基于深度学习的机器异常声音检测

A Machine Abnormal Sound Detection Based on Deep Learning

随着大规模工业生产的发展,机器设备的健康管理越来越重要。由于机器设备潜在的故障,机器异常声音的检测对工业生产的保障有待提高。不同的机器运作时发出的声音有规律性,可以根据这一特性判断机器是否处于一个正常运作状态,通过对机器运作时的声音特征进行研究,提出一种基于深度学习的机器异常声音的检测,通过对声音特征的提取,经过模型的训练,判断机器是否处于异常状态,防患于未然。首先对数据集通过等高梅尔滤波器处理后提取出对数Mel谱作为声音特征,之后针对实际中异常声音的缺失等问题,使用mobilenetv2对声音模型进行训练,通过模型输出的逻辑回归值来计算异常分数和确定异常阈值。经过对比分析,表明对原始音频进行特征提取后训练的模型,机器异常声音检测性能有所提升。

With the development of large-scale industrial production, the health management of machinery and equipment is becoming increasingly important. Due to the potential failure of machinery and equipment, the detection of abnormal machine sounds needs to be improved. The sound emitted by different machines during operation is regular, and whether the machine is in a normal operation state can be judged according to this characteristic. Through the research on the sound characteristics of the machine during operation, a machine abnormal sound detection based on deep learning is proposed. Through the extraction of sound characteristics and the training of the model, whether the machine is in an abnormal state can be judged to prevent potential problems. Firstly, the data set is processed by the constant-height Mel filter to extract the logarithmic Mel spectrum as the sound feature, then the sound model is trained by mobilenetv2 to process the absence of abnormal sound in the data set, and the abnormal score and the abnormal threshold are calculated by the logistic regression value output by the model. After comparative analysis, it is shown that the machine abnormal sound detection performance of the model trained is improved by feature extraction of the original audio.