Deep learning model denoises noisy traffic data in a single real-time pass

Traffic engineers trying to forecast journey times in real time have long faced an awkward trade-off: the classical filters that clean noisy sensor data were designed to see an entire dataset at once, but a live forecasting system only ever holds a few hours of recent observations. A new study of Bluetooth-derived journey times from Greater Manchester shows that this mismatch quietly degrades predictions, and proposes a way around it. Instead of running a denoising filter online, the researchers train a deep learning model on raw, noisy windows to predict what an expert-selected filter would have produced, folding denoising and forecasting into one causal forward pass that meets real-time latency requirements.

The work, published in European Transport Research Review, was carried out by Mohammadmahdi Rahimiasl, Ynte Vanderhoydonc and Siegfried Mercelis within the European Union’s Horizon 2020 TANGENT project, which focuses on multimodal traffic management across European cities. Their starting point is an uncomfortable property of the data itself. Bluetooth sensors detect device MAC addresses without knowing the transport mode, match counts fluctuate over time, and physical obstructions can shrink detection range. The result is a journey-time series studded with missing values, outliers and sharp fluctuations. Across the 204 predefined routes recorded at 15-minute intervals over 2022, roughly 20 percent of entries are missing, and the aggregated statistics show a positively skewed, heavy-tailed distribution with substantial outlier influence.

Reliable journey-time forecasts matter for traffic management and traveler information, and the stakes are considerable. Citing an external report, the authors note that drivers in London lost an average of 101 hours to congestion in 2024, a 2 percent increase from the previous year, while across the UK the average was 62 hours, with an estimated national cost of 7.7 billion pounds in lost time. Better real-time prediction can support services such as live traffic information, multimodal travel tools and cooperative, connected and automated mobility, all priority areas identified by the European Commission.

The core problem the paper identifies is what happens when whole-series signal processing is squeezed into short, shifting windows. Real-time forecasting models typically work from one to three hours of recent observations, but wavelet shrinkage denoising and Butterworth low-pass filtering were conceived for full series or long segments. Applied independently to short sliding windows, wavelet transforms suffer edge artifacts and can give inconsistent values for the same timestamp depending on where the window sits, because the method depends on global coefficient context. Butterworth filtering, an infinite-impulse-response design, effectively resets its filter state at every segment boundary and can amplify boundary transients. Zero-phase filtering, which applies the filter forward and backward, is inherently non-causal and therefore unsuitable for real-time use, since it requires future samples that have not yet arrived.

To quantify the damage, the team benchmarked windowed filtering against an idealized whole-series baseline on the final 20 percent of the Manchester data, which was withheld from training and never seen by the models. Five wavelet configurations and six Butterworth parameter sets were tested. The discrepancies were substantial. Averaged over the five wavelet settings, windowed filtering produced a mean absolute error of 26.76 seconds against the whole-series reference, while deep learning surrogates trained to mimic the whole-series effect cut that to 12.12 seconds, a 54.71 percent reduction, and lowered root mean squared error from 48.21 to 21.28 seconds, a 55.86 percent reduction. The largest gain appeared in the noisiest, most edge-sensitive configuration, db2 at level 2, where the error fell from 31.64 to 10.71 seconds. For Butterworth filtering the picture was more nuanced: mild parameter sets were already well approximated by direct windowed filtering, but in the most difficult regime the windowed approach collapsed to an RMSE of 124.90 seconds while the learned surrogate recovered to 23.13 seconds.

Yet the surrogates exposed a second trap. Models trained only on denoised data generalized poorly when confronted with raw, noisy real-time inputs. The authors’ solution is what they call expert-guided denoising distillation. First, an expert selects a wavelet or Butterworth parameterization and applies it offline to the full series, producing a processed reference target. Then a forecasting model, in this case MTGNN, a multivariate graph-learning architecture that infers latent spatial structure directly from data, is trained on raw causal windows of the noisy series to predict the next twelve values of that processed target, at 15-minute resolution. The denoising rule stays explicit, configurable and auditable offline, while the learned model delivers the low-latency approximation online.

The payoff shows up in the head-to-head forecasting comparison. Relative to a standard pipeline trained on window-filtered data, the proposed procedure reduced average MAE by 36.85 percent and RMSE by 8.77 percent for wavelet targets, and MAE by 32.76 percent and RMSE by 16.19 percent for Butterworth targets, all measured against the whole-series processed reference. At inference time the system keeps a rolling buffer of raw observations, normalizes the input window and emits its forecast through a single causal forward pass, with no online wavelet decomposition, no forward-backward filtering and no whole-history storage. Benchmarked on the UTD19 Torino dataset using PyTorch on an NVIDIA A40 GPU, training took roughly five seconds per epoch, and mean single-forward-pass latency was 6.27 milliseconds on GPU and 39.53 milliseconds on CPU.

The authors are careful about what these numbers do and do not mean. Each processed target is defined by the chosen preprocessing rule, not a direct observation of the physical traffic state, so the reported gains demonstrate improved prediction of the expert-defined denoised reference rather than improved accuracy against unfiltered real traffic conditions. The operational value, they argue, depends on whether the selected denoising procedure preserves the traffic dynamics that matter in a given deployment. If the expert specification oversmooths the data and suppresses meaningful congestion dynamics, the surrogate will inherit that bias. They also flag a statistical caveat: repeated training runs were not performed because of computational constraints, so the variability of the reported gains remains unquantified.

To test whether the effects were specific to Bluetooth sensing, the team ran a controlled experiment on the UTD19 Torino traffic-flow subset, a complete 5-minute series spanning 399 sensors over roughly three weeks in 2016. They injected synthetic corruptions, including additive Gaussian noise, sparse impulsive spikes, a small linear drift on 30 percent of sensors, random point dropouts and contiguous blackout intervals, then compared windowed denoising in non-overlapping one-hour segments against whole-series processing. Whole-series processing consistently won, with Butterworth RMSE reductions of 29.82 to 36.94 percent and MAE reductions of 10.66 to 18.44 percent across configurations, and wavelet RMSE reductions of 12.31 to 20.90 percent. The authors stress that this benchmark is a controlled robustness check rather than external validation on naturally noisy data, and that the injected corruptions are simplified approximations of operational sensing errors, so the gains should not be assumed to transfer quantitatively to all real-world datasets.

The paper positions the method as an auditable deployment procedure rather than a new forecasting architecture, and it is candid about when the simpler alternative wins. If the denoising requirement is mild, the input series is stable and the filter behaves consistently on short windows, a classical pipeline may be preferable because it is easier to implement and explain. The distillation approach earns its offline cost, generating the processed target, training the surrogate and revalidating after any change to the denoising rule, only when short-window preprocessing materially changes the target signal, as observed for the wavelet configurations and the harder Butterworth settings. The authors recommend pairing any deployment with monitoring and periodic revalidation, since noise characteristics can drift as match rates, sensor coverage, seasonal patterns or infrastructure change. Future work, they suggest, should add operator-relevant criteria such as peak timing and congestion onset detection, stratification by match-count regimes, and additional forecasting backbones under the same explicit contract.

Subject of Research: Real-time journey-time forecasting and denoising of noisy Bluetooth traffic data using expert-guided deep learning distillation

Article Title: Real-time prediction and denoising: deep learning approaches for noisy traffic data

Article References: Rahimiasl, M., Vanderhoydonc, Y., & Mercelis, S. (2026). Real-time prediction and denoising: deep learning approaches for noisy traffic data. European Transport Research Review, 18(1), Article 105. https://doi.org/10.1186/s12544-026-00868-4

Image Credits: AI Generated

DOI: 10.1186/s12544-026-00868-4

Keywords: traffic forecasting, deep learning, denoising, Bluetooth sensors, journey time, MTGNN, wavelet denoising, Butterworth filtering, real-time prediction, Greater Manchester, UTD19, sliding windows