Assessment of a Hybrid HMM-GMM and LSTM Pipeline for State Identification and Predictive Maintenance in Shaft Alignment

Abstract

This thesis provides an implementation and evaluation of a predictive pipeline consisting of Long Short-Term Memory (LSTM) networks and Hidden Markov Model with Gaussian emissions (HMM-GMM) for predicting anomalies and operating states in shaft alignment. The study focuses on a motor-pump setup, utilizing the Easy-Laser XT770 laser measurement platform for data acquisition as well as external temperature readings. Monitoring shaft alignment is critical, as misalignment leads to increased energy consumption and accelerated wear on bearings and other mechanical components. Empirical data collected from a physical setup provided initial insights into real-world behavior of the system. However, the algorithms were development centered around simulated data. Consequently, a simulation environment was designed to emulate a variety of physical systems, providing diverse datasets for training and evaluation. In the proposed architecture, the LSTM model is employed for forecasting future alignment trends, while the HMM-GMM is utilized for state identification and anomaly detection. A marginalized variant of the HMM-GMM was also developed to perform these tasks without the continuous requirement of the laser measurement units. While the models yielded promising results in simulated environments, performance deviated when applied to empirical data, likely due to a model mismatch between the simulation and the physical system. The results suggest that the HMM-GMM can function as a preliminary framework for identifying states and anomalies. Although further refinement is needed before practical deployment, the study demonstrates the potential of using a single measurement session of healthy states to initialize the model for real-world applications.