Spectrum monitoring and analysis, and related methods, systems, and devices
Disclosed embodiments relate to ensemble wireless signal classification and systems and devices the incorporate the same. Some embodiments of ensemble wireless signal classification may include energy-based classification processes and machine learning-based classification processes. In some embodiments, incremental machine learning techniques may be incorporated to add new machine learning-based classifiers to a system or update existing machine learning-based classifiers.
1. A method of monitoring a wireless environment, comprising:
monitoring a wireless environment for known signal classes;
extracting one or more features from a signal detected in the wireless environment using one or more novelty detection models trained, at least in part, using known signals of the known signal classes;
training one or more binary classifier models using data indicative of noise and data indicative of a signal;
classifying the detected signal as noise or a signal using the one or more binary classifier models;
determining that the signal is unknown responsive to the one or more features being different from features of each of the known signal classes and classifying the detected signal as a signal class;
labeling the signal as an unknown signal responsive to the determination that the signal is unknown;
training a learned wireless-signal classifier model using the labeled unknown signal;
defining a new known signal class using feature representations of the learned wireless-signal classifier model; and
adding the new known signal class to the known signal classes used for monitoring the wireless environment.
2. The method of claim 1 , wherein the defining the new known signal class using feature representations of the learned wireless-signal classifier model comprises:
defining the new known signal class using cyclostationary feature representations of the learned wireless-signal classifier.
3. The method of claim 1 , wherein the defining the new known signal class using feature representations of the learned wireless-signal classifier model comprises:
defining the new known signal class using shift-invariant feature representations of the learned wireless-signal classifier.
4. The method of claim 3 , wherein the defining the new known signal class using the shift-invariant feature representations of the learned wireless-signal classifier comprises:
defining the new known signal class using an alpha-profile derived from a Spectral Correlation Function.
5. A system, comprising:
a radio;
a processor; and
a non-transitory computer-readable memory, wherein the non-transitory computer-readable memory is configured to store:
known signal classes;
signal data for detected known signals, unknown signals, and noise; and
instructions executable by the processor, the instructions adapted to enable the processor to perform operations for claim 1 .
6. A method of monitoring a wireless environment, comprising:
monitoring a wireless environment for known signal classes;
extracting one or more features from a signal detected in the wireless environment using one or more novelty detection models trained, at least in part, using known signals of the known signal classes;
determining that the detected signal is not a known signal class responsive to the one or more features being different from features of the known signal classes;
classifying the detected signal as noise or a signal using a learned binary classifier, wherein the learned binary classifier was trained using data sets comprising data indicative of noise and data indicative of a signal; and
labeling the detected signal as unknown responsive to the determination that the detected signal is not a known signal class and classifying the detected signal as a signal.
7. The method of claim 6 , wherein the learned noise/signal classifier model was trained using supervising learning techniques to distinguish between noise and signals using training data comprising noise and signals for a given frequency spectrum.
8. A signal monitoring system, the system comprising:
a feature set calculator configured to generate feature sets for signals detected in a deployed environment;
classifier models configured to classify the signals detected in the deployed environment responsive to the calculated feature sets, the classifiers comprising:
a learned binary classifier model configured to classify one or more signals as a respective noise class or a signal class, the learned binary classifier model trained using data labeled as noise data and data labeled as signal data; and
learned known signal classifier models, the learned known signal classifier models configured to sort the one or more signals into respective known signal classes responsive to a determination that a feature set for each of the one or more signals corresponds to a known signal class of the known signal classes;
a result merger configured to label one or more signals as unknown responsive to the determination that a feature set of each of the one or more signals does not correspond to a known signal class of the known signal classes and classifying the one or more signals as a signal class; and
a classification engine, configured to train a new learned known signal classifier using the labeled unknown signal.
9. The system of claim 8 , wherein the classification engine is configured to define a new known signal class using feature representations of the learned known signal classifier model.
10. The system of claim 8 , wherein the learned signal/noise classifier model is configured to classify the unknown signal as belonging to a signal class responsive to detecting the absence of noise.