Distributed acoustic sensing (DAS) is well-suited to detecting induced seismicity, but routine processing must cope with terabytes of data, heterogeneous backgrounds, and severe class imbalance. Classical STA/LTA—originally devised for single seismometers—requires adaptation to massive sensor arrays and modern acquisition geometries. We present a practical workflow that augments multi-channel coincidence STA/LTA with positive–unlabeled (PU) learning and distance-based ranking. From each trigger we compute compact temporal and spatial attributes tailored to DAS (e.g., temporal consistency/uniformity, spatial continuity, channel-gap statistics, coincidence efficiency), apply log transforms to mitigate heavy tails, and robustly scale with respect to the positive class. Without explicit negatives, we learn the geometry of the positive manifold from a small set of labeled microseismic examples and assign each unlabeled trigger an ensemble similarity score that combines distance to the positive centroid with k-nearest-neighbour proximity in standardized feature space. The system operates section-wise for long arrays with chunked I/O and caching, yielding auditable rankings that reduce tens of thousands of triggers to a few hundred candidates for expert review. We demonstrate the workflow on subsea DAS data and outline its extension to other acquisition geometries, including downhole and surface deployments. The approach is conservative, scalable, and interpretable: it prioritizes events similar to known positives while avoiding over-commitment to assumed negatives, and it is designed to accommodate future incorporation of geometry-specific priors as catalogues grow.