A Hybrid Generative and Transformer-Based Framework for Anomaly Detection in Industrial Sensor Time-Series for Predictive Maintenance

The hybrid anomaly detection framework presented in this paper is a development in the field of predictive maintenance in an industrial environment utilizing multivariate time series data gathered from multiple sensors. An anomaly detection framework solves the anomaly detection problems found in predictive maintenance applications where labeled data may not be available or is limited. In order to solve these problems, the development of a hybrid framework by combining Statistics, Attention-based models, and Deep Generative Models to learn both the distributional and temporal characteristics of sensors such as vibration and electrical current. More specifically, a state-of-the-art GAN was designed to estimate the normal distribution of vibration and electrical current data. This estimated normal distribution can be used to calculate an anomaly score through reconstruction error, Mahalanobis distance, and isolation tree reconstruction error using specific data from the GAN model. Anomaly detection was performed on long-range temporal dependencies through self-attention mechanisms using the Anomaly Transformer Architecture. Additional multi-scale anomaly patterns in industrial datasets are evaluated using the MSPG-SEN framework. The framework was evaluated on a total of four datasets (two labeled benchmark datasets: CWRU Bearing, Machine Failure, and two unlabeled real-world crane datasets). The results of the experiment indicate that the proposed framework achieves over a 98% true positive rate on the labeled benchmark datasets. The results from the unlabeled crane datasets indicate that different anomaly detection models produce partially overlapping but distinct anomaly patterns indicating different modeling paradigms demonstrate complementary aspects of anomalous behavior. These findings highlight that combining generative, attention-based and statistical methods increases the robustness of an anomaly detection system and can provide a wider array of detection patterns. The proposed hybrid method, therefore, has proven to be a useful system for predictive maintenance in real world applications, especially when there is limited ground truth or when ground truth is not available.