# KrakenSDR Signal Processor # # Copyright (C) 2018-2021 Carl Laufer, Tamás Pető # # This program is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # any later version. # # This program is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with this program. If not, see . # # # - coding: utf-8 -*- # Import built-in modules import sys import os import time import logging import threading import queue import math # Import optimization modules import numba as nb from numba import jit, njit from functools import lru_cache # Math support import numpy as np import numpy.linalg as lin #from numba import jit import pyfftw # Signal processing support import scipy from scipy import fft from scipy import signal from scipy.signal import correlate from scipy.signal import convolve from pyapril import channelPreparation as cp from pyapril import clutterCancellation as cc from pyapril import detector as det c_dtype = np.complex64 #import socket # UDP is useless to us because it cannot work over mobile internet # Init UDP #server = socket.socket(socket.AF_INET, socket.SOCK_DGRAM, socket.IPPROTO_UDP) #server.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEPORT, 1) # Enable broadcasting mode #server.setsockopt(socket.SOL_SOCKET, socket.SO_BROADCAST, 1) # Set a timeout so the socket does not block # indefinitely when trying to receive data. #server.settimeout(0.2) class SignalProcessor(threading.Thread): def __init__(self, data_que, module_receiver, logging_level=10): """ Parameters: ----------- :param: data_que: Que to communicate with the UI (web iface/Qt GUI) :param: module_receiver: Kraken SDR DoA DSP receiver modules """ super(SignalProcessor, self).__init__() self.logger = logging.getLogger(__name__) self.logger.setLevel(logging_level) root_path = os.path.dirname(os.path.dirname(os.path.realpath(__file__))) #doa_res_file_path = os.path.join(os.path.join(root_path,"_android_web","DOA_value.html")) #self.DOA_res_fd = open(doa_res_file_path,"w+") self.module_receiver = module_receiver self.data_que = data_que self.en_spectrum = False self.en_record = False self.en_DOA_estimation = True self.first_frame = 1 # Used to configure local variables from the header fields self.processed_signal = np.empty(0) # Squelch feature self.data_ready = False self.en_squelch = False self.squelch_threshold = 0.1 self.squelch_trigger_channel = 0 self.raw_signal_amplitude = np.empty(0) self.filt_signal = np.empty(0) self.squelch_mask = np.empty(0) # DOA processing options self.en_DOA_Bartlett = False self.en_DOA_Capon = False self.en_DOA_MEM = False self.en_DOA_MUSIC = False self.en_DOA_FB_avg = False self.DOA_offset = 0 self.DOA_inter_elem_space = 0.5 self.DOA_ant_alignment = "ULA" self.DOA_theta = np.linspace(0,359,360) # PR processing options self.PR_clutter_cancellation = "Wiener MRE" self.max_bistatic_range = 128 self.max_doppler = 256 self.en_PR = True # Processing parameters self.spectrum_window_size = 2048 #1024 self.spectrum_window = "hann" self.run_processing = False self.is_running = False self.channel_number = 4 # Update from header # Result vectors self.DOA_Bartlett_res = np.ones(181) self.DOA_Capon_res = np.ones(181) self.DOA_MEM_res = np.ones(181) self.DOA_MUSIC_res = np.ones(181) self.DOA_theta = np.arange(0,181,1) self.max_index = 0 self.max_frequency = 0 self.fft_signal_width = 0 self.DOA_theta = np.linspace(0,359,360) self.spectrum = None #np.ones((self.channel_number+2,N), dtype=np.float32) self.spectrum_upd_counter = 0 def run(self): """ Main processing thread """ pyfftw.config.NUM_THREADS = 4 scipy.fft.set_backend(pyfftw.interfaces.scipy_fft) pyfftw.interfaces.cache.enable() while True: self.is_running = False time.sleep(1) while self.run_processing: self.is_running = True que_data_packet = [] start_time = time.time() #-----> ACQUIRE NEW DATA FRAME <----- self.module_receiver.get_iq_online() # Check frame type for processing en_proc = (self.module_receiver.iq_header.frame_type == self.module_receiver.iq_header.FRAME_TYPE_DATA)# or \ #(self.module_receiver.iq_header.frame_type == self.module_receiver.iq_header.FRAME_TYPE_CAL)# For debug purposes """ You can enable here to process other frame types (such as call type frames) """ que_data_packet.append(['iq_header',self.module_receiver.iq_header]) self.logger.debug("IQ header has been put into the data que entity") # Configure processing parameteres based on the settings of the DAQ chain if self.first_frame: self.channel_number = self.module_receiver.iq_header.active_ant_chs self.spectrum_upd_counter = 0 self.spectrum = np.ones((self.channel_number+1, self.spectrum_window_size), dtype=np.float32) self.first_frame = 0 decimation_factor = 1 self.data_ready = False if en_proc: self.processed_signal = self.module_receiver.iq_samples self.data_ready = True first_decimation_factor = 1 #480 # TESTING: DSP side main decimation - significantly slower than NE10 but it works ok-ish #decimated_signal = signal.decimate(self.processed_signal, first_decimation_factor, n = 584, ftype='fir', zero_phase=True) #first_decimation_factor * 2, ftype='fir') #self.processed_signal = decimated_signal #.copy() #spectrum_signal = decimated_signal.copy() max_amplitude = -100 #max_ch = np.argmax(np.max(self.spectrum[1:self.module_receiver.iq_header.active_ant_chs+1,:], axis=1)) # Find the channel that had the max amplitude max_amplitude = 0 #np.max(self.spectrum[1+max_ch, :]) #Max amplitude out of all 5 channels #max_spectrum = self.spectrum[1+max_ch, :] #Send max ch to channel centering que_data_packet.append(['max_amplitude',max_amplitude]) #-----> SQUELCH PROCESSING <----- if self.en_squelch: self.data_ready = False self.processed_signal, decimation_factor, self.fft_signal_width, self.max_index = \ center_max_signal(self.processed_signal, self.spectrum[0,:], max_spectrum, self.module_receiver.daq_squelch_th_dB, self.module_receiver.iq_header.sampling_freq) #decimated_signal = [] #if(decimation_factor > 1): # decimated_signal = signal.decimate(self.processed_signal, decimation_factor, n = decimation_factor * 2, ftype='fir') # self.processed_signal = decimated_signal #.copy() #Only update if we're above the threshold if max_amplitude > self.module_receiver.daq_squelch_th_dB: self.data_ready = True #-----> SPECTRUM PROCESSING <----- if self.en_spectrum and self.data_ready: spectrum_samples = self.module_receiver.iq_samples #spectrum_signal #self.processed_signal #self.module_receiver.iq_samples #self.processed_signal N = self.spectrum_window_size N_perseg = 0 N_perseg = min(N, len(self.processed_signal[0,:])//25) N_perseg = N_perseg // 1 for m in range(self.channel_number): # Get power spectrum f, Pxx_den = signal.welch(self.processed_signal[m, :], self.module_receiver.iq_header.sampling_freq//first_decimation_factor, nperseg=N_perseg, nfft=N, noverlap=0, #int(N_perseg*0.25), detrend=False, return_onesided=False, window= ('tukey', 0.25), #tukey window gives better time resolution for squelching #self.spectrum_window, #('tukey', 0.25), #self.spectrum_window, #window=self.spectrum_window, scaling="spectrum") self.spectrum[1+m, :] = np.fft.fftshift(10*np.log10(Pxx_den)) #self.spectrum[1:self.module_receiver.iq_header.active_ant_chs+1,:] = np.fft.fftshift(10*np.log10(Pxx_den)) self.spectrum[0,:] = np.fft.fftshift(f) # Create signal window for plot # signal_window = np.ones(len(self.spectrum[1,:])) * -100 # signal_window[max(self.max_index - self.fft_signal_width//2, 0) : min(self.max_index + self.fft_signal_width//2, len(self.spectrum[1,:]))] = max(self.spectrum[1,:]) #signal_window = np.ones(len(max_spectrum)) * -100 #signal_window[max(self.max_index - self.fft_signal_width//2, 0) : min(self.max_index + self.fft_signal_width//2, len(max_spectrum))] = max(max_spectrum) #self.spectrum[self.channel_number+1, :] = signal_window #np.ones(len(spectrum[1,:])) * self.module_receiver.daq_squelch_th_dB # Plot threshold line que_data_packet.append(['spectrum', self.spectrum]) #-----> Passive Radar <----- conf_val = 0 theta_0 = 0 if self.en_PR and self.data_ready and self.channel_number > 1: ref_ch = self.module_receiver.iq_samples[0,:] surv_ch = self.module_receiver.iq_samples[1,:] td_filter_dimension = self.max_bistatic_range start = time.time() if self.PR_clutter_cancellation == "Wiener MRE": surv_ch, w = Wiener_SMI_MRE(ref_ch, surv_ch, td_filter_dimension) #surv_ch, w = cc.Wiener_SMI_MRE(ref_ch, surv_ch, td_filter_dimension) surv_ch = det.windowing(surv_ch, "Hamming") #surv_ch * signal.tukey(surv_ch.size, alpha=0.25) #det.windowing(surv_ch, "hamming") max_Doppler = self.max_doppler #256 max_range = self.max_bistatic_range #RD_matrix = det.cc_detector_ons(ref_ch, surv_ch, self.module_receiver.iq_header.sampling_freq, max_Doppler, max_range, verbose=0, Qt_obj=None) RD_matrix = cc_detector_ons(ref_ch, surv_ch, self.module_receiver.iq_header.sampling_freq, max_Doppler, max_range) end = time.time() #print("Time: " + str((end-start) * 1000)) que_data_packet.append(['RD_matrix', RD_matrix]) # Record IQ samples if self.en_record: # TODO: Implement IQ frame recording self.logger.error("Saving IQ samples to npy is obsolete, IQ Frame saving is currently not implemented") stop_time = time.time() que_data_packet.append(['update_rate', stop_time-start_time]) que_data_packet.append(['latency', int(stop_time*10**3)-self.module_receiver.iq_header.time_stamp]) # If the que is full, and data is ready (from squelching), clear the buffer immediately so that useful data has the priority if self.data_que.full() and self.data_ready: try: #self.logger.info("BUFFER WAS NOT EMPTY, EMPTYING NOW") self.data_que.get(False) #empty que if not taken yet so fresh data is put in except queue.Empty: #self.logger.info("DIDNT EMPTY") pass # Put data into buffer, but if there is no data because its a cal/trig wait frame etc, then only write if the buffer is empty # Otherwise just discard the data so that we don't overwrite good DATA frames. try: self.data_que.put(que_data_packet, False) # Must be non-blocking so DOA can update when dash browser window is closed except: # Discard data, UI couldn't consume fast enough pass """ start = time.time() end = time.time() thetime = ((end - start) * 1000) print ("Time elapsed: ", thetime) """ @jit(fastmath=True) def Wiener_SMI_MRE(ref_ch, surv_ch, K): """ Description: ------------ Performs Wiener filtering with applying the Minimum Redundance Estimation (MRE) technique. When using MRE, the autocorrelation matrix is not fully estimated, but only the first column. With this modification the required calculations can be reduced from KxK to K element. Parameters: ----------- :param K : Filter tap number :param ref_ch : Reference signal array :param surv_ch: Surveillance signal array :type K : int :type ref_ch : 1 x N complex numpy array :type surv_ch: 1 x N complex numpy array Return values: -------------- :return filt: Filtered surveillance channel :rtype filt: 1 x N complex numpy array :return None: Input parameters are not consistent """ N = ref_ch.shape[0] # Number of time samples R, r = pruned_correlation(ref_ch, surv_ch, K, N) R_mult = R_eye_memoize(K) w = fast_w(R, r, K, R_mult) #return surv_ch - np.convolve(ref_ch, w)[0:N], w # subtract the zero doppler clutter return surv_ch - signal.oaconvolve(ref_ch, w)[0:N], w # subtract the zero doppler clutter #oaconvolve saves us about 100-200 ms @njit(fastmath=True, parallel=True, cache=True) def fast_w(R, r, K, R_mult): # Complete the R matrix based on its Hermitian and Toeplitz property for k in range(1, K): R[:, k] = shift(R[:, 0], k) #R[:, K] = shift(R[:,0], K) R += np.transpose(np.conjugate(R)) R *= R_mult #(np.ones(K) - np.eye(K) * 0.5) #w = np.dot(lin.inv(R), r) # weight vector w = lin.inv(R) @ r #np.dot(lin.inv(R), r) # weight vector #matmul (@) may be slightly faster that np.dot for 1D, 2D arrays. # inverse and dot product run time : 1.1s for 2048*2048 matrix return w #Memoize ~50ms speedup? @lru_cache(maxsize=2) def R_eye_memoize(K): return (np.ones(K) - np.eye(K) * 0.5) #Modified pruned correlation, returns R and r directly and saves one FFT @jit(fastmath=True, cache=True) def pruned_correlation(ref_ch, surv_ch, clen, N): """ Description: ----------- Calculates the part of the correlation function of arrays with same size The total length of the cross-correlation function is 2*N-1, but this function calculates the values of the cross-correlation between [N-1 : N+clen-1] Parameters: ----------- :param x : input array :param y : input array :param clen: correlation length :type x: 1 x N complex numpy array :type y: 1 x N complex numpy array :type clen: int Return values: -------------- :return corr : part of the cross-correlation function :rtype corr : 1 x clen complex numpy array :return None : inconsistent array size """ R = np.zeros((clen, clen), dtype=c_dtype) # Autocorrelation mtx. # --calculation-- # set up input matrices pad zeros if not multiply of the correlation length cols = clen - 1 #(clen = Filter drowsimension) rows = np.int32(N / (cols)) + 1 zeropads = cols * rows - N x = np.pad(ref_ch, (0, zeropads)) # shaping inputs into matrices xp = np.reshape(x, (rows, cols)) # padding matrices for FFT ypp = np.vstack([xp[1:, :], np.zeros(cols, dtype=c_dtype)]) #vstack appears to be faster than pad yp = np.concatenate([xp, ypp], axis=1) #print("pruned corr xp shape: " + str(xp.shape)) #print("pruned corr yp shape: " + str(yp.shape)) # execute FFT on the matrices xpw = fft.fft(xp, n = 2*cols, axis=1, workers=4, overwrite_x=True) bpw = fft.fft(yp, axis=1, workers=4, overwrite_x=True) # magic formula which describes the unified equation of the universe # corr_batches = np.fliplr(fft.fftshift(fft.ifft(corr_mult(xpw, bpw), axis=1, workers=4, overwrite_x=True)).conj()[:, 0:clen]) corr_batches = fft.fftshift(fft.ifft(corr_mult(xpw, bpw), axis=1, workers=4, overwrite_x=True)).conj()[:, 0:clen] # sum each value in a column of the batched correlation matrix R[:,0] = np.fliplr([np.sum(corr_batches, axis=0)])[0] #calc r y = np.pad(surv_ch, (0, zeropads)) yp = np.reshape(y, (rows, cols)) ypp = np.vstack([yp[1:, :], np.zeros(cols, dtype=c_dtype)]) #vstack appears to be faster than pad yp = np.concatenate([yp, ypp], axis=1) bpw = fft.fft(yp, axis=1, workers=4, overwrite_x=True) #corr_batches = np.fliplr(fft.fftshift(fft.ifft(corr_mult(xpw, bpw), axis=1, workers=4, overwrite_x=True)).conj()[:, 0:clen]) corr_batches = fft.fftshift(fft.ifft(corr_mult(xpw, bpw), axis=1, workers=4, overwrite_x=True)).conj()[:, 0:clen] #r = np.sum(corr_batches, axis=0) r = np.fliplr([np.sum(corr_batches, axis=0)])[0] return R, r @njit(fastmath=True, cache=True) def shift(x, i): """ Description: ----------- Similar to np.roll function, but not circularly shift values Example: x = |x0|x1|...|xN-1| y = shift(x,2) x --> y: |0|0|x0|x1|...|xN-3| Parameters: ----------- :param:x : input array on which the roll will be performed :param i : delay value [sample] :type i :int :type x: N x 1 complex numpy array Return values: -------------- :return shifted : shifted version of x :rtype shifted: N x 1 complex numpy array """ N = x.shape[0] if np.abs(i) >= N: return np.zeros(N, dtype=c_dtype) if i == 0: return x shifted = np.roll(x, i) if i < 0: shifted[np.mod(N + i, N):] = np.zeros(np.abs(i), dtype=c_dtype) if i > 0: shifted[0:i] = np.zeros(np.abs(i), dtype=c_dtype) return shifted @njit(fastmath=True, parallel=True, cache=True) def resize_and_align(no_sub_tasks, ref_ch, surv_ch, fs, fD_max, r_max): surv_ch_align = np.reshape(surv_ch,(no_sub_tasks, r_max)) # shaping surveillance signal array into a matrix pad_zeros = np.expand_dims(np.zeros(r_max, dtype=c_dtype), axis=0) surv_ch_align = np.vstack((surv_ch_align, pad_zeros)) # padding one row of zeros into the surv matrix surv_ch_align = np.concatenate((surv_ch_align[0 : no_sub_tasks,:], surv_ch_align[1 : no_sub_tasks +1, :]), axis = 1) ref_ch_align = np.reshape(ref_ch, (no_sub_tasks, r_max)) # shaping reference signal array into a matrix pad_zeros = np.zeros((no_sub_tasks, r_max),dtype = c_dtype) ref_ch_align = np.concatenate((ref_ch_align, pad_zeros),axis = 1) # shaping return ref_ch_align, surv_ch_align @njit(fastmath=True, cache=True) def corr_mult(surv_fft, ref_fft): return np.multiply(surv_fft, ref_fft.conj()) @jit(fastmath=True, cache=True) def cc_detector_ons(ref_ch, surv_ch, fs, fD_max, r_max): """ Parameters: ----------- :param N: Range resolution - N must be a divisor of the input length :param F: Doppler resolution, F has a theoretical limit. If you break the limit, the output may repeat itself and get wrong results. F should be less than length/N otherwise use other method! Return values: -------------- :return None: Improper input parameters """ N = ref_ch.size # --> Set processing parameters fD_step = fs / (2 * N) # Doppler frequency step size (with zero padding) Doppler_freqs_size = int(fD_max / fD_step) no_sub_tasks = N // r_max # Allocate range-Doppler maxtrix mx = np.zeros((2*Doppler_freqs_size+1, r_max),dtype = c_dtype) #memoize_zeros((2*Doppler_freqs_size+1, r_max), c_dtype) #np.zeros((2*Doppler_freqs_size+1, r_max),dtype = nb.c8) ref_ch_align, surv_ch_align = resize_and_align(no_sub_tasks, ref_ch, surv_ch, fs, fD_max, r_max) #print("ref_ch_align shape: " + str(ref_ch_align.shape)) #print("surv_ch_align shape: " + str(surv_ch_align.shape)) # row wise fft on both channels ref_fft = fft.fft(ref_ch_align, axis = 1, overwrite_x=True, workers=4) #pyfftw.interfaces.numpy_fft.fft(ref_ch_align_a, axis = 1, overwrite_input=True, threads=4) #fft.fft(ref_ch_align_a, axis = 1, overwrite_x=True, workers=4) surv_fft = fft.fft(surv_ch_align, axis = 1, overwrite_x=True, workers=4) #pyfftw.interfaces.numpy_fft.fft(surv_ch_align_a, axis = 1, overwrite_input=True, threads=4) #fft.fft(surv_ch_align_a, axis = 1, overwrite_x=True, workers=4) corr = corr_mult(surv_fft, ref_fft) #np.multiply(surv_fft, ref_fft.conj()) corr = fft.ifft(corr,axis = 1, workers=4, overwrite_x=True) corr_a = pyfftw.empty_aligned(np.shape(corr), dtype=c_dtype) corr_a[:] = corr #.copy() # This is the most computationally intensive part ~120ms, overwrite_x=True gives a big speedup, not sure if it changes the result though... corr = fft.fft(corr_a, n=2* no_sub_tasks, axis = 0, workers=4, overwrite_x=True) # Setting the output size with "n=.." is faster than doing a concat first. # crop and fft shift mx[ 0 : Doppler_freqs_size, 0 : r_max] = corr[2*no_sub_tasks - Doppler_freqs_size : 2*no_sub_tasks, 0 : r_max] mx[Doppler_freqs_size : 2 * Doppler_freqs_size+1, 0 : r_max] = corr[ 0 : Doppler_freqs_size+1 , 0 : r_max] return mx