PCB plating solution health state monitoring and plating performance evaluation method and system

Abstract

The present application relates to a PCB plating solution health state monitoring and plating performance evaluation method and system, belonging to the technical field of PCB plating, the present application outputs a candidate causal path set by using a dynamic knowledge graph, combines the candidate causal path set output by the dynamic knowledge graph to evaluate the plating performance, generates an evaluation result, constructs a reasoning engine based on a large language model, and uses the reasoning engine to reason the evaluation result to generate a specific scheme for repair and process compensation. The present scheme uses the text reasoning capability of the large language model to convert the causal path into directly executable operation suggestions for process guidance personnel and automatic control equipment, significantly reducing the complexity and response time of expert diagnosis of plating quality problems, realizing online causal reasoning from dynamic multi-source perception to semantic machine reading comprehension, so as to quickly respond to adjustment strategies, reduce the rate of defective products in the plating process, and improve the quality of plating.

Description

Technical Field

[0001] This invention relates to the field of PCB electroplating technology, and in particular to a method and system for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance. Background Technology

[0002] As the support structure for electronic components and the provider of electrical interconnects, the manufacturing quality of printed circuit boards (PCBs) directly determines the reliability and performance of electronic products. In the manufacturing process of multilayer and high-density interconnect (HDI) PCBs, electroplating is a crucial step in achieving interlayer electrical conductivity, via metallization, and circuit pattern formation. With the miniaturization, high-frequency operation, and high integration of electronic products, PCB design is characterized by finer conductors, smaller via diameters, increased aspect ratios, and higher signal integrity requirements, posing unprecedented challenges to electroplating technology.

[0003] Traditional PCB electroplating is mainly based on the principle of direct current electroplating. On the surface of the holes and circuit areas where copper has been chemically deposited, an external current is applied to reduce and deposit copper ions, thereby obtaining a conductive coating of the required thickness. Currently, the most widely used process is acidic sulfate system electroplating, the core of which lies in controlling the deposition rate of different micro-regions through the synergistic effect of additives to achieve a uniform coating distribution.

[0004] In the PCB electroplating production process, abnormal fluctuations in the health status of the electroplating solution are often caused by a combination of factors, such as batch differences in raw materials, fluctuations in equipment supply, and changes in ambient temperature. Traditional manual experience judgment or single-parameter quantitative models are difficult to provide scientific and accurate root cause analysis. Therefore, it is urgent to comprehensively consider multiple factors to monitor the health status of the PCB electroplating solution and optimize the electroplating based on the monitoring results in order to reduce the generation of defective products. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and system for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for monitoring the health status of PCB electroplating solutions and evaluating their electroplating performance, characterized by comprising the following steps: Construct electroplating process entity nodes and process parameter attribute nodes, and construct a directed attribute graph structure based on the electroplating process entity nodes and process parameter attribute nodes to form a dynamic knowledge graph. A dynamic causal intensity matrix is ​​constructed for each sensor time series data, and a multi-hop graph attention network is used to model the dynamic causal intensity matrix. A set of candidate causal paths is output using a dynamic knowledge graph. The electroplating performance is evaluated by combining the candidate causal path set output from the dynamic knowledge graph, and the evaluation results are generated. A reasoning engine based on a large language model is constructed, and the reasoning engine is used to reason about the evaluation results to generate specific solutions for maintenance and process compensation.

[0007] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, electroplating process entity nodes and process parameter attribute nodes are constructed, specifically as follows: Define the process entity of PCB electroplating process, and divide the process entity of PCB electroplating process into process equipment layer, process medium layer, additive layer and process index layer. Electroplating process entity nodes are constructed based on the process equipment layer, and process parameter attribute nodes are constructed based on the process medium layer, additive layer, and process index layer. The process equipment layer is a hardware entity, including a main electroplating tank, circulating pump, filter, heat exchanger, rectifier, automatic dosing pump, and stirring device; the process medium layer includes Cu. 2+ Concentration, Ni 2+ Concentration, pH value, temperature, conductivity, dissolved oxygen concentration, and total organic carbon; The additives include inhibitors, accelerators, leveling agents, and chloride ions; the process indicator layers include deposition rate, coating thickness uniformity, concentration of additive degradation byproducts, and bath viscosity.

[0008] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a directed attribute graph structure is constructed based on the electroplating process entity nodes and process parameter attribute nodes to form a dynamic knowledge graph, specifically: Based on electrochemical theory and historical process documentation, direct causal relationships, inhibitory relationships, promoting relationships, and degradation-genesis relationships were obtained. Edge types are constructed based on the aforementioned direct causal relationships, inhibition relationships, promotion relationships, and degradation-generation relationships; A dynamic knowledge graph is constructed, and a directed attribute graph structure is built based on the edge type, electroplating process entity nodes, and process parameter attribute nodes. The directed attribute graph structure is input into the dynamic knowledge graph for knowledge representation.

[0009] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a dynamic causal intensity matrix is ​​constructed for the time series data of each sensor, specifically as follows: The Granger causality test algorithm is introduced, and sensor data information collected by the sensor at each time stamp is obtained. Sensor time series data is constructed based on the sensor data information collected by the sensor at each time stamp. The time series data of each sensor is input into the Granger causality test algorithm, and the Granger causality test algorithm is used to calculate the causal strength value of the bivariate within a continuous sliding time window. The causal intensity values ​​of the two variables within a continuous sliding time window are smoothed by an exponentially weighted moving average, and the causal intensity is used to capture the delayed response and asymmetric causal relationship between different processes. A causal intensity matrix is ​​constructed based on the causal intensity values ​​of the two variables within a continuous sliding time window, and the causal intensity matrix is ​​updated incrementally in real time using a sliding time window.

[0010] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, the dynamic causal intensity matrix is ​​modeled using a multi-hop graph attention network, specifically as follows: A multi-hop graph attention network is introduced, and the dynamic causal intensity matrix is ​​input into the multi-hop graph attention network. For each node and its neighboring nodes within a preset range, the attention coefficient is calculated using single-head attention and then normalized. After normalization calculation, the node features are updated. A multi-hop graph attention network is used to stack a preset number of layers. Each layer performs attention aggregation on the node features of the previous layer and calculates the credibility of the long-distance path. Based on the credibility of the long-range path, a reverse search is performed starting from the abnormal node. At each step, a part of the path is retained and sorted according to the current cumulative confidence. When the path length reaches a preset length or the cumulative confidence is lower than a preset threshold, the expansion stops. All long-range paths are counted, and a set of candidate causal paths is constructed based on these long-range paths. This set is then used as the output of the dynamic knowledge graph, thus completing the modeling process.

[0011] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a set of candidate causal paths is output using a dynamic knowledge graph, specifically: Acquire sensor time series data within a preset time period, and input the sensor time series data within the preset time period into the dynamic knowledge graph for knowledge matching; By matching knowledge, a set of candidate causal paths is output using a dynamic knowledge graph and displayed in a preset manner.

[0012] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, the electroplating performance is evaluated by combining the candidate causal path set output by the dynamic knowledge graph, and evaluation results are generated, specifically as follows: Set the range of electroplating performance evaluation index data, and determine whether the set of candidate causal paths is within the range of electroplating performance evaluation index data; When the set of candidate causal paths falls within the range of electroplating performance evaluation index data, an evaluation result indicating normal electroplating is generated. When the set of candidate causal paths is not within the range of electroplating performance evaluation index data, an evaluation result for electroplating anomaly is generated.

[0013] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a reasoning engine based on a large language model is constructed, specifically as follows: The sensor time series data is normalized, and an embedding vector for each parameter is generated using a pre-trained Transformer time series encoder. After positional encoding of all parameter embeddings, they are concatenated into a global embedding vector, which is then compressed to a fixed dimension by linear projection and input into the large language model. The output candidate path set is combined with the global embedding vector to calculate the path matching score. After the causal path is sorted on the candidate set by the semantic vector, the path with the highest matching score in the candidate path set is selected and input into the large language model. The path with the highest matching score in the candidate path set and the current semantic context are fed into the context retrieval enhancement generation module of the large language model, and finally the large language model is used to output the result in a chain-like reasoning manner.

[0014] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, the inference engine is used to infer the evaluation results and generate specific solutions for repair and process compensation, specifically as follows: When the evaluation result is an evaluation result of electroplating abnormality, the sensor time series data within a preset time and the candidate path set are input into the inference engine for inference. By leveraging a large language model inference engine and ontology knowledge in the field of electroplating chemistry, complex causal chains are transformed into actionable work order instructions, generating specific solutions for maintenance and process compensation, and supplementing them with optimization suggestion text described in natural language.

[0015] A second aspect of the present invention provides a PCB electroplating solution health status monitoring and electroplating performance evaluation system, including a memory and a processor. The memory includes a PCB electroplating solution health status monitoring and electroplating performance evaluation method program. When the processor executes the PCB electroplating solution health status monitoring and electroplating performance evaluation method program, it implements the steps of any of the PCB electroplating solution health status monitoring and electroplating performance evaluation methods described in the present invention.

[0016] This invention addresses the shortcomings of the prior art and has the following beneficial effects: This solution constructs a dynamic causal graph, combining Granger causality tests and graph attention networks to automatically identify complex nonlinear process causal chains, thereby enabling the monitoring of electroplating solution performance. Leveraging the text reasoning capabilities of a large language model, the causal paths are transformed into actionable operational suggestions for process instructors and automated control equipment. It significantly reduces the complexity and response time of expert diagnosis of electroplating quality problems, achieving dynamic multi-source perception and semantic machine reading comprehension-based online causal reasoning. This allows for rapid response and strategy adjustment, reducing the defect rate and improving electroplating quality. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0018] Figure 1 The overall flowchart of the method for monitoring the health status of PCB electroplating solution and evaluating electroplating performance is shown; Figure 2 A system block diagram of a PCB electroplating solution health status monitoring and electroplating performance evaluation system is shown. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0021] like Figure 1 As shown, the first aspect of the present invention provides a method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, characterized by comprising the following steps: Construct electroplating process entity nodes and process parameter attribute nodes, and build a directed attribute graph structure based on the electroplating process entity nodes and process parameter attribute nodes to form a dynamic knowledge graph. A dynamic causal intensity matrix is ​​constructed for each sensor time series data, and a multi-hop graph attention network is used to model the dynamic causal intensity matrix. A set of candidate causal paths is output using a dynamic knowledge graph. By combining a dynamic knowledge graph to output a set of candidate causal paths, electroplating performance is evaluated, and evaluation results are generated. A reasoning engine based on a large language model is constructed, and the reasoning engine is used to reason about the evaluation results to generate specific solutions for maintenance and process compensation.

[0022] It should be noted that this solution constructs a dynamic causal graph, combines Granger causality tests and graph attention networks to automatically identify complex nonlinear process causal chains, and leverages the text reasoning capabilities of a large language model to transform causal paths into actionable operational suggestions for process instructors and automated control equipment. This significantly reduces the complexity and response time of expert diagnosis of electroplating quality problems, achieving dynamic multi-source perception and semantic machine reading comprehension-based online causal reasoning. This enables rapid response and strategy adjustment, reducing the defect rate and improving electroplating quality.

[0023] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, electroplating process entity nodes and process parameter attribute nodes are constructed, specifically as follows: Define the process entity of PCB electroplating process, and divide the process entity of PCB electroplating process into process equipment layer, process medium layer, additive layer and process index layer. Electroplating process entity nodes are constructed based on the process equipment layer, and process parameter attribute nodes are constructed based on the process medium layer, additive layer, and process index layer. The process equipment layer consists of physical hardware, including the main electroplating tank, circulating pump, filter, heat exchanger, rectifier, automatic dosing pump, and stirring device; the process medium layer includes Cu... 2+ Concentration, Ni 2+ Concentration, pH value, temperature, conductivity, dissolved oxygen concentration, and total organic carbon; Additives include inhibitors, accelerators, leveling agents, and chloride ions; process indicators include deposition rate, coating thickness uniformity, concentration of additive degradation byproducts, and bath viscosity.

[0024] It should be noted that the direct causal effects include increased temperature leading to a faster deposition rate; inhibitory effects include increased inhibitor concentration leading to a slower deposition rate; accelerating effects include the synergistic effect of chloride ions as accelerators; and degradation-generative effects include the decomposition of additives producing byproducts. Since the physical nodes and process parameter attribute nodes of the electroplating process affect the performance of the electroplating solution, thus influencing its health status, this method can construct a directional anisotropy graph to monitor and evaluate the health status of the electroplating solution.

[0025] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a directed attribute graph structure is constructed based on the entity nodes of the electroplating process and the attribute nodes of the process parameters to form a dynamic knowledge graph, specifically: Based on electrochemical theory and historical process documentation, direct causal relationships, inhibitory relationships, promoting relationships, and degradation-genesis relationships were obtained. Edge types are constructed based on direct causal relationships, inhibitory relationships, promoting relationships, and degradation-generation relationships; Construct a dynamic knowledge graph, and build a directed attribute graph structure based on edge type, electroplating process entity nodes, and process parameter attribute nodes; The directed attribute graph structure is input into the dynamic knowledge graph for knowledge representation.

[0026] It should be noted that this method can construct a dynamic knowledge graph to represent the relationship between electroplating process entity nodes and process parameter attribute nodes, providing a theoretical basis for causal reasoning.

[0027] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a dynamic causal intensity matrix is ​​constructed for the time series data of each sensor, specifically as follows: The Granger causality test algorithm is introduced, and sensor data information collected by the sensor at each time stamp is obtained. Sensor time series data is constructed based on the sensor data information collected by the sensor at each time stamp. The time series data of each sensor is input into the Granger causality test algorithm, and the Granger causality test algorithm is used to calculate the causal strength value of the two variables within a continuous sliding time window. The causal intensity values ​​of the two variables within a continuous sliding time window are smoothed by an exponentially weighted moving average, and the causal intensity is used to capture the delayed response and asymmetric causal relationship between different processes. A causal intensity matrix is ​​constructed based on the causal intensity values ​​of the two variables within a continuous sliding time window, and the causal intensity matrix is ​​updated incrementally in real time using a sliding time window.

[0028] It should be noted that by continuously calculating the causal strength values ​​of the two variables within a continuous sliding time window, and evaluating the dynamic probabilistic causal strength in real time, this design can capture delayed responses and asymmetric causal relationships between different processes, solving the problem of lagged attribution bias caused by fixed static causal or simple correlation analysis. The dynamic causal strength matrix is ​​updated incrementally in real time using a sliding time window, without relying on offline training. It is fully applicable to the slow drift of electroplating process parameters in the field, solving the major pain point of traditional causal detection models requiring retraining after parameter changes due to bath aging / equipment wear. Two variables refer to two different electroplating process parameters or state variables, such as Cu... 2+ Concentration and deposition rate, temperature and Cu 2+ Concentration, accelerator (SPS) concentration and additive by-product generation, pH value and leveling agent adsorption efficiency, bath age and total organic carbon (TOC) accumulation.

[0029] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a multi-hop graph attention network is used to model the dynamic causal intensity matrix, specifically as follows: A multi-hop graph attention network is introduced, and a dynamic causal intensity matrix is ​​input into the multi-hop graph attention network. For each node and its neighboring nodes within a preset range, the attention coefficient is calculated using single-head attention and then normalized. After normalization calculation, the node features are updated. A multi-hop graph attention network is used to stack a preset number of layers. Each layer performs attention aggregation on the node features of the previous layer and calculates the credibility of the long-distance path. Based on the credibility of long-range paths, a reverse search is performed starting from the abnormal node. At each step, a part of the path is retained and sorted according to the current cumulative confidence. When the path length reaches the preset length or the cumulative confidence is lower than the preset threshold, the expansion stops. All long-range paths are counted, and a set of candidate causal paths is constructed based on these long-range paths. This set is then used as the output of the dynamic knowledge graph, thus completing the modeling process.

[0030] It should be noted that a multi-layer graph attention network is constructed by using the dynamic causal intensity matrix as the edge weight input, aggregating node feature information layer by layer to form a learnable attention coefficient weight distribution. The long-range dependency propagation strategy in graph neural networks is adopted, using a K-hop neighborhood update aggregation message passing mechanism to comprehensively consider the causal propagation of parameter nodes across time spans in the process flow. This method uses the confidence of long-range paths to perform a reverse search starting from the anomalous node, retaining a portion of the path at each step, thus achieving gradual energy attenuation of the source node's anomalous impact during propagation to the target node. This design effectively filters out pseudo-causal error paths formed by random jitter or noise during production, ultimately outputting a set of high-confidence candidate causal paths that can be directly used for inference in large language models, addressing the deficiency in existing knowledge graphs regarding the lack of reliable attenuation suppression of noise and interference paths. This provides a theoretical basis for monitoring the performance of electroplating solutions.

[0031] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a set of candidate causal paths is output using a dynamic knowledge graph, specifically: Acquire sensor time series data within a preset time period, and input the sensor time series data within the preset time period into a dynamic knowledge graph for knowledge matching; By matching knowledge, a set of candidate causal paths is output using a dynamic knowledge graph and displayed in a preset manner.

[0032] It should be noted that the set of candidate causal paths, such as "dosing pump wear → actual dosing deviation → Cu", is not applicable. 2+ Concentration fluctuations → additive ratio imbalance → rough coating.

[0033] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a set of candidate causal paths is output using a dynamic knowledge graph to evaluate electroplating performance and generate evaluation results, specifically: Set the range of electroplating performance evaluation index data, and determine whether the set of candidate causal paths is within the range of electroplating performance evaluation index data. For example: a set of candidate causal paths such as "dosing pump wear → actual dosage deviation → Cu" 2+ "Concentration fluctuation → additive ratio imbalance → coating roughness." The ratio imbalance has an evaluation threshold range, and roughness also has a threshold range. When a certain indicator of the candidate causal path set (such as ratio imbalance or coating roughness) is within the range of data for electroplating performance evaluation indicators, a normal electroplating evaluation result is generated. Conversely, if there is an anomaly in one of the data, an abnormal electroplating evaluation result will be generated.

[0034] When the set of candidate causal paths falls within the range of electroplating performance evaluation index data, a normal electroplating evaluation result is generated. When the set of candidate causal paths is not within the range of data for electroplating performance evaluation indicators, an evaluation result for electroplating anomalies is generated.

[0035] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, a reasoning engine based on a large language model is constructed, specifically as follows: The sensor time series data is normalized, and an embedding vector for each parameter is generated using a pre-trained Transformer time series encoder. After positional encoding of all parameter embeddings, they are concatenated into a global embedding vector, which is then compressed to a fixed dimension by linear projection and input into the large language model. The output candidate path set is combined with the global embedding vector to calculate the path matching score. After the causal path is sorted on the candidate set by the semantic vector, the path with the highest matching score in the candidate path set is selected and input into the large language model. The path with the highest matching score in the candidate path set and the current semantic context are fed into the context retrieval enhancement generation module of the large language model, and finally the large language model is used to output the result in a chain-like reasoning manner.

[0036] It should be noted that the Large Language Model Context Retrieval Enhancement Generation (RAG) module includes vector data, and the vector database stores historical successful diagnostic cases, chemical technical manuals, equipment operation SOPs, etc. The trend semantic encoder uses a pre-trained Transformer temporal encoding module to encode the collected real-time trend patterns of multiple parameters in the electroplating solution, transforming them into a semantic feature vector embedding layer suitable for LLM embedding. After the causal paths are sorted on the candidate set by semantic vectors, the optimal path and the current semantic context are fed into the Large Language Model Context Retrieval Enhancement Generation (RAG) module, which is loaded with chemical process optimization constraints. The Large Language Model outputs in a chain-like reasoning manner: for example, since [root cause parameter A] experiences [drift / mutation / impulse] at time T, [causal path B] causes [abnormal parameter C] to exceed the limit, it is recommended to adjust using [causal path A].

[0037] Furthermore, in the method for monitoring the health status of PCB electroplating solutions and evaluating electroplating performance, an inference engine is used to infer the evaluation results and generate specific solutions for maintenance and process compensation, specifically as follows: When the evaluation result is an evaluation result of electroplating abnormality, the sensor time series data within the preset time and the candidate path set are input into the inference engine for inference. By leveraging a large language model inference engine and ontology knowledge in the field of electroplating chemistry, complex causal chains are transformed into actionable work order instructions, generating specific solutions for maintenance and process compensation, and supplementing them with optimization suggestion text described in natural language.

[0038] It should be noted that the large language model inference engine, based on ontology knowledge in the electroplating chemistry field, transforms complex causal chains into actionable work order instructions, generates specific solutions for maintenance and process compensation, and supplements them with optimization suggestions described in natural language. This significantly reduces the analysis and judgment time for on-site engineers and improves operational efficiency. The model can also be continuously and adaptively updated. Through a continuous model update mechanism based on newly added causal data samples during equipment maintenance, a small number of new causal evidences with high confidence are selected to update the graph attention network weights, and the trend encoder is fine-tuned periodically. Simultaneously, the closed-loop evaluation (confirmation / denial) of the accuracy of anomaly attribution based on human feedback is recorded in the historical knowledge graph, forming a dynamically evolving knowledge accumulation that continuously improves the accuracy of future anomaly inferences.

[0039] In summary, this solution constructs a dynamic causal graph, combines Granger causality tests and graph attention networks to automatically identify complex nonlinear process causal chains, and leverages the text reasoning capabilities of a large language model to transform causal paths into actionable operational suggestions for process instructors and automated control equipment. It significantly reduces the complexity and response time of expert diagnosis of electroplating quality problems, achieving online causal reasoning from dynamic multi-source perception to semantic machine reading comprehension.

[0040] In addition, this method also includes: The current multimodal health status of the electroplating solution is obtained, including the concentration of each ion in the tank, the concentration of additives, temperature, pH and conductivity, and normalized high-dimensional vector representation. A diffusion generation model is constructed based on a deep neural network, and the distribution characteristics of the electroplating solution over time in historical working condition data are learned using the diffusion generation model. The diffusion model includes a forward diffusion process and a reverse diffusion process. In the forward diffusion process, noise is gradually added to the data until it completely degenerates into Gaussian noise. In the reverse diffusion process, the neural network is trained to iteratively remove noise from pure noise samples, reconstruct and generate the probability distribution of state trajectories under different intervention strategies in the future. After obtaining the real-time sensor health status, the trained diffusion model is used to quickly and efficiently predict the distribution of electroplating solution health status within a preset time under a given combination of candidate control commands. A quantum-inspired particle swarm optimization algorithm is introduced, using the predicted state given by the diffusion generation model as the input of the fitness function to construct the objective function of the optimization problem. The objective function minimizes the error norm between the controlled end-stage health index and the target health index, while optimizing the minimum energy / material total cost constraint of the control process. The output decision variables are the unit time addition rate of each agent, stirring speed, and temperature correction control parameters.

[0041] It should be noted that after obtaining the real-time sensor health status, the trained diffusion model is used to quickly and efficiently predict the distribution of the electroplating solution's health status at time H under a given combination of candidate control commands. The fundamental difference from traditional time series prediction (single-point prediction) is that the diffusion generation model outputs a probability distribution of the health status, which includes the probability of future multi-branch evolution. This distribution is particularly suitable for providing quantitative analysis of uncertainty margins for downstream decision optimization, enabling multi-scenario risk assessment in online optimal decision-making. Furthermore, the quantum-inspired particle swarm optimization algorithm is based on the concepts of quantum superposition and quantum potential wells, allowing the updated position of particles to be determined by a probability distribution function, fundamentally overcoming the weakness of traditional particle swarm optimization in getting trapped in local optima. The quantum potential well center point determination strategy of the quantum-inspired particle swarm optimization algorithm adapts to the situation of each particle swarm. The population size is flexibly adjustable, exhibiting superior exploration and distribution convergence capabilities compared to standard PSO in multi-objective strongly coupled optimization scenarios.

[0042] It should be noted that this method innovatively introduces a diffusion-generative model to describe the multi-branch state evolution distribution of nonlinear electroplating solutions under controlled and uncontrolled conditions, and combines it with QPSO to achieve dynamic solution of the optimal control quantity under high-dimensional constraints. This mechanism significantly improves the convergence speed and robustness of electroplating solutions under strongly coupled and complex operating conditions, and is particularly effective in the early correction and control of complex health problems such as additive ratio imbalance and accelerated aging at the end of the bath life.

[0043] For example, this embodiment adds an advanced control step. Once the engineer confirms the "process compensation" scheme output by the LLM, the system uses the instruction (e.g., "inhibitor acceleration rate increased by 1.2 times, duration 30 minutes") as a candidate control command. The pre-trained diffusion generation model receives the current health state vector and predicts the probability distribution of the electroplating solution's health state within 30 minutes after executing the command (e.g., 85% probability of recovery to normal, 10% probability of overcompensation, 5% probability of no response). The quantum-inspired particle swarm optimization algorithm uses this probability distribution as its fitness function and eventually converges to the optimal solution: suggesting "inhibitor rate increased by 1.1 times, duration 35 minutes, and rectifier current density increased by 2%" to balance the recovery effect and material cost. The system then sends the optimized final solution to the automatic dosing pump and rectifier for execution.

[0044] In addition, this method also includes: Multiple corrosion-resistant single-mode sensing optical fibers with polyimide or metal coatings are laid out inside the PCB electroplating tank along different depths and elevations. The temperature field data and indirect data of the flow field / stirring field of the electroplating solution are obtained by interactive measurement using a Brillouin optical time domain reflectometer and a Brillouin optical frequency domain analyzer. It should be noted that the temperature field data is affected by localized temperature differences in the electroplating bath caused by the thermal effects of current, the heat of chemical reactions, and external heating / cooling cycles, leading to uneven convective mass transfer in the flow field. The Brillouin frequency shift exhibits a highly linear relationship with temperature changes, enabling millimeter-level spatial resolution and a temperature detection accuracy of ±0.1℃.

[0045] Indirect data of flow field / stirring field: Using fiber-optic coupled vibration sensors, the local electroplating solution flow pattern and disturbance information are inferred from the Brillouin frequency shift signal, and mapped synchronously with the temperature field to completely reconstruct the dynamic diagram of dimensional thermal-fluid coupling within the tank.

[0046] Based on the temperature field data and indirect data of flow field / stirring field of the electroplating solution, the high current heating and bubble accumulation that often occur in the edge area are located and identified. A transparent observation window is set in the electroplating tank, and the optical probe of the digital coaxial holographic microscope is aimed at the vicinity of the surface of the cathode plate to be tested. Among them, the deployment of multi-layered spiral path optical fibers can reconstruct the internal temperature field and flow field anomalies of the plating solution in 3D all-time, accurately locate hot spots such as high current heating and bubble accumulation that often occur in the edge area, and identify health dead zones that cannot be detected by traditional discrete single-point monitoring.

[0047] The interference hologram of the object light wave and the reference light is collected by the charge coupler. The high-resolution amplitude and phase images are recovered by the digital refocusing algorithm. Based on the interference hologram of the object light wave and the reference light and the high-resolution amplitude and phase images, the cathode near-wall ion concentration field change / deposition morphology image is obtained. It should be noted that the high-resolution amplitude and phase images are recovered using a digital refocusing algorithm, and the phase information is linearly related to the optical path length (ion concentration × thickness). Specifically, based on the interference record hologram of the light wave and the reference light, the high-resolution amplitude and phase images can simultaneously obtain: the changes in the metal ion concentration gradient near the cathode plate and the dynamic adsorption mode of local additives (leveling agents, inhibitors, brighteners), as well as the dynamic evolution of the two-dimensional instantaneous morphology of the electrodeposition. Specifically, regarding the dynamic evolution of the two-dimensional instantaneous morphology of the electrodeposition: when the holographic reconstructed image exhibits characteristics such as irregular edge deposition, plating inclusions, and needle-like growth, the tendency for plating degradation caused by severe electroplating solution aging can be identified in advance, without waiting for destructive quality inspection in the later stages of PCB development.

[0048] The real-time thermal field and rheological characteristics of the electroplating solution obtained by fiber-optic distributed Brillouin thermometry and the deposition morphology images obtained by digital holographic microscopy were converted and unified into time series and two-dimensional normalized tensors, respectively. A joint assessment of the electroplating solution's health status and standardized tensor is conducted to pinpoint potential sources of failure, issuing tiered warnings and precise maintenance recommendations. Sources of failure include solutions degradation and localized nozzle blockage.

[0049] It should be noted that this solution solves the core problem of existing probe electrode contact sensors being prone to contamination and failure by using non-invasive physical optics in-situ methods of fiber optic BOTDR / A and DHM. It can continuously acquire multimodal differential information on global and local chemical dynamics of the plating solution in real time, improve the depth of full inspection and microscopic insight of PCB plating solution health monitoring, and greatly enhance the early warning capability for defects that are difficult to trace to their source in the electroplating process.

[0050] One specific implementation process is as follows: Data preparation: On the PCB horizontal electroplating line, Cu data is collected in real time using online sensors. 2+ Time-series data including concentration, temperature, inhibitor concentration, and deposition rate. Simultaneously, based on electrochemical knowledge, entity nodes (such as "rectifier" and "circulation pump") and attribute nodes ("Cu") are predefined. 2+"Concentration", "coating thickness uniformity", and establish edge relationships (such as "temperature -> deposition rate" as a promoting relationship).

[0051] Causal path discovery: Assume that in the t-th time window, the system detects an anomaly in the "coating roughness" index. The system calls the Granger causality test to calculate the causal strength of all variables on "coating roughness" over the past N time windows, forming a dynamic matrix. This matrix is ​​input into a two-layer graph attention network. The first layer aggregates information from direct neighbors (such as "additive ratio"), and the second layer aggregates information from second-order neighbors (such as "dosing pump wear"). The calculated confidence score of the long-range path from "dosing pump wear" to "coating roughness" is 0.92, which is higher than the threshold of 0.75. The reverse search preserves this path: "dosing pump wear → actual dosage deviation → Cu". 2+ Concentration fluctuations → additive ratio imbalance → rough coating.

[0052] Performance evaluation: The system determines that the "additive ratio imbalance" index in this path has exceeded the preset normal range threshold, so the system generates an evaluation result of "electroplating abnormality".

[0053] LLM Inference and Solution Output: The system feeds the above causal path, the raw data of the most recent hour, and the "dosing pump maintenance manual" and "additive compensation formula" retrieved from the vector database into the fine-tuned large language model.

[0054] LLM outputs the following text: Diagnostic conclusion: Impeller wear was detected in the main dosing pump (ID: P-102), causing periodic fluctuations in the actual inhibitor dosage (evidence), which in turn triggered Cu... 2+ An imbalance between the concentration and the ratio of inhibitors ultimately leads to excessive coating roughness (abnormality).

[0055] Repair solution: Immediately inspect and replace the impeller seal of dosing pump P-102.

[0056] Process compensation: After the repair is completed, it is recommended to add the inhibitor at a rate of 1.2 times according to the formula table C-3 for 30 minutes, and at the same time increase the cathode swing amplitude by 10% to promote the homogenization of the diffusion layer.

[0057] like Figure 2 As shown, the second aspect of the present invention provides a PCB electroplating solution health status monitoring and electroplating performance evaluation system, including a memory and a processor. The memory includes a PCB electroplating solution health status monitoring and electroplating performance evaluation method program. When the PCB electroplating solution health status monitoring and electroplating performance evaluation method program is executed by the processor, it implements the steps of any one of the PCB electroplating solution health status monitoring and electroplating performance evaluation methods.

[0058] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0059] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0060] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0061] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0062] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0063] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for monitoring the health status of PCB electroplating solutions and evaluating their electroplating performance, characterized in that, Includes the following steps: Construct electroplating process entity nodes and process parameter attribute nodes, and construct a directed attribute graph structure based on the electroplating process entity nodes and process parameter attribute nodes to form a dynamic knowledge graph. A dynamic causal intensity matrix is ​​constructed for each sensor time series data, and a multi-hop graph attention network is used to model the dynamic causal intensity matrix. A set of candidate causal paths is output using a dynamic knowledge graph. The electroplating performance is evaluated by combining the candidate causal path set output from the dynamic knowledge graph, and the evaluation results are generated. A reasoning engine based on a large language model is constructed, and the reasoning engine is used to reason about the evaluation results to generate specific solutions for maintenance and process compensation.

2. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, Construct the electroplating process entity node and process parameter attribute node as follows: Define the process entity of PCB electroplating process, and divide the process entity of PCB electroplating process into process equipment layer, process medium layer, additive layer and process index layer. Electroplating process entity nodes are constructed based on the process equipment layer, and process parameter attribute nodes are constructed based on the process medium layer, additive layer, and process index layer. The process equipment layer is a hardware entity, including a main electroplating tank, a circulating pump, a filter, a heat exchanger, a rectifier, an automatic dosing pump, and a stirring device; The process medium layer includes Cu 2+ Concentration, Ni 2+ Concentration, pH value, temperature, conductivity, dissolved oxygen concentration, and total organic carbon; The additives include inhibitors, accelerators, leveling agents, and chloride ions; the process indicator layers include deposition rate, coating thickness uniformity, concentration of additive degradation byproducts, and bath viscosity.

3. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, Based on the electroplating process entity nodes and process parameter attribute nodes, a directed attribute graph structure is constructed to form a dynamic knowledge graph, specifically: Based on electrochemical theory and historical process documents, direct causal relationships, inhibitory relationships, promoting relationships, and degradation-genesis relationships were obtained; Edge types are constructed based on the aforementioned direct causal relationships, inhibition relationships, promotion relationships, and degradation-generation relationships; A dynamic knowledge graph is constructed, and a directed attribute graph structure is built based on the edge type, electroplating process entity nodes, and process parameter attribute nodes. The directed attribute graph structure is input into the dynamic knowledge graph for knowledge representation.

4. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, A dynamic causal intensity matrix is ​​constructed for each sensor's time series data, specifically as follows: The Granger causality test algorithm is introduced, and sensor data information collected by the sensor at each time stamp is obtained. Sensor time series data is constructed based on the sensor data information collected by the sensor at each time stamp. The time series data of each sensor is input into the Granger causality test algorithm, and the Granger causality test algorithm is used to calculate the causal strength value of the bivariate within a continuous sliding time window. The causal intensity values ​​of the two variables within a continuous sliding time window are smoothed by an exponentially weighted moving average, and the causal intensity is used to capture the delayed response and asymmetric causal relationship between different processes. A causal intensity matrix is ​​constructed based on the causal intensity values ​​of the two variables within a continuous sliding time window, and the causal intensity matrix is ​​updated incrementally in real time using a sliding time window.

5. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, The dynamic causal intensity matrix is ​​modeled using a multi-hop graph attention network, specifically as follows: A multi-hop graph attention network is introduced, and the dynamic causal intensity matrix is ​​input into the multi-hop graph attention network. For each node and its neighboring nodes within a preset range, the attention coefficient is calculated using single-head attention and then normalized. After normalization calculation, the node features are updated. A multi-hop graph attention network is used to stack a preset number of layers. Each layer performs attention aggregation on the node features of the previous layer and calculates the credibility of the long-distance path. Based on the credibility of the long-range path, a reverse search is performed starting from the abnormal node. At each step, a part of the path is retained and sorted according to the current cumulative confidence. When the path length reaches a preset length or the cumulative confidence is lower than a preset threshold, the expansion stops. All long-range paths are counted, and a set of candidate causal paths is constructed based on these long-range paths. This set is then used as the output of the dynamic knowledge graph, thus completing the modeling process.

6. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, The dynamic knowledge graph is used to output a set of candidate causal paths, specifically: Acquire sensor time series data within a preset time period, and input the sensor time series data within the preset time period into the dynamic knowledge graph for knowledge matching; By matching knowledge, a set of candidate causal paths is output using a dynamic knowledge graph and displayed in a preset manner.

7. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, Electroplating performance is evaluated by combining the candidate causal path set output from the dynamic knowledge graph, and the evaluation results are generated, specifically as follows: Set the range of electroplating performance evaluation index data, and determine whether the set of candidate causal paths is within the range of electroplating performance evaluation index data; When the set of candidate causal paths falls within the range of electroplating performance evaluation index data, an evaluation result indicating normal electroplating is generated. When the set of candidate causal paths is not within the range of electroplating performance evaluation index data, an evaluation result for electroplating anomaly is generated.

8. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, The inference engine based on a large language model is constructed as follows: The sensor time series data is normalized, and an embedding vector for each parameter is generated using a pre-trained Transformer time series encoder. After positional encoding of all parameter embeddings, they are concatenated into a global embedding vector, which is then compressed to a fixed dimension by linear projection and input into the large language model. The output candidate path set is combined with the global embedding vector to calculate the path matching score. After the causal path is sorted on the candidate set by the semantic vector, the path with the highest matching score in the candidate path set is selected and input into the large language model. The path with the highest matching score in the candidate path set and the current semantic context are fed into the context retrieval enhancement generation module of the large language model, and finally the large language model is used to output the result in a chain-like reasoning manner.

9. The method for monitoring the health status and evaluating the electroplating performance of PCB electroplating solution according to claim 1, characterized in that, The inference engine is used to reason about the evaluation results to generate specific solutions for maintenance and process compensation, specifically: When the evaluation result is an evaluation result of electroplating abnormality, the sensor time series data within a preset time and the candidate path set are input into the inference engine for inference. By leveraging a large language model inference engine and ontology knowledge in the field of electroplating chemistry, complex causal chains are transformed into actionable work order instructions, generating specific solutions for maintenance and process compensation, and supplementing them with optimization suggestion text described in natural language.

10. A PCB electroplating solution health status monitoring and electroplating performance evaluation system, characterized in that, The device includes a memory and a processor. The memory includes a program for monitoring the health status of PCB electroplating solution and evaluating its electroplating performance. When the processor executes the program for monitoring the health status of PCB electroplating solution and evaluating its electroplating performance, it implements the steps of the method for monitoring the health status of PCB electroplating solution and evaluating its electroplating performance as described in any one of claims 1-9.

Images