Characterization of the Collagen Extraction Manufacturing Process Using Markov Chains and Artificial Neural Networks

Modelling a process requires a priori knowledge of the possible causal relationships between the study variables and the response, especially when the raw material is waste. A new approach to characterizing a manufacturing process is presented to extend data that are difficult to obtain by direct experimentation. This research aims to identify industrial applications for derivatives from animal by-products, such as tilapia skin, to develop an alternative methodology for issues related to collagen. Empirical data from a design of experiments (DoE) fed a Markov chain Monte Carlo algorithm in an artificial neural network (MCMC-ANN) for data imputation, numerical simulation, and training. The DoE model presented a coefficient of 86.92% in the observed data in the first stage of the methodology. In the second stage, the predictive model MCMC-ANN presented a mean square error (MSE) of 0.00171, and a standard deviation (SD) of 0.00133. The model was tested using validation data obtained outside the training process. The SD of empirical validation of MCMC-ANN were 1.7890, 0.2051 and 0.0919 for three random points. The product of this research is a four-dimensional predictive model that estimates the yield with decreasing material, energy resources, and the time required for the collagen extraction process from the skin waste of tilapia. The proposed tool obtains a minimum adjustment with a deviation whose tolerance is valid for predicting the collagen extraction yield. MCMC-ANN can be applied to any case that requires a deeper analysis with more data or experiments.