Accurate network traffic prediction enables network operators to formulate effective resource allocation strategies. However, traditional prediction methods often fail to capture and process the complex nonlinear relationships within network traffic data, resulting in limited prediction accuracy. To address this challenge, this paper proposes a novel network traffic prediction method, GCCTNet. This method combines data decomposition with a hybrid prediction model to efficiently extract multi-frequency features from network traffic data and capture both local and global characteristics to enhance prediction accuracy. Specifically, GCCTNet employs Gaussian filters and the CEEMDAN method to decompose long-term trends and short-term fluctuations across frequencies in the data. Based on this decomposition, a multi-branch prediction model integrating Convolutional Neural Networks (CNN) and Transformer is used to model and predict each component. Experimental results demonstrate that GCCTNet outperforms other baseline models in terms of prediction accuracy on the Wikipedia and university campus authentication interface datasets, validating its effectiveness across various network traffic scenarios.

Large language models (LLMs) perform exceptionally well in text generation tasks, but they often face issues of outdated or inaccurate content (i.e., "hallucination") when handling dynamic real-world data. Retrieval-augmented generation (RAG) alleviates this limitation by incorporating external knowledge; however, significant challenges remain when processing long or complex documents, particularly in capturing document-level semantic context.To address this, we propose a novel hierarchical summarization framework—Chain of Summarization (CoS). This framework combines two complementary strategies that significantly enhance model adaptability and computational efficiency: (1) a multi-clustering summarization strategy (e.g., Gaussian Mixture Models (GMM) and hierarchical clustering); and (2) a sequential summarization strategy combined with GMM. Experimental results show that the CoS framework, through its multi-clustering strategy, significantly outperforms single clustering methods, achieving a 10% performance improvement over baseline methods on the QASPER dataset. Additionally, the combination of sequential summarization and GMM reduced processing time by 50%, while maintaining high performance on the NarrativeQA dataset, with its ROUGE L score improving by approximately 15.7% compared to NaiveRAG. These results demonstrate the significant application potential of the CoS framework in RAG tasks, providing an efficient, scalable, and robust solution for handling long-document processing challenges.

When the rotating armature permanent magnet motor (RAPMM) works, there is an electromagnetic excitation force, whose frequency is generated by the number of rotor slots and the rotational frequency. It is observed that no frequency component is related to the number of magnetic poles. For the purpose of further reducing the vibration of the RAPMM shell, a non-magnetic metal ring is placed on the inner surface of the stator permanent magnet, and it is shown by analytical calculation and simulation of finite element that the non-magnetic metal ring is able to weaken the high-frequency time harmonic of the air gap magnetic field. As a result, it reduces the high-frequency component of the electromagnetic excitation force, and significantly suppresses the vibration of the corresponding frequency on the stator. It has been verified by experiments that the built-in stator copper ring is known as an effective measure in reducing the vibration of the rotating armature permanent magnet synchronous motor.