Robot Anomaly Detection Using a Boosted Time Series Prediction Model

Early robot anomalous behaviour detection is critical to avoid potential risks in robot working environments, enabling safer manipulation operations. Data-driven solutions have achieved prominent progress due to the advance of deep neural networks. In this paper, we propose a boosted time series prediction model based on dilated convolutional neural network (DCNN) and long short-term memory (LSTM) to detect anomalous behaviours during robot manipulation operations. The proposed method utilizes the merits of DCNN and LSTM by a self-attention mechanism to enhance the ability to identify local and global dependencies across time series. Even the magnitude of the anomalous signals is similar to the normal signals, the proposed method is still able to identify the anomalies by learning the local and global dependencies. We have conducted experiments to demonstrate the proposed boosted model for robot anomaly detection.