Research And Application Of Deep Learning In Time Series Classification

Time series exist everywhere,analyzing it can help people discover the essence of things.The analysis of time series mainly includes two directions: prediction and classification.The current research focuses on prediction,and there is not much research on classification problems and usually stays on traditional methods.Traditional time series classification methods are based on the similarity of time domain,shape,and change.They are more sensitive to time phase drift and noise.In this article,deep learning methods are used to study the classification of time series.The main work and contributions are as follows:(1)For the classification of time series,this article designs two deep learning models: Time series classification model based on long short-term memory(LSTM),time series classification model based on positional encoding and multi-head attention.For the structure and key parameters of the designed models,this article evaluates in two actual scenarios of autistic(ASD)children's time series gesture classification and four-axis UAV flight attitude classification.The data set used in the evaluation is collected in real application scenarios.(2)In this article,the designed deep learning models are compared and applied in the scene of ASD children's time series gesture classification.For the recognition of ASD children's time series gestures,this article designs 4 noise reduction methods.Experiments show that these methods improve the accuracy by 5%.In addition,in order to recognize multiple ASD children,this article designs a multi-target tracking algorithm based on skeleton data.In the comparative experiments,the LSTM model designed in this article achieved the best accuracy of 90%.And in the ASD child gesture recognition test in the real environment,an average accuracy of 84.1% was achieved.(3)In this article,the designed deep learning models are compared and applied in the scene of four-axis UAV flight attitude classification.To recognize the flight attitude of four-axis UAV,this article divides 4 regular attitudes and 4 dangerous attitudes.Comparative experiments show that the encoder model designed in this paper has achieved the best recognition accuracy of 90%.In addition,this article takes the offline analysis of the four-axis UAV ”black box” data as the background and develops an application to classify attitudes.Based on this application,the UAV attitude recognition test in the actual scene was carried out,and finally an average accuracy of 84.8% was achieved.The deep learning models designed in this article have achieved good performance in two application scenarios with high social value.This provides a theoretical basis and experimental basis for the use of deep learning methods to study time series classification problems in the future.