Sensor-based remote healthcare monitoring can be used for the timely detection of adverse health events in people living with long-term health conditions, and help reduce preventable hospitalization. Current anomaly detection approaches in a real-world setting are challenged by noisy data and unreliable event annotation. Inspired by the conceptual simplicity and recent applications of negative sample-free contrastive learning in computer vision, we propose a lightweight, self-supervised model to extract noise-adaptive representations from multidimensional sensor data. We use the contrastive loss between the more granular observation data and a corresponding learnable, lower temporal resolution augmentation, and use the learned representations for anomaly detection. Learning to adjust this "contrast factor" enables the model to identify and leverage the most informative temporal features at different scales, enhancing its ability to discern underlying patterns amidst noise. Our model outperformed comparable representation learning algorithms in detecting agitation and fall events across three distinct participant cohorts in a real-world study of people living with dementia in their homes. We further used the representations to create a "spatiotemporal attention map” to focus on the source of anomaly and offer explainability. Our approach is domain-agnostic and can be used in wider healthcare, industrial and urban sensor settings.