Recurrent Neural Detection of Time–Frequency Overlapped . . .
71
3.2 Model Training for Noiseless Prediction
During the research, it was found that when the SINR is relatively small, the
network extract the signal features hardly; therefore, the predictive signal and
the original signal are very different. The implementation of speech enhancement uses LSTM by learning the correlation between noisy signals and noiseless
signals in [12]; hence, we are inspired by speech enhancement to propose a novel
training method which can improve the prediction performance under low SINR
conditions. We call this prediction process as noiseless prediction.
In noiseless prediction, the training data is the received signal with noise,
but the input label during training is the corresponding noiseless signal sampling
point. This method makes the network focus only on the correlation of the signal
itself, regardless of the irregularity of noise; therefore, the output is closer to the
noiseless signal. Figure 3 depicts the digital modulation signal predicted by BILSTM, when the SINR = 6 db. The first picture is the original signal without
noise for comparison, and then the second is the predictive signal when use
noise signal as label, and the third is the predictive signal when use noiseless
signal as label. From the figure, we can see that noiseless prediction model has
better predictive performance than the noisy prediction model.
3.3 Prediction Performance
To measure the effectiveness of different neural network, the mean absolute errors
(MAE) and mean absolute percentage errors (MAPE) are computed in Eqs. 5
and 6:
MAE =
1
n
n
i=1
|x i −
x i | ,
(5)
MAPE =
1
n
n
i=1
x i −
x i
x i
.
(6)
where x i is the actual signal sampling point at i th ,
x i is the predictive signal
sampling point, and n is the total number of test signal sampling points.
All the predictive result of different models in this section trained and tested
multiple times to eliminate outliers. We compare the prediction performance of
LSTM, BI-LSTM and the case of noiseless prediction in Table 1.
We can see from Table 1 that the prediction result is getting better with
the increase of SINR, but the influence becomes less obvious when the SINR
reaches a certain level. This proves that Gaussian white noise does interfere with
the learning ability of the network. In addition, the performance of BI-LSTM
model is better than that of LSTM, although not obvious, which indicates that
the sequence correlation after a certain signal sampling point will affect the
prediction result of model. Finally, we can also see that noiseless prediction
obtains the best results, especially under low SINR condition. This proves that
noiseless prediction does have good denoise ability.
71
3.2 Model Training for Noiseless Prediction
During the research, it was found that when the SINR is relatively small, the
network extract the signal features hardly; therefore, the predictive signal and
the original signal are very different. The implementation of speech enhancement uses LSTM by learning the correlation between noisy signals and noiseless
signals in [12]; hence, we are inspired by speech enhancement to propose a novel
training method which can improve the prediction performance under low SINR
conditions. We call this prediction process as noiseless prediction.
In noiseless prediction, the training data is the received signal with noise,
but the input label during training is the corresponding noiseless signal sampling
point. This method makes the network focus only on the correlation of the signal
itself, regardless of the irregularity of noise; therefore, the output is closer to the
noiseless signal. Figure 3 depicts the digital modulation signal predicted by BILSTM, when the SINR = 6 db. The first picture is the original signal without
noise for comparison, and then the second is the predictive signal when use
noise signal as label, and the third is the predictive signal when use noiseless
signal as label. From the figure, we can see that noiseless prediction model has
better predictive performance than the noisy prediction model.
3.3 Prediction Performance
To measure the effectiveness of different neural network, the mean absolute errors
(MAE) and mean absolute percentage errors (MAPE) are computed in Eqs. 5
and 6:
MAE =
1
n
n
i=1
|x i −
x i | ,
(5)
MAPE =
1
n
n
i=1
x i −
x i
x i
.
(6)
where x i is the actual signal sampling point at i th ,
x i is the predictive signal
sampling point, and n is the total number of test signal sampling points.
All the predictive result of different models in this section trained and tested
multiple times to eliminate outliers. We compare the prediction performance of
LSTM, BI-LSTM and the case of noiseless prediction in Table 1.
We can see from Table 1 that the prediction result is getting better with
the increase of SINR, but the influence becomes less obvious when the SINR
reaches a certain level. This proves that Gaussian white noise does interfere with
the learning ability of the network. In addition, the performance of BI-LSTM
model is better than that of LSTM, although not obvious, which indicates that
the sequence correlation after a certain signal sampling point will affect the
prediction result of model. Finally, we can also see that noiseless prediction
obtains the best results, especially under low SINR condition. This proves that
noiseless prediction does have good denoise ability.