<p>The safe and stable operation of power plants is seriously threatened by boiler tube leakage accidents, making accurate detection of boiler tube leakages essential. However, minor leakages are difficult to detect under the influence of strong hot-state background noise in boilers, and the variety and quantity of available leakage acoustic signals on site are very limited. An equivalent criterion based on sound pressure amplitude is derived, and a method based on the proportional relationship of effective sound pressure is employed to map the hot-state background noise to the experimental environment, thereby constructing a leakage acoustic detection platform that is almost equivalent to the boiler environment. On the platform, various types of leakage acoustic signals with noise are collected, effectively addressing the issue of sample scarcity. To fully utilize the information from multi-channel acoustic sensors and enhance the detection accuracy of minor leakages, a calculation method based on weighted Kullback–Leibler (KL) distance is proposed to select the time-domain and frequency-domain statistical features and wavelet packet energy (WPE) features from four-channel acoustic signals. Experimental results show that the proposed feature selection method combined with a support vector machine (SVM) outperforms the detection method using an acoustic signal energy threshold, with a detection accuracy of 97.00% versus 65.36%. Furthermore, when detecting acoustic data from leakage source locations that were not involved in the training dataset, as well as data collected from the cold-state boiler environment, the detection accuracies of the proposed method are 92.67% and 93.53%, with the narrowest confidence intervals. The results demonstrate that the proposed method exhibits superior generalization capability compared to the feature selection method based on Relief-F and KL distance.</p>