Compensation method for circuit forming process of high-density printed circuit board

Abstract

The invention discloses a high-density printed circuit board circuit forming process compensation method, which comprises the following steps of: firstly, collecting and standardizing structures and detection data of multiple batches and multiple process sections, and constructing a multi-target compensation optimization model through multi-dimensional parameter integration and feature extraction; the multi-objective compensation optimization model fuses objectives such as size precision, cost, electrical and thermal behaviors, establishes inter-objective collaborative constraints, introduces a reinforcement learning algorithm, dynamically adjusts weights of all the objectives, continuously optimizes compensation amount distribution in combination with process simulation feedback, finally fuses the optimal compensation weight into actual process control and collects effect feedback, and finally achieves the purpose of optimizing the optimal compensation weight. According to the method, high-precision and multi-target adaptive optimization of process compensation is realized, and the forming quality and the production efficiency of the circuit board are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-density printed circuit board manufacturing process optimization, and particularly relates to a high-density printed circuit board line forming process compensation method. BACKGROUND

[0002] High-density printed circuit board (HDI PCB) has become an important foundation of modern electronic manufacturing industry due to its fine lines, multiple layers, and high manufacturing precision. With the growing demand for high-function circuit boards in emerging fields such as 5G communication, Internet of Things, and artificial intelligence, the level of intelligence and refinement of HDI PCB production process is steadily improving. The current line forming process compensation technology is widely used to improve the yield of PCB finished products, and it comprehensively balances multiple target indicators such as dimensional accuracy, line width and spacing control, electrical performance, and thermal reliability, which is a key link to achieve high-quality production of high-end circuit boards.

[0003] However, the compensation optimization technology of the existing PCB line forming process still has the following limitations: (1) Multi-objective conflict trade-off limitation. In traditional methods, different target weights are usually set statically by humans, which is difficult to fully respond to the dynamic changes in target importance caused by batch-to-batch, device-to-device, material-to-material, structural level, or special process changes. When there is a conflict between dimensional accuracy, process feasibility, cost, electrical performance, and other targets, the weight adjustment and coordination mechanism lags behind, and the compensation strategy cannot obtain a systematic and globally optimal trade-off solution; (2) Limited local optimal and generalization ability. The current multi-objective optimization scheme is limited to rule base or non-iterative optimizer, and can only obtain the optimal solution under local constraints. The generalization ability for batch-to-batch, unconventional processes, or newly introduced devices is not strong, leading to slow response when process abnormalities occur; (3) Lack of adaptive closed loop and continuous evolution mechanism. The current compensation amount generation process relies on human experience or static models, and it is difficult to dynamically adjust model parameters according to new production batches and process execution feedback in actual application. The closed-loop feedback and self-learning mechanism between compensation strategy and finished product quality is weak, which affects the process continuous optimization capability. SUMMARY

[0004] The present application provides a high-density printed circuit board line forming process compensation method to solve the above technical problems.

[0005] The technical solution of the present application is as follows: a high-density printed circuit board line forming process compensation method, comprising: S1: Collecting process parameter data and finished product detection data from multiple batches and different process stages in the production process of high-density printed circuit boards to obtain multi-dimensional original data sets such as dimensional accuracy, line width, line spacing, electrical performance, and thermal behavior; S2: standardizing the original data set, including removing outliers, data normalization and process scene label division, to obtain a pre-processed process parameter set for different production batches and equipment working conditions; S3: based on the pre-processed process parameter set, a multi-objective compensation optimization model is constructed, the target functions of size accuracy, manufacturing feasibility, cost optimization, electrical performance and thermal behavior are parameterized defined, and a multi-objective collaborative constraint relationship is established; S4: input the multi-objective compensation optimization model into the reinforcement learning agent, use deep Q network or policy gradient algorithm, let the weight distribution vector as the action space, the comprehensive compensation simulation result as the reward signal, initialize the dynamic weight adjustment mechanism; S5: for different production batches, process labels and structure differences, the reinforcement learning agent dynamically adjusts the multi-objective weight distribution vector, and outputs the preliminary compensation amount distribution result under each target; S6: based on the preliminary compensation amount distribution result, the compensation amount calculation process is automatically deduced, and each compensation amount calculated is input into the process simulation module to perform global performance simulation evaluation on forming error, qualification rate and process window; S7: input the process simulation evaluation result into the reinforcement learning reward and punishment function, adjust the weight distribution strategy that does not meet the overall forming yield and multi-objective collaborative constraint, and continuously iterate and optimize until the global optimal or Pareto optimal comprehensive compensation weight vector is output; S8: fuse the comprehensive compensation weight vector optimized by reinforcement learning with the process parameters to generate the final comprehensive compensation amount under the multi-objective collaborative optimization of the current batch, and output the compensation strategy list; S9: apply the final comprehensive compensation amount to the actual PCB forming process execution unit, and collect the feedback data of size accuracy, electrical performance and batch yield in the execution result to incrementally train the reinforcement learning model, forming a self-learning and continuous optimization closed loop.

[0006] The high-density printed circuit board line forming process compensation method provided by the application has the following beneficial effects: (1) Unlike the existing multi-objective compensation scheme which relies on static weight manual adjustment, linear optimization or traditional heuristic algorithm, the application first introduces a reinforcement learning mechanism with deep Q network (DQN) or policy gradient as the core, dynamically models the weight distribution of multi-objectives such as size accuracy, manufacturing feasibility, cost optimization, electrical performance and thermal behavior. Through automated data collection and simulation closed loop feedback, the reinforcement learning agent can adaptively adjust the priority between each target, realize the optimal collaborative distribution of weight conflicts between cross-targets and cross-scenes, and finally obtain the compensation strategy optimized by multiple objectives; (2) The application seamlessly integrates original process parameters, detection results, compensation simulation and finished product feedback through the construction of batch MES data interface, automatic process and defect detection, simulation and AI optimization of the whole process data flow link. The reinforcement learning agent continuously takes the actual production feedback data as the reward / punishment signal, completes the incremental self-optimization training, so that the compensation strategy can dynamically evolve with the production batch, process material and equipment state, and always maintain the optimal adaptability. This closed-loop mechanism significantly overcomes the defects of existing compensation strategies, such as heavy dependence on historical experience and manual intervention, and difficulty in self-optimization, and improves the autonomous evolution and promotion robustness of the compensation model; (3) The application is suitable for single process and single product type, and can automatically output the optimal weight and compensation amount distribution scheme according to different batches, different equipment and different product structures. Whether batch fluctuation or new material is introduced, high-quality compensation self-adaptation can be realized. BRIEF DESCRIPTION OF DRAWINGS

[0007] Fig. 1 A flowchart of a high-density printed circuit board line forming process compensation method of the application; Fig. 2 A sub-flowchart of a high-density printed circuit board line forming process compensation method of the application; Fig. 3 Another sub-flowchart of a high-density printed circuit board line forming process compensation method of the application. DETAILED DESCRIPTION

[0008] In order to make the purpose and advantages of the application clearer and more apparent, the application will be further described below with reference to examples; it should be understood that the specific examples described herein are only used to explain the application, and do not limit the application.

[0009] The preferred implementation methods of the application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these implementation methods are only used to explain the technical principles of the application, and are not intended to limit the protection scope of the application.

[0010] As used herein, the singular forms "a," "an," and "the" also include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term "comprising" or "including" or "having" or the like specifies the presence of stated features, integers, steps, operations, components, parts, or combinations thereof, but does not preclude the presence or addition of one or more other features, integers, steps, operations, components, parts, or combinations thereof. Meanwhile, the term "and / or" used in the specification includes any and all combinations of the related listed items.

[0011] Referring to Figs. 1-3 As shown in the drawings, a high-density printed circuit board line forming process compensation method comprises: S1: Collecting process parameter data and finished product detection data from multiple batches and different process stages in the production process of high-density printed circuit boards to obtain multi-dimensional original data sets of size accuracy, line width, line spacing, electrical performance, thermal behavior, etc. S2: Standardizing the original data set, including removing outliers, data normalization and process scene label division, to obtain a pre-processed process parameter set for different production batches and equipment working conditions; S3: Based on the pre-processed process parameter set, a multi-objective compensation optimization model is constructed, the size accuracy, manufacturing feasibility, cost optimization, electrical performance and thermal behavior are parameterized and defined, and a multi-objective collaborative constraint relationship is established; S4: Input the multi-objective compensation optimization model into the reinforcement learning agent, use deep Q network or policy gradient algorithm, let the weight distribution vector as the action space, the comprehensive compensation simulation result as the reward signal, initialize the dynamic weight adjustment mechanism; S5: For different production batches, process labels and structure differences, the reinforcement learning agent dynamically adjusts the multi-objective weight distribution vector, and outputs the preliminary compensation amount distribution result under each target; S6: Based on the preliminary compensation amount distribution result, the compensation amount calculation process is automatically deduced, and each calculated compensation amount is input into the process simulation module to perform global performance simulation evaluation on forming error, pass rate and process window; S7: The process simulation evaluation result is input into the reinforcement learning reward and punishment function, the weight distribution strategy that does not meet the overall forming yield and multi-objective collaborative constraint is adjusted, and continuous iteration optimization is carried out until the global optimal or Pareto optimal comprehensive compensation weight vector is output; S8: The comprehensive compensation weight vector optimized by reinforcement learning is fused with the process parameters to generate the final comprehensive compensation amount under the multi-objective collaborative optimization of the current batch, and the compensation strategy list is output; S9: The final comprehensive compensation amount is applied to the actual PCB forming process execution unit, and the feedback data of size accuracy, electrical performance and batch yield rate in the execution result are collected to incrementally train the reinforcement learning model, forming a self-learning and continuous optimization closed loop.

[0012] The step S1: Collecting process parameter data and finished product detection data from multiple batches and different process stages in the production process of high-density printed circuit boards to obtain multi-dimensional original data sets of size accuracy, line width, line spacing, electrical performance, thermal behavior, etc. Specifically includes: S1.1: Schedule and screen the production batch information in the high-density printed circuit board production line, call the historical batch, process section, and equipment number basic process parameters based on the manufacturing execution system (MES), realize batch-level management of process parameter data source to ensure the pertinence and traceability of subsequent data collection; The input data includes historical batch information, process section code, and equipment number basic process parameters of the high-density printed circuit board production line, which are provided in real time by the manufacturing execution system (MES) data interface; A production batch scheduling and screening algorithm (parameters: batch history record, process section identifier, and equipment unique number) is used to realize target screening of the batch to be collected, and to exclude irrelevant batches and abnormal batches; Further, through the MES data correlation mapping mechanism (parameters: process section code, equipment number, and batch timestamp), a multiple data binding of batch parameters and process section and equipment is established to realize batch-level process data traceability; Further, through the batch distribution statistical analysis method (parameters: batch quantity distribution, process section priority, and equipment running state), a process parameter distribution matrix of each batch in the high-density printed circuit board production process is generated for target batch locking in subsequent data collection; Further, through the batch screening result driven scheduling control algorithm, the target batch information is synchronized to the data collection scheduling module to realize automatic scheduling and task allocation of the data collection process; A batch-level process parameter metadata management mechanism (parameters: batch unique code, process section, equipment information, and material batch) is used to encode the metadata of all batches to be collected to ensure that the process parameter data source has complete traceability attributes and unique index; Through batch-level screening and metadata management methods, the results of the foregoing steps are converted into process parameter data sources with high consistency, traceability, and batch management, realizing the technical effects of targeted and controllable data collection; For example, for a high-density printed circuit board factory No. 9 production line, the historical records from batch 202404 to batch 202405 are selected as the input batch range. The MES system interface is used for scheduling, each batch contains 6 process sections such as pressing, etching, and silk printing, each section has process codes such as P01, E02, and S03, equipment numbers 001-006, and material batches MD202404A and MD202404B. Through production batch scheduling and screening, abnormal batches (such as fault batch 202404X) are excluded, the total number of batches is locked to 8, and the batch parameter distribution matrix is generated as follows: Through the MES data correlation and batch-level metadata encoding, the following batch metadata management table is formed: Based on this coding, the integrated data traceability and control of production batch, process section, equipment and materials are realized, and it serves as the target batch list for subsequent parameter collection such as dimensional accuracy, line width and line spacing. In actual implementation, the data collection coverage rate after batch screening reaches 100%, and all collected data can be accurately traced back through the unique batch code, realizing the completeness and high controllability of data; S1.2: Based on automatic detection equipment and online sensing system, real-time acquisition of process parameter information at key nodes of the selected batch (such as pressing, etching, silk printing, through-hole copper plating, etc.), multi-point synchronous collection of structure parameters such as dimensional accuracy, line width, line spacing and environmental parameters such as temperature, current and voltage through process parameter collection interface to generate structure process parameter raw data; S1.3: For the aforementioned structure process parameter raw data, obtain product detection data through online defect detection system and automatic optical inspection (AOI) equipment, including dimensional deviation, electrical performance parameters (impedance, insulation resistance, etc.), thermal behavior parameters (thermal expansion coefficient, peel strength), etc., to form detection process parameter raw data; S1.4: Integrate the structure process parameter raw data and the detection process parameter raw data at the batch level, encode different process sections, equipment, material batches, etc. using data tagging mechanism to realize the unified collection of multi-dimensional process parameter metadata set; S1.5: Apply process mapping relationship model to the collected multi-dimensional process parameter metadata set to perform parameter mapping and feature expansion processing, supplement and extract advanced process features such as lamination relationship, dielectric thickness, micro-pore density, etc., to realize the conversion of raw data to multi-dimensional process feature data set, and provide high-dimensional and structured data basis for subsequent standardized preprocessing and modeling analysis.

[0013] The step S2: standardizing the raw data set, including removing outliers, data normalization and process scenario label division, to obtain a preprocessed process parameter set for different production batches and equipment conditions. Specifically, it includes: S2.1: Perform automatic outlier detection algorithm on each process parameter (including dimensional accuracy, line width, line spacing, electrical performance, thermal behavior, etc.) in the raw data set, such as 3σ statistical method or based on isolation forest model, to mark and remove outliers that significantly deviate from historical distribution, to obtain a high-density printed circuit board production process parameter intermediate data set with outliers removed; S2.2: Take the high-density printed circuit board production process parameter intermediate data set with outliers removed as input, apply Z-score normalization or max-min normalization algorithm to perform numerical standardization processing on various process parameter features, map parameters of different dimensions and different units to a unified distribution interval, and output the normalized process parameter feature vector set; S2.3: Based on the normalized process parameter feature vector set, use clustering algorithms (such as K-means clustering or clustering based on Gaussian mixture model) to analyze the multi-dimensional feature space of process parameters, and divide the multi-dimensional process scene labels according to production batch, equipment category and process window state, etc. attributes, to form a process parameter label set with scene recognition ability; S2.4: Consistency check and deduplication processing on the process parameter label set, introduce multi-label fusion judgment mechanism (such as voting mechanism or confidence weighting algorithm), handle the problem of multi-label overlap and unclear attribution caused by batch distribution noise, and output the consistent process label parameter set with high confidence label; S2.5: Joint processing of consistent process label parameter set and normalized feature vector set, generating high-consistency pre-processing process parameter set for different production batches and equipment working conditions through label binding mechanism, providing standard input data for parameter input and weight adjustment mechanism of subsequent multi-objective compensation optimization model.

[0014] The step S3: based on the pre-processing process parameter set, construct a multi-objective compensation optimization model, parameterize the definition of size accuracy, manufacturing feasibility, cost optimization, electrical performance and thermal behavior, etc. objective function, and establish a multi-objective collaborative constraint relationship. As shown in Fig. 2 Specifically includes: S3.1: Process feature extraction processing is performed on the key process parameters (such as size accuracy, line width, line spacing, etc.) in the pre-processing process parameter set to obtain the feature parameter set related to each objective function, and the parameter input basis is established for the multi-objective compensation optimization model; S3.2: Based on the obtained feature parameter set, use parameterized modeling algorithm (such as vector parameterization, normalized weighting, etc.) to parameterize the definition of size accuracy, manufacturing feasibility, cost optimization, electrical performance and thermal behavior, etc. each objective function, to clearly express the scientific calculation expression of each objective function; S3.3: Using the parameterized objective function and its dependent feature parameter set, use the target-dependent mapping algorithm to analyze the constraint relationship and mutual influence between each objective function, and obtain the coupling matrix between objectives in the multi-objective compensation optimization model, to provide systematic calculation basis for constraint relationship establishment; S3.4: On the basis of the multi-objective coupling matrix, a collaborative constraint modeling method (such as Lagrange multiplier, penalty function, etc.) is used to build the collaborative constraint relationship between size accuracy, manufacturing feasibility, cost optimization, electrical performance and thermal behavior of multiple objectives, so as to obtain a constraint system that can be used for global collaborative optimization; S3.5: The above-mentioned collaborative constraint system is integrated with the parameterized objective function set, and a data structure generation method oriented to the multi-objective optimization framework is used to output the systematic model description of the multi-objective compensation optimization model, laying an integrated data foundation for subsequent reinforcement learning driven weight adjustment and compensation scheme simulation.

[0015] The step S4: input the multi-objective compensation optimization model into the reinforcement learning agent, use deep Q network or policy gradient algorithm, let the weight distribution vector as the action space, the comprehensive compensation simulation result as the reward signal, initialize the dynamic weight adjustment mechanism. As shown in Fig. 3 Specifically includes: S4.1: The parameterized objective function (such as size accuracy, manufacturing feasibility, cost optimization, etc.) in the multi-objective compensation optimization model constructed is structurally analyzed, and the target weight distribution vector is extracted to form the initial input of the action space of the reinforcement learning agent; S4.2: Based on the above-mentioned target weight distribution vector, the policy network of the reinforcement learning agent is initialized by using deep Q network or policy gradient algorithm, and the weight adjustment mechanism is set so that the mechanism can be directly associated with and respond to the parameter weight change of the multi-objective compensation optimization model; S4.3: The weight distribution vector defined by the parameterization of the multi-objective compensation optimization model is input into the reinforcement learning agent, and the output results (such as compensation amount simulation pass rate, forming quality, process window, etc.) of the comprehensive compensation simulation module are used as the reward signal source of the reinforcement learning agent, and a training data pair is constructed; S4.4: The reward function of the reinforcement learning agent is normalized to ensure that the multi-objective collaborative constraint parameters including size accuracy, manufacturing feasibility, cost optimization, electrical performance, etc. are accurately mapped to the reward signal vector as the professional term basis for initializing the dynamic weight adjustment mechanism; S4.5: Based on the weight distribution action space and reward signal of the reinforcement learning agent, the first weight distribution optimization iteration is performed to generate a preliminary compensation weight adjustment scheme with multi-objective collaborative constraint characteristics, providing a data-driven basis for subsequent dynamic adjustment and comprehensive optimization.

[0016] The step S5: For different production batches, process labels and structure differences, the multi-objective weight distribution vector is dynamically adjusted by the reinforcement learning agent to output the preliminary compensation amount distribution result under each target. Specifically includes: S5.1: Input the pre-processing process parameter set obtained in the previous stage into the multi-objective compensation optimization model, generate the corresponding process scene matching feature vector based on the production batch identification, process label and structure characteristic information, to realize the initial condition configuration of weight distribution under multiple scene process conditions; S5.2: Input the process scene matching feature vector into the reinforcement learning agent, initialize the weight distribution vector through the deep Q network or policy gradient algorithm, to obtain the initial state of the multi-objective weight suitable for the current process scene; S5.3: Based on the process scene matching feature vector and the initialized weight distribution state, execute the policy decision iteration process of the reinforcement learning agent, dynamically adjust the multi-objective weight distribution vector parameters to adapt to the multi-dimensional differences of production batch, process label, structure characteristics, etc., and realize the self-learning optimization of weight distribution parameters; With the process scene matching feature vector and the initial weight distribution state as input, the reinforcement learning agent (such as deep Q network DQN or policy gradient method) is used to dynamically execute the multi-objective collaborative weight adjustment; The reinforcement learning policy iteration algorithm (parameters: current batch process feature vector, initial weight distribution, learning rate, discount factor, etc.) is used to realize the action policy selection and update of the multi-objective weight distribution vector; Further, through the policy evaluation mechanism, the process simulation feedback (including size accuracy, electrical performance, manufacturing feasibility, cost optimization, etc.) corresponding to each weight distribution behavior is normalized as a reward signal according to the following formula: Wherein, is the total reward of the th decision, is the current policy distribution weight of the th objective, is the process simulation evaluation result of the th objective under the weight distribution vector ; The greedy strategy or policy gradient update method is used to update the weight distribution policy according to the reward ; The gradient ascent or Q value update is as follows: Wherein, is the learning rate, is the reward gradient of the current weight distribution; Further, through the policy network / value function convergence judgment, when the loss function is lower than the threshold or the weight distribution tends to be stable in several rounds of simulation, the optimal multi-objective weight distribution result under the current batch, process label and structure characteristics is output. Through the dynamic weight adjustment and self-learning optimization mechanism described above, the process characteristics are adaptively coupled with historical feedback, realizing personalized collaborative compensation weight distribution results for multiple batches and multiple process scenarios, providing structured input for subsequent compensation amount calculation and objective function decomposition; For example, for a certain high-end HDI PCB sample, the input batch is 20240601A, the process label is multi-layer fine line etching, and the structure characteristics are inner layer micro-hole density 0.8 / mm 2 , medium plate thickness 0.075mm, the initial weight vector is [0.25, 0.25, 0.2, 0.2, 0.1] (dimensional accuracy, manufacturing feasibility, cost optimization, electrical performance, thermal behavior). Using DQN algorithm, setting learning rate 0.01, discount factor 0.9, simulation evaluation function Based on the process simulation output, it is measured that the dimensional error is 10μm, the manufacturing compliance rate is 98.5%, the unit cost is 6.8 yuan / pcs, the impedance deviation is ±5%, and the peel strength is ≥1.8kgf / cm. After 20 rounds of policy iteration, the reward function is calculated by the aforementioned weighted formula, and after 5 times of optimization convergence, the weight is updated to [0.32, 0.2, 0.18, 0.22, 0.08], the dimensional accuracy weight is preferentially improved, and the electrical performance and manufacturing feasibility are considered. The final output of this batch of personalized multi-objective compensation weight distribution is assigned to each target and the compensation amount is simulated in detail, and the process yield after verification is improved to 99.1%, the dimensional error is converged to 8μm, and the electrical performance compliance rate is improved to 99%; S5.4: The multi-objective weight distribution vector output by the reinforcement learning intelligent agent is processed by objective function decomposition, and the weight distribution parameters are assigned to the dimensional accuracy target, the manufacturing feasibility target, the cost optimization target, the electrical performance target and the thermal behavior target, forming a preliminary compensation weight configuration set for each target function; S5.5: The preliminary compensation weight configuration set is subjected to normalization test and process constraint consistency verification, and a multi-objective collaborative constraint relationship model is used to determine the feasibility of the distribution result. The preliminary weight distribution result that passes the test is output as the preliminary compensation amount distribution result matrix corresponding to each target function in the current process batch, and is used by the subsequent compensation amount calculation and simulation evaluation module.

[0017] The step S6: based on the preliminary compensation amount distribution result, automatically deduce the compensation amount calculation process, input the calculated compensation amount into the process simulation module, and perform global performance simulation evaluation on the forming error, compliance rate and process window. Specifically, it includes: S6.1: uniformly format the preliminary compensation amount allocation results to form a compensation parameter set containing dimensional accuracy compensation amount, manufacturing feasibility compensation amount, electrical performance compensation amount and thermal behavior compensation amount, and provide input conditions for the standard data interface of the subsequent process simulation module; S6.2: based on the compensation parameter set, use a multi-objective high-dimensional mapping algorithm to model the correlation between the compensation amounts to obtain a coupling parameter matrix of the compensation parameter set under multi-objective collaborative constraints, and provide multi-dependent variable inputs for simulation calculation; S6.3: input the coupling parameter matrix into the multi-objective process simulation module, and use the comprehensive performance simulation engine to quantitatively simulate the dimensional accuracy error to obtain the forming error distribution parameter as the simulation output parameter of the dimensional target; S6.4: in the process simulation module, based on the forming error distribution parameter and the manufacturing feasibility compensation amount, evaluate the process compliance rate under each compensation amount allocation through statistical analysis and process window simulation algorithm, and form multi-objective compliance rate evaluation data; S6.5: fuse the forming error distribution parameter and the multi-objective compliance rate evaluation data, combine the simulation results of the electrical performance compensation amount and the thermal behavior compensation amount, and use a comprehensive performance evaluation function to analyze the process window in multiple dimensions to obtain a global performance simulation evaluation report, which provides feedback for subsequent dynamic optimization of compensation weights.

[0018] The step S7: input the process simulation evaluation results into the reinforcement learning reward and punishment function, adjust the weight allocation strategy that does not meet the overall forming yield and multi-objective collaborative constraints, and continuously iterate and optimize until the global optimal or Pareto optimal comprehensive compensation weight vector is output. Specifically, it includes: S7.1: data extraction and professional classification analysis of forming error, compliance rate and process window performance indicators in process simulation evaluation results are performed as input factors of the reinforcement learning reward and punishment function, so as to realize quantitative evaluation of the effect of compensation weight allocation scheme and provide objective performance basis for subsequent weight optimization steps; S7.2: based on the extracted process simulation performance indicators, the reinforcement learning reward and punishment function is used to calculate the comprehensive reward signal of the current compensation weight allocation strategy, which clearly reflects the overall compensation advantages and disadvantages under the multi-objective collaborative constraints of dimensional accuracy, electrical performance, manufacturing feasibility and cost optimization, and outputs the reward / punishment feedback results; S7.3: reinforcement learning strategy gradient or deep Q network iterative update is performed on the compensation weight allocation vector, and the target compensation weights are automatically optimized under the driving of the reward / punishment feedback results, realizing adaptive dynamic weight adjustment and obtaining a new compensation weight allocation scheme after optimization; The compensation weight allocation vector, the reward / punishment feedback results and the multi-objective collaborative constraint parameters are input into the reinforcement learning model. A parameterized optimization of the compensation weight distribution vector is implemented by using a policy gradient algorithm (parameters: current weight distribution vector, reward signal, learning rate, gradient of the objective function); Further, an action selection of the weight distribution strategy is implemented by using a deep Q-network algorithm (parameters: state space is the history trajectory of the weight distribution, action space is the adjustment scale of each target weight, reward function is the evaluation result of the multi-objective process performance), and a value function update is performed according to the reward / punishment feedback; A dynamic self-adaptive adjustment of the weight distribution parameters is implemented by using a gradient update formula of the weight distribution vector; wherein, is a parameter of the compensation weight distribution vector, is a learning rate, is a reward function value of the reinforcement learning, is a reward function gradient under the current parameter; Further, the updated weight distribution scheme is mapped to a feasible objective function domain by using weight normalization and constraint correction processing to avoid constraint violation; An objective function decomposition algorithm is used to assign the optimized weight distribution vector to each target compensation weight parameter domain of the dimensional accuracy, manufacturing feasibility, electrical performance, cost optimization and thermal behavior, etc. Through the adaptive iteration mechanism of the reinforcement learning model, the weight optimization is continuously performed according to the reward / punishment feedback results, the continuous iteration and evolution of the weight distribution scheme are realized, and finally a new compensation weight distribution scheme with collaborative optimization characteristics is output; By using reinforcement learning algorithms such as policy gradient and deep Q-network, the optimized compensation weight distribution vector is converted into new process compensation parameters, and a self-learning dynamic adjustment of the multi-objective collaborative compensation strategy is realized; Exemplary, in the actual high-density PCB mass production scene, the reinforcement learning agent takes the historical 20 batches of process simulation data as the state space, and sets the initial weight distribution: size accuracy 0.35, manufacturing feasibility 0.25, electrical performance 0.15, thermal behavior 0.10, cost optimization 0.15, and each weight is normalized to 1. The policy gradient learning rate is set to 0.05, and the reward function is "compensation simulation compliance rate 0.4 + molding error reverse score 0.3 + process window coverage rate 0.2 + production cost reverse score 0.1". The weight distribution vector is adjusted using the gradient update formula: if the simulation reward is limited, the gradient descent optimization will increase the size accuracy weight to 0.38, decrease the manufacturing feasibility to 0.23, increase the electrical performance to 0.17, keep the thermal behavior, and slightly decrease the cost optimization to 0.12. After 10 iterations, the output compensation weight distribution is [0.40, 0.22, 0.18, 0.10, 0.10], the process simulation compliance rate is improved from 92.2% to 96.5%, the average molding error is reduced by 8%, and the cost increase is controlled within the warning line. This scheme is corrected by constraint every round to ensure that the size, electrical and manufacturing feasibility are not lower than the previous indicators. The final output of the optimized compensation weight distribution scheme realizes the global performance improvement under multi-objective constraints, and provides reliable parameter basis for subsequent compensation strategy generation and actual process execution; S7.4: Take the new round of optimized compensation weight distribution scheme as input, re-derive the compensation amount calculation process, and input the obtained compensation amount into the process simulation module to obtain the next round of molding error, compliance rate and process window simulation evaluation index, forming a closed-loop causal interaction between process performance and compensation strategy; S7.5: In the process of continuous iteration and optimization, all historical process simulation feedback indicators and compensation weight distribution schemes are summarized and multi-dimensional performance boundary analysis is performed, and a multi-objective optimal solution screening algorithm is used to finally output the global optimal or Pareto optimal comprehensive compensation weight vector, providing high-reliability decision basis for subsequent compensation strategy list and process execution unit.

[0019] The step S8: fuse the comprehensive compensation weight vector optimized by reinforcement learning with the process parameters to generate the final comprehensive compensation amount under the multi-objective collaborative optimization of the current batch, and output the compensation strategy list. Specifically, it includes: S8.1: Obtain the comprehensive compensation weight vector optimized by reinforcement learning and the current batch of standardized process parameter set, and perform structured mapping on the two types of data based on a data fusion algorithm to construct a compensation parameter input matrix and serve as the input condition of the compensation calculation function; S8.2: Based on the compensation parameter input matrix, a multi-objective compensation amount calculation model (such as a weighted linear model, a multi-objective nonlinear programming algorithm) is used to calculate the preliminary compensation amount output of each target for size accuracy, manufacturing feasibility, electrical performance, etc. professional terms, and obtain the target compensation amount distribution result; S8.3: The target compensation amount distribution result is prioritized and the constraint condition is checked, and a collaborative constraint verification algorithm (such as multi-objective weight normalization, Pareto effectiveness judgment) is applied to generate a collaborative optimization verification report to determine whether the compensation scheme meets the multi-objective performance requirements; S8.4: According to the collaborative optimization verification report, the final compensation amount set that meets the multi-objective performance requirements is subjected to data format processing to generate a standardized compensation strategy list, and the parameter archiving content of each target such as size accuracy compensation and electrical performance compensation is marked; S8.5: The standardized compensation strategy list and each final compensation amount are associated with the current batch process number and transmitted to the molding process execution unit and monitoring module to realize the application of the process compensation strategy under multi-objective collaboration, and form a technical closed-loop interface data.

[0020] The step S9: The final comprehensive compensation amount is applied to the actual PCB molding process execution unit, and the feedback data such as size accuracy, electrical performance and batch yield in the execution result are collected to strengthen the learning model for incremental training, forming a self-learning and continuous optimization closed loop. Specifically includes: S9.1: The comprehensive compensation amount optimized by reinforcement learning is applied and deployed, and the comprehensive compensation amount is embedded in the PCB molding process control parameter group as the direct input of the molding equipment control system to realize the molding process parameter adjustment based on multi-objective dynamic collaborative optimization; S9.2: Based on the molding process control parameter group, the molding execution unit is driven to perform line molding operation on each batch of PCB, and the production batch identifier and process scene context generated during the execution process are recorded in real time to ensure subsequent data tracing and batch controllability; S9.3: Collect the size accuracy, electrical performance and batch yield of the PCB after applying the comprehensive compensation amount, and establish a quality feedback data set for the specific production batch as the data input for incremental training of reinforcement learning; S9.4: The batch quality feedback data set is subjected to data standardization and format alignment, and the data preprocessing algorithm is used to eliminate abnormal values and noise data, and scene label mapping is performed based on the molding process control parameter group to obtain a high-consistency model training input data stream; S9.5: input the format-aligned model training input data stream into the reinforcement learning model, perform incremental training, adjust the weight distribution mechanism through the reinforcement learning agent, continuously optimize the model decision-making ability combined with newly collected finished product quality data, and improve the adaptability of subsequent comprehensive compensation amount output to actual molding quality; S9.6: based on the reinforcement learning model after incremental training, dynamically update the comprehensive compensation amount generation strategy, and generate self-learning adjustment instructions for the molding process parameter group to realize the continuous evolution of the compensation strategy and the process self-optimization closed loop.

[0021] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without deviating from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

[0022] The above description is only the preferred embodiments of the present application and is not intended to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included within the protection scope of the present application.

Claims

1. A method of compensating for line formation in a high-density printed circuit board process, characterized by, The method comprises the following steps: S1: Collecting process parameter data and finished product detection data from multiple batches and different process stages in the production process of high-density printed circuit boards, obtaining a multi-dimensional original data set; S2: Standardizing the multi-dimensional original data set to obtain a pre-processed process parameter set; S3: Based on the pre-processed process parameter set, a multi-objective compensation optimization model is constructed, the objective function is parameterized defined, and a multi-objective collaborative constraint relationship is established; S4: Input the multi-objective compensation optimization model into the reinforcement learning agent, let the weight distribution vector be the action space, and the comprehensive compensation simulation result be the reward signal, initialize the dynamic weight adjustment mechanism; S5: For different production batches, process labels and structure differences, the reinforcement learning agent dynamically adjusts the multi-objective weight distribution vector, and outputs the preliminary compensation amount distribution result; S6: Based on the preliminary compensation amount distribution result, the compensation amount calculation process is automatically deduced, and the calculated compensation amounts are input into the process simulation module to perform global performance simulation evaluation on the forming error, the qualification rate and the process window; S7: The process simulation evaluation result is input into the reinforcement learning reward and punishment function, and the weight distribution strategy that does not meet the overall forming yield and multi-objective collaborative constraint is adjusted for continuous iteration and optimization; S8: The comprehensive compensation weight vector optimized by reinforcement learning is fused with the process parameters to generate the final comprehensive compensation amount under the multi-objective collaborative optimization of the current batch, and the compensation strategy list is output.

2. The method for compensating for high-density printed circuit board circuit forming process according to claim 1, characterized in that, The step S8 further comprises: S9: Apply the final comprehensive compensation amount to the actual PCB forming process execution unit, collect feedback data in the execution result to perform incremental training on the reinforcement learning model, and form a self-learning and continuous optimization closed loop.

3. The method for compensating for high-density printed circuit board circuit forming process according to claim 1, characterized in that, The step S1 specifically comprises: The production batch information in the high-density printed circuit board production line is scheduled and screened, and the basic process parameters are called based on the manufacturing execution system; Based on the automatic detection equipment and the online sensing system, the process parameter information of the key nodes of the selected batch is obtained in real time, the multi-point synchronous collection of structure parameters and environment parameters is performed through the process parameter collection interface, and the structure process parameter original data is generated; Through the online defect detection system and the automatic optical detection equipment, the finished product detection data is obtained from the structure process parameter original data, and the detection process parameter original data is formed; The structure process parameter original data and the detection process parameter original data are associated and integrated at the batch level, the meta information is encoded through the data labeling mechanism, and the multi-dimensional process parameter metadata set is uniformly collected; The collected multi-dimensional process parameter metadata set is applied to the process mapping relationship model to perform parameter mapping and feature expansion processing, and high-level process features are supplemented and extracted.

4. The method for compensating for high-density printed circuit board circuit forming process according to claim 3, characterized in that, In the step S1, the manufacturing execution system MES interface is used for data tracing of production batch, process section, equipment number and material batch, and the target batch is automatically locked and the data collection process is automatically scheduled through batch scheduling screening, metadata coding and batch distribution matrix analysis.

5. The method for compensating for high-density printed circuit board circuit forming process according to claim 1, characterized in that, The step S2 specifically comprises: An automatic outlier detection algorithm is performed on each process parameter in the original data set, and outliers significantly deviating from the historical distribution are marked and removed to obtain an intermediate data set of high-density printed circuit board production process parameters without outliers; The intermediate data set of high-density printed circuit board production process parameters without outliers is inputted to perform numerical standardization processing on various process parameter features, and an output is obtained in the form of a set of normalized process parameter feature vectors; Based on the set of normalized process parameter feature vectors, a multi-dimensional feature space analysis is performed on the process parameters, a multi-dimensional process scene label division is performed, and a process parameter label set is formed; The process parameter label set is subjected to consistency verification and deduplication processing, a multi-label fusion judgment mechanism is introduced, problems of multi-label overlap and unclear attribution caused by batch distribution noise are processed, and a consistent process label parameter set with high confidence labels is outputted; The consistent process label parameter set and the normalized feature vector set are jointly processed to generate a high-consistency preprocessed process parameter set for different production batches and equipment working conditions.

6. The method for compensating for high-density printed circuit board circuit forming process according to claim 5, characterized in that, The standardization processing includes performing automatic outlier detection on process parameters, using a 3σ statistical method or an isolation forest model, performing numerical standardization on the parameters after outlier removal using a Z-Score normalization or a maximum and minimum value normalization algorithm, and performing multi-dimensional feature space analysis and label division on the process parameters using a clustering algorithm to output a consistent process label parameter set.

7. The method for compensating for high-density printed circuit board circuit forming process according to claim 1, characterized in that, The step S3 specifically includes: Process feature extraction processing is performed on key process parameters in the preprocessed process parameter set to obtain a feature parameter set related to each objective function; Based on the feature parameter set, a parameterized modeling algorithm is used to parameterize and define each objective function; Using the parameterized objective function and its dependent feature parameter set, a target-dependent mapping algorithm is used to analyze the constraint relationship and mutual influence between each objective function to obtain a target coupling matrix in the multi-objective compensation optimization model; Based on the target coupling matrix, a collaborative constraint modeling method is used to construct the collaborative constraint relationship between multiple objectives to obtain a collaborative constraint system that can be used for global collaborative optimization; The collaborative constraint system is integrated with the parameterized objective function set, and a data structure generation method for a multi-objective optimization framework is used to output a systematic model description of the multi-objective compensation optimization model.

8. The method for compensating for high-density printed circuit board circuit forming process according to claim 1, characterized in that, The step S4 specifically includes: The parameterized objective function in the constructed multi-objective compensation optimization model is structurally analyzed to extract a target weight distribution vector to form an initial input of the action space of the reinforcement learning agent; Based on the target weight distribution vector, a deep Q network or a policy gradient algorithm is used to initialize the policy network of the reinforcement learning agent and set a weight adjustment mechanism; The weight distribution vector obtained by parameterizing the multi-objective compensation optimization model is inputted into the reinforcement learning agent, and the output result of the comprehensive compensation simulation module is used as the reward signal source of the reinforcement learning agent to construct a training data pair; The process performance indicators in the reward function of the reinforcement learning agent are normalized, and the multi-objective collaborative constraint parameters are mapped to the reward signal vector; Based on the weight allocation action space and the reward signal of the reinforcement learning agent, a first weight allocation optimization iteration is performed to generate a preliminary compensation weight adjustment scheme with multi-objective collaborative constraint characteristics.

9. The method of claim 1, wherein the method is used in a high density printed circuit board routing process. The step S5 specifically includes: ​ The obtained pre-processing process parameter set and the multi-objective compensation optimization model are input processed, and based on the production batch identifier, the process label and the structure feature information, a corresponding process scene matching feature vector is generated; The process scene matching feature vector is input into the reinforcement learning agent, and the weight allocation vector is initialized by a deep Q network or a policy gradient algorithm to obtain a multi-objective weight initial state adapted to the current process scene; Based on the process scene matching feature vector and the initialized weight allocation state, a policy decision iteration process of the reinforcement learning agent is performed to dynamically adjust the multi-objective weight allocation vector parameters; The multi-objective weight allocation vector output by the reinforcement learning agent is subjected to objective function decomposition processing, the weight allocation parameters are assigned to the dimensional accuracy target, the manufacturing feasibility target, the cost optimization target, the electrical performance target and the thermal behavior target, and a preliminary compensation weight configuration set for each target function is formed; The preliminary compensation weight configuration set is subjected to normalization inspection and process constraint consistency verification, a multi-objective collaborative constraint relationship model is used to determine the feasibility of the allocation result, and the preliminary weight allocation result that passes the inspection is output as a preliminary compensation amount allocation result matrix corresponding to each target function under the current process batch.

10. A method for compensating for high-density printed circuit board circuit forming process according to claim 9, characterized in that, In the step S5, for different production batches, process labels and structure features, the weight allocation is dynamically adjusted based on the scene matching feature vector through the policy decision process of the reinforcement learning agent, the individualized collaborative compensation weight allocation is performed for multiple batches and multiple process scenes, and the reward function is optimized iteratively after normalization by using a greedy strategy or a policy gradient method.

Images