Measurement: Sciencedirect
Measurement: Sciencedirect
Measurement: Sciencedirect
Measurement
journal homepage: www.elsevier.com/locate/measurement
A R T I C LE I N FO A B S T R A C T
Keywords: A belt weigher is widely used in industrial production and trade settlement; however, it is difficult to maintain its
Error compensation nominal measuring accuracy in service. With the belt weigher indication, average speed of belt, variation in belt
BP neural network sag, running deviation of belt, environmental temperature, and humidity as inputs, and the control instrument
Automatic weighing indication as output, a BP neural network model to compensate the automatic weighing error of the belt weigher
Dynamic material experiment
is built. We obtained the sample data depending on the test system for the type evaluation of the belt weigher in
Belt weigher
the Jiangsu Institute of Metrology, as well as supplemental relevant parameters for the monitoring devices. The
BP model is trained and validated using MATLAB. The validation results show that the absolute value of the
maximum relative automatic weighing error output by the BP model is less than 0.5%, significantly lower than
the error before using the BP model for compensation. The BP model is effective, feasible, and practical for
compensating the automatic weighing error of the belt weigher.
⁎
Corresponding author at: Jiangsu Institute of Metrology, Nanjing, China.
E-mail address: lby_8441@sina.com (B. Li).
https://doi.org/10.1016/j.measurement.2018.07.080
Received 11 August 2017; Received in revised form 9 July 2018; Accepted 25 July 2018
Available online 29 July 2018
0263-2241/ © 2018 Published by Elsevier Ltd.
B. Li et al. Measurement 129 (2018) 625–632
difficult to analyse. Ideally, the load cell bears the real load value of the
material on the weighing segment. In fact, owing to the conveyor belt
with properties of a viscoelastic material, it will respond to changes
when affected by external forces, and subsequently causes the load cell
to receive unexpected signals and output a signal of weighing gravity
WL̂ (t ) that is different from WL (t ), thus causing the error of weighing
gravity. The sources of error of the weighing gravity contain the ‘belt
effect’, vibration of belt, running deviation of belt, sag resistance,
centrifugal force, and running resistance of belt. Although the sources
of error of the weighing gravity are complicated, most of them can be
reduced, eliminated, or consolidated by appropriate methods. Finally,
they can be approximately summarised as the influence of tension and
running deviation of the belt [9–12].
Fig. 1. Assembly diagram of belt weigher.
The sources of automatic weighing error can be summarised in five (5) Environmental error
categories as follows [4]:
The belt weigher is often used in relatively poor conditions, e.g.
(1) Error of weighing gravity unstable temperature and humidity . The volatile temperature and
humidity will change the belt status and will significantly affect the
According to literature research, the error of weighing gravity stability and linearity of the weighing result of the belt weigher [13].
constitutes the largest source of automatic weighing error, i.e. ap- Temperature may also affect the electrical performance of the belt
proximately 85% of the total automatic weighing error [11]. The error weigher body and the performance of the load cell. Therefore, the
of weighing gravity has a complicated composition and is the most temperature and humidity should be monitored.
626
B. Li et al. Measurement 129 (2018) 625–632
Fig. 3. Typical model of neurons. Fig. 5. Schematic diagram of running deviation of belt.
627
B. Li et al. Measurement 129 (2018) 625–632
f (x ) = x (4)
j= i+k +α (5)
Fig. 6. Block diagram of BP model. The self-built BP model is trained by the built-in toolbox of the
forward neural network (feedforwardnet) under the operation en-
with nonlinearity, differentiability, and monotonicity can meet the vironment of MATLAB R2011b. The maximum number of training
basic requirements of the BP model [14]; in addition, it can compress (net.trainParam.epochs) is set as 1000, the learning rate
the continuous input real value between -1 and 1 and output 0 mean (net.trainParam.lr) is set as 0.01, and the minimum error of training
values to effectively guarantee that the update of the weight value is a target (net.trainParam.goal) is set as 0.0001. The feedforwardnet
gentle iterative change rather than a sawtooth mutation. The training toolbox will automatically divide the input data into training data,
speed and calculation precision of the tansig function are outstanding. validation data, and testing data with the ratios of 70%, 15%, and 15%,
The prototypes of the tansig function are as follows: respectively. To guarantee the number of training samples, the ratios
are changed to 85%, 8%, and 7%, respectively.
2
f (x ) = −1
1 + e−2x (3)
4. Acquisition of sample data
The training process of the BP neural network is as follows: the
input layer values spread separately to the hidden layer after network We obtained the experimental sample data depending on the test
computing, and subsequently spread to the output layer similarly. system for the type evaluation of the belt weigher in the National Type
Finally, the network calculates the error between the output value and Evaluation Laboratory of Automatic Weighing Instrument of the
sample value, and the network continues training until the error Jiangsu Institute of Metrology, as well as supplemental relevant para-
reaches a small set value. The purelin function can maintain the am- meters for the monitoring devices.
plification and reduction of arbitrary previous range values, and eases
the comparison between the output value and sample value. Therefore, 4.1. Test system for type evaluation of belt weigher
we selected the purelin function as the activation function of the output
layer. The system shown in Fig. 7 is equipped with eight conveyor belts
The prototypes of the purelin function are as follows: with different bandwidths. The maximum flow rate is up to 7200 t/h,
628
B. Li et al. Measurement 129 (2018) 625–632
the highest belt speed is 4 m/s. Its bandwidth, material flow rate, belt
speed, etc., can accommodate the belt weighers typically used in the
market. The system can real-time control the parameters such as uni-
formity of material delivery, speed of belt and feed flow, and meet the
experimental requirements of a belt weigher of each accuracy class. It
uses wear-resistant, high-hardness, and non-friable ceramic granules as
the experimental
material, with a total capacity of 150 tons. It is also equipped with
five hopper weighers as the control instruments. The system fully
complies with the requirements of material experiment stated in the
International Recommendations and Chinese National Procedures of
Belt Weigher.
629
B. Li et al. Measurement 129 (2018) 625–632
matrixandthethresholdvalue , andsubsequentlynormalisethedata. and The trained BP model uses the five retained groups of sample data to
the threshold value, and subsequently normalise the data. validate its effect. Table 1 shows the validation results and the com-
Step 3: calculate the weighted sums of the input of the eight neurons pensation results of the automatic weighing error using the BP model.
in the hidden layer and the output value of the model. In Table 1, the output value of the BP model is closer to the true
Step 4: the mean square error (MSE) of the output value of the value of the material. The absolute value of the maximum relative
model and the sample value are calculated. If the MSE does not automatic weighing error output by the BP model is less than 0.5%, and
reach the training target, the error needs to be propagated backward is significantly lower than the error before using the BP model for
by differentiating the MSE with respect to the weight to obtain the compensation. The validation results show that the BP model exhibits
gradient.. effectiveness, high precision, and better generalisation ability to var-
Step 5: update the weight matrix as follows: iations in belt tension, belt speed, running deviation, temperature, and
humidity. In addition to the belt weigher with the same weighing
(1) W= (0)W - u × D(1) (6)
principle but different bandwidth, new experimental sample data can
be obtained and the BP model can be subsequently retrained, such that
Step 6: the BP model recalculates using the updated weights and the BP model has universal applicability.
iterates continuously until the MSE is less than the training target. However, limited by the cost of the dynamic material experiment on
the test system for the type evaluation of the belt weigher, we obtained
Under the MATLAB software platform and according to the training only 65 groups of sample data experimentally. Therefore, the re-
process above, the BP model was simulated and trained several times. presentativeness of the sample data is deficient. Moreover, it cannot
Subsequently, the ideal performance curve of the model error (Fig. 11), artificially change the temperature and humidity in the laboratory;
and the regression curve of the model error (Fig. 12) of the BP model therefore, the training sample is unbalanced relative to the temperature
are obtained. Figs. 11 and 12 show that the training results achieve the and humidity. For future work, we will continually collect sample data
training target after the seventh iteration, as well as the small number with different environment temperatures and humidities during dif-
of iterations and short training time. ferent seasons to improve the accuracy and generalisation of the BP
In Fig. 11, the abscissa represents the number of iterations and the model, to achieve the goal of maintaining the error in-service within its
ordinate represents the MSE between the output value of the BP model nominal range for a long time.
and the sample value. The smaller the MSE, the closer is the output
value of the BP model to the sample value. The ‘train’ blue line shows 6. Conclusions
the MSE performance in each iteration in the BP training process; the
‘validation’ green line shows the MSE performance in each iteration in (1) The automatic weighing process of a belt weigher can be modelled
the BP cross validating process; the ‘test’ red line shows the MSE per- using the BP neural network.
formance in each iteration in the BP testing process. The ‘best’ dotted (2) A simple and practical dynamic totalised method of belt weigher is
line shows that the training results are the most ideal after the seventh proposed based on the BP neural network, while complex me-
iteration. The ‘goal’ dotted line represents one of the set targets for the chanical analysis and cross influence of parameters are avoided.
BP model to stop training. (3) The average belt speed, average variation in belt sag, average
The BP model synchronously predicts the training sample, valida- running deviation of the belt, and the environment temperature and
tion sample, testing sample, and all samples while training by the re- humidity during automatic weighing are influencing factors of the
gression equation. Fig. 12 shows the regression results. The abscissa weighing results. The material experiment primarily depended on
represents the target value; the ordinate represents the output value of the test system for the type evaluation of the belt weigher in the
the BP model; ‘R’ is the regression coefficient. The closer is the value of National Type Evaluation Laboratory of Automatic Weighing
‘R’ to one, the better is the BP model and the higher is the prediction Instrument of the Jiangsu Institute of Metrology. We used the
accuracy.
630
B. Li et al. Measurement 129 (2018) 625–632
631
B. Li et al. Measurement 129 (2018) 625–632
Application, North China Electric Power University, 2008 (Ph.D Thesis). [15] H.B. Bahar, D.H. Horrocks, Dynamic weight estimation using an artificial neural
[10] Fansheng Wang, Study on the Flexible Multi-body Dynamics Analysis Method of network, Artificial Intell. Eng. 12 (1–2) (1998) 135–139.
Belt Conveyor, China University of Mining and Technology, 2010 (Ph.D Thesis). [16] Jie Yan, Yan Zhao, Junli Zhang, Study on static nonlinear error compensating for
[11] Kun Chen, Static And Dynamic Analysis On The Measurement System Of An weighing sensor based on BP neural network, Chinese J. Sensors Actuators 21 (6)
Electronic Belt Scale, Anhui Polytechnic University, China, 2014 (M.D Thesis). (2008) 1025–1028.
[12] Liang Zhu, Study on the Accuracy Compensation of Continuous Weighing System of [17] Yufeng Zhuang, Hu. Xiaojin, Yu. Zhai, Nonlinear compensation of micro scale
Bulk Materials, Nanjing University of Science & Technology, China, 2015 (Ph.D capsule dynamic condition weighing unit based on BP neural network model,
Thesis). Chinese J. Sci. Instrument 35 (8) (2014) 1914–1920.
[13] Yanqiang Yuan, Discussion of error theory and durability of Belt scale, 11th [18] Song Mei, Yuee Zhou, et al., The relational analysis of sag and tension based on the
Weighing Technology Seminar, Nanjing, China, (2012). method of least square curve fitting, Manuf. Automation 19 (2012) 123–126 (In
[14] Martin T. Hagan, Howard B. Demuth, Mark H. Beale, Neural Network Design, PWS Chinese).
Publishing Company, USA, 1996.
632