Vibration-Based Failure Prediction in Machinery: Novel Application of Time Series Decomposition and Neural Networks

Predictive maintenance (PdM) is critical in reducing downtime and minimizing operational costs in industrial settings. This study introduces a novel framework for vibration-based failure prediction by combining Seasonal and Trend Decomposition using LOESS (STL) and deep learning models, specifically Long Short-Term Memory (LSTM) networks. The framework processes multivariate vibration data collected from two-roller mills in a tire manufacturing facility. STL is employed to decompose the data into trend, seasonal, and residual components, enabling noise reduction and enhanced feature extraction. These components serve as inputs to LSTM models for predicting machine condition scores. A comprehensive analysis compared five deep learning architectures: LSTM, Bidirectional LSTM, Gated Recurrent Unit (GRU), Deeper LSTM, and Conv1D LSTM. The GRU and Conv1D LSTM demonstrated superior performance across multiple vibration axes, achieving low errors and high R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values. Results highlight the effectiveness of STL in preparing vibration data for sequential learning and the robust predictive capabilities of deep learning models. This study provides a scalable approach to PdM, enhancing reliability and efficiency in industrial machinery maintenance. Future work includes extending the framework to other machinery types and real-time monitoring systems.