Incorporating prior knowledge into self-supervised representation learning for long PHM signal

Prognostics and Health Management (PHM) is a discipline that monitors, diagnoses, and predicts the health status of complex systems. Representation learning plays a pivotal role in converting sensor data into informative features that accurately reflect the system's health status. However, PHM datasets pose unique challenges, including long and nonlinear signal data, limited quantity, and inferior quality. To overcome these challenges, we introduce the Prior Knowledge Enhanced Self-Supervised Learning Framework (PKESSLF). PKESSLF employs the Patch Long Signal Transformer (PLST) as an encoder, effectively mitigating the impact of signal nonlinearity and nonstationarity. This is achieved by maintaining local features through a patch-based approach, resulting in improved efficiency when handling long signals.Furthermore, the framework integrates self-supervised learning enhanced by prior knowledge (PKESSL), synchronizing masked pretraining and contrast learning to align data representation with prior knowledge. PKESSL adeptly captures localized and broad-ranging characteristics inherent in the data. The inclusion of a Prior Attention Pooler assigns dynamic weights to signal patches based on prior knowledge, enhancing the model's aggregation effect on long sequences. PKESSLF achieves effective modeling and extraction of system representation information, providing a flexible and efficient approach for pretraining and finetuning PLST in scenarios with functional prior knowledge. Incorporating prior knowledge holds great promise in addressing the limitations of data-driven methods in PHM, as evidenced by significant performance enhancements on the N-CMAPPS datasets.