Research On Technology Health Meanagement For Gas Sensor System

Gas sensor system is the device used for gas concentration information collection.The measurement quality affects the working state of the system tremendous.However,due to the large amount of detection,the harsh working environment,the sensor system will be affected by the environment inevitably,which lead to the decline of system reliability.If the fault data is used to make decision,greater harmful will be brought.The fault of gas sensor system has the characteristics of concealment,correlation and unpredictability.It is difficult to detect all the faults according to the output value of sensor system.Taking the gas metal oxide semiconductor sensor as an example,it is difficult to improve the system reliability from the material and process.There are a large number of unnecessary maintenances in regular maintenance,which is a passive and inefficient maintenance method.There are often multiple sensors failure at the same time or continuously,which cannot meet the requirements by simply using hardware redundancy method.The self-validating method can only realize the fault detection,which cannot realize the fault pre-caution.This method can only passively detect fault,and partially improve the system reliability.Therefore,it is necessary to monitor and evaluate the abnormal state of the sensor system to improve the reliability of the system and the credibility of the output data.In this paper,gas sensor system is taken as the research target.The research focused on the problems of incipient fault diagnosis,health prediction under strong interference and health management decision-making under hybrid multi-attribute information of health management.For the incipient fault diagnosis of sensor system,the paper focuses on data-driven and t-distributed Stochastic Neighbor Embedding(t-SNE)and random forest(RF)is used to achieve fault diagnosis based on the fault mode analysis of sensor system.The long tail effect of t-distribution is applied to increase the between-class-distance of different faults and reduce the inference of mode mixing caused by incipient.The double randomness of random forest can effectively reduce the risk of over fitting and improve the incipient fault diagnosis of sensor system in accuracy.As a result,the separability of different faults is increased.Then random forest is used to classify faults,which improved the accuracy of fault diagnosis.The experimental results show that the accuracy of the proposed method is 4.2% higher than that of the traditional method such as PCA and LDA.Aimed at the health prediction under the condition of strong interference,the paper focuses on data-driven and a kind of health prediction method named unascertained deep ensemble relevance vector machine is proposed.In this paper,health evaluation index with strong anti-interference and small sample predictor are studied.Unascertained deep soft sensor is applied to obtain the anti-interference health reliability degree of sensor system.By utilizing multi-level unascertained evaluation index set,the influence of strong interference data to health reliability degree(HRD)is effectively suppressed.This method can evaluate the measurement state quantitatively.Compared with the traditional method such as grey theory and linear regression,the accuracy of the method is improved by 9% in the case of strong interference.HRD is taken as the evaluation index and Bootstrap method is used as the theoretical framework to construct relevant vector machine ensemble predictor model with moving window technology and bias function,which improve the prediction accuracy in the case of small samples furtherly.The experimental results show that the fault prediction error of this method is as low as 0.6%.For the health management decision-making problem of sensor system under the condition of hybrid multi-attribute information,the paper focus on the reliability of the system and a health management decision-making method based on multi-experts grey group decision-making is proposed,which is used to solve the problem of low reliability of traditional decision-making methods such as D-S evidence theory and fuzzy set theory.By using the method of grey group decision,combined with the mixed information of historical data,maintenance probability and overhaul rate,the corresponding maintenance level of system health management and the optimal decision suggestions are given.Compared with the traditional decision-making method,this method effectively improves the accuracy of health management decision-making under the condition of mixed multi-attribute information by 1.5%.At the same time,aiming at corrective maintenance,the correlation characteristics between sensors is used to construct the prediction model based on cross-correlation extremely learning machine to realize multi-fault recovery,which ensure that the system can continue to work even some sensor is failure and the minimum recovery error can reach 0.5%.To verify the feasibility of the fault diagnosis,health prediction and health management decision-making method of the gas sensor system proposed in this paper,the experimental verification platform is designed and implemented.Taking the air pollution gas sensor system as the research object,the incipient fault diagnosis of the sensor system,such as sensor drift,pulse impact and other fault types,multi-level health assessment and long-term prediction under strong interference are tested.Under this condition,combined with historical data,maintenance rate and other hybrid information,the health management decision-making methods under those two conditions were verified respectively.After the fault recovery for maintenance,the measurement accuracy of the system can be no decline or less decrease.The effectiveness and feasibility of the proposed method for incipient fault diagnosis,health prediction under strong interference and health management decision-making method with mixed multi-attribute information are verified.