Observations Drift, Structures Remain: Structural Pretraining with Time Alignment for Electromagnetic Signals

Wenjin Gui, Junyu Shen, Haibo Xu, Yuchuang Sun, Luqing Luo, Wupeng Xie, Qian Huang, Fengxiang Wang, Yangang Sun, Maosong SunarXiv 2026 · arXiv

TL;DR

Current self-supervised pretraining for electromagnetic (EM) signals often fails because models learn to "cheat" by focusing on observation-specific noise and sampling rates rather than the underlying signal structure. This paper introduces a framework that aligns signals in physical time and replaces raw waveform reconstruction with the prediction of discrete "structural units," leading to significantly better generalization across diverse wireless tasks and hardware settings.

What Problem Does This Paper Address?

In the world of wireless communication and radar, we are increasingly using foundation-model-style pretraining on massive amounts of I/Q (In-phase and Quadrature) data. However, an EM signal is a cocktail of three things: the original source structure (what we want to learn), the propagation environment (noise and fading), and the acquisition process (sampling rates).

The authors identify a failure mode they call observation-biased pretraining. When models are trained to simply reconstruct raw waveforms (like a Masked Autoencoder), they often take "shortcuts." They find it easier to minimize loss by learning the specific sampling rate or the unique noise profile of a dataset rather than the reusable signal structures. As shown in the paper's diagnosis, standard pretrained models tend to cluster representations by which dataset they came from rather than what modulation they are using.

Figure 1: Motivation and diagnosis of observation-biased EM pretraining. Heterogeneous EM observations arise from source generation, propagation, and acquisition, causing appearance and temporal mismatches. Reconstruction objectives can easily exploit these observation-specific cues, leading representations to organize by observation conditions rather than reusable structures. The right panels summarize the resulting sampling-rate extrapolation degradation and the improved downstream performance achieved by our pretraining method.

Methods

To break these "observational shortcuts," the authors propose two main innovations:

  1. RoPE-fs (Physical-Time-Aware RoPE): Standard positional encodings use discrete token indices (1, 2, 3...). But in EM signals, "index 10" means something different if you sample at 1 MHz versus 10 MHz. RoPE-fs normalizes these indices into physical-time coordinates. This ensures that a specific temporal pattern (like a pulse width) looks the same to the model regardless of the sampling rate used to capture it.
  2. Structural Unit Discovery: Instead of predicting continuous waveform values, the model predicts discrete labels.
  3. Atomic Units: The authors use K-means clustering on signal patches to create a base "alphabet" of local waveform primitives.
  4. Emergent Units (ESUs): Using a technique borrowed from Natural Language Processing called Byte-Pair Encoding (BPE), they merge frequent sequences of atomic units into longer, variable-length "words" that capture complex signal compositions.

The final objective is Masked Multi-level Structural Unit Prediction: the model sees a corrupted signal and must guess which atomic and emergent units are hidden behind the mask.

Figure 3: Overview of structural pretraining with time alignment. I/Q patches are encoded with RoPE-fs, while masked prediction of atomic and emergent structural-unit labels provides multi-level structural supervision instead of raw waveform reconstruction.

Datasets

The authors pretrain their model on a balanced subset of 1.62 million samples from the EMind suite, which includes diverse sources like RadarCommDataset, WiSig, and HisarMod2019.1.

For evaluation, they test across nine benchmarks covering:

  • Automatic Modulation Classification (AMC): RML2016.10A/B/04C, RML2018.01A.
  • Radar Waveform Classification (RWC): Panoradio.
  • Wireless Interference Identification (WII): RadChar, EM-Infer-Comm.
  • Radio Frequency Fingerprinting (RFFI): EM-AIS.

Key Insights and Ideas

  • Reconstruction is too easy: In heterogeneous data, the "noise" or "sampling grid" is often more predictable than the actual signal structure. By moving to discrete targets, the model is forced to ignore low-level variations and focus on recurring structural patterns.
  • Signals have a "Vocabulary": By using BPE, the authors show that EM signals can be decomposed into a hierarchy of units, much like letters forming words. Visualizing these units reveals they capture specific physical behaviors, like phase rotations or amplitude edges.
  • Time Alignment is Critical: Without RoPE-fs, a model trained on one sampling rate is essentially "blind" when tested on another. Normalizing to physical time allows for sampling-rate extrapolation, where a model performs well on hardware settings it never saw during training.

Key Results

The proposed method consistently outperformed existing EM foundation models like SpectrumFM and EMind across all benchmarks:

  • Full Supervision: Achieved significant gains, particularly in challenging tasks like RFFI (improving accuracy by +11.83%) and Radar classification (+13.16%).
  • Few-Shot Learning: When only 50-100 samples per class were available, the method showed robust transferability, outperforming baselines by up to 4 percentage points.
  • Robustness: The model maintained high performance even when the sampling rate metadata was slightly inaccurate (up to 10% error), proving the stability of the RoPE-fs approach.

Limitations and Future Work

The authors note three primary limitations:

  • Two-Stage Pipeline: The structural vocabulary is discovered via clustering before the main pretraining begins. A more elegant, end-to-end approach for vocabulary learning remains an open challenge.
  • Physical Interpretability: While the units capture clear waveform patterns, their direct mapping to specific modulation components or complex physical mechanisms needs deeper investigation.
  • Granularity Selection: The sizes of the atomic and emergent vocabularies are currently chosen empirically and might not be optimal for every possible EM task.

Summary updated

Abstract

Electromagnetic signal modeling is increasingly moving toward self-supervised pretraining on heterogeneous I/Q observations. However, an observed waveform is jointly shaped by signal generation, propagation environments, and acquisition processes, which mix reusable structural regularities with observation-specific variations. Existing methods usually pursue unified pretraining with reconstruction objectives on mixed observations. This heterogeneity creates two key mismatches for reconstruction pretraining: (1) appearance mismatch, induced by propagation-dependent waveform variations; and (2) temporal mismatch, induced by acquisition-dependent sampling coordinates. These mismatches make observation-specific appearances and sampling coordinates easier to reconstruct than stable signal structures, inducing observational shortcuts. We identify this failure mode as observation-biased pretraining. To address this, we propose Structural Pretraining with Time Alignment for heterogeneous EM signals. It mitigates temporal mismatch with RoPE-fs, which maps token positions to sampling-rate-normalized physical-time positions, and reduces reliance on appearance shortcuts by replacing waveform reconstruction with masked prediction of discrete structural units. These units are organized into atomic structural units and emergent structural units that capture variable-length compositional patterns beyond fixed patch boundaries. Across diverse EM benchmarks, our method improves both full-supervision and few-shot adaptation over existing EM pretraining baselines, showing that the learned representations better preserve structural consistency when observations drift in waveform appearance and temporal scale.
BibTeX
@misc{gui2026observations,
  author = {Wenjin Gui and Junyu Shen and Haibo Xu and Yuchuang Sun and Luqing Luo and Wupeng Xie and Qian Huang and Fengxiang Wang and Yangang Sun and Maosong Sun},
  title = {{Observations Drift, Structures Remain: Structural Pretraining with Time Alignment for Electromagnetic Signals}},
  year = {2026},
  eprint = {2611.15566},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15566},
}
Figure