Proposing New Data Whitening Method through Application of LSTM Method for Fault Condition Prediction in Bearing Data

In the context of big data, data preprocessing, especially data whitening, plays a pivotal role in improving the effectiveness of AI models. While traditional whitening methods suffer from severe computational bottlenecks, arising from the intractable inversion of large-scale covariance matrices, it is possible to overcome these limitations by employing iterative, low-rank methods for whitening. Iterative, low-rank methods for whitening overcome these limitations and have been shown to produce whitening of large-scale covariance matrices. This paper introduces a novel, computationally efficient data whitening framework that does not require direct matrix inversion but instead leverages the wide-sense stationarity (WSS) of large random data sets. The primary contributions are: (1) Maximum Likelihood Estimation (MLE) of WSS Gaussian processes for ease in statistical modeling, and (2) the introduction of an approximation method when the covariance matrix eigenvalues and eigenvectors are directly learned from the Power Spectral Density (PSD) and Discrete-Time Fourier Transform (DTFT), respectively. This preprocessed data is then passed into a proposed predictive modeling framework that consists of wavelet denoising, Particle Swarm Optimization (PSO) for optimal delay selection, and a Long Short-Term Memory (LSTM) neural network (with 400 hidden units and an Adam optimizer). The framework was validated using a real-world bearing vibration dataset. The results of the experiments show that the proposed preprocessing strategy, when used with the LSTM model, can predict bearing failure with nearly 90% accuracy and up to 120 minutes of advance notice. This work offers a scalable, low-complexity solution for real-time predictive maintenance in industrial big data applications that can reduce computational load without compromising the reliability of the diagnosis and early detection of faults.

# Computer software # Mining engineering. Metallurgy # data whitening # long short-term memory networks # fault condition prediction # rolling element bearings # big data preprocessing

Nasrin Rezaei, Andreas Rehkopf (2026). "Proposing New Data Whitening Method through Application of LSTM Method for Fault Condition Prediction in Bearing Data". Advances in Engineering and Intelligence Systems. https://doi.org/10.22034/aeis.2026.575941.1510