电子器件散热风扇气动噪声管道声学模态截止控制技术

深度学习输入特征的选择直接影响其分类性能,为了进一步提高基于深度学习的鸟类物种识别模型的分类性能,该文提出一种多特征融合识别方法。该方法首先通过短时傅里叶变换、梅尔倒谱变换和线性调频小波变换分别计算得到鸣声信号的3种语图样本集,然后分别利用3种语图样本集训练3个基于VGG16迁移的单一特征模型,将3个模型的输出进行自适应加权求和实现融合,并修正了加权交叉熵函数以克服样本不平衡的问题,最后对语图进行分类实现鸟类物种的识别。以ICML4B鸣声库的35种鸟类为研究对象,对比了4种模型的平均识别准确率(MAP),结果表明特征融合模型较单一特征模型的MAP最大提高了0.307;选择输入语图的持续时间分别为100 ms、300 ms以及500 ms,对比不同持续时间下4种模型的测试MAP值,结果表明持续时间为300 ms时4种模型的MAP值均为最高;对比了不同信噪比下4种模型的识别效果,多特征融合模型的识别准确率随着信噪比的下降降低最少。说明在选择合适的语图持续时间后,该文提出的特征融合模型能得到更高的识别准确率,具有一定的抗噪能力,且训练参数少,更适合于少样本鸟类的识别。

Aerodynamic noise control technology of electronic device cooling fan based on duct acoustic mode cutoff

An experimental study on the aerodynamic noise and noise reduction method of a variable-speed axial flow fan for electronic device heat dissipation is conducted. The far-field radiation noise at different rotational speed is measured by a microphone array uniformly distributed along the circumference in the contour plane of the fan axis. The logarithmic relationship between the total SPL and the rotational speed verifies that the main aerodynamic noise of the fan belongs to the dipole source noise, and the spectrum analysis indicates that the discrete tonal noise is the dominant noise. Based on the duct mode cutoff in duct acoustic theory, the effects on the aerodynamic noise of installing a circular short duct at the inlet and outlet of the fan are studied, pointing out that the short ducts with different lengths on the different sides have different effects on the far-field noise. At rated fan speed, installing a 2 cm duct at the inlet can reduce the average total SPL by 4.1 dB(A) at the far field of 1 m, which is significant. Mode measurement results show that in this case the main mode amplitude corresponding to the discrete tone is greatly reduced, and the reduction of discrete tonal noise leads to a significant reduction of the total SPL. This method provides a new way for cooling fan noise reduction.