With the widespread application of industrial Internet of Things technology in the state monitoring of moving machinery, the low efficiency of real-time fusion of multi-source heterogeneous sensor data and the imbalance of cross-node collaborative computing resource allocation have seriously restricted the accurate perception and rapid response capabilities of the health status of machinery under complex working conditions. To this end, this study proposes a hierarchical-dynamic-distributed collaborative processing method. First, a hierarchical data fusion architecture based on spatiotemporal feature encoding is constructed, and a feature weight dynamic allocation algorithm based on the attention mechanism is designed. Secondly, a distributed collaborative computing strategy based on edge computing is developed, and the computing tasks are dynamically divided in combination with the equipment state prediction model. Finally, a parameter sharing mechanism based on federated learning is designed to break the dependence on communication bandwidth. The experimental results show that the proposed method significantly improves the accuracy and real-time performance of data fusion under different data volume conditions. The processing delay of the dynamic weighted fusion method reaches 0.058 s when the data volume is large, and the accuracy of data fusion is improved by about 10%. In addition, the distributed collaborative computing strategy based on federated learning effectively reduces the communication overhead. These experimental results verify the effectiveness and feasibility of this research method in the industrial Internet of Things environment.