摘要
针对现有输煤传送机托辊故障检测方法存在识别精度较低、抗环境干扰能力较差、无法长期稳定运行等问题,提出了一种基于融合信号(TFM)及多输入一维卷积神经网络(MI-1DCNN)的输煤传送机托辊故障检测方法。首先,通过拾音器采集输煤传送机沿线托辊运行的音频信号,采用dB4小波无偏风险估计阈值降噪法对信号进行预处理,消除背景噪声,提高信噪比。然后,对降噪音频信号的时域、频域和梅尔频率倒谱系数(MFCC)及其一阶二阶差分系数进行归一化处理,并进行拼接,得到特征TFM。最后,将TFM输入到多尺度卷积核的MI-1DCNN模型,在网络通道末端进行特征融合,通过Softmax函数完成对正常托辊和故障托辊的分类识别。以某煤矿实际采集的输煤传送机托辊音频信号样本对TFM-MI-1DCNN模型进行试验,结果表明:故障托辊平均识别准确率达98.65%,较改进小波阈值降噪-反向传播-径向基函数网络、MFCC-K邻近方法-支持向量机的平均识别准确率分别提高了1.50%和1.03%。现场应用结果表明:该方法下故障托辊平均识别准确率为98.4%,说明该方法适用于现场应用。
The existing fault detection methods for belt conveyor idler have the problems of low recognition precision, poor anti-interference capability and inability to operate stably over a long period of time. In order to solve the above problems, a fault detection method for belt conveyor idler based on time-frequency-MFCC(TFM)and multi-input one-dimensional convolutional neural network(MI-1DCNN) is proposed. Firstly, the pickup collects the audio signal of the coal conveyor idler running along the line. The dB 4 wavelet unbiased risk estimation threshold noise reduction method is used to preprocess the signal to eliminate the background noise and improve the signal-to-noise ratio. Secondly, the time domain, frequency domain and Mel frequency cepstrum coefficient(MFCC), and the first and second order difference coefficient of the noise reduction audio signal are normalized respectively, and finally assembled to obtain the feature TFM. Finally, that TFM signals are input into a MI-1DCNN model with a multi-scale convolution kernel. The feature fusion is carried out at the end of a network channel. The classification and identification of the normal idler and the fault idler are completed through a Softmax function. The TFM-MI-1DCNN model is tested with the audio signal samples of coal conveyor idler collected in a coal mine. The results show that the average recognition accuracy of the fault idler is 98.65%. The average recognition accuracy is improved by 1.50% and 1.03% compared to the improved wavelet threshold denoising-backpropagation-radial basis function network and MFCC-K-nearest neighbor algorithm-support vector machine. The false detection rate is only 0.194%. The results of field application show that the average dentification accuracy of the proposed method is 98.4%, indicating that the proposed method is suitable for field application.
作者
武国平
WU Guoping(CHN Energy Zhunneng Energy Group Co.,Ltd.,Ordos 017000,China)
出处
《工矿自动化》
CSCD
北大核心
2023年第2期149-156,共8页
Journal Of Mine Automation
基金
国家自然科学基金项目(51905459)。
关键词
输煤传送机
智能巡检机器人
托辊
音频信号
小波阈值降噪
MFCC
多输入一维卷积神经网络
belt conveyor
intelligent inspection robot
idler
audio signal
wavelet threshold denoising
MFCC
multi-input one-dimensional convolutional neural network
