Design And Implementation Of Online Monitoring Platform For Scientific Instrument Operation Based On Deep Learning

Scientific instruments are the key propellers for the development of scientific research.The total asset value of scientific research instruments and equipment in all types of schools at all levels in the country continues to grow,and the number of instruments increases rapidly.However,most of the monitoring of scientific instruments is still in the stage of manual statistics,only by recording the use of instruments or laboratories to represent the real operation of the instruments.The way is cumbersome and inefficient,and the data is unreliable.It is impossible to meet the requirements of refined and intelligent scientific instrument management.Therefore,a digital transformation of the way scientific instruments are monitored is necessary.Effective monitoring is the key to effective management.Aiming at the current status of scientific instrument management,this thesis designs and implements a scientific instrument operation online monitoring platform based on deep learning,aiming to provide a reliable data basis for the digital transformation of scientific instrument management,and has been applied to Zhejiang Dayi equipment management services Collaboration platform to monitor the operating status of more than 10,000 scientific instruments.The platform collects the operating current data of scientific instruments themselves through non-invasive detectors,and builds a clustered persistent layer for data storage.The business system is modularized based on the micro-service architecture.The display layer uses the web page and We Chat applet as the carrier.The platform provides reliable and stable scientific instrument data services,basically covers the monitoring needs of managers,and focus management efforts on the instruments themselves.The state recognition algorithm of scientific instruments is the key to the monitoring of scientific instruments,and it is applied to the recognition of operating states and abnormal states.In view of the situation that the cost of model training caused by the rapid increase in the number of instruments,this thesis proposes the idea of using a class of scientific instruments for model training.This thesis takes gas chromatograph as an example,carries out time-domain feature extraction and designs a method for constructing scientific instrument operation data sets based on feature selection.In terms of operating status recognition,this thesis builds a model around Long Short-Term Memory networks(LSTM),and uses Convolutional Neural Networks(CNN)and Attention Mechanism(Attention)to optimize the feature extraction process,and the Transformer-LSTM model performs the best.The final result is that the accuracy reaches 96.06%,the macro-F score reaches 0.961,and the Kappa coefficient reaches 0.937.In terms of abnormal state recognition,this thesis designs a abnormal state recognition model based on Particle Swarm Optimization(PSO)and One-Class Support Vector Machine(OCSVM)for the situation where there are few real abnormal data,and combined with LSTM to form a three-level abnormal state recognition model.The final result is that the accuracy reaches 94.69%,and the macroF score reaches 0.958.This model can minimize the false alarm of identifying the normal state as an abnormal state and realizes hierarchical processing of known abnormal data.Through multi-angle testing and universality analysis,it is proved that the data set construction method and model training ideas designed in this thesis have certain practicability in the state identification of scientific instruments.