Qualitative and quantitative near infrared analysis using artificial neural networks

The performance of artificial neural networks (ANNs) for near infrared analysis was studied. Two ANNs were used in this research: (1) a linear neuron and (2) a backpropagation network.; The first paper discusses the design of back propagation networks and a method of calibration and validation. The optimum network architecture was chosen to be the one which gave the best performance among the different network architectures tested. Model calibration and validation for ANNs were divided into: (1) a new training approach for ANNs, (2) calibration model and (3) true performance. The neural networks models were trained by dividing the data into training sets and tuning sets using 5-fold cross validation. As the training process proceeded, the MSE of the tuning set was recorded. The minimum MSE of the tuning set and its corresponding epoch number were determined. Then, a network was trained to the same epoch number using all the available data. A calibration model was built using all the available data. The true performance of the model was determined by dividing the data using 10-fold cross validation. The training procedure was applied for each training set. The true performance was determined as the average of the ten testing sets performance.; In the second paper, the quantitative performance of ANNs models were compared to multiple linear regression. Data set A and B were used as the basis of estimating nicotine in tobacco. For data set A, the MSE of the calibrated regression model and its true performance (0.0105, 0.0122, respectively) were better than the backpropagation network (0.0117, 0.0142) and the linear neuron (0.0130, 0.0130). For data set B, the backpropagation network (0.0256, 0.0384) outperformed both the linear neuron (0.0478, 0.0592) and the regression model (0.0478, 0.0592) for both the calibration model and its true performance.; In the third paper, the performance of ANNs was compared to a quadratic discriminant analysis model. The correct classification rate for classifying Burley and flue-cured tobacco (data set C) was (100%, 100%) using discriminant analysis followed by (99.38%, 99.39%) using backpropagation network for the calibration model and it's performance. The linear neuron model gave (95.19%, 99.26%). The same three models were used to identify native Burley tobacco (data set D). The results for the calibration model and its true performance were (100%, 100%) for discriminant analysis, (89.12%, 88.46%) for backpropagation network and (80.68%, 79.62%) for the linear neuron model.