Anomaly Detection in Multivariate Time Series Using Reconstruction Autoencoder with Mahalanobis Distance

Sammanfattning

In modern industrial automation, the continuous monitoring of high-dimensional sensor data is vital for maintaining operational reliability and preventing hardware failure. Although deep learning architectures such as the standard reconstruction autoencoder have emerged as prominent unsupervised anomaly detection methods, they suffer from a fundamental limitation arising from the manifold hypothesis: these models frequently reconstruct samples that lie close to the learned data manifold, enabling subtle anomalies to evade detection. To address this limitation, this study evaluates a hybrid anomaly detection framework in which the conventional reconstruction error is replaced by the Mahalanobis distance computed within the latent space of the autoencoder. The model was trained on an industrial data set spanning three months of runtime of a rotary encoder. Experimental evaluation reveals a clear performance trade-off between the two approaches: the reconstruction autoencoder achieves higher detection rates on isolated point anomalies, whereas the Mahalanobis autoencoder is the stronger detector for complex, correlation-breaking collective anomalies. The study further evaluates whether autoencoder-based models provide meaningfully higher detection performance than a static per-channel threshold baseline. The autoencoder variants substantially outperform this baseline on collective anomalies, whereas on point anomalies, the baseline achieves comparable results at large magnitudes. These findings suggest that the Mahalanobis autoencoder constitutes a more robust alternative for industrial monitoring applications in which subtle, multivariate anomaly patterns are of primary concern. Furthermore, an ensemble model combining both scoring functions achieves the most balanced performance across anomaly types, indicating that the two approaches are complementary and that their combination represents the most viable candidate for practical deployment.