Anomaly Detection of Under-Over Current Faults in Magnetorheological Damper Suspensions using Variational Autoencoders

Detecting anomalies in complex sensor-based systems, such as Magnetorheological (MR) dampers, is critical for ensuring operational reliability and safety. Traditional anomaly detection methods often struggle with the high-dimensional, noisy nature of MR damper data, particularly when anomalies are induced through variable voltage inputs. In this study, we propose a novel hybrid approach using a Variational Autoencoder (VAE) with an integrated Multilayer Perceptron (MLP) classifier to directly classify anomalies within the VAE’s latent space. This end-to-end model jointly optimizes unsupervised representation learning and supervised anomaly classification, effectively balancing reconstruction, regularization, and classification objectives. We evaluate the model on MR damper data collected from position and force sensors, with anomalies induced by over- and under-voltage scenarios. Experimental results demonstrate that our integrated VAE-MLP framework outperforms traditional methods, such as standalone PCA, Isolation Forest, and conventional Autoencoders, in both classification accuracy and robustness to noise. Additionally, we show that the inclusion of an MLP during training enhances the interpretability of the VAE’s latent space, clustering anomalies distinctly from normal operational data. This approach not only achieves superior anomaly detection performance but also provides a promising framework for hybrid representation and classification in high-dimensional sensor data applications.