Data analysis based evaporation process monitoring method, system and medium

Abstract

The application relates to the technical field of evaporation process monitoring, and discloses an evaporation process monitoring method and system based on data analysis and a medium. The method comprises the following steps: collecting data of sensors in an evaporation cavity, and obtaining process parameter data through filtering and standardization processing; performing time-frequency analysis and correlation calculation on the process parameter data, and extracting key parameter features; setting a dynamic threshold based on the key parameter features, calculating a deviation degree to obtain an abnormal index; and generating a control scheme by adjusting a radio frequency power and a substrate temperature in real time according to the abnormal index. The application can analyze complex correlation relationships between multi-source sensing data in real time, identify abnormal patterns in a timely manner, predict potential problems before abnormal conditions are obviously manifested, and automatically adjust process parameters through an intelligent control strategy, so that the technical defects of a monitoring system in the prior art, such as response lag and the inability to adjust process parameters in real time, are overcome, and the stability of an evaporation process and the consistency of product quality are improved.

Description

Technical Field

[0001] This application relates to the field of vapor deposition process monitoring technology, and in particular to a data analysis-based vapor deposition process monitoring method, system and medium. Background Technology

[0002] Evaporation deposition is a crucial modern thin-film deposition method, widely used in the manufacturing processes of semiconductor devices, optical coatings, and solar cells. During evaporation, raw materials are deposited onto the substrate surface in a vacuum environment through heating and evaporation or sputtering, forming a thin-film structure with specific functions. Traditional evaporation process monitoring relies primarily on empirical parameter settings and manual intervention, controlled by pre-set process parameters such as chamber temperature, pressure, and RF power. With technological advancements, automated monitoring systems have begun to be applied to evaporation production, employing fixed threshold methods to monitor key parameters and triggering alarms when parameters exceed preset ranges. Some advanced systems incorporate simple data analysis methods, such as moving averages and linear regression, to pre-process process parameters, reducing false alarms and improving monitoring accuracy.

[0003] However, existing technologies have the following significant shortcomings: First, parameter collection is scattered, making it difficult to form a unified data standard. Data from each monitoring point is stored in isolation, lacking integrated analysis. Second, although the evaporation process generates a large amount of data, the analysis methods are too simplistic, mostly remaining at the stage of comparing single parameter thresholds, failing to uncover deep-seated correlations and complex patterns between parameters. Third, the monitoring system has a lag in response, unable to adjust process parameters in real time, resulting in significant delays in identifying and handling abnormal situations, leading to unnecessary waste of materials and energy. Fourth, existing systems lack self-learning capabilities, have a low recognition rate for new abnormal patterns, and are difficult to adapt to changes in process conditions. Fifth, the inter-system coordination is poor, with monitoring, analysis, and control links being isolated, making it difficult to form a complete optimization loop. These problems are particularly prominent in high-precision scenarios, such as the difficulty in meeting the requirements for uniform thickness control and optical performance stability of micro- and nano-scale thin films.

[0004] This application solves the core problem that traditional monitoring systems cannot achieve accurate anomaly detection and automatic control. In this application, a technical system is constructed from raw data processing to deep feature extraction, and then to anomaly detection and classification based on deep learning, which solves the technical problems of high data dimensionality, complex patterns and diverse anomaly types in the vapor deposition process. Summary of the Invention

[0005] This application provides a data analysis-based method, system, and medium for monitoring the vapor deposition process. It can analyze the complex correlations between multi-source sensor data in real time, identify abnormal patterns in a timely manner, predict potential problems before abnormal conditions become obvious, and automatically adjust process parameters through intelligent control strategies. This overcomes the technical defects of existing monitoring systems, such as slow response and inability to adjust process parameters in real time, thereby improving the stability of the vapor deposition process and the consistency of product quality.

[0006] In a first aspect, this application provides a data analysis-based method for monitoring the vapor deposition process. The method includes: collecting raw data from temperature sensors, pressure sensors, power monitors, and thickness monitoring systems installed within the vapor deposition chamber; obtaining vapor deposition process parameter data through filtering and standardization; performing time-frequency analysis and parameter correlation calculations on the vapor deposition process parameter data to extract the variation characteristics of the evaporation source temperature, chamber pressure, and deposition rate, thereby obtaining key parameter characteristics of the vapor deposition process; based on the key parameter characteristics, performing deviation analysis by setting dynamic thresholds to calculate the degree of deviation between real-time parameter values ​​and standard values, thereby obtaining abnormal indicators of the vapor deposition process; and adjusting the RF power input and substrate temperature control in real time according to the abnormal indicators of the vapor deposition process, generating correction control commands, and obtaining a target control scheme for thin film growth within the vapor deposition chamber.

[0007] In a first embodiment of the first aspect, the raw data collected by the temperature sensor, pressure sensor, power monitor, and thickness monitoring system installed in the vapor deposition chamber is processed through filtering and standardization to obtain vapor deposition process parameter data, including:

[0008] Temperature data is collected by temperature sensors installed at the evaporation source, substrate and cavity wall. The raw temperature distribution data is obtained by combining local regression model with sliding window analysis.

[0009] Vacuum data is collected from pressure sensors installed at different locations in the cavity. The raw pressure gradient data is obtained by multi-point averaging and outlier removal.

[0010] The power input data recorded in real time by the power monitor is sampled at a fixed frequency and marked with abrupt change points to obtain the raw power fluctuation data.

[0011] Wavelet transform filtering technology is applied to the original temperature distribution data, original pressure gradient data, and original power fluctuation data to retain signal characteristics while removing high-frequency noise, resulting in filtered parameter data.

[0012] The filtered parameter data is subjected to zero-mean normalization to convert parameters of different dimensions into a standard distribution format, thereby obtaining standardized parameter data.

[0013] By aligning sensor data with different sampling frequencies in the standardized parameter data using timestamps, a unified time reference is established, and the first derivative and moving average features are extracted to obtain the vapor deposition process parameter data.

[0014] In the second embodiment of the first aspect, the step of performing time-frequency analysis and parameter correlation calculation on the vapor deposition process parameter data to extract the variation characteristics of evaporation source temperature, cavity pressure, and deposition rate, and obtaining key parameter characteristics of the vapor deposition process, includes:

[0015] The vapor deposition process parameter data is divided into fixed windows, and the statistical moment characteristics and trend characteristics of evaporation source temperature, cavity pressure and deposition rate in each window are calculated to obtain the time domain characteristics of the parameters.

[0016] A fast Fourier transform is performed on the vapor deposition process parameter data to extract the power spectral density curve, identify the periodic fluctuation components of the evaporation source temperature, cavity pressure, and deposition rate, and obtain the parameter frequency domain characteristics.

[0017] Wavelet transform or Hilbert transform is applied to the vapor deposition process parameter data to capture the time-varying frequency characteristics of the evaporation source temperature, cavity pressure and deposition rate, and obtain the time-frequency characteristics of the parameters;

[0018] The sample entropy and multi-scale entropy values ​​are calculated for the time-domain features, frequency-domain features, and time-frequency features of the parameters to quantify the complexity of parameter fluctuations and obtain parameter complexity features.

[0019] Based on the time-domain characteristics, frequency-domain characteristics, time-frequency characteristics, and complexity characteristics of the parameters, the Pearson correlation coefficient and mutual information value between the evaporation source temperature, cavity pressure, and deposition rate are calculated to obtain the parameter correlation matrix.

[0020] Partial correlation analysis and conditional independence tests are applied to the parameter correlation matrix to eliminate the influence of indirect correlation, determine the causal relationship and time delay between parameters, and obtain the key parameter characteristics of the vapor deposition process.

[0021] In the third embodiment of the first aspect, the step of performing deviation analysis based on the key parameter characteristics of the vapor deposition process by setting a dynamic threshold, calculating the degree of deviation between the real-time parameter value and the standard value, and obtaining the abnormal index of the vapor deposition process includes:

[0022] Cluster analysis was performed on the key parameter characteristics of the vapor deposition process, and the process operation mode was divided using a Gaussian mixture model to obtain a standard mode library for the vapor deposition process.

[0023] For each process mode in the standard mode library of the vapor deposition process, the characteristic multidimensional distribution parameters, including the mean vector and covariance matrix, are calculated to obtain the mode statistical characterization;

[0024] For the key parameters of the vapor deposition process collected in real time, the Mahalanobis distance and Hotling statistic compared with the statistical characterization of the model are calculated to obtain the parameter deviation.

[0025] Based on the parameter deviation, a deep autoencoder network is constructed to encode and decode the input features, calculate the reconstruction error, and obtain the abnormal reconstruction loss.

[0026] For the aforementioned abnormal reconstruction loss, an adaptive threshold is set based on the distribution characteristics of historical data to distinguish between normal fluctuations and abnormal deviations, thereby obtaining an abnormality level classification;

[0027] Based on the aforementioned anomaly classification, the anomaly start time, duration, severity score, and contribution of related parameters are calculated to obtain the anomaly index of the vapor deposition process.

[0028] In the fourth embodiment of the first aspect, the step of constructing a deep autoencoder network based on the parameter deviation, encoding and decoding the input features, calculating the reconstruction error, and obtaining the anomaly reconstruction loss includes:

[0029] The parameter deviation and key parameter features of the vapor deposition process are concatenated to form an input vector, which is then fed into the input layer of a deep autoencoder network to obtain the features to be processed.

[0030] The features to be processed are compressed using a four-layer convolutional encoder. Each convolutional kernel in the encoder is 3×3 in size, and the number of channels in the convolutional layers is 32, 64, 128, and 256 respectively. An activation function and pooling operation are applied after each convolutional layer to obtain compressed features.

[0031] The compressed features are further dimensionality reduced by a fully connected layer, with the number of nodes being 512, 256, 128, and 64 respectively, to obtain the latent space feature representation;

[0032] The latent space feature representation is restored by a four-layer transposed convolutional decoder. The transposed convolutional decoder adopts a structure symmetrical to the encoder, with the number of channels being 256, 128, 64, and 32 respectively, to obtain the reconstructed features.

[0033] Calculate the Euclidean distance and cosine similarity between the reconstructed features and the original input features, construct a weighted loss function, and obtain the feature reconstruction loss.

[0034] The feature reconstruction loss is compared and analyzed with the reconstruction loss of time-series continuous samples to identify abnormal mutation points and gradual trends, and the abnormal reconstruction loss is obtained.

[0035] In the fifth implementation of the first aspect, the step of determining the anomaly reconstruction loss by setting an adaptive threshold based on historical data distribution characteristics to distinguish between normal fluctuations and abnormal deviations, thereby obtaining an anomaly level classification, includes:

[0036] The probability distribution of historically accumulated anomaly reconstruction loss data is fitted, and a nonparametric probability density function is constructed using the kernel density estimation method to obtain the loss distribution model;

[0037] Based on the loss distribution model, the reconstruction loss values ​​of multiple quantiles are calculated and set as different levels of judgment criteria to obtain a set of static thresholds;

[0038] For the static threshold set, combined with the parameter distribution of the evaporation process in the start-up, stabilization and end stages, a piecewise linear transformation is applied to obtain the stage adaptive threshold.

[0039] For the aforementioned stage adaptive threshold, a sliding window averaging and exponential smoothing method is applied to eliminate the impact of short-term fluctuations and obtain a smoothed adaptive threshold.

[0040] The real-time anomaly reconstruction loss is compared with the smooth adaptive threshold, and the anomaly level is obtained by classifying the degree to which it exceeds the threshold of different level quantiles.

[0041] The initial anomaly level is processed in a time series, and the anomaly level is updated according to the anomaly duration and the accumulated anomaly value to obtain the anomaly level classification.

[0042] In the sixth embodiment of the first aspect, the step of adjusting the RF power input and substrate temperature control in real time based on the abnormal indicators of the evaporation process, generating a correction control command, and obtaining a target control scheme for thin film growth within the evaporation cavity includes:

[0043] By integrating the abnormal indicators of the vapor deposition process with the historical operation data of the vapor deposition process, a digital twin model of the vapor deposition process is constructed to simulate the influence of different control parameters on thin film growth and obtain the influence mapping of control parameters.

[0044] Based on the control parameter influence mapping, a control strategy network is trained using a strategy gradient algorithm. Abnormal indicators of the evaporation process are taken as inputs, and the adjustment amounts of RF power and substrate temperature are output to obtain the parameter control strategy.

[0045] Physical constraints are checked on the parameter control strategy to limit the range and rate of change of RF power and substrate temperature, preventing system instability caused by drastic adjustments, thus obtaining a safe control strategy.

[0046] Based on the aforementioned safety control strategy, large-scale parameter adjustments are decomposed into multiple small-step sequences to form a smooth-transition control trajectory, resulting in a progressive control sequence.

[0047] The progressive control sequence is converted into instruction signals in the standard format of the industrial control interface, and written into the programmable logic controller of the vapor deposition equipment through the communication interface to obtain the execution control instructions;

[0048] The execution results of the control commands are monitored in real time, and the changes in abnormal indicators before and after parameter adjustment are recorded to form control effect evaluation data, thereby obtaining the target control scheme for thin film growth in the vapor deposition chamber.

[0049] Secondly, this application provides a data analysis-based vapor deposition process monitoring system, the data analysis-based vapor deposition process monitoring system comprising:

[0050] The data acquisition module is used to collect raw data from the temperature sensor, pressure sensor, power monitor and thickness monitoring system installed in the vapor deposition chamber, and obtain vapor deposition process parameter data through filtering and standardization.

[0051] The calculation module is used to perform time-frequency analysis and parameter correlation calculation on the vapor deposition process parameter data, extract the variation characteristics of evaporation source temperature, cavity pressure and deposition rate, and obtain the key parameter characteristics of the vapor deposition process.

[0052] The setting module is used to perform deviation analysis by setting dynamic thresholds based on the key parameter characteristics of the vapor deposition process, calculate the degree of deviation between the real-time value of the parameter and the standard value, and obtain the abnormal index of the vapor deposition process.

[0053] The adjustment module is used to adjust the RF power input and substrate temperature control in real time according to the abnormal indicators of the evaporation process, generate correction control commands, and obtain the target control scheme for thin film growth in the evaporation cavity.

[0054] Thirdly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described data analysis-based vapor deposition process monitoring method.

[0055] The technical solution provided in this application effectively solves the problems of inconsistent data formats and severe noise interference in traditional monitoring by filtering and standardizing the raw data collected by various sensors in the vapor deposition chamber, thereby improving data quality and usability. Time-frequency analysis and parameter correlation calculation of vapor deposition process parameter data not only extract parameter features from multiple dimensions such as time domain, frequency domain, and time-frequency domain, but also analyze the correlation and causal relationships between parameters. This overcomes the limitations of traditional methods that only focus on a single parameter and ignore complex relationships between parameters, enabling the system to identify complex process state changes. Dynamic thresholds are set based on the key parameter characteristics of the vapor deposition process for deviation analysis, overcoming the shortcomings of fixed thresholds being unsuitable for different process stages, achieving adaptive anomaly detection, and significantly reducing false alarm and false negative rates. Real-time adjustments are made to the RF power input and substrate temperature control based on abnormal indicators of the vapor deposition process, forming a closed-loop control system. This solves the problem of traditional monitoring systems lacking automatic adjustment mechanisms after detecting anomalies, significantly shortening the response time to abnormal states. In specific applications of vapor deposition processes, the deep learning and reinforcement learning algorithms employed in this invention contribute intelligent perception and autonomous decision-making capabilities to the overall solution: The deep autoencoder network, by learning low-dimensional representations of normal process parameters, can accurately capture minute anomaly patterns, exhibiting extremely high sensitivity to subtle fluctuations in deposition rate and film thickness uniformity; the policy gradient algorithm autonomously learns the optimal control strategy based on historical optimization experience, automatically adjusting RF power and substrate temperature parameters according to anomaly indicators, achieving human-like intelligent control capabilities; the kernel density estimation and piecewise linear transformation algorithms in dynamic threshold calculation adapt to the changing characteristics of parameter distribution at each stage of the vapor deposition process, enabling the system to accurately distinguish between normal fluctuations and abnormal deviations; the wavelet transform in time-frequency analysis captures the time-varying frequency characteristics of vapor deposition parameters, revealing dynamic patterns that are difficult to detect using traditional spectrum analysis. These algorithmic features work synergistically to significantly improve the stability of the vapor deposition process and the consistency of product quality, while reducing material and energy waste caused by abnormal conditions and lowering production costs. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of one embodiment of the data analysis-based vapor deposition process monitoring method in this application.

[0058] Figure 2 This is a schematic diagram of one embodiment of the data analysis-based vapor deposition process monitoring system in this application. Detailed Implementation

[0059] This application provides a data analysis-based method, system, and medium for monitoring a vapor deposition process. The terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0060] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the data analysis-based vapor deposition process monitoring method in this application includes:

[0061] Step S101: Collect raw data from the temperature sensor, pressure sensor, power monitor, and thickness monitoring system installed in the vapor deposition chamber, and obtain vapor deposition process parameter data through filtering and standardization.

[0062] Step S102: Perform time-frequency analysis and parameter correlation calculation on the vapor deposition process parameter data, extract the variation characteristics of evaporation source temperature, cavity pressure and deposition rate, and obtain the key parameter characteristics of the vapor deposition process;

[0063] Step S103: Based on the key parameter characteristics of the vapor deposition process, deviation analysis is performed by setting a dynamic threshold to calculate the degree of deviation between the real-time parameter value and the standard value, and to obtain the abnormal index of the vapor deposition process.

[0064] Step S104: Based on the abnormal indicators of the evaporation process, adjust the RF power input and substrate temperature control in real time, generate correction control commands, and obtain the target control scheme for thin film growth in the evaporation cavity.

[0065] It is understood that the executing entity of this application can be a data analysis-based vapor deposition process monitoring system, or it can be a terminal or a server; no specific limitation is made here. This application's embodiments use a server as an example for illustration.

[0066] Specifically, the process of acquiring raw data from temperature sensors, pressure sensors, power monitors, and thickness monitoring systems installed within the vapor deposition chamber, and obtaining vapor deposition process parameter data through filtering and standardization, is implemented as follows: Thermocouple temperature sensors are installed at the evaporation source, substrate, and chamber wall of the vapor deposition equipment; pressure sensors are installed at different locations within the chamber; and a power monitor and a real-time thin film thickness monitoring system based on the ellipsometer principle are configured. Temperature data acquired by the temperature sensors is processed using a local regression model combined with sliding window analysis. The sliding window length is 30 seconds, and the window overlap rate is 50%. The local regression model uses quadratic polynomial fitting to identify and mark abrupt change points. Vacuum degree data from the pressure sensors is processed using multi-point averaging sampling and outlier removal methods. Multi-point averaging takes the average of ten sampling points, and outlier removal uses the median absolute deviation method, marking data points deviating more than three times the median as outliers and removing them. Power input data from the power monitor is processed using fixed-frequency sampling and abrupt change point marking. The fixed-frequency sampling frequency is 100Hz, and abrupt change point marking uses a cumulative sum control chart method to detect power abrupt change events. The raw data were processed using wavelet transform filtering technology. The wavelet transform employed the Daubechies wavelet basis, with a decomposition level of 3, preserving signal characteristics while removing high-frequency noise. The filtered parameter data underwent zero-mean normalization, using the formula (x-μ) / σ, where x is the original data, μ is the mean, and σ is the standard deviation, uniformly converting parameters of different dimensions into a standard distribution format. Finally, sensor data from different sampling frequencies were aligned using timestamps to establish a unified time reference, and the first derivative and moving average features were extracted to obtain the vapor deposition process parameter data.

[0067] The detailed process of performing time-frequency analysis and parameter correlation calculations on vapor deposition process parameter data to extract the variation characteristics of evaporation source temperature, cavity pressure, and deposition rate, and obtaining the key parameter characteristics of the vapor deposition process, is as follows: The vapor deposition process parameter data is divided into fixed windows with a window length of 30 seconds. The statistical moment characteristics (mean, standard deviation, skewness, kurtosis) and trend characteristics (monotonicity, periodicity, inflection point density) of evaporation source temperature, cavity pressure, and deposition rate within each window are calculated to obtain the time-domain characteristics of the parameters. Then, a fast Fourier transform is performed on the vapor deposition process parameter data to analyze the power spectral density curves and identify the periodic fluctuation components of evaporation source temperature, cavity pressure, and deposition rate, obtaining the frequency-domain characteristics of the parameters. Next, wavelet transform or Hilbert transform is applied to capture the time-varying frequency characteristics of the parameters, obtaining the time-frequency characteristics of the parameters. Sample entropy and multi-scale entropy values ​​are calculated for these characteristics to quantify the complexity of parameter fluctuations. The sample entropy calculation uses a parameter setting with an embedding dimension of 2 and a tolerance of 0.2 times the standard deviation. The multi-scale entropy is calculated using a scale factor ranging from 1 to 20, obtaining the parameter complexity characteristics. Based on these characteristics, the Pearson correlation coefficient and mutual information values ​​among the evaporation source temperature, cavity pressure, and deposition rate are calculated to generate a parameter correlation matrix. Partial correlation analysis and conditional independence tests are applied to the parameter correlation matrix to eliminate the influence of indirect correlations, determine the causal relationships and time delays among the parameters, and finally obtain the key parameter characteristics of the evaporation process.

[0068] Based on the key parameter characteristics of the vapor deposition process, deviation analysis is performed by setting dynamic thresholds to calculate the degree of deviation between real-time parameter values ​​and standard values, thus obtaining the abnormal indicators of the vapor deposition process. The specific process is as follows: First, cluster analysis is performed on the key parameter characteristics of the vapor deposition process. A Gaussian mixture model is used to divide the process operation modes, and the number of clusters is automatically determined using the Bayesian information criterion, resulting in a standard mode library for the vapor deposition process. For each process mode in the standard mode library, the multidimensional distribution parameters of the features are calculated, including the mean vector and covariance matrix, to obtain the mode statistical representation. For the key parameter characteristics of the vapor deposition process collected in real time, the Mahalanobis distance and Hotelling statistic with the mode statistical representation are calculated to obtain the parameter deviation. Based on the parameter deviation, a deep autoencoder network is constructed to encode and decode the input features, and the reconstruction error is calculated to obtain the abnormal reconstruction loss. This deep autoencoder network structure includes four convolutional encoder layers for compression. Each convolutional kernel has a size of 3×3, and the number of channels is 32, 64, 128, and 256 respectively. LeakyReLU activation function and max pooling operation are applied after each convolutional layer to obtain compressed features. The compressed features are further dimensionality-reduced through fully connected layers to obtain latent space feature representations. These latent space features are then reconstructed using a four-layer transposed convolutional decoder to obtain the reconstructed features. Euclidean distance and cosine similarity are calculated between the reconstructed features and the original input features to construct a weighted loss function, yielding the feature reconstruction loss. Furthermore, anomalous abrupt changes and gradual trends are identified through comparative analysis to obtain the anomaly reconstruction loss.

[0069] An adaptive threshold is set for abnormal reconstruction loss based on the distribution characteristics of historical data to distinguish between normal fluctuations and abnormal deviations, resulting in an anomaly level classification. The specific process includes fitting the probability distribution of historically accumulated abnormal reconstruction loss data, constructing a nonparametric probability density function using kernel density estimation, calculating reconstruction loss values ​​at multiple quantiles as judgment criteria for different levels, and obtaining a static threshold set. Combining the parameter distribution characteristics of different stages of the evaporation process, piecewise linear transformation is applied to obtain stage-adaptive thresholds. Sliding window averaging and exponential smoothing methods are applied to the stage-adaptive thresholds to eliminate the influence of short-term fluctuations, resulting in smoothed adaptive thresholds. Real-time abnormal reconstruction loss is compared with the smoothed adaptive thresholds, classified according to the degree to which it exceeds the thresholds of different quantiles, and the anomaly level is updated based on the duration of the anomaly and the accumulated anomaly value, ultimately obtaining an anomaly index for the evaporation process. Based on the anomaly index, real-time adjustments are made to the RF power input and substrate temperature control, generating corrective control commands to obtain the target control scheme for thin film growth within the evaporation chamber. The detailed process is as follows: The anomaly index of the evaporation process is integrated with historical operating data of the evaporation process to construct a digital twin model of the evaporation process, simulating the influence of different control parameters on thin film growth, and obtaining the control parameter influence mapping. Based on the influence mapping of control parameters, a policy gradient algorithm is used to train the control policy network. Anomalies in the evaporation process are taken as input, and the outputs are adjustments to the RF power and substrate temperature, resulting in a parameter control policy. Physical constraints are then applied to the parameter control policy to limit the range and rate of change of the RF power and substrate temperature, preventing abrupt adjustments that could lead to system instability, thus obtaining a safe control policy.

[0070] Based on a safety control strategy, large-scale parameter adjustments are decomposed into a series of small steps to form a smooth-transition control trajectory, resulting in a progressive control sequence. This progressive control sequence is converted into instruction signals in the standard format of an industrial control interface and written to the programmable logic controller (PLC) of the vapor deposition equipment via a communication interface to execute the control instructions. The execution results of the control instructions are monitored in real time, recording changes in abnormal indicators before and after parameter adjustments to generate control effect evaluation data, thus obtaining the target control scheme for thin film growth within the vapor deposition chamber.

[0071] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0072] Temperature data is collected by temperature sensors installed at the evaporation source, substrate and cavity wall. The raw temperature distribution data is obtained by combining local regression model with sliding window analysis.

[0073] Vacuum data is collected from pressure sensors installed at different locations in the cavity. The raw pressure gradient data is obtained by multi-point averaging and outlier removal.

[0074] The power input data recorded in real time by the power monitor is sampled at a fixed frequency and marked with abrupt change points to obtain the raw power fluctuation data.

[0075] Wavelet transform filtering technology is applied to the original temperature distribution data, original pressure gradient data, and original power fluctuation data to retain signal characteristics while removing high-frequency noise, resulting in filtered parameter data.

[0076] The filtered parameter data is subjected to zero-mean normalization to convert parameters of different dimensions into a standard distribution format, thereby obtaining standardized parameter data.

[0077] By aligning sensor data with different sampling frequencies in the standardized parameter data using timestamps, a unified time reference is established, and the first derivative and moving average features are extracted to obtain the vapor deposition process parameter data.

[0078] Specifically, temperature data is collected by temperature sensors installed at the evaporation source, substrate, and cavity wall. Raw temperature distribution data is obtained through a local regression model combined with sliding window analysis. The temperature sensors include K-type thermocouples with a range of 20℃ to 1200℃ and a sampling frequency of 10Hz. The local regression model is a non-parametric regression method that obtains a smooth curve by weighted fitting of local regions near each data point. In the implementation, LOESS (Locally Weighted Scatter Smoothing) is used to process the temperature data, with a sliding window length of 30 seconds and a window overlap rate of 50%. The LOESS algorithm first selects a quadratic polynomial fitting function within each window, then applies a cubic kernel function to assign weights to the data points within the window. The weights decrease as the distance between the data point and the center point increases, thereby achieving smoothing of local data and identification of abrupt changes. Vacuum degree data is collected by pressure sensors installed at different locations within the cavity. Raw pressure gradient data is obtained through multi-point averaging sampling and outlier removal. The pressure sensors are capacitive diaphragm vacuum gauges with a sampling frequency of 5Hz. Multi-point averaging takes the arithmetic mean of multiple consecutive sampling points as a single data point, reducing the impact of random fluctuations. In practice, every 10 original sampling points are merged into one data point. Outlier removal uses the median absolute deviation (MAD) method, calculated as follows: first, the median of all data is calculated; then, the absolute deviation of each data point from the median is calculated; finally, the median of these deviations (i.e., the MAD value) is calculated. If a data point deviates from the median by more than 3 times the MAD value, it is marked as an outlier and removed, replaced by the average of the preceding and following data points.

[0079] The power input data recorded in real time by the power monitor is used to obtain the raw power fluctuation data through fixed-frequency sampling and abrupt change point marking. The power monitor records the RF power input value with a range of 0-5kW and a raw sampling frequency of 1kHz. Fixed-frequency sampling involves resampling the raw data acquired at high frequencies at fixed time intervals; in actual processing, it is downsampled to 100Hz. Abrupt change point marking uses the CUSUM (cumulative sum) control chart method. By calculating the cumulative deviation of the data sequence, abrupt change point is marked when the cumulative deviation exceeds a preset threshold. The CUSUM algorithm consists of two parts: rise detection and fall detection. The functions S+(t)=max[0, S+(t-1)+(x(t)-μ0-k)] and S-(t)=max[0, S-(t-1)-(x(t)-μ0+k)] are calculated respectively, where μ0 is the baseline mean and k is a sensitivity parameter set to half the standard deviation. Abrupt change point is marked when S+ or S- exceeds the threshold h (set to 5 times the standard deviation).

[0080] Wavelet transform filtering was applied to the original data of temperature distribution, pressure gradient, and power fluctuation to preserve signal characteristics while removing high-frequency noise, resulting in filtered parameter data. Wavelet transform is a time-frequency analysis method suitable for processing non-stationary signals, providing both time and frequency domain information. In practice, the Daubechies wavelet basis (db4) was used with a three-level decomposition. Wavelet decomposition separates the signal into approximation coefficients and detail coefficients. Noise is removed by thresholding the detail coefficients (retaining large coefficients and setting small coefficients to zero), and then the signal is reconstructed. A general thresholding method was used. This method effectively removes high-frequency electromagnetic interference and noise caused by equipment vibration during the vapor deposition process, while preserving the true characteristics of temperature, pressure, and power changes. Zero-mean normalization was then applied to the filtered parameter data, converting parameters of different dimensions into a standardized distribution format to obtain standardized parameter data. Zero-mean normalization, also known as Z-score normalization, is calculated by subtracting the mean from the original data and then dividing by the standard deviation, resulting in a mean of 0 and a standard deviation of 1 for the processed data. In practice, the mean and standard deviation are calculated separately for each parameter series, and then a standardization formula is applied. This step unifies parameters with different dimensions and numerical ranges, such as temperature (°C), pressure (Pa), and power (W), to the same scale, facilitating subsequent unified processing and comparative analysis.

[0081] By aligning sensor data with different sampling frequencies in the standardized parameter data using timestamps, a unified time reference is established, and first-order derivative and moving average features are extracted to obtain the vapor deposition process parameter data. Since different sensors have different sampling frequencies (temperature 10Hz, pressure 5Hz, power 100Hz), a unified time reference is needed for data alignment. In actual processing, the lowest sampling frequency (5Hz) is selected as the reference, and other sensor data are adjusted to the same time point through linear interpolation or downsampling. The first-order derivative feature is obtained by calculating the difference between adjacent time points, reflecting the rate of parameter change; the moving average feature uses a 10-point sliding window to calculate the mean, reflecting the smoothing trend of the parameters. These time-series features enrich the data dimensions and provide a foundation for subsequent pattern recognition and anomaly detection.

[0082] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0083] The vapor deposition process parameter data is divided into fixed windows, and the statistical moment characteristics and trend characteristics of evaporation source temperature, cavity pressure and deposition rate in each window are calculated to obtain the time domain characteristics of the parameters.

[0084] A fast Fourier transform is performed on the vapor deposition process parameter data to extract the power spectral density curve, identify the periodic fluctuation components of the evaporation source temperature, cavity pressure, and deposition rate, and obtain the parameter frequency domain characteristics.

[0085] Wavelet transform or Hilbert transform is applied to the vapor deposition process parameter data to capture the time-varying frequency characteristics of the evaporation source temperature, cavity pressure and deposition rate, and obtain the time-frequency characteristics of the parameters;

[0086] The sample entropy and multi-scale entropy values ​​are calculated for the time-domain features, frequency-domain features, and time-frequency features of the parameters to quantify the complexity of parameter fluctuations and obtain parameter complexity features.

[0087] Based on the time-domain characteristics, frequency-domain characteristics, time-frequency characteristics, and complexity characteristics of the parameters, the Pearson correlation coefficient and mutual information value between the evaporation source temperature, cavity pressure, and deposition rate are calculated to obtain the parameter correlation matrix.

[0088] Partial correlation analysis and conditional independence tests are applied to the parameter correlation matrix to eliminate the influence of indirect correlation, determine the causal relationship and time delay between parameters, and obtain the key parameter characteristics of the vapor deposition process.

[0089] Specifically, the evaporation process parameter data is divided into fixed windows. The statistical moment characteristics and trend characteristics of the evaporation source temperature, chamber pressure, and deposition rate within each window are calculated to obtain the time-domain characteristics of the parameters. Fixed window segmentation divides the time-series data into multiple continuous data segments of a fixed length, with each window length set to 30 seconds and an overlap rate of 50% between adjacent windows to ensure the capture of continuous changes. Within each window, four statistical moment characteristics of the evaporation source temperature, chamber pressure, and deposition rate are calculated: mean (average level of the data), standard deviation (dispersion of the data), skewness (asymmetry of the data distribution), and kurtosis (peaks or flatness of the data distribution). Simultaneously, trend characteristics are calculated, including monotonicity (using the Mann-Kendall trend test to determine if there is an upward or downward trend), periodicity (detecting repetitive patterns of the signal through the autocorrelation function), and inflection point density (calculating the number density of points where the first derivative of the curve changes sign, reflecting the frequency of fluctuations). Fast Fourier Transform (FFT) was performed on the vapor deposition process parameter data to extract the power spectral density curve, identify the periodic fluctuations in evaporation source temperature, chamber pressure, and deposition rate, and obtain the frequency domain characteristics of the parameters. FFT is an efficient algorithm for converting time-domain signals to the frequency domain, revealing the hidden periodic components within the signal. The processing first preprocesses the data, including detrending (subtracting the linear fitting curve) and windowing (using a Hanning window to reduce spectral leakage). Then, the FFT algorithm is applied to convert the preprocessed data to the frequency domain, calculate the power spectral density, and obtain the energy distribution of different frequency components. Features such as the main frequency (the frequency corresponding to the energy peak), frequency bandwidth (half-width at half maximum), and energy distribution characteristics (the proportion of energy in different frequency ranges) are extracted from the power spectral density curve to identify the periodic components and frequency distribution characteristics in the data.

[0090] Wavelet transform or Hilbert transform is applied to vapor deposition process parameter data to capture the time-varying frequency characteristics of evaporation source temperature, cavity pressure, and deposition rate, thus obtaining the time-frequency features of the parameters. Wavelet transform is a time-frequency analysis tool suitable for analyzing non-stationary signals, providing both time and frequency domain information simultaneously. In practical processing, continuous wavelet transform is used, with Morlet wavelets chosen as the wavelet basis function for their excellent time-frequency localization capabilities. By calculating wavelet coefficients at different scales, a wavelet spectrum is formed, displaying the frequency distribution of the signal at different time points. Time-varying frequency features, including instantaneous frequency (the dominant frequency changing with time), energy concentration regions, and frequency change slopes, are extracted from the wavelet spectrum to describe the dynamic frequency characteristics of the parameters. Hilbert transform can also be used for similar purposes, extracting instantaneous frequency and instantaneous amplitude information by calculating the analytic signal.

[0091] Sample entropy and multi-scale entropy values ​​are calculated for the time-domain, frequency-domain, and time-frequency characteristics of the parameters to quantify the complexity of parameter fluctuations and obtain parameter complexity features. Sample entropy is a method to measure the complexity or irregularity of a time series; a higher value indicates a more complex or random sequence. The calculation process includes setting the embedding dimension m (usually 2 or 3) and the similarity tolerance r (usually 0.2 times the standard deviation), counting the number of patterns in the sequence that satisfy the similarity condition, and then calculating the negative logarithm of the conditional probability. Multi-scale entropy, on the other hand, coarsely processes the original sequence at different scales and then calculates the sample entropy at each scale, reflecting the complexity of the signal at different time scales. By calculating the sample entropy and multi-scale entropy (usually scale factors from 1 to 20) for each parameter feature, a quantitative index of parameter fluctuation complexity is obtained, helping to identify the complexity and regular changes in the system state.

[0092] Based on the time-domain, frequency-domain, time-frequency, and complexity characteristics of the parameters, the Pearson correlation coefficient and mutual information value among evaporation source temperature, cavity pressure, and deposition rate are calculated to obtain the parameter correlation matrix. The Pearson correlation coefficient measures the degree of linear correlation between two variables, ranging from -1 to 1, with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation; a larger absolute value indicates a stronger correlation. The calculation process involves multiplying the covariance of each pair of parameter features by its respective standard deviation. The mutual information value measures the non-linear dependence between two variables, calculated as the expected logarithmic value of the ratio of the joint probability distribution to the marginal probability distribution of the two variables. A larger mutual information value indicates more shared information between the two variables. By calculating the Pearson correlation coefficient and mutual information value among all parameter feature pairs, a parameter correlation matrix is ​​constructed to demonstrate the strength of linear and non-linear correlations between the parameters.

[0093] Partial correlation analysis and conditional independence tests are applied to the parameter correlation matrix to eliminate the influence of indirect correlations, determine the causal relationship and time delay between parameters, and obtain the key parameter characteristics of the vapor deposition process. Partial correlation analysis calculates the direct correlation between two variables by controlling for the influence of other variables, eliminating indirect correlations caused by intermediate variables. The calculation process is based on conditional probability theory, solving for the correlation coefficient between the two variables after controlling for other variables. Conditional independence tests use statistical hypothesis testing methods to determine whether two variables are independent given a third variable; commonly used methods include the G² test based on conditional mutual information or Fisher's Z-transform test. Furthermore, to identify the time delay relationship between parameters, cross-correlation functions and Granger causality tests are applied. The cross-correlation function calculates the correlation between two sequences at different time lags, finding the time lag value corresponding to the maximum correlation, representing the time delay between parameters. The Granger causality test constructs an autoregressive model to determine whether one time series helps predict the future value of another time series, determining the causal direction between variables.

[0094] Based on the above analysis, key parameter characteristics of the evaporation process are identified, including important characteristic values ​​of each parameter, direct correlations between parameters, causal directions, and time delay information, providing a foundation for subsequent anomaly detection and control. Taking the high-precision optical filter film evaporation process as an example, the evaporation process parameter data of a single production process are analyzed. First, the data is divided into 30-second windows. The calculated mean temperature of the evaporation source is 1080℃, with a standard deviation of 8.5℃, a skewness of 0.12 (slightly right-skewed), and a kurtosis of 2.8 (close to a normal distribution); the mean chamber pressure is 2.3×10^-4 Pa, with a standard deviation of 3.7×10^-5 Pa; and the mean deposition rate is 0.5 nm / s, with a standard deviation of 0.08 nm / s. FFT analysis reveals a major fluctuation component in the temperature data with a period of approximately 120 seconds, and a significant peak in the power spectral density. Wavelet transform analysis shows that this periodic fluctuation is most pronounced within 20-30 seconds after the evaporation source power is adjusted, and then gradually decays. Sample entropy calculations show that the sample entropy of the temperature sequence under normal operating conditions is 0.62, while it increases to 0.85 during abnormal fluctuations, indicating increased system complexity. The parameter correlation matrix shows that the Pearson correlation coefficient between evaporation source temperature and deposition rate is 0.83, with a mutual information value of 0.76, indicating a strong correlation; the correlation coefficient between evaporation source temperature and cavity pressure is 0.65, with a mutual information value of 0.58, indicating a moderate correlation. Further partial correlation analysis eliminated the influence of indirect correlations, finding that after controlling the evaporation source temperature, the partial correlation coefficient between cavity pressure and deposition rate decreased to 0.23, indicating that their correlation is mainly transmitted through temperature changes. Cross-correlation analysis determined a 5-second time delay between power adjustment and temperature change, and a 3-second time delay between temperature change and deposition rate change. This multi-dimensional feature extraction and correlation analysis overcomes the limitations of traditional methods in uncovering deep-seated relationships between parameters.

[0095] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0096] Cluster analysis was performed on the key parameter characteristics of the vapor deposition process, and the process operation mode was divided using a Gaussian mixture model to obtain a standard mode library for the vapor deposition process.

[0097] For each process mode in the standard mode library of the vapor deposition process, the characteristic multidimensional distribution parameters, including the mean vector and covariance matrix, are calculated to obtain the mode statistical characterization;

[0098] For the key parameters of the vapor deposition process collected in real time, the Mahalanobis distance and Hotling statistic compared with the statistical characterization of the model are calculated to obtain the parameter deviation.

[0099] Based on the parameter deviation, a deep autoencoder network is constructed to encode and decode the input features, calculate the reconstruction error, and obtain the abnormal reconstruction loss.

[0100] For the aforementioned abnormal reconstruction loss, an adaptive threshold is set based on the distribution characteristics of historical data to distinguish between normal fluctuations and abnormal deviations, thereby obtaining an abnormality level classification;

[0101] Based on the aforementioned anomaly classification, the anomaly start time, duration, severity score, and contribution of related parameters are calculated to obtain the anomaly index of the vapor deposition process.

[0102] Specifically, cluster analysis is performed on the key parameter characteristics of the vapor deposition process, and a Gaussian mixture model is used to classify the process operation modes, resulting in a standard pattern library for the vapor deposition process. A Gaussian mixture model is a probabilistic model that assumes the data is generated by a mixture of multiple Gaussian distributions, each representing a data pattern. In implementation, firstly, a suitable number of clusters is determined, and the optimal number of clusters is automatically determined using the Bayesian information criterion. The Bayesian information criterion comprehensively considers model fit and complexity, seeking a balance between the two to avoid overfitting. Then, the model parameters are initialized, including the weights, mean vectors, and covariance matrices of each Gaussian distribution, using the K-means++ algorithm. Next, the model parameters are iteratively optimized using the expectation-maximization algorithm. This algorithm consists of two steps: the expectation step calculates the posterior probability of each data point belonging to each Gaussian distribution, and the maximization step updates the parameters of each Gaussian distribution. The iterative process continues until the parameters converge or the maximum number of iterations is reached. Through Gaussian mixture model clustering, the key parameter characteristics of the vapor deposition process are divided into multiple process operation modes, such as the start-up stage, stable deposition stage, and power adjustment stage, forming a standard pattern library for the vapor deposition process. For each process mode in the standard pattern library for vapor deposition, multidimensional distribution parameters of features are calculated, including the mean vector and covariance matrix, to obtain a statistical characterization of the mode. The mean vector represents the average level of each parameter feature, and is calculated by taking the arithmetic mean of all data points for each mode. The covariance matrix describes the correlation structure and degree of variation among parameter features, and is calculated by calculating the covariance for each pair of features to form a covariance matrix. The covariance represents the degree to which two variables change together; a positive value indicates a positive correlation, a negative value indicates a negative correlation, and zero indicates no correlation. The diagonal elements are the variances of each feature. By calculating the mean vector and covariance matrix of each process mode, a statistical characterization of the mode is established, providing a benchmark for subsequent anomaly detection. In addition, physical interpretation and semantic annotation of each mode are combined with expert knowledge, such as identifying a certain mode as "film stabilization period" or "temperature adjustment period," to enhance the interpretability of the mode.

[0103] For key parameters of the real-time vapor deposition process, the Mahalanobis distance and Hotelling's statistic, statistically represented by the patterns, are calculated to obtain the parameter deviation. Mahalanobis distance measures the distance from a data point to the distribution center, considering the correlation between features and suitable for multidimensional data analysis. The calculation method involves multiplying the difference vector between the data point and the mean by the inverse of the covariance matrix, then multiplying by the transpose of the difference vector, and finally taking the square root to obtain the Mahalanobis distance. Hotelling's T² statistic is the square of the Mahalanobis distance multiplied by the sample size, used for multivariate hypothesis testing to assess whether data points significantly deviate from the distribution center. In practical processing, the most likely process mode is first identified by calculating the probability density of real-time data and each mode, selecting the mode with the highest probability density as the current mode. Then, the Mahalanobis distance and Hotelling's T² statistic, statistically represented by the real-time data and the current mode, are calculated, and the results are compared with an adaptive threshold to determine if there is a significant deviation. The threshold setting considers the multidimensionality of the data and the pattern covariance structure, commonly using critical values ​​based on chi-square or F-distribution, with a confidence level set at 95%. In this way, parameter deviation is obtained, reflecting the degree to which real-time data deviates from the normal working mode. Based on the parameter deviation, a deep autoencoder network is constructed to encode and decode the input features, calculate the reconstruction error, and obtain the anomaly reconstruction loss. A deep autoencoder is an unsupervised learning neural network consisting of an encoder and a decoder. Its goal is to learn the latent space representation of the input data and reconstruct the original input. The encoder uses a four-layer convolutional neural network with 32, 64, 128, and 256 kernels per layer, a kernel size of 3×3, and LeakyReLU activation to prevent gradient vanishing. Max pooling is applied after each convolutional layer to reduce feature dimensionality and extract key features. The decoder uses a structure symmetrical to the encoder, reconstructing the original input through transposed convolution (also known as deconvolution). Normal data is used during training, aiming to minimize the reconstruction loss function, typically choosing mean squared error. After training, the network can reconstruct normal mode data well, but performs poorly on anomaly data. In the anomaly detection phase, real-time data is input into the trained autoencoder, and the error between the input and the reconstructed output is calculated as the anomaly reconstruction loss. The calculation methods include the mean or weighted sum of point-to-point errors, taking into account the importance of different features. Since outlier data often leads to larger reconstruction errors, monitoring the magnitude of the reconstruction error can effectively detect anomalies.

[0104] For abnormal reconstruction losses, adaptive thresholds are set based on the distribution characteristics of historical data to distinguish between normal fluctuations and abnormal deviations, resulting in an anomaly level classification. First, the probability distribution of historically accumulated abnormal reconstruction loss data is fitted, and a nonparametric probability density function is constructed using kernel density estimation to obtain a loss distribution model. Kernel density estimation is a nonparametric statistical method that estimates probability density by superimposing kernel functions (such as Gaussian kernels), avoiding strong assumptions about the data distribution pattern. Based on the fitted loss distribution model, reconstruction loss values ​​at different probability quantiles, such as the 75%, 90%, 95%, and 99% quantiles, are calculated as different level judgment benchmarks, resulting in a static threshold set. To adapt to different stages of the evaporation process, piecewise linear transformation is applied to the static threshold set, considering the parameter distribution characteristics of the start-up, stabilization, and finish-up stages, to obtain stage-adaptive thresholds. The piecewise linear transformation considers the differences in normal fluctuation amplitudes at different stages, making the thresholds more targeted. To eliminate the impact of short-term fluctuations, sliding window averaging and exponential smoothing methods are applied to the stage-adaptive thresholds to obtain smoothed adaptive thresholds. The real-time anomaly reconstruction loss is compared with a smoothing adaptive threshold, and the anomaly is classified according to the degree to which it exceeds different quantile thresholds. For example, exceeding 90% of the quantiles indicates a minor anomaly, exceeding 95% indicates a moderate anomaly, and exceeding 99% indicates a severe anomaly, thus obtaining the initial anomaly level. Finally, considering the temporal characteristics of the anomaly, the initial anomaly level is processed temporally, and the anomaly level is updated based on the duration of the anomaly and the cumulative anomaly value to obtain the final anomaly level classification. Based on the anomaly level classification, the anomaly start time, duration, severity score, and contribution of associated parameters are calculated to obtain the anomaly index of the vapor deposition process. The anomaly start time is determined by backtracking the anomaly reconstruction loss sequence to find the time point when the minor anomaly threshold is first exceeded. The duration is calculated as the length of time from the onset of the anomaly to its disappearance (the loss value falls back to the normal range). The severity score is a comprehensive index that considers the anomaly level, duration, and impact range, and a weighted algorithm is used to calculate the score, typically ranging from 1 to 10 levels. The contribution analysis of associated parameters aims to identify the key parameters that cause the anomaly. This is achieved by calculating the reconstruction error contribution of each original parameter or by using SHAP values ​​to analyze the contribution of each parameter to the anomaly. The SHAP value, based on the Shapley value in game theory, can fairly evaluate the contribution of each feature to the model output. By calculating these anomalous feature indices, a comprehensive set of anomaly indicators for the vapor deposition process is obtained, providing a basis for process control and adjustment.

[0105] In one specific embodiment, the process of performing the encoding and decoding steps of the input features may specifically include the following steps:

[0106] The parameter deviation and key parameter features of the vapor deposition process are concatenated to form an input vector, which is then fed into the input layer of a deep autoencoder network to obtain the features to be processed.

[0107] The features to be processed are compressed using a four-layer convolutional encoder. Each convolutional kernel in the encoder is 3×3 in size, and the number of channels in the convolutional layers is 32, 64, 128, and 256 respectively. An activation function and pooling operation are applied after each convolutional layer to obtain compressed features.

[0108] The compressed features are further dimensionality reduced by a fully connected layer, with the number of nodes being 512, 256, 128, and 64 respectively, to obtain the latent space feature representation;

[0109] The latent space feature representation is restored by a four-layer transposed convolutional decoder. The transposed convolutional decoder adopts a structure symmetrical to the encoder, with the number of channels being 256, 128, 64, and 32 respectively, to obtain the reconstructed features.

[0110] Calculate the Euclidean distance and cosine similarity between the reconstructed features and the original input features, construct a weighted loss function, and obtain the feature reconstruction loss.

[0111] The feature reconstruction loss is compared and analyzed with the reconstruction loss of time-series continuous samples to identify abnormal mutation points and gradual trends, and the abnormal reconstruction loss is obtained.

[0112] Specifically, the parameter deviation and key parameter features of the vapor deposition process are concatenated to form an input vector, which is then fed into the input layer of a deep autoencoder network to obtain the features to be processed. Feature concatenation is the process of combining multiple related features into a unified vector. In this method, the parameter deviation calculated using Mahalanobis distance and Hotelling statistics is merged with the key parameter features of the vapor deposition process. In practice, the parameter deviation is directly connected as a scalar value to the key parameter feature vector to form an augmented feature vector. For example, the evaporation source temperature feature set, the cavity pressure feature set, and the deposition rate feature set each contain multiple time-domain, frequency-domain, and time-frequency features, which, together with the deviation index, constitute a complete input vector. To ensure that the numerical scales of each dimension of the input vector are similar, normalization is applied to map each feature value to the [0,1] interval. The input vector is usually reorganized into a two-dimensional tensor form to facilitate processing by convolutional neural networks. For example, a one-dimensional vector can be reshaped into a matrix form, preserving the spatial or temporal relationships of the original features to obtain the features to be processed. The features to be processed are compressed using a four-layer convolutional encoder. Each layer in the encoder has a 3×3 kernel size, with the number of channels in each layer being 32, 64, 128, and 256 respectively. An activation function and pooling operation are applied after each convolutional layer to obtain the compressed features. The convolutional encoder is the encoding part of the deep autoencoder, extracting a high-level representation of the input features through convolutional operations. The first convolutional layer uses 32 3×3 kernels to convolve the input features with a stride of 1 and "same" padding to maintain the feature map size. The convolution operation performs a dot product operation with the input features through a sliding window to capture local feature patterns. After convolution, the LeakyReLU activation function is applied, with the form f(x) = max(0.01x, x). Compared to the traditional ReLU activation function, LeakyReLU has a smaller slope in the negative region, avoiding the "dead neuron" problem. After activation, a 2×2 max pooling operation is applied with a stride of 2, halving the feature map size while preserving salient features. The second convolutional layer uses 64 3×3 convolutional kernels to process the feature map output from the first layer. The convolution, activation, and pooling settings are the same as the first layer. The third layer uses 128 convolutional kernels, and the fourth layer uses 256 convolutional kernels, increasing the number of feature channels layer by layer to extract more abstract feature representations. After four layers of convolution and pooling operations, the feature map size is significantly reduced while the number of channels increases, achieving information compression and abstract representation, resulting in compressed features.

[0113] The compressed features are further reduced in dimensionality using fully connected layers, with the number of nodes decreasing sequentially to 512, 256, 128, and 64, resulting in the latent space feature representation. The fully connected layers further compress the spatial features extracted by the convolutional layers into a compact one-dimensional vector representation. The compressed features after four convolutional layers are flattened into a one-dimensional vector and then connected to the first fully connected layer, which contains 512 neurons. Each neuron is fully connected to all elements of the input vector, and a weighted sum is calculated before applying the LeakyReLU activation function. To prevent overfitting, Dropout regularization is applied after the fully connected layer, randomly discarding a portion of the neuron outputs; the Dropout ratio is set to 0.3. The second fully connected layer contains 256 neurons and also applies the LeakyReLU activation function and Dropout. The third and fourth layers contain 128 and 64 neurons respectively, forming a progressively decreasing "funnel" structure. The final 64-dimensional vector is the latent space feature representation, also known as the "latent encoding" or "bottleneck layer," which is a highly compressed and abstract representation of the entire input feature. The latent space feature dimension is much smaller than the original input, forcing the network to learn the essential structure of the data and discard redundant and noisy information. Because normal data has inherent regularity, its latent space representation is usually more compact and consistent, while anomalous data is difficult to compress effectively, exhibiting significantly different latent space features. The latent space feature representation is restored through a four-layer transposed convolutional decoder. The transposed convolutional decoder adopts a structure symmetrical to the encoder, with channel numbers of 256, 128, 64, and 32 respectively, resulting in reconstructed features. Transposed convolution is the inverse operation of convolution, mapping low-dimensional features back to high-dimensional space; it is also called deconvolution or de-convolution. First, the 64-dimensional latent space features are reshaped into a small-sized feature map and connected to the first transposed convolutional layer. This layer uses 256 3×3 transposed convolutional kernels with a stride of 2. By inserting blank pixels and applying convolution operations, the feature map size is enlarged. The LeakyReLU activation function is also applied after the transposed convolution. The second layer uses 128 transposed convolutional kernels, the third layer uses 64, and the fourth layer uses 32, with the number of channels decreasing layer by layer, symmetrical to the encoder. The four transposed convolutional operations progressively restore the spatial dimensions of the feature map, ultimately outputting reconstructed features with the same size as the original input. The decoder's structure follows the encoder's mirror principle to ensure a balanced information flow. For example, pooling operations downsample in the encoder, while transposed convolutions upsample in the decoder; as the number of channels increases in the encoder, the number of channels decreases in the decoder. This symmetrical design facilitates efficient information reconstruction, and ideally, the final reconstructed features should closely approximate the original input features.

[0114] The Euclidean distance and cosine similarity are calculated between the reconstructed features and the original input features to construct a weighted loss function, resulting in the feature reconstruction loss. Euclidean distance measures the straight-line distance between two vectors in Euclidean space, calculated by taking the square root of the sum of the squares of the differences between corresponding elements. Specifically, the difference between each corresponding element of the reconstructed feature and the original input feature is calculated, squared, summed, and the square root is taken to obtain the overall reconstruction error. Cosine similarity measures the similarity in direction between two vectors, calculated by dividing the vector dot product by the product of their respective magnitudes. The cosine similarity range is [-1, 1], with values ​​closer to 1 indicating greater similarity. Both measures have their advantages: Euclidean distance is sensitive to magnitude differences, while cosine similarity emphasizes directional consistency. The two measures are combined to construct a weighted loss function, with the formula L = α × Euclidean distance - β × cosine similarity, where α and β are weighting coefficients, adjusted according to the specific application; typically α = 0.7 and β = 0.3. This combined loss function considers both the accuracy of feature reconstruction and the structural similarity of feature vectors, improving the sensitivity and accuracy of anomaly detection. The calculated feature reconstruction loss reflects the degree of deviation between the input data and the normal pattern; a larger loss value indicates a more significant deviation.

[0115] This study compares and analyzes the feature reconstruction loss with the reconstruction loss of continuous temporal samples to identify anomalous abrupt changes and gradual trends, thus obtaining the anomalous reconstruction loss. Continuous temporal samples refer to multiple samples collected sequentially over time. The comparative analysis involves comparing the reconstruction loss of the current sample with that of historical samples. First, a time series of reconstruction losses is constructed, recording and storing the feature reconstruction loss at each time point. Then, a sliding window technique is applied, with a window length typically between 10 and 20 time points, to calculate the statistical characteristics of the reconstruction loss within the window, such as mean, standard deviation, and trend slope. For identifying anomalous abrupt changes, a change point detection algorithm is used to calculate significant changes in reconstruction loss between adjacent windows. If the difference in the mean loss between adjacent windows exceeds three times the standard deviation, it is marked as an abrupt change. For identifying gradual trends, linear regression is used to estimate the slope of the loss value within the window. A slope significantly greater than zero indicates an increasing degree of anomalousness, while a slope significantly less than zero indicates that the system is recovering to normal. By comprehensively considering the absolute value, relative change, and trend change of the reconstruction loss, an anomalous reconstruction loss index is constructed. This index not only reflects the degree of anomalousness in the current state but also includes information on the evolution trend of the anomalousness. The anomaly reconstruction loss metric is a comprehensive score that takes into account both point anomalies (instantaneous deviations) and contextual anomalies (abnormal change patterns), thereby enhancing the robustness and accuracy of anomaly detection.

[0116] In one specific embodiment, the process of performing the anomaly reconstruction loss step may specifically include the following steps:

[0117] The probability distribution of historically accumulated anomaly reconstruction loss data is fitted, and a nonparametric probability density function is constructed using the kernel density estimation method to obtain the loss distribution model;

[0118] Based on the loss distribution model, the reconstruction loss values ​​of multiple quantiles are calculated and set as different levels of judgment criteria to obtain a set of static thresholds;

[0119] For the static threshold set, combined with the parameter distribution of the evaporation process in the start-up, stabilization and end stages, a piecewise linear transformation is applied to obtain the stage adaptive threshold.

[0120] For the aforementioned stage adaptive threshold, a sliding window averaging and exponential smoothing method is applied to eliminate the impact of short-term fluctuations and obtain a smoothed adaptive threshold.

[0121] The real-time anomaly reconstruction loss is compared with the smooth adaptive threshold, and the anomaly level is obtained by classifying the degree to which it exceeds the threshold of different level quantiles.

[0122] The initial anomaly level is processed in a time series, and the anomaly level is updated according to the anomaly duration and the accumulated anomaly value to obtain the anomaly level classification.

[0123] Specifically, the probability distribution of historically accumulated anomaly reconstruction loss data is fitted, and a nonparametric probability density function is constructed using kernel density estimation to obtain a loss distribution model. Kernel density estimation is a nonparametric statistical method that does not pre-determine the data distribution form but directly learns the distribution characteristics from the data itself. In the implementation process, accumulated anomaly reconstruction loss data under normal operating conditions is extracted from the historical database, typically containing thousands to tens of thousands of data points. The basic idea of ​​kernel density estimation is to treat each observation point as a source of probability quality and smoothly distribute these qualities through a kernel function. In the specific calculation, an appropriate kernel function (usually a Gaussian kernel) and bandwidth parameter are selected. The bandwidth controls the degree of smoothing; too small a bandwidth leads to overfitting, and too large a bandwidth leads to underfitting. Generally, the Silverman rule is used to automatically determine the optimal bandwidth. By placing the kernel function at each data point and summing the results, the overall probability density function is obtained, forming a continuous and smooth distribution curve that accurately reflects the distribution characteristics of anomaly reconstruction loss, including multimodal characteristics and tail behavior. Based on the loss distribution model, the reconstruction loss values ​​at multiple quantiles are calculated and set as different levels of judgment criteria to obtain a static threshold set. Quantiles are points that divide a probability distribution into equally probable regions. Typical quantiles include 75% (minor anomaly threshold), 90% (moderate anomaly threshold), 95% (severe anomaly threshold), and 99% (extreme anomaly threshold). The calculation process is based on a reverse lookup of the cumulative distribution function. This involves integrating the probability density function of the loss distribution model to obtain the cumulative distribution function, and then finding the loss value corresponding to the cumulative probability reaching a specified percentage. For example, the 95% quantile means that 95% of normal data have a loss value less than this threshold, and data points with losses exceeding this threshold have a 95% probability of being anomalies. By calculating different quantiles, multi-level anomaly judgment criteria are established, forming a static threshold set. This threshold setting method based on probability distribution is more statistically grounded than simple fixed thresholds, automatically adjusting the judgment criteria according to the actual data distribution characteristics, and reducing human intervention.

[0124] For the static threshold set, combined with the parameter distribution of the evaporation process in the start-up, stabilization, and termination stages, a piecewise linear transformation is applied to obtain stage-adaptive thresholds. The evaporation process is generally divided into three main stages: the start-up stage (initial heating and evaporation stabilization process), the stabilization stage (main film formation process), and the termination stage (power reduction and cooling process). The parameter fluctuation characteristics differ significantly between these stages; the start-up and termination stages exhibit larger parameter variations, while the stabilization stage requires high stability. The piecewise linear transformation is a method for adjusting static thresholds based on the characteristics of each process stage. The basic principle is to apply different transformation coefficients to different stages. In practice, the current process stage is first identified based on the status judgment of key parameters such as evaporation power, chamber temperature, and deposition rate. Then, threshold adjustment coefficients are set for different stages. During the start-up stage, due to normal fluctuations, the threshold is relaxed by 20%-30%; during the stabilization stage, strict control is required, so the threshold remains unchanged or is slightly tightened by 5%-10%; the termination stage also appropriately relaxes the threshold. Through stage identification functions and linear transformation functions, the static thresholds are mapped to dynamic thresholds that match the process stages, making anomaly detection more consistent with actual process characteristics.

[0125] For stage-adaptive thresholds, sliding window averaging and exponential smoothing are applied to eliminate the impact of short-term fluctuations, resulting in a smoothed adaptive threshold. Sliding window averaging is a commonly used time-series data smoothing method that replaces the original value by calculating the arithmetic mean of data within a fixed-length window. In implementation, an appropriate window length (usually 10-20 data points) is set, and sliding window averaging is applied to the stage-adaptive threshold sequence. For the smoothing threshold at time point t, the average of the thresholds within half the window length before and after t is calculated. Exponential smoothing assigns exponentially decreasing weights to data at different time points, with the most recent data having the highest weight, which gradually decreases over time. In implementation, an appropriate smoothing coefficient α is chosen (usually between 0.1 and 0.3), and the current smoothed value is equal to the weighted combination of the previous smoothed value and the current observation. Exponential smoothing is particularly suitable for handling data with trending changes, preserving the data's trend while smoothing noise. Combining these two smoothing methods—sliding window averaging for handling local fluctuations and exponential smoothing for tracking long-term trends—results in a smoothed adaptive threshold that eliminates the impact of short-term fluctuations and adapts to the gradual changes in process parameters.

[0126] The real-time anomaly reconstruction loss is compared with a smoothing adaptive threshold, and the anomaly level is determined by classifying the degree to which it exceeds different quantile thresholds. The real-time anomaly reconstruction loss is the current time-stamped loss value calculated by a deep autoencoder network, while the smoothing adaptive threshold is a multi-level threshold sequence obtained through the aforementioned steps. The comparison process first calculates the normalized deviation, i.e., the ratio of the real-time loss value to the current threshold, and then determines the initial anomaly level based on the magnitude of the deviation and the threshold level exceeded. Anomalies are typically classified into three levels: minor anomalies (loss value exceeds the 75% quantile threshold but not the 90% quantile threshold), moderate anomalies (loss value exceeds the 90% quantile threshold but not the 95% quantile threshold), and severe anomalies (loss value exceeds the 95% quantile threshold). For extreme cases, such as loss values ​​exceeding the 99% quantile threshold, anomalies are marked as extreme anomalies. The initial anomaly level is an instantaneous evaluation result calculated based on a single time point, reflecting the degree of anomaly in the current state and providing a basis for subsequent time-series processing.

[0127] The initial anomaly level is processed temporally. Based on the duration of the anomaly and the cumulative anomaly value, the anomaly level is updated to obtain the anomaly level classification. This temporal processing considers the persistence and cumulative effect of the anomaly state, avoiding over-alarming of isolated anomalies and under-alarming of continuous anomalies. During implementation, an anomaly state buffer is maintained, recording the anomaly level sequence within a recent period (e.g., the past 30 minutes). Temporal analysis is performed on the anomaly levels within the buffer, including duration statistics and cumulative anomaly value calculation. Duration statistics calculate the duration of consecutive anomalies; if consecutive moderate anomalies occur for a specified time (e.g., 60 seconds), the anomaly level is upgraded to a severe anomaly. The cumulative anomaly value calculation considers the cumulative effect of anomalies, performing a weighted sum of the anomaly values ​​in the buffer, with the weights decaying over time. When the cumulative anomaly value exceeds a set threshold, the anomaly level is upgraded even if a single anomaly is not severe. Furthermore, the trend of anomaly changes is considered; trend analysis is performed on the anomaly level sequence. If the anomaly severity continues to increase, the warning is upgraded; if the anomaly severity continues to decrease, a downgrade is considered. These timing processing rules enable more accurate anomaly classification, reduce false alarms and missed alarms, and improve the reliability and usability of the monitoring system.

[0128] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0129] By integrating the abnormal indicators of the vapor deposition process with the historical operation data of the vapor deposition process, a digital twin model of the vapor deposition process is constructed to simulate the influence of different control parameters on thin film growth and obtain the influence mapping of control parameters.

[0130] Based on the control parameter influence mapping, a control strategy network is trained using a strategy gradient algorithm. Abnormal indicators of the evaporation process are taken as inputs, and the adjustment amounts of RF power and substrate temperature are output to obtain the parameter control strategy.

[0131] Physical constraints are checked on the parameter control strategy to limit the range and rate of change of RF power and substrate temperature, preventing system instability caused by drastic adjustments, thus obtaining a safe control strategy.

[0132] Based on the aforementioned safety control strategy, large-scale parameter adjustments are decomposed into multiple small-step sequences to form a smooth-transition control trajectory, resulting in a progressive control sequence.

[0133] The progressive control sequence is converted into instruction signals in the standard format of the industrial control interface, and written into the programmable logic controller of the vapor deposition equipment through the communication interface to obtain the execution control instructions;

[0134] The execution results of the control commands are monitored in real time, and the changes in abnormal indicators before and after parameter adjustment are recorded to form control effect evaluation data, thereby obtaining the target control scheme for thin film growth in the vapor deposition chamber.

[0135] Specifically, abnormal indicators of the evaporation process are integrated with historical operational data to construct a digital twin model of the evaporation process. This model simulates the impact of different control parameters on thin film growth, obtaining a control parameter influence mapping. The digital twin model is a virtual replica of the physical system, capable of reflecting real-time changes in the system's state. The construction process first extracts historical operational data from a database, including process parameters (such as evaporation source temperature, chamber pressure, and substrate temperature), control parameters (such as RF power and gas flow rate), and quality indicators (such as thin film thickness uniformity and optical performance). Data preprocessing includes outlier removal, missing value imputation, and standardization. Then, a hybrid model architecture is constructed, integrating the physical model and the data-driven model. The physical model, based on evaporation kinetics, particle transport, and film formation theory, describes the theoretical relationships between parameters; the data-driven model uses deep neural networks to learn complex nonlinear relationships in the data. Model training employs a hybrid loss function, simultaneously considering prediction errors and the degree of violation of physical constraints, achieving a deep fusion of physical information and data. The final established digital twin model can accurately simulate the dynamic behavior of the vapor deposition process. By inputting different combinations of control parameters (such as RF power and substrate temperature), it predicts the thin film growth process and final quality, forming a control parameter influence mapping. Based on the control parameter influence mapping, a policy gradient algorithm is used to train the control policy network. The abnormal indicators of the vapor deposition process are used as inputs, and the outputs are the adjustment amounts of RF power and substrate temperature, resulting in the parameter control policy. The policy gradient algorithm is a reinforcement learning method that directly optimizes the control policy rather than the value function. First, the vapor deposition process is modeled as a Markov decision process. The state space consists of abnormal indicators and key process parameters, the action space consists of the adjustment amounts of RF power and substrate temperature, and the reward function comprehensively considers thin film quality, energy consumption, and equipment stability. Specifically, the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is used, which includes a policy network and a dual-Q network. The policy network consists of four fully connected layers. The input layer receives the current state (anomaly indicators and process parameters). The number of nodes in the intermediate hidden layers are 400, 300, 200, and 100 respectively, with ReLU as the activation function. The output layer uses the tanh activation function to map the output to [-1, 1], and then maps it to the actual action range through a linear transformation. The Q-network structure is similar, but the input includes both state and action, and the output is a single Q-value. The training process employs experience replay and target network techniques to reduce sample correlation and improve stability. The experience pool capacity is set to 1 million, with 128 batches sampled each time, and a discount factor of 0.99. To enhance the exploration capability, adaptive Gaussian noise is added to the policy network parameters. Training iterations continue until the performance indicators converge or the preset number of iterations is reached, ultimately obtaining a parameter control strategy that can output optimal control adjustments for different anomaly conditions.

[0136] Physical constraint checks are performed on parameter control strategies to limit the range and rate of change of RF power and substrate temperature, preventing system instability caused by abrupt adjustments and resulting in a safe control strategy. Physical constraint checks are a crucial step in ensuring that the control strategy conforms to actual process limitations. First, safe operating ranges for RF power and substrate temperature are defined, typically based on equipment specifications and process requirements. For example, the RF power range is 10%-90% of the rated power, and the substrate temperature range is room temperature to the equipment's maximum withstand temperature. Then, parameter change rate limits are defined to prevent system oscillation caused by abrupt adjustments. Typical limits include an RF power change rate not exceeding 5% of the rated power per second and a substrate temperature change rate not exceeding 20°C per minute. The constraint check process consists of two steps: range checking and rate checking. The range check ensures that the target value of the adjusted parameters output by the control strategy is within a safe range; if exceeded, it is truncated to the boundary value. The rate check ensures that the parameter change rate does not exceed the limit value; if exceeded, it is adjusted at the maximum permissible rate. Furthermore, parameter interaction constraint checks are also required, as certain parameter combinations may lead to unsafe conditions, such as high power at high temperatures potentially damaging the equipment. By applying these constraints to the strategy network output, a safe control strategy that satisfies both the control objective and physical limitations is obtained.

[0137] Based on a safety control strategy, large-amplitude parameter adjustments are decomposed into a series of smaller steps to form a smooth transition control trajectory, resulting in a progressive control sequence. The progressive control sequence aims to achieve smooth parameter adjustments, avoiding abrupt changes and oscillations in the system response. In practice, it is first determined whether the control adjustment is large-amplitude. The criterion is that the adjustment amount exceeds a preset threshold, such as an RF power adjustment exceeding 10% of the rated power or a substrate temperature adjustment exceeding 30°C. For large-amplitude adjustments, a trajectory planning method is used to decompose them into multiple smaller steps. Trajectory planning considers the starting point (current parameter value), the target point (target parameter value), and the system's dynamic characteristics to design a suitable transition trajectory. Common methods include linear interpolation, polynomial interpolation, and S-curve interpolation. Linear interpolation is the simplest, evenly distributing the adjustment amount across multiple time steps; polynomial interpolation enables a smooth transition between the start and end points, reducing abrupt acceleration changes; S-curve interpolation further limits jerk, providing the smoothest transition. The time step size and total transition time are determined based on the adjustment amplitude and system response characteristics; typically, large-amplitude adjustments are decomposed into 5-10 time steps. For example, adjusting the RF power by 500W can be broken down into five small adjustments of 100W each, with a 10-second interval between each step, forming a smooth transition trajectory of 50 seconds. For small-amplitude adjustments, the original plan can be executed directly without decomposition. This ultimately forms a series of control command sequences that are continuous in time and gradual in amplitude, i.e., a progressive control sequence. The progressive control sequence is converted into instruction signals in the standard format of the industrial control interface and written to the programmable logic controller of the vapor deposition equipment through the communication interface to obtain the execution control instructions. The standard format of the industrial control interface is a common data exchange format for industrial control systems. Common standards include OPC UA (OPC Unified Architecture), Modbus, and Profinet. The conversion process first identifies the type of control system and interface protocol used by the vapor deposition equipment, and then encapsulates the control parameters into standard data packets according to the protocol specifications. For example, for a system using the OPC UA protocol, the RF power and substrate temperature adjustment values ​​need to be encapsulated into OPC UA data nodes, with the correct data type, range, and unit set; for the Modbus protocol, they need to be mapped to specific register addresses. Once the data packet is prepared, it is sent to the programmable logic controller (PLC) of the vapor deposition equipment via a physical layer communication interface such as Ethernet, serial port, or fieldbus. The communication process must handle issues such as timeout retransmission, data verification, and error recovery to ensure reliable instruction delivery. Upon receiving the instruction, the PLC converts the digital instruction into analog control signals, driving the power control unit and temperature control unit to adjust the actual parameters. The entire communication and execution process requires strict timing control and status feedback, typically employing a handshake mechanism to confirm the instruction execution status and form a complete execution control instruction flow.

[0138] The execution results of control commands are monitored in real time, and changes in abnormal indicators before and after parameter adjustments are recorded to form control effect evaluation data, thus obtaining the target control scheme for thin film growth in the evaporation chamber. Real-time monitoring requires the collection of multi-source data, including actual values ​​of control parameters (such as actual readings of RF power and substrate temperature), process parameter responses (such as changes in evaporation source temperature and chamber pressure), and changes in abnormal indicators (such as real-time values ​​of reconstruction loss and deviation). The data acquisition frequency is determined according to the control cycle, typically 1-10 times per second. After preprocessing, the monitoring data is stored in a time-series database. Preprocessing includes noise filtering, outlier detection, and data alignment. Then, the effect of parameter adjustment is evaluated, and key indicators are calculated, including: response time (the time delay from issuing the command to the start of significant parameter change), settling time (the time required to reach a new steady state after parameter adjustment), overshoot (the degree to which the parameter briefly exceeds the target value), and improvement in abnormal indicators (the magnitude of changes in abnormal indicators before and after adjustment). The evaluation results are used to verify the effectiveness of the control strategy and adjust the parameters of the control strategy. If an abnormal indicator does not improve significantly after an adjustment, it indicates that the current control strategy is not effective enough, and the adjustment range needs to be increased or other control parameters should be tried. If the abnormal indicator quickly returns to normal, it indicates that the control strategy is effective. Through continuous monitoring and evaluation, the combination of control parameters and adjustment strategies are continuously optimized, ultimately forming a target control scheme library applicable to different abnormal situations, guiding automatic control decisions in subsequent production processes.

[0139] The data analysis-based vapor deposition process monitoring method in the embodiments of this application has been described above. The data analysis-based vapor deposition process monitoring system in the embodiments of this application is described below. Please refer to [link to relevant documentation]. Figure 2 One embodiment of the data analysis-based vapor deposition process monitoring system in this application includes:

[0140] The data acquisition module 201 is used to acquire raw data from the temperature sensor, pressure sensor, power monitor and thickness monitoring system installed in the vapor deposition chamber, and obtain vapor deposition process parameter data through filtering and standardization.

[0141] The calculation module 202 is used to perform time-frequency analysis and parameter correlation calculation on the vapor deposition process parameter data, extract the variation characteristics of evaporation source temperature, cavity pressure and deposition rate, and obtain the key parameter characteristics of the vapor deposition process.

[0142] The setting module 203 is used to perform deviation analysis by setting a dynamic threshold based on the key parameter characteristics of the vapor deposition process, calculate the degree of deviation between the real-time value of the parameter and the standard value, and obtain the abnormal index of the vapor deposition process.

[0143] The adjustment module 204 is used to adjust the RF power input and substrate temperature control in real time according to the abnormal indicators of the evaporation process, generate correction control commands, and obtain the target control scheme for thin film growth in the evaporation cavity.

[0144] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the data analysis-based vapor deposition process monitoring method.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data analysis-based method for monitoring a vapor deposition process, characterized in that, The method includes: Raw data is collected from temperature sensors, pressure sensors, power monitors, and thickness monitoring systems installed inside the vapor deposition chamber. Through filtering and standardization, vapor deposition process parameter data is obtained. Time-frequency analysis and parameter correlation calculations were performed on the vapor deposition process parameter data to extract the variation characteristics of evaporation source temperature, cavity pressure, and deposition rate, obtaining key parameter characteristics of the vapor deposition process. This included: dividing the vapor deposition process parameter data into fixed windows and calculating the statistical moment characteristics and trend characteristics of evaporation source temperature, cavity pressure, and deposition rate within each window to obtain the time-domain characteristics of the parameters; performing a fast Fourier transform on the vapor deposition process parameter data to extract the power spectral density curve and identify the periodic fluctuation components of evaporation source temperature, cavity pressure, and deposition rate to obtain the frequency-domain characteristics of the parameters; and applying wavelet transform or Hilbert transform to the vapor deposition process parameter data to capture the evaporation source temperature... The time-varying frequency characteristics of temperature, cavity pressure, and deposition rate are used to obtain the time-frequency characteristics of the parameters. Sample entropy and multi-scale entropy values ​​are calculated for the time-domain, frequency-domain, and time-frequency characteristics of the parameters to quantify the complexity of parameter fluctuations, thus obtaining the parameter complexity characteristics. Based on the time-domain, frequency-domain, and time-frequency characteristics of the parameters, the Pearson correlation coefficient and mutual information value between the evaporation source temperature, cavity pressure, and deposition rate are calculated to obtain the parameter correlation matrix. Partial correlation analysis and conditional independence tests are applied to the parameter correlation matrix to eliminate the influence of indirect correlations, determine the causal relationship and time delay between parameters, and obtain the key parameter characteristics of the evaporation process. Based on the key parameter characteristics of the vapor deposition process, deviation analysis is performed by setting dynamic thresholds to calculate the degree of deviation between the real-time parameter values ​​and the standard values, thereby obtaining abnormal indicators of the vapor deposition process. Based on the abnormal indicators of the evaporation process, the RF power input and substrate temperature control are adjusted in real time to generate correction control commands, thereby obtaining the target control scheme for thin film growth in the evaporation cavity.

2. The data analysis-based vapor deposition process monitoring method according to claim 1, characterized in that, The raw data collected by the temperature sensor, pressure sensor, power monitor, and thickness monitoring system installed in the evaporation chamber are filtered and standardized to obtain evaporation process parameter data, including: Temperature data is collected by temperature sensors installed at the evaporation source, substrate and cavity wall. The raw temperature distribution data is obtained by combining local regression model with sliding window analysis. Vacuum data is collected from pressure sensors installed at different locations in the cavity. The raw pressure gradient data is obtained by multi-point averaging and outlier removal. The power input data recorded in real time by the power monitor is sampled at a fixed frequency and marked with abrupt change points to obtain the raw power fluctuation data. Wavelet transform filtering technology is applied to the original temperature distribution data, original pressure gradient data, and original power fluctuation data to retain signal characteristics while removing high-frequency noise, resulting in filtered parameter data. The filtered parameter data is subjected to zero-mean normalization to convert parameters of different dimensions into a standard distribution format, thereby obtaining standardized parameter data. By aligning sensor data with different sampling frequencies in the standardized parameter data using timestamps, a unified time reference is established, and the first derivative and moving average features are extracted to obtain the vapor deposition process parameter data.

3. The data analysis-based vapor deposition process monitoring method according to claim 1, characterized in that, Based on the key parameter characteristics of the vapor deposition process, deviation analysis is performed by setting dynamic thresholds to calculate the degree of deviation between the real-time parameter values ​​and the standard values, thereby obtaining abnormal indicators of the vapor deposition process, including: Cluster analysis was performed on the key parameter characteristics of the vapor deposition process, and the process operation mode was divided using a Gaussian mixture model to obtain a standard mode library for the vapor deposition process. For each process mode in the standard mode library of the vapor deposition process, the characteristic multidimensional distribution parameters, including the mean vector and covariance matrix, are calculated to obtain the mode statistical characterization; For the key parameters of the vapor deposition process collected in real time, the Mahalanobis distance and Hotling statistic compared with the statistical characterization of the model are calculated to obtain the parameter deviation. Based on the parameter deviation, a deep autoencoder network is constructed to encode and decode the input features, calculate the reconstruction error, and obtain the abnormal reconstruction loss. For the aforementioned abnormal reconstruction loss, an adaptive threshold is set based on the distribution characteristics of historical data to distinguish between normal fluctuations and abnormal deviations, thereby obtaining an abnormality level classification; Based on the aforementioned anomaly classification, the anomaly start time, duration, severity score, and contribution of related parameters are calculated to obtain the anomaly index of the vapor deposition process.

4. The data analysis-based vapor deposition process monitoring method according to claim 3, characterized in that, The process of constructing a deep autoencoder network based on the parameter deviation, encoding and decoding the input features, calculating the reconstruction error, and obtaining the anomaly reconstruction loss includes: The parameter deviation and key parameter features of the vapor deposition process are concatenated to form an input vector, which is then fed into the input layer of a deep autoencoder network to obtain the features to be processed. The features to be processed are compressed using a four-layer convolutional encoder. Each convolutional kernel in the encoder is 3×3 in size, and the number of channels in the convolutional layers is 32, 64, 128, and 256 respectively. An activation function and pooling operation are applied after each convolutional layer to obtain compressed features. The compressed features are further dimensionality reduced by a fully connected layer, with the number of nodes being 512, 256, 128, and 64 respectively, to obtain the latent space feature representation; The latent space feature representation is restored by a four-layer transposed convolutional decoder. The transposed convolutional decoder adopts a structure symmetrical to the encoder, with the number of channels being 256, 128, 64, and 32 respectively, to obtain the reconstructed features. Calculate the Euclidean distance and cosine similarity between the reconstructed features and the original input features, construct a weighted loss function, and obtain the feature reconstruction loss. The feature reconstruction loss is compared and analyzed with the reconstruction loss of time-series continuous samples to identify abnormal mutation points and gradual trends, and the abnormal reconstruction loss is obtained.

5. The data analysis-based vapor deposition process monitoring method according to claim 3, characterized in that, The abnormal reconstruction loss is determined by setting an adaptive threshold based on the distribution characteristics of historical data to distinguish between normal fluctuations and abnormal deviations, resulting in an anomaly level classification, including: The probability distribution of historically accumulated anomaly reconstruction loss data is fitted, and a nonparametric probability density function is constructed using the kernel density estimation method to obtain the loss distribution model; Based on the loss distribution model, the reconstruction loss values ​​of multiple quantiles are calculated and set as different levels of judgment criteria to obtain a set of static thresholds; For the static threshold set, combined with the parameter distribution of the evaporation process in the start-up, stabilization and end stages, a piecewise linear transformation is applied to obtain the stage adaptive threshold. For the aforementioned stage adaptive threshold, a sliding window averaging and exponential smoothing method is applied to eliminate the impact of short-term fluctuations and obtain a smoothed adaptive threshold. The real-time anomaly reconstruction loss is compared with the smooth adaptive threshold, and the anomaly level is obtained by classifying the degree to which it exceeds the threshold of different level quantiles. The initial anomaly level is processed in a time series, and the anomaly level is updated according to the anomaly duration and the accumulated anomaly value to obtain the anomaly level classification.

6. The data analysis-based method for monitoring the vapor deposition process according to claim 1, characterized in that, The step involves adjusting the RF power input and substrate temperature control in real time based on the abnormal indicators of the evaporation process, generating correction control commands, and obtaining a target control scheme for thin film growth within the evaporation cavity, including: By integrating the abnormal indicators of the vapor deposition process with the historical operation data of the vapor deposition process, a digital twin model of the vapor deposition process is constructed to simulate the influence of different control parameters on thin film growth and obtain the influence mapping of control parameters. Based on the control parameter influence mapping, a control strategy network is trained using a strategy gradient algorithm. Abnormal indicators of the evaporation process are taken as inputs, and the adjustment amounts of RF power and substrate temperature are output to obtain the parameter control strategy. Physical constraints are checked on the parameter control strategy to limit the range and rate of change of RF power and substrate temperature, preventing system instability caused by drastic adjustments, thus obtaining a safe control strategy. Based on the aforementioned safety control strategy, large-scale parameter adjustments are decomposed into multiple small-step sequences to form a smooth-transition control trajectory, resulting in a progressive control sequence. The progressive control sequence is converted into instruction signals in the standard format of the industrial control interface, and written into the programmable logic controller of the vapor deposition equipment through the communication interface to obtain the execution control instructions; The execution results of the control commands are monitored in real time, and the changes in abnormal indicators before and after parameter adjustment are recorded to form control effect evaluation data, thereby obtaining the target control scheme for thin film growth in the vapor deposition chamber.

7. A data analysis-based evaporation process monitoring system, characterized in that, For implementing the data analysis-based vapor deposition process monitoring method as described in any one of claims 1 to 6, the data analysis-based vapor deposition process monitoring system comprises: The data acquisition module is used to collect raw data from the temperature sensor, pressure sensor, power monitor and thickness monitoring system installed in the vapor deposition chamber, and obtain vapor deposition process parameter data through filtering and standardization. The calculation module is used to perform time-frequency analysis and parameter correlation calculation on the vapor deposition process parameter data, extract the variation characteristics of evaporation source temperature, cavity pressure and deposition rate, and obtain the key parameter characteristics of the vapor deposition process. The setting module is used to perform deviation analysis by setting dynamic thresholds based on the key parameter characteristics of the vapor deposition process, calculate the degree of deviation between the real-time value of the parameter and the standard value, and obtain the abnormal index of the vapor deposition process. The adjustment module is used to adjust the RF power input and substrate temperature control in real time according to the abnormal indicators of the evaporation process, generate correction control commands, and obtain the target control scheme for thin film growth in the evaporation cavity.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the data analysis-based vapor deposition process monitoring method as described in any one of claims 1 to 6.

Images