Federated Transformer Graphs for Secure Multisite Predictive Maintenance Across Industrial Sensor Networks
Аннотация
This study introduces Federated Transformer Graphs (FTG), a framework for multisite predictive maintenance that protects user privacy and combines graph relational learning with temporal attention, enabled by industrial sensor networks. In addition, each node must know how to use transformer attention to duplicate long-range deterioration, how to document interactions between sensors using graph convolutions, and how to integrate multi-modal sensors in the local area. Even if there isn't a central place to store raw data, a fusion block can nevertheless make predictive representations. A secure federated layer might use homomorphic encryption and differential privacy to aggregate changes made to many sites. This lets people work together while keeping their information secret. FTG can find faults and estimate RUL at once with multi-location benchmarks. With good time-series, graph, and table baselines, such an outcome is possible. FTG decreases the Miss@1% FA to 20.3% at the operational low-false-alarm threshold, which is a 9.6 percentage point improvement over the next best. It also has an AUPRC value of 0.873 and a classification AUROC rating of 0.962. The MAE and RMSE drop to 5.3 and 8.3, respectively, because of FTG. It makes the probability more accurate (Pinball <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\tau=0.9= 4.2$</tex>) and makes sure that uncertainty is distributed out evenly (Coverage <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\approx 0.93$</tex> with small gaps). This results in an approach that is both effective and secure for Industry 4.0. Inference operates on a millisecond scale; therefore, the procedure is still accurate even if the distribution changes.