MMoE-LSTM-based diesel engine nitrogen oxide emission prediction method

Abstract

The invention discloses a diesel engine nitrogen oxide emission prediction method based on an MMoE-LSTM. The method comprises the following steps: firstly, decomposing nitrogen oxide emission time sequence data of the diesel engine by using a completely integrated empirical mode decomposition and adaptive noise algorithm to obtain a series of subsequences; secondly, in combination with a sample entropy method and a K-means clustering method, reconstructing the sub-sequence into three sub-sequences with physical significance; then, a prediction model based on MMoE-LSTM is constructed, the method can accurately capture features in diesel engine nitrogen oxide emission sequence data, and the prediction precision of the three subsequences is improved through a soft sharing mechanism; and finally, superposing the prediction results of the subsequences to obtain a final prediction result of the nitrogen oxide emission of the diesel engine. According to the method, decomposition and denoising of diesel engine nitrogen oxide data can be effectively achieved, and high efficiency and precision are achieved for diesel engine nitrogen oxide emission prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of diesel engine testing, and in particular to a method for predicting nitrogen oxide emissions from a diesel engine based on MMoE-LSTM. Background Art

[0002] According to the 2023 China Mobile Source Environmental Management Annual Report, the national motor vehicle nitrogen oxide emissions in 2022 reached 5.267 million tons, of which diesel vehicles accounted for more than 80% of the total automobile emissions. Therefore, diesel vehicles are the main source of nitrogen oxide pollution in my country. Selective catalytic reduction technology (SCR) can effectively reduce the nitrogen oxide emissions of diesel engines, and the prediction of nitrogen oxide emissions from diesel engines can help SCR adjust the injection amount of urea according to the actual nitrogen oxide emissions, thereby reducing nitrogen oxide emissions to meet the National VI emission standards.

[0003] In the traditional method of diesel engine nitrogen oxide emission prediction, the prediction process is usually based on the analysis of physical and chemical properties and relies on empirical or semi-empirical assumptions, which requires the predictor to have a deep understanding of the complex physical and chemical processes inside the diesel engine, thereby increasing the time and human resources required for prediction. Therefore, how to improve the accuracy and efficiency of nitrogen oxide emission prediction has become a technical challenge that needs to be solved urgently.

[0004] In recent years, machine learning algorithms have been widely used in various fields and have achieved good practical results. Because the model established by machine learning does not need to consider the complex physical and chemical knowledge in diesel engines, its application in the field of diesel engine nitrogen oxide emission prediction is becoming more and more extensive. Considering the nonlinear and instability characteristics of diesel engine nitrogen oxide emission data, the CEEMDAN method is used to effectively decompose these data. However, when processing the decomposed nitrogen oxide emission data, traditional methods often predict each component independently and ignore the relationship between components. In order to solve this problem, the present invention introduces the idea of ​​multi-task learning. A major advantage of multi-task learning is that it can capture the intrinsic relationship between components. In the field of multi-task learning, the multi-gate expert mixture network (MMoE) dynamically selects different expert combinations according to input features by introducing an independent gated network for each task. This design allows different tasks to use experts in a diversified way, so as to better capture the relationship between tasks and further improve the performance and efficiency of the prediction model. In addition, LSTM networks have now been widely used in diesel engine nitrogen oxide emissions. LSTM networks can learn long sequence dependencies and perform well in processing time series data. Therefore, the present invention combines MMoE and LSTM to propose a new prediction model that can explore the dependencies of long sequences and the relationships between subtasks. Summary of the invention

[0005] The purpose of the present invention is to provide a diesel engine nitrogen oxide emission prediction method based on MMoE-LSTM to improve the diesel engine nitrogen oxide emission prediction accuracy in order to solve the problem of low diesel engine nitrogen oxide emission prediction accuracy.

[0006] To achieve the above object, the present invention provides a diesel engine nitrogen oxide emission prediction method based on MMoE-LSTM, comprising the following steps:

[0007] Step S1: obtaining diesel engine nitrogen oxide emission time series data, and decomposing the diesel engine nitrogen oxide emission time series data into multiple intrinsic mode functions (IMFs) and final residuals (RES) using a fully integrated empirical mode decomposition and adaptive noise method (CEEMDAN);

[0008] Step S2: Calculate the entropy value of each intrinsic mode function using the sample entropy method, and use the K-means method to classify and aggregate according to the calculated entropy value, and reconstruct it into high-frequency component, low-frequency component and trend component;

[0009] Step S3: The above three subsequences are used to form a data set, and the prediction of each subsequence is set as a subtask to build a multi-gate hybrid expert combined with long short-term memory network (MMoE-LSTM) prediction model;

[0010] Step S4: Input the data set into the MMoE-LSTM prediction model to obtain the prediction results of each subtask, that is, the prediction results of the three components;

[0011] Step S5: Accumulate the prediction results of the components to obtain the diesel engine nitrogen oxide emission prediction result.

[0012] Optionally, the step of decomposing nitrogen oxide emission time series data by CEEMDAN in step S1 is specifically as follows:

[0013] A1: Let x(t) be the original NOx emission data time series, ω n (t) is the Gaussian white noise added in the nth experiment. The time series of the nth decomposed NOx emission data can be expressed as x n (t) = x(t) + ε 0 ω n (t), (n=1,2,…,k),ε 0 is the Gaussian white noise weighting coefficient, k is the number of empirical mode decomposition, EMD k (·) is called the kth IMF obtained by the empirical mode decomposition (EMD) algorithm;

[0014] The first IMF component is obtained by EMD decomposition Calculate the CEEMDAN decomposition to get the unique residual signal r1 (t), the calculation formula is:

[0015]

[0016] A2: Continue to add Gaussian white noise to the residual r 1 (t), repeat the above steps to obtain and the new residual r 2 (t), the calculation formula is:

[0017]

[0018] A3: Repeat the above steps to calculate the kth residual signal and the k+1th IMF component. The calculation formula is:

[0019]

[0020] A4: Until the residual components obtained cannot be further decomposed, all the components and the final trend term of CEEMDAN decomposition are obtained, and the original sequence x(t) is divided into k subsequences and a residual sequence, as shown below:

[0021]

[0022] Where: k represents the nth modal component; R(t) represents the final residual signal.

[0023] Optionally, in step S2, the IMFs obtained by decomposing CEEMDAN are reconstructed into high-frequency components, low-frequency components and trend components, and the reconstruction step is specifically as follows:

[0024] B1: For the time series composed of data decomposed by CEEMDAN In order, a vector sequence of dimension p is formed, denoted as X p (i) as follows:

[0025]

[0026] Define vector X p (i) and X p (j) The distance d[X p (i), X p (j)] is the absolute value of the maximum difference between the two corresponding elements, and the calculation formula is as follows:

[0027]

[0028] B2: For a given X p (i), statistics X p (i) With X pThe number of j between (j) that is less than the similarity tolerance r is denoted as B i , and is ratioed with the number of distances Np, recorded as The calculation formula is as follows:

[0029]

[0030] Based on this, we define B p (r):

[0031]

[0032] B3: Increase the dimension to p+1, repeat the above formula, and calculate B p+1 (r), as follows:

[0033]

[0034] Since the length of the time series N is finite rather than infinite, the sample entropy is calculated as follows:

[0035]

[0036] The sample entropy value of each IMF is calculated by the sample entropy method. The k-means clustering algorithm is used to divide the IMF into high-frequency components, low-frequency components and trend components according to the frequency characteristics.

[0037] Optionally, in step S3, the constructed MMoE-LSTM prediction model is:

[0038] MMoE is a multi-task learning model that selects different expert networks through multiple gating networks to learn the connections and differences between different tasks. The expression of the MMoE model is as follows:

[0039] y k =h k (f k (x))

[0040] Among them, y k represents the output of the kth task, h k (·) is the output layer function of the kth task, f k (x) represents the intermediate result after being processed by the expert network. Next, the processing result of the expert network is explained:

[0041]

[0042] f k (x) is the output of all expert networks f i The weighted sum of (x), weight Determined by the gating network:

[0043] g k (x) = softmax(W gk x)

[0044] The output of the gating network is the weights of all experts, W gk ∈R n×d represents a trainable matrix, n is the number of experts, and d is the dimension of the feature;

[0045] When building the expert network of MMoE, the LSTM network can be combined to build an expert network that can remember and process time series information. Through this design, the expert network can better capture and utilize the feature associations between different tasks, thereby improving the performance and generalization ability of the model in multi-task learning scenarios. The LSTM network consists of an input gate, an output gate, and a forget gate: the forget gate is responsible for calculating which information in the hidden state and cell state of the current time step should be forgotten or retained; the input gate is responsible for controlling how much of the current input information can enter the cell state and affect the update of the memory unit; the output gate is responsible for determining what information the hidden state of the current time step should contain. The specific calculation process is as follows:

[0046] The forget gate fi determines whether it is necessary to forget the information in the unit state at the previous time;

[0047] f t =σ(W f [γ t-1 , x t ]+b f ),

[0048] Input Gate i t , determine the new information added to the cell state, The input gate i t The candidate values ​​generated;

[0049] i t =σ(W i [γ t-1 , x t ]+b i ),

[0050]

[0051] Combined with forget gate f i and output gate i t The output updates the unit state;

[0052]

[0053] Output gate o t , output unit state value; γ t is the hidden state output;

[0054] o t =σ(W o [γ t-1 , x t ]+b o ),

[0055] γ t =o t tanh(C t ),

[0056] In the formula, x t , represents the input at time step t; γ t-1 and γ t Represent the previous state and current state of the hidden layer respectively; C t-1 and C t They are the previous and current unit storage information respectively; W f , W i , W o , W C is the weight matrix; b f , b i , b C , b o is the offset vector; σ is the sigmoid activation function. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flowchart of the method for predicting diesel engine nitrogen oxide emissions based on MMoE-LSTM of the present invention;

[0058] Figure 2 This is the CEEMDAN decomposition result diagram of this embodiment;

[0059] Figure 3 This is a sample entropy reorganization result diagram of this embodiment;

[0060] Figure 4 This is a diagram of the recombinant component prediction result of this embodiment;

[0061] Figure 5 This is a diagram showing the nitrogen oxide emission prediction results of this embodiment. DETAILED DESCRIPTION

[0062] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention. It should be particularly pointed out that the accompanying drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention.

[0063] The present invention provides a diesel engine nitrogen oxide emission prediction method based on MMoE-LSTM.

[0064] Reference Figure 1 , showing the process of the present invention, specifically the following steps:

[0065] Step S1: obtaining diesel engine nitrogen oxide emission time series data, and decomposing the diesel engine nitrogen oxide emission time series data into multiple IMFs using a fully integrated empirical mode decomposition and adaptive noise method (CEEMDAN);

[0066] Step S2: Calculate the entropy value of each IMF using the sample entropy method, and use the K-means method to classify and aggregate the calculated entropy values, and reconstruct them into high-frequency components, low-frequency components, and trend components;

[0067] Step S3: The above three subsequences are used to form a data set, and the prediction of each subsequence is set as a subtask to build a multi-gate hybrid expert combined with long short-term memory network prediction model;

[0068] Step S4: Input the data set into the MMoE-LSTM prediction model to obtain the prediction result of each subtask, that is, the prediction result of each subsequence;

[0069] Step S5: Accumulate the prediction results of the subsequences to obtain the nitrogen oxide emission prediction result.

[0070] Reference Figure 2 , showing the CEEMDAN decomposition results of diesel engine NOx time series data. The specific steps of CEEMDAN decomposition of NOx emission time series data are:

[0071] A1: Let x(t) be the original NOx emission data time series, ω n (t) is the Gaussian white noise added in the nth experiment. The time series of the nth decomposed NOx emission data can be expressed as x n (t) = x(t) + ε 0 ω n (t), (n=1,2,…,k),ε 0 is the Gaussian white noise weighting coefficient, and k is the number of empirical mode decomposition. EMD k (·) is called the kth IMF mode obtained by the EMD algorithm.

[0072] Through EMD decomposition, we get the first IMF component Calculate the CEEMDAN decomposition to get the unique residual signal r 1 (t), the calculation formula is:

[0073]

[0074] A2: Continue to add Gaussian white noise to the residual r 1(t), repeat the above steps to get and the new residual r 2 (t), as follows:

[0075]

[0076] A3: Repeat the above steps to calculate the kth residual signal and the k+1th IMF component as follows:

[0077]

[0078] A4: Until the residual components obtained cannot be further decomposed, all components and the final trend term of CEEMDAN decomposition are obtained. Then the original sequence x(t) is divided into k subsequences and a residual sequence, as shown below:

[0079]

[0080] Where: k represents the nth modal component; R(t) represents the final residual signal.

[0081] Reference Figure 3 , showing the sample entropy reorganization results. The multiple subsequences obtained by decomposition are reconstructed into three meaningful sequences. The specific reconstruction steps are:

[0082] B1: For the time series composed of data decomposed by CEEMDAN In order, a vector sequence of dimension p is formed, denoted as X p (i) as follows:

[0083]

[0084] Define vector X p (i) and X p (j) The distance d[X p (i), X p (j)] is the absolute value of the maximum difference between the corresponding elements of the two, as shown below:

[0085]

[0086] B2: For a given X p (i), statistics X p (i) With X p The number of j between (j) that is less than the similarity tolerance r is denoted as B i , and is ratioed with the number of distances Np, recorded as As shown below:

[0087]

[0088] Based on this, we define B p (r):

[0089]

[0090] B3: Increase the dimension to p+1, repeat the above, and calculate B p+1 (r), as follows:

[0091]

[0092] Since the length of the time series N is finite rather than infinite, the sample entropy can be calculated as follows:

[0093]

[0094] The sample entropy value of each IMF is calculated by the sample entropy method. The k-means clustering algorithm is used to divide the IMF into high-frequency components, low-frequency components and trend components according to the frequency characteristics.

[0095] Reference Figure 4 , showing the prediction results of the MMoE-LSTM prediction model for high-frequency components, low-frequency components and trend components. This example uses the last 20% of the data as the verification result. The prediction model of multi-gate hybrid experts combined with long short-term memory network is:

[0096] MMoE is a multi-task learning model that selects different expert networks through multiple gating networks to learn the connections and differences between different tasks. The expression of the MMoE model is as follows:

[0097] y k =h k (f k (x))

[0098] Among them, y k represents the output of the kth task, h k (·) is the output layer function of the kth task, f k (x) represents the intermediate result after being processed by the expert network. Next, the processing result of the expert network is explained:

[0099]

[0100] Among them, f k (x) is the weighted sum of all expert network outputs fi(x), with weight Determined by the gating network:

[0101] g k (x) = softmax(W gk x)

[0102] The output of the gating network is the weights of all experts, W gk ∈R n×d Represents a trainable matrix, n is the number of experts, and d is the dimension of the feature.

[0103] When building the expert network of MMoE, combining with the LSTM network can build an expert network that can remember and process time series information. Through this design, the expert network can better capture and utilize the feature associations between different tasks, thereby improving the performance and generalization ability of the model in multi-task learning scenarios. The LSTM network consists of an input gate, an output gate, and a forget gate: the forget gate is responsible for calculating which information in the hidden state and cell state of the current time step should be forgotten or retained; the input gate is responsible for controlling how much of the current input information can enter the cell state and affect the update of the memory unit; the output gate is responsible for determining what information the hidden state of the current time step should contain. The specific calculation process is as follows:

[0104] The forget gate fi decides whether it is necessary to forget the information in the unit state at the previous time.

[0105] f t =σ(W f [γ t-1 , x t ]+b f ),

[0106] Input Gate i t , determine the new information added to the cell state, The input gate i t The resulting candidate values.

[0107] i t =σ(W i [γ t-1 , x t ]+b i ),

[0108]

[0109] Combined with forget gate f i and output gate i t The output of updates the unit state.

[0110]

[0111] Output gate o t , output unit state value; γ t is the hidden state output.

[0112] o t =σ(W o [γ t-1, x t ]+b o ),

[0113] γ t =o t tanh(C t ),

[0114] In the formula, x t , represents the input at time step t; γ t-1 and γ t Represent the previous state and current state of the hidden layer respectively; C t-1 and C t They are the previous and current unit storage information respectively; W f , W i , W o , W C is the weight matrix; b f , b i , b C , b o is the offset vector; σ is the sigmoid activation function.

[0115] Reference Figure 5 , showing the final result of the MMoE-LSTM prediction model. This embodiment selects the last 20% of the data as the verification result. The high-frequency prediction results, low-frequency prediction results and trend prediction results are accumulated to obtain the prediction results of diesel engine nitrogen oxide emissions. It is particularly pointed out that the data used in this invention is provided by Guangxi Yuchai Machinery Group Co., Ltd.

[0116] The above descriptions are detailed descriptions and illustrations of the preferred embodiments of the present invention, but these descriptions are not intended to limit the scope of protection required by the present invention. All equivalent changes or modified modifications accomplished under the technical teachings suggested by the present invention should fall within the scope of patent protection covered by the present invention.

Claims

1. A diesel engine nitrogen oxide emission prediction method based on MMoE-LSTM, characterized in that: The following steps are involved: Step S1: obtaining diesel engine nitrogen oxide emission time series data, and decomposing the diesel engine nitrogen oxide emission time series data into multiple intrinsic mode functions (IMFs) and final residuals (RES) using a fully integrated empirical mode decomposition and adaptive noise method (CEEMDAN); Step S2: Calculate the entropy value of each intrinsic mode function using the sample entropy method, and use the K-means method to classify and aggregate according to the calculated entropy value, and reconstruct it into high-frequency component, low-frequency component and trend component; Step S3: The above three subsequences are used to form a data set, and the prediction of each subsequence is set as a subtask to build a multi-gate hybrid expert combined with long short-term memory network (MMoE-LSTM) prediction model; Step S4: Input the data set into the MMoE-LSTM prediction model to obtain the prediction results of each subtask, that is, the prediction results of the three components; Step S5: Accumulate the prediction results of the components to obtain the diesel engine nitrogen oxide emission prediction result.

2. The method for predicting diesel engine nitrogen oxide emissions based on MMoE-LSTM according to claim 1, characterized in that: The CEEMDAN decomposition of nitrogen oxide emission time series data in step S1 is specifically as follows: A1: Let x(t) be the original NOx emission data time series, ω n (t) is the Gaussian white noise added in the nth experiment. The time series of the nth decomposed NOx emission data can be expressed as x n (t) = x(t) + ε0ω n (t), (n=1,2,…,k), ε0 is the weighting coefficient of Gaussian white noise, k is the number of empirical mode decomposition, EMD k (·) is called the kth IMF obtained by the empirical mode decomposition (EMD) algorithm; The first IMF component is obtained by EMD decomposition Calculate the CEEMDAN decomposition to get the unique residual signal r1(t), the calculation formula is: A2: Continue to add Gaussian white noise to the residual r1(t), repeat the above steps, and get And the new residual r2(t), calculated as: A3: Repeat the above steps to calculate the kth residual signal and the k+1th IMF component. The calculation formula is: A4: Until the residual components obtained cannot be further decomposed, all the components and the final trend term of CEEMDAN decomposition are obtained, and the original sequence x(t) is divided into k subsequences and a residual sequence, as shown below: Where: k represents the nth modal component; R(t) represents the final residual signal.

3. The method for predicting diesel engine nitrogen oxide emissions based on MMoE-LSTM according to claim 1, characterized in that: In step S2, the IMFs decomposed by CEEMDAN are reconstructed into high-frequency components, low-frequency components and trend components. The reconstruction steps are specifically as follows: B1: For the time series composed of data decomposed by CEEMDAN In order, a vector sequence of dimension p is formed, denoted as X p (i) as follows: Define vector X p (i) and X p (j) The distance d[X p (i), X p (j)] is the absolute value of the maximum difference between the two corresponding elements, and the calculation formula is as follows: B2: For a given X p (i), statistics X p (i) With X p The number of j between (j) that is less than the similarity tolerance r is denoted as B i , and is ratioed with the number of distances Np, recorded as The calculation formula is as follows: Based on this, we define B p (r): B3: Increase the dimension to p+1, repeat the above formula, and calculate B p+1 (r), as follows: Since the length of the time series N is finite rather than infinite, the sample entropy is calculated as follows: The sample entropy value of each IMF is calculated by the sample entropy method. The k-means clustering algorithm is used to divide the IMF into high-frequency components, low-frequency components and trend components according to the frequency characteristics.

4. The method for predicting diesel engine nitrogen oxide emissions based on MMoE-LSTM according to claim 1, characterized in that: In step S3, the constructed MMoE-LSTM prediction model is: MMoE is a multi-task learning model that selects different expert networks through multiple gating networks to learn the connections and differences between different tasks. The expression of the MMoE model is as follows: y k =h k (f k (x)) Among them, y k represents the output of the kth task, h k (·) is the output layer function of the kth task, f k (x) represents the intermediate result after being processed by the expert network. Next, the processing result of the expert network is explained: f k (x) is the output of all expert networks f i The weighted sum of (x), weight Determined by the gating network: g k (x)=softmax(W gk x) The output of the gating network is the weights of all experts, W gk ∈R n×d represents a trainable matrix, n is the number of experts, and d is the dimension of the feature; When building the expert network of MMoE, the LSTM network can be combined to build an expert network that can remember and process time series information. Through this design, the expert network can better capture and utilize the feature associations between different tasks, thereby improving the performance and generalization ability of the model in multi-task learning scenarios. The LSTM network consists of an input gate, an output gate, and a forget gate: the forget gate is responsible for calculating which information in the hidden state and cell state of the current time step should be forgotten or retained; the input gate is responsible for controlling how much of the current input information can enter the cell state and affect the update of the memory unit; the output gate is responsible for determining what information the hidden state of the current time step should contain. The specific calculation process is as follows: The forget gate fi determines whether it is necessary to forget the information in the unit state at the previous time; f t =σ(W f [c t-1 ,x t ]+b f ), Input Gate i t , determine the new information added to the cell state, The input gate i t The candidate values ​​generated; I t =σ(W i [c t-1 ,x t ]+b i ), Combined with forget gate f i and output gate i t The output updates the unit state; Output gate o t , output unit state value; γ t is the hidden state output; the t =σ(W o [c t-1 ,x t ]+b o ), γ t =o t tanh(C t ), In the formula, x t , represents the input at time step t; γ t-1 and γ t Represent the previous state and current state of the hidden layer respectively; C t-1 and C t They are the previous and current unit storage information respectively; W f , W i , W o , W C is the weight matrix; b f 、b i 、b C 、b o is the offset vector; σ is the sigmoid activation function.

Images