A method and system for incremental learning of control parameters of a cut tobacco dryer

Abstract

The present invention provides a method and system for incremental learning of the control parameters of a cut tobacco dryer. After preprocessing the original data of the working conditions of the cut tobacco dryer according to the production time and batches, a first standard data set that meets the preset conditions is screened out. After integrating the cut tobacco control parameter feature vector at the current moment in the first standard data set with the environmental parameters at the current moment into a cut tobacco control parameter feature vector, it is input into the incrementally learned random forest model to obtain the control parameter value at the current moment. By combining with the preset control parameter value, the optimized control parameter value is determined, and the cut tobacco dryer is controlled to work through the optimized control parameter value. Compared with the prior art, the present invention introduces the environmental parameters that directly affect the cut tobacco drying effect, and at the same time optimizes the random forest model by means of incremental learning, realizing precise control of the cut tobacco drying process and ensuring the stability and consistency of the cut tobacco drying effect.

Description

Technical Field

[0001] The present invention relates to the technical field of cigarette cut tobacco processing, and particularly to a method and system for incremental learning of drying machine control parameters. Background Art

[0002] In the tobacco cut tobacco processing link, the drying process plays a decisive role in the quality of tobacco. In a production environment with non-constant temperature and humidity, environmental temperature and humidity are one of the key factors affecting the drying effect. Usually, the fluctuations are relatively large, which often directly affect the drying speed of tobacco, the internal moisture distribution, and the final quality. Traditional adjustment of drying machine control parameters mostly relies on the experience and trial-and-error of operators, lacking scientific and systematic real-time optimization methods, resulting in poor stability of the quality of the tobacco cut tobacco drying production process and high production costs.

[0003] Traditional drying control methods mainly rely on empirical parameters and are difficult to effectively identify the fluctuations of environmental temperature and humidity, resulting in unstable control effects of the moisture content of cut tobacco after drying and difficult to guarantee the quality of tobacco products. Although PID control can perform basic temperature and humidity adaptive adjustment, the control logic is simple and it is difficult to cope with the complex changes in the drying process. Manual control is limited by the experience and skill level of operators and cannot achieve precise and stable control. Although the linear model can model some processes, there are many non-linear factors involved in tobacco drying, resulting in limited model prediction accuracy.

[0004] Although the existing control methods have achieved control of the drying process to a certain extent, they have the following limitations: (1) Lack of differential treatment of tobacco characteristics. Different brands and batches of tobacco have unique physical and chemical characteristics, and their drying processes respond differently to environmental temperature and humidity. Therefore, general control and optimization methods have poor effects in practical applications. (2) The existing methods are difficult to perform precise control according to the real-time changes of environmental temperature and humidity. The drying process is a dynamically changing process, and the changes in environmental temperature and humidity will directly affect the drying effect. However, the existing methods often cannot obtain and analyze environmental temperature and humidity data in real time, resulting in the inability to adjust control parameters in a timely manner, thereby affecting the stability and consistency of the drying effect.

[0005] Therefore, it is an urgent problem for those skilled in the art to provide a method and system for incremental learning of drying machine control parameters to solve the above technical problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for incremental learning of drying machine control parameters, which has clear logic, is safe, effective, reliable and easy to operate, can improve the control and optimization effects in practical applications, achieve precise control of the drying process, and ensure the stability and consistency of the drying effect.

[0007] Based on the above purpose, the technical solution provided by the present invention is as follows:

[0008] An incremental learning method for the control parameters of a cut tobacco dryer, comprising the following steps:

[0009] After obtaining the original data of the cut tobacco drying conditions, preprocess it according to the production time and batches to obtain a cut tobacco drying condition data set;

[0010] Select a first standard data set that meets the preset conditions from the cut tobacco drying condition data set;

[0011] Obtain the cut tobacco drying parameter feature vector at the current moment from the first standard data set, and integrate it with the environmental parameters at the current moment to form a cut tobacco drying control parameter feature vector;

[0012] Input the cut tobacco drying parameter feature vector into the incrementally learned random forest model to obtain the control parameter value at the current moment, and determine the optimized control parameter value according to the control parameter value at the current moment and the preset control parameter value.

[0013] Preferably, the step of preprocessing according to the production time and batches after obtaining the original data of the cut tobacco drying conditions to obtain a cut tobacco drying condition data set includes the following steps:

[0014] After obtaining the original data of the cut tobacco drying conditions, divide it according to the production time and batches to obtain the cut tobacco drying condition data of multiple batches;

[0015] Clean the cut tobacco drying condition data of each batch to obtain the first data of multiple batches;

[0016] Process the first data by a translation search method based on the maximum correlation coefficient to obtain the second data of multiple batches;

[0017] Sort the second data of multiple batches according to the batches and brands and perform vectorization processing to form a cut tobacco drying condition data set.

[0018] Preferably, the step of selecting a first standard data set that meets the preset conditions from the cut tobacco drying condition data set includes the following steps:

[0019] Calculate the process capability index of the second data of each batch at a first preset time interval respectively;

[0020] Select the third data of multiple batches whose process capability index is not greater than the preset process capability threshold;

[0021] Secondarily select the third data of each batch in a preset proportion from high to low according to the process capability index as the first standard data set.

[0022] Preferably, the incrementally learned random forest model is determined by the following steps:

[0023] After obtaining the historical data of the cut tobacco drying conditions, the sample feature set is obtained according to the preprocessing method of the original data of the cut tobacco drying conditions;

[0024] Add the trained basic regression tree to the initial random forest model to obtain the optimized random forest model;

[0025] Evaluate whether the optimized random forest model is better than the initial random forest model according to the preset evaluation index and the sample feature set;

[0026] If so, use the optimized random forest model as the random forest model after incremental learning;

[0027] If not, adjust the incremental learning parameters in the optimized random forest model until the obtained optimized random forest model is better than the initial random forest model.

[0028] Preferably, after obtaining the historical data of the cut tobacco drying conditions, the sample feature set is obtained according to the preprocessing method of the original data of the cut tobacco drying conditions, including the following steps:

[0029] Obtain multiple historical cut tobacco drying parameter feature vectors, and integrate them with historical environmental parameters to form a historical data set;

[0030] According to the screening method of the first standard data set, screen out the second standard data set that meets the preset conditions from the historical data set;

[0031] After data cleaning the second standard data set, perform vectorization processing to form the sample feature set.

[0032] Preferably, the trained basic regression tree is obtained through the following steps:

[0033] According to the historical environmental parameters in the sample feature set, calculate the mean squared error attenuation value after the split of the parent node of the split point;

[0034] Loop the above steps until the maximum mean squared error attenuation value is obtained, and split the parent node corresponding to the maximum mean squared error attenuation value;

[0035] According to the historical cut tobacco drying parameter feature vectors in the sample feature set, calculate the impurity of the mean squared error value, divide the parent node corresponding to the minimum mean squared error value impurity into left and right child nodes, and recursively split the left and right child nodes to obtain the trained basic regression tree.

[0036] Preferably, the calculating and obtaining the mean squared error attenuation value after the split of the parent node of the split point according to the historical environmental parameters in the sample feature set includes the following steps:

[0037] A preset segmentation point is used to segment the historical environmental parameters in the sample feature set into first historical environmental data and second historical environmental data;

[0038] Calculate the first prediction value and the second prediction value corresponding to the first historical environmental data and the second historical environmental data respectively;

[0039] Calculate the mean square error value of the sub-nodes of the segmentation point according to the first prediction value, the second prediction value, the first historical environmental data and the second historical environmental data;

[0040] According to the mean square error value of the parent node of the segmentation point and the mean square error value of the sub-nodes of the segmentation point, obtain the mean square error attenuation value after the splitting of the parent node of the segmentation point.

[0041] Preferably, the evaluation of whether the optimized random forest model is superior to the initial random forest model according to the preset evaluation index and the sample feature set includes the following steps:

[0042] Select the mean square error value as the evaluation index;

[0043] Evaluate the initial random forest model according to the mean square error value and the sample feature set to obtain the first performance score;

[0044] Evaluate the optimized random forest model according to the mean square error value and the sample feature set to obtain the second performance score;

[0045] Judge whether the second performance score is greater than the first performance score.

[0046] Preferably, the determination of the optimized control parameter value according to the current moment control parameter value and the preset control parameter value is specifically determined by the following formula:

[0047] ;

[0048] 0,1]

[0049] Wherein, is the optimized control parameter value, is the current moment control parameter value, is the preset control parameter value, is the control parameter weight.

[0050] A control parameter incremental learning system for a cut tobacco dryer, comprising:

[0051] An acquisition module, configured to acquire the original data of the cut tobacco drying process and preprocess it according to the production time and batches to obtain a cut tobacco drying process data set;

[0052] A screening module for screening a first standard data set that meets preset conditions from the set of tobacco drying operation data;

[0053] A vectorization module for obtaining a tobacco drying parameter feature vector at the current moment from the first standard data set and integrating it with the environmental parameters at the current moment into a tobacco drying control parameter feature vector;

[0054] An optimization control module for inputting the tobacco drying parameter feature vector into a random forest model after incremental learning to obtain a control parameter value at the current moment, and determining an optimized control parameter value according to the control parameter value at the current moment and a preset control parameter value.

[0055] The present invention provides a method for incremental learning of control parameters of a tobacco dryer. After preprocessing the obtained original tobacco dryer operation data according to production time and batches, a first standard data set that meets preset conditions is screened out; after integrating the tobacco drying parameter feature vector at the current moment in the first standard data set with the environmental parameters at the current moment into a tobacco drying control parameter feature vector, it is input into a random forest model after incremental learning to obtain a control parameter value at the current moment, and an optimized control parameter value is determined in combination with a preset control parameter value, and the tobacco dryer is controlled to operate by the optimized control parameter value.

[0056] Compared with the prior art, the present invention introduces environmental parameters that directly affect the tobacco drying effect, and at the same time optimizes the random forest model through incremental learning. By analyzing the characteristics of cut tobacco reflected by real-time data and the fluctuations of environmental parameters, the control parameters are dynamically adjusted to achieve precise control of the tobacco drying process and ensure the stability and consistency of the tobacco drying effect.

[0057] The present invention also provides a system for incremental learning of control parameters of a tobacco dryer. Since it belongs to the same technical concept as this method and solves the same technical problems, it should have the same technical effects and will not be elaborated here. Description of the Drawings

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0059] Figure 1 It is a flowchart of a method for incremental learning of control parameters of a tobacco dryer provided by an embodiment of the present invention;

[0060] Figure 2 It is a flowchart of step S1 provided by an embodiment of the present invention;

[0061] Figure 3 It is a flowchart of step S2 provided by an embodiment of the present invention;

[0062] Figure 4 It is a flowchart of determining a random forest model after incremental learning provided by an embodiment of the present invention;

[0063] Figure 5 It is a flowchart of step C1 provided by an embodiment of the present invention;

[0064] Figure 6 It is a flowchart of obtaining a trained basic regression tree provided by an embodiment of the present invention;

[0065] Figure 7 It is a flowchart of step E1 provided by an embodiment of the present invention;

[0066] Figure 8 It is a flowchart of step C3 provided by an embodiment of the present invention;

[0067] Figure 9 It is a schematic structural diagram of an incremental learning system for the control parameters of a cut tobacco dryer provided by an embodiment of the present invention. Detailed implementation manners

[0068] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0069] The embodiments of the present invention are written in a progressive manner.

[0070] The embodiments of the present invention provide a method and system for incremental learning of the control parameters of a cut tobacco dryer. It mainly solves the technical problems in the prior art that the control and optimization have poor effects in actual applications, and the cut tobacco drying process cannot be accurately controlled, resulting in poor stability and consistency of the cut tobacco drying effect.

[0071] As Figure 1 shown, a method for incremental learning of the control parameters of a cut tobacco dryer includes the following steps:

[0072] S1. After obtaining the original data of the cut tobacco drying conditions, preprocess it according to the production time and batches to obtain a cut tobacco drying condition data set;

[0073] S2. Screen out a first standard data set that meets the preset conditions from the cut tobacco drying condition data set;

[0074] S3. Obtain the cut tobacco drying parameter feature vector at the current moment from the first standard data set, and integrate it with the environmental parameters at the current moment into a cut tobacco drying control parameter feature vector;

[0075] S4. Input the feature vector of the cut tobacco drying parameters into the randomly forest model after incremental learning to obtain the control parameter value at the current moment, and determine the optimized control parameter value according to the control parameter value at the current moment and the preset control parameter value.

[0076] In step S1, obtain the original data of the cut tobacco drying working condition from the cut tobacco drying control system, sort it in ascending order of production time, and perform preprocessing in batches to obtain the cut tobacco drying working condition data set.

[0077] In this embodiment, the data acquisition system is used to collect the raw material characteristic information, the set value standard of the outlet moisture content, and the production process parameters, and online identify the required parameter information, including: the inlet moisture content, the material flow rate, the steam flow rate, the cumulative material flow rate, the steam pressure, and the steam dryness.

[0078] In step S2, screen out the relevant data that meet the conditions from the cut tobacco drying working condition data set and combine them into the first standard data set.

[0079] In this embodiment, in the cut tobacco drying production, there are obvious differences in the production states at different time periods. The parameters and control variables corresponding to the time periods with poor control indicators have poor guiding effects on the control of the cut tobacco drying outlet moisture. Eliminate this part of the data and retain the high-quality data such as the parameters and control variables corresponding to the time periods with good control indicators.

[0080] In step S3, obtain the cut tobacco drying parameter feature vector at the current moment from the first standard data set, and integrate it with the environmental parameters at the current moment collected by the data acquisition system to form the cut tobacco drying control parameter feature vector, preparing for the input model calculation.

[0081] In this embodiment, obtain the production process brand number at the current moment, denoted as , and obtain the cut tobacco drying parameter feature vector at the current moment from the first standard data set according to the brand number, denoted as , Among them, = , represents the parameters related to the equipment and sensors in the cut tobacco drying production, represents the control target of the cut tobacco drying outlet moisture, represents the control variables, usually the frequency of the hot air blower, the output opening of the moisture exhaust damper, etc.; collect the environmental temperature and humidity information of the current cut tobacco drying production workshop, denoted as , (representing the outdoor temperature, outdoor humidity, indoor temperature, and indoor humidity respectively). Then combine with the cut tobacco drying parameter feature vector at the current moment to obtain , after vectorization, it is transformed into the characteristic vector of the control parameters of the drying wire;

[0082] In step S4, the characteristic vector of the control parameter of the tofu drying is input into the random forest model after incremental learning to obtain the control parameter value at the current moment, and the optimized control parameter value is determined in combination with the preset control parameter value;

[0083] In this embodiment, the current control parameter value is the hot air blower frequency setting value at the current moment, and the preset control parameter value is the hot air blower frequency setting value of the head material of the previous batch of the same brand.

[0084] like Figure 2 As shown, preferably, step S1 includes the following steps:

[0085] A1. After obtaining the original data of the fiber drying condition, divide it according to the production time and batch to obtain the fiber drying condition data of multiple batches;

[0086] A2. Data cleaning of the drying condition data of each batch of tofu to obtain the first data of multiple batches;

[0087] A3. Processing the first data by a translation search method based on a maximum correlation coefficient to obtain a plurality of batches of second data;

[0088] A4. The second data of multiple batches are sorted by batch and brand and then quantized to form a tofu drying condition data set.

[0089] In step A1, the parameter information required for the torrefaction condition is obtained from the Taos time series database of the torrefaction control system at dawn every day, and the information is sorted in positive order according to the torrefaction production time. After the sorting is completed, the information is divided according to the production time and batch number to obtain the torrefaction condition data of multiple batches;

[0090] In step A2, the tofu drying condition data of each batch is cleaned, and the head and tail data and abnormal data are removed to obtain high-quality data in the steady-state production stage, i.e., the first data;

[0091] In this embodiment, the production data of each batch is respectively freed of null values ​​(NULL) and abnormal values ​​(normal data range: , is the data mean, is the data standard deviation); the data with large moisture deviation in the head and tail materials are eliminated (outside the outlet moisture setting value ±0.5) to ensure that the obtained data set is high-quality data in the steady-state production stage;

[0092] In step A3, the first data is processed according to a translation search method based on a maximum correlation coefficient to obtain multiple batches of second data, that is, the delay time of some parameters in the first data is adjusted;

[0093] It should be noted that it is assumed that there are two related variables A and B. The change of A will affect the change of B, and the influence is not necessarily linear. At the same time, this influence has hysteresis. As can be seen from the description of MIC, MIC is very effective for both linear and non-linear correlation relationships. However, MIC can only calculate the data corresponding in real time in the time domain. If the correlation relationship between A and B has a time delay, the calculation result of MIC will no longer be definite and effective. The translation search method based on MIC can effectively capture this type of correlation relationship. According to the characteristics and scenarios of the data, an appropriate translation search window and translation step size can be set to translate A or B forward or backward. Calculate MIC once for each translation. The maximum MIC calculated within the size of the search window is denoted as MIC, and the number of steps to obtain MIC multiplied by the step size is the time delay between A and B;

[0094] The maximal information coefficient (MIC) is developed on the basis of mutual information. The MIC method can quickly evaluate different types of associations, so as to discover a wide range of relationship types.

[0095] The larger the correlation coefficient in mutual information, the stronger the correlation between the two variables. Correlation analysis is used to study the relationship between quantitative data, including whether there is a relationship and the degree of closeness of the relationship, etc.

[0096] In this embodiment, during the moisture control of the cut tobacco dryer, it takes about 240 seconds for the material to be conveyed from the inlet conveyor to the outlet of the cut tobacco dryer, resulting in different time periods corresponding to the detection of different parameters of the same material varying with the arrival of the material. The translation search method based on the maximum correlation coefficient is used to delay different parameters detected by the sensor (displace the characteristic variable and the target outlet moisture content downward in sequence according to the transmission time of 240 seconds of the cut tobacco drum, and calculate the corresponding correlation coefficients during the process of the delay from 0 second to 240 seconds respectively, and obtain the delay time corresponding to the maximum correlation coefficient), so as to align the data of each parameter at the time level;

[0097] In step A4, after sorting multiple batches of second data according to batches and brands, vectorization processing is performed to convert the data type to obtain a cut tobacco drying condition data set;

[0098] It should be noted that data vectorization is a technology that converts non-numerical data into numerical data. It is widely used in the fields of machine learning and deep learning. Its advantages are: unifying data representation, facilitating subsequent machine learning modeling; extracting effective features of data, enhancing the learning ability of the model; supporting complex data types such as text, images, audio, etc.; providing a basis for the end-to-end machine learning process.

[0099] In this embodiment, after sorting the aligned second data by brand in batches, the corresponding feature types are processed and converted into corresponding strings and floating-point numbers (i.e., data vectorization processing) to form a data set of the cut tobacco drying working conditions.

[0100] As Figure 3 shown, preferably, step S2 includes the following steps:

[0101] B1. Calculate the process capability index of the second data of each batch at a first preset time interval respectively;

[0102] B2. Screen out the third data of multiple batches whose process capability index is not greater than the preset process capability threshold;

[0103] B3. Secondary screen out the third data of each batch with a preset proportion from high to low according to the process capability index as the first standard data set.

[0104] In steps B1 to B3, by calculating the process capability index of each batch at a preset time interval, high-quality data such as parameters and control variables corresponding to the time periods with better control indicators are screened out from the data set of the cut tobacco drying working conditions; that is, the effect of the control indicators is represented by the process capability index.

[0105] It should be noted that the process capability index CPK refers to the degree to which the process capability meets the requirements of product quality standards (such as specification range, etc.). It is also called the process capability index, which refers to the actual processing ability of the process in a certain period of time under the control state (stable state). It is the inherent ability of the process, or the ability of the process to ensure quality. The process referred to here is the process of the comprehensive action of five basic quality factors: operator, machine, raw material, process method, and production environment, that is, the production process of product quality;

[0106] In this embodiment, the first preset time is selected as 180 seconds, the process capability index of the outlet moisture at different time intervals is calculated according to the time interval of 180 seconds, and then the process capability indexes are sorted. The data with a process capability index less than or equal to 1.33 (preset process capability threshold) are excluded, and the top 30% (preset proportion) of the data with high process capability indexes of each batch are left as the first standard data set.

[0107] As Figure 4 shown, preferably, the random forest model after incremental learning is determined through the following steps:

[0108] C1. After obtaining the historical data of the cut tobacco drying working conditions, obtain the sample feature set according to the preprocessing method of the original data of the cut tobacco drying working conditions;

[0109] C2. Add the trained basic regression tree to the initial random forest model to obtain an optimized random forest model;

[0110] C3. Evaluate whether the optimized random forest model is better than the initial random forest model according to the preset evaluation index and the sample feature set;

[0111] C41. If so, use the optimized random forest model as the random forest model after incremental learning;

[0112] C42. If not, adjust the incremental learning parameters in the optimized random forest model until the obtained optimized random forest model is better than the initial random forest model.

[0113] In step C1, obtain the historical data of the cut tobacco drying working conditions from the Taos time series database of the cut tobacco drying control system, and obtain the corresponding sample feature set according to the preprocessing method of the original data of the cut tobacco drying working conditions;

[0114] In this embodiment, obtain the historical data of the previous 5 minutes of the current moment, and process each piece of data into a feature vector using the processing method of step S3 to obtain a sample data set;

[0115] In step C2, construct an initial random forest model, add the trained basic regression tree (i.e., incremental learning) to obtain an optimized random forest model;

[0116] In step C3, evaluate the model performance of the initial random forest model and the optimized random forest model according to the preset evaluation index and the sample data set;

[0117] In steps C41 and C42, when the performance of the optimized random forest model is better than the performance of the initial random forest model, use it as the random forest model after incremental learning; when the performance of the initial random forest model is better than the performance of the optimized random forest model, continuously adjust the incremental learning parameters until the performance of the optimized random forest model is better than the performance of the initial random forest model and then stop.

[0118] In this embodiment, step C2 is specifically: on the basis of the initial random forest model add the basic learner regression tree ; then perform incremental learning training on the basic learner regression tree according to the process from step E1 to step E3 to obtain an optimized random forest model ;

[0119] The prediction result of the optimized random forest model is the average of the prediction results of all trees. Assume that the original forest contains trees, and after adding trees, the predicted value of the optimized random forest is:

[0120] ;

[0121] ;

[0122] in, is the first The predicted value of a tree, is the first The predicted value of a tree, is the weight of the experience value and incremental learning result, usually 0.5, the weight of the predicted value of the reinforcement incremental learning;

[0123] Step C42 is specifically as follows: when the performance of the initial random forest model is better than that of the optimized random forest model, the incremental learning parameters, i.e., weights, are continuously adjusted. , repeat steps C2 to C3 until the performance of the optimized random forest model is better than that of the initial random forest model.

[0124] like Figure 5 As shown, preferably, step C1 comprises the following steps:

[0125] D1. Obtain multiple historical silk-cooking parameter feature vectors and integrate them with historical environmental parameters to form a historical data set;

[0126] D2. According to the screening method of the first standard data set, a second standard data set that meets the preset conditions is screened from the historical data set;

[0127] D3. Data cleaning The second standard data set is vectorized to form a sample feature set.

[0128] In step D1 to step D3, the historical wire-drying parameter feature vector of the wire-drying condition is obtained from the Taos time series database of the wire-drying control system, and the historical environmental parameters stored in the data acquisition system are integrated into the historical wire-drying control parameter feature vector; then, the second standard data set is screened out by the process capability index screening method in step B1 to step B3; data cleaning is performed in the manner of step A2, and after removing the head and tail data and abnormal data, the numerical data is converted into floating point type and the string data is converted into classification index in the manner of step A4 to construct a sample feature set;

[0129] In this embodiment, step D3 specifically includes: removing abnormal values ​​and null values ​​from the data in the second standard data set, converting numerical data into floating point data, converting string data into classification indicators, and constructing a new sample set. ,Feature Set ,for , the sample feature set is segmented into subsets by considering all its possible values .

[0130] As Figure 6 shown, preferably, the trained basic regression tree in step C2 is obtained through the following steps:

[0131] E1. Calculate and obtain the mean square error attenuation value after the split of the parent node of the split point according to the historical environment parameters in the sample feature set;

[0132] E2. Loop the above steps until the maximum mean square error attenuation value is obtained, and split the parent node corresponding to the maximum mean square error attenuation value;

[0133] E3. Calculate and obtain the impurity of the mean square error value according to the historical cut tobacco drying parameter feature vector in the sample feature set, divide the parent node corresponding to the minimum impurity of the mean square error value into left and right child nodes, recursively split the left and right child nodes, and obtain the trained basic regression tree.

[0134] From step E1 to step E3, the splitting of the decision tree is carried out in the way of CART tree as a whole. However, considering the requirement of identifying environmental parameters (temperature and humidity information), when processing features, the environmental temperature and humidity data P is used as the leading data to be processed first to enhance its weight; after the environmental temperature and humidity features are processed, for the subsets of the sample feature set, for other features, the impurity after the split of the regression tree is also calculated according to MSE, and the optimal split point is selected for other features in the way of the optimal MSE.

[0135] It should be noted that the mean square error (MSE) is a measure reflecting the degree of difference between the estimator and the estimated quantity.

[0136] In this embodiment, step E3 is specifically: calculate the impurity after the split of the regression tree according to MSE through the following formula:

[0137] ;

[0138] where is the mean square error on the subset .

[0139] Other features select the optimal split point in the way of the optimal MSE. After obtaining , determine the best split feature through the following formula :

[0140] ;

[0141] After determining the best split feature After that, according to the value of this feature, the current node is divided into left and right child nodes, and the left and right child nodes are recursively split.

[0142] As Figure 7 shown, preferably, step E1 includes the following steps:

[0143] F1. Preset a splitting point, and split the historical environmental parameters in the sample feature set into first historical environmental data and second historical environmental data;

[0144] F2. Calculate the first predicted value and the second predicted value corresponding to the first historical environmental data and the second historical environmental data respectively;

[0145] F3. Calculate the mean square error value of the child nodes of the splitting point according to the first predicted value, the second predicted value, the first historical environmental data and the second historical environmental data;

[0146] F4. Obtain the mean square error attenuation value after splitting the parent node of the splitting point according to the mean square error value of the parent node of the splitting point and the mean square error value of the child nodes of the splitting point.

[0147] Steps F1 to F4 split the historical environmental parameters into first historical environmental data and second historical environmental data by setting a splitting point, substitute them into the random forest model to obtain the corresponding first predicted value and second predicted value; calculate the mean square error value of the child nodes of the splitting point according to the first historical environmental data, the first predicted value, the second environmental data and the second predicted value; take the difference between the mean square error value of the parent node of the splitting point and the mean square error value of the child nodes of the splitting point as the mean square error attenuation value after splitting the parent node of the splitting point.

[0148] In this embodiment, taking the outdoor temperature as an example, usually the temperature distribution in the data set is 10~35°C, and the median 22.5 is selected as the splitting point, and the feature data is divided into and ( 2.5, 2.5), and the control quantity is predicted;

[0149] Through the following formula, calculate the predicted values and of the control quantity under the conditions of and respectively:

[0150] ;

[0151] Then, calculate under this division rule through the following formula:

[0152] ;

[0153] Calculate the of the parent node specifically as follows:

[0154] Calculate the mean temperature before splitting: ;

[0155] Calculate based on the data samples before splitting);

[0156] And calculate the MSE attenuation value after splitting: - 。 ;

[0157] Step E2 is specifically as follows: According to the process rules of steps F1 to F4, find different splitting points in a loop, and find the splitting point that maximizes the MSE attenuation value for splitting.

[0158] As Figure 8 shown, preferably, step C3 includes the following steps:

[0159] G1. Select the mean square error value as the evaluation index;

[0160] G2. Evaluate the initial random forest model according to the mean square error value and the sample feature set to obtain the first performance score;

[0161] G3. Evaluate the optimized random forest model according to the mean square error value and the sample feature set to obtain the second performance score;

[0162] G4. Judge whether the second performance score is greater than the first performance score.

[0163] In steps G1 to G4, using MSE as the evaluation index, input the sample data set into the initial random forest model and the optimized random forest model respectively to obtain the corresponding first performance score and second performance score, and judge whether the second performance score is greater than the first performance score;

[0164] In this embodiment, calculate the initial model performance score according to and the sample data set, use the same evaluation index and data set to evaluate the optimized random forest model to obtain the performance score ; Judge whether it is greater than (where the performance score ). )

[0165] Preferably, in step S4, the optimized control parameter value is determined according to the current control parameter value and the preset control parameter value, and is specifically determined by the following formula:

[0166] ;

[0167] 0,1];

[0168] Among them, is the optimized control parameter value, is the control parameter value at the current moment, is the preset control parameter value, is the control parameter weight.

[0169] In the actual application process, the cut tobacco parameter feature vector is input into the random forest model after incremental learning to obtain the set value of the hot air blower frequency at the current moment , and at the same time, the set values of the hot air blower frequencies of the head materials of the previous batches of the same brand are loaded, denoted as . Combining the fitting results of the random forest model after incremental learning and the set values of the head materials of the previous batches of the same brand, a weighted optimized hot air blower frequency is given.

[0170] After step S4, communication with the PLC is established through the OPC Serve, and the optimized hot air blower frequency is sent to the PLC program segment for implementation control, and the PLC program is further corrected by combining real-time data incremental learning.

[0171] After implementing the above method through the PLC program, the unsteady state time of cut tobacco is reduced from 151 s to 128 s, a total reduction of 15.2%; the unsteady state dry head amount is reduced from 11.495 to 7.8, a reduction of 32%, and the control effect is significantly improved. The intelligent control of the cut tobacco machine process parameters is realized, the time of the unsteady state stage of the cut tobacco head material is reduced, the cut tobacco loss is reduced, the cut tobacco utilization rate is improved, and the production efficiency of the enterprise is increased.

[0172] As Figure 9 shown, a cut tobacco machine control parameter incremental learning system includes:

[0173] An acquisition module, configured to preprocess the original data of the cut tobacco working condition according to the production time and batches to obtain a cut tobacco working condition data set;

[0174] A screening module, configured to screen out a first standard data set that meets the preset conditions from the cut tobacco working condition data set;

[0175] A vectorization module, configured to obtain a feature vector of the cut tobacco drying parameters at the current moment from a first standard dataset, and integrate it with the environmental parameters at the current moment into a feature vector of the cut tobacco drying control parameters;

[0176] An optimization control module, configured to input the feature vector of the cut tobacco drying parameters into the random forest model after incremental learning to obtain the control parameter value at the current moment, and determine the optimized control parameter value according to the control parameter value at the current moment and the preset control parameter value.

[0177] In the actual application process, a cut tobacco dryer control parameter incremental learning system is also disclosed. Each module in this system corresponds to each step in the implementation method, solves the same technical problems, and has the same beneficial effects, which will not be elaborated here.

[0178] In the embodiments provided in the present application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are only illustrative. For example, the division of modules is only a logical function division. In actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed with each other can be through some interfaces, and the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0179] In addition, in each embodiment of the present invention, each functional module can be all integrated in one processor, or each module can be separately used as a device, or two or more modules can be integrated in one device; each functional module in each embodiment of the present invention can be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.

[0180] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed through program instructions and related hardware. The foregoing program instructions can be stored in a computer-readable storage medium. When the program instructions are executed, the steps including the above method embodiments are executed; and the foregoing storage medium includes: various media that can store program codes such as mobile storage devices, read-only memories (ROMs), magnetic disks, or optical discs.

[0181] It should be understood that in the present application, if "system", "device", "unit", and / or "module" are used, it is only a method for distinguishing different components, elements, parts, portions, or assemblies at different levels. However, if other words can achieve the same purpose, then the word can be replaced by other expressions.

[0182] As shown in this application and the claims, unless the context clearly indicates otherwise, words such as "a", "an", "one", and / or "the" are not specifically singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. A method or device may also include other steps or elements. An element defined by the statement "comprising one..." does not exclude the existence of other identical elements in the process, method, article, or device that includes the element.

[0183] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features.

[0184] If a flowchart is used in this application, the flowchart is used to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the previous or subsequent operations are not necessarily executed precisely in sequence. On the contrary, the steps can be processed in reverse order or simultaneously. At the same time, other operations can also be added to these processes, or one or more steps can be removed from these processes.

[0185] The above has introduced in detail a method and system for incremental learning of control parameters of a cut tobacco dryer provided by the present invention. The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather will conform to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for incremental learning of control parameters of a tofu drying machine, characterized in that: The steps include: After obtaining the raw data of the fiber drying condition, pre-processing is performed according to the production time and batch to obtain a fiber drying condition data set; Filtering out a first standard data set that meets a preset condition from the tofu drying condition data set; Obtaining a current moment's tow-bread baking parameter feature vector from the first standard data set, and integrating it with the current moment's environmental parameters to form a tow-bread baking control parameter feature vector; Inputting the tofu baking parameter feature vector into the random forest model after incremental learning to obtain the control parameter value at the current moment, and determining the optimized control parameter value according to the control parameter value at the current moment and the preset control parameter value; The method of obtaining the raw data of the fiber drying condition and pre-processing it according to the production time and batch to obtain the fiber drying condition data set includes the following steps: After obtaining the original data of the fiber drying condition, the data is divided according to the production time and batches to obtain the fiber drying condition data of multiple batches; Data cleaning of the tofu drying condition data of each batch to obtain first data of multiple batches; Processing the first data by a translation search method based on a maximum correlation coefficient to obtain a plurality of batches of second data; The second data of the plurality of batches are sorted by batch and brand and then quantized to form a tow-baking condition data set; The random forest model after incremental learning is determined by the following steps: After acquiring the historical data of the tofu drying condition, a sample feature set is obtained according to the preprocessing method of the original data of the tofu drying condition; Add the trained basic regression tree to the initial random forest model to obtain the optimized random forest model; Evaluate whether the optimized random forest model is better than the initial random forest model according to the preset evaluation index and the sample feature set; If yes, the optimized random forest model is used as the random forest model after incremental learning; If not, the incremental learning parameters in the optimized random forest model are adjusted until the obtained optimized random forest model is better than the initial random forest model.

2. The incremental learning method for control parameters of a tofu drying machine according to claim 1, characterized in that: The step of selecting a first standard data set that meets a preset condition from the fiber drying condition data set comprises the following steps: Calculating the process capability index of each batch of the second data at a first preset time interval respectively; Screening out a plurality of batches of third data whose process capability index is not greater than a preset process capability threshold; According to the process capability index from high to low, a preset proportion of the third data of each batch is secondary screened out as the first standard data set.

3. The incremental learning method for control parameters of a tofu drying machine according to claim 2, characterized in that: After the historical data of the fiber drying condition is obtained, the sample feature set is obtained according to the preprocessing method of the original data of the fiber drying condition, including the following steps: Acquire multiple historical tofu drying parameter feature vectors, and integrate them with historical environmental parameters to form a historical data set; According to the screening method of the first standard data set, a second standard data set that meets the preset conditions is screened from the historical data set; After data cleaning, the second standard data set is vectorized to form the sample feature set.

4. The incremental learning method for control parameters of a tofu drying machine according to claim 3, characterized in that: The trained basic regression tree is obtained by the following steps: According to the historical environment parameters in the sample feature set, a mean square error attenuation value after the parent node of the segmentation point is split is calculated; The above steps are repeated until the maximum value of the mean square error attenuation is obtained, and the parent node corresponding to the maximum value of the mean square error attenuation is split; According to the historical tow-baking parameter feature vector in the sample feature set, the mean square error value impurity is calculated, the parent node corresponding to the minimum mean square error value impurity is divided into left and right child nodes, and the left and right child nodes are recursively split to obtain the trained basic regression tree.

5. The incremental learning method for control parameters of a tofu drying machine according to claim 4, characterized in that: The step of calculating and obtaining the mean square error attenuation value after the parent node of the segmentation point is split according to the historical environment parameters in the sample feature set comprises the following steps: Preset a segmentation point to segment the historical environment parameters in the sample feature set into first historical environment data and second historical environment data; Calculating a first prediction value and a second prediction value corresponding to the first historical environment data and the second historical environment data respectively; Calculating a mean square error value of a segmentation point subnode according to the first prediction value and the second prediction value and the first historical environment data and the second historical environment data; According to the mean square error value of the parent node of the segmentation point and the mean square error value of the child node of the segmentation point, the mean square error attenuation value after the parent node of the segmentation point is split is obtained.

6. The incremental learning method for control parameters of a tofu drying machine according to claim 1, characterized in that: The step of evaluating whether the optimized random forest model is better than the initial random forest model according to the preset evaluation index and the sample feature set comprises the following steps: Select the mean square error value as the evaluation index; Evaluating the initial random forest model according to the mean square error value and the sample feature set to obtain a first performance score; Evaluating the optimized random forest model according to the mean square error value and the sample feature set to obtain a second performance score; It is determined whether the second performance score is greater than the first performance score.

7. The incremental learning method for control parameters of a tofu drying machine according to claim 1, characterized in that: The optimized control parameter value is determined according to the current control parameter value and the preset control parameter value, and is specifically determined by the following formula: ； 0,1] in, is the optimized control parameter value, is the control parameter value at the current moment, is the preset control parameter value, To control the parameter weight.

8. A control parameter incremental learning system for a tofu drying machine, characterized in that: include: An acquisition module is used to acquire the raw data of the fiber drying condition and pre-process it according to the production time and batch to obtain a fiber drying condition data set; A screening module, used for screening out a first standard data set that meets a preset condition from the tofu drying condition data set; A vectorization module, used for obtaining a characteristic vector of wire-baking parameters at the current moment from the first standard data set, and integrating it with the environmental parameters at the current moment into a characteristic vector of wire-baking control parameters; An optimization control module, used for inputting the tow baking parameter feature vector into the random forest model after incremental learning to obtain the control parameter value at the current moment, and determining the optimized control parameter value according to the control parameter value at the current moment and the preset control parameter value; The acquisition module is specifically used to acquire the original data of the tofu drying condition and then divide it according to production time and batches to acquire the tofu drying condition data of multiple batches; Data cleaning of the tofu drying condition data of each batch to obtain first data of multiple batches; Processing the first data by a translation search method based on a maximum correlation coefficient to obtain a plurality of batches of second data; The second data of the plurality of batches are sorted by batch and brand and then quantized to form a tow-baking condition data set; The random forest model after incremental learning is determined by the following steps: After acquiring the historical data of the tofu drying condition, a sample feature set is obtained according to the preprocessing method of the original data of the tofu drying condition; Add the trained basic regression tree to the initial random forest model to obtain the optimized random forest model; Evaluate whether the optimized random forest model is better than the initial random forest model according to the preset evaluation index and the sample feature set; If yes, the optimized random forest model is used as the random forest model after incremental learning; If not, the incremental learning parameters in the optimized random forest model are adjusted until the obtained optimized random forest model is better than the initial random forest model.

Images