Anomaly detection and removal strategies for in-line permittivity sensor signal used in bioprocesses

TL;DR

This paper presents a new method for detecting and removing signal anomalies in real-time bioprocess sensors, improving process control.

Contribution

A novel three-step algorithm for real-time anomaly detection and removal in permittivity sensor signals is proposed.

Findings

The method achieves an F1-score of 0.79 using a static threshold and rolling aggregate transformer.

The algorithm is computationally efficient and suitable for real-time applications.

Signal preprocessing and threshold-based detection improve anomaly handling in dynamic bioprocess signals.

Abstract

In-line sensors, which are crucial for real-time (bio-) process monitoring, can suffer from anomalies. These signal spikes and shifts compromise process control. Due to the dynamic and non-stationary nature of bioprocess signals, addressing these issues requires specialized preprocessing. However, existing anomaly detection methods often fail for real-time applications.

This study addresses a common yet critical issue: developing a robust and easy-to-implement algorithm for real-time anomaly detection and removal for in-line permittivity sensor measurement. Recombinant Pichia pastoris cultivations served as a case study. Trivial approaches, such as moving average filtering, do not adequately capture the complexity of the problem. However, our method provides a structured solution through three consecutive steps: 1) Signal preprocessing to reduce noise and eliminate context dependency; 2) Anomaly detection using threshold-based identification; 3) Validation and removal of identified anomalies.

We demonstrate that our approach effectively detects and removes anomalies by compensating signal shift value, while remaining computationally efficient and practical for real-time use. It achieves an F1-score of 0.79 with a static threshold of 1.06 pF/cm and a double rolling aggregate transformer using window sizes w1 = 1 and w2 = 15. This flexible and scalable algorithm has the potential to bridge a crucial gap in process real-time analytics and control.

Figures