Prediction of Polyphonic Alarm Sound by Deep Neural Networks
Prediction of Polyphonic Alarm Sound by Deep Neural Networks
复制标题
通过深度神经网络预测和弦警报声
DOI：
10.11239/jsmbe.60.8
复制
发表时间：
2022
影响因子：
--
通讯作者：
黒田 知宏
中科院分区：
文献类型：
--
作者：
岸本 和昌 竹村 匡正;杉山 治;小島 諒介;八上 全弘;南部 雅幸;藤井 清孝;黒田 知宏
在医院里，当医务人员在远处或封闭的房间里没有注意到警报响起时，可能会发生事故。在许多医院中，患者监护设备通过网络连接到医院信息系统，但有些医疗设备由于不产生任何外部输出而无法连接。如果工作人员能够在一定距离内检测到病房的警报，他们就可以提供更有效和主动的医疗服务。在本研究中，使用单耳麦克风收集报警声音，并使用深度神经网络构建机器学习分类器。分类器被评估使用复调报警声音的模拟数据集，叠加与医院病房的环境声音。从四个设备收集数据，并使用Mel滤波器组（MFB）和自定义滤波器组（CFB）创建两个对数谱图训练数据集。此外，基于这四种设备的组合，为16个类别开发了两个分类器。一个分类器在MFB上训练，另一个在CFB上训练。分类器对模拟数据集的信噪比（SNR）分别为30、20、10和0 dB。在信噪比为0 dB的情况下，CFB分类器的微F1评分为72.7%，roc曲线下面积为0.963。该微F1得分比在MFB上训练的分类器得分高4.5分。此外，环境声音（不含所有装置的类别）的误认率为1.2%。因此，分类器不能可靠地区分报警声音和环境声音，但提出了作为通知系统的可能性。
Accidents may occur in hospitals when the medical staff fail to notice the alarm ringing at a distance or in a closed room. In many hospitals, patient monitoring devices are connected to the hospital information system through a network, but some medical devices cannot be connected because they do not produce any external output. If the staff can detect the alarm ringing in a hospital room from some distance, they can provide more efficient and proactive medical care. In this study, alarm sounds were collected using a monaural microphone, and a machine learning classifier was constructed using deep neural networks. The classifier was evaluated using a simulation dataset of polyphonic alarm sounds, superimposed with the environmental sounds of a hospital ward. Data were collected from four devices, and two training datasets were created with a logarithm spectrogram using Mel filter bank (MFB) and custom filter bank (CFB). In addition, two classifiers were developed for 16 classes based on a combination of the four devices. One classifier was trained on MFB and the other on CFB. The classifiers evaluated the simulation dataset with a signal-to-noise ratio (SNR) of 30, 20, 10, and 0 dB. The classifier trained on CFB had a micro F1 score of 72.7% and an area-under-the-ROC-curve of 0.963 at an SNR of 0 dB. This micro F1 score was 4.5 points higher than that of the score of the classifier trained on MFB. In addition, the misidentification rate of the environmental sounds (class without all devices) was 1.2%. Therefore, the classifier could not reliably distinguish between the alarm sound and environmental sounds, but the possibility as a notification system was presented.