Diesel engine modeling based on recurrent neural networks for a hardware-in-the-loop simulation system of diesel generator sets

Diesel engine modeling based on recurrent neural networks for a hardware-in-the-loop simulation system of diesel generator sets

Accepted Manuscript Diesel engine modeling based on recurrent neural networks for a hardware-in-the-loop simulation system of diesel generator sets M...

1MB Sizes 0 Downloads 60 Views

Accepted Manuscript

Diesel engine modeling based on recurrent neural networks for a hardware-in-the-loop simulation system of diesel generator sets Mingxin Yu , Xiaoying Tang , Yingzi Lin , Xiangzhou Wang PII: DOI: Reference:

S0925-2312(17)31916-1 10.1016/j.neucom.2017.12.054 NEUCOM 19188

To appear in:

Neurocomputing

Received date: Revised date: Accepted date:

8 May 2017 18 November 2017 18 December 2017

Please cite this article as: Mingxin Yu , Xiaoying Tang , Yingzi Lin , Xiangzhou Wang , Diesel engine modeling based on recurrent neural networks for a hardware-in-the-loop simulation system of diesel generator sets, Neurocomputing (2017), doi: 10.1016/j.neucom.2017.12.054

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Diesel engine modeling based on recurrent neural networks for a hardware-in-theloop simulation system of diesel generator sets Mingxin Yua, b, Xiaoying Tanga, *, Yingzi Linb, Xiangzhou Wangc a

School of Life Sciences, Beijing Institute of Technology, 5 South Zhongguancun

b

CR IP T

Street, Haidian District, Beijing 100081, China Intelligent Human-Machine Systems Lab, College of Engineering, Northeastern University, 360 Huntington Ave, Boston, MA 02115, USA c

School of Automation, Beijing Institute of Technology, 5 South Zhongguancun Street,

AN US

Haidian District, Beijing 100081, China

Mingxin Yu: [email protected],

Xiaoying Tang: [email protected],

M

Yingzi Lin: [email protected],

ED

Xiangzhou Wang: [email protected]

PT

*

Correspondence information: Full name: Xiaoying Tang,

CE

Affiliation: School of Life Sciences, Beijing Institute of Technology, Detailed

AC

permanent address: 708 room, 5 building, 5 South Zhongguancun Street, Haidian District, Beijing 100081, China, Email address: [email protected],

Telephone number: +86-13301318097, Fax number: +86-010-68918820

1

ACCEPTED MANUSCRIPT

Abstract The electronic speed governors are widely used in diesel generator sets (DGS). To develop and debug electronic speed governor, the best option is to build a hardware-inthe-loop (HIL) simulation system. In the HIL simulation system, the physical diesel

CR IP T

engine is replaced with its mathematical model for reducing the cost and producing less emissions. To meet the requirement of closing to the real environment, the performance of mathematical model representatives is very important. This paper presents a diesel engine modeling method based on recurrent neural networks (RNNs). This

AN US

mathematical model is identified and estimated using the real data from one physical DGS. The experimental results showed that the proposed model accurately reproduced the diesel engine output characteristics with the changes of electrical power loads. To validate the proposed model, the simulation experiment was conducted on the

M

established HIL simulation system. In the simulation experiment, the rack displacement

ED

and rotational speed were measured from the physical part of the HIL simulation system. The simulation result has been confirmed that the proposed model could well simulate

PT

the loading and unloading processes of the DGS.

AC

CE

Keywords: HIL, RNNs, Diesel Engine, Diesel Generator Sets, Dynamic Characteristics

2

ACCEPTED MANUSCRIPT

1. Introduction As electrical power supplies, the diesel generator sets (DGS) are widely used in many fields, such as agriculture, marine electrical power plate, mobile power station, emergency system, etc. The performance of its electronic speed governor is directly

CR IP T

responsible for the AC (Alternating Current) power outputs quality of the DGS, since the rotational speed of the diesel engine in an AC power generation system is proportional to the AC power frequency. Therefore, the electronic speed governor and its control functions need to be optimized and tested for each specific diesel governor

AN US

project. To reduce the number of speed governor tests in a real diesel generator, which are very expensive and produces large amounts of emissions, HIL (Hardware-in-theloop) simulation systems are used to develop and validate the electronic speed governor unit [1]. In this way, expensive test bench hours can be partly replaced by less

M

expensive HIL hours.

ED

The key component of this HIL simulation system is the diesel engine modeling, which is able to simulate and reproduce the static and dynamic characteristics of the

PT

diesel engine. In particular, it means that this model can accurately reflect the relationship between the output (rotational speed) and inputs (the rack of displacement

CE

and electrical power loads) in real time applications. According to literatures,

AC

researchers have proposed several of mathematical modeling methods for the diesel engine simulation [2]. Those methods can be classified into two categories: mechanism modeling and system identification. Mechanism modeling methods need to know the professional knowledge about the diesel engine [3-4]. Generally, the diesel engine is regarded as both a thermodynamic device and a kinematical device, so a variety of dynamic processes occur along with the

3

ACCEPTED MANUSCRIPT

running of the diesel engine, such as intake and exhaust dynamic, fueling dynamic, combustion and heat release, heat transfer and power production. Based on those complicated processes, the mechanism modeling methods need to consider making the model simplifications and some hypotheses for reducing the dimensions of complexity,

CR IP T

such as linearization model [5], filling-and-emptying model [6], and quasi-steady state model [7], etc. However, in certain cases it is very hard to give an optimal solution for the diesel engine model. More importantly, the limitation of its computation time makes it very hard to apply those models on the HIL simulation system, which satisfies with

AN US

the real-time requirement [8-9]. For those reasons, therefore, models derived from the mechanism of the diesel engine cannot well meet the requirements of high accuracy and short run-time at the same time.

System identification method can be seen as an experiment modeling method. It has a

M

strong ability to synthesize input-output information of dynamic systems. Therefore,

ED

this method is widely used to establish models for the nonlinear system. Typical modeling methods include nonlinear auto regressive mode with exogenous (NARX) and

PT

Artificial Neural Networks (ANNs). Liu et al. used the NARX modeling method to establish a single-input signal-output system for the diesel engine [10]. However, this

CE

system ignored the interaction between electrical power loads and rotational speed of

AC

the diesel engine. Ai et al. and Jiang et al. proposed a two-input one-output NARX model, which fully considered the influence of the power loads exerted on the diesel engine models [11-12]. Additionally, their proposed models have been successfully applied on a HIL simulation system. Artificial Neural Networks have proved that it can be used as an effective modeling method for the diesel engine, because it not only satisfies with their ability to manage

4

ACCEPTED MANUSCRIPT

system’s complexity, but also with their capability of real-time calculations. Therefore, many researchers have taken the ANN approach to predict characteristics of the diesel engine and applied it in practical engineering, such as predicting engine exhaust emissions, engine fault diagnosis, engine dynamic characteristics, etc. Clark et al. [33],

CR IP T

in an effort to predict NOx emissions of the diesel engine, showed that ANN models trained by axle torque and axle speed as input variables were well able to complete the task of NOx emissions prediction. Shanmugam et al. [34] deals with ANN modeling to predict the performance and exhaust emissions of a diesel engine at different load

predict

the

emission

and

AN US

conditions. Similar work provided by Kannan [35], an ANN model was adopted to combustion

characteristics

of

a

diesel

engine.

Mohammadhassani et al. [36] also used the ANNs to model the relationship between NOx emissions and operating parameters of a diesel engine, such as various engine

M

speeds, mass fuel injection rates, and intake temperatures. For fault diagnosis of diesel

ED

engine, the ANNs achieved a good performance. Zhu [37] taken the ANNs to build the relationship between vibration characteristics and engine working conditions for

PT

researching fault diagnosis of the diesel engine. In order to analyze the influence of the vibration on the diesel engine, Dehkordi et al. [38] applied the ANNs in building the

CE

relationship between the various fuel blends and vibration effects of a diesel engine.

AC

Artificial Neural Networks are also one of the most well-known methods to predict dynamic characteristics of the diesel engine. Mohd Noor et al. [39] used an artificial neural networks modelling of a diesel engine to predict the output torque, brake power, brake specific fuel consumption and exhaust gas temperature. To simulate the processes of the diesel engine cylinder, Brzozowska et al. [40] proposed an ANN model and well reproduced the processes. With the aid of the ANNs technique, Billings et al. [13]

5

ACCEPTED MANUSCRIPT

presented a dynamic model for the diesel engine, which used the fuel injection quantity and power loads as two inputs, the output power and the rotational speed as two outputs. Arsie et al. [14] used the recurrent neural networks for simulating air-fuel ratio dynamics in spark ignition engine, and then the authors further performed a new

CR IP T

research that adopted neural networks to build a diesel engine model for analyzing and estimating pollutant emissions. Not only that, artificial neural networks have been used in different research areas of automotive applications [15-19]. In addition, the good performance of Artificial Neural Networks could also be well used as a control method

AN US

to apply in mechanical and robotic fields [41-43].

Based on above investigations, the goal of this study is to take the ANN technique as a modeling approach to build a diesel engine model, which can accurately simulate the diesel engine output characteristics. More specifically, this paper presents a diesel

M

engine model capable of reproducing the diesel engine speed changes when loading and

ED

unloading electrical power. To meet the requirement, we proposed a diesel engine model based on recurrent neural networks (RNNs) modeling method. The RNN with

PT

feedback connections have been demonstrated to be well suited for modeling nonlinear system on time series tasks [20-24]. Because the feedback connections are considered

CE

among the neurons, a form of memory is incorporated in RNN. To show its superiority,

AC

we recorded the compared results with other identification modeling methods. Based on the proposed model, this study successfully applied it on the established hardware-inthe-loop simulation system. The simulation result demonstrated that the proposed model could well simulate the dynamic characteristics of the diesel engine. This paper is arranged as follows. After an overview of basic structure of the diesel generator, the experimental setup and data acquisition are described in Section 2.

6

ACCEPTED MANUSCRIPT

Section 3 presents the architecture of the diesel engine model and parameters optimization method. Experimental results for the proposed diesel engine model is reported in Section 4. Section 5 gives the simulation results of the diesel engine model run on a HIL simulation system. The conclusions are provided in Section 6.

CR IP T

2. Experimental Data The structure diagram of the DGS is illustrated in Fig. 1. It is mainly consist of the diesel engine, electronic governor, generator, transmission system and power load bank. The electronic governor is used to drive the actuator for controlling the fuel quantity in

AN US

order to changing the rotational speed of the diesel engine. The DGS used in this paper

ED

M

is with AC380V@50Hz voltage output and rated power 30kW.

Fig. 1. The structure diagram of the DGS.

PT

2.1. Setup

According to above DGS basic structure, the rotational speed of the diesel engine is

CE

not only determined by the actuator, but also influenced by electrical power loads

AC

exerted on the diesel engine. Therefore, the DGS can be seen as a system with two inputs and one output. The two inputs are the rack displacement and the electrical power load, respectively. Fuel injection system driven by the actuator is measured by the rack displacement. The output is the rotational speed of the diesel engine. To build the diesel engine model, the actual data of rotational speed, the rack displacement, and electrical power loads should be acquired by the experiment. Fig. 2

7

ACCEPTED MANUSCRIPT

shows the block diagram of the experimental system. The components used in this experimental system are listed in Table 1. In Fig. 2 the actuator with an embedded rack displacement sensor is used to drive injector rocker, and the signal (green line) from the embedded displacement sensor is sent to the data acquisition card. Two magnetic speed

CR IP T

pickup sensors are mounted on the flywheel housing of the diesel engine. One is for the electronic speed governor, another for the data acquisition card (blue line). A resistive load bank provides equivalent power loads for the DGS. The load bank is configured by manually to produce different power loads exerted on the generator. The power loads

AN US

and the temporal information of loading and unloading power can be sent to data acquisition card (red line). Therefore, the computer can record real-time data of the rotational speed, the rack displacement, and the power loads online for the identification of the diesel engine model.

M

Table 1 The components used in the experimental system. Model

Manufacturer

BFM1011

DEUTZ, Germany

TF30-4

Lanzhou Electric Corporation, China

ABD225

GAC, USA

Governor

ESD-5100

GAC, USA

Speed Pickup

SZCB-01

Wuxi Longsheng Technology Co., LTS, China

Load Bank

RCL30kW AC Load Bank Kaixiang Technology Beijing Co., LTD, China

DAQ

USB-6341

NI, USA

Computer

Thinkpad T400

Lenovo, China

Generator

AC

CE

Actuator

PT

Diesel Engine

ED

Devices

8

AN US

CR IP T

ACCEPTED MANUSCRIPT

Fig. 2. The block diagram of experimental system for data acquisition.

M

2.2. Procedure and Data

The aim of this study is to train the diesel engine model using experimental data that

ED

can reproduce dynamic performances of the DGS under different power loading and

PT

unloading conditions. For this purpose, we performed experiments representing typical power loads. Specially, the size of power loads applied to the diesel engine was

CE

specified to 25%, 50%, 75%, and 100% of its rated power, i.e., 7.5kW, 15kW, 22.5kW, and 30kW, respectively. Before the commencement of the experiment, the rated speed

AC

of the diesel engine was set into the electronic speed governor. In this study, the rated speed of the diesel engine was 1236 rpm with 0.1 second sampling period. The collected experimental data includes the rotational speed, the rack displacement,

and the size of power loads. To obtain the representations of different power loads, a succession of sudden loading and unloading experiments was performed on the DGS. An example of those experimental data is presented in Fig. 3. For this study, a total of

9

ACCEPTED MANUSCRIPT

16 experiments was used to train, validate and test the proposed diesel engine model. Specially, 12 data sets were selected as the training data, 2 data sets as the validation

PT

ED

M

AN US

CR IP T

data, and 2 data sets as the testing data.

Fig. 3. An example set of experimental data.

CE

3. Methods

This section describes the basic theory and methods used for building the diesel

AC

engine model. Firstly, the NARX RNN architecture is given, its purpose is to compare with our architecture. Then, the modeling architecture base on the RNNs for this study is proposed. Lastly, with the proposed architecture, the diesel engine model and learning method are presented, respectively. 3.1. NARX RNN An important class of discrete-time nonlinear model is the nonlinear auto regressive

10

ACCEPTED MANUSCRIPT

mode with exogenous (NARX) defined by the following equation: y  k   f  y  k  1 , y  k  2  





, y k  n y , u  k  , u  k  1 , u  k  2  ,

, u  k  nu   

(1)

where y  k  and u  k  represent inputs and outputs of the system, respectively;





CR IP T

y k  n y and u  k  nu  are the previous values of output and input signals; k is the

discrete time index; n u and n y represent the input order and output order of the system, respectively; f    is a nonlinear function. When the function f    is expressed by a multilayer perception (MLP), the resulting system is called NARX recurrent neural

AC

CE

PT

ED

M

AN US

network [25]. Its architecture is shown in Fig. 4.

Fig. 4. The architecture of NARX RNN.

The left two rectangles represent tapped delay lines ( TDL ). The input signals enter the TDL and pass through N-1 delays. The output of the TDL is a N-dimensional vector made up of the current and previous input signals. The number of delays in the TDL for

11

ACCEPTED MANUSCRIPT

the input variable u  t  is n u and for the output variable y  t  is n y . Based on the architecture of NARX RNN, next subsection proposes a modified model. It represents a recurrent neural network (RNN), which fits the purpose of non-linearity of the problem. Its feature of recurrence enables the model to take into account not only the precedent

CR IP T

state of the system output, but also the internal state of each neuron. 3.2. Proposed Architecture

According to the knowledge of NARX RNN architecture, we learned that the feedback connection was just established between the output layer and the input layer.

AN US

In this study, we proposed a network architecture with external and internal connections. Fig. 5 shows the architecture of the designed RNN with input layer, layer 1 (hidden layer) and layer 2 (output layer). The RNN has a real-valued time-delay which is so

M

called Tapped Delay Lines ( TDL ). In the proposed architecture, three types of delays are introduced with the descriptions of input delays ( TDLin ), output delays ( TDL fb1,2 )

ED

and internal delays ( TDL fb1,1 and TDL fb 2,2 ). The input TDLin allows to delay the input

PT

(red line) to the RNN, the output TDL fb1,2 adds a recurrent connection (green line) from output layer to input layer, and the internal TDL fb1,1 and TDL fb 2,2 add recurrent

CE

connections (blue lines) from outputs of layer 1 and layer 2 to themselves, respectively.

AC

Thereby, the proposed neural network depends not only on the current input, but also previous inputs, outputs, and internal feedback states. The equations for three types of TDL are represented as:

TDLin   p  t  1 , p  t  2  ,

, p  t  nin 

TDL fb1, 2  a 2  t  1 , a 2  t  2  ,

12



(2)



, a 2 t  n fb1, 2  

(3)

ACCEPTED MANUSCRIPT

1 TDL fb1,1  a1 t  1 , a  t  2  , 

2 TDL fb 2, 2  a 2  t  1 , a  t  2  , 





, a1 t  n fb1,1  



(4)



, a 2 t  n fb 2, 2  

(5)

where p  t  is the input vector at time t ; a1 is the output of layer 1 at time t; a 2 is the

CR IP T

output of layer 2 at time t; nin , n fb1,2 , n fb1,1 , and n fb2,2 define the lag space dimensions

ED

M

AN US

of inputs, outputs and internal feedback variables, respectively.

PT

Fig. 5. The proposed architecture for recurrent neural networks.

CE

There are two inputs for the diesel engine model in the proposed architecture. Therefore, the vector length of R is two. S 1 and S 2 are the number of neurons in layer

AC

1 and layer 2. n1  t  and n 2  t  are the input signal of the transfer function f 1 and f 2 , respectively. The equations for the computation of the net input n1  t  and n 2  t  are shown as follows:

13

ACCEPTED MANUSCRIPT

1 1,1 n  t   IW  p  t  ; p  t  1 ;

  a  t  TDL fb    b

1 a t  TDL fb1,1 

 LW1,2  a 2  t  1 ; 

2

where  p  t  ; p  t  1 ;



a 2  t  1 ; 

from

p  t  TDLin  is a vector which is built from input vector at

layer



(7)



1

delayed



in



a1 t  TDL fb1, 1  is a vector which

the

value

of

TDL fb1, 1

to

itself;

AN US

built



2 2 a t  TDL fb 2,2   b

time t and consecutive tapped delay time; a1  t  1 ;  is

1

1,2

2 2,1 1 2,2 2 n  t   LW a  t   LW a  t  1 ;

(6)

CR IP T

 LW1,1  a1  t  1 ; 

p  t  TDL in  

a 2 t  TDL fb1, 2  is a vector which is built from layer 2 delayed in the value

of TDL fb1, 2 to layer 1; a1  t  1 ; 





a1 t  TDL fb 2, 2  is a vector which is built from layer

M

2 delayed in the value of TDL fb1, 1 to itself; IW1, 1 is the input weight; LW1, 1 is the

ED

internal feedback weight in layer 1; LW1, 2 is the external feedback weight in layer 2.

PT

LW 2, 1 is the layer weight between layer 2 and layer 1; LW 2, 2 is the internal feedback

weight in layer 2; b1 and b 2 are the bias vectors for layer 1 and 2, respectively.

CE

Based on above equations, the outputs of layer 1 and 2 are calculated as shown in the

AC

following equation:



am t   f m nm t 



where f is activation function in layer m , m  1, 2 . 3.3. Diesel Engine Modeling and Learning Method

14

(8)

ACCEPTED MANUSCRIPT

Learned from the proposed RNN architecture, the diesel model implemented in the present work can be represented in the following form:

p  t   d  t  , d  t  1 ,

1,2

a



, a1,1 t  n f 1,1 |   





 t   a1,2  t  1|   ,

, a1,2 t  n f 1,2 |   

 t   a 2,2  t  1|   ,

, a 2,2 t  n f 2,2 |   

2,2





(10)

(11)

(12)

(13)

M

a



 t   a1,1  t  1|   ,

, l t  l TDL 

AN US

1,1

a

, d  t  d TDL  , l  t  , l  t  1 ,

(9)

CR IP T

 nh n f 1,1 n f 1,2  ni  1,1 1, 1 1,2 1, 2 1   LW i2, 1 f 1   IW 1,1 ij p j  t    LW ik a k  t    LW il a l  t  b i  j 1   i 1 k 1 l 1   y t |    f 2  2,2  nf  2 2, 2    LW 2,2  m a m t   b  m1 

where y , d , and l represent the estimated rotation speed, the rack displacement, and

ED

the size of power load, respectively; n h is the number of nodes in layer 1, n i the

PT

number of system input nodes, n f 1,1 the number of nodes in internal feedback of layer 1,

CE

n f 1,2 the number of nodes in external feedback of layer 2, n f 2,2 the number of nodes in

internal feedback of layer 2; LW i2, 1 represents the weight between the i th node in layer

AC

1 and the nodes in layer 2, LW 2,2 the weight between the m th node in internal m 1,1 1,2 feedback of layer 2 and the nodes in layer 2, while IW 1,1 ij , LW ik , and LW il are the

weights connecting the j th, k th, and l th input node and the i th node in layer 1, respectively; b 2 is the bias value between the node in layer 1 and layer 2, b1i the bias value between the i th node in input layer and layer 1. The weights and bias values are

15

ACCEPTED MANUSCRIPT

the components of the parameters vector; p j  t  is the j th element of the input vector (10), a1,k 1  t  the k th element of the internal feedback vector (11), a1,l 2  t  the l th external feedback vector (12), and a 2,m 2  t  the m th internal feedback vector (13). The

CR IP T

activation f 1    in layer 1 is the nonlinear Log-sigmoid function. In layer 2 the linear activation function f 2    has been considered. The parameters vector for all weights and biases is expressed as:

  [ LW 12,1, , LW 2,1 , IW 1,1 , IW 2,1 , LW 1,1 , IW 1,1 , LW 1,2 , IW 1,2 , 11 , 11 , 11 , LW ,

, IW

2,2

nm

1 1

,b ,

n hni ,b ,b ] n 1

nhnk

2

AN US

nh

2,2 1

h

n hnl

(14)

The parameters identification of the proposed diesel engine model is performed through a optimization process. Its purpose to minimize the learning error by adapting

M

the values of the network weights and biases. In this study, a second order method based

ED

on the Levenberg-Marquardt optimization algorithm has been considered [8]. To limit the occurrence of overfitting, a regularization term is added into the cost function shown

CE

PT

as follows [9]:

E   





2 1 N 1 y  t |    y  t      2 N t 1 2

(15)

AC

where y  t  is training data (experimental data),  represents the weight decay which is set to 0.01. The cost function E is computed over the time horizon N . The weights and biases are adjusted at each iteration of the optimization process (i.e., training epoch). At each epoch the RNN evaluates the complete set of inputs belonging to the training data. This procedure is known as batch learning. The learning process continues until some

16

ACCEPTED MANUSCRIPT

stopping criterion are reached. This study has three stopping criterions: the first one is reached if the current number of iterations surpasses maximum number 500; the second one is reached if the computation value of Eq. 15 has an error less than the minimum established 10 6 ; and last one is gradient value of the error function reaches the

CR IP T

minimum error 10 4 . Algorithm 1 describes these criterion. Algorithm 1. Stopping criteria

if the iteration I is greater than the maximum number 500 then Stop the training process.

AN US

else if E is less than 10 6 then Stop the training process else if E is less than 10 4 then Stop the training process else continue training

M

end if

ED

4. Results

PT

To test the performance of the proposed architecture, firstly, the evaluation methods are proposed. Secondly, the process of parameters selection for the diesel engine model

CE

is given in detail. Thirdly, using those optimized parameters, experimental results for simulating the speed changes under different power loads are reported. Lastly, to show

AC

the advantage of the proposed RNN architecture, this section presents comparison results between the proposed RNN model and three alternative models. 4.1. Performance Measurements There are many customary measures in the performance evaluation of RNN on time series data throughout the literature [26]. In accordance to suitable comparisons between

17

ACCEPTED MANUSCRIPT

the proposed method and those reported in the literature, we chosen the most commonly used metrics, which include the mean squared error (MSE), root mean squared error (RMSE) and normalized mean squared error (NMSE). The equations for three metrics are given in the following:

RMSE 

1 N

  y i  yˆ  N

i 1

1 N

2

(16)

i

  y i  yˆ  N

i 1

 

i

     2

AN US

 N   i 1 y i  yˆi NMSE   N   i 1 y i  y i 

2

CR IP T

MSE 

2

(17)

(18)

where y i , yˆ i , and y i are the observed data, predicted data and average of observed

M

data, respectively. N is the total number of data points.

ED

4.2. Parameters Selection for the Proposed Architecture

PT

In order to determine the best RNN architecture, we performed a set of experiments to test different architectures. In these experiments, we used the training and validation

CE

sets described in Section 2.2. Specially, the training data was used to train each chosen RNN architecture, and the validation data for the parameters selection, i.e., the number

AC

of hidden neurons and time delays, which aimed to select an appropriate RNN structure that could better meet the desired behavior of diesel generator and low computational cost.

Through these experiments we need to determine the number of hidden neurons ( H ), the number of time delays presented in the input time delay line ( i ), in the output time

18

ACCEPTED MANUSCRIPT

delay line ( j ) and in the internal time delay line ( k ). To achieve the objectives, these experiments were performed under different combinations with the number of hidden neurons and time delay lines. The variation range of the number of hidden neuron was set to [10, 40], number of time delays presented in the input time delay line to [1, 4],

CR IP T

number of time delays presented in the output time delay line to [1, 4], and number of time delays presented in the internal time delay line to [1, 4], respectively.

Based on these setting ranges, we carried on a series of experiments to search an optimal combination. Specifically, we first fixed the number of delay lines i  1 , j  1 ,

AN US

and k  1 , respectively. Then the number of hidden neurons increased from 10 to 40 with 1 interval. Using the metrics presented in Subsection 4.1, the best results could be achieved. Afterward, we varied the fixed number of delay lines to make different

M

combinations, and repeated above experimental process with increasing the number for hidden neurons.

ED

According to the experimental results, we observed that when the number of neurons

PT

increased, the output precision of the proposed model improved. However, when the number of hidden neurons was higher than 25, the metrics of the output did not increase

CE

rather little decrease. This fact showed that the RNN started to represent over-fitting of the data with more than 25 neurons. In addition, with the increased number of hidden

AC

neurons, the computation time was also considerably increased. Therefore, we selected a configuration of twenty-five hidden neurons which yielded the desired performance with low computational cost. For the input time delay line and output time delay line, we found that the performance of the proposed RNN reduced when using a small number of delay lines as 1 and 2. When we increased the number of delay lines to 3 and 4, the RNN’s

19

ACCEPTED MANUSCRIPT

performance was improved. However, after adding the number of delay lines more than 4, its performance was not improved. For the internal time delay line, we also found that when the number of delay lines was more than 2, the performance was not improved. Likewise, considering the computation cost, we selected to use three delays in the inputs

CR IP T

time delay line and two delays in the internal time delay line. Table 2 lists the performance of the selected RNN (i.e., H  25, i  3, j  3, k  2 ). Table 2 Selected parameters for the proposed architecture. i

j

k

MSE

RME

25

3

3

2

1.2647

1.1246

4.3. Experimental Results

AN US

H

NMSE 0.0390

M

Based on the chosen RNN architecture, this section gives the experimental results on the testing data from Section 2.2. Here, the training and validation data were not used.

ED

The experiments performed in this study focus on reproducing the dynamic process of the diesel engine model with different representative characteristics under various

PT

electrical power loads, i.e., 7.5kW, 15kW, 22.5kW, and 30kW. In order to show and

CE

illustrate the dynamic processes, each testing data was divided into four segments. Each segment contains the data of lasting 20s. Within the 20s period, the electrical power was

AC

first exerted on the diesel engine model at the beginning phase, and then discharged from the diesel engine model. Fig. 6 presents the predicted rotational speed of the diesel engine model produced by

the proposed RNN and measured (observed) data under 7.5kW, 15kW, 22.5kW, and 30kW loads. It can be seen in Fig. 6 that the predicted outputs of the model perform well and are in good agreement with the measured data. More specifically, we observe

20

ACCEPTED MANUSCRIPT

that the dynamic parts for loading and unloading conditions are perfectly reproduced, as shown in Fig. 6 (a), (c), (d). For the dynamic part when unloading electrical power in Fig.6 (b), it can be seen that the model performs a little differences between measured data and predicted data. But we still thought that its performance could satisfy with this

PT

ED

M

AN US

CR IP T

study’s requirement.

CE

Fig. 6. Predicted and measured rotational speed curves under (a) 7.5kW load, (b) 15kW

AC

load, (c) 22.5kW load, and (d) 30kW load. The data extracted from one set of the testing data.

To further evaluate the proposed diesel engine model, Table 3 reports the results

represented by three metrics for the testing data. It can be seen that the experimental result for 30kW load has the best for prediction error. In the development of electronic

21

ACCEPTED MANUSCRIPT

speed governor, generally, it is very important to test the performance under the full power load. Table 3 Performances of the proposed model under 7.5kW, 15kW, 22.5kW, and 30kW loads. MSE

RME

NMSE

7.5

1.2385

1.1129

0.0419

15

1.2967

1.1387

22.5

0.9479

0.9736

30

0.7327

0.8560

CR IP T

Power loads(kW)

0.0424 0.0288

AN US

0.0170

4.4. Comparisons with Other Methods

To better describe the advantages of the proposed RNN architecture, this section reports the performance comparisons with other methods including Polynomial NARX

M

[11], Multi Layer Perception Feed Forward Neural Network [27], and NRAX RNN

ED

[28]. These picking methods have all been successfully applied on time series prediction in different fields [29-32]. In this study, all methods are run on the same training,

PT

validation and testing data from Section 2.2. The same performance measurements was used to evaluate the prediction performance of the diesel engine model. The objective of

CE

this study is to reproduce the dynamic characteristics of diesel generators with the

AC

varying of the electrical power loads. Based on this purpose, Table 4 to Table 7 present the performance comparisons under 7.5kW, 15kW, 22.5kW, and 30kW loads, respectively. From those results, it can be seen in four tables that the performance of our proposed RNN in MSE, RME, and NMSE metrics achieves the best prediction performance comparing with other three methods. In particular, when the 22.5kW and 30kW loads exerted on this model, our method obtains 0.0288 and 0.0170 of NMSE.

22

ACCEPTED MANUSCRIPT

Table 4 Comparison results for four modeling methods under 7.5kW load. MSE

RME

NMSE

Polynomial NARX

2.7808

1.6676

0.1070

MLPFF

3.3126

1.8200

0.1544

NARX RNN

2.9035

1.7040

0.1140

Proposed RNN

1.2385

1.1129

CR IP T

Methods

0.0419

As for the polynomial NARX, this method has been successfully used in our previous research and achieved a good prediction performance [11]. In this study, it can be seen

AN US

in Tables 6 and 7 for 22.5kW and 30kW loads that the polynomial NRAX performs well on the testing data, comparing with NARX RNN and MLPFF. In particular, its result is the same as one obtained by our proposed RNN under 30kW load. As for the

M

MLPFF, it has the worst performance for prediction on the diesel engine model in all

ED

methods.

Table 5 Comparison results for four modeling methods under 15kW load. MSE

PT

Methods

RME

NMSE

1.9834

1.4083

0.0850

MLPFF

2.8861

1.6989

0.1100

2.1267

1.4583

0.0884

1.2967

1.1387

0.0424

CE

Polynomial NARX

AC

NARX RNN

Proposed RNN

Comparing with the MLPFF, it can be seen in these tables that the prediction

performance by the NARX RNN is better than the MLPFF. For this result, the NARX RNN is demonstrated to be well suited for modeling nonlinear system for time series tasks. More specifically, the NARX network is a recurrent dynamic network, with

23

ACCEPTED MANUSCRIPT

feedback connections form the output layer. As for our proposed RNN architecture, it can have copied of any neurons in the network from the previous time-step, not just the neuron in the last layer. This characteristic is equivalent to a fully connected RNN. It may allow the RNN to further influence the prediction of data at future iterations. Based

three methods on the time series tasks.

CR IP T

on these results, it can be concluded that the proposed method is competitive with other

Table 6 Comparison results for four modeling methods under 22.5kW load. MSE

RME

Polynomial NARX

1.1879

1.0899

MLPFF

1.8710

NARX RNN

1.4026

Proposed RNN

0.9479

AN US

Methods

NMSE 0.0350 0.0743

1.1843

0.0503

0.9736

0.0288

M

1.3678

Table 7 Comparison results for four modeling methods under 30kW load.

MLPFF

RME

NMSE

0.7327

0.8559

0.0170

1.4226

1.1927

0.0542

1.1238

1.0601

0.0310

0.7327

0.8560

0.0170

PT

Polynomial NARX

MSE

ED

Methods

CE

NARX RNN

AC

Proposed RNN

5. The HIL Simulation System for the DGS This section presents the Hardware-In-The-Loop simulation system with the

proposed diesel engine model. The framework of the HIL simulation system first is introduced. Then the loading and unloading experiments are run on the HIL simulation system for testing its performance.

24

ACCEPTED MANUSCRIPT

5.1. System Setup Fig. 7 shows the framework of the HIL simulation system. The DGS model (software) are run by the host PC which is connected to the physical objects of the HIL through the data acquisition card and signal conditioning board. The physical objects include an AC

CR IP T

servomotor system, two speed pickups, an electronic governor, an electronic actuator, etc. Table 8 lists the components which are used in the HIL system of the DGS. The physical objects are not simulated in the model, its purpose is to get a more accurate

AC

CE

PT

ED

M

AN US

representations and dynamic characteristics of a diesel generator.

Fig. 7. Diagram of the HIL simulation system of the DGS.

The AC servomotor system is employed to imitate the rotational speed of the diesel

engine. The signal of rotational speed captured by the speed pickup is sent into the electronic speed governor. Based on the setting speed reference, the electronic governor drives the actuator, which is used as an engine fuel control rack, through a linear motor

25

ACCEPTED MANUSCRIPT

mounted on the electronic actuator. The displacement of the actuator is measured by a DC LDVT sensor as the rack displacement. The rack displacement signal is conditioned through the signal conditioner and sent to host PC through the data acquisition card.

Devices

Model

Three-phase AC Servo Driver

GS0300A

Three-phase AC Servomotor

120MB300C

Electromechanical Actuator

ABD225G

CR IP T

Table 8 The components used in the HIL system. Manufacturer

Speed Governor

ESD-2200

GAC, USA

Speed Pickup

SZCB-01

Wuxi Longsheng Technology Co., LTS, China

Displacement Sensor

DC-SE 2000

York Instrument

DAQ

USB-6341

NI, USA

W540

Lenovo, China

SYNTRON

ED

M

AN US

GAC, USA

PT

Computer

SYNTRON

For the impacts of generator’s power loads on the rotational speed of the diesel

CE

engine, it is embodied by the input variable of the proposed model. The loading and

AC

unloading processes of the DGS can be realized by changing the value of power loads in the model. The rack displacement and simulated power loads are taken as two inputs of the model, and the rotational speed is able to be updated by the model at the speed of 10Hz. Then the computed speed signal is sent to the data acquisition card and converted to a voltage signal proportional to the output of the model by the signal conditioner. Once the servomotor system receives the rotational speed signal, it is controlled by the

26

ACCEPTED MANUSCRIPT

servo driver and runs at the same speed as the output of the model. The speed pick-up mounted on the shaft of the servomotor detects the rotational speed. The speed signal is transmitted to the electronic speed governor as the speed feedback. Lastly, the electronic speed governor can regulate the rack displacement to maintain the rotational speed of

CR IP T

servomotor as a constant. From Fig. 7, it can be seen that the HIL simulation of the DGS is a closed-loop system. Therefore, the overall system works as a real diesel generator so that the users are able to develop and debug electronic speed governor for the DGS.

AN US

5.2. Simulation Results

This section presents the simulation results run on the HIL system with the proposed diesel engine model. The performance of the HIL system is targeted on reproducing the

M

dynamic characteristics of the DGS when loading and unloading electrical loads. Therefore, we performed four sets of experimental procedure corresponding to 7.5kW,

ED

15kW, 22.5kW, and 30kW loads, i.e., four simulation epochs. To objectively evaluate

PT

the performance of the HIL system, we selected two metrics, which included the settling time (ST) and the transient regulation ratio (TRR). The equations for two metrics are

AC

CE

given in the following:

ST L  T LW  T LS

(19)

ST UL  T ULW  T ULS

(20)

TRR L 

S max  S r 100% Sr

(21)

TRRUL 

S min  S r 100% Sr

(22)

27

ACCEPTED MANUSCRIPT

where ST L and ST UL represent the settling time corresponding to power loads and unloads; T LW and T ULW represent the time spent how long the rotational speed close to the reference speed after loading and unloading power; T LS and T ULS are the time when power loads exerted on the DGS and power unloads discharged from the DGS;

CR IP T

TRR L and TRRUL are the transient regulation ratios corresponding to power loads and

unloads; S max and S min are the maximum and minimum rotational speed during each

AC

CE

PT

ED

M

AN US

simulation epoch; S r is the reference speed.

Fig. 8. Speed curves of the HIL simulation system under (a) 7.5kW load, (b) 15kW load, (c) 22.5kW load, and (d) 30kW load.

In this simulation, the electronic speed governor used in this experiment is ESD-2200, GAC. The reference speed was set into 1236 rpm. The software for this HIL system was programmed by LabVIEW 2011. The simulation result for each epoch is shown in Fig.

28

ACCEPTED MANUSCRIPT

8. The performance measures for the HIL system are required as follows: the settling time is less than 3 seconds and the transient regulation ratio of the speed is less than 5%. The results of the performance measures are reported in Table 9. From Fig. 8 and Table 9, it is worth nothing that the proposed diesel engine model is performed well on the

CR IP T

HIL simulation system and successfully reproduces the dynamic parts when loading and unloading electrical power.

Table 9 Performances of the diesel engine model run on the HIL simulation system. Loading Condition

Unloading Condition

ST L (s)

TRR L (%)

T ULS (s)

ST UL (s)

TRRUL (%)

7.5

2.2

1.3

1.71

10.9

1.3

1.76

15

2.3

1.2

2.18

10.8

1.1

2.08

22.5

2.4

1.7

2.90

8.8

1.6

2.77

30

2.3

1.7

13.5

1.9

3.37

AN US

T LS (s)

M

Power Loads(kW)

3.31

ED

Specially, in four simulation epochs, the most important index for the HIL system

PT

was in full loading and unloading conditions. Under the control of ESD-2200 speed governor, the full power load was applied to the HIL system at 2.3 second and removed

CE

at 13.5 second. The settling time of the system was the 1.7 second and 1.9 second for loading power and unloading power, respectively. The transient regulation ratios of

AC

loading power and unloading power were 3.31% and 3.37%, respectively. The HIL system are satisfied with the requirements of reproducing the dynamic characteristics of real DGS when loading power and unloading power. The achieved results are in line with the requirements for on-board application of the development and debug of the electronic speed governor for the DGS.

29

ACCEPTED MANUSCRIPT

7. Conclusions In this paper, a diesel engine model based on recurrent neural networks has been proposed to reproduce and simulate the dynamic characteristics with the changes of electrical power loads. In order to meet the requirements on both the real-time and the

CR IP T

accuracy, this study gives the schematic of parameters selection. The experimental results and performance comparisons are reported in detail, providing evidence that the proposed model is effective at reproducing the loading and unloading processes of the DGS. Using the proposed model, the simulation experiment is conducted on the

AN US

established hardware-in-the-loop simulation system. The simulation results provide a conclusion that it can well simulate the speed changes of the DGS with power loads and unloads. Therefore, the HIL simulation system can work as a real environment for developing and debugging the electronic speed governor for the DGS.

M

Furthermore, from the reported settling time and transient regulation ratio of the

ED

simulation system, we believe that the HIL simulation system with the proposed mathematical model for the diesel engine can provide enabling technology to

PT

applications in the future. Additionally, future works focus on capturing the data from other types of diesel generators. Its aim is to increase the types of the DGS models. The

CE

simulation software for the proposed mathematical model was completed in Matlab

AC

2013b, and the software for the HIL simulation system was programmed by LabVIEW 2011. All programs run on a laptop with 2.40GHz i7-4700MQ CPU and 8GB DDR3 RAM.

30

ACCEPTED MANUSCRIPT

Acknowledgments This work was supported by National Natural Science Foundation of China [grant number 81471743]; U.S. National Science Foundation (NSF) [grant number 0954579, 1333524]; and Zhejiang University State Key Laboratory Open Funding [grant number

AC

CE

PT

ED

M

AN US

CR IP T

GZKF-201512]. We thank all the participants who have participated in this work.

31

ACCEPTED MANUSCRIPT

References [1] F. Maroteaux, C. Saad, Diesel engine combustion modeling for hardware in the loop applications: Effects of ignition delay time model, Energy. 57(2013) 641-652. https://doi.org/10.1016/j.energy.2013.03.098.

CR IP T

[2] P. Navratil, J. Balate, Simulation of Control of Multi-Variable Control Loop: Stream Turbine, International Journal of Simulation Modelling. 10(2011) 38-45. https://doi.org/10.2507/IJSIMM10(2)1.167.

[3] A. Palianos, A. Achir, R. Stobart, N. Langlois, H. Chafouk, Dynamic feedback

AN US

linearization based control synthesis of the turbocharged diesel engine. In: American Control Conference, New York, USA, 2007, pp. 4407-4412. https://doi.org/10.2507/IJSIMM10(2)1.167.

[4] M. Jankovic, M. Jankovic, I. Kolamnovsky, Constructive Lynapunov Control

M

System Technology, IEEE Transactions on Control System Technology. 8(2000) 288-

ED

299. https://doi.org/10.1109/87.826800.

[5] M. Ayadi, N. Langlois, H. Chafouk, Polynomial control of nonlinear turbocharged

PT

diesel engine model, In: Proceedings of the IEEE International Conference on Industrial Technology, 2004, pp. 1384-1389. https://doi.org/10.1109/ICIT.2004.1490763.

CE

[6] W.C. Lan, T. Katagi, T. Hashimoto, Quasi steady state simulation of the diesel

AC

engine transient performance and design of mechatronics governor, Bulletin of marine engineering society in Japan. 24(1996) 1-13. [7] D.E. Winterbone, Dynamic model of TC engine for TF evaluation. SAE Paper, 770124.

32

ACCEPTED MANUSCRIPT

[8] B. Maass, R. Stobart, J.M. Deng, Diesel Engine Emissions Prediction Using Parallel Neural Networks, American Control Conference, Louis, Mo, USA, June 10-12, 2009, pp. 1122-1127. https://doi.org/10.1109/ACC.2009.5160119. [9] J. Leuchter, P. Bauer, V. Rercha, V. Hajek, Dynamic Behavior Modeling and

CR IP T

Verification of Advanced Electrical-Generator Set Concept, IEEE Transactions on Industrial Electronics. 56(2009) 266-279. https://doi.org/10.1109/TIE.2008.2009517. [10] B. Liu, Q.C. Lu, Z.H. Jin, S.F. Nie, System Identification of Locomotive Diesel

Engines with Autoregressive Neural Network, In: Proceeding of 4th IEEE Conference

AN US

on Industrial Electronics and Applications, 2009, pp. 3417-3421. https://doi.org/10.1109/ICIEA.2009.5138836.

[11] C.L. Jiang, X.Z. Wang, S.H. Zheng, M.X. Yu, On Hardware-In-The-Loop simulation system of diesel generator set, In: Proceedings of the 33rd Chinese Control

M

Conference, July 28-30, Nanjing, China, 2014, pp. 6229 – 6234.

ED

https://doi.org/10.1109/ChiCC.2014.6896011. [12] S.S. Ai, M. Bo, S.H. Zheng, M.X. Yu, The Hardware-In-The-Loop simulation

PT

system of the diesel generator set based on the NARMAX model, In: Proceedings of the 2nd International Conference on Compute Science and Electronics Engineering

CE

(ICCSEE), Beijing, China, 2013, pp. 1080-1083.

AC

https://doi.org/10.2991/iccsee.2013.271. [13] S.A. Billings and S. Chen, The identification of linear and non-linear models of a turbocharged automotive diesel engine, Mechanical Systems and Signal Processing. 3(1989) 123-142. https://doi.org/10.1016/0888-3270(89)90012-5. [14] I. Arsie, C. Pianese, M. Sorrentino, A procedure to enhance identification of recurrent neural networks for simulating air-fuel ratio dynamics in SI engines,

33

ACCEPTED MANUSCRIPT

Engineering Applications of Artificial Intelligence. 19(2006) 65-77. https://doi.org/10.1016/j.engappai.2005.06.003. [15] I. Arsie, A. Cricchio, M.D. Cesare, F. Lazzarini, C. Pianese, M. Sorrentino, Neural network models for virtual sensing of NOx emissions in automotive diesel engines with

https://doi.org/10.1016/j.conengprac.2017.01.005.

CR IP T

least square-based adaption, Control Engineering Practice. 61(2017) 11-20.

[16] G. Kushwaha, S. Saraswati, Air Path Identification of Turbocharged Diesel Engine using RNN, International Conference on Industrial Instrumentation and Control (ICIC),

AN US

College of Engineering Pune, India, May 28-30, 2015. https://doi.org/10.1109/IIC.2015.7150954.

[17] V.M. Janakiraman, X.L. Nguyen, D. Assanis, Nonlinear identification of a gasoline HCCI engine using neural networks coupled with principle component analysis,

M

Applied Soft Computing. 13(2013) 2375-2389.

ED

https://doi.org/10.1016/j.asoc.2013.01.006.

[18] M. Hafner, M. Schüler, O. Nelles, R. Isermann, Fast Neural Networks for Diesel

PT

Engine Control Design, Control Engineering Practice. 8(2000) 1211-1221.

CE

https://doi.org/10.1016/S0967-0661(00)00057-5. [19] M. Ouladsine, G. Bloch, X. Dovifaaz, Neural Modelling and Control of the diesel

AC

engine Pollution Constrains, Journal of Intelligent and Robotic Systems. 4(2004) 157171. https://doi.org/10.1007/s10846-005-3806-y · Source: DBLP. [20] J. Koutník, K. Greff, F. Gomez, J. Schmidhuber, A Clockwork RNN, Proceedings of the 31st International Conference on Machine Learning, Beijing, China, 2014, pp. 1863-1871.

34

ACCEPTED MANUSCRIPT

[21] E. Diaconescu, The use of NARX Neural Networks to predict Chaotic Time Series, WSEAS Transactions on Computer Research. 3(3) (2008) 182-191. [22] C.H. Wu, J.W. Wang, J.T. Liu, W.Y. Liu, Recurrent neural network based recommendation for time heterogeneous feedback, Knowledge-Based Systems. 109(1)

CR IP T

(2016) 90 – 103. https://doi.org/10.1016/j.knosys.2016.06.028. [23] K. Raza, M. Alam, Recurrent neural network based hybrid model for

reconstructing gene regulatory network, Computational Biology and Chemistry. 64(2016) 232-334. https://doi.org/10.1016/j.compbiolchem.2016.08.002.

AN US

[24] S.K. Biswas, N. Sinha, B. Purkayastha, L. Marbaniang, Weather prediction by

recurrent neural network dynamics, International Journal of Intelligent Engineering Informatics. 2(3) (2014) 166-180. http://dx.doi.org/10.1504/IJIEI.2014.066208. [25] T. Lin, B.G. Home, P. Tiňo, L. Giles, Learning Long-Term Dependencies in

M

NARX Recurrent Neural Networks, IEEE Transactions on Neural Networks. 7(1996)

ED

1329 – 1338. http://dx.doi.org/10.1109/72.548162. [26] C. Smith, Y.C. J, Evolutionary multi-objective generation of recurrent neural

PT

network ensembles for time series prediction, Neurocomputing. 143(2014) 301-311.

CE

https://doi.org/10.1016/j.neucom.2014.05.062. [27] S. Agarwalla, K.K. Sarma, Machine learning based sample extraction for automatic

AC

speech recognition using dialectal Assamese speech, Neural Networks. 78(2016) 97-111. https://doi.org/10.1016/j.neunet.2015.12.010. [28] C.A. Duchanoy, M.A. Moreno-Armendáriz, L. Urbina, C.A. Cruz-Villar, H. Calvo, J.J. Rubio, A novel recurrent neural network soft sensor via a differential evolution training algorithm for the tire contact patch, Neurocomputing. 235(2017) 71-82. https://doi.org/10.1016/j.neucom.2016.12.060.

35

ACCEPTED MANUSCRIPT

[29] D. Brezak, T. Bacel, D. Majetic, J, Kasac, B. Novakovic, A comparison of feedforward and recurrent neural network in time series forecasting, in: 2012 IEEE Conference on Computational Intelligence for Financial Engineering Economics (CIFEr), 2012, pp. 1-6. https://doi.org/10.1109/CIFEr.2012.6327793.

CR IP T

[30] A.A. Ferreira, T.B. Ludermir, R.R. de Aquino, Comparing recurrent networks for time-series forecasting, in: The 2012 International Joint Conference on Neural

Networks (IJCNN), 2012, pp. 1-8. https://doi.org/10.1109/IJCNN.2012.6252459. [31] M. Nørgaddrd, O. Ravn, N.L. Poulsen, L.K. Hansen, Neural Networks for

AN US

Modelling and Control of Dynamic Systems. Springer, Berlin, 2002. 0-48.

[32] S. Li, S. Chen, B. Liu, Y. Li, Y. Liang, Decentralized kinematic control of a class of collaborative redundant manipulators via recurrent neural networks, Neurocomputing. 91(2012) 1-10. https://doi.org/10.1016/j.neucom.2012.01.034.

M

[33] N. Clark, A. Tehranian, R. Jarrett, R. Nine, Translation of Distance-Specific

ED

Emissions Rates between Different Heavy Duty Vehicle Chassis Test Schedules, SAE Technical Paper. 1(2002) 1754. https://doi.org/10.4271/2002-01-1754

PT

[34] P. Shanmugam, V. Sivakumar, A. Murugesan, M. Ilangkumaran, Performance and Exhaust Emissions of a Diesel Engine Using Hybrid Fuel with an Artificial Neural

CE

Network, Energy Sources, Part A: Recovery, Utilization, and Environment Effects.

AC

33(2011) 1440-1450. http://dx.doi.org/10.1080/15567036.2010.539085. [35] G.R. Kannan, Artificial neural network approach to investigate the effect of injection pressure and timing on diesel engine fuelled with diestrol, International Journal of Oil, Gas, and Coal Technology. 11(2016). https://doi.org/10.1504/IJOGCT.2016.074286.

36

ACCEPTED MANUSCRIPT

[36] J. Mohammadhassani, Sh. Khalilarya, M. Solimanpur, A. Dadvand, Prediction of NOx Emissions from a Direct Injection Diesel Engine Using Artificial Neural Network. 2012(2012). http://dx.doi.org/10.1155/2012/830365 [37] J.Y. Zhu, Marine Diesel Engine Condition Monitoring by Use of BP Neural

Scientists, Hong Kong, China, 2009, pp. 18-20.

CR IP T

Network, Proceedings of the International MultiConference of Engineers and Computer

[38] S.H.H.F. Dhkordi, Simulation of small diesel engine vibration using artificial

neural network, International Journal of Agriculture and Crop Sciences. 5(2013) 2084-

AN US

2090.

[39] C.W. Mohd Noor, R. Mamat, G. Najafi, M.H. Mat Yasin, C.K. Ihsan, M.M. Noor, Prediction of marine diesel engine performance by using artificial neural network model, Journal of Mechanical Engineering and Sciences. 10(2016) 2289-4659.

M

http://dx.doi.org/10.15282/jmes.10.1.2016.15.0183.

ED

[40] L. Brzonzowska, K. Brzozowski, J. Nowakoski, An Application of Artificial Neural Network to Diesel Engine Modelling, IEEE Conference: Intelligent Data

PT

Acquisition and Advanced Computing Systems: Technology and Applications, Sofia, Bulgaria, 2007, pp. 18-20. https://doi.org/10.1109/IDAACS.2005.282958.

CE

[41] C.G. Yang, J. Luo, Y.P. Pan, Z. Liu, C.Y. Su, Personalized Variable Gain Control

AC

With Tremor Attenuation for Robot Teleoperation, IEEE Transactions on Systems, Man, and Cybernetics: Systems. PP(2017) 1-12. http://dx.doi.org/10.1109/TSMC.2017.2694020. [42] C.G. Yang, X.J. Wang, Z.J. Li, Y.N. Li, C.Y. Su, Teleoperation Control Based on Combination of Wave Variable and Neural Networks. IEEE Transactions on Systems,

37

ACCEPTED MANUSCRIPT

Man, and Cybernetics: Systems. 47(2017) 2125-2136. http://dx.doi.org/10.1109/TSMC.2016.2615061. [43] C.G. Yang, Y.M. Jiang, Z.J. Li, W. He, C.Y. Su, Neural Control of Bimanual Robots With Guaranteed Global Stability and Motion Precision, IEEE Transactions on

AC

CE

PT

ED

M

AN US

http://dx.doi.org/10.1109/TII.2016.2612646.

CR IP T

Industrial Informatics. 13(2017) 1162-1171.

38

ACCEPTED MANUSCRIPT

Figure captions Fig. 1. The structure diagram of the DGS. Fig. 2. The block diagram of experimental system for data acquisition. Fig. 3. An example set of experimental data.

CR IP T

Fig. 4. The architecture of NARX RNN. Fig. 5. The proposed architecture for recurrent neural networks.

Fig. 6. Predicted and measured rotational speed curves under (a) 7.5kW load, (b) 15kW load, (c) 22.5kW load, and (d) 30kW load. The data extracted from one set of the testing

AN US

data. Fig. 7. Diagram of the HIL simulation system of the DGS.

Fig. 8. Speed curves of the HIL simulation system under (a) 7.5kW load, (b) 15kW load,

M

(c) 22.5kW load, and (d) 30kW load.

ED

Authors’ Biography and Pictures Mingxin Yu obtained his B.S. degree in Computer Science &

PT

Technology from North University of China, in 2006, the M.S.

AC

CE

degree in Electronic Engineering, in 2010, and Ph.D. degree in Control Science & Engineering, in 2015, from Beijing Institute of Technology, China. In 2016, he was a visiting research scientist at Northeastern University, Boston, USA. Currently he is a postdoctoral researcher working at Beijing Institute of

Technology, China. His research interests include neural networks, pattern recognitions, image processing, and human-computer interaction.

39

ACCEPTED MANUSCRIPT

Xiaoying Tang obtained the Ph.D. degree in Electronic Engineering from Beijing Institute of Technology, China. She is a Professor with School of life Sciences, Beijing Institute of Technology, China, where she directs Biomedical Engineering Laboratory. Her research has been funded by the Natural Science Foundation of China, National 863 program, and industries. Her

CR IP T

area of expertise includes: neural networks, biomedical image and signal processing, fMRI brain functional imaging, and pattern recognition.

Yingzi Lin obtained the Ph.D. degree in mechanical engineering

AN US

from University of Saskatchewan, Saskatoon, SK, Canada, and another Ph.D. degree in vehicle engineering from China Agricultural University, Beijing, China.

She is an Assistant Professor with Department of Mechanical and Industrial Engineering, Northeastern University, Boston,

M

MA, where she directs the Intelligent Human-Machine Systems Laboratory. Her research has been funded by the National Science Foundation, Natural

ED

Sciences and Engineering Research Council of Canada, and industries. Her area of expertise includes: driver-vehicle systems, human-centered intelligent machine systems,

PT

and human machine interface design.

Dr. Lin is the Chair of the Virtual Environments Technical Group of the Human

CE

Factors and Ergonomics Society (HFES). She is on the committees of the

AC

Transportation Research Board (TRB) of the National Academy of Sciences.

Xiangzhou Wang obtained the Ph.D. degree in department of automatic control from Beijing Institute of Technology, China, in 1996. During 1996 to 1998 he was a researcher at the postdoctoral center of Tsinghua University. He currently is a professor with the School of Automation in Beijing Institute of Technology, where he directs the Measurement Technology and Automatic Equipment Laboratory. His area of expertise include

40

ACCEPTED MANUSCRIPT

measurement technology, intelligent control, intelligent instrument, diesel generator

AC

CE

PT

ED

M

AN US

CR IP T

control, and hydraulic drive & control.

41