Dual-domain Context-aware network in time series classification for identifying agricultural machinery trajectory operation mode

Agricultural machinery trajectory operation mode identification aims to determine the spatiotemporal distribution characteristics of agricultural machinery trajectories, thereby understanding the machinery's activity at each time step. It is a multivariate time series classification task. However, existing methods are limited to capturing feature interactions only within local trajectory points, without fully utilizing the global contextual information hidden within the trajectory. To effectively model the contextual information of trajectories to accurately identify operation modes pattern of agricultural machinery, we propose a dual-domain context-aware network (DCANet). First, we propose a multi-angle feature enhancement method that uses physical kinematics formulas and mathematical statistics methods to expand the feature set of trajectories, thereby fully exploiting the inherent information. We then design a frequency-domain context-aware module consisting of two components: a local awareness bottleneck module and a learnable frequency-domain attention context module. The former is used to model the local dependencies. Based on this, the latter uses a discrete wavelet transform and a multi-head attention mechanism to calculate the correlation and dependency of trajectory features in multiple frequency subspaces, adaptively analysing trajectory feature changes in the time-frequency domain. Finally, we design an information aggregation context module. Based on the construction of the spatiotemporal relationship graph of agricultural machinery trajectory points, we combine it with a multi-filter graph convolution operator to capture the multi-scale features of trajectory points in different channels. To validate DCANet, we evaluated it on 3 datasets against 12 related methods. The results demonstrate that DCANet achieves superior performance, with accuracies of 89.03%, 90.24%, and 90.47%, and F1-scores of 71.81%, 84.62%, and 90.43%, respectively.