Data-Driven Sparse Reconstruction Framework for Anomaly Detection in Smart Manufacturing Environment

Reliable anomaly detection of sensor measurements is essential for monitoring complex systems in smart manufacturing environments. However, conventional approaches often require predefined statistical assumptions or explicit knowledge of physical measurement models. Therefore, a data-driven sparse reconstruction framework (DSRF) is proposed for anomaly detection without requiring explicit noise distribution or separate identification of a physical measurement model. The proposed framework decomposes observation deviations into a reconstructed normal measurement component, a sparse anomaly residual, and a remaining reconstruction residual. The DSRF is evaluated using current and vibration measurements from robot welding systems. Normal measurements are divided into training, validation, and independent test subsets, while labeled abnormal samples are reserved for final evaluation. Training data are used to learn an effective sensing matrix and determine anomaly thresholds. At the 99th-percentile threshold, DSRF achieves 100% detection with no false negatives and accuracies of 99.76% and 98.38% for the current and vibration datasets, respectively. Comparative experiments with PCA reconstruction and an autoencoder show similarly high classification performance while DSRF only provides a larger threshold-normalized separation and explicitly estimates sparse anomaly residuals. The anomaly residual also provides frequency feature-level information for interpreting prominent abnormal deviations. These results demonstrate the feasibility of DSRF for anomaly detection and residual-based interpretation without explicit noise assumptions or separately identified physical system models.