Process Optimization Method Based on Principal Component Analysis and Neural Network

Abstract

The embodiment of the present invention provides a process optimization method based on principal component analysis and neural network, which belongs to the field of process modeling of big data, including: obtaining sample point data at a site in the process; using principal component analysis to perform dimensionality reduction processing to obtain feature data corresponding to the site in the process; using principal component analysis to determine the weight of the feature data corresponding to the site in the process; determining the main site in the process according to the weight of the feature data corresponding to the site in the process; inputting the sample point data at the main site in the process into a recurrent neural network for training, taking the percentage of reducing sugar in the dry matter of syrup as the target, training the recurrent neural network, and determining the importance value of the main site in the process according to the weight matrix of the recurrent neural network; determining the site in the process for optimizing the process parameters according to the importance value of the main site in the process. The present invention uses process parameters to adjust the main sites in the process, which can optimize production costs and processing efficiency.

Description

Technical Field

[0001] The present invention relates to the field of process modeling of big data, and in particular to a process optimization method based on principal component analysis and neural network. Background Art

[0002] With the rapid development of industry, the scale and output of factories are constantly expanding, and the variety of products is increasing. There are many difficulties in the process of upgrading and renovating old factories or adjusting and optimizing parameters for market demand. Many industrial processes are developing rapidly, and their products are widely used in food, chemical, pharmaceutical and other industrial processes, playing an increasingly important role in the development of my country's national economy. For example, the process of producing sugar products using starch as raw material is collectively referred to as starch sugar process. The starch sugar process includes four-stage liquefaction section, saccharification section, glucose filtration section, decolorization section, four-stage separation section, evaporation section and isomerization section. The entire processing process involves a long process route and complex process control. Due to the complex raw material system and frequent changes in production conditions, it is impossible to make rapid parameter adjustment instructions and optimization plans for product quality fluctuations, which often leads to a series of problems such as unstable product quality and low raw material utilization. In the starch sugar process, the double enzyme process is the main production process of starch sugar. In the dual enzyme process, starch is first hydrolyzed into dextrin and oligosaccharide products with smaller molecular weight by liquefaction enzymes, and then further hydrolyzed into glucose by saccharification enzymes. The final products include crystalline glucose, whole sugar, glucose-maltose syrup and maltose syrup, etc. This is because the starch sugar production process is affected by a large number of operating factors (process sites) such as pH value, saccharification time, and enzyme addition. Relying solely on the experience of engineers to adjust the parameters will lead to uneven product quality.

[0003] On the basis of the general implementation of digital monitoring and recording in factories, we need to fully tap the large amount of useful information contained in the data itself, develop and utilize the potential value hidden behind the data, and help production management departments quickly identify the key parameters that affect product quality through big data analysis. This is the technical problem we need to solve. Summary of the invention

[0004] The purpose of the embodiment of the present invention is to provide a process optimization method based on principal component analysis and neural network, which mainly solves how to determine the location in the process for optimizing process parameters to improve the efficiency of process adjustment.

[0005] In order to achieve the above object, an embodiment of the present invention provides a process optimization method based on principal component analysis and neural network, comprising:

[0006] Obtain sample point data at the process location;

[0007] The process midpoint and the sample point data at the process midpoint are subjected to dimensionality reduction processing using principal component analysis to obtain feature data corresponding to the process midpoint;

[0008] Using principal component analysis to determine the weights of characteristic data corresponding to the sites in the process;

[0009] Determining the main sites in the process according to the weights of the characteristic data corresponding to the sites in the process;

[0010] Inputting the sample point data at the main sites in the process into a recurrent neural network for training, taking the percentage of reducing sugar in the dry matter of syrup as a target, training the recurrent neural network, and determining the importance values ​​of the main sites in the process according to the weight matrix of the recurrent neural network;

[0011] The sites in the process for optimizing process parameters are determined according to the importance values ​​of the main sites in the process.

[0012] Optionally, determining the sites in the process for optimizing process parameters according to the importance values ​​of the main sites in the process includes:

[0013] According to the importance values ​​of the main sites in the process, they are sorted from large to small to obtain an importance value sequence;

[0014] The first N positions of the importance value sequence were analyzed using the dual-enzyme starch sugar processing technology to obtain the positions in the process for optimizing process parameters.

[0015] Optionally, before the sample point data at the process midpoint and the process midpoint are subjected to dimensionality reduction processing by using the principal component analysis method, the following is further included:

[0016] The sample point data at the process location are subjected to noise preprocessing.

[0017] Optionally, the sample point data at the process midpoint and the process midpoint are subjected to dimensionality reduction processing using principal component analysis to obtain feature data corresponding to the process midpoint, including:

[0018] The sample point data at the process site and the variables of the correlation formed by the process site are orthogonally transformed, and the variables obtained by the orthogonal transformation are used as the feature data corresponding to the process site.

[0019] Optionally, the method of using principal component analysis to determine the weights of characteristic data corresponding to sites in the process includes:

[0020] Using principal component analysis to determine the variable component matrix, principal component eigenvalue and principal component variance contribution rate of the characteristic data corresponding to the sites in the process;

[0021] Calculate the ratio of the variable component matrix to the arithmetic square root of the principal component eigenvalue as the component score coefficient matrix;

[0022] Taking the principal component variance contribution rate as the weight, the component score coefficient matrix is ​​weighted averaged to obtain the comprehensive score model coefficient;

[0023] The weights of the characteristic data corresponding to the sites in the process are determined according to the comprehensive score model coefficients.

[0024] Optionally, determining the weight of the characteristic data corresponding to the site in the process according to the comprehensive score model coefficient includes:

[0025] The comprehensive score model coefficients are normalized, and the processed results are used as the weights of the characteristic data corresponding to the sites in the process.

[0026] Optionally, determining the main sites in the process according to the weights of the feature data corresponding to the sites in the process includes:

[0027] The weights of the characteristic data corresponding to the sites in the process are sorted from large to small to obtain a ranking sequence, and the sites corresponding to the first M sites in the ranking sequence are taken as the main sites in the process.

[0028] Optionally, the recurrent neural network includes LSTM and artificial neural network layers;

[0029] The sample point data at the main points in the process are input into a recurrent neural network for training, and the percentage of reducing sugar in the dry matter of the syrup is taken as a target, and the recurrent neural network is trained, including:

[0030] The sample point data at the main points in the process are respectively input into the LSTM and artificial neural network layers, and the LSTM and artificial neural network layers are trained with the percentage of reducing sugar in the dry matter of the syrup as the target.

[0031] Optionally, the step of calculating the weight matrix of the recurrent neural network includes:

[0032] Perform matrix multiplication on the matrices composed of the weights of each neuron in the artificial neural network layer after each training to obtain the weight matrix of the artificial neural network layer;

[0033] The comprehensive weight matrix of the LSTM is matrix-multiplied with the weight matrix of the artificial neural network layer to obtain the weight matrix of the recurrent neural network.

[0034] Optionally, the step of calculating the weight matrix of the recurrent neural network includes:

[0035] The sample point data at the main sites in the process are input into the artificial neural network layer in batches for training, and the average value of the weights of the neurons after the training of the sample point data at the main sites in each batch of the process is completed is used as the batch weights of the neurons in the batch training; the matrix composed of all the batch weights is matrix-multiplied to obtain the weight matrix of the artificial neural network layer;

[0036] The comprehensive weight matrix of the LSTM is matrix-multiplied with the weight matrix of the artificial neural network layer to obtain the weight matrix of the recurrent neural network.

[0037] Optionally, determining the importance values ​​of main sites in the process according to a weight matrix of a recurrent neural network includes:

[0038] The elements in the weight matrix of the recurrent neural network are normalized to obtain the importance values ​​of the main sites in the process.

[0039] Optionally, determining the sites in the process for optimizing process parameters according to the importance values ​​of the main sites in the process includes:

[0040] The importance values ​​of the main sites in the process are sorted into an importance value sequence, and the sites corresponding to the first K bits of the importance value sequence are used as sites in the process for optimizing process parameters.

[0041] Through the above technical solution, the main sites that affect the process are screened out through principal component analysis. On the other hand, the process can be modeled based on big data using a recurrent neural network based on these main sites. The recurrent neural network can effectively process time series data, achieve accurate prediction of target parameters, and further obtain key control sites (main sites in the process). The application of key control sites (main sites in the process) is of great significance for guiding the adjustment of on-site process parameters, which can optimize production costs and processing efficiency while ensuring product quality.

[0042] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:

[0044] Figure 1 is a schematic diagram of the structure of an artificial neural network layer provided in an embodiment;

[0045] Figure 2 This is an LSTM training effect diagram provided in the embodiment;

[0046] Figure 3 It is a box plot of importance values ​​corresponding to each site with the top 10 importance values ​​in 10 trainings provided in the embodiment;

[0047] Figure 4 It is a structural diagram of an LSTM network provided in an embodiment. DETAILED DESCRIPTION

[0048] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.

[0049] Example 1

[0050] This example selects a double-enzyme starch sugar processing technology as a specific optimization implementation, and provides a double-enzyme starch sugar processing technology optimization method based on principal component analysis and neural network.

[0051] A dual-enzyme starch sugar processing technology optimization method based on principal component analysis and neural network includes:

[0052] Obtain the sample point data on the site in the dual enzyme starch sugar processing process; use the principal component analysis method to reduce the dimension of the site in the dual enzyme starch sugar processing process and the sample point data on the site in the dual enzyme starch sugar processing process to obtain the characteristic data corresponding to the site in the dual enzyme starch sugar processing process; use the principal component analysis method to determine the weight of the characteristic data corresponding to the site in the dual enzyme starch sugar processing process; determine the main site in the dual enzyme starch sugar processing process according to the weight of the characteristic data corresponding to the site in the dual enzyme starch sugar processing process; input the sample point data on the main site in the dual enzyme starch sugar processing process into a recurrent neural network for training, take the percentage of reducing sugar in syrup dry matter as the target, train the recurrent neural network, and determine the importance value of the main site in the dual enzyme starch sugar processing process according to the weight matrix of the recurrent neural network; determine the site in the dual enzyme starch sugar processing process for optimizing the parameters of the dual enzyme starch sugar processing process according to the importance value of the main site in the dual enzyme starch sugar processing process. By using the principal component analysis method to reduce the dimension of a large number of operating sites, the main sites that have a greater impact on the process are selected. Furthermore, in view of the large lag, nonlinearity, and strong coupling of industrial data, a recurrent neural network that is good at processing time series data is introduced for modeling, and the target parameters are predicted using the selected main sites as its input. This patent selects the DE value as the target. The DE value represents the percentage of reducing sugar in the dry matter of the syrup, and is an important indicator for evaluating the saccharification effect of the liquid. Finally, an extended weighted connection method is used to calculate the importance of these main sites, sort the importance, and screen out the key sites with high importance. The process flow is optimized by controlling and adjusting these key sites.

[0053] Specifically, the site selection in the dual-enzyme starch sugar processing process collects information from 165 sites from the DCS system and the field test record database, and each site contains 13,600 sample point data. These 165 sites are located in the upstream section of the DE value detection point, and the upstream section mainly includes four liquefaction sections and saccharification sections; the information of the 165 sites is the detection and operation sites in the utility system in the above upstream section.

[0054] Optionally, before the sample point data at the sites in the dual-enzyme starch sugar processing technology and the sites in the dual-enzyme starch sugar processing technology are subjected to dimensionality reduction processing using the principal component analysis method, it also includes: preprocessing the sample point data at the sites in the dual-enzyme starch sugar processing technology. Optionally, noise preprocessing includes screening and eliminating outliers, data noise reduction, data supplementation, and error correction. For manually recorded data, automatic time division is achieved, and linear interpolation is used for data supplementation; for data with a time interval less than the benchmark interval, the method of extracting the average value is used for data extraction; for missing values, linear interpolation is used, and the missing values ​​are calculated according to the following formula:

[0055]

[0056] Among them, (x0, y0), (x1, y1) represent known points in the data, x represents the time point corresponding to the missing value, and y represents the supplementary value at that time point.

[0057] The box plot method is used to screen and remove outliers, and the data of all time periods are placed in the same box, that is, this box contains all sampling points. For the box Ai=[Xi,1,Xi,2,…Xi,N], all the values ​​in the box are arranged from small to large to obtain an ordered box A'i=[X'i,1,X'i,2,…X'i,N], where Xi represents the site i and N represents the number of sampling points;

[0058] The box plot method described above detects outliers by quartiles, and the calculation formula for the position of the quartiles is as follows:

[0059]

[0060] Where l represents the quartile number, Pl represents the position of the quartile, N represents the number of sampling points, and the corresponding quartile is calculated according to the position of the quartile; the lower quartile Q1 is also called the "smaller quartile", which is equal to the number at the 25th% after all the values ​​in the sample are arranged from small to large. The lower quartile Q1 is:

[0061]

[0062] in It means to round P1 down. Indicates that P1 is rounded up, and N* indicates a positive integer;

[0063] The upper quartile Q3 is also called the "larger quartile", which is equal to the number at the 25th percentile after all values ​​in the sample are arranged from small to large. The lower quartile Q3 is:

[0064]

[0065] in Indicates that P3 is rounded down. Indicates rounding up P3, N* indicates a positive integer;

[0066] The lower limit Qmin and upper limit Qmax of the non-abnormal value are calculated by the interquartile range IQR, and the IQR calculation formula is: IQR = Q3-Q1;

[0067] Then the calculation formulas for Qmin and Qmax are:

[0068] Qmin =Q1-m·IQR

[0069] Q max =Q3+m·IQR;

[0070] The value of m can be changed according to the data situation and the number of data to be retained. Generally, the value of m is 1.5.

[0071] Data denoising uses the "(2n+1 point) simple moving average" method to smooth and filter the data for denoising. The central idea is to find the average value yi' instead of yi by finding the n data before and after yi (yi–n,…,yi–1,yi,…yi+n) centered on yi. The formula is as follows:

[0072]

[0073] Further, the sample point data on the site in the double enzyme starch sugar processing technology after noise pretreatment and the site in the double enzyme starch sugar processing technology are subjected to dimensionality reduction processing using the principal component analysis method to obtain the feature data corresponding to the site in the double enzyme starch sugar processing technology. Specifically, the sample point data on the site in the double enzyme starch sugar processing technology and the variables of the correlation formed by the site in the double enzyme starch sugar processing technology are orthogonally transformed, and the variables obtained by the orthogonal transformation are used as the feature data corresponding to the site in the double enzyme starch sugar processing technology. Principal component analysis is the most widely used data dimensionality reduction method. A set of variables that may have correlations are converted into a set of linearly unrelated variables through orthogonal transformation. The converted set of variables is called the principal component, that is, the k-dimensional feature is reconstructed by the principal component analysis method on the basis of the original n-dimensional feature, so as to achieve the purpose of dimensionality reduction. In this embodiment, the information of 165 sites and the corresponding sites contain 13,600 sample point data, and a new set of feature matrices constructed by the principal component analysis method are used as the feature data corresponding to the site in the double enzyme starch sugar processing technology.

[0074] Furthermore, the characteristic data corresponding to the site in the dual-enzyme starch sugar processing technology cannot directly reflect the importance of the input site, so it is necessary to obtain the weight of each site in the dual-enzyme starch sugar processing technology, and the weight of each site in the dual-enzyme starch sugar processing technology is also the weight of the characteristic data corresponding to the site. Specifically, the principal component analysis method is used to determine the variable component matrix, principal component eigenvalue and principal component variance contribution rate of the characteristic data corresponding to the site in the dual-enzyme starch sugar processing technology; the ratio of the arithmetic square root of the variable component matrix and the principal component eigenvalue is calculated as the component score coefficient matrix; the principal component variance contribution rate is used as the weight, and the component score coefficient matrix is ​​weighted and averaged to obtain the comprehensive score model coefficient. The weight of the characteristic data corresponding to the site in the dual-enzyme starch sugar processing technology is determined according to the comprehensive score model coefficient. Optionally, the comprehensive score model coefficient is normalized, and the processed result is used as the weight of the characteristic data corresponding to the site in the dual-enzyme starch sugar processing technology.

[0075] Specifically, the main sites in the dual-enzyme starch sugar processing technology are determined according to the weights of the characteristic data corresponding to the sites in the dual-enzyme starch sugar processing technology, including:

[0076] The weights of the characteristic data corresponding to the sites in the dual-enzyme starch sugar processing technology are sorted from large to small to obtain a ranking sequence, and the sites corresponding to the first M sites in the ranking sequence are used as the main sites in the dual-enzyme starch sugar processing technology. The greater the weight of the characteristic data corresponding to the site in the dual-enzyme starch sugar processing technology, the more important the corresponding site is. In this embodiment, the top 50 sites with the largest weights are selected as the main sites in the dual-enzyme starch sugar processing technology.

[0077] Furthermore, the sample point data at the main sites in the 50 dual-enzyme starch sugar processing techniques are input into the recurrent neural network (LSTM and artificial neural network layer) for training, and the percentage of reducing sugar in the dry matter of the syrup is used as the target to train the recurrent neural network. Specifically, the sample point data at the main sites in the 50 dual-enzyme starch sugar processing techniques are input into the LSTM and artificial neural network layers respectively, and the percentage of reducing sugar in the dry matter of the syrup is used as the target to train the LSTM and artificial neural network layers. The long short-term memory model (LSTM) architecture is used for training the recurrent neural network. The LSTM model introduces the concept of "gate" on the basis of the traditional recurrent neural network model, thereby solving the gradient vanishing and gradient exploding problems existing in the recurrent neural network model; the network is composed of an LSTM layer plus an artificial neural network layer, wherein the structure of the artificial neural network layer is as follows: Figure 1 As shown, the structural principle of the LSTM layer is as follows Figure 4The modeling is based on keras, and the relevant parameters are set as follows: the time step is 5, the number of iterations is set to 300, the root mean square error (mse) is used as the loss function, and the optimizer is selected as adam. After the training is completed, the model mse is 0.16%. The training iteration process is as follows Figure 2 As shown, it can be seen that the model has neither overfitting problem nor high accuracy.

[0078] Optionally, in a possible embodiment, the step of calculating the weight matrix of the recurrent neural network includes:

[0079] The matrix composed of the weights of each neuron in the artificial neural network layer after each training is matrix-multiplied to obtain the weight matrix of the artificial neural network layer; the comprehensive weight matrix of the LSTM is matrix-multiplied with the weight matrix of the artificial neural network layer to obtain the weight matrix of the recurrent neural network.

[0080] Optionally, in another possible embodiment, since the weights in the recurrent neural network training process are random, the recurrent neural network is trained 15 times, and the weight average (batch weight) is taken for analysis and calculation. Specifically, the calculation steps of the weight matrix of the recurrent neural network include:

[0081] The sample point data at the main sites in the dual-enzyme starch sugar processing technology are input into the artificial neural network layer in batches for training, and the average value of the weights of the neurons after the training of each batch of sample point data at the main sites in the dual-enzyme starch sugar processing technology is completed is used as the batch weights of the neurons in the batch training; the matrices composed of all the batch weights are matrix-multiplied to obtain the weight matrix of the artificial neural network layer, and then the comprehensive weight matrix of the LSTM is matrix-multiplied with the weight matrix of the artificial neural network layer to obtain the weight matrix of the recurrent neural network.

[0082] Specifically, the formula of the weight matrix of the artificial neural network layer is:

[0083]

[0084] Among them, Oi represents the neuron output value, w ij and v jk represents the hidden layer weight, and I and J represent the corresponding number of variables. This formula is obtained by normalizing the weight of each neuron and then multiplying the weight matrix to obtain the weight matrix of the artificial neural network layer.

[0085] The solution of the comprehensive weight matrix of LSTM is as follows:

[0086] The LSTM layer includes forget gate, input gate, candidate gate, memory gate and output gate. The operation mechanism of each gate is introduced below:

[0087] For the LSTM layer, this layer includes forget gate, input gate, candidate gate, memory gate and output gate. The operation mechanism of each gate is introduced below:

[0088] Forget Gate t =σ(W f ·[h t-1 ,x t ]+b f ), where W f represents the weight of the forget gate, b f represents the forget gate threshold, σ represents the sigmoid function;

[0089] Input Gate i t =σ(W i ·[h t-1 ,x t ]+b i ), where W i represents the weight of the forget gate, b i represents the forget gate threshold;

[0090] Candidate Gate Where W c represents the candidate gate weight, b c represents the forget gate threshold;

[0091] Memory Gate Indicates that matrices are multiplied element by element;

[0092] Output Gate O t =σ(W o ·[h t-1 ,x t ]+b o ), where W o represents the output gate weight, b o represents the output gate threshold;

[0093] The final output value is: h represents the output value, and x represents the input value, where the subscripts of h and x represent different times.

[0094] According to the principle of LSTM layer, using the idea of ​​Olden weight connection method, the calculation formula of LSTM layer comprehensive weight matrix can be derived as follows: Where W LSTM represents the comprehensive weight of the LSTM layer, W i Nor , W c Nor , W o Nor Respectively represent W f , Wi , W c , W o The normalized value.

[0095] Finally, the comprehensive weight matrix W of the above LSTM LSTM Matrix multiplication is performed with the weight matrix of the artificial neural network layer to obtain a weight matrix of a recurrent neural network. Then, the elements in the weight matrix of the recurrent neural network are normalized, and the elements in the normalized weight matrix of the recurrent neural network correspond to the importance values ​​of the main sites in the dual-enzyme starch sugar processing technology.

[0096] The importance values ​​of the main sites in the dual-enzyme starch sugar processing technology are sorted into an importance value sequence, and the sites corresponding to the first 10 of the importance value sequence are used as sites for optimizing the dual-enzyme starch sugar processing process parameters in the dual-enzyme starch sugar processing technology. The first 10 sites are shown in Table 1.

[0097] Serial number Location Site meaning Importance value X1 Glucoamylase Glucoamylase flow 0.169629 X2 TI1113_5 Liquid column temperature measurement 0.059181 X3 TC1107_2.MV Primary flash non-condensable gas regulating valve 0.051036 X4 TC1111_2.MV Secondary flash non-condensable gas regulating valve 0.040574 X5 SIC6120.MV Acid frequency control for isomerization 0.039223 X6 FIQ1125 Liquefaction enzyme flow accumulation 0.035748 X7 DSQ1104 Starch milk to dry flow accumulation 0.035432 X8 FIQ1105 One jet steam inflow accumulation 0.035225 X9 DSQ1103 The total dry flow rate of starch emulsion 0.034611 X10 SI1109B Secondary injection signal feedback 2 0.033754

[0098] Table 1: Detailed information of the top ten sites

[0099] The importance values ​​of the top 10 points in 10 trainings are plotted in a box plot, such as Figure 3 As shown. Figure 3 It can be seen that although the weights fluctuate during each training process, they remain stable as a whole, proving the effectiveness of the extended weight connection method. After process analysis of the 10 sites found, it can be seen that there is an obvious mechanism-explainable relationship between the sites such as saccharifying enzyme flow, liquefaction column temperature measurement, liquefaction enzyme flow accumulation, starch milk absolute dry flow accumulation, starch milk absolute dry flow accumulation and DE value, that is, the quality of starch sugar products; and some sites such as the primary flash evaporation non-condensable gas regulating valve, the secondary flash evaporation non-condensable gas regulating valve, the frequency control of isomerization acid, the accumulation of the first spray steam inlet flow, the secondary injection signal feedback and the DE value. There is no obvious mechanism-explainable relationship between these sites, which requires the factory to pay more attention and verify through field tests. The selected key sites can assist the factory in adjusting the parameters of the production process, thereby optimizing the starch sugar production process, improving the quality of starch sugar, and maintaining the safety of enterprise production. The present invention has a simple principle and good scalability. It can also be applied to other process industry modeling and key site screening processes, and has a very broad application prospect.

[0100] The optional implementation modes of the embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above implementation modes. Within the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical scheme of the embodiments of the present invention, and these simple modifications all belong to the protection scope of the embodiments of the present invention.

[0101] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe various possible combinations.

[0102] In addition, various implementation modes of the embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the embodiments of the present invention, they should also be regarded as the contents disclosed by the embodiments of the present invention.

Claims

1. A process optimization method based on principal component analysis and neural network, characterized in that: include: Obtain sample point data at the process location; The process midpoint and the sample point data at the process midpoint are subjected to dimensionality reduction processing using principal component analysis to obtain feature data corresponding to the process midpoint; Using principal component analysis to determine the weights of characteristic data corresponding to the sites in the process; Determining the main sites in the process according to the weights of the characteristic data corresponding to the sites in the process; Inputting the sample point data at the main sites in the process into a recurrent neural network for training, taking the percentage of reducing sugar in the dry matter of syrup as a target, training the recurrent neural network, and determining the importance values ​​of the main sites in the process according to the weight matrix of the recurrent neural network; The sites in the process for optimizing process parameters are determined according to the importance values ​​of the main sites in the process.

2. The process optimization method according to claim 1, characterized in that: The step of determining the sites in the process for optimizing process parameters according to the importance values ​​of the main sites in the process includes: According to the importance values ​​of the main sites in the process, they are sorted from large to small to obtain an importance value sequence; The first N positions of the importance value sequence were analyzed using the dual-enzyme starch sugar processing technology to obtain the positions in the process for optimizing process parameters.

3. The process optimization method according to claim 1, characterized in that: Before the process midpoint and the sample point data on the process midpoint are subjected to dimensionality reduction processing using the principal component analysis method, the following steps are also included: The sample point data at the process location are subjected to noise preprocessing.

4. The process optimization method according to claim 1, characterized in that: The process midpoint and the sample point data on the process midpoint are subjected to dimensionality reduction processing using principal component analysis to obtain feature data corresponding to the process midpoint, including: The variables of the correlation formed by combining the process midpoint and the sample point data on the process midpoint are orthogonally transformed, and the variables obtained by the orthogonal transformation are used as the characteristic data corresponding to the process midpoint.

5. The process optimization method according to claim 1, characterized in that: The method of using the principal component analysis method to determine the weight of the characteristic data corresponding to the sites in the process includes: Using principal component analysis to determine the variable component matrix, principal component eigenvalue and principal component variance contribution rate of the characteristic data corresponding to the sites in the process; Calculate the ratio of the variable component matrix to the arithmetic square root of the principal component eigenvalue as the component score coefficient matrix; Taking the principal component variance contribution rate as the weight, the component score coefficient matrix is ​​weighted averaged to obtain the comprehensive score model coefficient; The weights of the characteristic data corresponding to the sites in the process are determined according to the comprehensive score model coefficients.

6. The process optimization method according to claim 5, characterized in that: Determining the weight of the characteristic data corresponding to the site in the process according to the comprehensive score model coefficient includes: The comprehensive score model coefficients are normalized, and the processed results are used as the weights of the characteristic data corresponding to the sites in the process.

7. The process optimization method according to claim 6, characterized in that: Determining the main sites in the process according to the weights of the characteristic data corresponding to the sites in the process includes: The weights of the characteristic data corresponding to the sites in the process are sorted from large to small to obtain a ranking sequence, and the sites corresponding to the first M sites in the ranking sequence are taken as the main sites in the process.

8. The process optimization method according to claim 1, characterized in that: The recurrent neural network includes LSTM and artificial neural network layers; The sample point data at the main points in the process are input into a recurrent neural network for training, and the percentage of reducing sugar in the dry matter of the syrup is taken as a target, and the recurrent neural network is trained, including: The sample point data at the main points in the process are respectively input into the LSTM and artificial neural network layers, and the LSTM and artificial neural network layers are trained with the percentage of reducing sugar in the dry matter of the syrup as the target.

9. The process optimization method according to claim 8, characterized in that: The calculation steps of the weight matrix of the recurrent neural network include: Perform matrix multiplication on the matrices composed of the weights of each neuron in the artificial neural network layer after each training to obtain the weight matrix of the artificial neural network layer; The comprehensive weight matrix of the LSTM is matrix-multiplied with the weight matrix of the artificial neural network layer to obtain the weight matrix of the recurrent neural network.

10. The process optimization method according to claim 8, characterized in that: The calculation steps of the weight matrix of the recurrent neural network include: The sample point data at the main sites in the process are input into the artificial neural network layer in batches for training, and the average value of the weights of the neurons after the training of the sample point data at the main sites in each batch of the process is completed is used as the batch weights of the neurons in the batch training; the matrix composed of all the batch weights is matrix-multiplied to obtain the weight matrix of the artificial neural network layer; The comprehensive weight matrix of the LSTM is matrix-multiplied with the weight matrix of the artificial neural network layer to obtain the weight matrix of the recurrent neural network.

11. The process optimization method according to claim 9 or 10, characterized in that: Determining the importance values ​​of the main sites in the process according to the weight matrix of the recurrent neural network includes: The elements in the weight matrix of the recurrent neural network are normalized to obtain the importance values ​​of the main sites in the process.

12. The process optimization method according to claim 11, characterized in that: The step of determining the sites in the process for optimizing process parameters according to the importance values ​​of the main sites in the process includes: The importance values ​​of the main sites in the process are sorted to obtain an importance value sequence, and the sites corresponding to the first K positions of the importance value sequence are used as sites in the process for optimizing process parameters.

Images