MAMBA-CROSS: Bidirectional State Space Models with Channel-Temporal Attention for Remaining Useful Life Prediction

Thatipally, Arun Reddy (2025) MAMBA-CROSS: Bidirectional State Space Models with Channel-Temporal Attention for Remaining Useful Life Prediction. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

Predictive maintenance has become essential for modern manufacturing systems to minimize downtime and optimize maintenance schedules. Current deep learning approaches, particularly Transformer-based models, achieve high accuracy but suffer from quadratic computational complexity O(n²), limiting their deployment for long sensor sequences and real-time applications. This paper proposed MAMBA-CROSS, a new architecture that integrates selective state space modeling with Channel-Temporal Attention, achieving linear complexity of O(n) with competitive accuracy.

The proposed approach overcomes the computation complexity of existing approaches using the following three novel components: bidirectional Mamba layers for effective sequence data modeling, Channel-Temporal Attention for modeling inter-sensor correlations, and a hybrid model combining local and global patterns. The proposed model is assessed on the NASA C-MAPSS Turbofan engine dataset, which is the benchmark dataset for the remaining useful life prediction problem.

MAMBA-CROSS performed with RMSE of 16.17 cycles and $R^2$ of 0.6868, showing improvement of 16.5% over state-of-the-art approaches. The model performed better than the baseline approaches on the four sub-data sets with differing complexity of operation, with predictions ranging within 10 cycles of the actual data 67% of the time. Most notably, the model's complexity allows it to handle large sequences, which cannot otherwise be handled by the quadratic complexity approach.

This is the first successful implementation of Mamba architectures for the problem of predictive maintenance, showing that state space models can efficiently trade off between accuracy and computations for the problem of prognostics. The results open new possibilities for implementation of real-time systems and edge computing that cannot be solved using the Transformer architecture.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Agarwal, Bharat UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	09 Sep 2026 10:32
Last Modified:	09 Sep 2026 10:32
URI:	https://norma.ncirl.ie/id/eprint/9921

Actions (login required)

View Item