Sensor-Based Fault Diagnosis and Prognosis of Neurophysiological States: A Transformer Autoencoder Approach to EEG...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Neurophysiological state monitoring is crucial for understanding and managing neurological conditions.
• Existing methods may lack the granularity to capture dynamic changes across therapeutic interventions.
• Sensor-based condition monitoring principles offer a novel approach to analyzing complex biological signals like EEG.

Purpose of the Study:

• To develop and evaluate a sensor-based condition monitoring framework for diagnosing and prognosing neurophysiological states using electroencephalographic (EEG) signals.
• To compare the efficacy of a Variational Autoencoder (VAE) against a Transformer-based Autoencoder in modeling EEG dynamics across therapeutic phases.
• To assess the framework's ability to identify deviations from stable states and track therapeutic stage progression.

Main Methods:

• Utilized a deep learning architecture comparing a VAE and a Transformer Autoencoder for latent representation modeling of EEG signals.
• Analyzed EEG data across three therapeutic phases: pre-intervention, during intervention, and post-intervention.
• Applied sensor-based fault diagnosis principles, treating state deviations as diagnostic indicators and phase transitions as progression markers.

Main Results:

• The Transformer Autoencoder demonstrated superior performance in capturing cross-band spectral relationships via its self-attention mechanism.
• This resulted in denser within-phase clusters and improved separation between pre-, during-, and post-intervention EEG states.
• Statistically significant effects were observed between pre- and during-intervention phases (ηpartial2=0.0388) and pre- and post-intervention phases (ηpartial2=0.0470), driven by specific brain rhythms.

Conclusions:

• The sensor-based latent state monitoring framework provides interpretable, data-driven insights for condition and phase transition assessment.
• The Transformer Autoencoder effectively models complex EEG dynamics, outperforming the VAE in distinguishing therapeutic stages.
• The framework shows potential applicability beyond clinical domains to industrial condition monitoring and fault diagnosis tasks, offering qualitative indicators.