采用基于神经网络的时间序列预测器的传感器故障诊断新方法

Sensor Failure Detection Based on Time Series Predictor Using Neural Network

本文提出一种采用基于神经网络的时间序列预测器的传感器故障诊断新方法。它先用神经网络对传感器输出序列建立预测模型（简称神经网络预测模型），然后利用神经网络预测模型预测传感器输出值和传感器实际输出之差判断传感器是否发生故障。与其它方法相比，本文方法只需单一传感器输出信号即可诊断相应的传感器是否发生故障。本文通过对一控制系统中转速传感器漂移故障诊断过程阐述了这种方法。

This paper proposes a new method to detect sensor failures through an artifical neural network based time series predictor for single sensor.The procedure of the proposed method can be descriped as follows:1.Training a kind of neural network based time series predictor for a target sensor using a series of simulation values of the sensor in certain working environment.2.The built predictor model is validated by a correlation analysis method.3.The discrepance between the estimation of the predictor for the target sensor and its actual value in on line condition may be used to determine whether sensor is working normaly or not with a given decision threshold.Compared with other methods,it need only single sensor signal to detect sensor failure.At the end of this paper,the correctness of the method is proved by a simulation result for a rotating sensor in an automotive engine.