Contrastive learning is a mainstream self-supervised method for time series classification, which can learn rich representations from unlabeled data. Although existing time series classification methods have shown promising performance, exploiting temporal information and dealing with uncertainty noise is still quite challenging. To this end, we propose a hybrid contrastive learning network (HCLNet) for time series classification, which combines disturbance and temporal contrastive mechanisms to capture time series representations. Specifically, we first devise a dual data augmentation module to obtain different views by strong and weak augmentation with disturbance, which can remove the noise and enhance the consistency of representations. Building on this foundation, we design a disturbance contrastive module to capture robustness representations within time series data. A temporal contrastive module is then used to learn global temporal representations. This module attempts to replace MLP in Transformer with Kolmogorov-Arnold Network (KAN) layer to enhance the model’s generalization capability and parameter efficiency. Meanwhile, we design an adaptive hybrid method to integrate different contrastive learning losses effectively. We conduct experiments on three widely-used time series datasets to evaluate the performance of the HCLNet model in time series classification.