Time-series Data Feature Search Based On CNN Algorithm Research And Implementation

Time series data reflects the changing process between time and transactions,and mining hidden patterns in time series data has become very important.The primary goal of this subject research is to develop a waveform matching algorithm to solve the curve search problem of locating a data segment in a large amount of time series data that matches a specific waveform.This thesis main research work is as follows:1)The CNN extraction of curve features is too random,and the curve targets to be searched are temporarily provided without any additional training samples.This article uses relevant time series datasets as training samples to solve the problem of excessively random curve feature extraction,and uses twin neural networks to solve the problem of few samples.2)To address the issue of poor curve matching performance.This article designs a curve feature matching algorithm,which explains the complete algorithm using mathematical principles from curve segmentation,data fitting,feature selection,feature recognition,and matching degree evaluation,and constructs a system model diagram of the algorithm in this article;Secondly,various coefficients in the algorithm that need to be verified through experiments were studied,such as weight values,range of feature segments,order of fitting values,etc;Finally,relevant experiments were set up to determine all undetermined problems in the algorithm one by one.In this thesis,the complete algorithm is used on a spacecraft thruster temperature data and ECG data set.Through experimental verification,The algorithm has better accuracy,but the search takes a long time.Using the UCR time series test set,the color histogram,the gray-scale co-occurrence matrix based on texture features and the CNN-based curve feature matching algorithm were compared and analyzed.The accuracy rate of the CNN-based curve feature matching algorithm was 95.3%,which was higher than The color histogram and grayscale cooccurrence matrix are 45.3% and 37% higher,respectively;The F1 measurement value is 95.3%,which is 28.7% and 32.9% higher than the color histogram and gray level co-occurrence matrix,respectively;In this thesis,a comprehensive evaluation index is defined for accuracy and efficiency,The comprehensive score of the CNN-based curve feature matching algorithm is0.2796,which is 0.0406 and 0.1036 higher than the color histogram and gray level co-occurrence matrix,respectively;The above results verify that the algorithm in this thesis is better than color histogram and texture based gray level co-occurrence matrix in curve matching.