A prediction of the frequency of non-periodic signals based on convolutional neural networks
The problem on creation of mathematical support for construction of forecast models based on convolutional neural networks is solved in the work. A method is proposed for using convolutional neural networks to predict the frequency of non-periodic signals. To determine the frequency of the signal, i...
Saved in:
| Date: | 2018 |
|---|---|
| Main Authors: | , , |
| Format: | Article |
| Language: | Ukrainian |
| Published: |
Інститут проблем реєстрації інформації НАН України
2018
|
| Subjects: | |
| Online Access: | http://drsp.ipri.kiev.ua/article/view/158515 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | Data Recording, Storage & Processing |
Institution
Data Recording, Storage & Processing| Summary: | The problem on creation of mathematical support for construction of forecast models based on convolutional neural networks is solved in the work. A method is proposed for using convolutional neural networks to predict the frequency of non-periodic signals. To determine the frequency of the signal, it was divided into parts, after which a fast Fourier transform was used for each part. The spectrograms obtained after the transform are used as inputs to the learning of the neural network. The output value depends on the presence or absence of a frequency that is above the critical value on the predicted interval. The first layer of the neural network uses a three-dimensional convolution, and on the next layers - a onedimensional convolution. Between the convolutional layers, there are subsampling layers used to accelerate learning and prevent retraining. The neural network contains two output neurons which determine the presence of a frequency that exceeds the critical value. The practical task of predicting the frequency of vibration of aircraft engines during their tests is solved. The construction of different neural network models, their training and testing on the data that were collected from vibration sensors during the testing of the aircraft engine has been performed. To increase the amount of data, augmentation is used. To do this, several copies of the signal with changed frequencies are added. The models constructed differ in the amount of data used and in the forecasting time. Comparison of the test results of all the models has been performed. The maximum forecasting time that can be achieved with the proposed method is determined. This time is enough for the pilot to react and change the flight mode or to land the helicopter. |
|---|