<p>Time series motifs are repeating patterns in time series data. They have emerged as important primitives for exploring massive time series and are also useful as a subroutine in existing time series data mining tasks. In multivariate data, motifs often occur only in a subspace, i.e., a subset of the attributes, yet many existing methods search for motifs in the full-dimensional space. Moreover, these methods often assume perfect alignment in time, ignoring potential temporal variations across attributes. We propose SubTSMD, a novel method for discovering subspace motifs in multivariate time series data. SubTSMD first discovers motifs independently within each attribute and then incrementally combines them in a bottom-up fashion to form subspace motifs. Unlike prior approaches, SubTSMD allows for non-exact temporal matches across attributes, enabling more robust motif discovery in real-world applications. Additionally, we introduce the first evaluation metric for subspace motif discovery that assesses whether the discovered and ground truth motifs span the same subspaces, unlike existing metrics, which only evaluate temporal alignment. SubTSMD significantly outperforms existing motif discovery methods on a newly curated benchmark for subspace motif discovery, comprising 6 synthetic and 14 real-world classification datasets, each containing 250 time series, for a total of 5000 time series. We showcase how SubTSMD uncovers meaningful patterns in real-world applications and robustly discovers motifs in the presence of noise.</p>