Enhancing time series forecasting under noisy and data-scarce conditions: simulation, denoising, and few-shot learning approaches

Jin, Zhongtian 2026. Enhancing time series forecasting under noisy and data-scarce conditions: simulation, denoising, and few-shot learning approaches. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (4MB) | Preview
	PDF (Cardiff University Electronic Publication Form) Restricted to Repository staff only Download (544kB)

Abstract

Accurate prediction of the remaining useful life (RUL) is a cornerstone of Prognostics and Health Management (PHM) for ensuring industrial reliability. This is particularly critical for rolling element bearings, which serve as essential yet failure-prone components in rotating machinery. While data-driven methods have shown promise, their practical deployment is often hindered by complex signal degradation, heavy industrial noise, and the acute scarcity of labelled run-to-failure data. To address these challenges, this thesis develops a modular, data-efficient framework designed to enhance forecasting performance by integrating advanced multi-scale feature simulation, adaptive denoising under varying signal-to-noise ratios, and frequency-aware learning for data-scarce regimes. The research contributions are threefold. First, a dynamic attention-enhanced feature learning approach is proposed to disentangle latent degradation signatures from complex vibration signals. By leveraging Continuous Wavelet Transform (CWT) for time-frequency manifold construction and a transformer-based attention mechanism, the model effectively captures the multi-scale spatial patterns and long-range temporal dependencies inherent in degradation processes. Second, to mitigate the impact of heterogeneous noise, a noise-conditioned generative and predictive architecture is developed. This involves a Noise-Conditioned Convolutional Denoising Autoencoder (NC-CDAE) that adaptively suppresses diverse interference through conditional modulation, integrated with a Temporal Attention Transformer (TAT) for robust trend modelling. Furthermore, a Conditional Generative Adversarial Network (cGAN) is used to generate degradation signals under different noise conditions, providing additional controlled data for evaluating forecasting performance across varying noise regimes. Third, to tackle the challenge of extreme sample scarcity, a Frequency-Aware Few-Shot Forecasting method is introduced. This approach integrates Wavelet Packet Decomposition (WPD) as a frequency-structured inductive bias with a Cross-Scale Attention Transformer. By decomposing vibration signals into multiple frequency components and modelling their cross-scale interactions, the method constrains the representation space and supports RUL forecasting when only a limited number of training trajectories are available. The proposed framework is validated through a combination of synthetic datasets and industrial benchmarks (e.g., XJTU-SY and PRONOSTIA), providing a comprehensive assessment across various degradation scenarios. The results demonstrate improved RUL prediction performance under the noisy and data-constrained conditions considered in this study. The proposed methods therefore provide a systematic approach for investigating bearing RUL forecasting across different data conditions and offer potential support for predictive maintenance applications.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Engineering
Uncontrolled Keywords:	1. Remaining Useful Life (RUL) Prediction 2. Predictive Maintenance 3. Bearing Prognostics 4. Time Series Forecasting 5. Noise-Conditioned Denoising 6. Deep learning
Date of First Compliant Deposit:	4 September 2026
Last Modified:	07 Sep 2026 09:01
URI:	https://orca.cardiff.ac.uk/id/eprint/189336

Actions (repository staff only)

Edit Item