Abstract: Neural networks based on large-scale time series can accurately capture cross scale spatiotemporal dependency features and achieve adaptive memory decay, thereby efficiently completing real-time fusion, semantic compression, and trend prediction of heterogeneous sensor data streams, providing algorithm support for fast decision-making on the edge side. In response to the ultra long, ultra wide, and distributed heterogeneous time series data generated by modern infrastructure, recent research has mainly focused on three cutting-edge directions:efficient attention mechanisms and frequency domain enhanced Transformer models, state space models and their structured variants, and Transformer architectures based on local partitioning. The related achievements are widely applied in typical intelligent perception scenarios such as smart manufacturing, smart transportation, and wearable health monitoring. Specifically, this article first proposes a unified evaluation framework that systematically outlines key elements such as data types and performance indicators. Secondly, review and analyze the evolution of neural networks based on large-scale time series, and delve into the theoretical foundations of each method. Finally, an analysis is conducted from multiple dimensions such as model architecture complexity, memory capacity, and inference throughput, providing a general methodology and engineering reference for cross domain research to assist in algorithm selection and hardware implementation of intelligent perception systems and their upper level decision-making.