PaperlayerView exact revision on arXiv

Source-linked AI summary

Rad-R: A Raw-ADC Radar Dataset and Capture-Invariant SSM for Hardware-Fault Diagnosis

Mainak Mallick, Junghwan Yim, Seung-Kyum Choi

arXiv:2608.30896v1cs.CV

TL;DR

Automotive radar-fault diagnosis lacks public data combining raw ADC, controlled physical faults, calibrated severities, and synchronised sensor measurements. The paper introduces Rad-R and evaluates raw-IQ Mamba models across multiple protocols, finding that capture-invariant RadrNet-DS-CI leads the controlled cross-severity benchmark while the dataset’s single-session scope limits generalisation claims.

  • Problem

    Public radar datasets do not combine raw ADC with controlled hardware-fault conditions and independently measured severities, despite the need to study deployment degradation.

  • Method

    The paper constructs Rad-R and benchmarks raw-IQ RadrNet models alongside seven vision backbones under within-clip, anytime, few-shot cross-capture, and controlled cross-severity protocols.

  • Results

    RadrNet-DS-CI ranks first on controlled cross-severity diagnosis at 0.663 macro-F1 versus 0.628 for the strongest RD-CNN, while the RadrNet family leads anytime and few-shot budgets.

  • Takeaways & Limitations

    Capture-invariant raw-IQ encoding improves cross-severity diagnosis, with magnitude standardisation accounting for most of the reported gain.

  • Takeaways & Limitations

    Rad-R is single-session, and the reported checkpoints are oracle-selected on evaluation splits, making current numbers upper bounds rather than deployment estimates.

Abstract

from arXiv · show

Automotive mmWave radar can develop vibration, antenna misalignment, radome blockage, and receive-channel degradation that corrupt the signal before perception begins. Data for these faults are scarce because each condition must be induced and measured on physical hardware. We introduce Rad-R, a raw-ADC dataset captured with a 4-chip 77GHz TI MMWCAS-RF-EVM cascade (192 virtual channels). Unlike existing raw-radar datasets, Rad-R pairs each recording with a controlled hardware fault at a calibrated severity, an independent physical severity measurement, and frame-synchronised IMU, temperature, GPS, and camera streams. Rad-R is a single-session dataset, so our generalisation claims are confined to a controlled cross-severity protocol in which train and test use physically distinct captures. A reproducible benchmark evaluates seven representative vision backbones and the proposed raw-IQ Mamba SSM (RadrNet) under within-clip, chirp-wise anytime, few-shot cross-capture, and controlled cross-severity protocols. Within-clip performance is near-saturated ($>0.98$ macro-F1), whereas cross-severity generalisation remains difficult: the absolute-phase RadrNet-DS falls to $0.49$ macro-F1. RadrNet-DS-CI replaces absolute phase with per-frame-standardised magnitude and relative chirp-to-chirp phase and ranks first on the controlled benchmark ($0.663$ vs. $0.628$ for the strongest RD-CNN; three seeds); the RadrNet family also leads on the anytime and few-shot budgets. A descriptive cross-modal analysis further finds that radar micro-Doppler covaries with independently measured IMU vibration energy (pooled Spearman $ρ=0.41$ across conditions). The complete dataset and code will be released publicly under permissive licences.

1. Introduction

Rad-R addresses the lack of public raw-ADC radar data pairing controlled physical faults with calibrated severity and synchronised sensor streams. The paper couples this dataset with raw-IQ RadrNet models and shows that capture-invariant encoding improves controlled cross-severity diagnosis.

  • Existing datasets do not combine raw ADC with controlled hardware-fault annotations and a sensor-state evaluation suite.
  • Rad-R contains raw 77 GHz ADC from a 4-chip cascade radar with 192 virtual channels and synchronised IMU, temperature, GPS, and camera streams.
  • RadrNet processes raw complex IQ with a hierarchical 3-stage Mamba, while RadrNet-DS adds an RD-map CNN through an input-dependent gate.
  • Single-stream cross-severity F1 rises from 0.545 to 0.612 with capture-invariant encoding, while RadrNet-DS-CI exceeds the strongest RD-CNN by +0.035 mean macro-F1.
  • RadrNet-DS-CI wins the controlled benchmark, and RadrNet leads ResNet-18 across all six chirp budgets and DeiT-III from five labelled target frames per class.

2. Related Work

Prior radar datasets mainly provide raw ADC under nominal conditions or processed representations, while radar-fault studies lack a public raw-ADC benchmark. RadrNet is presented as the first hierarchical Mamba evaluated on complex raw IQ for chirp-wise anytime and fault diagnosis settings.

  • Existing public datasets do not pair raw ADC with hardware-fault labels, and radar degradation studies lack a public raw-ADC benchmark.
  • Prior methods include interference mitigation on RD maps, RD-to-ADC distillation, Mamba for RD detection, and transformers for radar tensors.
  • RadrNet is described as the first hierarchical Mamba evaluated on complex raw IQ for chirp-wise anytime and fault diagnosis.

3. The Rad-R Dataset

Rad-R is a physically captured, multimodal dataset covering four hardware faults at independently gated severities, with raw radar recordings and synchronised companion sensors. Its protocols distinguish confounded within-clip evaluation from cross-capture and controlled cross-severity testing, whose scope remains limited by the single-session design.

  • 3.1. Hardware and Capture Setup: The capture rig combines a 4-chip 77 GHz cascade radar, 192 virtual channels, cameras, IMU, GPS, temperature probes, Raspberry Pi, and storage.
  • 3.1. Hardware and Capture Setup: Each capture uses 256 ADC samples per chirp, 64 loops per frame, and a 100 ms frame period in TDM-MIMO mode.
  • 3.2. Fault Taxonomy and Induction Procedures: Four faults are physically induced at mild and severe levels, with independent gates measuring vibration, yaw misalignment, blockage, and receive-channel degradation.
  • 3.3. Dataset Contents: Rad-R contains approximately 9 capture runs, 27,000 frames, and 324 GB of raw ADC, plus a balanced 1,800-frame training cache and synchronised sensor signatures.
  • 3.4. Evaluation Splits: The within-clip split shares scene, sensor, and configuration between training and testing, so it is treated as a confounded upper bound rather than an independent generalisation test.
  • 3.4. Evaluation Splits: Controlled cross-severity testing trains on one severity and tests on the other using physically distinct captures, while held-out day, device, and scene evaluation remains unsupported.

4. Models and RadrNet

RadrNet models raw complex IQ with hierarchical Mamba stages that capture range, Doppler, and virtual-array structure. Its capture-invariant variants standardise magnitude and encode relative chirp-to-chirp phase, optionally fusing an RD-CNN stream through an input-dependent gate.

  • RadrNet: RadrNet processes raw complex IQ tensors with a hierarchical 3-stage Mamba and multi-scale features for fault and severity prediction.The input shape is (B, 2, C, S, R), representing real/imaginary parts, chirps, ADC samples, and virtual channels.
  • Capture-invariant RadrNet: The capture-invariant encoding uses per-frame-standardised magnitude and relative chirp-to-chirp phase represented by sine and cosine channels.Differencing cancels constant global and per-channel phase offsets, while standardisation removes global magnitude scale.
  • RadrNet: The IQ stream tokenises neighbouring encoded samples with a strided 1-D convolution before sequential Mamba stages scan joint range–slow-time, Doppler, and cross-range structure.The tokeniser uses N = CS/8 = 2048 tokens with dimension d = 256.
  • Fusion variants: The RD branch computes a compact CNN representation from released range–Doppler maps and combines it with the IQ representation through gated projection.The RD stream uses four strided convolutions followed by global average pooling.
  • Fusion variants: RadrNet-DS-CI combines the invariant IQ representation with a compact RD-CNN representation using an input-dependent 2-way gate before dual prediction heads.RadrNet-CI omits the RD stream; the heads predict five fault classes and three severity levels.

5. Experiments and Results

The experiments compare radar models under within-clip, anytime, few-shot, and controlled cross-severity protocols. Within-clip scores are near-saturated, while capture-disjoint cross-severity evaluation is substantially harder and favours capture-invariant RadrNet variants.

  • 5.1. Within-Clip Split as Context: Within-clip models reach 0.974 macro-F1 for DeiT-III and Swin-T, while raw-IQ baselines reach 0.989–0.994 macro-F1.Because train and test share scene and sensor, this split is treated as a contextual upper bound rather than evidence of generalisation.
  • 5.2. Anytime / Chirp-Wise Inference: RadrNet-DS takes 5.5–6.2 ms/frame versus 2.1–4.6 ms/frame for ResNet-18, trading roughly 2× wall-time for +0.5–0.7 macro-F1 at small chirp budgets.Both remain below the 100 ms frame deadline.
  • 5.3. Few-Shot Cross-Capture Adaptation: From k=5 labelled target frames per class, RadrNet-DS reaches 0.755 versus 0.562 for ResNet-18 and 0.711 for DeiT-III.At k=1, RadrNet-DS is statistically tied with ResNet-18 at 0.584 versus 0.574.
  • 5.4. Controlled Cross-Severity Generalisation: Controlled cross-severity macro-F1 falls to 0.485 for RadrNet-DS and 0.545 for RadrNet, while RadrNet-CI reaches 0.612±0.015 and RadrNet-DS-CI ranks first at 0.663 ± 0.009.RadrNet-DS-CI exceeds ResNet-18 at 0.628±0.015 by +0.035 and is higher at each of three seeds.
  • 5.2. Anytime / Chirp-Wise Inference: At k=64 in the separate anytime harness, RadrNet-DS-CI scores 0.601, ahead of ResNet-18 at 0.591, RadrNet-CI at 0.580, and RadrNet-DS at 0.441.These prefix-evaluation values are not interchangeable with the 0.663 full-frame aggregate.

6. Conclusion

Rad-R pairs raw ADC with physically induced, independently calibrated radar faults and synchronised companion sensors, while its controlled benchmark shows that capture-invariant encoding improves cross-severity robustness. However, single-session data and oracle checkpoint selection constrain the reported results as upper bounds rather than deployment estimates.

  • RadrNet-DS-CI is strongest on the controlled challenge, reaching 0.663 macro-F1 versus 0.628 for the best RD-CNN.The capture-invariant encoding removes constant phase offsets and global gain by construction.
  • Magnitude standardisation accounts for most of the capture-invariance gain, while relative phase cancels constant phase offsets.
  • The dataset and code will be released publicly under permissive licences.
  • Single-session data and oracle-selected checkpoints make the current numbers upper bounds rather than deployment estimates.These limitations restrict generalisation claims and introduce checkpoint-selection bias.
  • Rad-R uniquely pairs raw ADC with controlled physical faults, independently calibrated severities, and synchronised companion sensors.

C. Sample Radar Visualizations

The supplementary visualisations provide expanded per-fault and per-capture views that align radar representations with camera and companion-sensor streams. Together, the dashboards expose both signal characteristics and synchronisation context for each recording.

  • Each dashboard combines a range profile, range–Doppler map, micro-Doppler spectrum, channel power, camera streams, and an IMU/temperature/GPS strip.
  • The upper-left dashboard block shows chirp-mean range, display-normalised range–Doppler, Doppler power, and power across 192 virtual channels.
  • Camera streams are shown at the radar timestamp marked on the lower sensor strip.The monochrome and colour frames are included with faces and licence plates Gaussian-blurred.

D. Datasheet for Rad-R

Rad-R is a physically controlled, multimodal radar-fault dataset whose instances contain raw IQ, derived radar products, synchronised sensors, and protocol-based fault and severity labels. Its single-device, single-day, single-scene sample supports fault-aware research but is not sufficient for unvalidated safety-critical deployment.

  • Rad-R contains approximately 26,800 frames from nine five-minute capture runs spanning healthy operation and four fault types at two severity levels.The training cache subsamples 1,800 frames, while the complete raw archive is approximately 26,800 frames.
  • The dataset is a sample containing one representative capture per fault-severity cell in a single outdoor/lab scene with one vehicle.
  • Each frame includes raw complex IQ, a pre-computed 224×224 dB-scaled range–Doppler map, and synchronised IMU, temperature, GPS, and camera references.
  • Fault labels cover healthy, vibration, misalignment, blockage, and Rx degradation, while severity labels encode S0 healthy, S1 mild, and S2 severe.Labels are determined by the physical induction protocol rather than post-hoc human annotation.
  • Rad-R should not train safety-critical production monitoring systems without extensive in-field validation because its single-day, single-device scope is narrow.

F. Hyperparameters and Training Setup

The benchmark standardises core optimisation settings across architectures while documenting model-specific configurations and capture-invariance augmentations. Reported within-clip and cross-severity scores require caution because saved checkpoints were selected directly on evaluation splits.

  • Within-clip and cross-severity checkpoints retain the epoch with maximum test fault macro-F1 without a separate validation partition.Their reported values are therefore oracle-selected upper bounds, whereas few-shot adaptation uses a fixed 30-step schedule.
  • RadrNet uses embedding dimension 256, three stages of depth 2, state size 16, convolution width 4, expansion 2, patch size 8, and dropout 0.3.
  • Capture-invariance training augments inputs with random global phase rotation, per-receiver amplitude scaling, receiver dropout, and per-frame standardisation.

G. Within-Clip Anytime Benchmark: Full Table

The benchmark evaluates radar models under anytime and controlled cross-severity settings, while the proposed invariant encoding removes capture-specific magnitude and phase cues. Per-fold results show that RadrNet-DS-CI substantially improves the difficult severe-to-mild direction.

  • Within-Clip Anytime Benchmark: The anytime benchmark reports fault macro-F1 statistics with five-seed means, standard deviations, bootstrap confidence intervals, and paired significance tests.Table S3 provides the complete per-budget values for the within-clip chirp-wise evaluation.
  • Capture-Invariant Encoding: Absolute phase can expose capture identity because global and per-channel phase offsets remain nearly constant within recordings but differ between recordings.This creates a shortcut under within-clip evaluation and motivates the invariant representation for cross-capture testing.
  • Capture-Invariant Encoding: RadrNet-DS-CI removes absolute phase by combining per-frame-standardised magnitude with relative chirp-to-chirp phase encoded as sine and cosine.The invariant input uses 3R channels instead of RadrNet’s 4R channels and does not rely on training-time augmentation for invariance.
  • Controlled Cross-Severity Evaluation: Fold B trains on mild S1 and tests on severe S2 in one direction, while the reverse severe-to-mild direction produces the main failure for off-the-shelf raw-IQ models.RadrNet-DS scores only 0.392 in Fold B, whereas RadrNet-DS-CI reaches 0.678 there.
  • Controlled Cross-Severity Evaluation: RadrNet-DS-CI recovers Fold B to 0.678 while keeping Fold A competitive, making the reverse-severity fold the largest source of its overall gain.The two folds use physically distinct training and testing captures by swapping the train and test severities.

H.4. Encoding Ablation and Per-Fault Breakdown

The ablation identifies per-frame magnitude standardisation as the dominant contributor to capture-invariant cross-severity performance, while relative phase helps selectively in the harder severe-to-mild direction. The anytime and few-shot results expose trade-offs between early raw-IQ evidence, full-frame RD evidence, and label efficiency.

  • Encoding Ablation: 0.631 mean fault macro-F1 is achieved by magnitude standardisation alone, compared with 0.545 for the absolute-phase baseline.Magnitude-only encoding also raises blockage recall from 23.4% to 67.3%.
  • Encoding Ablation: 0.399 mean fault macro-F1 from relative phase alone is below the absolute-phase baseline, and adding it to magnitude lowers the mean from 0.631 to 0.612 within the seed spread.Its contribution is concentrated in the harder severe-to-mild Fold B rather than the overall mean.
  • Per-Fault Breakdown: Relative phase improves Fold B from 0.538 to 0.578, lifting blockage recall from 76.8% to 89.3% and Rx-degradation recall from 72.7% to 91.8%.The same addition costs Fold A performance, which falls from 0.724 to 0.647, mainly through misalignment recall dropping from 93.7% to 66.3%.
  • RD-Branch Analysis: The RD branch changes behavior across encodings: it collapses absolute-phase Fold-B Rx-degradation recall from 98.8% to 17.3% but improves invariant Fold-B vibration recall from 31.5% to 85.5%.For the invariant encoding, it also raises misalignment recall from 40.8% to 64.0%, with blockage recall decreasing from 89.3% to 77.7%.
  • Anytime Under Cross-Severity: 0.479 fault macro-F1 is reached by RadrNet-CI using 4 of 64 chirps, with performance increasing monotonically across the cross-severity anytime budget.RadrNet-DS-CI is weaker at partial budgets and reaches its highest prefix-sweep score only at k=64.
  • Few-Shot Evaluation: 0.732 fault macro-F1 at k=5 places RadrNet-DS-CI above DeiT-III’s 0.711, while absolute-phase RadrNet-DS remains the most label-efficient model.Few-shot comparisons use a common adaptation protocol, but seed counts differ across models.

J. Per-Fault Error Analysis under Cross-Severity Shift

Under controlled cross-severity shift, blockage and misalignment are the main per-fault weaknesses, while capture-invariant models retain the strongest overall accuracy-efficiency trade-off. Confusion toward healthy remains the principal source of headroom.

  • Per-fault confusion: 39% blockage recall for RadrNet-DS-CI and 30% for ResNet-18 make blockage the hardest fault under cross-severity shift.The RD-CNN also confuses 29% of blockage frames with receive degradation.
  • Per-fault confusion: 66% misalignment recall for RadrNet-DS-CI trails ResNet-18 at 85%, with one-third of invariant-model misalignment frames classified as healthy.Both models also read over one-third of blockage frames as healthy.
  • Per-fault confusion: 86–89% vibration recovery and 86% receive-degradation recovery contrast with the weaker blockage and misalignment recalls.These values are reported from pooled confusion matrices over three seeds and both fold directions at k=64.
  • Cross-modal analysis: The cross-modal analysis reports pooled Spearman ρ = 0.41 between radar off-centre Doppler energy and independently measured IMU vibration energy.Because frames are temporally correlated within only three condition-level captures, the plot supports a between-condition trend rather than a capture-level significance claim.
  • Accuracy–efficiency trade-off: 0.663 controlled cross-severity macro-F1 makes RadrNet-DS-CI the top model at 4.7 M parameters, below ResNet-18’s 11 M parameters.Capture-invariant models are Pareto-optimal, and parameter count alone does not explain the ranking.
  • Within-clip versus cross-severity: 0.977 → 0.485 is the within-clip to cross-severity decline for absolute-phase RadrNet-DS, the largest drop among the compared models.The re-ranking is consistent with capture-specific phase and gain acting as within-clip shortcuts.
Loading 2608.30896v1…