Time Series Classification (TSC) has achieved significant results in improving the efficiency and safety of life and production. This paper aims to enhance the accuracy of TSC by encoding one-dimensional time series into two-dimensional Recurrence Plots (RP) to obtain richer texture information. However, existing RPs face issues such as multi-scale issues, tendency confusion, and information redundancy, which hinder their application in TSC tasks. To address these problems, this paper proposes a Large Receptive Field Network (LRFN). LRFN advocates for extracting global information by enlarging the receptive field to overcome the multi-scale issues of RPs. It further addresses other problems by applying Multi-scale Signed RP (MSRP). The method encodes the time series as MSRP and then constructs multiple Global-Local Convolution (GLC) modules. These modules obtain medium, local, and global receptive fields through the Dilated Residual Module, SelfCalibrated Convolution, and Atrous Spatial Pyramid Pooling, respectively. By integrating three different qualities and sizes of receptive fields, comprehensive global information is acquired to address the multi-scale problem. Experimental results show that LRFN can further improve the accuracy of TSC tasks. LRFN demonstrated the best comprehensive performance in comparison to 13 baseline methods across 43 different domains in the UCR dataset. It performed optimally on 20 datasets, with an average accuracy of 91.3%. Additionally, receptive field visualization experiments further validate the effectiveness of LRFN.

Multivariate time series classification (MTSC) is a significant research topic in the realm of data mining, with broad applications in different industries, including healthcare, finance, meteorology, and traffic. While existing studies have designed many classifiers based on LSTMs, CNNs, and Transformer, the sophisticated architectures raise concerns regarding efficiency in computation. Additionally, most methods concentrate on a single dimension, typically temporal patterns, without fully considering multi-dimensional information such as the independence and interactions across variables that are essential in multivariate settings. To address these challenges, this article introduces FreConvNet, a lightweight convolutional network integrated with frequency adaptive learning. Inheriting the modular design paradigm of Transformer to achieve multi-view modeling of multivariate time series. FreConvNet consists of two key components: the frequency adaptive block (FAB) and the convolutional feed-forward network (ConvFFN). The FAB leverages the Fourier Transform in conjunction with adaptive filters to capture both long-term and short-term dependencies in the temporal dimension. Following that, ConvFFN captures cross-variable and cross-feature interactions by controlling inter-channel information flow through grouped pointwise convolutions, while introducing non-linearity to enhance representational capacity. Extensive experiments conducted on the well-known UEA archive validate that FreConvNet outperforms existing convolution-based, Transformer-based, and hybrid methods in classification performance and offers a computationally efficient solution.

Recently, time series classification has attracted significant interest. One of the most promising recent approaches is the shapelet transform, which offers two main advantages over traditional approaches: optimization of the shapelet selection process and the flexible integration of different classifiers. However, the high time complexity of identifying shapelets hinders its application in real-time data processing. To overcome this drawback, we propose an adaptive shapelet selection algorithm (ASS). In our method, we first identify Import Data Points (IDPs) for every time series and select the subsequences between two different IDPs as shapelet candidates. We then adaptively select the k best shapelets using ASS. Our experimental results demonstrate that ASS outperforms all other relevant classification methods.