无监督异常声音检测
我们提出的无监督异常声音检测鲁棒训练框架获得了DCASE2022 Task2的第一名!
We propose a robust training framework for anomalous sound detection, which includes feature preprocessing, model pretraining, joint loss, and anomaly score selection. The experimental results show that our anomalous sound detection model outperforms the official model, with an average performance improvement of 22.08% based on the official scoring method.




声事件定位与检测SELD
在 ICASSP 2022 L3DAS22 挑战赛中采用了多特征融合深度学习模型,同时利用Ambisonics格式音频数据增广技术克服了模型的过拟合现象,有效的提升了系统性能。最终提交模型得到0.592的T2Metric得分,获得了全球第二名,论文ICASSP2022 L3DAS22 Challenge: Ensemble of ResNet-Conformers with Ambisonics Data Augmentation for Sound Event Localization and Detection 被ICASSP2022 收录。
基于音频的扇叶故障检测
通过采集叶片转动的音频数据,首次完成了基于音频的扇叶故障检测,该方案无侵入式伤害,具有实时性,符合工业级应用。
Wind turbine blade fault detection by acoustic analysis: preliminary results, IEEE International Conference on Signal Processing, Communications and Computing (ICSPCC), Xi’an, China, Aug. 2021.

