System and method for modulation and signal classification

A method for generating low-complexity radio digital signal processing (DSP) and artificial intelligence (AI)-processing via Spectrum Correlation Function (SCF) using Approximate Discrete Fourier Transform (ADFT) with linear computational complexity O(N) for modulation and signal classification, the method comprising:

• a) receiving a digital input signal comprising in-phase (I) and quadrature (Q) components from a receiver array operatively coupled to a phased antenna array;
• b) normalizing the I and Q components, thereby generating a normalized signal with zero mean and unit power;
• c) computing SCF by applying a sparse, factorized ADFT to cyclic autocorrelation of the normalized signal, thereby generating a SCF feature matrix;
• d) inputting the SCF feature matrix for training a neural network model;
• and
• e) outputting a modulation classification label corresponding to type of modulation present in the digital input signal.

A system for generating low-complexity radio digital signal processing (DSP) and artificial intelligence (AI)-processing via Spectrum Correlation Function (SCF) using Approximate Discrete Fourier Transform (ADFT) with linear computational complexity O(N) for modulation and signal classification, the system comprising:

• a phased antenna array;
• a receiver array in operable communication with the phased antenna array;
• a processor;
• and
• a machine-readable medium in operable communication with the processor and having instructions stored thereon that, when executed by the processor, perform the following steps:
• a) receiving a digital input signal comprising in-phase (I) and quadrature (Q) components from the receiver array;
• b) normalizing the I and Q components, thereby generating a normalized signal with zero mean and unit power;
• c) computing SCF by applying a sparse, factorized ADFT to cyclic autocorrelation of the normalized signal, thereby generating a SCF feature matrix;
• d) inputting the SCF feature matrix for training a neural network model;
• and
• e) outputting a modulation classification label corresponding to type of modulation present in the digital input signal,
• the sparse, factorized ADFT being based on sparse factorized transformation configured to reduce computational complexity to approximately O(N),
• the sparse, factorized ADFT comprising a sequence of sparse matrix operations, including diagonal and permutation matrices,
• step b) comprising the sub-steps of:
• b1) removing a zero-frequency component by subtracting a mean of each of the I and Q components over a defined window;
• b2) calculating a signal magnitude for each complex sample formed by the I and Q components at a single point in time;
• and
• b3) scaling each complex sample to generate the normalized signal with zero mean and unit power,
• step c) comprising the sub-steps of:
• c1) segmenting the digital input signal into a plurality of time-domain analysis windows, each time-domain analysis window being overlapping or non-overlapping;
• c2) applying a Hamming window function to each time-domain analysis window to mitigate spectral leakage, thereby generating windowed segments;
• c3) applying a first ADFT to each windowed segment to obtain frequency-domain segments;
• c4) performing frequency down-conversion by mixing each frequency-domain segment with a complex local oscillator signal to shift the frequency-domain segments to baseband, thereby generating pairs of baseband-shifted frequency-domain segments;
• c5) performing conjugate multiplication between the pairs of baseband-shifted frequency-domain segments to extract spectral correlation values, thereby generating conjugate multiplied segments;
• c6) applying a second ADFT to the spectral correlation values, thereby generating SCF values over a range of cycle frequencies and spectral frequencies through SCF computation;
• and
• c7) generating a SCF feature matrix by mapping the SCF values across dimensions of the cycle frequencies and the spectral frequencies, thereby characterizing cyclostationary properties of the digital input signal,
• the first ADFT being configured to transform each windowed segment from the time domain to the frequency domain,
• the second ADFT being configured to operate over correlation delays corresponding to the conjugate multiplied segments, thereby extracting cyclostationary features,
• the SCF values being generated by associating accumulated spectral correlation values with corresponding cycle frequencies and spectral frequencies,
• step c) being performed in real time on streaming input data using high-bandwidth hardware designed to reduce power consumption, and
• step c) being executed on an edge computing device comprising an embedded processor, a software-defined radio (SDR), and a mobile radio terminal.