Process Monitoring of Batch Bioreactor Process Using Optimized Data Based Model

Monitoring is an important aspect in industrial processes to ensure process safety, and maintain high product quality. Two important steps of process monitoring are fault detection and diagnosis. Data-based fault detection techniques such as multivariate statistical method have been widely discussed in the literature, and in this work, we have used partial least square (PLS), an input-output model to analyze the batch data [1]. In batch processes, the data is first arranged in a multiway fashion according to batches for easier access to train and test the PLS model. Most of industrial data have noise, correlated variables, and unknown disturbances, and thus we have introduced a filter by using multiscale based wavelet functions to de-noise, de-correlate and normalize the dataset [2-3]. Orthogonal wavelet functions decompose the original data signal into multiscale approximate, and detailed scale and additional statistical filter is applied to remove unknown disturbances, and then the filtered signals are reconstructed to form global signals.

Most of the industrial process are nonlinear in nature and application of linear PLS will lead to mismatch between model and process data, hence in this work we have used kernel extension of PLS technique as our nominal data-based model. Kernel partial least square (KPLS) is first trained with normal operating historical data to compute all the training score vectors. KPLS model is applied on the global scale to the reconstructed dataset to generate model residue or error, and fault detection decision is made by performing statistical analysis on the residue. In literature conventional statistical test, and test have been studied but they have been inadequate to diagnose the fault, and show inferior performance while detecting multiple faults compared to advanced statistical test like generalized likelihood ratio test (GLRT) [4]. GLRT solves composite hypothesis test, and provides an ability to analyze the multivariate data in time and also along individual variables, and thus improving fault detection performance for detecting multiple faults. As soon as the fault is detected, it is necessary to find the variable that presents the major deviation from its expected value; therefore, a contribution plot with GLRT is performed for diagnosis.

To further improve performance of KPLS model, selection of kernel function and its parameter is optimized, and in this work we have used multi-objective genetic algorithm to optimize selection of kernel function and its parameter value by minimizing missed detection rate and false alarm rate. KPLS being an input-output model, can also be used as nonlinear regression tool to predict the product concentration from online sensor measurements and thus providing an additional option to monitor the concentrations that are not available by sensor measurement. The developed model is applied to batch bioreactor process; experimental data of batch production of beta-carotene via fermentation is used to describe glucose consumption, metabolic product formation and depletion, and the beta-carotene production in the Saccharomyces cerevisiae strain mutant SM14, with 20 g/l glucose as the carbon source. Experiments are performed to validate the results obtained from the model; the process monitoring results by our developed model shows superior performance compared to conventional fault detection techniques.

Keywords: KPLS, GLRT, wavelet function, fault detection, batch processes, bioreactor.

References:

1. MacGregor, J.F., and T. Kourti. “Statistical Process Control of Multivariate Processes.” Control Engineering Practice 3, no. 3 (March 1995): 403–14.
2. Bakshi, Bhavik R. “Multiscale PCA with Application to Multivariate Statistical Process Monitoring.” American Institute of Chemical Engineers. AIChE Journal 44, no. 7 (July 1998): 1596.
3. Botre, Chiranjivi, Majdi Mansouri, M. Nazmul Karim, Hazem Nounou, and Mohamed Nounou. “Multiscale PLS-Based GLRT for Fault Detection of Chemical Processes.” Journal of Loss Prevention in the Process Industries 46 (March 2017): 143–53
4. Mansouri, M., Nounou, M., Nounou, H., Karim, N., 2016. "Kernel pca-based glrt for nonlinear fault detection of chemical processes." Journal of Loss Prevention in the Process Industries 26 (1), 129–139.