Data-driven Prediction Model For Small Sample Data Of Metal Materials Atmospheric Corrosion.

Metal atmospheric corrosion can cause huge economic property losses to the country and society and may induce accidents,resulting in more serious personal safety losses.By establishing an effective prediction model,it can not only greatly reduce the economic loss caused by atmospheric corrosion,but also realize the corrosion knowledge mining,and provide qualitative and quantitative analysis results for corrosion research,which has an important practical engineering significance.In addition,for metal atmospheric corrosion data,especially for outdoor test samples,the complex features contained therein also pose challenges for corresponding predictive modeling.Therefore,how to select the model according to the characteristics of atmospheric corrosion data and improve it to obtain higher prediction accuracy also has theoretical research significance.This thesis collects the corrosion data of two commonly used metal materials(carbon steel and low alloy steel)in real outdoor atmospheric environment as the research object.Based on the characteristics,the collected corrosion samples are classified into time-univariate corrosion rate sequence data sets and multivariate-corrosion rate data sets including environmental factors,materials and time.The thesis analyzes the characteristics of the two types of data sets and conducts the following three parts:(1)For the two types of atmospheric corrosion sequence data,two prediction models based on grey theory are proposed.Firstly,for the monotonic sequence of non-equal interval sampling,the thesis proposes a combination of generalized inverse harmonic mean weakening buffer operator(GCHM_WBO)and non-equal interval GM(1,1).The monotonicity of GCHM_WBO is proved and a GCHM_WBO parameter search algorithm with re-adjustment process based on historical data are designed.The proposed algorithm can effectively reduce the influence of data disturbance in the sequence and improve the long-term prediction accuracy of corrosion.Secondly,for the monotonic sequence with differential oscillation,the thesis proposes a nonlinear grey Bernoulli model based on regularized genetic algorithm(RGANGBM(1,1)).The algorithm is for solving NGBM(1,1)parameters and the problem of less modeling data.The regularized genetic algorithm is designed to optimize the NGBM(1,1)parameters under super-small samples,and finally the prediction accuracy is improved.(2)As for the practical characteristics of multivariate datasets of low-alloy steel atmospheric corrosion,the thesis proposes a random forest dynamic integrated selection algorithm(RF-WKNNs)based on weighted K-nearest neighbors.On the basis of the random forest(RF),combined with the importance of each input variable on the corrosion effect,the weighted K-nearest neighbor algorithm(WKNNs)and dynamic integration selection process are constructed to improve the combination of tree models in random forests.The purpose of improving the prediction accuracy of the model is achieved finally.Subsequently,this thesis uses the proposed algorithm to mine some knowledge in the field of corrosion,and gives the influence changing of corrosion input variables along with time and the qualitative and quantitative analysis of various environmental factors on corrosion.(3)In order to further improve the prediction accuracy,this thesis combines RF-WKNNs and cForest model to propose a fully cascade-connected dynamic ensemble selection forest algorithm(DCCF-WKNNs)with deep structure.Firstly,based on cForest,the algorithm changes the hierarchical connection mode of the deep model into the full connection mode,making full use of the extracted information of each layer.Secondly,the random forest model in cForest is replaced by RF-WKNNs algorithm to improve the prediction accuracy of the model.Finally,with the advantage of the deep structure of the proposed model,the qualitative and quantitative analysis of the effects of environmental variables on corrosion are conducted.This thesis provides a new and novel research idea and method for the inter-discipline field of material corrosion and computer science.