Self-supervised deep representation learning adversarial autoencoder: Deep domain adaptation for data-based fault diagnostics.

Self-supervised deep representation learning adversarial autoencoder: Deep domain adaptation for data-based fault diagnostics.

Self-supervised deep representation learning adversarial autoencoder: Deep domain adaptation for data-based fault diagnostics. 150 150 UKAEA Opendata
UKAEA-RACE-PR(23)03

Self-supervised deep representation learning adversarial autoencoder: Deep domain adaptation for data-based fault diagnostics.

Calibration data for condition monitoring (CM), often originates from equipment running in different environment conditions, and operational settings. Due to the uneven distribution of the data, the performance of traditional machine learning approaches for CM can easily be skewed in favour of operating conditions with larger data distributions. This paper presents a novel unsupervised machine learning methodology for addressing the problem of domain adaptation for CM. A self-supervised deep representation learning adversarial autoencoder (SR-AAE) is proposed to model the latent space as a sum of two vectors: a categorical cluster identifier and a Gaussian distribution style vector. SR-AAE is regularized using a proposed self-supervision method which recycles the data samples back through the same network in order to strengthen the performance of the encoder model. Fault diagnostics is accomplished in two stages: do main adaptation by SR-AAE, and fault diagnostics by temporal variation shift monitoring of the flattened reconstruction error by principal component analysis (PCA). The proposed methodology is evaluated on FD004 turbofan engine degradation datasets from Commercial Modular AeroPropulsion System Simulation (C-MAPSS). The results demonstrate that the proposed methodology is able to learn clear disentangled representations of the operating conditions in the latent space, and the approach shows excellent results in the task of domain adaptation for eliminating bias in the reconstruction error estimation. The results show 99.524% accuracy in binary classification of healthy/faulty state, without any prior knowledge of the faulty state. Our work represents a novel approach towards fault diagnostics in cases where only data from the healthy state is present.

Collection:
Journals
Journal:
IEEE Transactions on Pattern Analysis and Machine Intelligence
Publisher:
IEEE