68
Q. Wu et al.
actual communication environment. When these interfering signals are superimposed on the original signal, the time and the frequency domain features do not
change significantly. We call these interfering signals as time–frequency overlapped signals. Obviously, the time–frequency overlapped signals are hard to
detect by above methods.
The research found that the neural network can better extract the time and
frequency domain feature of signals. The feature can be used as signal fingerprint
to classify signals [4,5]; therefore, many interference detection methods based on
deep learning emerged in recent years. A sequential autoencoder framework is
introduced to distinguish normal and interfering signal in [6]. They calculate
the difference between original signals and reconstructed signals by autoencoder
framework. In [7,8], predicting future signals from known signals by utilizing the
prediction function of recurrent neural network and then taking the difference
between original and predictive signals as a feature to do interference detection.
However, these methods have not good detection performance under low signalto-interference-plus-noise ratio (SINR) conditions, especially for time–frequency
overlapped interfering signals.
In this paper, we present a bidirectional recurrent neural network (BI-RNN)based interference detection structure, which can utilize the correlation between
the front and back sampling points to predict the data of the intermediate position. The structure can do bidirectional training and has the better prediction
performance; hence, the difference between predictive and original signals at the
interference-containing part is larger. And then in prediction process, we adopt
special training labels for noise reduction; therefore, the detection accuracy is
improved under low SINR condition. Finally, in the detection process, we use
the feature correlation classification instead of simply taking a decision threshold, which reduces the contingency of the decision process and can completely
identify the complete interfering signal.
The rest of this paper is organized as follows. In Sect. 2, the interfering signal
and the principle of interference detection are described. Then, the neural network model and the special training method are explained in Sect. 3. Section 4
shows the performance of interference detection and analysis of influencing factors. Finally, Sect. 5 concludes the paper and suggests future work.
2 Problem Formulation
When there is no interference, the received signal can be represented by Eqs. 1,
and 2 and the received signal contains interference.
r(t) = s(t) + n(t),
(1)
r j (t) = s(t) + n(t) + j(t).
(2)
where s(t) is a carrier signal transmitted by transmitter, n(t) represents noise,
and j(t) is a interfering signal(In this paper, we consider j(t) as the time–
frequency overlapped interfering signal.) The aim of interference detection is
to find the exact position of j(t).
Q. Wu et al.
actual communication environment. When these interfering signals are superimposed on the original signal, the time and the frequency domain features do not
change significantly. We call these interfering signals as time–frequency overlapped signals. Obviously, the time–frequency overlapped signals are hard to
detect by above methods.
The research found that the neural network can better extract the time and
frequency domain feature of signals. The feature can be used as signal fingerprint
to classify signals [4,5]; therefore, many interference detection methods based on
deep learning emerged in recent years. A sequential autoencoder framework is
introduced to distinguish normal and interfering signal in [6]. They calculate
the difference between original signals and reconstructed signals by autoencoder
framework. In [7,8], predicting future signals from known signals by utilizing the
prediction function of recurrent neural network and then taking the difference
between original and predictive signals as a feature to do interference detection.
However, these methods have not good detection performance under low signalto-interference-plus-noise ratio (SINR) conditions, especially for time–frequency
overlapped interfering signals.
In this paper, we present a bidirectional recurrent neural network (BI-RNN)based interference detection structure, which can utilize the correlation between
the front and back sampling points to predict the data of the intermediate position. The structure can do bidirectional training and has the better prediction
performance; hence, the difference between predictive and original signals at the
interference-containing part is larger. And then in prediction process, we adopt
special training labels for noise reduction; therefore, the detection accuracy is
improved under low SINR condition. Finally, in the detection process, we use
the feature correlation classification instead of simply taking a decision threshold, which reduces the contingency of the decision process and can completely
identify the complete interfering signal.
The rest of this paper is organized as follows. In Sect. 2, the interfering signal
and the principle of interference detection are described. Then, the neural network model and the special training method are explained in Sect. 3. Section 4
shows the performance of interference detection and analysis of influencing factors. Finally, Sect. 5 concludes the paper and suggests future work.
2 Problem Formulation
When there is no interference, the received signal can be represented by Eqs. 1,
and 2 and the received signal contains interference.
r(t) = s(t) + n(t),
(1)
r j (t) = s(t) + n(t) + j(t).
(2)
where s(t) is a carrier signal transmitted by transmitter, n(t) represents noise,
and j(t) is a interfering signal(In this paper, we consider j(t) as the time–
frequency overlapped interfering signal.) The aim of interference detection is
to find the exact position of j(t).