Fixed‐Length Representation of Varying‐Length Multivariate Time Series With Application to Aerial Target Activity Classification
Zihao Song, Yan Zhou, Wei Dong Cheng, Futai Liang, Chenhao Zhang, Kai Yuan · The Journal of Engineering · 2025
ABSTRACT The present work aims at developing a fixed‐length representation for multivariate time series with varying lengths and apply it to the classification of aerial target activities. Numerous machine learning classification methods rely on fixed‐length sequences, which makes them not directly applicable to target feature time series of varying lengths. Processing techniques like resampling, truncation and padding may lead to information loss and the introduction of invalid data. Our proposed representation method can acquire fixed‐length feature vectors for target feature series of different lengths, which are well‐suited for use in machine learning classification approaches. Specifically, we employ the Gaussian mixture model (GMM) to group the aggregated window data, which is obtained through fixed‐length sliding windows and the largest triangle three buckets (LTTB) method, into a specified number of clusters. The window membership rate within each cluster serves as a feature in the representation. The experiments on the classification of varying‐length multivariate time series in aerial target activity show that using the proposed representation as the input feature vector of machine learning methods significantly outperforms other fixed‐length representation methods in terms of performance and efficiency.