基于随机卷积核与孤立森林柱塞泵异常检测方法

Anomaly Detection Method of Axial Pump Based on Random Convolution Kernel and Isolated Forest

液压柱塞泵出口压力脉动信号近似于周期信号,有比较明确的物理含义,携带着丰富的泵的健康状态信息,是柱塞泵健康管理比较理想的信号源。利用柱塞泵出口压力上的脉动成分,提出一种基于随机卷积核的柱塞泵异常检测方法,只需柱塞泵在正常工况下的压力脉动数据即可具备检测异常压力脉动数据的能力。该方法包括波形划分、异常数据段检测2个阶段:采用基于动态时间规整(Dynamic Time Warping, DTW)数据划分算法对柱塞泵压力脉动原始数据进行分割,获取压力脉动数据段以构建数据集;基于大量一维随机卷积核提取特征,获取正常状态下压力脉动数据段特征;使用孤立森林算法对基于随机卷积核提取的特征进行异常检测。该方法在真实数据集上的表现,表明其对于异常波形的判断有优异的表现,且查准率较单一孤立森林算法提升了6.3%。

The pressure pulsation signal at the outlet of the hydraulic plunger pump is a pseudo periodic signal with explicit physical significance, and contains abundant information about the pump’s health condition. It is an ideal signal source for the health management of the axial piston pump. In this paper, an anomaly detection method for a axial piston pump based on a random convolution kernel is proposed by using the pressure pulsation signals. The method only needs the pressure pulsation data of the axial piston pump under normal working conditions to have the ability to detect anomaly. The method includes two steps: Using the DTW-based data division algorithm to segment the plunger pump pressure pulsation raw data, and obtaining pressure pulsation data segments to construct a data set;Extracting features based on a large number of one-dimensional random convolution kernels, to obtain the characteristics of the pressure pulsation data segment under normal conditions;use the isolated forest algorithm to perform anomaly detection based on the features extracted by random convolution kernels. The performance of the method on the real data set shows that the method is more accurate indetecting abnormal waveforms, and the precision is 6.3% higher than that of the single isolated forest algorithm.