基于改进BP网络的冷水机组故障诊断

Research on Fault Diagnosis of Chillers Based on Improved BP Network

采用常规神经网络进行冷水机组的故障检测与诊断,存在整体检测率低或完全无法检测的现象。为了提高冷水机组故障检测效率及诊断精度,本文提出了一种基于贝叶斯正则化的改进神经网络故障检测策略。由于BP神经网络存在泛化能力差的缺陷,对神经网络进行贝叶斯正则化,从而提高模型的检测效率。贝叶斯算法通过限制神经网络权值,使网络反应更加光滑,模型更精确。通过利用ASHRAE Project提供的数据对FDD(fault detection and diagnosis)策略进行验证,检测率明显提高。

The overall detection rate using conventional neural networks to detect and diagnose the chillers＇ fault is low,even this method can＇t detect the fault completely. In order to improve the fault detection and diagnostic accuracy of chiller,an improved neural network fault detection strategy based on Bayesian regularization is proposed. Due to the defects of poor generalization ability of BP neural network,the neural network based on Bayesian regularization can improve the detection efficiency of the model. Bayesian algorithm by limiting the weights of the neural network makes the network more smooth,which make the model more precise. Validation of FDD（ fault detection and diagnosis） strategy through using ASHRAE Project data shows that the detection rate is improved obviously.