Research On Time Series Data Classification Methods Based On Deep Learning

Time series data refers to the data recorded by observing certain things or phenomena at a given sampling frequency within a certain time range.Time series data classification judges the state of things or phenomena described by these data.It has a wide range of applications in the fields of equipment fault diagnosis,human activity recognition,and medical auxiliary diagnosis.With the development of the internet of things technology,a variety of time series data in different application fields are growing explosively,and various applications have higher requirements for classification performance.The traditional time series data classification method based on feature similarity is unable to satisfy the corresponding requirements.With the development of artificial intelligence technology,the analysis method combined with deep learning provides new ideas for improving the performance of time series data classification.Different from the data usually processed by deep learning in the fields of computer vision,natural language processing and speech recognition,time series data is closely related to specific application fields.Therefore,it is necessary to carry out research on time series data feature representation methods,time series data classification methods and time series data augmentation methods.The existing time series data feature representation methods rely on prior knowledge of related fields and specific application scenarios,which have certain limitations and poor performance in time series data classification.The time series data feature representation method based on the relative position matrix utilizes the relative position between any two elements of the time series data,and converts the original one-dimensional time series data into a two-dimensional image by constructing a relative position matrix.The converted image retains the patterns and features of the original time series data,and has high intra-class similarity and low inter-class similarity,which helps the following deep learning based classification models to learn the mapping between features and classes more easily.The performance analysis results show that on the UCR(University of California Riverside)time series data standard data set,when use the same deep learning model,this method is better than the same type of feature representation methods GAF-MTF(Gramian Angular FieldMarkov Transition Field)and RP(Recurrence Plot),and the classification performance(average error rate)improved 7.4% and 6.9% respectively.The relative position matrix feature representation method converts one-dimensional time series data into images.The improvement of classification performance comes at the cost of more memory consumption and runtime,and the original modes and features of some time series data may be destroyed after the feature representation.The time series data classification model based on multi-scale attention convolutional neural network adopts an end-to-end model,and the multi-scale feature maps containing short-term,mid-term and long-term dependencies of time series data is obtained by multi-scale convolution,and then establishes an attention mechanism to readjust the weights of the multi-scale feature maps,selectively pays attention to important information and ignores unimportant information according to the value of the weights,continuously enhances the useful feature maps while suppressing the less useful feature maps in the iterative learning process.This method helps to improve the recognition ability and generalization performance of the classification model.The performance analysis results show that on the UCR time series data standard data set,the classification performance(average error rate)of this method is 2.9% higher than that of the existing deep learning model with only parameter tuning,and at most improved 16.6%on single data set.Time series data often comes from different application fields,and the labeling of samples requires experienced domain experts,which makes it more difficult to obtain highquality sample data.The data augmentation method based on stepwise improving generative adversarial network enriches the training set by generating high-quality fake samples to augment the original time series data set,so as to improve the classification performance of the existing deep learning based classification methods on time series data with limited sample.This method designs a generator and discriminator structure suitable for time series data,and utilizes the trend samples generated by the time series data trend information extraction method to perform stepwise training,first learns the main trend distribution of the original time series data,and then learns the minor detail distribution of the original time series data.During the training process,the gradient penalty is adopted to perform Lipschitz constraint to approximate the distribution of the original time series data,so that the generator can fit the original time series data,and increase the diversity of the generated samples and the stability of the training process while retaining the trend of the original time series data.The performance analysis results show that on the UCR time series data standard data set with limited samples,after augment the training set based on this method,the classification performance(average error rate)of FCN(Fully Convolutional Network),Res Net(Residual Network)and MACNN(Multi-scale Attention Convolutional Neural Network)improved1.9%,1.5% and 0.8% respectively,and at most improved 5.5%,3.4% and 3.4% respectively on single data set.