Data Fusion and Artificial Neural Networks for Biomass Instrumentation

The ability of Artificial Neural Networks (ANN) to learn from experience rather than from mechanistic descriptions is making them the preferred choice to model processes with intricate variable interrelations. Some of these processes can be found in the area of biotechnology. Of particular interest is the robust estimation of biomass in the production of an antibiotic. A desirable feature is a way to integrate several measurements into a single robust estimator. Several feed-forward neural networks have been chosen to do the estimations using the Backpropagation/Levenberg-Marquardt learning algorithm. A two-stage estimation system is proposed. Several experiments have been carried out to test the generalisation capabilities and performance of the system. Tests were also carried out in the presence of noise, positive drift, negative drift and sudden failure of certain sensors. The results show the estimations are almost insensitive to the perturbations applied. In this work we present the proposed estimation system, the results of the system when tested on new fermentor data and the results when the system is subjected to simulated perturbations. Introduction. Within the monitoring and control tasks required to optimise fermentor operation, on-line monitoring of all variables would be the best solution, as off-line methods mean loss of information density, delay in getting results and normally require greater human effort. Two barriers to the availability of on-line measurements for all process variables include absence of suitable detection methods, and the uneconomically high cost of some of the well known methods [1]. Hence, some required measurements in the fermentation broth are determined by off-line analytical methods from samples taken manually, although on-line measurements would be desirable. One of the characteristics of biochemical processes is the complex interrelations among the variables of the process. This has lead researchers to create 'software sensors' whereby, by means of a computer program, variables are estimated from the information gathered by other measurements [2] [3] [4]. We are using this technique together with the fusion of several data sources to estimate the variable of interest. The fusion of data and estimation with the above mentioned techniques has been successfully accomplished by Leal et al [4]. Our aim is to estimate biomass in the given fermentation process using ANN and sensor datafusion and thus bridge the gap left due to the lack of on-line measurements of this variable. In this work we present further proof of the robustness of the system applying simulated perturbations to real data. Using ANN and data fusion we try to acquire a general model of a type of fermentation process and one that will hardly fail with instrumentation errors. The fusion of data from different sensors, be it redundant or complementary, will add new valuable information that would otherwise be unavailable. The need of data fusion arises because normally the information gathered is incomplete, uncertain, prone to failure or imprecise [5]. The information is generally modelled and integrated using several possible methods one of which is the ANN approach. The data is put through a preprocessing stage prior to its application to the ANN in order to facilitate convergence. The details of the pre-processing of the data, the design of the ANN and the first tests estimating biomass will be left for the reader to consult in our last publication [4]. We will now go further in our proposed fusing systems, the results and the results of the tests with simulated perturbations. Sensor data fusion for biomass estimation Most of the inferential estimation applications in the literature are of biomass and are drawn from the readings of carbon dioxide evolution rate (CER) and fermentation age. If the carbon dioxide 1 Departamento de Electrónica, Sistemas e Informática, ITESO, Periférico Sur 8585, Tlaquepaque, Jalisco, México 2 Manchester Biotechnology Centre, UMIST, Manchester M60 1QD, UK. 3 Department of Instrumentation and Analytical Science, Faraday Building, UMIST, Manchester M60 1QD, UK.