Constraint-Optimized Process Control System and Method Using Predictive Maintenance Risk Index

Abstract

The present invention provides a system and method for realizing process stabilization and equipment protection by calculating the amount of change in APC control parameters through constraint optimization based on a risk index combining RUL, a prediction uncertainty upper limit (U_upper), a health indicator (HI), and a rate of change in RUL, and applying it as a ramping rule, and by performing closed-loop operation including a rollback after switching to a conservative mode according to a hysteresis threshold and applying it, and simultaneously optimizing a maintenance schedule under production constraints.

Description

Technology Field

[0001] The present invention relates to Advanced Process Control (APC) technology for semiconductors, display devices, and general manufacturing facilities, and relates to a system and method for calculating the amount of change in APC control parameters through constraint optimization based on a risk index combining Remaining Useful Life (RUL), upper limit of prediction uncertainty (U_upper), Health Index (HI), and the rate of change of RUL, applying ramping, closed-loop operation including switching to a conservative mode and rollback according to a hysteresis threshold, and simultaneously optimizing maintenance schedules under production constraints.

[0002] The definitions of key terms used in this specification are as shown in the following table, and these variables are used in the calculation of the Predictive Maintenance (PHM)-based risk index.

[0003] terminology definition APC Advanced Process Control refers to advanced process control. RUL Remaining Useful Life refers to the remaining lifespan. U_upper It refers to the upper limit of prediction uncertainty. HI Health Index refers to a health indicator PHM Prognostics and Health Management refers to prediction and health management. Background Technology

[0005] Conventional APCs (Run-to-Run, Model Predictive Control (MPC), VM / RI, FDC, etc.) focused on predicting process results and parameter correction, but had the following problems.

[0006] First, conventional APCs did not consider the deterioration state of the equipment, resulting in excessive parameter correction in equipment with worn parts, which led to quality overshoot and equipment damage.

[0007] Second, while conventional Predictive Maintenance (PdM) provides Remaining Ultimate Life (RUL) estimates, it fails to quantify the uncertainty inherent in RUL predictions, leading to prediction errors propagating to the control system and causing quality instability. In particular, although a safety margin must be secured as the confidence interval of the RUL prediction widens, no specific method has been presented to incorporate this into the calculation of control change amounts.

[0008] Third, conventional technology has failed to provide a structure that calculates a comprehensive risk index integrating the component's available life (RUL), degradation trend (Health Index), prediction uncertainty (U_upper), and rate of change of RUL, and integrates this in real-time due to constraints in calculating the amount of control change.

[0009] Fourth, conventional technology failed to integrate performance verification after control parameter changes, automatic rollback, hysteresis-based conservative mode switching, and gradual ramping application into a single closed loop, resulting in a lack of stability and reliability of the control system.

[0010] Fifth, conventional technology performed changes to process control parameters and preventive maintenance (PM) scheduling independently, which had the problem of not being able to simultaneously optimize both under production planning constraints.

[0011] Therefore, a system was required to calculate a risk index in real time by combining RUL, forecast uncertainty, health indicators, and the rate of change of RUL, reflect this in the control change amount constraint, integrate ramping, verification, rollback, and mode switching into a closed loop, and simultaneously optimize control changes and preventive maintenance schedules. Prior art literature

[0012] Korean Patent Registration No. KR102672799B1 (System and method for matching and analyzing real-time process data of semiconductor equipment). - Abstract: Predicting the probability of failure by matching and learning real-time process data and failure records. - Differentiating feature of the present invention: Optimizing the amount of control change 60% more accurately using a risk index combining RUL, U_upper, and HI, excluding hysteresis and rollback. Korean Patent Registration No. KR102464762B1 (Predictive maintenance method for equipment failure using artificial intelligence based on empirical shared knowledge). - Abstract: Predictive maintenance procedure using IoT, AI, and experiential knowledge. - Differentiating feature of the present invention: Absence of a risk index, uncertainty upper limit, constraint optimization, mode switching, and rollback structure that safely converts RUL into the amount of control change. Korean Published Patent Application (A) (Related to predictive maintenance device / model, example of published application) - Abstract: Configuration of a predictive maintenance model including encoder and feature extraction. - This Differentiating features of the invention: Undisclosed structure for converting RUL into a risk index and applying APC parameters to constraint optimization and hysteresis. U.S. Patent No. US10430719B2 (Process control techniques for semiconductor manufacturing). - Summary: Semiconductor process parameter correction framework (including ML). - Differentiating features of the invention: Undisclosed risk index combining RUL, U_upper, and HI, and the resulting constraint optimization, hysteresis, conservative mode, and simultaneous optimization (including PM). U.S. Patent No. US7349753B2 (Adjusting manufacturing process control parameter using updated process threshold). - Summary: Dynamic threshold-based recipe / parameter adjustment (R2R). - Differentiating features of the invention: Not linked to PHM / RUL, application of change amount upper limit based on uncertainty upper limit (U_upper), and exclusion of mode switching / rollback / simultaneous optimization.U.S. Patent No. US7477960B2 (Fault detection and classification (FDC) using a run-to-run controller). - Summary: FDC utilizing a run-to-run controller. - Differentiation of the invention: Fault detection-centric, PHM-based risk index-linked control, and absence of hysteresis / conservative mode / rollback / simultaneous optimization. Japanese Patent No. JP5292602B2 (Advanced process control system and …). - Summary: Integrating VM / RI into a run-to-run APC to consider quality and latency. - Differentiation of the invention: RUL / U_upper / HI are not integrated due to control change amount constraints, and mode switching / rollback / simultaneous optimization are not included. The present invention improves downtime by 10%. U.S. Patent Publication US20080233662A1 (Advanced Process Control for Semiconductor Processing). - Abstract: General APC procedure and configuration. - Differentiating features of the present invention: RUL / U_upper-based risk index, constraint optimization, mode switching excluded. The present invention improves quality stability by 50%. Chinese Patent Registration No. CN112947287B (Semiconductor Equipment Intelligent Control System Based on Predictive Maintenance). - Abstract: Predictive maintenance and automatic process parameter adjustment system based on IoT sensor data. - Differentiating features of the present invention: (1) Does not quantify prediction uncertainty (U_upper) and uses only simple RUL threshold values. (2) Does not disclose a hysteresis-based mode switching and automatic rollback structure for the risk index. (3) Absence of a function for simultaneous optimization of control changes and PM schedules. The present invention integrates U_upper as a constraint and includes a closed-loop verification-rollback structure, thereby improving quality stability by more than 50%.Chinese Patent Registration No. CN113467398A (Process Control Method Based on Health Index). - Abstract: Process control method based on the Health Index. - Differentiating features of the present invention: (1) Uses only HI and does not have a comprehensive risk index combining RUL, U_upper, and dRUL / dt. (2) Does not include ramping and rollback mechanisms. (3) Absence of a constraint optimization structure. The present invention calculates a risk index by weighted combination of four elements (RUL, dRUL / dt, U_upper, HI) and reflects this in constraint optimization to reduce overshoot by 60%.

[0013] A survey of run-to-run control for batch processes, ISA Transactions, 2018. - Summary: Comprehensive overview of R2R control structures and case studies (EWMA, d-EWMA, MPC, OAQC, ANN) in semiconductor processes, etc. - Comparison: Integrated closed-loop for RUL / U_upper / HI / hysteresis not presented. This invention demonstrates a 60% reduction in overshoot and is proven by experimental data. Predictive Maintenance in semiconductor manufacturing, SemiEngineering article, 2015 / Updated. - Abstract: Trends and challenges (data, models, organizations) in the adoption of PdM in semiconductor equipment. - Comparison: At the level of a PdM overview, the structure for simultaneous optimization of APC parameter constraints, mode switching, rollback, and PM converted from RUL to a risk index is not disclosed. Predictive Maintenance: A Reality for Semiconductor Manufacturing, Critical Manufacturing blog, 2024. - Abstract: Overview and benefits of PdM implementation based on an MES / IoT data platform. - Comparison: The structure for APC closed-loop that limits the control change amount to the U_upper function and links it to hysteresis, conservative mode, rollback, and simultaneous PM optimization is not disclosed. Prognostics uncertainty reduction by right-time prediction of RUL, Eksploatacja i Niezawodność, 2021. - Abstract: HMM / PHM Proposal of a RUL uncertainty reduction technique utilizing [the method]. - Contrast: Limited to uncertainty handling itself, it does not cover control integration that secures real-time process stability by applying parameters through constraint optimization, ramping, hysteresis-based mode switching, and rollback. Uncertainty-aware Remaining Useful Life predictor, arXiv, 2021.- Abstract: Uncertainty-aware RUL estimation framework. - Comparison: Calculation of control change amount based on RUL / U_upper / HI combined risk index and simultaneous PM optimization not disclosed. A survey of run-to-run control algorithms in high-mix manufacturing, Survey, Multiple years. - Abstract: Trends in R2R control algorithms in high-mix environments. - Comparison: PdM / RUL linkage, risk index, conservative mode, rollback, simultaneous PM optimization not disclosed. Offset-free ARX-based adaptive MPC with steady-state constraints, ISA Transactions, 2021. - Abstract: Offset-free adaptive MPC technique. - Comparison: APC design perspective, but not an integration of constraints, mode switching, and simultaneous PM optimization based on predictive maintenance RUL / uncertainty-based risk index. Predictive Maintenance in Fab Tools: Enhancing Efficiency, 2025. - Abstract: Introduction to semiconductor fab PdM concepts and effects (non-academic). - Comparison: Specific control linkage structure and safety constraints / PM No concurrent optimization. The Impact of Uncertainty on a Production Line, Management Science, 1990. - Abstract: Analysis of the impact of uncertainty on a multi-stage production line (theoretical model). - Comparison: No direct disclosure of the RUL·U_upper·HI-based risk index of the present invention and the concurrent optimization of APC constraint optimization, mode switching, rollback, and PM (cited as background theory). Predictive Maintenance in Semiconductor Manufacturing, IEEE / Other Reviews, 2025. - Abstract: Comprehensive review and case studies of ML-based PdM in the semiconductor field. - Comparison: As a general theory of PdM, it does not cover the concurrent optimization of safety constraints, hysteresis, rollback, and PM of risk index-based APC correction. The problem to be solved

[0014] It provides a control change amount calculation structure that ensures compliance with quality specifications and equipment limits based on a risk index (R) combining the uncertainty upper limit (U_upper), health indicator (HI), and rate of change of RUL; secures process stability and equipment protection through a closed loop including hysteresis, conservative mode, ramping, and rollback; and improves availability and quality by simultaneously optimizing control changes and preventive maintenance (PM) schedules under production constraints. means of solving the problem

[0015] The present invention includes the following components.

[0016] RUL prediction module (100) that calculates RUL and upper limit of prediction uncertainty (U_upper) using sensor data (current, voltage, pressure, temperature, vibration, gas flow rate, RF Power, etc.);

[0017] A soundness index calculation unit (115) that calculates HI using vibration, electric, or thermal-based features or residual indicators;

[0018] A risk index calculation unit (120) that defines R = w1·(RUL_thr / RUL) + w2·max(0, -dRUL / dt) + w3·U_upper + w4·HI and applies steping and hysteresis;

[0019] A constraint optimization-based APC correction unit (130) that calculates ΔParam (ΔGain, ΔOffset, ΔRecipe, ΔGas, ΔRF, ΔTset, etc.) under quality specifications, equipment tolerance, rate of change limit, and U_upper function upper limit constraints and applies it as ramping;

[0020] An application, verification, and rollback module (135) that performs automatic rollback and ramping re-application when KPI deteriorates or U_upper spikes after application;

[0021] Conservative mode switching unit (140) that switches to Conservative Mode when a dangerous state persists to perform gain reduction, output upper limit, recipe change rate upper limit, and protection interlock reinforcement, and returns to normal mode under return conditions;

[0022] A maintenance scheduler unit (150) that generates PM using RUL, trend, and risk persistence as triggers and optimizes it into CP-SAT or MILP under production plan, part lead time, and equipment availability constraints; and visualizes R, RUL, U_upper, HI, ΔParam, mode, and PM plans on an operation dashboard (160) and maintains an audit log. Effects of the invention

[0023] The present invention has the following remarkable effects:

[0024] First, by reflecting the upper limit of prediction uncertainty (U_upper) in the constraint on the amount of control change, over-correction occurring in systems that use only conventional RUL is prevented. Conventional technology uses only point estimates of RUL prediction values, so prediction errors propagate to the control system and cause quality overshoot. However, the present invention quantifies the upper limit of prediction uncertainty and reflects it as a constraint in Equation 3 (|ΔParam| ≤ ΔParam_max × (1 - k × U_upper)), thereby limiting the amount of change conservatively as the uncertainty increases, which has the effect of reducing quality overshoot by 60%.

[0025] Second, a comprehensive risk index (R) is calculated by combining RUL, the rate of change of RUL, prediction uncertainty, and health indicators, and by applying a hysteresis threshold-based state transition, switching chattering and false alarms are suppressed compared to conventional technology that uses only a single indicator. In the example, as a result of applying a hysteresis width (θ_up - θ_down = 0.30), the frequency of mode switching was reduced by 70% compared to the conventional method, which has the effect of improving the stability of the control system.

[0026] Third, by integrating ramping, verification, automatic rollback, and conservative mode switching into a single closed loop, the process can be automatically restored to its previous state in the event of quality deterioration or equipment malfunction after a control change, thereby maintaining process stability without human intervention. In the example, a total of 23 automatic rollbacks were performed over a period of six months, which prevented potential scrap generation in advance and reduced yield loss by 0.5%.

[0027] Fourth, by simultaneously optimizing process control parameter changes and preventive maintenance (PM) schedules under production planning constraints, it resolves conflicts between control changes and maintenance plans that were previously performed independently, and has the effect of improving equipment availability.

[0028] In the example, unexpected downtime was reduced by 10 to 18%, and the PM schedule conflict rate was maintained at less than 5%, achieving a production plan compliance rate of 95% or higher.

[0029] Fifth, the steady-state quality deviation decreased by 20%, and the process capability index (Cpk) improved from 1.33 to 1.67, which has the effect of improving quality stability and customer satisfaction.

[0030] The above effects are the results of a 6-month demonstration targeting plasma etching equipment and can be equally applied to semiconductors, displays, and general manufacturing equipment. Brief explanation of the drawing

[0031] FIG. 1 is a block diagram showing the overall architecture of a predictive maintenance linked process control system according to one embodiment of the present invention, showing the connection relationships and data flow between a sensor data collection unit (110), a RUL prediction module (100), a prediction uncertainty upper limit calculation unit (105), a health indicator calculation unit (115), a risk index calculation unit (120), an APC constraint optimization-based correction unit (130), an application / verification / rollback module (135), a conservative mode switching unit (140), a maintenance scheduler (150), and an operation dashboard (160). FIG. 2 is a flowchart showing the remaining lifespan (RUL), upper limit of prediction uncertainty (U_upper), and health index (HI) calculation pipeline according to the present invention, and details the data processing process of sensor data collection, feature engineering, application of RUL prediction model, uncertainty estimation, health index calculation, and risk index calculation. FIG. 3 is a detailed flowchart illustrating a closed-loop control flow including the calculation of a control parameter change amount (ΔParam) based on a risk index (R) according to the present invention, the application of ramping, hysteresis-based mode switching, performance verification, and rollback execution, and includes an objective function and constraint conditions of the constraint optimization process and conservative mode switching conditions. FIG. 4 is a graph showing the control curve of the amount of change in control parameters (ΔParam) at each stage (normal / caution / danger) according to the risk index (R) according to the present invention, and the state transition and return process by hysteresis threshold values ​​(θ_down, θ_up), showing the conservative mode region and the path of change in the risk index over time. Specific details for implementing the invention

[0032] Preferred embodiments of the present invention will be described in detail with reference to the attached drawings. However, the technical concept of the present invention is not limited to the following embodiments. Those skilled in the art who understand the technical concept of the present invention may easily propose other inventions that are inferior or other embodiments included within the scope of the present invention by adding, changing, or deleting other components within the same technical concept, and such are also to be considered to be included within the scope of the present invention.

[0033] FIG. 1 is an overall configuration diagram of a PHM-linked intelligent process control system according to an embodiment of the present invention. Referring to FIG. 1, the system of the present invention comprises a sensor data collection unit (110), a RUL prediction module (100), a prediction uncertainty upper limit calculation unit (105), a health indicator calculation unit (115), a risk index calculation unit (120), a constraint optimization-based APC correction unit (130), an application / verification / rollback module (135), a conservative mode switching unit (140), a maintenance scheduler (150), and an operation dashboard (160).

[0034] Each of the above components is connected in a closed-loop form, and the data flow proceeds in the following order: (a) Data collection and time synchronization → (b) Preprocessing and feature extraction → (c) Calculation of RUL and upper uncertainty limit (U_upper) → (d) Calculation of health index (HI) → (e) Calculation of risk index (R), stepping, and application of hysteresis → (f) Optimization of process parameter change amount (ΔParam) and application of ramping → (g) Verification after application and rollback if necessary → (h) Switching to or returning to conservative mode → (i) Simultaneous optimization of preventive maintenance plan. The functions and operations of each component are described in detail below.

[0035] The sensor data collection unit (110) is a component that collects data in real time from a plurality of sensors attached to the manufacturing equipment, and includes a vibration sensor, a temperature sensor, a current / voltage sensor, a pressure sensor, a flow sensor, a spectrum analysis sensor, etc. In one embodiment, the data collection period can be set within the range of 1 second to 1 millisecond, and preferably is set to a period of 100 milliseconds.

[0036] The sensor data collection unit (110) includes a time synchronization module, and the time synchronization is performed based on the Precision Time Protocol (PTP) or IEEE 1588 standard. In one embodiment, the time synchronization error between multiple sensors is configured to be maintained at less than 1 millisecond, which is to ensure the consistency of multiple sensor data.

[0037] The sensor data collection unit (110) further includes a missing value processing module. The missing value processing module applies a forward fill method, but only when the number of consecutive missing samples is 5 or less, and when it exceeds 5, it masks the corresponding section and applies padding or mask attention techniques when inputting to a subsequent model.

[0038] The sensor data collection unit (110) further includes an outlier removal module. In one embodiment, the outlier removal is performed according to the 3σ (sigma) rule or the IQR (Inter-Quartile Range) criterion (1.5×IQR), and in another embodiment, a machine learning-based outlier detection model such as One-Class SVM (Support Vector Machine) or Isolation Forest may be used in combination. The outlier removal policy may be selectively applied depending on equipment characteristics and process conditions.

[0039] The sensor data collection unit (110) further includes a feature extraction module, and the feature extraction module extracts the following features from the collected raw sensor data:

[0040] (1) Time series features: include a moving average, a moving standard deviation, a moving maximum / minimum value, and a time delay embedding vector. In one embodiment, the window size of the moving statistics is set within the range of 10 to 100 samples, preferably set to 30 samples.

[0041] (2) Frequency domain features: Includes energy, spectral centroid, harmonic ratio, etc. of a specific frequency band extracted through the Fast Fourier Transform (FFT). In one embodiment, for a vibration signal, the signal is divided into a low frequency band (0.100 Hz), a mid frequency band (100 Hz / 1 kHz), and a high frequency band (1 kHz or higher), and energy for each band is calculated.

[0042] (3) Time-frequency feature: Includes the Root Mean Square (RMS) value of the spectrogram generated through the Continuous Wavelet Transform (CWT) or Short-Time Fourier Transform (STFT).

[0043] (4) Power-related features: include voltage-current phase difference, peak power to average power ratio, power factor, etc.

[0044] (5) Thermal features: Includes temperature difference between sensors (ΔT), rate of change of temperature (dT / dt), thermal residual relative to steady state, etc.

[0045] (6) Oscillation features: Includes high-order statistics such as envelope signals and kurtosis through envelope analysis.

[0046] (7) Process-specific features: include the time constant of the gas flow rate response, the pressure stabilization time, and the transient response characteristics of the RF power.

[0047] The above features undergo a normalization process, and in one embodiment, Min-Max scaling is applied, while in another embodiment, Robust scaling (based on median and IQR) may be applied. The normalized features are composed of feature vectors and input into the RUL prediction module (100).

[0048] The RUL prediction module (100) includes a machine learning or deep learning-based prediction model. In one embodiment, the prediction model is a sequence model and includes one or more of LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), TCN (Temporal Convolutional Network), or 1D-CNN (1-Dimensional Convolutional Neural Network).

[0049] In another embodiment, the prediction model includes XGBoost (eXtreme Gradient Boosting) or LightGBM (Light Gradient Boosting Machine) as a tree-based model. In yet another embodiment, a hybrid structure combining a sequence model and a tree model may be adopted, in which case a stacking or blending ensemble technique is applied.

[0050] When training the above prediction model, labels are generated based on the equipment's usage history, component replacement times, and failure logs. The RUL value is defined as the remaining time (in hours or operating cycles) from the current point in time to the point of failure or replacement. In one embodiment, weighted loss is applied to sections where equipment degradation progresses non-linearly (e.g., sections where failure is imminent) to improve prediction accuracy in the later stages.

[0051] Huber loss or Quantile loss can be used as the loss function. In particular, for Quantile loss, setting τ=0.9 enables learning that is favorable for upper bound estimation. Here, τ represents the quantile parameter.

[0052] The prediction uncertainty upper limit calculation unit (105) is a component that quantifies the upper limit value among the confidence intervals of the RUL prediction value. One of the key features of the present invention is that it does not simply use the point estimate of the RUL, but explicitly quantifies the uncertainty inherent in the prediction and reflects it in decision-making.

[0053] In one embodiment, the prediction uncertainty upper limit calculation unit (105) applies the Monte Carlo Dropout (MC-Dropout) technique. Specifically, a distribution of RUL prediction values ​​is generated by performing N iterations (e.g., N=30 times) of inference with the dropout layer enabled in the trained neural network model. The N is determined experimentally and, in one embodiment, is optimized based on 500 validation data points. From the generated prediction distribution, the upper p percentile (e.g., 90th percentile or 95th percentile) is adopted as U_upper.

[0054] In another embodiment, the prediction uncertainty upper limit calculation unit (105) applies an ensemble technique. K independently trained models (e.g., K=3 to 5) are used to predict RUL for each, and an upper limit is calculated from the distribution of the predicted values. The K is optimized based on experimental data and is determined by considering model diversity and computational cost.

[0055] The prediction uncertainty upper limit calculation unit (105) may further include a calibration module. The calibration module applies Temperature Scaling or Isotonic Regression techniques to correct the prediction confidence so that it matches the actual accuracy. This improves the reliability of decision-making by resolving the problem of overconfidence or underconfidence.

[0056] Validation of the RUL prediction model is performed using time-series preserving cross-validation (rolling-origin evaluation). Additionally, generalization performance on untrained equipment is checked through equipment-specific hold-out validation.

[0057] As an optional embodiment, the prediction uncertainty upper limit calculation unit (105) may include a concept drift detection module. The drift detection module monitors changes in the input data distribution or prediction error distribution in real time by applying the Page-Hinkley test, ADWIN (ADaptive WINdowing), or DDM (Drift Detection Method). When drift is detected, model retraining or policy parameter readjustment is triggered.

[0058] The health indicator calculation unit (115) is a component that calculates a health indicator (HI) representing the health status of the equipment from multiple sensor signals. The health indicator is calculated independently of the RUL prediction and serves to complement the model prediction by directly reflecting changes in physical sensor signals.

[0059] The soundness indicator calculation unit (115) calculates the following candidate indicators:

[0060] (1) Energy sum of specific frequency bands of vibration signals

[0061] (2) Envelope-Kurtosis Combination Index: Kurtosis value calculated from the envelope of the vibration signal

[0062] (3) Current harmonic ratio: Rate of increase of harmonic components compared to normal operation

[0063] (4) Voltage-current phase deviation: Absolute value of deviation relative to normal phase angle

[0064] (5) Temperature deviation (ΔT) and thermal residual: Deviation of the measured temperature from the design reference temperature

[0065] (6) Cumulative RMS of model prediction residuals: The value of the accumulated short-term prediction error of the RUL prediction model

[0067] Each of the above candidate indicators is normalized individually. In one embodiment, Min-Max normalization is applied, and in another embodiment, Robust scaling may be applied. The normalized indicators are weighted linearly combined according to the following formula to produce the final health indicator HI:

[0068] [Mathematical Formula 1]

[0069] HI = Σ(i=1 to M) v_i · h_i

[0070] Here,

[0071] M: Number of candidate indicators

[0072] v_i: Weight of the i-th indicator (Σv_i = 1)

[0073] h_i: i-th normalized candidate indicator

[0074] The above weight v_i can be set by an expert based on equipment type, process characteristics, and the results of historical failure data analysis, or can be automatically optimized through Bayesian optimization or grid search.

[0075] The integrity index calculation unit (115) may further include an environment correction module. The environment correction module is intended to eliminate the influence of changes in process conditions, such as chamber internal pressure, external temperature, and recipe changes, on the integrity index. In one embodiment, a regression model is constructed with process state variables as independent variables, and the residuals are used as the corrected integrity index. In another embodiment, a conditional standardization technique may be applied.

[0076] The risk index calculation unit (120) calculates a risk index (R), which is a single scalar value representing the risk level of the equipment, by combining the RUL prediction value, the upper limit of prediction uncertainty (U_upper), the health indicator (HI), and the rate of change of RUL. The risk index is defined by the following mathematical formula:

[0077] [Mathematical Formula 2]

[0078] R = w1·(RUL_thr / RUL) + w2·max(0, -dRUL / dt) + w3·U_upper + w4·HI

[0079] Here,

[0080] RUL_thr: Reference remaining life threshold (e.g., 48 hours)

[0081] RUL: Currently predicted remaining lifespan

[0082] dRUL / dt: Time rate of change of RUL (negative value indicates accelerated degradation)

[0083] U_upper: Upper limit of prediction uncertainty

[0084] HI: Health Indicator

[0085] w1, w2, w3, w4: Weights of each term (typically normalized to w1+w2+w3+w4=1)

[0086] Each term of the above mathematical formula has the following physical meaning:

[0087] Paragraph 1 w1·(RUL_thr / RUL): Reflects that risk increases as RUL length decreases. Increases sharply if RUL decreases below RUL_thr.

[0088] Paragraph 2 w2·max(0, -dRUL / dt): The faster the rate of RUL decrease (i.e., the more accelerated the deterioration), the higher the risk. A positive rate of change (improvement) is treated as 0.

[0089] Paragraph 3 w3·U_upper: The greater the uncertainty of the prediction, the more conservatively the risk is assessed. This is a key feature of the present invention.

[0090] Paragraph 4 w4·HI: The higher the physical sensor-based health indicator, the higher the risk.

[0091] In one embodiment, the weights are set to w1=0.4, w2=0.2, w3=0.3, and w4=0.1. However, the weights can be adjusted according to the equipment type, process importance, and safety requirements, and can be automatically tuned through Bayesian optimization or grid search to minimize the deterioration of Key Performance Indicators (KPIs).

[0092] The risk index calculation unit (120) classifies the calculated risk index R into a plurality of discrete steps. In one embodiment, the risk steps are defined as follows:

[0093] Normal stage: R < θ1

[0094] Caution stage: θ1 ≤ R < θ2

[0095] Risk Level: R ≥ θ2

[0096] Here, θ1 and θ2 are step classification threshold values, which are set based on equipment-specific quality specifications, safety margins, and historical degradation trend data. In one embodiment, θ1 is set to 0.5 and θ2 to 1.0, but this is merely an example and the actual values ​​may vary by line.

[0097] The risk index calculation unit (120) includes hysteresis logic. The hysteresis logic is intended to ensure the stability of the system by preventing frequent switching of risk levels. Specifically, different threshold values ​​are applied when the level increases and decreases:

[0098] Rise threshold: θ_up (typically θ_up = θ2)

[0099] Descent threshold: θ_down (typically θ_down < θ2)

[0100] For example, if θ_up=1.0 and θ_down=0.4 are set, R ≥ 1.0 must be required to transition from the caution stage to the danger stage, but R < 0.4 must be required to return from the danger stage to the caution stage. This prevents chattering.

[0101] Additionally, the risk index calculation unit (120) applies time-based filtering. Specifically, for a phase transition to be confirmed, the new phase must be maintained for at least a minimum duration (T_min). In one embodiment, T_min is set to 30 minutes. Similarly, if the risk phase is maintained for at least a duration judgment time (T_risk), a conservative mode or preventive maintenance trigger is activated. In one embodiment, T_risk is set to 10 minutes.

[0102] As an optional embodiment, the risk index calculation unit (120) may additionally consider a safety margin metric. The safety margin metric is a value that quantifies the distance of the current state relative to quality specifications (LSL, USL) or physical limits of the facility. By merging this with the risk index R, the sensitivity of the decision can be dynamically adjusted.

[0103] The constraint optimization-based APC correction unit (130) calculates the optimal change amount (ΔParam) of the process parameter by considering the risk index R and the uncertainty upper limit U_upper. This serves as a core component of the present invention and plays a role in organically linking PHM and APC.

[0104] The above optimization problem is formulated as follows:

[0105] [Mathematical Formula 3]

[0106] minimize J(ΔParam)

[0107] subject to:

[0108] (i) Quality ∈ [LSL, USL]

[0109] (ii) Param_min ≤ Param ≤ Param_max

[0110] (iii) |d(Param) / dt| ≤ ρ_max

[0111] (iv) |ΔParam| ≤ g(U_upper)

[0112] (v) Resource constraints

[0113] Here, the decision variable ΔParam includes the following:

[0114] ΔGain: Amount of change in control loop gain

[0115] ΔOffset: Amount of change in the setpoint offset

[0116] ΔRecipe: Amount of change in recipe parameters

[0117] ΔGas: Amount of change in gas flow rate

[0118] ΔRF: Change in RF power

[0119] ΔTset: Amount of change in temperature setpoint

[0120] The objective function J(ΔParam) is defined by the following mathematical formula:

[0121] [Mathematical Formula 4]

[0122] J = α·E[(QualityDev)^2] + β·Stress(ΔParam) + γ·ThroughputLoss(ΔParam) + ζ·Energy(ΔParam)

[0123] Here,

[0124] E[(QualityDev)^2]: Squared expected value of quality deviation (quality stability term)

[0125] Stress(ΔParam): Stress exerted on the equipment by a parameter change (equipment load term)

[0126] ThroughputLoss(ΔParam): Productivity loss due to parameter change (productivity term)

[0127] Energy(ΔParam): Increase in energy consumption (Energy efficiency term)

[0128] α, β, γ, ζ: Weighting coefficients of each term

[0129] In one embodiment, α=0.5, β=0.3, and γ=0.2 are set, and the energy term may be omitted (ζ=0). However, the weighting factor can be adjusted according to the process objective and the results of the economic evaluation.

[0130] Constraint (i) ensures compliance with quality specifications. Here, LSL (Lower Specification Limit) and USL (Upper Specification Limit) represent the acceptable quality range of the product and restrict the optimization result so as not to violate it.

[0131] Constraint (ii) specifies upper and lower limits for parameters for safety and facility protection. For example, RF power cannot exceed physical output limits, and the temperature must be set within the thermal limits of the chamber material.

[0132] Constraint (iii) is a ramping constraint that limits the rate of change of the parameter per hour. This is to prevent process instability, quality degradation, and equipment damage caused by rapid parameter changes. The ramping limit ρ_max is set differently for each type of parameter. In one embodiment, ρ_max(RF) is set to 1% / s, ρ_max(Gas) to 0.5% / s, and ρ_max(Tset) to 0.2°C / min.

[0133] Constraint (iv) is one of the key features of the present invention, defining the upper limit of the parameter change amount as a function of the upper limit of the prediction uncertainty (U_upper):

[0134] [Mathematical Formula 5]

[0135] |ΔParam| ≤ g(U_upper)

[0136] Here, the function g satisfies the following properties:

[0137] g'(U_upper) > 0: As U_upper increases (i.e., as prediction uncertainty increases), the upper bound on parameter change decreases.

[0138] Monotonically decreasing relationship: Control more conservatively when uncertainty is high

[0139] The specific form of the function g can be implemented as one of the following:

[0140] (a) Linear form:

[0141] g(U_upper) = a·(U_max - U_upper) + b

[0142] Here, a and b are design parameters, and U_max is the maximum value of uncertainty

[0143] (b) Logarithmic form:

[0144] g(U_upper) = a·ln(1 + U_max - U_upper) + b

[0145] Adjust sensitivity to increased uncertainty using a logarithmic function

[0146] (c) Hierarchical form:

[0147] g(U_upper) = min{g_quality(U_upper), g_safety(U_upper)}

[0148] Adopt the stricter value between quality constraints and safety constraints

[0149] (d) Neural network approximation form:

[0150] Approximate the g function through a neural network based on training data

[0151] Constraint (v) is a resource constraint and includes upper limits for gas supply, power load, cooling capacity, etc.

[0152] CP-SAT (Constraint Programming - Satisfiability) or MILP (Mixed Integer Linear Programming) techniques are used as solutions to the above optimization problem. In one embodiment, the CP-SAT solver of Google OR-Tools is applied.

[0153] When large-scale variables and constraints exist, the following two-step approach can be applied:

[0154] Step 1: Generate an initial solution using an experience-based policy table or heuristics.

[0155] Step 2: Starting from the initial solution above, perform optimization within the time limit to improve the solution.

[0156] To ensure real-time performance, a sliding horizon technique is applied. In one embodiment, optimization is re-performed at intervals of 5 to 10 minutes, and the previous solution is used as an initial value to reduce computation time. When multiple devices are operated simultaneously, optimization operations for each device are executed in parallel in a distributed computing environment.

[0157] The application, verification, and rollback module (135) applies the parameter change amount (ΔParam) calculated by the constraint optimization-based APC correction unit (130) to the actual equipment, while performing a ramping algorithm to prevent sudden changes.

[0158] The ramping algorithm described above is a method of gradually changing parameters in intervals up to a target parameter value. In one embodiment, linear ramping is applied so that the parameter is updated according to the following mathematical formula:

[0159] [Mathematical Formula 6]

[0160] Param(t) = Param(t0) + (ΔParam / T_ramp) · (t - t0)

[0161] Here,

[0162] Param(t0): Initial parameter value

[0163] ΔParam: Target change amount

[0164] T_ramp: Ramping duration

[0165] t: Current time

[0166] In other embodiments, exponential ramping may be applied, in which case the rate of change is large initially and decreases as it approaches the target value. The ramping rate limit ρ_max may be set differently for each parameter, for example, RF power can be changed quickly, but temperature must be changed slowly.

[0167] Verification after applying 8.2

[0168] The application, verification, and rollback module (135) monitors changes in KPIs during a verification window (W_v) after applying parameter changes. In one embodiment, W_v is set to 10 minutes, and the following indicators are evaluated during this period:

[0169] Quality indicators: CD (Critical Dimension) deviation, thickness uniformity, defect density, etc.

[0170] Productivity Indicators: Processing Time (throughput), Yield

[0171] Safety Indicators: Temperature, Pressure, Vibration Levels

[0172] The statistical significance of the above KPI change is evaluated using the Z-Score or CUSUM (Cumulative Sum) technique. In one embodiment, if the Z-Score exceeds a threshold value (Z_thr), it is determined to be an abnormal signal. For example, Z_thr can be set to 2.5.

[0173] Apply, verify, and rollback (135) cancels the parameter change and returns to the previous stable state if one or more of the following rollback trigger conditions are satisfied:

[0174] (i) KPI deterioration: Z-Score exceeds Z_thr (e.g., Z > 2.5)

[0175] (ii) Uncertainty surge: The growth rate of U_upper exceeds the threshold (u_thr) (e.g., U_upper > 0.9)

[0176] (iii) Insufficient Safety Margin: If a rollback is triggered when the safety margin relative to quality or facility limits is below the threshold value (m_thr), the following procedures are executed sequentially:

[0177] Step 1: Immediately return the parameters to the previous stable point.

[0178] Step 2: Reduce the ramping rate limit ρ_max (e.g., ρ_max ← 0.5 × ρ_max)

[0179] Step 3: Recalculate ΔParam by applying more conservative constraints (e.g., adjusting the coefficients of the g(U_upper) function).

[0180] Step 4: Reapply the recalculated ΔParam using step-wise ramping

[0181] All parameter changes, validation results, and rollback events are recorded in the audit log along with timestamps. The aforementioned audit log is utilized for post-hoc analysis, regulatory compliance, and system improvement.

[0182] The conservative mode switching unit (140) is a component that conservatively changes the control strategy to ensure equipment protection and safety when the risk level of the equipment is high or the RUL is very short. This is an operating mode that prioritizes stability and failure prevention over active performance optimization.

[0183] The conservative mode switching unit (140) switches to conservative mode when one or more of the following conditions are satisfied:

[0184] (i) Persistence of risk level: A state where the risk index R is θ2 or higher persists for T_risk or longer (e.g., R ≥ 1.0 persists for 10 minutes or longer)

[0185] (ii) RUL threshold: RUL decreases below the lower limit (RUL_low) (e.g., RUL ≤ 24 hours)

[0186] 9.3 Control Policy in Conservative Mode

[0187] In conservative mode, the following control policies are applied:

[0188] (1) Control gain reduction:

[0189] k ← η · k, where 0 < η < 1

[0190] Reduce the gain of the control loop to smooth out the control action. In one embodiment, η is set to 0.7.

[0191] (2) Output upper limit:

[0192] P_max ← μ · P_max, where 0 < μ < 1

[0193] Limits the maximum output of RF power, gas flow rate, etc. In one embodiment, μ is set to 0.8.

[0194] (3) Restriction on recipe change rate enhancement:

[0195] Further reduce the ramping rate limit ρ_max (e.g., ρ_max ← 0.5 × ρ_max)

[0196] (4) Enable additional protection interlocks:

[0197] Set interlock thresholds for temperature, pressure, vibration, etc., more strictly.

[0198] In conservative mode, the upper limit of the ΔParam of the constraint (iv) is further reduced, and the priority of the preventive maintenance scheduler (150) is increased so that the PM plan can be triggered early.

[0199] 9.4 Conditions for Returning to Conservative Mode

[0200] The conservative mode switching unit (140) returns to normal mode when the following conditions are satisfied:

[0201] (i) Decline in risk index: R < θ_down condition persists for T_min or longer (e.g., R < 0.4 persists for 30 minutes or longer)

[0202] (ii) Restoring Safety Margin: Ensuring Sufficient Safety Margin Against Quality and Equipment Limits

[0203] Hysteresis logic is applied even upon return to prevent frequent mode switching.

[0204] The maintenance scheduler (150) creates a preventive maintenance (PM) task when one or more of the following conditions are satisfied:

[0205] (i) RUL threshold: RUL ≤ RUL_low (e.g., 24 hours)

[0206] (ii) Sharp decline in RUL: Rate of change in RUL |dRUL / dt| ≥ λ (e.g., 0.02 h -1 )

[0207] (iii) Risk accumulation: Cumulative duration of risk level (R ≥ θ2) ΣT ≥ T_sum

[0208] (iv) PM Window: Hazardous state persists for longer than the PM trigger window τ (e.g., τ=10 minutes)

[0209] The maintenance scheduler (150) does not simply generate PM, but calculates an optimal PM schedule that simultaneously achieves the following objectives:

[0210] Objective: Maximize equipment availability and minimize maintenance costs

[0211] Pharmaceuticals:

[0212] Production Schedule: Prevents conflicts with campaign and lot schedules

[0213] Parts lead time: The time when replacement parts are available for procurement

[0214] Worker shift schedule: Availability of maintenance personnel

[0215] Equipment Availability: Upper limit on the number of equipment that can be maintained simultaneously

[0216] Concurrent work restrictions: Certain equipment cannot be maintained simultaneously (e.g., equipment using shared utilities)

[0217] One of the key features of the present invention is that the PM schedule and the process parameter change plan (ΔParam applied schedule) are **simultaneous coherently optimized**. This prevents the following conflicts:

[0218] Preventing the inefficiency of changing parameters immediately before PM

[0219] Prevents PM from intervening while parameter changes are being applied

[0220] Comprehensive coordination of production plans and maintenance / control change schedules

[0221] The above concurrent optimization problem is solved using CP-SAT or MILP techniques. The objective function can be defined as follows:

[0222] [Mathematical Formula 7]

[0223] minimize (C_downtime · Downtime + C_maint · MaintCost - Benefit_availability · Availability)

[0224] Here,

[0225] C_downtime: Downtime unit cost

[0226] C_maint: Maintenance unit cost

[0227] Benefit_availability: Benefit in units of availability

[0228] In large-scale multi-equipment environments, scalability of the solution can be ensured through column generation or Lagrangian relaxation techniques.

[0229] 10.3 Dynamic Adjustment of Maintenance Schedules

[0230] The maintenance scheduler (150) dynamically adjusts the PM schedule according to changes in real-time RUL prediction values ​​and risk indices. For example, if the RUL decreases faster than expected, the PM schedule is advanced, and conversely, if the deterioration progresses gradually, the PM is postponed to maximize production opportunities.

[0231] However, safety limits apply to PM postponement, and restrictions are imposed to ensure that PM is executed before the RUL falls below an absolute lower limit (e.g., 12 hours).

[0232] 11. Automatic parameter tuning and drift support

[0233] 11.1 Parameters Targeted for Automatic Tuning

[0234] The system of the present invention includes a plurality of policy parameters, the optimal values ​​of which vary depending on the equipment type, process characteristics, and operating environment. The parameters subject to automatic tuning include the following:

[0235] θ1, θ2: Risk index level classification threshold values

[0236] θ_up, θ_down: Hysteresis threshold

[0237] T_min, T_risk, τ: Time-based decision parameters

[0238] λ: RUL rate of change threshold value

[0239] ρ_max: Ramping rate limit

[0240] RUL_low: Preventive maintenance trigger RUL lower limit

[0241] Coefficients of the g function: Parameters of the uncertainty-change upper bound mapping function

[0242] 11.2 Automatic Tuning Method

[0243] Automatic tuning of the above parameters is performed through Bayesian optimization or reinforcement learning assistance policies. The objective function of optimization includes the following:

[0244] Minimizing KPI deterioration: Reducing quality variance and defect rates

[0245] Minimizing Downtime: Reducing downtime due to unexpected breakdowns

[0246] Minimizing maintenance costs: Preventing excessive preventive maintenance

[0247] When applying Bayesian optimization, a surrogate model based on a Gaussian process is constructed, and parameter combinations that maximize the acquisition function (e.g., Expected Improvement) are sequentially explored.

[0248] When using reinforcement learning as an auxiliary policy, the policy network is trained by defining parameter selection as the action, the system state as the state, and KPI improvement as the reward.

[0249] The manufacturing environment changes over time (e.g., recipe changes, part replacements, changes in environmental conditions). These changes can lead to a degradation in the performance of RUL prediction models and control policies, a phenomenon known as concept drift.

[0250] The system of the present invention includes the following drift response mechanism:

[0251] (1) Drift detection: Real-time monitoring of changes in input data distribution or prediction error distribution by applying statistical techniques such as Page-Hinkley test, ADWIN, and DDM.

[0252] (2) Model Retraining: If drift is detected, retrain the RUL prediction model using recent data. You can choose between incremental learning or full retraining.

[0253] (3) Policy Reset: Re-run automatic parameter tuning to derive policy parameters adapted to the changed environment

[0254] (4) Enhanced safety limit: Temporarily switch to conservative mode immediately after drift detection to prevent a decrease in prediction reliability.

[0255] The operation dashboard (160) is a user interface that visualizes the operation status of the system in real time. The dashboard displays the following information:

[0256] Risk Index (R) and Risk Level: Intuitive display via color code (Green / Yellow / Red)

[0257] Remaining Life (RUL): Displayed in hours or operating cycles, includes trend graph

[0258] Upper Limit of Prediction Uncertainty (U_upper): Confidence Interval Visualization

[0259] Health Indicator (HI): Individual values ​​and weighted sum of each candidate indicator

[0260] Process Parameter Change Amount (ΔParam): Changes currently being applied, ramping progress

[0261] Driving Mode: Displays Normal Mode or Conservative Mode

[0262] Preventive Maintenance Plan: Scheduled PM Schedule and Priorities

[0263] KPI Status: Quality, Productivity, and Energy Efficiency Indicators

[0264] Alarm and Event Logs: Recording events such as rollback, mode switching, and drift detection

[0265] The operations dashboard (160) applies Role-Based Access Control (RBAC). User permissions are classified as follows:

[0266] Read Permission: Dashboard viewing only, general monitoring personnel

[0267] Approve Authority: Approve parameter changes or PM plans proposed by the system, Process Engineer

[0268] Modify Authority: Manual adjustment of policy parameters, change of alarm thresholds, Senior Engineer

[0269] Administrative Authority (Admin): System settings, user management, run model retraining, system administrator

[0270] Significant changes (e.g., manual changes to policy parameters, forced adjustments to PM schedules) are double-checked through electronic signatures or approval workflows.

[0271] 12.3 Security and Compliance

[0272] The system of the present invention includes the following security and compliance mechanisms:

[0273] (1) Data De-identification / Pseudonymization: If personal or sensitive information is included, storage and transmission after de-identification processing

[0274] (2) Access control: Restricting access through network-level firewalls, VPNs, and account authentication (2-factor authentication).

[0275] (3) Audit Log: Record all user actions, system decisions, parameter changes, and alarm occurrences as immutable logs with timestamps. These logs are used for post-audits, regulatory compliance, and root cause analysis.

[0276] (4) Differential Privacy (Optional): If the privacy of data contributors needs to be protected during model training, the differential privacy technique is applied optionally.

[0277] (5) Source Data Protection: Raw sensor data is stored only within the manufacturing site's internal system and is not transmitted to an external network. Only feature or aggregated data is transmitted when necessary.

[0278] The above security policy is designed to comply with information security standards of the semiconductor / display industry (e.g., ISO / IEC 27001) and data protection regulations of each country (e.g., GDPR, CCPA).

[0279] Hereinafter, as a preferred embodiment of the present invention, an example of applying the system of the present invention to plasma etching equipment of a semiconductor manufacturing line will be described in detail.

[0280] The target equipment is an ICP (Inductively Coupled Plasma) etcher that processes 300mm wafers. The installed sensors are as follows:

[0281] Three vibration sensors: attached to the pump, RF matching box, and chamber wall.

[0282] 5 temperature sensors: upper electrode, lower electrode, chamber wall, cooling water inlet / outlet

[0283] 2 Current / Voltage Sensors: RF Power Supply and Bias Power Supply

[0284] Two pressure sensors: inside the chamber and in the vacuum line

[0285] 4 gas flow meters: Etching gas (CF4, O2), carrier gas (Ar, N2)

[0286] The data collection cycle was set to 100ms.

[0287] 13.2 Feature Extraction and RUL Model

[0288] The feature extraction module generates the following features:

[0289] Vibration: Low-frequency band (0~50Hz) energy, Kurtosis

[0290] Temperature: Temperature difference between electrodes (ΔT), rate of change in cooling water temperature

[0291] Power: RF voltage-current phase difference, reflected power ratio

[0292] Pressure: Pressure stabilization time constant

[0293] Gas: Flow rate response delay time

[0294] A total of 23 features were extracted.

[0295] An LSTM (2 layers, 128 units each) model was adopted as the RUL prediction model. It predicts RUL by taking feature sequences from 60 past time points (100ms × 60 = 6 seconds) as input. The training data used was 18 months of operational data (including replacement history) from 10 machines of the same model.

[0296] The prediction uncertainty upper limit calculation unit (105) generates a prediction distribution by applying MC-Dropout (N=30 times) and adopts the 90th percentile as U_upper.

[0297] The soundness indicator calculation unit (115) calculated the following six candidate indicators:

[0298] h1: Vibrational low-frequency energy (weight v1=0.25)

[0299] h2: Accumulated temperature deviation (weight v2=0.15)

[0300] h3: RF phase deviation (weight v3=0.20)

[0301] h4: Pressure response delay (weight v4=0.10)

[0302] h5: Gas flow imbalance (weight v5=0.10)

[0303] h6: Model predicted residual RMS (weight v6=0.20)

[0304] The risk index was calculated using the following weights:

[0305] R = 0.4·(48h / RUL) + 0.2·max(0, -dRUL / dt) + 0.3·U_upper + 0.1·HI

[0306] The step classification threshold was set to θ1=0.5 and θ2=1.0, and the hysteresis threshold was set to θ_up=1.0 and θ_down=0.4.

[0307] The constraint optimization-based APC correction unit (130) set the following five parameters as optimization targets:

[0308] ΔRF: RF power change (±10% range, ρ_max=1% / s)

[0309] ΔBias: Amount of change in bias voltage (±15% range, ρ_max=0.8% / s)

[0310] ΔGas_CF4: Change in CF4 gas flow rate (±5% range, ρ_max=0.5% / s)

[0311] ΔGas_O2: Change in O2 gas flow rate (±8% range, ρ_max=0.5% / s)

[0312] ΔTset: Change amount of lower electrode temperature setting (±3°C range, ρ_max=0.2°C / min)

[0313] The objective function was defined as a weighted sum of minimizing CD (Critical Dimension) deviation, minimizing parameter change stress, and minimizing productivity loss (α=0.5, β=0.3, γ=0.2).

[0314] The g function of constraint (iv) is implemented in linear form:

[0315] g(U_upper) = 10·(1.0 - U_upper) + 2 [Unit: %]

[0316] That is, when U_upper=0.5, the upper limit of ΔParam is limited to 7%, and when U_upper=0.9, the upper limit of ΔParam is limited to 3%.

[0317] The conditions for entering conservative mode were set as follows:

[0318] If the R ≥ 1.0 condition persists for 10 minutes or more, or

[0319] RUL ≤ 24 hours

[0320] In conservative mode, the following policies were applied:

[0321] RF power gain reduced by 0.7x (η=0.7)

[0322] RF power is limited to 80% (μ=0.8)

[0323] Reduce all ramping rates by 50%

[0324] The maintenance scheduler (150) manages the following maintenance tasks:

[0325] Upper electrode replacement (Estimated cycle: 1,000 hours, Duration: 4 hours)

[0326] RF Matching Box Inspection (Estimated Cycle: 500 hours, Duration: 1 hour)

[0327] O-ring replacement (Estimated cycle: 2,000 hours, Duration: 2 hours)

[0328] The PM trigger conditions were set as follows:

[0329] RUL ≤ 24 hours, or

[0330] |dRUL / dt| ≥ 0.02 h -1 , or

[0331] R ≥ 1.0, state accumulation time ΣT ≥ 60 minutes

[0332] As a result of applying the system of the present invention to an actual production line for six months, the following performance improvements were confirmed:

[0333] 60% Reduction in CD Overshoot: Quality overshoot reduced by 60% compared to conventional methods when changing parameters

[0334] 20% reduction in CD deviation: 20% improvement in steady-state quality deviation

[0335] 10–18% reduction in unexpected downtime: Reduced downtime due to sudden breakdowns

[0336] Achieved 95% PM Plan Compliance: PM Execution Without Conflicts with Production Plan

[0337] Although the above performance may vary depending on process conditions, production line characteristics, and the initial state of the equipment, significant improvement compared to conventional technology has been confirmed.

[0338] The above embodiment is merely a preferred example of the present invention, and various modifications are possible within the scope of the technical concept of the present invention.

[0339] Variation Example 1: Expansion of Risk Index Items

[0340] Environmental variables (e.g., ambient temperature, humidity) or recipe complexity indicators can be included as additional terms in the mathematical formula for calculating the risk index R. In this case, Equation 2 is expanded as follows:

[0341] R = w1·(RUL_thr / RUL) + w2·max(0, -dRUL / dt) + w3·U_upper + w4·HI + w5·EnvFactor + w6·RecipeComplexity

[0342] Variation Example 2: Changing the form of the g function

[0343] The g(U_upper) function of constraint (iv) can be approximated through a neural network in addition to linear, logarithmic, and hierarchical methods. In this case, a non-linear relationship between U_upper and a safe ΔParam upper bound is learned based on past operational data.

[0344] Variation Example 3: Replacement of Optimization Solution

[0345] As a solution for the constraint optimization-based APC correction unit (130) and maintenance scheduler (150), methods such as Quadratic Programming (QP), Second-Order Cone Programming (SOCP), or metaheuristics (genetic algorithm, particle swarm optimization) can be used instead.

[0346] Variant Example 4: Multipurpose Extension of PM Optimization

[0347] The objective function of the maintenance scheduler (150) can be expanded from a single objective to a multi-objective one, and a multi-objective evolutionary algorithm such as NSGA-II or MOEA / D can be applied. In this case, availability, cost, quality, and safety are set as independent objectives, and a set of Pareto optimal solutions is derived.

[0348] Variation Example 5: Expansion of applicable equipment

[0349] Although the above embodiments were intended for plasma etching equipment, the present invention is applicable to all core process equipment for semiconductor and display manufacturing, such as CVD (Chemical Vapor Deposition), PVD (Physical Vapor Deposition), CMP (Chemical Mechanical Polishing), lithography, and cleaning equipment.

[0350] Variant Example 6: Distributed System Implementation

[0351] When applying the system to multiple production lines or the entire fab, it can be implemented as a cloud-based distributed system. In this case, the data collection unit (110) of each piece of equipment operates as an edge computing node, and high-computation tasks such as RUL prediction and optimization are performed on a central server or the cloud. Industrial applicability

[0352] The present invention can be widely applied to the management of core equipment in precision manufacturing industries such as semiconductors, displays, solar cells, MEMS, and power semiconductors. In particular, in a fab environment with a large number of expensive equipment assets, it can simultaneously achieve the effects of preventing unexpected failures, improving quality stability, maximizing productivity, and reducing maintenance costs.

[0353] Furthermore, the present invention has broad applicability across various industries, as it can be provided by equipment manufacturers (OEMs) as an embedded PHM-APC solution in their products or introduced by manufacturing site operators as a retrofit to existing equipment. Explanation of the symbols

[0354] 100: RUL Prediction Module 105: Prediction Uncertainty Upper Limit Calculator 110: Sensor data acquisition unit 115: Soundness Indicator Calculation Department 120: Risk Index Calculation Unit 130: Constraint Optimization-Based APC Correction Unit 135: Application, Verification, and Rollback Module 140: Conservative mode switching section 150: Maintenance Scheduler 160: Operations Dashboard

Claims

Claim 1 In a predictive maintenance linked process control system that links Advanced Process Control (APC) and predictive maintenance, a sensor data collection unit (110) that collects at least one data among vibration, temperature, current, voltage, pressure, and flow rate in real time from a plurality of sensors attached to manufacturing equipment; and a RUL prediction module unit (100) that predicts the remaining useful life (RUL) based on the data collected from the sensor data collection unit (110). A prediction uncertainty upper limit calculation unit (105) that calculates a prediction uncertainty upper limit (U_upper: an upper boundary value of the RUL prediction distribution, normalized to a range of 0 to 1, meaning that the closer to 1, the higher the prediction uncertainty) based on the prediction result of the RUL prediction module (100); a health index calculation unit (115) that calculates a health index (HI: Health Index, normalized to a range of 0 to 1, meaning that the larger the value, the greater the deterioration of the equipment condition) using vibration, electrical, and thermal-based features or residual indicators extracted from the sensor data; a risk index (R) that calculates the calculated RUL, U_upper, HI, and RUL change rate (dRUL / dt) by weighted combination according to the following mathematical formula 2, wherein the calculated risk index (R) is updated at every control cycle and fed back in real-time as a control parameter constraint condition of the following constraint optimization-based APC correction unit (130), and for the risk index (R), an increase threshold (θ_up) and A risk index calculation unit (120) that determines entry into and return to a caution or risk stage from a normal stage by separately setting a falling threshold (θ_down) (provided that it is set to satisfy the hysteresis relationship θ_up > θ_down + 0.1), and applying hysteresis such that it switches to a higher stage only when the risk index (R) is θ_up or greater and returns to a lower stage only when it is less than θ_down;[Equation 2] R = w₁·(RUL_thr / RUL) + w₂·max(0, -dRUL / dt) + w₃·U_upper + w₄·HI(where w₁, w₂, w₃, and w₄ are weighting coefficients that adjust the contribution of each predictive maintenance indicator, where w₁+w₂+w₃+w₄=1 and each is greater than 0; RUL_thr is a pre-set remaining life threshold for each piece of equipment; dRUL / dt is the rate of change of remaining life with respect to time, where a negative value indicates rapid deterioration; and U_upper and HI are as defined above). Along with the constraints of Equation 3 below, the quality specification compliance constraint (LSL ≤ Quality ≤ USL), the equipment tolerance range constraint (Param_min ≤ Param ≤ Param_max), and the parameter change rate limit constraint (|d(Param) / dt| ≤ A constraint optimization-based APC correction unit (130) that calculates a control parameter change amount ΔParam (including at least one of ΔGain, ΔOffset, ΔRecipe, ΔGas, ΔRF, and ΔTset) using a Continuous Constrained Optimization method to simultaneously satisfy ρ_max), and gradually applies the calculated change amount through a ramping function;[Mathematical Equation 3] |ΔParam| ≤ ΔParam_max × (1 - k × U_upper)(wherein ΔParam_max is the maximum allowable change amount per facility, and k is a sensitivity coefficient (design parameter) of 0.1 or greater and 1.0 or less that is pre-set by an engineer, which is a safety factor designed so that the allowable change amount converges to 0 as U_upper approaches 1); an application, verification, and rollback module (135) including a closed-loop verification logic that, if the process KPI (a predefined process quality key performance indicator including at least one of yield, process capability index (Cpk), and defect rate) deteriorates after applying the control parameter change amount ΔParam calculated by the constraint optimization-based APC correction unit (130), or if the U_upper increases by 30% or more compared to immediately before the parameter change, automatically rollback to the previous parameter state, reduce and reset the ramping speed limit, and reapply the recalculated control parameter change amount ΔParam as gradual ramping; A conservative mode switching unit (140) that, when the risk index (R) is determined to be above the rising threshold (θ_up) for N consecutive times or more (N is an integer between 2 and 10), switches to a conservative mode to perform at least one of reducing the control gain, limiting the maximum output, limiting the recipe change rate, and strengthening the protection interlock, and returns to a normal mode when the risk index (R) returns below the falling threshold (θ_down) for N consecutive times or more and is maintained for a predetermined minimum holding time (T_min: a time preset to ensure the stability of the mode switch);A predictive maintenance-linked process control system characterized by comprising: a maintenance scheduler (150) that generates a preventive maintenance (PM) request using the above-mentioned risk index (R), RUL trend, and risk duration as triggers, and simultaneously optimizes the application schedule of the control parameter change amount ΔParam and the PM schedule using CP-SAT (Constraint Programming - Satisfiability) or MILP (Mixed Integer Linear Programming) under production plan, part lead time, and equipment availability constraints; and an operation dashboard (160) that visualizes the above-mentioned risk index (R), RUL, U_upper, HI, control parameter change amount ΔParam, operation mode, and preventive maintenance (PM) plan, and maintains an audit log for each operation. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete Claim 6 A predictive maintenance linked process control system according to claim 1, wherein the RUL and U_upper calculations are performed by an ensemble of at least one sequence model (LSTM, GRU, 1D-CNN) or a tree-based model (XGBoost, LightGBM), and the U_upper is adopted as the 90th percentile or 95th percentile of the RUL prediction value distribution generated by iterative inference with dropout enabled, or as the upper limit of the RUL prediction value distribution generated by an ensemble of multiple independently trained models, and reliability is corrected through calibration. Claim 7 A predictive maintenance interlocking process control system according to claim 1, wherein the health index (HI) is calculated based on at least one of a vibration-based feature, a voltage-current phase, a temperature rise (ΔT), or a residual-based index. Claim 8 A predictive maintenance linked process control system according to claim 1, wherein the risk stage is classified into normal, caution, and risk, and corresponding to each stage, at least one of Gain, Offset, Recipe, Gas Flow, RF Power, and temperature setpoint is automatically corrected according to a policy table or the constraint optimization solution. Claim 9 A predictive maintenance linked process control system according to claim 1, wherein the upper limit of ΔParam is defined as a monotonically decreasing function of the uncertainty upper limit (U_upper), and the amount of change is automatically reduced as the prediction uncertainty increases. Claim 10 A predictive maintenance linked process control system according to claim 1, characterized in that the maintenance scheduler triggers preventive maintenance (PM) when the remaining life decreases below a threshold value, when the rate of change of RUL exceeds a reference value, or when a dangerous condition persists for longer than a set time, and optimizes using the CP-SAT or MILP algorithm under production planning, part lead time, and equipment availability constraints. Claim 11 In a predictive maintenance-linked process control method that links Advanced Process Control (APC) with predictive maintenance, the method comprises: (a) a sensor data collection unit (110) collecting sensor data from a plurality of sensors attached to manufacturing equipment; a RUL prediction module unit (100) predicting the remaining lifespan (RUL) from the collected data; a prediction uncertainty upper limit calculation unit (105) calculating the prediction uncertainty upper limit (U_upper) based on the prediction result; and a health index calculation unit (115) calculating the health index (HI); (b) a risk index calculation unit (120) calculating a risk index (R) by weighted combining the calculated RUL, U_upper, HI, and RUL change rate (dRUL / dt) according to the following mathematical formula 2, wherein the calculated risk index (R) is updated at every control cycle and fed back in real-time to the control parameter constraints of the following constraint optimization-based APC correction unit (130), and the rising threshold (θ_up) and falling Critical(θ_down, where θ_up > θ_down + 0.Step of applying hysteresis according to 1); [Mathematical Formula 2] R = w₁·(RUL_thr / RUL) + w₂·max(0, -dRUL / dt) + w₃·U_upper + w₄·HI (where w₁, w₂, w₃, and w₄ are weighting coefficients that adjust the contribution of each predictive maintenance indicator, where w₁+w₂+w₃+w₄=1 and each is greater than 0; RUL_thr is a remaining life threshold value pre-set for each facility; dRUL / dt is the rate of change of the remaining life with respect to time, where a negative value indicates rapid deterioration; U_upper is the upper boundary value of the RUL prediction distribution, normalized to a range of 0 to 1; and HI is a health indicator, where a larger value indicates greater deterioration of the facility condition) (c) A PC correction unit (130) based on constraint optimization A step of calculating the amount of change in control parameters ΔParam (including at least one of ΔGain, ΔOffset, ΔRecipe, ΔGas, ΔRF, and ΔTset) under the constraints of the following Equation 3 and the quality specifications, equipment tolerance range, and parameter change rate limit constraints, and gradually applying the calculated amount of change through a ramping function; [Equation 3] |ΔParam| ≤ ΔParam_max × (1 - k × U_upper)(where ΔParam_max is the maximum allowable amount of change per facility, and k is 0.1 or greater, which is pre-set by an engineer.(d) A safety factor designed such that the allowable change amount converges to 0 as U_upper approaches 1, as a sensitivity coefficient (design parameter) of 0 or less; (d) If the application, verification, and rollback module (135) deteriorates in the process KPI (a predefined process quality key performance indicator including at least one of yield, process capability index (Cpk), and defect rate) after the application of the control parameter change amount ΔParam, or if the U_upper increases by 30% or more compared to immediately before the parameter change, the application, verification, and rollback module performs an automatic rollback, reduces and resets the ramping speed limit, and reapplies the recalculated control parameter change amount ΔParam as a gradual ramping step; (e) If the conservative mode switching unit (140) determines that the risk index (R) is above the rise threshold (θ_up) for N consecutive times or more (N is an integer between 2 and 10), it switches to a conservative mode to reduce the control gain, limit the maximum output, limit the recipe change rate, and implement a protection interlock. A predictive maintenance linked process control method characterized by comprising: a step of performing at least one of reinforcement, and returning to a normal mode when the risk index (R) returns to below the falling threshold (θ_down) for N consecutive times or more and is maintained for a predetermined minimum holding time (T_min: a time preset to ensure the stability of the mode switch); and (f) a maintenance scheduler (150) generates a preventive maintenance (PM) request using the risk index (R), RUL trend, and risk duration as triggers, and simultaneously optimizes the application schedule of the control parameter change amount ΔParam and the PM schedule using CP-SAT or MILP under production plans, part lead times, and equipment availability constraints. Claim 12 delete Claim 13 delete Claim 14 A predictive maintenance linked process control method according to claim 11, wherein the conservative mode comprises at least one of control gain reduction, maximum output limit, recipe change rate limit, and protection interlock reinforcement, wherein the control gain reduction is performed by multiplying the control gain by a reduction factor (η) greater than 0 and less than 1, the maximum output limit is performed by multiplying the maximum output upper limit by a limit factor (μ) greater than 0 and less than 1, and the method returns to a normal mode when the risk index (R) returns to a value below the down threshold (θ_down) for N consecutive times or more, and the state persists for a minimum holding time (T_min) or longer. Claim 15 A predictive maintenance linked process control method according to claim 11, characterized in that the maintenance schedule is calculated using a CP-SAT or MILP algorithm with constraints of production plan, part lead time, and equipment availability, and is simultaneously optimized to be consistent with the control parameter change plan.

Images