Small sample photovoltaic hot spot identification method based on deep stack type hybrid self-encoding network

Abstract

The invention discloses a small sample photovoltaic hot spot identification method based on a deep stack type hybrid self-encoding network. The method comprises the following steps: carrying out image preprocessing on an acquired photovoltaic infrared image to obtain a small sample hot spot image data set; firstly, pre-training DAE by using a small sample hot spot image data set without a label, and when the reconstruction error of input and output is minimum, indicating that training is completed; taking image features extracted by a DAE hidden layer as input of SAE, making the pre-trained DAE, SAE and AE cascaded, and adding a Softmax classifier to form a deep stack type hybrid self-encoding network model; and inputting the labeled small sample hot spot image data set into the deep stack type hybrid self-encoding network, carrying out fine adjustment on the model by using a back propagation algorithm, and carrying out prediction through a classifier to obtain a hot spot identification result. The method has strong feature extraction and expression capabilities, and can overcome an over-fitting phenomenon caused by insufficient sample size so as to improve the hot spot recognition and positioning accuracy of the model.

Description

technical field [0001] The invention relates to the technical field of small-sample photovoltaic hotspot identification, in particular to a small-sample photovoltaic hotspot identification method based on a deep stacked hybrid self-encoding network. Background technique [0002] There are mainly the following problems in the field of hot spot recognition of photovoltaic infrared images: (1) the lack of original samples makes it impossible to directly use the traditional data-driven classification neural network; (2) the number of cells with hot spots is relatively small compared with other states, Therefore, there is an unbalanced sample set, which leads to the prediction result of the model focusing on the category with a large number of samples, or even giving up the prediction of a small number of samples. (3) Due to the small sample size, over-fitting phenomenon will occur in the training of the network model, resulting in a low accuracy rate of hot spot recognition. ...

AI Technical Summary

Problems solved by technology

[0002] There are mainly the following problems in the field of hot spot recognition of photovoltaic infrared images: (1) the lack of original samples makes it impossible to directly use the traditional data-driven classification neural network; (2) the number of cells with hot spots is relatively small compared with other states, Therefore, there is an unbalanced sample set, which leads to the prediction result of the model focusing on the category with a large number of samples, or even giving up the prediction of a small number of samples.

(3) Due to the small sample size, over-fitting phenomenon will occur in the training of the network model, resulting in low accuracy of hot spot recognition

[0003] For the problem of insufficient sample size, data enhancement methods can be used, such as cropping, rotation, GAN generation, etc. The sample set has been expanded, but some images are distorted, resulting in a low accuracy of hot spot recognition; for the phenomenon of unbalanced sample sets There are two methods of oversampling and undersampling. Oversampling balances the data set by increasing the number of positive samples, but there is a risk of overfitting

Undersampling extracts the same number of samples from the infrared image of the normal working condition battery as the fault infrared heat map, and balances the data set by reducing the number of negative samples. This method has the problem that the sample data cannot be fully utilized.

For the problem of low model accuracy, the deep transfer learning method is adopted, which has high recognition accuracy, low false detection rate, and strong generalization ability, but the training time is long

Convolutional neural network model based on GAN has slow convergence speed, long training time and relatively low accuracy

Images