Unsupervised Transform-View Alignment for Raw-Only Time Series Representation Learning

Deterministic transforms reorganize time series so that periodicity, recurrence geometry, and repeated motifs become easier to isolate. We introduce Multi-transform Multi-view Feature Alignment (MMFA), which uses the resulting auxiliary views to train a raw representation. During pretraining, the frequency view, image views, and symbolic view pass through auxiliary encoders. Raw-anchor alignment then pairs each auxiliary representation with the raw representation by sample identity. Encoder-level variance regularization preserves aggregate spread in the raw representation. Projection-head variance regularization maintains spread in the projected representations, and projection-head covariance regularization reduces their redundancy. Inference uses only the raw encoder. For a finite batch, the exact raw-anchor disagreement identity shows that raw-anchor alignment controls average pairwise disagreement among all views within a factor of two. Finite-loss raw-representation spread bounds further link auxiliary-view spread and raw-anchor alignment loss to aggregate spread in the raw representation. These bounds are independent of representation dimension. Across 30 UEA classification datasets, MMFA achieves the strongest average rank among unsupervised representation learning methods and the best accuracy on 21 datasets. In single-transform and leave-one-out ablations, full MMFA ranks 3.02 versus 5.25--7.55 for every reduced variant, with Holm-corrected p <= 0.003, demonstrating that the views contribute jointly. The raw encoder reaches 0.6241 F1 on UCR anomaly detection and records nine wins and one tie against CSL on each clustering metric across 12 subsets. Together, these results connect auxiliary-view structure and raw-anchor alignment with encoder-level variance regularization and projection-head variance and covariance regularization.