Farklı çikolataların reolojik, tekstürel, duyusal özelliklerinin ve diğer bazı kalite kriterlerinin yapay sinir ağları ile tahminlenmesi

Prediction of rheological, textural, sensory properties and some other quality criteria of different chocolates using artificial neural networks

PDF İndir

Tez No: 1017646
Yazar: NURCANAN AYDIN
Danışmanlar: PROF. DR. FİLİZ ALTAY
Tez Türü: Doktora
Konular: Gıda Mühendisliği, Food Engineering
Anahtar Kelimeler: Belirtilmemiş.
Yıl: 2026
Dil: Türkçe
Üniversite: İstanbul Teknik Üniversitesi
Enstitü: Lisansüstü Eğitim Enstitüsü
Ana Bilim Dalı: Gıda Mühendisliği Ana Bilim Dalı
Bilim Dalı: Gıda Mühendisliği Bilim Dalı
Sayfa Sayısı: Belirtilmemiş.

Özet

Bu çalışmada, makine öğrenmesinin bir alt dalı olan yapay sinir ağları ile farklı çikolata örneklerinin kompozisyonlarına bağlı olarak sertlik, viskozite, renk, erime özelliklerini ve duyusal parametre skorlarını yüksek doğrulukla tahmin edebilen model geliştirmek amaçlanmıştır. Araştırma, başlıca dört aşamadan oluşmuştur. Birinci aşamada, farklı kompozisyolara sahip yerli ve ithal markalardan oluşan beyaz, sütlü ve bitter çikolata örnekleri pazaryerlerinden temin edilmiştir. Etiket bilgilerine göre, çikolata bileşenleri varlık/yokluk durumlarına göre 1-0 şeklinde kodlanmış ikili değişkenler olarak, toplam yağ ve toplam kakao katıları oranları da sürekli değişkenler olarak modelde kullanılacak datasetinin giriş verileri oluşturulmuştur. İkinci aşamada, her bir örnekte tekerrürlü olmak kaydıyla sertlik, viskozite, renk ve erime parametreleri, duyusal analizleri gerçekleştirilmiştir. Elde edilen deneysel sonuçlar ile modelde kullanılacak datasetinin çıktı verileri oluşturulmuştur. Çikolata kompozisyonları ile elde edilen deneysel veriler arasında istatistiksel anlamlılık ve korelasyon değerlendirilmiştir. Üçüncü aşamada, yapay sinir ağları eğitim sürecinde çeşitli öğrenme algoritmaları ve ince ayar hiperparametreleri deneme/yanılma yoluyla özelleştirilmiş bir Matlab kodu kullanılarak performansları değerlendirilmiştir. Dördüncü ve son aşamada ise, en iyi eğitim ve tahmin performansını sağlayan hiperparametreler belirlenerek modelin başarısı ve güvenilirliği çeşitli testler ile değerlendirilmiştir. Elde edilen sonuçlara göre, en güvenilir ve stabil eğitim süreci Levenberg-Marquardt eğitim algoritması ile elde edilmiştir. Bu algoritma ile eğitim sonucunda, geliştirilen sinir ağı modeli yalnızca çikolata kompozisyonuna bağlı olarak sertlik ve viskozite değerlerini, renk ve erime parametrelerini, duyusal skorları sırasıyla 0,95, 0,97, 0,95, 0,84, 0,23 R2 değerleri ile tahmin edebilmiştir. Model, kompozisyon ile çikolatanın sertlik, viskozite, erime ve renk değerleri arasındaki ilişkiyi başarılı bir şekilde yakalayabilmiş ve güvenilir tahmin performansları sergilemiştir. Bunun yanısıra, duyusal analiz sürecinde panelistlerin subjektif puanlamaları modelin kompozisyon ile duyusal skorlar arasındaki ilişkiyi kaçırarak bu farklılıkları gürültü olarak nitelendirmesine neden olmuştur. Bu nedenle, duyusal analiz skorlarının tahmini için model iyileştirmeleri gerekmektedir. Elde edilen sonuçların, yeni çikolata formülasyonlarının hızlı bir şekilde değerlendirilmesini sağlayarak Ar-Ge ve üretim süreçlerinde endüstriye katkı sağlama potansiyeli bulunmaktadır.

Özet (Çeviri)

This study aims to develop a model that can predict the hardness, viscosity, color, melting properties, and sensory parameter scores of different chocolate samples based on their compositions with a high degree of accuracy, using artificial neural networks, a subfield of machine learning. The research consisted of four main stages. In the first stage, white, milk, and dark chocolate samples with different compositions from domestic and imported brands were obtained from the market. According to the label information, the input dataset to be used in the model was created as binary variables coded as 1-0 according to the presence/absence of chocolate components, and total fat and total cocoa solids percentages were created as continuous variables. In the second stage, hardness, viscosity, color, and melting parameters, as well as sensory parameters, were analyzed by performing iterations for each sample. The output dataset to be used in the model was created from the obtained experimental results. Statistical significance and correlation between chocolate compositions and the obtained experimental data were evaluated. In the third stage, the prediction performance of the artificial neural networks was evaluated using a customized Matlab code that included various learning algorithms and fine-tuning hyperparameters through trial and error during the training process. In the fourth and final stage, hyperparameters that provided the best training and prediction performance were identified, and the success and reliability of the model were evaluated through various tests. The dataset used in the model ranges from 102 to 272 elements. To eliminate the disadvantages of working with small datasets, the dataset was divided into training, validation, and test sets. A 5-fold cross-validation technique was used on the validation dataset. This ensured more generalizable results compared to a single training-test split. Furthermore, the entire validation set was evaluated in the model as both a training and test set. A multilayer, feedforward, backpropagation artificial neural network was used in the modeling of the chocolate quality parameter prediction algorithm. For this purpose, artificial neural network architectures with at least one hidden layer were developed, and between 1 and 20 neurons were tested in the input layer. The hyperparameters (epoch number, learning rate, momentum, etc.) that produced the best training and validation results were determined using the Random Search technique. Seeds were added to the model to ensure consistent results each time. The entire dataset entering the model was normalized from min. to max. to bring it to the same plane between 0 and 1. Therefore, tansig and logsig were chosen as activation codes in the hidden layer, and purelin in the output layer. The performance of the traininglm, trainbr, traincfg, trainoss, trainingd, and trainingdm learning algorithms, known to be more compatible with small datasets, was evaluated. The chocolate input components with high correlation to quality parameters were identified using feature selection techniques. This dimensionality reduction process aimed to prevent inflated and misleading results. For this purpose, dimensionality reduction was applied separately to each quality parameter, and modeling was performed with the components that affected the output. The prediction success of the developed models was determined using metrics. The R2 metric, which expresses how well the model represents the structure in the data, and the MSE (mean squared error), which expresses the difference in error between the actual value and the predicted value, were evaluated. Accordingly, the model with the highest R2 value and the lowest MSE value was labeled as successful. In addition, to thoroughly evaluate the model's success and reliability, learning curve, calibration plot, error/residual plot, uncertainty plot at a 95% confidence interval, and permutation plots were obtained. Relative importance analysis plots showing how much weight the model gives to each component when making decisions, and sensitivity analysis plots showing how effective/sensitive even the smallest change in inputs is on the output were also obtained. With these plots, it was observed to what extent the developed model could capture the interaction between real-world chocolate input components and quality parameters. In all quality parameters, the Levenberg-Marquardt training algorithm yielded the most reliable and stable training process among the applied learning algorithms. The neural network model developed with this algorithm successfully predicted the hardness values of chocolate samples based solely on composition with high accuracy, achieving 95% R2 and 0.03 MSE values. The topology providing these values has 8 neurons in the input layer, 10 neurons in a single hidden layer, 1 neuron in the output layer, and logsig activation. The model, exhibiting a balanced learning curve, reached very small error rates in the first few iterations and stabilized. The learning and validation curves showed parallelism and did not exhibit overfitting risk. Model predictions are consistent with the calibration curve. A few outliers observed in the test set are thought to be due to extreme compositions. The model's predictions have a very narrow band range in the uncertainty graph with a 95% confidence interval. The model was able to detect outliers within the confidence interval. Permutation testing showed that the model's success was not coincidental. According to the relative importance analysis table, the model assigned the largest weights to total fat percentage, total cocoa solids percentage, and whey powder inputs when predicting hardness values. The sensitivity analysis table indicates that changes in total fat percentage, total cocoa solids percentage, and cocoa mass components have an effect on hardness. The model predicted viscosity values with high accuracy, with 97% R2 and 0.009 MSE values, depending on the composition. The topology that provided these values has 13 neurons in the input layer, 4 neurons in a single hidden layer, 1 neuron in the output layer, and logsig activation. Having a balanced learning curve, the model achieved very small error rates in the first few iterations and stabilized. Although the learning and validation curves show parallelism, it was observed that the validation curve tends to rise after the 25th iteration. The early stopping technique applied at this point eliminates the risk of overfitting. Model predictions are consistent with the calibration curve. Two outliers observed in the test set are thought to be due to extreme compositions, deviations from the standard recipe-viscosity range, and smoothness. In the uncertainty prediction range graph, outliers observed in the calibration graph are noticeable. Values outside these outliers remain within the confidence interval, indicating that the model makes predictions confidently. Model predictions did not exhibit systematic error. However, it is observed that the model tends to underestimate and overestimate the actual values in two respects. The permutation plot shows that the model's success is statistically significant. According to the relative importance analysis table, the model gives the highest weight values to the total fat percentage, total cocoa solids percentage, and whey powder inputs when predicting viscosity values. According to the sensitivity analysis table, it is understood that changes in total fat percentage, total cocoa solids percentage, and cocoa mass components have an effect on viscosity. The model has a high success rate in predicting hardness and viscosity values, especially depending on the chocolate composition, and has captured the physicochemical interactions of the components in chocolate. The model predicted color values with high accuracy, with 95% R2 and 0.001 MSE values, depending on the composition. The topology that provided these values has 10 neurons in the input layer, 12 neurons in a single hidden layer, 3 neurons in the output layer, and logsig activation. According to the model's learning curve, training and validation errors decreased after the first two iterations. However, it is observed that the validation error increased after the 4th iteration. Early stopping technique is important in these cases. The model has a high prediction rate for the L*, a*, and b* color parameters. Although there is no systematic error, it has been observed that the model's success is not coincidental. Relative significance analysis shows that the model gives high significance scores to the total fat and cocoa mass components in predicting color parameters. However, sensitivity analysis showed that total fat, cocoa mass, and whole milk powder components have an effect on color parameters. The model was able to predict the melting parameters of chocolate based on composition with 84% R2 and 0.014 MSE values. The topology that provided these values has 12 neurons in the input layer, 8 neurons in a single hidden layer, 4 neurons in the output layer, and logsig activation. According to the learning curve, in the first 5 iterations, the model minimized the error very quickly, rapidly learning the relationship between composition information and melting parameters. After the 8th iteration, validation and training errors balanced out. The observed in the validation error suggest that there may be samples with very different thermal characteristics. The model has a high prediction rate for the Tonset, Tpeak, and ΔH melting parameters. However, this rate remained at 60% for the Tend parameter prediction. The end of melting (Tend) is the point at which the highest melting point crystals in the chocolate matrix melt. This stage is significantly affected not only by the composition but also by the chocolate's storage process and tempering success. This caused noise in the dataset structure, affecting the model's performance. Total fat and total cocoa solids have a dominant relative importance on all melting parameters and influence them positively. Changes in these components alter the density of the crystal lattice structure, thus dominating all melting parameters. Components that lower melting temperatures, such as whole milk powder, indicate a eutectic effect in chocolate. Total cocoa solids (%) and total fat (%) are the most dominant components with the highest sensitivity scores on all melting parameters. According to the sensitivity analysis, dairy components play a decisive role, especially in the later stages of melting. The model has an R2 of 23% and an MSE of 0.036 in predicting sensory parameter scores based on composition. The topology providing these values has 10 neurons in the input layer, 20 neurons in a single hidden layer, 11 neurons in the output layer, and logsig activation. 20 neurons make the model complex, and overlearning issues may arise. However, the training R2 of 0.40 and the test R2 of 0.23, along with the difference between the validation error and training error curves, while indicating a slight risk of rote learning, suggest that the main problem is insufficient learning. The model's training R2 being around 0.4 indicates that the model architecture or data is inadequate for this problem. The 10 features selected through feature selection and dimensionality reduction may not provide sufficient information to explain the relationship between chocolate composition and sensory parameters. Furthermore, the subjective scoring of the panelists during the sensory analysis process caused the model to overlook the relationship between composition and sensory scores, classifying these differences as noise. Instrumental measurements, expert panelist ratings, and various machine learning techniques can be used to estimate sensory analysis scores. The results obtained have the potential to contribute to the industry's R&D and production processes by optimizing the need for hundreds of trials in new product development, enabling the evaluation of chocolate formulations in seconds in a fast and non-destructive way.

Benzer Tezler

Tez No
592508
Effects of some soy products on rheological, functional and sensory properties of milk chocolate
Bazı soya ürünlerinin sütlü çikolatanın reolojik, fonksiyonel ve duyusal özellikleri üzerine etkileri
İZLEM CANSU MECLİS
Yüksek Lisans
İngilizce
2019
Gıda Mühendisliği Hacettepe Üniversitesi
Gıda Mühendisliği Ana Bilim Dalı
DR. ÖĞR. ÜYESİ ELİF YOLAÇANER
DOÇ. DR. HALİL MECİT ÖZTOP
Tez No
152
Buğdayda farklı melezleme teknikleri kullanarak tohum tutma oranının saptanılması
Başlık çevirisi yok
İSMAİL TÜZÜN
Yüksek Lisans
Türkçe
1986
Ziraat Uludağ Üniversitesi
Tarla Bitkileri Ana Bilim Dalı
PROF. DR. HALİS RUHİ EKİNGEN
Tez No
252
İzmir'de bazı yolların kapatılarak yaya bölgeleri oluşturma da kent peyzajını geliştirme açısından yeniden planlanması üzerinde araştırmalar
Başlık çevirisi yok
BAHAR TÜRKYILMAZ
Yüksek Lisans
Türkçe
1985
Şehircilik ve Bölge Planlama Ege Üniversitesi
Şehir ve Bölge Planlama Ana Bilim Dalı
Tez No
102
Diyabetus mellitus'da gözlenen mikrovasküler komplikasyonların prostaglandinler ile olan ilişkisi
The Relation of microvaskular complications observed in diabetes mellitus with prostoglandins
CANAN NEBİGİL
Doktora
Türkçe
1985
Eczacılık ve Farmakoloji Cumhuriyet Üniversitesi
Farmakoloji Ana Bilim Dalı
YRD. DOÇ. DR. YUSUF SARIOĞLU
Tez No
202
Bazı esterlerin kinetik incelenmesi ve termodinamik parametrelerin belirlenmesi
Kinetic study and detesmination of thermodnamic parameters of some esters
İBRAHİM TAŞ
Yüksek Lisans
Türkçe
1986
Kimya Mühendisliği Uludağ Üniversitesi
Kimya Ana Bilim Dalı
PROF. DR. MUSTAFA CEBE

Geri Dön