Anomaly detection based on time series data from industrial automatic sewing machines

Industrial automatic sewing machines are used for the production of technical textiles and safety critical work pieces such as parachutes or safety belts. Hence, the production does not tolerate sewing errors within the final product. Unfortunately, errors like skip stitches or illusive sewing may occur even on correctly adjusted production machines due to their complex kinematics. A recently developed sensor for such production systems is able to measure the lower thread rotation, a key parameter of the sewing process. This opens up the possibility to analyze the sewing quality on the basis of time series data whilst the work piece is still in production, thus being able to instantly detect anomalous sewing operations which leads to reduced cost for quality control and improves safety. In this research we analyze the suitability of existing Neural Network based Machine Learning algorithms to detect anomalies within the available process data. We give an overview regarding the Machine Learning pipeline, the tested Neural Network types and the experimental results. We show that several Neural Network architectures are able to detect anomalies within the sewing process.