The modelling of IR emission spectra and solid rocket motor parameters using neural networks and partial least squares

Hamp, Niko (Stellenbosch : University of Stellenbosch, 2003-04)

Abstract

ENGLISH ABSTRACT: The emission spectrum measured in the middle infrared (IR) band from the plume of a rocket can be used to identify rockets and track inbound missiles. It is useful to test the stealth properties of the IR fingerprint of a rocket during its design phase without needing to spend excessive amounts of money on field trials. The modelled predictions of the IR spectra from selected rocket motor design parameters therefore bear significant benefits in reducing the development costs. In a recent doctorate study it was found that a fundamental approach including quantum-mechanical and computational fluid dynamics (CFD) models was not feasible. This is first of all due to the complexity of the systems and secondly due to the inadequate calculation speeds of even the most sophisticated modern computers. A solution was subsequently investigated by use of the ‘black-box’ model of a multi-layer perceptron feed-forward neural network with a single hidden layer consisting of 146 nodes. The input layer of the neural network consists of 18 rocket motor design parameters and the output layer consists of 146 IR absorbance variables in the range from 2 to 5 μm wavelengths. The results appeared promising for future investigations. The available data consist of only 18 different types of rocket motors due to the high costs of generating the data. The 18 rocket motor types fall into two different design classes, the double base (DB) and composite (C) propellant types. The sparseness of the data is a constraint in building adequate models of such a multivariate nature. The IR irradiance spectra data set consists of numerous repeat measurements made per rocket motor type. The repeat measurements form the pure error component of the data, which adds stability to training and provides lack-of-fit ANOVA capabilities. The emphasis in this dissertation is on comparing the feed-forward neural network model to the linear and neural network partial least squares (PLS) modelling techniques. The objective is to find a possibly more intuitive and more accurate model that effectively generalises the input-output relationships of the data. PLS models are known to be robust due to the exclusion of redundant information from projections made to primary latent variables, similarly to principal components (PCA) regression. The neural network PLS techniques include feed-forward sigmoidal neural network PLS (NNPLS) and radial-basis functions PLS (RBFPLS). The NNPLS and RBFPLS algorithms make use of neural networks to find non-linear functional relationships for the inner PLS models of the NIPALS algorithm. Error-based neural network PLS (EBNNPLS) and radial-basis function network PLS (EBRBFPLS) are also briefly investigated, as these techniques make use of non-linear projections to latent variables. A modification to the orthogonal least squares (OLS) training algorithm of radial-basis functions is developed and applied. The adaptive spread OLS algorithm (ASOLS) allows for the iterative adaptation of the Gaussian spread parameters found in the radial-basis transfer functions. Over-fitting from over-parameterisation is controlled by making use of leaveone- out cross-validation and the calculation of pseudo-degrees of freedom. After cross-validation the overall model is built by training on the entire data set. This is done by making use of the optimum parameterisation obtained from cross-validation. Cross-validation also gives an indication of how well a model can predict data unseen during training. The reverse problem of modelling the rocket propellant chemical compositions and the rocket physical design parameters from the IR irradiance spectra is also investigated. This problem bears familiarity to the field of spectral multivariate calibration. The applications in this field readily make use of PLS and neural network modelling. The reverse problem is investigated with the same modelling techniques applied to the forward modelling problem. The forward modelling results (IR spectrum predictions) show that the feedforward neural network complexity can be reduced to two hidden nodes in a single hidden layer. The NNPLS model with eleven latent dimensions outperforms all the other models with a maximum average R2-value of 0.75 across all output variables for unseen data from cross-validation. The explained variance for the output data of the overall model is 94.34%. The corresponding explained variance of the input data is 99.8%. The RBFPLS models built using the ASOLS training algorithm for the training of the radialbasis function inner models outperforms those using K-means and OLS training algorithms. The lack-of-fit ANOVA tests show that there is reason to doubt the adequacy of the NNPLS model. The modelling results however show promise for future development on larger, more representative data sets. The reverse modelling results show that the feed-forward neural network model, NNPLS and RBFPLS models produce similar results superior to the linear PLS model. The RBFPLS model with ASOLS inner model training and 5 latent dimensions stands out slightly as the best model. It is found that it is feasible to separately find the optimum model complexity (number of latent dimensions) for each output variable. The average R2-value across all output variables for unseen data is 0.43. The average R2-value for the overall model is 0.68. There are output variables with R2-values of over 0.8. The forward and reverse modelling results further show that dimensional reduction in the case of PLS does produce the best models. It is found that the input-output relationships are not highly non-linear. The non-linearities are largely responsible for the compensation of both the DB- and C-class rocket motor designs predictions within the overall model predictions. For this reason it is suggested that future models can be developed by making use of a simpler, more linear model for each rocket class after a class identification step. This approach however requires additional data that must be acquired.

AFRIKAANSE OPSOMMING: Die emissiespektra van die uitlaatpluime van vuurpyle in die middel-infrarooi (IR) band kan gebruik word om die vuurpyle te herken en om inkomende vuurpyle op te spoor. Dit is nuttig om die uitstralingseienskappe van ‘n vuurpyl se IR afdruk te toets, sonder om groot bedrae geld op veldtoetse te spandeer. Die gemodelleerde IR spektrale voorspellings vir ‘n bepaalde stel vuurpylmotor ontwerpsparameters kan dus grootliks bydra om motorontwikkelingskostes te bemoei. In ‘n onlangse doktorale studie is gevind dat ‘n fundamentele benadering van kwantum-meganiese en vloeidinamika-modelle nie lewensvatbaar is nie. Dit is hoofsaaklik as gevolg van die onvoldoende vermoë van selfs die mees gesofistikeerde moderne rekenaars. ‘n Moontlike oplossing tot die probleem is ondersoek deur gebruik te maak van ‘n multilaag perseptron voorwaartse neurale netwerk met 146 nodes in ‘n enkele versteekte laag. Die laag van invoer veranderlikes bestaan uit agtien vuurpylmotor ontwerpsparameters en die uitvoerlaag bestaan uit 146 IR-absorbansie veranderlikes in die reeks golflengtes vanaf 2 tot 5 μm. Dit het voorgekom dat die resultate belowend lyk vir toekomstige ondersoeke. Weens die hoë kostes om die data te genereer bestaan die beskikbare data uit slegs agtien verskillende tipes vuurpylmotors. Die agtien vuurpyl tipes val verder binne twee ontwerpsklasse, naamlik die dubbelbasis (DB) en saamgestelde (C) dryfmiddeltipes. Die yl data bemoeilik die bou van doeltreffende multiveranderlike modelle. Die datastel van IR uitstralingspektra bestaan uit herhaalde metings per vuurpyltipe. Die herhaalde metings vorm die suiwer fout komponent van die data. Dit verskaf stabilitieit tot die opleiding op die data en verder die vermoë om ‘n analise van variansie (ANOVA) op die data uit te voer. In hierdie tesis lê die klem op die vergelyking tussen die voorwaartse neurale netwerk en die lineêre en neurale netwerk parsiële kleinste kwadrate (PLS) modelleringstegnieke. Die doel is om ‘n moontlik meer insiggewende en akkurate model te vind wat effektief die in- en uitvoer verhoudings kan veralgemeen. Dit is bekend dat PLS modelle meer robuus kan wees weens die weglating van oortollige inligting deur projeksies op hoof latente veranderlikes. Dit is analoog aan hoofkomponente (PCA) regressie. Die neurale netwerk PLS-tegnieke sluit in voorwaartse sigmoïdale neurale netwerk PLS (NNPLS) en radiale-basis funksies PLS (RBFPLS). Die NNPLS en RBFPLS algoritmes maak gebruik van die neurale netwerke om nie-lineêre funksionele verbande te kry vir die binne PLS-modelle van die nie-lineêre iteratiewe parsiële kleinste kwadrate (NIPALS) algoritme. Die fout-gebaseerde neurale netwerk PLS (EBNNPLS) en radiale-basis funksies PLS (EBRBFPLS) is ook weens hulle nie-lineêre projeksies na latente veranderlikes kortiliks ondersoek. ‘n Aanpassing tot die ortogonale kleinste kwadrate (OLS) opleidingsalgoritme vir radiale-basis funksies is ontwikkel en toegepas. Die aangepaste algoritme (ASOLS) behels die iteratiewe aanpassing van die verspreidingsparameters binne die Gauss-funksies van die radiale-basis transformasie funksies. Die oormatige parameterisering van ‘n model word beheer deur kruisvalidering met enkele weglatings en die berekening van pseudo-vryheidsgrade. Na kruisvalidering word die algehele model gebou deur opleiding op die volledige datastel. Dit word gedoen deur van die optimale parameterisering gebruik te maak wat deur kruisvalidering bepaal is. Kruisvalidering gee ook ‘n goeie aanduiding van hoe goed ‘n model ongesiende data kan voorspel. Die modellering van die vuurpyle se chemiese en fisiese ontwerpsparameters (omgekeerde probleem) is ook ondersoek. Hierdie probleem is verwant aan die veld van spektrale multiveranderlike kalibrasie. Die toepassings in die veld maak gebruik van PLS en neurale netwerk modelle. Die omgekeerde probleem word dus ondersoek met dieselfde modelleringstegnieke wat gebruik is vir die voorwaartse probleem. Die voorwaartse modelleringsresultate (IR voorspellings) toon dat die kompleksiteit van die voorwaartse neurale netwerk tot twee versteekte nodes in ‘n enkele versteekte laag gereduseer kan word. Die NNPLS model met elf latente dimensies vaar die beste van alle modelle, met ‘n maksimum R2-waarde van 0.75 oor alle uitvoer veranderlikes vir die ongesiende data (kruisvalidering). Die verklaarde variansie vir die uitvoer data vanaf die algehele model is 94.34%. Die verklaarde variansie van die ooreenstemmende invoer data is 99.8%. Die RBFPLS modelle wat gebou is deur van die ASOLS algoritme gebruik te maak om die PLS binne modelle op te lei, vaar beter in vergelyking met die K-gemiddeldes en OLS opleidingsalgoritmes. Die toetse wat ‘n ‘tekort-aan-passing’ ANOVA behels, toon dat daar rede is om die geskiktheid van die NNPLS model te wantrou. Die modelleringsresultate lyk egter belowend vir die toekomstige ontwikkeling van modelle op groter, meer verteenwoordigde datastelle. Die omgekeerde modellering toon dat die voorwaartse neurale netwerk, NNPLS en RBFPLS modelle soortgelyke resultate produseer wat die lineêre PLS model s’n oortref. Die RBFPLS model met ASOLS opleiding van die PLS binne modelle word beskou as die beste model. Dit is lewensvatbaar om die optimale modelkompleksiteite van elke uitvoerveranderlike individueel te bepaal. Die gemiddelde R2-waarde oor alle uitvoerveranderlikes vir ongesiende data is 0.43. Die gemiddelde R2-waarde vir die algehele model is 0.68. Daar is van die uitvoer veranderlikes wat R2-waardes van 0.8 oortref. Die voor- en terugwaartse modelleringsresultate toon verder dat dimensionele reduksie in die geval van PLS die beste modelle lewer. Daar is ook gevind dat die nie-lineêriteite grootliks vergoed vir die voorspellings van beide DB- en Ctipe vuurpylmotors binne die algehele model. Om die rede word voorgestel dat toekomstige modelle ontwikkel kan word deur gebruik te maak van eenvoudiger, meer lineêre modelle vir elke vuurpylklas nadat ‘n klasidentifikasiestap uitgevoer is. Die benadering benodig egter addisionele praktiese data wat verkry moet word.

Thesis (MScIng)--University of Stellenbosch, 2003.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/16334
This item appears in the following collections:
Find Full text

The following license files are associated with this item: