| The new generation of Product Geometry Specifications（GPS）standardizes the indication of linear dimensions.For cylindrical parts,in addition to the traditional two-point sizes,global and calculated sizes are also specified.The global sizes include the least squares diameter,the minimum circumscribed diameter,the maximum inscribed diameter and the minimax diameter,and the calculated sizes include the circumference diameter,the area diameter and the volume diameter.The application of the sizes above in the mechanical manufacturing industry will face pressing challenges of measurement and limit deviation selection,the systematic research work was carried out in this thesis.Based on the extraction strategy and the relevant literatures,the evaluation models for cylindricity error,roundness error,global sizes and calculated sizes were established.On the basis of the basic principles of the Seagull Optimization Algorithm,the algorithm was improved and the Improved Seagull Optimization Algorithm（ISOA）was used for the evaluation of cylindricity error,roundness error and global sizes.The ISOA algorithm flow was given and the evaluation program for the above errors and sizes was developed by MATLAB.Based on BP neural network and support vector machine（SVM）,the relationship problem between the actual sizes’feature values,cylindricity（or roundness）error and global（or calculated）sizes was studied,and seven relational models were established.The parameters of BP neural network and SVM were optimized using ISOA,and the relational model training program was programmed.The actual sizes’feature values and cylindricity（roundness）error of the tested part are taken as the input values of the above models,and the output value is the indirect measurement value of the global（calculated）size of the tested sample.Based on the standard tolerances,basic deviations,cylindricity tolerance,roundness tolerance of linear size,the prediction method of the global and calculated sizes’limit deviations of cylindrical parts based on the above relationship model was proposed.Using Talyrond 585LT cylindricity measuring instrument,129 external cylindrical（shaft）and 69 inner cylindrical（hole）specimens were extracted with 6789 roundness profiles,and the systematic errors in the extracted roundness profiles were eliminated by using standard columns and standard holes.After eliminating the systematic errors,the roundness profiles were evaluated for the global sizes,calculated sizes,cylindricity errors and roundness errors by four evaluation methods（least squares method,minimum circumscribed method,maximum inscribed method and minimum zone method）by using the developed evaluation program.The actual sizes’feature values,cylindricity（or roundness）error obtained from the above cylindrical specimens were as input values,the global size or calculated size were as target values,and the training of seven relational models were achieved by using BP neural networks（based on different training functions）and SVM parameters of which were optimized by ISOA.The training results showed that:（1）The mean square error（MSE）of the relational model trained based on BP neural network is smaller than the MSE of the relational model constructed based on SVM,but the training time of the former is larger than that of the latter;（2）The training results of BP neural network-based relational model with training function‘trainlm’is better than training function‘trainbr’;（3）The MSE of the relationship model between the actual sizes’feature values,cylindricity error and global size or volume diameter obtained based on BP neural network training are less than 9.25×10-7 mm;（4）The MSE of the relationship model between the actual size feature value,roundness error and calculated size obtained based on BP neural network training is less than 3.07×10-7mm.The accuracy of the above relationship model meets the measurement accuracy requirements for global and calculated sizes of cylindrical parts.Based on the BP neural network relationship models obtained from the above training,the limit deviations of global and calculated sizes for standard tolerance IT5～IT8,cylindricity（roundness）tolerance of Grade 5～8,h（H）and other 7 kinds of shaft（hole）basic deviation codes,nominal size>18～400mm of linear sizes were predicted,and the corresponding global and calculated size limit deviations of cylindrical parts were given.The indirect measurement and limit deviation prediction method of global and calculated sizes based on neural network relational model proposed in this thesis has some theoretical support and practical value for the implementation of global and calculated sizes in manufacturing industry. |
