MFGAN: Multimodal Fusion for Industrial Anomaly Detection Using Attention-Based Autoencoder and Generative...

Area of Science:

• Industrial IoT
• Machine Learning
• Sensor Data Fusion

Background:

• Industrial operations rely on anomaly detection for safety and efficiency.
• Increasingly complex and multimodal sensor data from IoT devices challenges traditional single-source methods.
• Existing anomaly detection techniques struggle to leverage diverse industrial data streams effectively.

Purpose of the Study:

• To develop an advanced anomaly detection model for industrial environments utilizing multimodal temporal data.
• To effectively capture and fuse information from diverse sensor sources for improved anomaly identification.
• To address the limitations of single-source anomaly detection in complex industrial settings.

Main Methods:

• Proposed a novel model integrating an attention-based autoencoder (AAE) and a generative adversarial network (GAN).
• The AAE captures time-series dependencies and features within individual data modalities.
• GAN introduces adversarial regularization to enhance the reconstruction of normal time-series data.

Main Results:

• Extensive experiments conducted on real industrial data, including distributed control system (DCS) measurements and acoustic signals.
• The proposed model demonstrated superior performance compared to the state-of-the-art TimesNet.
• Achieved a 5.6% improvement in F1 score for anomaly detection.

Conclusions:

• The integrated AAE-GAN model effectively fuses multimodal temporal data for robust industrial anomaly detection.
• The model's ability to capture complex data dependencies and enhance reconstruction significantly improves detection accuracy.
• This approach offers a promising solution for enhancing the safety and efficiency of industrial operations through advanced anomaly detection.