Polysilicon production equipment control method and device, server and storage medium

Abstract

The application provides a polysilicon production equipment control method and device, a server and a storage medium. The polysilicon production equipment control method comprises the following steps: obtaining a target parameter; performing empirical mode decomposition (EMD) and principal component analysis on the target parameter to obtain a stationary parameter of the target parameter; obtaining a prediction sample of a principal component parameter of the target parameter and a prediction sample of a robust distance parameter of the target parameter; determining a prediction result of an equipment parameter of a target device and a process parameter of an upstream and a downstream of the target device according to the stationary parameter of the target parameter, the prediction sample of the principal component parameter of the target parameter and the prediction sample of the robust distance parameter of the target parameter; and adjusting the equipment parameter of the target device and / or the process parameter of the upstream and the downstream of the target device according to the prediction result. The application can improve the prediction accuracy of the equipment parameter and further realize accurate control of the target device and / or the upstream and the downstream of the target device.

Description

Technical Field

[0001] This invention relates to the field of equipment health management technology, and in particular to a control method, device, server, and storage medium for polysilicon production equipment. Background Technology

[0002] Equipment fault diagnosis and early warning are crucial in the large-scale chemical industry. This is not only to ensure the safe operation of equipment and timely maintenance under abnormal conditions, but also to guarantee the safe and stable operation of chemical processes. Integrating the condition monitoring of critical equipment with upstream and downstream processes facilitates a close link between equipment management and process control, thereby improving the equipment health management system and enhancing its reliability, usability, safety, and efficiency.

[0003] Currently, for equipment parameters, various industries generally build correlation models based on known historical data, and these models can be directly used for real-time monitoring of parameter updates. However, accurate prediction of future conditions is still not possible.

[0004] In polysilicon manufacturing processes, current predictions rely solely on rough lags and trends in end-stage samples to forecast future performance. While methods like multiple linear regression and principal component analysis for noise reduction can correct these biases to some extent, precise predictions are ultimately unattainable. Therefore, a comprehensive approach is needed to achieve accurate predictions over a specific future timeframe.

[0005] For future predictions, the Long Short-Term Memory (LSTM) model is the most suitable model. However, due to equipment failures, 82% of them are caused by the influence of upstream process conditions. Therefore, the relationship between equipment and process must be taken into account, and this relationship should be used to optimize single parameters.

[0006] Patent CN111784068A provides a solution for noise reduction of power load parameters using a combination of Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA). For each component, stable components are predicted using multiple linear regression, while unstable components are predicted using LSTM. This method allows for detailed analysis of data from each power load sampling point. However, fundamentally, each power load parameter remains an "information silo," meaning no connection is established between them. The multiple linear regression only establishes connections between components within each power load parameter, failing to explain the connections between different power load sampling points. Therefore, to be applicable to large-scale chemical processes such as polysilicon manufacturing, an improved solution is needed to meet specific requirements and enhance prediction accuracy. Summary of the Invention

[0007] This invention provides a method, apparatus, server, and storage medium for controlling polysilicon production equipment, in order to solve the problem of poor prediction accuracy of equipment parameters in the prior art.

[0008] To address the aforementioned technical problems, the embodiments of the present invention provide the following technical solutions.

[0009] This invention provides a method for controlling polysilicon production equipment, comprising:

[0010] Obtain target parameters, which include equipment parameters of the target device and upstream and downstream process parameters of the target device;

[0011] The target parameters are processed by empirical mode decomposition (EMD) and principal component analysis to obtain stationary parameters of the target parameters;

[0012] Obtain predicted samples of the principal component parameters of the target parameter and predicted samples of the robust distance parameter of the target parameter;

[0013] Based on the stationary parameters of the target parameters, the predicted samples of the principal component parameters of the target parameters, and the predicted samples of the robust distance parameters of the target parameters, the equipment parameters of the target equipment and the predicted results of the upstream and downstream process parameters of the target equipment are determined.

[0014] Based on the prediction results, the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment are adjusted.

[0015] Optionally, obtaining prediction samples of the principal component parameters of the target parameter and prediction samples of the robust distance parameter of the target parameter includes:

[0016] Based on the stationary parameters, the load matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equations, the correlation model between the target device and its upstream and downstream, the future prediction samples of the response parameters, and the future prediction samples of the principal parameters are obtained.

[0017] Based on the robust distance of the target parameter, the covariance matrix of the target parameter, the eigenvalue of the stationary parameter, the future prediction sample of the response parameter, and the future prediction sample of the principal parameter, the prediction sample of the robust distance parameter of the target parameter is obtained.

[0018] Based on the loading matrix of the stationary parameter, the eigenvalues ​​of the stationary parameter, the future prediction samples of the response parameter, the future prediction samples of the principal parameter, and the principal component equation, the prediction samples of the principal component parameters of the target parameter are obtained.

[0019] Wherein, the main parameters are the parameters among the target parameters that affect the target equipment and the upstream and downstream processes of the target equipment; the response parameters are the parameters among the target parameters other than the main parameters.

[0020] Optionally, based on the stationary parameters, the loading matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equations, the principal parameters in the target parameters, the response parameters in the target parameters, the correlation model between the target device and its upstream and downstream components, future prediction samples of the response parameters, and future prediction samples of the principal parameters are obtained, including:

[0021] Dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix and eigenvalues ​​of the stationary parameters.

[0022] Based on the loading matrix of the stationary parameters, the principal component equation is obtained, and based on the principal component equation, the principal parameters and the response parameters are obtained;

[0023] Based on the main parameters and the response parameters, construct an association model between the target device and its upstream and downstream partners;

[0024] Based on the residual components of the response parameters and the Long Short-Term Memory (LSTM) model, future prediction samples of the response parameters are obtained; the residual components of the response parameters are obtained by principal component analysis dimensionality reduction processing of the intrinsic modal component (IMF) and the residual components of the response parameters according to a first preset cumulative contribution rate threshold; the IMF and the residual components of the response parameters are obtained by EMD processing of the residual components of the response parameters.

[0025] Based on the future prediction samples of the response parameters and the correlation model, the future prediction samples of the main parameters are obtained.

[0026] Optionally, dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix of the stationary parameters and the eigenvalues ​​of the stationary parameters, including:

[0027] Based on the target process parameters and historical target process parameters, an augmented matrix is ​​constructed;

[0028] Based on the second preset cumulative contribution rate threshold, sparse principal component analysis is performed on the augmented matrix to obtain the loading matrix of the stationary parameter and the eigenvalues ​​of the stationary parameter.

[0029] Optionally, empirical mode decomposition (EMD) and principal component analysis are performed on the target parameters to obtain stationary parameters of the target parameters, including:

[0030] Interpolation and correlation processing are performed on the sample data of each target parameter to obtain the preprocessed target parameters.

[0031] The preprocessed target parameters are subjected to Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA) to obtain stationary parameters of the target parameters.

[0032] Optionally, interpolation processing is performed based on the sample data in each of the target parameters, including:

[0033] If the number of consecutively missing sample data in the first target parameter is within a first preset range, the consecutively missing sample data is interpolated and filled using linear interpolation.

[0034] If the number of consecutively missing sample data in the first target parameter is within the second preset range, the consecutively missing sample data is interpolated and filled using the third preset range of sample data adjacent to the consecutively missing sample data, the fourth preset range of sample data adjacent to the consecutively missing sample data, a preset bidirectional sliding data window algorithm, and a linear scaling algorithm.

[0035] If the number of consecutively missing sample data in the first target parameter is within the fifth preset range, the first target parameter is discarded.

[0036] Wherein, the first target parameter is any one of the target parameters;

[0037] The number of the first preset range is less than the number of the second preset range, and the number of the second preset range is less than the number of the fifth preset range.

[0038] Optionally, the preprocessed target parameters are subjected to Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA) to obtain stationary parameters of the target parameters, including:

[0039] EMD processing is performed on the sample data in the second target parameter to obtain the intrinsic modal component (IMF) and residual component of the second target parameter.

[0040] Based on the first preset cumulative contribution rate threshold, principal component analysis is performed on the IMF of the second target parameter and the residual components of the second target parameter to reduce the dimensionality and obtain the residual components of the second target parameter.

[0041] The remaining components of each second objective parameter are combined to obtain the stationary parameters of the second objective parameter;

[0042] Wherein, the second target parameter is any one of the preprocessed target parameters;

[0043] The residual component of the second target parameter is the component of the second target parameter minus the IMF of the second target parameter.

[0044] This invention also provides a control device for polysilicon production equipment, comprising:

[0045] The first acquisition module is used to acquire target parameters, which include equipment parameters of the target device and upstream and downstream process parameters of the target device;

[0046] The first processing module is used to perform Empirical Mode Decomposition (EMD) and Principal Component Analysis on the target parameters to obtain stationary parameters of the target parameters.

[0047] The second acquisition module is used to acquire prediction samples of the stationary parameters of the target parameters, prediction samples of the principal component parameters of the target parameters, and prediction samples of the robust distance parameters of the target parameters.

[0048] The determination module is used to determine the equipment parameters of the target equipment and the prediction results of the upstream and downstream process parameters of the target equipment based on the prediction samples of the stationary parameters of the target parameters, the prediction samples of the principal component parameters of the target parameters, and the prediction samples of the robust distance parameters of the target parameters.

[0049] The adjustment module is used to adjust the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment based on the prediction results.

[0050] This invention also provides a server, comprising: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps in the polysilicon production equipment control method as described above.

[0051] This invention also provides a readable storage medium storing a program, which, when executed by a processor, implements the steps in the polysilicon production equipment control method as described above.

[0052] The beneficial effects of this invention are:

[0053] The present invention obtains target parameters (including equipment parameters of the target device and upstream and downstream process parameters of the target device), performs EMD and principal component analysis on the target parameters to obtain stationary parameters of the target parameters, and obtains prediction results of equipment parameters of the target device and upstream and downstream process parameters of the target device based on prediction samples of stationary parameters, principal component parameters, and robust distance parameters of the target parameters. This improves the prediction accuracy of equipment parameters, and adjustments can be made to the equipment parameters of the target device and / or upstream and downstream process parameters based on the prediction results, thereby further achieving accurate control of the target device and / or its upstream and downstream components. Attached Figure Description

[0054] Figure 1 A flowchart illustrating the control method for polysilicon production equipment provided in an embodiment of the present invention;

[0055] Figure 2 This is a flowchart illustrating the Empirical Mode Decomposition (EMD) process performed on preprocessed target parameters according to an embodiment of the present invention.

[0056] Figure 3 A timing diagram showing the main parameter X18 and its parsed values ​​provided in the embodiments of the present invention;

[0057] Figure 4 A timing diagram showing the main parameter X111 and its parsed values ​​provided in the embodiments of the present invention;

[0058] Figure 5 A timing diagram showing the main parameter X25 and its resolved values ​​provided in the embodiments of the present invention;

[0059] Figure 6 A timing diagram showing the main parameter X29 and its parsed values ​​provided in the embodiments of the present invention;

[0060] Figure 7A timing diagram showing the main parameter X310 and its resolved values ​​provided in the embodiments of the present invention;

[0061] Figure 8 A timing diagram showing the main parameter X312 and its resolved values ​​provided in the embodiments of the present invention;

[0062] Figure 9 This represents a robust distance time series diagram of the full sample provided in the embodiments of the present invention;

[0063] Figure 10 This represents the principal component time series diagram of the full sample provided in the embodiments of the present invention;

[0064] Figure 11 This is a flowchart illustrating the specific process of controlling polysilicon production equipment according to an embodiment of the present invention.

[0065] Figure 12 This is a flowchart illustrating the specific process of obtaining prediction samples of principal component parameters and prediction samples of robust distance parameters of the target parameters according to an embodiment of the present invention.

[0066] Figure 13 This is a schematic diagram illustrating the structure of the device for future prediction based on weighted device master data provided in an embodiment of the present invention.

[0067] Figure 14 This is a schematic diagram of the structure of the control device for polysilicon production equipment provided in an embodiment of the present invention;

[0068] Figure 15 This is a schematic diagram of the server structure provided in an embodiment of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0071] To address the problem of poor prediction accuracy of equipment parameters in existing technologies, this invention provides a control method, apparatus, server, and storage medium for polysilicon production equipment.

[0072] like Figure 1 As shown, an embodiment of the present invention provides a method for controlling polysilicon production equipment, including:

[0073] Step 101: Obtain target parameters, which include the equipment parameters of the target device and the upstream and downstream process parameters of the target device.

[0074] In this embodiment, the target device is the reflux pump to be predicted. The upstream and downstream process parameters of the target device are the upstream and downstream process parameters of the reflux pump, and the target parameters are collected according to the location and sample data. The sample data is the set of values ​​of the device parameters and process parameters at a certain time point.

[0075] For example, taking the washing and dust removal process in the cold hydrogenation workshop as an example in this embodiment, this step analyzes the equipment parameters of the reflux pump to be predicted in the washing and dust removal process, as well as the process parameters at the upstream and downstream points of the reflux pump.

[0076] The equipment parameters can be the operating parameters of the reflux pump, such as current and amplitude, while the process parameters can be material-related parameters, such as flow rate, temperature, pressure, and liquid level. A schematic table of the equipment parameters of the reflux pump and the upstream and downstream process parameters of the reflux pump is shown in Table 1 below, including target parameters X1, X2, X3, ..., X12.

[0077] Table 1. Equipment parameters of the reflux pump and upstream and downstream process parameters of the reflux pump.

[0078]

[0079]

[0080] Step 102: Perform Empirical Mode Decomposition (EMD) and Principal Component Analysis on the target parameters to obtain the stationary parameters of the target parameters.

[0081] In this embodiment, to address the issue of significant fluctuations in all process and equipment parameters due to the unstable operation of the reflux pump and the entire washing and dust removal process, each target parameter is processed using a combination of EMD and principal component analysis to obtain the stationary parameter X of the target parameter. 1’ X 2’ X 3’ ...、X n’ .

[0082] Step 103: Obtain prediction samples of the stationary parameters of the target parameter, prediction samples of the principal component parameters of the target parameter, and prediction samples of the robust distance parameters of the target parameter.

[0083] The predicted sample of the robust distance parameter of the target parameter can also be called the predicted sample of the stability of the target parameter.

[0084] Step 104: Based on the predicted samples of the stationary parameters of the target parameters, the predicted samples of the principal component parameters of the target parameters, and the predicted samples of the robust distance parameters of the target parameters, determine the equipment parameters of the target equipment and the predicted results of the upstream and downstream process parameters of the target equipment.

[0085] In this step, the predicted samples of the stationary parameters of the target parameters, the predicted samples of the principal component parameters of the target parameters, and the predicted samples of the robust distance parameters of the target parameters are integrated and used together as the prediction results of the equipment parameters of the target equipment and the upstream and downstream process parameters of the target equipment.

[0086] Through the above steps 101-104, the future prediction of the equipment parameters of the target equipment and the upstream and downstream process parameters of the target equipment can be achieved with relatively high accuracy.

[0087] Step 105: Based on the prediction results, adjust the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment.

[0088] In this step, guidance is provided for adjusting the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment based on the changing trends of the above prediction results.

[0089] It should be noted that in the automatic control of chemical processes, process parameters exhibit time-series characteristics including lag, proportional, integral, and derivative properties. Therefore, the time-series characteristics of equipment parameters and process parameters need to be taken into account. In an optional embodiment of the present invention, step 102 includes:

[0090] To ensure the temporal integrity of the equipment parameters of the target device and the upstream and downstream process parameters of the target device, the data collection is based on factors such as missing samples and the degree of correlation between parameters (e.g., correlation R). 2 In cases where the value is greater than 0.8, preprocessing operations are performed on the target parameters. This involves interpolation and correlation processing based on the sample data of each target parameter to obtain the preprocessed target parameters, ensuring the temporal integrity of the target parameters.

[0091] In response to the situation where the unstable operation of the reflux pump and the entire washing and dust removal process causes large fluctuations in all process parameters and equipment parameters, empirical mode decomposition (EMD) is performed on the pre-processed target parameters to obtain stable parameters of the target parameters.

[0092] Furthermore, due to network fluctuations and other issues during data export, the sample data of the target parameters often contains some missing values ​​(represented by "*" in the exported file). To address this missing data, interpolation is performed based on the sample data for each target parameter to imputate the missing values, thereby ensuring the temporal integrity of each target parameter. This interpolation imputation based on the sample data for each target parameter includes:

[0093] If the number of consecutively missing sample data in the first target parameter is within the first preset range, the consecutively missing sample data is interpolated and filled using linear interpolation. Optionally, the first preset range is 1 to 10 consecutive sample data, that is, if 1 to 10 consecutive sample data (sample points) are missing in the first target parameter, the missing data is filled directly using linear interpolation.

[0094] When the number of consecutively missing sample data in the first target parameter is within the second preset range, the consecutively missing sample data is interpolated and filled using the third preset range of sample data adjacent to the consecutively missing sample data before it, the fourth preset range of sample data adjacent to the consecutively missing sample data after it, a preset bidirectional sliding data window algorithm, and a linear scaling algorithm. The first preset range is smaller than the second preset range. Optionally, the second preset range is 11-50 sample data, the third preset range is 5-10 sample data, and the fourth preset range is 5-10 sample data. That is, when 11-50 consecutively missing sample data are found in the first target parameter, the average of the 5-10 sample data adjacent to the missing sample data before it and the 5-10 sample data adjacent to the missing sample data after it is taken, and the predicted value of each missing sample data is obtained using a bidirectional sliding data window. The final missing data filling value is obtained using a linear scaling method and the predicted value.

[0095] If the number of consecutively missing sample data in the first target parameter is within the fifth preset range, the first target parameter is discarded. The second preset range is less than the fifth preset range. Optionally, the fifth preset range is greater than 50 sample data. That is, since the selected target parameters usually take into account the completeness of the sample data, if there are many missing sample data, the parameter will be deleted. Therefore, if the number of consecutively missing sample data in the first target parameter is greater than 50 sample data, the first target parameter is discarded or deleted.

[0096] Wherein, the first target parameter is any one of the target parameters, and for each of the target parameters, the sample data of each target parameter can be interpolated and filled in according to the above process.

[0097] It should be noted that the reason for maintaining the integrity of the time series is mainly for the purpose of future time series prediction.

[0098] Further, the preprocessed target parameters are subjected to Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA) to obtain stationary parameters of the target parameters, including:

[0099] The sample data in the second target parameter are subjected to EMD processing to obtain the intrinsic mode functions (IMF) and residual components of the second target parameter. That is, taking each preprocessed target parameter as a unit, EMD decomposition is performed on each preprocessed target parameter (i.e., the second target parameter) to obtain several IMFs and several residual components. Among them, one target parameter corresponds to one IMF and one residual component.

[0100] Based on the first preset cumulative contribution rate threshold, principal component analysis is performed on the IMF of the second target parameter and the residual components of the second target parameter to reduce the dimensionality, thereby obtaining the residual components of the second target parameter. The first preset cumulative contribution rate threshold is 90%. The residual components of each second target parameter are merged to obtain the stationary parameters of the second target parameter. That is, principal component analysis is performed on the IMF of the second target parameter and the residual components of the second target parameter respectively. Under the condition of the first preset cumulative contribution rate threshold of 90%, the interference components are removed and the residual components are merged to obtain the stationary parameters of the second target parameter.

[0101] The second target parameter is any one of the preprocessed target parameters. For each target parameter, EMD processing can be performed according to the above process to obtain a stationary parameter of the target parameter.

[0102] Specifically, such as Figure 2 As shown, the specific steps for performing Empirical Mode Decomposition (EMD) on the preprocessed target parameters are as follows:

[0103] Step 201: Obtain the target parameters X1, X2, X3..., Xn;

[0104] Step 202: Calculate the IMF components of each target parameter as IMF. i1 IMF j2 IMF k3 ..., IMF qn (Where, suffixes i, j, k, q, etc., represent a specific intrinsic modal component of a target parameter, and suffixes 1, 2, 3, ..., n belong to the target parameters of each mode.) In addition to the IMF, each target parameter also has a residual component, denoted as Res1, Res2, Res3, ..., Res... n When the number of objective parameters is 12, n = 12.

[0105] Step 203: Perform principal component analysis (PCA) dimensionality reduction on all IMFs and residual components of each target parameter, taking each parameter as a unit, to remove (reduce) interference components;

[0106] Step 204: Obtain the corresponding reduced IMF i1 The reduced IMF j2 The reduced IMF k3 ... the reduced IMF qn .

[0107] The measure to reduce the number of components is not to directly merge them into a number of fewer principal components through principal component analysis, but to determine the number of principal components for each target parameter through principal component dimensionality reduction analysis, and to reduce the number of IMF components according to the order of IMF components, so that the number of remaining components for each target parameter is equal to the number of principal components of the target parameter.

[0108] Step 205: Taking each parameter in the target parameters as a unit, combine the remaining components of each parameter to obtain the stationary parameter of the target parameters, denoted as X. 1’ X 2’ X 3’ ...、X n’ .

[0109] It should be noted that the residual component of the second target parameter is the remaining part of the second target parameter after removing the IMF of the second target parameter. The residual component of the second target parameter is the remaining component (including the residual component) after removing the interference component from the second target parameter.

[0110] After obtaining the residual components of each parameter in the target parameters, the residual components and residual components of each parameter are retained separately for subsequent applications. This separate retention is necessary because the residual components of some parameters will be used for future time series prediction.

[0111] The Empirical Mode Decomposition (EMD) process applied to the preprocessed target parameters can solve the noise reduction problem once and for all (it only requires reducing the first 1-3 IMFs based on the cumulative contribution rate threshold of principal component analysis, and the components that are filtered out are the interference components). A cumulative contribution rate threshold of 90% is based on experimental results, ensuring that the data remains undistorted.

[0112] In this process, principal component analysis (PCA) is performed on a per-parameter basis for dimensionality reduction. It does not involve combining all components of all parameters into a single set for dimensionality reduction, nor does it directly use the reduced principal components as the final reduced components. For example, when performing EMD decomposition and dimensionality reduction on preprocessed X1 sample data, and then performing EMD decomposition on preprocessed X2 sample data, dimensionality reduction is only applied to all components of preprocessed X2, not by adding preprocessed X2 components. Interference components are removed progressively, starting with IMF1 components, until the total number of remaining components equals the number of reduced principal components. That is, the interference components are the first 1-3 components of each parameter.

[0113] By performing Empirical Mode Decomposition (EMD) on the preprocessed target parameters as described above, the problems of difficult noise reduction, high subjectivity, and large workload can be solved, and data noise reduction can be achieved to construct healthy data.

[0114] Based on the future prediction method using full equipment master data, this invention provides the following model building and future time series prediction method based on parameter operating condition interaction and dynamic characteristics. Specifically, in an optional embodiment of this invention, obtaining prediction samples of stationary parameters of the target parameter, prediction samples of principal component parameters of the target parameter, and prediction samples of robust distance parameters of the target parameter includes:

[0115] Based on the stationary parameters, the load matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equations, the correlation model between the target device and its upstream and downstream, the future prediction samples of the response parameters, and the future prediction samples of the principal parameters are obtained.

[0116] Based on the robust distance of the target parameter, the covariance matrix of the target parameter, the eigenvalue of the stationary parameter, the future prediction sample of the response parameter, and the future prediction sample of the principal parameter, the prediction sample of the robust distance parameter of the target parameter is obtained.

[0117] The robust distance of the target parameter is an optimized model of Mahalanobis distance, which is an optimized model based on the fast minimum covariance determinant algorithm. In this embodiment, the calculation formula of the robust distance is equivalent to the calculation formula of the Mahalanobis distance:

[0118]

[0119] Where, x m Let S be the m-th sample in the stationary parameters of the target parameter, S be the covariance matrix, T be the mean matrix, T represent matrix transpose, and n represent the number of historical data samples. In other words, the robust distance is obtained based on the basis of historical data of n samples.

[0120] After obtaining the robust distance of the target parameter, the covariance matrix S of the target parameter, and the mean matrix T of the target parameter using the above formula, the predicted sample of the robust distance parameter of the target parameter (future prediction result) is obtained based on the robust distance of the target parameter, the covariance matrix S of the target parameter, the mean matrix T of the target parameter, the future prediction sample of the response parameter, and the future prediction sample of the main parameter. The predicted sample of this stable parameter is then used as the future prediction result of the overall stability of the target equipment and its upstream and downstream processes.

[0121] The steps described above for obtaining the predicted samples of the robust distance parameters of the target parameters are key steps in the overall stability analysis.

[0122] Based on the loading matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the future prediction samples of the response parameters, the future prediction samples of the principal parameters, and the principal component equation, the predicted samples of the principal component parameters of the target parameter are obtained. Specifically, using the loading matrix and eigenvalues ​​of the stationary parameters, the future prediction samples of the principal parameters and the future prediction samples of the response parameters of the target parameter are substituted into the principal component equation to calculate the principal component parameters, which are then used as the predicted samples (future prediction results) of the principal component parameters of the target parameter. The specific operation involves using X... fu P = T fu Determine the principal component values ​​of the future prediction samples, where X fu T is the original sample matrix containing future predicted samples. fu It contains the principal component values ​​of the future prediction samples obtained from the analysis, where P represents the loading matrix;

[0123] The steps described above for obtaining the predicted samples of the principal component parameters of the target parameters indirectly enable the future prediction of the principal component parameters of the target equipment and its upstream and downstream components.

[0124] Wherein, the main parameters are the parameters among the target parameters that affect the target equipment and the upstream and downstream processes of the target equipment; the response parameters are the parameters among the target parameters other than the main parameters.

[0125] Further, based on the stationary parameters, the loading matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equations, the correlation model between the target device and its upstream and downstream components, the future prediction samples of the response parameters, and the future prediction samples of the principal parameters are obtained, including:

[0126] Dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix and eigenvalues ​​of the stationary parameters; the principal component equation is obtained based on the loading matrix of the stationary parameters, and the principal parameters and response parameters are obtained based on the principal component equation.

[0127] It should be noted that dynamic sparse principal component analysis combines dynamic principal component analysis and sparse principal component analysis. The former takes into account the dynamic characteristics of the data by establishing an augmented matrix; the latter not only achieves centralized analysis of important parameters but also reduces the amount of data processing by concentrating and sparsifying the loading matrix and reducing non-zero elements.

[0128] In this embodiment, dynamic sparse principal component analysis is performed on stationary parameters to obtain the load matrix and eigenvalues ​​of the analysis case. The principal component equation is then obtained based on the load matrix. The sample parameters existing in the principal component equation are used to determine the main parameters affecting the target equipment and upstream and downstream processes, as well as the response parameters other than the main parameters.

[0129] Among them, the loading matrix, score matrix, and eigenvalues ​​are important results obtained from principal component analysis. The loading matrix reflects the importance of the corresponding parameters, the score matrix reflects the result obtained after dimensionality reduction, and the eigenvalues ​​reflect the importance of each principal component to the principal component analysis model. Specifically, the method for determining the main parameters is as follows: Based on the principal component equations obtained from the loading matrix and score matrix, the following equations are used:

[0130]

[0131] Among them, the above a ij x is an element in the load matrix. j Let PC be the stationary parameter of the j-th index, n be the number of all stationary parameters, and PC be the stationary parameter of the j-th index. iLet be the parameter of the i-th principal component, and k be the total number of principal component parameters. Finally, determine the PC of each principal component parameter. i The largest |a in the expression ij | Corresponding parameter x j These are the principal parameters. However, in the context of dynamic sparse principal component analysis, the method for determining the principal parameters differs slightly from that of principal component analysis: values ​​of 'a' with a value of 0 are selected. ij The corresponding stationary parameter x j It is directly used as the main parameter.

[0132] Based on the main parameters and the response parameters, a correlation model is constructed between the target device and its upstream and downstream components. Specifically, a multiple linear regression model is constructed between the main parameters and the response parameters, using the main parameters as units. The model is then filtered based on significance and goodness of fit to obtain the correlation model between the target device and its upstream and downstream components. The main parameters, response parameters, and correlation model are then backed up. The reason for backing up the correlation model separately is that it can also be used to analyze important parameters in the target parameters. Specifically, it allows us to examine the degree to which the important parameter deviates from the analytical value and analyze the problems that occur at the corresponding points of the important parameter.

[0133] The steps described above for establishing the correlation model are crucial for establishing the connection between the target equipment and its upstream and downstream parameters (i.e., the coupling between equipment and process conditions). The multiple linear regression model is obtained by linearly fitting the principal parameters to the response parameters. Screening is performed based on significance and goodness of fit. Significance is determined by the t-distribution test value of the coefficient analysis and the F-distribution test value of the variance analysis (the corresponding P-value (i.e., the probability value of the corresponding t-distribution and F-distribution) is less than a first preset threshold, which is considered significant; optionally, the first preset threshold is 0.05). The goodness of fit is determined by the percentage of the multiple linear regression model error to the total error, R-Sq. An R-Sq greater than a second preset threshold is considered a good fit; optionally, the second preset threshold is 0.8. If neither the significance judgment nor the goodness of fit analysis is satisfied, the corresponding multiple linear regression model is rejected. If all multiple linear regression models are invalid, the target parameters are re-obtained, i.e., the process restarts from step 101.

[0134] Furthermore, since this multiple linear regression model is built on the basis of an augmented matrix, it also includes the dynamic characteristics of the parameters themselves. In the embodiment where the target device is a reflux pump, a total of 6 multiple linear regression models are established (corresponding to the main parameters X18 temperature 1, X111 flow control 3, X25 pump current, X29 temperature 5, X310 flow control 1, and X312 flow control 4, respectively, where the subscript prefix points to the lag factor, that is, the subscript prefix 1 points to the parameter with no lag, the subscript prefix 2 points to the parameter with a lag factor of 1, and the subscript prefix 3 points to the parameter with a lag factor of 2), and all of them meet the requirements of significance and good fit, and together serve as the correlation model for the future prediction of the target parameters.

[0135] Based on the residual components of the response parameters and the Long Short-Term Memory (LSTM) model, future prediction samples of the response parameters are obtained. The residual components of the response parameters are obtained by principal component analysis (PCA) dimensionality reduction processing of the intrinsic modal components (IMF) and residual components of the response parameters according to a first preset cumulative contribution rate threshold. The IMF and residual components of the response parameters are obtained by EMD processing of the residual components of the response parameters. Specifically, LSTM time-series prediction is performed on all residual components of the response parameters to obtain future prediction samples of all residual components of the response parameters. Then, the future prediction samples of all residual components are merged for each response parameter to obtain future prediction samples of all response parameters. The process of determining the residual components of the response parameters is consistent with the steps of performing EMD processing on the second target parameter described above, and will not be repeated here.

[0136] Based on the future prediction samples of the response parameters and the correlation model, the future prediction samples of the main parameters are obtained, that is, the future prediction samples of the response parameters are substituted into the above correlation model to obtain the future prediction samples of the main parameters.

[0137] The LSTM model is obtained using the training and test sets of the residual components of the response parameters. Specifically, based on the fitted values ​​of the training and test sets of the residual components of the response parameters, the future value of each residual component of the response parameters is obtained by using a sliding data window. The future values ​​of the obtained residual components are then merged to obtain the future prediction samples of the response parameters.

[0138] The ratio of the training set to the test set can be determined by the user. Optionally, the ratio is 8:2. Future prediction samples for each remaining component of the response parameters are not included in the test set parsing samples.

[0139] Specifically, due to the high randomness of the LSTM model fitting results, this patent employs a method of repeatedly selecting the optimal target index. Three metrics are used to evaluate the LSTM model: MSE (mean squared error), MAE (mean absolute error), and R-value (goodness of fit). 2 Because under the same training samples, the MSE, MAE, and R-values ​​of different models built repeatedly vary. 2 All three are different, and their trends are correlated; therefore, the goodness of fit R is used here. 2 As the target indicator, select the largest R 2 The LSTM model corresponding to the value.

[0140] It should be noted that the LSTM model is a special type of neural network model that can solve the gradient vanishing and gradient exploding problems during long sequence training. The above steps predict the future of each IMF component and residual component, which can reduce the error compared to predicting the future of all parameters as a whole.

[0141] Further, dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix of the stationary parameters and the eigenvalues ​​of the stationary parameters, including:

[0142] Based on the stationary parameters and historical stationary parameters, construct an augmented matrix;

[0143] Based on the second preset cumulative contribution rate threshold, sparse principal component analysis is performed on the augmented matrix to obtain the loading matrix of the stationary parameter and the eigenvalues ​​of the stationary parameter.

[0144] Specifically, after obtaining the aforementioned stationary parameter X 1’ X 2’ X 3’ ...、X n’ Then, parameter samples with a time shift l lag (l is the lag factor, typically 2) are incorporated into it to obtain the augmented matrix of the stationary parameters of the target parameter. Constructing the augmented matrix involves building an augmented matrix of historical data and supplementing the corresponding regions with samples shifted forward l steps (l is the lag factor) (i.e., historical stationary parameters). Specifically, the augmented matrix is:

[0145] X = [X] t ,X t-1 ,X t-2 ,...,X t-l ]

[0146] Among them, X t For stationary parameter samples from time t to t+n, determining the lag factor of l is quite complex, but based on actual process experience, the lag factor of l is taken as 2.

[0147] Optionally, the second preset cumulative contribution rate threshold is 85%.

[0148] Sparse principal component analysis is used to reduce the dimensionality of the augmented matrix of the stationary parameters of the target parameter, requiring a cumulative contribution rate of 85%.

[0149] Dynamic sparse principal component analysis combines dynamic principal component analysis with sparse principal component analysis. The former takes into account the dynamic characteristics of the data by establishing an augmented matrix, while the latter reduces non-zero elements by concentrating and sparsifying the loading matrix, thereby not only enabling concentrated analysis of important parameters but also reducing the amount of data processing.

[0150] The sparse principal component analysis dimensionality reduction described herein is based on principal component analysis dimensionality reduction. It utilizes ridge regression and Lasso regression to sparsify the loading matrix, resulting in a new loading matrix. This minimizes the number of non-zero elements in each column of the principal component vector while accurately extracting data information. This can be understood as centralizing the parameters.

[0151] The load matrix and eigenvalues ​​of the stationary parameters of the target parameters obtained from the above sparse principal component analysis dimensionality reduction are used as backups.

[0152] The loading matrix is ​​the loading matrix processed by ridge regression and Lasso regression. In addition, the mean and variance obtained from the augmented matrix are also used as backups.

[0153] Specifically, during the construction of the augmented matrix, the lag factor l was set to 2, and the load matrix sparsification was performed using ridge regression and Lasso regression. In the embodiment of the circulating pump A, the original parameters were 12, and after augmented matrix processing, 36 parameters were involved. After dimensionality reduction (cumulative contribution rate threshold 85%), there were 5 principal component parameters.

[0154] The dynamic sparse principal component analysis described in this embodiment involves performing dynamic sparse principal component analysis on the stationary parameters, determining the principal parameters through the loading matrix, and the dynamic sparse principal component analysis method takes into account both the dynamic characteristics of dynamic principal component analysis and the sparse principal component analysis methods such as ridge regression and Lasso regression on the loading matrix.

[0155] Specifically, the method for determining the principal parameters is not to find the maximum absolute value of the corresponding load value in each principal component equation, but to directly collect the parameters corresponding to each non-zero load element, integrate them together, remove duplicate parameters, and summarize the final principal parameters.

[0156] In this embodiment of the invention, based on the detailed prediction of each parameter, the relationship between each parameter is established to achieve the coupling of equipment and process conditions. Considering the dynamic characteristics of the parameters, the dynamic relationship between them in time series is established, thereby reducing the amount of data processing during the analysis process.

[0157] This invention also provides a specific embodiment, the details of which are as follows:

[0158] Twelve target parameters were selected for a circulating pump (reflux pump) A in the cold hydrogenation process: X1 Pressure 1, X2 Pressure 2, X3 Pressure 3, X4 Pressure 4, X5 Pump A Current, X6 Liquid Level 4, X7 Liquid Level 5, X8 Temperature 1, X9 Temperature 5, X10 Flow Control 1, X11 Flow Control 3, and X12 Flow Control 4. A total of 1032 samples were selected from April 20, 2022 to June 1, 2022, at hourly intervals. Due to a fault between April 25th and 28th, fluctuations occurred, necessitating noise reduction. After EMD decomposition and Principal Component Analysis (PCA) dimensionality reduction to remove components (cumulative contribution rate 90%, this threshold was experimentally obtained; when the cumulative contribution rate threshold is below 90%, historical data is not completely optimized into healthy sample data), optimized parameters (preprocessed target parameters) were obtained. The processing procedure is shown in Table 2 below. (The effect of noise reduction will be clearly demonstrated in the subsequent robust distance analysis).

[0159] Table 2. Processing Procedure Diagram

[0160]

[0161] Based on the above parameters, dynamic sparse principal component analysis for dimensionality reduction is then performed:

[0162] First, construct the augmented matrix. Specifically:

[0163] The lag factor l is set to 2. Parameter X 1i Sample selection X 1i The 3rd to 1032nd sample; parameter X 2i Sample selection X i The 2nd to 1031st sample; parameter X 3i Sample selection X i The first 1030 samples (the prefixes 1, 2, and 3 of the above parameters refer to the parameters of the 1st, 2nd, and 3rd lag factors, respectively (the 1st has no lag), and i represents the index of the original parameters). Together, they construct the augmented matrix:

[0164] [X 11 X 12 ...X 111 X 112 X 21 X22 ...X 211 X 212 X 31 X 32 ...X 311 X 312 ]

[0165] Secondly, based on the augmented matrix described above, sparse principal component analysis is performed for dimensionality reduction. Specifically, principal component analysis is first conducted, and the principal component parameters are determined to be 5 when the cumulative contribution rate is greater than 85%, corresponding to 5 principal component equations. Based on the distribution of non-zero elements in the loading matrix, the main parameters are summarized as shown in Table 3 below:

[0166] Table 3. Schematic diagram of main components

[0167] PC1 X210, X212, X310, X312 PC2 X15, X25, X35 PC3 X18, X19, X111, X28, X29, X38, X39 PC4 X18, X110, X112, X210, X212 PC5 X18, X110, X111, X28, X210, X211, X38, X310, X311 Total of main parameters X18, X111, X25, X29, X310, X312

[0168] Based on the above six main parameters, a multiple linear regression model was performed, and the results (showing the case with only one main parameter) are as follows:

[0169] X18=2.34-2.4X11+5.35X12-5.7X13+4.08X14+0.09X15+0.0038X16-0.137X17+0.6850X19+0.1371 X110+0.1000X111+0.0103X112+9.5X21-11.2X22+1.0X23-0.77X24-0.16X25+0.0173X26+0.205X2 7+0.1846X28-0.2837X29-0.1709X210-0.2226X211-0.0530X212+2.8X31+7.24X32-4.8X33-1.80X 34+0.07X35-0.0421X36-0.055X37+0.2337X38+0.0133X39+0.1247X310+0.1443X311+0.0354X312

[0170] The R-squared value of the above equation is 97.82%, and the p-values ​​corresponding to the t-test and F-test values ​​of each parameter are all less than 0.05.

[0171] Based on the R-sq test results calculated from the actual six regression equations, the linear regression model with the above six main parameters was selected as a backup.

[0172] For the remaining components (a total of 324), LSTM modeling was performed. Of the 1032 samples, 80% were used for the training set and 20% for the test set. Future time series prediction was also performed, forecasting 24 future samples (i.e., samples from the next day). In this embodiment, the LSTM(a) function from the keras.layers database in Python was used for LSTM modeling, where 'a' was set to 128, and the number of training samples was reflected in the batch_size. The LSTM modeling was repeated 50 times, and the model with the highest R² fit was selected. In fact, the highest R² fit for each component's LSTM modeling was above 90%.

[0173] The component values ​​of all 24 predicted samples are combined according to their corresponding parameters. The results are then substituted into the linear regression model described above, yielding a time series plot of each principal parameter and its analytical value. Specifically, the time series plot of the principal parameter X18 and its analytical value is shown below. Figure 3 As shown, the timing diagram of the main parameter X111 and its analytical values ​​is as follows: Figure 4 As shown, the timing diagram of the main parameter X25 and its resolved values ​​is as follows: Figure 5 As shown, the timing diagram of the main parameter X29 and its resolved values ​​is as follows: Figure 6 As shown, the timing diagram of the main parameter X310 and its resolved values ​​is as follows: Figure 7 As shown, the timing diagram of the main parameter X312 and its analytical values ​​is as follows: Figure 8 As shown. By Figures 3 to 8 It is evident that the fitting effect is good, the fitting result is good, and the future prediction result is foreseeable; in fact, the result matches its changing trend. Specifically: the flow rate 3X111 does indeed show an increase from approximately 75m³. 3 / h decreased to 70m 3 / h rises again to 75m 3 The flow rate is 3X111, and the temperature (5X29), flow rate (1X310), and flow rate (4X312) also show an increasing trend, while the current of pump A (X25) shows a slight decreasing trend.

[0174] Furthermore, by combining the LSTM prediction results of each component of the above response parameters and the prediction results obtained by analyzing the main parameters through a linear regression model, a robust distance model is established. This model is then integrated with the corresponding 24 sets of future prediction samples to obtain a robust distance time series plot of the entire sample (including 24 future time series prediction results), as shown below. Figure 9 As shown. The threshold value is... p=36. This indicates that the robust distances of the denoised data did not exceed the threshold, meaning the denoised samples were all in a healthy state. It can also be observed that the robust distances of the last 24 samples did not exceed the threshold either. In fact, no malfunctions or operational errors occurred on June 2nd; the equipment operated smoothly.

[0175] Finally, the merged results (including 24 future prediction samples) are substituted into the dynamic sparse principal component analysis model to obtain the future prediction principal component samples, such as... Figure 10 As shown, Figure 10 The image shows the principal component time series plots for the entire sample (including 24 future time series predictions). The selected area represents the 24 future time series predictions. The scatter plots of the first three principal components for each of the 24 future prediction samples are also shown. Based on the scatter plots, it can be observed that PC1 deviates significantly from 0, indicating a substantial impact from flow rate, in conjunction with the main parameters. However, its values ​​are similar to those of principal components in similar time periods, suggesting a locally stable state.

[0176] The above method enables future prediction of various parameters (primary parameters and response parameters) in combination with dynamic and interactive characteristics, as well as future prediction of corresponding stability and principal component values, providing a reference for adjusting the operation of the reflux pump.

[0177] The following is combined with Figure 11 The specific flow of the polysilicon production equipment control method provided in the embodiments of the present invention is as follows:

[0178] Focusing on pump A in the cold hydrogenation workshop, equipment parameters and upstream and downstream process parameters of the target equipment were collected. Based on the on-site fault conditions, outlier removal was initially performed, and the target parameter samples were preprocessed. Through EMD decomposition of all target parameters, each target parameter was divided into several components. Principal component analysis (PCA) was performed on each target parameter as the principal component to reduce dimensionality. Under a cumulative contribution rate threshold of 90%, the number of components was reduced to obtain stationary parameters. All remaining components were backed up, and the components of each target parameter were merged and subjected to dynamic sparse principal component analysis to determine the principal parameters. The remaining parameters were used as response parameters. Simultaneously, the loading matrix and eigenvalues ​​were determined and reserved for later use. A multiple linear regression model was constructed between the principal parameters and the response parameters, and the models were screened based on significance and good fit. LSTM time-series prediction was performed on the backed-up components of the response parameters to obtain future... The system calculates the predicted samples and merges them according to their respective parameters. These samples are then substituted into the retained multiple linear regression model to obtain the analytical values ​​of the principal parameters at the same time as the predicted future samples of the corresponding response parameters. Robust distances, the final covariance matrix S, and the mean matrix T are obtained from the training set samples. These parameters are then used to calculate the robust distances between the test set samples and the predicted future samples, and the results of the predicted future samples of the robust distances are used as the prediction results. Using the determined loading matrix and eigenvalues, the predicted future samples are substituted into principal component analysis to calculate the corresponding score samples, which are then used as the prediction results of the principal component parameters. Finally, the system integrates the stationary parameters, the predicted future samples of each principal component parameter, and the predicted future samples of stability (robust distance) to provide the prediction results for the upstream and downstream processes of the reflux pump. Based on these trends, the system provides guidance for adjusting the operation of the reflux pump.

[0179] The following is combined with Figure 12 The specific process for obtaining the predicted samples of the principal component parameters and the predicted samples of the robust distance parameters of the target parameters is as follows:

[0180] Step 1201: Obtain the stationary parameter of the target parameter, denoted as X. 1’ X 2’ X 3’ ...、X n’ ;

[0181] Step 1202: Construct an augmented matrix, incorporating the parameter samples from the previous two time shifts;

[0182] Step 1203: Perform sparse principal component analysis to reduce the dimensionality of the augmented matrix, requiring a cumulative contribution rate of 85%;

[0183] Step 1204: Obtain the loading matrix and eigenvalues ​​of the stationary parameter sample from the sparse principal component analysis dimensionality reduction as backups, and use the resulting final loading matrix and eigenvalues ​​as backups;

[0184] Step 1205: Obtain the principal component equation based on the load matrix, identify the sample parameters existing in the principal component equation as the main parameters, and back them up. These parameters will be used to construct the correlation model of the reflux pump and upstream and downstream processes.

[0185] Step 1206: Obtain the response parameters other than the main parameters from the target parameters.

[0186] Step 1207: Perform multiple linear fitting based on the response parameters and main parameters to construct a correlation model between the reflux pump and the upstream and downstream processes;

[0187] Step 1208: Obtain the remaining components of the backup response parameters;

[0188] Step 1209: Obtain several multiple linear regression models, and screen the relevant models through significance test and goodness-of-fit test;

[0189] Step 1210: Determine the training set and test set for each remaining component of each response parameter, and derive the LSTM model. Obtain the fitted values ​​of the training set and test set under the model, and obtain the future values ​​through future prediction. Merge the obtained future values ​​of the remaining component to obtain the future prediction samples of the reflux pump and upstream and downstream process response parameters.

[0190] Step 1211: Back up the obtained multiple linear regression model and correlation model;

[0191] Step 1212: Substitute the analysis results of the response parameters to obtain the analysis results of the main parameters;

[0192] Step 1213: Merge the analytical values ​​and future prediction samples of the training set and test set of each component as the analytical result of the response parameter;

[0193] Step 1214: Using the backup final load matrix and eigenvalues, along with the future prediction results samples of all parameters, substitute them into the above dynamic sparse principal component model to solve for the principal component values ​​of the corresponding samples, which are then used as the future prediction results of the principal component parameters.

[0194] Based on the above embodiments, this invention also provides an apparatus for future prediction based on weighted device master data, such as... Figure 13 As shown, it specifically includes:

[0195] The collection module 1301 is mainly used to collect parameter samples from various points on site.

[0196] The preprocessing module 1302, while retaining the interlock values ​​and warning values ​​of each parameter, incorporates algorithms for EMD decomposition, PCA dimensionality reduction, component screening, and merging. It is used to perform EMD decomposition on each parameter and PCA dimensionality reduction to screen out interference components, thereby performing noise reduction processing.

[0197] The dynamic sparse principal component analysis module 1303 is used to construct an augmented matrix from preprocessed historical data samples, perform sparse principal component analysis, and output the corresponding mean, variance, loading matrix, score matrix, and eigenvalues. These values ​​are then used to construct a dynamic sparse principal component model.

[0198] Modeling module 1304 mainly includes multiple linear regression algorithm and LSTM algorithm, as well as corresponding methods for judging the effectiveness of the model. The multiple linear regression algorithm only requires input of target parameter information and response parameter samples, while the LSTM algorithm requires manual adjustment of the input repetition number and the number of training sets.

[0199] The prediction module 1305 mainly receives the model established by the modeling module 1304 and uses the model and the number of samples input for future time series prediction to perform full-parameter future time series prediction.

[0200] The stability module 1306 mainly determines whether the data is stable by calculating the robust distance of all parameters and all samples and setting a threshold.

[0201] The coordination and invocation module 1307 is mainly used to retrieve and allocate resources from the storage module 1308 as needed for corresponding instructions from the root modules 1301 to 1306.

[0202] Storage module 1308 is mainly used to store samples and corresponding calculation examples, steps, and fault point markers.

[0203] Display module 1309 is mainly used to display the time series graphs of various parameters and indicators, as well as the health status of the parameters.

[0204] Output module 1310 is used to output and export the processing results of each module.

[0205] The method provided in this invention addresses the challenging issue of data denoising in polycrystalline silicon processes, facilitating the construction of healthy data. Based on refined predictions of various parameters, it establishes connections between them. Considering the dynamic characteristics of the parameters, it establishes their dynamic temporal relationships, ultimately achieving accurate future prediction of master data for all equipment. This invention utilizes robust Mahalanobis distance to improve fault detection efficiency and constructs a robust space for more accurate predictions. This model is crucial for assessing overall stability. For future time-series predictions, it goes beyond a single LSTM method, constructing a correlation model through multiple linear regression, thus making future predictions more accurate. Dynamic sparse principal component analysis is applied to the denoised data, considering not only the dynamic characteristics of the parameters but also facilitating the centralized analysis of important parameters, reducing data processing volume, and simplifying modeling. Once this process is implemented, subsequent processing incorporates the dynamic characteristics of the parameters. Future prediction of principal components is achieved indirectly. Specifically, this prediction does not directly predict the principal components over time. Instead, it uses the loading matrix of dynamic parameters (or the principal component equation) as a basis to predict the required parameters and then substitutes them into the equation to obtain analytical values. This process reveals that this method of predicting the future principal components incorporates the future states of each parameter, increasing accuracy and bypassing the rapid distortion inherent in traditional predictions. It considers the relationships between parameters while eliminating the need to establish correlations between individual components of each parameter (i.e., the stationary components are predicted using MLR multiple linear regression). The use of dynamic sparse principal component analysis takes into account the dynamic characteristics of the parameters, which is a significant improvement. Furthermore, a robust distance model is constructed to assess overall stability.

[0206] like Figure 14 As shown, this embodiment of the invention also provides a control device for polysilicon production equipment, comprising:

[0207] The first acquisition module 1401 is used to acquire target parameters, which include equipment parameters of the target device and upstream and downstream process parameters of the target device;

[0208] The first processing module 1402 is used to perform empirical mode decomposition (EMD) and principal component analysis on the target parameters to obtain stationary parameters of the target parameters.

[0209] The second acquisition module 1403 is used to acquire prediction samples of the stationary parameters of the target parameters, prediction samples of the principal component parameters of the target parameters, and prediction samples of the robust distance parameters of the target parameters.

[0210] The determination module 1404 is used to determine the equipment parameters of the target equipment and the prediction results of the upstream and downstream process parameters of the target equipment based on the prediction samples of the stationary parameters of the target parameters, the prediction samples of the principal component parameters of the target parameters and the prediction samples of the robust distance parameters of the target parameters.

[0211] The adjustment module 1405 is used to adjust the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment according to the prediction results.

[0212] Optionally, the second acquisition module 1403 includes:

[0213] The first processing unit is configured to obtain, based on the stationary parameters, the load matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equation, the correlation model between the target device and its upstream and downstream, future prediction samples of the response parameters in the target parameters, and future prediction samples of the main parameters in the target parameters;

[0214] The second processing unit is used to obtain a prediction sample of the robust distance parameter of the target parameter based on the robust distance of the target parameter, the covariance matrix of the target parameter, the eigenvalue of the stationary parameter, the future prediction sample of the response parameter, and the future prediction sample of the principal parameter.

[0215] The third processing unit is used to obtain the predicted samples of the principal component parameters of the target parameter based on the load matrix of the stationary parameter, the eigenvalues ​​of the stationary parameter, the future prediction samples of the response parameter, the future prediction samples of the principal parameters, and the principal component equation.

[0216] Wherein, the main parameters are the parameters among the target parameters that affect the target equipment and the upstream and downstream processes of the target equipment; the response parameters are the parameters among the target parameters other than the main parameters.

[0217] Optionally, the first processing unit is specifically used for:

[0218] Dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix and eigenvalues ​​of the stationary parameters.

[0219] Based on the loading matrix of the stationary parameters, the principal component equation is obtained, and based on the principal component equation, the principal parameters and the response parameters are obtained;

[0220] Based on the main parameters and the response parameters, construct an association model between the target device and its upstream and downstream partners;

[0221] Based on the residual components of the response parameters and the Long Short-Term Memory (LSTM) model, future prediction samples of the response parameters are obtained; the residual components of the response parameters are obtained by principal component analysis dimensionality reduction processing of the intrinsic modal component (IMF) and the residual components of the response parameters according to a first preset cumulative contribution rate threshold; the IMF and the residual components of the response parameters are obtained by EMD processing of the residual components of the response parameters.

[0222] Based on the future prediction samples of the response parameters and the correlation model, the future prediction samples of the main parameters are obtained.

[0223] Optionally, the first processing unit is specifically used for:

[0224] Based on the target process parameters and historical target process parameters, an augmented matrix is ​​constructed;

[0225] Based on the second preset cumulative contribution rate threshold, sparse principal component analysis is performed on the augmented matrix to obtain the loading matrix of the stationary parameter and the eigenvalues ​​of the stationary parameter.

[0226] Optionally, the first processing module 1402 includes:

[0227] The fourth processing unit is used to perform interpolation and missing data processing based on the sample data in each target parameter and to perform correlation processing based on the sample data in each target parameter to obtain the preprocessed target parameters.

[0228] The fifth processing unit is used to perform empirical mode decomposition (EMD) and principal component analysis on the preprocessed target parameters to obtain stationary parameters of the target parameters.

[0229] Optionally, the fourth processing unit is specifically used for:

[0230] If the number of consecutively missing sample data in the first target parameter is within a first preset range, the consecutively missing sample data is interpolated and filled using linear interpolation.

[0231] If the number of consecutively missing sample data in the first target parameter is within the second preset range, the consecutively missing sample data is interpolated and filled using the third preset range of sample data adjacent to the consecutively missing sample data, the fourth preset range of sample data adjacent to the consecutively missing sample data, a preset bidirectional sliding data window algorithm, and a linear scaling algorithm.

[0232] If the number of consecutively missing sample data in the first target parameter is within the fifth preset range, the first target parameter is discarded.

[0233] Wherein, the first target parameter is any one of the target parameters;

[0234] The number of the first preset range is less than the number of the second preset range, and the number of the second preset range is less than the number of the fifth preset range.

[0235] Optionally, the fifth processing unit is specifically used for:

[0236] EMD processing is performed on the sample data in the second target parameter to obtain the intrinsic modal component (IMF) and residual component of the second target parameter.

[0237] Based on the first preset cumulative contribution rate threshold, principal component analysis is performed on the IMF of the second target parameter and the residual components of the second target parameter to reduce the dimensionality and obtain the residual components of the second target parameter.

[0238] The remaining components of each second objective parameter are combined to obtain the stationary parameters of the second objective parameter;

[0239] Wherein, the second target parameter is any one of the preprocessed target parameters;

[0240] The residual component of the second target parameter is the component of the second target parameter minus the IMF of the second target parameter.

[0241] It should be noted that the polysilicon production equipment control device provided in the embodiments of the present invention is a device capable of executing the above-described polysilicon production equipment control method. Therefore, all embodiments of the above-described polysilicon production equipment control method are applicable to this device and can achieve the same or similar technical effects.

[0242] like Figure 15 As shown, this embodiment of the invention also provides a server, including: a processor 1501; and a memory 1503 connected to the processor 1501 via a bus interface 1502, the memory 1503 being used to store programs and data used by the processor 1501 when performing operations, and the processor 1501 calling and executing the programs and data stored in the memory 1503.

[0243] The transceiver 1504 is connected to the bus interface 1502 and is used to receive and send data under the control of the processor 1501. Specifically, the processor 1501 is used to read the program in the memory 1503, and the processor 1501 executes the following processes:

[0244] Obtain target parameters, which include equipment parameters of the target device and upstream and downstream process parameters of the target device;

[0245] The target parameters are processed by empirical mode decomposition (EMD) and principal component analysis to obtain stationary parameters of the target parameters;

[0246] Obtain prediction samples of the principal component parameters of the target parameter and prediction samples of the robust distance parameter of the target parameter;

[0247] Based on the stationary parameters of the target parameters, the predicted samples of the principal component parameters of the target parameters, and the predicted samples of the robust distance parameters of the target parameters, the equipment parameters of the target equipment and the predicted results of the upstream and downstream process parameters of the target equipment are determined.

[0248] Based on the prediction results, the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment are adjusted.

[0249] Optionally, the processor 1501 is configured to:

[0250] Based on the stationary parameters, the load matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equations, the correlation model between the target device and its upstream and downstream, the future prediction samples of the response parameters in the target parameters, and the future prediction samples of the main parameters in the target parameters are obtained.

[0251] Based on the robust distance of the target parameter, the covariance matrix of the target parameter, the eigenvalue of the stationary parameter, the future prediction sample of the response parameter, and the future prediction sample of the principal parameter, the prediction sample of the robust distance parameter of the target parameter is obtained.

[0252] Based on the loading matrix of the stationary parameter, the eigenvalues ​​of the stationary parameter, the future prediction samples of the response parameter, the future prediction samples of the principal parameter, and the principal component equation, the prediction samples of the principal component parameters of the target parameter are obtained.

[0253] Wherein, the main parameters are the parameters among the target parameters that affect the target equipment and the upstream and downstream processes of the target equipment; the response parameters are the parameters among the target parameters other than the main parameters.

[0254] Optionally, the processor 1501 is specifically used for:

[0255] Dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix and eigenvalues ​​of the stationary parameters.

[0256] Based on the loading matrix of the stationary parameters, the principal component equation is obtained, and based on the principal component equation, the principal parameters and the response parameters are obtained;

[0257] Based on the main parameters and the response parameters, construct an association model between the target device and its upstream and downstream partners;

[0258] Based on the residual components of the response parameters and the Long Short-Term Memory (LSTM) model, future prediction samples of the response parameters are obtained; the residual components of the response parameters are obtained by principal component analysis dimensionality reduction processing of the intrinsic modal component (IMF) and the residual components of the response parameters according to a first preset cumulative contribution rate threshold; the IMF and the residual components of the response parameters are obtained by EMD processing of the residual components of the response parameters.

[0259] Based on the future prediction samples of the response parameters and the correlation model, the future prediction samples of the main parameters are obtained.

[0260] Optionally, the processor 1501 is specifically used for:

[0261] Based on the target process parameters and historical target process parameters, an augmented matrix is ​​constructed;

[0262] Based on the second preset cumulative contribution rate threshold, sparse principal component analysis is performed on the augmented matrix to obtain the loading matrix of the stationary parameter and the eigenvalues ​​of the stationary parameter.

[0263] Optionally, the processor 1501 is configured to:

[0264] Interpolation and correlation processing are performed on the sample data of each target parameter to obtain the preprocessed target parameters.

[0265] The preprocessed target parameters are subjected to Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA) to obtain stationary parameters of the target parameters.

[0266] Optionally, the processor 1501 is specifically used for:

[0267] If the number of consecutively missing sample data in the first target parameter is within a first preset range, the consecutively missing sample data is interpolated and filled using linear interpolation.

[0268] If the number of consecutively missing sample data in the first target parameter is within the second preset range, the consecutively missing sample data is interpolated and filled using the third preset range of sample data adjacent to the consecutively missing sample data, the fourth preset range of sample data adjacent to the consecutively missing sample data, a preset bidirectional sliding data window algorithm, and a linear scaling algorithm.

[0269] If the number of consecutively missing sample data in the first target parameter is within the fifth preset range, the first target parameter is discarded.

[0270] Wherein, the first target parameter is any one of the target parameters;

[0271] The number of the first preset range is less than the number of the second preset range, and the number of the second preset range is less than the number of the fifth preset range.

[0272] Optionally, the processor 1501 is specifically used for:

[0273] EMD processing is performed on the sample data in the second target parameter to obtain the intrinsic modal component (IMF) and residual component of the second target parameter.

[0274] Based on the first preset cumulative contribution rate threshold, principal component analysis is performed on the IMF of the second target parameter and the residual components of the second target parameter to reduce the dimensionality and obtain the residual components of the second target parameter.

[0275] The remaining components of each second objective parameter are combined to obtain the stationary parameters of the second objective parameter;

[0276] Wherein, the second target parameter is any one of the preprocessed target parameters;

[0277] The residual component of the second target parameter is the component of the second target parameter minus the IMF of the second target parameter.

[0278] Among them, Figure 15In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 1501) and memory (memory 1503). The bus architecture may also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides a user interface 1505. A transceiver 1504 may be multiple elements, including transmitters and receivers, providing units for communicating with various other devices over a transmission medium. Processor 1501 is responsible for managing the bus architecture and general processing, and memory 1503 may store data used by processor 1501 during operation.

[0279] In addition, specific embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the polysilicon production equipment control method as described above.

[0280] The above describes the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also within the scope of protection of the present invention.

Claims

1. A control method for polysilicon production equipment, characterized in that, include: Obtain target parameters, which include equipment parameters of the target device and upstream and downstream process parameters of the target device; The target parameters are processed by empirical mode decomposition (EMD) and principal component analysis to obtain stationary parameters of the target parameters; Obtain prediction samples of the stationary parameters of the target parameter, prediction samples of the principal component parameters of the target parameter, and prediction samples of the robust distance parameters of the target parameter; Based on the prediction samples of the stationary parameters of the target parameters, the prediction samples of the principal component parameters of the target parameters, and the prediction samples of the robust distance parameters of the target parameters, the equipment parameters of the target equipment and the prediction results of the upstream and downstream process parameters of the target equipment are determined. Based on the prediction results, the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment are adjusted; The process of obtaining prediction samples of the principal component parameters of the target parameter and prediction samples of the robust distance parameter of the target parameter includes: Based on the stationary parameters, the following are obtained: the load matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equation, the principal parameters in the target parameters, the response parameters in the target parameters, the correlation model between the target device and its upstream and downstream, the future prediction samples of the response parameters in the target parameters, and the future prediction samples of the principal parameters in the target parameters. Based on the robust distance of the target parameter, the covariance matrix of the target parameter, the eigenvalue of the stationary parameter, the future prediction sample of the response parameter, and the future prediction sample of the principal parameter, the prediction sample of the robust distance parameter of the target parameter is obtained. Based on the loading matrix of the stationary parameter, the eigenvalues ​​of the stationary parameter, the future prediction samples of the response parameter, the future prediction samples of the principal parameter, and the principal component equation, the prediction samples of the principal component parameters of the target parameter are obtained. Wherein, the main parameters are the parameters among the target parameters that affect the target equipment and the upstream and downstream processes of the target equipment; the response parameters are the parameters among the target parameters other than the main parameters.

2. The method according to claim 1, characterized in that, Based on the stationary parameters, the following are obtained: the loading matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equations, the principal parameters in the target parameters, the response parameters in the target parameters, the correlation model between the target device and its upstream and downstream components, future prediction samples of the response parameters in the target parameters, and future prediction samples of the principal parameters in the target parameters, including: Dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix and eigenvalues ​​of the stationary parameters. Based on the loading matrix of the stationary parameters, the principal component equation is obtained, and based on the principal component equation, the principal parameters and the response parameters are obtained; Based on the main parameters and the response parameters, construct an association model between the target device and its upstream and downstream partners; Based on the residual components of the response parameters and the Long Short-Term Memory (LSTM) model, future prediction samples of the response parameters are obtained; the residual components of the response parameters are obtained by principal component analysis dimensionality reduction processing of the intrinsic modal component (IMF) and the residual components of the response parameters according to a first preset cumulative contribution rate threshold; the IMF and the residual components of the response parameters are obtained by EMD processing of the residual components of the response parameters. Based on the future prediction samples of the response parameters and the correlation model, the future prediction samples of the main parameters are obtained.

3. The method according to claim 2, characterized in that, Dynamic sparse principal component analysis is performed on the stationary parameters to obtain the loading matrix and eigenvalues ​​of the stationary parameters, including: Based on the target parameters and historical target process parameters, an augmented matrix is ​​constructed; Based on the second preset cumulative contribution rate threshold, sparse principal component analysis is performed on the augmented matrix to obtain the loading matrix of the stationary parameter and the eigenvalues ​​of the stationary parameter.

4. The method according to claim 1, characterized in that, Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA) are performed on the target parameters to obtain stationary parameters of the target parameters, including: Interpolation and correlation processing are performed on the sample data of each target parameter to obtain the preprocessed target parameters. The preprocessed target parameters are subjected to Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA) to obtain stationary parameters of the target parameters.

5. The method according to claim 4, characterized in that, Interpolation and gap filling are performed based on the sample data in each of the target parameters, including: If the number of consecutively missing sample data in the first target parameter is within a first preset range, the consecutively missing sample data is interpolated and filled using linear interpolation. If the number of consecutively missing sample data in the first target parameter is within the second preset range, the consecutively missing sample data is interpolated and filled using the third preset range of sample data adjacent to the consecutively missing sample data, the fourth preset range of sample data adjacent to the consecutively missing sample data, a preset bidirectional sliding data window algorithm, and a linear scaling algorithm. If the number of consecutively missing sample data in the first target parameter is within the fifth preset range, the first target parameter is discarded. Wherein, the first target parameter is any one of the target parameters; The number of the first preset range is less than the number of the second preset range, and the number of the second preset range is less than the number of the fifth preset range.

6. The method according to claim 4, characterized in that, The preprocessed target parameters are subjected to Empirical Mode Decomposition (EMD) and Principal Component Analysis (PCA) to obtain stationary parameters of the target parameters, including: EMD processing is performed on the sample data in the second target parameter to obtain the intrinsic modal component (IMF) and residual component of the second target parameter. Based on the first preset cumulative contribution rate threshold, principal component analysis is performed on the IMF of the second target parameter and the residual components of the second target parameter to reduce the dimensionality and obtain the residual components of the second target parameter. The remaining components of each second objective parameter are combined to obtain the stationary parameters of the second objective parameter; Wherein, the second target parameter is any one of the preprocessed target parameters; The residual component of the second target parameter is the component of the second target parameter minus the IMF of the second target parameter.

7. A control device for polysilicon production equipment, characterized in that, include: The first acquisition module is used to acquire target parameters, which include equipment parameters of the target device and upstream and downstream process parameters of the target device; The first processing module is used to perform Empirical Mode Decomposition (EMD) and Principal Component Analysis on the target parameters to obtain stationary parameters of the target parameters. The second acquisition module is used to acquire prediction samples of the stationary parameters of the target parameters, prediction samples of the principal component parameters of the target parameters, and prediction samples of the robust distance parameters of the target parameters. The determination module is used to determine the equipment parameters of the target equipment and the prediction results of the upstream and downstream process parameters of the target equipment based on the prediction samples of the stationary parameters of the target parameters, the prediction samples of the principal component parameters of the target parameters, and the prediction samples of the robust distance parameters of the target parameters. An adjustment module is used to adjust the equipment parameters of the target equipment and / or the upstream and downstream process parameters of the target equipment based on the prediction results. The second acquisition module includes: The first processing unit is configured to obtain, based on the stationary parameters, the load matrix of the stationary parameters, the eigenvalues ​​of the stationary parameters, the principal component equation, the principal parameters in the target parameters, the response parameters in the target parameters, the correlation model between the target device and its upstream and downstream, future prediction samples of the response parameters in the target parameters, and future prediction samples of the principal parameters in the target parameters; The second processing unit is used to obtain a prediction sample of the robust distance parameter of the target parameter based on the robust distance of the target parameter, the covariance matrix of the target parameter, the eigenvalue of the stationary parameter, the future prediction sample of the response parameter, and the future prediction sample of the principal parameter. The third processing unit is used to obtain the predicted samples of the principal component parameters of the target parameter based on the load matrix of the stationary parameter, the eigenvalues ​​of the stationary parameter, the future prediction samples of the response parameter, the future prediction samples of the principal parameters, and the principal component equation. Wherein, the main parameters are the parameters among the target parameters that affect the target equipment and the upstream and downstream processes of the target equipment; the response parameters are the parameters among the target parameters other than the main parameters.

8. A server, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps in the polysilicon production equipment control method as described in any one of claims 1 to 6.

9. A readable storage medium, characterized in that, The readable storage medium stores a program that, when executed by a processor, implements the steps of the polysilicon production equipment control method as described in any one of claims 1 to 6.

Images