Neural Network Model for Identification of Material Creep Curves Using Cuda Technologies

Abstract

This pa­per addres­ses the prob­lem of iden­tif­ying rhe­olo­gi­cal pa­ra­me­ters of wo­od using ar­ti­fi­ci­al neu­ral net­works with pa­ral­lel le­ar­ning al­go­rithm using Python prog­ram­ming lan­gua­ge, Cha­iner fra­me­work and CU­DA techno­logy. An in­tel­li­gent system for iden­ti­fi­ca­ti­on of rhe­olo­gi­cal pa­ra­me­ters of wo­od has be­en de­ve­lo­ped. The system cre­ated con­ta­ins the most user-fri­endly in­ter­fa­ce, all the ne­ces­sary set of to­ols for au­to­ma­ti­on of the pro­cess of vis­ua­li­za­ti­on and analysis of da­ta. In the pro­cess of cre­ati­on of the in­tel­lec­tu­al system, the fol­lo­wing tasks we­re en­vi­sa­ged: to carry out the analysis of ar­ti­fi­ci­al in­tel­li­gen­ce systems and the analysis of tra­ining of ar­ti­fi­ci­al neu­ral net­works, in par­ti­cu­lar mul­ti­la­yer neu­ral net­works of di­rect pro­pa­ga­ti­on, re­cur­rent neu­ral net­works and the Ko­ho­nen neu­ral net­work; exa­mi­ne the struc­tu­re of the Cha­iner fra­me­work and its in­te­rac­ti­on with CU­DA; to con­duct exis­ting clo­ud techno­lo­gi­es to ac­complish the task; to con­duct the analysis of al­go­rithms of stu­di­es of ar­ti­fi­ci­al neu­ron net­works, the­ir mat­he­ma­ti­cal pro­vi­ding; to imple­ment pa­ral­le­li­za­ti­on of le­ar­ning al­go­rithms and to de­ve­lop the ne­ces­sary softwa­re. Using Cha­iner al­lows you to cre­ate a me­mory po­ol for GPU me­mory al­lo­ca­ti­on. To avo­id me­mory al­lo­ca­ti­on and era­su­re du­ring com­pu­ting, Cha­iner pro­vi­des the abi­lity to use the CuPy me­mory po­ol as a stan­dard me­mory al­lo­ca­ti­on wit­ho­ut de­aling with me­mory al­lo­ca­ti­on. An in­tel­lec­tu­al system to de­ter­mi­ne the physi­cal and mec­ha­ni­cal pa­ra­me­ters of a mat­he­ma­ti­cal mo­del of non-isot­her­mal mo­is­tu­re transfer and vis­co­elas­tic de­for­ma­ti­on of ca­pil­lary-po­ro­us ma­te­ri­als was de­ve­lo­ped. It pro­vi­des the op­por­tu­nity to iden­tify pa­ra­me­ters of the ker­nels of cre­ep and re­la­xa­ti­on that is writ­ten as a li­ne­ar com­bi­na­ti­on of ex­po­nen­ti­al ope­ra­tors. The pro­po­sed al­go­rithm of appro­xi­ma­ti­on and ob­ta­ined cal­cu­la­ted ra­ti­os of rhe­olo­gi­cal be­ha­vi­or of wo­od by me­ans of mul­ti­la­yer neu­ral net­work with ex­po­nen­ti­al ac­ti­va­ti­on functi­ons in hid­den la­yers al­lows to incre­ase the ac­cu­racy of appro­xi­ma­ti­on of ex­pe­ri­men­tal cre­ep da­ta. The de­ve­lo­ped mat­he­ma­ti­cal mo­dels can be used to cre­ate an au­to­ma­ted systems of fi­ni­te-dif­fe­ren­ce cal­cu­la­ti­on of tem­pe­ra­tu­re and mo­is­tu­re con­tent, stress com­po­nents du­ring the drying of ca­pil­lary-po­ro­us ma­te­ri­als with ta­king in­to ac­co­unt the techno­lo­gi­cal pa­ra­me­ters of the drying agent.

Development of software and algorithms of parallel learning of artificial neural networks using CUDA technologies

“…In works [1,2], the features of the software implementation of the learning procedure for artificial neural networks using a graphics processor were considered and an acceleration was obtained in comparison with learning using a central processor. Since the goal was to study the learning process itself, only classical datasets and simple ANN models were used.…”