ZSTT: A Zero-Shot Time-Series Prediction Model Based on Transformer and Its Application to Hydraulic Actuator Performance Degradation Prediction
Jiashan Gao, Shaoping Wang, Rentong Chen, Chao Zhang, Enrico Zio, Yuwei Zhang, Yaohui Lu
Abstract
Most of the existing time-series prediction methods depend on large training datasets, which restricts their applicability in industrial applications with limited or no target-system data. This occurs in various instances of performance degradation prediction based on multisource sensor data. To address this issue, this article proposes a zero-shot time-series prediction (ZSTT) model based on the transformer framework, referred to as ZSTT. It enables multisource time-series prediction without requiring pretraining on the target dataset. Approximate Bayesian inference is first considered to increase the generalization ability of the model. A learnable pattern module, which employs a Gumbel-distribution-based dynamic masking mechanism, is then introduced to optimize joint training across multisource datasets. Finally, a real degradation test of the hydraulic actuator is conducted to assess the performance of ZSTT. Its validation through comprehensive experiments is also conducted on benchmark datasets. A comparison with state-of-the-art methods is conducted, demonstrating the performance of ZSTT in terms of higher accuracy and robustness of prediction. Furthermore, ablation studies were performed to validate the contribution of each individual module. The case study demonstrates the high performance of ZSTT in multisensor data prediction and performance degradation prediction.