Browsing by Author "Peng, Wei"

Now showing 1 - 20 of 61

Open Access
A decision-making system for medical transportation mode using machine learning methods
(Faculty of Graduate Studies and Research, University of Regina, 2023-09) Khodabakhshi, Sahar; Peng, Wei; Stilling, Denise; Wang, Zhanle (Gerald); Zeng, Fanhua
There is number of complicated operations in freight transportation system to cover customer demands in the world. Nowadays, companies have huge competition to fulfill customer needs and get a higher level of performance in freight transportation. Transportation mode has been considered as one of the components that has influence on service levels of freight transportation. Road, sea, air are popular modes of transportation which have different features and unique benefits. They also have various costs, different emissions in environment and risks in society. People use these transportation modes based on their needs, but they do have some advantages and disadvantages. When we are dealing with a lot of shipment transactions in a company, making decision for choosing the best option will not be easy. Companies face with challenges since there are numerous factors effecting shipment mode selection. Moreover, the number of low-volume and high-frequency shipments has also increased due to increased demand diversity, shorter product life cycles, and increased agile customer response. Consequently, logistics costs are increasing for those shippers who need to export a small number of products abroad. As a result, researchers have been actively focusing on this matter, which has a significant impact on a country's social and economic situation. This research aims to develop a hybrid approach to create a shipment selection model with a case study of pharmaceutical drugs by machine learning algorithms, checking the accuracy of predictions and using multi-criteria decision-making methods (MCDM) for validation of our work. Several different features of the dataset including shipping cost, country of origin and destination, cargo weight, cargo dimensions, etc., are given to decision tree, Random forest, logistic regression, XGboost and SVM machine learning algorithms so that we can predict the best shipping method by land, air, or sea. Then, using different criteria F1 score, Recall and precision, accuracy score we measured the accuracy of the forecast and finally, we validated the research method by MCDM methods SAW, MARCOS, TOPSIS, MULTIMOORA and VIKOR. After being familiar with all important factors, tools, research gap in literature review, we realized that choosing one machine algorithm is not enough to get an accurate result and we used the most popular ones. data science scored important features influencing transportation modes and used machine learning techniques to learn the factors and the relationships between them to increase the accuracy of the pharmaceuticals drugs shipment selection system. By MCDM we found XGboost as the best machine learning algorithm to predict the shipment mode with the average performance evaluation of 84 percentage then random forest, decision tree, SVM and LR respectively.
Open Access
Absorption capacity of carbon dioxide in aqueous solution of 1,2-bis(3-aminopropylamino) ethane and Dytek EP diamine: Experimental measurements and simulation with the E-NRTL model
(Faculty of Graduate Studies and Research, University of Regina, 2024-12) Fallah, Abbas; Henni, Amr; Peng, Wei; Khan, Sharfuddin
The increasing threat of climate change has elevated the importance of carbon dioxide (CO2) capture technologies. This thesis explores the solubility of CO2 on aqueous solution of two novel amines 1,2-Bis(3-AminoPropylamino) Ethane and Dytek EP diamine at two different temperatures of 313.15 K and 333.15 K, and two different concentrations of 10 wt% and 30 wt%. These amines were selected for their potential to enhance CO2 absorption efficiency and reduce energy consumption in carbon capture and storage (CCS) processes to provide valuable data for developing more efficient CO2 capture systems. Utilizing the Electrolyte Non-Random Two-Liquid (eNRTL) model for the liquid phase and the RK equation of state for the gas phase, the research includes extensive thermodynamic modelling to simulate the experimental data and predict the behaviour of these amines in CO2 capture processes. The binary e-NRTL and molecule–ion pair parameters were obtained by regression. The overall percentage of the average absolute deviation (%AAD) between the experimental and estimated values for the temperature, pressure, and mole fractions are 0.006%, 0.052% and 0.015%, respectively, for 1,2-Bis(3-AminoPropylamino) Ethane, and similarly, 0.197%, 0.093%, 0.105% for Dytek EP diamine. 1,2-Bis(3-AminoPropylamino) Ethane showed superior solubility performance concerning other amines studied in the literature due to its high molecular weight and four amine groups in its structure, which increased its reactivity and decreased its steric hindrance. Dytek EP diamine had a moderate performance due to its lower molecular weight and the presence of only two amino groups and a methyl group in the structure, creating a steric hindrance and decreasing its capacity.
Open Access
An intelligent system approach for predicting the risk of heart failure
(Faculty of Graduate Studies and Research, University of Regina, 2023-08) Raihan Khan Rabbi, Imran; Peng, Wei; Mehrandezh, Mehran; Khondoker, Mohammad; Jia, Na
Heart failure is a chronic, progressive condition in which the heart muscle is unable to pump enough blood to meet the body’s needs for blood and oxygen. It is a severe and long-term condition and there are several complications from heart failure that include irregular heartbeat, sudden cardiac arrest, heart valve problems, pulmonary hypertension, kidney damage, liver damage, malnutrition etc. According to the World Health Organization (WHO), the number one cause of death in cardiovascular diseases (CVD) is estimated at 17.9 million a year, which accounts for 31% of all deaths worldwide. The majority of heart patients are diagnosed at a stage of high risk since the early screening and diagnosis of any heart disease is complicated and the particular medical exams are expensive. The current therapeutic approaches lose their effectiveness at this time, which can have deadly repercussions. To lower the mortality rate, novel methods for the early identification of cardiac disease are therefore vital. The research aims to create intelligent systems that can help doctors identify heart disease more quickly and affordably. The likelihood of a patient's survival will rise with the early discovery of potential damage in the system of the heart. The thesis provides a Fuzzy Inference System approach and Feed Forward Back Propagation Neural Network approach to develop intelligent systems based on some input parameters. There are so many factors that can affect the system of the heart. This research uses eleven major parameters to predict the risk of heart failure. The primary outcome of this study is that modelling based on artificial intelligence approaches is far more successful than what is currently available in the medical field for the early detection of heart disease. The performance of the developed systems has been evaluated by a confusion matrix based on 221 datasets collected from a valid source. The obtained result demonstrates that the performance parameters of the FIS model provide superior results compared to the ANN model. The developed FIS system's accuracy, precision, sensitivity, and specificity are 90.50%, 90.91%, 90.50% and 90.31%, respectively. A Graphical User Interference (GUI) is developed using the MATLAB App designer tool to facilitate the system’s practical applicability for the end-users.
Open Access
Analyzing the effectiveness of Covid-19 vaccines among different age groups using multinomial logistic regression model
(Faculty of Graduate Studies and Research, University of Regina, 2023-05) Khalid, Arfa; Deng, Dianliang; Volodin, Andrei; Peng, Wei
This study is conducted to evaluate the effectiveness of Covid-19 vaccines in different age groups in Saskatchewan, Canada. Data was collected between September 2021 and December 2021, and a statistical method called multinomial logistic regression was used to analyze the relationships between multiple categorical variables. In this study, the categorical variables were the age groups and the vaccination status (fully vaccinated cases, partially vaccinated cases, and unvaccinated cases) of the individuals with the interaction effect of rate of cases. The mathematical proof for the multinomial logistic regression model with interaction effect was derived in this study. The study demonstrated the effectiveness of Covid-19 vaccines among vaccinated age groups and provided theory and practical application of the multinomial logistic regression model. Results show that there is a statistically significant impact of age group and vaccination status on the effectiveness of Covid-19 cases in Saskatchewan. Specifically, there is a difference in vaccine effectiveness based on age groups and vaccination status. The findings of this study provide crucial insights for policymakers and public health officials to optimize vaccination rollout strategies and control the spread of Covid-19. Overall, this study represents an important step in the ongoing efforts to understand the effectiveness of Covid-19 vaccines and to develop policies and interventions that can help mitigate the pandemic impact.
Open Access
Application of Machine Learning Techniques for the Classification of Lower Back Pain in Human Body
(Faculty of Graduate Studies and Research, University of Regina, 2019-11) Sharma, Shubham; Mayorga, Rene; Peng, Wei; Kabir, Golam
Advancement of technology in the field of medical science is providing promising results in this modern era. Intelligent systems designed especially for this sector not only help doctors to solve complex situations but also take comparatively lesser computational time which can be a really important factor as longer time in critical cases may lead to death of a patient. These days, everyone experiences long waiting time to see a doctor. It would be really desirable if an Intelligent System makes the work easy for a doctor by giving accurate decisions after processing patient’s data, reducing examination time for the current patient and hence reducing waiting sessions for other patients. Currently, “Lower back pain is one of the biggest problems being faced by more than 80% of the population at least once during their lifetime” [1]. Its diagnosis at early stages is necessary in order to find a proper cure. Along with Conventional Medical Diagnostic Systems, Various Non-Conventional techniques are used for the successful classification of Lower Back Pain symptoms categorised as normal and abnormal. Naïve Bayes, Artificial Neural Networks, Logistic Regression, Deep Learning, Fast Large Margin, Random Forest, Gradient Boosted Trees, Multi-Layer Perceptron, K-Nearest Neighbour, Decision Tree and Support Vector Machine methods are most suitable machine learning techniques which can classify given dataset with good accuracy. The aim of this research is the application of several machine learning techniques to correctly classify Spine Dataset and finding best technique among those in terms of Accuracy, Precision, Sensitivity, Specificity, and F-measure [2]. Original dataset is taken from website named Kaggle (https://www.kaggle.com/). This dataset is normalized first and then an Automatic Feature Engineering technique has been implemented on the dataset to extract the most important features to do the correct classification. Training of each model is performed using II featured data and after training, each algorithm is tested and hence performance is calculated and compared. After analysing results, it is found that for the problem considered the Logistic Regression algorithm is the best classifier in terms of Accuracy giving 90.91% accurate results on test data followed by an Artificial Neural Network algorithm whose accuracy is 88.64%. In terms of Precision calculation, the Logistic Regression is best and the ANN Classifier is second best algorithm. Taking Sensitivity into Consideration, the Fast-Large Margin is best. ANN Classifier is best in terms of Specificity. Logistic Regression provides best results in terms of AUC (Area under Curve).
Open Access
Application of Predictive Analytics in Estimating Mechanical Properties for Investment Castings
(Faculty of Graduate Studies and Research, University of Regina, 2019-09) Virdi, Jaspalsingh Karamsingh; Peng, Wei; Yow, Kin-Choong; Mayorga, Rene; Deng, Dianliang
The investment cast parts are widely consumed in aviation, automotive, power generation, biomedical, and ship-building industries. The investment casting process is known for manufacturing complex shape parts with near-net finish. This process starts by making wax pattern similar to the shape of finished cast part. These patterns are assembled on a tree structure. A ceramic mold is formed around the tree assembly by repeatedly coating the ceramic mold with ceramic slurry and sand until the desired ceramic mold strength is obtained. The wax is then removed from the ceramic mold by heating it in the furnace. The molten metal is poured into the heated ceramic molds. Once the metal is cooled down, metal parts are extracted from the broken ceramic mold. This high-quality casting should have the required dimension, surface finish, mechanical properties, and should be defect free to provide better serviceability. Industrial quality control system measures mechanical properties through destructive testing after the manufacturing of cast components. The investment casting foundries faces a large amount of rejection or recycling that makes casting process less efficient due to wasted time, money, manpower, and raw material. Controlling the investment casting process is very difficult because it contains several sub-processes. This sub-processes leads to large number of parameters from the processing conditions and compositions. These process parameters are constantly changing from one batch to another during the production process, resulting in highly variable mechanical properties as well as finished castings. So, there is a need to develop an efficient system that can estimate the mechanical properties of investment casting parts before physical production. This thesis proposed a prediction system to ensure better quality control such that casting production meets targeted mechanical properties. Moreover, also shows research efforts in application of feature selection for finding the significant processing parameters, which affect the mechanical properties of investment casting parts. In this prediction system, several machine learning models including Multiple Linear Regression model (MLR), K-Nearest Neighbor (KNN), Random Forest (RF), and Extreme gradient boosting (Xgboost) were employed to predict individual mechanical properties (such as elongation percentage, yield strength, and ultimate tensile strength). Here, aforementioned methods are called as Data Learning Models (DLMs), they were trained with a large amount of foundry data collected for stainless steel automotive investment casting parts. Various feature selection techniques used in this proposed model include Boruta, Variable Selection Using Random Forest (VSURF), and Recursive Feature Elimination (RFE). They were employed to reduce the redundancy in the large dataset and find a best reduced dataset which can improve the prediction capability of DLMs. The performance of prediction models using the reduced datasets have been compared to the original dataset. The results showed that the reduced dataset generated from RFE maximizes the prediction accuracy. Xgboost provides the best accuracy for predicting mechanical properties when using the reduced dataset generated from RFE. The applications showed that the proposed prediction model is robust enough to handle noisy foundry data. Xgboost can easily be implemented and used by an operator for property forecasting. This data-driven machine learning models has demonstrated an effective way to optimize and control the investment casting process. The future work in this direction can lead to smart manufacturing with online monitoring and control of properties.
Open Access
Assessment of Greenhouse Gas Emissions From Spring Wheat Cropping System in Saskatchewan
(Faculty of Graduate Studies and Research, University of Regina, 2019-04) Shi, Yarong; Huang, Guo (Gordon); Peng, Wei; Wu, Peng; Yang, Boting
The assessment of GHG emissions from spring wheat cropping system in Saskatchewan was conducted in this study. A general emission assessment model was developed. The main sources of GHG included emissions from farming operations, emissions from the manufacturing and transportation of N/P fertilizer, emissions from herbicides usage, and direct and indirect emissions from agricultural land. A case study based on spring wheat in Saskatchewan was then investigated. The results show that the total GHG emissions mainly come from the manufacture, storage, delivery, and application of nitrogen and environmental conditions have a significant effect on the total GHG emissions. In addition, a factorial analysis has been applied to evaluate the impact of uncertain parameters on system performance. In the case study of spring wheat crop in Saskatchewan, the total GHG emission is around 3,358,451 Mg CO2-eq. The results indicate that GHG emission is affected by environmental condition. From north to south Saskatchewan, the total GHG emission is decreasing as humidity decreased. The total GHG emission is mainly related to N fertilizer application, which accounts for about 80% of total emission. Through applying multivariate factorial analysis, the effects of uncertain parameters were identified, and the main effects and their interactions were also examined. The results have important implications for our efforts to evaluate the GHG emission from agricultural activities.
Open Access
An Automating Interpretation System of Industrial Radiographic Digital Images Used in Nondestructive Testing
(Faculty of Graduate Studies and Research, University of Regina, 2019-09) Alqahtani, Abdullah Falah; Hussein, Esam; Peng, Wei; Zhang, Lei; Kabir, Golam; Yao, Yiyu
This thesis presents a method for automating the interpretation of industrial radiographic digital images used in nondestructive testing of subsurface defects. The goal of this study is to develop a system for detecting and identifying defects in welding processes from digital radiographic images. The proposed approach consists of three main stages: digital image processing, feature extraction, and pattern recognition. Twelve features were selected in a process to classify welding defects. Three well-known classifiers were applied in the stage of the classification process: Support Vector Machine (SVM), k-nearest neighbor (KNN) and artificial neural networks classifiers (ANN). A confusion Matrix was used to analyze the performance of the methods. Numerical experimental results confirmed the reliability and feasibility of the proposed model for detecting and locating and separating defect from non-defect indications.
Open Access
Cellular network KPI prediction on simulated 5G-NR V2N traffic dataset using machine learning
(Faculty of Graduate Studies and Research, University of Regina, 2023-03) Pusapati, Suryanarayanaraju; Peng, Wei; Khan, Sharfuddin; Maciag, Timothy; Wang, Zhanle
The arrival of 5G has brought a promise of better connectivity for users, but also a challenge for cellular networks to maintain high-quality service and energy efficiency. To optimize the network and meet user demands, a resource management system is used to allocate resources in the 5G Radio Access Network (RAN). However, manual tuning of this system is complex and time-consuming. By predicting the future behavior of Network Key Performance Indexes (KPIs) of the 5G network using Artificial Intelligence (AI) and its subfield, Machine Learning (ML), this study can automate the operations of the Resource Management system, improve resource allocation, and satisfy QoS requirements while optimizing energy consumption. However, to develop a better performing ML model, a high-quality dataset is essential. Since there is a lack of open datasets available on 5G systems, many researchers rely on synthetically generated datasets. This thesis work utilized 5G simulation tool to simulate 5G New Radio (NR) Vehicle-to-Network (V2N) communication using OMNeT++ and SUMO simulators. The NR V2N communication was simulated in a Regina downtown scenario using the proposed simulation framework, and the simulation results were processed using the developed new_df Python module into synthetic datasets that were validated by comparing with technical specifications to ensure their quality. The synthetic datasets were then used to develop proposed Network KPI prediction models using ML. Three ML models are trained and tested, which can predict multiple KPIs, bi-directional Signal to Interference and Noise Ratio (SINR) and classify uplink Channel Quality Indicator (CQI) respectively. The multi-output regression models have shown outstanding performance with MSE as low as 0.002, and the multi-class classification model has a high accuracy. In summary, this study contributes to the development of efficient and automated Resource Management systems for 5G networks using AI and ML techniques. An open source V2N simulation framework was developed using OMNeT++ and SUMO simulators that can simulate 5G-NR V2N communication in a realistic urban scenario. Moreover, a new_df Python function was developed for processing simulation results into an aggregated dataset and spatiotemporal dataset, providing a high-quality dataset that can be used to train and test ML models for predicting Network KPIs of the 5G network.
Open Access
Classification of soil surface texture using high-resolution RGB images captured under uncontrolled field conditions
(Faculty of Graduate Studies and Research, University of Regina, 2023-09) Babalola, Ekunayo-Oluwabami Oreoluwa; Bais, Abdul; Wang, Zhanle (Gerald); Peng, Wei
Understanding the properties of soil and its impact on the environment and farming practices requires accurately classifying its texture. Accurate soil texture classification can optimize soil nutrient levels and improve land management. This study proposes a framework that uses images captured under Uncontrolled Field Conditions (UFC) to classify soil texture for farmlands accurately. UFC images are captured in varying ambient light and environmental conditions, which can introduce unwanted elements such as shadows and varying lighting. Our framework uses image-processing techniques, texture-enhancing methods, and deep learning to process and classify these soils accurately. First, we process the soil using semantic segmentation to eliminate all non-soil pixels. We compare Segmentation Network (SegNet), U-shaped Neural Network (UNet), Pyramid Scene Parsing Network (PSPNet), and DeepLab v3+ models to choose the best for semantic segmentation. The trained segmentation model produces masks used to eliminate non-soil pixels from the images. This process produces new images with random 0 pixel clusters that would negatively disrupt texture information, and so next, we split the new images to eliminate all 0 pixel clusters while preserving only soil pixels. We then perform texture enhancement on the images before feeding them into the classification network. We design and use an improved network called EfficientCNN for classification to use a reduced number of parameters while producing maximum accuracy. We also compare this model with Residual Network (ResNet 50), EfficientNetB7 and Inception v3. EfficientCNN architecture uses just 5.9 million parameters and produces an accuracy of 84.783%, while Inception v3 uses 21.7 million parameters and produces an accuracy of 85.621%. EfficientCNN produces only 0.838% less accuracy than Inception v3. Our results contribute to agriculture and soil science studies.
Open Access
Classifying ovarian cancer using machine learning methods
(Faculty of Graduate Studies and Research, University of Regina, 2023-11) Rahman, Rushda; Peng, Wei; Henni, Amr; Muthu, S. D. Jacob
Ovarian cancer is one of the most fatal cancers for women nowadays. It is ranked as fifth most common cancer deaths among women resulting more deaths than any other cancers in female reproductive system. According to Canadian Cancer society that about 3000 ovarian cancer patients were detected, and among them 1950 patients died in 2022 which indicating more than 50% of mortality rate. Ovarian cancer is mainly generated from cancerous ovarian tumour. So, it is very important to classify cancerous tumour from noncancerous tumour to prevent false positive for ovarian cancer. Moreover, if cancerous tumour is diagnosed in early stage, it can be prevented from spreading and thus survival rate for ovarian cancer can be increased. Also, by separating cancer patients from benign tumour patients, it will be easier for doctors to know the stages of the cancer and know patient’s prognosis and life expectancy. The principal and initial objective of this thesis is building a feasible system using Artificial Intelligence which is easy to use and compatible to classify ovarian cancer. Proposed study will give a new non-conventional way to classify ovarian cancer from ovarian tumour which will be affordable for the patients. Moreover, one of the primary benefits of this study is that doctors/physicians can detect ovarian cancer with only blood test/ serum test. There is no need to do any expensive tests such as: ultrasound, MRI or CT-Scan. The main concept of this research is the application of several machine learning techniques to correctly classify ovarian cancer and finding best technique among those in terms of Accuracy, Precision, Sensitivity, and Specificity. Original dataset is taken from website named Kaggle (https://www.kaggle.com/). This dataset is screened, cleaned and normalized first and then expert’s advice has been taken to extract the most important features to do the correct classification. Later, a correlation test has been done for better understanding of the relations and independency among the input features. 10 input features have been selected including age, menopause, CA-125, AFP, NEU etc. From correlation test result 7 inputs were taken again and a comparison had been made between 10 inputs and 7 inputs. And the output is TYPE which denotes 1 for benign ovarian tumour and 0 for ovarian cancer. Four machine learning models have been used for classification and they are, ANN, SVM, Naïve Bayes, and k-NN. Training of each model is performed and after training, each algorithm is tested and hence performance is calculated and compared. After analysing results, it is found that for the problem considered, the Artificial Neural Network (ANN) is the best classifier in terms of accuracy giving 85.91% accurate results on test data whereas SVM, NB and k-NN gave accuracy of 76.05%, 83.09% and 76.06% respectively. In terms of sensitivity and precision calculation, Naïve Bayes is best, and the ANN Classifier is second best algorithm. Taking specificity into Consideration, the ANN is best with 87.50%. Keywords: Machine Learning Classifier, Ovarian Cancer, Benign Ovarian Tumour, Artificial Intelligence, Artificial Neural Network (ANN), Support Vector Machine (SVM), Naïve Bayes (NB), k-nearest Neighbour (k-NN), Confusion Matrix.
Open Access
Collision Free and Energitaclly Optimized Motion Planning of Manipulators in Partially-Known Environment Using Modified D* Life Algorithm
(Faculty of Graduate Studies and Research, University of Regina, 2016-12) Feizollahi, Amir; Mayorga, Rene; Peng, Wei; Ismail, Mohamed
Robotics is a relatively young field of studies in modern technology and it has tremendously grown during the past fifty years. Manipulators are categorized as a group of robots designed to accomplish the manipulation tasks without direct contact by a human. Generating an optimized algorithm to transform high-level motion tasks to low-level description that are understandable by the robots, including manipulators, has been one of the most interesting problems in this field. Motion planning in robotics, is referred to the process of breaking down a desired movement task into discrete motions in order to satisfy some specific criteria and optimize some certain variables during the motion from the start point to the goal point. Although various types of problems in robot motion planning have been investigated; trajectory planning for the manipulators in partially-known environment with respect to the amount of energy consumption has not been fully addressed. This aspect of motion planning can be of great importance in optimizing the path planner algorithms while satisfying their collision free attribute. This study attempts to develop a comprehensive mathematical model for robot-actuator system dynamics to include the energy consumption level as a variable in the cost function of the path planner algorithm. For this purpose, the manipulator’s general equation of motion is derived using the Lagrange equations. Same approach has been taken to develop the equation of motion of the robot’s actuators which are considered to be DC motors at each joint of the manipulator. The fourth order of Runge-Kutta algorithm is used to solve the coupled governing differential equations of the system. The output of the modeling phase of the study is fed to the graph search algorithm as an input. Graph search algorithm deals with the start-to-goal node problem in which the best path in a network of nodes is desired. D* Lite algorithm is one the most well-known algorithms in dealing efficiently with the partially-known environment motion planning problems. The advantages of D* Lite algorithm over the other famous algorithms such as A* are investigated, the best path generation procedure is thoroughly discussed, and the implementation of the algorithm is deliberated. A modification on D* Lite algorithm is proposed here to enhance the efficiency of the best path generation method and the alternative procedure is provided correspondingly. A MATLAB framework is designed to define the robot and its workspace and the manipulator equation of motion is developed using MATLAB classes to transform the user input values into the state space variables of the system. Several scenarios have been simulated using the developed framework to verify the path planner effectiveness in avoiding the pre-known and partially-known obstacles while minimizing the amount of consumed energy by the manipulator during its motion from the start node to the goal node.
Open Access
Comparison of Random Forests, Support Vector Machine and Artificial Neural Network Methods for Agriculture Land Cover Classification
(Faculty of Graduate Studies and Research, University of Regina, 2022-04-13) Zhou, Xin; Wu, Peng; McCullum, Kevin; Veawab, Amornvadee; Han, Todd; Peng, Wei
Land cover classification is critical in remote sensing. Reliable classification on land cover is required to address a wide range of environmental issues. Over recent years, the application of machine learning techniques in remote sensing has attracted wide attention. Machine learning has the capacity to classify land cover in remotely sensed photos effectively and efficiently. Machine learning techniques, such as Random Forests (RF), Support Vector Machines (SVM) and Artificial Neural Networks (ANN) can be applied in land cover classification. However, putting a machine learning categorization system in place is not easy, especially in the field of agricultural land classifications. Very limited research can be found using the machine learning techniques for agriculture land cover classifications. In the Canadian prairies, the cropping systems are based on simplified input-driven production of annual crops, which could affect land cover classification. To investigate the performance of machine learning techniques, the present research is conducted in the southern prairie region of Saskatchewan. The satellite images used for classification are from Sentinel-2, which is a publicly data from the European Space Agency. In total, 133,080 samples were analyzed using stratified random sampling, divided into training (70%) and test (30%) subsets. The accuracy is assessed by a variety of indicators. It is found that RF has the highest overall accuracy while the SVM has the lowest accuracy. The ANN has more advantages compared to others. Future agricultural land cover classifications can be based on the current research.
Open Access
Computation of inverse kinematics of redundant manipulator using particle swarm optimization algorithm and its combination with artificial neural networks
(Faculty of Graduate Studies and Research, University of Regina, 2023-03) Monfared, Pedram; Peng, Wei; Henni, Amr; Kabir, Golam
Nowadays, the world of the industry is not working without the robotics’ manipulators. Logically, robotic manipulators are very old topics and lots of researches have been done on them. However, with implementing recent methodologies on them, many contributions have been achieved, which even become a part of human’s life. The objective of this research is to achieve more precise position for end effector of rescue robotic manipulator in real time within reasonable calculation time. In this regard, all the inverse kinematics formulation of the 2, 3, and 4-links manipulator are derived, and be solved by Particle Swarm Optimization (PSO) method. It should be mentioned that this optimization method will be used to solve the inverse kinematics problem in real-time mode which consequently making the results more accurate. The presented PSO method is implemented on a rigid 3-link manipulator with three rigid revolute joints. It is pointed out that the application of the robot manipulator in this paper is used as a rescue robot for the first time, which has stronger environmental adaptability and real-time performance. It is shown in this study, the precision of PSO method is about 10-3.6 (0.0002511) within 50 iterations which is a lot better than previous research (i.e. an ANN application did the same study and used the same platform only can generate a result with the error of 0.045589 in 1000 epochs). However, as the PSO method is real-time, the presented PSO method will consume a lot more time than the other methods such as ANNs. For instance, the PSO method took 2 minutes and 10 seconds to solve an inverse kinematics problem comparing with 36.10 seconds that solved by an ANNs. Therefore, author developed a combination approach that mixing the PSO method and ANN technique, namely PSO-ANN, to solve rescue robotic manipulator in real time, which can generate high accuracy results in a short testing time. Due to the aforementioned description, the contribution of this study is to reach to more precise methodology with less error due to the previous studies and also presenting the methodology that can solve all the issues in regards of the run-time of the real-time methods.
Open Access
Covid-19 (Coronavirus) Disease Diagnosis Using Fuzzy Inference System
(Faculty of Graduate Studies and Research, University of Regina, 2021-12) Jani, Bhargav Nikhilbhai; Mayorga, Rene; Peng, Wei; Yao, Jing Tao
Intelligent Systems (IS) are called technologically advanced machines. IS includes a variety of techniques that can deal with uncertainty and complex problems. This research study aims to diagnose the coronavirus (COVID-19) diseases symptoms using an Intelligent Systems approach. The COVID-19 epidemic is currently one of the most deadly diseases in the world. The pandemic has affected human life and global economy adversely at a large scale. It is spreading rapidly through humans, causing severe health issues and death worldwide. This virus affects different people in different ways. A large number of people are not aware of being infected with COVID- 19, as some people are asymptomatic, or show initial symptoms that could be confused with a mild illness, such as having a cough or fever. There are over 3 million deaths currently caused by COVID-19, with the number increasing regularly. Advanced healthcare and error-free disease diagnosis are ubiquitous now-a-days as a result of technological advancement. Better future and healthy lifestyle are majorly depending on sensible and latest health care facilities. Large number of research and articles prove the efficiency and effectiveness of Intelligent Systems being used for diagnosing various symptoms of heart disease, cancer and diabetes. In this Thesis, an Intelligent System encompassing five Mamdani FISs is presented to help the COVID-19 disease diagnosis. This advanced Intelligent System will help to reduce the uncertainty as well as ease the diagnosing process. The proposed FISs can provide fuzzification results for the considered sub-systems with multiple inputs. In addition, a user-friendly interface is implemented for ease of interaction between a human and the proposed Intelligent System using the MATLAB software. The designed Intelligent System can decide the possible risk of getting COVID-19 disease. In each of the FISs, the fuzzification rules and database play a wide role. The 1st FIS considers four (Age, Medical Supplements, Immunity Strength, Previous Medical History) factors as input parameters to find out the “Disease Tendency” of COVID-19. Similarly, the input (Temperature, Tiredness, Dry Cough, Sore Throat) factors for the 2nd FIS, yield the “Most Common” symptoms. The 3rd FIS considers 4 input (Diarrhea, Headache, Conjunctivitis, Loss of Taste) factors yielding “Less Common” symptoms. The 4th FIS considers also 4 input (Breathing Difficulties, Chest Pain, Loss of Speech/Movement, Cholesterol Level) factors to yield “Serious Common” symptoms. Finally, the last FIS considers as inputs the: “Disease Tendency”, the “Most Common” symptoms, the “Less Common” symptoms, and the “Serious Common” symptoms; to yield as a result the disease likelihood (Consider changing serious common to something else. It doesn't really make sense. If you're just talking about more serious symptoms, maybe just have that category be called "Serious" instead of "Serious Common") Therefore, the overall proposed Intelligent System considers a total of 16 factors as input variables. It is important to notice the novel consideration of the Cholesterol Level as a factor in the “Serious Common” symptoms FIS module. It is mandatory to diagnose a disease in early stage to control and halt its spread. The proposed IS will help diagnose the COVID-19 symptoms and affection in early stage. Moreover, the advanced FIS is beneficial for an individual to diagnose disease by him/herself and extremely helpful in such places and societies where it is almost impossible to find the supply of physicians for the timely treatment of any medical disease.
Open Access
Dealing With Uncertainty in Engineering and Management Practices
(Faculty of Graduate Studies and Research, University of Regina, 2011-07) Peng, Wei; Mayorga, Rene; Mehrandezh, Mehran; Henni, Amr; Deng, Dianliang; Hipel, Keith W.
A set of methodologies is proposed for dealing with uncertainties in five fields. These fields are: (1) traffic noise impact assessment, (2) hydraulic reliability assessment and reliability based optimization, (3) binary linear programming, (4) real-time multiple source water blending optimization, and (5) process control of an industrial rotary kiln. The proposed methods are applied to several engineering and management cases to demonstrate their explicabilities and advantages. In the field (1), an integrated approach is presented to assess traffic noise impact under uncertainty (Peng and Mayorga, 2008). Three uncertain inputs, namely, traffic flow, traffic speed and traffic components, are represented by probability distributions. Monte Carlo simulation is performed to generate these noise distributions. Further, fuzzy sets and binary fuzzy relations are employed in the qualitative assessment. Finally, the quantification of noise impact is evaluated using the probability analysis. In the field (2), two innovative approaches are developed under uncertainty (Peng and Mayorga, 2010e). One is to assess hydraulic reliability that accounting for the deterioration of both structural integrity and hydraulic capacity of each pipe; another is to design a reliability- based optimal rehabilitation/upgrade schedule that considering both hydraulic failure potential and mechanical failure potential. In these two approaches, all uncertain hydraulic parameters are treated as random values. The main methodologies used are: Monte Carlo simulation, EPANET simulation, genetic algorithms, Shamir and Howard’s exponential model, threshold break rate model, and two-stage optimization model. Eventually, two universal codes, the hydraulic reliability assessment code and the long-term schedule code,were written in MATLAB and linked with EPANET. In the field (3), an interval coefficient fuzzy binary linear programming (IFBLP) and its solution are built under uncertainty (Peng and Mayorga, 2010c, 2010d). In the IFBLP, the parameter uncertainties are represented by the interval coefficients, and the model structure uncertainties are reflected by the fuzzy constraints and a fuzzy goal. The solution includes a defuzzification process and a crisping process. An alpha-cut technique is utilized for the defuzzification process, and an interval linear programming algorithm is used to the crisping process. One mixed technique (links the alpha-cut technique and min-operator technique) is used to determine a single optimal alpha value on a defuzzified crisp-coefficient BLP. Finally, the IFBLP is converted into two extreme crisping BLP models: a best optimum model and a worst optimum model. Uncertainties in the field (4) include the modeling uncertainty and dynamic input uncertainty (Peng et al, 2010a, 2010b). This dissertation provides a fuzzy multiple response surface methodology (FMRSM) to deal with these kinds of uncertainties. In the FMRSM, the experimental data sets are fitted into the first quadratic models and their residuals are fitted into the second quadratic models; the multiple objectives are optimized using a fuzzy optimization method. Six scenarios are designed based on a real-time operation. The results show the FMRSM is a robust, computational efficient and overall optimization approach for the real-time multi-objective nonlinear optimization problems. In the field (5), a dual-response-surface-based process control (DRSPC) programming is developed to address the uncertainty and dynamic calcination process (Peng, et al, 2010f). Several response surface models are appropriately fitted for an industrial rotary kiln. The proposed approach is applied on a real case. The application shows that the proposed approach can rapidly provide the optimal and robust outputs to the industrial rotary kiln. Other properties of the proposed approach include a solution for the time delay problem and a statistical elimination of measurement errors.
Open Access
Design and development of a Virtual Window with industrial and civil applications
(Faculty of Graduate Studies and Research, University of Regina, 2023-08) Ansari, Saeed; Wang, Shanle; Paranjape, Raman; Peng, Wei
The recent advances in deep learning algorithms and computer vision applications have prompted research into their application in various problem domains, including industrial and civil applications. This area o ers a unique opportunity to develop applications that were not previously Possible without computer vision technology. One such application is the replacement of conventional windows with digital smart windows that provide end-users with access to natural scenery. This technology has potential uses in military contexts and meetings, among others. This research focuses on a novel approach to implementing this technology in industrial settings using a real-time face detection algorithm to display natural scenery on a digital screen in a way that provides a user experience similar to looking out a real window. The study involved testing multiple methods and algorithms to identify the fastest and most ef- cient approach that can be implemented purely through software methods, thereby reducing the need for costly sensors and hardware like what you may nd in VRs with expensive sensors. Given the paramount signi cance of both execution speed and detection accuracy in our project, we have made a deliberate decision to utilize the most e cient real-time face detection algorithm available. This algorithm, known as Yolov7, excels in achieving swift processing while maintaining a high level of preci- sion.Details of how this model excels compared to other real-time models can be found in Chapter 3. YOLO represents a singular stage detector that adeptly handles object identi cation and classi cation within a single iteration of the network. Although various single stage detection models exist, YOLO consistently demonstrates supe- rior performance in terms of both speed and accuracy. By approaching the detection task as a single-shot regression method for identifying bounding boxes, YOLO mod- els exhibit remarkable swiftness and compactness, rendering them highly amenable to e cient training and deployment, particularly on resource-constrained edge devices. The algorithm employed in this study was utilized for the dual objectives of detect- ing facial features and identifying landmarks. To cater to the speci c requirements of the project, a custom dataset was employed during the training phase of the algo- rithm. Rather than undertaking the task of creating a novel algorithm, our approach involved identifying the latest and most e cient algorithm, which we subsequently employed in our application. Consequently, we were able to divert more resources towards enhancing the software capabilities of our application, such as accurately estimating the user's head orientation and focusing on related aspects.
Open Access
Develop innovative methodology to optimally fill in missing values and predict progression on multiple sclerosis
(Faculty of Graduate Studies and Research, University of Regina, 2024-11) Pilehvari, Shima; Peng, Wei; Shirif, Ezeddin; Khan, Sharfuddin; Fan, Lisa; Bui, Francis
Applying Machine Learning (ML) to predict and track Multiple Sclerosis (MS) progression is a significant advancement in medical research, with the potential to enhance patient outcomes. Accurate MS prediction enables personalized treatment, timely interventions, and improved quality of life by slowing disease progression and preventing complications. This research aims to deepen our understanding of MS by developing ML models and comprehensive risk assessments to support early prognosis, guide treatment strategies, and reduce disease impact. A major challenge in medical research, especially in predicting MS progression, is effectively managing missing data in MS datasets. This study introduces an innovative sequential Multi-Imputation (MI) bootstrapping method to address the challenge of missing data in MS datasets. Initially, several ML algorithms, including k-Nearest Neighbors (kNN), Random Forest (RF), and Multilayer Perceptron (MLP), are evaluated for imputation efficiency. RF and MLP perform best, achieving overall accuracies of 92% and 91.5%, respectively, in handling missing data more accurately than other models. Given the effectiveness of RF and MLP in capturing complex patterns in data, these models are selected for further development. The next step applies Multi-Imputation (MI) bootstrapping in a sequential manner, prioritizing features based on the strength of their relationships, as determined by Pearson correlation analysis. This statistical technique identifies features with the highest correlations, ensuring that attributes with stronger relationships with other attributes, are imputed first. These imputed features then inform the next imputation in the sequence, cooperating with the subsequent ranked feature in the order. Bootstrapping, a resampling technique that involves replacement, creates multiple training datasets by repeatedly sampling from the original data, enhancing the robustness of the imputation process. The proposed sequential imputation method integrates bootstrapping with RF, achieving an accuracy up to 97 % for MS datasets. This iterative approach effectively imputes missing data attributes while accounting for feature significance and relationships. The results also show that prioritizing normalization improves scaling impact, and that the significant features in the original dataset are crucial to the accuracy of MS missing data estimations. These findings provide valuable insights into effective imputation techniques for MS prediction, offering a foundation for future improvements in handling missing data in specific datasets. In addition, this study solves the common overfitting problem caused by data imbalance through a comprehensive method combining feature extraction, undersampling, Synthetic Minority Oversampling Technique (SMOTE) and optimal threshold method. Support Vector Machine (SVM), Logistic Regression (LogR), Decision Tree (DT), RF, KNN, MLP and Naive Bayes (NB) are used for prognostic modeling while examining risk factor associations. The results showed that the proposed method prevented overfitting during model training and developed a robust MS progression prognosis model, achieving a prediction accuracy of 98%, particularly for SVM and MLP The methods proposed in this dissertation can help develop more concise guidelines for the medical research communities and improve their evaluation processes. These innovations not only advance prognostic analysis in MS, but also pave the way for future research focused on optimizing patient outcomes and treatment strategies.
Open Access
Development of a Collaborative Decision-Making Framework to Improve the Patients’ Service Quality in the Intensive Care Unit During the COVID-19 Pandemic
(Faculty of Graduate Studies and Research, University of Regina, 2020-12) Siva Kumar, Gowthaman; Kabir, Golam; Almehdawe, Eman; Henni, Amr; Peng, Wei
Healthcare is one of the biggest and complex service sectors, where decisions need to be made quickly, accurately and efficiently including multiple stake-holders. Especially during this COVID-19 pandemic, service improvement decisions in Intensive Care Units (ICU) are considered to be a critical factor. Many implemented solutions in the healthcare sector become unsustainable in the long term due to a lack of consensus among the stake-holders. In this thesis, a collaborative decision-making framework is developed to improve the Patients’ Service Quality in the ICU including multiple stake-holders. The key criteria, alternatives that can advance the service quality in the ICU are identified from in-depth literature review. In this research work, the best-worst method (BWM) is integrated with Multi-Actor Multi-Criteria Analysis (MAMCA) method to capture the stake-holder’s opinion and preference value. Based on criteria weight and preference value the best-ranked alternative for each stake-holder is determined. However, to find out the consensus solution a Multi-Objective Linear Programming (MOLP) is integrated with BWM-MAMCA methods. The MOLP considers every criterion as an objective function including all stake-holder’s criteria to find a consensus solution. Two scenarios are considered in this research work, scenario 1 considers all selected criteria under each stake-holder, whereas in scenario 2, every stake-holder selects their preferred set of criteria based on their work background. Both scenarios’ results suggest that alternative A1 (Hiring Part-time physicians and medical staff) is the best consensus solution to improve the patients’ service quality in the ICU by satisfying every stake-holder’s requirement. The developed integrated framework even allows stake-holder groups (multiple participants under each stake-holder group) to participate in the group decision-making process to select an effective strategy that can improve the patients’ service quality.
Open Access
Development of a pellet extruder with co-axial nozzle for 3D printing using inflatable extrudates
(Faculty of Graduate Studies and Research, University of Regina, 2024-08) Habib, Md Ahsanul; Khondoker, Mohammad; Muthu, SD Jacob; Peng, Wei
Additive manufacturing (AM) has emerged as one of the core components of the fourth industrial revolution, Industry 4.0. Among others, the extrusion AM (EAM) of thermoplastic materials has been named as the most widely adopted technology. Fused filament fabrication (FFF) relies on the commercial availability of expensive filaments; hence pellet extruder-based EAM techniques are desired. Large-format EAM systems would benefit from the ability to print lightweight objects with less materials and lower power consumption which can be possible by using hollow extrudates rather than solid extrudates to print objects. In this work, we designed a custom extruder head and developed an EAM system that allows the extrusion of inflatable hollow extrudates of a relatively wide material choice. By incorporating a co-axial nozzle-needle system, a thermoplastic shell was extruded while the hollow core was generated by using pressurized Nitrogen gas. The ability to print using hollow extrudates with controllable inflation allows printing objects with gradient part density with different degrees of mechanical properties. In this article, the effect of different process parameters namely, extrusion temperature, extrusion speed, and gas pressure were studied using poly-lactic acid (PLA) pellets. Initially, a set of preliminary tests was conducted to identify the maximum and minimum ranges of these parameters that result in consistent hollow extrudates. Later, the parameters were varied to understand how they affect the core diameter and shell thickness of the hollow extrudates. These findings were supported by analyses of microscopic images taken under an optical microscope. In the next phase of our experiment, we printed an inflated cylindrical part using the process parameters derived from the initial set of experiments. We carefully compared the results with the data obtained earlier to ensure accuracy and consistency. Finally, we successfully printed an object with varying densities in different sections. Keywords: Additive Manufacturing; Extrusion Additive Manufacturing; Hollow Extrudates; Pellet Extrusion; Fused Filament Fabrication.