Information Technology Reference
In-Depth Information
β
d
E
σ
ij
(
k
+1)=
σ
ij
(
k
)
−
d
σ
ij
,j
=1
,
2
,...,m
i
(9)
where,
i
=1
,
2
,...,n,β>
0 is the learning rate.
The training process of neural networks is that constantly adjust the weight
coecient between each layer until the anticipation error of the neural network
output meets the application requirements and save the learning fuzzy neural
network data. After learning, download the fuzzy neural network weights and
thresholds to the decision maker.
4 Simulation Experiment and Analysis
4.1 Selection of Sample Data
The data are collected from real-time operating on greenhouse spraying site to
constitute training samples and train the constructed fuzzy neural network.
Analyzing the real-time operating records, it is concluded that the actual
quantity of spraying is determined by operators spraying experience combin-
ing with influence factors including the crop area, distance, damage level. The
operation data, to some extent, reflects the operator's experience and strategy.
Choose
3
of the sample data to train the fuzzy neural network (the rest
3
as the
test samples), and make the trained network memory these experience and au-
tomatically generate a series of fuzzy rules, then the reasonable rate of spraying
can be decided. Parts of sample dates are shown in Table 1.
4.2 Fuzzy Neural Network Training
In this paper, the model is built up by MATLAB neural network toolbox func-
tion. Select 7, 7, 7 as each input fuzzy division and Gauss function as membership
function. In learning algorithm, the training goal error is defined as 0.001, and
the training step is defined as 1000. The momentum and adaptive gradient de-
scent training function Traindx is used as training function. Then the MATLAB
model of fuzzy neural network is established.
Results of network training are shown in Fig. 4. It is indicated that the de-
viation between the actual training output and the expectancy output is very
small, the fuzzy neural network reflects the decision rules well and the training
result is more ideal. Target error of the network, after 191 steps training, meets
the convergence requirement.
4.3 Experimental Verification
The spraying mobile robot system for greenhouse studied in this paper is shown
in Fig. 5. The system is consisted of a mobile robot, a computer, a water tank for
pesticide, an image collector, an ultrasonic sensor, a decision maker, a controller,
an electromagnetic valve, a flow sensor, a sprayer, etc.
Search WWH ::
Custom Search