• Proposes a multisensor contrast neural network that leverages cross-sensor similarity to mine health-indicative representations from unlabeled sensor data.
• Uses ResNet18 to map multisensor features into a co-occurrence space and applies alternate contrast learning to capture degradation stages.
• Achieves accurate RUL prediction with scarce labeled data by fine-tuning similar representations, reducing dependency on expensive failure data.
• On a public bearing dataset, reduces mean absolute percentage error by at least 0.058 and improves score by at least 0.122 over state-of-the-art methods.