王煜林, 周登极(), 郝佳瑞, 黄大文

上海交通大学 机械与动力工程学院, 上海 200240

收稿日期:2020-04-03出版日期:2021-06-28发布日期:2021-06-30
通讯作者:周登极E-mail:1516761299@sjtu.edu.cn
作者简介:王煜林(1997-),男,辽宁省大连市人,硕士生,主要研究方向为动力系统数据驱动建模
基金资助:国家自然科学基金资助项目(51706132)

A Compressor Power Soft-Sensing Method Based on Interpretable Neural Network Model

WANG Yulin, ZHOU Dengji(), HAO Jiarui, HUANG Dawen

School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China

Received:2020-04-03Online:2021-06-28Published:2021-06-30
Contact:ZHOU Dengji E-mail:1516761299@sjtu.edu.cn

摘要/Abstract

摘要： 为在保证测量的准确性和高效性的同时,降低软测量方法对数据集的依赖性,提出一种基于可解释神经网络的压缩机功率软测量方法.实验中,在使用泛化性良好的数据集进行训练时,可解释神经网络模型在测试集上的均方根误差为 0.0094,相比反向传播(BP)神经网络模型降低了1.1%.在使用泛化性较差的数据集进行训练时,可解释神经网络模型在测试集上的均方根误差为 0.0128,相比BP神经网络模型降低了79.8%.实验结果表明,基于可解释神经网络的压缩机功率软测量方法不但具有较高的准确率,且在使用泛化性较差的数据集进行训练时,依然能够保持较高的测量性能.
关键词: 可解释性, 神经网络, 软测量方法, 压缩机
Abstract: In order to ensure the accuracy and efficiency of measurement, and reduce the dependence of the soft sensing on dataset, a soft-sensing method of compressor power based on interpretable neural network is proposed. When training on a dataset with good generalization in the experiment, the root mean squared error(RMSE) of the interpretable neural network model on the test set is 0.0094, which is 1.1% lower than that of the back propagation(BP) neural network model. When training on a dataset with poor generalization, the RMSE of the interpretable neural network model on the test set is 0.0128, which is 79.8% lower than that of the BP neural network model. The experimental results show that the soft-sensing method based on interpretable neural network not only has a high accuracy rate, but also can maintain a good measurement performance when training on a dataset with poor generalization.
Key words: interpretability, neural network, soft-sensing method, compressor


PDF全文下载地址:

点我下载PDF