Browsing by Author "Kumarika, B. M. T."

Now showing 1 - 5 of 5

Development of rapid detection strip for amines from other organic functional groups
(Faculty of Science, University of Kelaniya Sri Lanka, 2024) Ravindu, M. A. Y.; Maheshani, Y. K. D. C.; Kumarika, B. M. T.; Wanniarachchi, D. D. C. de S.
Identifying organic compounds in a laboratory requires a lot of chemicals and hence, the process is expensive. To address the challenges of controlling costs and reducing chemical waste, an investigation into the integration of chemistry with computer science techniques has been initiated. This approach emphasizes the significance and innovative aspects of the research. The research focuses on predicting Organic Compounds using both color strips and machine learning methods. A disposable strip was designed with ten separate holes, each serving as a colorimetric indicator. The first hole does not contain any chemical, from the second hole FeCl3, Chromic Acid, CuCl2, FCP, Methyl Orange, Phenol Red, Bromophenol Blue, Thymol Blue, Bromocresol Green were in holes respectively. These sensor indicators react with Functional group, causing distinctive color changes. RGB values from colorimetric strips were extracted as the dataset using ImageJ, an image analysis software, which analyzed photos of the sensor strip to obtain the RGB values for each hole. Two methods were used to classify compounds. Initially, the dataset containing RGB values of every compound was subjected to Principal Component Analysis (PCA) to evaluate the sensor array's intrinsic capacity for distinguishing between distinct categories of organic compounds. Second, specific chemicals were categorized using their RGB profiles because of the development of machine learning algorithms. It was shown that alcohol, ester, aldehyde, ketone, carboxylic acid could not be effectively separated using a single-color value (red, green, or blue) using PCA. But in the green value PCA plots, amines frequently formed unique clusters that allowed for their independent identification. Using PCA-derived green values, the K-Nearest Neighbors (KNN) model proved to be the most effective among all models for classifying chemicals as amines or non-amines, with an accuracy of 94%, recall of 95%, and precision of 95%. The KNN model achieved 99% training accuracy by adding additional amine and non-amine chemicals (27 and 26 respectively) to the training dataset. This study demonstrates the potential of RGB data for chemical identification, particularly for amines, suggesting that the colorimetric sensor array can be used as an identification strip for amine compounds in environmental samples and for educational purposes. Clustering mixtures like Carboxylic Acid-Aldehyde, Alcohol-Ketone into different categories was shown to be a substantial issue. Mixtures' color patterns frequently matched the dominating component's (amine in an amine-alcohol-ester combination, for example). This shows that in complex samples, clear categorization is made difficult by solvent effects or inter-component interactions.
Machine learning analysis of colorimetric sensor arrays for amine classification
(Faculty of Science, University of Kelaniya Sri Lanka, 2024) Maheshani, Y. K. D. C.; Ravindu, M. A. Y.; Wanniarachchi, D. D. C.; Kumarika, B. M. T.
This research explores the application of machine learning for the classification of amines using colorimetric sensor arrays. Colorimetric sensor arrays are chemical sensors that detect substances through color changes, offering simplicity and ease of use for rapid on-site analysis. The identification of amines plays a crucial role in various fields due to their significant impact on health, safety, environmental quality, and industrial efficiency. In particular, the fish industry relies heavily on the identification of amines to ensure product quality and safety. Amines are often associated with spoilage and degradation in fish products. In this study, the identification of the presence or absence of amines in colorimetric sensor array images was performed using machine learning. The sensor array consists of several metal-based and acid-base indicators. Colorimetric sensor array images in strip format were utilized. The images, of varying quality, were in JPEG/JPG and PNG formats. The RGB values were extracted from 76 images of 10-hole strips and 38 for each class: Amine, and Not Amine. The dataset was split randomly into two subsets: 75% for the training set and 25% for the validation set. A separate dataset was used to test the model. This study explores various machine learning models, including Knearest neighbors (KNN), decision tree, and support vector machine (SVM). The KNN model achieved promising results with an accuracy of 94.74% on both training and validation datasets, an average precision of 95%, an average recall of 95%, and an average F1 score of 95% demonstrating effective classification capabilities across both training and validation sets. The confusion matrix method was employed to evaluate the model's performance. Hyperparameter tuning was conducted to optimize model performance and to avoid overfitting and underfitting, techniques such as selecting the optimal number of neighbors, and choosing the appropriate distance metric were employed. A user-friendly web prototype was developed to demonstrate the practical application of the model. After building the prototype, the model was tested on previously unseen data using a web prototype. The model achieved 90% accuracy on this new data. Future work aims to expand the dataset and include additional compounds to enhance the model's robustness and utility. This study highlights the potential of machine learning in advancing the detection and classification of chemical substances, contributing to various fields requiring precise amine identification. This makes it possible for people in different areas to use technology and understand the results without having to be experts.
Real-time system for place recognition by interpreting Sri Lankan sign language into text using machine learning approach.
(Faculty of Science, University of Kelaniya Sri Lanka, 2023) Perera, J. D. K. N.; Kumarika, B. M. T.
Sri Lankan sign language (SSL) serves as a vital visual-gestural communication system for the deaf community in Sri Lanka. However, effective communication between hearing-impaired individuals and the general population is limited due to challenges in understanding sign language. Further interpreting dynamic signs is more challenging due to the complexity involved in the sequence of unique expressions in SSL. To address this issue, a novel SSL to Sinhala text interpreting technology was developed, focusing specifically on dynamic signs associated with Sri Lankan locations. The research dataset encompassed a dynamic sign dataset comprised of 30 videos per category, each precisely divided into 30 frames. This robust dataset further strengthened the effectiveness of our approach in enhancing location-specific dynamic sign recognition. This study contributes to bridging a critical gap in Sri Lankan sign language recognition by assessing our model's performance on dynamic signs across three and five distinct locations. To recognise these dynamic gestures, a vision-based approach was chosen, providing a simpler and cost-effective solution compared to sensor-based systems. The study integrated Media Pipe and a Long short-term memory (LSTM) neural network as part of a combined methodology to enable gesture detection and interpretation. By leveraging these techniques, a camera-based, low-cost solution was successfully developed for interpreting SSL’s dynamic gestures. The study systematically tested multiple models, tuning LSTM and dense layers with varying neurons, resulting in an optimal model. Following rigorous 50-epoch training, our model exhibited an accuracy of 98.89% for dynamic signs across three distinct locations and an accuracy of 93.33% for dynamic signs across five locations. Cross-validation techniques were employed to assess the system’s performance and ensure its generalizability across different datasets. By validating the system through cross-validation, its robustness and reliability were tested, enabling a more accurate interpretation of SSL’s dynamic gestures. The proposed SSL to Sinhala text interpreting technology has the potential to significantly improve communication between hearing-impaired individuals and the general population in Sri Lanka. By leveraging vision-based methods and incorporating dynamic gesture recognition, this technology can enhance the accessibility and inclusivity of communication for the deaf community. Further research and enhancements are being carried out to expand the system’s capabilities for more place recognition and address the challenges associated with dynamic gestures and facial expressions in SSL recognition.
Smart Cricket: Strategic batting place prediction and player ranking with ensemble learning
(Faculty of Science, University of Kelaniya Sri Lanka, 2024) Rajapaksha, R. A. R. S.; Kumarika, B. M. T.
Cricket is a global sport with three formats, of which T20 is the fastest format. A team’s batting order decides the outcome of the match as each player must play according to their position. Therefore, assigning roles to batsmen according to their position is crucial. Previous studies have focused on win prediction, team selection, player performance evaluation, and player classification using machine learning. However, there appears to be a lack of research on classifying batsmen according to their positions and ranking players within each category. This study aims to classify players based on their roles in the batting order, as top-order, middle-order, lower-order, or tail-ender batsmen in the T20 format, and rank players within each category using classification probabilities. Data were collected from www.cricinfo.com, encompassing 347 players from countries that participate in Test-format cricket. For this study, 15 batting features were used for model building. Additionally, the dataset was split into training and testing sets, with 75% of the data used for training and 25% for testing. The training set was used for model development and the testing set for performance evaluation. Initially, classification was done using machine learning models such as Naïve Bayes, Random Forest, Decision Tree, SVM, and KNN. Then, ensemble learning techniques including boosting, voting, and stacking on different combinations of base models were employed to improve model performance. The best performance was observed with a stacking model that had seven base models. This model had a training accuracy of 97.69% and a testing accuracy of 95.40%. It also had a precision of 95.16%, recall of 95.35%, and F1-score of 95.23%. The findings of this study show that ensemble models can classify batsmen accurately according to their positions and rank players within each category. This classification and ranking system is highly valuable for selectors, coaches, captains, and team management for team selection, forming batting lineups, and tasks like strategic player replacement. Also, this study is beneficial for players to understand and improve their performance. Furthermore, this classification is helpful in franchise leagues, particularly for player selection in auctions. Additionally, this study proves that using ensemble models with different techniques and multiple base models can significantly enhance predictive performance. Future work will focus on improving the predictions with scheduled and continued training.
Unlocking the potential of convolutional neural networks for precise classification of finger pulse waves in diabetic patients and healthy individuals
(Faculty of Science, University of Kelaniya Sri Lanka, 2023) Gunathilaka, P. A. D. H. J.; Kumarika, B. M. T.; Jayathilaka, K. M. D. C.; Perera, D.; Liyanage, J.A.; Kalingamudali, S. R. D.
Pulse wave analysis (PWA) is a valuable technique for assessing the cardiovascular health of diabetic patients. However, it encounters several challenges, including the complexity of pulse wave signals and the need for standardization and validation of measurement methods. Convolutional Neural Networks (CNNs) play a crucial role in addressing these challenges by offering a robust and accurate approach to classifying pulse wave images. Pulse wave analysis offers a cost-effective, time-efficient, highly accurate, and non-invasive method for diagnosing diabetes-related cardiovascular issues. This study aims to investigate the effectiveness of CNN in classifying finger pulse wave images to accurately distinguish between diabetic and non-diabetic subjects, thus enabling non-invasive diabetes diagnosis. The study's methodology comprises four main steps: data collection, data preprocessing, CNN model development, and model evaluation. Primary data, including finger pulse waves, blood pressure, mean arterial pressure, oxygen saturation, and pulse rate, were acquired from the multipara patient monitor. Subsequently, single pulse wave cycles from 50 healthy individuals and 50 diabetes patients were subjected to preprocessing. The CNN model was developed through data collection, preprocessing, and the creation of its architecture, followed by compilation, training, and evaluation, ultimately achieving a 92% accuracy in classifying pulse wave images for non-invasive diabetes diagnosis. Descriptive statistics were used to summarize participants' demographic and clinical data, revealing no significant differences in age, gender, or body mass index between the two groups. The model's ability to discriminate based on pulse wave images highlights its potential for noninvasive diabetes diagnosis. In order to improve accuracy in future work, increasing the dataset size and conducting hyperparameter tuning will be essential for optimizing the CNN model.