Multi-mode non-Gaussian variational autoencoder network with missing sources for anomaly detection of complex...

Area of Science:

• Engineering
• Data Science
• Machine Learning

Background:

• Anomaly detection is critical for the safety of complex electromechanical systems.
• Intelligent, unsupervised methods are increasingly vital for analyzing large industrial datasets.
• Standard Variational Autoencoders (VAEs) struggle with non-Gaussian data common in electromechanical equipment.

Purpose of the Study:

• To develop an advanced anomaly detection method for non-Gaussian vibration signals.
• To address limitations of standard VAEs in detecting anomalies without prior fault samples or knowledge.
• To enhance the reliability and accuracy of anomaly detection in industrial equipment.

Main Methods:

• Proposed a Multi-Mode Non-Gaussian Variational Autoencoder (MNVAE).
• Encoder maps input to a Gaussian mixture distribution; Householder Flow enhances latent features.
• Decoder utilizes Weibull distribution for non-Gaussian signal reconstruction and anomaly discrimination.

Main Results:

• MNVAE achieved superior performance compared to six other methods across diverse datasets.
• Demonstrated robustness in detecting anomalies in unknown distribution vibration signals.
• Experimental validation confirmed the effectiveness of the proposed improvements.

Conclusions:

• The MNVAE offers a powerful solution for anomaly detection in non-Gaussian industrial data.
• The method excels in scenarios lacking fault samples or prior knowledge.
• MNVAE significantly advances the safety and monitoring capabilities for electromechanical equipment.