Automatic control method and device for silk making process

Abstract

The invention discloses an automatic control method and device for a silk making process, and the method comprises the steps: receiving real-time input data in silk making production, including an equipment export process index which is a fixed value; inputting the real-time input data into a plurality of first prediction models to obtain a prediction value of each first prediction model; wherein different first prediction models adopt different algorithms, and the output values of all the first prediction models are all automatic control parameter values in the tobacco shred making technological process; fusing all prediction values to obtain a final automatic control parameter value; and transmitting the final automatic control parameters to a corresponding control actuator. The automatic control actuator is directly controlled by constructing the reverse model, a control rate conversion model does not need to be added between the equipment export process index and the automatic control parameter, linkage between the equipment export process index and the automatic control parameter is simplified, and the cooperation efficiency is improved.

Description

Technical Field

[0001] The present application relates to the technical field of cigarette shred production, and more specifically, to an automatic control method and device for a shred production process. Background Art

[0002] At present, the machine learning method used in silk making is mainly based on forward modeling. The relationship between the independent variables and the dependent variables in forward modeling is shown in the figure below: Figure 1 As shown, the model input is the control value X of the actuator of the closed-loop feedback control loop. f1 and other independent variables X i1 (such as raw material characteristics, equipment parameters, environmental conditions, etc.), the model output is the equipment outlet process indicator variable Y1 (such as equipment outlet material moisture, temperature, other indicator values ​​or equipment status characterization values, etc.) to predict the product quality at the equipment outlet, optimize the production process, and reduce resource waste. Although the predicted value output by the forward modeling model can achieve a high prediction accuracy, its prediction result cannot directly guide the electronic control to perform precise closed-loop feedback control. An additional control rate conversion model must be added between the forward modeling model predicted output value and the electronic control closed-loop feedback control to achieve a cascade effect by combining the closed-loop feedback realized by the machine learning forward modeling model with the automatic control actuator.

[0003] However, currently, the additional control rate conversion models are mostly simple linear models, which makes it difficult for the forward prediction model with high fit to play its advantages in the automatic control end, thus making the application of the forward modeling model subject to many restrictions. Summary of the invention

[0004] The present application provides an automatic control method and device for a silk making process, which uses the output data of the prediction model in the prior art as a part of the input of the prediction model of the present application, and uses a part of the input data of the prediction model in the prior art - the automatic control parameter value, as the output value of the prediction model of the present application, to construct a reverse model. The automatic control actuator can be directly controlled by the prediction module of the present application, without adding a control rate conversion model between the equipment outlet process indicators and the automatic control parameters, thereby simplifying the linkage between the two and improving the collaborative efficiency.

[0005] The present application provides an automatic control method for a silk making process, comprising:

[0006] Receive real-time input data in silk production, including equipment export process indicators, which are fixed values;

[0007] Input the real-time input data into a plurality of first prediction models to obtain the prediction value of each first prediction model; wherein different first prediction models adopt different algorithms, and the output values ​​of all first prediction models are the automatic control parameter values ​​in the silk making process;

[0008] Fusion of all predicted values ​​to obtain the final automatic control parameter value;

[0009] The final automatic control parameters are transmitted to the corresponding control actuator.

[0010] Preferably, the automatic control method of the silk making process further comprises retraining the corresponding first prediction model using a data set formed by actual values ​​of the equipment outlet process indicators and theoretical output values ​​of each first prediction model.

[0011] Preferably, obtaining a plurality of first prediction models specifically includes:

[0012] Get historical training data;

[0013] Using a plurality of model fitting algorithms to respectively construct second prediction models, and using historical training data to train each second prediction model, and using a preset number of second prediction models with the highest model determination coefficient as the first prediction model;

[0014] Among them, multiple model fitting algorithms come from at least two types of machine learning algorithms.

[0015] Preferably, obtaining historical training data specifically includes:

[0016] Perform time-lag alignment on the collected original historical data, where the original historical data includes input data and output data;

[0017] The aligned historical data is divided into units of preset time intervals, and first data corresponding to the preset time interval is obtained according to the historical data in each preset time interval to form a first data set in units of preset time intervals;

[0018] Remove outliers from the first data set to obtain a second data set;

[0019] The second data set is used as historical training data.

[0020] Preferably, obtaining historical training data further includes:

[0021] Taking batches as units, the input data in the second data set are verified to have a normal distribution;

[0022] If the verification fails, the outliers in the first data set are removed and the normal distribution verification is performed again;

[0023] If the normal distribution of all input data is verified, the current data set will be used as historical training data.

[0024] The present application also provides an automatic control device for a silk-making process, comprising a receiving module, a prediction module, a parameter value determination module and a control module;

[0025] The receiving module is used to receive real-time input data in silk production, including equipment export process indicators, which are fixed values;

[0026] The prediction module is used to input real-time input data into multiple first prediction models to obtain the prediction value of each first prediction model; wherein different first prediction models use different algorithms, and the output values ​​of all first prediction models are automatic control parameter values ​​in the silk making process;

[0027] The parameter value determination module is used to fuse all predicted values ​​to obtain the final automatic control parameter value;

[0028] The control module is used to transmit the final automatic control parameters to the corresponding control actuator.

[0029] Preferably, the automatic control device for the silk making process further comprises an optimization module, which is used to retrain the corresponding first prediction model using a data set formed by actual values ​​of the equipment outlet process indicators and theoretical output values ​​of each first prediction model.

[0030] Preferably, the automatic control device for the silk making process further comprises a model acquisition module, which comprises a training data acquisition module and a model selection module;

[0031] The training data acquisition module is used to acquire historical training data;

[0032] The model selection module is used to construct the second prediction models respectively using multiple model fitting algorithms, and train each second prediction model using historical training data, and select a preset number of second prediction models with the highest model determination coefficient as the first prediction model;

[0033] Among them, multiple model fitting algorithms come from at least two types of machine learning algorithms.

[0034] Preferably, the training data acquisition module includes an alignment module, a segmentation module, an outlier removal module and a data determination module;

[0035] The alignment module is used to perform time-lag alignment on the collected original historical data, where the original historical data includes input data and output data;

[0036] The segmentation module is used to segment the aligned historical data in units of preset time intervals, and obtain first data corresponding to the preset time interval based on the historical data in each preset time interval to form a first data set in units of preset time intervals;

[0037] The outlier removal module is used to remove outliers in the first data set to obtain a second data set;

[0038] The data determination module is used to use the second data set as historical training data.

[0039] Preferably, the training data acquisition module further comprises a verification module, which is used to perform normal distribution verification on the input data in the second data set in batches;

[0040] The data determination module is used to use the current data set as historical training data when the normal distribution verification of all input data is passed.

[0041] Other features and advantages of the present application will become apparent from the following detailed description of exemplary embodiments of the present application with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the application and, together with the description, serve to explain the principles of the application.

[0043] Figure 1 is a schematic diagram of a forward model of the prior art;

[0044] Figure 2 A flow chart of the automatic control method of the silk making process provided in this application;

[0045] Figure 3 A schematic diagram of the first prediction model provided in this application;

[0046] Figure 4 An example of the first data set provided for this application;

[0047] Figure 5 An example of a second data set provided for this application;

[0048] Figure 6 This is a structural diagram of the automatic control device for the silk making process provided in this application. DETAILED DESCRIPTION

[0049] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present application.

[0050] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present application, its application, or uses.

[0051] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0052] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0053] The present application provides an automatic control method and device for a silk making process, which uses the output data of the prediction model in the prior art as a part of the input of the prediction model of the present application, and uses a part of the input data of the prediction model in the prior art - the automatic control parameter value, as the output value of the prediction model of the present application, to construct a reverse model. The automatic control actuator can be directly controlled by the prediction module of the present application, without adding a control rate conversion model between the equipment outlet process indicators and the automatic control parameters, thereby simplifying the linkage between the two and improving the collaborative efficiency.

[0054] like Figure 2 As shown, the automatic control method of the silk making process provided by the present application includes:

[0055] S210: receiving real-time input data in silk production.

[0056] Please refer to Figure 3 , the real-time input data includes the equipment outlet process index Y, which is the material measurement index output by the single silk-making equipment, including parameters such as the moisture content, temperature, other index values ​​or equipment status characterization values ​​of the equipment outlet material, such as the moisture content of the loose leaf outlet material, the temperature value of the loose leaf outlet material, the moisture content of the feed outlet material, the moisture content of the drying outlet material, etc. The equipment outlet process index is usually obtained through a sensor or a measuring device. It should be noted that in this application, the equipment outlet process index used as input data is a fixed value Y 固定 .

[0057] Real-time input data also includes other independent variables X i , which are factors related to the equipment control process indicators in the upstream and downstream process flows of the silk-making equipment, such as the characteristics of various raw materials, equipment parameters, environmental conditions, etc. Other independent variables X i is the actual measured value.

[0058] S220: Input the real-time input data into a plurality of first prediction models to obtain a prediction value of each first prediction model.

[0059] Different first prediction models use different algorithms, and the output values ​​of all first prediction models are automatic control parameter values ​​in the silk making process.

[0060] There is a closed-loop feedback control loop based on the automatic control principle inside the silk-making equipment. There are one or more control actuators in the closed-loop feedback control loop to adjust the equipment outlet process indicators, that is, the electric control program outputs the automatic control parameter value of the control actuator to directly drive the equipment to change, so as to achieve the purpose of adjusting the equipment outlet process indicators. Among them, the automatic control parameter value X f It includes the frequency of the water pump for loosening and moistening leaves, the steam diaphragm valve for the hot air pipeline for loosening and moistening leaves, the damper actuator for the hot air pipeline for loosening and moistening leaves, the frequency of the feed pump, the wall temperature of the wire drying cylinder, the diaphragm valve for the main hot steam pipeline for wire drying, etc.

[0061] As an embodiment, obtaining multiple first prediction models specifically includes:

[0062] P1: Get historical training data.

[0063] As an embodiment, obtaining historical training data specifically includes:

[0064] Q1: Perform time-lag alignment on the collected original historical data, which includes the actual measured values ​​of the input data and output data used for the prediction model.

[0065] Q2: Split the aligned historical data into units of preset time intervals, and obtain first data corresponding to the preset time interval based on the historical data in each preset time interval to form a first data set in units of preset time intervals.

[0066] As an embodiment, the time interval (for example, 2 seconds) for collecting historical data is used as the minimum granularity time unit t2, and the preset time interval t1 (for example, 10 seconds) is greater than the minimum granularity time unit.

[0067] As an embodiment, if the preset time interval t1 / the minimum granularity time unit t2 is an integer, the arithmetic mean of the integer number of historical data within the preset time interval is used as the first data corresponding to the preset time interval.

[0068] If the preset time interval t1 / the minimum granularity time unit t2 is not an integer, the arithmetic mean within the preset time interval is calculated in a sliding window manner as follows:

[0069] The calculation method of window size N is:

[0070]

[0071] Among them, TRUNC(*) is a rounding function, which will truncate the decimal part of the quotient value and return an integer.

[0072] The arithmetic mean S in the current window 当前窗口 The calculation method is:

[0073]

[0074] Among them, the g value is the variable number value of the starting point of the current window, and X i is the value of the i-th variable in the current window.

[0075] Considering that the N+1th variable value (counting from the beginning of the current window) is included in the calculation of the next window, the arithmetic mean S of the next window is 下一窗口 The calculation method is:

[0076]

[0077] Figure 4 is an example of the first data set thus obtained. Figure 4 The data in the same row are the arithmetic means of all variables.

[0078] Q3: Remove outliers from the first data set and obtain the second data set.

[0079] Specifically, firstly, the fault downtime data and the data of the test production batches are removed according to the past production records, and then the historical trend graph is drawn for all the variable data to remove the outlier data that deviates from the majority distribution area.

[0080] like Figure 5 As shown in the figure, if a variable in a row of data is an outlier data, the data in that row will be deleted, that is, all the data in that row will be removed.

[0081] Q4: Use the second dataset as historical training data.

[0082] Preferably, between Q3 and Q4, the process further includes performing normal distribution verification on the input data in the second data set in batches. If the verification fails, outliers in the first data set are removed and normal distribution verification is performed again.

[0083] If the normal distribution of all input data is verified, the current data set will be used as historical training data.

[0084] Specifically, the host operation status data of equipment production, preheating, and finishing are intercepted, and a single batch of data is used as the verification unit. The input variables in all single batches in the second data set are tested for normality one by one. If the normality test p≥0.05, the data obeys the normal distribution, indicating that the input variable data distribution is reasonable; if the normality test p<0.05, the data does not obey the normal distribution, indicating that the input variable data distribution is unreasonable, and return to step Q3 to continue removing outliers until the normality test p≥0.05.

[0085] P2: Use multiple model fitting algorithms to construct second prediction models respectively, and use historical training data to train each second prediction model, and use a preset number of second prediction models with the highest model determination coefficient as the first prediction model.

[0086] Among them, multiple model fitting algorithms come from at least two types of machine learning algorithms.

[0087] As an embodiment, two types of machine learning algorithms, namely, improved neural network and improved decision tree, are selected to fit the second prediction model. As an embodiment, the improved neural network algorithm includes BP neural network, convolutional neural network (CNN), residual network (ResNet), long short-term memory network (LSTM) and generative adversarial network (GAN). The improved decision tree algorithm includes random forest, CART decision tree, gradient boosting decision tree, XGBoost and lightweight gradient boosting machine.

[0088] During training, one row of data is used as a sample size. As an example, 50% of random data is used to fit the model, and 50% of the data is used to verify the fitting accuracy.

[0089] The coefficient of determination R for different model fitting algorithms in the improved neural network 2 Sort the values ​​from high to low, select R 2 As an example, if R 2 If there are more than 2 models with a value greater than 0.75, select at most 2 and only select R 2 The number a of the two highest models, i.e. the screened improved neural network models, satisfies 0≤a≤2.

[0090] The coefficient of determination R for different model fitting algorithms in the improved decision tree 2 Sort the values ​​from high to low, select R 2 As an example, if R 2 If there are more than 2 models with a value greater than 0.75, select at most 2 and only select R 2 The number b of the two highest models, i.e. the screened decision tree improved models, satisfies 0≤b≤2.

[0091] In step S220, all input data are input into each first prediction model respectively to obtain the prediction value output by each first prediction model, that is, the automatic control parameter value.

[0092] S230: All predicted values ​​are integrated to obtain the final automatic control parameter value.

[0093] As an embodiment, all predicted values ​​are integrated to obtain the final automatic control parameter value, which specifically includes:

[0094] If the number of prediction models is greater than the first threshold (for example, 3), that is, 3≤a+b≤4, the class voting method is used to obtain the final automatic control parameter value, that is, the mean of all prediction values ​​is calculated, and then the prediction value with the largest difference from the mean is removed, and finally the arithmetic mean of the remaining prediction values ​​is used as the final automatic control parameter value.

[0095] If the number of prediction models is 2, the arithmetic mean of the two prediction values ​​is calculated as the final automatic control parameter value.

[0096] If the number of prediction models is 1, the prediction value of the prediction model is used as the final automatic control parameter value.

[0097] Preferably, in order for the first prediction model to directly drive the electrical automation control of the equipment and ensure that the feedback adjustment is not too frequent to cause overshoot, the first prediction model is calculated once every preset period (for example, 30 seconds), and other input independent variables Xi are smoothed and filtered within the preset period.

[0098] S240: Transmit the final automatic control parameters to the corresponding control actuator.

[0099] Since the equipment status is dynamically changing, and the control rules based on data mining are easily affected by factors such as weather, environment, energy, and operating methods, it is necessary to maintain the prediction model to dynamically optimize the model hyperparameters according to the actual conditions of the most recently produced batches. Based on this consideration, preferably, the automatic control method of the silk making process provided by the present application also includes:

[0100] The corresponding first prediction model is retrained using a data set formed by the actual value of the equipment outlet process indicator and the theoretical output value of each first prediction model.

[0101] As an embodiment, obtaining a theoretical output value of the first prediction model specifically includes:

[0102] R1: Calculate the mean Y of the actual values ​​of the equipment export process indicators in each batch in the third data set in batches. 实际 With fixed value Y 固定 The difference between them is taken as the first deviation of the equipment export process indicators.

[0103] R2: Calculate the second deviation between the mean of the predicted values ​​and the mean of the actual values ​​of the final automatic control parameter values ​​in each batch in the third data set in batch units.

[0104] R3: Fitting the linear relationship between the first deviation and the second deviation based on the third data set.

[0105] From the perspective of automatic control principle, the relationship between the equipment export process index and the automatic control parameter is nonlinear, and the automatic control parameter and the equipment export process index are in a first-order lag or second-order lag transfer function. However, from the perspective of the stable and controlled stage of equipment production, when the equipment export process index tends to the stable small area value change range, the equipment export process index and the automatic control parameter can be approximately simplified to a linear k proportional function:

[0106] X f 理论 -X f 实际 =(V 固定 -Y 实际 )*k+b (4)

[0107] Among them, k is the linear proportional coefficient between the automatic control parameters and the equipment outlet process indicators, and b is the bias of the linear function.

[0108] In many application scenarios, the bias b tends to be close to 0, and the above formula (4) is often simplified to:

[0109] X f 理论 -X f 实际 =(Y 固定 -Y 实际 )*k (5)

[0110] When reverse modeling is performed based on material moisture, when the equipment outlet process indicator is the material moisture at the leaf moistening outlet and the automatic control parameter is the water addition flow rate, the above formula (5) is satisfied and can be directly used for fitting.

[0111] R4: The mean X of the actual values ​​of the automatic control parameter values ​​in each batch in the fourth data set f 实际 , the difference DV between the mean and fixed value of the actual value of the equipment export process index, and the linear relationship shown in formula (4) reversely infer the theoretical prediction value X of the final automatic control parameter value f 理论 , as the theoretical output value of the first prediction model for the batch.

[0112] Therefore, the mean of the actual values ​​of the equipment export process indicators in each batch in the fourth data set and the corresponding theoretical output values ​​are combined into a sample data to form a data set, and each first prediction model is retrained using the data set to correct the first prediction model.

[0113] Preferably, after the actual production is completed, if the DV value is greater than the set threshold Fa, the mean of the actual values ​​of the equipment export process indicators obtained for the batch and the corresponding theoretical output value are taken as a sample data and stored in the data stack. When there are more than 50 data samples in the data stack, the sample data in the data stack are used to retrain each first prediction model and fine-tune the super parameters. The data stack maintains the principle of first-in-first-out. After each retraining, 25 sample data are randomly discarded until the data stack is filled with 50 samples again, and the super parameters of the first prediction model are corrected again.

[0114] Based on the above, if Figure 6 As shown, the automatic control device for the silk making process provided in the present application includes a receiving module 610 , a prediction module 620 , a parameter value determination module 630 and a control module 640 .

[0115] The receiving module 610 is used to receive real-time input data in silk production, including equipment export process indicators, which are fixed values.

[0116] The prediction module 620 is used to input real-time input data into multiple first prediction models to obtain the prediction value of each first prediction model. Different first prediction models use different algorithms, and the output values ​​of all first prediction models are automatic control parameter values ​​in the silk making process.

[0117] The parameter value determination module 630 is used to fuse all the predicted values ​​to obtain the final automatic control parameter value.

[0118] The control module 640 is used to transmit the final automatic control parameters to the corresponding control actuator.

[0119] Preferably, the automatic control device for the silk making process also includes an optimization module 650, which is used to retrain the corresponding first prediction model using a data set formed by actual values ​​of equipment outlet process indicators and theoretical output values ​​of each first prediction model.

[0120] Preferably, the automatic control device for the silk making process further includes a model acquisition module 660 , and the model acquisition module 660 includes a training data acquisition module 6601 and a model selection module 6602 .

[0121] The training data acquisition module 6601 is used to acquire historical training data.

[0122] The model selection module 6602 is used to construct the second prediction models respectively using multiple model fitting algorithms, and train each second prediction model using historical training data, and use a preset number of second prediction models with the highest model determination coefficient as the first prediction model.

[0123] Among them, multiple model fitting algorithms come from at least two types of machine learning algorithms.

[0124] Preferably, the training data acquisition module 6601 includes an alignment module, a segmentation module, an outlier removal module and a data determination module.

[0125] The alignment module is used to perform time-lag alignment on the collected original historical data, where the original historical data includes input data and output data.

[0126] The segmentation module is used to segment the aligned historical data in units of preset time intervals, and obtain first data corresponding to the preset time interval according to the historical data in each preset time interval to form a first data set in units of preset time intervals.

[0127] The outlier removal module is used to remove outliers in the first data set to obtain a second data set.

[0128] The data determination module is used to use the second data set as historical training data.

[0129] Preferably, the training data acquisition module 6601 also includes a verification module, which is used to perform normal distribution verification on the input data in the second data set in batches.

[0130] The data determination module is used to use the current data set as historical training data when the normal distribution verification of all input data is passed.

[0131] Preferably, the parameter value determination module 630 is used to calculate the mean of all prediction values ​​when the number of prediction models is greater than a threshold, then remove the prediction value with the largest difference from the mean, and finally take the arithmetic mean of the remaining prediction values ​​as the final automatic control parameter value.

[0132] Preferably, the optimization module 650 includes a theoretical value acquisition module 6501, and the theoretical value acquisition module 6501 includes a first deviation calculation module, a second deviation calculation module, a linear fitting module and an inference module.

[0133] The first deviation calculation module is used to calculate the difference between the mean of the actual value of the equipment export process indicator in each batch in the third data set and the fixed value in batch units as the first deviation of the equipment export process indicator.

[0134] The second deviation calculation module is used to calculate, in batch units, a second deviation between a mean of predicted values ​​and a mean of actual values ​​of the final automatic control parameter values ​​in each batch in the third data set.

[0135] The linear fitting module is used to fit the linear relationship between the first deviation and the second deviation based on the third data set.

[0136] The inference module is used to reversely infer the theoretical predicted value of the final automatic control parameter value based on the mean of the actual values ​​of the automatic control parameter values ​​in each batch in the fourth data set, the difference between the mean of the actual values ​​of the equipment export process indicators and the fixed value, and the linear relationship, as the theoretical output value of the first prediction model for the batch.

[0137] The beneficial effects of this application are as follows:

[0138] 1. This application uses data mining and machine learning technology to reversely establish a machine learning model, directly using the control quantity of the electric control actuator as the predicted output value of the model, and using the equipment export process indicators in the form of fixed values ​​as the input data of the prediction model for real-time prediction, and on this basis, continuous model optimization is performed to ensure that the prediction model fitted by machine learning can directly control the control actuator, directly cascade machine learning modeling with automatic control, and ensure that the high accuracy of machine learning reverse modeling prediction is directly reflected in the automatic control feedback effect. The essence of the reverse modeling method being superior to forward modeling is that there is no need to add a control rate conversion model, and there is no need to solve the mechanism model and transfer function between the automatic control parameter value and the equipment export process indicator, which can simplify the method and collaborative efficiency of the linkage between machine learning and automatic control.

[0139] 2. This application pre-processes the historical data of the reverse model in units of preset time intervals, and performs sliding average processing on the data within the unit time interval, and at the same time takes into account the problem that the unit time interval and the minimum granularity time unit cannot be divided evenly, and uses the sliding window algorithm to ensure data integrity and data smoothing characteristics. At the same time, the unit time interval and data smoothing processing method suitable for the silk-making equipment process ensure the characterization characteristics and fitting accuracy when fitting the model using historical data, and prevent the impact of mutation data on model accuracy.

[0140] 3. This application removes erroneous data by removing outliers from historical data, including silk-making equipment failure shutdown data and test production batch data. At the same time, the rationality of the independent variable data is verified using the normal distribution test tool on a batch basis, ensuring the scientific nature of the data and the final preprocessing effect of feature engineering.

[0141] 4. This application uses a variety of improved neural network models and decision tree improved models to train prediction models, and selects multiple models with the highest prediction accuracy to form a model cluster. At the same time, it uses ensemble learning method modeling and class voting method for real-time prediction, and integrates the prediction deviation values ​​of multiple single models to ensure the accuracy of reverse modeling prediction.

[0142] 5. This application uses a linear fitting function to establish the relationship between the equipment export process index and the automatic control parameters, and forms an efficient, fast and accurate model continuous optimization method. At the same time, the super parameter optimization uses batch data as the sample unit, and retraining is performed only when the DV value exceeds the threshold. At the same time, a first-in-first-out data stack and a method of randomly discarding half of the training samples are designed to correct the model, ensuring the efficiency of retraining and the convergence effect of training optimization. At the same time, randomness also avoids the problem of retraining entering the local optimal solution too early.

[0143] Although some specific embodiments of the present application have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are only for illustration, not for limiting the scope of the present application. It should be understood by those skilled in the art that the above embodiments may be modified without departing from the scope and spirit of the present application. The scope of the present application is defined by the appended claims.

Claims

1. An automatic control method for a silk making process, characterized in that: include: Receiving real-time input data in silk production, including equipment export process indicators, wherein the equipment export process indicators are fixed values; Input the real-time input data into a plurality of first prediction models to obtain a prediction value of each first prediction model; wherein different first prediction models use different algorithms, and the output values ​​of all first prediction models are automatic control parameter values ​​in the silk making process; Fusion of all predicted values ​​to obtain the final automatic control parameter value; The final automatic control parameters are transmitted to the corresponding control actuator.

2. The automatic control method for silk making process according to claim 1, characterized in that: It also includes retraining the corresponding first prediction model using a data set formed by actual values ​​of equipment outlet process indicators and theoretical output values ​​of each first prediction model.

3. The automatic control method for silk making process according to claim 1, characterized in that: Obtaining the multiple first prediction models specifically includes: Get historical training data; Using a plurality of model fitting algorithms to respectively construct second prediction models, and using the historical training data to train each second prediction model, and using a preset number of second prediction models with the highest model determination coefficient as the first prediction model; Wherein, the multiple model fitting algorithms come from at least two types of machine learning algorithms.

4. The automatic control method for silk making process according to claim 3, characterized in that: Obtain historical training data, including: Performing time-lag alignment on the collected original historical data, wherein the original historical data includes input data and output data; The aligned historical data is divided into units of preset time intervals, and first data corresponding to the preset time interval is obtained according to the historical data in each preset time interval, so as to form a first data set in units of preset time intervals; removing outliers in the first data set to obtain a second data set; The second data set is used as the historical training data.

5. The automatic control method for silk making process according to claim 4, characterized in that: Get historical training data, including: Performing normal distribution verification on the input data in the second data set in batches; If the verification fails, the outliers in the first data set are removed and the normal distribution verification is performed again; If the normal distribution of all input data is verified to be passed, the current data set is used as the historical training data.

6. An automatic control device for silk making process, characterized in that: It includes a receiving module, a prediction module, a parameter value determination module and a control module; The receiving module is used to receive real-time input data in silk production, including equipment export process indicators, which are fixed values; The prediction module is used to input the real-time input data into multiple first prediction models to obtain the prediction value of each first prediction model; wherein different first prediction models use different algorithms, and the output values ​​of all first prediction models are automatic control parameter values ​​in the silk making process; The parameter value determination module is used to fuse all predicted values ​​to obtain the final automatic control parameter value; The control module is used to transmit the final automatic control parameters to the corresponding control actuator.

7. The automatic control device for silk making process according to claim 6, characterized in that: It also includes an optimization module, which is used to retrain the corresponding first prediction model using a data set formed by actual values ​​of equipment outlet process indicators and theoretical output values ​​of each first prediction model.

8. The automatic control device for silk making process according to claim 6, characterized in that: It also includes a model acquisition module, which includes a training data acquisition module and a model selection module; The training data acquisition module is used to acquire historical training data; The model selection module is used to respectively construct second prediction models using multiple model fitting algorithms, and train each second prediction model using the historical training data, and use a preset number of second prediction models with the highest model determination coefficient as the first prediction model; Wherein, the multiple model fitting algorithms come from at least two types of machine learning algorithms.

9. The automatic control device for silk making process according to claim 8, characterized in that: The training data acquisition module includes an alignment module, a segmentation module, an outlier removal module and a data determination module; The alignment module is used to perform time-lag alignment on the collected original historical data, wherein the original historical data includes input data and output data; The segmentation module is used to segment the aligned historical data in units of preset time intervals, and obtain first data corresponding to the preset time interval based on the historical data in each preset time interval to form a first data set in units of preset time intervals; The outlier removal module is used to remove outliers in the first data set to obtain a second data set; The data determination module is used to use the second data set as the historical training data.

10. The automatic control device for silk making process according to claim 9, characterized in that: The training data acquisition module further includes a verification module, which is used to perform normal distribution verification on the input data in the second data set in batches; The data determination module is used to use the current data set as the historical training data when the normal distribution verification of all input data is passed.

Images