Multisensor contrast neural network for remaining useful life prediction of rolling bearings under scarce labeled data

Abstract

Predicting remaining useful life (RUL) of bearings under scarce labeled data is significant for intelligent manufacturing. Current approaches typically encounter the challenge that different degradation stages have similar behaviors in multisensor scenarios. Given that cross-sensor similarity improves the discrimination of degradation features, we propose a multisensor contrast method for RUL prediction under scarce RUL-labeled data, in which we use cross-sensor similarity to mine multisensor similar representations that indicate machine health condition from rich unlabeled sensor data in a co-occurrence space. Specifically, we use ResNet18 to span the features of different sensors into the co-occurrence space. We then obtain multisensor similar representations of abundant unlabeled data through alternate contrast based on cross-sensor similarity in the co-occurrence space. The multisensor similar representations indicate the machine degradation stage. Finally, we focus on fine-tuning these similar representations to achieve RUL prediction with limited labeled sensor data. The proposed method is evaluated on a publicly available bearing dataset, and the results show that the mean absolute percentage error is reduced by at least 0.058, and the score is improved by at least 0.122 compared with those of state-of-the-art methods.

1. Introduction

Remaining useful life (RUL) prediction for bearings aims to forecast the time duration from current operation until failure of the bearings (Wen et al., 2019). As a critical component of intelligent machine health management (Tao et al., 2018; Souza et al., 2021; Wang WJ et al., 2022), RUL prediction for bearings can assist in reducing maintenance costs and preventing significant losses from accidents, thereby improving competitiveness.

Traditional approaches for RUL prediction for bearings can be broadly classified as model-based and statistics-based methods. Model-based methods (Morales-Espejel and Gabelli, 2020) require extensive domain knowledge to build physical models that accurately reflect machine degradation. However, obtaining domain knowledge is challenging, and building accurate physical models is difficult due to the complex system structure and operating environment. On the other hand, statistics-based methods (Li Q et al., 2022) focus on building a stochastic model that describes the degradation process to predict RUL based on monitored machine degradation variables. Nevertheless, these methods have limited processing ability on low-quality data.

In recent years, using deep learning for RUL prediction for bearings has become a research trend due to its effectiveness in improving prediction accuracy. Data-driven approaches (Wang B et al., 2020; Wang X et al., 2021) use deep learning models to establish potential relationships between machine monitoring data and RUL labels or degradation labels. These approaches having powerful degradation feature extraction capabilities could effectively handle massive and complex structured data and reduce the need for domain knowledge. Compared to model-based and statistical-based approaches, data-driven approaches heavily rely on RUL-labeled data, posing challenges in practical applications due to the scarcity of degradation data with RUL labels.