Neural network approach for identifying a radar signal in the presence of noise

A system for training and improving a radar signal detection system with an unlabeled training sequence of an unknown radar signal and noise to reconstruct a radar signal if present in radiation with noise, the radar signal detection system including an antenna for receiving radiation in an analog format and an analog-to-digital converter for digitally sampling the received analog radiation, the unlabeled training sequence including a digitally sampled sequence of an unknown radar radiation and noise, the digitally sampled sequence including a plurality of sampled values describing a radar signal and the noise, the plurality of sampled values divided into a plurality of windows, the system for training and improving the radar signal detection system comprising:

• an encoding neural network, comprising:
• a weighted encoding long short term memory (LSTM) compressing each window of the plurality of windows over the sampled sequence into a respective interim context vector;
• and a weighted first linear fully connected layer receiving the respective interim context vector and generating a respective context vector having a fixed dimension less than an incoming dimension of each window of the plurality of windows over the sampled sequence;
• a decoding neural network, comprising:
• a hyperbolic tangent module for receiving the respective context vector and outputting a hyperbolic tangent of the respective context vector;
• a weighted decoding LSTM receiving and decompressing the hyperbolic tangent of the respective context vector for each window of the plurality of windows, beginning with a zero estimate for a last reconstructed value in the window, into an interim sequence which is time reversed from the window over the sampled sequence;
• and a gating neural network, comprising:
• a weighted multi-layer perceptron receiving the interim sequence and producing a confidence sequence from a sigmoidal output;
• a binary mask threshold masking the confidence sequence;
• a weighted second linear fully connected layer generating ungated values of an ungated sequence by applying an activation to respective interim vectors in the interim sequence;
• a multiplier receiving the ungated sequence and the masked confidence sequence and generating a respective reconstructed value in the reconstructed sequence for each sampled value in the sampled sequence as a product of a respective confidence value for the sampled value in the confidence sequence and a respective ungated value for the sampled valued in the ungated sequence;
• and wherein the reconstructed sequence feeds back to a next estimate for a reconstructed value in the decoding neural network;
• and a trainer in communication with the encoding neural network, the decoding neural network and the gating neural network, the trainer comprising:
• an error loss between the sampled sequence and the reconstructed sequence;
• and a gradient descent backpropagation receiving the calculated error loss and adjusting weighting of:
• the weighted encoding LSTM;
• the weighted first linear fully connected layer;
• the weighted decoding LSTM;
• the weighted multi-layer perceptron;
• and the weighted second linear fully connected layer;
• and the trainer optimizing the calculated error loss via the adjusted weighting performed by the gradient descent backpropagation module;
• the system outputting a trained encoding neural network, a trained decoding neural network and a trained gating neural network for use in a radar signal detection system.

A method for training a self-supervised machine-learning system to improve a radar signal detection system used to sample radiation, detect an unknown radar signal and reconstruct the unknown radar signal even in the presence of noise, the method comprising:

• providing the radar signal detection system, comprising:
• a radiation receiver, comprising:
• an antenna for receiving the radiation in an analog format;
• an analog-to-digital converter for converting the received analog radiation into a digitally formatted plurality of sampled values of a sampled sequence, the sampled sequence describing the unknown radar signal and the noise;
• and the receiver further segmenting the plurality of sampled values into a plurality of windows;
• providing the self-supervised machine-learning system in communication with the radar signal detection system, the self-supervised machine-learning system, comprising:
• an encoding neural network, comprising:
• a weighted encoding long short term memory (LSTM) compressing each window of the plurality of windows over the sampled sequence into a respective interim context vector;
• and a weighted first linear fully connected layer receiving the respective interim context vector and generating a respective context vector having a fixed dimension less than an incoming dimension of each window of the plurality of windows over the sampled sequence;
• a decoding neural network, comprising:
• a hyperbolic tangent module for receiving the respective context vector and outputting a hyperbolic tangent of the respective context vector;
• a weighted decoding LSTM receiving and decompressing the hyperbolic tangent of the respective context vector for each window of the plurality of windows, beginning with a zero estimate for a last reconstructed value in the window, into an interim sequence which is time reversed from the window over the sampled sequence;
• and a gating neural network, comprising:
• a weighted multi-layer perceptron receiving the interim sequence and producing a confidence sequence from a sigmoidal output;
• a binary mask threshold masking the confidence sequence;
• a weighted second linear fully connected layer generating ungated values of an ungated sequence by applying an activation to respective interim vectors in the interim sequence;
• a multiplier receiving the ungated sequence and the masked confidence sequence and generating a respective reconstructed value in the reconstructed sequence for each sampled value in the sampled sequence as a product of a respective confidence value for the sampled value in the confidence sequence and a respective ungated value for the sampled valued in the ungated sequence;
• and wherein the reconstructed sequence feeds back to a next estimate for a reconstructed value in the decoding neural network;
• and a trainer in communication with the encoding neural network, the decoding neural network and the gating neural network, the trainer comprising:
• an error loss between the sampled sequence and the reconstructed sequence;
• and a gradient descent backpropagation receiving the calculated error loss and adjusting weighting of:
• the weighted encoding LSTM;
• the weighted first linear fully connected layer;
• the weighted decoding LSTM;
• the weighted multi-layer perceptron;
• and the weighted second linear fully connected layer;
• and the trainer optimizing the calculated error loss via the adjusted weighting performed by the gradient descent backpropagation module;
• the self-supervised machine-learning system outputting a trained encoding neural network, a trained decoding neural network and a trained gating neural network for use in the radar signal detection system.