Fault detection algorithm in Auto Encoders (AE), through domain generalization strategies, using Artificial Intelligence (AI)

This work presents the application of domain generalization (DG) strategies in unsupervised deep learning methods, employing a Long Short-Term Memory (LSTM) and Autoencoder (AE) architecture for the detection of incipient faults in critical equipment at a mining desalination plant, within a predictive maintenance (PdM) framework. The case study is based on real operational data from 12 high-pressure and booster pumps over 2 to 3 years, using sensorbased monitoring of vibration, temperature, speed, and pressure. The initial phase involved basic exploratory analysis through correlation matrices, Principal Component Analysis (PCA), and histograms. Next, an unsupervised deep learning algorithm was implemented using only the LSTM-AE model, with parameter tuning focused on improving signal reconstruction quality. In the following stage, a hybrid model combining LSTM-AE and DG techniques was developed. The domain generalization strategies investigated include Domain-Adversarial Neural Networks (DANN), Domain Alignment (DA), and Learning Disentangled Representations (LDR). These approaches were assessed using performance metrics to compare the LSTM-AE baseline against the enhanced DG-integrated models. The results show a modest improvement when applying domain generalization. This enhancement is attributed to the model’s ability to capture domain-invariant features during training, leading to better generalization and improved anomaly detection performance compared to the standalone LSTM-AE approach.