An end-point detection method for beluga whistle signals under burst pulse interferences

Aiming at the problem that the detection accuracy of the beluga whistle is reduced due to a large number of burst pulse interferences under captive conditions, an unsupervised whistle detection method based on the tunable Q-factor wavelet transform (TQWT) and adaptive spectral clustering is proposed. To eliminate the interference of silent time and low intensity pulses, the high Q-factor wavelet decomposition is used to decompose each original signal frame, and the average wavelet coefficient is used as the test statistics of the whistle rough detection. On this basis, local density adaptive spectral clustering is designed for the remaining frames to further distinguish whistles and strong pulse interference for complete beluga whistle detection. The performance of the detector is tested with real signals of beluga whales, and its F₁-score is calculated. The results show that the detector achieves F₁-score of 0.9487 and 0.9429 under two test signals, which is obviously higher than the detector composed of k-means clustering or traditional spectral clustering in the case of large data, and has higher robustness.