Contribution
Anomaly Detection in Ball Screw Drives Using Acceleration Data
* Presenting author
Abstract:
Currently, ball screw drives are mainly inspected through haptic and auditory assessments by trained specialists, evaluating up to 500 units per eight-hour shift. This approach maintains stable quality but is limited by human error, lacks comprehensive protocols, and poses challenges in knowledge transfer.To address these limitations, a test stand is being developed to replicate the manual inspection process. Acceleration sensors capture haptic data relevant to quality assessment, which is initially analyzed using classical engineering methods in the time and frequency domains. Subsequently, a binary classification of ball screw drive conditions is performed using a Support Vector Machine. For anomaly detection, a Long Short-Term Memory autoencoder, a neural network model suited for detecting and classifying anomalies in sequential or time-dependent data, is employed to link anomalies to specific damage patterns. Trained on flawless components, it learns to reconstruct typical patterns. The autoencoder reconstructs normal data, while anomalies result in elevated reconstruction errors, enabling a clear distinction between normal and deviant data, facilitating anomaly identification.This project aims to advance production engineering by shifting from parameter measurement to AI-based anomaly detection, digitally capturing experiential knowledge. The test stand shortens inspection time and establishes a foundation for standardized quality control in production.