BIT Technology Research Series • Signal Processing & RF

AI-Driven Radio Frequency Spectrum Quality Analysis Using a Software-Defined Receiver: ML Noise Classification and Spectral Anomaly Detection in the 24 MHz – 1766 MHz Band

Ing. Stanislav Pittner

BIT Technology s.r.o., Trstínska cesta 9, 917 01 Trnava, Slovakia

Published: September 5, 2025 Measurement period: April – August 2025 Version 1.0

DOI: 10.5281/bittechnology.2025.tr006 (preprint)

Abstract

We present an AI-driven system for continuous radio frequency spectrum analysis in the 24 MHz – 1766 MHz band based on a software-defined receiver (SDR) RTL2832U with an R820T tuner. The system combines a GNU Radio signal processing pipeline with ML models for automatic modulation classification (accuracy > 92 %), noise type classification, and spectral anomaly identification with latency < 2 seconds. On a dataset of 2.4 million spectral snapshots collected during 5 months of monitoring, we evaluate four ML architectures: CNN (accuracy 94.1 %), Random Forest (91.8 %), SVM (89.3 %), and Autoencoder for unsupervised anomaly detection (F1 = 0.87). The system monitors aviation communication (118–136 MHz, BTS and VIE airports), FM broadcasting (87.5–108 MHz), NB-IoT Guard Band B20 (791–821 MHz), and LoRa (868 MHz). We present a dual-use analysis (civilian and security applications) and a comparison with commercial spectrum analyzers.

Keywords: SDR, RTL-SDR, spectrum monitoring, machine learning, modulation classification, anomaly detection, signal processing, cognitive radio, dual-use

1. Introduction

The radio frequency spectrum is an increasingly congested resource — with the growth of IoT devices, 5G NR, and unlicensed bands, the risk of interference and signal quality degradation is rising. Traditional spectrum monitoring requires specialized equipment (Rohde & Schwarz, Keysight) in the price range of 10,000 – 100,000 EUR, which limits widespread deployment.

The Software-Defined Radio (SDR) paradigm (Mitola, IEEE 1995) enables the implementation of radio systems in software, dramatically reducing costs. An RTL-SDR dongle (RTL2832U + R820T tuner) at ~25 EUR provides a receiver covering 24 MHz – 1766 MHz with an 8-bit ADC and 2.4 MHz bandwidth. Combined with ML models for automatic classification and anomaly detection, it creates a cost-effective platform for intelligent spectrum monitoring.

O'Shea et al. (IEEE TCCN 2018) demonstrated that CNNs achieve significantly better results in spectrum occupancy detection compared to traditional energy detection methods. Rajendran et al. (IEEE Access 2018) showed accuracy > 95 % for automatic modulation classification at SNR > 5 dB.

This study evaluates: (1) the performance of ML models for modulation and noise classification on RTL-SDR data, (2) real-time anomaly detection capability, and (3) dual use of the system in civilian and security contexts.

2. System Architecture

SDR AI Signal Analyzer — Processing Pipeline

RTL-SDR
R820T Tuner

→

GNU Radio
IQ Sampling

→

Feature Extract
PSD, SNR, Kurtosis

→

ML Classifier
CNN / RF / SVM

→

Anomaly Detect
Autoencoder

→

Dashboard
Waterfall + Alerts

RTL2832U • 24–1766 MHz • 2.4 MHz bandwidth • 8-bit ADC • External antenna + bias-tee

3. Data and Methods

3.1 Monitored Frequency Bands

Table 1. Monitored frequency bands and their characteristics.

Band	Frequency	Modulation	Monitoring Purpose	Typical SNR
Aviation communication	118 – 136 MHz	AM (DSB)	BTS/VIE airport safety	15–35 dB
FM radio	87.5 – 108 MHz	FM (WFM)	Broadcast quality	25–55 dB
DAB+	174 – 230 MHz	OFDM	Digital broadcasting	10–30 dB
NB-IoT B20	791 – 821 MHz	OFDMA (SC-FDMA)	IoT coverage	5–20 dB
LoRa ISM	868 MHz	CSS (Chirp)	IoT sensors	−5–15 dB
ADS-B	1090 MHz	PPM	Aircraft tracking	10–25 dB

3.2 Feature Set

Table 2. Extracted spectral features for ML models.

Feature	Description	Dimension	Relevance
PSD (Power Spectral Density)	FFT-based spectral power density	256 bins	Primary feature
Spectral Flatness	Wiener entropy (noise vs. signal)	1	Noise classification
Bandwidth (3dB/10dB)	Signal bandwidth	2	Modulation ID
SNR	Signal-to-Noise Ratio	1	Signal quality
Kurtosis	Impulsive signal characteristic	1	Impulsive noise
Crest Factor	Peak-to-RMS ratio	1	Modulation ID
IQ histogram	I/Q sample distribution	64 bins	Constellation analysis

3.3 ML Models

Table 3. Compared ML architectures.

Model	Input	Parameters	Task Type
CNN (1D-Conv)	PSD + IQ histogram (320-dim)	1.2M	Modulation classification
Random Forest	Extracted features (6-dim)	500 trees	Modulation classification
SVM (RBF kernel)	Extracted features (6-dim)	—	Modulation classification
Autoencoder (Conv)	PSD (256-dim)	0.8M	Anomaly detection
LSTM	PSD time-series (32 steps)	0.5M	Temporal anomaly

Dataset: 2.4 million spectral snapshots from 5 months of continuous monitoring. Split: 70/15/15 (train/val/test). Data augmentation: SNR jitter (±5 dB), frequency offset (±1 kHz).

4. Results

4.1 Modulation Classification

Figure 1. Classification accuracy by modulation type for the CNN model (overall accuracy 94.1 %). AM achieves the highest accuracy (96.3 %) due to its simple envelope characteristic.

Table 4. Overall classification accuracy by model (test set, n = 360,000).

Model	Accuracy	Precision	Recall	F1	Inference (ms)
CNN (1D-Conv)	94.1%	0.93	0.94	0.935	2.1
Random Forest	91.8%	0.91	0.92	0.913	0.3
SVM (RBF)	89.3%	0.88	0.89	0.887	1.8
LSTM	90.7%	0.90	0.91	0.904	8.4

4.2 Anomaly Detection

Figure 2. SNR estimation accuracy by frequency band (CNN model). The aviation band achieves the best accuracy due to high SNR and simple AM modulation.

Table 5. Anomaly detection performance (Autoencoder, threshold = 95th percentile reconstruction error).

Anomaly Type	Precision	Recall	F1	False Alarm Rate
Narrowband interference (jammer)	0.94	0.91	0.924	2.1%
Impulsive noise	0.89	0.86	0.874	3.8%
Illegal transmitter	0.87	0.82	0.844	4.2%
Multipath fading	0.78	0.74	0.759	6.1%
Overall (weighted average)	0.88	0.85	0.865	3.7%

5. Dual Use

Table 6. Civilian and security applications of the system.

Domain	Civilian Applications	Security Applications
Aviation band	ATC communication quality monitoring	GPS spoofing detection, jammer detection
FM radio	Broadcast quality audit, coverage mapping	Detection of illegal (pirate) transmitters
NB-IoT / LoRa	IoT coverage verification, Smart City	Critical IoT infrastructure monitoring
Wideband	EMC compliance testing	SIGINT, situational awareness, CBRN sensors

6. Comparison with Alternatives

Table 7. Comparison of the SDR AI system with commercial solutions.

Device	Range	Bandwidth	ADC	AI/ML	Price
RTL-SDR + AI (ours)	24–1766 MHz	2.4 MHz	8-bit	CNN + AE	~50 EUR
HackRF One	1–6000 MHz	20 MHz	8-bit	No (GNU Radio)	~350 EUR
Airspy R2	24–1800 MHz	10 MHz	12-bit	No	~200 EUR
R&S FSV3000	10 Hz–44 GHz	200 MHz	14-bit	Optional	~50,000 EUR
Keysight N9000B	9 kHz–26.5 GHz	160 MHz	14-bit	Optional	~30,000 EUR

7. Recommendations for Future Research

Table 8. Proposed system extensions.

Strategy	Predicted Impact	Complexity
GPU-accelerated processing (CUDA)	10× throughput, real-time wideband	Low
HackRF upgrade (1–6 GHz, 20 MHz BW)	5G NR monitoring, WiFi 6 analysis	Low
Distributed sensor network (5+ nodes)	Transmitter geolocation, coverage map	Medium
Deep learning protocol identification	Automatic protocol ID (GSM, LTE, 5G)	Medium
FPGA real-time processing (RFSoC)	GHz bandwidth, ns latency	High

8. Conclusions

1. CNN achieves 94.1 % accuracy in modulation classification on RTL-SDR data, which is comparable to results on synthetic datasets (O'Shea et al. 2018) and confirms the viability of low-cost SDR for AI-driven monitoring.

2. Autoencoder anomaly detection with F1 = 0.87 and a false alarm rate of 3.7 % enables reliable detection of jammers, illegal transmitters, and impulsive noise in real time.

3. RTL-SDR at ~50 EUR combined with ML models provides 85–90 % of the functionality of commercial spectrum analyzers at ~0.1 % of their cost for monitoring and screening applications.

4. Dual use (civilian + security) increases the system's value — from Smart City IoT monitoring through aviation safety to SIGINT situational awareness.

5. Main limitation: 8-bit ADC and 2.4 MHz bandwidth of the RTL-SDR restricts dynamic range and wideband analysis. An upgrade to HackRF (20 MHz BW) or Airspy (12-bit) would significantly expand capability at a reasonable cost.

References

Mitola, J. (1995). The software radio architecture. IEEE Communications Magazine, 33(5), 26–38.
Haykin, S. (2005). Cognitive radio: brain-empowered wireless communications. IEEE JSAC, 23(2), 201–220.
O'Shea, T.J., et al. (2018). Over-the-air deep learning based radio signal classification. IEEE JSAC, 36(1), 132–141.
Rajendran, S., et al. (2018). Deep learning models for wireless signal classification. IEEE Access, 6, 18642–18652.
Yucek, T. & Arslan, H. (2009). A survey of spectrum sensing algorithms for cognitive radio. IEEE Communications Surveys, 11(1), 116–130.
West, N.E. & O'Shea, T. (2017). Deep architectures for modulation recognition. IEEE DySPAN 2017.
Liu, X., et al. (2017). Data-driven spectrum sensing using deep learning. IEEE Communications Letters, 21(12), 2614–2617.
Selim, A., et al. (2018). Spectrum monitoring for 5G using deep learning. IEEE ICC 2018.
Stewart, R.W. (2015). Software Defined Radio using MATLAB & Simulink and the RTL-SDR. Strathclyde Academic Media.
GNU Radio Project (2024). GNU Radio Manual and Reference. gnuradio.org.
RTL-SDR Blog (2024). RTL-SDR V4 Specifications and Applications. rtl-sdr.com.
ETSI (2020). EN 303 413: Satellite Earth Stations and Systems; GNSS based location systems. ETSI.

Hardware: RTL2832U + R820T tuner, external dipole antenna, Raspberry Pi 4 (8 GB).

Project reference: bittechnology.bemooore.com/referencie/sdr-receiver-ai

Author: Ing. Stanislav Pittner, CEO of BIT Technology s.r.o.