Multimodal Time-Series Fusion Model and Reproducible Benchmark for Industrial Equipment Fault Prediction

This project was released by daliMa-hub on GitHub on May 23, 2026 (link: https://github.com/daliMa-hub/robust-industrial-timeseries-fusion). Its core is to provide a robust multimodal time-series fusion model and reproducible benchmark testing framework for industrial equipment fault prediction, aiming to address the limitations of traditional single-data-source methods and support predictive maintenance in smart manufacturing.

Predictive maintenance of industrial equipment is a core component of smart manufacturing, but traditional methods relying on single data sources struggle to fully capture equipment operating status. With the development of the Industrial Internet of Things (IIoT), multimodal data (vibration signals, temperature readings, current/voltage, etc.) are generated, but fusion faces four major challenges: data heterogeneity (differences in sampling frequency, dimension, noise), complex nonlinear correlations between modalities, scarcity of fault-labeled samples leading to overfitting, and industrial on-site environmental interference affecting data quality.

The project adopts multimodal fusion strategies: early fusion preserves the integrity of original information, cross-modal attention modules learn modality weights, and combines CNN and Transformer to capture local features and long-term dependencies. For robustness enhancement, it uses time-series data augmentation (random scaling, jittering, occlusion, etc.), adaptive filtering for denoising, and domain adaptation to support cross-equipment migration, improving the model's anti-interference ability.

The project provides a complete benchmark testing framework: standardized datasets based on real industrial scenarios, unified training/validation/test splits to ensure experimental comparability, support for multi-dimensional evaluation such as accuracy, recall, F1-score, AUC, etc., and also provides an ablation experiment interface to facilitate analysis of each component's contribution, helping with research reproduction and comparison.

Applied to predictive maintenance, it can reduce downtime, optimize maintenance costs, and extend equipment lifespan; it provides support for Industry 4.0 such as data-driven decision-making, edge deployment friendliness (lightweight model), interpretability (feature importance analysis), etc.; in practice, it lowers R&D thresholds, improves model reliability, promotes domain standardization, and fosters open-source collaboration.

Technical highlights include cross-modal feature learning modules (automatically discovering implicit modality correlations), lightweight design (adapting to edge devices), and continuous learning support (online updates to adapt to working condition changes); in the future, with the deepening of industrial digitalization, multimodal time-series fusion technology will be more widely applied, and the open-source implementation of this project provides a foundation for academia and industry.