基于LSTM-DHMM的MOSFET器件健康状态识别与故障时间预测

doi:10.12263/DZXB.20210047

摘要/Abstract

摘要：

针对MOSFET（Metal-Oxide-Semiconductor Field-Effect Transistor）器件故障预测与健康管理问题，提出了一种长短时记忆（Long Short-Term Memory, LSTM）算法与离散隐马尔可夫模型（Discrete Hidden Markov Model, DHMM）相结合的故障预测新方法.该方法利用LSTM算法预测器件状态发展趋势；用自回归（AutoRegressive, AR）模型提取故障信息特征；以DHMM建立特征向量和退化等级之间的映射关系；在LSTM-DHMM模型预测结果的基础上，结合失效阈值排除虚警并预测故障时间，预测误差小于10%，精度较高.与GRU-DHMM（Gated Recurrent Unit Discrete Hidden Markov Model）、GRU-SVM（Gated Recurrent Unit Support Vector Machine）、LSTM-SVM（Long Short-Term Memory Support Vector Machine）方法进行对比分析，结果表明，LSTM-DHMM的预测准确率高于其他三种方案，能有效识别实验器件健康状态、较好预测故障时间，具有有效性和优越性.

关键词: 故障预测与健康管理, MOSFET（Metal-Oxide-Semiconductor Field-Effect Transistor）, 长短时序列, 离散隐马尔可夫模型, 自回归模型, 故障时间

Abstract:

Aiming at the problem of MOSFET（Metal-Oxide-Semiconductor Field-Effect Transistor） device prognostic and health management, a fault prediction method combining long short term memory(LSTM) algorithm and discrete hidden Markov model(DHMM) is proposed to identify the health status and predict the fault time of MOSFET devices. In this method, LSTM algorithm is used to predict the development trend of device state; autoregressive(AR) model is used as the feature extraction method; DHMM is used to establish the mapping relationship between feature vector and degradation level; based on the prediction results of LSTM-DHMM model, false alarm is eliminated and fault time is predicted by combining with the failure threshold. The prediction error is less than 10% and the accuracy is high. Compared with single-stress GRU-DHMM（Gated Recurrent Unit Discrete Hidden Markov Model）、GRU-SVM（Gated Recurrent Unit Support Vector Machine） and LSTM-SVM（Long Short-Term Memory Support Vector Machine）, the proposed method is superior to the other four schemes in prediction accuracy and rationality, the results show that the prediction accuracy of the proposed method is higher than that of the other three schemes, and the proposed method can effectively identify the health state of the experimental devices and predict the fault time well, which is effective and superior.

Key words: prognostic and health management, MOSFET（Metal-Oxide-Semiconductor Field-Effect Transistor）, long short-term memory, discrete hidden Markov model, autoregressive model, fault time

中图分类号:

TH165+.3

张明宇, 王琦, 于洋. 基于LSTM-DHMM的MOSFET器件健康状态识别与故障时间预测[J]. 电子学报, 2022, 50(3): 643-651.

Ming-yu ZHANG, Qi WANG, Yang YU . Health Status Identification and Fault Time Prediction of MOSFET Device Based on LSTM-DHMM[J]. Acta Electronica Sinica, 2022, 50(3): 643-651.

图/表 16

参考文献 16

1	樊浩,李兴文,苏海博,等.基于主成分分析—支持向量机优化模型的断路器故障诊断方法研究[J].高压电器, 2020, 56(6): 143-151.
	FANHao, LIXing-wen, SUHai-bo, et al. Research on circuit breaker fault diagnosis method based on principal component analysis⁃support vector machine optimization model[J]. High Voltage Apparatus, 2020, 56(6): 143-151. (in Chinese)
2	刘磊,龙兵,刘震.两种多故障诊断算法的性能比较研究[J].电子测量与仪器学报, 2011, 25(1): 75-80.
	LIULei, LONGBing, LIUZhen. Research on performance comparison of two MFD algorithms[J]. Journal of Electronic Measurement and Instrument, 2011, 25(1): 75-80.
3	KALGRENP W, BAYBUTTM, GINARTA, et al. Application of prognostic health management in digital electronic systems[C]//2007 IEEE Aerospace Conference.Big Sky, MT, USA: IEEE, 2007: 1-9.
4	MENONS, JINX, CHOWT W S, et al. Evaluating covariance in prognostic and system health management applications[J]. Mechanical Systems & Signal Processing, 2015, 58-59(6): 206-217.
5	VASANA S S, PECHTM. Electronic circuit health estimation through kernel learning[J]. IEEE Transactions on Industrial Electronics, 2018, 65(2): 1585-1594.
6	RAMAKRISHNANA, PECHTM. A life consumption monitoring methodology for electronic systems[J]. IEEE Transactions on Components and Packaging Technologies, 2003, 26(3): 625-634.
7	陈颖,高蕾,康锐.基于故障物理的电子产品可靠性仿真分析方法[J].中国电子科学研究院学报, 2013, 8(5): 444-448.
	CHENYing, GAOLei, KANGRui. Research on reliability simulation prediction of electronic product based on physics of failure method[J]. Journal of CAEIT, 2013, 8(5):444-448. (in Chinese)
8	TSUIK L, CHENN, ZHOUQ, et al. Prognostics and health management: a review on data driven approaches[J]. Mathematical Problems in Engineering, 2015, 2015(6): 1-17.
9	杨立峰,吕卫民,肖阳.基于故障机理和伪失效寿命的电子产品剩余寿命预测[J].海军航空工程学院学报, 2017, 32(2): 246-250.
	YANGLi-feng, Wei-minLÜ, XIAOYang. Residual life prediction for electronic products based on fault mechanism and pseudo-failure data[J]. Journal of Naval Aeronautical and Astronautical University, 2017, 32(2): 246-250. (in Chinese)
10	周永道,王会琦,吕王勇.时间序列分析及应用[M]. 北京：高等教育出版, 2015.
	ZHOUYong-dao, WANGHui-qi, Wang-yongLÜ. Time Series Analysis and Application[M]. Beijing: Higher Education Press,2015. (in Chinese)
11	郝俊虎,胡毅.基于XGBoost和自回归模型的轴承故障诊断和预警方法研究[J].组合机床与自动化加工技术, 2020, 552(2): 140-142, 157.
	HAOJun-hu, HUYi. Research on bearing fault diagnosis and warning based on XGBoost and autoregressive algorithm[J]. Modular Machine Tool & Automatic Manufacturing Technique, 2020, 552(2):140-142, 157. (in Chinese)
12	伊恩.古德费洛, 约书亚.本吉奥, 亚伦.库维尔. 深度学习[M].北京: 人民邮电出版社, 2017. 187-190.
	IanGoodfellow, YoshuaBengio, AaronCourville. Deep Learning[M]. Beijing: Posts&Telecom Press Co,Ltd, 2017. 187-190. (in Chinese)
13	张峰,王东,石现峰.振动信号Burg谱估计算法的性能优化研究[J].计算机仿真, 2017, 34(4): 262-266.
	ZHANGFen, WANGDong, SHIXian-feng. Study on performance optimization of vibration signal burg spectral estimation algorithm[J]. Computer Simulation, 2017, 34(4): 262-266. (in Chinese)
14	RABINERL R. A tutorial on hidden Markov models and selected applications in speech recognition[J]. Proceedings of the IEEE, 1989, 77(2): 257-286.
15	于劲松,刘浩,张平,等.故障预测算法稳定性实时评估方法[J].北京航空航天大学学报,2014,40(9): 1208-1212.
	YUJin-song, LiuHao, Zhangping, et al. Real-time evaluation method for stability of fault prognostic algorithm[J]. Journal of Beijing University of Aeronautics and Astronautics, 2014, 40(9): 1208-1212.
16	TESTAA, CAROS D, PANARELLOS, et al. Stress analysis and lifetime estimation on power MOSFETs for automotive ABS systems[C]//Power Electronics Specialists Conference. Rhodes, Greece: IEEE Press, 2008: 1169-1175.

实验序号	优化算法
实验序号	Adam	SGDM	RMSProp
1	0.00814	0.00832	0.00951
2	0.00912	0.02174	0.01846
3	0.00898	0.01944	0.01633
4	0.00665	0.00741	0.00786
5	0.00814	0.00997	0.01021

实验序号	优化算法
实验序号	Adam	SGDM	RMSProp
1	0.00814	0.00832	0.00951
2	0.00912	0.02174	0.01846
3	0.00898	0.01944	0.01633
4	0.00665	0.00741	0.00786
5	0.00814	0.00997	0.01021

ACF	PACF	ACF	PACF
0.990	0.990	0.854	-0.004
0.980	0.020	0.848	-0.010
0.972	0.070	0.842	-0.002
0.964	-0.006	0.836	-0.009
0.953	-0.113	0.831	0.008
0.943	0.010	0.825	-0.006
0.935	0.048	0.819	-0.003
0.926	0.022	0.813	0.009
0.917	-0.022	0.807	0.000
0.909	0.024	0.801	-0.009
0.901	0.038	0.796	0.008
0.894	0.018	0.790	-0.004
0.887	0.005	0.784	0.002
0.881	0.041	0.778	-0.013
0.875	0.016	0.773	0.006
0.870	0.013	0.767	-0.012
0.864	0.002	0.761	0.006
0.859	0.005	0.755	-0.014

ACF	PACF	ACF	PACF
0.990	0.990	0.854	-0.004
0.980	0.020	0.848	-0.010
0.972	0.070	0.842	-0.002
0.964	-0.006	0.836	-0.009
0.953	-0.113	0.831	0.008
0.943	0.010	0.825	-0.006
0.935	0.048	0.819	-0.003
0.926	0.022	0.813	0.009
0.917	-0.022	0.807	0.000
0.909	0.024	0.801	-0.009
0.901	0.038	0.796	0.008
0.894	0.018	0.790	-0.004
0.887	0.005	0.784	0.002
0.881	0.041	0.778	-0.013
0.875	0.016	0.773	0.006
0.870	0.013	0.767	-0.012
0.864	0.002	0.761	0.006
0.859	0.005	0.755	-0.014

实验序号	三种退化状态DHMM输出的对数似然概率			识别结果
实验序号	正常损耗期	衰退期	严重退化期	识别结果
1	-54.41	-45.42	-33.04	严重退化期
2	-129.05	-133.25	-173.86	正常损耗期
3	-78.91	-79.17	-145.01	正常损耗期
4	-116.78	-116.39	-167.30	衰退期
5	-178.76	-186.99	-137.32	严重退化期