Coffee Global Quality Estimation Using Multilayer Perceptron
This application uses artificial neural networks for qualifying coffee batches or coffee brands from a set of sensors based on conductive polymers, which were developed by EMBRAPA (Brazilian Agricultural Research Corporation).
- PDF / 244,797 Bytes
- 5 Pages / 439.37 x 666.142 pts Page_size
- 66 Downloads / 202 Views
Coffee Global Quality Estimation Using Multilayer Perceptron
11.1
Introduction
This application uses artificial neural networks for qualifying coffee batches or coffee brands from a set of sensors based on conductive polymers, which were developed by EMBRAPA (Brazilian Agricultural Research Corporation). The set, known as the “electronic tongue,” is shown in Fig. 11.1. This device enables a quality characterization of the analyzed coffee, which is something traditionally done by professional tasters by giving the sample a graded score between 1 and 10. The electronic tongue performs capacitive reactance measurements in the samples (in a liquid environment) using 55 distinct frequencies. Using artificial neural networks, it is desired to extract, from this information, all those features that could be used for creating a model capable of assigning grades to coffee batches or coffee brands. Therefore, this problem consists basically in fitting a curve whose domain is defined by frequencies from the operation of the electronic tongue, which gives a measure of the coffee quality as a response. The response is tabulated on a scale from 1 to 10. These values are related to those assigned by specialized tasters, being the score ten to the maximum quality level.
11.2
MLP Network Characteristics
The input signals were normalized between [0; 1] for a better output conditioning. The ANN used in this application was a multilayer perceptron (MLP). Several MLP topologies were tested, finally leading to two candidates with good results, this is, Topology T1 (with one hidden layer) and Topology T2 (with two hidden layers), whose configurations are shown as follows: © Springer International Publishing Switzerland 2017 I.N. da Silva et al., Artificial Neural Networks, DOI 10.1007/978-3-319-43162-8_11
209
210
11
Coffee Global Quality Estimation Using Multilayer Perceptron
Fig. 11.1 The “Electronic Tongue” set
• TOPOLOGY T1 – Input Layer ! 55 input signals, related to the capacitance measurements in 55 different frequencies. – Hidden Neural Layer ! 28 neurons. – Output Neural Layer ! 1 neuron. • TOPOLOGY T2 – Input Layer ! 55 input signals, related to the capacitance measurements in 55 different frequencies. – First Hidden Neural Layer ! 10 neurons. – Second Hidden Neural Layer ! 2 neurons. – Output Neural Layer ! 1 neuron. The Backpropagation algorithm was used as the training algorithm in the learning phase. Cross-validation with a dataset made up of 236 examples was used in the first place. From the data set, 90 % of the data were used for training and 10 % of the data for validation. Next, a k-fold cross-validation was applied to increase the generalization ability of the network. Samples were divided into 10 subsets, having 6 subsets with 24 samples and 4 subsets with 23 samples. The learning rate used was 0.1. The specified number of iterations for the network convergence was 500 epochs. The tested topologies can be better understood in Figs. 11.2 and 11.3.
11.3
Computational Results
Capacitive reactance measurements
21
Data Loading...
