Welcome To The Intelligent Manufacturing and Data Science Laboratory!

Chinese| English
current location: News > Laboratory news > Content

Machine learning

【source: | Date:2017年11月21日 】

Main Research Points

Massive data including various types of streaming data, audio, video and document are increasingly produced by modern intelligent manufacturing, smart grid and biomedical systems every time. We research on the application of machine learning method in data science: extract data features from massive unstructed heterogeneous data to predict the prospective process evolution such as health monitoring and failure prediction of intelligent manufacturing equipment, smart grid and biomedical systems. More significantly, from searching for the correlation and causality between data of different layers and different measuring points, as well as mining the hidden nodes and detonating nodes in the process, we can reveal deeper mechanism and regularity.

 

Main Research Progresses

1. By using compressed sensing, eight genes that are possible to regulate circadian rhythms are mined from 22800 Arabidopsis genes. Later, two genes are proved that they have regulation function by the empirical study of Department of Botany in University of Cambridge. The results are published on Plant Cell, cited more than 40 times by Nature and its series, Cell and other journals with impact factor greater than 10, as well as a number of members of American Academy of Sciences.

2. By using sparse Bayesian learning method, we make a prediction of deformation of complex surfaces in multi-axis CNC machining, then transmit feedback to the spindle controller. Based on that, a predictive control method of deformation compensation and flutter suppression is designed, and it is successfully applied to thin-walled machining process in Wuxi turbine blade factory manufacturing aero-engine which the surface roughness of thin-walled workpieces is significantly decreasing.

3. We revealed the relevance and causality among different nodes in smart grid by introducing the compressed sensing method and the power flow formulation. The fault location and prediction of smart grid is achieved by data detection based on a small number of nodes.

4. By using the Powerball optimization method, the traditional Newton's method in machine learning process is speeded up by more than 1000 times. An optimization problem which should spend one week to be computated by a 4-core CPU, can be completed in a few minutes on a laptop by our optimization algorithm.

 

Application Prospects 

The failure prediction, life estimation, correlation and causality analysis between measurement points of intelligent manufacturing equipment. Intrusion locating, detection of hidden nodes and detonating node of Smart Grid. The mechanisms mining of gene and protein in gene regulation network.