A small sample depth learning method based on knowledge transfer of shallow model

Abstract

The invention discloses a small sample depth learning method based on knowledge transfer of a shallow model. The invention firstly preprocesses the data, and then transforms the original signal into different transform domains according to the prior knowledge and expert experience of the related field, and calculates the artificial features. According to artificial features, different shallow models are selected and trained based on a small amount of labeled sample data. According to classification accuracy / prediction error and other indicators, different shallow models with different featurecombinations are screened to form candidate model pool. Then, based on the candidate model pool, the model is selected to predict the unlabeled samples, and the prediction tags are obtained, and a plurality of prediction tags are fused. The prediction tags are combined with a small number of existing tagged samples to construct the training set. Deep neural network structure is designed for the specific task, and the training is based on the above-mentioned mixed training set. The validity of the proposed method is verified by the rotating machinery fault diagnosis data set.

Description

technical field [0001] The invention belongs to the field of machine learning, and in particular relates to a small-sample deep learning method based on shallow model knowledge transfer. Background technique [0002] In recent years, deep neural networks, especially convolutional neural networks, have made remarkable progress in image recognition, speech recognition and other fields. Compared with classical machine learning methods, deep learning methods can unify feature extraction and classifier training in the form of deep neural networks for modeling, and realize end-to-end learning. [0003] Based on deep learning to classify samples, when a large number of samples are used for training, the characteristics of samples can be automatically learned, which greatly improves the classification accuracy; however, the recognition and classification effect in the case of small samples is not optimistic. The multi-layer structure and a large number of network parameters of the ...

AI Technical Summary

Problems solved by technology

However, the above method has the following disadvantages: 1) the sample is not balanced

Unbalanced samples will make it difficult for unsupervised or semi-supervised learning to use and learn the information and characteristics of high-value samples, resulting in weak feature representation

2) The above process is usually a data-driven end-to-end training method, which lacks embedding and guidance of expert knowledge

Images