Real time system based artificial neural network for prediction shelf-life of biscuit with dielectric sensor

Abstract

Biscuit is type of dried food which found frequently on expired condition in market, therefore prediction method should be implemented in avoidance of this condition. Recently, prediction of shelf-life of biscuit has been done by laboratory test, but this test has some disadvantages, i.e.: time-consuming, expensive, also required trained panelists, complex equipment and suitable ambience. One alternative method is the dielectric sensor integrated with Artificial Neural Networks (ANN) in an intelligent real-time system. The aim of this research is to select the measurement attribute dielectric (dielectric parameters, the condition of the circuit and frequency), to formulate design of sensors can measure the value of selected parameters, to design an ANN model, and to integrate sensor and ANN dielectric to predict the shelf-life of biscuits in real time. Selection dielectric parameters and conditions of the circuit (CC/Constant current or CV/Constant Voltage) performed with Linear Regression, Feature Selection (Relief) and Artificial Neural Networks. Dielectric parameters for determining the best shelf life of biscuits are parallel capacitance (Cp). Cp constantly in the top five at three selection method, especially in ANN. Whereas the best condition of the circuit's for determining the shelf life of biscuits are CC, because always at the first sequence of three methods of selection. Selection of the frequency range 42 Hz - 5 MHz done graphically. Frequency range is able to differentiate a good shelf-life is a biscuit with a low frequency range from 42 to 351.65 Hz and the high frequency of 4721.13 to 50054.56 Hz. Selection results of the previous stage for dielectric parameter, the condition of the circuit and frequency is used for sensor design formulation basic. Approach the design tools in accordance with the results of the selection is a RC circuit. Capacitance (C) be the parameters of the sensor dielectric designed. The frequency of this sensor set 5-6 kHz, because the results are more sensitive reading capacitance values of different shelf-life and type biscuits. Designing neural network model is the next stage. This model is designed by using a data input capacitance (C), frequency (f) and Dielectric Constanta (k) from the sensor readings. Moreover, the input data on ANN is a type of packaging, type of biscuit and actual shelf-life (in packaging). ANN learning algorithm is Back propagation. For training process, trial and error is processed in the activation function, learning function, amount of nodes per hidden layer variation, and number of hidden layer. The best training architecture (lowest MSE and highest R) is ANN with the 5 hidden layer, 10 nodes per hidden layer, the hidden layer activation function tansig, the output layer activation function purelin, learning function trainlm and 86 epoch. ANN combined with dielectric parameters has lowest MSE (9,8648 x 10-5) and R highest (99.80%) in the training performance. ANN models are integrated with dielectric sensor using microcontroller AVR ATMega 8535. This integration to measure in real time the dielectric value (variable input) of an unknown product shelf life and readable on computer. To facilitate interaction with the user, an interface is built with MATLAB GUI toolbox. So, integration of ANN and capacitance sensor has the ability to predict the shelf-life of biscuit in real time. The performance of intelligent real-time system in predict the shelf-life of biscuit is enough accurate. Verification get wafers, cookies, crackers and hard biscuits have consecutive a shelf life prediction of 517, 488, 338 and 377 days. The actual expiration biscuits fourth are 486 days. For the next research, the learning process of all parameters in parallel ANN architecture design can be tried to get a more accurate performance (MSE <0.0001 and R> 99%) with a shorter time. ANN performance improvement can be tried also for more number of hidden layer and nodes with a smaller goal (<10-4). Another alternative to increase the performance of ANN is add data (at least 10) at varying frequencies for each sample and shelf-life date. Sensor frequency be set at one value can also be tried to reduce the fluctuations in the measured capacitance value and improve the performance of ANN testing and shelf-life prediction. Input data shelf-life at ANN, can be tried with value from laboratory version and not from manufacture version. It is for predictive value more valid. Correlation dielectric properties with AW (water activity) can be studied so that the application of this system can be more extensive. Dielectric parameter phase can be tried as an alternative basis of measurement sensors, so that increasing accuracy in the future.