Failure Detection in Sensors via Variational Autoencoders and Image-Based Feature Representation

Area of Science:

• Engineering
• Data Science
• Machine Learning

Background:

• Current sensor failure detection methods struggle with multiple failure modes and diverse data.
• Limitations in existing approaches can negatively impact decision-making and overall system performance.
• There is a need for more adaptable and robust sensor monitoring solutions.

Purpose of the Study:

• To develop a novel approach for detecting multiple sensor failures.
• To improve the resilience and adaptability of sensor monitoring systems.
• To enhance the interpretability of sensor failure detection.

Main Methods:

• Transforming sensor data into image-based feature representations (e.g., mean, variance, kurtosis, entropy, skewness, correlation).
• Utilizing a Convolutional Variational Autoencoder (CVAE) model trained on these image representations.
• Developing a Health Index (HI) based on reconstruction error and CVAE latent space for failure detection and visualization.

Main Results:

• The proposed approach demonstrated encouraging results in detecting diverse configurations of faulty sensors.
• A complementary HI and visualization tool derived from the CVAE latent space improved interpretability.
• The method was successfully illustrated in an aeronautical industrial case study with complex electromechanical system data.

Conclusions:

• The novel image-based feature representation and CVAE model effectively detect multiple sensor failures.
• The approach offers enhanced flexibility and resilience compared to existing methods.
• The technique shows significant potential for real-world applications, particularly in complex industrial systems.