution map, extracting the historical detection data within the range of r = 5 mm around them, constructing a local Gaussian process model, and obtaining the local prediction model for anomaly point correction; based on the local prediction model, correcting the marked potential anomaly points, calculating the correction value, and obtaining the material property prediction values after anomaly point correction; according to the time series change characteristics of the material property prediction values, calculating the material property gradient, and applying an adaptive time window smoothing filter for processing to obtain the smoothed material property change curve; combining the smoothed material property change curve with the spatial position information, and adopting an improved bilateral filtering algorithm for spatio-temporal joint filtering to obtain the filtering result that eliminates noise while maintaining the sharpness of the material property boundary; constructing the spatial distribution confidence map of the material property based on the filtering result, calculating the confidence score for each prediction point, and obtaining the material property confidence distribution representing the prediction reliability; according to the material property confidence distribution, performing adaptive enhanced sampling on the low confidence region (S(p) < 0.6) to obtain the enhanced material property sampling point set; performing data fusion on the enhanced material property sampling point set and the raw material property prediction values, adopting confidence-based weighted average, and at the same time introducing spatial consistency constraints, and solving through the alternating direction multiplier method to obtain the final material property distribution that satisfies spatial smoothness, forming the enhanced material property prediction values for state space model construction.

[0037] In one example, according to the material property prediction values, a state space model of the double optical path laser system is constructed. By solving the Lyapunov equation to establish the control Lyapunov function, the constraint conditions for ensuring the stability of the system are obtained, including: Constructing the state vector of the double optical path laser system, where the state vector includes the position, velocity, and acceleration information of the detection optical path and the scribing optical path, the control input vector includes the drive signals of the two optical paths, and the system output vector is the actual scribing position, obtaining the basic variable set of the state space representation; Based on the basic variable set of the state space representation and the material property prediction values, constructing a nonlinear state transition function to obtain the nonlinear system dynamic equation; Performing first-order Taylor expansion linearization processing on the nonlinear system dynamic equation at the current operating point, calculating the Jacobian matrix, and obtaining the time-varying linear model; Select a positive definite weight matrix based on a time-varying linear model, and obtain a positive definite matrix by solving a discrete-time Lyapunov equation to construct a quadratic-form-based control Lyapunov function; Design Lyapunov stability constraints according to the control Lyapunov function, and transform the Lyapunov stability constraints into an explicit constraint form with respect to the control input to obtain control constraints that ensure system stability.

[0038] In this example, in order to construct the state vector of the dual-path laser system and establish a complete state description, the state vector should include the key dynamic information of the detection path and the scribing path to ensure that the system can accurately simulate its motion characteristics. Let the system state vector be which contains the position, velocity, and acceleration information of the detection path and the scribing path, that is:

[0039] where, and represent the positions of the detection path and the scribing path respectively, and represent the velocities of the two paths, and represent the accelerations of the two paths. To control the motion of the two paths, define the control input vector which contains the drive signals for the detection path and the scribing path:

[0040] where, is the control input for the detection path, is the control input for the scribing path. The system output vector is set to the actual scribing position that is:

[0041] Through this step, the basic variable set of the state-space representation is completely defined, where the system state consists of position, velocity, and acceleration, the control input is used to adjust the system dynamics, and the output directly corresponds to the target position of scribing. Based on the basic variable set of the state-space representation and the predicted values of material properties, construct a non-linear state transition function to describe the true dynamic behavior of the system. Since the material properties during scribing affect the absorption and dissipation of laser energy, thus affecting the scribing quality, the state transition function depends not only on the control input but also on the material property parameters. Let the material property vector be:

[0042] where, is the material reflectivity, is the scattering coefficient, is the thermal conductivity, is the heat capacity of the material, is the material density. The nonlinear state equation of the system is expressed as:

[0043] where, is the nonlinear state transition function, and its specific form depends on the physical characteristics of the laser system. For example, the absorption rate of the material will affect the actual energy input of the scribing optical path, thereby changing the dynamic behavior of the position and velocity. Since the analytical solution of the nonlinear system is relatively complex, in order to implement model predictive control, a first-order Taylor expansion is performed at the current operating point to obtain a linear approximation model. Let the current operating point be , perform a Taylor expansion on the state transition equation and ignore the high-order terms to obtain:

[0044] where,

[0045] The matrix is the partial derivative of the system state with respect to the state variable, describing the internal dynamic changes of the system. The matrix reflects the influence of the control input on the system state, while the matrix represents how the change of material properties affects the system behavior. This time-varying linear model can be updated at each time step to adapt to the state evolution under different material conditions. To ensure the stability of the system under different operating conditions, a positive definite weight matrix is selected based on the time-varying linear model, and a positive definite matrix is obtained by solving the discrete-time Lyapunov equation to construct a control Lyapunov function. The control Lyapunov function adopts a quadratic form:

[0046] where, is a positive definite matrix, obtained by solving the discrete-time Lyapunov equation:

[0047] where is the given positive definite weight matrix, ensuring that the Lyapunov function can decrease over time, thereby guaranteeing the convergence of the system. Based on the control Lyapunov function, a Lyapunov stability constraint is designed to ensure that the system remains stable throughout the prediction process. The Lyapunov stability condition requires:

[0048] where, is a positive number to ensure that the system state converges over time. Substitute the state equation into the stability constraint and transform it into an explicit constraint on the control input, we get:

[0049] Substitute the linearized state equation, we can get:

[0050] This constraint is used in the model predictive control optimization process to ensure that the control input obtained by optimization does not cause the system to diverge.

[0051] In one example, substitute the constraint conditions and the predicted values of material properties into the model predictive control optimization problem, and solve the quadratic programming problem by the interior point method to obtain the optimal control sequence, including: Divide the surface of the solar cell into m×n grid cells, assign a corresponding material property vector to each grid cell, and determine the target path set according to the battery scribing design requirements to obtain a gridded battery model for path planning; Establish an energy density model based on the gridded battery model and the predicted values of material properties, and design a laser power adaptive adjustment function and a scanning speed adaptive adjustment function according to the predicted values of material properties and the energy density model; Construct a model predictive control optimization objective function, and combine the optimization objective function and the constraint conditions to construct a power and speed quadratic programming problem for the laser power adaptive adjustment function and the scanning speed adaptive adjustment function; Apply the interior point method to solve the power and speed quadratic programming problem, use the barrier function to transform the constraint into an unconstrained problem, and find the optimal solution through Newton's method iteration to obtain the optimal control sequence including laser power and scanning speed parameters.

[0052] In this example, the surface of the solar cell is divided into grid cells, each grid cell corresponds to a spatial region, and a material property vector is assigned to each grid cell to accurately describe the physical properties of the region. Let the material property vector be:

[0053] where, represents the material property vector of the grid cell located in the th row and the th column, is the material reflectivity, is the scattering coefficient is the thermal conductivity, is the material heat capacity, is the material density. These parameters have an important impact on the interaction between the laser and the material. Therefore, when performing laser scribing, the laser power and scanning speed are dynamically adjusted according to the material properties of different grid cells to ensure uniform scribing quality. Based on the gridded battery model and the predicted values of material properties, an energy density model is established to describe the distribution of laser energy on the material. The energy density of the laser is affected by the laser power, scanning speed, and spot diameter. Let the energy density be:

[0054] where is the laser power at the grid cell, is the scanning speed, is the laser spot diameter. To ensure scribing uniformity, an adaptive adjustment function for laser power and an adaptive adjustment function for scanning speed are designed according to the predicted values of material properties. The power adjustment function should consider the reflectivity and thermal conductivity of the material to compensate for the difference in the absorption rate of laser energy by different materials. Let the power adjustment function be:

[0055] where is the reference power, and are the reflectivity and thermal conductivity of the reference material, and are the adjustment coefficients. The scanning speed adjustment function needs to consider the heat capacity and density of the material to ensure the uniformity of energy distribution. Let the speed adjustment function be:

[0056] where is the reference scanning speed, is the heat capacity of the reference material, and are the speed adjustment coefficients. To optimize the laser power and scanning speed, a model predictive control optimization objective function is constructed. This objective function should simultaneously consider the scribing trajectory tracking accuracy, scribing uniformity, and system control stability. Let the optimization objective function be:

[0057] where is the scribing position at the current moment, is the target trajectory, is the control input (including laser power and scanning speed), is the weight matrix of the trajectory error, is the weight matrix of the control input, The weight matrix for the terminal state. To make the optimization process more compliant with physical constraints, a combined laser power adaptive adjustment function and scanning speed adaptive adjustment function are used to construct an optimization framework that includes a quadratic programming problem for power and speed. For the quadratic programming problem of power and speed, the interior point method is used for solution. The interior point method introduces a barrier function, converts the constrained optimization problem into an unconstrained problem, and iteratively searches for the optimal solution through the Newton method. Let the optimization variables be:

[0058] Then the optimization problem is formulated as:

[0059] where, represents the upper bound of the inequality constraint, is the constraint coefficient, is the barrier parameter. By introducing the barrier function, it is ensured that the optimization variables are always within the feasible region and gradually approach the optimal solution during the iteration process. The iterative update of the optimization solution uses the Newton method:

[0060] where, is the Hessian matrix, is the gradient of the objective function, is the step size parameter. Through continuous iterative optimization, an optimal control sequence including laser power and scanning speed parameters is obtained, thereby ensuring uniform energy distribution during the scribing process and improving the overall control performance of the system.

[0061] In one example, the laser detection method of the laser scribing machine further includes: Construct a dual-light-path collaborative control architecture including a path actuator, a feedback corrector, and a dynamic adjuster, decompose the optimal control sequence, and obtain a detection light-path control signal and a scribing light-path control signal; Calculate the leading distance based on the system processing delay, and drive the detection light path to run ahead of the scribing light path according to the leading distance through the detection light-path control signal, obtaining a dual-light-path operation mode that is spatially leading and temporally synchronized; Based on the dual-light-path operation mode and the real-time reflection signal collected by the detection light path on the surface of the solar cell, perform fast feature extraction on the real-time reflection signal, update the model parameters of the material property predictor through the recursive least squares method, and obtain the target property predictor; Compare the actual material properties collected by the detection light path with the predicted material properties of the target property predictor, and calculate the error vector characterizing the prediction accuracy; Dynamically adjust the weight matrix in the control Lyapunov function according to the error vector, and at the same time finely adjust the laser power and scanning speed of the scribing light path to obtain control parameters considering the prediction error. Clock synchronization is performed on the detection optical path and the scribing optical path according to the control parameters considering the prediction error. At the same time, a Kalman filter is introduced for system state estimation to achieve precise scribing on the surface of the solar cell.

[0062] In this example, the optimal control sequence is decomposed to separately generate the detection optical path control signal and the scribing optical path control signal, realizing the independent drive and coordinated control of the dual optical paths. Let the optimal control sequence be , where represents the optimal control input at the -th time step. Since the motion paths of the detection optical path and the scribing optical path are different, but their synchronization in time needs to be ensured, the control sequence is decoupled. Define the control signal of the detection optical path as , and the control signal of the scribing optical path as , then:

[0063] where and respectively represent the functions for path decomposition of the control sequence, ensuring that the detection optical path leads in space while the scribing optical path maintains an accurate trajectory following. To ensure that the detection optical path can obtain the material characteristics of the solar cell surface before the scribing optical path, the processing delay of the system is calculated, and the leading distance of the detection optical path is determined according to this delay. Assume that the total processing delay of the system is , then the leading distance between the detection optical path and the scribing optical path is determined by the following relationship:

[0064] where is the scanning speed of the scribing optical path. Through this calculation, the detection optical path operates with the required lead to ensure that the collected material characteristic data can be used to adjust the control strategy before the scribing optical path is executed. On this basis, based on the dual optical path operation mode and the real-time reflection signal collected by the detection optical path, fast feature extraction is performed on the collected data to obtain a real-time estimated value of the current material characteristics. Let the original reflection signal collected by the photodetector of the detection optical path be , then after preprocessing, the normalized reflectance signal is extracted:

[0065] where is the reference light intensity. The parameters of the material characteristic predictor are updated online by the recursive least squares method to adapt to the dynamic changes of material characteristics in different regions. Assume that the parameter vector of the material characteristic predictor is , then its update formula is:

[0066] Among them, is the gain matrix, is the actual material property observed in the detection optical path, is the current eigenvector. By continuously updating , the material property predictor can be dynamically adjusted to more accurately describe the change of material properties. After obtaining the target property predictor, the actual material properties collected by the detection optical path are compared with the output of the predictor to calculate the prediction error vector , which reflects the accuracy of the predictor. In order to improve the control accuracy of the system, the weight matrix in the control Lyapunov function is dynamically adjusted according to the error vector, and at the same time, the laser power and scanning speed of the scribing optical path are finely adjusted to compensate for the deviation caused by the prediction error. Let the control Lyapunov function be:

[0067] Among them, is a positive definite matrix, and its update is affected by the error vector , and the update rule is:

[0068] Among them, is the adjustment coefficient. The laser power and the scanning speed of the scribing optical path are adjusted according to the error vector:

[0069]

[0070] Among them, , , , are the prediction errors of reflectivity, thermal conductivity, heat capacity and density respectively, is the adjustment coefficient. After calculating the adjusted control parameters, it is necessary to ensure the clock synchronization of the detection optical path and the scribing optical path to ensure their coordinated operation. In order to improve the robustness of the system state estimation, a Kalman filter is introduced for state estimation. Let the system state be , and the observed value be , then the state update equation of the Kalman filter is:

[0071]

[0072] Among them, and are the process noise and measurement noise respectively, is the state transition matrix, is the control input matrix, is the observation matrix. Through the Kalman filter, the optimal estimate of the system state is obtained and used to adjust the control strategies of the detection optical path and the scribing optical path.

[0073] Referring to Figure 2 , this embodiment provides a laser detection device for a laser scribing machine, including: An acquisition module 1, which is used to perform laser scanning on the surface of a solar cell, acquire the reflected light intensity signal, and perform adaptive domain transformation processing on the reflected light intensity signal to obtain a target feature matrix; A training module 2, which is used to train a material property predictor with a hybrid architecture of a bidirectional long short-term memory network and a temporal convolutional network based on the target feature matrix, and predict the material property parameters on the scribing path to obtain a material property prediction value; A construction module 3, which is used to construct a state space model of a dual-optical-path laser system according to the material property prediction value, establish a control Lyapunov function by solving the Lyapunov equation, and obtain the constraint conditions for ensuring system stability; A solution module 4, which is used to substitute the constraint conditions and the material property prediction value into the model predictive control optimization problem, and solve the quadratic programming problem by the interior point method to obtain an optimal control sequence.

[0074] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the description in the above method embodiment, and details are not described herein again.

[0075] Referring to Figure 3 , this embodiment of the present invention also provides a computer device, which may be a server, and its internal structure may be as Figure 3 shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface, and a database connected through 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 through a network connection. The computer program, when executed by the processor, implements the above method.

[0076] Those skilled in the art can understand that Figure 3 the structure shown in

[0077] An embodiment of the present invention further 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.

[0078] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments 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 many 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), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0079] It should be noted that in this article, the terms "include", "comprise", or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, apparatus, article, or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, apparatus, article, or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, apparatus, article, or method including that element.

[0080] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A laser detection method for a laser scribing machine, characterized in that, Including: Laser-scan the surface of a solar cell, collect the reflected light intensity signal, and perform an adaptive domain transformation on the reflected light intensity signal to obtain a target feature matrix; Train a material property predictor with a hybrid architecture of bidirectional long short-term memory network and temporal convolutional network based on the target feature matrix, and predict the material property parameters on the scribing path to obtain material property prediction values; Construct a state-space model of the dual-path laser system according to the material property prediction values, establish a control Lyapunov function by solving the Lyapunov equation, and obtain the constraint conditions for ensuring system stability; Substitute the constraint conditions and the material property prediction values into the model predictive control optimization problem, and solve the quadratic programming problem by the interior point method to obtain the optimal control sequence.

2. The laser detection method of the laser scribing machine according to claim 1, wherein The laser-scanning the surface of the solar cell, collecting the reflected light intensity signal, and performing an adaptive domain transformation on the reflected light intensity signal to obtain a target feature matrix includes: Construct a dual-path laser system including a detection optical path and a scribing optical path, scan the surface of the solar cell using the detection optical path, and collect the reflected light intensity signal through a photodetector in the detection optical path; Perform analog-to-digital conversion on the reflected light intensity signal to obtain a digitized reflected light intensity signal, and perform filtering on the digitized reflected light intensity signal to obtain a filtered reflected light intensity signal; Normalize the filtered reflected light intensity signal by dividing it by the reflected light intensity of the reference substrate to obtain a normalized reflectivity signal; Perform time-domain analysis on the normalized reflectivity signal to obtain time-domain characteristic parameters including signal amplitude, rise time, fall time, and duration. At the same time, perform a fast Fourier transform on the normalized reflectivity signal to obtain frequency-domain characteristics, and combine the time-domain characteristic parameters and the frequency-domain characteristics into a signal feature matrix; Input the signal feature matrix into a deep neural network for adaptive domain transformation to eliminate signal differences under different material conditions and obtain a target feature matrix.

3. The laser detection method of the laser scribing machine according to claim 2, characterized in that, The inputting the signal feature matrix into a deep neural network for adaptive domain transformation to eliminate signal differences under different material conditions and obtain a target feature matrix includes: Input the signal feature matrix into the first fully connected layer of the deep neural network for linear transformation to obtain the first-layer output features; Input the first-layer output features into the second fully connected layer of the deep neural network for weight matrix and bias vector calculation to obtain the second-layer output features; Apply batch normalization to the second-layer output features to obtain normalized features; Input the normalized features into the third fully connected layer of the deep neural network, perform matrix multiplication and bias addition to obtain the third-layer output features; Input the third-layer output features into the fourth fully connected layer of the deep neural network and perform linear transformation to obtain a feature vector with a dimension of 128; Apply random masking and random zeroing to the feature vector with a dimension of 128 to prevent overfitting, and then map the feature vector with a dimension of 128 to a d-dimensional space through a linear projection matrix to obtain a target feature matrix.

4. The laser detection method of the laser scribing machine according to claim 1, characterized in that, Training a material property predictor with a hybrid architecture of bidirectional long short-term memory network and temporal convolutional network based on the target feature matrix, and predicting the material property parameters on the scribing path to obtain the material property prediction values, including: Constructing a material property predictor, the material property predictor including a bidirectional long short-term memory network and a temporal convolutional network, and organizing the target feature matrix into a training sample set according to the temporal relationship; Training the material property predictor based on the training sample set by applying a multi-task learning strategy to obtain a trained material property predictor; Inputting the historical target feature matrix sequence on the scribing path to be predicted into the trained material property predictor to obtain a dual-path feature representation; Performing a feature fusion operation on the dual-path feature representation to obtain a fused feature representation; Inputting the fused feature representation into a multi-task output module, and respectively calculating the material reflectivity, scattering coefficient, thermal conductivity, heat capacity, and material density to obtain the material property prediction values.

5. The laser detection method of the laser scribing machine according to claim 1, characterized in that, Constructing a state space model of the dual-path laser system according to the material property prediction values, establishing a control Lyapunov function by solving the Lyapunov equation, and obtaining the constraint conditions for ensuring the system stability, including: Constructing a state vector of the dual-path laser system, the state vector including the position, velocity, and acceleration information of the detection optical path and the scribing optical path, the control input vector including the drive signals of the two optical paths, and the system output vector being the actual scribing position, to obtain a basic variable set represented by the state space; Constructing a non-linear state transition function based on the basic variable set represented by the state space and the material property prediction values to obtain a non-linear system dynamic equation; Performing a first-order Taylor expansion linearization process on the non-linear system dynamic equation at the current operating point, calculating the Jacobian matrix, and obtaining a time-varying linear model; Selecting a positive definite weight matrix based on the time-varying linear model, and obtaining a positive definite matrix by solving the discrete-time Lyapunov equation to construct a quadratic-form based control Lyapunov function; Designing a Lyapunov stability constraint according to the control Lyapunov function, and transforming the Lyapunov stability constraint into an explicit constraint form regarding the control input to obtain a control constraint for ensuring the system stability.

6. The laser detection method of the laser scribing machine according to claim 1, characterized in that, Substituting the constraint conditions and the material property prediction values into a model predictive control optimization problem, and solving the quadratic programming problem by the interior point method to obtain an optimal control sequence, including: Dividing the surface of the solar cell into m×n grid cells, assigning corresponding material property vectors to each grid cell, and determining a target path set according to the battery scribing design requirements to obtain a grid-based battery model for path planning; Establishing an energy density model based on the grid-based battery model and the material property prediction values, and designing a laser power adaptive adjustment function and a scanning speed adaptive adjustment function according to the material property prediction values and the energy density model; Construct an optimization objective function for model predictive control, and construct a quadratic programming problem for the power and speed of the laser power adaptive adjustment function and the scanning speed adaptive adjustment function by combining the optimization objective function and the constraint conditions; Apply the interior point method to solve the quadratic programming problem for the power and speed. Use the barrier function to transform the constraints into an unconstrained problem, and find the optimal solution through Newton's method iteration to obtain an optimal control sequence including laser power and scanning speed parameters.

7. The laser detection method of the laser scribing machine according to claim 1, characterized in that The laser detection method of the laser scribing machine further includes: Construct a dual-light-path collaborative control architecture including a path actuator, a feedback corrector, and a dynamic adjuster, decompose the optimal control sequence to obtain a detection light-path control signal and a scribing light-path control signal; Calculate the leading distance based on the system processing delay, and drive the detection light path to run ahead of the scribing light path through the detection light-path control signal according to the leading distance to obtain a dual-light-path operation mode that is spatially leading and temporally synchronized; Based on the dual-light-path operation mode and the real-time reflection signal collected by the detection light path on the surface of the solar cell, perform fast feature extraction on the real-time reflection signal, and update the model parameters of the material property predictor through the recursive least squares method to obtain a target property predictor; Compare the actual material properties collected by the detection light path with the predicted material properties of the target property predictor, and calculate an error vector representing the prediction accuracy; Dynamically adjust the weight matrix in the control Lyapunov function according to the error vector, and at the same time finely adjust the laser power and scanning speed of the scribing light path to obtain control parameters considering the prediction error; Synchronize the clocks of the detection light path and the scribing light path according to the control parameters considering the prediction error, and at the same time introduce a Kalman filter for system state estimation to achieve precise scribing on the surface of the solar cell.

8. A laser detection device for a laser scribing machine, characterized in that, For implementing the steps of the method according to any one of claims 1 to 7, the laser detection device of the laser scribing machine includes: An acquisition module for laser scanning the surface of the solar cell, acquiring the reflected light intensity signal, and performing adaptive domain transformation processing on the reflected light intensity signal to obtain a target feature matrix; A training module for training a material property predictor with a hybrid architecture of a bidirectional long short-term memory network and a temporal convolutional network based on the target feature matrix, and predicting the material property parameters on the scribing path to obtain a material property prediction value; A construction module for constructing a state space model of the dual-light-path laser system according to the material property prediction value, and establishing a control Lyapunov function by solving the Lyapunov equation to obtain constraint conditions for ensuring system stability; A solution module for substituting the constraint conditions and the material property prediction value into the model predictive control optimization problem, and solving the quadratic programming problem through the interior point method to obtain an optimal control sequence.

9. A computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 7.

Images