SteelNet: Multimodal Representation Learning for Industrial Process Optimization

ICLR 2026 Conference Submission21239 Authors

Keywords: Multimodal Representation Learning, Cross-Modal Alignment, Steel Rolling Mills, Deep Learning, Casual Inference

Abstract: Steel rolling mills must continuously monitor various sensors like vibration probes, thermocouples, and flow meters to ensure safe operations and maintain product quality. However, the harsh industrial condition sometimes lead to sensor failures, unclear signals, and incomplete data. Traditional monitoring systems often struggle in these environments, makes it challenging to detect early signs of problems and predict failures. To this end, we propose SteelNet: A Multimodal Representation Learning Framework designed for robust learning from various industrial sensor data. SteelNet incorporates the cross-modal alignment and modality dropout strategies that enable consistent representation learning even when modalities are partially missing. The core problem being solved is improving equipment availability and optimizing process parameters. This framework allows for the early detection of critical events by combining information from multiple sensors, and effectively handling missing or data, which are common in industrial environments. By improving the reliability of anomaly detection and predictive insights, SteelNet not only strengthens fault tolerance but it also support better decision making in the process optimization. Although developed for steel rolling mills, it's applicability extends to real-world scenarios and other industry setups.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 21239