CASSAD: Chroma-Augmented Semi-Supervised Anomaly Detection for Conveyor Belt Idlers

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Conveyor idlers are critical for industrial efficiency, but their maintenance is challenging.
• Traditional fault detection methods require extensive labelled data, which is difficult to obtain in industrial settings.
• Rarity of faults and labor-intensive labeling hinder supervised learning approaches.

Purpose of the Study:

• To propose a novel semi-supervised anomaly detection method for conveyor idler fault detection.
• To address the challenge of limited labelled data in industrial fault diagnosis.
• To develop a robust system for identifying idler anomalies with minimal labeled examples.

Main Methods:

• Developed the Chroma-Augmented Semi-Supervised Anomaly Detection (CASSAD) method, centered on One-Class SVM (OC-SVM).
• Extracted chroma features (CENS, CQT, chroma STFT) from idler sounds, reducing complexity using mean and standard deviation.
• Augmented the dataset with additive white Gaussian noise (AWGN) and compared performance against Local Outlier Factor (LOF) and Isolation Forest (iForest).

Main Results:

• CASSAD achieved a 96% Area Under the Curve (AUC) and 91% accuracy on an industrial idler sound dataset.
• The proposed method significantly outperformed a baseline autoencoder and other traditional anomaly detection models.
• Demonstrated robust anomaly detection capabilities even with minimal labelled data.

Conclusions:

• CASSAD offers a practical and effective solution for conveyor idler fault detection in industries with limited labelled datasets.
• The method's reliance on semi-supervised learning makes it suitable for real-world industrial applications.
• Highlights the potential of sound analysis combined with advanced machine learning for predictive maintenance.