TDMamba: Temporal Regression Model Based on Enhanced Mamba Architecture and Cross-Domain Feature Fusion for Industrial Predictive Maintenance

The widespread adoption of industrial big data in production systems has led to increasing demands for accurate predictive maintenance. Inaccurate forecasts can result in significant losses in industrial operations. To address this challenge, we propose TDMamba, a Mamba-based temporal regression model tailored for industrial applications. The model integrates temporal convolution and wavelet transform in dual path. The key focus is on the wavelet transform domain, which better captures local dependencies in time series through wavelet transformation, complemented by Mamba's ability to capture long-term dependencies. At the same time, a temporal convolution component is added, utilizing effective channel-wise feature interaction to enhance the extraction of complex features. We validate our approach on two real-world datasets, including a civil aviation engine lubrication system and a renewable energy forecasting challenge. Experimental results demonstrate that TDMamba consistently outperforms baseline approaches in both prediction accuracy and robustness.