Anomaly detection in industrial sensor data is challenging as sensor readings are frequently affected by routine operations, leading to sudden changes that may not indicate actual issues. This makes it difficult to distinguish between normal and anomalous behavior. With a few expert-labeled anomalies, we aim to leverage these sparse labels to improve sensor anomaly detection. Besides the issue of limited labels, since these labels are collected from heterogeneous sensors across different machines, we need a framework that can learn general anomaly patterns across sensors and then adapt to the unique behavior of each individual sensor. In this paper, we propose a weakly-supervised multi-sensor anomaly detection (WMAD) framework that leverages deep networks, including foundation models, to construct a data-enclosing hypersphere, effectively separating normal from anomalous time windows. By incorporating two-level importance sampling and meta-learning, WMAD effectively handles both label sparsity and sensor heterogeneity. The experiment shows that our method outperforms state-of-art competing methods on both a large proprietary industrial amperage dataset spanning over 700K hours of time-series data from Amazon and a public telemetry dataset.