http://dlib.iit.ac.lk/xmlui/handle/123456789/492 2026-04-21T17:08:18Z http://dlib.iit.ac.lk/xmlui/handle/123456789/540 SVM Based Part of Speech Tagger For Sinhala Language Wijerathna, Y.A.D.S.S. Natural language processing is providing the ability of understanding human language as it is spoken to a computer programme.”It is a component of computer science, linguistics and artificial intelligence. Building NLP application is a difficult task because human speech is not always specific. NLP is a process of developing a system that can read text and translate between one human language and another. In performing NLP task part of Speech tagging is a basic requirement which needs to catered appropriately.” Part of Speech (POS) tagging is the task”of labelling each word in a sentence with its appropriate syntactic category called part of speech. POS tagging is a very important pre processing task for language processing activities. “ Sinhala is the native”language of the Sinhalese people who make up the largest ethnic group of Sri Lanka. Sinhala is a morphologically rich language in the Indic family. While this makes it harder to build an accurate tagger without a good morphological pre-processor, it provides a compelling reason for attempting to build a POS tagger for Sinhala. However, due to poverty in both linguistic and economic capital, Sinhala, in the perspective of Natural Language Processing tools and research, remains a resource-poor language.” This paper is an initiative to overcome above issue in lack of NLP tools for Sinhala language, here discusses the task of POS tagging for Sinhala language using Support Vector Machine (SVM). The POS tagger has been developed using a tag set of 30 POS tags, defined for the Sinhala languages by University of Moratuwa and used the SVMTool developed by Jesus Gimenez and Lluıs Marquez.” The accuracy of available Sinhala Part-Of-Speech taggers, which are based on Hidden Markov Models, still falls far behind state of the art. Our Support Vector Machine based tagger achieved an overall accuracy of 85.68% for known words. 2020-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/539 Automated Cloud Pattern Identifier from Satellite Images Wijayathilaka, Hansamali In this millennial climate changes have become a hot topic and researchers are trying to address many possible problems regarding our planet earth and its atmosphere. Clouds play a major role when it comes to atmosphere. Moreover, clouds have been helpful in weather forecasting, photovoltaic fields, disaster recovery and etc. Scientists who are doing research regarding the world’s ever-changing atmosphere they have faced a need of identifying cloud patterns. In order to overcome the need these scientists have carried out cloud pattern recognition through crowd sourcing, which means they have gathered several scientists and they have identified several cloud patterns based on their appearance and several other scientific factors and thereafter this group have identified these cloud patterns from satellite images manually. Identifying cloud patterns manually take a considerable amount of time. As a solution to this the Author came up with an idea of automating the cloud pattern identification process from satellite images. This system would automate the manual process of identifying cloud patterns from satellite images using segmentation and classification methods along with Deep Learning algorithms. This system is first of its kind in Python language and no other existing systems covers and end to end workflow of automating cloud pattern identification. It is capable of identifying cloud patterns in a single image as well as in multiple images. It is available as a desktop solution along with GUIs to carry out the necessary user actions and finally the results could be exported to the local machine. 2020-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/538 Elevating The Accuracy Level Of Predictive Maintenance in Field Services Management Weerakkody, W.A.S.S. Maintenance costs are highly substantial, manufacturing plants tend to spend huge amount of money to have maintenance activities steady as it can directly impact on key factors of the production process; productivity and profitability. However, at the same time plants losing another considerable amount of money as a result of unnecessary or improperly carried out maintenance. The foremost reason for unproductive maintenance is non-existence of factual data to quantify the actual desire for maintenance of manufacturing equipment. In many instances, maintenance scheduling is done based on statistical trend data or on the actual failure of machinery. Common judgement towards maintenance costs has been; “Maintenance is a necessary evil” or “Nothing can be done to improve maintenance costs”. It is said that, industry is currently undergoing what it is considered to be the fourth industrial revolution. Various utilization of condition monitoring tools was presented alongside the progress of IoT and Cloud based technologies, which led to the concepts such as smart cities and smart factories. Huge advancement of condition monitoring tools and their ability to collect numerous real time data has opened a door to new maintenance stagey; Predictive maintenance, which in simply terms means, do the maintenance only when it is needed. Even though the concept of predictive maintenance has been there in the industrial domain since the 1990s, Recent developments of IoT and machine learning has redefined it. Over the last decade, machine learning has made number of contributions in the domain of predictive maintenance and has earned itself a prominent role. But the overuse of feature engineering on data obtained from condition monitoring tools have bounded machine learning capabilities of predictive maintenance models to a designated domain they initiated. In order to address this scenario, researches on deep learning approaches which are able to extract features automatically were encouraged. Goal of this research is to take a deep learning approach to build a reusable predictive maintenance model irrespective of domain or the physical process of plant machinery 2020-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/530 Samaritan: A Real CCTM Stream Analytics Engine For Suspect Detection Perera, Ninda Most of the public areas such as roads, airports, bus stands, railway stations and office buildings are covered with huge network of CCTV cameras nowadays and enormous amount of video surveillance data is produced every moment. Since these devices do not have any intelligence, manual monitoring by human is required to detect any alarming activity. But it is not very feasible due to human resources limitation. Since there is no mechanism to monitor all these CCTV feeds manually, lot of terrorist activities and criminal activities have not been prevented before it is too late. Therefore, the requirement to analyze large amount of surveillance data in real-time is increasing day by day. Some of the researchers have implemented intelligent frameworks which were capable of identifying suspicious behaviors of persons by analyzing a video. Also, there have been a lot of researches for facial recognition. But still these systems have not been enhanced to analyses video streams of broad CCTV networks in real-time. Therefore, failure to analyze large amounts of CCTV streaming data for known suspects and monitor human activities in real-time have become a major problem. However, the proposed engine, which is introduced as Samaritan, addresses the problem stated earlier through an intelligent real time data processing engine which analyses CCTV streaming data in real time and identify pre-known faces and suspicious activities conducted by criminals or terrorists. Also, it is capable of alerting the authorities, in case any suspicious activity is detected in CCTV feeds In the proposed system, large amounts of video streaming data are handled by a centralized consumer, which is built on top of Apache Kafka. Then deep learning techniques were used to perform facial recognition and human action classification for video streams. First of all, the system trains a model from the face images of known suspects and then those faces will be recognized in CCTV surveillance video streams which are fed in to the real-time video analytics engine. Next, Samaritan classifies suspicious activities in CCTV video feeds based on a deep learning model that was pre trained with suspicious activity datasets. Then the proposed system is able to identify suspicious activities of CCTV streaming data. By combining these the features in to the real time big data CCTV stream analytics engine, it is able to identify suspicious activities of a CCTV surveillance videos and known faces in real-time and inform the authorities when necessary. Finally, the evaluations describe that the proposed system is efficient and accurate. 2020-01-01T00:00:00Z