Recurrent Neural Detection of Time–Frequency Overlapped . . .
73
During the detection process, it is found that when the bias of each sampling
point is used as the feature to train SVDD classifier, it is hard to determine the
specific location of interference signal. So the feature is windowed according to
the time step and is detected according to the waveform and correlations of the
plurality of feature points to overcome the contingency. The windowing process
is expressed as Fig. 5a. The length of the window (Wlen) we selected is 10, and
the step size (Step) is 1 here. The experimental results are shown in Fig. 5b. The
part marked with purple in the first picture is the interference-containing signal.
The second picture shows the result of test. The part with a value of −1 is the
signal that is judged to be interference, and the part with a value of 1 is the
signal that is judged to be normal. It can be seen that most of the interference
signal sampling points can be detected.
(a)
(b)
Fig. 5. a is the windowing process, and b is the interference detection performance at
12 db (QPSK modulation)
In order to evaluate performance comprehensively, four typical experimental
scenarios are set as follows, which represent various types of interference that may
occur in the actual communication system. In addition, to reduce the influence
of the amplitude, we have done the amplitude normalization on each signal.
• Normal signal is interfered by interference signal with different modulation
mode:
1. The 16QAM signal is normal and the superimposed QPSK signal with the
same sampling rate, symbol rate, and carrier rate is interference (QAMQPSK).
2. The 16QAM signal is normal and the superimposed FM signal is interference
(QAM-FM).
3. The FM signal is normal, and the superimposed DSSS signal is interference
(FM-DSSS).
• Normal signal is interfered by interference signal with different symbol rates:
4. The 16QAM signal is normal, and the superimposed 16QAM signal with only
different symbol rate is interference (QAM-QAM).
73
During the detection process, it is found that when the bias of each sampling
point is used as the feature to train SVDD classifier, it is hard to determine the
specific location of interference signal. So the feature is windowed according to
the time step and is detected according to the waveform and correlations of the
plurality of feature points to overcome the contingency. The windowing process
is expressed as Fig. 5a. The length of the window (Wlen) we selected is 10, and
the step size (Step) is 1 here. The experimental results are shown in Fig. 5b. The
part marked with purple in the first picture is the interference-containing signal.
The second picture shows the result of test. The part with a value of −1 is the
signal that is judged to be interference, and the part with a value of 1 is the
signal that is judged to be normal. It can be seen that most of the interference
signal sampling points can be detected.
(a)
(b)
Fig. 5. a is the windowing process, and b is the interference detection performance at
12 db (QPSK modulation)
In order to evaluate performance comprehensively, four typical experimental
scenarios are set as follows, which represent various types of interference that may
occur in the actual communication system. In addition, to reduce the influence
of the amplitude, we have done the amplitude normalization on each signal.
• Normal signal is interfered by interference signal with different modulation
mode:
1. The 16QAM signal is normal and the superimposed QPSK signal with the
same sampling rate, symbol rate, and carrier rate is interference (QAMQPSK).
2. The 16QAM signal is normal and the superimposed FM signal is interference
(QAM-FM).
3. The FM signal is normal, and the superimposed DSSS signal is interference
(FM-DSSS).
• Normal signal is interfered by interference signal with different symbol rates:
4. The 16QAM signal is normal, and the superimposed 16QAM signal with only
different symbol rate is interference (QAM-QAM).