Ieroham S. Baruch

Neural prediction of hydrocarbon degradation profiles developed in a biopile

Abstract The biochemical and physical nature of the degradation process in biopile systems is very complex and difficult to describe analytically, thus, neural network modeling and simulation can be of great help. Predictive feedforward neural models (FFNMs) have been commonly used to capture the dynamic phenomena of biological systems by a learning process, but the large number of input/output variables and the vast connectivity of the neural network makes it very time consuming. This paper proposes the use of a recurrent neural network model (RNNM) to predict biodegradation profiles of hydrocarbons contained in an aged polluted soil. The proposed multi-input multi-output RNNM has eight inputs, five outputs, 13 neurons in the hidden layer, and global and local feedbacks. The weight update learning algorithm is a version of dynamic backpropagation. The approximation error for the last epoch of learning is below 1.25% and the total time of learning is about 101 epochs. The learning process is applied to the kinetics of residual hydrocarbons, pH, carbon dioxide, oxygen consumption and moisture obtained with different operational conditions of air flow, and temperature; the kinetics are analyzed at four heights of the columns. The low learning error approximation makes the RNNM interesting to facilitate the study of complex biological processes in a short time.

A soft sensor based on online biomass measurements for the glucose estimation and control of fed-batch cultures of Bacillus thuringiensis

On bioprocess engineering, experimental measurements are always a costly part of the modeling effort; therefore, there is a constant need to develop cheaper, simpler, and more efficient methodologies to exploit the information available. The aim of the present work was to develop a soft sensor with the capacity to perform reliable substrate predictions and control in microbial cultures of the fed-batch type, using mainly microbial growth data. This objective was achieved using dielectric spectroscopy technology for online monitoring of microbial growth and hybrid neural networks for online prediction of substrate concentration. The glucose estimator was integrated to a fuzzy logic controller to control the substrate concentration in a fed-batch experiment. Dielectric spectroscopy is a technology sensitive to the air volume fraction in the culture media and the turbulence generated by the agitation; however, the introduction of a polynomial function for the calibration of the permittivity signal allowed biomass estimations with an approximation error of 2%. The methodology presented in this work was successfully implemented for the glucose prediction and control of a fed-batch culture of Bacillus thuringiensis with an approximation error of 6%.

Application of Kalman Filter Recurrent Neural Network for Identification and Control of Hydrocarbon Biodegradation Bioprocess

Biological treatment is attractive as a potentially low-cost technology, which converts toxic organic contaminants into CO2 and biomass. Since the 70’s, this technology has been applied for the hydrocarbon degradation, and today, it is considered as the best alternative to clean up polluted soils. For this bioprocess, one challenge is to provide enough O2 and nutrients to enable rapid conversion of contaminants by either indigenous microorganisms or inoculated species. Another challenge is to achieve efficient contact between the active micro-organisms and the contaminants, which may represented a problem with in-situ treatment. An attractive alternative to overcome this problem is to apply a biological treatment in slurry phase using Horizontal Rotating Drum (HRD). Nowadays, semi empirical HRD models, based on the Monod equation, have been developed to predict micro-organism growth as a function of available contaminants concentration. However, as the application of such models requires experimental work for calculating the kinetics parameters involved, so an alternative modeling technique is required. The Kalman Filter Recurrent Neural Network Model (KF RNNM) offers many advantages as the possibility to approximate complex non linear high order multivariable processes, as the biodegradation process is. The KF RNNM has been applied for measurement data filtering and parameters plus state estimation of hydrocarbons biodegradation process, contained in polluted slurry, treated in a rotating bioreactor. Then the KF RNNM is simplified and used to design a Sliding Mode Control (SMC) of two-input two-output high order nonlinear plant. The KF RNNM learning algorithm is the dynamic Backpropagation one (BP). The graphical simulation results of the system approximation, and indirect adaptive neural control, exhibited a good convergence, and precise reference tracking.

Hybrid Recurrent Neural Network for Nonlinear Hybrid Dynamical Systems Identification

This paper is devoted to the development of a Neural Network Hybrid Identification Framework for unknown Nonlinear Hybrid Dynamical Systems. The proposal is based in the well known Recurrent Trainable Neural Networks Identifiers. In a first instance, the unknown hybrid system is considered like a black-box where by using only hybrid input-output data an approximated model is found. In a second instance, by considering that the hybrid output of the unknown hybrid system is triggered by a defined set of hypersurfaces we extent the approach identification by introducing a Hybrid Recurrent Trainable Neural Network Identifier. The effectiveness of the proposed approach is shown using a commutable pendulum example.

A Recurrent Neural Network for Controlling a Fed-Batch Fermentation of B. t.

The paper proposed to use a new Recurrent Neural Network Model (RNNM) to stabilize fermentation process of Bacillus thuringiensis from fermentation kinetic data. The multi-input multi-output RNNM proposed, have ten inputs, six outputs, sixteen neurones in the hidden layer, and also global and local feedbacks. The weight update learning algorithm designed, is a version of the well known backpropagation through time algorithm, directed to the RNNM learning. The approximation error for the last epoch of learning is about 2% and the total time of learning is 201 epochs, where the size of epoch is 115 iterations.