TSINR: Capturing Temporal Continuity via Implicit Neural Representations for Time Series Anomaly Detection

TL;DR

The paper tackles unsupervised time series anomaly detection when training data may contain unlabeled anomalies. It proposes TSINR, which uses implicit neural representations with spectral bias and a transformer to predict INR parameters, incorporating a tripartite decomposition into trend, seasonal, and residual components. A frozen pre-trained LLM encoder augments anomaly signals in both time and channel dimensions, enabling a simple reconstruction-error-based criterion to identify anomalies. Empirical results across seven multivariate and one univariate benchmarks show TSINR outperforms state-of-the-art reconstruction-based methods, with ablations validating each component and visual analyses illustrating its sensitivity to discontinuous anomalies.

Abstract

Time series anomaly detection aims to identify unusual patterns in data or deviations from systems' expected behavior. The reconstruction-based methods are the mainstream in this task, which learn point-wise representation via unsupervised learning. However, the unlabeled anomaly points in training data may cause these reconstruction-based methods to learn and reconstruct anomalous data, resulting in the challenge of capturing normal patterns. In this paper, we propose a time series anomaly detection method based on implicit neural representation (INR) reconstruction, named TSINR, to address this challenge. Due to the property of spectral bias, TSINR enables prioritizing low-frequency signals and exhibiting poorer performance on high-frequency abnormal data. Specifically, we adopt INR to parameterize time series data as a continuous function and employ a transformer-based architecture to predict the INR of given data. As a result, the proposed TSINR method achieves the advantage of capturing the temporal continuity and thus is more sensitive to discontinuous anomaly data. In addition, we further design a novel form of INR continuous function to learn inter- and intra-channel information, and leverage a pre-trained large language model to amplify the intense fluctuations in anomalies. Extensive experiments demonstrate that TSINR achieves superior overall performance on both univariate and multivariate time series anomaly detection benchmarks compared to other state-of-the-art reconstruction-based methods. Our codes are available.

Figures (9)

• Figure 1: (a) The diagram of INR for time series data. (b) The spectral bias property of INR to prioritize the low-frequency signals is advantageous for accomplishing time series data anomaly detection tasks.
• Figure 2: The overall workflow of the proposed TSINR method. The INR tokens predicted by the transformer encoder are the parameters of the INR continuous function. And the input of the INR continuous function is the timestamp $t$.
• Figure 3: A sample of the proposed group-based architecture. It contains 2 global layers and 2 group layers. In this case, each group corresponds to one variable.
• Figure 4: The visual analysis of the decomposition components on synthetic trend and seasonal datasets.
• Figure 5: The visual analysis of the non-spike anomaly segment in the real-world PSM dataset.