Vibration-Based Fault Detection in a Wire Drawing Machine Using an LSTM Deep Learning Model
6th International Paris Applied Sciences Congress, Paris, Fransa, 12 - 17 Temmuz 2026, ss.67-75, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Paris
- Basıldığı Ülke: Fransa
- Sayfa Sayıları: ss.67-75
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Recep Tayyip Erdoğan Üniversitesi Adresli: Evet
Özet
Predictive maintenance has become an essential strategy for improving reliability, reducing unexpected downtime, and optimizing maintenance planning in modern manufacturing systems. In industrial production lines, early detection of abnormal machine behaviour plays a critical role in maintaining stable operation and preventing costly failures. Among various condition monitoring techniques, vibration analysis is widely used because it can capture mechanical changes associated with wear, friction, and structural degradation in rotating machinery. However, conventional vibration monitoring approaches often rely on handcrafted statistical indicators or fixed threshold values, which may not adequately capture complex temporal patterns in non-stationary industrial signals. This study proposes a vibration-based fault detection framework for an industrial wire drawing machine using a deep learning approach based on Long Short-Term Memory (LSTM) networks. The investigated wire drawing system operates under continuously varying mechanical loads and tribological conditions, which may lead to gradual degradation of components such as drawing dies, lubrication systems, and drive elements. To monitor these conditions, vibration signals were collected from eight production blocks of the line and analysed through a data-driven signal processing pipeline implemented in Python. The acquired vibration signals were segmented into fixed-length time windows and normalized to reduce the influence of operating condition variations. Instead of converting the signals into time–frequency images, the proposed approach directly utilizes the sequential structure of the vibration data. The segmented sequences were then used as input to an LSTM autoencoder network, which learns temporal dependencies and degradation-related patterns within the vibration signals and flags abnormal behaviour through reconstruction error. Experimental results obtained from approximately 47,700 sensor readings collected over a seven-week monitoring period demonstrate that the LSTM-based model can effectively detect abnormal vibration behaviour and distinguish normal from abnormal operating conditions, reaching an accuracy of 96.4% and an F1-score of 0.646 when evaluated against an independent statistical reference, since verified fault records were not available for this dataset. The results highlight the potential of time-series deep learning methods for developing practical redictive maintenance solutions in industrial wire drawing systems.