Unlocking Long-Term Temporal Patterns: TCAE for Anomaly Detection in Multivariate Time Series Data

Anomaly detection in multivariate time series is extremely important in modern industrial systems to mitigate attacks and minimize system downtime. Anomalies are often seen as subtle deviations from established normal patterns. The challenge of learning extended temporal patterns in time series remains unresolved, hindering effective anomaly detection. To address the above challenge in this study, a novel approach, termed Temporal Convolutional Auto Encoder (TCAE), is introduced. TCAE utilizes Temporal Convolution Networks (TCN) and employs casual convolutions and dilations to effectively simulate long-term dependency in sequential data, taking advantage of its temporality and large fields. The autoencoder is trained on normal operations to learn the temporal dependencies present in the input time series. Two anomaly detection strategies employing the Local Outlier Factor (LOF) and thresholding are investigated. A supervised grid search technique is employed to determine the threshold, optimizing the model's performance. The thresholding technique demonstrates a performance improvement of over 20% when compared to the average performance of other baseline models.