Development of an Anomaly Detection Model for CNC Machining Using Attention-LSTM Autoencoder

In this study, an Attention-based Long Short-Term Memory (LSTM) Autoencoder model is proposed for semi-supervised anomaly detection in CNC turning processes. The model addresses challenges in highly imbalanced industrial environments where abnormal data account for less than 1% of total samples. Spindle current signals were collected from real machining operations, preprocessed using Root Mean Square transformation, and feature-extracted via the Time Series Feature Extraction Library (TSFEL). The data were reconstructed into sliding windows to preserve temporal dependencies. The proposed hybrid architecture combines an LSTM encoder-decoder, an Attention module, and a parallel classifier trained through a hybrid loss function integrating reconstruction and classification errors. Experimental results showed significant improvements over conventional Autoencoder and LSTM-Autoencoder models, achieving an accuracy and a recall of 0.99. Although the model exhibited near-perfect ROC AUC values, this was intentionally designed to prioritize recall—detecting even subtle tool anomalies over avoiding false alarms. The model effectively generalized across different tools without retraining, confirming its applicability as a unified anomaly detection module. These findings demonstrate strong potential for real-time predictive maintenance and cost-efficient smart manufacturing, especially in small and medium-sized enterprises.