列车轴温监测数据软测量方法
Soft measurement method for temperature monitoring data of train axle

为解决监测数据缺失导致的轴温监测系统误诊和漏诊率较高的问题,提出了一种基于数据特征分析的轴温监测数据软测量方法; 通过轴温监测点的布局与相关性分析,确定了监测数据软测量的源数据范围; 采用自组织特征映射算法,通过对源数据归一化、优胜区域定义与隶属度优化,实现了轴温数据本征维数确定与数据聚类; 引入多维尺度分析方法,通过数据间距的相似性量化与距离矩阵特征值分解,实现了轴温数据的类内降维; 采用多维尺度分析方法对类间降维数据再次降维,提出了一种分步式降维方法,构建了信息量最大化与计算量最小化的平衡策略; 采用深度学习栈式自编码器方法提取类间降维数据的内部特征,构建了缺失轴温数据的软测量模型。研究结果表明:基于降维数据的软测量方法的时间效率比基于原始数据的软测量方法高14.25%; 2种方法的精度相当,当一维数据缺失时,数据软测量的平均精度可达99.83%; 当二维数据缺失时,平均精度可达99.75%; 当三或四维数据缺失时,平均精度均可达99.16%; 在满足最大允许误差2.5%、误差容忍度1.0%条件的情况下,针对任意缺失维度不高于四维的情况,提出的方法可有效地实现高精度与高效率的缺失数据恢复。
To solve the high misdiagnosis and missing report rates of existing axle temperature monitoring system caused by data missing, a soft measurement method based on the characteristic analysis of axle temperature monitoring data was proposed. By analyzing the layout and correlation of axle temperature monitoring points, the source data range of soft measurement for monitoring data was determined. By using the self-organizing feature mapping algorithm, the intrinsic dimension of axle temperature data was determined. The data were clustered through source data normalization, winning region definition and membership degree optimization. By introducing the multi-dimensional scaling analysis method, the dimension reduction of axle temperature data in the internal class was realized through data space quantization and similarity as well as eigenvalue decomposition of the distance matrix. The multi-dimensional scaling analysis method was reused to reduce the dimension of data between external classes. A step-by-step dimension reduction method was proposed, and an quilibrium strategy was built to maximize the amount of information and minimize the calculations. The stacked auto encoder of a deep learning method was utilized to extract the internal characteristics of data between external classes, and a soft measurement model of axle temperature missing data was constructed. Research result shows that the time efficiency of the soft measurement method based on the dimension reduced data is 14.25% higher than that based on original data. The accuracies of the two methods maintain a similar level. When one-dimensional data is missed, the average accuracy of softmeasurement data can reach 99.83%. When two-dimensional data is missed, the average accuracy can achieve 99.75%. When three- or four-dimensional data is missed, the average accuracy can arrive at 99.16%. Under the conditions that the maximum allowable error and error tolerance are 2.5% and 1.0%, respectively, as long as the dimension of missing data is not higher than four, the method can effectively