GAN-Based Data Augmentation Strategy for Sensor Anomaly Detection in Industrial Robots

In the current industry world, the industrial robot has emerged as a critical device to make the manufacturing process more efficient through automation. However, abnormal operation of industrial robots caused by sensor failures may interrupt the entire manufacturing process, thereby increasing production costs. In this paper, we first propose a domain-specific framework consisting of offline training and online inference to effectively detect anomalies in the scenario of industrial robotic sensors. In a nutshell, the framework can be identified in three folds: (1) the offline training obtaining historical data from the database to train the time series and anomaly detection models; (2) the online inference deploying the offline trained models for online anomaly detection in real-time; (3) the incremental learning updating the online model for a new type of anomalies. We then propose an improved Generative Adversarial Networks (GANs) named MSGAN with the adaptive update strategy mechanism based on WGAN-GP to generate fake anomaly samples, improving anomaly detection accuracy. Specifically, the Wasserstein distance with the gradient penalty is introduced to improve training stability and sample quality. Moreover, adapting the generator's complexity in robotic sensors, an adaptive update strategy based on the loss change ratio is adopted to speed up training convergence. Extensive experiments based on real robotic sensor datasets demonstrate the effectiveness of the proposed framework. Moreover, the results demonstrate that the detection accuracy can be improved by the synthetic samples based on the proposed MSGAN algorithm.