| In the field of time series analysis,deep learning models have been widely used for classification tasks,and have shown significant improvement compared to traditional algorithms.However,the foundation of deep learning theory is based on repeated training with massive amounts of data to capture features.In real life,due to the impact of many objective factors such as time,cost,environment,etc.,the actual research samples often have insufficient data,that is,small sample data sets.Therefore,how to improve the the accuracy of time series data classification under small sample conditions has become an important and challenging issue.This article proposes a method called GCGAN for enhancing small-sample time series data classification performance,which is often hindered by insufficient data.The GCGAN model is based on an improved GAN architecture and aims to increase the amount of data by generating additional samples,thereby improving the classification accuracy.The GCGAN model is designed to address the characteristics of time series data and employs a GRU network as the generator to simulate the details distribution of the original data and generate higher quality fake samples.The discriminator is designed with two stacked 1D convolutional neural networks(CNN)to enhance its discriminative power and facilitate the optimization of the generator.In order to avoid the problem of gradient disappearance,Wasserstein distance is used instead of the original JS divergence to measure the similarity between the generated and real distributions,and a penalty term is introduced to satisfy the Lipschitz continuity condition.Experimental results on the UCR small-sample dataset show that the GCGAN method effectively improves the accuracy of deep learning models,achieving up to a 4.4% increase in classification accuracy and a 3.2% increase in F1-score,indicating the effectiveness of the GCGAN method for data augmentation.To improve the classification accuracy of time series data with limited samples,it is possible to optimize the structure of the classification model in addition to data augmentation.This article presents a BILSTM-FCN based classification model that combines Bi LSTM with FCN to learn long-term dependencies and capture local features,resulting in improved recognition and generalization ability.Compared with other mainstream classification models,the BILSTM-FCN model shows excellent feature learning ability and achieves good results in multiple evaluations.Additionally,the article combines the GCGAN method with the BILSTM-FCN method to propose a hybrid deep learning model classification method,which improves the classification performance of time-series data through two aspects of data augmentation and network structure optimization.The experimental results show that the hybrid deep learning model’s classification accuracy and average ranking on the UCR small sample dataset are among the best,and the classification error rate is reduced by a maximum of 0.056 compared to the basic BILSTM-FCN model,indicating the proposed method is effective and feasible. |
