A Controlled Synthetic Benchmark for Multi-Label Transformation Detection in Signal and Trajectory Pairs

Comparing paired signals and trajectories is commonly framed as distance or alignment estimation, whereas controlled diagnostics may require identifying concurrent transformations. We formulate the supervised prediction of 19 applied-transformation labels, which need not all produce observable changes, and evaluate 5 1D and fifteen axis-wise 2D scenario profiles. Each profile uses 12,000 pairs of length-256 sequences per seed across five seeds (42, 142, 242, 342, and 442). A classifier using the full feature set is compared with a classifier using nine established similarity/alignment measures, and four raw-pair deep models are evaluated in two 2D X+Y scenarios. Across the 20 profiles, the unweighted mean Macro-F1 values for full and similarity-only features were, respectively, 0.8193 and 0.5677 on validation, 0.8102 and 0.5589 on test samples from training-represented families, and 0.7381 versus 0.4566 on the specified held-out synthetic families—pulse and trend_mix in 1D, and spiral alone in 2D. In 1D, the held-out-family Macro-F1 for full features decreased from 0.9317 for one transformation to 0.6536 for four. ResNet1D had the highest mean held-out-family Macro-F1 among the four evaluated deep models, reaching 0.7779 and 0.4961 in the single-transformation and mixed zero-to-four-transformation X+Y scenarios, respectively, compared with 0.8542 and 0.7026 for full features. A separate subject-disjoint, twelve-label evaluation used injected transformations on observational base trajectories from the 2014 white stork dataset; seed-42 test Macro-F1 scores were 0.7285, 0.7234 and 0.7267 for X-only, Y-only and axis-wise X+Y, respectively, on 255 windows from 17 white storks excluded from training and validation. These controlled labels do not establish natural-label field performance. The results establish a controlled benchmark for transformation-label prediction; generalization is restricted to these synthetic holdouts and does not establish performance on arbitrary unseen trajectory families.