Characterization of oil palm fruitlets using artificial neural network

Accurate data of the dielectric properties of oil palm fruitlets and the development of appropriate models are central to the quest of quality sensing and characterization, and Artificial Neural Network (ANN) and Adaptive Neurofuzzy Inference Systems (ANFIS) are becoming increasingly relevant for this purpose owing to their excellent pattern matching and generalization ability. In this study, a Layer Sensitivity-Based Artificial Neural Network (LSB_ANN) and a Multi-Adaptive Neurofuzzy Inference System (MultiANFIS) were designed to characterize oil palm fruitlets and to model the dielectric phenomena of microwave interacting with oil palm fruitlets within the frequency range of 2-4GHz. The LSB_ANN has a unique weight update mechanism which employs network layer input-output sensitivity analysis. The inputs of the networks are the frequency, the magnitude of the reflection coefficient and the phase of the reflection coefficient while the outputs are the dielectric constant, the loss factor and the oil content. The training data for the models were obtained from dielectric and moisture content measurements and the obtained data were fitted into the quasi-static wave Equations and optimized using MATLAB complex root finding technique to obtain the normalized conductance, susceptance and the complex permittivity of the fruitlets. To further validate the generalization accuracy of the LSB_ANN, its performance was compared with that of a Multi-ANFIS network as well as those of three different ANN training algorithms: Levenberg Marquardt (LM) algorithm, Resilient Backpropagation (RP) algorithm and Gradient Descent with Adaptive learning rate (GDA). Having a testing Variance-For (VAF) of 97.81 and Root Mean Square Error of 3.97, the LSB_ANN was found to possess a better post training generalization ability than the LM, RP and GDA algorithms which had VAF of 93.57, 96.26 and 94.09 respectively, and RMSE of 4.14, 4.38, and 7.98 respectively. The results also showed that contrary to the widely reported gap between the accuracy of the LM algorithm and other feed forward neural network training algorithms, the RP trained network performed as good as that of the LM algorithm for the range of data considered. A user friendly neural network based Graphical User Interface (GUI) was designed suitable for rapid determination of the dielectric constant and percentage oil content of oil palm fruitlets from measured magnitude and phase of reflection coefficient within a frequency range of 2-4GHz.

Preview Text FK 2014 121IR.pdf Download (1MB) | Preview

Actions (login required)

View Item