Anomaly Detection and Classification for Process Surveillance of Industrial Robot-assisted Applications

Abstract

With the increasing automation of production lines in small and medium-sized companies, the dependency on intelligent and collaborative robots is growing. At the same time, small- scale production orders and a high degree of variance in the products lead to volatile robot programming with an ongoing lack of know-how. While there are promising approaches for volatile robot programming, the detection and handling of error states is still difficult. To avoid the manual programming of error detection and handling processes, there are model-based approaches which, though, require a high degree of expertise and are inflexible and time-consuming to implement in new applications. This thesis therefore presents the Multi-Scale Convolutional Variational Autoencoder as a data-based modeling approach for the detection of anomalies in a robotic system, using unsupervised deep learning methods. Furthermore, the model applicability for the classification of detected errors is examined. First, the convolutional neural network and the autoencoder model are effectively integrated to construct a two-dimensional convolutional variational autoencoder model. Secondly, a multi-scale sliding window algorithm is combined with a correlation computation for data enhancement. Thirdly, a reconstruction error of the input is calculated and compared with an error threshold to determine anomalous states in the input data. Finally, the latent representation of the input data is analyzed for clustering. The model is evaluated against a public robot dataset and the results are discussed with regards to its suitability for error detection and as a suitable basis for classification.