CN109034388A - A kind of prediction model of cigarette material and mainstream smoke constituents based on Genetic Algorithm Optimized Neural Network - Google Patents

Abstract

A prediction model of cigarette material and mainstream smoke components based on a genetic algorithm optimized neural network, including the following steps: sample preprocessing; obtaining the optimal weight and threshold parameters of the neural network based on the genetic algorithm; obtaining the optimal weight and threshold parameters based on the genetic algorithm. Weights and thresholds, construct a neural network, and train the neural network model; verify the trained neural network model, and evaluate the actual effect of the model. Compared with the neural network without optimization algorithm, the neural network model based on genetic algorithm optimization first uses the genetic algorithm to select the weight and threshold that minimize the model error as the initial parameters of the training neural network, which can avoid the model from falling into a local optimal solution , but the global optimal solution cannot be obtained.

Description

Cigarette material and mainstream smoke components based on genetic algorithm optimization neural network predictive model

technical field

The invention relates to a predictive model, which belongs to the field of industrial design and production. The genetic algorithm is used to optimize the neural network to establish the mapping relationship between cigarette materials and mainstream smoke components, so as to rationally adjust the formula of cigarette materials for the cigarette manufacturing industry and produce cigarettes that meet specifications. A model is provided, specifically a neural network-based predictive model for cigarette material and mainstream smoke components.

Background technique

In the tobacco manufacturing industry, the composition and content of harmful gases released by tobacco has always been a concern of tobacco manufacturing companies and consumers. In order to optimize product design, speed up production efficiency, and produce qualified and low-hazard products, the tobacco manufacturing industry needs to use statistics-related knowledge to conduct canonical correlation analysis, principal component analysis, and partial correlation analysis on tobacco raw materials, and obtain the results related to mainstream smoke. The most relevant cigarette material is then modeled appropriately for the principal components and mainstream smoke from the analysis to assist in material extraction and production during tobacco manufacturing. With the development of information technology, the computing and processing capabilities of computers have been continuously improved, and it has become a trend to use computers to conduct statistical analysis, processing and modeling of data. Computers can be used for rapid analysis and modeling, but traditional statistical analysis methods are based on a small amount of data and people's prior knowledge for data analysis and modeling. This modeling method is highly targeted, and the model's Applicability is poor.

In recent years, deep learning has risen again. Relying on complex network structure and self-learning ability, neural network can better fit various seemingly irregular data, which not only provides a new modeling method for data modeling with no obvious rules. , and the fitting effect is better than the traditional statistical modeling method. BP neural network is a learning algorithm that generalizes the learning rules and performs weight training on nonlinear differentiable functions. The learning process is divided into two stages: (1) information forward propagation; (2) error back propagation. In the training process of BP neural network, if the difference between the output result of one time and the expected output result exceeds a certain standard, and the number of training times does not reach the set maximum number of training times, then the error is returned in some form , and correct the weights of each layer, so that the output value is close to the expected output value step by step, until the error is reduced to an acceptable level or the number of training times reaches the preset number of times. At present, neural network-related technologies have begun to be applied in the model of the relationship between cigarette sensory evaluation and mainstream smoke components, but there is no precedent for using genetic algorithm to optimize BP neural network to establish a model of cigarette materials and mainstream smoke.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention proposes a method for establishing a prediction model of cigarette materials and mainstream smoke components based on genetic algorithm optimization of BP neural network.

The invention discloses a method for establishing a prediction model and a use model of cigarette materials and mainstream smoke components. First, the genetic algorithm is used to select the best initial weight and threshold based on the structure of the neural network; The initial weights and thresholds train the neural network to avoid the model from falling into the local optimal solution instead of the global optimal solution; finally, based on the trained model, use real data as input to evaluate whether the model can be applied to actual production.

The technical solution adopted by the present invention is: a prediction model of cigarette material and mainstream smoke components based on genetic algorithm optimization neural network, characterized in that it comprises the following steps:

Step 1: Sample preprocessing;

Step 2: Obtain the optimal weight and threshold parameters of the neural network based on the genetic algorithm;

Step 3: Construct a neural network based on the optimal weight and threshold obtained by the genetic algorithm, and train the neural network model;

Step 4: Verify the trained neural network model and evaluate the actual effect of the model.

Further, the specific implementation of step 1 includes the following sub-steps:

Step 1.1: Remove data records with missing values and data that obviously do not conform to reality in the sample;

Step 1.2: The data in the original data is normalized according to the following expression, x represents the original data, y represents the data after normalization, Minvalue and MaxValue represent the minimum and maximum values of the original data, y=(x- MinValue)/(MaxValue-Minvalue);

Step 1.3: Divide the normalized data into training set and test set.

Further, the specific implementation of step 2 includes the following sub-steps:

Step 2.1: Define the constant parameters of the neural network, the number of input layer nodes INPUT_NODE, the number of hidden layer nodes LAYER1_NODE, the number of output layer nodes OUTPUT_NODE, the basic learning rate LEARNING_RATE_BASE, the decay rate LEARNING_RATE_DECAY of the learning rate, and the regularization item describing the complexity of the model in The coefficient REAULARIZATION_RATE in the loss function, the number of training rounds TRAINING_SETPS, and the moving average decay rate MOVING_AVERAGE_DECAY;

Step 2.2: Define the activation function tf.nn.relu(), the loss function loss(), and the optimization algorithm tf.train.GradientDescentOptimizer() for optimizing the loss function;

Step 2.3: Set the evolutionary algebra, population size, crossover and mutation probability of the population, and encode the initial weight and threshold of neurons in real numbers;

Step 2.4: Calculate the fitness of the population, and select the best individuals from the current population;

Step 2.5: The population is selected, crossed, and mutated;

Step 2.6: Judging whether the optimization goal is reached, if the optimization goal is reached, select the best weight and threshold to execute the next step, otherwise execute step 2.4;

Further, the specific implementation of step 3 includes the following sub-steps:

Step 3.1: Construct a BP neural network based on the optimal weights and thresholds and other neural network parameters obtained in the previous step;

Step 3.2: Train the neural network model;

Step 3.3: Calculate the error between the output value of the trained neural network model and the real value in the data set, and update the weights and thresholds of each layer of the neural network layer by layer according to the error;

Step 3.4: Determine whether the obtained model has reached the expected accuracy or the maximum number of iterations, if so, execute the next step, otherwise return to step 3.2;

Further, the specific implementation of step 4 includes the following sub-steps:

Step 4.1: Input the data in the test set into the neural network model obtained in step 3, and obtain the corresponding output value;

Step 4.2: Calculate the error between the model output value and the real value. If the error meets the expected requirements, the model can be applied in actual production. Otherwise, return to step 2, adjust various parameters, and retrain the model.

The beneficial effects and characteristics of the present invention are: the present invention is the first predictive model based on genetic algorithm optimization of BP neural network from cigarette material to mainstream smoke components. The traditional linear regression fitting curve often underfits when dealing with the nonlinear mapping relationship between multiple inputs and multiple outputs. The BP network can learn and store a large number of input-output pattern mapping relationships without revealing the mathematical equations describing the mapping relationship in advance. In the process of using the BP neural network to train the model, the weights are repeatedly corrected and the model is continuously optimized, which can make the model fit the training data with arbitrary precision. There is a certain complex nonlinear mapping relationship between cigarette materials and mainstream smoke components, so using genetic algorithm to optimize BP neural network to establish a prediction model between cigarette materials and mainstream smoke is not only more suitable than traditional regression methods, but also because Unoptimized neural network model.

Description of drawings

Fig. 1 is the principle framework structure of the embodiment of the present invention;

Figure 2 is the fitness change curve during the process of selecting the optimal neural network initialization parameters based on the genetic algorithm;

Figure 3 and Figure 4 are the comparison charts of the prediction errors of tar oil and CO without using the neural network model optimized by the genetic algorithm;

Figure 5 and Figure 6 are comparisons of the prediction accuracy of tar oil and CO by the neural network model optimized by genetic algorithm.

Detailed ways

The present invention will be further described below in conjunction with accompanying drawing:

In this example, the entire experimental data set is divided into two parts, one part is used as the training set of BP neural network, and the other part is used as the test set of BP neural network. The data in the training set takes 2/3-4/5 of the total data volume . Definition D is an n×m data set, each row represents a data record, and each column represents an attribute. Let sets A and B denote the training set and test set obtained by dividing the dataset, respectively. First, based on the training set A, the optimal neural network weights and thresholds are obtained based on the genetic algorithm, and then the optimal neural network weights and thresholds are used as the initial parameters for training the BP neural network model to train the neural network. When the output of the network is in line with the expected If the difference between the output results reaches a certain standard or reaches the maximum number of iterations, it is considered that the BP neural network has been trained, and then the test set B is used to test the prediction accuracy of the trained model on the new data set.

The invention proposes a method for establishing a prediction model and a use model of cigarette materials and mainstream smoke components based on genetic algorithm optimization BP neural network. Including the following steps (refer to Figure 1):

Step 1: Sample preprocessing;

Step 1.1: Remove data records with missing values and data that obviously do not conform to reality in the sample;

Step 1.2: The data in the original data is normalized according to the following expression, x represents the original data, y represents the data after normalization, Minvalue and MaxValue represent the minimum and maximum values of the original data, y=(x- MinValue)/(MaxValue-Minvalue);

Step 1.3: Divide the normalized data into training set and test set.

The specific implementation process of the embodiment is described as follows:

For the value x of each attribute in each data record in the data set, the normalized value y is obtained according to the method of y=(x-MinValue)/(MaxValue-Minvalue), and each calculated y value is used as the corresponding attribute The value of , thus forming a new n×m data set.

Divide 2/3-4/5 of the data records in the normalized data set into the training set, and divide the remaining data into the test set. Avoid the influence of factors considered in the process of data set division on the experimental results.

Step 2: Obtain the optimal weight and threshold parameters of the neural network based on the genetic algorithm;

Step 2.1: Define the constant parameters of the neural network, the number of input layer nodes INPUT_NODE, the number of hidden layer nodes LAYER1_NODE, the number of output layer nodes OUTPUT_NODE, the basic learning rate LEARNING_RATE_BASE, the decay rate LEARNING_RATE_DECAY of the learning rate, and the regularization item describing the complexity of the model in The coefficient REAULARIZATION_RATE in the loss function, the number of training rounds TRAINING_SETPS, and the moving average decay rate MOVING_AVERAGE_DECAY;

Step 2.2: Define the activation function tf.nn.relu(), the loss function loss(), and the optimization algorithm tf.train.GradientDescentOptimizer() for optimizing the loss function;

Step 2.3: Set the evolutionary algebra, population size, crossover and mutation probability of the population, and encode the initial weight and threshold of neurons in real numbers;

Step 2.4: Calculate the fitness of the population, and select the best individuals from the current population;

Step 2.5: The population is selected, crossed, and mutated;

Step 2.6: Judging whether the optimization goal is reached, if the optimization goal is reached, select the best weight and threshold to execute the next step, otherwise execute step 2.4;

The specific implementation process of the embodiment is described as follows:

Before using the genetic algorithm to obtain the optimal weight and threshold, first determine the number of input layer nodes INPUT_NODE of the neural network, the number of hidden layers and the number of hidden layer nodes LAYER1_NODE, the number of output layer nodes OUTPUT_NODE and a series of parameters required by the neural network and The parameters required in the genetic algorithm, such as: evolutionary algebra, population size, etc. Then initialize the initial weights and thresholds of neurons to encode real numbers, calculate the fitness of the population one by one, and perform neuron selection, crossover and mutation operations in sequence. Every time a round of selection, crossover and mutation is performed, the population is updated, and the fitness of the population needs to be recalculated. The fitness of the population is an index that describes the degree of approximation of the population to the optimal solution of the problem. Here, the fitness of each individual is calculated by calculating the sum of the errors of all the data in the training set and the predicted value of the model. According to the fitness of individuals, some individuals with smaller fitness values are selected to form a new population. In the new population, selection, crossover and mutation are continued to make the population iterate continuously and approach the optimal solution improperly. After each calculation of the fitness, save the neuron that obtained the best fitness in all previous iterations, and the final best neuron will be used as the initial weight and threshold for the next step of training the final neural network model.

Step 3: Construct a neural network based on the optimal weight and threshold obtained by the genetic algorithm, and train the neural network model;

Step 3.1: Construct a BP neural network based on the optimal weights and thresholds and other neural network parameters obtained in the previous step;

Step 3.2: Train the neural network model;

Step 3.3: Calculate the error between the output value of the trained neural network model and the real value in the data set, and update the weights and thresholds of each layer of the neural network layer by layer according to the error;

Step 3.4: Determine whether the obtained model has reached the expected accuracy or the maximum number of iterations, if so, execute the next step, otherwise return to step 3.2;

The specific implementation process of the embodiment is described as follows:

Initialize the initial weights and thresholds of the neural network according to the best neurons obtained in step 2. During the training process, the weights of each layer will be gradually adjusted according to the difference between the output value and the expected output value until the best state according to the current network structure is obtained. The training process of the BP neural network not only requires multiple rounds of training, but also may adjust the number of hidden layers of the network and the number of nodes in each layer many times to continuously optimize the output results.

In the hidden layer and output layer of the neural network, it is necessary to use a nonlinear function as the activation function to make the model establish a nonlinear mapping. In the output layer, in order to evaluate the quality of model training, a loss function is defined to evaluate the utility of the model. In the initial stage of network training, in order to increase the training speed, an exponential descent strategy is used to adjust the weights to achieve rapid convergence. In order to avoid over-fitting of the model, a regularization term describing the complexity of the model is added to the loss function. After several iterations, an optimization algorithm is used to optimize the weights of each layer in the network.

After the network structure, activation function, loss function and optimization algorithm are determined, the neural network can be trained based on the training data set. In the process of training the BP neural network, the weights of the nodes in each layer will be continuously adjusted inside the neural network until the best training result under the current network configuration is achieved. At this time, the training of the BP neural network is not over. After the training is completed under the configuration of the neural network, it should be determined according to the experimental results whether to adjust the structure of the neural network and some parameters before retraining to obtain the best performance. network model.

Step 4: Verify the trained neural network model and evaluate the actual effect of the model. .

Step 4.1: Input the data in the test set into the neural network model obtained in step 3, and obtain the corresponding output value;

Step 4.2: Calculate the error between the model output value and the real value. If the error meets the expected requirements, the model can be applied in actual production. Otherwise, return to step 2, adjust various parameters, and retrain the model.

After training the neural network, use the test set to check whether the trained model can be applied to actual production. If the difference between the predicted value based on the BP neural network in the test set and the output value in the test set is within a specified range , it is considered that the trained BP neural network model can be applied to actual production; otherwise, the BP neural network model needs to be rebuilt and trained.

Through the comparison of Fig. 3, Fig. 4, Fig. 5, and Fig. 6, the accuracy of the predicted values of tar, nicotine and CO obtained by using the neural network model optimized by genetic algorithm in this patent is higher than that of the neural network optimized by genetic algorithm. The accuracy of the predicted values of tar, nicotine and CO obtained by the model has been greatly improved, which shows that the technical solution adopted in this patent can indeed achieve better practical results, and will have a positive effect on assisting in the extraction and production of tobacco during the manufacturing process. effect.

The basic principles and main features of the present invention and the advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited by the above-mentioned embodiments. The above-mentioned embodiments and descriptions only illustrate the structural relationship and principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention There are also various changes and improvements which fall within the scope of the claimed invention. The protection scope of the present invention is defined by the appended claims and their equivalents.

Claims (5)

1. A predictive model based on genetic algorithm optimization neural network cigarette material and mainstream smoke composition, is characterized in that, comprises the following steps: Step 1: Sample preprocessing; Step 2: Obtain the optimal weight and threshold parameters of the neural network based on the genetic algorithm; Step 3: Construct a neural network based on the optimal weight and threshold obtained by the genetic algorithm, and train the neural network model; Step 4: Verify the trained neural network model and evaluate the actual effect of the model. 2. The prediction model based on the genetic algorithm optimization neural network of cigarette material and mainstream smoke composition according to claim 1, it is characterized in that, the specific realization of step 1 comprises the following sub-steps: Step 1.1: Remove data records with missing values and data that obviously do not conform to reality in the sample; Step 1.2: The data in the original data is normalized according to the following expression, x represents the original data, y represents the data after normalization, Minvalue and MaxValue represent the minimum and maximum values of the original data, y=(x- MinValue)/(MaxValue-Minvalue); Step 1.3: Divide the normalized data into training set and test set. 3. The prediction model of cigarette material and mainstream smoke composition based on genetic algorithm optimization neural network according to claim 1, is characterized in that, the specific realization of step 2 comprises the following sub-steps: Step 2.1: Define the constant parameters of the neural network, the number of input layer nodes INPUT_NODE, the number of hidden layer nodes LAYER1_NODE, the number of output layer nodes OUTPUT_NODE, the basic learning rate LEARNING_RATE_BASE, the decay rate LEARNING_RATE_DECAY of the learning rate, and the regularization item describing the complexity of the model in The coefficient REAULARIZATION_RATE in the loss function, the number of training rounds TRAINING_SETPS, and the moving average decay rate MOVING_AVERAGE_DECAY; Step 2.2: Define the activation function tf.nn.relu(), the loss function loss(), and the optimization algorithm tf.train.GradientDescentOptimizer() for optimizing the loss function; Step 2.3: Set the evolutionary algebra, population size, crossover and mutation probability of the population, and encode the initial weight and threshold of neurons in real numbers; Step 2.4: Calculate the fitness of the population, and select the best individuals from the current population; Step 2.5: The population is selected, crossed, and mutated; Step 2.6: Judging whether the optimization goal is reached, if the optimization goal is reached, select the best weight and threshold and execute the next step, otherwise execute step 2.4. 4. the prediction model based on the genetic algorithm optimization neural network of cigarette material and mainstream smoke composition according to claim 1, is characterized in that, the specific realization of step 3 comprises the following sub-steps: Step 3.1: Construct a BP neural network based on the optimal weights and thresholds and other neural network parameters obtained in the previous step; Step 3.2: Train the neural network model; Step 3.3: Calculate the error between the output value of the trained neural network model and the real value in the data set, and update the weights and thresholds of each layer of the neural network layer by layer according to the error; Step 3.4: Determine whether the obtained model has reached the expected accuracy or the maximum number of iterations, if so, execute the next step, otherwise return to step 3.2. 5. The prediction model of cigarette material and mainstream smoke composition based on genetic algorithm optimization neural network according to claim 1, is characterized in that, the specific realization of step 4 comprises the following sub-steps: Step 4.1: Input the data in the test set into the neural network model obtained in step 3, and obtain the corresponding output value; Step 4.2: Calculate the error between the model output value and the real value. If the error meets the expected requirements, the model can be applied in actual production. Otherwise, return to step 2, adjust various parameters, and retrain the model.