Generality-aware self-supervised transformer for multivariate time series anomaly detection

Efficient identification of anomalies within multivariate time series data holds significant relevance in contemporary industrial settings. The challenge lies in swiftly and accurately pinpointing anomalous data points. This challenge is further compounded by factors such as the absence of labeled anomalies, data volatility, and the need for ultra-fast inference times. While previous approaches have introduced advanced deep learning models to address these challenges, comprehensive efforts to tackle all these issues simultaneously have been limited. Recent developments in unsupervised learning-based models have demonstrated remarkable performance. However, many of these models rely on reconstruction error as an anomaly score, making them sensitive to unseen normal data patterns. To address this limitation, we propose a novel framework, generality-aware self-supervised transformer for multivariate time series anomaly detection, which utilizes a transformer that effectively generalizes normal data patterns through self-knowledge distillation. Furthermore, we incorporate an auxiliary decoder to compute generality-based anomaly scores, thereby enhancing the differentiation between anomalous and normal data points in testing datasets. In our study, encompassing a diverse range of publicly available datasets and our own extracted data from linear motion (LM) guides and reducers built to model the vertical and rotational motions of robots, we establish the superior anomaly detection performance of our framework compared to existing state-of-the-art models. Notably, we verify that this improved performance is achieved while also considering time efficiency.