Artificial neural nets(ANNs) were applied to nonlinear bioprocess systems to overcome limitations in the classical mathematical approach. The mathematical approach requires time-consuming and difficult optimization calculations to obtain a mathematical model for the process under consideration, which often covers only a limited range of operating conditions. As an alternative to this classical approach, ANNs give a simple black box model from the knowledge of inputs and outputs for a given system. ANNs have a good nonlinear interpolation capability and do not require any assumptions on the model structure.
A self-organizing feature map(SOFM), a type of ANN was used for modelling of batch yeast cultures. The model was constructed by training the neural net with experimental data of specific rates such as specific growth rate, specific substrate consumption rate, and specific net production rate of ethanol. Estimates of state variables such as cell, glucose, and ethanol concentrations were obtained from the neural net model and differential mass balance equations via integration. They were compared with the experimental data. The neural net model showed a good modeling accuracy and interpolation capability, and could model the process from minimal amount of knowledge on the process because it formed a state space model which did not depend on time and initial conditions. The ANN also had the capability of noise insensitive state estimation. By using the SOFM, a simpler and easier modeling would be possible because of its simple network structure and no necessity for data pretreatment. Its application can be extended to on-line estimation of state variables in fed-batch and continuous processes.
A gas sensor system for gas monitoring was constructed and a SOFM was used for the analysis of the sensor signals. An operating software for the system, which could control the experimental devices and could perform on-line data acquisition was also developed. Quartz crystal microbalance (QCM) was used as the transducer on which a sensing material, a pure lipid film or a mixture of lipids was deposited by using the Langmuir-Blodgett(LB) method. In the purpose of identifying pure gases, a 6-channel gas sensor system was constructed. The overall gas sensor system was automated by using computer interfaces and the operating software. For pure gases, the SOFM was trained by using the frequency changes from the 6 channels for a particular gas as the input pattern and the identification number for the gas as the target. For nine pure gases and a binary mixture of methanol and butanol, The trained SOFM showed a good gas identification capability, higher than 90%. The SOFM also showed a potential for identifying the composition of a binary mixture.
A simulation study was done for the possibility of the application of a multi-channel gas sensor with ANNs to identify gas mixtures. This study was to contribute to the development of on-line biochemical gas sensors for process monitoring that has not been developed yet. Two ANNs were tested to know whether the ANNs could recognize and quantify the concentration of each component. A simulated system was studied in which a SOFM and a backpropagation neural network(BPNN) were applied to the frequency data for a 2- and a 3-component gas mixtures, which were measured by a 3-channel gas sensor. The virtual 3-component mixture system generates frequency changes of a higher nonlinearity than the 2-component system, which were used as input patterns to the ANNs. The ANNs showed a possibility for quantification of gases for the 2-component model system despite the nonlinearity involved. For the 3-component model system, The SOFM showed a better potential than BPNN although its accuracy should be improved through further studies. For a system with more than 4 components, more work on physical arrangement of gas sensor system such as the selection of LB membranes with a specificity rather than on ANNs would be required.
