Abstract: Ice cracking signals, which serve as acoustic fingerprints of the dynamic evolution of polar ice sheets, possess significant monitoring values, while traditional detection methods face challenges such as insufficient sensitivity in low signal-to-noise ratio environments and excessive dependence on annotated data. This paper proposes a semi-supervised learning framework based on multi-component feature fusion for ice cracking signal detection, aimed at enhancing automated detection capabilities under limited annotation conditions. First, two three-dimensional acoustic field feature fusion schemes, i.e. mixed waveform diagrams and stacked time-frequency representations, are proposed and compared to optimize the utilization of three-dimensional elastic acoustic field information. Then, a dual-branch iterative optimization architecture with teacher-student models is constructed, where a semi-supervised learning pseudo-labeling strategy is introduced to leverage unannotated data, thereby enhancing the model generalization capability while significantly reducing dependency on manual annotations. Through experiments conducted in high-latitude ice regions using YOLOX as a benchmark model, the effectiveness of the proposed technique is validated. With the same amount of manually annotated data, the detection performance is improved by 26.2% compared to the benchmark; when the annotated data is reduced by 50%, the proposed method maintains a performance comparable to fully-supervised models (achievement rate of 94.5%). The ROC curve analysis further confirms that the stacked time-frequency feature fusion scheme exhibits distinct advantages in ice cracking signal detection tasks.