Machine Learning RF Signal Detection with uSDR and the WSDR Platform

Machine learning is becoming an increasingly important tool for RF signal analysis.

Traditional RF analysis relies on well-understood mathematical tools: FFTs, filters, demodulators, synchronization loops, correlation methods, and protocol-specific decoders. These tools remain essential. But machine learning can add another layer by helping classify patterns, recognize modulation types, identify signal behavior, and support anomaly detection.

Wavelet Lab has been exploring how machine learning can be used with uSDR and the WSDR platform for RF signal detection and modulation recognition.

Starting with modulation recognition

RF machine learning is a broad and difficult problem. Instead of trying to solve every RF classification challenge at once, Wavelet Lab started with a more focused problem: modulation recognition.

The approach converts an I/Q sample stream into normalized vector I/Q diagrams. These diagrams can then be processed by a neural network in a way that is similar to image recognition.

This is a practical starting point because modulation types often produce recognizable constellation patterns, even when the final system will need more advanced models, better datasets, and stronger validation.

Browser-connected RF intelligence

The WSDR platform creates an interesting environment for RF machine learning because it connects SDR hardware, visualization, data handling, and application logic through a browser-based workflow.

A machine-learning RF application can include:

• SDR hardware collecting live I/Q samples
• preprocessing blocks preparing signal data
• constellation or spectrogram generation
• backend training workflows
• frontend inference using trained models
• browser-based visualization and user interaction

This makes it possible to prototype RF intelligence features without forcing every user to build a full native ML and SDR stack from scratch.

Backend training and frontend inference

A practical ML workflow can separate training from inference.

The backend can handle model training, dataset management, and storage of trained results. The frontend can apply a trained model to live or recorded I/Q data.

This architecture is useful because training is usually heavier and may require more compute resources, while inference can be lighter and more interactive.

For SDR users, this means trained models can potentially be reused across different applications, signal sources, and browser-based workflows.

Dataset direction

Machine learning performance depends heavily on dataset quality.

For RF systems, datasets must capture real signal variation: noise, gain changes, frequency offset, multipath, interference, different sampling rates, and hardware-specific behavior.

Wavelet Lab's direction includes support for structured signal data workflows, including formats such as SigMF for I/Q recordings and metadata. This helps make RF datasets easier to organize, share, and use for training or testing.

Applications

Machine learning for RF signal detection can support:

• modulation recognition
• signal classification
• RF anomaly detection
• spectrum monitoring
• protocol identification research
• interference detection
• distributed sensing
• training and education
• signal intelligence  workflows
• automated RF monitoring

From research to deployable workflows

The long-term value is not only in recognizing a single modulation type. The larger goal is to combine SDR hardware, browser-based tools, datasets, and ML inference into workflows that help users understand the RF environment faster.

This is especially relevant for distributed SDR systems, monitoring applications, and future RF situational awareness tools.

uSDR and WSDR provide a compact platform for experimenting with these ideas and moving toward more practical AI-assisted RF applications.