Animal sensory systems, such as vision and acoustic sensing, have evolved over millions of years to detect, identify, and respond to novel events that could pose a threat or deliver a reward. As a result, when a new sound or sight is observed, most animals will make an immediate classification as a friend, foe, or neutral without an explicit model or expert knowledge of the environment.
Instead, animals rely on previously learned low-level environmental features that generate activity in the different layers of neurons within the sensory cortex. As the information propagates through layers of the cortex, the concepts that the neurons are sensitive to become more and more abstract. Decisions based on these hierarchical features allow the animal to make the friend-foe classification. This decision can be made without prior knowledge of the exact input properties and in the presence of noise.
Similarly, an automatic modulation classification (AMC) of radio signals can be modeled on animals moving within a natural environment. AMC is the intermediate step between signal detection and demodulation and is a critical task of an intelligent radio receiver, with many civilian and military applications.
AMC allows for the successful transmission of approved signals and the blocking or jamming of unapproved signals in a crowded transmission environment. With no knowledge of the transmitted data and many unknown parameters at the receiver, such as the signal power, frequency, phase offsets, and timing information, blind identification of the modulation is a difficult task and classification is generally not feasible as most existing methods require prior information regarding the modulation mechanism.
Leveraging knowledge of animal sensory perception, Navy researchers have developed a system forautomatically classifying input signals having unknown feature types.
In their approach, features associated with different observed signals having different class types are learned by a neural network. The neural network then recognizes features of the input signals having unknown class types that at least partially match some of the characteristics associated with different observed signals having respective different known class types.
The neural network determines probabilities that each of the input signals has each of the known modulation types based on strengths of matches between the recognized features of the input signals and the features associated with different observed signals. The neural network classifies each of the input signals as having one of the respective different known class types based on a highest determined probability.
The unsupervised training method of this neural network allows for much greater flexibility in terms of incorporating unusual characteristics of environmental noise, accommodating signals for which no detailed model may be available, and in adapting to changes in environmental or signal characteristics over time through the use of online learning techniques.
- Accurate signal classification system applicable in noisy environments can be considered to be biologically inspired in that it learns modulation types much in the same way an animal might learn patterns without advance knowledge
- System differs significantly from conventional approaches to AMC in that it is model-free and expert-free on the input side, and requires only example signals to detect future signals modulated in the same manner
- US patent 10,003,483 available for license
- Potential for collaboration with Navy researchers