Spelling suggestions: "subject:"ensory evaluation - atatistical methods"" "subject:"ensory evaluation - astatistical methods""
1 |
Prediction of consumer liking from trained sensory panel information: evaluation of artificial neural networks (ANN)Krishnamurthy, Raju, Chemical Sciences & Engineering, Faculty of Engineering, UNSW January 2007 (has links)
This study set out to establish artificial neural networks (ANN) as an alternate to regression methods (multiple linear, principal components and partial least squares regression) to predict consumer liking from trained sensory panel data. The study has two parts viz., I) Flavour study - evaluation of ANNs to predict consumer flavour preferences from trained sensory panel data and 2) Fragrance study ??? evaluation of different ANN architectures to predict consumer fragrance liking from trained sensory panel data. In this study, a multi-layer feedforward neural network architecture with input, hidden and output layer(s) was designed. The back-propagation algorithm was utilised in training of neural networks. The network learning parameters such as learning rate and momentum rate were optimised by the grid experiments for a fixed number of learning cycles. In flavour study, ANNs were trained using the trained sensory panel raw data as well as transformed data. The networks trained with sensory panel raw data achieved 98% correct learning, whereas the testing was within the range of 28 -35%. A suitable transformation methods were applied to reduce the variations in trained sensory panel raw data. The networks trained with transformed sensory panel data achieved between 80-90% correct learning and 80-95% correct testing. In fragrance study, ANNs were trained using the trained sensory panel raw data as well as principal component data. The networks trained with sensory panel raw data achieved 100% correct learning, and testing was in a range of 70-94%. Principal component analysis was applied to reduce redundancy in the trained sensory panel data. The networks trained with principal component data achieved about 100% correct learning and 90% correct testing. It was shown that due to its excellent noise tolerance property and ability to predict more than one type of consumer liking using a single model, the ANN approach promises to be an effective modelling tool.
|
2 |
Evaluating the properties of sensory tests using computer intensive and biplot methodologiesMeintjes, M. M. (Maria Magdalena) 03 1900 (has links)
Assignment (MComm)--University of Stellenbosch, 2007. / ENGLISH ABSTRACT: This study is the result of part-time work done at a product development centre. The organisation extensively makes use of trained panels in sensory trials designed to asses the quality of its product. Although standard statistical procedures are used for analysing the results arising from these trials, circumstances necessitate deviations from the prescribed protocols. Therefore the validity of conclusions drawn as a result of these testing procedures might be questionable. This assignment deals with these questions.
Sensory trials are vital in the development of new products, control of quality levels and the exploration of improvement in current products. Standard test procedures used to explore such questions exist but are in practice often implemented by investigators who have little or no statistical background. Thus test methods are implemented as black boxes and procedures are used blindly without checking all the appropriate assumptions and other statistical requirements. The specific product under consideration often warrants certain modifications to the standard methodology. These changes may have some unknown effect on the obtained results and therefore should be scrutinized to ensure that the results remain valid.
The aim of this study is to investigate the distribution and other characteristics of sensory data, comparing the hypothesised, observed and bootstrap distributions. Furthermore, the standard testing methods used to analyse sensory data sets will be evaluated. After comparing these methods, alternative testing methods may be introduced and then tested using newly generated data sets.
Graphical displays are also useful to get an overall impression of the data under consideration. Biplots are especially useful in the investigation of multivariate sensory data. The underlying relationships among attributes and their combined effect on the panellists’ decisions can be visually investigated by constructing a biplot. Results obtained by implementing biplot methods are compared to those of sensory tests, i.e. whether a significant difference between objects will correspond to large distances between the points representing objects in the display. In conclusion some recommendations are made as to how the organisation under consideration should implement sensory procedures in future trials. However, these proposals are preliminary and further research is necessary before final adoption. Some issues for further investigation are suggested. / AFRIKAANSE OPSOMMING: Hierdie studie spruit uit deeltydse werk by ’n produk-ontwikkeling-sentrum. Die organisasie maak in al hul sensoriese proewe rakende die kwaliteit van hul produkte op groot skaal gebruik van opgeleide panele. Alhoewel standaard prosedures ingespan word om die resultate te analiseer, noodsaak sekere omstandighede dat die voorgeskrewe protokol in ’n aangepaste vorm geïmplementeer word. Dié aanpassings mag meebring dat gevolgtrekkings gebaseer op resultate ongeldig is. Hierdie werkstuk ondersoek bogenoemde probleem.
Sensoriese proewe is noodsaaklik in kwaliteitbeheer, die verbetering van bestaande produkte, asook die ontwikkeling van nuwe produkte. Daar bestaan standaard toets- prosedures om vraagstukke te verken, maar dié word dikwels toegepas deur navorsers met min of geen statistiese kennis. Dit lei daartoe dat toetsprosedures blindelings geïmplementeer en resultate geïnterpreteer word sonder om die nodige aannames en ander statistiese vereistes na te gaan. Alhoewel ’n spesifieke produk die wysiging van die standaard metode kan regverdig, kan hierdie veranderinge ’n groot invloed op die resultate hê. Dus moet die geldigheid van die resultate noukeurig ondersoek word.
Die doel van hierdie studie is om die verdeling sowel as ander eienskappe van sensoriese data te bestudeer, deur die verdeling onder die nulhipotese sowel as die waargenome- en skoenlusverdelings te beskou. Verder geniet die standaard toetsprosedure, tans in gebruik om sensoriese data te analiseer, ook aandag. Na afloop hiervan word alternatiewe toetsprosedures voorgestel en dié geëvalueer op nuut gegenereerde datastelle.
Grafiese voorstellings is ook nuttig om ’n geheelbeeld te kry van die data onder bespreking. Bistippings is veral handig om meerdimensionele sensoriese data te bestudeer. Die onderliggende verband tussen die kenmerke van ’n produk sowel as hul gekombineerde effek op ’n paneel se besluit, kan hierdeur visueel ondersoek word. Resultate verkry in die voorstellings word vergelyk met dié van sensoriese toetsprosedures om vas te stel of statisties betekenisvolle verskille in ’n produk korrespondeer met groot afstande tussen die relevante punte in die bistippingsvoorstelling.
Ten slotte word sekere aanbevelings rakende die implementering van sensoriese proewe in die toekoms aan die betrokke organisasie gemaak. Hierdie aanbevelings word gemaak op grond van die voorafgaande ondersoeke, maar verdere navorsing is nodig voor die finale aanvaarding daarvan. Waar moontlik, word voorstelle vir verdere ondersoeke gedoen.
|
3 |
A simulation tool for evaluating sensory data analysis methodsNaini, Shuo 09 May 2003 (has links)
In cross-cultural studies, respondents from specific cultures may have
different product preferences and scale usage. Combining data from different
cultures will result in departures from the basic assumptions of analysis of variance
(ANOVA) and loss of power in testing capability of finding product and culture
differences. However, the result of violations on power of ANOVA is unknown by
sensory researchers. The objectives of this research were by simulating consumer
product evaluation data, to evaluate the robustness and testing power of ANOVA
under different cross-cultural situations.
The study was conducted in two parts. First, an Empirical Logit simulation
model was employed for generating sensory data. This model included respondent,
product, consumer segment and product by segment interaction effects. Four
underlying distributions: Binomial, Beta-Binomial, Hypergeometric, and Beta-Hypergeometric were used to increase or decrease the dispersion of the responses.
Alternatively, instead of using these four distributions, the same applications were
achieved by a binning step. The entire simulation procedure including the
Empirical Logit model and the binning step was called Discrete Empirical Logit
model. In the second part of the study, the Discrete Empirical Logit model was
chosen to generate specified data sets under six different cross-cultural cases. After
analyzing these data sets by ANOVA reduced and full models, the empirical power
of ANOVA under different cases was calculated and compared.
The results showed that both Beta-Hypergeometric and Discrete Empirical
Logit were flexible on simulating sensory responses, but the Discrete Empirical
Logit was relatively simple to use. Comparing with the ANOVA reduced model,
the full model gave better information on evaluating the case that segments differ in
product preferences. This suggested segmentation was very important in cross-cultural
data analysis. Under the situations that sample sizes were equal and
respondents performed consistently within segment (MSE ≈ 1), ANOVA was very
robust to different scale usage, losing at worst 18% in power.
From the scope of this study, we recommend using the ANOVA full model
in the cross-cultural research. Results from different cultures could be combined
when consistency within segments was high. / Graduation date: 2003
|
4 |
Avaliação do desempenho de quatro metodos de escalonamento em testes sensoriais de aceitação utilizando modelos normais aditivos de analise da variancia e mapas internos de preferencia / Assessing the performance of four methods of phasing in tests of sensory acceptance additives using standard models of analysis of variance and internal maps of preferenceMontes Villanueva, Nilda Doris 31 July 2003 (has links)
Orientadores: Maria Aparecida A. Pereira da Silva, Ademir Jose Petenate / Tese (doutorado) - Universidade Estadual de Campinas, Faculdade de Engenharia de Alimentos / Made available in DSpace on 2018-08-03T16:42:14Z (GMT). No. of bitstreams: 1
MontesVillanueva_NildaDoris_D.pdf: 7571939 bytes, checksum: 5b97da35754c8719f94ee3b21e0cf955 (MD5)
Previous issue date: 2003 / Resumo: Em testes sensoriais, a análise dos dados geralmente é realizada através de algum modelo ANOVA. Estes modelos pressupõem que as respostas experimentais sejam: i) independentes, ii) normalmente distribuídas, m) homoscedásticas (variâncias iguais) e, iv) provenientes de uma mesma escala de medida (aditividade). Os principais problemas na análise de dados sensoriais através de modelos ANOV A referem-se aos dois últimos pressupostos. A homogeneidade das variâncias não pode ser assegurada devido à existência de pelo menos duas fontes potenciais de variabilidade dos dados, quais sejam: provadores e tratamentos. Por outro lado, a aditividade pode ser violada quando um provador utiliza faixas consistentemente mais (ou menos) amplas da escala para expressar a sua impressão sobre o produto. A maneira pessoal com que cada provador utiliza a escala para avaliar os produtos, chama-se de variação da expansibilidade entre provadores. Tanto a falta de homogeneidade das variâncias como a não aditividade do modelo, acarretam conseqüências sérias na obtenção do verdadeiro nível de significância para o efeito dos tratamentos, podendo afetar adversamente as comparações entre as médias dos tratamentos e comprometer seriamente tanto a interpretação dos resultados fornecidos pelo experimento como a validade do modelo ANOVA. Em testes com consumidores, escalas tradicionais como a escala hedônica de 9 pontos freqüentemente apresentam a seguinte problemática: i) geram dados que freqüentemente não satisfazem os pressupostos estatísticos de normalidade, aditividade e homoscedasticidade exigidos nos modelos ANOVA, ii) oferecem pouca liberdade aos consumidores para expressarem. suas percepções sensoriais, devido ao limitado número de categorias, m) induzem efeitos numéricos e contextuais no julgamento dos provadores e, iv) os valores numéricos associados às suas categorias, embora numericamente possuam intervalos iguais, não refletem iguais diferenças em percepção. Das metodologias utilizadas em testes sensoriais com consumidores, a escala hedônica de 9 pontos, é sem dúvida, a mais utilizada. Porém, em função da problemática anteriormente mencionada, surge a necessidade de serem pesquisadas escalas alternativas que possuam um melhor desempenho que a escala hedônica tradicional, tanto quando os dados são analisados através de modelos ANOVA como quando os mesmos são analisados através de métodos multivariados como Mapa Interno de Preferência - MDPREF. De um modo geral, o objetivo do presente trabalho foi pesquisar o desempenho de duas escalas alternativas em estudos com consumidores, quais sejam: escala autoajustável e escala hedônica híbrida, comparando-as com métodos afetivos tradicionais como a escala de ordenação e escala hedônica de 9 pontos. Para isso, três experimentos foram realizados conforme descrito a seguir: o primeiro experimento foi realizado com o objetivo de se avaliar em condições reais de teste de consumidor, o desempenho da escala autoajustável em relação à escala hedônica de 9 pontos e escala de ordenação, utilizando-se os seguintes critérios: i) diferenças em expansibilidade entre provadores, ii) poder discriminativo e, iii) adequação dos dados coletados por cada escala aos pressupostos do modelo ANOVA. Três marcas comerciais de confeitos foram avaliadas por 288 consumidores. Os resultados obtidos através das escalas hedônica de 9 pontos e autoajustável foram analisados através de ANOVA e os resultados da escala de ordenação, através do teste de Friedman. Os valores de pFamostra. pFprovador e QMresíduo fornecidos pela ANOVA de cada escala, foram respectivamente utilizados para avaliar o poder discriminativo, a expansibilidade dos provadores e a variabilidade residual dos dados. Teste de Tukey foi também aplicado para análise do poder discriminativo de cada escala. A normalidade dos dados foi verificada através do cálculo dos Coeficientes de assimetria e curtose, gráfico de probabilidade normal e teste de Kolmogorov-Smirnov. A homoscedasticidade, foi avaliada através de gráficos de dispersão e teste de Levene. Os resultados mostraram que a escala autoajustável foi efetiva para tratar o problema da expansibilidade entre provadores e da desigualdade das variâncias, porém, os resíduos mostraram moderados desvios da normalidade. A escala hedônica de 9 pontos apresentou problemas de heteroscedasticidade. As escalas autoajustável e de ordenação apresentaram o menor e o maior poder discriminativo respectivamente. Apesar dos problemas detectados, as três escalas apresentaram as mesmas tendências de preferência dos produtos avaliados. O segundo experimento foi realizado com o objetivo de se avaliar o desempenho da escala hedônica híbrida em estudos com consumidores, comparando-a à escala hedônica de 9 pontos, escala autoajustável, e escala de ordenação; através dos seguintes critérios: i) variabilidade das respostas sensoriais, ii) poder discriminativo, iii) adequação dos dados às suposições dos modelos ANOVA e, iv) facilidade de uso pelos consumidores. Cinco marcas de suco de laranja foram avaliadas por 80 consumidores, divididos em quatro grupos de 20 indivíduos cada. Todos os indivíduos avaliaram todas as amostras através de todas as escalas em 4 diferentes sessões de degustação. Um delineamento em quadrado latino 4x4, foi utilizado para controlar o efeito de ordem de apresentação das escalas e avaliar sem vícios a facilidade de uso das mesmas. Para cada escala, a ordem de apresentação das amostras e efeitos residuais ("carry-over") foram balanceados. Os resultados obtidos através das escalas hedônica tradicional, híbrida e autoajustável foram avaliados através de ANOVA. A normalidade dos dados foi verificada através do teste de Shapiro-Wilks, a homoscedasticidade através do teste de Brown-Forsythe e a aditividade, através do teste de Tukey para um grau de liberdade. Os valores de pFamostra, pFprovador e QMresíduo fornecidos pela ANOVA de cada escala, foram respectivamente utilizados para avaliar o poder discriminativo, a expansibilidade dos provadores e a variabilidade residual dos dados. O teste de REGWQ foi também aplicado para análise do poder discriminativo de cada escala. Os resultados obtidos através da escala de ordenação foram avaliados pelo teste de Friedman e, a facilidade de uso das escalas por testes de Cochran-Mantel-Haenszel. Os resultados sugeriram uma superioridade da escala hedônica híbrida sobre as escalas hedônica estruturada e a utoaj ustável , tanto em função do poder discriminativo como da adequação dos dados às suposições de normalidade e homoscedasticidade. A despeito dos dados da escala autoajustável terem apresentado maior variabilidade e sérios desvios da normalidade dos resíduos, o poder discriminativo desta escala foi ligeiramente superior ao da escala hedônica estruturada. A escala de ordenação apresentou o menor poder discriminativo em relação às demais. As escalas hedônicas estruturada e híbrida foram consideradas significativamente (p:S;0,01) mais fáceis de serem utilizadas que a autoajustável, não havendo diferença (p:s;O,OS) entre as duas primeiras. Finalmente, o objetivo do terceiro experimento foi avaliar o desempenho das escalas hedônica estruturada, hedônica híbrida e autoajustável na construção de Mapas Internos de Preferência - MDPREF. Nesta pesquisa, a aceitação global de 8 marcas comerciais de vinho tinto, a maioria deles varietal Cabernet Sauvignon, foi avaliada por 112 consumidores. Foram utilizados delineamentos experimentais balanceados para ordem de apresentação das escalas, ordem de apresentação das amostras e efeitos residuais. Os dados foram analisados através de ANOV A e MDPREF. O critério de avaliação do desempenho da cada escala baseou-se no número de consumidores significativamente ajustados (ps O,OS) e no grau de segmentação dos produtos e dos consumidores produzidos pelo MDPREF. Os resultados sugeriram uma superioridade da escala híbrida sobre a escala hedônica tradicional e autoajustável. O MDPREF gerado pelos dados da escala híbrida produziu um maior número de dimensões significativas de preferência (pS O,OS), trazendo como decorrência, uma porcentagem de 79,S% consumidores significativamente ajustados (pS O,OS), enquanto a escala autoajustável ajustou S4,S% dos consumidores e a escala hedônica S1,8%. Em geral a escala hedônica de 9 pontos apresentou um desempenho inferior ao das demais escalas. Os resultados do presente estudo sugerem fortemente que a escala hedônica híbrida é uma ferramenta válida e eficiente que pode ser utilizada na coleta de dados associados a estudos com consumidores, tanto quando eles forem analisados através de modelos normais para análise da variância como através da metodologia de Mapa Interno de Preferência / Abstract: In sensory tests, the basie statistieal toei for analyzing data is almost invariably some sort of analysis of variance models. These models presuppose that the experimental responses are: i) independent, ii) normally distributed, iii) homoscedastie (have equal varianees) and, iv) seores are on the same scale of measurement (additivity). The main problems arising from the analysis of sensory data using ANOVA models are related to the last two assumptions. Homogeneity of error variance is not assured, espeeially as there are at least two potential sources of heterogeneity: treatments and assessors. On the other hand, the additivity could be violated if one assessor used a eonsistently larger (or smaller) portion of the scale range, scoring more (or less) expansively than other assessors to express his opinion of the produet. The individual way in whieh eaeh panelist uses the scale to evaluate the produets is known as the differential expansiveness of seoring between assessors. 80th the laek of homogeneity of the variances and the non-additivity of the model, result in serious consequenees in obtaining a true levei of significance for the effect of the treatments and may adversely affeet the eomparison of treatment means. The non-additivity can seriously affeet and possible invalidate the analysis of variance and the interpretation of the results that it provides. In consumer tests, traditional scales sueh as the nine-point hedonie scale frequently present the following problems: i) they do not satisfy the statistical assumptions of independenee, normalityand homoscedastieity required by ANOVA models; ii) they give little freedom to the individuais to express their perceptions, due to the limited number of categories; iii) they induce numerical and contextual effects in the judgments by the panelists and, iv) the difference between numerical values associated with the categories do not reflect equivalent differenees in perception. Of the methodologies used in sensory tests with consumers, the 9-point hedonie scale is undoubtedly the most widely used. However, considering the previously mentioned problem, there is a need to investigate alternative scales providing better performanee than the traditional hedonie scale, both when the data are analyzed by ANOVA models and multivariate methods such as the Internal Preference Map - MDPREF. In general the objective of this research was to investigate the performance of two alternative scales in consumer studies, these being the self-adjusting scale and the hybrid hedonic scale, comparing them with traditional affective methods such as the ranking scale and the 9-point hedonic scale. With this objective three experiments were carried out as follows: The first experiment was carried out with the objective of evaluating the performance of the self-adjusting scale as compared to the 9-point hedonic scale and ranking scale under real consumer test conditions, using the following criteria: i) differential expansiveness between assessors, ii) discriminating power and, iii) compliance of the data collected by each scale with the ANOVA assumptions. Three commercial brands of candy were evaluated by 288 consumers. The results obtained from the 9-point hedonic and self-adjusting scales were analyzed by ANOVA and those of the ranking test by Friedman's test. The values for pFsample, pFassessor and QMerror provided by ANOV A for each scale, were used respectively to evaluate the discriminating power, the expansiveness of scoring between assessors and the data variability. Tukey's test was also applied to analyze the discriminating power of each scale. Normal probability plots, Kolmogorov-Smirnov test and coefficients of skewness and kurtosis checked data normality. Homoscedasticity was evaluated by scatter plots and the Levene test. The results showed that the self-adjusting scale was effective to deal with differential assessor expansiveness and produced homogeneous variances, however the residuais showed moderate deviations from normality. The 9-point hedonic scale showed problems with heteroscedasticity. Rank and the self-adjusting scales showed the highest and the lowest discriminating powers, respectively. Despite the problems detected, the three scales presented the same tendencies for preference amongst the products tested. The second experiment was carried out with the objective of evaluating the performance of the hybrid hedonic scale in consumer studies, comparing it with the 9point hedonic, the self-adjusting and the ranking scales, using the following criteria: i) variability of sensory response, ii) discriminative power, iii) data adequacy to the assumptions of ANOVA models and, iv) ease of use. Eighty consumers, divided into four groups of 20 individuais each, evaluated tive brands of orange juice. Ali the individuais evaluated ali the samples using ali the scales, in 4 distinct tasting sessions. A 4 x 4 Latin square design was used to control the effect of the order of presentation of the scales and evaluate their ease of use without biases. For each scale the presentation order and carry-over were balanced. The results obtained using the traditional hedonic, hybrid hedonic and self-adjusting scales were evaluated using ANOVA. Data normality was evaluated using the Shapiro-Wilks test, homoscedasticity by the Brown-Forsythe's test and the Tukey's one degree of freedom test for non-additivity. The values for pFsample, pFassessor and QMerror, provided by ANOVA for each scale, were used respectively to evaluate discriminating power, expansiveness between assessors and data variability. The REGWF test was also applied to analyze the discriminative power of each scale. The results obtained from the ranking test were evaluated by Friedman's test and the ease of use of the scales by the Cochran-Mantel-Haenszel tests. The results indicated the superiority of the hybrid hedonic scale as compared to the structured hedonic and self-adjusting scales, both with respect to discriminative power and to data adequacy to the assumptions of normality and homoscedasticity. The self-adjusting scale presented a slightly greater discriminative power than the structured hedonic scale, despite the former having presented data with a greater variability and lack of normality of the residuais. Of ali the methods, the ranking test presented the least discriminative power. The structured and hybrid hedonic scales were considered to be signiticantly (psO.01) easier to use than the self-adjusting scale, there being no difference (pSO.O5) between these first two scales. Finally, the objective of the third experiment was to evaluate the performance of the nine-point hedonic, hybrid hedonic and self-adjusting scales in the segmentation of samples and consumers using Internal Preference Mapping methodology. One hundred and twelve consumers evaluated the overall acceptability of 8 commercial brands of red wine, the majority being Cabernet Sauvignon. The effects of presentation order -scales and samples- and carry over effects were balanced. The data were analyzed by ANOVA and MDPREF. Scale performance was evaluated using as criteria: number of significant dimensions in the MDREF (psO.O5), number of consumers significantly adjusted (psO.O5) and the degree of segmentation of the products and consumers. The results suggested a superiority of the hybrid scale over the traditional hedonic and self-adjusting scales. The MDPREF generated by the hybrid scale data produced the greatest number of significant dimensions (p=5%), yielding 79.5% of the consumers significantly adjusted (p=5%), while the MDPREF generated by the self-adjusting scale adjusted 54.5% of the consumers and that of the hedonic scale, 51.8%. Overall, the 9-point hedonic scale showed the worst performance in relation to the other scales examined. The results of this study strongly suggest that the hybrid hedonic scale is a valid and efficient tool for use in data collection associated with consumer studies, both when analyzed by normal models for the analysis of variance and by Internal Preference Mapping methodology / Doutorado / Doutor em Alimentos e Nutrição
|
Page generated in 0.7845 seconds