Small sample transfer learning method for aircraft electric signal classification and identification

Abstract

The invention relates to a transfer learning method for classifying and identifying small sample electric signals of an aircraft, and the method comprises the following steps: acquiring and transmitting source domain signals (102), and sending the source domain signals into a feature extraction basic module source domain multi-scale residual error convolution module (103) and a source domain maximum pooling layer (104); sending parameters in the basic module to a multi-scale residual convolution module (110) of the target domain and a maximum pooling layer (111) of the target domain, and carrying out the back propagation on a global average pooling layer (112) of the target domain, a random discarding layer (113) of the target domain and a full connection layer (114) of the target domain in the training process to update network parameters (217); and ending the training (221) until the iteration step number update meets the target domain iteration step number judgment module condition (220). By means of the method, the deep learning problem of small sample data in the actual testing process is effectively solved, and the accuracy of aircraft signal classification and recognition is remarkably improved.

Description

technical field [0001] The invention relates to a small-sample transfer learning method for classification and identification of aircraft electrical signals, which can be applied to the classification and identification of small-sample signals. Background technique [0002] In some areas of the aerospace field, few-shot learning is an important frontier direction in the field of deep learning. Since deep learning relies on a large amount of labeled data, the cost of time and money for the actual measurement experiment of the aircraft is extremely high, and a large amount of data cannot be collected. It is guaranteed that deep learning is prone to overfitting and it is difficult to extract enough features. However, small-sample transfer learning can effectively solve the problem of sample data volume and avoid over-fitting, and can use a small number of labeled samples to extract high-quality data features to achieve the purpose of signal classification and recognition. The...

AI Technical Summary

Problems solved by technology

Since deep learning relies on a large amount of labeled data, the cost of time and money for the actual measurement experiment of the aircraft is extremely high, and a large amount of data cannot be collected. Guaranteed, so deep learning is prone to overfitting and it is difficult to extract enough features

Images