Abstract:Accurate prediction of the Remaining Useful Life (RUL) of rolling bearings is crucial in industrial production, yet existing models often struggle with limited generalization capabilities due to their inability to fully process all vibration signal patterns. We introduce a novel multi-input autoregressive model to address this challenge in RUL prediction for bearings. Our approach uniquely integrates vibration signals with previously predicted Health Indicator (HI) values, employing feature fusion to output current window HI values. Through autoregressive iterations, the model attains a global receptive field, effectively overcoming the limitations in generalization. Furthermore, we innovatively incorporate a segmentation method and multiple training iterations to mitigate error accumulation in autoregressive models. Empirical evaluation on the PMH2012 dataset demonstrates that our model, compared to other backbone networks using similar autoregressive approaches, achieves significantly lower Root Mean Square Error (RMSE) and Score. Notably, it outperforms traditional autoregressive models that use label values as inputs and non-autoregressive networks, showing superior generalization abilities with a marked lead in RMSE and Score metrics.
What problem does this paper attempt to address?
### Problems the Paper Attempts to Solve
This paper aims to address the issue of insufficient generalization ability in the prediction of the remaining useful life (RUL) of rolling bearings. Existing models often fail to fully capture all information when processing vibration signal patterns, leading to limited predictive performance in practical applications. Specifically, the paper proposes a new multi-input autoregressive model that outputs the health indicator (HI) value of the current window by integrating vibration signals and previously predicted HI values. This model obtains a global receptive field through autoregressive iteration, effectively overcoming the generalization limitations of existing models. Additionally, the paper introduces a segmentation method and multiple training iterations to reduce error accumulation in the autoregressive model.
### Background and Motivation
In modern industrial production, fault diagnosis and health management (PHM) of mechanical equipment are crucial for improving the reliability and efficiency of factory operations. Rolling bearings, as key components of modern production equipment, have high requirements for reliability and safety and are one of the main causes of mechanical failures. Therefore, predicting and preventing bearing failures and estimating their remaining useful life (RUL) is particularly important. However, due to various factors such as workload, operating mode, material defects, and operating temperature, the RUL prediction task is very challenging.
### Limitations of Existing Methods
Currently, RUL prediction methods mainly include model-based methods, data-driven methods, and hybrid methods. Compared to model-based methods, data-driven methods can learn degradation patterns from sensor signals through machine learning or deep learning to establish predictive models. With the rapid development of sensing technology, obtaining large amounts of industrial data (such as vibration signals) has become increasingly easy. When the data volume is large, data-driven methods significantly outperform the other two methods.
### Proposed New Method
The paper proposes a new multi-input autoregressive model, which is unique in the following ways:
1. **Multi-input Fusion**: It not only processes vibration signals as input but also integrates previously predicted health indicator (HI) values, outputting the HI value of the current window through feature fusion.
2. **Autoregressive Iteration**: Through autoregressive iteration, the model obtains a global receptive field, capable of remembering early degradation signals and maintaining memory even in the later degradation stages.
3. **Segmentation Method**: The complete bearing data is divided into multiple independent samples, each trained separately to reduce error accumulation.
4. **Multiple Training Iterations**: Multiple training iterations are performed at the initial time point to minimize errors, followed by standard single training iterations for subsequent time points.
### Experimental Validation
The paper conducts empirical evaluations on the PMH2012 dataset. The results show that, compared to other backbone networks using similar autoregressive methods, traditional autoregressive models, and non-autoregressive networks, this model demonstrates significant advantages in terms of root mean square error (RMSE) and score metrics, indicating stronger generalization ability.
### Main Contributions
1. **Proposed a New Multi-input Autoregressive Model**: This model can integrate vibration signals and previously predicted health indicator (HI) values, outputting the HI value of the current window through feature fusion.
2. **Adopted Segmentation Method**: Reduces error accumulation in the autoregressive model.
3. **Innovative Multiple Training Process**: Addresses the initial point prediction bias issue.
4. **Empirical Validation**: Experimental results on the PMH2012 dataset show that this model significantly outperforms existing methods in terms of generalization ability and prediction accuracy.