Applied Research On Machine Learning Technologies For Spectrum Data

Applying machine learning technology in industrial environments can reduce production costs,increase production efficiency,and save human resources.It is a viable path for the development of advanced manufacturing.On the one hand,the development of machine learning technologies and the improvement of the computational performance make machine learning techniques more and more widely used in industrial environments;on the other hand,the inherit characteristics of data in industrial environments also bring challenges to the applications of existing machine learning technologies.The first problem is the diversity and complexity of data in the industrial environment.Applying machine learning techniques in such environment requires adequate data transformation and feature construction.In this paper,we focus on different types of spectrum data in industrial environments and do the following works:First,in the abnormal fiber detection task,we discuss the most basic frequency features,namely histogram based features.Three different image histogram features are applied to the abnormal fiber detection task of cotton image for comparison,and the differences and characteristics of the three features are analyzed.We select the optimal scheme and make improvements,and finally,good experiment results are obtained and a systematic solution is proposed.Second,in the task of using satellite remote sensing data for dynamic voyage planning,we discuss the construction and application of frequency features based on spectrum decomposition.For anisotropic matrices,we propose a multi-angle rotation approximate rank feature,which effectively characterizes the anisotropy of matrices.Its application shows that this method can accurately label the MODIS satellite remote sensing images and provide effective support for the subsequent voyage planning tasks.Third,in the task of using sensor signals for transmission equipment fault detection,we discuss the construction and application of frequency features based on information theory.For the high-speed transmission equipment vibration signal collected by sensors,we use information entropy and combine certain domain knowledge to transform the mechanical vibration spectrum,construct its information entropy features,and involve the ensemble learning techniques to verify the effectiveness of features and solve the task requirements.Finally,positive results are obtained on certain types of equipment.