基于核密度估计的故障诊断信号非均匀量化方法

Inhomogeneous Quantization Method of Fault Diagnosis Signal Based on Kernel Density Estimation

在工业物联网的某些故障诊断场景中,由于缺少电信网络覆盖,采集信号通过远距离无线电(Long Range Radio,LoRa)技术实现无线可靠回传,但其较低的传输速率会限制故障诊断的精度。针对LoRa窄带宽的技术限制,提出了采样频率和量化分辨率固定条件下的时域信号非均匀量化方案。首先,通过建立基于核密度估计(Kernel Density Estimation,KDE)的非参数拟合模型,重点研究了带宽受限场景中合适的核函数类型和带宽确定准则,拟合生成传感信号幅度的概率密度函数(Probability Density Function,PDF)。其次,以PDF为输入,以最小化量化噪声为目标函数,通过非线性规划,输出最佳的一组非均匀量化电平值。其特点在于针对出现频次最高的时域幅度,采用更小的量化间隔,实现量化噪声的最小化。最后,以轴流风机状态检测为例进行了实验,结果表明,基座松动和轴承故障对量化电平的影响更大。随着量化分辨率的增加,KDE量化逐渐趋近均匀量化,相较于高斯量化的优势逐渐缩小。因此,提出的KDE量化方案适合窄带宽条件下的非均匀量化,可提高信道利用率,并在传输带宽和量化噪声之间取得折中。

In some fault diagnosis scenarios of industrial Internet of things,due to the lack of telecommunication network coverage,the collected signals are wirelessly and reliably backhauled by long range radio(LoRa)technology,but its lower transmission rate can limit the accuracy of fault diagnosis.An inhomogeneous quantization scheme for time-domain signals under the conditions of fixed sampling frequency and quantization resolution is proposed to address the technical limitations of the narrow bandwidth of LoRa.Firstly,the probability density function(PDF)is fitted to generate the sensed signal amplitude by establishing a nonparametric fitting model based on kernel density estimation(KDE),focusing on the suitable type of kernel function and bandwidth determination criteria in bandwidth-constrained scenarios.Next,using PDF as input and minimizing the quantization noise as the objective function,the optimal set of inhomogeneous quantization level values is output through nonlinear programming.It is characterized by using smaller quantization intervals for the time-domain amplitude with the highest frequency of occurrence to minimize the quantization noise.Finally,taking axial fan condition detection as an example,the experimental results show that the base loosening and bearing faults have a greater impact on the quantization level.With the increase of quantization resolution,the KDE quantization gradually converges to the uniform quantization,and the advantage over Gaussian quantization is gradually reduced.Therefore,the proposed KDE quantization scheme is suitable for inhomogeneous quantization under narrow bandwidth conditions,which can improve channel utilization and achieve a compromise between transmission bandwidth and quantization noise.