Positive and Negative Recognition of Deep Neural Network Knowledge Distillation Based on Real-Time PCR Time Series Data
Objective To propose a new technique for positive and negative recognition based on deep neural network knowledge distillation based on real-time polymerase chain reaction(PCR)time series data.Methods Neural networks to accurately classify time series were used,which could increase robustness and reduce the influence of outliers on model results.Deep neural networks were compressed by knowledge distillation,the demand for computational power of neural networks was reduced to adapt to the operating environment of low computational resources.Results 302260 pieces of data from AutoMolec 3000 automatic nucleic acid purification and real-time fluorescent PCR analysis system were collected,the data set was divided into training set and test set according to 80%and 20%ratio,and the fully convolutional networks(FCN)and long short-term memory(LSTM)after knowledge distillation were verified on the test set.When the accuracy,sensitivity,specificity and F1 were all higher than 0.999,the FCN model could shrink 21.3 times,and the LSTM model parameters could be reduced by 12.8 times.The prediction time of 60452 samples took 5.6900 and 2.2516 s,respectively.Conclusion The model can guarantee the overall accuracy and performance of the model,and meet the requirement of low algorithm requirements for the deployment environment.
real-time fluorescent PCRtime series classificationdeep neural networkknowledge of distillation
国家科技期刊平台
NSTL
万方数据
