THE ARTIFICIAL NEURAL NETWORKS AND THE DRYING PROCESS MODELING

Artificial Neural Networks (ANN) becomes an important tool in modellin g. designing and controlling a chemical or biochemical process. becaus e of its learning and generalisation properties that confer it the ful l poser of a self organising system. A neural network is formed of syn thetic neurones. grouped in layers (input. output and hidden). Each on e has as task to process the signals received from its dendrites accor ding to its threshold function (the main step when processing informat ion): after that. the answer outputs through its axon to the rest of t he neurones (($) under bar 1). This stands for the ability of the net to use the information stored into the private neurones weights. In th e learning phase. the derivative of the threshold function plays. on a lmost every case. a key role in matching the answer of the net with th e correct output data learning set, providing that any steep descent l earning rule is used (($) under bar 1). It is obvious that choosing a suitable threshold function is an essential step in having an appropri ate neural network. By far. the most used threshold function is the we ll-known sigmoid f(x) = 1/1+epsilon-x. The authors fully examined the impact over the performance of a given neural network (input, hidden a nd output layers kept the same] of changing this function with another sigmoid: more versatile [GRAPHICS] The shape of this function heavily modifies when changes in alpha and/or beta occur. This new threshold function seems to be more promising due to its ability to match each n euron's needs changing alpha and beta accordingly. The training data w as a set of experimentally obtained points regarding the sebacic acid' s drying as powder with hot air. the neural network learning rule bein g the back-propagation algorithm (($) under bar 1). The learning rate for the new threshold function is drastically affected bg alpha and be ta values, and for these reasons, the following step made by the autho rs was to adjust, during the Learning phase, alpha and beta for each n euron.