列车定位单元ODO故障分析及HMM的应用研究

Application of Hidden Markov Model for Fault Analysis of ODO in Train Location Unit

列车运行过程中车轮会出现打滑抱死等现象,导致ODO(odometers)呈现功能故障状态,针对这一问题,进行了列车定位单元ODO故障分析及HMM(hidden Markov models)的应用研究.首先利用神经网络方法对ODO故障状态进行诊断;然后,引入基于HMM的故障诊断方法,从模式识别角度出发,建立ODO隐藏的故障状态模型,通过输入观测序列与分类器中的HMM匹配,得出ODO的故障诊断结果;最后,通过遗传算法对HMM中的参数训练部分进行改进.实例分析结果表明:利用神经网络进行故障诊断的故障识别率为33.3%,基于HMM的故障诊断方法对于正常和故障状态的诊断精度可达100%,总体诊断精度可达95%,利用遗传算法进行参数改进后经仿真对比表明:在训练速度上遗传算法可以较快到达稳态,训练精度提高了86%;在高噪声背景下用神经网络方法实现故障诊断具有明显缺陷,利用遗传算法可以改进B-W(Baum-Welch)算法易陷入局部最优的缺陷,基于HMM的故障诊断方法相较于神经网络方法有更高的准确性.

During the running of a t ra in , the lock-and-slip state of a wheel may occur and h en c e , an Odometers （ODO） will temporarily display the state of functional failure. To address this is su e, fault analysis of ODO based on the application of hidden Markov model （ HMM） has been carried out. Firstly, an algorithm based on neural networks （NNs） was applied for the fault diagnosis of ODO. A fault diagnosis method based on HMM was then studied from the perspective of pattern recognition. The observation of the quantization coding sequence was introduced into the HMM classifier for establishing the hidden fault states of the ODO model with the aid of training d a ta. By entering the observation sequence matching with the HMM classifier, the effective state of ODO was obtained. Finally, genetic algorithms （GAs） were employed for obtaining the parameters of HMM, rather than employing theBaum-Welch （ B - W ） algorithm. The obtained case study results show that the fault identification accuracy based on NN is 33. 3 % , the diagnos-tic accuracy based on H M M can reach up to 100% between normal and fault conditions, and the overall diagnostic accuracy is 95% . The training speed of GA can help in reaching the steady state quickly and the training accuracy can be improved by 86% . Furthermore, the obtained results show that the classification ability of N N is comparatively weak and cannot meet the required needs of classification when there is high noise in the data. The G A can improve on the drawbacks of the B-W algorithm that in turn can easily fall into the local optimum. The fault diagnosis method based on HMM has higher accuracy when compared with the NN-based method.