Big data cannot effectively extract features or perform suitable intelligent analysis in a complex, multi-source, heterogeneous, and multi-dimensional setting, surpassed by dimensional redundancy of data, the lack of correlation between features and lack of ability to generalize their models. The paper presents a proposal of a multi-layered big data feature extraction and intelligent analysis platform that employs deep learning (DL). It has a deep adaptive feature mining of spatiotemporal data by building a feature extraction module relying on Convolutional Neural Network (CNN) and Self-Attention Mechanism (SAM). The overall workflow of the platform is as follows: (1)A data acquisition and preprocessing module that standardizes the multi-source information and filters noises; (2) A feature extraction layer that finds local and time-series features with joint modeling on CNN and Long Short-Term Memory (LSTM); (3) A feature fusion layer that optimizes the distribution of features by use of SAM; (4) An intelligent analysis layer that performs the task decision-making relying on ensemble regression and classification models. A 1.2TB traffic monitoring and energy consumption data were experimented. Results showed a 9.3% improvement in feature extraction accuracy, a 32.5% reduction in average inference time, and an overall analysis accuracy of 96.1%, outperforming existing platforms. The findings demonstrate that this platform possesses strong generalization and real-time analysis capabilities in high-dimensional and complex scenarios.
Wenjing Kang (Thu,) studied this question.