Paper Title
Anomaly Detection for Automotive Test Bench Engines Using Machine Learning Techniques

Abstract
Modern automotive engine test bench systems generate massive volumes of high-dimensional time-series data from hundreds of sensor channels, making automated anomaly detection crucial for ensuring engine quality and reliability. Traditional rule-based and statistical methods struggle to capture complex multivariate temporal patterns often producing excessive false alarms or missing subtle anomalies. This paper proposes a two-stage anomaly detection framework that combines Long Short-Term Memory (LSTM) autoencoders with K-Means clustering to address these challenges. The LSTM autoencoder learns compressed temporal representations of normal operational behavior from reference engines effectively filtering noise andcapturing sequential dependencies. K-Means clustering then identifies three distinct operational prototypes per channel, providing explicit reference points that capture the natural diversity of normal behavior across different load conditions and operational states. This hybrid approach balances the representational power of deep learning with the interpretability of clustering-based methods. The framework employs a comprehensive preprocessing pipeline incorporating cycle quality filtering and channel selection to ensure high-quality training data. Empirical evaluation on real engine test bench data demonstrates an average accuracy of 87.1% with 100% precision on evaluated channels indicating strong potential for reducing false alarms in engine hardware testing environment while maintaining computational efficiency suitable for large-scale industrial deployment. Keywords - Anomaly detection, LSTM autoencoder, K-Means clustering, automotive test bench, engine diagnostics, multivariate temporal patterns, reference-based detection, unsupervised learning