Development of machine-learning based anomaly detection system for manufacturing

- Autoencoder-LOF model -

Abstract

IoT is widely applied in the modern manufacturing system to acquire the time-series data from each individual equipment. Based on such time-series data, a variety of anomaly detection methods have been proposed for improvement in safety and productivity. However, conventional methods have some problems such as insufficient supervised anomaly data and need of extra procedure of feature extraction. As a solution, we developed a machine-learning based anomaly detection model which combines AE (Autoencoder) with LOF (Local Outlier Factor) in a simple way. This model requires the normal data only for AE at the training phase, to extract the features from the raw time-series data. At the detection phase, LOF is applied to calculate the deviation of input data in feature space. The model has been verified by use of the phase-shifted sine waves generated artificially. As a result, we conclude our model can discern the normal data from anomalies.