Vacuum coating valve intelligent control system based on self-adaptive PID (Proportion Integration Differentiation) algorithm

Abstract

The invention discloses a vacuum coating valve intelligent control system based on a self-adaptive PID algorithm, and the system comprises a multi-source process data collection module which is used for collecting and preprocessing multi-source process data, and forming multivariable time sequence data; the variable diagram construction module is used for mapping the multivariable time sequence data into a vacuum coating variable diagram; the improved CSPNet model parameter generation module is used for outputting a candidate proportionality coefficient, a candidate integral coefficient and a candidate differential coefficient through cross-stage feature extraction and working condition prediction; the initial control quantity generation module is used for calculating an initial control quantity; the self-adaptive PID parameter updating module is used for obtaining the updated PID parameters; and the PID control execution module is used for closed-loop intelligent adjustment of the vacuum degree and the process gas flow. According to the invention, through combination of the improved CSPNet model and the adaptive PID adjustment, intelligent and stable control of the vacuum coating valve is realized.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and in particular to an intelligent control system for vacuum coating valves based on an adaptive PID algorithm. Background Technology

[0002] In the manufacturing processes of semiconductors, LCD panels, solar cells, and LED / OLED products, vacuum coating is a crucial step in forming functional thin film layers and improving product performance. Various processes, including PECVD, MPCVD, and LPCVD, are commonly employed. These processes demand high precision in controlling chamber vacuum and process gas flow rates, as well as high dynamic response speed and process repeatability. Typically, a feedback control loop is constructed using vacuum valves, vacuum sensors, and mass flow meters. Currently, most vacuum coating control valves in production lines rely on imported products, while domestically produced vacuum control valves and control systems are relatively weak. Their ability to rapidly adjust vacuum signals within the 0–1000 Torr range under complex process conditions is limited, especially in multi-stage processes and multi-layer film deposition scenarios, often resulting in insufficient response speed and low control precision. As coating equipment evolves towards larger substrate sizes, high-throughput production, and precise multi-gas formulation, the shortcomings of traditional vacuum valve control systems in terms of stability, adjustment precision, and adaptability to different machine types and process requirements are becoming increasingly prominent, making it difficult to meet the intelligent and precise control requirements of next-generation high-end vacuum coating equipment.

[0003] Current vacuum coating valve control systems mostly employ traditional PID or empirical adaptive PID algorithms. PID parameters are typically manually tuned by engineers through repeated trial coatings under specific machine conditions and process formulations. When process formulations, carrier loads, temperatures, or pipeline pressure drops change, the original parameters are difficult to adapt in a timely manner. This leads to problems such as large overshoot, large steady-state fluctuations, and even oscillations in vacuum levels and process gas flow rates under nonlinear and time-varying conditions. Traditional PID algorithms and simple adaptive strategies struggle to fully utilize multi-source data and lack the ability to deeply model and mine operational characteristics. Current technologies lack a complete system solution for adaptive intelligent closed-loop control of vacuum coating valves that can extract features from multi-source process data, predict optimal PID parameters online using deep learning models, and combine adaptive updates decoupled by P / I / D channels with event-triggered updates. This solution fails to effectively address the challenge of balancing rapid response and high stability under complex vacuum coating conditions.

[0004] Therefore, how to provide an intelligent control system for vacuum coating valves based on an adaptive PID algorithm is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent control system for vacuum coating valves based on an adaptive PID algorithm. This invention comprehensively utilizes industrial automation and deep learning methods, including multi-source process data acquisition and preprocessing, variable graph construction and improved CSPNet feature extraction, embedded physical model-based operating condition prediction, and adaptive PID parameter updates decoupled by P / I / D channels. It fully describes the process from acquiring multi-source data such as vacuum level, process gas flow rate, valve opening, control variables, and process setpoints; constructing a vacuum coating variable graph; performing cross-stage feature extraction of the graph structure and physical model fusion using an improved CSPNet model; generating candidate PID parameters; and then combining event detection and parallel adaptive updates of P, I, and D channels to obtain updated PID parameters, driving the vacuum coating valve to achieve closed-loop intelligent regulation. This achieves rapid response and high stability control of the vacuum coating valve under nonlinear and time-varying operating conditions. Compared with existing traditional PID control that relies on manual experience tuning, it has advantages such as strong parameter adaptability, high control accuracy, good operating condition adaptability, and ease of engineering application deployment.

[0006] An intelligent control system for a vacuum coating valve based on an adaptive PID algorithm according to an embodiment of the present invention includes the following modules:

[0007] The multi-source process data acquisition module is used to acquire multi-source process data and preprocess it to form multivariate time series data;

[0008] The variable graph construction module is used to map multivariate time series data into a vacuum coating variable graph based on physical structure relationships and coupling relationships;

[0009] An improved CSPNet model parameter generation module is used to receive a vacuum coating variable map and output candidate proportional coefficients, candidate integral coefficients, and candidate differential coefficients through cross-stage feature extraction and working condition prediction.

[0010] The initial control quantity generation module is used to calculate the initial control quantity based on the control error and according to the candidate proportional coefficient, candidate integral coefficient and candidate derivative coefficient;

[0011] The adaptive PID parameter update module is used to correct the candidate PID parameters according to the initial control quantity using a three-channel adaptive method, so as to obtain the updated PID parameters.

[0012] The PID control execution module is used to recalculate the control quantity according to the updated PID parameters and drive the vacuum coating valve execution platform to achieve closed-loop intelligent regulation of vacuum degree and process gas flow.

[0013] Optionally, modules can be integrated using the following methods:

[0014] Multi-source process data from the vacuum coating equipment were collected and preprocessed to obtain multivariate time series data;

[0015] Based on the physical structure relationship of the controlled object, the variables in the multivariate time series data are used as graph nodes, and a variable graph topology is established according to the coupling relationship between the variables to obtain the vacuum coating variable graph.

[0016] An improved CSPNet model is constructed, which takes the vacuum coating variable graph as input. Through cross-stage feature extraction of the graph structure and condition prediction of the embedded physical model, the final feature representation of the current vacuum coating condition is obtained. Candidate PID parameters, including candidate proportional coefficients, candidate integral coefficients and candidate derivative coefficients, are obtained by the hierarchical output head module.

[0017] The control error is calculated based on the process setpoint and the real-time vacuum measurement value, and the control quantity is calculated according to the candidate proportional coefficient, candidate integral coefficient and candidate derivative coefficient.

[0018] The control error and control quantity are input into the adaptive PID algorithm. Event detection determines whether to update the parameters. When it is determined that an update is needed, the candidate PID parameters are updated using P-channel adaptive, I-channel adaptive and D-channel adaptive methods to obtain the updated PID parameters.

[0019] The control quantity is recalculated based on the updated PID parameters and output to the vacuum coating valve execution platform to perform closed-loop regulation of vacuum degree and process gas flow.

[0020] Optionally, the multi-source process data includes vacuum level, process gas flow rate, valve opening degree, control quantity, and process setpoint.

[0021] Optionally, obtaining multivariate time series data includes:

[0022] Multi-source raw process data is collected by using vacuum sensors, process gas flow meters, valve opening feedback devices, and controller acquisition interfaces installed on the vacuum coating equipment.

[0023] A unified sampling period is set for the raw process data from multiple sources. Various signals are resampled and time-aligned according to the sampling time axis, and abnormal data is removed, filtered, and smoothed.

[0024] The data is scaled or standardized according to the variable dimension, and the processed vacuum degree, process gas flow rate, valve opening degree, control quantity and process set value are combined according to each sampling time to construct multivariate time series data.

[0025] Optionally, obtaining the vacuum coating variable map includes:

[0026] Define each type of variable in the multivariate time series data as a graph node, and assign a unique identifier to each graph node;

[0027] Based on the physical transmission relationships between pressure and process gas flow rate, pressure and valve opening, process gas flow rate and valve opening, pressure and pump current, and pressure and control quantity, we determine whether there is a direct coupling relationship between each variable. We establish graph edges between two graph nodes with direct coupling relationships and do not establish graph edges between two graph nodes without direct coupling relationships, thus obtaining a variable graph topology structure containing a set of graph nodes and a set of graph edges.

[0028] The time series data corresponding to each graph node is used as node features and associated with the topology of the variable graph to form a vacuum coating variable graph that includes node identifiers, node feature sequences, and connection relationships between nodes.

[0029] Optionally, the candidate PID parameters obtained by the hierarchical output head module, including candidate proportional coefficients, candidate integral coefficients, and candidate derivative coefficients, include:

[0030] Construct an improved CSPNet model, including a node-level cross-stage module, a physical model module, and a hierarchical output header module;

[0031] The node feature sequence and topological connection relationship of the vacuum coating variable graph are input into the node-level cross-stage module. At each graph node, the node features are split into main branch features and bypass branch features according to the channel dimension. One-dimensional convolution and residual operation are performed on the main branch features in sequence, and linear transformation is performed on the bypass branch features. At the exit of each cross-stage unit, channel splicing and channel compression convolution are performed on the main branch features and bypass branch features to obtain node-level features.

[0032] Based on the topology of the vacuum coating variable graph, graph convolution and adjacency aggregation operations are performed on the node-level features of each graph node. The first node-level features of adjacent nodes are aggregated to obtain multivariate coupled operating condition features that simultaneously represent the coupling relationship between vacuum degree, process gas flow rate, valve opening, control quantity and process set value. Channel splitting, main branch convolution extraction, bypass branch retention and graph structure aggregation operations are repeatedly performed along the network depth direction to form cross-stage feature extraction of graph structure and obtain operating condition feature representation.

[0033] The physical model module input is formed by combining the operating condition feature representation with the physical parameters, control quantities and process settings of vacuum coating. The discrete update equation of the dynamic relationship between vacuum degree and process gas flow rate is used to calculate the vacuum degree and process gas flow rate within the prediction time step to obtain the physical prediction features. The physical prediction features and the operating condition feature representation are concatenated in the channel dimension and processed by channel compression convolution. Graph structure feature extraction and physical prediction feature fusion are alternately performed between the node-level cross-stage module and the physical model module to obtain the final feature representation of the current vacuum coating operating condition.

[0034] The final feature representation is pooled and concatenated in the time dimension and graph node dimension to obtain the global operating condition feature vector. The global operating condition feature vector is then input into the hierarchical output head module. In the coarse parameter layer, the first fully connected sub-network is used to generate multiple sets of coarse PID parameter prototypes. In the intermediate parameter layer, the second fully connected sub-network is used to locally modify each coarse PID parameter prototype to generate intermediate PID parameters. In the fine parameter layer, the third fully connected sub-network is used to output candidate PID parameters, which include candidate proportional coefficients, candidate integral coefficients, and candidate derivative coefficients.

[0035] Optionally, the step of calculating the control error based on the process setpoint and the real-time vacuum measurement, and calculating the control quantity according to the candidate proportional coefficient, candidate integral coefficient, and candidate derivative coefficient, includes:

[0036] Within each control sampling cycle, acquire the vacuum process setting value and real-time vacuum measurement value of the corresponding process stage, subtract the real-time vacuum measurement value from the vacuum process setting value to obtain the current control error, and record the current control error, the control error of the previous sampling cycle, and the control error of the previous sampling cycle.

[0037] The candidate proportional coefficient, candidate integral coefficient, and candidate derivative coefficient are multiplied by the current control error and the historical control error respectively, and then weighted and summed to obtain the control increment for the current sampling period. The control increment is then added to the control quantity of the previous sampling period to obtain the control quantity for the current control period.

[0038] Optionally, obtaining the updated PID parameters includes:

[0039] In each control sampling cycle, the current control error, error threshold, vacuum level, fluctuation threshold, and current process stage identifier are read. The absolute value of the current control error is compared with the error threshold, the fluctuation index of the vacuum level is compared with the fluctuation threshold, and the process stage identifier is used to determine whether a process stage switch has occurred.

[0040] If the absolute value of the current control error is greater than any of the preset error thresholds, or if the vacuum fluctuation index is greater than any of the fluctuation thresholds, or if a process stage identifier switch is detected, a PID parameter update event is determined to have occurred. If none of the three conditions are met, a PID parameter update event is determined not to have occurred.

[0041] When a PID parameter update event is detected, the candidate proportional coefficient, candidate integral coefficient, and candidate derivative coefficient, together with the current control error, historical control error, and control quantity, are input into the P-channel adaptive unit, I-channel adaptive unit, and D-channel adaptive unit. The P-channel adaptive unit generates the updated proportional coefficient based on the candidate proportional coefficient and the error amplitude and response speed information. The I-channel adaptive unit generates the updated integral coefficient based on the candidate integral coefficient and the steady-state error and error accumulation information. The D-channel adaptive unit generates the updated derivative coefficient based on the candidate derivative coefficient and the error change rate and output oscillation information.

[0042] When it is determined that no PID parameter update event has occurred, the proportional coefficient, integral coefficient, and derivative coefficient of the previous control sampling period are directly used as the PID parameters of the current control sampling period, and the PID parameters are not updated.

[0043] When a PID parameter update event is detected, the updated proportional coefficient, integral coefficient, and derivative coefficient output by the P-channel adaptive unit, I-channel adaptive unit, and D-channel adaptive unit are written into the PID controller as new PID parameters.

[0044] Optionally, the closed-loop regulation of vacuum level and process gas flow rate includes:

[0045] At the beginning of each control sampling cycle, the proportional coefficient, integral coefficient, and derivative coefficient currently in use are read from the PID controller. The vacuum process setpoint is subtracted from the real-time vacuum measurement value at the same moment to obtain the current control error. The current control error, the control error of the previous sampling cycle, and the control error of the previous sampling cycle are recorded.

[0046] The proportional, integral, and derivative coefficients are used to calculate the control error, resulting in the proportional control quantity, integral control quantity, and derivative control quantity. The proportional control quantity is equal to the product of the proportional coefficient and the current control error. The integral control quantity is equal to the product of the integral coefficient and the sum of the current control error and the historical control error over time. The derivative control quantity is equal to the product of the derivative coefficient and the ratio of the difference between the current control error and the control error of the previous sampling period. The proportional control quantity, integral control quantity, and derivative control quantity are added together to obtain the control quantity for the current control sampling period.

[0047] The control quantity is output as a valve control command to the vacuum coating valve execution platform, driving the vacuum coating valve opening to increase or decrease, and the vacuum degree and process gas flow rate in the vacuum coating chamber are adjusted in a closed loop around the vacuum degree process set value within the continuous control sampling cycle.

[0048] The beneficial effects of this invention are:

[0049] This invention significantly enhances the feature perception and parameter adaptation capabilities of vacuum coating valves under complex operating conditions by introducing an improved CSPNet model driven by multi-source process data and variable graph topology modeling. By mapping vacuum level, process gas flow rate, valve opening, control variables, and process setpoints to a vacuum coating variable graph, and performing cross-stage feature extraction within the graph structure, followed by alternating fusion with a vacuum dynamics physical model, operating condition features reflecting multi-variable coupling relationships and dynamic characteristics are obtained. Hierarchical output heads are used to generate candidate PID parameters in stages, enabling precise prediction of proportional, integral, and derivative coefficients. Compared to existing control methods relying on manual experience tuning and fixed-structure PID controllers, this invention effectively reduces overshoot and steady-state fluctuations in vacuum level and process gas flow rate under nonlinear and time-varying conditions, improves response speed and control accuracy, and significantly enhances adaptability and stability to multi-stage coating processes.

[0050] This invention constructs an event-triggered and parallel adaptive closed-loop intelligent control structure by setting up an event detection and P / I / D three-channel adaptive PID parameter update mechanism, achieving a balance between control performance and system stability. The PID parameter update is determined based on error threshold, vacuum fluctuation, and process stage changes, avoiding the impact of frequent minor parameter adjustments on system stability. The P-channel, I-channel, and D-channel adaptive units respectively modify candidate PID parameters for error amplitude and response speed, steady-state error and error accumulation, and error change rate and oscillation degree, ensuring that the proportional, integral, and derivative actions each achieve optimal adjustment matching the operating conditions. Compared with existing single adaptive law or fixed parameter control, this invention effectively reduces manual parameter tuning and trial plating times, shortens equipment start-up and changeover debugging time, improves coating yield and process repeatability, and facilitates widespread application and engineering implementation in vacuum coating production lines. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 The flowchart shows a smart control system for vacuum coating valves based on an adaptive PID algorithm proposed in this invention.

[0053] Figure 2 This is a structural block diagram of a smart control method for vacuum coating valves based on an adaptive PID algorithm proposed in this invention.

[0054] Figure 3 This is a functional diagram of the improved CSPNet model for an intelligent control method for vacuum coating valves based on an adaptive PID algorithm proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figure 1 A smart control system for vacuum coating valves based on an adaptive PID algorithm includes the following modules:

[0057] The multi-source process data acquisition module is used to acquire multi-source process data and preprocess it to form multivariate time series data;

[0058] The variable graph construction module is used to map multivariate time series data into a vacuum coating variable graph based on physical structure relationships and coupling relationships;

[0059] An improved CSPNet model parameter generation module is used to receive a vacuum coating variable map and output candidate proportional coefficients, candidate integral coefficients, and candidate differential coefficients through cross-stage feature extraction and working condition prediction.

[0060] The initial control quantity generation module is used to calculate the initial control quantity based on the control error and according to the candidate proportional coefficient, candidate integral coefficient and candidate derivative coefficient;

[0061] The adaptive PID parameter update module is used to correct the candidate PID parameters according to the initial control quantity using a three-channel adaptive method, so as to obtain the updated PID parameters.

[0062] The PID control execution module is used to recalculate the control quantity according to the updated PID parameters and drive the vacuum coating valve execution platform to achieve closed-loop intelligent regulation of vacuum degree and process gas flow.

[0063] refer to Figure 2 and Figure 3 A smart control method for vacuum coating valves based on an adaptive PID algorithm, comprising:

[0064] Multi-source process data from the vacuum coating equipment were collected and preprocessed to obtain multivariate time series data;

[0065] Based on the physical structure relationship of the controlled object, the variables in the multivariate time series data are used as graph nodes, and a variable graph topology is established according to the coupling relationship between the variables to obtain the vacuum coating variable graph.

[0066] An improved CSPNet model is constructed, which takes the vacuum coating variable graph as input. Through cross-stage feature extraction of the graph structure and condition prediction of the embedded physical model, the final feature representation of the current vacuum coating condition is obtained. Candidate PID parameters, including candidate proportional coefficients, candidate integral coefficients and candidate derivative coefficients, are obtained by the hierarchical output head module.

[0067] The control error is calculated based on the process setpoint and the real-time vacuum measurement value, and the control quantity is calculated according to the candidate proportional coefficient, candidate integral coefficient and candidate derivative coefficient.

[0068] The control error and control quantity are input into the adaptive PID algorithm. Event detection determines whether to update the parameters. When it is determined that an update is needed, the candidate PID parameters are updated using P-channel adaptive, I-channel adaptive and D-channel adaptive methods to obtain the updated PID parameters.

[0069] The control quantity is recalculated based on the updated PID parameters and output to the vacuum coating valve execution platform to perform closed-loop regulation of vacuum degree and process gas flow.

[0070] In this embodiment, the multi-source process data includes vacuum level, process gas flow rate, valve opening degree, control quantity, and process setpoint.

[0071] In this embodiment, obtaining multivariate time series data includes:

[0072] Multi-source raw process data is collected by using vacuum sensors, process gas flow meters, valve opening feedback devices, and controller acquisition interfaces installed on the vacuum coating equipment.

[0073] A unified sampling period is set for the raw process data from multiple sources. Various signals are resampled and time-aligned according to the sampling time axis, and abnormal data is removed, filtered, and smoothed.

[0074] The data is scaled or standardized according to the variable dimension, and the processed vacuum degree, process gas flow rate, valve opening degree, control quantity and process set value are combined according to each sampling time to construct multivariate time series data.

[0075] In this embodiment, obtaining the vacuum coating variable map includes:

[0076] Define each type of variable in the multivariate time series data as a graph node, and assign a unique identifier to each graph node;

[0077] Based on the physical transmission relationships between pressure and process gas flow rate, pressure and valve opening, process gas flow rate and valve opening, pressure and pump current, and pressure and control quantity, we determine whether there is a direct coupling relationship between each variable. We establish graph edges between two graph nodes with direct coupling relationships and do not establish graph edges between two graph nodes without direct coupling relationships, thus obtaining a variable graph topology structure containing a set of graph nodes and a set of graph edges.

[0078] The time series data corresponding to each graph node is used as node features and associated with the topology of the variable graph to form a vacuum coating variable graph that includes node identifiers, node feature sequences, and connection relationships between nodes.

[0079] In this embodiment, the process of obtaining candidate PID parameters from the hierarchical output head module includes candidate proportional coefficients, candidate integral coefficients, and candidate derivative coefficients, comprising:

[0080] An improved CSPNet model is constructed, including a node-level cross-stage module, a physical model module, and a hierarchical output header module, wherein:

[0081] The improved CSPNet model is constructed using a modular, serial structure. It uses the vacuum-deposited variable graph as input to connect node-level cross-stage modules. At the output end, the feature tensor is fed directly into the physical model module as an intermediate output. The features processed by the physical model module are then returned as input and serialized to the node-level cross-stage modules. The network is stacked alternately in the depth direction to form a unified backbone network. The output end of the backbone network is connected to the hierarchical output head module, which maps the output of the backbone network to form the overall structure of the improved CSPNet model.

[0082] The node feature sequence and topological connection relationship of the vacuum coating variable graph are input into the node-level cross-stage module. At each graph node, the node features are split into main branch features and bypass branch features according to the channel dimension. The main branch features are sequentially subjected to one-dimensional convolution and residual operation, and the bypass branch features are subjected to linear transformation. At the exit of each cross-stage unit, the main branch features and bypass branch features are subjected to channel splicing and channel compression convolution to obtain node-level features. Specifically, the node-level features are obtained as follows:

[0083] The vacuum coating variable graph is used as the graph structure input. The time window length is determined based on the control sampling period and the dominant time constants of vacuum degree and process gas flow of the vacuum coating equipment. The time window length is selected by response analysis and simulation experiments of historical process data to cover the number of sampling points of the three closed-loop adjustment time constants. Within the time window, the multidimensional time series data of each graph node is organized into a node feature sequence according to the time dimension and channel dimension, and input into the node-level cross-stage module along with the node set and edge set in the variable graph topology.

[0084] In the node-level cross-stage module, for each graph node, the node features received at the current layer are divided into two parts according to the channel dimension: main branch features and bypass branch features. The main branch features are used for processing by the convolutional residual unit, and the bypass branch features are used for cross-stage splicing and retention. The main branch features pass through several one-dimensional convolutional units and residual units along the time dimension. The convolution kernel size, convolution stride and number of convolutional layers of the one-dimensional convolution are determined by offline experiments on historical data based on the typical rise time, oscillation period and sampling period of vacuum degree and process gas flow rate, so that the convolution receptive field covers the main dynamic change range. The residual unit adds the convolution output to the input of the same size to obtain the main branch residual output features.

[0085] The bypass branch features are linearly transformed using a one-dimensional one-to-one convolution. The number of output channels of the one-to-one convolution is determined by offline participation and computational complexity constraints based on the number of channels of the main branch residual output features and the target compression ratio, so that the bypass branch matches the main branch residual output features in terms of the number of channels and numerical scale.

[0086] At the current cross-stage unit exit, the residual output features of the main branch and the bypass branch features after linear transformation are concatenated in the channel dimension to obtain the concatenated feature tensor. Then, a channel compression convolution consisting of one-to-one convolution, normalization and activation function is applied. The number of output channels of the channel compression convolution is set through multiple sets of comparative experiments based on the input dimension of the graph convolutional layer and the overall model complexity to form the node-level features of the graph node in the current layer.

[0087] Based on the topology of the vacuum coating variable graph, graph convolution and adjacency aggregation operations are performed on the node-level features of each graph node. The first node-level features of adjacent nodes are aggregated to obtain multivariate coupled operating condition features that simultaneously represent the coupling relationship between vacuum degree, process gas flow rate, valve opening, control quantity and process set value. Channel splitting, main branch convolution extraction, bypass branch retention and graph structure aggregation operations are repeatedly performed along the network depth direction to form cross-stage feature extraction of graph structure and obtain operating condition feature representation.

[0088] The physical model module input is formed by combining the operating condition feature representation with the physical parameters, control variables, and process setpoints of vacuum coating. A discrete update equation based on the dynamic relationship between vacuum degree and process gas flow rate is used to calculate the vacuum degree and process gas flow rate within the prediction time step, yielding physical prediction features. These physical prediction features are then concatenated with the operating condition feature representation along the channel dimension and processed by channel compression convolution. At the node level, graph structure feature extraction and physical prediction feature fusion are alternately performed between the cross-stage module and the physical model module to obtain the final feature representation of the current vacuum coating operating condition. Specifically, obtaining the final feature representation of the current vacuum coating operating condition involves:

[0089] The operating condition characteristics are combined with the physical parameters, control quantities, and process settings of vacuum coating to form the input of the physical model module. The physical parameters include the effective volume of the vacuum chamber, the flow resistance coefficient of the pipeline, the rated pumping speed of the pump, and the leakage coefficient corresponding to the actual equipment structure. The control quantity is the valve control command output to the vacuum coating valve execution platform within the current control sampling period. The process setting is the target vacuum degree corresponding to the current process stage. By splicing the operating condition characteristics with the physical parameters, control quantities, and process settings in the feature dimension, the physical model input characteristics describing the state and driving conditions of the vacuum coating system are obtained.

[0090] The physical model module establishes discrete update equations based on the input characteristics of the physical model. The discrete update equations are based on mass conservation and flow balance. They express the change in the amount of gas in the vacuum chamber within a unit time interval as the difference between the gas inflow and gas outflow. Then, based on the relationship between the effective volume of the vacuum chamber and the gas state, the change in the amount of gas is converted into the change in vacuum degree. The continuous time form is discretized by using the sampling period to obtain the discrete update relationship. The prediction time step is selected by comparing the simulation of different prediction step configurations on historical data based on the dominant time constant of the vacuum system, the control response time and the computational resource constraints. This ensures that the prediction time range covers the short-term dynamic change range that the controller needs to pay attention to.

[0091] By using discrete update equations, starting from the current moment, the vacuum level and process gas flow rate of each prediction step are iteratively calculated within the preset prediction time step. The vacuum level and process gas flow rate sequences of each prediction step are combined in time order to form physical prediction features of future time vacuum dynamic changes. The physical prediction features are then linearly transformed to reduce dimensionality, and the number of channels is matched with the representation of operating conditions.

[0092] The physical prediction features and the corresponding working condition features are concatenated in the channel dimension to obtain the fused feature tensor. Channel compression convolution operation is applied to the fused feature tensor, and the channel dimension is compressed and rearranged to generate the fused intermediate features. The intermediate features are used as the input of the next layer node-level cross-stage module, so that the node-level cross-stage module and the physical model module alternately perform graph structure feature extraction and physical prediction feature fusion in the network depth direction. After multiple alternations, the final feature representation of the current vacuum coating working condition is obtained.

[0093] The final feature representation is pooled and concatenated in the time dimension and graph node dimension to obtain the global operating condition feature vector. The global operating condition feature vector is then input into the hierarchical output head module. In the coarse parameter layer, the first fully connected sub-network is used to generate multiple sets of coarse PID parameter prototypes. In the intermediate parameter layer, the second fully connected sub-network is used to locally modify each coarse PID parameter prototype to generate intermediate PID parameters. In the fine parameter layer, the third fully connected sub-network is used to output candidate PID parameters, which include candidate proportional coefficients, candidate integral coefficients, and candidate derivative coefficients.

[0094] In this embodiment, the step of calculating the control error based on the process setpoint and the real-time vacuum measurement value, and calculating the control quantity according to the candidate proportional coefficient, candidate integral coefficient, and candidate derivative coefficient, includes:

[0095] Within each control sampling cycle, acquire the vacuum process setting value and real-time vacuum measurement value of the corresponding process stage, subtract the real-time vacuum measurement value from the vacuum process setting value to obtain the current control error, and record the current control error, the control error of the previous sampling cycle, and the control error of the previous sampling cycle.

[0096] The candidate proportional coefficient, candidate integral coefficient, and candidate derivative coefficient are multiplied by the current control error and the historical control error respectively, and then weighted and summed to obtain the control increment for the current sampling period. The control increment is then added to the control quantity of the previous sampling period to obtain the control quantity for the current control period.

[0097] In this embodiment, obtaining the updated PID parameters includes:

[0098] In each control sampling cycle, the current control error, error threshold, vacuum level, fluctuation threshold, and current process stage identifier are read. The absolute value of the current control error is compared with the error threshold, the fluctuation index of the vacuum level is compared with the fluctuation threshold, and the process stage identifier is used to determine whether a process stage switch has occurred.

[0099] If the absolute value of the current control error is greater than any of the preset error thresholds, or if the vacuum fluctuation index is greater than any of the fluctuation thresholds, or if a process stage identifier switch is detected, a PID parameter update event is determined to have occurred. If none of the three conditions are met, a PID parameter update event is determined not to have occurred.

[0100] When a PID parameter update event is detected, the candidate proportional coefficient, candidate integral coefficient, and candidate derivative coefficient, along with performance-related data consisting of the current control error, historical control error, and control input, are input into the P-channel adaptive unit, I-channel adaptive unit, and D-channel adaptive unit. The P-channel adaptive unit generates the updated proportional coefficient based on the candidate proportional coefficient, error amplitude, and response speed information. The I-channel adaptive unit generates the updated integral coefficient based on the candidate integral coefficient, steady-state error, and error accumulation information. The D-channel adaptive unit generates the updated derivative coefficient based on the candidate derivative coefficient, error change rate, and output oscillation information. Where:

[0101] The P-channel adaptive unit extracts the absolute value of the error, the maximum error, the average error, and the rate of change of the error from the current control error and the historical control error. It concatenates these with the candidate proportional coefficients to form the proportional channel input vector. The vector is then mapped through several layers of fully connected networks to obtain the proportional gain adjustment. Finally, it is combined with the candidate proportional coefficients in a weighted manner to generate the updated proportional coefficients.

[0102] The I-channel adaptive unit calculates the cumulative error, average deviation, and continuous deviation direction from the current control error and historical control error, and concatenates them with candidate integral coefficients to form the integral channel input vector. It then outputs the integral gain adjustment through a fully connected network and combines it with the candidate integral coefficients to obtain the updated integral coefficients.

[0103] The D-channel adaptive unit uses the difference between the current control error and the control error of the previous sampling period, along with the vacuum degree, to concatenate the candidate differential coefficients into a differential channel input vector. This vector is then passed through a fully connected network to obtain the differential gain adjustment, which is then combined with the candidate differential coefficients to generate updated differential coefficients.

[0104] When it is determined that no PID parameter update event has occurred, the proportional coefficient, integral coefficient, and derivative coefficient of the previous control sampling period are directly used as the PID parameters of the current control sampling period, and the PID parameters are not updated.

[0105] When a PID parameter update event is detected, the updated proportional coefficient, integral coefficient, and derivative coefficient output by the P-channel adaptive unit, I-channel adaptive unit, and D-channel adaptive unit are written into the PID controller as new PID parameters.

[0106] In this embodiment, the closed-loop regulation of vacuum degree and process gas flow rate includes:

[0107] At the beginning of each control sampling cycle, the proportional coefficient, integral coefficient, and derivative coefficient currently in use are read from the PID controller. The vacuum process setpoint is subtracted from the real-time vacuum measurement value at the same moment to obtain the current control error. The current control error, the control error of the previous sampling cycle, and the control error of the previous sampling cycle are recorded.

[0108] The proportional, integral, and derivative coefficients are used to calculate the control error, resulting in the proportional control quantity, integral control quantity, and derivative control quantity. The proportional control quantity is equal to the product of the proportional coefficient and the current control error. The integral control quantity is equal to the product of the integral coefficient and the sum of the current control error and the historical control error over time. The derivative control quantity is equal to the product of the derivative coefficient and the ratio of the difference between the current control error and the control error of the previous sampling period. The proportional control quantity, integral control quantity, and derivative control quantity are added together to obtain the control quantity for the current control sampling period.

[0109] The control quantity is output as a valve control command to the vacuum coating valve execution platform, driving the vacuum coating valve opening to increase or decrease, and the vacuum degree and process gas flow rate in the vacuum coating chamber are adjusted in a closed loop around the vacuum degree process set value within the continuous control sampling cycle.

[0110] Example 1:

[0111] To verify the feasibility of this invention in practice, it was applied to a process upgrade project for an 8-inch power device PECVD vacuum coating production line. The production line has high requirements for segmented pressure control within the 0-1000 Torr range and the flow ratio of various process gases in multiple stages, including chamber evacuation, stabilization coating, gas switching, and venting. The original system used traditional PID control valves and manual experience-based tuning. This frequently resulted in significant pressure overshoot during the evacuation stage, slow pressure rise during the stabilization coating stage, and large fluctuations in process gas flow rate when there were changes in formula switching, carrier loading, or pipeline pressure loss. Some batches even required multiple trial coatings and parameter adjustments to meet thickness uniformity and refractive index requirements, leading to long setup times, significant raw material waste, and difficulty in improving the overall equipment uptime. Without replacing the chamber and sensors, the original vacuum valve control cabinet was upgraded by adding the multi-source data acquisition unit of this invention, an improved CSPNet intelligent controller, and an adaptive PID execution unit. This enables intelligent adaptive control of vacuum level and process gas flow rate under complex and time-varying conditions.

[0112] In practical applications on the production line, vacuum level, process gas flow rate, valve opening, control input, and process setpoints are collected in real time by field sensors and controllers. After preprocessing, these data are used to construct a variable graph reflecting the coupling relationship between the vacuum chamber, the pump, and the intake branch within the controller. This graph is then used to input an improved CSPNet model. The model is online to extract and fuse features through a node-level cross-stage structure and an embedded vacuum dynamics physical model, resulting in the final feature representation of the current operating condition. A set of candidate PID parameters is then generated by the hierarchical output head. The controller calculates the control error based on the process setpoint and vacuum level measurement. It first calculates the initial control input using the candidate PID parameters and then uses an event detection module to determine whether parameter updates are triggered. When an error continuously exceeds the limit, the vacuum level fluctuates more, or a process stage switch is detected, the P, I, and D three-channel adaptive units are activated to correct the proportional, integral, and derivative coefficients, respectively, resulting in updated PID parameters that are written to the PID controller. Within the sampling period, the valve drive is calculated according to the updated PID parameters, enabling the pumping stage to quickly approach the set pressure, the coating stage to maintain the target vacuum level, and the gas switching stage to transition smoothly, while reducing the impact of frequent small parameter adjustments on system stability.

[0113] To verify the beneficial effects of this invention, comparative experiments were conducted on multiple batches of production data for the same product and formulation in a PECVD production line from March to May 2025. After offline training using historical data, the system was directly run online. By comparing the adjustment time to reach the target vacuum level during the evacuation stage, the overshoot and steady-state fluctuation of the chamber pressure during the stabilization coating stage, the short-term fluctuation of the process gas flow rate, the setup time after formulation switching, and the yield rate, it was observed that the system using this invention provides more stable vacuum and gas flow control under conditions of multi-stage processes, varying carrier loads, and long-term continuous operation. The time to reach the set value is significantly shortened, steady-state deviation and fluctuation are significantly reduced, the number of setups is reduced, and the overall yield rate and equipment utilization are improved.

[0114] Table 1 Comparison of Valve Control Performance under PECVD Vacuum Coating

[0115] Indicator Items Traditional PID control Traditional adaptive PID Fuzzy PID control Method of the present invention Evacuation time (s) 1.30 0.88 0.80 0.60 Pressure overshoot (%) 9.5 5.6 4.9 3.2 Pressure deviation (mTorr) 1.9 1.2 1.0 0.6 Pressure fluctuation (mTorr) 0.85 0.50 0.42 0.25 Traffic fluctuation (%) 2.8 1.7 1.4 0.9 Setup time (min) 30.0 20.5 17.3 11.8 Number of trial plating batches (batch / month) 24 16 12 7 Yield rate (%) 92.1 95.0 96.1 97.8

[0116] As shown in Table 1, the intelligent control of this invention has the most significant advantages in terms of pumping speed and pressure overshoot. The pumping time of traditional PID is 1.30s, that of traditional adaptive PID is 0.88s, and that of fuzzy PID is 0.80s, while this invention further reduces the pumping time to 0.60s. In terms of pressure overshoot, the traditional PID is 9.5%, that of traditional adaptive PID is 5.6%, and that of fuzzy PID is 4.9%, while this invention reduces it to 3.2%. At the same time, the pressure deviation is also reduced from 1.9 mTorr of traditional PID to 0.6 mTorr. This indicates that under the same process conditions, this invention can achieve the target vacuum level faster and get closer to the set value.

[0117] In terms of process stability, this invention also outperforms the comparative methods in terms of pressure and flow rate fluctuations. Regarding pressure fluctuations, the traditional PID method has a fluctuation of 0.85 mTorr, the traditional adaptive PID method has a fluctuation of 0.50 mTorr, and the fuzzy PID method has a fluctuation of 0.42 mTorr, while this invention further reduces it to 0.25 mTorr. Regarding flow rate fluctuations, the traditional PID method has a fluctuation of 2.8%, the traditional adaptive PID method has a fluctuation of 1.7%, and the fuzzy PID method has a fluctuation of 1.4%, while this invention controls it to 0.9%. This indicates that by using the improved CSPNet model to predict parameters and combining it with three-channel adaptive updates, the fluctuations in vacuum degree and process gas flow rate during the stable coating stage are significantly reduced, resulting in a smoother and more stable process curve.

[0118] From an engineering benefit perspective, this invention brings significant advantages in terms of setup time, number of trial plating operations, and yield. Setup time has been reduced from 30.0 min with traditional PID to 20.5 min with traditional adaptive PID and 17.3 min with fuzzy PID, while this invention further reduces it to 11.8 min. The number of trial plating batches per month has decreased from 24 to 16 with traditional adaptive PID and 12 with fuzzy PID, while this invention only requires 7 batches. The yield has increased from 92.1% with traditional PID to 95.0% with traditional adaptive PID and 96.1% with fuzzy PID, while this invention achieves 97.8%. These data demonstrate that this invention, while ensuring improved control performance, effectively reduces parameter adjustment and trial plating costs, significantly enhancing the stable output capability of the production line.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A vacuum coating valve intelligent control system based on an adaptive PID algorithm, characterized in that, The method comprises the following modules: A multi-source process data acquisition module is configured to acquire multi-source process data and perform preprocessing to form multivariate time series data; A variable graph construction module is configured to map the multivariate time series data to a vacuum coating variable graph according to physical structure relationships and coupling relationships; An improved CSPNet model parameter generation module is configured to receive the vacuum coating variable graph and output candidate proportional coefficients, candidate integral coefficients and candidate differential coefficients through cross-stage feature extraction and working condition prediction; An initial control quantity generation module is configured to calculate an initial control quantity according to a control error and the candidate proportional coefficients, candidate integral coefficients and candidate differential coefficients; An adaptive PID parameter update module is configured to correct the candidate PID parameters through a three-channel adaptive method according to the initial control quantity to obtain updated PID parameters; A PID control execution module is configured to recalculate the control quantity according to the updated PID parameters and drive a vacuum coating valve execution platform to perform closed-loop intelligent adjustment on the vacuum degree and the process gas flow.

2. The intelligent control method of the vacuum coating valve based on the adaptive PID algorithm is applied to the intelligent control system of the vacuum coating valve based on the adaptive PID algorithm in claim 1, characterized in that, The method comprises the following steps: Acquiring multi-source process data of a vacuum coating device and performing preprocessing to obtain multivariate time series data; According to the physical structure relationship of the controlled object, the variables in the multivariate time series data are taken as graph nodes, and the variable graph topology structure is established according to the coupling relationship between the variables to obtain a vacuum coating variable graph; An improved CSPNet model is constructed, the vacuum coating variable graph is taken as input, the cross-stage feature extraction of the graph structure and the working condition prediction of the embedded physical model are performed to obtain the final feature representation of the current vacuum coating working condition, and the candidate PID parameters, including the candidate proportional coefficient, the candidate integral coefficient and the candidate differential coefficient, are obtained by the hierarchical output head module; The control error is calculated according to the process set value and the real-time vacuum degree measurement value, and the control quantity is calculated according to the candidate proportional coefficient, the candidate integral coefficient and the candidate differential coefficient; The control error and the control quantity are input into the adaptive PID algorithm, and it is determined whether to update the parameters through event detection, and when it is determined that the parameters need to be updated, the P-channel adaptive, I-channel adaptive and D-channel adaptive are used to update the candidate PID parameters to obtain the updated PID parameters; The control quantity is recalculated according to the updated PID parameters, and the control quantity is output to the vacuum coating valve execution platform to perform closed-loop adjustment on the vacuum degree and the process gas flow. 3.The intelligent control method of vacuum coating valve based on adaptive PID algorithm according to claim 2, characterized in that, The multi-source process data comprises a vacuum degree, a process gas flow, a valve opening, a control quantity and a process set value.

4. The intelligent control method of vacuum coating valve based on adaptive PID algorithm according to claim 2, characterized in that, The multivariate time series data is obtained by: Acquiring multi-source original process data through a vacuum degree sensor, a process gas flow meter, a valve opening feedback device and a controller acquisition interface arranged on the vacuum coating device; Setting a unified sampling period for the multi-source original process data, resampling and time aligning the signals according to the sampling time axis, and performing abnormal data rejection, filtering and smoothing processing; Scaling or standardizing the variables, and combining the processed vacuum degree, process gas flow, valve opening, control quantity and process set value according to each sampling time to construct the multivariate time series data.

5. The intelligent control method of vacuum coating valve based on adaptive PID algorithm according to claim 2, characterized in that, The obtained vacuum coating variable graph comprises: Each type of variable in the multivariate time series data is defined as a graph node, and each graph node is assigned a unique identifier; According to the physical transmission relationship between pressure and process gas flow, pressure and valve opening, process gas flow and valve opening, pressure and pump current, and pressure and control quantity, it is judged whether there is a direct coupling relationship between each variable, a graph edge is established between two graph nodes with a direct coupling relationship, and no graph edge is established between two graph nodes without a direct coupling relationship, to obtain a variable graph topology structure comprising a graph node set and a graph edge set; The time series data corresponding to each graph node is taken as node features, and the variable graph topology structure is associated to form a vacuum coating variable graph comprising node identifiers, node feature sequences, and connection relationships between nodes.

6. The intelligent control method of vacuum coating valve based on adaptive PID algorithm according to claim 2, characterized in that, The candidate PID parameter obtained by the hierarchical output head module includes a candidate proportional coefficient, a candidate integral coefficient, and a candidate differential coefficient, which comprises: An improved CSPNet model is constructed, including a node-level cross-stage module, a physical model module, and a hierarchical output head module; The node feature sequences and topological connection relationships of the vacuum coating variable graph are input into the node-level cross-stage module, the node features are split into main branch features and bypass branch features according to the channel dimension at each graph node, one-dimensional convolution and residual operation are performed on the main branch features in turn, linear transformation is performed on the bypass branch features, and channel splicing and channel compression convolution are performed on the main branch features and the bypass branch features at the outlet of each cross-stage unit to obtain node-level features; According to the topological structure of the vacuum coating variable graph, graph convolution and adjacency aggregation operations are performed on the node-level features of each graph node, the node-level first features of adjacent nodes are aggregated to obtain multivariate coupling working condition features representing the coupling relationship between vacuum degree, process gas flow, valve opening, control quantity, and process set value, and channel splitting, main branch convolution extraction, bypass branch reservation, and graph structure aggregation operations are repeatedly performed along the network depth direction to form cross-stage feature extraction of the graph structure, to obtain working condition feature representation; The working condition feature representation, the vacuum coating physical parameters, the control quantity, and the process set value are combined to form the input of the physical model module, the discrete update equation of the dynamic relationship between the vacuum degree and the process gas flow is used to calculate the vacuum degree and the process gas flow in the prediction time step to obtain physical prediction features, the physical prediction features and the working condition feature representation are spliced in the channel dimension and processed by channel compression convolution, and the graph structure feature extraction and the physical prediction feature fusion are alternately performed between the node-level cross-stage module and the physical model module to obtain the final feature representation of the current vacuum coating working condition. The final feature representation is pooled and spliced in the time dimension and the graph node dimension to obtain a global working condition feature vector, and the global working condition feature vector is input into a hierarchical output head module, a first fully connected subnetwork is used to generate a plurality of groups of coarse level PID parameter prototypes at a coarse level parameter layer, a second fully connected subnetwork is used to locally correct each coarse level PID parameter prototype to generate a medium level PID parameter at a medium level parameter layer, and a third fully connected subnetwork is used to output a candidate PID parameter at a fine level parameter layer. The candidate PID parameter includes a candidate proportional coefficient, a candidate integral coefficient and a candidate differential coefficient.

7. The intelligent control method of vacuum coating valve based on adaptive PID algorithm according to claim 2, characterized in that, The control error is calculated according to the process set value and the real-time vacuum degree measurement value, and the control amount is calculated according to the candidate proportional coefficient, the candidate integral coefficient and the candidate differential coefficient, including: In each control sampling period, the vacuum degree process set value and the real-time vacuum degree measurement value corresponding to the process stage are obtained, the vacuum degree process set value is subtracted from the real-time vacuum degree measurement value to obtain the current control error, and the current control error, the last sampling period control error and the last sampling period control error are recorded; The candidate proportional coefficient, the candidate integral coefficient and the candidate differential coefficient are multiplied by the current control error and the historical control error respectively and are weighted and summed to obtain the control increment of the current sampling period, and the control increment is added to the control amount of the last sampling period to obtain the control amount of the current control period. 8.The intelligent control method of vacuum coating valve based on adaptive PID algorithm according to claim 2, characterized in that, The updated PID parameter is obtained, including: In each control sampling period, the current control error, the error threshold, the vacuum degree, the fluctuation threshold and the current process stage identifier are read, the absolute value of the current control error is compared with the error threshold, the fluctuation index of the vacuum degree is compared with the fluctuation threshold, and whether the process stage switching occurs is determined according to the process stage identifier; If the absolute value of the current control error is greater than any one of the preset error threshold, or the fluctuation index of the vacuum degree is greater than any one of the fluctuation threshold, or the process stage identifier is detected to switch, it is determined that the PID parameter updating event occurs, and when the three conditions are not met, it is determined that the PID parameter updating event does not occur; When it is determined that the PID parameter updating event occurs, the candidate proportional coefficient, the candidate integral coefficient and the candidate differential coefficient are input into the P channel adaptive unit, the I channel adaptive unit and the D channel adaptive unit together with the performance related data composed of the current control error, the historical control error and the control amount, the P channel adaptive unit generates the updated proportional coefficient based on the candidate proportional coefficient and the error amplitude and the response speed information, the I channel adaptive unit generates the updated integral coefficient based on the candidate integral coefficient and the steady state error and the error accumulation information, and the D channel adaptive unit generates the updated differential coefficient based on the candidate differential coefficient and the error change rate and the output oscillation information; When it is determined that the PID parameter updating event does not occur, the proportional coefficient, the integral coefficient and the differential coefficient of the last control sampling period are directly used as the PID parameters of the current control sampling period, and the PID parameters are not updated. When determining that a PID parameter updating event occurs, write the updated proportional coefficient, integral coefficient and differential coefficient output by the P channel adaptive unit, the I channel adaptive unit and the D channel adaptive unit into the PID controller as new PID parameters. 9.The intelligent control method of vacuum coating valve based on adaptive PID algorithm according to claim 2, characterized in that, The closed-loop regulation of the vacuum degree and the process gas flow comprises: At the beginning of each control sampling period, read the currently used proportional coefficient, integral coefficient and differential coefficient from the PID controller, subtract the real-time vacuum degree measured value at the same time from the vacuum degree process set value to obtain a current control error, and record the current control error, the last sampling period control error and the last sampling period control error; Use the proportional coefficient, the integral coefficient and the differential coefficient to operate the control error, calculate the proportional control amount, the integral control amount and the differential control amount, the proportional control amount is equal to the product of the proportional coefficient and the current control error, the integral control amount is equal to the product of the integral coefficient and the sum of the current control error and the historical control error in time, and the differential control amount is equal to the product of the differential coefficient and the ratio of the current control error minus the last sampling period control error, and add the proportional control amount, the integral control amount and the differential control amount to obtain the control amount of the current control sampling period; Output the control amount as a valve control instruction to the vacuum coating valve execution platform, drive the vacuum coating valve opening to increase or decrease, and perform closed-loop regulation of the vacuum degree and the process gas flow in the vacuum coating cavity around the vacuum degree process set value in the continuous control sampling period.

Images