Ensemble-based adaptive soft sensor for fault-tolerant biomass monitoring.

The accuracy and precision of soft sensors depend strongly on the reliability of underlying model inputs. These inputs (particularly readings of hardware sensors) are frequently subject to faults. This study aims to develop an adaptive soft sensor capable of reliable and robust biomass concentration predictions in the presence of faulty model inputs for a Pichia pastoris bioprocess. Hence, three soft sensor submodels were developed based on three independent model inputs (base addition, CO2 production, and mid-infrared spectrum). An ensemble-based algorithm combined the submodels to form an ensemble model, that is, an adaptive soft sensor, to achieve fault-tolerant prediction. The algorithm's basic steps are as follows: the initial determination of submodel reliability is followed by selecting appropriate submodels to generate a reliable prediction via variance-based weighting of the submodels. The adaptive soft sensor demonstrated high robustness and accuracy in biomass prediction in the presence of multiple simulated sensor faults (RMSE = 0.43 g L-1) and multiple real sensor faults (RMSE = 0.70 g L-1).

Challenges in the Development of Soft Sensors for Bioprocesses: A Critical Review.

Among the greatest challenges in soft sensor development for bioprocesses are variable process lengths, multiple process phases, and erroneous model inputs due to sensor faults. This review article describes these three challenges and critically discusses the corresponding solution approaches from a data scientist's perspective. This main part of the article is preceded by an overview of the status quo in the development and application of soft sensors. The scope of this article is mainly the upstream part of bioprocesses, although the solution approaches are in most cases also applicable to the downstream part. Variable process lengths are accounted for by data synchronization techniques such as indicator variables, curve registration, and dynamic time warping. Multiple process phases are partitioned by trajectory or correlation-based phase detection, enabling phase-adaptive modeling. Sensor faults are detected by symptom signals, pattern recognition, or by changing contributions of the corresponding sensor to a process model. According to the current state of the literature, tolerance to sensor faults remains the greatest challenge in soft sensor development, especially in the presence of variable process lengths and multiple process phases.

Biomass soft sensor for a Pichia pastoris fed-batch process based on phase detection and hybrid modeling.

A common control strategy for the production of recombinant proteins in Pichia pastoris using the alcohol oxidase 1 (AOX1) promotor is to separate the bioprocess into two main phases: biomass generation on glycerol and protein production via methanol induction. This study reports the establishment of a soft sensor for the prediction of biomass concentration that adapts automatically to these distinct phases. A hybrid approach combining mechanistic (carbon balance) and data-driven modeling (multiple linear regression) is used for this purpose. The model parameters are dynamically adapted according to the current process phase using a multilevel phase detection algorithm. This algorithm is based on the online data of CO2 in the off-gas (absolute value and first derivative) and cumulative base feed. The evaluation of the model resulted in a mean relative prediction error of 5.52% and R² of .96 for the entire process. The resulting model was implemented as a soft sensor for the online monitoring of the P. pastoris bioprocess. The soft sensor can be used for quality control and as input to process control systems, for example, for methanol control.

Online sensor validation in sensor networks for bioprocess monitoring using swarm intelligence.

Sensor faults can impede the functionality of monitoring and control systems for bioprocesses. Hence, suitable systems need to be developed to validate the sensor readings prior to their use in monitoring and control systems. This study presents a novel approach for online validation of sensor readings. The basic idea is to compare the original sensor reading with predictions for this sensor reading based on the remaining sensor network's information. Deviations between original and predicted sensor readings are used to indicate sensor faults. Since especially batch processes show varying lengths and different phases (e.g., lag and exponential phase), prediction models that are dependent on process time are necessary. The binary particle swarm optimization algorithm is used to select the best prediction models for each time step. A regularization approach is utilized to avoid overfitting. Models with high complexity and prediction errors are penalized, resulting in optimal predictions for the sensor reading at each time step (5% mean relative prediction error). The sensor reliability is calculated by the Kullback-Leibler divergence between the distribution of model-based predictions and the distribution of a moving window of original sensor readings (moving window size = 10 readings). The developed system allows for the online detection of sensor faults. This is especially important when sensor data are used as input to soft sensors for critical quality attributes or the process control system. The proof-of-concept is exemplarily shown for a turbidity sensor that is used to monitor a Pichia pastoris-batch process.

Biomass estimation in Pichia pastoris cultures by combined single-wavelength fluorescence measurements.

In this work, the evolution of different biogenic fluorophores involved in the metabolism of Pichia pastoris was determined at four different single-wavelength pairs (excitation/emission) during batch culture in microwell plates and used for effective and reliable biomass estimation by means of chemometric tools. The chemometric tools for biomass estimation were multiple linear regression (MLR), partial least squares regression (PLSR), and principal component regression (PCR). Variable importance in the projection (VIP) scores were used to rate the importance of model input variables, indicating tryptophan as the most important variable for biomass estimation. A direct correlation between the single fluorescence signals of tryptophan and biomass was additionally set up. Results indicate a successful fitting of the MLR, PLSR, PCR, and direct tryptophan correlation models for the present case and confirm the relevance of biogenic fluorophores for bioprocess state variables monitoring. The root mean squared error of prediction (RMSEP) between the predicted and measured values for the validation batches was 0.017, 0.023, 0.025, and 0.049 g L(-1) dry cell weight for MLR, PLSR, PCR, and direct tryptophan correlation, respectively. The presented approach of indirectly measuring biomass based on combined single-wavelength fluorescence measurements can be used for the development of a low-cost alternative to 2D fluorescence spectroscopy. Biotechnol. Bioeng. 2016;113: 2394-2402. © 2016 Wiley Periodicals, Inc.