Researchers have developed a machine learning framework that pairs generative diffusion models with convolutional neural networks to improve the accuracy of estimating the first arrival path delay in wireless communications. When estimating the delay of the earliest arriving signal path with a convolutional neural network, engineers typically face a severe data bottleneck because training reliable models requires large volumes of labeled samples that capture the diverse multipath fading conditions of physical environments. Collecting such extensive channel measurements in real-world propagation settings remains exceptionally expensive. To overcome this practical shortage, the research team designed a specialized data augmentation pipeline that converts one-dimensional cross-correlation sequences into two-dimensional images, generates synthetic training samples using a diffusion model, and decodes them back into sequence data to train convolutional estimators.
In wireless signal propagation, an emitted waveform travels across multiple paths, arriving at a receiver as a complex superposition of resolvable multipath components resulting from reflection, refraction, and scattering. Among these incoming components, the earliest arrival is designated as the first arrival path, which corresponds directly to the physical line-of-sight propagation path between the transmitter and the receiver. Because this first component experiences the lowest propagation loss and provides the highest reception quality, it serves as the foundational reference for high-precision wireless positioning, synchronization, and ranging across fifth-generation and sixth-generation communication networks. Accurate identification of this delay governs ranging precision and clock alignment in complex indoor settings and urban canyons, where signal reflections off walls and physical obstacles separate the earliest path from subsequent paths by merely a few nanoseconds.
Conventional signal processing algorithms such as the Multipath-Estimating Delay-Lock Loop, cross-correlation detectors, and matched filters often degrade substantially in dense multipath environments. Classical approaches like the delay-lock loop suffer from high computational complexity and require prior knowledge of the total number of multipath components, a parameter rarely available in dynamic scenarios. While deep learning models can automatically extract latent features from dense multipath channels without requiring prior statistical channel orders, their performance hinges on massive collections of accurately labeled training sequences. Alternative generative methods such as Generative Adversarial Networks have been explored for channel augmentation, but their adversarial training dynamics frequently suffer from instability and mode collapse, often necessitating supplementary empirical measurements to rebalance synthetic datasets.
To establish a stable generative augmentation pipeline, the authors turned to generative diffusion models, which learn target data distributions through a gradual denoising Markov chain. While diffusion architectures excel at generating high-fidelity two-dimensional image structures with low risk of mode collapse, their underlying convolutional operations rely on grid-like spatial structures that are absent in raw one-dimensional signal sequences. To resolve this structural incompatibility, the investigators introduced an encode-diffuse-decode methodology. Time-domain signal features are first obtained by computing the cross-correlation function between transmitted reference sequences and received signals. The resulting one-dimensional cross-correlation sequences are then rescaled and mathematically transformed into two-dimensional Gramian Angular Field images, casting the temporal sequences into polar coordinates where angular values preserve temporal dependencies across matrix positions.
A critical architectural innovation in this sequence-to-image encoding involves the treatment of supervisory ground truth labels. Rather than generating signal features in isolation and requiring a secondary labeling mechanism, the researchers embedded both the cross-correlation sequence values and the first arrival path delay offset label directly into a unified Gramian Angular Field matrix. In this label-inscribed self-supervised setup, the sequence features occupy specific coordinate relations while the delay label is encoded directly into pixel intensity values. During training, the generative diffusion model is forced to reconstruct both the temporal correlation structure of the channel signal and the precise gray-level spatial position of the target label. Consequently, newly sampled synthetic images naturally incorporate valid regression labels upon generation, eliminating the need for separate manual annotation or external label generation algorithms.
The generative diffusion process operates through two distinct phases: a forward noise addition process and a reverse denoising process. During forward diffusion, Gaussian noise is incrementally added to an encoded training image across discrete steps according to a defined variance schedule, eventually transforming the image into standard normal noise. In the reverse process, a neural network structured around a U-Net architecture with an attention mechanism is trained to estimate and subtract the added noise step by step. Once trained, the model draws random vectors from a standard Gaussian distribution and iteratively denoises them to produce novel, synthetic Gramian Angular Field images that reflect the underlying data distribution of the physical channel sequences without reproducing exact duplicates.
Transforming generated images back into numerical sequences suitable for convolutional neural network training requires an exact mathematical inversion. The authors utilized a mathematical property of the Gramian Angular Field transformation, wherein the diagonal elements of the image matrix directly relate to the squared values of the normalized sequence data points. By constraining the original normalized sequence values to a specific positive numerical range between zero and one, the transformation ensures a unique inverse mapping. The decoding algorithm extracts the diagonal matrix elements from each synthetic image and reconstructs one-dimensional sequence values and corresponding delay labels, creating a fully formed synthetic dataset ready for downstream model training.
The downstream estimation task relies on a custom convolutional neural network designed to infer the first arrival path delay offset from processed cross-correlation sequences. The network architecture incorporates two convolutional layers, five residual blocks, and two fully connected output layers, utilizing a linear rectified linear unit activation with a parameter of 0.02 across intermediate layers. The network receives normalized cross-correlation sequences and optimizes its parameters using mean squared error loss against the true delay offsets. By feeding the network cross-correlation sequences concentrated around detected peaks, the framework reduces processing overhead while allowing convolutional filters to learn the functional mapping between signal peak deformations and exact propagation delays.
To evaluate the framework, the research team conducted systematic numerical simulations across multiple standard wireless channel formulations, including deterministic Zadoff-Chu sequences, Rayleigh fading channels representing non-line-of-sight scattering environments, and Rician fading channels with varying line-of-sight ratios. Initial baseline datasets of 1,000 sequences were encoded into grayscale Gramian Angular Field images and processed by the diffusion model. Dimensionality reduction via t-distributed stochastic neighbor embedding and kernel density estimation verified that the synthetic sequence distributions closely matched the empirical distributions of the original training data, confirming that the diffusion model successfully captured channel characteristics without mode collapse.
Experimental optimization of the diffusion model parameters demonstrated that generation quality and subsequent delay estimation accuracy depended strongly on the diffusion step budget and noise schedule parameters. Testing across step counts revealed that synthetic samples generated using 1,000 diffusion steps yielded the highest delay estimation accuracy, whereas reducing the budget to 700 steps or fewer introduced structural distortions into the decoded sequences. Similarly, evaluating noise variance schedules indicated that ramping from a starting value of 0.0001 to a terminal value of 0.02 produced optimal synthetic data fidelity. Sensitivity sweeps across convolutional network hyperparameters further showed that a convolutional kernel size of 7 by 7 and a depth of 5 residual blocks achieved the highest delay estimation accuracy.
Comparative evaluations against insufficient baseline datasets confirmed the substantial benefits of the diffusion-based augmentation strategy. Using Zadoff-Chu sequences evaluated across signal-to-noise ratios from 6 decibels to 20 decibels, the convolutional neural network trained on an augmented dataset of 5,000 samples consistently outperformed models trained solely on the original 1,000 samples. When assessed across error thresholds ranging from 0.1 to 0.5 sampling points, the model trained on augmented data achieved higher correct estimation rates across all noise conditions. Evaluating different expansion volumes revealed that estimation accuracy improved steadily as the training set expanded from 1,000 to 5,000 samples, beyond which marginal performance gains saturated, identifying an augmentation target of approximately 5,000 to 6,000 sequences as optimal for balancing accuracy against computational cost.
Further benchmark comparisons demonstrated that the proposed convolutional architecture achieved higher delay estimation accuracy than competing convolutional recurrent neural network baselines across multiple signal-to-noise levels. Additional experiments on Rayleigh and Rician fading sequences confirmed that the generative framework transferred effectively across stochastic multipath environments, maintaining superior delay estimation accuracy even under severe noise conditions down to negative 10 decibels. The simulation findings verify that combining Gramian Angular Field encoding with generative diffusion models provides a reliable method for expanding scarce wireless channel data, significantly boosting deep-learning-based first arrival path delay estimation accuracy without requiring costly physical data collection campaigns.
Subject of Research: Deep learning data augmentation for wireless signal arrival path delay estimation
Article Title: Gdm-assisted CNN framework for time domain first arrival path delay estimation in integrated sensing and communications
Article References: Quan, J., Feng, L., Wang, X., Yang, Z., Li, W., Liu, G., & Zhao, C. (2026). Gdm-assisted CNN framework for time domain first arrival path delay estimation in integrated sensing and communications. Journal on Advances in Signal Processing, 2026(1), Article 58. https://doi.org/10.1186/s13634-026-01338-8
Image Credits: AI Generated
DOI: 10.1186/s13634-026-01338-8
Keywords: first arrival path, generative diffusion model, convolutional neural network, data augmentation, gramian angular field, cross-correlation function, wireless positioning, multipath fading, channel estimation, time domain estimation, zadoff-chu sequence, rayleigh fading, rician fading, signal synchronization, synthetic channel data, u-net denoising
Cite Scienmag News
Cassandra Pierce. (November 20, 2026). Diffusion Model Synthetic Data Enhances Convolutional Network Signal Arrival Delay Estimation. Scienmag. https://scienmag.com/diffusion-model-synthetic-data-enhances-convolutional-network-signal-arrival-delay-estimation/
Cassandra Pierce. "Diffusion Model Synthetic Data Enhances Convolutional Network Signal Arrival Delay Estimation." Scienmag, 20 November 2026, https://scienmag.com/diffusion-model-synthetic-data-enhances-convolutional-network-signal-arrival-delay-estimation/. Accessed 20 November 2026.
Cassandra Pierce. "Diffusion Model Synthetic Data Enhances Convolutional Network Signal Arrival Delay Estimation." Scienmag. November 20, 2026. https://scienmag.com/diffusion-model-synthetic-data-enhances-convolutional-network-signal-arrival-delay-estimation/








