Self-regulation method and system for coating machine in failure stage

Abstract

The application discloses a kind of coating machine in fault stage self-regulation method and system, and the application relates to the technical field of self-regulation, carries out fault analysis to multidimensional state data, determines current fault type and fault severity in fault analysis process, and generates corresponding fault feature vector;Fault feature vector is input to the preset self-regulation decision space, in the self-regulation decision space, optimization matching is carried out in combination with coating process knowledge graph, and the accuracy of self-regulation strategy is improved.Process parameters and actuating mechanism of coating machine are dynamically reconstructed and synergistically compensated, and the running state of the coating machine after real-time monitoring and control is obtained, so as to obtain corresponding state feedback data, and determine the control efficiency of the coating machine in combination with the working history and current working data of the coating machine, realize online iterative correction of self-regulation decision space, until the coating machine returns to normal state.

Description

Technical Field

[0001] This invention relates to the field of autonomous control technology, and in particular to an autonomous control method and system for a coating machine during a fault phase. Background Technology

[0002] With the increasing application of vacuum coating technology in high-end optical devices, semiconductor chips, and precision machining, the requirements for film thickness uniformity, adhesion, and process consistency are becoming increasingly stringent. As a core process equipment, the vacuum coating machine involves multiple physical field coupling processes such as heat and mass transfer, plasma physics, gas discharge, and mechanical transmission under high-temperature vacuum conditions.

[0003] When a coating machine encounters a fault, it typically uses preset fixed rules or simple PID feedback adjustment, failing to construct an autonomous control decision space driven by reinforcement learning and coating process mechanism. Due to the lack of a dynamic knowledge graph directly driven by fault feature vectors, the system struggles to perform global strategy optimization and matching under complex operating conditions, resulting in insufficient accuracy of the output autonomous control strategy. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and system for autonomous control of a coating machine during a fault phase.

[0005] This invention provides a method for autonomous control of a coating machine during a fault phase, comprising: Based on a multi-source heterogeneous sensor network deployed at key nodes of the coating machine, multi-dimensional state data during the coating process is collected in real time. Fault analysis is performed on the multi-dimensional state data. During the fault analysis, the current fault type and fault severity are determined, and the corresponding fault feature vector is generated. The fault feature vector is input into the preset autonomous control decision space, and the coating process knowledge graph is combined in the autonomous control decision space to perform optimization matching, thereby determining the corresponding autonomous control strategy. Based on the identification of autonomous control strategies, multiple sub-control commands are determined and sent to the controller of the coating machine. The process parameters and actuators of the coating machine are dynamically reconstructed and coordinated to compensate for control, so as to maintain the process stability of the coating machine during the fault transition phase. The system monitors the operating status of the coating machine in real time after adjustment, thereby obtaining the corresponding status feedback data. It then combines the working history of the coating machine with the current working data to determine the adjustment efficiency of the coating machine. Based on the adjustment efficiency results, the autonomous adjustment decision space is iteratively corrected online until the coating machine returns to normal.

[0006] This invention provides an autonomous control system for a coating machine during a fault phase. This system is applied to the aforementioned autonomous control method for a coating machine during a fault phase. The autonomous control system for the coating machine during a fault phase includes: The fault analysis module is used to collect multi-dimensional state data in real time during the coating process based on a multi-source heterogeneous sensor network deployed at key nodes of the coating machine, perform fault analysis on the multi-dimensional state data, determine the current fault type and severity during the fault analysis process, and generate the corresponding fault feature vector. The control strategy module is used to input the fault feature vector into the preset autonomous control decision space, and to perform optimization matching in the autonomous control decision space by combining the coating process knowledge graph, so as to determine the corresponding autonomous control strategy. The compensation control module is used to identify multiple sub-control commands based on the autonomous control strategy, and send the multiple sub-control commands to the controller of the coating machine to dynamically reconstruct and coordinate the process parameters and actuators of the coating machine to maintain the process stability of the coating machine during the fault transition phase. The performance result module is used to monitor the operating status of the coating machine after regulation in real time, thereby obtaining the corresponding status feedback data. It combines the working history of the coating machine and the current working data to determine the regulation performance result of the coating machine. Based on the regulation performance result, the autonomous regulation decision space is iteratively corrected online until the coating machine returns to normal.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) Based on the multi-source heterogeneous sensor network deployed at key nodes of the coating machine, multi-dimensional state data during the coating process is collected in real time. Fault analysis is performed on the multi-dimensional state data. During the fault analysis, the current fault type and fault severity are determined, and the corresponding fault feature vector is generated. The fault feature vector is input into the preset autonomous control decision space. In this autonomous control decision space, the coating process knowledge graph is combined to perform optimization matching, thereby determining the corresponding autonomous control strategy. The introduction of fault feature vector further controls the autonomous control decision space and the coating process knowledge graph, improving the accuracy of the autonomous control strategy.

[0008] (2) Based on the identification of autonomous control strategy, multiple sub-control commands are determined and sent to the controller of the coating machine. The process parameters and actuators of the coating machine are dynamically reconstructed and coordinated to maintain the process stability of the coating machine during the fault transition stage. The operating status of the coating machine after control is monitored in real time to obtain the corresponding status feedback data. The control efficiency result of the coating machine is determined by combining the working history of the coating machine and the current working data. The multiple sub-control commands are further controlled, and the control efficiency result of the coating machine is fully considered. The autonomous control decision space is realized to perform online iterative correction until the coating machine returns to normal. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the autonomous control method of the coating machine during the fault stage in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the autonomous control system of the coating machine in the fault stage according to an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 2 An autonomous control method for a coating machine during a fault phase, applied to autonomous control scenarios; the autonomous control method for a coating machine during a fault phase includes: Step S11: Based on the multi-source heterogeneous sensor network deployed at key nodes of the coating machine, multi-dimensional state data during the coating process is collected in real time. Fault analysis is performed on the multi-dimensional state data. During the fault analysis, the current fault type and fault severity are determined, and the corresponding fault feature vector is generated. Step S12: Input the fault feature vector into the preset autonomous control decision space, and perform optimization matching in the autonomous control decision space in combination with the coating process knowledge graph to determine the corresponding autonomous control strategy. Step S13: Based on the identification of the autonomous control strategy, determine multiple sub-control instructions, and send the multiple sub-control instructions to the controller of the coating machine to dynamically reconstruct and coordinately compensate the process parameters and actuators of the coating machine to maintain the process stability of the coating machine during the fault transition phase. Step S14: Monitor the operating status of the coating machine after adjustment in real time to obtain the corresponding status feedback data, and determine the adjustment efficiency result of the coating machine by combining the working history of the coating machine and the current working data. Based on the adjustment efficiency result, perform online iterative correction of the autonomous adjustment decision space until the coating machine returns to normal.

[0012] In step S11, the specific steps are as follows: S111: Mark multiple key nodes of the coating machine. Multiple key nodes are deployed with corresponding multi-source heterogeneous sensor networks. The multi-source heterogeneous sensor network collects multi-dimensional status data in real time during the coating process. The multi-dimensional status data includes acoustic emission signals, thermal field distribution data and electrical parameters. S112: Utilize edge computing nodes to perform temporal alignment and cross-modal feature fusion on multidimensional state data. During the fusion stage, extract the correlation features of multidimensional state data in spatial topology and the evolution features in time series. Combine this with a multi-source tracing mechanism to determine the corresponding fault source and thus determine the current fault type. S113: Dynamically evaluate based on the spread trend of the fault source and historical operating data to output the corresponding fault severity. Then, encode the fault severity and fault type using high-dimensional tensor quantization to generate the corresponding fault feature vector, which includes temporal features, spatial features and severity probability distribution.

[0013] In the embodiments of this application, multiple key nodes of the coating machine are marked, and a corresponding multi-source heterogeneous sensor network is deployed at the multiple key nodes. The multi-source heterogeneous sensor network collects multi-dimensional state data in real time during the coating process. The multi-dimensional state data includes acoustic emission signals, thermal field distribution data and electrical parameters, thus introducing multi-dimensional state data.

[0014] At this point, for the key node marking and multi-source heterogeneous sensor network deployment stage, based on the physical structure of the coating machine and the heat and mass transfer mechanism of the process, the coating machine system is divided into four core areas: vacuum acquisition and maintenance node, target source / evaporation source excitation node, substrate transfer and movement node, and process gas distribution node.

[0015] At the vacuum acquisition and maintenance node, Pirani gauges and cold cathode ionization complexation sensors are marked and deployed on the chamber walls and at each level of the extraction port; at the target source excitation node, Hall current sensors and high-frequency voltage probes are marked and deployed on the cooling water circuit and target power supply feed line of the magnetron sputtering target holder; at the substrate transmission node, photoelectric encoders and triaxial accelerometers are marked and deployed on the end of the planetary carrier drive motor shaft; at the process gas distribution node, piezoelectric pressure sensors are marked and deployed on the downstream pipeline of the mass flow controller; furthermore, acoustic emission sensor arrays and infrared thermal imager probes are marked and deployed in the top-view blind area and side wall flanges inside the chamber, thereby constructing a multi-source heterogeneous sensor network covering mechanical, electrical, thermal, and acoustic multiple physical fields.

[0016] For the real-time acquisition of multidimensional state data, the sensor network deployed above synchronously acquires multidimensional state data during the coating process at a preset acquisition frequency, specifically including acoustic emission signals, thermal field distribution data, and electrical parameters. For the acquisition of acoustic emission signals, the acoustic emission sensor array captures broadband elastic stress wave signals generated by the evolution of microcracks in the target material, film peeling, and abnormal arc discharge at a megahertz sampling rate.

[0017] For the acquisition of thermal field distribution data, the infrared thermal imager probe and fiber optic temperature sensor network acquire two-dimensional matrix data of the target surface temperature gradient field distribution, substrate holder thermal radiation distribution, and cavity wall thermal stress field distribution by low-frequency periodic scanning.

[0018] For the acquisition of electrical parameters, Hall current sensors and voltage probes use high-frequency synchronous sampling to acquire transient waveforms of sputtering target current and voltage, transient spikes of arc discharge, and power feeder impedance fluctuation data. The raw analog signals acquired by each sensor node are amplified, filtered, and converted from analog to digital by a pre-amplifier signal conditioning circuit. These digital signals are then aggregated in real time to the edge computing node via industrial Ethernet or high-speed fieldbus. This completes the synchronous acquisition and pre-processing of multi-source heterogeneous and multi-dimensional state data, providing a high-fidelity underlying data foundation for subsequent cross-modal feature fusion and fault source tracing.

[0019] Specifically, during normal operation of the coating machine, a flow rate reduction fault occurred due to local scaling in the cooling water circuit of the magnetron sputtering target. In the early stage of the fault evolution, the infrared thermal imager in the thermal field distribution data first captured the abnormal thermal field distribution in a local area of ​​the target surface and identified local high-temperature hot spots caused by uneven cooling of the target surface. As the local thermal stress increased sharply, the acoustic emission sensor array collected high-frequency acoustic emission signals of microcracks induced by uneven thermal expansion of the target material in real time.

[0020] Meanwhile, due to the increased thermionic emission on the target surface caused by local high temperature, and the abnormal magnetic field distribution, the electrical parameter acquisition system captured high-frequency spike fluctuations in the target current and abnormal transient voltage drops. In this scenario, the multi-dimensional state data of "thermal field hot spot distribution - acoustic high-frequency stress wave - electrical transient fluctuation" synchronously and in real time collected by the multi-source heterogeneous sensor network together constitute the complete physical manifestation of the target cooling abnormality fault, and is aggregated in real time to trigger the subsequent fault analysis and feature vector generation process.

[0021] Furthermore, edge computing nodes are used to perform temporal alignment and cross-modal feature fusion on multidimensional state data. During the fusion stage, the correlation features of multidimensional state data in spatial topology and the evolution features in time series are extracted. Combined with a multi-source tracing mechanism, the corresponding fault source is determined, and the current fault type is determined, which fully considers the fault source.

[0022] At this point, the computing nodes deployed at the edge of the coating machine receive the converged multidimensional state data stream, perform microsecond-level synchronization calibration of the sampling clock of each sensor node based on a high-precision hardware timestamp protocol, and use piecewise linear interpolation or dynamic time warping algorithms to precisely align heterogeneous data with different sampling frequencies, such as acoustic emission signals, thermal field distribution data, and electrical parameters, on the time axis.

[0023] The time-aligned multidimensional data is input into a pre-defined hybrid neural network model. The temporal convolutional network is used to extract the evolution features of the multidimensional state data in the time series, such as identifying the time delay correlation between the slow drift of the temperature field and the transient burst of acoustic stress waves. At the same time, the graph neural network or spatial convolutional network is used, with the physical topology of the key components of the coating machine as graph nodes and the heat transfer and electrical connection relationship between components as edges, to extract the correlation features of the multidimensional state data in the spatial topology, such as the heat conduction path from the hot spot distribution of the thermal field to adjacent components and the propagation coupling features of electrical parameter fluctuations in the feeder network, thereby outputting a cross-modal fusion feature matrix containing spatiotemporal joint representation.

[0024] The aforementioned cross-modal fusion feature matrix is ​​input into a preset multi-source inference engine, which integrates temporal source inference based on temporal causal reasoning, spatial source inference based on topological networks and graph reasoning, and parametric source inference mechanism based on multivariate correlation and partial least squares regression.

[0025] Among them, temporal tracing establishes the origin and propagation time chain of the fault event by analyzing the sequence of changes in various state parameters; spatial tracing locates the physical origin of the abnormal signal based on the three-dimensional topology map of the coating machine, and determines the physical node where the fault first occurred; parameter tracing reveals the fundamental physical variable causing the anomaly by decoupling the mutual interference between multiple variables. Through the comprehensive cross-validation and confidence fusion of this triple tracing mechanism, the accompanying disturbances caused by cascading effects are eliminated, and the fundamental physical location or functional module causing the current anomaly is accurately located, thereby outputting a unique fault source identifier and its corresponding confidence assessment parameters.

[0026] Based on the fault source identifiers determined by multiple source tracing and their associated cross-modal fusion feature matrices, these are input into a fault classification model pre-trained on a large historical fault sample database and equipment mechanism model. This model uses lightweight multilayer perceptron or support vector machine classification algorithms to map the fused features to a pre-constructed coating machine fault category space, which covers categories such as mechanical wear, vacuum leakage, electrical breakdown, thermal control failure, and process gas anomalies. The classification model calculates the membership degree between the current feature and the standard feature patterns of various faults, and outputs the fault category with the highest probability distribution, thus completing the accurate mapping from the underlying physical fault source to the high-level semantic fault type.

[0027] Specifically, during the operation of the coating machine, a local scaling failure occurred in the cooling water circuit of the magnetron sputtering target; the edge computing node performs microsecond-level time alignment and cross-modal fusion on the target surface local high temperature hot spot data collected by the infrared thermal imager, the microcrack high frequency stress wave data collected by the acoustic emission sensor, and the target current burr fluctuation data collected by the Hall sensor.

[0028] During the fusion process, the temporal convolutional network extracted the temporal evolution characteristics of the thermal field temperature slowly rising to a certain threshold, followed by an acoustic stress wave burst and transient current fluctuation. The graph neural network extracted the spatial correlation characteristics of the thermal field anomaly being strictly limited to a local area of ​​the target material and spreading to the cooling water inlet side. The multiple source tracing mechanism was then activated: temporal source tracing determined that the temperature anomaly occurred earlier than the acoustic and electrical anomalies, confirming that the thermal field failure was the first event; spatial source tracing, combined with the internal pipeline topology map of the coating machine, located the source of the anomaly to the magnetron target cooling water circuit rather than the vacuum chamber or substrate holder; parameter source tracing decoupled the negative correlation between the decrease in cooling water flow and the increase in target temperature; the multiple source tracing comprehensively confirmed that the source of the fault was "local scaling of the magnetron target cooling water leading to a decrease in flow".

[0029] The classification model accurately classifies the fault source as a "target source thermal control failure" fault type based on the fusion feature pattern of "local thermal field anomaly - thermal stress-induced microcracks - increased thermionic emission leading to abnormal current," thus providing a precise decision-making basis for subsequent severity assessment and autonomous control strategy generation. Therefore, dynamic assessment is performed based on the fault source's diffusion trend and historical operating data to output the corresponding fault severity. The fault severity and fault type are then encoded using high-dimensional tensor quantization to generate a corresponding fault feature vector. This fault feature vector includes temporal features, spatial features, and a severity probability distribution, thus introducing the fault feature vector.

[0030] At this point, after identifying the root cause of the fault through multiple tracing mechanisms, the edge computing node calls the embedded reduced-order multiphysics dynamics model of the coating machine. Using the actual physical state parameters of the current fault source as the initial boundary conditions, and combining the multiphysics coupling mechanism, such as the heat conduction equation, the fluid network equation and the electromagnetic field equation, the state disturbance propagation path and time evolution trajectory of the fault source on the spatial topology network are calculated, thereby constructing a predictive sequence of the fault propagation trend.

[0031] Meanwhile, edge computing nodes retrieve historical operating condition data records from local storage units that are the same as the current fault source or in the same wear cycle stage, extract the fault evolution time sequence characteristics and critical failure threshold parameters under historical operating conditions, and align the current real-time fault evolution data with the fault propagation trajectory in historical operating conditions by performing similarity measurement and time scale scaling based on dynamic time warping. By using historical prior distribution to perform data-driven online deviation correction on the propagation trend predicted based on the mechanism model, high-fidelity forward prediction of the dynamic evolution process of faults spreading from local origin to global spread can be achieved.

[0032] Based on the above dynamically corrected fault propagation trend prediction results, a multi-dimensional severity quantification index system is introduced. This index system covers the real-time deviation of the current fault source from the normal operating range, the maximum predicted impact on "key process evaluation indicators, such as film thickness uniformity deviation rate and deposition rate attenuation ratio" within the preset future prediction time window, and the remaining safety margin time before reaching the equipment hard protection shutdown threshold. By inputting the above multi-dimensional prediction indexes into a preset fuzzy logic comprehensive evaluator or a kernel function-based regression model, continuous values ​​representing the current degree of fault deterioration are output.

[0033] To characterize the confidence level of the assessment results under the conditions of sensor noise and model uncertainty, the continuous numerical value is further mapped to a severity estimate that follows a specific probability distribution based on the statistical variance of historical operating data and the quality assessment results of current real-time data. The output includes severity probability distribution parameters containing the expected mean and variance.

[0034] Distributed representation learning technology is used to perform dense vector embedding mapping on discrete fault type labels. By vectorizing the topological distance between the fault type node and its adjacent fault mechanism node and component relationship node in the pre-trained coating machine fault knowledge graph space, a low-dimensional dense embedding vector representing the semantic connotation of the fault type is generated. At the same time, statistical feature extraction and multiquantile discretization bucket encoding are performed on the severity probability distribution containing the expected mean and variance to generate an equal-length severity probability distribution vector.

[0035] The temporal feature matrix representing the evolutionary characteristics of multidimensional state data in time series, extracted by the hybrid neural network model in step S112 above, and the spatial feature matrix representing the spatial topological correlation characteristics, are concatenated with the above-mentioned fault type embedding vector and severity probability distribution vector to generate a high-dimensional tensor-quantized fault feature vector with a unified dimension that integrates the evolutionary laws of the time dimension, the coupling correlation of the spatial dimension, and the global deterioration risk quantification assessment.

[0036] Specifically, the coating machine has been traced back to the source of a local scaling failure in the cooling water circuit of the magnetron sputtering target. The failure source has been classified as "target source thermal control failure". During the dynamic evaluation phase, the edge computing node calls the target material heat conduction and cooling fluid network model. Using the currently measured decrease in cooling water inlet flow rate and the local target surface temperature gradient as boundary conditions, it predicts that the local thermal blockage effect caused by scaling will spread along the target substrate to the sputtering surface. It is expected that within the next ten minutes, the overall target surface temperature will cross the critical threshold point of a sharp increase in thermionic emission.

[0037] Searching historical operating condition records for the same scaling stage in the database revealed that the historical evolution trajectory indicated frequent arc discharges would occur approximately 8 minutes after the current flow deviation rate. Combining mechanism prediction with historical data correction, the expected mean of the current fault severity was 0.75 (out of 1.0), with a variance of 0.02, indicating a stage where the fault is on the verge of severe deterioration but has not yet triggered hard shutdown protection. In the coding stage, the discrete fault type "target source thermal control failure" was mapped to a knowledge graph to obtain its dense embedding vector. The severity mean and variance were discretized into probability distribution vectors, and tensor concatenated with the temporal evolution feature matrix reflecting "slow temperature rise followed by acoustic stress wave bursts and transient current fluctuations" extracted in step S112, and the spatial correlation feature matrix reflecting "thermal field anomalies confined to a local area of ​​the target material and spreading towards the cooling water inlet side." This resulted in a high-dimensional fault feature vector containing complete physical evolution, serving as a precise input data base for subsequent autonomous control decision-making space optimization matching.

[0038] In step S12, the specific steps are as follows: S121: Input the fault feature vector into the crop and map it to the preset autonomous control decision space. The autonomous control decision space is constructed by a dynamic knowledge graph driven by reinforcement learning and coating process mechanism. The nodes of the dynamic knowledge graph include process parameter nodes, equipment status nodes and control action nodes. The edge weights between nodes represent the strength of process causal relationship and the confidence of historical control success. S122: Graph attention network is used to determine the matching degree between the fault feature vector and each node in the dynamic knowledge graph. In the local subgraph region where the matching degree reaches the preset threshold, the strategy optimization mechanism is triggered. The action space is explored in the subgraph region through the strategy gradient mechanism. Combined with the current process constraints, the expected benefits and system safety margin of each candidate control action are determined, and then the corresponding autonomous control strategy is output. The autonomous control strategy includes the process parameter adjustment range and the action sequence of the actuator.

[0039] In the embodiments of this application, the fault feature vector is used as the input quantity and mapped to a preset autonomous control decision space. The autonomous control decision space is constructed by a dynamic knowledge graph driven by reinforcement learning and coating process mechanism. The nodes of the dynamic knowledge graph include process parameter nodes, equipment status nodes and control action nodes. The edge weights between nodes represent the strength of process causal relationship and the confidence of historical control success. The dynamic knowledge graph is introduced to construct the space.

[0040] At this point, the "high-dimensional tensor fault feature vector: a concatenated tensor of the i.e., sequential feature matrix, spatial feature matrix, fault type embedding vector, and severity probability distribution vector" generated in step S113 above is used as the input data base and input to the pre-set autonomous control decision space within the edge computing node. This mapping process adopts a feature space mapping algorithm based on the attention mechanism. By calculating the dot product correlation score between the input fault feature vector and the preset state nodes in the decision space, the high-dimensional tensor is projected onto the corresponding state subspace region of the decision space. The Softmax normalization function is used to generate the initial membership probability distribution of the current fault feature vector in each state node in the decision space, thereby completing the accurate semantic mapping from the underlying data features to the state of the high-level decision space, and determining the state search anchor point for subsequent graph-based optimization matching.

[0041] The core of this autonomous control decision space consists of a dynamic knowledge graph. The construction of this graph adopts a dual-drive paradigm that integrates mechanism priors and data-driven approaches. On the mechanism-driven side, based on the multi-physics field coupling mechanism model of heat and mass transfer, gas discharge and thin film growth of the coating machine, hard physical constraint boundaries between process parameters and equipment status are established, such as the positive causal transmission relationship between target power loading and chamber temperature rise.

[0042] On the reinforcement learning-driven side, deep Q-networks or near-end policy optimization algorithms are used to continuously learn through trial and error in simulation environments and historical control data. Successful control experiences are transformed into connection topologies between graph nodes. The nodes of this dynamic knowledge graph are strictly divided into three types of entities: "process parameter nodes" representing adjustable physical quantities, such as argon flow rate setpoint, target power limit, and substrate revolution speed; "equipment status nodes" representing the equipment's operating conditions, such as current vacuum level range, target temperature gradient state, and plasma impedance state; and "control action nodes" representing actuator action commands, such as increasing cooling water pump speed, triggering pulse arc extinguishing mode, and reducing substrate transmission duty cycle. The nodes in the graph are connected by directed edges. The edge weights from equipment status nodes to control action nodes are characterized by the Q-value or policy probability output by the reinforcement learning algorithm, representing their necessity. The edge weights from control action nodes to process parameter nodes are characterized by the physical influence coefficients by the mechanistic model, thus forming a mesh decision topology with a closed loop of physical causal logic.

[0043] Each connection edge in the dynamic knowledge graph is assigned a dynamic, time-varying composite weight, which is formed by the weighted fusion of the first weight component and the second weight component. The first weight component represents the strength of the process causal relationship, which is calculated based on the multiphysics mechanism model and historical operating data through partial least squares regression or sensitivity analysis. It is used to quantify the magnitude and direction of the influence of a certain control action node on a specific process parameter node, such as the reduction coefficient of the deposition rate due to a 1% decrease in target power.

[0044] The second weight component represents the confidence level of historical control success. It is updated based on the feedback results of the actual control effectiveness during the online iteration process. When a certain control action successfully maintains the stability of the process under a specific state and does not cause secondary risks, the confidence level of its corresponding edge weight increases. Conversely, when it causes process oscillation or causes new failures, the confidence level decreases. Through this composite edge weight mechanism, the knowledge graph can dynamically eliminate control paths that fail due to equipment aging or process drift, and strengthen the control strategy link that is most suitable and robust within the current coating machine life cycle.

[0045] Specifically, the coating machine has confirmed a local scaling failure in the magnetic target cooling water circuit and outputs a fault feature vector containing the type "target source thermal control failure" and a severity of 0.75. In step S121, the fault feature vector is input into the autonomous control decision space and mapped by the attention mechanism to locate the equipment status node in the graph that represents "local overheating and thermal stress accumulation on the target surface".

[0046] Around this state node, the dynamic knowledge graph activates multiple candidate regulatory action nodes connected to it based on the joint driving logic of reinforcement learning and mechanism model: the first is the node of "reducing the average power of the target and increasing the pulse turn-off time", the second is the node of "reducing the substrate revolution speed to compensate for the decrease in deposition rate", and the third is the node of "increasing the process argon flow rate to enhance the plasma cooling effect".

[0047] At this point, the edge weights connecting these action nodes to process parameter nodes such as "target temperature" and "film thickness uniformity" come into play. For example, the edge weight from "reducing target power" to "target temperature decrease" has extremely high process causal correlation strength, based on the heat conduction mechanism, and has extremely high historical control success confidence because historical records show that this action has successfully suppressed thermal runaway multiple times, resulting in a large second weight component. On the other hand, although the edge weight from "increasing argon flow rate" to "target temperature decrease" has a certain causal strength, its historical control success confidence is low because it has caused arc initiation risk in historical records. Through this dynamic weight mechanism, the graph prioritizes guiding the control action node "reducing target average power and increasing pulse turn-off time" with high-confidence links at the current severity level of 0.75. This provides a highly reliable topology search foundation for subsequent steps S12 to combine the knowledge graph for optimization matching and finally determine the optimal autonomous control strategy.

[0048] Furthermore, a graph attention network is used to determine the matching degree between the fault feature vector and each node in the dynamic knowledge graph. A strategy optimization mechanism is triggered in the local subgraph region where the matching degree reaches a preset threshold. The action space is explored in the subgraph region through the strategy gradient mechanism. The expected benefits and system safety margin of each candidate control action are determined in combination with the current process constraints. Then, the corresponding autonomous control strategy is output. The autonomous control strategy includes the process parameter adjustment range and the action sequence of the actuator. The fault feature vector is introduced to further control the autonomous control decision space and the coating process knowledge graph, thereby improving the accuracy of the autonomous control strategy.

[0049] At this point, the edge computing node uses the high-dimensional tensor fault feature vector generated in the previous steps as a query vector and inputs it into the pre-trained graph attention network. The graph attention network uses the "each node: process parameter node, equipment status node, and control action node" in the dynamic knowledge graph as key vector and value vector. By calculating the dot product correlation between the query vector and the key vector of each node and normalizing it through the Softmax function, it outputs the attention weight coefficient of each node, thereby representing the dynamic matching degree between the current fault feature vector and each node in the graph.

[0050] The system performs threshold discrimination on the matching degree, filters out the set of nodes whose attention weight coefficient is greater than the preset matching threshold, and extracts the local subgraph region containing the set of nodes and their adjacent first-order and second-order nodes based on the original topological edge connection relationship of the graph. Within this local subgraph region, the system formally triggers the strategy optimization mechanism, which converges the subsequent strategy search space from the global graph to the local subgraph range that is strongly related to the current fault situation, thereby significantly reducing the dimension of the action space and improving the real-time performance of decision response.

[0051] Within a defined local subgraph region, the system employs a policy gradient-based reinforcement learning algorithm to explore the action space. The "policy network, such as an actor network based on Gaussian distribution," takes the state characteristics and edge weights of each node within the local subgraph region as input and outputs the probability distribution of candidate control actions and their continuous parameter combinations that can be executed in the current state. During this exploration, the system introduces multi-dimensional constraints based on the coating process mechanism, such as the thermal limit temperature of the vacuum chamber, the maximum current carrying capacity of the target power supply, and the minimum stable rotation speed of the substrate transmission mechanical structure. Action combinations that violate these constraints are masked through a masking mechanism or subject to a large penalty coefficient, ensuring that the exploration process occurs within safe physical boundaries.

[0052] For candidate control actions that meet the constraints, the system uses the commentator network combined with the strength of process causal relationships and the confidence of historical control success in the aforementioned dynamic knowledge graph to predict the "expected process stability benefits, such as film thickness deviation convergence" of the system during the fault transition phase after executing the action combination, as well as the system safety margin after execution, i.e. the distance of the predicted state from the hard protection threshold, thus forming an evaluation system with the joint goal of maximizing expected benefits and optimizing safety margin.

[0053] Based on the joint objective evaluation results of candidate control action combinations using the policy gradient mechanism, the system adopts a non-maximum greedy strategy or a multi-objective optimization screening method based on the Pareto front to select the action combination with the highest expected return and the system safety margin meeting the preset safety standard from the candidate action set as the optimal control strategy.

[0054] Furthermore, the system decodes the optimal control strategy into a blueprint of specific control instructions at the underlying level. This blueprint clearly defines the "adjustment range of process parameters, such as the upper and lower limits of the flow setpoint of the mass flow controller, the percentage reduction of the target power supply and the pulse frequency adjustment range" and the "action sequence of the actuator, such as the timing execution link of first adjusting the process gas, then reducing the pressure and finally changing the substrate speed." This outputs a structured autonomous control strategy, providing a direct decision-making basis for the decomposition and issuance of multiple sub-instructions in subsequent steps.

[0055] Specifically, the coating machine has been identified as experiencing a "local scaling in the magnetron target cooling water circuit" fault, with a severity of 0.75, and has been mapped to a device status node in the dynamic knowledge graph representing "local overheating and thermal stress accumulation on the target surface." In step S122, the graph attention network calculates the matching degree between this high-dimensional fault feature vector and adjacent control action nodes in the graph, such as "reducing the average target power," "activating the pulse arc extinguishing mode," "reducing the substrate speed," and "increasing the cooling water pump speed." It is found that the matching weights of the first three with the current thermal stress and the decline in film quality all exceed the preset threshold of 0.8. Therefore, a local subgraph containing the above nodes is extracted, and an optimization mechanism is triggered. Within the subgraph region, a strategy gradient mechanism explores different action combinations: during the exploration process, the current process constraints of the coating machine are considered, such as the substrate speed must not be lower than the minimum process threshold to prevent abnormal film crystallinity, and the violation action of "substrate speed dropping to extremely low" is eliminated.

[0056] The expected benefits of the candidate action "reducing the target power by 20% and simultaneously activating the pulse arc extinguishing mode" are assessed as convergence of film thickness fluctuations and reduction of thermal stress. The system safety margin is assessed as having a 30% temperature rise margin between the target temperature and the hard trip threshold. Taking into account both expected benefits and safety margin, the system ultimately outputs an autonomous control strategy. This strategy includes adjusting the process parameters by reducing the average power of the target power supply by 15% to 20% and extending the pulse turn-off time to 50 microseconds. The actuator action sequence is to prioritize triggering the target power supply control mode switch and smoothly reduce the target power setpoint within 2 seconds, thus completing the strategy optimization and output.

[0057] In step S13, the specific steps are as follows: S131: Dynamically identify the autonomous control strategy, determine multiple autonomous control items during the identification process, and convert them in combination with the current working data of the coating substrate to output the corresponding multi-level control sequence. The multi-level control sequence includes multiple first-level sub-control instructions and multiple second-level sub-control instructions. Input each sub-control instruction into the model predictive controller, and further combine the dynamic mechanism information of the coating process to make rolling predictions on the film deposition rate and uniformity of future multi-steps, and dynamically optimize and adjust the heating power, working gas flow rate and bias parameters to achieve process parameter adjustment. S132: Sends each second-level sub-control command to the coating machine controller, drives the baffle position adjustment mechanism, substrate rotation mechanism or spare target material switching mechanism to perform hardware-level redundancy reconstruction. During execution, a cross-coupling compensation controller is introduced to perform real-time collaborative compensation for the timing error and nonlinear coupling interference between process parameter adjustment and mechanical action of the actuator, so as to suppress the sudden change of process parameters during the fault transition stage and maintain the dynamic balance of the vacuum environment.

[0058] In the embodiments of this application, the autonomous control strategy is dynamically identified. During the identification process, multiple autonomous control items are determined and converted in combination with the current working data of the coating substrate to output a corresponding multi-level control sequence. The multi-level control sequence includes multiple first-level sub-control instructions and multiple second-level sub-control instructions. Each sub-control instruction is input into the model predictive controller, and the dynamic mechanism information of the coating process is further combined to predict the film deposition rate and uniformity of future multi-steps. The heating power, working gas flow rate and bias voltage parameters are dynamically optimized and adjusted to achieve process parameter adjustment, thus introducing process parameter adjustment.

[0059] At this point, after receiving the autonomous control strategy output in step S122, the edge computing node calls the strategy parsing engine to perform semantic recognition and structured decomposition of the strategy to determine the multiple autonomous control items contained in the strategy, such as target power limiting items, pulse parameter change items, substrate transmission speed compensation items, etc.

[0060] The system acquires real-time operating data of the coating machine, including the current chamber vacuum baseline, the current actual output power of the target power supply, the current actual revolution speed of the substrate holder, and the current process gas flow rate setpoint. It uses these autonomously adjustable parameters as target setpoints and calculates the dynamic deviation between them based on the current operating data. Based on this deviation, the system converts it into a corresponding multi-level control sequence using a preset kinematic trajectory planning or proportional-integral-derivative control law. This multi-level control sequence is clearly divided into multiple first-level sub-control commands and multiple second-level sub-control commands. The first-level sub-control commands represent macroscopic fast-variable control commands that play a decisive role in system safety, such as commands to directly cut off or significantly limit target power output, or commands to trigger electrical protection modes, used to curb fault deterioration in the shortest possible time. The second-level sub-control commands represent slow-variable control commands that play a dominant role in process quality compensation, such as commands to gradually compensate for substrate speed changes, or commands to fine-tune gas flow rates, used to maintain process stability after fault suppression.

[0061] The multiple first-level and second-level sub-control commands generated above are synchronously input into the model prediction controller deployed on the edge node. The model prediction controller has an embedded dynamic mechanism information model based on the heat and mass transfer dynamics of the coating process, plasma sheath physics and gas discharge mechanism to reflect the nonlinear coupling relationship between process parameters and thin film growth.

[0062] The controller uses the above-mentioned sub-control commands as control variables, takes the current real-time state data as the initial condition, performs forward multi-step rolling prediction within the preset future prediction time window, calculates and outputs the predicted trajectory of the internal state variables of the chamber within a finite number of future steps, and then predicts and evaluates the fluctuation trend of film deposition rate and the evolution trajectory of film thickness distribution uniformity deviation in the future multi-step process, thereby realizing a forward-looking quantitative evaluation of the control effect before the actual issuance of control actions.

[0063] Based on the rolling prediction results of the above model predictive controller, the controller constructs a constrained optimization solution model with the objective function of minimizing the future multi-step film deposition rate and uniformity deviation, and with the constraints of the heating power safety limit, the working gas flow rate adjustment rate limit, and the bias voltage limit range. By solving this optimization problem in each control cycle, the optimal control action increment sequence at the current moment is obtained, thereby dynamically optimizing and adjusting "heating power, for example, reducing the substrate heating power to compensate for the overall thermal balance when the cooling capacity decreases", "working gas flow, for example, fine-tuning the argon flow rate to maintain plasma density and thus stabilize the deposition rate", and "substrate bias parameters, for example, adjusting the bias voltage to optimize film adhesion and uniformity", ensuring that the process parameters are always in the optimal compensation state during the fault transition phase.

[0064] Specifically, the coating machine has been diagnosed with a localized scaling fault in the magnetically controlled target cooling water circuit and has output an autonomous control strategy including "reducing the average target power by 15% to 20%, extending the pulse off-time, and compensating for substrate speed." In step S131, the system identifies this strategy and determines three autonomous control items: "reducing target power," "changing pulse parameters," and "compensating for substrate speed." Based on the current operating data (current target power 10kW, current substrate speed 10rpm), the system generates a multi-level control sequence: The first-level sub-control command is to "smoothly reduce the average target power from 10kW to 8.2kW within 500 milliseconds and extend the pulse turn-off time from 20 microseconds to 50 microseconds" in order to quickly suppress thermal stress and arc extinguishing; The second-level sub-control command is to "gradually reduce the substrate speed from 10 rpm to 9 rpm over the next 5 seconds" to compensate for the decrease in deposition rate caused by the power reduction; these commands are input into the model prediction controller, which is based on the plasma discharge and deposition rate mechanism model embedded in the controller.

[0065] If only the first and second level instructions are executed, the film deposition rate will decrease by 12% within the next 10 seconds due to the decrease in argon ionization, and the film thickness uniformity deviation will increase to ±4% due to the change in thermal field distribution. To avoid this deterioration, the model predictive controller is optimized and dynamically generates a third adjustment action under the constraint of the upper limit of gas flow rate adjustment, namely, "synchronously increasing the working gas argon flow rate from 50 sccm to 55 sccm to improve plasma density" and "simultaneously fine-tuning the substrate bias voltage from -100V to -110V to improve ion bombardment energy distribution", thereby offsetting the decrease in deposition rate and maintaining film thickness uniformity, and finally completing the intelligent reconstruction and collaborative compensation adjustment of process parameters under the local scaling fault state.

[0066] Furthermore, each second-level sub-control command is sent to the controller of the coating machine to drive the baffle position adjustment mechanism, substrate rotation mechanism or spare target material switching mechanism to perform hardware-level redundancy reconstruction. During the execution process, a cross-coupling compensation controller is introduced to perform real-time collaborative compensation for the timing error and nonlinear coupling interference between process parameter adjustment and mechanical action of the actuator, so as to suppress the sudden change of process parameters during the fault transition stage and maintain the dynamic balance of the vacuum environment, fully considering the dynamic balance.

[0067] At this time, the second-level sub-control instruction sequence generated by the edge computing node after dynamic optimization by the model predictive controller is sent to the underlying programmable logic controller or motion control unit of the coating machine in the form of high real-time messages via industrial Ethernet or high-speed fieldbus. This second-level sub-control instruction is not a simple parameter setpoint, but a mechanical execution logic that includes the target physical position, motion speed and acceleration / deceleration curve.

[0068] After receiving the command, the controller drives each redundant actuator to perform physical reconstruction: for the baffle position adjustment mechanism, the opening and closing angle or deflection position of the shielding baffle in the vacuum chamber is adjusted by stepper motor or servo cylinder to compensate for the uneven spatial distribution of film thickness caused by local faults through physical shielding; for the substrate rotation mechanism, the orbital and rotational speed ratio of the planetary carrier or single-axis tray is changed by adjusting the output frequency or torque setpoint of the servo driver, and the deposition flux is averaged by reconstructing the spatial motion trajectory of the substrate; for the standby target switching mechanism, when the main target fails due to a fault and the process cannot be maintained, the magnetic coupling driver or linear guide inside the chamber is triggered to translate or flip the standby target in the standby position to the sputtering station, realizing the hardware-level switching of the physical deposition source.

[0069] Since the response time constants of electrical parameters, such as target power and gas flow rate, are typically in the millisecond range, while the action delay and inertial time constants of mechanical actuators, such as baffle flipping and base speed change, are typically in the hundreds of milliseconds or even seconds, there will inevitably be significant timing errors between the two during the fault transition phase. To address this, the system introduces a cross-coupled compensation controller. This controller uses the expected process state trajectory predicted by the aforementioned model as a reference benchmark, collects the actual position feedback of the mechanical mechanism and the actual response values ​​of the electrical parameters in real time, and constructs a timing error evaluation function.

[0070] Based on this, in response to the nonlinear coupling interference caused by mechanical motion lag, such as local gas flow field disturbance and vacuum pressure fluctuation caused by baffle movement, which in turn affects plasma discharge stability, the controller designs a feedforward-feedback composite control law to send feedforward compensation increments to the mass flow meter and target power supply in advance at the same time as the mechanical motion starts, so as to offset the foreseeable mechanical-process coupling disturbance, and achieve millisecond-level timing alignment and nonlinear decoupling compensation between electronic control and mechanical motion.

[0071] Through the closed-loop regulation of the aforementioned cross-coupled compensation controller, the system applies active damping to "key process parameters, such as deposition rate and chamber pressure" during the parallel process of hardware reconstruction and parameter adjustment. When a sudden change in the pumping resistance inside the vacuum chamber or a sudden change in the venting area is detected due to mechanical action, the compensation controller dynamically adjusts the opening of the vacuum pump valve or the fine-tuning flow rate of the process gas to form a dynamic pressure balance closed loop in the vacuum environment. This constrains the peak value of process parameter mutations caused by the heterogeneous actions of multiple actuators during the fault transition phase within the tolerance window allowed by the process, avoiding film composition segregation or adhesion failure caused by step drop or overshoot of process parameters.

[0072] Specifically, the coating machine has experienced a local scaling failure in the cooling water circuit of the magnetron target. The aforementioned S131 step has reduced the average power of the target and extended the pulse off time through the first-level sub-control command, but the prediction shows that the deposition rate will still decrease and result in insufficient film thickness. In the S132 step, the system sends the second-level sub-control command to the bottom controller as "the substrate holder speed is gradually reduced from 10 rpm to 9 rpm" and "the baffle in the trigger cavity is deflected by 15 degrees to block the non-uniform deposition area".

[0073] During execution, the mechanical flipping of the baffle lags behind the electrical decrease in target power, and the displacement of the baffle changes the airflow distribution in the cavity, causing transient fluctuations in vacuum pressure and nonlinear jumps in discharge impedance. At this time, the cross-coupled compensation controller is activated. It calculates the actual position lag of the baffle motor in real time, and at the instant the baffle moves, it feeds forward to compensate for the increments to "mass flow meter: fine-tune argon flow rate by 2 sccm" and "bias power supply: fine-tune substrate bias voltage -5V" to compensate for the local decrease in plasma density caused by the baffle obstruction.

[0074] Meanwhile, in response to the micro-leakage of the bearing gas seal caused by the substrate rotation deceleration, the controller dynamically adjusts the opening of the pumping speed valve of the front vacuum pump to maintain the dynamic balance of the chamber working pressure at 0.8Pa. Through the above-mentioned hardware-level redundant reconstruction and cross-coupling collaborative compensation, the coating machine smoothly completed the mechanical and electrical reconstruction during the scaling fault transition stage, successfully suppressed the transient changes in deposition rate and vacuum degree, maintained the process stability during the fault transition stage, and provided a stable system control base for the performance monitoring and online correction of the decision space in the subsequent S14 step.

[0075] In step S14, the specific steps are as follows: S141: Real-time monitoring of the operating status of the coating machine after regulation, outputting corresponding status feedback signals during the monitoring process, determining the status feedback data of the coating machine after regulation based on the analysis of the status feedback signals, synchronizing the status feedback data to the digital twin space of the coating machine, dynamically evolving the status feedback data, and determining the regulation efficiency result of the coating machine by combining the working history of the coating machine and the current working data. The regulation efficiency result includes the film quality retention rate, equipment thermal stress change and energy consumption fluctuation characteristics. S142: If the control efficiency result meets the preset efficiency threshold, the current control is determined to be effective, and the fault characteristics and control process are used as experience playback data to iteratively update the edge weights of the knowledge graph and the parameters of the reinforcement learning network in the autonomous control decision space online, so as to guide the coating machine to maintain normal status.

[0076] In the embodiments of this application, the operating status of the coating machine after regulation is monitored in real time. During the monitoring process, a corresponding status feedback signal is output. The status feedback data of the coating machine after regulation is determined based on the analysis of the status feedback signal. The status feedback data is synchronized to the digital twin space of the coating machine. The status feedback data is dynamically evolved. The regulation efficiency result of the coating machine is determined by combining the working history of the coating machine and the current working data. The regulation efficiency result includes the film quality retention rate, equipment thermal stress change and energy consumption fluctuation characteristics, thus introducing the regulation efficiency result.

[0077] At this point, after the actuator completes hardware-level redundancy reconstruction and parameter collaborative compensation, the system continuously calls the multi-source heterogeneous sensor network deployed in the aforementioned S111 step to monitor the operating status of the regulated coating machine vacuum system, target source excitation system, substrate transmission system, and process gas distribution system in real time at a preset high-frequency sampling rate.

[0078] Each sensor converts the collected physical quantities, such as target surface temperature residual, cavity vacuum fluctuation, actual substrate rotation speed, target current-voltage transient waveform, and actual bias feedback value, into digital or analog state feedback signals. The edge computing node uses a signal conditioning module and data parsing protocol to filter and reduce noise, perform unit conversion and scaling on the state feedback signal, and extracts quantitative state feedback data that characterizes the current physical operation of the equipment, ensuring that the data truly reflects the equipment's operating boundary conditions after the intervention of the autonomous control strategy.

[0079] The system relies on the Industrial Internet or high-speed fieldbus to synchronize the parsed status feedback data to the digital twin space of the coating machine deployed in the cloud or local edge server in a low-latency mode. The digital twin space is embedded with a high-fidelity mechanism model of multi-physics coupling that is isomorphic to the physical coating machine. After receiving the real-time status feedback data, the twin space uses the data as dynamic boundary conditions and initial input to drive the internal finite element analysis and computational fluid dynamics model to perform dynamic evolution calculations with high step size.

[0080] This dynamic evolution process not only reconstructs the real-time thermal, electromagnetic, and flow field distributions of the current equipment in virtual space, but also predicts the evolution trajectory of the internal state variables of the equipment within a limited time window under the current control command through state space deduction. Thus, it realizes a high-fidelity mirror mapping and forward-looking deduction of the operating status of the physical coating machine after control in the virtual dimension.

[0081] The system retrieves the coating machine's operational history data stored in the local database. This operational history data includes the cumulative runtime of the equipment since installation and commissioning, the life wear curves of "key components, such as the target material, vacuum pump, and sealing rings," historical maintenance records, and archives of each failure evolution. The system compares and performs a goodness-of-fit analysis of the predicted trajectory output by the twin space dynamic evolution with the currently collected real-time operational data, and compares it with the baseline health status and historical evolution trends of similar failures in the aforementioned operational history data.

[0082] Through a pre-set performance evaluation algorithm model, the system comprehensively and quantitatively calculates the actual performance of the current autonomous control strategy in suppressing fault deterioration, maintaining process stability, and ensuring equipment safety. Finally, it outputs quantitative control performance results, which specifically include "film quality retention rate: characterizing the degree to which the film thickness uniformity and deposition rate deviate from the target value after control", "equipment thermal stress change: characterizing the thermal stress growth rate and thermal fatigue accumulation increment of key components during the fault transition stage", and "energy consumption fluctuation characteristics: characterizing the fluctuation variance and abnormal increment of the overall system energy consumption during degraded operation or redundant reconfiguration". This provides multi-dimensional quantitative evaluation basis for online correction of the decision space.

[0083] Specifically, the coating machine has experienced a local scaling failure in the magnetic target cooling water circuit, and the system has executed the first and second level control commands to reduce the average power of the target and simultaneously activate the pulse arc extinguishing mode, as well as trigger the baffle deflection and substrate deceleration.

[0084] The multi-source sensor network monitors the equipment status in real time after regulation and outputs status feedback signals such as "target temperature slowly drops from 85℃ to 72℃ and the fluctuation converges", "cavity vacuum stabilizes at 0.8Pa and there is no abnormal gas outflow fluctuation", "substrate speed is steadily maintained at 9rpm", and "baffle position sensor feedback deflection angle is 15 degrees". This data is synchronized to the digital twin space. The twin space uses the target temperature of 72℃ and the baffle deflection of 15 degrees as boundary conditions to dynamically calculate the changing trend of plasma density distribution in the cavity and thermal stress field on the target surface.

[0085] Furthermore, the system retrieved the working history data of the coating machine and found that the equipment had been running for 3000 hours, and the target material's service life was in the middle to late stage. Historical records showed that this stage was extremely sensitive to thermal stress. Combining the current working data after adjustment, the efficiency evaluation model calculated that the film quality retention rate after adjustment was 94%, the film thickness deviation was controlled within ±4%, which met the process tolerance, and the equipment thermal stress change index showed that the local thermal stress on the target surface decreased and no new microcracks were generated. The cumulative increase in thermal fatigue did not exceed the threshold. However, the energy consumption fluctuation characteristics showed that the instantaneous reactive power increased due to the pulse arc extinguishing mode and the baffle action, and the energy consumption fluctuation variance increased by 12%. Based on the above indicators, the system determined that the current autonomous control strategy was effective in maintaining process stability and equipment safety, but there was still room for optimization in energy efficiency control, thus providing a clear direction for the iterative correction of the subsequent decision space.

[0086] Furthermore, if the control effectiveness meets the preset effectiveness threshold, the current control is deemed effective. The fault characteristics and control process are used as experience playback data to iteratively update the edge weights of the knowledge graph and the parameters of the reinforcement learning network in the autonomous control decision space online, so as to guide the coating machine to maintain normal status. This realizes online iterative correction of the autonomous control decision space until the coating machine returns to normal status.

[0087] At this point, the edge computing node extracts the "regulation efficiency results" output from the aforementioned step S141: namely, film quality retention rate, equipment thermal stress change, and energy consumption fluctuation characteristics, into the efficiency evaluation comparator. The comparator is preset with a multi-dimensional efficiency threshold range established based on coating process standards and equipment safety specifications, including the "lower limit threshold" for film quality retention rate, such as not less than 90%, the "safety limit threshold" for equipment thermal stress change, such as thermal stress increment being less than 10% of the material yield strength, and the "tolerance upper limit" for energy consumption fluctuation characteristics, such as energy consumption fluctuation variance being less than 15% of the rated value.

[0088] If and only if all three performance indicators fall within the preset threshold range, the system logic determines that the current autonomous control strategy has successfully overcome the fault disturbance and maintained the stability of the process transition phase. Then, it triggers the effective control flag and removes the emergency intervention state of the system, guiding the coating machine to maintain a normal long-term operating state according to the reconstructed process parameters and hardware topology.

[0089] After determining that the control is effective, the system automatically triggers a data archiving mechanism to structurally extract and encapsulate key data segments from the entire fault evolution cycle. This experience replay data includes at least the initial "high-dimensional tensor fault feature vector: covering temporal features, spatial features, and severity probability distribution," the "autonomous control strategy: including the process parameter adjustment range and actuator action sequence" matched by the autonomous control decision space, the first and second-level sub-control instruction sets generated by the model predictive controller, and the final state feedback data and control performance results. This diverse and heterogeneous data is pieced together into a complete Markov decision process trajectory, along with an action reward value calculated based on the actual performance results, and stored in a high-priority experience replay pool on edge computing nodes or in the cloud. This preserves on-site control samples with high real-world value and strong guiding significance.

[0090] The system samples the encapsulated experience data in batches or in a single instance from the experience replay pool, driving the autonomous control decision space to perform backpropagation and parameter updates. For the dynamic knowledge graph, the system uses the actual control effectiveness results to adaptively correct the edge weights between the control action nodes and process parameter nodes on the fault propagation path in the graph: if a certain control action contributes positive performance gains, its second weight component representing the confidence of historical control success is increased; otherwise, it is decreased.

[0091] For reinforcement learning networks, the system inputs the states, actions, and reward values ​​from the experience replay data into the loss function. It then fine-tunes the network weight parameters using an optimization algorithm based on stochastic gradient descent to minimize the error between the current policy output and the actual optimal action. This online iterative update mechanism enables the decision space model to adaptively evolve as the physical components of the coating machine age and the process environment drifts, continuously approaching the true physical optimal solution. As a result, when encountering similar or derivative faults in the future, it can output the optimal control strategy more quickly and accurately, guiding the coating machine to autonomously recover after experiencing fault impacts and maintain a normal and efficient working state for a long time.

[0092] Specifically, the coating machine previously experienced a local scaling failure in the cooling water circuit of the magnetron target. The system has implemented control strategies such as reducing the average power of the target, activating the pulse arc extinguishing mode, and deflecting the baffle. In step S142, the system reads the performance results output by S141: the film quality retention rate is 94%, which is higher than the preset lower limit threshold of 90%. The equipment thermal stress change index shows that the thermal stress on the target surface has decreased and there are no new microcracks, which is lower than the safety limit threshold. The energy consumption fluctuation variance has increased by 12%, which is lower than the tolerance upper limit of 15%. Based on this, the system determines that the current control is effective, and the coating machine maintains the current degraded operation state to continue to complete the coating process of this batch.

[0093] Meanwhile, the system will encapsulate the control strategies, including "target cooling fouling fault feature vector", "reducing target power by 15% and deflecting the baffle by 15 degrees", and the performance results, including "final film retention rate of 94% and energy consumption increase of 12%", into empirical data and store them in the playback pool.

[0094] During the online iteration phase, although the 15-degree deflection of the baffle compensated for the film thickness, it also caused a 12% increase in energy consumption. When updating the edge weights of the knowledge graph, the system fine-tuned the confidence of the edge weights between the "baffle deflection" action and the "energy consumption fluctuation" parameter. This energy consumption penalty factor was also introduced in the subsequent reinforcement learning network parameter update. Through this online learning, when the coating machine shows slight signs of fouling again in subsequent operations, the updated decision space will give priority to compensating for the process by fine-tuning the substrate speed and gas flow rate without deflecting the baffle. This will further suppress energy consumption fluctuations while maintaining film quality, and enable the coating machine to autonomously evolve and maintain its long-term state during the fault evolution cycle.

[0095] Please see Figure 2 The autonomous control system for the coating machine during a fault phase is applied to the aforementioned autonomous control method for the coating machine during a fault phase; the autonomous control system for the coating machine during a fault phase includes: The fault analysis module 21 is used to collect multi-dimensional state data in real time during the coating process based on a multi-source heterogeneous sensor network deployed at key nodes of the coating machine, perform fault analysis on the multi-dimensional state data, determine the current fault type and severity during the fault analysis process, and generate the corresponding fault feature vector. The control strategy module 22 is used to input the fault feature vector into the preset autonomous control decision space, and to perform optimization matching in the autonomous control decision space in combination with the coating process knowledge graph, so as to determine the corresponding autonomous control strategy. The compensation control module 23 is used to identify multiple sub-control commands based on the identification of autonomous control strategies, and send the multiple sub-control commands to the controller of the coating machine to dynamically reconstruct and coordinately compensate the process parameters and actuators of the coating machine to maintain the process stability of the coating machine during the fault transition phase. The performance result module 24 is used to monitor the operating status of the coating machine after regulation in real time, thereby obtaining the corresponding status feedback data, and determining the regulation performance result of the coating machine by combining the working history of the coating machine and the current working data. Based on the regulation performance result, the autonomous regulation decision space is iteratively corrected online until the coating machine returns to normal.

[0096] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for autonomous control of a coating machine during a fault phase, characterized in that, include: Based on a multi-source heterogeneous sensor network deployed at key nodes of the coating machine, multi-dimensional state data during the coating process is collected in real time. Fault analysis is performed on the multi-dimensional state data. During the fault analysis, the current fault type and fault severity are determined, and the corresponding fault feature vector is generated. The fault feature vector is input into the preset autonomous control decision space, and the coating process knowledge graph is combined in the autonomous control decision space to perform optimization matching, thereby determining the corresponding autonomous control strategy. Based on the identification of autonomous control strategies, multiple sub-control commands are determined and sent to the controller of the coating machine. The process parameters and actuators of the coating machine are dynamically reconstructed and coordinated to compensate for control, so as to maintain the process stability of the coating machine during the fault transition phase. The system monitors the operating status of the coating machine in real time after adjustment, thereby obtaining the corresponding status feedback data. It then combines the working history of the coating machine with the current working data to determine the adjustment efficiency of the coating machine. Based on the adjustment efficiency results, the autonomous adjustment decision space is iteratively corrected online until the coating machine returns to normal.

2. The autonomous control method for a coating machine during a fault phase according to claim 1, characterized in that, The multi-source heterogeneous sensor network deployed at key nodes of the coating machine collects multi-dimensional state data in real time during the coating process, performs fault analysis on the multi-dimensional state data, determines the current fault type and severity during the fault analysis process, and generates corresponding fault feature vectors, including: The coating machine is marked with multiple key nodes, and a corresponding multi-source heterogeneous sensor network is deployed at each key node. This multi-source heterogeneous sensor network collects multi-dimensional state data in real time during the coating process. The multi-dimensional state data includes acoustic emission signals, thermal field distribution data, and electrical parameters.

3. The autonomous control method for a coating machine during a fault phase according to claim 2, characterized in that, The multi-source heterogeneous sensor network deployed at key nodes of the coating machine collects multi-dimensional state data in real time during the coating process, performs fault analysis on the multi-dimensional state data, determines the current fault type and severity during the fault analysis process, and generates corresponding fault feature vectors, and also includes: Edge computing nodes are used to perform time-series alignment and cross-modal feature fusion on multidimensional state data. During the fusion stage, the correlation features of multidimensional state data in spatial topology and the evolution features in time series are extracted. Combined with a multi-source tracing mechanism, the corresponding fault source is determined, and then the current fault type is determined. Dynamic evaluation is performed based on the spread trend of the fault source and historical operating data to output the corresponding fault severity. The fault severity and fault type are then encoded using high-dimensional tensor quantization to generate the corresponding fault feature vector, which includes temporal features, spatial features, and severity probability distribution.

4. The autonomous control method for a coating machine during a fault phase according to claim 1, characterized in that, The step of inputting the fault feature vector into a preset autonomous control decision space, and then performing optimization matching in this autonomous control decision space in conjunction with the coating process knowledge graph to determine the corresponding autonomous control strategy includes: The fault feature vector is used as input and mapped to a pre-defined autonomous control decision space. The autonomous control decision space is constructed from a dynamic knowledge graph driven by reinforcement learning and coating process mechanism. The nodes of the dynamic knowledge graph include process parameter nodes, equipment status nodes and control action nodes. The edge weights between nodes represent the strength of process causal relationship and the confidence of historical control success.

5. The autonomous control method for a coating machine during a fault phase according to claim 4, characterized in that, The step of inputting the fault feature vector into a preset autonomous control decision space, and performing optimization matching in this autonomous control decision space in conjunction with the coating process knowledge graph to determine the corresponding autonomous control strategy, also includes: A graph attention network is used to determine the matching degree between the fault feature vector and each node in the dynamic knowledge graph. A strategy optimization mechanism is triggered in the local subgraph region where the matching degree reaches a preset threshold. The action space is explored in the subgraph region through the strategy gradient mechanism. The expected benefits and system safety margin of each candidate control action are determined in combination with the current process constraints. Then, the corresponding autonomous control strategy is output. The autonomous control strategy includes the process parameter adjustment range and the action sequence of the actuator.

6. The autonomous control method for a coating machine during a fault phase according to claim 1, characterized in that, The process involves identifying multiple sub-control commands based on an autonomous control strategy, and then sending these commands to the coating machine's controller. This allows for dynamic reconfiguration and collaborative compensation control of the coating machine's process parameters and actuators to maintain process stability during fault transitions. This includes: The autonomous control strategy is dynamically identified, and multiple autonomous control items are determined during the identification process. These items are then converted in conjunction with the current working data of the coating substrate to output a corresponding multi-level control sequence. This multi-level control sequence includes multiple first-level sub-control commands and multiple second-level sub-control commands. Each sub-control command is input into the model predictive controller, which further combines the dynamic mechanism information of the coating process to make rolling predictions of the film deposition rate and uniformity in the future multi-step process. The heating power, working gas flow rate, and bias voltage parameters are dynamically optimized and adjusted to achieve process parameter adjustment.

7. The autonomous control method for a coating machine during a fault phase according to claim 6, characterized in that, The method of identifying multiple sub-control commands based on the autonomous control strategy and issuing these commands to the coating machine's controller to dynamically reconstruct and coordinately compensate for the coating machine's process parameters and actuators, thereby maintaining the coating machine's process stability during fault transition phases, also includes: Each second-level sub-control command is sent to the controller of the coating machine to drive the baffle position adjustment mechanism, substrate rotation mechanism or spare target material switching mechanism to perform hardware-level redundancy reconstruction. During the execution process, a cross-coupling compensation controller is introduced to perform real-time collaborative compensation for the timing error and nonlinear coupling interference between process parameter adjustment and mechanical action of the actuator, so as to suppress the sudden change of process parameters during the fault transition stage and maintain the dynamic balance of the vacuum environment.

8. The autonomous control method for a coating machine during a fault phase according to claim 1, characterized in that, The real-time monitoring and control of the coating machine's operating status obtains corresponding status feedback data. This data, combined with the machine's operational history and current data, determines the control efficiency of the coating machine. Based on this efficiency result, the autonomous control decision space is iteratively corrected online until the coating machine returns to normal operation. This includes: The operating status of the coating machine after regulation is monitored in real time. During the monitoring process, the corresponding status feedback signal is output. The status feedback data of the coating machine after regulation is determined by analyzing the status feedback signal. The status feedback data is synchronized to the digital twin space of the coating machine. The status feedback data is dynamically evolved. The regulation efficiency result of the coating machine is determined by combining the working history of the coating machine and the current working data. The regulation efficiency result includes the film quality retention rate, equipment thermal stress change and energy consumption fluctuation characteristics.

9. The autonomous control method for a coating machine during a fault phase according to claim 8, characterized in that, The method of real-time monitoring and control of the coating machine's operating status to obtain corresponding status feedback data, and combining the coating machine's working history and current working data to determine the control efficiency result of the coating machine, and performing online iterative correction of the autonomous control decision space based on the control efficiency result until the coating machine returns to normal status, also includes: If the control effectiveness meets the preset effectiveness threshold, the current control is deemed effective. The fault characteristics and control process are used as experience playback data to iteratively update the edge weights of the knowledge graph and the parameters of the reinforcement learning network in the autonomous control decision space online, so as to guide the coating machine to maintain normal operation.

10. An autonomous control system for a coating machine during a fault phase, characterized in that, The autonomous control system of the coating machine during the fault stage is applied to the autonomous control method of the coating machine during the fault stage as described in any one of claims 1-9; The autonomous control system of the coating machine during the fault phase includes: The fault analysis module is used to collect multi-dimensional state data in real time during the coating process based on a multi-source heterogeneous sensor network deployed at key nodes of the coating machine, perform fault analysis on the multi-dimensional state data, determine the current fault type and severity during the fault analysis process, and generate the corresponding fault feature vector. The control strategy module is used to input the fault feature vector into the preset autonomous control decision space, and to perform optimization matching in the autonomous control decision space by combining the coating process knowledge graph, so as to determine the corresponding autonomous control strategy. The compensation control module is used to identify multiple sub-control commands based on the autonomous control strategy, and send the multiple sub-control commands to the controller of the coating machine to dynamically reconstruct and coordinate the process parameters and actuators of the coating machine to maintain the process stability of the coating machine during the fault transition phase. The performance result module is used to monitor the operating status of the coating machine after regulation in real time, thereby obtaining the corresponding status feedback data. It combines the working history of the coating machine and the current working data to determine the regulation performance result of the coating machine. Based on the regulation performance result, the autonomous regulation decision space is iteratively corrected online until the coating machine returns to normal.

Images