基于改进退化隐马尔可夫模型的设备健康诊断与寿命预测研究

针对隐马尔可夫模型在进行设备健康诊断时与实际存在较大偏差的问题,提出了一种以似幂关系加速退化为核心的改进退化隐马尔可夫模型(DGHMM).首先,引入退化因子描述设备衰退过程,提出的似幂关系加速退化较常规指数式加速退化而言,能更好地描述设备服役期间随着役龄增加性能的逐步下降.其次,以全局搜索能力相对较强的改进遗传算法代替常规EM算法进行参数估计,克服了EM算法易陷入局部最优的局限性.同时,针对隐马尔可夫模型时间上须服从指数分布而不能直接用于寿命预测的局限性问题,提出了一种以近似算法与Viterbi算法为基础的贪婪近似法,以寻求最大概率剩余观测为目的,动态地寻求最大概率剩余状态路径,对设备剩余寿命进行预测.最后,通过美国卡特彼勒公司液压泵数据集对所提出的方法进行验证评价.结果表明,基于改进退化隐马尔可夫模型的设备健康诊断与寿命预测方法在描绘设备退化、设备状态诊断准确率方面更加有效,在剩余寿命预测上亦为可行.

Equipment health diagnostics and prognostics method based on improved degenerated HMM

In order to solve the problem of large deviation between hidden Markov model and actual equipment health diagnosis, this paper developed an improved degenerated hidden Markov model(DGHMM) with a core of the quasi power relation. Firstly, the model adopted the degradation factors, modeling the process of recession for the equipment''s continuous decrease in performance. Compared with the conventional exponential accelerated degradation, the quasi power relation accelerated degradation could better describe the process that the performance of the equipment decreases gradually with the increase of service age. Then, the improved genetic algorithm could replace the conventional EM algorithm for parameters'' estimation, which overcame the limitation that the EM algorithm was easy to fall into local optimization. At the same time, in terms of the limitation of life prediction problem as a result of the hidden Markov model must obey exponential distribution, an algorithm named greed & approximation based on approximation algorithm and Viterbi algorithm came out, and to seek maximum probability remaining observation, for the purpose of seeking maximum probability dynamically surplus state path, to predict the residual life of equipment. Finally, it validated and evaluated the proposed method with the data set of caterpillar hydraulic pumps. The results show that the method of equipment health diagnosis and life prediction based on the improved degraded hidden Markov model is more effective in describing equipment''s degeneration and the accuracy of equipment state diagnosis, and is also feasible in the prediction of residual life.