A Signal Cognition Method Based on Multi-View Evidence and Hierarchical Probabilistic Outputs Without Prior Information

Abstract

Signal cognition without prior information is often required in electromagnetic situational awareness, cognitive radio, and signals intelligence (SIGINT). This paper proposes a blind signal cognition method based on multi-view evidence and hierarchical probabilistic outputs. A fixed-length in-phase/quadrature (I/Q) data segment is used as one cognition window. Multidimensional physical features are extracted from the time, frequency, and time-frequency domains and at different time scales. A staged multi-layer perceptron (MLP) then produces support probabilities for the reference background, signal process, basic signal form, signal family, and fine-grained signal type. Complete upstream probability vectors are also passed to later stages, allowing the model to retain broader cognition results when the evidence is insufficient for a fine-grained distinction. Simulations involving 73 signal types show that, at signal-to-noise ratios (SNRs) of 0 dB or higher, the average recognition rates for basic signal form, signal family, and fine-grained type reach 92.40%, 87.63%, and 87.15%, respectively. The results indicate that combining multi-view evidence with a hierarchical soft cascade provides a feasible approach to multi-class signal cognition when no prior information about the input signal is available.