Research On One-dimensional Continuous Time-series Dataanalysis Method Based On Transformer

One-dimensional time series data can be divided into continuous time series data and discrete time series data,which can be regarded as an extension of a regular two-dimensional feature matrix into the time domain.The main characteristic of continuous time series data is that the waveform is smooth.In the research direction of one-dimensional continuous data,common tasks include classification,detection(detecting the position of the signal),and predicting trends.The two main requirements are high analysis accuracy and fast inference speed.This thesis is based on the Transformer model and focuses on onedimensional continuous time series data.It conducts research around three basic tasks(classification,detection,and prediction)and explores detailed issues of four types of one-dimensional continuous data with representativeness and practical value.These four types of data are:magnetic anomaly detection,remote sensing image classification,medical field arrhythmia diagnosis,and material science field coating life prediction.The research content of this thesis is as follows:(1)The Shifted-Grad Transformer(SGT)is proposed,in which the Shifted-Grad Block module expands the gradient feature construction space of one-dimensional data and constructs gradient features from nearest neighbors to offset nearest neighbors.This simplifies the cumbersome data preprocessing steps and significantly improves the accuracy and speed of classification by directly feeding one-dimensional continuous data into the SGT network.At the same time,SGT improves the Transformer and,based on the Vision Transformer model,improves the patch split construction structure to combine the Transformer with the Shifted-Grad Block and enhance the long-distance dependence of one-dimensional data.SGT was first applied to magnetic anomaly signal classification and showed good performance on magnetic data,with a classification accuracy of 99.01% and a processing speed of 2.8×103 inputs per second,exceeding the best method by 1.57 percentage points and nearly 28 times faster.For magnetic anomaly signal detection,the detection task is transformed into a classification task through a sliding window and a post-processing dynamic programming algorithm is used to effectively eliminate false alarms in sliding window detection results at ()time complexity,effectively achieving magnetic anomaly signal target detection.The code for this framework has been opensourced1.This thesis continues to study some classic problems and public datasets of other one-dimensional continuous data.Through extensive experimental comparisons on the Pavia University hyperspectral image classification public dataset(spatially disjoint samples),the Flevoland dataset for polarimetric synthetic aperture radar(SAR)image classification,and the MIT-BIH public dataset for arrhythmia diagnosis,it is proven that SGT is a general framework for processing one-dimensional continuous data.(2)This thesis also studies the prediction problem of one-dimensional continuous time series data.Based on the Vision Transformer,a Time embedding module is introduced and a dual-branch Bidirectional LSTM network is used to achieve prediction of one-dimensional continuous time series data and enhance the influence of time features on prediction results.Taking the coating performance degradation prediction problem in the material field as an actual application scenario,the new model is compared with Informer and MLP-Mixer in experiments and shows better prediction ability.In summary,this thesis’ s work shows that the analysis model of onedimensional continuous time series data based on Transformer has important research significance and practical value in multiple related fields.