A Cloud Computing-Based Automated Production Process Optimization Method and System for Synthetic Mica

Abstract

This invention relates to the field of mica automation technology, and more particularly to a cloud-based method and system for optimizing the automated production process of synthetic mica. The method includes the following steps: deploying a multi-source industrial sensor array on the synthetic mica production line to collect data throughout the entire process, constructing a preliminary industrial dataset for mica; using a cloud platform to extract and calculate the statistical and frequency domain features of mica synthesis from the preliminary industrial dataset, and fitting distribution parameters to obtain spatial gradient feature data of mica; performing tensor concatenation on the spatial gradient feature data of mica to obtain a mica generation feature matrix; and creating a closed-loop process from multi-source data acquisition, feature extraction and fusion, mechanism model embedding, dynamic state representation to action optimization generation, effectively improving the intelligence level of the mica production process and enhancing the system's adaptability and responsiveness to complex and changing production environments.

Description

Technical Field

[0001] This invention relates to the field of mica automation technology, and in particular to a method and system for optimizing the automated production process of synthetic mica based on cloud computing. Background Technology

[0002] Most existing technologies employ static parameter configuration methods, lacking the ability to model various nonlinear coupling factors in the production process (such as the dynamic relationship between crystal growth rate and temperature, supersaturation), making it difficult to cope with the dynamic optimization needs under real-world conditions such as production disturbances, raw material fluctuations, or changes in operating conditions. Secondly, at the data level, there are problems such as limited data acquisition dimensions, insufficient spatiotemporal resolution, and ineffective collaboration of distributed data, failing to achieve global perception and accurate modeling of the entire production line status. Thirdly, most current systems lack advanced intelligent analysis functions, only able to make control logic judgments within a limited range based on rule bases or expert experience, lacking self-learning and strategy optimization mechanisms based on historical data and real-time feedback. Furthermore, existing methods generally neglect the deep integration of physical mechanism models and data-driven models, resulting in limited prediction accuracy and difficulty in fully utilizing the fundamental laws such as crystal growth kinetics to improve model credibility. Finally, production optimization usually remains at the offline decision-making or manual intervention stage, lacking an integrated, closed-loop optimization-execution-evaluation process, resulting in significant bottlenecks in response speed and control efficiency, making it difficult to meet the demands of high-performance, high-consistency, and flexible production of modern synthetic mica. Therefore, when faced with high-dimensional dynamic data, highly coupled processes, and multi-objective optimization tasks, existing technologies urgently need a systematic approach that integrates various intelligent computing methods such as cloud computing, mechanism modeling, spatiotemporal feature analysis, and strategy sampling to overcome the above limitations. Summary of the Invention

[0003] Therefore, it is necessary to provide a cloud computing-based method and system for optimizing the automated production process of synthetic mica to solve at least one of the aforementioned technical problems.

[0004] To achieve the above objectives, a cloud computing-based automated production process optimization method for synthetic mica is proposed, comprising the following steps:

[0005] Step S1: Deploy a multi-source industrial sensor array on the synthetic mica production line to collect data throughout the entire process and construct a preliminary industrial dataset for mica.

[0006] Step S2: Utilize the cloud platform to extract and calculate the statistical and frequency domain features of mica synthesis from the preliminary industrial dataset of mica, and fit the distribution parameters to obtain the spatial gradient feature data of mica; perform tensor concatenation on the spatial gradient feature data of mica to obtain the mica generation feature matrix;

[0007] Step S3: Based on the mica generation feature matrix, the crystal growth kinetic equation of kiln heating is embedded to obtain the basic framework of the mechanism model; the mica spatial gradient feature data of the mica generation feature matrix is ​​used to predict the mica quality to obtain the theoretical quality value of the automated production line; the basic framework of the mechanism model and the theoretical quality value of the automated production line are fused with crystal purity to obtain the predicted value of mica mixed quality.

[0008] Step S4: Perform state space construction processing on the mica spatial gradient feature data of the mica generation feature matrix to obtain the mica state vector; set according to the preliminary industrial dataset of mica, and use the mica mixing quality prediction value to sample the crystal nucleus-mica melt to obtain the mica automated production optimization data.

[0009] The beneficial effects of this invention lie in achieving high-precision acquisition of data from the entire mica synthesis production line through the deployment of a multi-source industrial sensor array. This constructs a preliminary industrial dataset covering the entire production process, enriching the data dimensions and spatiotemporal coverage, and providing a detailed raw data foundation for subsequent analysis. Based on the data processing capabilities of the cloud platform, the system performs in-depth mining of the statistical and frequency domain features of the preliminary industrial dataset. Statistical calculations and distributed parameter fitting techniques are used to extract mica spatial gradient feature data, fully capturing the detailed features of spatial changes during mica production. Tensor splicing is used to structure the multidimensional spatial gradient feature data into a mica generation feature matrix, achieving efficient fusion and representation of multidimensional information, providing a solid data carrier for the numerical simulation of complex physical processes. Based on this feature matrix, a basic framework for a mechanistic model is constructed by embedding crystal growth kinetic equations, achieving a close integration of physical mechanisms and data features, and enhancing the model's explanatory and predictive capabilities for the mica growth process. By utilizing mica spatial gradient feature data for quality prediction, theoretical quality values ​​for automated production lines are obtained. Combined with a mechanistic model framework, a fusion processing method is used to generate a mica hybrid quality prediction value, improving the accuracy and robustness of the prediction results. State space construction transforms multi-dimensional features into dynamic state vectors, endowing the system with the ability to represent production states over time. Using nucleus-mica melt sampling, production optimization action proposals are generated based on the hybrid quality prediction values, enhancing the ability to explore the action space and enabling dynamic adjustment and optimized control of production parameters. Overall, this technology systematically constructs a closed-loop process from multi-source data acquisition, feature extraction and fusion, mechanistic model embedding, dynamic state representation to action optimization generation at the data level, effectively improving the intelligence level of the mica production process and enhancing the system's adaptability and responsiveness to complex and changing production environments. Attached Figure Description

[0010] Figure 1 A schematic diagram illustrating the steps of an optimized automated production process for synthetic mica based on cloud computing;

[0011] Figure 2 A schematic diagram showing the predicted mixed mass of mica;

[0012] Figure 3 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2.

[0013] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0015] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0016] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0017] To achieve the above objectives, please refer to Figures 1 to 3 A method for optimizing the automated production process of synthetic mica based on cloud computing, the method comprising the following steps:

[0018] Step S1: Deploy a multi-source industrial sensor array on the synthetic mica production line to collect data throughout the entire process and construct a preliminary industrial dataset for mica.

[0019] In this embodiment of the invention, based on the analysis of the mica synthesis process, key process nodes such as raw material transportation, pyrolysis reaction, crystal growth, cooling and solidification, and peeling are selected as data acquisition support points. Different types of industrial-grade sensor arrays are deployed at each node. These sensors include thermocouples, infrared thermometers, piezoelectric accelerometers, capacitive hygrometers, X-ray thickness analyzers, and image spectral acquisition equipment, covering multiple data dimensions such as temperature, pressure, vibration, humidity, density, crystal morphology, and structure. To ensure data consistency across time scales and upstream / downstream correlation, a synchronous clock system based on industrial Ethernet or 5G private network communication is introduced to achieve high-frequency time synchronization sampling for all sensor nodes, and the sampling frequency is uniformly mapped to a millisecond-level time series stream. Simultaneously, an edge node caching mechanism is used to perform short-term buffering, preliminary filtering, and labeling of high-throughput raw data, including marking signal anomalies, dividing key time periods into process segments, and batch mapping of sample numbers. The collected data is organized hierarchically according to the data source and finally uploaded to a private cloud or industrial internet platform through a data gateway to construct a "mica preliminary industrial dataset" with a four-element structure of time stamp, process segment, measurement point location and sampling dimension.

[0020] Step S2: Utilize the cloud platform to extract and calculate the statistical and frequency domain features of mica synthesis from the preliminary industrial dataset of mica, and fit the distribution parameters to obtain the spatial gradient feature data of mica; perform tensor concatenation on the spatial gradient feature data of mica to obtain the mica generation feature matrix;

[0021] In this embodiment of the invention, based on time-series data such as temperature, humidity, pressure, vibration spectrum, and crystal structure images collected by different sensors, a sliding window and time segmentation method is used to construct multi-scale sampling intervals. Within each interval, statistical characteristics are calculated for each dimension of the data, including the extraction of indicators such as mean, standard deviation, skewness, kurtosis, extreme value symmetry, and first / second derivative trends, forming a basic time-domain feature vector. Simultaneously, for variables with vibrational or periodic fluctuation characteristics, such as vibrational acceleration signals within the synthetic reaction chamber and thermal field disturbance curves, frequency domain analysis techniques such as Fast Fourier Transform (FFT) and Power Spectral Density Estimation (PSD) are used to extract the main frequency components, principal energy bandwidth, frequency offset, and harmonic distribution characteristics, further expanding the frequency domain feature dimensions. After completing the dual-channel feature extraction in both the statistical and frequency domains, spatial gradient fitting is performed on the data features of different sampling points, combining the spatial coordinate information corresponding to the data source and the physical layout of the sensors. A multidimensional gradient field is constructed, and the fitting model is mainly based on the estimation of multivariate Gaussian distribution parameters. The covariance structure, gradient direction field, and local density fluctuation index are calculated to obtain "mica spatial gradient feature data" that reflects the spatial differences between sensor arrays. On this basis, third-order and higher tensor stitching techniques are used to uniformly encapsulate different sampling time periods, spatial locations, and feature type dimensions into a tensor structure, forming a unified "mica generation feature matrix".

[0022] Step S3: Based on the mica generation feature matrix, the crystal growth kinetic equation of kiln heating is embedded to obtain the basic framework of the mechanism model; the mica spatial gradient feature data of the mica generation feature matrix is ​​used to predict the mica quality to obtain the theoretical quality value of the automated production line; the basic framework of the mechanism model and the theoretical quality value of the automated production line are fused with crystal purity to obtain the predicted value of mica mixed quality.

[0023] In this embodiment of the invention, a constructed "mica generation feature matrix" is used as input. This matrix integrates spatial distribution, time series, and multidimensional statistical frequency domain features, possessing strong data integrity and structural expressiveness. Based on this, key physical variables in the crystal growth process are selected as driving factors, such as heat flux density gradient, reaction chamber temperature perturbation frequency, raw material feed rate, and crystal plane structure perturbation parameters. The model equations are embedded by matching these variables with the kinetic expressions in existing crystal growth theories, such as the growth rate equation and the diffusion-convection-deposition synergistic equation. Specifically, symbolic mapping or parametric interpolation methods are used to map the corresponding dimensions in the data matrix to the physical terms in the kinetic equations. This ensures that the embedded mechanistic model framework retains the interrelationships between the velocity field, temperature field, and concentration field during crystal formation, while also possessing the ability to be dynamically solved using sample features. Subsequently, based on this mechanistic model, spatial gradient feature data was used as input variables, and methods such as multivariate regression fitting, spline interpolation, or nonlinear kernel function transformation were introduced to construct a theoretical quality prediction model. This model outputs theoretical quality values ​​for automated mica production lines under specific operating conditions, manifested as quality indicators such as particle size distribution, crystal plane integrity, and defect rate. After obtaining the theoretical quality values ​​for the automated production line, a hybrid quality prediction mechanism was constructed to further integrate the differences between actual industrial fluctuations and mechanistic deductions. This mechanism involves analyzing the error distribution of prediction results driven by theoretical values ​​and historical data, introducing a Bayesian confidence fusion strategy or a weighted residual correction method to adjust the output results of the mechanistic model, and finally generating a "mica hybrid quality prediction value" that characterizes the fusion result of the theoretical mechanism and data behavior.

[0024] Step S4: Perform state space construction processing on the mica spatial gradient feature data of the mica generation feature matrix to obtain the mica state vector; set according to the preliminary industrial dataset of mica, and use the mica mixing quality prediction value to sample the crystal nucleus-mica melt to obtain the mica automated production optimization data.

[0025] In this embodiment of the invention, based on the spatial gradient feature data in the "mica generation feature matrix" formed in steps S2 and S3, dynamic evolution variables related to process control are extracted, and a state transition structure is established according to the time series window sliding mechanism to form a high-dimensional time-space sequence data set. Subsequently, in the state modeling stage, the feature matrix is ​​structurally compressed by methods such as principal component dimensionality reduction, covariance analysis, or autoencoder compression to extract the feature subspace that best reflects the process evolution trend. Logical mapping is then performed by combining the location and physical meaning of each sensor source to construct a "mica state vector" containing state description quantities, state change rates, and control sensitivity factors. This state vector reflects the instantaneous operating status of the process system under the current operating conditions and is the core input basis for subsequent action decisions. On this basis, a crystal nucleus-mica melt sampling method is introduced to establish an optimal control action proposal mechanism. That is, under the conditions of historical quality distribution and current mixed quality prediction, a multidimensional normal distribution space is constructed. This space is centered on the state vector and has the prediction error covariance matrix as the distribution structure. A series of feasible control action sets are generated using a Gaussian random variable sampling mechanism. During the sampling process, inefficient samples are excluded by introducing dynamic weight correction and rejection sampling mechanisms to ensure that the generated actions are actually feasible within the process physical boundaries and equipment constraints. Furthermore, a heuristic function based on Expected Improvement (EI) or Lower Confidence Bound (LCB) is introduced to sort the samples according to their expected improvement in prediction quality and select action proposals that represent the optimal strategy.

[0026] Preferably, step S1 includes the following steps:

[0027] Step S11: Deploy a K-type thermocouple with a range of 0-1600℃, an accuracy of ±1.5℃, and a sampling frequency of 10Hz on the inner wall of the melting furnace to collect melting temperature gradient data;

[0028] Step S12: Deploy a laser particle size analyzer with a measurement range of 0.1-1000μm, a resolution of 0.1μm, and a sampling frequency of 1Hz in the crystal growth region to collect time-series data on crystal size distribution;

[0029] Step S13: Deploy an infrared thermal imager with a frame rate of 30fps in the cooling section to collect data on the cooling stress distribution of the mica sheet;

[0030] Step S14: Use wavelet threshold denoising algorithm to denoise the melting temperature gradient data, crystal size distribution time series data and mica sheet cooling stress distribution data, and construct a preliminary industrial dataset of mica.

[0031] In this embodiment of the invention, K-type thermocouples are arranged on the inner wall of the melting furnace to dynamically sample the melting temperature field under high-temperature conditions. These thermocouples have a wide measurement range of 0–1600℃ and a temperature measurement accuracy of ±1.5℃, recording continuously at a frequency of 10Hz. The collected results are presented as time-series melting temperature gradient data, reflecting the thermal field change pattern in the synthesis region. Next, a laser particle size analyzer is deployed in the crystal growth region. This device has a measurement range covering 0.1–1000μm, a resolution of 0.1μm, and a sampling frequency of 1Hz. It can acquire real-time particle size distribution data of crystal particle size evolution over time, forming a continuous particle size sequence and frequency relationship pair, revealing the dynamic process of crystal nucleation and growth. The third part involves deploying a 30 frames per second infrared thermal imager in the cooling section to acquire high-frame-rate thermal images of the stress thermal field distribution of mica sheets during the cooling process, obtaining an infrared spectrum sequence with spatial dimensions. This sequence can be overlaid over time to construct a dynamic distribution tensor of the sheet cooling stress. Since the three types of data originate from different physical mechanisms and sampling frequencies, their original signals all contain sensor noise, environmental disturbances, and system errors. To improve data quality, a wavelet threshold denoising algorithm is uniformly used for preprocessing. Specifically, for melt temperature gradient data and crystal grain size sequence data, multi-scale wavelet decomposition is used, and a soft threshold function is applied to each frequency band to filter out high-frequency interference signals. For infrared thermal image sequences, a two-dimensional wavelet transform is performed on the pixel intensity of each frame of the image to extract the low-frequency approximation and high-frequency disturbance details, and then a dynamic threshold is set based on the local mean of the thermal distribution, followed by denoising and reconstruction. After denoising, the three types of data are archived under a unified time base and structured format, and the original industrial dataset "Mica Preliminary Industrial Dataset" is constructed using process segment and sampling type as index metadata. This dataset has the characteristics of temporal integrity, spatial structure consistency, and signal quality stability, providing standardized input for subsequent feature extraction and modeling.

[0032] Preferably, step S2 includes the following steps:

[0033] Step S21: Use the cloud platform to align the mica preliminary industrial dataset with a time axis of a maximum alignment error of 50ms to obtain mica industrial aligned data;

[0034] Step S22: Extract and calculate the statistical and frequency domain features of mica synthesis from the industrial alignment data of mica, and fit the distribution parameters to obtain the spatial gradient feature data of mica.

[0035] Step S23: Extract spatial gradient features from the mica spatial gradient feature data and perform tensor concatenation to obtain the mica generation feature matrix.

[0036] In this embodiment of the invention, for heterogeneous time-series data from different sensors concentrated in the initial industrial dataset of mica, a multi-channel time index table is constructed using the unified timestamp management mechanism of the cloud platform, with a maximum allowable alignment error of 50ms. Through methods such as time window sliding matching, interpolation correction, and hysteresis compensation, the time axes of different signal sources are aligned to ensure strict synchronization of each data channel at key process time points, thereby forming "mica industrial aligned data" with time consistency. Based on this, the feature extraction stage proceeds, calculating statistical features for each channel of the aligned data, including indicators describing central tendency and distribution structure such as mean, variance, skewness, kurtosis, range, zero crossover rate, and fluctuation factor. Simultaneously, frequency domain analysis methods based on FFT and power spectral density (PSD) are introduced to extract features such as frequency peak, spectral width, and dominant frequency energy percentage to capture periodic, oscillatory, and non-steady-state characteristics in the data. Subsequently, based on the extracted statistical and frequency domain features, maximum likelihood estimation or the method of moments is used to fit the distribution of these features over time segments, prioritizing typical industrial signal probability models such as Gaussian, log-normal, or Cauchy distributions to obtain the distribution parameter matrix corresponding to each sampling point. Next, various feature data with clearly defined spatial locations are mapped into a three-dimensional physical structure space. A spatial gradient structure is constructed based on the geometric arrangement of the sampling points in the process flow. By calculating the gradient transformation rate, cooperative change direction, and anomalous offset between spatially adjacent points, a set of spatially resolvable "mica spatial gradient feature data" is formed. Finally, this feature dataset is tensor-quantized and stitched together according to time, space, and feature dimensions. Standardized padding, multidimensional coordinate encoding, and structural mapping methods are used to construct a unified "mica generation feature matrix."

[0037] Preferably, step S22 includes the following steps:

[0038] Step S221: Calculate the mean, variance, skewness, and kurtosis of the mica industrial alignment data according to a window length of 60s to obtain the statistical feature vector of the melting temperature.

[0039] Step S222: Perform a fast Fourier transform on the mica sheet cooling stress distribution data of the mica industrial alignment data, and obtain the pressure frequency domain feature vector by dividing the gradient frequency band into 32 bins and the spectral entropy.

[0040] Step S223: Perform spatial gradient calculation of the infrared hot-line sequence based on the statistical feature vector of melting temperature and the frequency domain feature vector of pressure, and construct the distribution parameters for density fitting by combining the time series data of crystal size distribution, to obtain the spatial gradient feature data of mica.

[0041] In this embodiment of the invention, for the melting temperature gradient data, a sliding window slice process is performed according to a fixed window length of 60 seconds. Within each time window, its first-order, second-order, and higher-order statistical characteristic indices are calculated, namely mean, variance, skewness, and kurtosis. These indices reflect the central trend, dispersion, asymmetric distribution, and peak characteristics of temperature fluctuations within that time period. Finally, a sequence of melting temperature statistical characteristic vectors is constructed using the window number as an index. In the frequency domain processing stage, for the mica sheet cooling stress distribution data, a Fast Fourier Transform (FFT) is first performed to convert the time-domain signal into a spectral form to reveal its frequency component composition and energy distribution characteristics. Subsequently, the spectral energy is binned according to a gradient frequency band binning strategy, with a bin number of 32. The entropy value of the spectral energy distribution, i.e., spectral entropy, is calculated to quantify the complexity and uncertainty of the signal's frequency domain structure. Finally, a pressure frequency domain feature vector is output. Next, in the spatial gradient construction stage, based on the aforementioned statistical feature vector of melting temperature and the frequency domain feature vector of pressure, combined with the spatial coordinate data extracted from the infrared thermographic sequence, the temperature-stress coupling relationship between adjacent pixels or sampling points is differentially calculated using the first-order spatial gradient calculation formula to form a temperature-stress joint gradient tensor. This tensor is then standardized and subjected to boundary constraints to enhance spatial consistency. Furthermore, time-series data of crystal size distribution is introduced as a conditional variable. Kernel density estimation (KDE) or maximum likelihood estimation methods are used to fit the conditional probability density of the gradient data, constructing a joint distribution model of the temperature-stress-crystal three factors. Core parameters of the fitted distribution (such as position parameters, scale parameters, skewness parameters, etc.) are extracted, ultimately yielding "mica spatial gradient feature data" with probabilistic structural expressive power and spatial topological characteristics.

[0042] Preferably, step S23 includes the following steps:

[0043] Step S231: Perform polynomial feature expansion processing on the statistical feature vector of melting temperature in the mica spatial gradient feature data to obtain expanded melting feature data;

[0044] Step S232: Perform logarithmic transformation on the time series data of crystal size distribution in the mica spatial gradient feature data to obtain logarithmic distribution feature parameters;

[0045] Step S233: Perform topological mapping on the extended molten feature data and logarithmic distribution feature parameters according to the time axis and spatial axis, respectively, and use the pressure frequency domain feature vector to perform dimensional concatenation to obtain the mica generation feature matrix.

[0046] In this embodiment of the invention, a polynomial feature expansion process is performed on the statistical feature vector of melting temperature. Specifically, second- and third-order nonlinear combination terms are calculated based on the original statistics (such as mean, variance, skewness, and kurtosis). Nonlinear interaction information between features is introduced by constructing higher-order terms. This expansion process can be implemented using a polynomial kernel function or a feature combination function, ultimately resulting in "extended melting feature data" that is more expressive in the feature space. Next, in step S232, a logarithmic transformation is performed on the time-series data of crystal size distribution. The purpose is to reduce the long-tail effect of skewed distributions, making the original crystal grain size data, which has power-law or log-normal distribution characteristics, more symmetrical, thus making it more suitable for subsequent density function fitting and modeling analysis. Finally, "logarithmic distribution feature parameters," such as logarithmic mean and logarithmic standard deviation, are extracted. In step S233, the extended melting feature data and logarithmic distribution feature parameters are topologically mapped along the time and spatial axes. This mapping is based on the acquisition timestamps and spatial arrangement indices, transforming them into a structured matrix in a unified spatiotemporal coordinate system. This is typically achieved using tensor coordinate encoding, spatial heatmap interpolation, or sliding window stacking methods to ensure consistent index structure and arrangement order for different physical quantities within the same coordinate system. Subsequently, the pressure frequency domain feature vector is introduced as external channel information and appended along the feature dimension to the aforementioned mapping result during the stitching operation, ultimately achieving the fusion and stitching of cross-source features.

[0047] Preferably, step S3 includes the following steps:

[0048] Step S31: Based on the mica generation feature matrix, embed the crystal growth kinetic equations for kiln heating to obtain the basic framework of the mechanism model, wherein the crystal growth kinetic equations include:

[0049]

[0050] Where D is the crystal size, S is the supersaturation, RT is the temperature, t is the time, and k0 and E are also present. a And n are the parameters to be calibrated;

[0051] Step S32: Perform mica quality prediction on the mica spatial gradient feature data of the mica generation feature matrix to obtain the theoretical quality value of the automated production line;

[0052] Step S33: Perform parameter initialization based on the basic framework of the mechanistic model, wherein the initial parameters of the parameters to be calibrated are k0 = 0.15 and E... a With n = 8.314 and n = 1.2, an initial mechanism model is obtained, and the output of the initial mechanism model is obtained to obtain the output data of the mica calibration model.

[0053] Step S34: Perform weighted fusion of the output data of the mica calibration model and the theoretical quality value of the automated production line to obtain the mica mixed quality prediction value, wherein the weighted fusion ratio is 7:3.

[0054] In this embodiment of the invention, the constructed "mica generation feature matrix" is used as input and embedded into a crystal growth kinetics model, which employs the classic supersaturation-driven growth formula: This formula maps the causal relationship between thermodynamic driving forces and molecular attachment rates. Therefore, feature vectors highly correlated with variables such as S, T, and crystal size change rate need to be extracted from the feature matrix as model inputs or initial boundary conditions. Next, quality-label supervised regression is performed on the spatial gradient feature variables in the generated feature matrix. This involves training a lightweight regressor (such as linear regression or support vector regression) using historical label data to predict crystal quality values, thus achieving personalized calibration of the physical model. For example, k0 is set to 0.15, and E... a Given 8.314 and n = 1.2, and based on the time series structure provided by the generated feature matrix, the time derivative is calculated using the finite difference method or the least squares method. An estimate is made, and then the result is substituted into the formula and compared with the theoretical quality value of the automated production line. A parameter inversion process based on gradient descent or a genetic algorithm optimization process is then performed to correct the model parameters. Finally, in step S34, the theoretical quality value of the automated production line obtained based on the data prediction model is weighted and fused with the physical quality value derived from the modified dynamics model. The fusion method is weighted linear summation, and the weight ratio is set to 7:3, indicating that data-driven prediction plays a major role in the fusion.

[0055] Of particular importance, step S32 includes:

[0056] Based on the mica spatial gradient feature data generated from the mica feature matrix, a spatiotemporal feature slice sequence is obtained by slicing the mica spatial gradient feature data in the time dimension.

[0057] The spatiotemporal feature slice sequence is enhanced in the frequency domain and then compressed to obtain a compressed feature matrix.

[0058] Based on the compressed feature matrix, gated fusion processing for fusion anomalies is performed, and physical constraint correction is applied to obtain the theoretical quality value of the automated production line.

[0059] In this embodiment of the invention, based on the feature matrix generated from mica, the spatial gradient feature data of mica is extracted, and the data sequence is sliced ​​according to the time dimension to form a series of slice sequences with spatiotemporal feature attributes. Time-dimensional slicing divides continuous time-series data into multiple discrete time periods, capturing the dynamic evolution of mica features over time, thereby constructing a spatiotemporally coupled multidimensional data structure. Subsequently, the obtained spatiotemporal feature slice sequences undergo frequency domain enhancement processing, specifically including performing Fourier transform or wavelet transform on the spatiotemporal data sequence to transform it from the time domain to the frequency domain to highlight key frequency components and periodic features. By enhancing the frequency domain signal strength, important frequency information is amplified and interference components are suppressed. After frequency domain enhancement, a feature compression step is implemented, using dimensionality reduction algorithms such as Principal Component Analysis (PCA), Singular Value Decomposition (SVD), or autoencoders to compress the high-dimensional spatiotemporal feature data, reducing redundant information and noise interference, and extracting a representative, low-dimensional compressed feature matrix for subsequent calculation and storage. Based on this compressed feature matrix, a gating mechanism is applied for anomaly processing during fusion. Gated fusion, by introducing a weight adjustment unit, effectively identifies and suppresses anomalous data or noise interference. When fusing different time slices and spatial gradient features, it achieves adaptive filtering of anomalous information, enhancing the stability and accuracy of the fusion results. Subsequently, physical constraint correction is performed on the fusion results. Based on theoretical constraints related to mica production and physical properties, such as mass conservation and thermodynamic equilibrium, the fused data is corrected to ensure that the obtained data conforms to actual physical laws and eliminates unreasonable deviations. Finally, through the above processing steps, a mass value that meets theoretical expectations is obtained. This mass value, based on multi-level analysis and fusion correction of spatiotemporal characteristics, forms an accurate theoretical estimate of mica mass.

[0060] Of particular importance, step S34 includes:

[0061] The residual calculation process is performed on the output data of the mica calibration model and the theoretical quality value of the automated production line to obtain the prediction error of the mechanism model;

[0062] The theoretical prediction error data is obtained by performing residual calculations based on the theoretical quality values ​​of the automated production line.

[0063] The prediction errors of the mechanistic model and the theoretical prediction errors were statistically analyzed using a sliding window and then weighted and fused to obtain the predicted value of the mica mixing quality, with a weighting ratio of 7:3.

[0064] As an example of the present invention, reference is made to... Figure 2 The diagram shown illustrates the predicted mixing quality of mica.

[0065] In this embodiment of the invention, point-by-point residual calculation is performed based on the output data of the mica calibration model and the theoretical quality value of the automated production line. This involves subtracting the numerical differences between the two values ​​at the same sampling time or spatial location to obtain the prediction error sequence of the mechanistic model. This residual reflects the deviation and error characteristics of the mechanistic model in fitting or predicting mica quality. Simultaneously, to ensure the stability and consistency of the theoretical quality value of the automated production line itself, self-residual calculation is performed using its historical or reference data to obtain theoretical prediction error data. This process typically involves quantifying the difference between the theoretical quality value of the automated production line and its benchmark or expected reference value, reflecting the fluctuations in prediction error within the theoretical model. Subsequently, the mechanistic model prediction error sequence and the theoretical prediction error data are input into a sliding window statistical analysis module. By setting a fixed-length sliding window, the residual data is segmented and statistically analyzed within the time or spatial dimensions. The mean, variance, and other statistical quantities of the residuals within each window are calculated to capture the local variation trend and fluctuation characteristics of the error. This statistical analysis helps identify the time-varying nature of the error and abnormal fluctuation regions, providing data support for subsequent fusion processing. Based on this, a weighted fusion technique is applied to merge the sliding window statistical results. Specifically, according to a preset weight ratio of 7:3, the prediction errors of the mechanistic model and the theoretical prediction errors are weighted and combined proportionally to generate the final predicted value of the mica mixture mass. Weighted fusion, through linear weighting or other fusion algorithms, balances and adjusts the two types of error information, aiming to comprehensively utilize the physical advantages of the mechanistic model and the numerical prediction characteristics of the theoretical model to improve the overall accuracy and robustness of the prediction.

[0066] As an example of the present invention, reference is made to Figure 3 As shown, step S4 in this example includes:

[0067] Step S41: Perform state space construction processing on the mica spatial gradient feature data of the mica generating feature matrix to obtain the mica state vector;

[0068] Step S42: Based on the preliminary industrial dataset of mica, set the temperature action set and the cooling action set. Specifically, the preliminary industrial dataset of mica is set by discretizing the melting temperature gradient data by 10℃ to obtain the temperature action set; and by discretizing the cooling rate of the mica sheet cooling stress distribution data by 5℃ / min to obtain the cooling action set.

[0069] Step S43: Use the mica mixing quality prediction value to sample the crystal nucleus-mica melt to obtain mica automated production optimization data.

[0070] In this embodiment of the invention, "mica spatial gradient feature data" is extracted from the "mica generation feature matrix." This data already includes multi-source statistical features (such as the mean melting temperature, frequency domain entropy of cooling stress, crystal size distribution parameters, etc.) and a topological structure based on tensor structure mapping. To map these high-dimensional feature vectors into state representations usable for strategy generation, dimensionality reduction techniques such as Principal Component Analysis (PCA), t-SNE, and AutoEncoder are used to extract key variables to form a "mica state vector." This vector serves as an instantaneous state description of the production system, constituting the state space. To define the action space, the original data related to controllable variables in the "preliminary mica industrial dataset" needs to be quantified: the melting temperature gradient data is discretized at 10°C intervals to construct a temperature action set; the cooling rate on the time axis of the cooling stress distribution data is extracted and then segmented at 5°C / min to construct a cooling action set. The two are combined to form a joint action space. From the above action space, the action combination with the optimal response potential is selected, and strategy sampling is performed using the "predicted mica mixing quality value" as the objective function. Specifically, the Gaussian policy is used, which involves constructing a normal distribution probability model for the action space and then performing probability sampling or maximum a posteriori sampling based on this model to select the optimal action as "mica automated production optimization data".

[0071] Preferably, step S43 includes the following steps:

[0072] Based on the predicted quality of mica mixture, a multi-objective reward component is designed, and the L2 norm is calculated to obtain the mica stability penalty component.

[0073] By sampling the mica stability penalty component and the predicted mica mixing quality using crystal nuclei-mica melt, we can obtain optimized data for automated mica production.

[0074] In this embodiment of the invention, the error and deviation of the mica mixing quality prediction value are weighted to reflect the priority and constraint relationships of different objectives. Subsequently, for the multi-objective reward component, the L2 norm calculation method is introduced. By summing and square-taking the sum of squares of the reward component vectors, the overall deviation and fluctuation are measured. This process effectively reflects the stability and consistency level of the mica mixing quality prediction. Based on the L2 norm calculation results, a stability penalty component is formed. This component serves as a constraint indicator for the fluctuation of the mica mixing quality prediction error, aiming to control and limit the uncertainty and deviation trend in the prediction process. On this basis, the stability penalty component and the mica mixing quality prediction value are jointly processed, and data sampling is performed using the nucleus-mica melt sampling method. The nucleus-mica melt sampling method introduces random perturbations conforming to a Gaussian distribution to jointly sample the predicted value and the penalty component, generating a series of production optimization action proposals. This sampling process relies on a probability distribution model to continuously explore the sampling space, ensuring that the action proposals cover diversity and feasibility, while balancing production optimization objectives and quality stability constraints.

[0075] Preferably, the nucleus-mica melt sampling of the mica stability penalty component and the predicted mica mixing quality includes the following:

[0076] The mica stability penalty component is forward-propagated to obtain mica state value estimation data;

[0077] By using the predicted quality of mica mixtures to sample the crystal nuclei-mica melt, we can obtain optimized data for automated mica production.

[0078] Based on the mica state value estimation data, the mica automated production optimization data is re-optimized throughout the entire process, and a machine instruction set is constructed to obtain the mica automated production process control instruction set.

[0079] Based on the mica automated production process control instruction set, the operating equipment is simulated and a report is generated to obtain the mica automated production process optimization report.

[0080] In this embodiment of the invention, calculations are performed in a neural network or deep learning model via a forward propagation mechanism to obtain value estimation data of the mica state. This value estimation data is a quantitative assessment of the current mica production state, reflecting the comprehensive performance of the system under given penalty conditions. Subsequently, based on the predicted mica hybrid quality value, a nucleus-mica melt sampling method is used to explore the action space. By introducing random perturbations conforming to a Gaussian distribution around the predicted value, diverse and probabilistic production optimization action proposals are generated, aiming to cover potential optimization schemes. Next, these action proposals are re-optimized throughout the entire process using the obtained mica state value estimation data. The process includes performance evaluation, selection, and correction of the action sequence to ensure that the generated optimized actions not only meet the quality prediction target but also take into account system stability and production efficiency. Based on the action re-optimization, a machine instruction set is constructed. This instruction set encodes the optimized production actions in structured data form, covering specific operation commands such as equipment control instructions and process parameter adjustments, forming the process control foundation for automated mica production. Finally, based on the constructed automated production process control instruction set, the operation process of the actual operating equipment is simulated. Through simulation calculations and state feedback data collection, a mica automated production process optimization report is generated. This report systematically summarizes the changes, optimization effects, and potential problems of various indicators in the production process, forming structured data output to support subsequent analysis and decision-making.

[0081] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0082] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A cloud computing based synthetic mica automated production process optimization method, characterized by, Comprising the following steps: Step S1: Deploy a multi-source industrial sensor array on the synthetic mica production line to collect full-process data and build a preliminary industrial dataset for mica; Step S2: Use the cloud platform to extract statistical features and frequency domain features of the mica synthesis from the preliminary industrial dataset for mica, and fit the distribution parameters to obtain mica spatial gradient feature data; perform tensor splicing on the mica spatial gradient feature data to obtain a mica generated feature matrix; Step S2 comprises the following steps: Step S21: Use the cloud platform to align the time axis of the preliminary industrial dataset for mica with a maximum alignment error of 50ms to obtain industrial aligned data for mica; Step S22: Extract statistical features and frequency domain features of the mica synthesis from the industrial aligned data for mica, and fit the distribution parameters to obtain mica spatial gradient feature data; Step S22 comprises the following steps: Step S221: Calculate the mean, variance, skewness and kurtosis of the melt temperature gradient data according to a window length of 60s from the industrial aligned data for mica to obtain a melt temperature statistical feature vector; Step S222: Perform a fast Fourier transform on the mica flake layer cooling stress distribution data of the industrial aligned data for mica, and perform a spectral entropy with a gradient frequency band bin number of 32 to obtain a pressure frequency domain feature vector; Step S223: Perform infrared thermal line sequence spatial gradient operation based on the melt temperature statistical feature vector and the pressure frequency domain feature vector, and combine the crystal size distribution time series data to construct the distribution parameters of the density fitting to obtain the mica spatial gradient feature data; Step S23: Extract spatial gradient features from the mica spatial gradient feature data, and perform tensor splicing to obtain a mica generated feature matrix; Step S3: Perform embedding processing of the crystal growth kinetics equation of the kiln heating based on the mica generated feature matrix to obtain a mechanism model basic framework; perform quality evaluation of the raw material batch of the mica production line on the mica spatial gradient feature data of the mica generated feature matrix to obtain a theoretical quality value of the automated production line; perform fusion processing of the crystal purity based on the mechanism model basic framework and the theoretical quality value of the automated production line to obtain a mica mixed quality prediction value; Step S4: Perform state space construction processing on the mica spatial gradient feature data of the mica generated feature matrix to obtain a mica state vector; set up the preliminary industrial dataset for mica, and perform crystal nucleus-mica melt sampling using the mica mixed quality prediction value to obtain mica automated production optimization data.

2. The cloud computing based synthetic mica automated production process optimization method as claimed in claim 1, wherein, Step S1 comprises the following steps: Step S11: Deploy a K-type thermocouple with a range of 0-1600°C, an accuracy of ±1.5°C, and a sampling frequency of 10Hz on the inner wall of the melting furnace to collect melt temperature gradient data; Step S12: Deploy a laser particle size analyzer with a measurement range of 0.1-1000μm, a resolution of 0.1μm, and a sampling frequency of 1Hz in the crystal growth zone to collect crystal size distribution time series data; Step S13: Deploy a thermal imager with an infrared frame rate of 30fps in the cooling section to collect mica flake layer cooling stress distribution data; Step S14: The melt temperature gradient data, the crystal size distribution time series data, and the mica sheet layer cooling stress distribution data are denoised by using a wavelet threshold denoising algorithm, and a mica preliminary industrial data set is constructed.

3. The cloud computing based synthetic mica automated production process optimization method as claimed in claim 1, wherein, Step S23 includes the following steps: Step S231: The melt temperature statistical feature vector in the mica spatial gradient feature data is subjected to polynomial feature expansion processing to obtain expanded melt feature data; Step S232: The crystal size distribution time series data in the mica spatial gradient feature data are subjected to logarithmic transformation processing to obtain logarithmic distribution feature parameters; Step S233: The expanded melt feature data and the logarithmic distribution feature parameters are respectively subjected to topological mapping according to the time axis and the space axis direction, and are subjected to dimension splicing by using the pressure frequency domain feature vector to obtain a mica generation feature matrix.

4. The cloud computing based synthetic mica automated production process optimization method as claimed in claim 1, wherein, Step S3 includes the following steps: Step S31: A crystal growth kinetics equation embedding process is performed on the mica generation feature matrix based on the heating of the kiln to obtain a mechanism model basic framework, wherein the crystal growth kinetics equation includes: ； wherein, is the crystal size, is the supersaturation, is the temperature, is the time, , and is the parameter to be calibrated; Step S32: The mica quality is predicted based on the mica spatial gradient feature data of the mica generation feature matrix to obtain a theoretical quality value of the automatic production line; Step S33: parameter initialization processing is performed according to the mechanism model basic framework, wherein the initial parameters of the parameters to be calibrated are 0.15, 8.314, and 1.2, to obtain an initial mechanism model, and the initial mechanism model is used for output to obtain mica calibration model output data. Step S34: Gradient feature data fusion is performed according to the mica calibration model output data and the theoretical quality value of the automatic production line to obtain a mixed quality prediction value of the mica, wherein the weighting fusion ratio is 7:

3.

5. The cloud computing based synthetic mica automated production process optimization method as claimed in claim 1, wherein, Step S4 includes the following steps: Step S41: A state space construction process is performed on the mica spatial gradient feature data of the mica generation feature matrix to obtain a mica state vector; Step S42: A temperature action set and a cooling action set are obtained according to the mica preliminary industrial data set, wherein the setting of the mica preliminary industrial data set is specifically that the melt temperature gradient data is subjected to 10°C discretization processing to obtain the temperature action set, and the cooling rate of the mica sheet layer cooling stress distribution data is subjected to 5°C / min discretization processing to obtain the cooling action set; Step S43: Crystal nucleus-mica melt sampling is performed by using the mixed quality prediction value of the mica to obtain mica automatic production optimization data.

6. The cloud computing based synthetic mica automated production process optimization method as claimed in claim 5, wherein, Step S43 includes: A multi-objective reward component is designed according to the mixed quality prediction value of the mica, and an L2 norm calculation is performed to obtain a mica stability penalty component; Crystal nucleus-mica melt sampling is performed on the mica stability penalty component and the mixed quality prediction value of the mica to obtain mica automatic production optimization data.

7. The cloud computing based synthetic mica automated production process optimization method as claimed in claim 6, wherein, Crystal nucleus-mica melt sampling on the mica stability penalty component and the mixed quality prediction value of the mica includes: Forward propagation is performed on the mica stability penalty component to obtain mica state value estimation data; Crystal nucleus-mica melt sampling is performed by using the mixed quality prediction value of the mica to obtain mica automatic production optimization data; The mica automatic production optimization data is subjected to full-process re-optimization according to the mica state value estimation data, and a machine instruction set is constructed to obtain a mica automatic production process control instruction set; Based on the mica automatic production process control instruction set, the running equipment is simulated to run, and a report is generated to obtain a mica automatic production process optimization report.

8. A cloud computing based synthetic mica automated production process optimization system, characterized by, The application discloses a cloud-computing-based synthetic mica automatic production process optimization method and system. A data acquisition and construction module is configured to acquire full-process data by deploying a multi-source industrial sensor array on a synthetic mica production line and to construct a preliminary industrial data set of the synthetic mica. A feature extraction and feature matrix generation module is configured to extract statistical features and frequency domain features of the preliminary industrial data set of the synthetic mica by using a cloud platform, to fit distribution parameters, and to obtain synthetic mica spatial gradient feature data; and to obtain a synthetic mica generated feature matrix by tensor splicing the synthetic mica spatial gradient feature data. A mechanism modeling and quality prediction module is configured to embed a crystal growth kinetics equation of a kiln heating process based on the synthetic mica generated feature matrix, to obtain a mechanism model basic framework; to predict the quality of the synthetic mica based on the synthetic mica spatial gradient feature data of the synthetic mica generated feature matrix, to obtain a theoretical quality value of an automatic production line; and to fuse the mechanism model basic framework and the theoretical quality value of the automatic production line to obtain a synthetic mica mixed quality prediction value. An optimization strategy generation module is configured to construct a state space based on the synthetic mica spatial gradient feature data of the synthetic mica generated feature matrix, to obtain a synthetic mica state vector; to set the preliminary industrial data set of the synthetic mica; and to sample a crystal nucleus-synthetic mica melt by using the synthetic mica mixed quality prediction value, to obtain synthetic mica automatic production optimization data.

Images