Using different Machine Learning algorithms to predict characteristics during the vacuum drying of fruits and vegetables
Drying is one of the most important food processing methods to improve the post-harvest shelf life which could greatly reduce the global food waste. Pulsed vacuum drying, alternative food drying strategy, has been proven by researches to be highly energy efficient, particularly when comparing to continuous vacuum drying of food products and is particularly suitable for drying of heat sensitive fruits and vegetables. Predicting drying characteristics using data driven methods are proven to be useful in developing sustainable and energy efficient drying methods. In addition to this, it could also help in real-time monitoring and control of drying processes which could result in better quality dried products.Using machine learning (ML) for predicting drying characteristics has been a bustling field of research in recent years as it could help developing more sustainable and energy efficient drying methods that would be a boon for businesses operating in the food industry. With that in mind, this work has been designed with an aim to use three different machine learning algorithms, namely linear regression (LR), artificial neural networks (ANN) and support vector regression (SVR) to predict the drying kinetics. The prediction of important drying kinetics, which is in this case, the moisture ratio, is performed by training the machine learning algorithms using the data obtained from experiments. Using this historical data, the models has been optimised using various techniques to improve their prediction accuracy.
Cross validation technique has been used to reduce the effect of overfitting of the ML models with the experimental data. An important contribution of this research is the extensive work done on the application of artificial neural networks for prediction. The ANN model’s performance is greatly dependent on the hyperparameters such as the activation and training functions. Therefore, five different training functions namely, Bayesian regularisation (brnn), resilient backpropagation with backtracking (rprop+), resilient backpropagation without backtracking (rprop-), smallest absolute gradient (sag) and smallest learning rate (slr) were used to train the model. An interesting thing to note is that the Bayesian regularised neural network resulted in poor predictive accuracy and a huge margin of error. One hidden layer and linear activation function for the output layer are considered for all the models based on previous studies.
In addition to this, six activation functions were analysed and the best was chosen based on the minimum cross validated RMSE and maximum R2. These activation functions include, Elliot symmetric sigmoid (elliotsig), radial basis (radbas), rectified linear unit (reLu), logistic sigmoid (sigmoid), softplus, and hyperbolic tangent sigmoid (tansig). With that in mind and assuming sigmoid activation function for the hidden layer, rprop- and radbas have been identified as the best training and activation functions, respectively. The appropriate number of hidden neurons to obtain the best predictive accuracy from the model has been found with the rprop- training function and radbas activation function. With the obtained parameters, the ANN model was trained and used for predicting the moisture ratio.
To validate the claims from literatures that support vector regression (SVR) has better generalisation ability than ANN, a SVR model was created. Radias basis kernel function was chosen and to tune the hyperparameters such as cost and sigma, Bayesian optimisation algorithm was used. With the tuned hyperparameters, the SVR model was trained and the moisture ratio was predicted.
