SVM + CFAR + OR: A Hybrid Machine-Learning-CFAR Receiver for Low-False-Alarm PRACH Preamble Detection in 5G NR and Beyond

arXiv:2611.07685 · eess.SP · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "SVM + CFAR + OR".

Dev: Here's the gist of the paper. The paper introduces SVM + CFAR + OR as a hybrid receiver architecture that combines an analytical CFAR detector with a score-thresholded machine learning classifier via a logical OR rule to detect low-false-alarm PRACH preambles in 5G NR and beyond.

Rosa: Pruning redundant reasoning tokens The physical random access channel (PRACH) is the initial access channel between the user equipment and the base station, and the ability to initiate transmission in a cellular network depends on its reliable operation.

Dev: The growing device density in 5G and beyond-5G networks, massive machine-type communication, and the trend towards a machine-learning-driven radio interface make analytical preamble detectors increasingly limited in their ability to adapt to complex and dynamic radio environments, particularly at low signal-to-noise ratios.

Rosa: This article recasts preamble detection as a 64-class classification problem and introduces two model-agnostic contributions: a robust adaptive threshold that realizes the constant-falsealarm-rate (CFAR) principle directly in the domain of a classifier’s decision scores (MAD-CFAR), and a hybrid receiver that fuses the learned detector with a classical CFAR detector on the power delay profile through a logical OR rule.

Title and authors: Rosa: So we're looking at this paper now, "SVM + CFAR + OR: A Hybrid Machine-Learning-CFAR Receiver for Low-False-Alarm PRACH Preamble Detection in 5G NR and Beyond" Rosa The title tells you right away that they're combining a few different things to build a receiver that can spot those preamble signals really well even when the network gets super crowded or the signal is weak Dev I like that part about it being for low-false-alarm detection, because in these new 5G and beyond networks, you need to be able to find those initial access signals reliably under tough conditions Taro It sounds like they’re tackling the problem of getting reliable initial access when things get really dense or noisy Rosa The authors are Filip Ferenc and Michał Szczachor from Wrocław University of Science and Technology at Nokia, which is interesting since Nokia is right there in the hardware world Dev That suggests this isn't just theoretical; they're working with real physical layer challenges Taro <ref:2611.07685#pg1,SVM + CFAR + OR: A Hybrid Machine-Learning-CFAR Receiver for Low-False>

The paper's summary: Rosa: Basically, what this paper does is take preamble detection, which used to rely on older analytical methods that struggle in complex radio environments, and they turn it into a classification problem for a machine learning model Rosa They introduce two main things: first, this adaptive thresholding called MAD-CFAR which applies the constant false alarm rate principle right to the decision scores of the classifier Dev And second, they build a hybrid receiver that uses a logical OR rule to merge the output from that learned detector with a classical CFAR detector on the power delay profile Taro It’s essentially using both methods together for detection Rosa This is trying to make it robust so you get good results whether the conditions are perfect or really messy Dev

The paper's improvements: Rosa: The main improvement they highlight is how this combination handles those low signal-to-noise ratio situations very well because the MAD-CFAR threshold drives false alarms to zero on a twenty-eight million position noise campaign Rosa This means that even when it’s just noise, the learned part of the system won't trigger falsely Dev And they also claim this logical OR fusion guarantees zero missed detection regression relative to the analytical reference detector Taro That’s important because it means if one method fails, you still get a detection from the other one Rosa They showed this across six propagation channels and two hundred fifty-eight conformance-style tests, and they got several decibels of low signal-to-noise ratio sensitivity without any missed detections compared to the analytical branch Dev

Conclusion: Rosa: So to wrap up, the authors of "SVM + CFAR + OR: A Hybrid Machine-Learning-CFAR Receiver for Low-False-Alarm PRACH Preamble Detection in 5G NR and Beyond" have shown a way to use score domain thresholding with MAD-CFAR to keep false alarms near zero on a big noise test Rosa And the logical OR fusion ensures you don't miss detections compared to the older analytical detector Dev This whole approach, which they call SVM + CFAR + OR, is model agnostic because it works whether you use an SVM or a one-D Convolutional Neural Network Taro It’s about making sure that a cheaper learned branch can actually be used effectively if you control the false alarm rate exactly how you want it to be controlled Rosa <ref:2611.07685#pg2,1-D Convolutional Neural Network>

Dev: And I think the point they make about model independence is key; it means we don't have to commit to one specific learning algorithm if we can just impose the same false alarm control structure on top of any classifier Taro I agree, that makes a lot of sense for developing flexible systems Rosa So, what’s next for them? They are looking at making the one-D CNN branch even more efficient to lower its computational cost while keeping that high accuracy they achieved on the dataset Dev <ref:2611.07685#pg2>

Taro: And I wonder about the limitations they mentioned; they said when you train the learned branch without considering carrier frequency offset, it degrades quite a bit under real residual frequency offset, which is order zero point four to one point zero in those scenarios Rosa That suggests that while the structure is sound, the data needs to be more comprehensive for real-world deployment Dev

Rosa: And another limitation they pointed out is that because the logical OR rule can't lower the false alarm rate below what the analytical detector has, it means in some extreme scenarios they still see those sporadic false alarm spikes and exceed a false alarm probability of one to three thousand percent in about thirty-three of two hundred fifty-eight conformance scenarios Dev

Taro: That’s a bit concerning for the safety case; if you're relying on that analytical branch, you inherit those spikes when things get really extreme Rosa But they did manage to keep the overall false alarm probability below one to three thousand percent in about two hundred twenty-five of those scenarios Dev

Taro: It seems like a trade-off between perfect reliability and achieving better sensitivity in low signal-to-noise ratio situations is what you're dealing with Rosa But the cost comparison was interesting too; they showed that the SVM is quite large, about four point six gigabytes and has a latency of two point eight one milliseconds, which is way slower than what they achieved with the CNN Dev

Dev: That’s a significant difference in terms of hardware requirements; their CNN branch required less than an order of magnitude less compute power and memory for the same accuracy Rosa So, while the structure is robust, you can actually use a smaller model if you optimize for cost or latency Taro

Rosa: And they did show that this lightweight one-D CNN was only about seven point one percent more accurate than their SVM offline accuracy, which is not a huge gap in terms of practical performance gain Dev So the main message here is that the MAD-CFAR threshold and the OR hybrid are genuinely classifier agnostic because they impose FAR control on both branches exactly the same way Taro <ref:2611.07685#pg2>

Rosa: It really confirms that you can deploy a lower-cost learned branch if you just treat its detection sensitivity as an optimization problem rather than something that fundamentally limits how good the method is Dev That’s a big shift in how we think about deploying these systems Taro

Wroclaw University of Science and Technology · Nokia

eess.SP

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: 12 page, 10 figures

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 87/100

The gist: The paper introduces SVM + CFAR + OR as a hybrid receiver architecture that combines an analytical CFAR detector with a score-thresholded machine learning classifier via a logical OR rule to detect…

Key concepts

MAD-CFAR
This is a robust adaptive thresholding technique applied directly to the raw decision scores output by the machine learning classifier. It is designed to force the false alarm rate of the learned branch to near zero during noise calibration, making it model-agnostic and effective for controlling false alarms in score space.
Logical OR Fusion
This is a decision fusion rule where a detection is flagged if either the analytical detector or the machine learning detector reports one. This guarantees that there are no missed detections relative to the analytical reference detector, ensuring reliability while inheriting its low standard false alarm rate.
PRACH Preamble Detection
This involves identifying specific sequences (preambles) transmitted by a user device during initial access in 5G networks. The system classifies these preambles into 64 categories based on their signal features, allowing the receiver to determine if a preamble is present or noise is dominating the channel.
Score-Domain Thresholding
Instead of using raw signal power, this method applies the thresholding mechanism to the internal scores generated by the classifier (like SVM logits). This allows for direct control over false alarms based on how confident the model is in its classification, optimizing performance through score space.

Terminology

Summary

The paper introduces SVM + CFAR + OR as a hybrid receiver architecture that combines an analytical CFAR detector with a score-thresholded machine learning classifier via a logical OR rule to detect low-false-alarm PRACH preambles in 5G NR and beyond. p. 0

Where the paper says it

The growing device density in 5G and beyond-5G networks, massive machine-type communication, and the trend towards a machine-learning-driven radio interface make analytical preamble detectors increasingly limited in their ability to adapt to complex and dynamic radio environments, particularly at low signal-to-noise ratios. This article recasts preamble detection as a 64-class classification problem and introduces two model-agnostic contributions: a robust adaptive threshold that realizes the constant-falsealarm-rate (CFAR) principle directly in the domain of a classifier’s decision scores (MAD-CFAR), and a hybrid receiver that fuses the learned detector with a classical CFAR detector on the power delay profile through a logical OR rule.

Pruning redundant reasoning tokens

The physical random access channel (PRACH) is the initial access channel between the user equipment and the base station, and the ability to initiate transmission in a cellular network depends on its reliable operation. The growing device density in 5G and beyond-5G networks, massive machine-type communication, and the trend towards a machine-learning-driven radio interface make analytical preamble detectors increasingly limited in their ability to adapt to complex and dynamic radio environments, particularly at low signal-to-noise ratios. p. 0

Where the paper says it

SVM + CFAR + OR: A Hybrid Machine-Learning-CFAR Receiver for Low-False-Alarm PRACH Preamble Detection in 5G NR and Beyond Filip Ferenc∗ Faculty of Information and Communication Technology, Wrocław University of Science and Technology, Wrocław, Poland Michał Szczachor Nokia, Wrocław, Poland Abstract The physical random access channel (PRACH) is the initial access channel between the user equipment and the base station, and the ability to initiate transmission in a cellular network depends on its reliable operation. The growing device density in 5G and beyond-5G networks, massive machine-type communication, and the trend towards a machine-learning-driven radio interface make analytical preamble detectors increasingly limited in their ability to adapt to complex and dynamic radio environments, particularly at low signal-to-noise ratios.

This article recasts preamble detection as a 64-class classification problem and introduces two model-agnostic contributions: a robust adaptive threshold that realizes the constant-falsealarm-rate (CFAR) principle directly in the domain of a classifier’s decision scores (MAD-CFAR), and a hybrid receiver that fuses the learned detector with a classical CFAR detector on the power delay profile through a logical OR rule. p. 0
Where the paper says it

The growing device density in 5G and beyond-5G networks, massive machine-type communication, and the trend towards a machine-learning-driven radio interface make analytical preamble detectors increasingly limited in their ability to adapt to complex and dynamic radio environments, particularly at low signal-to-noise ratios. This article recasts preamble detection as a 64-class classification problem and introduces two model-agnostic contributions: a robust adaptive threshold that realizes the constant-falsealarm-rate (CFAR) principle directly in the domain of a classifier’s decision scores (MAD-CFAR), and a hybrid receiver that fuses the learned detector with a classical CFAR detector on the power delay profile through a logical OR rule.

How the training signal is built

The training dataset is generated with a professional LTE/5G physical-layer simulator provided by Nokia, implemented in MATLAB and reproducing a full LTE/5G network environment. Signals are recorded for all 64 preambles across 41 SNR points from −20 to +20 dB in a topology of two base stations, each equipped with two receive antennas. p. 2

Where the paper says it

2 Dataset and Dimensionality Reduction The training dataset is generated with a professional LTE/5G physical-layer simulator provided by Nokia, implemented in MATLAB and reproducing a full LTE/5G network environment. Signals are recorded for all 64 preambles across 41 SNR points from −20 to +20 dB in a topology of two base stations, each equipped with two receive antennas.

With a 10 ms access-slot step over a one-second simulation, each base station contributes 100 occasions per SNR point; treating the two base stations and their individual antenna paths as independent observations yields about one million signals in total (64×41×100×2×2 ≈ 1. p. 2
Where the paper says it

Signals are recorded for all 64 preambles across 41 SNR points from −20 to +20 dB in a topology of two base stations, each equipped with two receive antennas. With a 10 ms access-slot step over a one-second simulation, each base station contributes 100 occasions per SNR point; treating the two base stations and their individual antenna paths as independent observations yields about one million signals in total (64×41×100×2×2 ≈ 1.

05×106). Each complex preamble of 839 samples is decomposed into its real and imaginary parts, concatenated into a single feature vector of length D = 1678, and normalized before classification. p. 2
Where the paper says it

The classifier is trained on four propagation channels – AWGN, EPA, EVA, and ETU – whereas two further channels, TDLC300 and TDLD30, are held out for testing only. Each complex preamble of 839 samples is decomposed into its real and imaginary parts, concatenated into a single feature vector of length D = 1678, and normalized before classification.

Where the gains come from

The key methodological novelty is twofold: (i) an adaptive median absolute deviation CFAR threshold (MAD-CFAR) applied directly to raw classifier decision scores, driving false alarms to zero on a 28-million position noise campaign, and (ii) a logical OR decision fusion that guarantees zero missed-detection regression relative to the analytical reference detector. The MAD-CFAR threshold drives the learned branch’s false alarm rate to near zero on its noise calibration campaign, and the OR hybrid inherits the analytical detector’s low, standard-compliant false alarm rate while widening the range of correct detection. p. 0

Where the paper says it

0000) at substantially lower inference cost and memory footprint. The MAD-CFAR threshold drives the learned branch’s false alarm rate to near zero on its noise calibration campaign, and the OR hybrid inherits the analytical detector’s low, standard-compliant false alarm rate while widening the range of correct detection.

Across six propagation channels and 258 conformance-style tests, the SVM-based hybrid gains several decibels of low-SNR sensitivity with no missed-detection regression relative to the analytical branch, a property that follows directly from the OR construction. p. 0
Where the paper says it

Across six propagation channels and 258 conformance-style tests, the SVM-based hybrid gains several decibels of low-SNR sensitivity with no missed-detection regression relative to the analytical ∗Corresponding author: 263578@student. pl branch, a property that follows directly from the OR construction.

Rigorous SVM vs. p. 1

Where the paper says it

• Classifier-Agnostic ML + CFAR + OR Framework: A robust score-domain adaptive thresholding mechanism (MAD-CFAR) that enforces CFAR operating point control on raw decision scores, driving false alarms to zero regardless of the classifier architecture. • Rigorous SVM vs.

CNN Benchmarking

A head-tohead evaluation of an interpretable OvA SVM with PCA and a lightweight 1-D CNN on the identical 1.05-million sample dataset was conducted. p. 1

Where the paper says it

05-million sample dataset, supported by t-SNE and SHAP/LIME explainability analyses. • Commercial Conformance Validation: Comprehensive evaluation with Nokia’s commercial physical-layer simulator across six propagation channels and 258 conducted conformance scenarios under 3GPP TS 38.

The CNN achieves 99. p. 3
Where the paper says it

05-million sample dataset generated across identical channel conditions. Under five-fold stratified cross-validation, the CNN achieves 99.

29% test accuracy (macro-AUC of 1. p. 0
Where the paper says it

On a held-out test set the SVM reaches 98. 58% accuracy (macro-averaged area under the curve, AUC, of 0.

0000), matching and slightly exceeding the SVM offline accuracy. p. 3
Where the paper says it

29% test accuracy (macro-AUC of 1. 0000), matching and slightly exceeding the SVM offline accuracy.

Its 64 pre-sigmoid logits act directly as decision scores s, feeding the identical MAD-CFAR thresholding stage. p. 3
Where the paper says it

0000), matching and slightly exceeding the SVM offline accuracy. Its 64 pre-sigmoid logits act directly as decision scores s, feeding the identical MAD-CFAR thresholding stage.

The CNN is markedly cheaper, requiring far fewer arithmetic operations per occasion and a model over an order of magnitude smaller. p. 3
Where the paper says it

3 GB ≈ 112 MB – and markedly cheaper, requiring far fewer arithmetic operations per occasion and a model over an order of magnitude smaller, which is attractive for the real-time budget analyzed in Section 3. At the operational, false-alarmconstrained operating point, however, the two models prove complementary rather than one strictly dominating: the SVM-based hybrid remains the strongest variant in missed-detection rate, while the CNN offers comparable behavior in benign channels at a fraction of the cost (Section 5.

Hybrid PRACH Receiver

The hybrid receiver applies a per-preamble logical OR to the binary decisions of the analytical detector, d ana i, and the ML detector, d ML i: d hyb i = d ana i ∨ d ML i. When both detectors agree, or when only the analytical branch reports a detection, the analytical detection (with its exact TA) is kept. p. 5

Where the paper says it

The hybrid receiver applies a per-preamble logical OR to the binary decisions of the analytical detector, d ana i , and the ML detector, d ML i : d hyb i = d ana i ∨ d ML i . (4) When both detectors agree, or when only the analytical branch reports a detection, the analytical detection (with its exact TA) is kept.

When only the ML branch reports a detection, a new detection is added and its TA is estimated post hoc from the same PDP by mapping the preamble index to its (ZC sequence, cyclic shift) pair, extracting the search window, smoothing it, applying a peak-to-mean quality gate, locating the first path, and refining it with parabolic sub-sample interpolation. p. 5
Where the paper says it

(4) When both detectors agree, or when only the analytical branch reports a detection, the analytical detection (with its exact TA) is kept. When only the ML branch reports a detection, a new detection is added and its TA is estimated post hoc from the same PDP by mapping the preamble index to its (ZC sequence, cyclic shift) pair, extracting the search window, smoothing it, applying a peak-to-mean quality gate, locating the first path, and refining it with parabolic sub-sample interpolation; the result is scaled to NTA units per 3GPP TS 38.

Comparison with Existing Machine Learning Receivers

The qualitative comparison shows that prior ML PRACH literature often focuses on binary peak validation (k = 2) or timing assistance rather than direct 64-class preamble identification from raw IQ samples. The proposed receiver is unique in introducing the score-domain MAD-CFAR mechanism, providing analytical and calibrated operational FAR control directly in score space. p. 9

Where the paper says it

As shown in Table 6, prior ML PRACH receivers rely on fixed classifier decision thresholds or confidence scores without explicit CFAR control. The proposed receiver is unique in introducing the score-domain MAD-CFAR mechanism, providing analytical and calibrated operational FAR control directly in score space.

p. 9
Where the paper says it

As shown in Table 6, prior ML PRACH receivers rely on fixed classifier decision thresholds or confidence scores without explicit CFAR control. The proposed receiver is unique in introducing the score-domain MAD-CFAR mechanism, providing analytical and calibrated operational FAR control directly in score space.

A meaningful numerical comparison of MDR and FAR requires that all receivers evaluate identical channel models and Doppler spreads, which none of the reviewed publications satisfy. p. 9
Where the paper says it

Although all reviewed publications address PRACH reception, they formulate different learning tasks, adopt non-identical receiver reference architectures, and evaluate performance under conflicting experimental assumptions. A meaningful numerical comparison of MDR and FAR requires that all receivers evaluate: (i) identical channel models and Doppler spreads, (ii) identical SNR ranges and power offsets, (iii) identical PRACH preamble formats (Short vs.

Scientific Invalidity of Direct Quantitative Comparison & Meta-Analysis

Direct numerical comparison is not scientifically valid because all reviewed publications do not satisfy the combined conditions required for a meaningful numerical comparison of MDR and FAR. The proposed receiver is unique in introducing the score-domain MAD-CFAR mechanism, providing analytical and calibrated operational FAR control directly in score space. Future surveys on ML-based physical layer receivers should adopt a structured meta-comparison framework rather than comparing isolated numerical metrics. p. 9

Where the paper says it

Toward a Meta-Comparison Framework. Future surveys on ML-based physical layer receivers should adopt a structured meta-comparison framework rather than comparing isolated numerical metrics.

Conclusion

This work presented the SVM + CFAR + OR hybrid PRACH receiver and established a broader, classifieragnostic ML + CFAR + OR framework built around two key mechanisms: a robust adaptive MAD-CFAR threshold operating in the domain of raw classifier decision scores, and a logical-OR decision fusion with a classical PDP-CFAR detector. The MAD-CFAR mechanism drives the learned branch’s false alarm rate to zero on a 28-million-position noise calibration campaign, while the OR hybrid inherits the analytical detector’s standardcompliant false alarm rate while extending sensitivity by several decibels. p. 9

Where the paper says it

Conclusion This article presented the SVM + CFAR + OR hybrid PRACH receiver and established a broader, classifieragnostic ML + CFAR + OR framework built around two key mechanisms: a robust adaptive MAD-CFAR threshold operating in the domain of raw classifier decision scores, and a logical-OR decision fusion with a classical PDP-CFAR detector. The MAD-CFAR mechanism drives the learned branch’s false alarm rate to zero on a 28-million-position noise calibration campaign, while the OR hybrid inherits the analytical detector’s standardcompliant false alarm rate while extending sensitivity by several decibels.

Model independence within the ML + CFAR + OR framework was demonstrated by evaluating a lightweight 1-D CNN on the identical dataset, achieving 99. p. 1
Where the paper says it

To prove model independence within the ML + CFAR + OR framework, the learned branch is instantiated with both a kernel Support Vector Machine (SVM with PCA) and a lightweight 1-D Convolutional Neural Network (CNN), both benchmarked on an identical 1. 05-million sample dataset generated by a commercial Nokia physical-layer simulator.

29% offline accuracy at an order-of-magnitude lower compute and memory footprint. p. 0
Where the paper says it

29% accuracy (macro-AUC 1. 0000) at substantially lower inference cost and memory footprint.

The key takeaway is that the MADCFAR threshold and the OR hybrid are genuinely classifieragnostic: they impose FAR control on the CNN exactly as on the SVM, so a lower-cost learned branch is already viable and its detection sensitivity is an optimization problem, not a structural limitation of the method. p. 8
Where the paper says it

It is expected to narrow with a broader architecture search, inferencedomain-matched preprocessing, recalibration of (k, smin) on a larger multi-channel noise campaign, and additional training (Section 6). The key takeaway is that the MADCFAR threshold and the OR hybrid are genuinely classifieragnostic: they impose FAR control on the CNN exactly as on the SVM, so a lower-cost learned branch is already viable and its detection sensitivity is an optimization problem, not a structural limitation of the method.

Limitations

The study also exposes clear limitations; the learned branch is trained without carrier-frequencyoffset realizations and consequently degrades markedly under residual frequency offset (P min MD of order 0.4–1.0 in the affected scenarios). p. 1

Where the paper says it

05-million sample dataset, supported by t-SNE and SHAP/LIME explainability analyses. • Commercial Conformance Validation: Comprehensive evaluation with Nokia’s commercial physical-layer simulator across six propagation channels and 258 conducted conformance scenarios under 3GPP TS 38.

Because the OR rule cannot lower the false-alarm rate below the analytical baseline, the hybrid inherits the analytical detector’s sporadic FA spikes and exceeds PFA = 10−3 in a minority of the conformance scenarios (about 33 of 258). p. 6
Where the paper says it

The construction is practical because the score-thresholded ML branch has a near-zero false alarm rate (P ML FA ≈ 0), meaning the hybrid FAR effectively coincides with the reference analytical detector’s level across standard scenarios. However, in extreme multi-user or high power-offset scenarios where the analytical correlation detector experiences false alarm spikes (P ana FA > 10−3 ), the hard OR rule inherits these spikes.

On the cost side, the SVM is large (≈ 4. p. 7
Where the paper says it

Because the OR rule declares a preamble whenever either branch does, P hyb MD ≤ min(P ana MD , P ML MD) holds by construction; the empirical finding that the HYB_LOSS category never occurs therefore confirms that the implementation respects this guarantee rather than uncovering an unexpected property. The cost of the OR fusion instead appears on the false-alarm side, quantified in Table 4 below.

6 GB) and its 2.81 ms latency exceeds the gNB timing budget, whereas the CNN reduces both by more than an order of magnitude and, at a measured 258 µs per occasion on the same H100, meets the budget with margin. p. 4
Where the paper says it

The CNN, measured on the same H100 in fp32, processes a single PRACH occasion in a median of 258 µs (257 µs median, 267 µs at the 99th percentile) – about 3. 9× inside the 1 ms budget and roughly 10.

Future work

Future work includes closing the CNN’s operational sensitivity gap through a broader architecture search, inference-domain-matched preprocessing, and recalibration of the MAD-CFAR (k, smin) on a larger multi-channel noise campaign. A further direction is a confidence-aware fusion rule that combines the two branches by their soft, per-occasion confidence rather than their hard decisions. p. 6

Where the paper says it

However, in extreme multi-user or high power-offset scenarios where the analytical correlation detector experiences false alarm spikes (P ana FA > 10−3 ), the hard OR rule inherits these spikes. To eliminate inherited analytical spikes in commercial gNB DUs, generalized confidenceweighted soft fusion or PDP quality-gated OR rules can be integrated to suppress marginal analytical triggers when the ML classifier confidently attributes the occasion to noise.

Online adaptation of the threshold parameters and hardware-in-the-loop validation on real radio frequency complete the roadmap.

Acknowledgment

The authors thank Nokia for access to the LTE/5G physical-layer simulator. The paper is 2611. p. 0

Where the paper says it

To address these challenges, this article introduces SVM + CFAR + OR, a novel hybrid PRACH receiver architecture, and establishes a broader, model-agnostic ML + CFAR + OR framework. The proposed solution combines an analytical CFAR detector operating on the Power Delay 1 arXiv:2611.

07685. EVERY sentence must end with a reference token. p. 6
Where the paper says it

Input: received signal; PDP; classifier scores s (SVM decision values or CNN logits) Analytical branch: compute CFAR threshold on the PDP; obtain d ana i and exact TA ML branch: T ← max(se + k 1. 4826 MAD(s), smin); d ML i ← [ si > T ] for i = 1 to 64 do d hyb i ← d ana i ∨ d ML i if d ana i = 0 and d ML i = 1 then estimate TA for preamble i post hoc from the PDP end if end for return detections {d hyb i } and TA estimates is used, so every final decision is directly attributable to one component detector, keeping the receiver modular and fully interpretable.

--- Page 3 ---

The initial link establishment between user equipment (UE) and base stations (BS) in 5G New Radio (NR) and emerging sixth-generation (6G) networks relies on the random access (RA) procedure and the physical random access channel (PRACH). Standardized preambles constructed from Zadoff–Chu (ZC) sequences exhibit constant amplitude and zero cyclic autocorrelation, enabling analytical correlation receivers to detect transmitted preambles and estimate timing advance (TA). p. 0

Where the paper says it

Introduction The initial link establishment between user equipment (UE) and base stations (BS) in 5G New Radio (NR) and emerging sixth-generation (6G) networks relies on the random access (RA) procedure and the physical random access channel (PRACH). Standardized preambles constructed from Zadoff–Chu (ZC) sequences exhibit constant amplitude and zero cyclic autocorrelation, enabling analytical correlation receivers to detect transmitted preambles and estimate timing advance (TA).

However, target 6G network densities (10 7–10 8 devices/km2) under massive machine-type communication (mMTC) and AI-native radio architectures expose critical limitations of classical correlation detectors. p. 0
Where the paper says it

Standardized preambles constructed from Zadoff–Chu (ZC) sequences exhibit constant amplitude and zero cyclic autocorrelation, enabling analytical correlation receivers to detect transmitted preambles and estimate timing advance (TA). However, target 6G network densities (107–108 devices/km2 ) under massive machine-type communication (mMTC) and AI-native radio architectures expose critical limitations of classical correlation detectors.

Under low signal-to-noise ratios (SNR), severe multipath delay spreads, high Doppler shifts, and heavy multiple-access interference, analytical Constant False Alarm Rate (CFAR) detectors suffer from degraded sensitivity or excessive false alarms. p. 0
Where the paper says it

However, target 6G network densities (107–108 devices/km2 ) under massive machine-type communication (mMTC) and AI-native radio architectures expose critical limitations of classical correlation detectors. Under low signal-to-noise ratios (SNR), severe multipath delay spreads, high Doppler shifts, and heavy multiple-access interference, analytical Constant False Alarm Rate (CFAR) detectors suffer from degraded sensitivity or excessive false alarms.

Background

During contention-based random access, the UE selects one of 64 preambles generated from ZC root sequences and cyclic shifts, as defined in 3GPP TS 38.211 [7], and transmits it in a PRACH occasion. p. 1

Where the paper says it

1 PRACH and the Random Access Procedure During contention-based random access, the UE selects one of 64 preambles generated from ZC root sequences and cyclic shifts, as defined in 3GPP TS 38. 211 [7], and transmits it in a PRACH occasion.

The constant-amplitude, zero-autocorrelation property of ZC sequences ideally yields a sharp correlation peak whose position encodes the TA. p. 1
Where the paper says it

The BS receiver must decide which preambles – if any – are present and estimate their timing advance for the subsequent uplink grant. The constant-amplitude, zero-autocorrelation property of ZC sequences ideally yields a sharp correlation peak whose position encodes the TA.

The reference analytical detector correlates the received signal with each ZC root sequence, combines the perantenna correlations into a PDP, and applies an adaptive constant-false-alarm-rate (CFAR) threshold. p. 1
Where the paper says it

The constant-amplitude, zero-autocorrelation property of ZC sequences ideally yields a sharp correlation peak whose position encodes the TA. 2 Analytical Detector The reference analytical detector correlates the received signal with each ZC root sequence, combines the perantenna correlations into a PDP, and applies an adaptive constant-false-alarm-rate (CFAR) threshold.

This detector offers a predictable, analytically controlled FAR, but its sensitivity degrades when the modeling assumptions behind the CFAR threshold break down – for example, when the delay spread approaches the search window length or when inter-carrier interference (ICI) grows at high UE speed. p. 1
Where the paper says it

The threshold is set from a target false-alarm probability, the number of preambles, and the number of receive antennas; the position of the first path in the detected window yields the TA estimate. This detector offers a predictable, analytically controlled FAR, but its sensitivity degrades when the modeling assumptions behind the CFAR threshold break down – for example, when the delay spread approaches the search window length or when inter-carrier interference (ICI) grows at high UE speed.

Problem Formulation and Metrics

Each PRACH occasion is mapped to a feature vector, and preamble detection is posed as a 64-class classification problem with an additional noise/no-preamble outcome governed by adaptive score-domain thresholding. Two metrics dictate receiver evaluation: the missed detection rate (MDR), i. p. 1

Where the paper says it

1 Problem Formulation and Metrics Each PRACH occasion is mapped to a feature vector, and preamble detection is posed as a 64-class classification problem with an additional noise/no-preamble outcome governed by adaptive score-domain thresholding. Two metrics dictate receiver evaluation: the missed detection rate (MDR), i.

e., the probability of failing to detect a transmitted preamble, and the false alarm rate (FAR), i. p. 1
Where the paper says it

Two metrics dictate receiver evaluation: the missed detection rate (MDR), i. , the probability of failing to detect a transmitted preamble, and the false alarm rate (FAR), i.

e., the probability of declaring a preamble on a noise-only position. p. 1
Where the paper says it

, the probability of failing to detect a transmitted preamble, and the false alarm rate (FAR), i. , the probability of declaring a preamble on a noise-only position.

Per 3GPP base station conducted conformance requirements (TS 38. p. 1
Where the paper says it

In contrast, this work addresses direct 64-class preamble identification from raw IQ samples under strict 3GPP TS 38. 141 base station conducted conformance requirements.

141-1 [1]), an acceptable receiver must maintain PFA ≤ 10−3 while achieving a detection probability of at least 99%. p. 2
Where the paper says it

requirements (TS 38. 141-1 [1]), an acceptable receiver must maintain PFA ≤ 10−3 while achieving a detection probability of at least 99%.

Computational Complexity

Counting one fused multiply–add as 2 FLOPs, the PCA projection from D = 1678 to K = 515 costs FLOPsPCA ≈ 2DK ≈ 1.73 MFLOPs, whereas the degree-2 polynomial kernel evaluated over all support vectors of the 64 OvA classifiers costs FLOPskernel = Ncls N¯ SV (2K + 5) ≈ 1. p. 4

Where the paper says it

Counting one fused multiply–add as 2 FLOPs, the PCA projection from D = 1678 to K = 515 costs FLOPsPCA ≈ 2DK ≈ 1. 73 MFLOPs, whereas the degree-2 polynomial kernel evaluated over all support vectors of the 64 OvA classifiers costs FLOPskernel = Ncls N¯ SV (2K + 5) ≈ 1.

16 GFLOPs. The total is ≈ 1158 MFLOPs per prediction, of which the kernel evaluation accounts for 99. p. 4
Where the paper says it

16 GFLOPs [16, 17]. The total is ≈ 1158 MFLOPs per prediction, of which the kernel evaluation accounts for 99.

7%, PCA for 0. p. 1
Where the paper says it

The key methodological novelty is twofold: (i) an adaptive median absolute deviation CFAR threshold (MAD-CFAR) applied directly to raw classifier decision scores, driving false alarms to zero on a 28-million position noise campaign, and (ii) a logical OR decision fusion that guarantees zero missed-detection regression relative to the analytical reference detector. To prove model independence within the ML + CFAR + OR framework, the learned branch is instantiated with both a kernel Support Vector Machine (SVM with PCA) and a lightweight 1-D Convolutional Neural Network (CNN), both benchmarked on an identical 1.

15%, and argmax for under 0. p. 4
Where the paper says it

5 Computational Complexity Deployment in a real-time BS receiver requires the peroccasion inference cost to fit the gNB processing budget. The inference pipeline has three stages: PCA projection, One-vs-All kernel evaluation, and argmax.

001%. The CNN adopted in Section 3. p. 4
Where the paper says it

3 compares the proposed polynomial-kernel SVM with alternative classifiers: it is slightly cheaper than an RBF-kernel SVM (≈ 1506 MFLOPs) but far more expensive than compact neural models. The CNN adopted in Section 3.

3 (three convolutional and three dense layers) costs ≈ 154 MFLOPs per occasion (≈ 77 M multiply–accumulates: 97. p. 4
Where the paper says it

The CNN adopted in Section 3. 3 (three convolutional and three dense layers) costs ≈ 154 MFLOPs per occasion (≈ 77 M multiply–accumulates: 97.

7 MFLOPs in the convolutional layers and 56. p. 4
Where the paper says it

The CNN adopted in Section 3. 3 (three convolutional and three dense layers) costs ≈ 154 MFLOPs per occasion (≈ 77 M multiply–accumulates: 97.

1 MFLOPs in the dense layers) – roughly 8× below the SVM. p. 4
Where the paper says it

7 MFLOPs in the convolutional layers and 56. 1 MFLOPs in the dense layers) – roughly 8× below the SVM – and stores its ≈ 28 million parameters in only ≈ 112 MB, over an order of magnitude smaller than the SVM’s 2.

Noise Floor vs. p. 5

Where the paper says it

23% at pre-calibrated zero false alarms. Noise Floor vs.

Dynamic Dispersion

The threshold parameters are calibrated on a dedicated dataset, separate from the SVM training set, comprising a signal campaign (∼ 248 000 preamble occasions) and a noise-only campaign (∼ 28 million positions with no preamble transmitted). The selected operating point (k = 5. p. 5

Where the paper says it

35), the dynamic Figure 4: FAR versus MDR trade-off over all tested (k, smin) pairs. The selected operating point (k = 5.

5, smin = 11. p. 4
Where the paper says it

, s64] ⊤ of per-class classifier scores of a single PRACH occasion; these are SVM decision values or, equivalently, CNN logits, so the mechanism is agnostic to the underlying model. Because the k ·MAD term is dimensionless, only the two parameters (k, smin) are re-calibrated per classifier.

0) attains zero FAR; the alternative point (k = 5. p. 5
Where the paper says it

5, smin = 11. 0) attains zero FAR; the alternative point (k = 5.

5, smin = 4. p. 4
Where the paper says it

, s64] ⊤ of per-class classifier scores of a single PRACH occasion; these are SVM decision values or, equivalently, CNN logits, so the mechanism is agnostic to the underlying model. Because the k ·MAD term is dimensionless, only the two parameters (k, smin) are re-calibrated per classifier.

5) trades a higher detection rate for a nonzero FAR. p. 5
Where the paper says it

5, smin = 4. 5) trades a higher detection rate for a nonzero FAR.

The identical procedure applied to CNN logits yields k∗ = 3. p. 5
Where the paper says it

2 × 10−4 with continuously active MAD thresholding (Fig. The identical procedure applied to CNN logits yields k ∗ = 3.

0.

Table 4: Performance Comparison across Representative Conformance Scenarios

The hybrid receiver consistently matches the better component and, in the frequencyoffset case, strictly outperforms both (SNRMD improved by 1 dB), demonstrating OR decision gains. The hybrid keeps P max FA ≤ 10−3 in most cases, though in line with the union bound (7), it inherits the analytical detector’s occasional FA spikes in extreme scenarios. p. 8

Where the paper says it

0 HYB_BEST, HYB_TIE ∼3 explicit OR synergy (HYB better than either component) HYB_LOSS 0 HYB never degrades the ANA result tection through the SVM. The hybrid keeps P max FA ≤ 10−3 in most cases, though in line with the union bound (7), it inherits the analytical detector’s occasional FA spikes in extreme scenarios.

Table 3: Aggregated verdicts over the 258 conformancestyle scenarios

The hybrid keeps P max FA ≤ 10−3 in most cases, though in line with the union bound (7), it inherits the analytical detector’s occasional FA spikes in extreme scenarios. Across the full set, PFA ≤ 10−3 is satisfied in approximately 225 of the 258 scenarios. p. 8

Where the paper says it

The hybrid keeps P max FA ≤ 10−3 in most cases, though in line with the union bound (7), it inherits the analytical detector’s occasional FA spikes in extreme scenarios. Across the full set, PFA ≤ 10−3 is satisfied in approximately 225 of the 258 scenarios.

The cost of the OR fusion instead appears on the false-alarm side, quantified in Table 4 below. p. 7
Where the paper says it

Because the OR rule declares a preamble whenever either branch does, P hyb MD ≤ min(P ana MD , P ML MD) holds by construction; the empirical finding that the HYB_LOSS category never occurs therefore confirms that the implementation respects this guarantee rather than uncovering an unexpected property. The cost of the OR fusion instead appears on the false-alarm side, quantified in Table 4 below.

Table 5: Qualitative Comparison of Representative Machine-Learning-Based PRACH Receivers

This work Raw IQ / PCA samples 64 Preambles (k = 64) MAD-CFAR Several-dB MDR gain Primary Detector Reference.

Table 2: Faithfulness test: classification accuracy versus the number k of modified PCA components

The gap between the SHAP and random curves confirms that the attributions are faithful to the model. The off-diagonal SHAP coupling is a property of the trained OvA decision functions, not of the ZC sequences themselves. p. 4

Where the paper says it

0345 that holds for any distinct pair of prime-length ZC preambles. The two sequences are therefore orthogonal in the cyclic-correlation sense, refuting the hypothesis of a physical similarity: the off-diagonal SHAP coupling is a property of the trained OvA decision functions, not of the ZC sequences themselves.

Table 1: Learned detectors in the same 64-class formulation and dataset

The CNN is slightly more accurate than the SVM at a fraction of the inference and memory cost. The CNN is marginally more accurate – 0. p. 3

Where the paper says it

Table 1 summarizes the head-to-head comparison in the identical 64-class formulation. In offline classification the CNN is marginally more accurate – 0.

71 percentage points. p. 3
Where the paper says it

In offline classification the CNN is marginally more accurate – 0. 71 percentage points Table 1: Learned detectors in the same 64-class formulation and dataset.

**Figure 3: Inference cost (MFLOPs per prediction, log scale) of the proposed polynomial-kernel SVM against an RBF-kernel SVM (estimate), a two-layer MLP, and a three-layer 1-D CNN. p. 5

Where the paper says it

Figure 3: Inference cost (MFLOPs per prediction, log scale) of the proposed polynomial-kernel SVM against an RBF-kernel SVM (estimate), a two-layer MLP, and a three-layer 1-D CNN. The kernel evaluation dominates the SVM cost.

The kernel evaluation dominates the SVM cost. p. 5
Where the paper says it

Figure 3: Inference cost (MFLOPs per prediction, log scale) of the proposed polynomial-kernel SVM against an RBF-kernel SVM (estimate), a two-layer MLP, and a three-layer 1-D CNN. The kernel evaluation dominates the SVM cost.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Robust adaptive thresholding in score domain MAD-CFAR allows for enforcement of constant false alarm rate on raw classifier decision scores, driving false alarms to zero on a 28-million position noise campaign. This mechanism is model agnostic and applies directly to both SVM and CNN logits.

  2. Logical OR decision fusion guarantees a no missed-detection regression relative to the analytical reference detector, meaning the hybrid receiver has an MDR that obeys P hyb MD <= min(P ana MD, P ML MD). This prevents missed detections when one branch fails, ensuring reliability.

  3. The system can achieve superior low signal-to-noise ratio sensitivity by combining the analytical and learned branches; for instance, in EPA at SNR = -20 dB, the hybrid MDR is 0.644 while the analytical detector is 0.910, demonstrating a clear advantage.

  4. A lightweight 1D CNN branch can be deployed as a cost-effective alternative to SVM; it achieves high offline accuracy (99.29%) with an order of magnitude smaller model size and lower inference cost (154 MFLOPs per occasion versus 1158 MFLOPs for the SVM).

  5. The system can perform post hoc timing advance estimation for ML-only detections by mapping the preamble index to its ZC sequence pair and refining it with parabolic sub-sample interpolation. This allows for TA estimation when only the ML branch detects a preamble.

  6. The system offers interpretability of classifier decisions through SHAP and LIME analyses, showing that importance is spread across the entire sequence consistent with the ZC sequences' CAZAC property, verifying that the classifier exploits the global signal structure.

Abstract

The physical random access channel (PRACH) enables initial access between user equipment and the base station, making reliable preamble detection essential for cellular networks. Growing device density in 5G and beyond-5G networks, massive machine-type communication, and the shift towards machine-learning-driven radio interfaces challenge conventional analytical detectors, particularly in complex radio environments and at low signal-to-noise ratios. This article formulates preamble detection as a 64-class classification problem and introduces two model-agnostic contributions: a robust adaptive threshold implementing the constant false alarm rate (CFAR) principle in the classifier decision-score domain (MAD-CFAR), and a hybrid receiver combining a learned detector with a conventional power-delay-profile CFAR detector through a logical OR rule. The learned branch is evaluated using two models trained on the same GPU-generated dataset: a support vector machine (SVM) with principal component analysis and a lightweight one-dimensional convolutional neural network (CNN). On a held-out test set, the SVM achieves 98.58% accuracy (macro-AUC 0.9991), while the CNN reaches 99.29% accuracy (macro-AUC 1.0000) at lower inference cost and memory footprint. MAD-CFAR reduces the learned branch's false alarm rate to near zero on the noise calibration campaign. The OR hybrid retains the analytical detector's low, standard-compliant false alarm rate while extending detection capability. Across six propagation channels and 258 conformance-style tests, the SVM-based hybrid improves low-SNR sensitivity by several decibels without increasing missed detections relative to the analytical detector. Under the same false alarm constraint, the CNN is competitive in some channels but currently trails the SVM in high-delay-spread multipath conditions.

Sources

Related papers