Deep Convolutional Clustering-Based Time Series Anomaly Detection.

Area of Science:

• Industrial Process Monitoring
• Machine Learning
• Deep Learning

Background:

• Anomaly detection is crucial for industrial processes.
• Unlabeled data presents challenges for traditional methods.
• Autoencoder architectures are effective for dimensionality reduction and reconstruction.

Purpose of the Study:

• To develop a novel anomaly detection system for industrial processes using unlabeled data.
• To enhance autoencoder performance by modifying its latent space.
• To validate the approach on the Tennessee Eastman benchmark dataset.

Main Methods:

• A 1D-convolutional neural network-based deep autoencoder architecture was employed.
• The autoencoder's latent space was split into discriminative and reconstructive features.
• An auxiliary loss function using k-means clustering and a Top-K objective was introduced for the discriminative features.

Main Results:

• The proposed method demonstrated improved performance in anomaly detection and classification tasks.
• Ablation studies confirmed the effectiveness of the latent space separation and auxiliary loss.
• The approach showed potential for enhancing downstream tasks compared to standard autoencoders.

Conclusions:

• The novel autoencoder architecture with a split latent space offers a promising approach for anomaly detection in industrial settings.
• The method effectively utilizes unlabeled data and improves upon standard autoencoder techniques.
• Further applications in industrial process monitoring and control are suggested.