Engine transient fuel consumption machine learning modeling method based on data enhancement

Abstract

The invention provides an engine transient fuel consumption machine learning modeling method based on data enhancement, which aims to solve the problem of high-precision transient fuel consumption modeling under the condition of a small amount of sample data, and comprises the following steps of: firstly, constructing a generative adversarial network model based on a Nash equilibrium principle; new engine transient fuel consumption data is generated through a GAN generator to enhance a sample set, so that the sample size is increased, and the sample diversity is improved; and training an engine transient fuel consumption model by using the data enhancement sample set through long short term memory (LSTM) recursive network machine learning so as to improve the generalization ability and prediction accuracy of the model.

Description

Technical Field

[0001] The present invention relates to a method for modeling engine transient fuel consumption, and in particular to a method for modeling engine transient fuel consumption based on data enhancement. Background Art

[0002] In the face of increasingly severe environmental and energy problems, improving the energy efficiency of vehicles is of great significance to reducing global energy consumption. In traditional vehicle simulation models, only the steady-state diagram of the internal combustion engine is usually used to estimate fuel consumption, while ignoring the difference between transient and steady-state fuel consumption. In actual vehicle operating conditions, the difference between the actual fuel consumption of the engine and the fuel consumption calculated based on the steady-state diagram is usually between 2% and 8%, and in transient processes, this difference can even reach 30%. Therefore, being able to more accurately estimate transient fuel consumption is extremely critical to improving the accuracy of control strategies. With the advancement of computer technology, machine learning can greatly improve the accuracy of engine fuel consumption prediction, but relying on machine learning to build an engine transient fuel consumption model requires a large number of training samples to ensure the accuracy of the network after training, and obtaining a large amount of high-quality engine test data is both difficult and expensive.

[0003] The invention patent with publication number CN114781245A proposes a method and system for estimating transient automobile fuel consumption based on correction of strongly correlated parameters of fuel consumption. The core idea of ​​this method is to establish a fuel consumption model through steady-state data, and then use non-steady-state data for error correction to improve the accuracy of the model. The principal component analysis method, cluster analysis method and polynomial fitting method are used for data processing. The application premise of this method is that it needs to be supported by a large amount of steady-state and non-steady-state data, and it cannot solve the problem of modeling transient fuel consumption of engines when there is less sample data. Therefore, it is necessary to design a method that can adapt to a small amount of sample data for the establishment of the transient fuel consumption model of the engine.

[0004] The present invention proposes a machine learning modeling method for engine transient fuel consumption based on data enhancement, which can solve the problem of high-precision transient fuel consumption modeling with a small amount of sample data. Summary of the invention

[0005] In order to solve the problem of high-precision transient fuel consumption modeling with a small amount of sample data, this study proposes a machine learning modeling method for engine transient fuel consumption based on data enhancement. By building a generative adversarial network model and using a GAN generator to generate new engine transient fuel consumption data to enhance the sample set, the engine transient fuel consumption model is trained using the data enhanced sample set. The specific steps include:

[0006] The first step is to use the existing engine transient fuel consumption data to train the GAN generator model and generate new enhanced data. The specific steps include the following:

[0007] 1GAN generator solution

[0008] The structure of the Generative Adversarial Network (GAN) consists of a generator (G) and a discriminator (D). The purpose of the generator is to create data samples that are as realistic as possible. These samples will be input into the discriminator together with the real samples; the task of the discriminator is to accurately identify the real samples and reject the generated samples. In this process, the generator continuously learns to improve the quality of the generated samples, and the discriminator continuously improves its accuracy in identifying the true and the false. The loss function used by the discriminator D is the regularized cross entropy loss function of the binary classifier, and the calculation formula is:

[0009] l D =-[ylog(D(x))+(1-y)log(1-D(G(z)))](1)

[0010] Where D(x) is the output of the discriminator for the real data sample x, that is, the probability of identifying x as a real sample; G(z) is the output of the generator, representing the fake sample generated by the latent variable z; D(G(z)) is the output of the discriminator for these generated samples, that is, the probability of identifying the generated samples as real samples; y is the label of the sample, where y=1 for real samples and y=0 for generated samples.

[0011] The goal of the generator G is to maximize the loss function value of the discriminator D. Since the first term in formula (1) is independent of the generator, the loss function l of the generator G is G Can be simplified to:

[0012] l G =(1-y)log(1-D(G(z)))(2)

[0013] In this framework, the generator and the discriminator are in a competitive relationship. Their joint objective function V(D,G) can be expressed as

[0014]

[0015] In the formula, p data (x) represents the probability distribution of the real data sample x; p z (z) is the probability distribution of the latent variable z.

[0016] Because p g (x) is the probability distribution of the data generated by the variable z, which is specifically expressed as formula (4):

[0017] p g (x) = p z (G -1 (x))(G -1 )(x)(4)

[0018] Substituting formula (4) into formula (3), we can further obtain formula (5)

[0019]

[0020] Next, the optimal solution of D(x) in the objective function is calculated, as shown in formula (6):

[0021]

[0022] After solving, we get formula (7):

[0023]

[0024] Where D * (x) represents the optimal solution of D(x). data (x) = p g (x), D * (x) = 1 / 2. Substitute this optimal solution into formula (4) and optimize the generator G so that the objective function C(G) is expressed as

[0025]

[0026] Further reorganizing formula (8) into the KL divergence form of a continuous function, we get formula (9)

[0027]

[0028] KL divergence is a non-negative number. According to formula (9), when p data (x) = p g (x), C(G) reaches the global minimum value -log4.

[0029] 2GAN data enhancement model design

[0030] In order to process and generate high-quality structured data, it is necessary to design a generative adversarial network (GAN) model based on fully connected layers. The designed GAN model consists of a generator and a discriminator, both of which are composed of multiple fully connected layers. The core task of the generator is to map the random noise vector in the latent space to the data space to generate new data samples. Considering the length of sample data and computational efficiency, the generator is configured with four main fully connected layers, each layer is followed by batch normalization to stabilize the training process, and the LReLU activation function is used to alleviate the sparsity problem of the ordinary ReLU function. The output layer of the generator uses the tanh activation function to ensure that the value of the output data is in an appropriate range; the discriminator is also composed of multiple fully connected layers, and each layer also uses the LReLU activation function and batch normalization to enhance the discrimination ability of the model. In addition, in order to resist overfitting, some layers also integrate dropout technology.

[0031] 3. Training to generate enhanced data

[0032] Select an existing sample set, of which 80% is used as a training set and the remaining 20% ​​is used as a test set. In addition, a 4-fold cross-validation method is used to evaluate the performance of the model and adjust the parameters to ensure the generalization ability and stability of the model. During the training process, observe the number of iterations and loss function value changes of the generator and discriminator. When the loss function values ​​of both show a basically stable trend, it indicates that the model can reach a relatively balanced state after multiple iterations, thereby ensuring the quality of the generated samples and the discriminant ability of the discriminator.

[0033] The second step is to use the original data and the new enhanced data to jointly train the engine transient fuel consumption model

[0034] The test data generated by the above-mentioned generative adversarial network (GAN) is fused and reorganized with the original sample data to expand it into a data enhancement sample set. Subsequently, the long short-term memory recurrent network (LSTM) machine learning model is used to predict the instantaneous fuel consumption of the engine, thereby improving the prediction accuracy of the engine's transient fuel consumption.

[0035] LSTM is based on RNN and introduces a gating mechanism to control the flow of historical information. LSTM is mainly composed of three gates, including input gate, output gate and forget gate. The input gate is used to control the input x t and candidate values The remaining information is combined with the result of the forget gate to update the cell state c t The purpose of the output gate is to provide information about the cell state c t Some information is output as h t The forget gate is the one that decides to discard the old cell state c t-1 Get the old output h t-1 and input x t After that, the LSTM model calculates the input gate, output gate, forget gate and candidate value in sequence. The relevant formula is as follows:

[0036]

[0037] where δ(·) is the logistic function, tanh is the hyperbolic tangent function, W and U are the weight matrices of each gate, and b is the bias vector. In addition, the cell state c t The output f of the output gate t and the output i of the input gate t Update, the equation is as follows:

[0038]

[0039] Finally, the output h of the LSTM model is calculated based on the output gate and the cell statet , the calculation formula is:

[0040] h t =o t ⊙tanh(c t )(12)

[0041] In the LSTM model, historical information is extracted and stored by repeatedly updating the cell state c t Because the forget gate can selectively forget the accumulated historical information, the LSTM method has a larger memory size than the simple RNN and can achieve excellent performance when processing time series prediction. Based on this, the training of the engine transient fuel consumption model is implemented.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The data-enhanced machine learning modeling method for engine transient fuel consumption described in the present invention realizes high-precision transient fuel consumption modeling with a small amount of sample data;

[0044] 2. The data-enhanced machine learning modeling method for transient fuel consumption of an engine described in the present invention applies a generative adversarial network (GAN) to the acquisition of transient fuel consumption data, which can effectively alleviate the scarcity problem of high-quality engine data and improve the accuracy of the model;

[0045] 3. The data-enhanced engine transient fuel consumption machine learning modeling method described in the present invention combines a generative adversarial network (GAN) with a long short-term memory recurrent network (LSTM) machine learning, which not only takes advantage of LSTM in processing nonlinear and complex data, but also solves the problem that machine learning requires a large number of training samples to ensure the accuracy of network training. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be further described below in conjunction with the accompanying drawings:

[0047] Figure 1 A flowchart of a method for machine learning modeling of engine transient fuel consumption based on data enhancement according to the present invention;

[0048] Figure 2 A flow chart of a generative adversarial network in a data-enhanced engine transient fuel consumption machine learning modeling method according to the present invention;

[0049] Figure 3 The loss curves of the generator and the discriminator in the data-enhanced engine transient fuel consumption machine learning modeling method described in the present invention;

[0050] Figure 4A long short-term memory recursive network (LSTM) algorithm structure in a data-enhanced engine transient fuel consumption machine learning modeling method according to the present invention; DETAILED DESCRIPTION

[0051] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below with reference to the accompanying drawings.

[0052] See also Figure 1 , showing the flow chart of the engine transient fuel consumption machine learning modeling method based on data enhancement. First, a generative adversarial network model is constructed based on the Nash equilibrium principle, and new engine transient fuel consumption data is generated through the GAN generator to enhance the sample set, so as to increase the sample size and improve the sample diversity; then the data enhanced sample set is used to train the engine transient fuel consumption model through long short-term memory recurrent network (LSTM) machine learning to improve the generalization ability and prediction accuracy of the model.

[0053] The first step is to use the existing engine transient fuel consumption data to train the GAN generator model and generate new enhanced data. The specific steps include the following:

[0054] 1GAN generator solution

[0055] See also Figure 2 , the generative adversarial network (GAN) structure includes a generator (G) and a discriminator (D). The purpose of the generator is to create data samples that are as realistic as possible. These samples will be input into the discriminator together with the real samples; the task of the discriminator is to accurately identify the real samples and reject the generated samples. In this process, the generator continuously learns to improve the quality of the generated samples, and the discriminator continuously improves its accuracy in identifying authenticity. The discriminator D uses the regularized cross entropy loss function of the binary classifier as the loss function, and the calculation formula is:

[0056] l D =-[ylog(D(x))+(1-y)log(1-D(G(z)))](1)

[0057] Where D(x) is the output of the discriminator for the real data sample x, that is, the probability of identifying x as a real sample; G(z) is the output of the generator, representing the fake sample generated by the latent variable z; D(G(z)) is the output of the discriminator for these generated samples, that is, the probability of identifying the generated samples as real samples; y is the label of the sample, where y=1 for real samples and y=0 for generated samples.

[0058] The goal of the generator G is to maximize the loss function value of the discriminator D. Since the first term in formula (1) is independent of the generator, the loss function l of the generator G is G Can be simplified to:

[0059] l G =(1-y)log(1-D(G(z)))(2)

[0060] In this framework, the generator and the discriminator are in a competitive relationship. Their joint objective function V(D,G) can be expressed as

[0061]

[0062] In the formula, p data (x) represents the probability distribution of the real data sample x; p z (z) is the probability distribution of the latent variable z.

[0063] Because p g (x) is the probability distribution of the data generated by the variable z, which is specifically expressed as formula (4):

[0064] p g (x) = p z (G -1 (x))(G -1 )(x)(4)

[0065] Substituting formula (4) into formula (3), we can further obtain formula (5)

[0066]

[0067] Next, the optimal solution of D(x) in the objective function is calculated, as shown in formula (6):

[0068]

[0069] After solving, we get formula (7):

[0070]

[0071] Where D * (x) represents the optimal solution of D(x). data (x) = p g (x), D * (x) = 1 / 2. Substitute this optimal solution into formula (4) and optimize the generator G so that the objective function C(G) is expressed as

[0072]

[0073] Further reorganizing formula (8) into the KL divergence form of a continuous function, we get formula (9)

[0074]

[0075] KL divergence is a non-negative number. According to formula (9), when p data (x) = p g (x), C(G) reaches the global minimum value -log4.

[0076] 2GAN data enhancement model design

[0077] In order to process and generate high-quality structured data, it is necessary to design a generative adversarial network (GAN) model based on fully connected layers. The designed GAN model consists of a generator and a discriminator, both of which are composed of multiple fully connected layers. The core task of the generator is to map the random noise vector in the latent space to the data space to generate new data samples. Considering the length of sample data and computational efficiency, the generator is configured with four main fully connected layers, each layer is followed by batch normalization to stabilize the training process, and the LReLU activation function is used to alleviate the sparsity problem of the ordinary ReLU function. The output layer of the generator uses the tanh activation function to ensure that the value of the output data is in an appropriate range; the discriminator is also composed of multiple fully connected layers, and each layer also uses the LReLU activation function and batch normalization to enhance the discrimination ability of the model. In addition, in order to resist overfitting, some layers also integrate dropout technology.

[0078] 3. Training to generate enhanced data

[0079] An existing sample set is selected, of which 80% is used as a training set and the remaining 20% ​​is used as a test set. In addition, a 4-fold cross-validation method is used to evaluate the performance of the model and adjust the parameters to ensure the generalization ability and stability of the model.

[0080] See also Figure 3 During the training process, the number of iterations and the changes in the loss function values ​​of the generator and discriminator are observed. When the loss function values ​​of both show a basically stable trend, it indicates that the model can reach a relatively balanced state after multiple iterations, thereby ensuring the quality of the generated samples and the discriminative ability of the discriminator.

[0081] The second step is to use the original data and the new enhanced data to jointly train the engine transient fuel consumption model

[0082] The test data generated by the above-mentioned generative adversarial network (GAN) is fused and reorganized with the original sample data to expand it into a data enhancement sample set. Subsequently, the long short-term memory recurrent network (LSTM) machine learning model is used to predict the instantaneous fuel consumption of the engine, thereby improving the prediction accuracy of the engine's transient fuel consumption.

[0083] See also Figure 4 LSTM is mainly composed of three gates, including input gate, output gate and forget gate. The input gate is used to control the t and candidate values The remaining information is combined with the result of the forget gate to update the cell state c t The purpose of the output gate is to provide information about the cell state c t Some information is output as h t The forget gate is the one that decides to discard the old cell state c t-1 Get the old output h t-1 and input x t After that, the LSTM model calculates the input gate, output gate, forget gate and candidate value in sequence. The relevant formula is as follows:

[0084]

[0085] where δ(·) is the logistic function, tanh is the hyperbolic tangent function, W and U are the weight matrices of each gate, and b is the bias vector. In addition, the cell state c t The output f of the output gate t and the output i of the input gate t Update, the equation is as follows:

[0086]

[0087] Finally, the output h of the LSTM model is calculated based on the output gate and the cell state t , the calculation formula is:

[0088] h t =o t ⊙tanh(c t )(12)

[0089] In the LSTM model, historical information is extracted and stored by repeatedly updating the cell state c t Because the forget gate can selectively forget the accumulated historical information, the LSTM method has a larger memory size than the simple RNN and can achieve excellent performance when processing time series prediction. Based on this, the training of the engine transient fuel consumption model is implemented.

Claims

1. A machine learning modeling method for transient fuel consumption of an engine based on data enhancement, characterized in that: When there are few training samples, the accuracy of the prediction model is improved by enhancing the sample set, including the following: The first step is to use the existing engine transient fuel consumption data to train the GAN generator model and generate new enhanced data. The specific steps include the following: ① GAN generator solution The structure of Generative Adversarial Network (GAN) includes a generator (G) and a discriminator (D); the purpose of the generator is to create data samples that are as realistic as possible, which will be input into the discriminator together with the real samples; the task of the discriminator is to accurately identify the real samples and reject the generated samples; in this process, the generator continuously learns to improve the quality of the generated samples, and the discriminator continuously improves its accuracy in identifying the authenticity; the loss function used by the discriminator D is the regularized cross entropy loss function of the binary classifier, and the calculation formula is: l D n-[ylog(D(x))+(1-y)log(1-D(G(z)))] (1) Where D(x) is the output of the discriminator for the real data sample x, that is, the probability of identifying x as a real sample; G(z) is the output of the generator, representing the fake sample generated by the latent variable z; D(G(z)) is the output of the discriminator for these generated samples, that is, the probability of identifying the generated samples as real samples; y is the label of the sample, where y=1 for the real sample and y=0 for the generated sample; The goal of the generator G is to maximize the loss function value of the discriminator D; since the first term in formula (1) is independent of the generator, the loss function l of the generator G is G Simplified to: l G =(1-y)log(1-D(G(z))) (2) In this framework, the generator and the discriminator are in a competitive relationship; their joint objective function V(D,G) is expressed as In the formula, p data (x) represents the probability distribution of the real data sample x; p z (z) is the probability distribution of the latent variable z; Because p g (x) is the probability distribution of the data generated by the variable z, which is specifically expressed as formula (4): p g (x)=p z (G -1 (x))(G -1 )(x) (4) Substituting formula (4) into formula (3), we get formula (5) Next, the optimal solution of D(x) in the objective function is calculated, as shown in formula (6): After solving, we get formula (7): Where D * (x) represents the optimal solution of D(x); when p data (x) = p g (x), D * (x) = 1 / 2; Substitute this optimal solution into formula (4) and then optimize the generator G, so that the objective function C(G) is expressed as Further reorganizing formula (8) into the KL divergence form of a continuous function, we get formula (9) KL divergence is a non-negative number. According to formula (9), when p data (x) = p g (x), C(G) achieves the global minimum value -log4; ②GAN data enhancement model design In order to process and generate high-quality structured data, a generative adversarial network (GAN) model based on fully connected layers needs to be designed. The designed GAN model includes a generator and a discriminator, both of which are composed of multiple fully connected layers. The core task of the generator is to map the random noise vector in the latent space to the data space to generate new data samples. Considering the length of sample data and computational efficiency, the generator is configured with four main fully connected layers, each layer is followed by batch normalization to stabilize the training process, and the LReLU activation function is used to alleviate the sparsity problem of the ordinary ReLU function. The output layer of the generator uses the tanh activation function to ensure that the value of the output data is in an appropriate range. The discriminator is also composed of multiple fully connected layers. LReLU activation function and batch normalization are also used after each layer to enhance the discrimination ability of the model. In addition, in order to resist overfitting, some layers also integrate dropout technology. ③ Training generates enhanced data Select an existing sample set, 80% of which is used as a training set and the remaining 20% ​​is used as a test set; In addition, a 4-fold cross-validation method is used to evaluate the performance of the model and adjust the parameters to ensure the generalization ability and stability of the model. During the training process, the number of iterations and the change in the loss function value of the generator and discriminator are observed. When the loss function values ​​of both show a basically stable trend, it indicates that the model can reach a relatively balanced state after multiple iterations, thereby ensuring the quality of the generated samples and the discriminative ability of the discriminator. The second step is to use the original data and the new enhanced data to jointly train the engine transient fuel consumption model The test data generated by the above-mentioned generative adversarial network (GAN) is fused and reorganized with the original sample data to expand it into a data enhancement sample set; Subsequently, the long short-term memory recurrent network (LSTM) machine learning model is used to predict the instantaneous fuel consumption of the engine, thereby improving the prediction accuracy of the engine's transient fuel consumption; LSTM introduces a gating mechanism based on RNN to control the flow of historical information. LSTM is mainly composed of three gates, including input gate, output gate and forget gate. The input gate is used to control the input x t and candidate values The remaining information is combined with the result of the forget gate to update the cell state c t The purpose of the output gate is to provide information about the cell state c t Some information is output as h t ; The forget gate is used to decide to discard the old cell state c t-1 What information is in it; get the old output h t-1 and input x t After that, the LSTM model calculates the input gate, output gate, forget gate and candidate value in sequence. The relevant formula is as follows: where δ(·) is the logistic function, tanh is the hyperbolic tangent function, W and U are the weight matrices of each gate, and b is the bias vector; in addition, the cell state c t The output f of the output gate t and the output i of the input gate t Update, the equation is as follows: Finally, the output h of the LSTM model is calculated based on the output gate and the cell state t , the calculation formula is: h t =o t ⊙tanh(c t ) (12) In the LSTM model, historical information is extracted and stored by repeatedly updating the cell state c t This is achieved; since the forget gate can selectively forget the accumulated historical information, the memory size of the LSTM method is larger than that of the simple RNN, and excellent performance can be achieved when processing time series prediction; based on this, the training and construction of the engine transient fuel consumption model is realized.

Images