<p>Semi-supervised learning (SSL) time series classification (TSC) is more in line with actual data scenarios. Existing methods have gradually transitioned from focusing only on time-domain features to exploring time-frequency invariant representations via contrastive learning. In this paper, we find class invariance between time-domain and frequency-domain features. Therefore, we propose a semi-supervised Time-Frequency Fusion Consistency (TFFC) model for Time Series Classification (TSC). Specifically, for the labeled data, we conduct supervised TSC training. For the unlabeled data, we design a fine-grained time-frequency fusion consistency strategy and use the confidence prediction of the classifier on the time-domain features to generate pseudo-labels for constraining the class invariance of temporal and fused features. Meanwhile, we design a trick of domain discrepancy minimization to mitigate the confirmation bias of the classifier due to the imbalance of labeled/unlabeled data in SSL. We perform TSC experiments on eight time series datasets with different characteristics. TFFC significantly outperforms the state-of-the-art baselines in all cases of labeled/unlabeled data on each dataset, and has an overall mid-range classification efficiency, which could verify the effectiveness of TFFC.</p>