International Conference of Electrical, Automation and Mechanical Engineering (EAME 2015)

Study on Diesel Engine Status Analysis and Fault

Diagnosis Based on SAE J1939 Protocol
L.H. Cao, J.N. Li, X.L. Liu, J.L. Yu, Z.M. Wu, F.G. Chen
Control Engineering Key Laboratory
Chongqing Communication Institute
Chongqing,China

Abstract—SAE J1939 is one of the most widely used application II. SAE J1939 PROTOCOL
layer protocols in automotive industry based on CAN bus. SAE
Open System Interconnection Reference Model (OSI/RM)
J1939 defines vehicle application layer to make us read the status
is a networking model proposed by International Organization
of the vehicle easily. We propose an improved RBF neural
network fault classification algorithm to classify diesel engine for Standardization (IOS). OSI protocols stacks are split into
fault. Also we analyse the state of the vehicle from vehicle CAN seven layers. CAN protocol only defines physical layer and
bus and use MATLAB to write a program so that we can analyse data link layer. SAE J1939 protocol is based on CAN2.0B, and
the CAN bus data. According to CAN data obtained from a diesel defined network layer and application layer in addition.
engine, we got diesel engine fault data and use they to train A. Data Link Layer
PSO-RBF neural network, the resulting effect is better than PSO
neural network and we can use it to do a diesel engine fault Data link layer provides reliable data transmission between
diagnosis. physical connections, includes sending synchronization,
sequence control, error control and flow control which is
Keywords-SAE J1939; diesel engine; fault diagnosis; PSO-RBF necessary for CAN data frame [4]. In the CAN standard, MAC
neural network ID has two standards: CAN2.0A, which is 11-bit, and
I. INTRODUCTION CAN2.0B, which is 29-bit. The structure difference is mainly
in arbitration field and control field [5]. J1939 uses CAN2.0B
With the development of the automobile industry and communication mechanism. The physical layer follows
electronic control technology, getting data of vehicle fault is electrical characteristics, communication media, physical
no longer only rely on traditional method. Traditional methods interface of CAN protocol. In data link layer, J1939 redefines
such as extracting eigenvalues of the vibration signal, 29-bit ID of arbitration field and 64-bit data of data field
hydraulic waveforms to diagnosis engine fault. Due to it is corresponding to vehicle information [6].
analyzed by one data source, it has a high probability of
misdiagnosis. Therefore, vehicle fault diagnosis based on a Protocol data unit consists of seven parts: priority,
variety of data sources has become a research focus. extended data page, data page, PDU format, PDU specific
domain, source address and data fields. PDU is encapsulated
SAE J1939 developed by the American Society of in one or more CAN data frames to transmit to the other
Automotive Engineers, is the vehicle bus recommended network devices through physical medium. Each CAN data
practice used for communication and diagnostics among frame has only one PDU. The PDU format is as Table 1.
vehicle components, originally by the car and heavy-duty
truck industry in the United States [1]. SAE J1939 defines four TABLE I. PDU FORMAT OF SAE J1939 PROTOCOL
layers in the seven-layer OSI network model: physical layer,
data link layer, network layer, vehicle application layer, and PDU
Dat
Extende PDU specifi Source
the physical layer is based on CAN controller [2]. SAE Nam Priorit
d data
a
forma c addres
Data
J1939-71 developed for the application layer is to make the e y pag fields
page t domai s
e
data provided by source node can be used by the other nodes n
[3]. Abbr DAT
P EDP DP PF PS SA
. A
Engine fault diagnosis is also largely dependent on human bit 3 1 1 8 8 8 64
experience and traditional diagnostic methods currently, fault
diagnosis based on SAE J1939 protocol is still in the B. Application Layer
development stage. We import the data which obtained in the Application layer of J1939 is stipulated by J1939-71. It
operation of the vehicle into MATLAB. Analysis in MATLAB provides a way to access the OSI environment and describes
is effective that we can analyze the state of vehicle and packet format, data range, transmission repetition rate,
diagnose whether the vehicle has a potential fault. parameters naming and ISO Latin character through PGN
(Parameter Group Number) and SPN (Special Parameter
Number).

© 2015. The authors - Published by Atlantis Press 263

 A set of parameters called "parameter group" (PG), PGN is Import data
its unique identification number. PGN serves as an index in
the communication process. It shows the contents and purpose
of the message, determines packet type and number of frames. Change data

Here is an example, PGN 61444: Default Priority is 011,

Extended Data Page is 0, Data Page is 0, PDU Format is Get SPN in PGN
11110000, PDU Specific is 00000100 and Data fields as
follow: Next SPN
Next
Get data of SPN
TABLE II. SPNS OF PGN 61444
Start
Length Parameter Name SPN Yes
Position Is length 32? Adjust data
1.1 4bit Engine Torque Mode 899
Driver's Demand Engine - No
2 1byte 512
Percent Torque
Actual Engine - Percent Parsing for SPN
3 1byte 513
Torque
4-5 2byte Engine Speed 190 Write into
Source Address of SPN{}.OutM()
6 1byte Controlling Device for 1483
FIGURE I. FLOW DIAGRAM OF J1939 DATA CONVERSION
Engine Control
7.1 4bit Engine Starter Mode 1675
Engine Demand – Percent
8 1byte 2432
Torque
Parameters of PGN 61444 represent electronic engine
controller 1, which composed of seven specific parameters.
For example in SPN190, SPN 190: data length is 4 bits,
resolution is 16 states/4 bit and 0 offset, data range is 0 to 15,
operational range is also 0 to 15, type is Measured. If we got
the data of SPN 190 is 201B, it shows that the engine speed is
868rpm. The other data can be got as the same way. (1) SPN(190)

We can get a variety of vehicle status data based on SAE

J1939 protocol. It is convenient for us to have a
comprehensive understanding of the performance of vehicle.
But the data is too scattered and inconvenient to understand,
we need an analytical tool to analyze the data.
III. DATA ANALYSIS OF SAE J1939-71 IN MATLAB
The most important in J1939 protocol is PGN and SPN. (2) SPN(110)
We create two new cell type: PGN{}, SPN{}. PGN{} include:
Name, Transmission Repetition Rate, Data Length, Extended FIGURE II. SPN OUTPUT
Data Page, Data Page, PDUF or mat, PDU Specific, Default
After the analysis of the parameters we need, we can have
Priority, SPNNUM, SPN{}. SPN{} include: Name, Data
an overall grasp of the vehicle's condition. When we analyze
Length, Resolution, Data Range, Type.
engine failure, we can choose targeted fault parameters
After acquiring CAN data, we import the data into forming fault parameter table, then use different fault
MATLAB as cell type. Flow diagram of J1939 data diagnosis algorithm to do engine fault diagnosis.
conversion is as Figure 1.
IV. DIESEL ENGINE FAULT DIAGNOSIS BASED ON
SPN{}.Out M() is the output of the SPN at the moment. IMPROVED RBF NEURAL NETWORK
We get Diesel Exhaust Fluid injection unit input pipe A. RBF Neural Network
failure data of Cummins 6B diesel engine, imported into
MATLAB for analysis. Take SPN(190) and SPN(110) for RBF neural network is two forward lays network with a
example, the output shown in Figure 2. It shows the situation single hidden layer. The input layer consists of source node.
within this period of time the engine speed and the engine The number of hidden layer units depend on the fixed problem.
coolant temperature. Transformation from the input space to the hidden layer space
is nonlinear, and from the hidden layer to the output layer
space transform is linear. Hidden layer node consist of
functions like Gaussian function. Output layer node is
typically a simple linear function. The topology structure
shown in Figure 3.

264
 Each particle update their speed and position.
If it does not met the termination condition, then
return (2), otherwise, stop.
B. Engine Fault Diagnosis Based on Improved RBF Neural
Network
We got 84 sets of data of seven kinds of faults which
are got from CAN bus of Cummins 6B diesel engine in a
FIGURE III. THE TOPOLOGY STRUCTURE DIAGRAM OF RBF truck( A:Diesel Exhaust Fluid injection input unit has
NEURAL NETWORK
problem, B: fuel rail pressure of injector is below the
normal operating range, C:Diesel Exhaust Fluid reservoir
Transformation function for hidden node is radial basis
level is below the normal operating range, D:power
function. It takes Gaussian function normally:
losing when the ignition switch is turned on, E:engine
control module calibration memory damages, F:speed is
|| x − c j ||
Φ j = exp[ − ] unstable, intermittent or error which based on the number
2δ j2 j = 1,2,⋯, q of revolutions of the wheel, G:sensor of nitrogen oxides
(1)
exporting updates rate abnormally). After MATLAB
analysis, some of the data as follows:
i -th output of the network is:
TABLE III. FAULTS DATA OF DIESEL ENGINE
q
yi = ∑ ω ji Φ ( x, c j , δ j ) No. X1 X2 X3 X4 X5 X6 X7
Fault
j =i i = 1,2,⋯, m (2)
Code
1 1079 23 249.3 74 868 117 130 1
2 85 0 0.5 30.3 191 30 30 2
x is cj
the input vector, is center of the hidden layer 3 915 0 0 77 0 0 68 3
δ j ω j ,i 4 723 0 0 17.2 0 0 11 4
unit, is width, is the connection weights.
cj δ ω j ,i 5 906 0 173 76.9 701 100 100 5
Because the value of , , has a great influence j

6 1125 0 338 58.1 861 135 114 6

on the learning and training effectiveness, for how to
cj δ ω j ,i 7 1275 16 503 45.5 1684 15 61 7
determine , j , , we use PSO to optimize these … … … … … … … … …
three parameters.
84 1312 15 506 45.8 1669 14 60 7
Particle swarm optimization (PSO) algorithm deemed X1 is measured fuel rail pressure, X2 is speed, X3 is oil
individuals to particles in N-dimensional search space. pressure, X4 is coolant temperature, X5 is rotational speed, X6
Each particle follow current best individual and change is SCR outlet temperature and X7 is SCR inlet temperature.
their speed and location through sharing information and
find the optimal solution through continuous iteration Select 56 groups as the training sample, and the rest as a
eventually. The algorithm first randomly generated test sample. Use training samples to train PSO-RBF neural
particle swarm, each particle has a speed and position in network, the 7 input vector are X1 to X7. There are 7 kinds of
the N-dimensional solution space. Speed and position are faults, replaced by code 1 to 7. One hidden layer and
adjusting by their own condition and the experience of containing 25 hidden layer units. PSO parameters are set as
other particles. Finally, reaching the maximum number of follows: scale is 500, inertia weight factor is 0.9, learning
iterations or PSO searched the optimal fitness value [7]. factor c1 = c2 = 2 , termination condition is that error is less than
0.01 or reaching 100 generations. PSO-RBF and RBF are used
PSO-RBF algorithm is as follows: for training. The convergence curve lines is shown in Figure 4.
cj δ ω j ,i
Each particle consist of , j
, , random initial
its speed and position.
1 m yi′ − yi
f= ∑| |
N i=1 yi
Set as a fitness function to determine
p b es t
the best individual particles and global
gbest y′i yi
optimization . is the actual value and is output (1) PSO-RBF (2) RBF
value. FIGURE IV. THE CONVERGENCE CURVE LINES OF PSO-RBF AND
RBF ALGORITHMS
Compare the fitness value of each particle with best
f ( Pi ) > f ( pbest )
individual particles and global optimum. If , Convergence rate of PSO-RBF is faster. We use the trained
pbest = Pi f ( Pi ) > f ( g best ) gbest = Pi network to classify the remaining 28 groups, the results are as
then . If , then .
follows. It shows that PSO-RBF has higher accuracy and

265
shorter training time.

TABLE IV. THE DIAGNOSTIC RESULTS AND TRAINING TIME

PSO-RBF RBF
Accuracy 89.3% 78%
Time 1.636 s 5.785 s

V. SUMMARY
Rely on powerful data processing capability and
convenient language compiler environment of MATLAB, data
processing base on J1939 has a unique position. We are able to
process and analyze a variety of data to have comprehensive
and accurate understanding of information on the vehicle. This
can be an efficient analysis tool for vehicle detection and
repairing. After analyze the data, we use PSO-RBF neural
network algorithm to diagnose vehicle fault with a high
accuracy. The algorithm can be improved in the future to make
the accuracy to 100% and a shorter time.
ACKNOWLEDGEMENTS
Supported by the Outstanding Achievements
Transformation Funded Major Projects of Chongqing College
(Grant No. KJZH14112).
REFERENCES
[1] SAE J1939 Standards Collection, Recommended Practice for a Serial
Control and Communication Vehicle Network[S]. Society of
Automotive Engineers, 2003.
[2] Wu Jianfang. Research of Embedded Vehicle Information Terminal
Implementation [D]. Harbin: Harbin Engineering University, 2010.
[3] SAEJ1939-71 Vehicle Application Layer[S]. USA. SAE. 2006.
[4] Hu Yajuan. Design and Implementation of Vehicle Information Display
System Based on J1939 Protocol [D]. Chengdu: Southwest Jiaotong
University, 2008.
[5] Zeng Hongan. Embedded Linux C Programming [M]. Beijing:
Publishing House of Electronic Industry.
[6] Zhang Qianqin. Research on the Control System about Central Signals
in Electric Vehicle [D]. Beijing: Beijing Jiaotong University, 2010.
[7] Kennedy J, Eberhart R.C., Particle swarm optimization [J]. Proceedings
of IEEE International Conference on Neural Networks. 1995.
1942-1948.

266