Machine Learning

CRFS's Machine Learning product development is pioneering the application of machine learning and artificial intelligence technologies to better monitor and manage the radio spectrum.

Signal classification

The use of neural networks has changed the way signals are processed to determine modulation and signalling type. Mathematical tools to train neural networks, combined with large signal training data sets captured by CRFS networks, means that signal classifiers can be trained much more quickly and more accurately than using traditional techniques.

Accurate signal classification with the ability to easily update signal classifier databases with new signal types is key to many spectrum monitoring and management problems such as interference detection, threat awareness, and border and perimeter security. These applications rely on maximising threat or interference detection and identification, while minimising false alarms. Machine learning techniques are ideal for optimising this classification performance

How does it work?

During the training process, the performance of the neural net can be monitored using a confusion matrix as shown below. The input data for training neural nets is a large database of signal types collected from real world RF environments using CRFS distributed RF sensors located all over the world. The more signal types, with real world distortions such as multipath, the more effectively the neural net can be trained to perform in actual field deployments.

Machine learning confusion matrix

Confusion matrix for signal recognition

The CRFS training database uses “tagged” signals with known classifications to train the neural net. The training process adjusts individual neurones in the net so that they can identify the known signals correctly i.e. the classifier’s calculation and the “tag” match.

To test the neural network, a separate set of known signals not used for the training process is input to the classifier. The confusion matrix shows the known input signal type against the classifier’s calculation for every test and increments the appropriate element of the matrix with the answer against the test input. In other words, a 100% correct classifier would produce a confusion matrix with non-zero elements only along the leading diagonal. Generally, test scores of >95% accuracy can be achieved. These test scores improve with more varied and deeper training data sets. CRFS is actively building massive tagged signal training databases for this purpose.

machine learning neural net
Typical signal classifier neural network

The input layer of a neural net has a series of weightings from features extracted from signal IQ data e.g. 20 feature weightings, 20 input neurons. The output layer has a neuron for each signal type to be classified e.g. 30 signal types, 30 output neurons. The neural net selects the output signal type to reflect the most probable signal given the input features.


Machine Learning and RF Spectrum Intelligence Gathering White Paper