Bacterial Evolutionary Algorithm Based Autoencoder Architecture Search for Anomaly Detection

Hunor István Lukács, Tamás Fischl, János Botzheim

Abstract: This paper presents an architecture optimization approach, based on evolutionary algorithms, which helps in finding the optimal architecture depth, shortening the time that one has to invest in an uncertain number of test runs to find the right architecture. The provided algorithm is applied in the field of anomaly detection, which consists of searching for small amounts (ppm) of anomalies in large datasets using autoencoder models. Nowadays, as a result of the continuous digitization efforts visible in the industry, the data of semi-finished and finished products and production processes are becoming available in increasing quantities. A lot of data can be connected through databases, which offer the possibility of recognizing connections that have been hidden to this day. Quality assurance in industry also requires that deviations within limit values and specifications are investigated (outlier detection). The analyzed data is extracted from the production of a MEMS (Micro Electro Mechanical System) based inertial sensor used in the automotive industry, thus the paper reflects on real problems arising in industry, while the results of the suggested method are validated by domain experts. The presented method in this paper shows how the bacterial evolutionary algorithm combined with an abstraction, based on rational quadratic Bézier curves, can be used in the autoencoder architecture search, and thus be applied to anomaly detection challenges appearing in the industry.