A federated learning and decision transformer-based multi-source data fusion architecture for real-time industrial scheduling dashboards

Abstract

With the rapid development of industrial IoT and intelligent scheduling systems, real-time fusion of multi-source heterogeneous data and decision optimization have become critical challenges. Traditional centralized data collection methods face privacy leakage and communication bottlenecks, while federated learning (FL) protects data privacy through distributed model training but has limited temporal decision-making capabilities in dynamic scheduling scenarios. This study combines federated learning with Decision Transformers (DT) to propose a multi-source data fusion architecture for real-time industrial scheduling dashboards, addressing privacy protection and temporal prediction challenges in cross-departmental, cross-regional data collaboration. The system employs a FedAvg-based federated learning framework for local training at 20 distributed nodes and global model aggregation, enhancing model generalization while safeguarding data privacy. Leveraging Decision Transformers’ strengths in modeling long-range temporal dependencies, it dynamically optimizes dispatch strategies via trajectory-based sequence learning. Experiments were deployed in a simulated industrial scheduling environment incorporating five heterogeneous data sources (device sensor data, historical scheduling records, environmental parameters, network status data, and user demand data) with 100,000 temporal samples. We compared the proposed method with traditional centralized LSTM, single federated learning, XGBoost, LightGBM, and modern tabular deep learning models. Results demonstrate a 12.7% improvement in scheduling accuracy over the traditional centralized LSTM model, a 63% reduction in cumulative parameter-transmission volume relative to the FedAvg-GRU baseline under the same client participation schedule, and an empirical privacy-attack success rate below 0.05%. Under dynamic load fluctuation scenarios, the single-step inference latency remains below 200 ms, whereas the average time required to complete an end-to-end scheduling cycle is 3.1 min; the system processes more than 2000 real-time data streams per second with an average CPU utilization of 51.6% and a peak memory usage of 8.9 GB. This research offers a novel solution for multi-source data fusion in industrial scheduling dashboards, balancing privacy security with decision efficiency, and validates the effectiveness of integrating federated learning with decision-making time-series models.