Bearing Vibration Abnormal Detection Based on Improved Autoencoder Network

doi:10.3778/j.issn.1673-9418.2007042

Abstract

Abstract:

In recent years, autoencoders and neural network technologies have been widely studied and applied to abnormal data detection problems of industrial data such as bearing vibration, but there are still problems such as large training data, network parameter initialization, low training efficiency, poor detection effect and so on. To solve such problems, this paper presents an anomaly data detection method combining Mahalanobis distance and autoencoder network. There is a certain correlation between bearing vibration data characteristics, so the Mahalanobis distance of the data is used to quickly detect some abnormal data, which reduces the amount of training data for the self-encoding network. In this research, the autoencoder and the classifier are combined to construct the autoencoder network, which solves the problem of network parameter initialization and significantly improves the training efficiency. The Mahalanobis distance of the data is added to the data features, which improves the anomaly detection effect of the autoencoder network. The sparseness restriction is added to the autoencoder, and a structure that first enhances the dimensionality and then encodes samples is constructed. The structure enhances the feature learning ability and convergence of the autoencoder. Experimental results show that the method presented has better detection results than other abnormal detection methods for low-dimensional bearing vibration data, and it has certain stability and generalization ability.

Key words: autoencoder, bearing vibration, abnormal detection, Mahalanobis distance, autoencoder network

摘要：

近年来,自编码器和神经网络技术已被广泛研究并应用于轴承振动等工业数据的异常检测问题上,但仍存在着训练数据量大、网络参数初始化、训练效率较低、异常检测效果较差等问题。为解决上述问题,提出了一种结合马氏距离和自编码网络的异常检测方法。利用轴承振动数据特征之间具有一定相关性的特点,通过数据的马氏距离快速检测出部分异常数据,减少了自编码网络的训练数据量;用自编码器结合分类器构建自编码网络,解决了网络参数初始化问题并且显著提高了训练效率;将数据的马氏距离作为特征加入训练中提升了自编码网络的异常检测效果;在自编码器中加入稀疏性限制并构造先升维再编码的结构,增强了自编码器的特征学习能力和收敛性。实验结果表明,针对低维轴承振动数据,提出的方法较其他异常检测方法具有较好的检测效果且具有一定的稳定性和泛化能力。

关键词: 自编码器, 轴承振动, 异常检测, 马氏距离, 自编码网络

CLC Number:

TP391

LI Beibei, PENG Li. Bearing Vibration Abnormal Detection Based on Improved Autoencoder Network[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(1): 163-175.

李贝贝, 彭力. 基于改进自编码网络的轴承振动异常检测[J]. 计算机科学与探索, 2022, 16(1): 163-175.

Figures/Tables 32

Fig.1 Structure of AE

Fig.2 Improved structure of AE

Fig.3 Supervised fine-tuning

Fig.4 Bearing vibration abnormal detection model

Table 1 Evaluation index confusion matrix

实际类别	预测为异常数据	预测为正常数据
实际异常数据	TP	FN
实际正常数据	FP	TN

Fig.5 Bearing vibration data from dataset 1

Table 2 Size of dataset 1

类别标签	类别名	样本数	数据占比/%
0	正常数据	537	54.7
1	异常数据	445	45.3

Fig.6 Normal bearing vibration data

Fig.7 Mahalanobis distance for normal data

Fig.8 Mahalanobis distance for all data

Table 3 Uncertain dataset

类别标签	类别名	样本数	数据占比/%
0	正常数据	537	86.3
1	异常数据	85	13.7

Fig.9 Training loss changes with β

Fig.10 Training loss changes with ρ

Table 4 Parameter setting of autoencoder

参数	数值
学习率	0.001
L2正则化惩罚因子	0.001
稀疏惩罚项权重系数 $β$	0.2
稀疏常数 $ρ$	0.04
一次训练选取样本数	10
输出层激活函数	sigmoid
输入层及隐层激活函数	ReLU

Table 4 Parameter setting of autoencoder

参数	数值
学习率	0.001
L2正则化惩罚因子	0.001
稀疏惩罚项权重系数 $β$	0.2
稀疏常数 $ρ$	0.04
一次训练选取样本数	10
输出层激活函数	sigmoid
输入层及隐层激活函数	ReLU

Fig.11 Improved autoencoder network

Fig.12 Experimental result changes with dimension of U layer

Fig.13 Experimental result changes with dimension of h layer

Table 5 Comparison of training data with or without Mahalanobis distance

网络	Acc	Pre	Rec	F1
不加入马氏距离	0.875	0.875	1.000	0.934
加入马氏距离	0.976	0.973	1.000	0.986

Fig.14 Traditional autoencoder network

Fig.15 Comparison of autoencoder loss

Table 6 Comparison of autoencoder training time

网络	一次训练所需平均时间/μs	训练次数	训练所需总时间/μs
传统AE	113	50	5 650
改进AE	142	15	2 130

Fig.16 Comparison of autoencoder network loss

Fig.17 Comparison of autoencoder network accuracy

Table 7 Comparison of autoencoder network training time

网络	一次训练所需平均时间/μs	训练次数	训练所需总时间/μs
传统AN	121	60	7 260
改进AN	162	30	4 860

Table 8 Comparison between traditional autoencoder network and improved autoencoder network

网络	Acc	Pre	Rec	F1
传统AN	0.927	0.923	1.000	0.960
改进AN	0.989	1.000	0.979	0.989

Table 9 Experimental results comparison under different amounts of training data

训练数据量	网络	Acc	Pre	Rec	F1
20%	传统AN	0.883	0.882	1.000	0.937
20%	改进AN	0.891	0.889	1.000	0.941
40%	传统AN	0.870	0.869	1.000	0.930
40%	改进AN	0.934	0.931	1.000	0.964
60%	传统AN	0.869	0.867	1.000	0.929
60%	改进AN	0.942	0.940	1.000	0.968
80%	传统AN	0.890	0.888	1.000	0.941
80%	改进AN	0.977	0.976	0.999	0.987

Fig.18 Experimental result changes with amount of training data

Table 10 Comparison of anomaly detection algorithms with proposed method

检测方法	Acc	Pre	Rec	F1
iForest	0.644	0.606	0.998	0.754
KNN	0.931	0.927	1.000	0.962
LOF	0.920	0.965	0.885	0.923
SVM	0.960	0.964	1.000	0.981
K-means	0.968	0.982	1.000	0.981
DNN	0.976	0.973	1.000	0.986
本文方法	0.995	0.991	1.000	0.995

Fig.19 Bearing vibration data from dataset 2

Fig.20 Bearing vibration data from dataset 3

Table 11 Size of 3 datasets

数据集	标签	类别名	样本数	数据占比/%
数据集2	0	正常数据	1 464	75.3
数据集2	1	异常数据	479	24.7
数据集3	0	正常数据	5 936	93.9
数据集3	1	异常数据	388	6.1
XJTU-SY数据集	0	正常数据	29 124	78.8
XJTU-SY数据集	1	异常数据	7 836	21.2

Table 12 Experimental results on 3 datasets

数据集	Acc	Pre	Rec	F1
数据集2	0.985	1.000	0.980	0.990
数据集3	0.993	0.994	0.998	0.996
XJTU-SY数据集	0.984	0.982	1.000	0.991

References 21

[1]	LEI Y G, LI N P, GUO L, et al. Machinery health prognostics: a systematic review from data acquisition to RUL prediction[J]. Mechanical Systems and Signal Processing, 2018, 104:799-834. DOI URL
[2]	毛文涛, 田思雨, 窦智, 等. 一种基于深度迁移学习的滚动轴承早期故障在线检测方法[J/OL]. 自动化学报 [2020-05-10]. https://doi.org/10.16383/j.aas.c190593.
	MAO W T, TIAN S Y, DOU Z, et al. A new deep transfer learning-based online detection method of rolling bearing early fault[J/OL]. Acta Automatica Sinica [2020-05-10]. https://doi.org/10.16383/j.aas.c190593.
[3]	LING L, CUI X, WANG Y G, et al. A novel switching unscented Kalman filter method for remaining useful life prediction of rolling bearing[J]. Measurement, 2019, 135:678-684. DOI URL
[4]	程艳云, 张守超, 杨杨. 基于大数据的时间序列异常点检测研究[J]. 计算机技术与发展, 2016, 26(5):139-144.
	CHENG Y Y, ZHANG S C, YANG Y. Research on the detection of outliers in time series based on big data[J]. Computer Technology and Development, 2016, 26(5):139-144.
[5]	GUO W Y, JI Y, LUO Y, et al. Substation equipment 3D identification based on KNN classification of subspace feature vector[J]. Journal of Intelligent Systems, 2019, 28(5):807-819. DOI URL
[6]	刘芳, 齐建鹏, 于彦伟, 等. 基于密度的Top-n局部异常点快速检测算法[J]. 自动化学报, 2019, 45(9):1756-1771.
	LIU F, QI J P, YU Y W, et al. A fast algorithm for density-based top-n local outlier detection[J]. Acta Automatica Sinica, 2019, 45(9):1756-1771.
[7]	SCHÖLKOPF B, WILLIAMSON R C, SMOLA A J, et al. Support vector method for novelty detection[C]// Proceedings of the 12th International Conference on Neural Information Processing Systems, Denver, Nov 29-Dec 4, 2000. Cambridge: MIT Press, 2000: 582-588.
[8]	LIU F T, TING K M, ZHOU Z H. Isolation forest[C]// Proceedings of the 8th IEEE International Conference on Data Mining, Pisa, Dec 15-19, 2008. Washington: IEEE Computer Society, 2008: 413-422.
[9]	JIA F, LEI Y G, GUO L, et al. A neural network constructed by deep learning technique and its application to intelligent fault diagnosis of machines[J]. Neurocomputing, 2018, 272:619-628. DOI URL
[10]	RUMELHART D E, HINTON G E, WILLIAMS R J. Learning representations by back-propagating errors[J]. Nature, 1986, 323(6088):533-536. DOI URL
[11]	HINTON G E, SALAKHUTDINOV R R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786):504-507. DOI URL
[12]	NG A. Sparse autoencoder[R]. CS294A Lecture Notes, 2011: 1-19.
[13]	XU J, XIANG L, LIU Q S, et al. Stacked sparse autoencoder (SSAE) for nuclei detection on breast cancer histopathology images[J]. IEEE Transactions on Medical Imaging, 2016, 35(1):119-130. DOI URL
[14]	JIN W. A deep autoencoder based outlier detection for time series[C]// Proceedings of the 2018 3rd International Conference on Computer Science and Information Engineering, Xi’an, Sep 21-22, 2018. Piscataway: IEEE, 2018: 305-309.
[15]	张西宁, 向宙, 唐春华. 一种深度卷积自编码网络及其在滚动轴承故障诊断中的应用[J]. 西安交通大学学报, 2018, 52(7):1-8.
	ZHANG X N, XIANG Z, TANG C H. A deep convolutional auto-encoding neural network and its application in bearing fault diagnosis[J]. Journal of Xi’an Jiaotong University, 2018, 52(7):1-8.
[16]	IMANI M. Difference-based target detection using Mahala-nobis distance and spectral angle[J]. International Journal of Remote Sensing, 2019, 40(3/4):811-831. DOI URL
[17]	THEODORIDIS S. Neural networks and deep learning[M]// Machine Learning. Orlando: Academic Press, 2016: 875-936.
[18]	KULLBACK S, LEIBLER R A. On information and sufficiency[J]. The Annals of Mathematical Statistics, 1951, 22(1):79-86. DOI URL
[19]	COATES A, NG A Y, LEE H. An analysis of single-layer networks in unsupervised feature learning[C]// Proceedings of the 14th International Conference on Artificial Intelligence and Statistics, Fort Lauderdale, Apr 11-13, 2011: 215-233.
[20]	TOBON-MEJIA DIEGO A, MEDJAHER K, ZERHOUNI N. A data-driven failure prognostics method based on mixture of Gaussians hidden Markov models[J]. IEEE Transactions on Reliability, 2012, 61(2):491-503. DOI URL
[21]	WANG B, LEI Y G, LI N P, et al. A hybrid prognostics approach for estimating remaining useful life of rolling element bearings[J]. IEEE Transactions on Reliability, 2020, 69(1):401-412. DOI URL