Adaptive control method for pvd and micro-arc oxidation equipment and electronic equipment thereof

Abstract

This application discloses an adaptive control method and electronic device for PVD and micro-arc oxidation equipment. The method includes: constructing a parameter inversion model characterized by a forward mapping relationship between prior distributions of process parameters and physical constraints; establishing a causal topology based on the mapping relationship, and identifying anomaly types by calculating the joint probability density of the two hypotheses; executing different strategies according to the anomaly type: shielding faulty sensors and reconstructing signals when measurement anomalies occur, and calculating parameter corrections through online optimization when mechanistic anomalies occur; performing safety verification and smoothing filtering on the corrections based on safety constraint envelopes; and generating an interpretable report containing root causes, contribution levels, and maintenance guidelines. This application achieves accurate anomaly type differentiation and robust adaptive control through physical-data decoupling modeling and causal reasoning, significantly improving the stability, safety, and interpretability of processes in complex industrial environments.

Description

Technical Field

[0001] This application relates to the field of high-end surface treatment process control technology, specifically to an adaptive control method and electronic equipment for PVD and micro-arc oxidation equipment. Background Technology

[0002] In processes such as vacuum magnetron sputtering (PVD) and micro-arc oxidation, precise closed-loop control of parameters such as power supply, gas, and temperature is required to prepare high-performance films. Existing technologies monitor data through sensors and compare it with preset values; once a deviation is detected, the output is adjusted via a controller.

[0003] However, this control mode has a fundamental flaw: the system cannot distinguish the true cause of the monitoring data deviation. This deviation may originate from two completely different situations: 1) actual process drift (such as target etching, electrolyte aging), in which case parameter adjustment is necessary for compensation; 2) sensor malfunction (such as probe contamination, reading drift), in which case parameter adjustment is absolutely prohibited, otherwise it will disrupt the normal process.

[0004] Traditional PID or neural network controllers, lacking an understanding of physical causal relationships, are unable to identify the above differences. They often misjudge sensor malfunctions as process fluctuations, thus executing incorrect compensations, leading to production disruptions, film scrap, or even equipment downtime. Summary of the Invention

[0005] This application aims to overcome the shortcomings of the prior art and provide an adaptive control method and electronic device for PVD and micro-arc oxidation equipment, so as to solve the technical problem that existing PVD and micro-arc oxidation equipment cannot automatically distinguish between sensor measurement failures and actual process drift during operation, resulting in miscompensation and process runaway.

[0006] In a first aspect, this application provides an adaptive control method for PVD and micro-arc oxidation equipment, applied to PVD and micro-arc oxidation equipment, comprising the following steps: S1: Construct a physical-data decoupled process parameter inversion model. The model is characterized by the prior distribution of process parameters obtained by training based on historical operating data and the positive mapping relationship from process parameters to membrane performance obtained by training based on physical equation constraints, so as to invert process parameters according to target membrane performance. S2: Based on the physical associations in the positive mapping relationship, establish a causal topology structure for the device operating status; based on this causal topology structure, identify the anomaly type of the real-time monitoring data to obtain the identification results of the indication measurement anomaly or mechanism anomaly. S3: Execute the control strategy based on the identification result: if it is a measurement abnormality, then shield the monitoring data of the corresponding sensor and reconstruct the control signal based on the data of the remaining sensors; if it is a mechanism abnormality, then calculate the correction amount of the control parameters to compensate for process drift through an online optimization algorithm. S4: Perform safety verification and instruction filtering on the control parameter correction amount, and smoothly truncate or correct the out-of-bounds instruction according to the predefined equipment safety operation constraint envelope to generate the final execution instruction. S5: Based on the identification results, execution control strategy, and final execution instructions, generate an explanatory report that includes an explanation of the root cause of the anomaly, the contribution of parameter adjustment, and maintenance guidelines.

[0007] The construction of the process parameter inversion model includes the following steps: extracting multidimensional time-series process parameters from historical operating data, and obtaining the prior distribution of process parameters based on an unconditional diffusion model; obtaining paired data of process parameters and membrane performance, and obtaining a positive mapping relationship based on a physical information neural network, wherein the loss function of the physical information neural network includes residual terms of physical equations; and using a sequential Monte Carlo sampling algorithm, combined with the prior distribution of process parameters and the positive mapping relationship, to achieve the inversion calculation from target membrane performance to process parameters.

[0008] The specific steps for establishing the causal topology of the device operating state in S2 include: defining plasma density, actual etching depth of the target surface, and actual surface temperature of the substrate as first-type nodes; defining emission spectrometer light intensity reading, power supply voltage value, and infrared thermometer reading as second-type nodes; and establishing physical evolution edges between first-type nodes and measurement response edges from first-type nodes to second-type nodes based on plasma physics and thermodynamics principles.

[0009] The step S2, which identifies anomalies in real-time monitoring data to obtain identification results indicating measurement anomalies or mechanism anomalies, includes: for each second-type node, training a noise-resistant benchmark regression model based on historical normal data and constructing a probability distribution model of the prediction residual; when a deviation occurs in the monitoring data, calculating the joint probability density value of the deviation conforming to the physical linkage law and the joint probability density value conforming to the independent sensor fault, respectively; and outputting the identification result of measurement anomalies or mechanism anomalies by comparing the magnitude of the joint probability density value conforming to the physical linkage law and the joint probability density value conforming to the independent sensor fault.

[0010] The step in S3, which calculates the control parameter correction amount for compensating process drift using an online optimization algorithm, includes: constructing a time-varying loss function containing the current membrane performance deviation term and the control parameter adjustment magnitude penalty term; estimating the performance gradient direction using the finite difference method based on the correspondence between historical process parameters and membrane performance; and iteratively calculating the control parameter correction amount using a gradient descent algorithm based on the performance gradient direction and a preset learning rate.

[0011] Furthermore, S3 also includes an interference observation compensation step: calculating the difference between the actual film performance and the predicted performance based on the positive mapping relationship as a lumped interference estimate; and superimposing a feedforward compensation amount in the opposite direction to the lumped interference estimate into the control parameter correction amount.

[0012] The step in S4 of smoothly truncating or correcting the out-of-bounds instruction based on the predefined equipment safety operation constraint envelope includes: defining a safety operation constraint envelope in a multi-dimensional control parameter space based on equipment safety operation parameter constraints; inputting the control parameter correction amount into the equipment dynamic model for virtual simulation to predict the process state after execution; if the predicted state exceeds the safety operation constraint envelope, finding the optimal correction point within the safety boundary through a constraint optimization algorithm; and calculating a smooth transition instruction based on the optimal correction point to generate the final execution instruction.

[0013] Step S5 includes: capturing a state snapshot containing the current membrane quality deviation value, real-time readings of each sensor, and control command increments; calculating the marginal contribution value of each control parameter adjustment to the membrane performance deviation correction using a pre-trained contribution analysis model, and normalizing the marginal contribution value into a contribution weight; generating a natural language explanation text containing anomaly root cause descriptions and parameter adjustment contribution values ​​based on the anomaly type in the identification results, combined with the contribution weights and a physical semantic mapping table; when the identification results indicate a measurement anomaly, extracting standard maintenance actions and their standard time consumption from the maintenance action database according to the corresponding sensor type, accumulating them to obtain the estimated maintenance time, and adding the estimated maintenance time to the explanation report as maintenance guidance.

[0014] In a second aspect, this application provides an electronic device including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first aspect.

[0015] The beneficial effects of this application, compared to existing technologies, include an adaptive control method and electronic device for PVD and micro-arc oxidation equipment. The method includes: constructing a parameter inversion model characterized by a positive mapping relationship between prior distributions of process parameters and physical constraints; establishing a causal topology based on the mapping relationship; identifying anomaly types by calculating the joint probability density of the two hypotheses; executing different strategies according to the anomaly type: shielding faulty sensors and reconstructing signals when measurement anomalies occur; calculating parameter correction amounts through online optimization when mechanistic anomalies occur; performing safety verification and smoothing filtering on the correction amounts based on safety constraint envelopes; and generating an interpretable report containing root causes, contribution levels, and maintenance guidelines. This application, through physical-data decoupling modeling and causal reasoning, achieves accurate differentiation of anomaly types and robust adaptive control, significantly improving the stability, safety, and interpretability of processes in complex industrial environments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an adaptive control method for PVD and micro-arc oxidation equipment according to an embodiment of this application.

[0018] Figure 2 This is a flowchart illustrating step S1 of an adaptive control method for PVD and micro-arc oxidation equipment provided in an embodiment of this application.

[0019] Figure 3 This is a flowchart illustrating step S5 of an adaptive control method for PVD and micro-arc oxidation equipment provided in an embodiment of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0023] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0024] In high-end surface treatment fields such as vacuum coating and micro-arc oxidation, the stability and repeatability of the process directly determine the quality and performance of the final product. However, existing control technologies face a long-standing challenge: when monitoring data shows anomalies, it is impossible to accurately distinguish whether this is due to drift in the actual process parameters inside the equipment (such as target material wear and electrolyte aging) or simply a "false alarm" caused by sensor malfunction (such as probe contamination and reading drift). This lack of differentiation ability often leads to "misjudgment and misoperation" of the control system, either incorrectly adjusting the process when the sensor malfunctions, thus disrupting normal production, or failing to compensate in time when the actual process drifts, resulting in a decline in film quality. More seriously, blind adjustments may trigger the equipment's safety interlocks, leading to unplanned shutdowns. Therefore, this application provides an adaptive control method and electronic device for PVD and micro-arc oxidation equipment to solve the above problems.

[0025] See Figure 1 , Figure 1 This is a flowchart illustrating an adaptive control method for PVD and micro-arc oxidation equipment according to an embodiment of this application.

[0026] The method is applied to PVD and micro-arc oxidation equipment, and includes the following steps: S1: Construct a physical-data decoupled process parameter inversion model. The model is characterized by the prior distribution of process parameters obtained by training based on historical operating data and the positive mapping relationship from process parameters to membrane performance obtained by training based on physical equation constraints, so as to invert process parameters according to target membrane performance.

[0027] This step aims to build the core "prior knowledge base" for the entire control method—a process parameter inversion model capable of understanding the operating laws of equipment and physical causal relationships—providing a physical cognitive foundation for subsequent real-time control. (See also...) Figure 2 Step S1 specifically includes the following sub-steps: S11: Extract multi-dimensional time-series process parameters from historical operating data, and obtain the prior distribution of process parameters based on training an unconditional diffusion model; S12: Obtain paired data of process parameters and film performance, and obtain a positive mapping relationship based on physical information neural network training. The loss function of the physical information neural network includes the residual term of the physical equation. S13: By using the sequential Monte Carlo sampling algorithm, combined with the prior distribution of the process parameters and the forward mapping relationship, the inversion calculation from the target film performance to the process parameters can be realized.

[0028] Specifically, its implementation involves two parallel data-driven modeling processes: First, based on the large amount of low-cost historical process parameter logs (such as time-series data on target power, gas flow rate, and vacuum level) automatically generated during the long-term operation of PVD or micro-arc oxidation equipment, this method employs an unconditional diffusion model based on iterative denoising principles for training. The training process involves progressively adding Gaussian noise to the real parameter samples until they are completely disordered, then training the neural network to predict the noise component at each step in reverse, thereby enabling the model to grasp the prior distribution patterns of the equipment's process parameters, i.e., assessing the reasonable probability of any given set of parameters in the physical operating logic. Second, to establish a deterministic mapping from "process parameters" to "film performance," this method acquires a small amount of high-cost paired experimental data (i.e., specific process parameter settings and their resulting film performance indicators such as hardness and refractive index) to train a physical information neural network (a high-dimensional regression network). The key innovation of this network lies in the additional physical equation residual penalty term introduced into its training loss function: in each training iteration, the network's predicted output is substituted into the partial differential equation or empirical formula (such as the magnetron sputtering deposition rate formula) describing the process. If the prediction result violates physical laws (large equation residual), a large penalty value is generated, forcing the network weights to be adjusted. This allows the network to generalize according to physical principles even when data is scarce, avoiding errors that violate common sense. Ultimately, these two models—the prior distribution model representing the rationality of the parameters and the forward mapping network representing physical causality—together constitute the process parameter inversion model. This model achieves the inversion function through iterative optimization strategies such as sequential Monte Carlo sampling: when the target film performance is input, the sampling algorithm starts from random initial values, alternately uses the prior distribution model to "de-noise" update the current parameter vector (making it more reasonable), and uses the forward mapping network to calculate the difference between the predicted performance and the target performance and the physical equation residual, combining the gradients of both to fine-tune the parameters. This process is repeated until a set of reliable process parameters that conforms to the operating rules of the equipment, achieves the target performance, and satisfies the laws of physics is output.

[0029] S2: Based on the physical associations in the positive mapping relationship, establish a causal topology structure for the device operating status; based on this causal topology structure, identify the anomaly type of the real-time monitoring data to obtain the identification results of the indication measurement anomaly or mechanism anomaly.

[0030] All steps in this process are based on the physical understanding established by S1, constructing a real-time diagnostic mechanism capable of accurately tracing the root cause of anomalies. Its core is to establish and apply a causal topology of the device's operating status, and to perform root cause analysis of monitoring deviations through probabilistic calculations.

[0031] In practical implementation, firstly, based on the physical correlations revealed by the positive mapping relationship in S1, two types of nodes and their connection relationships are clearly defined to form a causal topology. Physical quantities that cannot be directly measured but determine the process state, such as the actual plasma density in the vacuum chamber, the actual etching depth on the target surface, and the actual surface temperature of the substrate, are defined as the first type of nodes. Values ​​that can be directly observed by sensors, such as the light intensity reading of the emission spectrometer, the power supply voltage value, and the infrared thermometer reading, are defined as the second type of nodes. Then, based on the fundamental principles of plasma physics and thermodynamics, directed connections are established between the nodes: firstly, physical evolution edges describing the mutual influence between the first type of nodes (e.g., changes in "actual target etching depth" lead to changes in "actual plasma impedance"); secondly, measurement response edges describing how the first type of nodes determine the values ​​of the second type of nodes (e.g., "actual plasma density" determines "OES light intensity reading").

[0032] Based on the completed topology construction, when deviations occur in real-time monitoring data, an anomaly identification process is initiated. First, for each second-type node (i.e., each sensor reading) in the causal topology, a noise-resistant baseline regression model is trained based on historical normal production data, using its associated first-type nodes (other process parameters) as input. This model preferably employs median regression or quantile regression algorithms, rather than traditional mean squared error regression, to effectively resist occasional outliers (noise) in historical data and fit a "standard operating curve" unaffected by anomalies. Subsequently, the residual between the actual observed values ​​in historical normal data and the predicted values ​​of this "standard operating curve" is calculated, and a probability distribution model of this residual is constructed using kernel density estimation techniques, thereby quantifying the range of "background noise" under normal operating conditions (e.g., determining that voltage reading fluctuations within ±2V are normal noise).

[0033] When online monitoring detects that the reading of the target sensor node (denoted as node X) suddenly deviates from the preset range, the identification process triggers the following two virtual deductions and probability calculations in parallel based on the above topology and distribution model: First deduction (path A: simulation mechanism abnormality): its premise is that all sensors are intact, and the deviation originates from a sudden change in the actual physical state of a certain type I node inside the equipment (e.g., the target material is suddenly poisoned).

[0034] The deduction logic is as follows: Based on the physical evolution edge and measurement response edge in the causal topology, the state change of the first type of node should cause a set of related second type of node readings to undergo a physically-compliant linkage change. For example, if the "true plasma density" decreases, not only should the "OES light intensity" reading decrease, but the "target voltage" reading should also typically increase accordingly. The system calculates the joint probability density value of the current combination of readings from all associated sensors that conforms to this physical linkage change law, denoted as ScoreMech. The second deduction (path B: simulated measurement anomaly): Its assumption is that the internal physical state of the equipment is completely normal, and the deviation is only due to the failure of the measurement mechanism of the "current sensor node X" itself (e.g., probe contamination). The deduction logic is: such faults are usually isolated and do not conform to the physical linkage law. That is, only the reading of the target node X is abnormal, while the readings of other sensors physically coupled with it should remain within the normal range. The system calculates the joint probability density value of the current observation data, denoted as ScoreMeas, under the assumption that only the target node X introduces a large independent external noise interference, while other nodes remain normal.

[0035] Finally, root cause determination is performed based on the maximum likelihood ratio: the ScoreMech (mechanistic anomaly score) and ScoreMeas (measurement anomaly score) are compared. If ScoreMeas is significantly greater than ScoreMech (e.g., by orders of magnitude), it is determined to be a "measurement anomaly" (sensor failure); if ScoreMech is significantly greater than ScoreMeas, it is determined to be a "mechanistic anomaly" (actual process drift). This determination result serves as the sole basis for switching the control mode in subsequent steps.

[0036] S3: Execute the control strategy based on the identification result: if it is a measurement abnormality, then shield the monitoring data of the corresponding sensor and reconstruct the control signal based on the data of the remaining sensors; if it is a mechanism abnormality, then calculate the control parameter correction amount to compensate for process drift through an online optimization algorithm.

[0037] This step, based on the precise anomaly type determination result output from step S2, executes two distinctly different, highly directional adaptive control strategies to fundamentally resolve the inherent contradiction between "anti-interference" and "response speed" in traditional control. This method abandons a single feedback loop and employs a hierarchical decision-making execution mechanism directly triggered by the determination result. Internally, a pre-set logic switch, based on the received "measurement anomaly" or "mechanism anomaly" flag, forces the control flow to two mutually exclusive processing paths, thereby ensuring that contradictory operations are not executed.

[0038] When the flag indicates "measurement anomaly" (sensor failure), the control method automatically enters the "fault masking and data reconstruction" mode. The core objective of this mode is to protect the production process from misleading data. Specifically, it immediately resets the data weight of the faulty sensor to zero or isolates the sensor data from the control loop. The controller no longer responds to any drastic jumps in its readings, preventing erroneous and drastic adjustments to power output or gas flow due to false readings. Simultaneously, using the causal topology established in S2, data from other normal and strongly correlated sensors (such as voltage, current, and temperature data) are input into a pre-trained noise-resistant regression model to calculate the "theoretical normal value" that the faulty sensor node should have in real time. This theoretical value is then input as a virtual observation to the feedback controller, misleading it into believing that the system is operating normally, maintaining stable process parameters, and providing a safe process hold window for manual intervention.

[0039] When the flag indicates "mechanism anomaly" (real process drift), the control method confirms that a real physical change has occurred in the process environment (such as target etching or electrolyte aging), and then switches to the "robust online optimization" mode. The core task of this mode is to find and drive the process parameters to approach the optimal operating point under the current new environment. Its specific implementation follows the calculation process as follows: First, a dynamic time-varying loss function is constructed. This function not only includes the real-time deviation term between the current process state (such as deposition rate) and the target film performance (such as target deposition rate), but also explicitly introduces a penalty term for the adjustment range of control parameters to prevent excessively rapid parameter adjustment from causing plasma instability or process detuning. Since the real physical environment is an unknown black box, this method uses the finite difference method to approximate the performance gradient direction at the current moment (i.e., quantify "how much the film deposition rate will increase if the sputtering power is slightly increased") using the process parameter changes and the actual film performance response data from the previous period. Then, based on this estimated performance gradient direction and the preset learning rate (step size), the required control parameter correction amount is iteratively calculated using the gradient descent algorithm. The iterative formula is: New parameter = Old parameter - Step size × Gradient. This allows the control parameters to iteratively approach along the negative direction of the performance gradient, automatically approaching the optimal process setpoint that can stably produce the target film performance under the current physical environment.

[0040] Furthermore, to address common but unmodeled external environmental disturbances in production (such as cooling water temperature fluctuations and grid voltage fluctuations), this step, after completing the aforementioned online optimization calculation, also superimposes an disturbance observation and compensation step. This step first compares the actual output membrane performance of the system (observable online or indirectly) with the performance value predicted based on the forward mapping relationship in S1, defining the difference as the "lumped disturbance estimate." This estimate includes all unmodeled external environmental influences. Then, based on the "theoretical correction" calculated by claim 6, an additional feedforward compensation component, opposite in direction and proportional in magnitude to the "lumped disturbance estimate," is added. Finally, the "theoretical correction" and the "compensation component" are synthesized to form the final "reference signal" sent to the power supply or flow controller. This mechanism ensures that the control method maintains high robustness and stability even when facing complex, time-varying external disturbances.

[0041] S4: Perform safety verification and instruction filtering on the control parameter correction amount, and smoothly truncate or correct the out-of-bounds instruction according to the predefined equipment safety operation constraint envelope to generate the final execution instruction.

[0042] This step serves as an indispensable physical safety filter within the entire adaptive control method. Its core function is to perform millisecond-level predictive safety checks and smooth corrections before any control command actually drives the device hardware. This mathematically ensures that all actions are within the absolute safe operating boundaries of the device, completely eliminating the risk of device shutdown or damage due to aggressive algorithm optimization. This step abandons the traditional post-event alarm mechanism and adopts a proactive, predictive safety filtering logic.

[0043] In practice, this step first digitally defines a clear and insurmountable safety operation constraint envelope based on the inherent physical characteristics, manufacturing specifications, and safe operating procedures of the equipment. This step specifically includes: calculating and delineating a closed, dynamically updatable geometrical safety zone within a multidimensional space composed of multiple control parameters (such as voltage, current, gas flow rate, and power), based on a series of hard constraint indicators such as the maximum output voltage of the power module (to prevent high-voltage breakdown of the insulation layer) and the maximum current rise rate (to prevent overcurrent protection tripping due to excessively rapid current changes), the minimum holding pressure of the vacuum system (to prevent plasma extinction due to excessively low pressure) and the maximum gas flow rate change rate (to prevent sudden gas flow from impacting or overloading the vacuum pump), and the maximum power density of the target material in the core process components (to prevent target melting or cracking due to localized overheating). Only when all control parameter state points fall within this zone is the equipment operation considered absolutely safe.

[0044] Subsequently, upon receiving the "process parameter correction amount" (i.e., the theoretically recommended adjustment values ​​for voltage, flow rate, etc.) calculated from step S3, this step does not immediately issue the correction. Instead, it initiates a high-speed virtual simulation and prediction process. This process uses this "correction amount" as input and feeds it into a predictive model capable of rapidly simulating the dynamic response of the equipment (this model can be built based on the equipment mechanism or data-driven approaches). It then simulates how the key state indicators of the equipment (such as instantaneous current, chamber pressure, target surface temperature, etc.) will change at the next sampling moment if the instruction is executed.

[0045] Next, a rigorous boundary violation determination is performed: the virtually deduced future device state is compared in real-time and across multiple dimensions with the aforementioned defined "safe operation constraint envelope." This will produce two clear predictive results: Scenario A (Safety Prediction): If all the future state points deduced are completely within the safety constraint envelope, then the instruction is deemed safe and can be directly released, allowing it to be issued as the final execution instruction.

[0046] Scenario B (Risk of Exceeding Boundaries Prediction): If the simulation results show that any one or more future state indicators will exceed or touch the boundary of the safety constraint envelope (for example, the prediction shows that the current will exceed the protection threshold instantaneously, or the gas pressure will fall below the limit for maintaining discharge), then the original instruction is determined to have a clear risk of exceeding the boundaries and must be intervened and corrected.

[0047] For instructions deemed risky, this step instantly activates its core instruction smoothing and truncation mechanism. The system then constructs and solves a miniature, real-time constraint optimization problem. The optimization objective is to find a new, modified instruction point that is mathematically as close as possible to the original theoretically optimized instruction generated in step S3 (to preserve the optimization intent to the greatest extent), but must be strictly within the "safe operating constraint envelope" (to satisfy all safety constraints). For specific calculation methods, efficient linear programming or bisection algorithms can be used. For example, the precise intersection of the line connecting the "current safe state point of the device" and the "ideal target state point provided by S3" with the boundary of the safety constraint envelope can be quickly calculated; this intersection point is the maximum allowable adjustment boundary within the safe range. Alternatively, the point closest to the ideal point in Euclidean distance can be directly searched within the safety envelope.

[0048] After the solution is completed, the system calculates a smoothing correction coefficient between 0 and 1. This coefficient quantifies the allowable adjustment range of the original instruction at the current moment. For example, if the original theoretical optimization instruction requires a voltage increase of 100V, but the safety boundary only allows an immediate increase of no more than 20V, the calculated correction coefficient might be 0.2. Finally, the original theoretical optimization instruction is weighted and smoothed based on this correction coefficient to generate an absolutely safe final execution instruction. Its execution logic can be expressed as: Final execution instruction = Stable instruction at the previous moment + Correction coefficient × (Original theoretical optimization instruction - Stable instruction at the previous moment). This processing method is equivalent to adding an intelligent smooth transition mechanism or instruction limiting and filtering module between the aggressive performance optimization instruction and the rigid physical safety boundary. It transforms the sudden and violent jumps that may cause equipment oscillation or shutdown into a smooth and gradual transition curve that closely follows the safety boundary, thereby optimizing process performance as much as possible while ensuring absolute equipment safety.

[0049] S5: Based on the identification results, execution control strategy, and final execution instructions, generate an explanatory report that includes an explanation of the root cause of the anomaly, the contribution of parameter adjustment, and maintenance guidelines.

[0050] This process aims to endow the entire adaptive control system with "interpretability" and "interactive guidance" capabilities. It seeks to transform the decision-making process, based on complex mathematical models and algorithms, from the preceding steps into a physical language that operators can intuitively understand and trust, along with actionable maintenance suggestions. This enhances the interpretability of the control process and establishes a trust-based human-machine collaborative loop. This method utilizes an integrated interpretation generation framework to ensure that every critical decision of the control system is clearly explained. See also... Figure 3 Specifically, it includes the following sub-steps: S51: Capture a status snapshot containing the current membrane quality deviation value, real-time readings of each sensor, and control command increments; S52: Calculate the marginal contribution value of each control parameter adjustment to the correction of film performance deviation through a pre-trained contribution analysis model, and normalize the marginal contribution value into contribution weight. S53: Based on the anomaly type in the identification results, and in conjunction with the contribution weight and physical semantic mapping table, generate a natural language explanation text containing anomaly root cause description and parameter adjustment contribution. S54: When the identification result indicates a measurement anomaly, standard maintenance actions and their standard time consumption are extracted from the maintenance action database according to the corresponding sensor type, and the estimated maintenance time is obtained by accumulating them. The estimated maintenance time is then added to the explanation report as a maintenance guide.

[0051] In practice, a complete "system state snapshot" is immediately captured and recorded the instant the system performs a critical action (e.g., S2 makes an anomaly judgment, S3 adjusts process parameters, or S4 makes a safety correction to the command). This snapshot structurally contains the core state vector at the current moment, specifically: key deviation values ​​of the current film quality (e.g., color difference reddish tint or deposition rate decrease detected through online spectral or thickness monitoring), real-time raw readings of all relevant sensors (e.g., values ​​of various vacuum gauges, spectrometers, voltmeters, ammeters, and thermometers), and the increment of control commands that the system has just calculated and is preparing to issue (specific adjustment actions such as "increase sputtering power setpoint by 50W" or "reduce argon flow rate by 2 sccm" are clearly recorded). These data, together with several sets of historical operating records highly similar to the current operating conditions retrieved from the historical database, constitute the "input-output correlation dataset" prepared for subsequent interpretation and analysis.

[0052] After obtaining the state snapshot data, a rapid feature contribution analysis is performed without retraining the model. A pre-trained, general-purpose "contribution analysis model" (based on techniques such as tabular models) is invoked. This model has learned the deep statistical correlation between massive changes in process parameters and changes in membrane performance. The prepared current snapshot data (as query samples) and similar historical data (as reference background) are input into this model. Through its internal pattern matching and statistical inference mechanisms, the model can directly simulate and answer a key question without accessing or calculating the gradients within the complex physical information neural network in S1: "If the current adjustment to a specific control parameter (such as power) is reversed, while keeping other parameters unchanged, how much will the currently observed membrane performance deviation (such as a decrease in deposition rate) increase?" Through this "counterfactual" simulation calculation for each parameter, the model quickly estimates the "marginal contribution value" of each control parameter adjustment to correcting the current performance deviation. Subsequently, the marginal contribution values ​​of all parameters are normalized to obtain the contribution weights presented intuitively as percentages (for example, the output result is: "In this operation to restore the deposition rate, the contribution of sputtering power adjustment is 85%, and the contribution of argon flow rate fine-tuning is 15%").

[0053] After obtaining the quantified contribution weights, this step enters the dynamic synthesis stage of the natural language interpretation report. Based on the precise anomaly type determination result ("measurement anomaly" or "mechanism anomaly") output from step S2, the system combines the calculated contribution weights and, by querying a pre-defined physical semantic mapping table, transforms the cold numbers and codes into expressions that workshop operators can understand. This mapping table establishes a direct correspondence between technical parameters and physical phenomena; for example, "increased power" maps to "compensating for target etching loss"; "decreased oxygen flow" maps to "inhibiting excessive film oxidation"; and "vacuum gauge reading ignored by the system" maps to "sensor suspected contamination, reading unreliable." Based on this, the system dynamically synthesizes text prompts: For scenarios with abnormal mechanisms: Generate a message such as "A 15% decrease in film deposition rate has been detected. This is determined to be a real process drift caused by target consumption (confidence level higher than 95%). Automatic compensation has been performed: increase sputtering power by 50W (main contributing factor, weight 85%) to enhance sputtering rate, and fine-tune argon flow rate by -1 sccm (secondary contributing factor, weight 15%) to maintain plasma stability." For scenarios involving measurement anomalies: A message such as "A logical conflict was detected between the reading of vacuum gauge A and the discharge voltage and spectral characteristics (99% confidence level). Vacuum gauge A has been determined to be an anomaly, and its data has been automatically masked. Current process control maintains operation based on the pressure value calculated from the voltage and spectrum, and the process status is stable." Finally, for scenarios identified as "measurement anomalies" requiring manual intervention (such as cleaning contaminated sensors), this step also provides a precise maintenance operation time estimation and guidance function. Based on the specific faulty sensor type, the system extracts a series of standard maintenance steps and their standard times from a pre-set database of standard maintenance action times (e.g., "Open the vacuum chamber door: 30 seconds," "Remove and clean the spectrometer lens: 300 seconds," "Re-evacuate to process requirements: 600 seconds"). Then, the time units of these standard maintenance steps are summed to calculate the estimated total downtime for maintenance. Finally, on the operator interface, a clear human-machine interface instruction is displayed, such as "It is recommended to perform 'Vacuum Gauge A Probe Cleaning' maintenance. The following steps are expected to be performed: ..., with a total time of approximately 15 minutes. Should it be scheduled for the next production break?" This quantitative forecast based on action decomposition is far more accurate than experience-based guessing. It can effectively help production managers with maintenance planning and production scheduling, avoiding production plan chaos caused by temporary and blind shutdowns. This upgrades the intelligent control system from a simple "automatic executor" to an intelligent auxiliary decision-making system that can guide manual maintenance and improve overall operation and maintenance efficiency.

[0054] Furthermore, this application provides an electronic device including at least one processor and a memory communicatively connected to the at least one processor. The memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other non-volatile memory technologies. The memory stores a computer program (or machine-executable instructions, software program modules, etc.), which is specifically configured to include instruction code for performing the steps of any of the foregoing method embodiments.

[0055] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or any combination thereof. When the processor loads and executes the computer program in the memory, it can realize the complete adaptive control flow described above: including building a physical-data decoupled process parameter inversion model, establishing and identifying anomaly types based on causal topology, executing submodal control strategies according to the identification results, performing safety verification and filtering on control commands, and generating interpretable reports. The processor interacts with various sensors (such as spectrometers, vacuum gauges, voltage and current probes) and actuators (such as power supplies, mass flow controllers, and temperature controllers) of the PVD or micro-arc oxidation equipment in real time through built-in or external communication interfaces (such as analog / digital input / output modules, fieldbus interfaces, industrial Ethernet interfaces, etc.), thereby forming a complete embedded or host computer control system.

[0056] In one specific embodiment, the electronic device may be embodied as a process controller or industrial computer built into the PVD or micro-arc oxidation equipment, or an edge computing gateway or server connected to the equipment via a network. The computer program stored within it, when running, will call and manage the various functional modules defined by the method, for example: Model management module: responsible for loading, updating or running online the process parameter inversion model and its sub-models (prior distribution model, forward mapping network) described in S1. Real-time diagnostic module: Executes step S2, continuously collects sensor data, maintains the causal topology, and performs anomaly probability calculation and judgment; Decision execution module: Executes step S3, switches control modes based on diagnostic results, and runs online optimization algorithms or data reconstruction logic; Safety monitoring module: Executes S4 steps, manages safety constraint envelopes, and performs instruction prediction and smoothing filtering; Human-Computer Interaction Module: Executes step S5, generates an explanation report, and provides an operation interface and maintenance guidelines.

[0057] Therefore, this electronic device, through a combination of hardware and software, transforms the abstract control logic of the aforementioned method embodiments into a concrete physical device that can operate stably and reliably in industrial settings. It is the key hardware carrier for realizing the adaptive control method of this application. Any general-purpose or special-purpose computing device containing the program instructions and capable of implementing the described method flow falls within the protection scope of this application.

[0058] It is conceivable that this application not only protects the method itself, but also the electronic device that implements the method, the beneficial effects of which will not be elaborated here.

[0059] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0060] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0061] The above are merely optional embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made based on the inventive concept of this application and the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the patent protection scope of this application.

Claims

1. An adaptive control method for PVD and micro-arc oxidation equipment, applied to PVD and micro-arc oxidation equipment, characterized in that, Includes the following steps: S1: Construct a physical-data decoupled process parameter inversion model. The model is characterized by the prior distribution of process parameters obtained by training based on historical operating data and the positive mapping relationship from process parameters to membrane performance obtained by training based on physical equation constraints, so as to invert process parameters according to target membrane performance. S2: Based on the physical associations in the positive mapping relationship, establish a causal topology structure for the device operating status; based on this causal topology structure, identify the anomaly type of the real-time monitoring data to obtain the identification results of the indication measurement anomaly or mechanism anomaly. S3: Execute the control strategy based on the identification result: if it is a measurement abnormality, then shield the monitoring data of the corresponding sensor and reconstruct the control signal based on the data of the remaining sensors; if it is a mechanism abnormality, then calculate the correction amount of the control parameters to compensate for process drift through an online optimization algorithm. S4: Perform safety verification and instruction filtering on the control parameter correction amount, and smoothly truncate or correct the out-of-bounds instruction according to the predefined equipment safety operation constraint envelope to generate the final execution instruction. S5: Based on the identification results, execution control strategy, and final execution instructions, generate an explanatory report that includes an explanation of the root cause of the anomaly, the contribution of parameter adjustment, and maintenance guidelines.

2. The method according to claim 1, characterized in that, Step S1 includes: Multidimensional time-series process parameters are extracted from historical operating data, and the prior distribution of process parameters is obtained by training an unconditional diffusion model. The process parameters and film properties are paired data, and a positive mapping relationship is obtained based on the training of a physical information neural network. The loss function of the physical information neural network includes the residual term of the physical equation. By using a sequential Monte Carlo sampling algorithm, combined with the prior distribution of the process parameters and the forward mapping relationship, the inversion calculation from the target film performance to the process parameters can be achieved.

3. The method according to claim 2, characterized in that, The unconditional diffusion model is obtained by iteratively adding Gaussian noise to the process parameter samples and training a neural network to predict the noise in reverse. The physical information neural network is a neural network that introduces physical equations as constraints into the loss function during the training process.

4. The method according to claim 1, characterized in that, The specific steps for establishing the causal topology structure of the device operating status in S2 include: Plasma density, true etching depth of target surface, and true surface temperature of substrate are defined as first-type nodes, and emission spectrometer light intensity reading, power supply voltage value, and infrared thermometer reading are defined as second-type nodes. Based on the principles of plasma physics and thermodynamics, physical evolution edges between first-type nodes and measurement response edges from first-type nodes to second-type nodes are established.

5. The method according to claim 4, characterized in that, The step in S2, which identifies the anomaly type of real-time monitoring data and obtains the identification result indicating measurement anomaly or mechanism anomaly, includes: For each second-class node, a noise-resistant baseline regression model is trained based on historical normal data, and a probability distribution model of the prediction residual is constructed. When the monitoring data deviates, calculate the joint probability density value that conforms to the physical linkage law and the joint probability density value that conforms to the independent fault of the sensor, respectively. By comparing the magnitude of the joint probability density value that conforms to the physical linkage law and the joint probability density value that conforms to the independent fault of the sensor, the identification result of measurement abnormality or mechanism abnormality is output.

6. The method according to claim 1, characterized in that, The step in S3, which calculates the control parameter correction amount for compensating process drift using an online optimization algorithm, includes: Construct a time-varying loss function that includes the current film performance deviation term and the penalty term for the adjustment magnitude of the control parameters; By utilizing the correlation between historical process parameters and film performance, the performance gradient direction is estimated using the finite difference method; Based on the performance gradient direction and the preset learning rate, the control parameter correction amount is iteratively calculated using the gradient descent algorithm.

7. The method according to claim 1, characterized in that, S3 also includes an interference observation compensation step: The difference between the actual film performance and the predicted performance based on the positive mapping relationship is calculated as a lumped interference estimate; The control parameter correction amount is superimposed with a feedforward compensation amount that is opposite to the direction of the lumped disturbance estimation.

8. The method according to claim 1, characterized in that, The step in S4 of smoothly truncating or correcting the out-of-bounds command based on the predefined equipment safety operation constraint envelope includes: Based on the constraints of equipment safety operation parameters, a safety operation constraint envelope is defined in a multidimensional control parameter space; The control parameter correction values ​​are input into the device dynamic model for virtual simulation to predict the process state after execution. If the predicted state exceeds the safe operating constraint envelope, the optimal correction point is found within the safe boundary using a constraint optimization algorithm. Based on the optimal correction point, a smooth transition instruction is calculated, and the final execution instruction is generated.

9. The method according to claim 1, characterized in that, Step S5 includes: Capture a status snapshot containing the current membrane quality deviation value, real-time readings of each sensor, and control command increments; The marginal contribution value of each control parameter adjustment to the correction of film performance deviation is calculated using a pre-trained contribution analysis model, and the marginal contribution value is normalized into a contribution weight. Based on the anomaly type in the identification results, and combined with the contribution weight and physical semantic mapping table, a natural language explanation text containing anomaly root cause description and parameter adjustment contribution is generated. When the identification result indicates a measurement anomaly, standard maintenance actions and their standard durations are extracted from the maintenance action database according to the corresponding sensor type, and the estimated maintenance time is obtained by accumulating them. The estimated maintenance time is then added to the explanation report as a maintenance guide.

10. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 9.

Images