doi:10.21311/002.31.11.03 An aesthetic evaluation model of agricultural machineries appearance design based on BP neural network Huiping Guo, Fuzeng Yang* College of Mechanical and Electronic Engineering, Northwest A&F University, Yangling 712100, China Abstract In order to research into nonlinear issues concerning aesthetic evaluation on appearance design of agricultural machineries, the paper used BP (Back propagation) neural network as a basis to have established a corresponding mathematic evaluation model. The assessment model was built up based on BP neural network, with 16 criteria values of aesthetic evaluation as an input layer and the comprehensive value as an output layer. Test results showed that among the 20 sample schemes, either the training values of 15 training samples or the test values of the 5 verification samples agreed with real ones. The research result thus provided methods of quantitative analysis of aesthetic measure, which helped perceive consumers aesthetic psychology better during product design and improve design efficiency as well. Keywords: Aesthetic evaluation, BP neural network, Tractors. 1.INTRODUCTION With economic development and improvement in living standards, in order to meet consumer demands at different psychological levels, product appearance design is supposed to match aesthetic needs in addition to realization of basic functions and usefulness. A beautiful appearance design attracts consumers attention and further sparks purchasing desires. What s more, aesthetic concepts contained in products can reinforce consumers impressions and upgrade their aesthetic experience. As a psychological response to product appearance design, aesthetic perception is deemed as emotional, subjective, and uncertain. The level of aesthetic perception, namely aesthetic measure, as an objective measurement is difficult for designers to achieve. Therefore, it is particularly significant to carry out quantitative theoretical research on aesthetic measure. Requirements for appearance design of agricultural machineries have ascended in the wake of increasing agricultural mechanization degrees, expanding worldwide sales market, and annually mounting inventories of agricultural machineries (Zhang et al., 2013). Innovativeness and brand identity for agricultural machineries has become a major direction of research and development of agricultural machineries, where appearance is the most representative part (Sun, 2013). When designing agricultural machineries, one used to give more priority to basic functions such that correlated manufacture can be completed, but lacks attention on appearance and agreeableness. Hence traditional impressions on agricultural machineries are concentrated on rough workmanship, heavy structures, and insufficient aesthetic beauty. The same impression is especially applicable to tractors (Zhang, 2010). For recent years, technological development regarding tractors has been stagnant. To bring new-type tractors into prominence, it is necessary to endow them with exquisite appearance that will draw purchasers attention and further arouse desires to buy (Zhu, 25
2005). Along with enrichment of social life and diversified tastes of life, single-shaped, monotone tractors have failed to satisfy aesthetic needs. Instead, strong contrasts arise between traditional agricultural machineries and consumer demands. Approaches to information of consumers potential aesthetic perception has become one of the important research directions in appearance design of agricultural machineries. Since the structures and functions of agricultural machineries are more complex than those of industrial products, corresponding research methods of appearance design and its evaluation are complicated still further. To study aesthetic measure of agricultural machinery appearance from the perspective of consumers judgements provides the access to realizing aesthetic evaluation on agricultural machineries and the basis for improving appearance design of agricultural machineries. Among studies on aesthetic measure analysis methods, Hua (1997)initiated the concept of aesthetic measure itself as well as product aesthetic measure. By conduction research on general industrial products, he analyzed four representatives of product aesthetic measure and its evaluation factors, and devised the function relations between aesthetic measure evaluation factors and integral evaluation. What s more, Hua determined the weighted value among those factors and summarized the procedure of aesthetic measure evaluation as well. Kou and Cui (2002, 2003a and 2003b) proposed the thought of quantitative evaluation on aesthetic measure and that of aesthetic space by means of describing appearance aesthetic measure with the help of concept of space as well as function relations. Gan and Xu (2010) put forward the idea that aesthetic factors were a crucial criterion in determining actual superiority and that the combination of technological beauty and formal beauty constituted the mainstay of evaluation standards. With a case study involving Apple computers, they also proposed the way to undertake fuzzy quantitative evaluation on aesthetic measure. Zhou et al., (2013) took interface design of products as an example to apply Kansei engineering to aesthetic evaluation. Through analysis, they determined aesthetic measure factors, established aesthetic measure computation system, and researched on superiority orders of scheme design with the help of grey correlation analysis as well, thus providing basis for comprehensive evaluation on aesthetic measure. Furniture is used more often for case studies of aesthetic measure. To practically optimize and choose schemes of appearance design for furniture, and to improve evaluation systems of furniture design, Chen and Lyu (2014) proposed that aesthetic indices should be used to assess design schemes. Based on Analytic Network Process (ANP), they established weight relations among evaluation indices, and realized automatic evaluation by virtue of SuperDecisions software. To quantify aesthetic values, and given appearance attributes for different furniture, Zeng and Liu (2010) built up weight relations by means of AHP, and designed corresponding models to assess and quantify aesthetic value systems of furniture as well. Products with simple structures are the mere objective of present studies, where linear methods are used to analyze quantitative relations between aesthetic decomposition indices and integral evaluations of aesthetic measure. However, there has yet been no attempt to evaluate aesthetic measure by building up any mathematical model. For tractors with complicated structures, quantitative analysis on aesthetic measure is still more intricate, posing new questions of aesthetic evaluation. The development of computer technology gives birth to intelligent algorithm, which has been applied to various research fields. Given that the fine attributes of BP neural network, namely selfadaptability, parallel processing, nonlinearity, fault-tolerance, and ratiocination, are what aesthetic evaluation requires of, BP neural network algorithm thus has the ability to handle such emerging issues in aesthetic evaluation as causal relationships and inaccurate data with possible contradictions and errors (Tang and Hu, 2008, Li et al., 2016). Therefore, the paper suggested to draw support from BP neural network algorithm in setting up a mathematical model that evaluated aesthetic measure of 26
tractors appearance design, intending to provide instructions for aesthetics-based appearance design of tractors. 2.RESEARCH FRAMEWORKS For the case study of wheeled tractors, a kind of typical agricultural machinery, in the paper, by analyzing characteristics of the wheeler tractor s appearance, the paper built up corresponding tier-one and -two evaluation criteria for its aesthetic measure. In addition, on-line evaluation on aesthetic measure of 20 sample schemes was conducted. Then, with BP neural network, a correlated model was established between each tier-two aesthetic evaluation values and integral aesthetic evaluation values. Figure 1 is the flow chart of aesthetic evaluation. Figure 1. The flow chart of aesthetic evaluation 3.CONSTRUCTING OF THE AESTHETIC EVALUATION SYSTEM 3.1 Analysizing factors of the aesthetic evaluation Aesthetic evaluation of appearance design is obtained through analysis on the premise that product design has satisfied functional requirements. According to structural features and functions, the paper established four tier-one criteria for aesthetic evaluation on appearance design of tractors: morphological aesthetic measure, color aesthetic measure, function aesthetic measure, and other aesthetic factors. Sixteen tiertwo criteria were determined, four of which were function aesthetic measure, three for color aesthetic measure, six for morphological aesthetic measure, and three for other aesthetic factors. 3.2 The method of obtaining test data for the aesthetic evaluation In the collection of over 100 pictures of different tractors appearance design, the paper selected out 30 clear ones with coordinated shooting angles and distinct appearance. With reference to morphological aesthetic measure, color aesthetic measure, function aesthetic measure, and other aesthetic factors, the paper classified and incorporated the 30 pictures into 20 sample schemes for research in the paper. The result of the first interview demonstrated that only those with relative professional backgrounds could understand appearance design factors concerning tractors well. Therefore, the respondents were chosen to be students majoring in machine design, agricultural machine design, or industrial design, agricultural machinery designers, and interested mass. 120 questionnaires were issued on platforms of Internet, WeChat 27
groups, and QQ groups, the percentage of valid questionnaires being 95%. Spss (Liu, 2014) analytic software was used for statistical analysis on collected evaluation data; 7- order Likert scale was applied to aesthetic evaluation, where 1 denoted the lowest aesthetic perception, 7 the highest aesthetic perception, and 4 the mean aesthetic perception. 18 experts in the industry were invited to give integral aesthetic evaluation under the hundred-mark system. Table 1 Aesthetic evaluation indices Assessment Objectives Evaluation onthe aesthetics measure First-class index level Functional aesthetic measure Morphological aesthetic measure Colour aesthetics Other aesthetics measure Second-class index level Conform to Cultivation practice Conform to Cultivation environment Explicit functional areas Continuous transition between functional areas Conform to functional requirement Functional unity Proper proportion Equilibrium and stability Unity in change Balanced contrast Conform to functional requirement Overall colour harmony Comply with the operating environment Detail processing Surface treatment Economical modeling 4.ESTABLISHING THE AESTHETIC EVALUATION MODEL BASED ON BP NEURAL NETWORK 4.1 Principles of BP neural network Initiated by a scientist group headed by Rumelhart and McCelland, BP neural network is a multi-layered feedforward network that is trained by back propagation algorithm, and has become one of the most widespread neural network models. The backpropagation learning algorithm can be divided into two phases: forward propagation of input messages and backward propagation of errors (the difference between the targeted and actual output values)(lazard, 2016). Neurons in the input layer are responsible for receiving outside input and transferring it to neurons in the middle layer. The middle layer is in charge of processing and converting internal information. According to demands for conversion, the middle layer can be alternatively designed as a single hidden layer or multiple hidden layers (Li et al., 2016). After all the information transferred from the middle layer to the output layer underwent further processing, a forward propagation of input messages is completed. The rest is outputting information processing results to the outside through the output layer. If actual output fails to agree with expected output, the stage of error backward propagation starts up. The backward propagation algorithm calculates the gradient of the error of the network regarding the 28
network's modifiable weights, and passes information stepwise to the hidden layer and the input layer. The procedure repeats, during which weights in various layers are modified and the neural network is trained. This process will not stop until output errors are reduced to an acceptable degree or until the number of training has reached the preset times. Figure 2 shows the working principle of BP neural network. Figure 2. The working principle of BP neural network 4.1 The aesthetic evaluation model of tractors appearance design based on BP neural network In constructing the aesthetic evaluation model based on BP neural network, it is necessary to determine the number of nodes in the input layer, the hidden layer, and the output layer, respectively. For the research in the paper, the 16 tier-two index values as the input information corresponded to 16 nodes in the input layer; the one integral evaluation value as the output result was converted to one node in the output layer. On the MatlabR2014 software platform, the aesthetic evaluation model of tractors appearance design based on BP neural network was established. Figure 3 shows the BP neural network structures concerning aesthetic evaluation. Figure 3. The BP neural network structures concerning aesthetic evaluation 29
5.SIMULATION AND VERIFICATION 5.1 Sample data 5.1.1 Training samples The sample data contained learning samples and test samples. 20 sample schemes of tractors were chosen as test data in the paper. Table 2 is the statistical results of 20 tiertwo indices. Table 3 is the integral evaluation value, where 15 samples were stochastically selected as learning samples and the rest as test samples. As could be seen from Table 3, the aesthetic evaluation order was: A>F>P>I>L>T>K>G>M>C>B>D>E>S>H>R>Q>O>J>N. Table 2 The statistical results of 20 tier-two indices The sample Values of the evaluation of target X1 X2 X3 X4 X5 X6 X7 X14 X15 X16 A 6.7 6.3 5.8 5.5 3.8 4.5 4.8 6.3 5.8 4.7 B 6.0 6.2 3.6 4.1 3.5 3.7 6.1 3.1 6.2 3.4 C 4.0 3.1 6.3 5.8 3.4 4.8 3.1 3.5 3.2 3.1 D 4.2 4.0 4.8 3.4 3.4 4.3 3.5 3.5 4.4 3.8 E 3.6 3.1 5.7 3.6 3.4 4.7 3.3 3.5 4.5 2.8 F 6.3 6.1 4.3 5.8 3.3 4.8 5.6 6.2 5.8 4.7 G 6.5 6.2 3.3 4.3 3.3 3.8 6.0 3.8 5.9 3.7 H 5.0 4.5 4.7 3.8 3.6 3.8 3.0 3.8 3.1 3.2 I 5.8 6.0 4.8 6.4 3.5 4.5 4.5 4.8 5.1 3.8 J 2.9 2.9 2.9 3.9 2.8 4.2 2.9 3.8 2.8 2.7 K 5.0 3.9 3.9 3.5 5.6 5.6 4.2 2.9 5.0 3.1 L 5.9 5.5 4.2 5.2 5.3 4.3 5.2 4.2 5.6 4.0 M 6.0 4.9 4.6 4.2 4.8 5.5 4.2 3.1 4.3 3.2 N 2.2 3.0 2.3 2.9 2.8 3.1 4.0 2.8 2.6 3.3 O 4.2 3.8 3.9 3.6 3.6 3.7 3.0 3.1 3.4 3.0 P 6.0 5.2 4.3 5.3 4.3 4.9 5.8 4.7 5.1 4.7 Q 4.3 4.1 3.4 4.2 3.5 4.2 3.2 2.4 3.5 3.8 R 5.1 3.8 4.1 3.8 3.8 4.1 3.1 2.1 3.4 3.9 S 3.2 4.7 5.1 3.7 4.1 5.2 3.8 3.4 2.6 3.9 T 4.8 5.1 5.2 4.9 3.9 5.4 6.0 4.1 3.8 4.5 5.1.2 Sample data processing Table 3 The integral evaluation values A B C D E F 88.1 74.2 68.2 64.4 63.6 86.3 G H R S T 72.3 59.8 59.6 61.7 75.8 Before the BP neural network underwent training, all of the sample data (including input data and output data) should be normalized. This measure was done to eliminate bizarre sample data, summarize the same sample data itself and its distribution characteristics, and accelerate learning efficiency as well as computation convergence efficiency. Antinormalization processing should be undertaken on simulation results so that the processed data reflected real results. Two linear transfer functions (Equation (1) and 30
Equation (2)) were used for normalization processing and anti-normalization processing, respectively. pn p p p max p min min (1) Where p n denoted normalized values, P represented original sample values, and p max and p min were the maximum original value and the minimum original value, respectively. q qn( q max q min) q min (2) Where q denoted anti-normalized values, q n represented original simulation values, and q max and q min were the maximum simulation value and the minimum simulation value, respectively. 5.2 Sample training and verification 16 tier-two evaluation values for the 15 sample schemes in Table 2 were the nodes in the input layer, and the integral evaluation value was the node in the output layer. After several trainings, it was obtained that: the number of nodes in the hidden layer was 8, the learning rate was 0.1; the targeted error value was 0.0000001, and the training time was 1000. The training results in Figure 4 agreed well with real values. Figure 4. Comparison between training results and real results Test samples were inputted in the trained BP neural networks, and corresponding training results could be obtained (as shown in Figure 5). Table 4 shows that the error rates of five training samples were 1.37, 1.07, 2.04, 3.96 and 4.73, respectively. The low error rates proved that the training results were reliable. 31
6.CONCLUSION Figure 5. Comparison between test values and real values Table 4 Error rates for test values and real values The samples Real results Testvalues Error rates% C 68.2 67.2656 1.37 A 88.1 89.0432 1.07 M 70.5 71.9384 2.04 G 72.3 69.4403 3.96 K 66.3 63.1602 4.73 By combining aesthetic characteristics with attributes of tractors appearance design, the paper establishes tier-one and -two criteria to conduct aesthetic evaluation on appearance design of wheeled tractors. With integral evaluation value by expert as the output layer, BP neural network acts as the basis of the proposed evaluation model in the paper, by which way quantitative aesthetic evaluation on appearance design of wheeled tractors is realized. 15 random samples of the 20 selected samples in the paper were tested with the established aesthetic evaluation model. The results show that test values and real values fit with each other, and that training values agree well with real values. That the error rates of five training samples were 1.37, 1.07, 2.04, 3.96 and 4.73, respectively, proves that the above aesthetic evaluation order maintains constant. This result weighs greatly in aesthetic studies of appearance design of products, and provides reference for appearance design as well. 7. REFERENCES Chen M.,LyuJ.H. (2014). Aesthetic Evaluation of Furniture Design Based on ANP Method, Applied Mechanics and Materials, 574(7), 318-323.Doi:10.4028/www.scientific.net/AMM.574.318. Cui J.B., Kou F.M. (2003a). Quantitative Study about Merchandise Esthetics(II), Journal of Gansu Normal College, 8(2), 17-18. Cui J.B., Kou F.M. (2003b). The Merchandise Esthetics Quantitative Analysis, Journal of Lanzhou University(Natural Science), 39(5), 25-29. Gan Q.C., Xu R.P. (2010). The Aesthetic Evaluation of the Product Design, Art of Design (Journal of Shandong University of Art and Design), (2), 75-78. Hua E. (1997). Research into Valuing Method of Degree of Beauty of Product. Bulletin of Science and Technology, 13(4), 245-249. 32
Hu J.B., Tang X.Q. (2004). BP Algorithm and Its Application in Artificial Neural Network, Information Technology, 28(4), 1-4. Kou F.M., Cui J.B. (2002). Quantitative Study about Merchandise Esthetics, Journal of Gansu Normal College, 7(5):54-56. Lazard M. (2016). Heat Transfer in a Semi-Transparent Parallelogram Shaped Medium. International Journal of Heat and Technology, 34(s2), S420-S424. Doi: 10.18280/ijht.34S232. Liu A.Y. (2014). Fundamental Analysis of SpssTutorial, Peking University Press-ISBN: 9787301246986, Beijing, China. Li M., Liao R., Luo W., Dong Y. (2016). Improved Aziz Prediction Model of Pressure Gradient for Multiphase Flow in Wells. International Journal of Heat and Technology, 34(3), 423-428. Doi: 10.18280/ijht.340311. Li Z.D., Guo J.K., Wang Z.J., Miao S.Y. (2016). Research on combined forecasting model for logistic material demand based on BP neural network and grey system theory, Revista de la Facultad de Ingenieria, 31(4), 210-220. Doi: 10.21311/002.31.4.19. Sun W.X. (2013). The Application Research of Modern Design Technology in Agricultural Machinery Engineering Design, Agriculture & Technology, 33(9), 32-32. Tang Y.C., Hu J.P. (2008). Application of Neural Network in Tourism Sustainable Development Capacity Assessment, Computer Engineering and Design, 29(18), 4815-4818. Zeng L., Liu D.P. (2010). A study on the model of furniture aesthetic value based on fuzzy AHP comprehensive evaluation. International Conference on Fuzzy Systems & Knowledge Discovery, (3), 1173-1175. Doi: 10.1109/FSKD.2010.5569152. Zhang Q.N.,Ren W.T., Cui H.G., Liu H.Y. (2013). Research of Implementation Effect of China AgriculturalMechanization Promotion Law. Journal of Agricultural Mechanization Research, (3), 249-252. Zhang J.M. (2010). The Exterior Design Research of the Agricultural Machinery. Journal of Agricultural Mechanization Research, (4):99-103. Zhou L., XueC.Q., Tang W.C., Li J., NiuY.F. (2013). Aesthetic Evaluation Method of Interface Elements Layout Design. Journal of Computer-Aided Design and Computer Graphics, 25(5), 758-766. Zhu S.C. (2005). Catching the Opportunity to Meet the Challenges 0f Tractor Market Internationalization. Farm Machinery, (11), 49-52. 33