Fractional-order neural network for detecting process deviations in optical fiber cable manufacturing

Area of Science:

• Industrial Manufacturing
• Data Science
• Materials Science

Background:

• Anomaly detection is crucial for cost reduction and quality improvement in industrial manufacturing.
• Fiber optic cable production is sensitive to small parameter deviations impacting optical properties.
• High-dimensional data and lack of labeled anomalies necessitate unsupervised learning approaches.

Purpose of the Study:

• To develop an advanced unsupervised learning model for anomaly detection in fiber optic cable manufacturing.
• To leverage fractional calculus within recurrent neural networks for improved temporal dependency modeling.
• To enhance the accuracy and capability of detecting subtle anomalies in complex manufacturing processes.

Main Methods:

• Proposed a Fractional Derivative-Long Short-Term Memory (FD-LSTM) network model.
• Implemented fractional order derivatives using the Grünwald-Letnikov method.
• Applied unsupervised learning for clustering and labeling production anomalies in high-dimensional data.

Main Results:

• The FD-LSTM model achieved 96.7% accuracy and a 0.93 F1-score on real production data.
• Achieved predictive performance above 95% across the evaluated dataset.
• Demonstrated improved separability of weak anomaly clusters and reduced misclassification compared to classical LSTM.

Conclusions:

• The FD-LSTM model effectively integrates fractional calculus into recurrent architectures for robust anomaly detection.
• Fractional order derivatives enhance the modeling of complex temporal dynamics in manufacturing processes.
• The proposed method offers a significant advancement for quality control in industrial settings.