Light guide plate processing parameter control method based on deep learning and reverse optimization

Abstract

The invention provides a light guide plate processing parameter control method based on deep learning and reverse optimization, and relates to the technical field of liquid crystal display panel manufacturing, and the method comprises the steps: obtaining historical data, and carrying out the time sequence preprocessing of the historical data; the historical data comprises a machining parameter serving as a control variable and a machining depth serving as a target variable; constructing a depth prediction model, and performing optimization training of model internal parameters on the depth prediction model based on historical data; and for the depth prediction model subjected to optimization training, a reverse optimization solution mechanism is adopted to carry out reverse optimization solution of processing parameters. According to the method, the optimal processing parameters are solved in a mode of combining deep learning and reverse optimization, so that the depth of the collision point stably reaches a target value.

Description

Technical Field

[0001] This invention relates to the field of liquid crystal display panel manufacturing technology, and more specifically, to a method for controlling light guide plate processing parameters based on deep learning and reverse optimization. Background Technology

[0002] The light guide plate is a core component of the backlight unit in a liquid crystal display module. Its function is to convert side-mounted point or line light sources (such as LEDs) into uniform planar light sources, providing stable and uniform backlight for the LCD screen. To achieve this function, tens of thousands of microscopic optical dots are typically formed on the surface of the light guide plate. The size, density, and depth distribution of these dots directly determine the uniformity and brightness of the backlight, making them key quality control factors in light guide plate manufacturing.

[0003] The impact-dot process is one of the mainstream and efficient methods for creating halftone dots in the industry. It involves using a high-frequency moving impact pin to strike the surface of a light guide plate, forming indentations with specific depths and shapes. The impact depth is a key process parameter that directly affects the light scattering ability of the halftone dots. However, in mass production, its long-term stability and consistency are difficult to guarantee. This is because the impact pin, as a direct contact consumable, inevitably wears down and becomes blunt after millions of high-frequency impacts. Therefore, even if processing parameters (such as the programmed dot diameter and impact pressure) remain constant, the actual impact depth will gradually become shallower over time, leading to deviations in the product's optical performance from the design goals and affecting yield.

[0004] To address these issues, existing technologies primarily employ two approaches: First, relying on experience for periodic manual intervention, adjusting processing parameters based on batch sampling results; this method is highly subjective, has a delayed response, and struggles to guarantee optimality. Second, introducing automated devices such as online visual inspection, providing feedback and adjustments only after defective products are detected; however, this reactive compensation logic fails to prevent defects from the outset. While there have been attempts to apply digital twins or artificial intelligence to equipment monitoring or product design, a generally lacking intelligent process control methods capable of accurately predicting the impact of impactor wear on machining depth and proactively and automatically solving for optimal compensation parameters during processing is still in development. Summary of the Invention

[0005] This invention provides a method for controlling the processing parameters of a light guide plate based on deep learning and reverse optimization. It solves the problem in existing technologies that the inability to accurately predict the impact of impactor wear on the processing depth makes it impossible to solve for the optimal processing parameters and thus ensure that the processing depth stably reaches the target value.

[0006] In a first aspect, embodiments of the present invention provide a method for controlling light guide plate processing parameters based on deep learning and reverse optimization, the method comprising the following steps:

[0007] Historical data is acquired and time-series preprocessed; the historical data includes processing parameters as control variables and processing depth as the result.

[0008] Build a deep prediction model and optimize the internal parameters of the deep prediction model based on historical data;

[0009] For the deep prediction model with optimized training, an inverse optimization solution mechanism is used to optimize the processing parameters.

[0010] In the above embodiments, the present invention solves for the optimal processing parameters by combining deep learning and inverse optimization, so that the impact point depth stably reaches the target value.

[0011] As some optional embodiments of this application, the processing parameters include the processing point diameter of the firing pin and the impact height.

[0012] In the above embodiments, the present invention uses the machining point diameter and impact height as machining parameters as control variables, which can accurately reflect the impact pin situation.

[0013] As some optional embodiments of this application, the depth prediction model includes an encoder and a decoder.

[0014] In the above embodiments, the present invention uses an encoder and a decoder to construct a depth prediction model, which can accurately reflect the nonlinear relationship between the wear state of the firing pin and the processing parameters and processing depth.

[0015] As some optional implementations of this application, the process of acquiring historical data and performing time-series preprocessing on the historical data is as follows:

[0016] Obtain historical data on the entire lifecycle production or long-term continuous production of multiple firing pins under a single specification;

[0017] Historical data is cleaned and then divided into time series to obtain supervised samples.

[0018] The continuous features of the supervised samples and the target variable are normalized to obtain the training sample set.

[0019] In the above embodiments, the present invention, through time-series preprocessing, can prepare a high-quality, formatted training sample set that avoids information leakage for subsequent model training.

[0020] As some optional embodiments of this application, the working principle of the depth prediction model is as follows:

[0021] Historical data is read by the encoder, compressed into a fixed-dimensional state vector, and output to the decoder; the state vector is a numerical representation of the current wear state of the firing pin.

[0022] The decoder receives the state vector and the current processing parameters, and predicts the corresponding processing depth based on the state vector and the current processing parameters.

[0023] In the above embodiments, the present invention reflects the nonlinear relationship between the wear state of the firing pin and the processing parameters and processing depth through a depth prediction model. The model not only considers the current processing parameters, but also makes full use of historical information to dynamically compensate for the impact of the firing pin wear on the processing results, thereby achieving higher accuracy prediction.

[0024] As one of the optional implementation methods of this application, the process of optimizing the internal parameters of the deep prediction model based on historical data is as follows:

[0025] The historical data of the training sample set and the current processing parameters are input into the depth prediction model in batches, and the corresponding processing depth is predicted by the depth prediction model.

[0026] The error between the predicted processing depth and the actual processing depth is quantified by a predefined loss function. A gradient optimizer is used to calculate the gradient of the loss function with respect to all learnable parameters within the model. The learnable parameters are then adjusted along the direction of gradient descent. Multiple iterations are performed on the entire training sample set until the value of the loss function converges.

[0027] In the above embodiments, the present invention trains the constructed deep prediction model with a large amount of historical data so that its internal learnable parameters (such as the weights of the neural network) can accurately capture the patterns in the data.

[0028] As some optional implementations of this application, for the deep prediction model with optimized training, the process of inverse optimization of processing parameters using an inverse optimization solution mechanism is as follows:

[0029] For deep prediction models with optimized training, all learnable parameters within the model are set to be non-trainable, and the processing parameters that need to be solved are defined as variable input parameters.

[0030] Set the expected target processing depth, define the error between the predicted depth and the target processing depth under the current variable input parameters, and use a loss function to quantify the error;

[0031] Initialize a gradient optimizer and iteratively optimize the variable input parameters using the gradient optimizer;

[0032] The gradient optimizer is used to iteratively update the variable input parameters so that the predicted depth in the next iteration is close to the target processing depth. The iteration ends when the error between the predicted depth and the target processing depth is less than a preset threshold. At this time, the corresponding variable input parameters are the optimal solution that can produce the target processing depth.

[0033] The optimal solution for the target machining depth is inversely normalized to obtain the machining parameters set by the production equipment.

[0034] In the above embodiments, the present invention utilizes the differentiability of deep learning models to solve for the optimal value by directly applying gradient descent to the model input parameters. Compared to traditional black-box optimization algorithms (such as particle swarm optimization and genetic algorithms) that require a large number of iterative evaluations, the present invention uses gradient information for efficient and deterministic optimization. While ensuring solution accuracy, it can significantly reduce solution time and meet the high timeliness requirements of industrial production.

[0035] In a second aspect, the present invention provides a light guide plate processing parameter control system based on deep learning and reverse optimization, the system comprising:

[0036] A data acquisition unit is used to acquire historical data and perform time-series preprocessing on the historical data; the historical data includes processing parameters as control variables and processing depth as a result.

[0037] A model training unit is used to construct a deep prediction model and optimize the internal parameters of the deep prediction model based on historical data.

[0038] The inverse optimization unit is used to perform inverse optimization of the processing parameters of the deep prediction model that has been optimized and trained.

[0039] In a third aspect, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method for controlling light guide plate processing parameters based on deep learning and reverse optimization.

[0040] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for controlling light guide plate processing parameters based on deep learning and reverse optimization.

[0041] The beneficial effects of this invention are as follows:

[0042] (1) This invention proposes a feedforward process parameter control method for the impact point depth of a light guide plate, constructing a feedforward control link of "state assessment - parameter solution - command execution". Its core lies in predicting and compensating for process deviations caused by impact pin wear before processing. This invention differs from the existing technology that relies on the detection of defective products and then making adjustments, shifting the quality control link from post-event correction to pre-event prevention, thus ensuring the consistency of product quality from the source.

[0043] (2) This invention uses an implicit state characterization method to model the time-varying characteristics of equipment. For the key variable of striker wear, which cannot be directly measured, a method for implicit state characterization through time series model is designed. This method does not rely on any physical wear model, but learns autonomously from historical data and generates a state vector that can accurately reflect the cumulative effect of wear, providing an effective modeling approach for dealing with time-varying variables that are difficult to measure directly.

[0044] (3) This invention applies a gradient-based inverse optimization algorithm to achieve rapid parameter solving. Utilizing the differentiability of deep learning models, the optimal value is solved by directly applying gradient descent to the model input parameters. Compared to traditional black-box optimization algorithms (such as particle swarm optimization and genetic algorithms) that require extensive iterative evaluation, this method uses gradient information for efficient and deterministic optimization. While ensuring solution accuracy, it can significantly reduce solution time, meeting the high timeliness requirements of industrial production.

[0045] (4) This invention automates the setting of key process parameters and objectifies the basis for decision-making. By constructing an end-to-end automated solution system, the traditional parameter tuning process, which relies on repeated trial and error based on the personal experience of engineers, is transformed into a data- and algorithm-driven, goal-oriented autonomous optimization process. The operator only needs to input the desired process result, and the system can autonomously output the optimal equipment instructions. This not only automates this process but also ensures that every decision is based on historical data and model reasoning, guaranteeing the objectivity and consistency of parameter settings. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of the light guide plate processing parameter control method according to an embodiment of the present invention.

[0048] Figure 2 This is a schematic diagram of the structure of the depth prediction model described in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of model training according to an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the reverse optimization solution as described in an embodiment of the present invention;

[0051] Figure 5 This is a graph showing the decrease in the loss function during the training process as described in an embodiment of the present invention;

[0052] Figure 6 This is a graph showing the reverse optimization solution results according to an embodiment of the present invention. Detailed Implementation

[0053] To better understand the above technical solutions, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solutions of the present invention, rather than limitations thereof. In the absence of conflict, the technical features in the embodiments and embodiments can be combined with each other.

[0054] It should also be understood that, in order to simplify the description of the invention and thus aid in the understanding of at least one embodiment, multiple features may sometimes be grouped into a single embodiment, drawing, or description thereof in the foregoing description of the embodiments of the invention. However, this method of disclosure does not imply that the subject matter of the invention requires more features than those mentioned in the claims. In fact, the embodiments contain fewer features than all the features of the single embodiment disclosed above.

[0055] To improve product quality consistency, increase yield, and reduce reliance on experience by considering time-varying factors such as consumable wear in real time and proactively optimizing processing parameters for the impact point depth of light guide plates, this invention proposes a light guide plate processing parameter control method based on deep learning and inverse optimization. By collecting historical time-series processing data, a time-series model capable of accurately predicting depth is trained to implicitly represent the dynamic wear state of the impact pin. Based on this, an inverse optimization solution mechanism is proposed, using the trained depth prediction model as a virtual physical system and employing an efficient gradient descent algorithm to quickly and automatically find the optimal processing parameters required to achieve the target depth. This method shifts the control mode from post-compensation to pre-prediction and optimization, achieving proactive, automated, and high-precision control of the impact point depth, effectively solving existing technical challenges.

[0056] Example 1

[0057] This invention provides a method for controlling light guide plate processing parameters based on deep learning and inverse optimization. Please refer to [link / reference]. Figure 1 , Figure 1 The flowchart of the method for controlling the processing parameters of the light guide plate is as follows:

[0058] (1) Obtain historical data and perform time-series preprocessing on the historical data.

[0059] In this embodiment of the invention, by collecting historical data and performing time-series preprocessing, a high-quality, formatted dataset that avoids information leakage can be provided for subsequent model training.

[0060] Specifically, the process for collecting historical data and preprocessing time series data is as follows:

[0061] (1.1) Data Collection: The study focuses on a single-specification firing pin, collecting historical data on the entire lifecycle production or long-term continuous production of multiple firing pins of that specification. This historical data includes, but is not limited to, processing parameters as control variables, and processing depth as the target variable. The processing parameters include the diameter of the firing pin's processing point and the impact height. A simplified data example of historical data for a single firing pin is shown in Table 1.

[0062] Machining point diameter (µm) Impact height (mm) Other parameters (...) Processing depth (µm) 52 1.20 … 3.98 52 1.20 … 3.97 … … … … 53 1.21 … 4.01 53 1.21 … 4.00

[0063] Table 1 is a simplified data example of historical data for a firing pin.

[0064] (1.2) Data cleaning: The original data is subjected to quality review and consistency processing, including outlier identification and removal, missing value handling, duplicate timestamp merging, and unification of units and dimensions, and rearranged according to the internal time order of the collider; specifically, the embodiments of the present invention do not limit the specific data cleaning methods.

[0065] (1.3) Constructing supervised samples: Set the window length L and step size s. The window length L is adjusted according to specific business needs and model performance, and the step size s is determined according to the amount of data. Generate samples independently for the time series of each firing pin according to the sliding window, and prohibit splicing across firing pins; for any time index t (satisfying t≥L), take the continuous historical data in the interval [t-L+1, t] as the historical sequence input, take the processing parameters at time t+1 as the current input, and take the processing depth at time t+1 as the supervision label; thus forming a sample triple (historical sequence, current processing parameters, current processing depth) for training the depth prediction model.

[0066] (1.4) Dataset partitioning: Using the spark pin number as the smallest partitioning unit, the spark pin set is randomly divided into a training sample set and a validation sample set (e.g., 80% and 20%). All samples of the same spark pin must belong to the same data subset. The same spark pin is prohibited from appearing in different sets to avoid data leakage. Furthermore, it is ensured that the validation sample set and the validation sample set each contain multiple spark pins and cover different production batches to improve the robustness and representativeness of the evaluation.

[0067] (1.5) Normalization: Using the training sample set as a benchmark, calculate the mean μ and standard deviation σ according to the feature dimension, and perform Z-Score normalization on all continuous features and the target variable:

[0068] z=(x-μ) / σ

[0069] Where x represents the training sample before normalization, and z represents the training sample after normalization. Specifically, the validation sample set is transformed using the statistics of the training sample set, and all spurs share the same set of training sample set statistics; mean μ and standard deviation σ are also saved so that denormalization can be performed during the inference phase.

[0070] (2) Construct a deep prediction model.

[0071] In this embodiment of the invention, this step aims to construct a deep learning model that can accurately reflect the nonlinear relationship between the wear state of the firing pin and the machining parameters and machining depth. This model employs an encoder-decoder architecture; please refer to [link to relevant documentation]. Figure 2 , Figure 2 This is a schematic diagram of the structure of the depth prediction model.

[0072] Specifically, the working principle of the depth prediction model is as follows:

[0073] (2.1) The encoder is responsible for processing the generated historical data. The encoder can employ models such as Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), or Transformer. That is, the encoder reads the historical data (historical sequence) and compresses its information into a fixed-dimensional state vector, which can be considered a numerical representation of the current wear state of the striker. The process can be summarized as follows:

[0074] C = Encoder(X1,X2,...,X) N )

[0075] Where X1, X2, ..., X N C represents the historical data, and C represents the state vector output by the encoder.

[0076] (2.2) The decoder part is a feedforward network, such as a multilayer perceptron (MLP). The decoder receives the state vector C output by the encoder and the processing parameters (e.g., processing point diameter) P set by the user for the next processing. N+1 And based on this, the corresponding processing depth can be predicted. The process can be summarized as follows:

[0077]

[0078] With this structure, the model not only considers the current machining parameters, but also makes full use of historical information to dynamically compensate for the impact of impact pin wear on the machining results, thereby achieving higher accuracy prediction.

[0079] (3) Optimize the internal parameters of the deep prediction model based on historical data.

[0080] In this embodiment of the invention, this step aims to train the constructed prediction model with a large amount of historical data so that its internal learnable parameters (such as the weights of the neural network) can accurately capture the patterns in the data.

[0081] Specifically, the process for optimizing the model's internal parameters during training is as follows:

[0082] (3.1) The optimized training process is a standard supervised learning paradigm. The training sample set is input into the depth prediction model in batches. The model outputs the predicted processing depth based on the historical sequence and the current processing parameters. Subsequently, the predicted depth is quantized using a predefined loss function. The error between the actual processing depth Y of the sample and the loss function Loss. Loss can be either mean absolute error or mean squared error, and its process can be summarized as follows:

[0083]

[0084] (3.2) Use a gradient-based optimization algorithm (e.g., the Adam optimizer) to calculate the gradient of the loss function with respect to all learnable parameters within the model, and make small adjustments to these parameters along the direction of gradient descent. See [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram illustrating the model training process. The above process is repeated multiple times across the entire training sample set until the loss function converges to a sufficiently small level and the model exhibits good generalization ability on independent validation sample sets. After training is complete, the model's internal parameters are considered optimized and fixed, forming a reliable virtual physics model for subsequent use.

[0085] (4) For the deep prediction model with optimized training, the reverse optimization solution mechanism is used to optimize the processing parameters.

[0086] In this embodiment of the invention, this step is the core of the invention, aiming to use the trained model to reverse-engineer the optimal processing parameters for a given target processing depth. Please refer to [link to relevant documentation]. Figure 4 , Figure 4 This is a schematic diagram of the reverse optimization solution.

[0087] Specifically, the reverse optimization solution process is as follows:

[0088] (4.1) For the deep prediction model with optimized training, all learnable parameters inside the model are set to be untrainable, and the processing parameters that need to be solved are defined as variable input tensors that need to be optimized.

[0089] (4.2) Set a target processing depth Y to be achieved target Then, a new optimization objective is defined: the model predicts depth under the current variable input parameters. With the target machining depth Y target The error between them is also quantified using loss functions such as mean squared error:

[0090]

[0091] (4.3) Initialize a gradient optimizer to iteratively optimize the variable input parameters. In each iteration, the gradient optimizer calculates the loss function. opt The gradient of the input parameter is used to update the parameter's value, thereby increasing the prediction depth for the next iteration. Closer to target depth Y target Specifically, let the input processing parameter to be solved be P. In the k-th iteration, the gradient of the loss function with respect to the input processing parameter P is calculated using the backpropagation algorithm:

[0092]

[0093] (4.4) Use this gradient to update the input processing parameters P. Taking stochastic gradient descent (SGD) as an example, the update formula is as follows: (Where η is the learning rate); if the Adam optimizer is used, the first and second moments of the gradient are further combined for adaptive updates. To ensure that the solution parameters are within the physical range that the device can execute, the parameters can be clipped to a preset effective range (e.g., [P]) after each update. min ,P max Within ]) . This process is repeated iteratively to increase the prediction depth for the next iteration. Closer to the target processing depth Y target .

[0094] (4.5) This iterative process continues until the loss function Loss opt The input parameter values ​​are either less than a preset threshold or the maximum number of iterations is reached. After the iteration ends, the obtained input parameter values ​​are the optimal solution that can produce the target processing depth. Finally, the standardized parameter values ​​obtained are denormalized to restore their true physical scale, and can then be used as the setpoints for production equipment in actual use.

[0095] Furthermore, as production continues and data gradually accumulates, to improve the accuracy of prediction and reverse engineering and to address potential data distribution or concept drift, steps (1) to (4) can be re-executed according to a preset cycle or triggering conditions (such as performance index decline, process change, material batch switching, equipment restart after maintenance, etc.). This includes cleaning and aligning new data, updating and normalizing training set statistics, retraining and calibrating the model, and redeploying. The aforementioned closed-loop iteration can improve the stability and robustness of the solution in long-term operation.

[0096] In the embodiments of the present invention, in order to more clearly demonstrate the purpose, technical solution and advantages of the present invention, the present invention will be described in detail below with specific data, in conjunction with the accompanying drawings and specific embodiments.

[0097] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0098] S10: Collection and time-series preprocessing of historical processing data.

[0099] ①) Data collection: Historical data of 11 firing pins of the same specification were collected during the production process. Each data point includes the processing parameters (processing point diameter) and the corresponding processing depth (average depth). Data from 4 other firing pins were used as an independent verification sample set.

[0100] ② Data cleaning: Check the collected data and remove records with null average depth.

[0101] ③ Constructing supervised samples: Set the historical window length L = 10, and perform sliding window sampling on the data sequence of each firing pin to construct samples. Each sample contains a sequence of length 11, where the data of the first 10 time steps (including the processing point diameter and average depth columns) are used as the historical input of the encoder, the processing point diameter of the 11th time step is used as the current input of the decoder, and the average depth of the 11th time step is used as the prediction target of the model (i.e., the supervision label).

[0102] ④ Dataset partitioning: The dataset was strictly partitioned according to the firing pin ID. All data from 11 firing pins were used for training, and all data from 4 firing pins were used for testing, ensuring no information leakage between the training sample set and the validation sample set.

[0103] ⑤ Normalization: Calculate the mean and standard deviation of the processing point diameter and average depth using only the training sample set data, and perform Z-Score normalization on all data in both the training and validation sample sets based on this. Save the mean and standard deviation for subsequent inverse optimization results.

[0104] S20: Construction of deep prediction models.

[0105] ① Encoder: A single-layer Long Short-Term Memory (LSTM) network is used. Its hidden layer dimension is set to 64. The encoder receives a normalized historical sequence numerical array containing 10 time steps and 2 dimensions as input. After processing the sequence, the encoder outputs a 64-dimensional state vector, which implicitly represents the current wear state of the firing pin.

[0106] ② Decoder: A multilayer perceptron (MLP) with two residual blocks is used. The decoder first maps a normalized machining point diameter value (i.e., the current machining setpoint) to a 64-dimensional vector, and then adds it to the state vector output by the encoder, fusing historical wear information with the current control command. The fused vector is processed by multiple residual network blocks and finally passes through a linear output layer to predict the normalized average depth.

[0107] S30: Optimization training of model internal parameters.

[0108] Specifically, the constructed model is trained using a Python-based deep learning framework.

[0109] ① Training configuration: The loss function is the mean absolute error. The optimizer is AdamW, and the initial learning rate is set to 1×10⁻⁶. -4 The weight decay coefficient is 1×10 -3 .

[0110] ② Training process: The maximum number of training epochs is set to 300, and the batch size is 64. During training, the validation set loss is monitored, and an early stopping mechanism is enabled: if the validation set loss does not improve significantly within 50 epochs, training is terminated early to prevent overfitting.

[0111] ③ Training Results: After training, the best-performing model was fixed and saved. This model was evaluated on an independent validation set, showing a mean absolute error (MAE) of 0.069 and a mean absolute percentage error (MAPE) of 1.75% (both at the original data scale). This result demonstrates that the model possesses high-precision predictive ability. Please refer to [link / reference]. Figure 5 , Figure 5 The graph shows the decrease in the loss function during training, illustrating the stable convergence process of the model.

[0112] S40: Inverse optimization solution of machining input parameters.

[0113] Based on the trained model, inverse optimization is performed to solve for the optimal processing point diameter at a specific target depth.

[0114] ① Optimization settings: Assume that the target average depth to be achieved in a certain production run is 3.95µm. The operator provides this target value, as well as the actual processing history data (processing point diameter and average depth) for the last 10 runs.

[0115] ② Reverse Solution: Load the trained depth prediction model and freeze all its internal parameters (weights) to prevent further training. Normalize the target depth value of 3.95µm and the historical data sequence. Initialize the processing point path to be solved as an optimizable numerical variable (e.g., initialize to 0). Define the optimization loss function as the mean squared error (MSE) between the model's predicted depth and the normalized target depth. Initialize an Adam optimizer, whose optimization object is only the processing point path variable to be solved, with a learning rate set to 1×10⁻⁶. -3 Perform iterative optimization. In each iteration, input historical data and the current processing point path into the model to obtain the predicted depth, calculate the MSE loss between the predicted depth and the target depth, and update the processing point path value based on the gradient of the loss. The iteration continues until the loss is less than a preset threshold of 1×10⁻⁶. -6 Or it may reach the maximum number of iterations (e.g., 10,000 times).

[0116] ③ Output Results: After the optimization process converges, the optimal machining point diameter under the normalized scale is obtained. Inverse normalization yields machining parameter settings with real physical meaning. For example, in this embodiment, to achieve a target depth of 3.95µm, the optimal machining point diameter solved by the system is 43.6µm. This value can be directly used to set the production equipment to compensate for impact pin wear and ensure accurate machining depth. To further verify the optimization results, a series of target depth values ​​(e.g., from 3.5µm to 4.5µm) can be solved in batches, and the relationship curve between the solved machining point diameter and the target average depth can be plotted. Please refer to [link to relevant documentation]. Figure 6 , Figure 6 The graph shows the results of the reverse optimization solution. It visually illustrates the machining point diameter parameters required to achieve different machining depths under the current firing pin condition.

[0117] Example 2

[0118] This invention provides a light guide plate processing parameter control system based on deep learning and reverse optimization. The system corresponds one-to-one with the method described in Embodiment 1, and includes:

[0119] A data acquisition unit is used to acquire historical data and perform time-series preprocessing on the historical data; the historical data includes processing parameters as control variables and processing depth as a result.

[0120] A model training unit is used to construct a deep prediction model and optimize the internal parameters of the deep prediction model based on historical data.

[0121] The inverse optimization unit is used to perform inverse optimization of the processing parameters of the deep prediction model that has been optimized and trained.

[0122] Example 3

[0123] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the light guide plate processing parameter control method based on deep learning and reverse optimization described in Embodiment 1.

[0124] The computer device provided in this embodiment can implement the method described in Embodiment 1. To avoid repetition, it will not be described again here.

[0125] Example 4

[0126] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the light guide plate processing parameter control method based on deep learning and reverse optimization described in Embodiment 1.

[0127] The computer-readable storage medium provided in this embodiment can implement the method described in Embodiment 1. To avoid repetition, it will not be described again here.

[0128] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0129] The memory can be used to store the computer program and / or modules. The processor, by running or executing the data stored in the memory, realizes various functions of the light guide plate processing parameter control system based on deep learning and reverse optimization in the invention. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart memory card, secure digital card, flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0130] If a light guide plate processing parameter control system based on deep learning and reverse optimization is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can also implement all or part of the processes in the methods of the above embodiments through a computer program that can be stored in a computer-readable storage medium. When executed by a processor, this computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, a dot carrier signal, a telecommunications signal, and a software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0131] The basic concepts of this invention have been described. Obviously, for those skilled in the art, the above detailed disclosure is merely illustrative and does not constitute a limitation of this specification. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this specification. Such modifications, improvements, and corrections are suggested in this specification and therefore remain within the spirit and scope of the exemplary embodiments described herein.

Claims

1. A method for controlling light guide plate processing parameters based on deep learning and inverse optimization, characterized in that, The method includes the following steps: Historical data is acquired and time-series preprocessed; the historical data includes processing parameters as control variables and processing depth as target variables. Build a deep prediction model and optimize the internal parameters of the deep prediction model based on historical data; For the deep prediction model with optimized training, an inverse optimization solution mechanism is used to optimize the processing parameters.

2. The method for controlling light guide plate processing parameters based on deep learning and inverse optimization according to claim 1, characterized in that, The processing parameters include the diameter of the processing point of the impact pin and the impact height.

3. The method for controlling light guide plate processing parameters based on deep learning and inverse optimization according to claim 1, characterized in that, The depth prediction model includes an encoder and a decoder.

4. The method for controlling light guide plate processing parameters based on deep learning and inverse optimization according to claim 1, characterized in that, The process of acquiring historical data and performing time-series preprocessing on it is as follows: Collect the full lifecycle production data or long-term continuous production data of multiple firing pins under a single specification to obtain historical data; Historical data is cleaned, and the cleaned historical data is then divided into time series to obtain supervised samples. The continuous features of the supervised samples and the target variable are normalized to obtain the training sample set.

5. The method for controlling light guide plate processing parameters based on deep learning and inverse optimization according to claim 1, characterized in that, The working principle of the depth prediction model is as follows: Historical data is read by the encoder, compressed into a fixed-dimensional state vector, and output to the decoder; the state vector is a numerical representation of the current wear state of the firing pin. The decoder receives the state vector and the current processing parameters, and predicts the corresponding processing depth based on the state vector and the current processing parameters.

6. The method for controlling light guide plate processing parameters based on deep learning and inverse optimization according to claim 4, characterized in that, The process of optimizing the internal parameters of a deep prediction model based on historical data is as follows: The historical data of the training sample set and the current processing parameters are input into the depth prediction model in batches, and the corresponding processing depth is predicted by the depth prediction model. The error between the predicted processing depth and the actual processing depth is quantified by a predefined loss function. A gradient optimizer is used to calculate the gradient of the loss function with respect to all learnable parameters within the model. The learnable parameters are then adjusted along the direction of gradient descent. Multiple iterations are performed on the entire training sample set until the value of the loss function converges.

7. The method for controlling light guide plate processing parameters based on deep learning and inverse optimization according to claim 1, characterized in that, For the optimized deep prediction model, the process of inverse optimization of the processing parameters using the inverse optimization solution mechanism is as follows: For deep prediction models with optimized training, all learnable parameters within the model are set to be non-trainable, and the processing parameters that need to be solved are defined as variable input parameters. Set the expected target processing depth, define the error between the predicted depth and the target processing depth under the current variable input parameters, and use a loss function to quantify the error; Initialize a gradient optimizer and iteratively optimize the variable input parameters using the gradient optimizer; The gradient optimizer is used to iteratively update the variable input parameters so that the predicted depth in the next iteration is close to the target processing depth. The iteration ends when the error between the predicted depth and the target processing depth is less than a preset threshold. At this time, the corresponding variable input parameters are the optimal solution that can produce the target processing depth. The optimal solution for the target machining depth is inversely normalized to obtain the machining parameters set by the production equipment.

8. A light guide plate processing parameter control system based on deep learning and inverse optimization, characterized in that, The system includes: A data acquisition unit is used to acquire historical data and perform time-series preprocessing on the historical data; the historical data includes processing parameters as control variables and processing depth as a result. A model training unit is used to construct a deep prediction model and optimize the internal parameters of the deep prediction model based on historical data. The inverse optimization unit is used to perform inverse optimization of the processing parameters of the deep prediction model that has been optimized and trained.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the light guide plate processing parameter control method based on deep learning and reverse optimization as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the light guide plate processing parameter control method based on deep learning and reverse optimization as described in any one of claims 1-7.

Images