With the continuous development of information technology and manufacturing level, large-scale machinery and equipment continue to develop towards high precision, high efficiency, and diverse and complex working conditions. Large machinery and equipment work under extreme conditions (high temperature, overload) and are prone to failures, resulting in serious economic losses. Prediction and Health Management (PHM) is an important technical means to ensure the reliability, safety and economy of operating equipment. It can monitor the operating status of key components of equipment in real time, so as to achieve equipment health management, prevent failures, and reduce equipment costs. The risk of accidents caused by failures, thereby saving maintenance costs and improving production efficiency.
In view of this, this paper aims to summarize the remaining service life of bearings in recent years
The research status of (RUL) prediction methods, and introduce the prediction methods of bearing RUL in detail: RUL prediction method based on physical model, data-driven RUL prediction method and fusion model RUL prediction method. By comparing the advantages and disadvantages of various methods, it is expected that There is an understanding of various bearing RUL prediction methods, and finally the future development trend is prospected.
1 Bearing RUL prediction method
There are many methods for predicting bearing RUL. However, different methods have their own advantages. There is no universal model for the so-called optimal bearing RUL prediction. RUL’s methods fall into 3 different categories, as shown in Figure 1.
1.1 Prediction methods based on physical models
The prediction method based on physical model usually needs to understand the failure mechanism of the measured object, and needs to master some knowledge of the failure mechanism of the mechanical system caused by various internal or external factors such as material properties, load and temperature, and then establish a physical description of the failure mechanism. Model, and verify the accuracy of the model through a large number of experiments and tests. As the failure mechanism is further understood, the model is continuously improved or revised to improve prediction accuracy. There are many physical models used for remaining life prediction. For example, Li et al. proposed an adaptive bearing life prediction method based on the Parise crack propagation model. Lei et al. used an improved Paris-Erdogan
The model describes the degradation process of the bearing and predicts the bearing RUL based on the particle filter algorithm. Using the Forman crack propagation law, Kirubarajan et al. proposed a life prediction model based on linear elastic fracture mechanics to calculate the remaining number of cycles when the bearing fails. Li et al.  improved the shortcomings of the exponential model, and used particle filtering to reduce the random error of the random process and realize the life prediction of rolling bearings. Considering that the Lundberg-Palmgren (L-P) model tends to underestimate bearing fatigue life, especially in low-load applications, Wei et al. proposed a new stress-based fatigue life model based on this model.
The above is the RUL prediction method based on the physical model. The key of this method is to establish a physical model or mathematical model that accurately describes the bearing degradation, so as to ensure a good prediction effect.
1.2 Data-driven forecasting methods
Currently, data-driven prediction methods are widely used in bearing RUL due to their universality and high accuracy. Data-driven methods can be divided into the following 3 types.
1.2.1 Statistical model method
When the statistical model method is used to predict the remaining service life of the bearing, the statistical model is modeled according to the theory of probability and statistics and the parameter evaluation is carried out using the full life cycle data of the bearing to obtain the degradation state of the bearing, thereby realizing the prediction of the bearing RUL. Kalman filter (KF) and particle filter (PF) are two common algorithms, which are widely used in statistical model methods. Qian et al.  proposed a method combining autoregressive model and Kalman filter, and the experimental results show that the method has better prediction results. Wen Juan et al. used a stochastic process model to model the bearing degradation process and used the unscented filtering algorithm to update the model parameters. This method reduces the particle degradation degree to a certain extent and improves the bearing life prediction accuracy. By improving the filtering algorithm, Wang Xiaotang proposed a multi-strategy co-evolutionary bearing life prediction method, which improved the particle weight and particle diversity, thereby improving the accuracy of bearing RUL prediction.
The statistical model method is to obtain the remaining service life by calculating the corresponding degradation index, and the calculation is relatively simple, but this type of model cannot modify the parameters of the model according to the change of the bearing or operating conditions, and the prediction result is poor.
1.2.2 Stochastic Model Method
The commonly used stochastic process model is Gaussian stochastic process model (Gaus-
A stochastic process model that describes various bearing failure degradation processes. The Gaussian stochastic process model is a cumulative damage process. Aye et al. proposed an optimized Gaussian process regression model through the combination of a single mean and a covariance function to predict the remaining life of low-speed bearings and achieve a low percentage error prediction of bearing RUL.
The Wiener process model, also known as the Brownian motion model with linearity, characterizes a non-monotonic degradation process and has been applied to model degradation processes and predict RUL of various equipment components. Jin Xiaohang et al. proposed a two-dimensional Wiener process bearing RUL prediction method, which effectively predicted the bearing’s RUL and had better prediction accuracy than the one-dimensional Wiener process model. Wang et al. proposed an improved Wiener process model with adaptive drift and diffusion for online RUL prediction models, which described the degradation process more accurately than other existing RUL prediction models. Stochastic models can describe the bearing health degradation process well, however, it is still a challenge to achieve accurate life prediction when the model takes into account the influence of multiple factors in the actual bearing operating environment.
1.2.3 Machine Learning Methods
As the core technology of artificial intelligence, machine learning, with its powerful learning ability, is very suitable for modeling the degradation model of complex mechanical systems in the era of big data, and is widely used in RUL prediction. Common machine learning methods include support vector machines, neural networks, and deep learning methods. The following summary
Several bearing RUL prediction methods based on machine learning methods are described.
Sun et al. proposed a bearing life prediction method based on support vector machine (SVM), which uses principal component analysis (PCA) to extract features from vibration signals, and optimizes the parameters of SVM through particle swarm optimization (PSO) to predict The results show that the method is more accurate. Liu et al. applied SVM to mechanical state prediction. Considering that the collected vibration data usually has multi-scale features, wavelet transform (WT) was introduced into the SVM model to reduce the influence of irregular features to simplify the original signal. The WT-SVM model is used to model the vibration signal of the double row bearing, and the experimental results show that the WT-SVM model is better than the single SVM model. Xi Lifeng et al.] used three time domain features and three frequency domain features and the newly proposed self-organizing map network (SOM) minimum quantization error feature to predict the remaining life of bearings based on the self-organizing map neural network model. Ren et al. used a deep convolutional neural (CNN) to realize the RUL of the bearing, extracted 64 frequency domain features as the input of the CNN, and trained it. The model structure has a total of 8 layers, and the method has higher accuracy.
In recent years, deep learning, a new branch of machine learning, has attracted more and more attention from researchers in the field of bearing fault diagnosis, and has also been recognized as a powerful tool for bearing health monitoring. It has broad applications in bearing life prediction and equipment failure prediction. space.
1.3 Prediction method of fusion model
The fusion model method combines 2 or more methods to predict the RUL of the bearing, overcomes the limitations of the traditional RUL method, maximizes the advantages of various prediction methods, and further improves the accuracy of the prediction results. Two methods of fusing the models are described below.
1.3.1 Method based on physical model and data-driven fusion
The method based on physical model and data-driven fusion is to fuse the above two prediction methods to improve the prediction accuracy. Yang Si combined the vehicle system dynamics model, the wheel-rail contact model and the optimized
Based on the Jendel wear model, a data-driven-model fusion-driven RUL prediction method is constructed to realize wheel wear prediction. This method improves the chemical resistance of the model and reduces the data and experiment costs. In order to bridge the gap between mechanism and data and take advantage of it, Zeng et al. proposed a new physics-based data-driven wheel wear modeling and prediction method. The model, based on the numerical prediction method of wheel degradation with alternating wear and deformation closed-loop, further evaluates the remaining useful life (RUL) of the wheel through point estimation and interval estimation.
1.3.2 Multiple data-driven fusion methods
A variety of data-driven fusion methods are to fuse different types of data-driven methods. This method can improve the generalization of the model and the prediction accuracy is greatly improved. Zhang et al. proposed a bearing residual service life prediction method based on Naive Bayes and Weibull distribution. The method uses Weibull distribution (WD) algorithm to fit the fluctuation characteristics of bearings in different decay stages. This feature is used as Naive Bayes ( NB) input to train the model, the experimental results show that the method is effective for bearing RUL prediction. Deutsch et al. proposed a method for RUL prediction of hybrid ceramic bearings integrating deep belief networks and particle filters.  proposed a bearing residual life prediction method based on SVM and artificial neural network (ANN), which is compared with random forest.
(RFR) model, lasso model (LASSO) and SVM model have higher
A variety of data-driven fusion methods give full play to the advantages of each model and solve the problem of remaining life prediction of equipment with different characteristics. This method has great potential and will become a research hotspot of bearing life prediction in the future.
Thanks for the article!