Self-supervised Health Representation Decomposition based on contrast learning

Accurately predicting the Remaining Useful Life (RUL) of equipment and diagnosing faults (FD) in Prognostics and Health Management (PHM) applications requires effective feature engineering. However, the large amount of time series data now available in industry is often unlabeled and contaminated by variable working conditions and noise, making it challenging for traditional feature engineering methods to extract meaningful system state representations from raw data. To address this issue, this paper presents a Self-supervised Health Representation Decomposition Learning(SHRDL) framework that is based on contrast learning. To extract effective representations from raw data with variable working conditions and noise, SHRDL incorporates an Attention-based Decomposition Network (ADN) as its encoder. During the contrast learning process, we incorporate cycle information as a priori and define a new loss function, the Cycle Information Modified Contrastive loss (CIMCL), which helps the model focus more on the contrast between hard samples. We evaluated SHRDL on three popular PHM datasets (N-CMAPPS engine dataset, NASA, and CALCE battery datasets) and found that it significantly improved RUL prediction and FD performance. Experimental results demonstrate that SHRDL can learn health representations from unlabeled data under variable working conditions and is robust to noise interference. •SHRDL tackles challenges in PHM: unlabeled data, varying conditions, and noise.•CIMCL loss leverages cycle info for dynamic contrast weights in pre-training.•Attention-based Decomposition Network (ADN) captures system state effectively.•Self-supervised learning framework adapts to unlabeled, noisy sensor data.•Delivers superior RUL prediction & FD performance on popular PHM datasets.