An MPCVD production line intelligent control method, device, medium and product

Abstract

An MPCVD production line intelligent control method, device, medium and product are provided, and relate to the field of diamond growth. In the method, an excitation response test is performed on an MPCVD device to obtain a device state drift matrix of the device hardware state; multi-modal data in a reaction cavity of the MPCVD device is collected; feature extraction and fusion are performed on the multi-modal data to construct a comprehensive feature vector; a target state vector corresponding to a current growth stage is determined; the comprehensive feature vector and the target state vector are compared to generate a difference feature vector; the comprehensive feature vector and the difference feature vector are input into a preset crystal growth process model for reasoning to obtain a preliminary adjustment instruction; a feedforward compensation operation is performed on the preliminary adjustment instruction based on the device state drift matrix to obtain a final control instruction, and the final control instruction is issued to each actuator of the MPCVD device. The technical solution provided in the application improves the control accuracy of the diamond growth process.

Description

Technical Field

[0001] This application relates to the field of diamond growth, specifically to an intelligent control method, device, medium, and product for an MPCVD production line. Background Technology

[0002] Microwave plasma chemical vapor deposition (MPCVD) is currently the mainstream process for preparing high-quality, large-size single-crystal diamond and other advanced semiconductor materials. Traditional MPCVD production line control relies on human experience and classic algorithms such as PID control. However, when faced with the complex nonlinear and high-hysteresis characteristics of the diamond growth process, the response is slow and optimal control is difficult to achieve. This approach typically intervenes only after the process has deviated significantly and caused crystal defects, resulting in inherent drawbacks such as poor production consistency and low yield.

[0003] To address this issue, existing technologies collect data using multimodal sensors, utilize AI models for real-time analysis and control, aggregate the data to the cloud for continuous model training and optimization, and then distribute the optimized model to the edge, forming a closed-loop intelligent system that improves the real-time performance and accuracy of control.

[0004] However, the above method implicitly assumes a core principle: the hardware state of MPCVD equipment remains constant. But in long-term production, the core components of the equipment inevitably experience performance degradation, physical wear, and other signs of aging. This gradual drift in equipment state alters the mapping relationship between process inputs and the actual growth state. Existing AI models cannot distinguish between normal process fluctuations and systematic shifts caused by equipment aging, leading to a gradual decline in model prediction accuracy and control effectiveness as equipment usage time increases, resulting in reduced accuracy in parameter control during diamond growth. Summary of the Invention

[0005] This application provides an intelligent control method, device, medium, and product for MPCVD production lines, which improves the control accuracy of the diamond growth process.

[0006] The first aspect of this application provides an intelligent control method for an MPCVD production line. The method includes: performing an excitation response test on the MPCVD equipment in the MPCVD production line to obtain a device state drift matrix of the equipment hardware state; collecting multimodal data within the reaction chamber of the MPCVD equipment during diamond growth; extracting and fusing features from the multimodal data to construct a comprehensive feature vector; obtaining the current growth stage of the diamond and determining the target state vector corresponding to the current growth stage; comparing the comprehensive feature vector and the target state vector to generate a difference feature vector; inputting the comprehensive feature vector and the difference feature vector into a preset crystal growth process model for inference to obtain a preliminary adjustment command; performing feedforward compensation calculation on the preliminary adjustment command based on the device state drift matrix to obtain a final control command, and issuing the final control command to each actuator of the MPCVD equipment.

[0007] By employing the above technical solutions, the equipment state drift matrix is ​​obtained through excitation response testing, enabling real-time monitoring of the evolution of the MPCVD equipment hardware state and providing accurate equipment characteristic data for subsequent compensation calculations. Collecting multimodal data within the reaction chamber during diamond growth comprehensively reflects key process parameters such as plasma state, temperature field, and pressure field. Feature extraction and fusion of the multimodal data to construct a comprehensive feature vector effectively integrates information from different physical quantities, forming a complete description of the current process state. Obtaining the current growth stage and determining the target state vector provides a clear optimization objective for process adjustment. Comparing the comprehensive feature vector and the target state vector to generate a difference feature vector accurately quantifies the degree and direction of deviation between the current state and the ideal state. Inputting the comprehensive feature vector and the difference feature vector into the crystal growth process model for inference yields preliminary adjustment commands, realizing intelligent decision-making based on the current state and deviation information. Feedforward compensation calculations are performed on the preliminary adjustment commands based on the equipment state drift matrix to obtain the final control commands, effectively offsetting the impact of factors such as equipment hardware aging and changes in response characteristics on control accuracy. This ensures that the actual output of each actuator is consistent with the expected value, improving the process stability and product quality consistency of the diamond growth process.

[0008] Optionally, the step of performing an excitation response test on the MPCVD equipment in the MPCVD production line to obtain the equipment state drift matrix of the equipment hardware state specifically includes: when the MPCVD equipment is in a preset reference state, executing an excitation response test program to obtain a reference transfer function; inputting a preset standard excitation signal sequence into the MPCVD equipment; collecting microwave system reflection coefficient and plasma emission spectrum data triggered by the preset standard excitation signal sequence as response output; performing Laplace transforms on the preset standard excitation signal sequence and the response output respectively to calculate the transfer function characterizing the current state of the equipment; and performing matrix difference operation between the transfer function and the reference transfer function to generate the equipment state drift matrix.

[0009] By employing the above technical solution, a reference transfer function is obtained by executing an excitation response test program under a preset reference state on the MPCVD equipment. This establishes a reference standard for the response characteristics of the equipment under ideal conditions, providing a benchmark for subsequent state assessment. A preset standard excitation signal sequence is input into the MPCVD equipment to ensure the standardization and repeatability of the test conditions. Microwave system reflection coefficient and plasma emission spectrum data are collected as response outputs. Key physical quantities that can sensitively reflect changes in equipment state are selected, where the microwave system reflection coefficient reflects the impedance matching state of the microwave transmission system, and the plasma emission spectrum data reflects the plasma excitation efficiency and energy transfer characteristics. The transfer function is calculated by performing Laplace transform on the preset standard excitation signal sequence and the response output, realizing the conversion from time-domain signals to frequency-domain system characteristics, and obtaining a complete mathematical model characterizing the current dynamic response characteristics of the equipment. A matrix difference operation is performed between the current transfer function and the reference transfer function to generate an equipment state drift matrix, which intuitively quantifies the deviation of the equipment response characteristics from the reference state. This drift matrix contains information such as the response delay changes, gain attenuation, and nonlinear characteristic evolution of each actuator, providing a precise basis for equipment state correction for feedforward compensation calculations.

[0010] Optionally, the step of extracting and fusing features from the multimodal data to construct a comprehensive feature vector specifically includes: extracting features from the plasma emission spectrum data, plasma sphere image data, microwave power data, and reaction chamber temperature and pressure data in the multimodal data to obtain spectral feature sub-vectors, image feature sub-vectors, and temporal statistical feature sub-vectors; calculating the correlation weight coefficients between the spectral feature sub-vectors, the image feature sub-vectors, and the temporal statistical feature sub-vectors and the current crystal growth stage based on an attention mechanism; and weighting and concatenating the spectral feature sub-vectors, the image feature sub-vectors, and the temporal statistical feature sub-vectors according to the correlation weight coefficients, and then normalizing them to obtain the comprehensive feature vector.

[0011] By employing the aforementioned technical solutions, feature extraction is performed on plasma emission spectrum data, plasma sphere image data, microwave power data, and reaction chamber temperature and pressure data from multimodal data, respectively, yielding spectral feature vectors, image feature vectors, and temporal statistical feature vectors. This achieves structured representation of information from different physical domains. The spectral feature vectors reflect the chemical composition and excitation state of the plasma, the image feature vectors reflect the spatial distribution of the plasma, and the temporal statistical feature vectors reflect the dynamic evolution of process parameters. Based on an attention mechanism, the correlation weight coefficients between each feature vector and the current crystal growth stage are calculated, enabling adaptive evaluation of feature importance. This allows for highlighting the most influential feature information at each growth stage; for example, more attention is paid to plasma sphere image features during the nucleation stage, while more attention is paid to spectral features and temporal statistical features such as temperature and pressure during the stable growth stage. The feature vectors are weighted, concatenated, and normalized according to the correlation weight coefficients to obtain a comprehensive feature vector. This effectively integrates multi-source heterogeneous information, avoiding information redundancy and the curse of dimensionality caused by simple concatenation. Simultaneously, the normalization process eliminates dimensional differences between different feature vectors, improving the numerical stability of the comprehensive feature vector and the accuracy of subsequent model inference.

[0012] Optionally, the step of inputting the comprehensive feature vector and the difference feature vector into a preset crystal growth process model for inference to obtain preliminary adjustment instructions specifically includes: stacking the comprehensive feature vector with multiple historical comprehensive feature vectors within a preset historical time window to construct a state tensor with time depth, and decomposing the difference feature vector to obtain amplitude and direction components; inputting the state tensor into a time-series encoding module embedded in the preset crystal growth process model, extracting time-dependent features through a bidirectional long short-term memory network to generate a time-series encoded vector; concatenating the time-series encoded vector with the direction components to obtain a joint inference input vector, and inputting the joint inference input vector into the strategy decision module of the preset crystal growth process model to output candidate adjustment actions and the probability distribution corresponding to the candidate adjustment actions; generating a sequence of adjustment actions to be executed based on the candidate adjustment actions and the probability distribution; performing feasibility constraint verification on the sequence of adjustment actions to be executed, the feasibility constraint verification including verification of the safe working range of the actuator and verification of the equipment response capability limitation; encoding the adjustment actions that pass the feasibility constraint verification to generate the preliminary adjustment instructions.

[0013] By employing the aforementioned technical solution, a state tensor is constructed by stacking the comprehensive feature vector with multiple historical comprehensive feature vectors within a preset historical time window. This introduces temporal context information, enabling the model to capture the dynamic evolution trend and temporal dependencies of the diamond growth process. The differential feature vector is decomposed into amplitude and direction components, achieving structured analysis of deviation information. The amplitude component represents the degree of deviation, while the direction component represents the specific dimension and trend of the deviation. The state tensor is input into the temporal encoding module, which extracts temporal dependency features through a bidirectional long short-term memory network to generate a temporal encoding vector. This effectively extracts long-term dependencies and historical evolution patterns in the diamond growth process, enhancing the model's understanding of the development trend of the process state. The temporal encoding vector is concatenated with the direction component to obtain a joint inference input vector, integrating current state, historical evolution, and deviation direction information, providing a more comprehensive input basis for strategy decision-making. The joint inference input vector is input into the strategy decision-making module, which outputs candidate adjustment actions and corresponding probability distributions, realizing the generation of multiple schemes and the quantification of uncertainty. Based on candidate adjustment actions and probability distribution, a sequence of adjustment actions to be executed is generated, and feasibility constraint verification is performed to ensure the executability and safety of adjustment instructions, and to avoid invalid instructions that exceed the working range of the actuator or the response capability of the equipment.

[0014] Optionally, the step of performing feedforward compensation calculation on the preliminary adjustment command based on the device state drift matrix to obtain the final control command specifically includes: parsing the preliminary adjustment command to generate an original adjustment vector containing the original adjustment parameters of each actuator; performing a Laplace transform on the original adjustment vector to obtain a frequency domain adjustment vector, and performing matrix multiplication on the frequency domain adjustment vector and the inverse matrix of the device state drift matrix to generate a frequency domain compensated adjustment vector; performing an inverse Laplace transform on the frequency domain compensated adjustment vector to generate a time domain compensated adjustment vector; calculating the timing advance of each adjustment command according to the device state drift matrix, and associating and binding each component value in the time domain compensated adjustment vector with the corresponding timing advance to generate the final control command.

[0015] By employing the above technical solution, the initial adjustment command is parsed to generate an initial adjustment vector containing the original adjustment parameters of each actuator, thus realizing a structured representation of the command information. A Laplace transform is performed on the initial adjustment vector to obtain a frequency domain adjustment vector, converting the adjustment command from the time domain to the frequency domain, providing a mathematical basis for subsequent frequency domain compensation calculations. A matrix multiplication operation is performed between the frequency domain adjustment vector and the inverse matrix of the equipment state drift matrix to generate a frequency domain compensated adjustment vector. This achieves precise compensation for equipment state drift in the frequency domain. This inverse matrix operation mathematically cancels the influence of changes in the equipment transfer function on the control effect, ensuring that even if the equipment response characteristics drift, the compensated command can still enable each actuator to produce the desired actual output. An inverse Laplace transform is performed on the frequency domain compensated adjustment vector to generate a time domain compensated adjustment vector. The compensation result is then converted back to the time domain, yielding time domain adjustment parameters suitable for practical control. The timing advance of each adjustment command is calculated based on the equipment state drift matrix, and the component values ​​in the time-domain compensated adjustment vector are associated and bound with the corresponding timing advance to generate the final control command. Timing coordination is performed to address the differences in response speed of different actuators, ensuring that the actual response of each actuator can reach the desired state synchronously in time, thereby improving the accuracy and real-time performance of multi-actuator collaborative control.

[0016] Optionally, before performing feature extraction and fusion on the multimodal data to construct a comprehensive feature vector, the method further includes: obtaining the original acquisition timestamp and sampling frequency corresponding to each modality in the multimodal data; calculating the statistical feature threshold of each modality within the sliding window based on the sliding window method, identifying and marking abnormal data points whose values ​​exceed the statistical feature threshold; and performing data repair and time-series alignment processing on the abnormal data points and missing data locations to obtain preprocessed multimodal data.

[0017] By employing the above technical solutions, the original acquisition timestamps and sampling frequencies of each modality in the multimodal data were obtained, thus understanding the time base and sampling characteristics of each modality and providing necessary time reference information for subsequent time alignment and anomaly detection. Based on the sliding window method, the statistical feature thresholds of each modality within the sliding window were calculated, establishing a dynamic anomaly judgment standard. This sliding window method can adapt to the normal fluctuation range of process parameters during diamond growth, avoiding misjudgments of normal process changes by the fixed threshold method. Abnormal data points with values ​​exceeding the statistical feature thresholds were identified and marked, accurately locating abnormal data caused by sensor failure, communication interference, or sudden disturbances, preventing abnormal data from contaminating subsequent feature extraction and model inference processes. Data repair and time alignment were performed on abnormal data points and missing data locations to obtain preprocessed multimodal data, improving the integrity, accuracy, and temporal consistency of the multimodal data. This provided high-quality data input for subsequent feature extraction and fusion, effectively avoiding feature distortion and control decision deviations caused by data quality issues, and ensuring the reliability of the entire intelligent control method.

[0018] Optionally, the step of performing data repair and time-series alignment processing on the abnormal data points and missing data locations to obtain preprocessed multimodal data specifically includes: using a cubic spline interpolation algorithm to interpolate and repair the abnormal data points and missing data locations based on the valid data adjacent to the abnormal data points, to obtain the repaired data; selecting the timestamp sequence corresponding to the modality data with the lowest sampling frequency in the multimodal data as a unified time reference, and aligning the other modality data using a nearest neighbor timestamp matching resampling method, to obtain the preprocessed multimodal data.

[0019] By adopting the above technical solution, the repaired data is obtained by interpolating and repairing the valid data adjacent to the abnormal data points using a cubic spline interpolation algorithm. This leverages the continuity and smoothness of local data. The cubic spline interpolation algorithm ensures the continuity of the first and second derivatives of the interpolation curve at the connection points. The generated repaired data closely matches the actual physical process in both numerical value and trend, avoiding the data abrupt changes and trend distortion problems caused by simple linear interpolation. The timestamp sequence corresponding to the modality with the lowest sampling frequency in the multimodal data is selected as a unified time reference. This ensures that all modality data have valid measurements after time alignment, avoiding information distortion caused by extensive interpolation of low-sampling-rate data when using a high sampling rate as a reference. The preprocessed multimodal data is obtained by resampling other modal data using the nearest neighbor timestamp matching method, which realizes the time synchronization of data with different sampling frequencies. This nearest neighbor matching method retains the original measurement data rather than the interpolated data, thus maximizing the authenticity of the data and making the preprocessed multimodal data consistent in the time dimension. This provides a high-quality data source with a unified time reference for subsequent multimodal feature extraction and fusion.

[0020] Secondly, embodiments of this application provide an intelligent control device for an MPCVD production line, the intelligent control device for an MPCVD production line comprising: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the intelligent control device for an MPCVD production line to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on an MPCVD production line intelligent control device, cause the MPCVD production line intelligent control device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on an intelligent control device for an MPCVD production line, cause the intelligent control device for the MPCVD production line to execute the method described in the first aspect and any possible implementation thereof.

[0023] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: 1. By introducing a device state drift matrix, the dynamic changes in the hardware state of MPCVD equipment are quantified. Combined with multimodal data acquisition and comprehensive feature vector construction, and utilizing a crystal growth process model based on differential feature vectors for inference, high-precision intelligent control of the diamond growth process is achieved. By incorporating device state drift compensation calculations into instruction generation, this scheme effectively addresses control deviations caused by equipment performance degradation, ensuring more precise adjustment of process parameters throughout the diamond growth process.

[0024] 2. By introducing sliding window anomaly detection, cubic spline interpolation repair, and time-series alignment techniques in the multimodal data processing stage, possible anomalies and missing data in the collected data are repaired, effectively reducing the impact of noise and bias on model accuracy, enhancing the temporal consistency of the comprehensive feature vector, and laying a data foundation for real-time process monitoring and accurate inference.

[0025] 3. By calculating the state difference of the equipment transfer function and performing feedforward compensation calculations, this scheme adjusts and optimizes the equipment state drift after the initial adjustment instructions are generated. At the same time, it performs feasibility constraint verification to ensure stable equipment operation, avoids damage to the equipment or crystal caused by control errors, and improves the stability and long-term consistency of the entire MPCVD production line. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating an intelligent control method for an MPCVD production line disclosed in an embodiment of this application; Figure 2 This is another schematic diagram of a method for intelligent control of an MPCVD production line disclosed in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an intelligent control device for an MPCVD production line provided in an embodiment of this application.

[0027] Explanation of reference numerals in the attached drawings: 301, Central Processing Unit; 302, Read-Only Memory; 303, Random Access Memory; 304, Bus; 305, Input / Output Interface; 306, Input Section; 307, Output Section; 308, Storage Section; 309, Communication Section; 310, Driver; 311, Removable Media. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0029] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0030] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple system devices refer to two or more system devices, and multiple screen terminals refer to two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0031] This application provides an intelligent control method for an MPCVD production line, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an intelligent control method for an MPCVD production line provided in an embodiment of this application. The method is applied to a server, which can execute the MPCVD production line intelligent control program. The server can be a single server, a server cluster consisting of multiple servers, or a cloud computing service center. The method includes steps S101 to S107, as follows: Step S101: Perform stimulus response test on the MPCVD equipment in the MPCVD production line to obtain the equipment state drift matrix of the equipment hardware status.

[0032] In step S101, the MPCVD production line refers to a complete microwave plasma chemical vapor deposition production line used for the mass production of diamonds. This production line includes multiple MPCVD machines and supporting auxiliary systems. MPCVD equipment, specifically microwave plasma chemical vapor deposition equipment, is the core equipment for diamond growth. Excitation response testing refers to the process of analyzing the dynamic characteristics of the MPCVD equipment by applying specific input signals and measuring the equipment's output response. The equipment state drift matrix is ​​a mathematical matrix that quantitatively describes the degree to which the hardware state of the MPCVD equipment deviates from its initial baseline state due to aging over time or environmental changes. This matrix typically represents the change in the equipment's transfer function in the frequency domain.

[0033] Specifically, the server first controls the MPCVD equipment under test to enter a preset baseline state where all parameters are standardized, such as maintaining a stable low-power plasma under specific gas ratios and pressures. In this state, the server executes an excitation response test program, calculating and storing a baseline transfer function characterizing the initial health state of the equipment. Subsequently, the server inputs a pre-designed sequence of standard excitation signals, such as frequency sweep signals or step signals of specific frequencies and amplitudes, to the various actuators of the MPCVD equipment, such as the microwave generator and gas mass flow controller. Simultaneously with the excitation, the server rapidly acquires the system response output triggered by this excitation signal sequence, primarily including data on the change in the microwave system's reflection coefficient and dynamic data on the plasma emission spectrum within the reaction cavity. The server performs Laplace transforms on both the input preset standard excitation signal sequence and the acquired response output data, converting the time-domain signals into a frequency-domain representation. By calculating the ratio of the frequency-domain output to the frequency-domain input, the server obtains the transfer function characterizing the current hardware state of the equipment. Finally, the server performs a matrix difference operation between the current state transfer function and the previously stored baseline transfer function; the result is the equipment state drift matrix.

[0034] In one possible implementation, an excitation response test is performed on the MPCVD equipment in the MPCVD production line to obtain the equipment state drift matrix of the equipment hardware state, specifically including steps S1011-S1015, as follows: Step S1011: When the MPCVD equipment is in a preset reference state, execute the excitation response test program to obtain the reference transfer function.

[0035] In step S1011, the preset reference state refers to a standardized and reproducible physical state in which the MPCVD equipment is located during initial calibration. The hardware performance of the equipment in this state is defined as an ideal reference. The excitation response test program is a software algorithm used to calculate the system transfer function. This program identifies the dynamic characteristics of the system by analyzing the relationship between the input and output signals. The reference transfer function is a mathematical model, usually represented in frequency domain polynomial or matrix form, used to accurately describe the dynamic response characteristics of the MPCVD equipment from input to output in the preset reference state.

[0036] Specifically, the server first issues a series of instructions to the MPCVD equipment, setting it to a preset baseline state. This preset baseline state is typically a stable operating point where the hardware is in optimal health after the equipment's first operation after leaving the factory or after maintenance. For example, under non-growth conditions, the chamber is filled with a specific proportion of argon gas, maintaining a low-power, stable plasma. After the equipment stabilizes, the server runs an excitation-response test program internally. This program generates a virtual ideal input signal and calculates the theoretical response output under this ideal input signal based on the equipment's design model. Subsequently, the program performs a Laplace transform on the virtual ideal input signal and the theoretical response output, and obtains the baseline transfer function by calculating the ratio of the transformed signals. This baseline transfer function is stored by the server as a standard for subsequent evaluation of changes in the equipment's state.

[0037] Step S1012: Input a preset standard excitation signal sequence into the MPCVD equipment.

[0038] In step S1012, the preset standard excitation signal sequence refers to a pre-designed time series signal containing specific waveform, amplitude, and frequency components. This signal is used as a probe and input into the actual MPCVD equipment to excite the dynamic response of the equipment.

[0039] Specifically, the server reads a preset standard excitation signal sequence from the configuration library. This preset standard excitation signal sequence is carefully designed to ensure that all dynamic modes of the device within the frequency band of interest are fully excited within a limited test time. Common signal forms include swept signals, i.e., continuous sinusoidal signals whose frequency varies linearly or logarithmically with time, or pseudo-random binary sequences that have approximately white noise characteristics in the frequency domain. The server converts this digitized preset standard excitation signal sequence into analog voltage or current signals via a digital-to-analog converter and applies them sequentially to the critical actuators of the MPCVD equipment, such as a microwave generator that controls microwave power or a gas mass flow controller that controls gas flow.

[0040] Step S1013: Collect microwave system reflection coefficient and plasma emission spectrum data induced by a preset standard excitation signal sequence as response output.

[0041] In step S1013, the microwave system reflection coefficient is a dimensionless parameter used to represent how much of the power input to the microwave system is reflected back due to impedance mismatch. It is a key indicator for measuring the coupling efficiency between the plasma and the microwave source. Plasma emission spectrum data refers to the wavelength and intensity distribution data of light emitted by excited atoms, ions, or molecules in the plasma when they transition back to lower energy levels, collected by a spectrometer. Response output refers to the complete set of data collected from sensors after a preset standard excitation signal sequence is applied to the MPCVD equipment, reflecting changes in the internal state of the equipment.

[0042] Specifically, during the same time period when the server inputs a preset standard excitation signal sequence to the MPCVD equipment, the server synchronously initiates a data acquisition task at a high sampling frequency. The server sends continuous query commands to the directional coupler and power meter coupled to the microwave system to acquire real-time readings of the forward and reflected power, and calculates the time-series data of the microwave system's reflection coefficient. Simultaneously, the server instructs a spectrometer installed outside the observation window of the reaction cavity to perform continuous exposure acquisition, obtaining the time-series of plasma emission spectral data. The server adds a high-precision timestamp to each frame of spectral data and each reflection coefficient value acquired to ensure that the response output data is precisely aligned in time with the input preset standard excitation signal sequence. The acquired microwave system reflection coefficient and plasma emission spectral data together constitute the response output.

[0043] Step S1014: Perform Laplace transforms on the preset standard excitation signal sequence and the response output respectively, and calculate the transfer function that characterizes the current state of the device.

[0044] In step S1014, the Laplace transform is an integral transform mathematical method used to convert a function or signal in the time domain into a function in the complex frequency domain, facilitating the analysis of the characteristics of linear time-invariant systems. Here, the transfer function refers to the mathematical model characterizing the current hardware state of the MPCVD device.

[0045] Specifically, the server first retrieves the preset standard excitation signal sequence used as input in step S1012 and the response output used as output in step S1013. The server applies a Laplace transform algorithm to the preset standard excitation signal sequence, a time-domain function, converting it into an input function in the frequency domain. Next, the server performs Laplace transforms on the microwave system reflection coefficient time-series data and plasma emission spectrum time-series data in the response output, converting them into output functions in the frequency domain. For multiple-input multiple-output (MIMO) systems, these input and output functions are typically represented as vectors. The server calculates the ratio of the frequency-domain output function to the frequency-domain input function by solving a system of linear equations; this ratio is the transfer function characterizing the current state of the device. This transfer function reflects the dynamic characteristics of the actual physical link from the excitation signal to the system response.

[0046] Step S1015: Perform matrix difference operation between the transfer function and the reference transfer function to generate the device state drift matrix.

[0047] In step S1015, matrix difference operation refers to the mathematical operation of subtracting corresponding elements between two or more matrices. The device state drift matrix is ​​a difference matrix, and each element of this matrix quantifies the deviation of a certain parameter of the current device transfer function from the corresponding parameter of the reference transfer function.

[0048] Specifically, the server retrieves the reference transfer function obtained in step S1011 from storage and represents it in matrix form. Simultaneously, the server also represents the transfer function representing the current state of the device, calculated in step S1014, in matrix form. The server performs a matrix difference operation, subtracting the reference transfer function matrix from the transfer function matrix representing the current state of the device. The resulting difference matrix is ​​the device state drift matrix. This device state drift matrix visually reveals the changes in the device hardware performance over time, such as gain attenuation, phase lag, or resonant frequency shift. The server stores the generated device state drift matrix for subsequent feedforward compensation of control commands.

[0049] Step S102: During the diamond growth process, collect multimodal data from the reaction chamber of the MPCVD equipment in the MPCVD production line.

[0050] In step S102, the diamond growth process refers to the complete process flow in which carbon source gas is decomposed and deposited to form diamond crystals within the reaction chamber of an MPCVD equipment using a chemical vapor deposition method. The reaction chamber refers to the core enclosed space within the MPCVD equipment where plasma discharge and diamond deposition occur. Multimodal data refers to a collection of multiple data streams from different sources or of different types, used to jointly describe the physical state of the diamond growth process from different dimensions. This data includes plasma emission spectrum data, plasma sphere image data, forward and reflected power data of the microwave system, and temperature and pressure data within the reaction chamber.

[0051] Specifically, after the diamond growth process begins, the server continuously communicates with various sensors deployed on the MPCVD production line to acquire multimodal data within the reaction chamber in real time. The server requests and receives plasma emission spectral data from the spectrometer to analyze the chemical composition and particle energy states within the plasma. The server acquires plasma sphere image data from an industrial camera mounted on the observation window to monitor the plasma's morphology, size, and location. Simultaneously, the server obtains forward and reflected power data from the microwave system's power meter and real-time temperature and pressure readings from temperature and pressure sensors within the reaction chamber. The server appends a precise timestamp to each acquired data point to ensure temporal synchronization and correlation of the modal data during subsequent data processing, thereby constructing a complete multimodal dataset describing the diamond growth process.

[0052] Step S103: Extract and fuse features from multimodal data to construct a comprehensive feature vector.

[0053] In step S103, the comprehensive feature vector refers to a numerical vector that has undergone dimensionality reduction and fusion processing, which can comprehensively and compactly represent the current growth state within the reaction chamber. This vector is the core basis for subsequent control decisions.

[0054] Specifically, the server first performs parallel feature extraction on the collected multimodal data. For plasma emission spectrum data, the server extracts spectral feature vectors by analyzing the intensity, full width at half maximum (FWHM), and intensity ratio of specific spectral lines, such as the Balmer series of hydrogen atoms or carbon-related spectral lines. For plasma sphere image data, the server uses a pre-trained convolutional neural network to extract deep visual features such as size, brightness, and texture, forming image feature vectors. For time-series data such as microwave power data and reaction chamber temperature and pressure, the server calculates statistical measures such as mean, variance, and slope within a sliding time window, forming time-series statistical feature vectors. Next, based on an attention mechanism algorithm, the server uses the current crystal growth stage information as a query condition and calculates the correlation weight coefficients between the spectral feature vector, image feature vector, and time-series statistical feature vector and the current growth stage. These correlation weight coefficients reflect the importance of the features of different modal data for determining the growth state at a specific growth stage. Based on the calculated correlation weight coefficients, the server weights and concatenates the spectral feature vector, image feature vector, and time-series statistical feature vector to form a high-dimensional fusion vector. Finally, the server normalizes the fused vector to eliminate the dimensional differences between different features, ultimately obtaining a comprehensive feature vector.

[0055] In one possible implementation, feature extraction and fusion are performed on the multimodal data to construct a comprehensive feature vector, specifically including steps S1031-S1033, as follows: Step S1031: Extract features from the plasma emission spectrum data, plasma ball image data, microwave power data, and reaction chamber temperature and pressure data in the multimodal data to obtain spectral feature sub-vectors, image feature sub-vectors, and time-series statistical feature sub-vectors.

[0056] In step S1031, multimodal data refers to a set of data collected from the MPCVD equipment that, while having different sources and structures, describe the same growth process. This includes plasma emission spectrum data, plasma sphere image data, microwave power data, and reaction chamber temperature and pressure data. Feature extraction refers to the process of converting the original, high-dimensional data into more representative and easily analyzable low-dimensional feature representations through a series of algorithms and transformations. Spectral feature sub-vectors, image feature sub-vectors, and time-series statistical feature sub-vectors are numerical vectors extracted from plasma emission spectrum data, plasma sphere image data, and microwave power and temperature / pressure time-series data, respectively, capable of quantifying the core information of their respective modes.

[0057] Specifically, the server first receives multimodal data from different sensors. For plasma emission spectrum data, the server identifies key spectral lines, such as the bands of hydrogen atoms (Hα, Hβ), carbon dimers (C2), and cyano groups (CN), calculates the peak intensity, integral area, full width at half maximum (FWHM), and intensity ratios between key spectral lines, and combines these calculated values ​​into a spectral feature sub-vector. For plasma sphere image data, the server uses a pre-trained convolutional neural network, such as ResNet or VGG, to process the plasma sphere images, extract deep semantic features, and uses the output of a specific layer in the network as an image feature sub-vector. For microwave power data and reaction chamber temperature and pressure data, the server analyzes these time-series data within a fixed time window, calculates statistics such as mean, standard deviation, maximum, minimum, skewness, and kurtosis within that time window, and concatenates all the statistics of the time-series data to form a time-series statistical feature sub-vector.

[0058] Step S1032: Based on the attention mechanism, calculate the correlation weight coefficients between the spectral feature vector, the image feature vector, the time-series statistical feature vector, and the current crystal growth stage.

[0059] In step S1032, the attention mechanism is a neural network technique that simulates human cognitive attention. This technique dynamically assigns different importance weights to different parts of the input based on the context. The current crystal growth stage refers to the specific period in the diamond growth process, such as the nucleation stage, rapid growth stage, or late adjustment stage. This stage information serves as context to guide the allocation of attention. The association weight coefficient is a value calculated by the attention mechanism, which quantifies the relative importance of the feature vector of each mode to describing the system state at the current crystal growth stage.

[0060] Specifically, the server takes the spectral feature vector, image feature vector, and temporal statistical feature vector obtained in step S1031 as input. The server also acquires information about the current crystal growth stage, which can be output by an independent stage discrimination model or pre-set according to the time nodes of the process formulation. The server encodes the current crystal growth stage information into a query vector. In a pre-trained attention network, the server calculates the similarity between this query vector and the spectral feature vector, image feature vector, and temporal statistical feature vector (which serve as keys), for example, through dot product operations. The calculated similarity score reflects the degree of association between the features of each modality and the stage target at the current growth stage. The server then normalizes these raw scores using a Softmax function to generate a set of association weight coefficients summing to 1, corresponding to the spectral feature vector, image feature vector, and temporal statistical feature vector, respectively.

[0061] Step S1033: The spectral feature vector, image feature vector, and temporal statistical feature vector are weighted and concatenated according to the correlation weight coefficient, and then normalized to obtain the comprehensive feature vector.

[0062] In step S1033, weighted concatenation is a feature fusion strategy. This strategy first multiplies each feature sub-vector by its corresponding weight coefficient to adjust its contribution, and then concatenates the adjusted feature sub-vectors into a longer vector. Normalization is a data preprocessing technique used to adjust the values ​​in the vector to a uniform scale range to eliminate the influence of different dimensions between features. The comprehensive feature vector is generated after weighted concatenation and normalization, and is ultimately used as a single vector to characterize the comprehensive state of the MPCVD system.

[0063] Specifically, the server obtains the spectral feature sub-vector, image feature sub-vector, and temporal statistical feature sub-vector generated in step S1031, as well as the correlation weight coefficients calculated in step S1032 that correspond one-to-one with these three sub-vectors. The server performs a weighting operation, multiplying each element of the spectral feature sub-vector by the corresponding correlation weight coefficient, and similarly performs the same operation on the image feature sub-vector and the temporal statistical feature sub-vector. Then, the server concatenates the weighted spectral feature sub-vector, the weighted image feature sub-vector, and the weighted temporal statistical feature sub-vector in a predetermined order to form a new vector with a higher dimension. Finally, the server normalizes this new vector, for example, using a min-max normalization method to linearly scale all values ​​in the vector to the range of 0 to 1, ultimately obtaining a comprehensive feature vector. This comprehensive feature vector will be used in subsequent growth quality prediction or process control models.

[0064] Step S104: Obtain the current growth stage of the diamond and determine the target state vector corresponding to the current growth stage.

[0065] In step S104, the current growth stage refers to a specific time period divided according to the diamond growth process formula, such as the nucleation stage, the thickening stage, or the lateral growth stage. Each stage has different process objectives and control requirements. The target state vector refers to the ideal comprehensive feature vector that is pre-set or dynamically optimized to achieve the optimal growth rate and crystal quality under a specific growth stage.

[0066] Specifically, the server determines the current growth stage of the diamond growth process based on the elapsed time of the diamond growth process or by analyzing the changing trends of the comprehensive feature vector itself. For example, by identifying features related to crystal nucleus density in the comprehensive feature vector, the server determines that the growth process has completed the nucleation stage and entered the thickening stage. After determining the current growth stage, the server accesses a database storing standard crystal growth process formulas. This database predefines an optimal target state vector for each growth stage. This target state vector is calibrated based on a large amount of experimental data and expert knowledge, representing the most ideal plasma state and growth environment at that stage. The server retrieves and reads the target state vector corresponding to the current growth stage from the database.

[0067] Step S105: Compare the integrated feature vector and the target state vector to generate a difference feature vector.

[0068] In step S105, the difference feature vector is a numerical vector generated by comparing the current actual state with the target ideal state, which can indicate the direction and magnitude of the deviation of the current state. This vector directly drives the subsequent adjustment action.

[0069] Specifically, after obtaining the comprehensive feature vector representing the current actual growth state and the target state vector representing the ideal growth state, the server performs vector subtraction. The server subtracts the target state vector from the comprehensive feature vector, and the result is the difference feature vector. Each dimension of this difference feature vector corresponds to the deviation between the actual state and the target state on a specific feature. The direction of the difference feature vector indicates the nature of the state deviation, such as whether the temperature is too high or the carbon concentration is too low; the magnitude of the difference feature vector quantifies the severity of the deviation.

[0070] In one possible implementation, the comprehensive feature vector and the difference feature vector are input into a preset crystal growth process model for inference to obtain preliminary adjustment instructions, specifically including steps S1051-S1056, as follows: Step S1051: Stack the comprehensive feature vector with multiple historical comprehensive feature vectors within the preset historical time window to construct a state tensor with time depth, and decompose the difference feature vector to obtain the amplitude component and the direction component.

[0071] In step S1051, the preset historical time series window refers to a fixed-length time interval that includes several consecutive sampling time points prior to the current moment. The state tensor is a three-dimensional numerical array composed of the current comprehensive feature vector and multiple historical comprehensive feature vectors within the preset historical time series window, stacked along the time dimension to capture the dynamic evolution of the system state. Decomposition refers to the process of breaking down a vector into its magnitude and direction. The magnitude component is a scalar value representing the length or magnitude of the difference feature vector, quantifying the overall degree of deviation of the system state from the target. The direction component is a unit vector representing the orientation of the difference feature vector in the multidimensional feature space, revealing the specific nature of the system state deviation.

[0072] Specifically, the server first obtains the composite feature vector at the current moment. Simultaneously, the server retrieves 10 consecutive historical composite feature vectors generated within a preset historical time window, such as the past 10 sampling periods, from internal storage. The server stacks these 10 historical composite feature vectors with the current composite feature vector in chronological order, forming a state tensor with 11 time steps. At the same time, the server obtains the input difference feature vector. The server calculates the Euclidean norm of this difference feature vector, and the result is the amplitude component. Subsequently, the server divides each element of the difference feature vector by this amplitude component, thus obtaining a unit vector with a magnitude of 1, which is the direction component.

[0073] Step S1052: Input the state tensor into the timing coding module embedded in the preset crystal growth process model, extract the timing-dependent features through the bidirectional long short-term memory network, and generate the timing coding vector.

[0074] In step S1052, the preset crystal growth process model is a deep learning model pre-trained with a large amount of historical growth data. This model can understand the complex relationship between process state and final crystal quality and make decisions accordingly. The timing encoding module is a component within the preset crystal growth process model, specifically designed to process and understand the dynamic characteristics of sequential data. The bidirectional long short-term memory network is a special type of recurrent neural network that, through forward and backward processing paths, can simultaneously capture past and future contextual information in the sequential data. Temporal dependency features refer to the correlations and patterns existing between data at different time points in time series data. The timing encoding vector is a fixed-length numerical vector that is a generalized representation obtained by the timing encoding module after processing the state tensor, condensing the evolution of the system state over time.

[0075] Specifically, the server takes the state tensor constructed in step S1051 as input and passes it to the timing encoding module embedded in the preset crystal growth process model. Inside the timing encoding module, a bidirectional long short-term memory (LSTM) network begins processing the state tensor. The forward LSM layer of the network starts from the earliest historical composite feature vector and processes it sequentially up to the current composite feature vector in chronological order, extracting forward temporal dependency features. Simultaneously, the reverse LSM layer of the network starts from the current composite feature vector and processes it sequentially up to the earliest historical composite feature vector in reverse chronological order, extracting reverse temporal dependency features. The server merges the feature representations finally output from the forward and reverse processing paths, for example, through vector concatenation or element-wise addition, to generate a timing-encoded vector containing rich dynamic information.

[0076] Step S1053: Concatenate the timing encoding vector with the directional component to obtain the joint inference input vector, and input the joint inference input vector into the strategy decision module of the preset crystal growth process model to output the candidate adjustment action and the probability distribution corresponding to the candidate adjustment action.

[0077] In step S1053, concatenation refers to the operation of joining two or more independent vectors end-to-end to form a longer vector. The joint inference input vector is generated through the concatenation operation and is a combined vector containing information on the dynamic evolution of the system and the current deviation direction. The strategy decision module is the part of the preset crystal growth process model responsible for outputting specific action suggestions. The candidate adjustment actions are generated by the strategy decision module and are a set of alternative operations that may improve the current growth state. The probability distribution is output along with the candidate adjustment actions, a set of probability values ​​corresponding to each candidate adjustment action, representing the model's confidence that each action can achieve a good effect.

[0078] Specifically, the server obtains the timing encoding vector generated in step S1052 and the direction component generated in step S1051. The server concatenates the timing encoding vector and the direction component to form a joint inference input vector. This joint inference input vector is input to the strategy decision module of the preset crystal growth process model. The strategy decision module typically consists of several fully connected neural network layers. This module performs nonlinear transformations and calculations on the joint inference input vector, ultimately outputting a set of discrete candidate adjustment actions, such as increasing the methane flow rate by 1 standard cubic centimeter per minute or decreasing the microwave power by 20 watts. Simultaneously, the output layer of the strategy decision module, typically a Softmax layer, calculates a probability value for each candidate adjustment action, and all probability values ​​constitute a probability distribution.

[0079] Step S1054: Generate a sequence of adjustment actions to be executed based on candidate adjustment actions and probability distribution.

[0080] In step S1054, the sequence of adjustment actions to be executed is selected from the candidate adjustment actions based on the probability distribution output by the model, and is a specific sequence of operation instructions to be actually issued to the equipment.

[0081] Specifically, the server receives candidate adjustment actions and their corresponding probability distributions from the strategy decision module. The server selects the final action based on a preset decision strategy. A common strategy is a greedy strategy, which directly selects the candidate adjustment action with the highest probability value in the probability distribution as the sequence of adjustment actions to be executed. For example, if the probability of the candidate adjustment action reducing methane flow is 0.9, while the probabilities of other actions are very low, the server will determine that reducing methane flow is the adjustment action to be executed. In more complex scenarios, the server may also employ a probability-based random sampling strategy to strike a balance between exploring new processes and utilizing existing optimal strategies. The finally selected one or more actions constitute the sequence of adjustment actions to be executed.

[0082] Step S1055: Perform feasibility constraint verification on the sequence of adjustment actions to be performed. The feasibility constraint verification includes verification of the safe working range of the actuator and verification of the equipment response capability limitations.

[0083] In step S1055, the feasibility constraint verification is a safety review process used to ensure that the adjustment actions generated by the model are physically safe and executable. The actuator safe operating range verification checks whether the target value of the adjustment action exceeds the minimum and maximum set range that the relevant hardware, such as a mass flow controller or power supply, can safely handle. The equipment response capability limit verification checks whether the rate of change required by the adjustment action exceeds the physical performance limits of the relevant hardware, such as the rate of temperature rise / fall or the valve opening / closing speed.

[0084] Specifically, the server reviews each action in the sequence of adjustment actions generated in step S1054. First, the server performs a safe operating range check on the actuators. The server queries a database storing all actuator parameters and compares the target value of the adjustment action with the corresponding safe operating upper and lower limits recorded in the database. For example, an action that sets the methane flow rate to 120 standard cubic centimeters per minute will fail the check if the corresponding mass flow controller's maximum range is 100 standard cubic centimeters per minute. Next, the server performs a device response capability limit check. The server calculates the implicit rate of change of the adjustment action and compares it with the maximum device response rate recorded in the database. For example, an action that requires raising the reaction chamber temperature by 50 degrees Celsius within one second will also fail the check if the maximum heating rate of the heating system is 10 degrees Celsius per second.

[0085] Step S1056: Encode the adjustment actions that have passed the feasibility constraint verification to generate preliminary adjustment instructions.

[0086] In step S1056, encoding is the process of converting verified, human-understandable adjustment actions into a machine language or communication protocol format that the MPCVD equipment's underlying controller can recognize and execute. The initial adjustment instructions are generated after encoding and can be directly sent to the final instructions of the equipment's hardware control system.

[0087] Specifically, the server acquires adjustment actions that have passed all feasibility constraint checks in step S1055. According to a predefined communication protocol, the server translates these high-level adjustment actions into low-level instruction strings or data packets. For example, an adjustment action to set the flow rate of gas channel 2 to 50.5 standard cubic centimeters per minute would be encoded by the server into an ASCII string conforming to a specific serial communication protocol, such as SET MFC2 FLOW 50.5. This encoded string or data packet is the initial adjustment instruction, which is temporarily stored and awaits transmission to the central controller of the MPCVD device for physical operation.

[0088] Step S106: Input the integrated feature vector and the difference feature vector into the preset crystal growth process model for reasoning to obtain preliminary adjustment instructions.

[0089] In step S106, the preset crystal growth process model refers to a deep learning model that has been pre-trained through reinforcement learning or supervised learning and is capable of inferring the optimal adjustment strategy based on the current growth state and the difference from the target. The preliminary adjustment instructions refer to the set of raw control instructions generated by the preset crystal growth process model, which are intended to correct state deviations but have not yet considered the impact of equipment hardware state drift.

[0090] Specifically, the server stacks the current integrated feature vector with the integrated feature vectors from several past time steps stored in the history buffer to construct a state tensor with temporal depth. Simultaneously, the server decomposes the difference feature vectors to obtain an amplitude component representing the severity of the deviation and a directional component representing the direction of the deviation. The server inputs the state tensor into a temporal encoding module embedded in a pre-defined crystal growth process model. This module, typically composed of a bidirectional long short-term memory network, is used to extract temporal dependency features of the growth state and generate a temporal encoded vector. Subsequently, the server concatenates this temporal encoded vector with the directional component of the difference feature vector to form a joint inference input vector. This joint inference input vector is input to the model's policy decision module, which outputs a series of candidate adjustment actions and the confidence probability distribution for each action. The server selects the optimal combination of adjustment actions based on the probability distribution, generating a sequence of adjustment actions to be executed. The server performs a feasibility constraint check on this sequence, checking whether each adjustment action exceeds the safe operating range of the corresponding actuator and whether the adjustment rate is within the equipment's response capability. The adjustment actions that pass the check are encoded into preliminary adjustment instructions containing specific parameter values.

[0091] In one possible implementation, the comprehensive feature vector and the difference feature vector are input into a preset crystal growth process model for inference to obtain preliminary adjustment instructions, specifically including steps S1061-S1066, as follows: Step S1061: Stack the comprehensive feature vector with multiple historical comprehensive feature vectors within the preset historical time window to construct a state tensor with time depth, and decompose the difference feature vector to obtain the amplitude component and the direction component.

[0092] In step S1061, the preset historical time series window refers to a fixed-length time interval that includes several consecutive sampling time points prior to the current moment. The state tensor is a three-dimensional numerical array composed of the current comprehensive feature vector and multiple historical comprehensive feature vectors within the preset historical time series window, stacked along the time dimension to capture the dynamic evolution of the system state. Decomposition refers to the process of breaking down a vector into its magnitude and direction. The magnitude component is a scalar value representing the length or magnitude of the difference feature vector, quantifying the overall degree of deviation of the system state from the target. The direction component is a unit vector representing the orientation of the difference feature vector in the multidimensional feature space, revealing the specific nature of the system state deviation.

[0093] Specifically, the server first obtains the composite feature vector at the current moment. Simultaneously, the server retrieves 10 consecutive historical composite feature vectors generated within a preset historical time window, such as the past 10 sampling periods, from internal storage. The server stacks these 10 historical composite feature vectors with the current composite feature vector in chronological order, forming a state tensor with 11 time steps. At the same time, the server obtains the input difference feature vector. The server calculates the Euclidean norm of this difference feature vector, and the result is the amplitude component. Subsequently, the server divides each element of the difference feature vector by this amplitude component, thus obtaining a unit vector with a magnitude of 1, which is the direction component.

[0094] Step S1062: Input the state tensor into the timing coding module embedded in the preset crystal growth process model, extract timing-dependent features through a bidirectional long short-term memory network, and generate a timing coding vector.

[0095] In step S1062, the preset crystal growth process model is a deep learning model pre-trained with a large amount of historical growth data. This model can understand the complex relationship between process state and final crystal quality and make decisions accordingly. The timing encoding module is a component within the preset crystal growth process model, specifically designed to process and understand the dynamic characteristics of sequential data. The bidirectional long short-term memory network is a special type of recurrent neural network that, through forward and backward processing paths, can simultaneously capture past and future contextual information in the sequential data. Temporal dependency features refer to the correlations and patterns existing between data at different time points in time series data. The timing encoding vector is a fixed-length numerical vector that is a generalized representation obtained by the timing encoding module after processing the state tensor, condensing the evolution of the system state over time.

[0096] Specifically, the server takes the state tensor constructed in step S1061 as input and passes it to the timing encoding module embedded in the preset crystal growth process model. Inside the timing encoding module, a bidirectional long short-term memory (LSTM) network begins processing the state tensor. The forward LSM layer of the network starts from the earliest historical composite feature vector and processes it sequentially up to the current composite feature vector in chronological order, extracting forward temporal dependency features. Simultaneously, the reverse LSM layer of the network starts from the current composite feature vector and processes it sequentially up to the earliest historical composite feature vector in reverse chronological order, extracting reverse temporal dependency features. The server merges the feature representations finally output from the forward and reverse processing paths, for example, through vector concatenation or element-wise addition, to generate a timing-encoded vector containing rich dynamic information.

[0097] Step S1063: Concatenate the timing encoding vector with the directional component to obtain the joint inference input vector, and input the joint inference input vector into the strategy decision module of the preset crystal growth process model to output the candidate adjustment action and the probability distribution corresponding to the candidate adjustment action.

[0098] In step S1063, splicing refers to the operation of connecting two or more independent vectors end-to-end to form a longer vector. The joint inference input vector is generated through the splicing operation and is a combined vector containing information on the dynamic evolution of the system and the current deviation direction. The strategy decision module is the part of the preset crystal growth process model responsible for outputting specific action suggestions. The candidate adjustment actions are generated by the strategy decision module and are a set of alternative operations that may improve the current growth state. The probability distribution is output along with the candidate adjustment actions, a set of probability values ​​corresponding to each candidate adjustment action, representing the model's confidence that each action can achieve a good effect.

[0099] Specifically, the server obtains the timing encoding vector generated in step S1062 and the direction component generated in step S1061. The server concatenates the timing encoding vector and the direction component to form a joint inference input vector. This joint inference input vector is input to the strategy decision module of the preset crystal growth process model. The strategy decision module typically consists of several fully connected neural network layers. This module performs nonlinear transformations and calculations on the joint inference input vector, ultimately outputting a set of discrete candidate adjustment actions, such as increasing the methane flow rate by 1 standard cubic centimeter per minute or decreasing the microwave power by 20 watts. Simultaneously, the output layer of the strategy decision module, typically a Softmax layer, calculates a probability value for each candidate adjustment action, and all probability values ​​constitute a probability distribution.

[0100] Step S1064: Generate a sequence of adjustment actions to be executed based on candidate adjustment actions and probability distribution.

[0101] In step S1064, the sequence of adjustment actions to be executed is selected from the candidate adjustment actions based on the probability distribution output by the model, and is a specific sequence of operation instructions to be actually issued to the equipment.

[0102] Specifically, the server receives candidate adjustment actions and their corresponding probability distributions from the strategy decision module. The server selects the final action based on a preset decision strategy. A common strategy is a greedy strategy, which directly selects the candidate adjustment action with the highest probability value in the probability distribution as the sequence of adjustment actions to be executed. For example, if the probability of the candidate adjustment action reducing methane flow is 0.9, while the probabilities of other actions are very low, the server will determine that reducing methane flow is the adjustment action to be executed. In more complex scenarios, the server may also employ a probability-based random sampling strategy to strike a balance between exploring new processes and utilizing existing optimal strategies. The finally selected one or more actions constitute the sequence of adjustment actions to be executed.

[0103] Step S1065: Perform feasibility constraint verification on the sequence of adjustment actions to be performed. The feasibility constraint verification includes verification of the safe working range of the actuator and verification of the equipment response capability limitations.

[0104] In step S1065, the feasibility constraint verification is a safety review process used to ensure that the adjustment actions generated by the model are physically safe and executable. The actuator safe operating range verification checks whether the target value of the adjustment action exceeds the minimum and maximum set range that the relevant hardware, such as a mass flow controller or power supply, can safely handle. The equipment response capability limit verification checks whether the rate of change required by the adjustment action exceeds the physical performance limits of the relevant hardware, such as the rate of temperature rise / fall or the valve opening / closing speed.

[0105] Specifically, the server reviews each action in the sequence of adjustment actions generated in step S1064. First, the server performs a safe operating range check on the actuators. The server queries a database storing all actuator parameters and compares the target value of the adjustment action with the corresponding safe operating upper and lower limits recorded in the database. For example, an action that sets the methane flow rate to 120 standard cubic centimeters per minute will fail the check if the corresponding mass flow controller's maximum range is 100 standard cubic centimeters per minute. Next, the server performs a device response capability limit check. The server calculates the implicit rate of change of the adjustment action and compares it with the maximum device response rate recorded in the database. For example, an action that requires raising the reaction chamber temperature by 50 degrees Celsius within one second will also fail the check if the maximum heating rate of the heating system is 10 degrees Celsius per second.

[0106] Step S1066: Encode the adjustment actions that have passed the feasibility constraint verification to generate preliminary adjustment instructions.

[0107] In step S1066, encoding is the process of converting verified, human-understandable adjustment actions into a machine language or communication protocol format that the MPCVD equipment's underlying controller can recognize and execute. The initial adjustment instructions are generated after encoding and can be directly sent to the final instructions of the equipment's hardware control system.

[0108] Specifically, the server acquires adjustment actions that have passed all feasibility constraint checks in step S1065. According to a predefined communication protocol, the server translates these high-level adjustment actions into low-level instruction strings or data packets. For example, an adjustment action to set the flow rate of gas channel 2 to 50.5 standard cubic centimeters per minute would be encoded by the server into an ASCII string conforming to a specific serial communication protocol, such as SET MFC2 FLOW 50.5. This encoded string or data packet is the initial adjustment instruction, which is temporarily stored and awaits transmission to the central controller of the MPCVD device for physical operation.

[0109] Step S107: Perform feedforward compensation calculation on the preliminary adjustment command based on the equipment state drift matrix to obtain the final control command, and send the final control command to each actuator of the MPCVD equipment.

[0110] In step S107, feedforward compensation calculation refers to a control method that predicts and actively compensates for the effects of known disturbances in the system. Here, disturbances specifically refer to hardware state drift caused by factors such as equipment aging. The final control command refers to the final control signal, after being corrected by the feedforward compensation calculation, that can accurately offset the effects of equipment state drift and is issued to the hardware for execution. The actuator refers to the physical component in the MPCVD equipment that directly executes the control command, such as a microwave generator, gas mass flow controller, or vacuum pump.

[0111] Specifically, the server first parses the preliminary adjustment commands generated by the preset crystal growth process model, extracts the original adjustment parameters for each actuator, and constructs an original adjustment vector. The server performs a Laplace transform on this original adjustment vector to convert it to the frequency domain, obtaining a frequency domain adjustment vector. Next, the server performs matrix multiplication on this frequency domain adjustment vector and the inverse matrix of the device state drift matrix obtained in step S101. This operation essentially corrects the commands in the frequency domain to offset the gain attenuation and phase lag caused by device drift; the result is a frequency-domain compensated adjustment vector. The server then performs an inverse Laplace transform on this frequency-domain compensated adjustment vector to convert it back to the time domain, generating a time-domain compensated adjustment vector. Furthermore, the server calculates the timing advance required to compensate for each adjustment command based on the device state drift matrix. Finally, the server associates and binds each component value in the time-domain compensated adjustment vector with its corresponding timing advance, generating a final control command. This final control command contains precisely amplitude and phase compensated adjustment parameters, and the server sends this final control command to each actuator corresponding to the MPCVD equipment for implementation.

[0112] In one possible implementation, the initial adjustment command is fedforward compensation calculation is performed on the device state drift matrix to obtain the final control command, specifically including steps S1071-S1074, as follows: Step S1071: Parse the preliminary adjustment command and generate an original adjustment vector containing the original adjustment parameters of each actuator.

[0113] In step S1071, parsing refers to the process of converting machine-readable preliminary adjustment instructions back into structured data, extracting the target actuators and corresponding parameter values ​​contained in the instructions. The raw adjustment parameters refer to the specific values ​​set for each actuator in the instructions, such as target power values ​​or target gas flow rates. The raw adjustment vector is a numerical vector composed of the raw adjustment parameters of all relevant actuators arranged in a predetermined order; this vector represents the complete adjustment intention of the system within that time step.

[0114] Specifically, the server receives preliminary adjustment instructions from previous steps. These instructions may be one or more encoded strings, such as setting the microwave power to 3000 watts and the methane flow rate to 28.0 standard cubic centimeters per minute. The server decodes these instruction strings according to its internal instruction parsing rules, identifying the target actuators as the microwave power source and the methane flow controller, and extracts the corresponding raw adjustment parameter values ​​of 3000 and 28.0. Following a fixed actuator order, such as [microwave power, cavity pressure, methane flow rate, hydrogen flow rate], the server places the extracted parameter values ​​in their corresponding positions. For actuators not mentioned in the instructions, their current settings are used to fill in the blanks. Finally, a complete list of values ​​containing the settings of all actuators is generated; this list is the raw adjustment vector.

[0115] Step S1072: Perform a Laplace transform on the original adjustment vector to obtain a frequency domain adjustment vector, and perform matrix multiplication on the frequency domain adjustment vector and the inverse of the device state drift matrix to generate a frequency domain compensated adjustment vector.

[0116] In step S1072, the Laplace transform is a mathematical integral transform used to convert the signal function in the time domain to the complex frequency domain for analysis, facilitating the processing of the system's dynamic response. The frequency domain adjustment vector is the frequency domain representation of the original adjustment vector after the Laplace transform. The equipment state drift matrix is ​​a transfer function matrix in the frequency domain that describes the dynamic coupling relationship between actuators and their own response hysteresis characteristics; this matrix quantifies the non-ideal dynamic behavior of the equipment. The inverse matrix is ​​a matrix whose product with the original matrix is ​​an identity matrix, used in control to counteract the system effects represented by the original matrix. Matrix multiplication is a standard matrix operation defined in linear algebra. The frequency domain compensated adjustment vector is the result vector obtained after pre-correcting the adjustment signal in the frequency domain to counteract the effects of equipment state drift.

[0117] Specifically, the server obtains the original adjustment vector generated in step S1071. The server treats each component of the original adjustment vector, i.e., the change process of each original adjustment parameter from its current value to its target value, as a step signal function in the time domain, and performs a Laplace transform on each step signal function, thereby transforming the entire original adjustment vector from the time domain to the frequency domain, obtaining a frequency domain adjustment vector. Simultaneously, the server loads the device state drift matrix, pre-identified by the system, from internal storage. The server then calculates the inverse of this device state drift matrix. Subsequently, the server performs matrix multiplication between this inverse matrix and the frequency domain adjustment vector. The purpose of this operation is to reverse-shape the control signal in the frequency domain, so that after the device's own drift effect, the final actual physical quantity can accurately reach the desired value. The result of the operation is the frequency domain compensated adjustment vector.

[0118] Step S1073: Perform an inverse Laplace transform on the frequency-domain compensated adjustment vector to generate a time-domain compensated adjustment vector.

[0119] In step S1073, the inverse Laplace transform is a mathematical process that converts a function in the complex frequency domain back to a function in the time domain. It is used to restore the frequency-domain processed signal to a time-dependent curve of physical quantity change. The time-domain compensated adjustment vector is the specific representation of the frequency-domain compensated adjustment signal in the time domain. Each component of this vector is no longer a single target value, but rather describes the precise trajectory of how the control parameter should change over time.

[0120] Specifically, the server obtains the frequency-domain compensated adjustment vector generated in step S1072. The server performs an inverse Laplace transform on each complex frequency-domain component of this vector, restoring the signal from the frequency domain back to the time domain. This inverse transform process converts the compensation operation performed in the frequency domain into a specific control signal waveform in the time domain. For example, an ideal step control command, after compensation and inverse transform, may be transformed into a signal curve with an initial overshoot that quickly stabilizes to the target value, or a signal curve that rises linearly with a specific slope, the purpose of which is to overcome system inertia and reach a stable state as quickly as possible. The vector obtained after the transformation is the time-domain compensated adjustment vector, where each component is a target value function with respect to time.

[0121] Step S1074: Calculate the timing advance of each adjustment command based on the equipment state drift matrix, and associate and bind each component value in the time-domain compensated adjustment vector with the corresponding timing advance to generate the final control command.

[0122] In step S1074, timing advance refers to the length of time by which the control action needs to be executed in advance to compensate for the response delay of the actuator and physical process. Association binding refers to pairing the compensated adjustment parameter value with the corresponding execution time point to form a timestamped instruction. The final control instruction is a complete set of instructions that combines the compensated parameter value and precise execution timing, and can be directly issued to the underlying hardware control system of the device.

[0123] Specifically, the server re-analyzes the device state drift matrix, extracting the pure time lag parameter for each actuator from the diagonal transfer function elements of the matrix. The value of this parameter is the timing advance of the corresponding adjustment command. For example, the analysis shows that the response delay for microwave power is 0.8 seconds, while the response delay for methane flow is 0.3 seconds. The server then obtains the time-domain compensated adjustment vector generated in step S1073. The server associates and binds each component function in the time-domain compensated adjustment vector with the calculated corresponding timing advance. This binding process generates a series of command sequences with precise execution timestamps. If the target effective time of an adjustment action is T, then the starting execution timestamp of the microwave power adjustment command will be set to T-0.8 seconds. The server packages all these timestamped compensated adjustment parameter functions to form the final control command.

[0124] Please refer to Figure 2 In one possible implementation, before performing feature extraction and fusion on the multimodal data to construct a comprehensive feature vector, the method further includes steps S201-S204, as follows: Step S201: Obtain the original acquisition timestamp and sampling frequency corresponding to each modality in the multimodal data.

[0125] In step S201, multimodal data refers to data in various formats collected from different types of sensors or information sources during the crystal growth process, such as spectral data from a spectrometer, temperature data from a thermocouple, and image data from a camera. The original acquisition timestamp refers to the precise moment when each independent data point is recorded by the acquisition system. The sampling frequency refers to the number of times a specific type of data source acquires data points per unit time; for example, if spectral data is acquired 10 times per second, the sampling frequency is 10 Hz.

[0126] Specifically, the server connects to the central data acquisition unit of the MPCVD system. The server receives data streams from various sensors from this unit. For each received data point, whether it's a set of intensity values ​​from the spectrometer or a reading from the temperature sensor, the server reads and stores the metadata transmitted along with that data point. The server extracts the original acquisition timestamp from the metadata. Simultaneously, the server accesses a pre-configured device parameter file that records the set sampling frequency for each sensor in the system. The server reads and records the sampling frequency corresponding to each modality of data; for example, the spectrometer's sampling frequency is 10 Hz, the temperature sensor's sampling frequency is 1 Hz, and the CCD camera's sampling frequency is 0.5 Hz.

[0127] Step S202: Based on the sliding window method, calculate the statistical feature threshold of each modality data within the sliding window, and identify and mark abnormal data points whose values ​​exceed the statistical feature threshold.

[0128] In step S202, the sliding window method is a data analysis technique that segments the data by moving a fixed-length window across the data sequence and performing local calculations on the data within each window. The statistical characteristic threshold is one or more boundary values ​​calculated based on the statistical properties of the data within the window, such as the mean and standard deviation, used to determine whether a data point is normal. Outlier data points are those whose values ​​exceed the range of the statistical characteristic threshold; these data points are generally considered invalid readings caused by sensor noise, signal interference, or transient malfunctions.

[0129] Specifically, the server processes the time series data for each modality separately. Taking temperature data as an example, the server sets a sliding window with a length of 50 data points. The server places this window at the beginning of the temperature data series and calculates the mean and standard deviation of these 50 data points. Based on the three-sigma criterion, the server sets the upper threshold as the mean plus three times the standard deviation and the lower threshold as the mean minus three times the standard deviation; these two values ​​constitute the statistical thresholds. The server checks the data point located at the center of the window. If the value of this data point is higher than the upper threshold or lower than the lower threshold, the server marks it as an outlier. After completing the check, the server moves the sliding window one data point to the right and repeats the process of calculating the statistical thresholds and marking outliers until the entire temperature data series has been traversed. The server performs the same operation for all other modal data, such as spectral data and pressure data.

[0130] Step S203: Based on the valid data adjacent to the abnormal data points, use the cubic spline interpolation algorithm to interpolate and repair the abnormal data points and missing data locations to obtain the repaired data.

[0131] In step S203, valid data refers to normal data points that were not marked as anomalous in step S202. Cubic spline interpolation is an advanced interpolation method that uses piecewise cubic polynomials to construct a smooth curve passing through a series of known data points, ensuring the continuity of the first and second derivatives of the curve at connection points. Missing data locations refer to blank points where no data points were collected at the expected timestamp. Repaired data refers to the complete dataset obtained after all anomalous data points and missing data locations in the original dataset have been replaced with new values ​​calculated by the interpolation algorithm.

[0132] Specifically, the server processes the data sequence marked with outliers in the preceding steps. When the server traverses to a location marked as an outlier, or discovers a missing data point by comparing timestamps, it searches for several valid data points immediately before and after that location, for example, two before and two after. The server takes the values ​​and timestamps of these four valid data points as input and applies a cubic spline interpolation algorithm. This algorithm generates a piecewise cubic polynomial function that smoothly passes through these four valid data points. Subsequently, the server substitutes the timestamps corresponding to the outlier or missing data point into this generated polynomial function for calculation, obtaining an interpolation result. The server uses this interpolation result to replace the original outlier or fill in the missing data point. The server repeats this repair process for all outliers and missing points in all modal data, and the final dataset is the repaired data.

[0133] Step S204: Select the timestamp sequence corresponding to the modality with the lowest sampling frequency in the multimodal data as a unified time reference, and align the other modality data using the nearest neighbor timestamp matching resampling method to obtain the preprocessed multimodal data.

[0134] In step S204, the unified time base is a common timestamp sequence. All modal data with different sampling frequencies will be adjusted to this time series to achieve temporal synchronization. The nearest neighbor timestamp matching resampling method is an alignment technique. For each time point in the unified time base, this method finds the data point in the data sequence to be aligned whose original timestamp is closest to that time point, and uses the value of that data point as the aligned value. The preprocessed multimodal data refers to the data of all modalities that have undergone anomaly repair and time alignment, forming a structured dataset that is completely synchronized in time.

[0135] Specifically, the server first checks the sampling frequency of all modal data acquired in step S201. Assume the temperature data sampling frequency is 1 Hz, the spectral data sampling frequency is 10 Hz, and the camera image data sampling frequency is 0.5 Hz. The server identifies the camera image data as having the lowest sampling frequency. Therefore, the server selects the timestamp sequence of the camera image data, [0s, 2s, 4s, 6s, ...], as the unified time reference. Next, the server resamples the modal data with higher sampling frequencies. Taking 10 Hz spectral data as an example, for the first time point 2s in the unified time reference, the server searches the timestamp sequence of the spectral data for the closest timestamp to 2s. Assuming a spectral data point with a timestamp of 2.01s is found, the server associates this spectral data acquired at 2.01s with the 2s time point in the unified time reference. The server repeats this nearest-neighbor timestamp matching process for each time point in the unified time reference and for all modal data requiring downsampling. Ultimately, all modal data have a one-to-one corresponding value on a unified time base, forming preprocessed multimodal data.

[0136] The following describes an intelligent control device for an MPCVD production line from the perspective of hardware processing in an embodiment of this invention. Please refer to [link to relevant documentation]. Figure 3 This is a schematic diagram of the structure of an intelligent control device for an MPCVD production line in an embodiment of this application.

[0137] It should be noted that, Figure 3 The structure of the intelligent control device for an MPCVD production line shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0138] like Figure 3 As shown, an intelligent control device for an MPCVD production line includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes according to a program stored in a Read-Only Memory (ROM) 302 or a program loaded from a storage section 308 into a Random Access Memory (RAM) 303, such as executing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for device operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0139] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including a hard disk, etc.; and communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0140] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0141] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0142] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0143] Specifically, an MPCVD production line intelligent control device according to this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements an MPCVD production line intelligent control method provided in the above embodiment.

[0144] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the MPCVD production line intelligent control device described in the above embodiments; or it may exist independently and not assembled into the MPCVD production line intelligent control device. The storage medium carries one or more computer programs, which, when executed by a processor of the MPCVD production line intelligent control device, cause the MPCVD production line intelligent control device to implement the MPCVD production line intelligent control method provided in the above embodiments.

Claims

1. A method for intelligent control of an MPCVD production line, characterized in that, The method includes: Perform stimulus response tests on the MPCVD equipment in the MPCVD production line to obtain the equipment state drift matrix of the equipment hardware status; During the diamond growth process, multimodal data are collected from the reaction chamber of the MPCVD equipment in the MPCVD production line. Feature extraction and fusion are performed on the multimodal data to construct a comprehensive feature vector; Obtain the current growth stage of the diamond and determine the target state vector corresponding to the current growth stage; The comprehensive feature vector and the target state vector are compared to generate a difference feature vector; The integrated feature vector and the difference feature vector are input into a preset crystal growth process model for reasoning to obtain preliminary adjustment instructions; Based on the device state drift matrix, the preliminary adjustment command is fed forward to obtain the final control command, and the final control command is sent to each actuator of the MPCVD device.

2. The method according to claim 1, characterized in that, The step of performing stimulus response testing on the MPCVD equipment in the MPCVD production line to obtain the equipment state drift matrix of the equipment hardware state specifically includes: When the MPCVD device is in a preset reference state, an excitation response test program is executed to obtain the reference transfer function; Input a preset standard excitation signal sequence into the MPCVD equipment; The microwave system reflection coefficient and plasma emission spectrum data triggered by the preset standard excitation signal sequence are collected as the response output; The Laplace transform is applied to the preset standard excitation signal sequence and the response output respectively to calculate the transfer function characterizing the current state of the device; The device state drift matrix is ​​generated by performing a matrix difference operation between the transfer function and the reference transfer function.

3. The method according to claim 1, characterized in that, The step of extracting and fusing features from the multimodal data to construct a comprehensive feature vector specifically includes: Feature extraction was performed on the plasma emission spectrum data, plasma sphere image data, microwave power data, and reaction chamber temperature and pressure data in the multimodal data to obtain spectral feature vectors, image feature vectors, and time-series statistical feature vectors. Based on the attention mechanism, the correlation weight coefficients between the spectral feature vector, the image feature vector, and the temporal statistical feature vector and the current crystal growth stage are calculated respectively. The spectral feature vector, the image feature vector, and the temporal statistical feature vector are weighted and concatenated according to the correlation weight coefficients, and then normalized to obtain the comprehensive feature vector.

4. The method according to claim 1, characterized in that, The step of inputting the comprehensive feature vector and the difference feature vector into a preset crystal growth process model for inference to obtain preliminary adjustment instructions specifically includes: The comprehensive feature vector is stacked with multiple historical comprehensive feature vectors within a preset historical time window to construct a state tensor with time depth, and the difference feature vector is decomposed to obtain amplitude components and direction components. The state tensor is input into the timing coding module embedded in the preset crystal growth process model, and timing-dependent features are extracted through a bidirectional long short-term memory network to generate a timing coding vector. The timing encoding vector is concatenated with the directional component to obtain a joint inference input vector, and the joint inference input vector is input to the strategy decision module of the preset crystal growth process model to output candidate adjustment actions and the probability distribution corresponding to the candidate adjustment actions; Based on the candidate adjustment actions and the probability distribution, a sequence of adjustment actions to be executed is generated; The feasibility constraint verification is performed on the sequence of adjustment actions to be executed. The feasibility constraint verification includes verification of the safe working range of the actuator and verification of the equipment response capability limitation. The adjustment actions that pass the feasibility constraint verification are encoded to generate the preliminary adjustment instructions.

5. The method according to claim 1, characterized in that, The step of performing feedforward compensation calculations on the preliminary adjustment commands based on the device state drift matrix to obtain the final control commands specifically includes: The initial adjustment command is parsed to generate an initial adjustment vector containing the initial adjustment parameters of each actuator; The original adjustment vector is subjected to a Laplace transform to obtain a frequency domain adjustment vector. The frequency domain adjustment vector is then multiplied with the inverse of the device state drift matrix to generate a frequency domain compensated adjustment vector. Perform an inverse Laplace transform on the frequency-domain compensated adjustment vector to generate a time-domain compensated adjustment vector; The timing advance of each adjustment command is calculated based on the device state drift matrix, and each component value in the time-domain compensated adjustment vector is associated and bound with the corresponding timing advance to generate the final control command.

6. The method according to claim 1, characterized in that, Before performing feature extraction and fusion on the multimodal data to construct a comprehensive feature vector, the method further includes: Obtain the original acquisition timestamp and sampling frequency corresponding to each modality in the multimodal data; Based on the sliding window method, the statistical feature thresholds of each modality data within the sliding window are calculated respectively, and abnormal data points whose values ​​exceed the statistical feature thresholds are identified and marked. For the abnormal data points and missing data locations, data repair and time-series alignment are performed to obtain preprocessed multimodal data.

7. The method according to claim 6, characterized in that, The process of performing data repair and time-series alignment on the abnormal data points and missing data locations to obtain preprocessed multimodal data specifically includes: Based on the valid data adjacent to the abnormal data points, a cubic spline interpolation algorithm is used to interpolate and repair the abnormal data points and the missing data locations to obtain the repaired data. The timestamp sequence corresponding to the modality with the lowest sampling frequency in the multimodal data is selected as a unified time reference, and the other modality data are aligned using a nearest neighbor timestamp matching resampling method to obtain the preprocessed multimodal data.

8. An intelligent control device for an MPCVD production line, characterized in that, The MPCVD production line intelligent control device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the MPCVD production line intelligent control device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the MPCVD production line intelligent control device, the MPCVD production line intelligent control device performs the method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program runs on the intelligent control device of the MPCVD production line, it causes the intelligent control device of the MPCVD production line to perform the method as described in any one of claims 1-7.

Images