Using Hierarchical Agglomerative Clustering in E-Nose for Coffee Aroma Profiling: Identification, Quantification, and Disease Detection

Zhang and Deng [8] proposed employing an Electronic Nose Application to identify odors inspired by extreme learning machines. The self-expression model (SEM) and the extreme learning machine (ELM) are two real-time verified methods for detecting aberrant odor 96 target samples and 24 aberrant samples were used in training and testing. The scents of perfume, flowery water, and fruits were used in Real-Time Abnormal Odor without Target Odor. The intended odor recognition rate is 90.67%, whereas the anomalous odor recognition rate is 91.67%. Fang et al. [9] proposed an olfactory algorithm in all features based smart Electronic Nose Application. The technique mixes one-dimensional convolutional and recurrent neural networks. Deep-learning-enabled sensor array code signs and recognition algorithms may aid in meeting the growing need for a large number of highly specialized gas sensors. Deep-learning-enabled e-nose code sign paradigm that automatically learns target-specific features and optimizes sensor materials and sampling procedures iteratively.

Liu et al. [10] proposed wine properties detection using a MOS sensor and Machine Learning Techniques algorithm using an Electronic Nose Application. Odor detection includes the area of production, fermentation process, vintage years, and variants of four groups of test runs, the extreme gradient boosting (Xgboost), support vector machine (SVM), random forest (RF), and backpropagation neural network (BPNN) datasets were utilized, and the results were achieved 94% by analyzing wine properties for 450 in 6 matrices, 600 in 6 matrices, 450 in 6 matrices, or 600 in 6 matrices, correspondingly. Tan and Xu proposed a review of food quality with related properties for e-nose and e-tongue determinations. Feature extraction and classification used ANN and CNN based on the Electronic Nose Application. A coffee sample size is type typically and was collected and tested for training and validation [11].

Mu et al. [12] developed Machine Learning Techniques and algorithms for detecting milk sources and estimating milk quality. Milk fats and proteins are enhanced with an estimated accuracy of 95% using Dairy Herd Improvement (DHI) analytical data that collects data on dairy cows, including their milk production, health, and reproduction, and a five-fold cross-validation technique dividing the data into five-fold, and then training the model on four of the folds and testing it on the remaining fold. This process is repeated five times, and the results from the five tests are averaged to get an estimate of the model's performance based on milk odor features using an Electronic Nose Application. Xu et al. [13] proposed an approach for psychiatric symptoms and volatile organic compounds in trans-diagnostic samples using an Electronic Nose Application. 1200 participant’s characteristics were collected based on age, sex, smoking status, BMI, depression PHQ-90, Anxiety OASISd, Addiction DAST-10e, Race Ethnicity, and Education datasets were collected and processed using Machine Learning Techniques and data analysis. Hayasaka et al. [14] proposed a Machine Learning Techniques-based strategy based on three types of gas water (H₂O), methanol (MeOH), and ethanol (MtOH) in a single graphene FET using e-nose. MFC1 and MFC2 controlled the background R.H., while MFC3 controlled the target gas concentration. Controlling overfitting and performing cross-validation tests ensured accuracy.

Dadi et al. [15] suggested a sensor array reduction for detecting odor utilizing e-nose. The cross-validation approach is used to calculate performance using the usual leave-one-group-out and group shuffling methods. The increased odor recognition power of e-noses obtained 95% performance from training and cross-validation. Ye et al. [16] developed recent developments towards smart Electronic Nose Applications, a Machine Learning Techniques approach. Advanced Machine Learning Techniques and methodologies are required for the e-nose to improve its performance and capabilities in a variety of applications, including robotics, food engineering, environmental monitoring, and medical diagnosis. Using Machine Learning Techniques and algorithms improved both qualitative and quantitative analysis outcomes. A Machine Learning Techniques-based prediction approach for detecting coffee quality from green to roasted coffee beans was proposed by Suarez-Peña et al. [17]. The neural network algorithm is compared to support vector classification. The findings of the analysis of green coffee bean attributes were predicted as high or low with a classification shoehorn accuracy of 81% using 10-fold stratified cross-validation procedures.

Harsono et al. [18] advocated employing e-nose in Machine Learning Techniques to recognize odors in original Arabica civet coffee. This study focuses on a blend of Arabica civet coffee and Robusta coffee, which results in nine different mixing combinations. A statistical computation would be determined to obtain parameter statistics. Furthermore, the classification approach used in this work is intended to discriminate between original Arabica civet coffee and original Robusta coffee. The KNN technique yields the best results, with an accuracy rating of 97.7% for nine classes as indicated in Table 2.

Caporaso et al. [19] suggested a hyperspectral imaging-based prediction method for single-roasted coffee beans. The current work used hyperspectral imaging (1000-2500 nm) to predict volatile chemicals in single roasted coffee beans, as assessed by Solid Phase Micro Extraction-Gas Chromatography-Mass Spectrometry and Gas Chromatography Olfactometry. Individual beans can be separated, resulting in batches of coffee with varying volatile flavor element combinations.

Hendrawan et al. [20] and colleagues suggested utilizing Deep Learning approaches to detect and categorize purity levels in luwak coffee green beans. The study sought to detect and classify the purity of Luwak coffee green beans into four categories: deficient (0-25%), low (25-50%), medium (50-75%), and high (75-100%). Based on the training and validation data, GoogleNet was recognized as the best CNN model with optimizer type Adam and a learning rate of 0.0001, with an accuracy of 89.65%. García et al. [21] proposed using computer vision to detect quality and flaws in green coffee beans. Color, morphology, shape, and size are all critical indicators of high-quality beans. To determine the quality of coffee beans, the k-nearest neighbor approach was applied. AI-based vision technique for choosing high-quality coffee beans to reduce production time and increase quality control. Angeloni et al. [22] proposed an expresso coffee aroma profile based on its properties and odorants. The odor of coffee arabica and canephora is evaluated using volatile analysis and chemical attributes to determine the granulomere of the coffee particles and the brew ratio, which can modify the aroma profile of the beverage. The optimized extraction procedure may enable greener EC preparation and consumption by using less coffee powder to generate the same amazing output.

Table 2. Summary of coffee variety and aroma

Thanarajan et al. [23] proposed using an Electronic Nose Application to control coffee output quality by odor categorization. Instant coffee samples are collected and analyzed depending on temperature, concentration, and brand, achieving 97.1% overall odor uniqueness. Raveena and Surendran [24] proposed using ResNet50 to classify coffee cherries. Six coffee stages are used for categorizing coffee cherries depending on coffee type and color variation. Using Deep CNN models, the author achieved 99.01% accuracy. Lee et al. [25] proposed an artificial intelligence-based coffee aroma recognition-based fingerprint extraction method. Various fragrances of freshly roasted coffee were collected, and performance was calculated based on individual coffee aroma signatures of authentic quality. Dewangan et al. [26] developed a smell detection model based on Machine Learning Techniques code-based methods. The Chi-square feature extraction technique was used to pick the most significant features in each dataset. The Long-method dataset with the various selected sets of metrics produced the most accuracy (100%) for all five methods; however, the Max voting method produced the lowest accuracy (91.45%) for the Feature-envy dataset with the specified twelve sets of metrics. Zhang et al. [27] suggested a data preprocessing strategy based on polynomial curve fitting (PCF), locally weighted regression (LWR), wavelet package correlation filter (WPCF), and mean filter (MF) for recognizing and detecting coffee odor. The well-known Support Vector Machine (SVM) classifier has a classification accuracy of 96.23%.

Gonzalez Viejo et al. [28] proposed the intensity of coffee aroma profile using an integrated electronic nose at a low cost. Aromatic chemicals in coffee and their representation in a principal component analysis. Overall 270 datasets were trained and validated to evaluate the Performance was assessed based on mean squared error (MSE). Tanveer et al. [29] proposed a Green requirement engineering for mobile application development based on IoT for gathering required mobile challenges. Section 3 discusses the Materials and Methods in the roasting process, and working of Electronic Nose application detection for coffee aromas.

To facilitate the advancement of a real-time, high-performance coffee aroma detection system. The experimental setup utilized in this study involved the utilization of a gas chromatograph (GC) to separate the volatile compounds present in coffee aroma. Additionally, a mass spectrometer (MS) was employed to identify the different compounds. The entire process was controlled by a computer, which also facilitated the storage and analysis of the acquired data. Furthermore, software specifically designed for data analysis and machine learning Tensor Flow was employed to process the collected data.

4.1 Collection of datasets

Finding specific datasets on coffee odors can be challenging as the availability of such datasets may be limited 415 coffee odor datasets are collected. Coffee odor samples are collected from various coffee factories based on stages of maturity, texture, aroma, and consistency. The dataset is divided according to their aroma flavors Class-A has a Floral Aroma and Sweet Aroma, Class-B has Acidic Aroma, Sulphurus Aroma, Class-C has a Roasted, Burnt Aroma, Class-D has Woody Spicy Aroma, Class-E has an Earthy Woody Aroma, and Class-F has Nutty Caramel Aroma is tested manually in Indian coffee manufacturing to compare and listed the final samples as indicated in Table 4. Used online datasets such as Google Search, Kaggle, and open data platforms. Visited various sensor analysis database institutions and collected aroma-related descriptive analysis. Browsed relevant research papers in search of datasets based on aroma and a variety of aromas.

Table 4. Odor samples for coffee datasets

4.2 Experiments

In our trials, we use Python alongside the Keras and Tensor Flow frameworks in our model. The network was trained on a system with an Intel Core i5-11400 CPU running at 2.60GHz. The research was conducted utilizing an NVIDIA RTX A5000 GPU, a 64-bit operating system, and 16GB of RAM. The specifics are listed in Table 5. For coffee roasting and grinding blade grinder and drum roaster are used.

Table 5. Experiment required

4.3 Coffee odor-processing

The preprocessing technique improves the quality of information acquired via electronic signal processing. Noise, drift, and undesired abnormalities may be present in signals. Signal processing methods such as amplification and filtering are used to boost signal quality. Filtering techniques (such as low-pass, high-pass, and band-pass filters) can remove noise or undesired frequency components, while amplification helps to improve signal intensity. Baseline correction is used to eliminate any regular drift or offset in sensor signals. This entails estimating and subtracting a baseline value from the signals to bring them all to the same reference level. Calibration is required to ensure consistent and dependable performance. The sensors are calibration models, which can then be used to quantify or classify unknown samples. Thiols, which are often found in trace concentrations in the low ppb range, can contribute to sulfur-like or skunky odors. Concentrations of phenols can range from a few ppb to a few ppm, with substances such as Guaiacol contributing to Smokey and phenolic overtones. Fruity esters add to the aroma complexity of the coffee, with concentrations ranging from a few ppb to a few ppm as indicated. Acids: Organic acids such as acetic acid, formic acid, and quinic acid can be found in concentrations ranging from a few ppm to tens of ppm, contributing to the acidity and flavor character of coffee.

4.4 Principal Components Analysis

Principal Component Analysis (PCA) is an approach for reducing dimensionality and its goal is to reduce a high-dimensional dataset to a lower-dimensional space while retaining the most relevant information or patterns in the data. Each major component is a linear combination of the initial variables. The first principal component is responsible for the greatest possible variance in the data, the second for the second largest variance, and further on. Eqs. (1)-(3) show the mean and SD.

$\mu_{-} \mathrm{j}=(1 / \mathrm{n}) * \Sigma\left(\mathrm{H}_{-} \mathrm{ij}\right)$        (1)

In data standardization, if the variables in the dataset have different scales, it is common practice to standardize them to have zero mean and unit variance. This step ensures that variables with larger scales do not dominate the analysis. To calculate data standardization, mean and variables are observed where H_ij represents the value of the jth variable for the ith observation, n data points.

$\sigma_{\mathrm{j}}=\operatorname{sqrt}\left(\left(\frac{1}{\mathrm{n}-1}\right) * \Sigma\left(\left(\mathrm{H}_{\mathrm{ij}}-\mu_{\mathrm{j}}\right)^2\right)\right)$            (2)

To calculate data standardization, the standard deviation (σ) for each variable across all the observations. Subtract the mean from each variable value and divide it by the standard deviation to obtain the standardized value for each observation and variable. Assemble the standardized values into an n x m matrix called H. Each column of H represents a standardized variable, and each row represents an observation.

$\mathrm{H}_{-} \mathrm{ij}($ standardized $)=\left(\mathrm{H}_{-} \mathrm{ij}-\mu_{-} \mathrm{j}\right) / \sigma_{-} \mathrm{j}$        (3)    

In Covariance Matrix Calculation provides information about the relationship between variables and how they vary by comparing them together. Let H be the standardized data matrix with dimensions n x m, where n is the number of observations and m is the number of variables. Eq. (4) where X^T is the transpose of X

$\mathrm{C}=\left(\mathrm{X}^{\wedge} \mathrm{T} * \mathrm{X}\right) /(\mathrm{n}-1)$            (4)

The eigenvectors in Eigen composition represent the amount of variance explained by each principal component, and the eigenvectors in Eigen composition represent the direction of each principal component. The eigenvectors are ranked in principle component selection according to their corresponding eigenvalues. The major components are the eigenvectors with the highest eigenvalues that explain the most variance in the data. The original dataset is projected onto the selected principal components in dimensionality reduction to obtain the lower-dimensional representation. This projection involves taking a dot product between the standardized data and the eigenvector. The projected data matrix Z has dimensions’ n x k and can be computed as, where * denotes matrix multiplication.

$\mathrm{Z}=\mathrm{X} * \mathrm{~V}$            (5)

4.4.1 Principal Components Analytical Pseudo code

1. Function PCA (dataset, num_components).
2. Standardize the dataset by subtracting the mean and dividing it by the standard deviation for each variable.
3. Compute the covariance matrix C of the standardized dataset.
4. Perform Eigen decomposition on C to obtain the eigenvectors and eigenvalues.
5. Sort the eigenvectors based on their corresponding eigenvalues in descending order.
6. Select the top num_components eigenvectors to form the principal components matrix V.
7. Project the standardized dataset onto the principal components.
8. Return the transformed data, which represents the lower-dimensional representation of the dataset.

Transformed Data $=$ Dataset $*$V        (6)

Mean is the average of all the values in a set of data. It is calculated by adding all the values and dividing by the number of values. Standard deviation is a measure of how spread out the values in a set of data is. It is calculated by taking the square root of the variance. P-value is a probability value that is used to determine statistical significance. It is calculated by comparing the observed results to the expected results under the null hypothesis.

4.5 Model architecture selection

Hierarchical Agglomerative Clustering (HAC) is a bottom-up clustering algorithm that recursively merges data points or clusters to form a hierarchical structure of clusters. After completing the Principal Components Analysis statistical dataset is moved into the initialization process where each data point is set as an individual cluster. Calculate the distance between each pair of clusters using Euclidean distance which is used to calculate the distance between two points (A1, A2) and (B1, B2) in an Euclidean space, S represents distance Eq. (7).

$\mathrm{S}=\operatorname{sqrt}\left((\mathrm{A} 2-\mathrm{A} 1)^2+(\mathrm{B} 2-\mathrm{B} 1)^2+(\mathrm{C} 2-\mathrm{C} 1)^2\right)$         (7)

3.png

Figure 3. Working flow of proposed work

To find the two closest merging or linkage criteria merge closest is used. Four closest are compared and used such as Single linkage for calculating the smallest pair of distances, Complete linkage for calculating the maximum smallest pair of distances, Average Linkage for calculating the smallest average pair of distances, and Ward’s linkage. After merging, try to reduce the rise in total within-cluster variation. After merging two clusters, update the distance matrix to reflect the distances that exist between the newly formed cluster and the existing clusters. The distance between the merged cluster and another cluster can be determined via several methodologies, such as the minimum distance, maximum distance, or average distance between their respective members. Repeat the merging and distance matrix updating stages until all data points or clusters are merged into a single cluster or an initial number of clusters is reached. The merging process represents the relationships between clusters at different levels of granularity and is called a Hierarchical cluster tree also called a dendrogram. To determine the desired number of clusters, the dendrogram can be visually analyzed. The final clusters are determined by cutting the dendrogram at a specified height or by a distance criterion as indicated in Figure 3.

4.6 Training and validation

Datasets of coffee aroma are collected based on features and voltaic components. The dataset was preprocessed by cleaning, normalizing, and transforming the aroma features as needed. Based on different aroma level factors such as coffee origin, roast level, and brewing method using Hierarchical Agglomerative Clustering [31, 32]. The dataset is separated into training and validation to fit the Hierarchical Agglomerative Clustering for analyzing coffee aroma and the trained model to make predictions on the validation dataset.

4.6.1 Training set

Initially, the data preparation process involves splitting the available coffee aroma 415 samples into a training set, validation set, and test set 291:62:62 ratios. Model initialization involves assigning each data point to its cluster so that each data point is considered a singleton cluster. Firstly, the distance between each pair of data points is calculated using Euclidean distance in each iteration two clusters are merged into a new cluster until all data points are merged. Finally, create a cluster hierarchy that represents the history of merging clusters, allowing for different levels of granularity in the clustering solution. A training loop operates to acquire and train the data to forecast the accurate value for each epoch. Forward propagation is the process of computing output values by moving input data from the input layer to the output layer across the network's layers. The loss calculation function determines the loss value using the expected output and the true labels as input. The loss value represents the algorithm's current performance on the given input. Back propagation calculates the gradient loss concerning the model parameters again loss calculation is done to maintain the data accuracy as indicated in Figure 4.

4.png

Figure 4. Coffee variety and flavors

4.6.2 Validation set

The validation set helps monitor and hyper-fine-tune the models and is an intermediate between the training and test sets. Clustering-based validation set has multiple performance evaluation metrics that can be used depending on the specific task and the nature of the data. The measurement of performance can be clustered using External Measurements: Rand Index (RI), Jaccard Coefficient, Calinski-Harabasz Index, Normalized Mutual Information (NMI), and Fowlkes-Mallows index. Internal Measurements: Silhouette Coefficient, Dunn Index, and Davies-Bouldin Index.

Silhouette Coefficient: This metric assesses how well each data point fits within its given cluster compared to others. It has a value between -1 and 1, with higher values indicating stronger clustering. The average distance between one point and all other points in the same class. This score represents the closeness of points in the same cluster. The average distance between a sample and the nearest cluster's other points. This score computes the distance between points in various clusters, where, a (i) represents the intra-clustering distance and b (i) represents the inter-clustering distance.

Silhouette coefficient$=\mathrm{b}(\mathrm{i})-\mathrm{a}(\mathrm{i}) / \max \{\mathrm{a}(\mathrm{i}), \mathrm{b}(\mathrm{i})\}$            (8)

Calinski-Harabasz Index: This is a clustering evaluation metric that is used to evaluate the quality of clustering results. It calculates the ratio of between-cluster to within-cluster dispersion. The index attempts to capture cluster compactness and separation, where CH is the Calinski-Harabasz Index. B is the dispersion between clusters. W is the dispersion inside the cluster. The total number of data points is denoted by N. The number of clusters is k.

$\mathrm{CH}=(\mathrm{B} / \mathrm{W}) *((\mathrm{~N}-\mathrm{K}) /(\mathrm{k}-1))$             (9)

Davies-Bouldin Index: Clustering evaluation metric that calculates the average similarity across clusters by considering intra-cluster and inter-cluster distances. It provides a quantitative evaluation of the clustering results' quality. Where DB represents the Davies-Bouldin Index and k represents the number of clusters. R (i, j) represents the average distance between clusters I and j. The average intra-cluster distance of cluster i is represented by R (j, i).

$D B=(1 / k) * \epsilon(\max (R(i, j)+R(i, j))) f$              (10)

Dunn Index: The ratio of the minimum inter-cluster distance to the maximum intra-cluster distance. The greater the Dunn Index, the superior the clustering result, where min-inter is the shortest distance between any two clusters and max-intra is the shortest distance between any two data points within the same cluster.

Dunn Index = min_inter $/$ max_intra           (11)

Normalized Mutual Information (NMI): calculates the normalized mutual information between two sets of labels, such as the predicted cluster labels and the true class labels. It provides a measure of how well the clustering aligns with the true underlying structure of the data where MI is the mutual information between the two label sets, H1 is the entropy of the first label set, and H2 is the entropy of the second label set.

$\mathrm{NMI}=(2 * \mathrm{MI}) /(\mathrm{H} 1+\mathrm{H} 2)$        (12)

Rand Index (RI): Calculates the similarity measure between two data clustering, where w is the number of data points in both clustering that belong to the same cluster (true positives). H is the number of data point pairs that belong to different clusters in both clustering (true negatives). V² is the total number of data point pairs.

$\mathrm{RI}=\frac{\mathrm{w}+\mathrm{h}}{\mathrm{v}^2}$         (13)

Jaccard Coefficient: The similarity measure quantifies the overlap between two sets, where C denotes the set of data points assigned to a specific cluster by the clustering algorithm. D is a set of data points that belong to a given class or ground truth cluster.

$J C=(\mid$ number of elements in common $\mid)$ /|total number of elements|                 (14)

Fowlkes-Mallows index (FMI): calculates the measure of similarity between two clustering. It combines precision and recall to evaluate the agreement between the two clustering where TP (True Positives) represents the number of pairs of data points that are in the same cluster in both clustering (agreements). FP (False Positives) represents the number of pairs of data points that are in the same cluster in the clustering result but not in the ground truth clustering. FN (False Negatives) represents the number of pairs of data points that are in the same cluster in the ground truth clustering but not in the clustering result.

$\mathrm{FMI}=(\mathrm{TP} / \sqrt{(\mathrm{TP}+\mathrm{FP})} * \sqrt{\mathrm{TP}+\mathrm{FN}})$          (15)

The initialize Clusters function creates individual clusters for each data point using Python code.

Function initialize Clusters (data):

Clusters= []

For each data point in data:

cluster=create a cluster (data Point)

Clusters. Append (cluster)

Return clusters

The compute_Similarity-Matrix function computes the similarity between each pair of clusters and stores the values in a similarity matrix using Python code.

Function compute_Similarity-Matrix (clusters)

Similarity matrix=empty matrix of size (num Clusters, num Clusters)

For i=1 to num Clusters

For j I I+1 to num Clusters

Similarity=compute Similarity (clusters [i], clusters [j])

Similarity matrix [i, j] =similarity

Similarity matrix [j, i] =similarity

Return similarity matrix

Function compute_Similarity (cluster1, cluster2)

The find Closest_Clusters function identifies the pair of clusters with the minimum distance in the similarity matrix using Python code.

Function find Closest-Clusters (similarity matrix)

Instance=infinity

Merge Indices= (0, 0)

For i=1 to num clusters

For j=i+1 to num clusters

Similarity matrix [i, j] <instance:

Instance=similarity matrix [i, j]

Merge Indices = (i, j)

Return merge Indices

The merge_Clusters function merges the closest pair of clusters into a new cluster, removes the original clusters from the list of clusters, and adds the new cluster using Python code.

Function merge clusters (clusters, mergeIndices)

cluster1=clusters [merge Indices [0]]

cluster2=clusters [merge Indices [1]]

Merged Cluster=create a cluster ()

Merged cluster. Add (cluster1.getData ())

Merged cluster. Add (cluster2.getData ())

Clusters. Remove (cluster1)

Clusters. Remove (cluster2)

Clusters. Append (merged Cluster)

External Measurements (Rand Index-RI, Jaccard Coefficient-JC, Calinski-Harabasz Index-CHI, Normalized Mutual Information-NMI, and Fowlkes-Mallows Index-FMI) for predicted and true values, the predicted cluster assignments, and the true class labels or ground truth as indicated in Figure 5. The measurements for RI, JC, CHI, NMI, and FMI. The "Predicted Value" column represents the values obtained from the clustering algorithm, while the "True Value" column represents the values from the ground truth or true class labels.

5.png

Figure 5. Clustering measurements for external values

Table 6. External measurements for predicted and true value

Table 7. Internal measurements for predicted and true value in clustering

In the Rand Index, the true value is calculated as 0.82 after 1st iteration and the predicted value is 0.75, the true value is calculated as 0.78 after 2nd iteration, and the predicted value is 0.75, the true value is calculated as 0.72 after the 3rd iteration and the predicted value as 0.75, the true value is calculated as 0.82 after the 4th iteration and the predicted value as 0.75, the true value is calculated as 0.85 after 5th iteration and predicted values as 0.79.

In Jaccard Coefficient, the true value is calculated as 0.73 after 1st iteration, and the predicted value is 0.65, the true value is calculated as 0.71 after 2nd iteration and the predicted value is 0.66, the true value is calculated as 0.74 after 3rd iteration and predicted values as 0.66, the true value is calculating as 0.73 after 4th iteration and predicted values as 0.67, the true value is calculating as 0.73 after 5th iteration and predicted values as 0.65. Similarly, for JC, CHI, NMI, and FMI true value calculation is indicated in Table 6. Internal measurements such as the Silhouette Coefficient, Dunn Index, and Davies-Bouldin Index involve the comparison between predicted and true values as indicated in Table 7. Instead, they assess the quality of the clustering results based on the internal characteristics of the data and the clustering solution itself. Therefore, these measurements are typically presented in a table format comparing predicted and true values based on Volatile compounds and Principal Components Analysis.

4.7 Performance evaluation

Through the internal and external values calculation predicted and true values are observed using Volatile compounds and Principal Components Analysis. Performance is evaluated based on cluster models. Accuracy, Precision, Recall, and F1 score are calculated as shown in Eqs. (16)-(19).

Accuracy is the proportion of data points that have been appropriately grouped. It is computed as follows:

Accuracy $=\frac{\text { number of correctly data points }}{\text { total number of data points }}$             (16)

Precision is defined as the percentage of data points labeled as belonging to a specific cluster that belongs to that cluster. It is determined as follows:

Precision$=$ (number of correctly classified data points)/(total number of data points classified as belongs to that cluster)                    (17)

Recall is the percentage of data points labeled as belonging to a specific cluster that belong to that cluster. It is determined as follows:

Recall$=$ (number of correctly classified datapoints)/(total number of data points that actually belongs to that cluster)                 (18)

The F1 score is the harmonic mean of precision and recall. It is determined as follows:

$F 1$ score $=2 * \frac{\text { Precision } * \text { recall }}{\text { Precision }+ \text { Recall }}$            (19) 

Classified as groups and subgroups of coffee aroma patterns that were already known. These groups could be classified on the kind of coffee, roasted level, and smell, and whether or not there are any diseases. These analyzed coffee flavor profiles into groups by using a method called hierarchical agglomerative clustering (HAC). They matched up the results of clustering with the known groups and subsets. The Jaccard index between the two sets of names was used to figure out how to do this. The Jaccard index is a way to figure out how much two sets are alike as shown in Eq. (14). Section 5 discusses the Results Analysis based on Principal Components Analysis for coffee aromas.

The primary objective of this study is to classify the aroma of coffee by considering both the roasting technique and the specific variety of coffee beans. In previous studies conducted before the roasting step, coffee berry stages were categorized to track the subsequent development of coffee aroma and identify any potential diseases. This study highlights the challenges of detecting many coffee aromas utilizing a Hierarchical Agglomerative Algorithm-created E-nose prototype. Principal Components Analysis compares organic volatile chemical aroma samples from different coffee aromas. Inspired by Machine Learning, external and internal parameters categorize coffee aroma. E-Nose's qualitative as well as quantitative performance and dependability have improved significantly after applying Machine Learning Techniques. The performance of testing samples was compared by accuracy. (1) Rand Index (RI), Jaccard Coefficient, Calinski-Harabasz Index, NMI, and Fowlkes-Mallows Index create external measures. (2) The Silhouette Coefficient, Dunn Index, and Davies-Bouldin Index are used to calculate internal measures and compare external and internal accuracy. Performance is evaluated based on cluster models. Accuracy, Precision, Recall, and F1 score achieved 97.08% accuracy in finding accurate coffee aroma. The future strategy will focus on coffee flavor with quality measurement in various stages of coffee progression using AI agents. This rigorous look at coffee aroma analysis and factory control can help manufacturers standardize operations, save money, and defend consumer rights.