Title: Real-time outlier detection based on DetMCD
Authors: Iwein Vranckx - KU Leuven (Belgium) [presenting]
Bart de Ketelaere - Katholieke Universiteit Leuven (Belgium)
Mia Hubert - KU Leuven (Belgium)
Peter Rousseeuw - KU Leuven (Belgium)
Abstract: Modern, state-of-the-art sorting machines can use robust covariance-based classification methods in order to separate the regular samples from outliers (raisins versus glass). Compared to well-known machine learning methods like deep neural networks and SVMs, robust statistical classifiers offer comparable classification efficiencies, in addition to being highly resistant against trainings-set contamination at the same time. However, industrial machines generate several gigabytes of spectroscopic measurements in milliseconds, frequently pushing the boundaries of currently available computational power. Due to the time criticalness of industrial classification tasks in day-to-day operations, combined with the vast amount of spectroscopic data, high performance is essential. The presented research therefore focuses specifically on the computational improvement of the DetMCD algorithm: a highly robust and deterministic estimator of location and scatter. To illustrate the performance of our accelerated DetMCD estimator, the algorithm is applied to industrial spectroscopic datasets of various food related products. We demonstrate the corresponding machine efficiency improvements and highlight the improved classification training times.