An AI model designed to untangle several radio interference signals at once was less reliable at finding every source as the mix grew crowded, according to a preprint marked under review in October 2026. In overlapping mixtures, its F1 score, a measure that balances precision and recall, fell from 93.9% with two sources to 74.6% with six. Recall, the share of actual sources the model found, fell from 0.93 to 0.69. Yet its type accuracy was 95.2%, and accuracy for the model's BOC modulation label was 96.5%.
A six-slot way to read a crowded spectrum
PI-SCAN takes a logarithmic power spectrogram, a map of signal energy across frequency and time, built from complex in-phase and quadrature, or IQ, samples. It turns that map into six unordered prediction slots. Each slot estimates whether a source is present, known as objectness, along with its interference type, center frequency, bandwidth, received signal strength and BOC modulation. It also produces a source-specific mask showing where that source appears in the spectrogram. Hungarian assignment is used to match predictions with sources without depending on their order.
The test data were synthetic mixtures generated in the Sionna ray-tracing environment, not measured field recordings. They contained one to six active interference sources drawn from 94 waveform configurations across six interference classes. Models were trained in batches of 32 for 100 epochs with Adam, and the reported results used a MobileNetV2-U-Net architecture without attention.
The trade-off in the scores
With all prediction objectives enabled, the reference configuration reported an F1 score of 74.8%. It classified interference type with 97.3% accuracy and BOC modulation with 98.4% accuracy. Its reported errors were 1.29 MHz for center frequency, 2.54 MHz for bandwidth and 3.19 dB for received signal strength. In the reported results, the full version paired a lower overall detection score with high scores on the characterization tasks.
A leaner configuration that predicted objectness, type and masks reported precision of 0.98, recall of 0.85 and an F1 score of 91.1%. Its type accuracy was 98.7%. Another configuration, with signal-strength estimation removed, reported an F1 score of 83.5%, count accuracy of 61.7%, frequency error of 1.12 MHz and bandwidth error of 1.62 MHz. Here, precision means the share of predicted sources that were correct, while recall measures the share of actual sources found.
Taken together, the configurations point to a trade-off between broad characterization and the detection measures reported in the test. The version with masks and type prediction scored better on F1 than the full objective, while the version without signal-strength estimation had different detection, frequency and bandwidth results. Those comparisons show how the objectives behaved in this setup, but they do not establish performance on real-world measurements.
Crowding, comparisons and speed
The number of overlapping sources was a clear fault line. F1 fell from 93.9% with two sources to 74.6% with six, and recall dropped from 0.93 to 0.69. Type accuracy and BOC accuracy in those overlapping tests were 95.2% and 96.5%. The reported scores suggest that recognizing the nature of a matched source remained easier than finding all the sources in the most crowded mixtures.
In the reference comparison, RCJ had F1 values of 90.9% to 92.2%, count accuracy of 64.5% to 75.1% and type accuracy of 96.5% to 99.0%. HFI's F1 ranged from 38.2% to 56.1%, while TN recorded 0.9% and 5.1% under two matching settings. The reduced PI-SCAN configuration was described as competitive with RCJ while also producing source-specific masks.
The masks offered a more detailed view of the output. In reported examples, IoU, a score for how closely predicted and target masks overlap, declined from 0.890 to 0.795. A related Dice score fell from 0.942 to 0.885, and precision from 0.969 to 0.869, while recall stayed above 0.90. Inference typically took about 8 to 9 milliseconds per sample on a GPU, compared with about 240 to 260 milliseconds on a CPU.
What the test leaves open
The biggest qualification is that the test never left the simulated world. The experiments were trained and evaluated on synthetic signals propagated in Sionna's ray-tracing environment, so they do not establish performance on real-world measurements. The fixed six-slot design also limits the model to at most six individually resolved sources per spectrogram; mixtures with more sources may leave some sources undetected or merge them.
The supplied manuscript is marked under review in October 2026. The work was carried out within the SAIP project, funded under code 50NP2603A, sponsored by BMWK and BMV, and supported by DLR, BNetzA and BKG.
Paper data and sources
Original title: Permutation-Invariant Time-Frequency Segmentation and Parameter Estimation of Multiple Overlapping Interference Signals
Authors: Lucas Heublein, Christian Wielenberg, Christopher Mutschler, Felix Ott
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-02
DOI: Not available
Original paper · Full text