b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/issn/15375110

Research Paper

Prediction of seed distribution in rectangular

vibrating tray using grey model and artificial
neural network

Zhan Zhao a,b,*, Mingzhi Jin a, Chunjie Tian a, Simon X. Yang b,**
a
Institute of Agriculture Engineering, Jiangsu University, Zhenjiang, 212013, China
b
School of Engineering, University of Guelph, Guelph, N1G 2W1, Canada

article info
To maintain good continuous working performance in a vacuum plate seeder, it is impor-
Article history: tant to monitor the distribution of seeds in real time and automatically adjust vibration
Received 27 June 2018 parameters accordingly. Seed motion in a rectangular vibrating tray with vibration varying
Received in revised form with time and interference by direction angle was simulated using the discrete element
30 August 2018 method (DEM). A plane model P was used to describe the variation of seed layer thickness.
Accepted 19 September 2018 Four square areas on the bottom of the tray were divided symmetrically near the four
corners to measure seed layer thickness, and a monitoring plane model Pm was established.
DEM simulation results showed that the models Pm and P had the similar change rules,
Keywords: although there were some differences in fitting parameters. There was obvious time delay in
Seed distribution the change of Pm compared with P. Therefore, a grey system model (GM) was adopted to
Prediction method predict the change of Pm, and two back-propagation (BP) neural networks which take GM
Vibrating tray prediction results as input parameters were developed respectively. Then, according to the
Discrete element method BP neural networks outputs, a prediction plane model Pp was proposed to predict the seed
Grey model distribution. Experiments were carried out on a test-rig to validate these predictions. The
Neural network seed distribution plane P was measured manually, the monitoring plane Pm was established
using seed layer thickness measurement results and the prediction plane Pp was established
using the GM and BP neural networks. The results indicated that the proposed method had
good precision and stability and provides the basis for the design of an automatic control
system for the vibrating tray to promote a uniform seed distribution.
© 2018 IAgrE. Published by Elsevier Ltd. All rights reserved.

the main cultivation technique since it produces strong

1. Introduction seedlings and extends the growth period for more than half a
month (He, Luo, Li, Wang, & Zhang, 2008). At present, more
Rice is the most important staple food for more than half of than 50% by area of rice cultivation in China is utilises hybrid
the global population. Nursery growing and transplanting is rice. Precision seeding has gradually become a major planting

* Corresponding author. Institute of Agriculture Engineering, Jiangsu University, Zhenjiang, 212013, China. Fax: þ86 (0)511 88797338.
** Corresponding author. Fax: þ1 519 8244120 (ext. 52437).
E-mail addresses: zhaozhan@ujs.edu.cn (Z. Zhao), syang@uoguelph.ca (S.X. Yang).
https://doi.org/10.1016/j.biosystemseng.2018.09.017
1537-5110/© 2018 IAgrE. Published by Elsevier Ltd. All rights reserved.
 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5 195

its precision seeding agronomic requirements, and damaged

Nomenclature seeds are a serious problem. Pneumatic metering technology
has been widely developed for precision seeding because it does
a Development coefficient of GM model
little damage to the seeds and has broad spectrum of applica-
A Vibration amplitude of seeds tray, mm
tions (Movahedi, Rrzvani, & Hemmat, 2016; Topakci, Karayel,
b Driving coefficient of GM model
Canakci, Furat, & Uzun, 2011; Yazgi & Degirmencioglu, 2014).
f Vibration frequency of seeds tray, Hz
Structure design, parameters optimisation and performance
n Sample number
experiments of different pneumatic seeders have been
h0 Average seed layer thickness, mm
researched extensively (Abdolahzare, Asoodar, Kazemi,
h1, h2, h3, h4 Seed layer thickness in four monitoring
Rahnama, & Mehdizadeh, 2016; Altikat, Celik, & Gozubuyuk,
areas, mm
€
2013; Fanigliulo & Pochi, 2011). Karayel, Barut, and Ozmerzi
hIJ Seed layer thickness in square regions, mm
(2004) determined the optimum vacuum pressure of a preci-
hj Output of neuron j in hidden layer
sion seeder and developed mathematical models by using some
L Number of neurons in hidden layer
physical properties of seeds such as mass, density, sphericity
nm Normal vector of monitoring plane
and projected area. Gaikwad and Sirohi (2008) designed a vac-
np Normal vector of seed distribution plane
uum seeder for sowing vegetable seeds in plug tray and opti-
nv Vibration direction vector of the tray
mised the operation parameters such as vacuum pressure and
N Number of input neurons
nozzle diameter. Onal,€ Degirmencioglu, and Yazgi (2012)
Ok Outputs of neuron k in output layer
established a theoretical model of seeds sowing uniformity
p Prediction step
performance. Ding, Liao, and Huang (2013) analysed the rela-
P Seed distribution plane model
tionship between seeding migration trajectory and working
Pm Monitoring plane model
pressure of a pneumatic precision seeder and Xing et al. (2017)
Pp Prediction plane model
developed a pneumatic seeder to meet the requirements for the
t Time, s
field seeding of hybrid rice.
Ts Vibration termination moment
As a precision seeding device for rice nursery seedlings, the
xI, yJ Coordinates of square regions
vacuum plate seeder is primarily composed of a rectangular
xm, ym Projection coordinates of nm in XY plane
suction plate and a seed tray (Liu & Song, 2004). The tray needs
xmþp, ymþp Prediction results of xm and ym
to vibrate at high frequency and low amplitude so that the
xp, yp Projection coordinates of np in XY plane
seeds in the tray are separated to reduce interactive forces and
xBP, yBP BP output of xp and yp
allow for easy and precise seed pick up. The slider-crank
X(0) Original numerical sequences
mechanism is the most commonly used device to drive the
X(1) First order AGO data
tray and reciprocate it along a guiding axis.
am Angle between projection of nm and X axis, 
The necessary condition for improving the continuous
amþp Prediction results of am, 
working performance of a vacuum plate seeder is that the
ap Angle between projection of np and X axis, 
seeds should be uniformly distributed in the vibrating tray,
av Angle between projection of nv and X axis, 
because this can maintain a uniform distance between the
aBP Angle calculated using xBP and yBP, 
suction nozzles in the vacuum plate and the vibrating seeds.
bm Angle between projection of nm and Y axis, 
However, under practical conditions, it is inevitable that there
bp Angle between projection of np and Y axis, 
is an uncertain angle between the vibration direction and the
bv Angle between projection of nv and Y axis, 
vertical direction, and that this angle usually varies with time.
gm Angle between nm and Z axis, 
This causes the seeds to flow in the tray resulting in an uneven
gmþp Prediction results of gm, 
distribution of seeds with time. In this case, it is hard to
gp Angle between np and Z axis, 
maintain the vertical distance between the suction nozzles in
gv Angle between nv and Z axis, 
different regions of the suction plate and the seeds and this
gt Time varying interference angle, 
can seriously restrict working performance (Gong, Chen, Li, &
gBP BP output of gp, 
Li, 2014). A feasible solution is to adjust the vibration direction
h Learning rate
of the tray to promote the motion of seeds orderly and uni-
h0 Initial learning rate
formly. This requires the real time prediction of seed distri-
vij Connection weighs between neuron i in input
bution. The prediction of material distribution in systems
layer and neuron j in hidden layer
under vibration is also widely used in agricultural cleaning
vjk Connection weighs between neuron j in hidden
and screening systems (Craessaerts, Saeys, Missotten, & De
layer and neuron k in output layer
Baerdemaeker, 2010).
The discrete element method (DEM) is an important tech-
nique for research into the granular dynamics of agricultural
technique to improve the yield because hybrid rice, which systems (Boac, Ambrose, Casada, Maghirang, & Maier, 2014;
uses the seeding technique, has a greater ability to tiller than Kruggel-Emden, Rickelt, Wirtz, & Scherer, 2008; Tijskens,
ordinary rice (Yang, Li, Ma, Sun, & Xu, 2014; Yi, Liu, Wang, & Ramon, & De Baerdemaeker, 2003). It can provide the motion
Tao, 2014). parameters of each granular unit during the whole calculation
Rice seed has irregular shape and has a frangible glume on process. In an earlier paper the influence of mechanical
the surface. Existing mechanical metering devices cannot meet properties on the impact behaviour of rice seeds excited by a
196 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5
reciprocating vibration was analysed using DEM, and a
monitoring method for seed mass in a small area was pro-
posed (Zhao, Wu, Yin, & Tang, 2015). In this study seed motion
in a rectangular vibrating tray is simulated, under vibration
with interference in direction angle. The objective is to
establish a seed distribution model according to seed distri-
bution state, and to propose a corresponding monitoring
method. Then, according to the monitoring results, an
improved prediction method which combines grey system
and artificial network was developed and the prediction
model optimised based on the error analysis and validated by
experiment to evaluate its prediction accuracy.
2. Materials
In this section, under the excitation of vibrating tray, the
motion of seeds with different layer thickness was simulated
using DEM. A plane model was used to describe the distribu-
tion of seeds in the tray. According to the central coordinates
and the corresponding seed layer thickness of all regions,
plane fitting models of seed distribution were established
using multiple regression method. The variation of seed dis-
tribution under the vibration with time varying interference in
direction angle was discussed.
2.1. DEM simulations
In the DEM, Newton's second law of motion and Euler's dy-
namics equation are usually used to describe the translational
and rotational movements of particulate assemblies. Inter-
action force models are established in the framework of
classical mechanics (Horabik & Molenda, 2016). In order to
represent the various shapes of agricultural seeds, a multi-
sphere approach method was adopted, in which a certain
number of spheres were “glued” together (Radvilaite, Ramı́res-
Go mez, & Kac ianauskas, 2016; Wang, Yu, Lv, Wang, & Fu,
2017). A commercial three-dimensional DEM code (EDEM®
2.5, EDEM, Edinburgh, UK) was used. According to the physical
properties of rice seeds, a triaxial ellipsoidal particle model
was established using multiple spheres. Its semi-axes were
2.85, 1.55 and 1.35 mm, and the mass of single seed was
26.5  103 g. Contact forces were calculated using the Hertz-
Mindlin model. Table 1 summarises the values of the material
properties required in the DEM.
Table 1 e Values of material properties used in the DEM.
Parameter Rice seed
Semi-axes (mm) 2.85  1.55  1.35
Mass (g) 26.5  103
Density (kg m3) 1080
Young's modulus (MPa) 375
Poisson's ratio 0.25
Coefficient of friction seed-seed 0.48
seed-tray / 0.32
Coefficient of restitution seed-seed 0.42
seed-tray / 0.48
Time step (s) 1  106
According to the structure of the vacuum plate seeder, the
size of rectangular tray model was 700 mm  420 mm and the
material was aluminium alloy # 7075. Taking the centre of the
tray as the origin, an inertial coordinate X0 ¼ [X Y Z] is estab-
lished, shown in Fig. 1. In ideal conditions, the tray is vertically
vibrated along the Z axis. But, some uncertain interference
factors such as the precision of installation and the horizontal
posture of the frame are inevitable in practical operations.
This usually leads to a small time varying change in vibration
direction. Here, the angle between the vibration direction
vector of the tray nv and the Z axis is defined as gv, and the
angles between the projection of nv and the X, Y axes are
defined as av, bv.
For rice nursery seeding operations, many experimental
results have indicated that the optimum vibration frequency f
and amplitude A of the tray are near 11 Hz and 4 mm, and the
average seed layer thickness h0 in the tray should be main-
tained in a range of 15e25 mm (Chen, Gong, Li, Li, & Xu, 2015).
Therefore, DEM simulations are carried out with f of 11 Hz and
A of 4 mm, h0 is selected as 15, 20 and 25 mm respectively. The
simulation process was divided into two stages. In the initial
10 s, the tray was sinusoidal vibrated in vertical direction.
Then, a continuous random time varying interference angle gt
was applied to the vibration direction nv from 10 to 70 s.
Variations of the vibration direction in DEM simulations are
given in Table 2.
2.2. Seed distribution model
The DEM simulation of seeds motion in the tray is shown in
Fig. 2. In order to analyse the seed distribution state, the tray
was divided into 10  6 small square regions.
Under the excitation of vibration, the rice seeds in the tray
are always in discrete motion state. Thus, the relationship
between number of seeds and the stable seed layer thickness
when the tray is static was firstly established. The stable seed
layer thickness in each region was then calculated by count-
ing the number of seeds. From DEM simulation of parameter #
2 (see Table 2), the calculated seed layer thickness in all re-
gions at time t of 10, 30 and 50 s is shown in Fig. 3. When the
tray was vibrated vertically, the seeds were always in vibra-
tional motion at their original locations and seed flow in the
tray could be ignored. As shown in Fig. 3 (a), seed distribution
in the tray can be well fitted using a horizontal plane model.
When the tray is vibrated with interference direction angle gt,
seeds will flow within the tray. The distribution presented
Tray
/
/
2800
72,000
0.33
/
/
Fig. 1 e Schematic diagram of seeds tray model.
 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5 197

Table 2 e Variations of the vibration direction in DEM simulations.

Time (s) Parameter# 1 (h0 ¼ 15 mm) Parameter# 2 (h0 ¼ 20 mm) Parameter# 3 (h0 ¼ 25 mm)
[av, bv, gv] ( ) [av, bv, gv] ( ) [av, bv, gv] ( )
0e10 [90, 90, 0] [90, 90, 0] [90, 90, 0]
10e15 [33.7, 56.3, 1.51] [56.3, 33.7, 1.50] [45.0, 45.0, 2.00]
15e20 [146.4, 56.3, 1.50] [64.6, 154.6, 1.94] [89.1, 0.9, 0.73]
20e25 [83.8, 6.2, 0.52] [19.4, 70.6,1.66] [47.7, 42.3, 2.49]
25e30 [71.6, 161.4, 1.21] [89.0, 179.0, 1.75] [39.2, 129.2, 1.55]
30e35 [1.2, 91.2, 0.33] [12.7, 77.3, 2.26] [88.9, 1.1, 1.00]
35e40 [81.1, 8.9, 2.48] [32.4, 57.6, 1.65] [33.2, 56.8, 1.13]
40e45 [11.9, 78.1, 2.30] [55.8, 34.2, 1.81] [70.2, 19.8, 1.37]
45e50 [11.9, 78.1, 0.20] [89.0, 1.0, 1.00] [169.6, 79.6, 1.69]
50e55 [67.5, 157.5, 1.02] [62.8, 152.8, 1.52] [89.7, 0.3, 0.97]
55e60 [0.33, 89.7, 0.26] [38.6, 128.6, 0.89] [89.9, 0.1, 2.47]
60e65 [0.06, 89.9, 1.46] [90.0, 0.0, 0.06] [89.8, 179.8, 1.52]
65e70 [176.7, 86.7, 0.83] [119.2, 29.2, 1.11] [119.2, 29.2, 0.06]

obvious irregular and time varying characteristics which were

Fig. 2 e Snapshot of DEM simulation.
comprehensively affected by the vibration parameters, seed
average layer thickness and physical properties. It is difficult
to establish an accurate uniform fitting equation. However, by
analysing the seed layer thickness in all regions at different
times, it is found that the distribution still can be approxi-
mately fitted using a plane model, although there are some
errors particularly in marginal regions of the tray, as shown in
Fig. 3 (b) and (c). According to the central coordinates of all
regions [xI, yJ] and the corresponding calculated seed layer
thickness hIJ, I ¼ 1, 2, …, 10 and J ¼ 1, 2, …, 6, plane fitting
models of seed distribution were established using multiple
regression method. The blue lines in Fig. 3 represent the fitting
plane obtained. The multiple regression results using the DEM

Fig. 3 e Fitting results of seeds distribution using a plane model: (a) t ¼ 10 s, (b) t ¼ 30 s, (c) t ¼ 50 s.
198 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5
simulation data showed that the root mean squared error
(RMSE) was less than 2.5 mm and the coefficient of determi-
nation was >0.9.
The seed distribution plane model P can be expressed by
P : Ax þ By þ Cz þ D ¼ 0
where the normal vector of the plane np ¼ [A, B, C], the angle
between the np and the Z axis is defined as gp, and the angles
between the projection of np and the X, Y axes are defined as
ap and bp. The fitting plane model P can be determined by gp
and ap because there is ap þ bp ¼ 90 . DEM simulation data was
sampled with frequency of 1 Hz. The seed distribution plane
models at different times were established according to vari-
ations of the seed layer thicknesses. The calculated variations
of gp and ap at different times are given in Fig. 4.
3. Methods
Four square areas on the base of the tray were divided to
measure the seed layer thickness. According to the mea-
surement results, a monitoring plane model Pm of seed dis-
tribution was established. By analysing the difference
between the measured seed distribution plane P and the
Fig. 4 e Comparation of the actual distribution plane P and the monitoring plane Pm: (a) Parameter # 1, (b) Parameter # 2,
(c) Parameter # 3.
monitoring plane Pm, an improved prediction method incor-
porating a grey system model and back-propagation (BP)
neural network is proposed.
3.1. Monitoring method
(1)
A seed layer thickness monitoring method and device for
measuring seeds impact force in a small area has been
developed (Zhao et al., 2015). Under vibration excitation, seeds
impact with the tray and generate a periodic force signal. In
order to obtain the stable impact force, a second-order But-
terworth low pass filtering calculation was used. The linear
relationship between the seed layer thickness and the stable
impact force was established.
In practice, it is impossible to monitor the seed layer
thickness in many regions. For the plane distribution model,
four 60 mm  60 mm square areas on the bottom of the tray
were delineated, shown in Fig. 1. The distances between the
centres of each area and the corresponding vertex in X-axis
and Y-axis directions were 120 and 95 mm. With the simula-
tion parameters given in Table 2 and the critical frequency of
low pass filter of 1 Hz, the received variations of seed layer
thickness h1, h2, h3 and h4 in four monitoring areas are shown
in Fig. 5. The measured seed distribution basically followed
 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5
Fig. 5 e Variations of seeds layer thickness in four monitoring areas: (a) Parameter # 1, (b) Parameter # 2, (c) Parameter # 3.
the plane model. Therefore, a plane model was used to fit the
seed distribution. According to the central coordinates of four
areas [xm1, ym1], [xm2, ym2], [xm3, ym3] and [xm4, ym4], a seed
distribution monitoring plane model Pm was established using
multiple regression, which was expressed as
Pm : Am x þ Bm y þ Cm z þ Dm ¼ 0
The normal vector of the plane is nm ¼ [Am, Bm, Cm]. The
angle between the nm and the Z axis is defined as gm, and the
angles between the projection of nm and the X, Y axes are
defined as am and bm. The multiple regression results showed
that the root mean squared error (RMSE) was generally less
than 2 mm. The variations of gm and am at different times are
given in Fig. 4.
The variation of measured seed distribution plane P was in
real time because it was established by counting the number
of seeds. The monitoring plane Pm was established by
measuring the seed layer thickness, and it was inevitable that
the signal processing caused some time delay. So, it can be
seen from Fig. 4 that the variations of gm and am both have
obvious time delays relative to gp and ap, but the delay period
was generally <3 s. The variation of gm basically reflected the
change rule of gp although there were some differences in the
value. Smaller gp indicated that the seed distribution was
uniform. With the increase of gp, it was shown that ap and am
have stable and consistent change rules.
199
3.2. Grey prediction model
To improve the real time performance of the measurement,
predicting the change of monitoring plane Pm was required.
With signal processing this can be calculated accurately.
However, seed motion under vibration excitation with inter-
(2)
ference from direction angle is complex and almost impos-
sible to establish using a defined mathematical model.
Therefore, the variation of seed distribution can be regarded
as a grey system.
Grey theory demonstrates the optimal ability of perform-
ing fitting predictions using small data sets and limited in-
formation (Bezuglov & Comert, 2016; Kayacan, Ulutas, &
Kaynak, 2010; Yin, 2013). It employs a data generation
method, rather than a statistical regulation, to obtain a more
regular generation sequence from the initial data. Grey pre-
dictions have been utilised in agriculture, industry, economics
and to study natural phenomenon. Although various types of
grey models have been developed, most of the research are
still focused on the first order with one variable grey model
GM(1,1) in prediction applications because the GM(1,1) has
good computational efficiency and also is the foundation of
other grey prediction models.
Assuming the obtained time data sequence with n samples
(time point) and named as original numerical sequences X(0) is
given by
200 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5

3.3. Neural network model


Xð0Þ ¼ xð0Þ ð1Þ; xð0Þ ð2Þ; …; xð0Þ ðnÞ (3)
A neural network was adopted to establish the relationship
The accumulated generating operation (AGO) technique is between GM prediction results and the measured distribution
then applied to reduce the randomisation of the raw data. plane model. As a typical artificial neural network, BP neural
These processed data become monotonic increase sequence networks can implement any complex nonlinear mapping
which complies with the solution of first order linear ordinary functions and approximate arbitrary nonlinear functions with
differential equation. Therefore, the solution curve would fit satisfactory accuracy (Atici, 2011). The network is trained by a
to the raw data with high precision. The first order AGO data supervised learning algorithm. During the learning process,
X(1)can thus be described as the errors are subsequently backward propagated to adjust

 the weights and thresholds of the connections between two
Xð1Þ ¼ xð1Þ ð1Þ; xð1Þ ð2Þ; …; xð1Þ ðnÞ (4)
layers to improve the accuracy of the output layer. A three-
where layer BP neural network is capable of approximating any
practical functions as long as the hidden layer has sufficient
X
k
neurons and the transfer function is nonlinear.
Xð1Þ ðkÞ ¼ xð0Þ ðiÞ; k ¼ 1; 2; …; n (5)
i¼1 Here, a two three-layer BP neural network including input
layer, hidden layer, and output layer was established. Due to
The first order grey differential equation is given as
ap and am are discontinuous at ±180 , the GM prediction re-
dxð1Þ sults of the projection coordinate of normal vector nm in XY
þ axð1Þ ¼ b (6)
dt plane xm, ym, …, xmþ3 and ymþ3 were normalised on the in-
terval [0 1], and then used as the input of the BP neural
The development coefficient a and the driving coefficient b
network, shown in Fig. 7(a). The normalised projection co-
can be determined from the data sequence using the least-
ordinates xp, yp of the normal vector np in XY plane were used
square technique. Its matrix form can be written as
to train the network. The sigmoid function gðxÞ ¼ 1=ð1 þ ex Þ
 1
T
½a; b ¼ BT B BT Y (7) was selected as the activation function of the hidden neurons
and the output of neuron j, denoted as hj, was calculated by
where
!
2   3 2 3 X
N

0:5 xð1Þ ð2Þ þ xð1Þ ð1Þ 1 xð0Þ ð2Þ hj ¼ g vji xi ; i ¼ 0; 1; ,,,; N; j ¼ 1; 2; ,,,; L (10)
6 0:5 xð1Þ ð3Þ þ xð1Þ ð2Þ 17 6 ð0Þ 7
7; Y ¼ 6 x ð3Þ 7
i¼0
B¼6
4
 /  /5 4 / 5
where vij is the connection weighs between neuron i in input
0:5 xð1Þ ðnÞ þ xð1Þ ðn  1Þ 1 xð0Þ ðnÞ
layer and the neuron j in hidden layer, xi is the ith input var-
By solving Eq. (6), the prediction formula can be presented iable, N is the number of input neurons and L is the number of
as neurons in hidden layer. The activation function of the output
 neurons is a linear function and the outputs Ok are given by
b ðnþp1Þ b
bð1Þ
x m ðn þ pÞ ¼
bð1Þ
x m ð0Þ  e þ (8)
a a X
l
Ok ¼ vjk hj ; k ¼ 1; 2 (11)
Hence, the desired prediction output at n step and p pre- j¼0

diction step can be calculated as

where O1 is the output xBP and O2 is the output yBP. Taking xp
 and yp as the desired output, the error function is defined as
b
bð0Þ
x m ðn þ pÞ ¼ bð0Þ
x m ð1Þ  ð1  ea Þeaðnþp1Þ (9)
a  2 2
E ¼ e2 =2 ¼ xp  xBP þ yp  yBP =2 (12)
The data gm is continuous and nonnegative which can be
used directly as original data for the GM prediction, but the The LMS algorithm was used to train the network to
data am is discontinuous at ±180 . So, the projection coor- compute the weights vji and vj that minimise over all error.
dinate [xm, ym] of the normal vector nm in XY plane is The changes of the connection weights are calculated as
adopted. After been normalised on the interval [0 1], they are  
vE
used as the input data of GM prediction models respectively. Dvji ¼ h ¼ hhj 1  hj xi vj e; i ¼ 0; 1; ,,,; N; j ¼ 1; 2; ,,,; L
vvji
Then, the prediction of am can be calculated according to the
(13)
GM prediction results of xm and ym. The rolling model is then
used to improve prediction accuracy. That is, when a new
vE
data is input and the oldest data should be updated simul- Dvkj ¼ h ¼ hhj e; j ¼ 0; 1; ,,,; L; k ¼ 1; 2 (14)
vvkj
taneously, and the GM model is rebuilt up. When the sample
number n is 5, the prediction step p is 1, 2 and 3 s, the pre- where, h is the learning rate.
diction results of gm and am are shown in Fig. 6. The Another single output BP model was established which is
subscript þ1, þ2 and þ3 indicates that the prediction step p shown in Fig. 7(b). The data gm, gmþ1, gmþ2 and gmþ3 were used
is 1, 2 and 3. It can be thus be seen that the GM can predict as the model inputs, and the error function was defined as
the variation tendency of gm and am in general, and the 2
prediction error will gradually increase with increasing pre- E ¼ gp  gBP =2 (15)
diction step.
 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5 201

Fig. 6 e GM prediction results of gm and am: (a) Parameter # 1, (b) Parameter # 2, (c) Parameter # 3.

It is important to determine the number of neurons in
hidden layer L for achieving the desired precision and good
generalisation performance. Based on the recommended
calculation method, an L value of 15 was determined by
repeated trials because the accuracy did not improve while
increasing the number of hidden neurons (Zhang, Yang,
Mittal, & Yi, 2002). The initial learning rate h0 was 0.005. To
avoid oscillation or divergence and improve training speed,
the learning rate h was select as an exponential decay func-
tion hðtÞ ¼ h0 expt=t .

Fig. 7 e Three-layer BP neural network: (a) prediction of ap, (b) prediction of gp.
202 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5

Fig. 8 e BP model prediction results of gp and ap: (a) Parameter # 1, (b) Parameter # 2, (c) Parameter # 3.

Finally, using the output xBP, yBP and gBP, a prediction plane
model Pp was established to describe the actual seed distri-
bution plane model P, which can be expressed by
Pp : singBP cosaBP x þ singBP sinaBP y þ cosgBP z þ DP ¼ 0 (16)
according to the error analysis with different BP model input
parameter combinations. Finally, experiments are carried out
on the test-rig to verify the feasibility and accuracy of the
proposed prediction method.
4.1. Prediction results

4. Results and discussion From three DEM simulations 70% data and the corresponding
GM prediction results were selected for training and the
In this section, the prediction results are calculated using the
remaining 30% data was used for testing the BP models. In
DEM simulation data. Then, the prediction model is optimised
order to analyse the influence of GM outputs on the prediction

Table 3 e Error analysis of BP model prediction results.

Number Input parameters Error ( )
gp < 1 1  gp < 2 2  gp < 3 3  gp Average
Data a#0 xm, ym 41.3 19.6 8.9 4.5 27.7
Data a#1 xm, ym, xmþ1, ymþ1 34.6 13.8 5.6 3.6 21.6
Data a#2 xm, ym, …, xmþ2, ymþ2 22.8 9.3 3.2 3.1 14.4
Data a#3 xm, ym, …, xmþ3, ymþ3 22.3 8.7 3.0 3.0 13.9
Data g#0 gm 0.35 0.31 0.57 0.68 0.366
Data g#1 gm, gm þ1, 0.25 0.25 0.36 0.41 0.266
Data g#2 gm, gm þ1, gm þ2 0.22 0.24 0.28 0.21 0.242
Data g#3 gm, gm þ1, gm þ2, gm þ3 0.22 0.23 0.26 0.18 0.237
 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5 203

precision, four different parameter combinations were

selected as input for the BP models, which are shown in Table
3. The outputs of aBP and gBP are shown in Fig. 8.
It can be seen that the prediction accuracy was signifi-
cantly improved by integrating the GM and BP neural network.
The GM output with prediction step p ¼ 3 had little effect on
the improving of prediction accuracy because the GM predic-
tion error increases with the increments of p. Therefore, the
ideal BP model input parameters could be chosen as the GM
output with p ¼ 0, 1 and 2. When the seeds average layer
thickness h0 was small, the randomness of seeds motion was
strong. This will increase the fluctuation of both the measured
seed distribution plane P and monitoring plane Pm, and result
in increasing of randomness in the error of the prediction
plane Pm. When the gp is small the seed distribution in the tray
is uniform and the difference between aBP and ap relatively
large. However, this has little effect on the prediction of seed
distribution. With increasing gp, the unevenness of seed dis-
tribution increased, and the difference between aBP and ap
rapidly reduced. When gp increases from 1 to 2 and 3 , the
corresponding differences gradually decreased from 22.8 to
9.3 and 3.2 , and the difference between gBP and gp could be
controlled <0.3 . Therefore, the proposed method can well
predict seed distribution state, which can thus provide a basis
for automatically adjusting the vibration parameters of the
tray to improve uniform seed distribution. The detailed error
analysis is given in Table 3.
The data of Parameters # 1 and 3 was used to train the BP
model, and the data of Parameters # 2 used for testing. The GM
output with p ¼ 0, 1 and 2 were used as BP model input. The
results indicate that the difference between gBP and gp was
also less than 0.3 , and the difference between aBP and ap were
less than 25 , 10 and 3.5 when the gp increased from 1 to 2
and 3 . This showed that the proposed method has good
generalisation performance.
4.2. Experimental results
The vacuum plate seeder used in experiments is shown in
Fig. 9. The size of rectangular seeds tray was
700 mm  420 mm, and multiple nozzles were installed on the
panel of suction plate. Driven by a motor and slider-crank
mechanism, the tray reciprocated along a guiding axis. By
adjusting a set of parallel mechanisms, the vibration direction
could be rotated along X and Y axes within the angle of ±10 .
Four sets of seeds mass monitoring devices were installed on
Fig. 10 e Output voltage of four monitoring devices.
the tray. The monitoring areas were 60 mm  60 mm, and the
installation positions were the same as shown in Fig. 1. The
seed layer thickness measurement error was <2.5% within the
range of 0e40 mm. Output voltages were recorded using a
programmable logic controller (PLC) system.
During the experiments, a specified mass of rice seeds
were weighed and laid evenly in the tray. The tray was then
vibrated with frequency f of 11 Hz and amplitude A of 4 mm.
The vibration direction nv was manually adjusted to cause the
seeds to flow across the tray. The received output voltages of
four monitoring devices are shown in Fig. 10.
It is impossible to measure the actual change process of
seed distribution state. So, during the experiments, the vi-
bration of the tray was rapidly stopped at a specified time Ts.
The tray was divided to 5  3 square regions to weigh the seeds
mass in each region. The measured seed distribution model P
at the moment of vibration termination was accordingly
established. After calibration using the linear relationship
between the output voltage and seed layer thickness, the
monitoring plane model Pm was established using h1, h2, h3
and h4 at time Ts. Then, the GM model was established using
h1, h2, h3 and h4 at time Ts, Ts1, …, Ts4. The GM outputs with
prediction steps p of 0, 1 and 2 were input into the trained BP
neural network to calculate gBP and aBP for establishing the
prediction plane model Pp. The comparison of models P, Pm
and Pp are shown in Fig. 11. By using GM and BP neural

Fig. 9 e Vacuum plate seeder using rectangular vibrating

tray. Fig. 11 e Comparison of prediction results.
204 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5
network, the error of prediction model Pp was effectively
reduced.
With the average seed layer thickness h0 of 15, 20 and
25 mm, experiments were repeated for 5 times for each
parameter. When the gp was greater than 1 , the prediction
results indicated that the difference between ap and aBP was
less than 7.5 and the difference between gp and gBP was less
than 0.25 .
5. Conclusions
DEM simulation results have shown that the seed distribution
in a vibrating tray can be fitted using a plane model P. By
monitoring the seed layer thickness in four square areas
located on the bottom of the tray, a monitoring plane model
Pm of seed distribution was established. Influenced by the
signal processing of seeds thickness measurement, the
change of monitoring plane Pm was obviously time delayed
with respect to the measured seed distribution plane P.
The comprehensive application of GM and BP was pro-
posed to predict the angle gp between the normal vector np of
plane model P and the Z axis, and the angle ap between the
projection of np and the X axis respectively.
Results for seed distribution prediction using DEM simu-
lation showed that optimal precision was received while using
the GM outputs with prediction step of 0, 1 and 2 as the BP
neural networks input parameters. When gp increased from 1
to 2 and 3 , the corresponding difference between ap and aBP
gradually decreased from 22.8 to 9.3 and 3.2 , and the dif-
ference between gp and gBP could be controlled below 0.3 .
Repeated experiments were carried out on vacuum plate
seeder test-rig. When gp was >1 , the obtained difference be-
tween ap and aBP was less than 7.5 and the difference be-
tween gp and gBP was less than 0.25 .
The paper thus provides an effective way for prediction of
seed distribution in a rectangular vibrating tray. This is
important for automatic adjusting the vibration parameters to
promote the uniform distribution of seeds in the tray, and to
improve the continuous working performance of a vacuum
plate seeder.
Acknowledgements
This work is supported by the National Natural Science
Foundation of China (No. 51775246), the Science Foundation of
Anhui Province (No. 1808085QE172) and a Project Funded by
the Priority Academic Program Development of Jiangsu Higher
Education Institutions (PADP).
references
Abdolahzare, Z., Asoodar, M. A., Kazemi, N., Rahnama, M., &
Mehdizadeh, S. A. (2016). Optimization of the most important
operational parameters of a pneumatic seeder using real-time
monitoring for cucumber and watermelon seeds. Journal of
Agricultural Machinery, 6(1), 35e48.
Altikat, S., Celik, A., & Gozubuyuk, Z. (2013). Effects of various no-
till seeders and stubble conditions on sowing performance
and seed emergence of common vetch. Soil and Tillage
Research, 126, 72e77.
Atici, U. (2011). Prediction of the strength of mineral admixture
concrete using multivariable regression analysis and an
artificial neural network. Expert Systems with Applications, 38,
9609e9618.
Bezuglov, A., & Comert, G. (2016). Short-term freeway traffic
parameter prediction: Application of grey system theory
models. Expert Systems with Applications, 62, 284e292.
Boac, J. M., Ambrose, R. P. K., Casada, M. E., Maghirang, R. G., &
Maier, D. E. (2014). Applications of discrete element method in
modeling of grain postharvest operations. Food Engineering
Reviews, 6(4), 128e149.
Chen, J., Gong, Z. Q., Li, Y. M., Li, J. H., & Xu, Y. (2015).
Experimental study on nursing seedlings of super rice
precision seeder device. Transactions of the Chinese Society for
Agricultural Machinery, 46(1), 73e78.
Craessaerts, G., Saeys, W., Missotten, B., & De Baerdemaeker, J.
(2010). Identification of the cleaning process on combine
harvesters, Part II: A fuzzy model for prediction of the sieve
losses. Biosystems Engineering, 106(2), 97e102.
Ding, Y. C., Liao, Q. X., & Huang, H. D. (2013). The seeding
migration trajectory extraction and analysis of pneumatic
precision metering device for rapeseed. Journal of Food
Agriculture and Environment, 11(1), 477e482.
Fanigliulo, R., & Pochi, D. (2011). Air-flow distribution efficiency of
a precision drill used in the sowing of different graded seeds.
Journal of Agricultural Science and Technology A, B1, 655e662.
Gaikwad, B. B., & Sirohi, N. P. S. (2008). Design of a low-cost
pneumatic seeder for nursery plug trays. Biosystems
Engineering, 99(10), 322e329.
Gong, Z. Q., Chen, J., Li, Y. M., & Li, J. H. (2014). Seed force in
airflow field of vacuum tray precision seeder device during
suction process of seeds. Transactions of the Chinese Society for
Agricultural Machinery, 45(6), 92e97.
He, R. Y., Luo, H. Y., Li, Y. T., Wang, X. H., & Zhang, L. (2008).
Comparison and analysis of different rice planting methods in
China. Transactions of the Chinese Society of Agricultural
Engineering, 24(1), 167e171.
Horabik, J., & Molenda, M. (2016). Parameters and contact models
for DEM simulations of agricultural granular materials: A
review. Biosystems Engineering, 147, 206e225.
€
Karayel, D., Barut, Z. B., & Ozmerzi, A. (2004). Mathematical
modelling of vacuum pressure on a precision seeder.
Biosystems Engineering, 87(4), 437e444.
Kayacan, E., Ulutas, B., & Kaynak, O. (2010). Grey system theory-
based models in time series prediction. Expert Systems with
Applications, 37(2), 1784e1789.
Kruggel-Emden, H., Rickelt, S., Wirtz, S., & Scherer, V. (2008). A
study on the validity of the multi-sphere discrete element
method. Powder Technology, 188(2), 153e165.
Liu, C. L., & Song, J. N. (2004). Influence of seed tray vibration on
work performance of suction-vibration type precision seed
device. Journal of China Agricultural University, 9(2), 12e14.
Movahedi, E., Rrzvani, M., & Hemmat, A. (2016). Design,
development and evaluation of a pneumatic seeder for
automatic planting of seeds in cellular trays. Journal of
Agricultural Machinery, 4(1), 65e72.
€
Onal, I., Degirmencioglu, A., & Yazgi, A. (2012). An evaluation of
seed spacing accuracy of a vacuum type precision metering
unit based on theoretical considerations and experiments.
Turkish Journal of Agriculture and Forestry, 36, 133e144.
Radvilaite, U., Ramı́res-Go  & Kac
 mez, A., ianauskas, R. (2016).
Determining the shape of agricultural materials using
spherical harmonics. Computers and Electronics in Agriculture,
128, 160e171.
 b i o s y s t e m s e n g i n e e r i n g 1 7 5 ( 2 0 1 8 ) 1 9 4 e2 0 5
Tijskens, E., Ramon, H., & De Baerdemaeker, J. (2003). Discrete
element modeling for process simulation in agriculture.
Journal of Sound and Vibration, 266, 493e514.
Topakci, M., Karayel, D., Canakci, M., Furat, S., & Uzun, B. (2011).
Sesame hill dropping performance of a vacuum seeder for
different tillage practices. Applied Engineering in Agriculture,
27(2), 203e209.
Wang, X. M., Yu, J. Q., Lv, F. Y., Wang, Y., & Fu, H. (2017). A multi-
sphere based modelling method for maize grain assemblies.
Applied Power Technology, 28, 584e595.
Xing, H., Wang, Z. M., Luo, X. W., Cao, X. M., Liu, C. B., & Zang, Y.
(2017). General structure design and field experiment of
pneumatic rice direct-seeder. International Journal of
Agricultural and Biological Engineering, 10(9), 31e42.
Yang, Z. Y., Li, N., Ma, J., Sun, Y. J., & Xu, H. (2014). High-yielding
traits of heavy panicle varieties under triangle planting
geometry: A new plant spatial configuration for hybrid rice in
China. Field Crops Research, 168, 135e147.
205
Yazgi, A., & Degirmencioglu, A. (2014). Measurement of seed
spacing uniformity performance of a precision metering unit
as function of the number of holes on vacuum plate.
Measurement, 56(6), 128e135.
Yi, S. J., Liu, Y. F., Wang, C., & Tao, G. X. (2014). Experimental study
on the performance of bowl-tray rice precision seeder.
International Journal of Agricultural and Biological Engineering,
7(1), 17e25.
Yin, M. S. (2013). Fifteen years of grey system theory research: A
historical review and bibliometric analysis. Expert Systems with
Applications, 40, 2767e2775.
Zhang, Q., Yang, S. X., Mittal, G. S., & Yi, S. J. (2002). Prediction of
performance indices and optimal parameters of rough rice drying
with neural networks. Biosystems Engineering, 83(3), 281e290.
Zhao, Z., Wu, Y. F., Yin, J. J., & Tang, Z. (2015). Monitoring method
of rice seeds mass in vibrating tray for vacuum-panel
precision seeder. Computers and Electronics in Agriculture, 114,
25e31.