CN112241609A - Real-time estimating system for NOx emission of diesel engine - Google Patents

Real-time estimating system for NOx emission of diesel engine Download PDF

Info

Publication number
CN112241609A
CN112241609A CN202011101762.2A CN202011101762A CN112241609A CN 112241609 A CN112241609 A CN 112241609A CN 202011101762 A CN202011101762 A CN 202011101762A CN 112241609 A CN112241609 A CN 112241609A
Authority
CN
China
Prior art keywords
nox
neural network
lstm
diesel engine
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011101762.2A
Other languages
Chinese (zh)
Inventor
宫洵
孙萌鸽
胡云峰
赵靖华
孙耀
张辉
陈虹
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jilin University
Original Assignee
Jilin University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Jilin University filed Critical Jilin University
Priority to CN202011101762.2A priority Critical patent/CN112241609A/en
Publication of CN112241609A publication Critical patent/CN112241609A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

A real-time NOx emission estimation system of a diesel engine belongs to the technical field of diesel engine control. The invention aims to select the pressure intensity of an intake manifold, the residual fraction of exhaust gas, the fuel injection quantity, the engine rotating speed and the EGR valve flow as input variables of an LSTM-RNN (least squares-Rich neural network) and train a diesel engine NOx emission real-time estimation system of the LSTM-RNN neural network. The method comprises the following steps: acquiring a data set related to the operating condition of the diesel engine, selecting input data of a NOx real-time estimation model, selecting a training set and a test set, selecting data for preprocessing, determining structural parameters of a recurrent neural network, and establishing an LSTM recurrent neural network NOx prediction model. According to the invention, the real-time accurate estimation of the NOx of the diesel engine is realized, and through the verification of a diesel engine rack, compared with the traditional BP neural network modeling, the estimation accuracy of the diesel engine NOx real-time estimation system provided by the invention is respectively improved by 17.82%, 50.35% and 32.34% under a steady-state working condition, a transient working condition and a comprehensive working condition.

Description

Real-time estimating system for NOx emission of diesel engine
Technical Field
The invention belongs to the technical field of diesel engine control.
Background
The problem of air quality has become more serious in recent years, and an important part of the causes of the air quality degradation is excess NOx (nitrogen oxides) in the air. Nitric acid and nitrate generated by combining NOx with water in the air are one of the main causes of acid rain, photochemical smog pollution can be caused under certain conditions, and if the NOx is inhaled into a human body, the NOx seriously harms the lung of the human body and causes various respiratory diseases. The treatment of NOx in the air is therefore an extremely important aspect in the abatement of atmospheric pollution.
Diesel vehicles are one of the major sources of NOx in air. The diesel engine is widely applied in the fields of transportation and the like due to the advantages of large torque and good economic performance, and NOx generated by diesel combustion makes diesel vehicles become the key point for treating atmospheric NOx. Currently, most areas in China have begun to implement national VI emission regulations, wherein the limit on the NOx emission of diesel vehicles is particularly strict, which is a great problem for the control of the NOx of the diesel vehicles. The prevention and control of the NOx emission mainly focuses on the control of a purification and post-treatment part in a diesel engine, and in order to effectively reduce the NOx emission of the diesel engine, closed-loop control is required to be carried out on the NOx emission, real-time NOx emission is fed back to a controller, the diesel engine is accurately controlled, and the accurate real-time measurement of the NOx is required to be realized. Currently, the real-time NOx sensor still stays in the laboratory research stage, and no technically mature NOx emission sensor on the market can measure NOx in real time, which makes the control of NOx to be bottleneck. There are several patents currently focusing on the study of NOx sensors and measurement devices. For example, in patent CN 106014571B, the corrected NOx sensor value or NOx MAP of the engine is used to determine the NOx emission value of the engine, including exhaust gas discharge pipe, exhaust gas pretreatment device, heating sampling pipeline, converter, nitrogen oxide gas analyzer, condensation tank, and pressure regulating valve; patent CN106841518B discloses a flue gas NOx concentration measurement method based on kalman filtering; patent CN 207114514U discloses a measuring device for NOx in exhaust pollutants of marine engines; patent CN 104897763B discloses a nitrogen oxygen sensor and a method for measuring the NOx content in exhaust gas.
However, research on nitrogen oxide sensors and measuring devices cannot accurately measure NOx in real time completely, and can only be used for off-line analysis of NOx emission of diesel engines. Therefore, in order to accurately control the NOx emission, the NOx emission is generally estimated in real time from the perspective of mechanism modeling, soft measurement of NOx is realized, and the current NOx emission is calculated through other related variables which are easy to measure in the diesel engine. This requires analysis of the cause of NOx generation starting from the mechanism of NOx generation. The combustion reaction in the diesel engine is very complex, related knowledge such as heat transfer science, combustion science and the like is involved, influence factors are more, and a generation model of the NOx has very strong nonlinear coupling characteristics, so that great challenges are brought to modeling of the diesel engine and design and solution of a model-based NOx controller.
Considering the great difficulty of modeling the NOx generation mechanism, some patents propose data-based methods. Patent CN 1047151420B discloses a dynamic soft measurement method for NOx emissions of utility boilers, which uses a particle swarm algorithm to model NOx emissions, thereby tracking the change of NOx emissions. Patent CN 110032747 a and patent CN 106680428A established a model of NOx emissions based on a support vector machine. Patent CN 108647483A models the NOx concentration of flue gas that thermal power plant boiler burning produced based on fuzzy tree, realizes its soft measurement. The NOx emission model obtained by the method considers the operation data of the diesel engine, however, the traditional modeling mode has many parameters, the generalization capability of the model still needs to be improved, and a distance exists for real-time accurate estimation of the NOx emission of the diesel engine.
Patent CN 104331736B uses RBF neural network to predict boiler NOx emissions. Neural networks can learn and store a large number of input-output mappings without prior disclosure of mathematical equations describing such mappings. With the wide application, the disadvantages of the method are revealed, such as the network convergence speed is slow, the overfitting is easily caused by more parameters, and the network structure is not selected.
At present, the rapid development of artificial intelligence provides a new idea for modeling of NOx, and key big data related to NOx emission can be extracted from an engine rack and modeled by adopting a deep learning technology. The concept of deep learning was proposed in 2006 by Hinton et al to discover a distributed feature representation of data by combining lower level features to form a more abstract higher level representation attribute class or feature. Compared with the traditional Neural Network, the Recurrent Neural Network (RNN) introduces the concept of deep learning into the structure of the Neural Network, which is a new data processing mode in recent years. The RNN variant Long Short Term Memory (LSTM) neural network adds a plurality of gate structures on the basis of the RNN, solves the problem of Long Term dependence in the learning process, further improves the adaptability of the neural network to data, and improves the estimation precision of output. The LSTM-RNN neural network has wide application in the aspects of language models, image processing, voice recognition, machine translation and the like in recent years, and the application of the LSTM-RNN neural network in the field of NOx emission of diesel engines is blank.
Disclosure of Invention
The invention aims to select the pressure intensity of an intake manifold, the residual fraction of exhaust gas, the fuel injection quantity, the engine rotating speed and the EGR valve flow as input variables of an LSTM-RNN (least squares-Rich neural network) and train a diesel engine NOx emission real-time estimation system of the LSTM-RNN neural network.
The method comprises the following steps:
step one, acquiring a data set related to the operating condition of a diesel engine;
selecting input data of a NOx real-time estimation model, selecting five indexes for predicting NOx emission, and pimIntake manifold pressure, F1Residual fraction of exhaust gas, QvQuantity of fuel injected, NeEngine speed, wEGREGR valve flow;
need to be extracted from the travel data setData matrix X ofinIs composed of
Figure BDA0002725611920000021
Where the first five terms are inputs for prediction and the sixth term is the model predicted NOx emissions cNOx；
Step three, selection of training set and test set
Extracting a data set XinThe first 80% of the training data are training data XtrainThe LSTM recurrent neural network prediction model is used for training NOx; the last 20% is test data XtestAfter the model is established, verifying the prediction precision of the model;
step four, selecting data for preprocessing
In order to improve the prediction accuracy of the NOx prediction model, normalization processing needs to be carried out on the selected data;
step five, determining structural parameters of the recurrent neural network, determining parameters of an input layer, a hidden layer and an output layer of the LSTM recurrent neural network on a TensorFlow platform, and selecting the structural parameters of the recurrent neural network
The method comprises the following steps of inputting a characteristic quantity 5, a hidden layer number 1, a hidden layer neuron number 2, an output characteristic quantity 1 and a time step length 1-2; step six: establishing an LSTM circulating neural network NOx prediction model, and carrying out LSTM-RNN neural network training on a TensorFlow platform comprises the following steps:
(1) reading training data and carrying out normalization processing
(2) Setting LSTM-RNN neural network structure parameters: the input dimension, the output dimension, the number of hidden layer units in the LSTM, the time step length and the input dimension are 5; the output dimension is 1; the LSTM comprises 1 hidden layer, the number of the hidden layer units is 2, the time step is 1 sampling step, and then the NOx emission data at the current moment can be estimated;
(3) setting deep learning parameters: learning rate and sequence segment batch size, wherein the learning rate is set to be 0.01, and the sequence segment batch size is set to be 100;
(4) training neural network parameters, in a circulating neural network structure, adopting a tanh function instead of a sigmoid function as an activation function of an output layer, wherein the sigmoid function is expressed as
Figure BDA0002725611920000031
Having a derivative function of
Figure BDA0002725611920000032
The loss function in the training process adopts a standard form of square loss, and the expression form of the square loss is
Figure BDA0002725611920000033
Wherein c isNOx(i) And
Figure BDA0002725611920000034
respectively, a measured value of NOx emissions at time i and an LSTM-RNN model estimated value.
Aiming at the problem that the NOx emission of the diesel engine is difficult to model by a mechanism, the invention provides a diesel engine NOx emission real-time estimation system based on an LSTM-RNN neural network, and the real-time accurate estimation of the NOx emission of the diesel engine is realized by mechanism analysis and data screening. Through the verification of a diesel engine rack, compared with the traditional BP neural network modeling, the diesel engine NOx real-time estimation system provided by the invention has the advantages that the estimation accuracy is respectively improved by 17.82%, 50.35% and 32.34% under the steady-state working condition, the transient working condition and the comprehensive working condition.
Drawings
FIG. 1 is a flow chart of an LSTM-RNN build diesel NOx emission estimation system;
FIG. 2 is a control block diagram of a real-time NOx emission estimation system applied to NOx emission control;
FIG. 3 is a schematic diagram of a single-layer structure of a conventional Recurrent Neural Network (RNN);
FIG. 4 is a schematic diagram of the LSTM neural network module structure of the present invention;
FIG. 5 is a schematic diagram of a multi-layer LSTM neural network architecture;
FIG. 6 is a comparison graph of steady state estimation effect of a diesel engine NOx emission real-time estimation system established based on an LSTM-RNN neural network under the condition that the engine speed is 1200RPM in the invention; FIG. 6a is a value comparison of NOx emissions after normalization; FIG. 6b is a comparison of NOx emission model versus predicted error values;
FIG. 7 is a graph comparing the transient estimation effect of the diesel engine NOx emission real-time estimation system based on the LSTM-RNN neural network at 2000RPM in the present invention; FIG. 7a is a value comparison of NOx emissions after normalization; FIG. 7b is a comparison of NOx emission model versus predicted error values;
FIG. 8 is a graph comparing the transient prediction effect of a diesel NOx emission model based on an LSTM-RNN neural network at 2400RPM for an engine of the present invention; FIG. 8a is a value comparison of NOx emissions after normalization; FIG. 8b is a comparison of NOx emission model versus predicted error values;
fig. 9 is a field diagram of a diesel gantry used for data sampling in the present invention.
Detailed Description
The detailed process of the invention is as follows:
the method comprises the following steps: and acquiring a data set related to the operating condition of the diesel engine. In the actual operation process of the diesel engine rack, time series data of relevant key indexes of a gas circuit and an oil circuit of the diesel engine are obtained, and then a real-time estimation model of NOx emission of the diesel engine is obtained through data set training. The CA4D32 diesel engine bench for acquiring the diesel engine data set of the invention is shown in FIG. 9, and the displacement is 3.17L and the compression ratio is 17: 1.
Step two: and selecting input data of the NOx real-time estimation model. Analyzing the correlation of the NOx emissions with the indicators in the data set obtained in step one. The related characteristic quantity can be selected from two modes of mechanism analysis and data analysis. And taking the selected characteristic quantity as input data of the NOx prediction model. Five criteria were selected for predicting NOx emissions, as shown in Table 1.
TABLE 1NOx predict related features
Figure BDA0002725611920000041
The fuel injection quantity, the engine speed and the EGR valve flow can be directly obtained from the engine running state, and the pressure intensity of an intake manifold and the exhaust gas residual fraction can also be obtained by measuring through sensors, so that the 5 characteristic input quantities of the selected NOx emission are easy to obtain, and the realizability of the invention is enhanced. Data matrix X to be extracted from a driving datasetinIs composed of
Figure BDA0002725611920000042
Where the first five terms are inputs for prediction and the sixth term is the model predicted NOx emissions cNOx。
Step three: and selecting a training set and a testing set. In order to improve the prediction accuracy of the prediction model under each operation condition, a data set X is extractedinThe first 80% of the training data are training data XtrainThe LSTM recurrent neural network prediction model is used for training NOx; the last 20% is test data XtestAnd after the model is established, verifying the prediction accuracy of the model.
Step four: and selecting data for preprocessing. In order to improve the prediction accuracy of the NOx prediction model, normalization processing needs to be performed on the selected data. The mapminmax function in MATLAB can realize the normalization and de-normalization processing of data, so that the preprocessing of the data can be realized by using the mapminmax function.
Step five: and determining structural parameters of the recurrent neural network. And determining parameters of an input layer, a hidden layer and an output layer of the LSTM recurrent neural network on a TensorFlow platform. The prediction accuracy of the prediction model increases with the increase of the complexity of the recurrent neural network structure, and the recurrent neural network structure is too complex, which results in the increase of the calculation amount and the reduction of the calculation efficiency. The recurrent neural network structure parameters chosen herein are therefore shown in table 2.
TABLE 2 LSTM recurrent neural network architecture parameters
Figure BDA0002725611920000043
Step six: and establishing an LSTM recurrent neural network NOx prediction model. Training data X using the model obtained in the previous steptrainThe LSTM diesel NOx emission model was trained on the TensorFlow platform. The traditional BP neural network adopts a static prediction method, however, the internal structure of the diesel engine is complex, the change of NOx emission has strong dynamics, and the NOx emission is difficult to be accurately predicted by the static prediction method, so that the method adopts a circulating neural network to predict the NOx emission. The LSTM recurrent neural network is additionally provided with three gate structures, namely a forgetting gate, an input gate and an output gate, on the basis of a common recurrent neural network, so that signals can be updated in real time in the prediction process, the problem of long-term dependence is solved on the basis of RNN, and a more accurate prediction result can be obtained.
In the invention, an LSTM-RNN neural network structure is trained based on a TensorFlow platform. TensorFlow is an open resource platform developed by Google Brain based on dataflow programming, and is widely applied to the open field of deep learning. The TensorFlow provides a Python programming interface, and the deep learning process is realized based on the Python language. The LSTM-RNN neural network training on the TensorFlow platform comprises the following steps:
(1) and reading training data and carrying out normalization processing.
(2) Setting LSTM-RNN neural network structure parameters: input dimension (Input dimension), Output dimension (Output dimension), number of Hidden layer units (Hidden units) in LSTM, and Time step (Time step). The input dimension is 5; the output dimension is 1; the LSTM includes 1 hidden layer, and the number of hidden layer units is 2. The time step in the invention is 1 sampling step, namely the NOx emission data at the current moment can be estimated (soft measurement).
(3) Setting deep learning parameters: learning rate (Learning rate), sequence piece Batch size (Batch size). In the invention, the learning rate is set to be 0.01, and the sequence segment batch size is set to be 100.
(4) And training the neural network parameters. In the recurrent neural network structure herein, the activation function of the output layer employs a tanh function instead of a sigmoid function. sigmoid function expression is
Figure BDA0002725611920000051
Having a derivative function of
Figure BDA0002725611920000052
Obviously, the function value domain is (0,1), the derivative range is (0,0.25), and in the neural network training process, the derivative is always a small decimal number smaller than 1, so that the gradient is continuously reduced until the gradient is close to 0, and the phenomenon of 'gradient disappearance' appears. In contrast, the tan h function has a value range of (-1,1), the derivative range is (0,1), and the derivative is closer to 1, which greatly slows down the gradient descent process. And the output value of the tanh function is symmetrical about the zero point, which will also accelerate the convergence process of the neural network. Therefore, the invention adopts the tanh function as the excitation function of the output layer of the neural network.
The loss function in the training process is in the standard form of Square loss (Square loss), which is expressed in the form of
Figure BDA0002725611920000053
Wherein c isNOx(i) And
Figure BDA0002725611920000054
respectively, a measured value of NOx emissions at time i and an LSTM-RNN model estimated value.
Verification of prediction effect of NOx prediction model
Using test data XtestThe measured data in (1) estimates the NOx emission, and the estimated result is compared with the actual NOx emission value, thereby verifying the estimation modelThe effectiveness of the model. The comparison of the measured and estimated values is shown in FIGS. 6, 7 and 8, where the NOx emission values are normalized relative NOx emission values cNOx,normThe expression is
Figure BDA0002725611920000055
I.e. the ratio of the current NOx emission value to the maximum NOx emission value.
The results were quantitatively analyzed as shown in Table 3. In the table, MAE, RMSE, and MEAP are Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE), respectively, and their expressions are Mean Absolute Percentage Error (MAPE), respectively
Figure BDA0002725611920000061
Compared with the traditional BP neural network, the LSTM-RNN has larger improvement on pre-estimation precision, and particularly under the transient working condition, the LSTM-RNN neural network can more accurately estimate NOx emission under the transient working condition. When the Mean Absolute Error (MAE) is used as an evaluation index, compared with a BP neural network, the estimation errors of the LSTM-RNN neural network under a steady-state working condition, a transient working condition and a comprehensive working condition are respectively reduced by 7.82%, 50.35% and 32.34%, and the estimation accuracy is greatly improved.
And carrying out real-time estimation on the NOx emission by using an LSTM-RNN model under the actual operation condition. In the running process of the diesel vehicle, the input characteristics of the vehicle are collected, and the trained LSTM-RNN model is used for estimating the NOx emission of the diesel engine in real time and carrying out NOx feedback control.
TABLE 3 prediction effect enhancement of LSTM-RNN neural network relative to BP neural network under various working conditions
Figure BDA0002725611920000062
FIG. 1 is a flow chart for establishing a diesel NOx emission estimation system based on LSTM-RNN. Firstly, acquiring a time sequence data set in the actual operation process of the diesel engine; therefore, through mechanism analysis and data analysis, data related to the NOx emission of the diesel vehicle is extracted from the acquired time sequence data and is used as input data of a prediction model; selecting training data and test data from the training data and performing data preprocessing; determining structural parameters of a recurrent neural network prediction model; then training the training data set by using a TensorFlow platform, thereby establishing an LSTM-RNN recurrent neural network NOx real-time estimation model; verifying the estimation accuracy of the model by using the test data set; and finally, after the verification is passed, the NOx emission of the diesel engine can be estimated in real time by using the real-time NOx emission estimation system of the diesel engine established by the invention.
FIG. 2 is a control block diagram of a real-time NOx emission estimating system applied to NOx emission control, wherein the real-time NOx emission estimating system is an LSTM-RNN-based real-time NOx emission estimating system for a diesel engine according to the present invention, and a portion indicated by a dotted line below the graph is a real-time NOx emission controller based on real-time NOx emission estimation, which feeds back NOx emission c through the estimating systemNOxAnd the expected output performance of the diesel engine, and each actuating mechanism of the diesel engine is optimized to achieve the aim of reducing the NOx emission of the diesel engine.
Fig. 3 is a schematic diagram of a single-layer structure of a conventional Recurrent Neural Network (RNN). Fig. 3(a) is an overall structure diagram, and fig. 3(b) is an expanded structure diagram thereof. Information passing through cell State CtCarrying out transmission among neurons in the same layer and obtaining output htThe expression is
Ct＝f(WCt-1+Uxt)
ht＝f(VCt) (7)
The recurrent neural networks differ from the traditional neural networks by: 1) weight sharing, wherein the neural network modules A in the figure 3(b) have the same structure uniformly, and U, V and W are the same; 2) each input is only connected with the corresponding neural network module, and no connection exists between the input and other neurons.
(IV) FIG. 4 is a schematic diagram of the LSTM neural network module structure of the present invention. Unlike traditional RNNs, LSTM networks add several "gate" structures, thus solving the problem of learning long-term dependence. In the repeated neural network module of RNN, unlike the conventional RNN, which has only one single structure of tanh, there are four structures in LSTM, and interaction is performed through three "gate" structures.
First, it is the forgetting gate layer that determines which states in the cell to forget (discard). The gate layer will read the output h of the neural network module at the last momentt-1And the model input quantity x at the current momenttTo perform an operation
ft＝σ(Wf[ht-1,xt]+bf) (8)
Where σ represents a threshold function sigmoid, the output may be limited to (0,1), which is expressed as
Figure BDA0002725611920000071
② an input gate layer, the function of the gate layer is to determine to input xtWhich new information is stored in the cell state.
The calculation performed in the door layer is
Figure BDA0002725611920000072
First, itDetermining which values need to be updated by a sigmoid layer, and then, a tanh layer creates an updated candidate value vector
Figure BDA0002725611920000073
Cell status through the above two gates, discarding information about the old state, updating the new information previously confirmed, and obtaining the new cell status C by the following formulat
Figure BDA0002725611920000074
Thirdly, the output gate determines the final output of the neural network module
ot＝σ(Wo[ht-1,xt]+bo)
ht＝ot·tanh(Ct) (12)
(V) FIG. 5 is a schematic diagram of a multi-layer LSTM neural network structure.
FIG. 6 is a comparison graph of steady state estimation effect of the diesel engine NOx emission real-time estimation system established based on the LSTM-RNN neural network under the condition that the engine speed is 1200RPM, namely, the system input is kept constant in a period of time. Where the NOx emissions are normalized values as shown in equation (5). As can be seen from the figure, the estimation effect of the LSTM-RNN is obviously better than that of the BP neural network, and the time lag is relatively small. Through calculation, MAE, RMSE and MAPE values of estimation errors of the LSTM-RNN mode are reduced by 17.82%, 3.27% and 33.76% on the basis of a BP neural network, and the estimation accuracy is greatly improved.
Seventhly, fig. 7 is a comparison graph of transient estimation effect of the diesel engine NOx emission real-time estimation system established based on the LSTM-RNN neural network under the condition that the engine speed is 2000RPM, namely, the system input changes frequently. Where the NOx emissions are normalized values as shown in equation (5). As can be seen from the figure, the estimation effect of the LSTM-RNN neural network is greatly improved on the basis of the BP neural network relatively. The rapidity and the accuracy of the method are far better than those of a BP neural network, and through calculation, the MAE, RMSE and MAPE values of the estimation error of the LSTM-RNN mode are reduced by 50.35%, 31.44% and 58.20% on the basis of the BP neural network, so that the requirement on the NOx emission estimation precision can be met.
(eighth) FIG. 8 is a comparison graph of transient prediction effect of a diesel engine NOx emission model established based on an LSTM-RNN neural network at 2400RPM, i.e., the system input will change frequently. Where the NOx emissions are normalized values as shown in equation (5). As can be seen from the figure, the improvement of the engine speed makes the estimation performance of the model face a great challenge, and a relatively large time lag is caused, which is inevitable in the process of estimating NOx emission at high speed, and on the basis, the LSTM-RNN neural network model can still maintain good estimation performance.
(ninth) FIG. 9 is a field diagram of a diesel gantry used for data sampling in the present invention. This engine model was CA4D32, with a displacement of 3.17L and a compression ratio of 17: 1.
Aiming at the problem that the NOx emission of the diesel engine is difficult to measure in real time, the invention provides a NOx real-time estimation system based on an LSTM-RNN neural network, and realizes real-time accurate estimation of the NOx emission of the diesel engine. (1) Compared with the traditional BP neural network, the diesel engine NOx real-time estimation system provided by the invention has the advantages that the estimation accuracy is respectively improved by 17.82%, 50.35% and 32.34% under the steady-state working condition, the transient working condition and the comprehensive working condition. The accuracy of NOx emission estimation is greatly improved, and an advanced technical support is provided for soft measurement of NOx emission of the diesel engine. (2) The real-time NOx estimation system is applied to the NOx control system, and the improvement of the estimation precision lays a foundation for a controller based on NOx feedback.
The invention provides a diesel engine NOx real-time estimation system based on a deep learning LSTM-RNN neural network by utilizing large operation data of a diesel engine rack, five measurable characteristic quantities of pressure intensity of an intake manifold, residual fraction of waste gas, fuel injection quantity, engine rotating speed and EGR valve flow are selected to represent NOx emission of a diesel engine, and accurate real-time estimation of NOx emission is realized.

Claims (1)

1. A real-time estimating system for NOx emission of a diesel engine comprises the following steps:
step one, acquiring a data set related to the operating condition of a diesel engine;
the method is characterized in that:
selecting input data of a NOx real-time estimation model, selecting five indexes for predicting NOx emission, and pimIntake manifold pressure, F1Residual fraction of exhaust gas, QvQuantity of fuel injected, NeEngine speed, wEGREGR valve flow;
data matrix X to be extracted from a driving datasetinIs composed of
Figure FDA0002725611910000011
Where the first five terms are inputs for prediction and the sixth term is the model predicted NOx emissions cNOx；
Step three, selection of training set and test set
Extracting a data set XinThe first 80% of the training data are training data XtrainThe LSTM recurrent neural network prediction model is used for training NOx;
the last 20% is test data XtestAfter the model is established, verifying the prediction precision of the model;
step four, selecting data for preprocessing
In order to improve the prediction accuracy of the NOx prediction model, normalization processing needs to be carried out on the selected data;
step five, determining structural parameters of the recurrent neural network, determining parameters of an input layer, a hidden layer and an output layer of the LSTM recurrent neural network on a TensorFlow platform, and selecting the structural parameters of the recurrent neural network
The method comprises the following steps of inputting a characteristic quantity 5, a hidden layer number 1, a hidden layer neuron number 2, an output characteristic quantity 1 and a time step length 1-2;
step six: establishing an LSTM circulating neural network NOx prediction model, and carrying out LSTM-RNN neural network training on a TensorFlow platform comprises the following steps:
(1) reading training data and carrying out normalization processing
(2) Setting LSTM-RNN neural network structure parameters: the input dimension, the output dimension, the number of hidden layer units in the LSTM, the time step length and the input dimension are 5; the output dimension is 1; the LSTM comprises 1 hidden layer, the number of the hidden layer units is 2, the time step is 1 sampling step, and then the NOx emission data at the current moment can be estimated;
(3) setting deep learning parameters: learning rate and sequence segment batch size, wherein the learning rate is set to be 0.01, and the sequence segment batch size is set to be 100;
(4) training neural network parameters, in a circulating neural network structure, adopting a tanh function instead of a sigmoid function as an activation function of an output layer, wherein the sigmoid function is expressed as
Figure FDA0002725611910000012
Having a derivative function of
Figure FDA0002725611910000013
The loss function in the training process adopts a standard form of square loss, and the expression form of the square loss is
Figure FDA0002725611910000021
Wherein c isNOx(i) And
Figure FDA0002725611910000022
respectively, a measured value of NOx emissions at time i and an LSTM-RNN model estimated value.
CN202011101762.2A 2020-10-15 2020-10-15 Real-time estimating system for NOx emission of diesel engine Pending CN112241609A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011101762.2A CN112241609A (en) 2020-10-15 2020-10-15 Real-time estimating system for NOx emission of diesel engine

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011101762.2A CN112241609A (en) 2020-10-15 2020-10-15 Real-time estimating system for NOx emission of diesel engine

Publications (1)

Publication Number Publication Date
CN112241609A true CN112241609A (en) 2021-01-19

Family

ID=74169119

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011101762.2A Pending CN112241609A (en) 2020-10-15 2020-10-15 Real-time estimating system for NOx emission of diesel engine

Country Status (1)

Country Link
CN (1) CN112241609A (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112861436A (en) * 2021-02-18 2021-05-28 天津大学 Real-time prediction method for engine emission
CN113554153A (en) * 2021-07-23 2021-10-26 潍柴动力股份有限公司 Method and device for predicting emission of nitrogen oxides, computer equipment and medium
CN114934848A (en) * 2022-06-16 2022-08-23 哈尔滨工程大学 Fuzzy neural network modeling method for optimizing control of combustion performance of diesel engine
CN115263505A (en) * 2022-08-01 2022-11-01 天津布尔科技有限公司 Vehicle NOX emission prediction method and device and storage medium
CN115324698A (en) * 2022-08-09 2022-11-11 上海交通大学 Diesel engine tail gas analysis sensing method based on neural network
CN116070791A (en) * 2023-03-27 2023-05-05 天津布尔科技有限公司 Diesel vehicle NO based on LSTM algorithm x Emission prediction method
CN116272363A (en) * 2023-05-17 2023-06-23 浙江浙能迈领环境科技有限公司 Ship exhaust gas mixed desulfurization system and method thereof
CN117272233A (en) * 2023-11-21 2023-12-22 中国汽车技术研究中心有限公司 Diesel engine emission prediction method, apparatus, and storage medium

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2001082128A (en) * 1999-07-12 2001-03-27 Toyota Motor Corp Exhaust emission control method for internal combustion engine
JP2005002968A (en) * 2003-06-16 2005-01-06 Mitsubishi Fuso Truck & Bus Corp Exhaust emission control device of internal combustion engine
CN102182575A (en) * 2011-01-27 2011-09-14 潍柴动力股份有限公司 Equipment and method for controlling air system of diesel engine
CN102588129A (en) * 2012-02-07 2012-07-18 上海艾铭思汽车控制系统有限公司 Optimization cooperative control method for discharge of nitrogen oxides and particles of high-pressure common-rail diesel
CN103410592A (en) * 2013-07-18 2013-11-27 武汉理工大学 Diesel NOx original emission load predicting method based on crankshaft angular velocity sensor
CN108519556A (en) * 2018-04-13 2018-09-11 重庆邮电大学 A kind of lithium ion battery SOC prediction techniques based on Recognition with Recurrent Neural Network
CN108520080A (en) * 2018-05-11 2018-09-11 武汉理工大学 Automatic system of marine diesel-generator failure predication and health status online evaluation system and method
CN108681747A (en) * 2018-05-11 2018-10-19 武汉理工大学 Rotary machinery fault diagnosis based on deep learning and condition monitoring system and method
CN109447373A (en) * 2018-11-16 2019-03-08 上海海事大学 Haze method is predicted based on the LSTM neural network of python platform
CN110297178A (en) * 2018-05-11 2019-10-01 宫文峰 Diesel generating set fault diagnosis and detection device and method based on deep learning
CN110702418A (en) * 2019-10-10 2020-01-17 山东超越数控电子股份有限公司 Aircraft engine fault prediction method

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2001082128A (en) * 1999-07-12 2001-03-27 Toyota Motor Corp Exhaust emission control method for internal combustion engine
JP2005002968A (en) * 2003-06-16 2005-01-06 Mitsubishi Fuso Truck & Bus Corp Exhaust emission control device of internal combustion engine
CN102182575A (en) * 2011-01-27 2011-09-14 潍柴动力股份有限公司 Equipment and method for controlling air system of diesel engine
CN102588129A (en) * 2012-02-07 2012-07-18 上海艾铭思汽车控制系统有限公司 Optimization cooperative control method for discharge of nitrogen oxides and particles of high-pressure common-rail diesel
CN103410592A (en) * 2013-07-18 2013-11-27 武汉理工大学 Diesel NOx original emission load predicting method based on crankshaft angular velocity sensor
CN108519556A (en) * 2018-04-13 2018-09-11 重庆邮电大学 A kind of lithium ion battery SOC prediction techniques based on Recognition with Recurrent Neural Network
CN108520080A (en) * 2018-05-11 2018-09-11 武汉理工大学 Automatic system of marine diesel-generator failure predication and health status online evaluation system and method
CN108681747A (en) * 2018-05-11 2018-10-19 武汉理工大学 Rotary machinery fault diagnosis based on deep learning and condition monitoring system and method
CN110297178A (en) * 2018-05-11 2019-10-01 宫文峰 Diesel generating set fault diagnosis and detection device and method based on deep learning
CN109447373A (en) * 2018-11-16 2019-03-08 上海海事大学 Haze method is predicted based on the LSTM neural network of python platform
CN110702418A (en) * 2019-10-10 2020-01-17 山东超越数控电子股份有限公司 Aircraft engine fault prediction method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
LIU RU- I XIANG 等: "Control of diesel soot and NOx emissions with a particulate trap and EGR", 《JOURNAL OF ENVIRONMENTAL SCIENCES》 *
孙萌鸽 等: "基于数据驱动模型的柴油发动机燃烧与排放一体化控制", 《PROCEEDINGS OF THE 39TH CHINESE CONTROL CONFERENCE》 *
戴金池等: "基于LSTM神经网络的柴油机NOx排放预测", 《内燃机学报》 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112861436A (en) * 2021-02-18 2021-05-28 天津大学 Real-time prediction method for engine emission
CN113554153A (en) * 2021-07-23 2021-10-26 潍柴动力股份有限公司 Method and device for predicting emission of nitrogen oxides, computer equipment and medium
CN114934848A (en) * 2022-06-16 2022-08-23 哈尔滨工程大学 Fuzzy neural network modeling method for optimizing control of combustion performance of diesel engine
CN115263505A (en) * 2022-08-01 2022-11-01 天津布尔科技有限公司 Vehicle NOX emission prediction method and device and storage medium
CN115324698A (en) * 2022-08-09 2022-11-11 上海交通大学 Diesel engine tail gas analysis sensing method based on neural network
CN115324698B (en) * 2022-08-09 2023-06-16 上海交通大学 Diesel engine tail gas analysis sensing method based on neural network
CN116070791A (en) * 2023-03-27 2023-05-05 天津布尔科技有限公司 Diesel vehicle NO based on LSTM algorithm x Emission prediction method
CN116272363A (en) * 2023-05-17 2023-06-23 浙江浙能迈领环境科技有限公司 Ship exhaust gas mixed desulfurization system and method thereof
CN117272233A (en) * 2023-11-21 2023-12-22 中国汽车技术研究中心有限公司 Diesel engine emission prediction method, apparatus, and storage medium

Similar Documents

Publication Publication Date Title
CN112241609A (en) Real-time estimating system for NOx emission of diesel engine
Dey et al. A hybrid ANN-Fuzzy approach for optimization of engine operating parameters of a CI engine fueled with diesel-palm biodiesel-ethanol blend
Li et al. Emissions modeling of a light-duty diesel engine for model-based control design using multi-layer perceptron neural networks
CN107203832B (en) Supercharged diesel engine optimal EGR performance evaluation method based on subjective and objective weighting-multi-objective grey decision-grey correlation analysis
Johri et al. Optimal energy management for a hybrid vehicle using neuro-dynamic programming to consider transient engine operation
Norouzi et al. Machine learning-based diesel engine-Out NOx reduction using a plug-in PD-type iterative learning control
Mariani et al. Diesel engine NOx emissions control: An advanced method for the O2 evaluation in the intake flow
Solmaz et al. Comparison of artificial neural network and fuzzy logic approaches for the prediction of in-cylinder pressure in a spark ignition engine
CN114037163A (en) Sewage treatment effluent quality early warning method based on dynamic weight PSO (particle swarm optimization) optimization BP (Back propagation) neural network
Liao et al. A comparative investigation of advanced machine learning methods for predicting transient emission characteristic of diesel engine
Ma et al. Improvement on the linear and nonlinear auto-regressive model for predicting the NOx emission of diesel engine
Bellone et al. Comparison of CNN and LSTM for Modeling Virtual Sensors in an Engine
Al-Himyari et al. Review of air fuel ratio prediction and control methods
Hao et al. Prediction of nitrogen oxide emission concentration in cement production process: a method of deep belief network with clustering and time series
CN112613237B (en) CFB unit NOx emission concentration prediction method based on LSTM
CN116741304A (en) Diesel engine original exhaust nitrogen oxide prediction method based on Internet of vehicles big data
CN116720144A (en) Free piston linear motor operation mode identification method based on data fusion and feature extraction
Warth Comparative investigation of mathematical methods for modeling and optimization of Common-Rail DI diesel engines
Mariani et al. Artificial intelligence methodologies for oxygen virtual sensing at diesel engine intake
Dimitrakopoulos Real-time Virtual Sensor for NOx emisions and stoichiometric air-fuel ratio λ of a Marine Diesel Engine Using Neural Networks
CN114282680A (en) Vehicle exhaust emission prediction method and system based on machine learning algorithm
Liao et al. Prediction of the transient emission characteristics from diesel engine using temporal convolutional networks
Sundaram et al. Modeling of transient gasoline engine emissions using data-driven modeling techniques
Ventura et al. Recurrent Neural Network to Estimate Intake Manifold O 2 Concentration in a Diesel Engine
CN113722837B (en) Method for establishing diesel vehicle fuel consumption and emission model according to data clusters of different working conditions

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20210119