1. Ph.D Theses
Permanent URI for this collectionhttps://idr.nitk.ac.in/handle/1/11
Browse
4 results
Search Results
Item Detection of Breast Cancer Using Thermal Patterns and Artificial Intelligence(National Institute of Technology Karnataka, Surathkal., 2024) Venkatapathy, Gonuguntla; N., GnanasekaranBreast cancer remains one of the leading causes of cancer-related deaths among women worldwide, emphasizing the crucial need for early detection to enhance treatment outcomes and survival rates. While advanced screening technologies have significantly benefited developed countries, underdeveloped and developing nations face significant challenges due to limited access and high costs. Traditional screening methods, though effective, are often invasive, uncomfortable, costly, and expose patients to harmful radiation. Present study explores the potential of thermography as a non-invasive, cost-effective adjunct tool for breast cancer screening. By utilizing advanced computational techniques and artificial intelligence, the study aims to improve the accuracy and reliability of breast cancer detection through thermal patterns on the surface of the breast. The initial phase of the research investigates the feasibility of using breast skin surface temperature variations, caused by underlying tumors, to estimate tumor size and location. A simplified two-dimensional numerical model is developed using COMSOL Multiphysics software to simulate breast thermal patterns resulting from underlying tumors. A dataset comprising of surface temperatures is generated that are correlated with tumor diameters and locations. This dataset is then used to train an artificial neural network, which demonstrates that thermography can serve as an effective adjunct screening tool alongside the more invasive gold standard technique, mammography. The study further develops a comprehensive numerical thermal image dataset through successive numerical simulations, addressing the absence of actual labelled thermal images. A three-dimensional breast model, representing a spherical tumor within a hemispherical breast, is created to simulate numerical thermal images for different sizes and locations of the tumor. Various machine learning regression models including linear regression, support vector regression, K-nearest neighbor regression, and decision tree regression, are evaluated. Among these, the decision tree regression model shows superior predictive performance, effectively distinguishing minor temperature variations that correspond to tumor characteristics. In the following phase, instead of numerical thermal image data a limited set of temperature data on the surface of the breast used to train a random forest machine learning algorithm. Random forest, which is an ensemble of decision trees, accurately estimates the tumor size and location, demonstrating the effectiveness of using temperature patterns data in breast cancer detection. The final phase of the research integrates real thermal imaging with deep learning to propose a novel, non-invasive breast cancer diagnosis framework. Using the database for Mastology Research with Infrared Image (DMR-IR) dataset, a specialized segmentation algorithm is developed to identify regions of interest within thermal images. These segmented images are then used to train a convolutional neural network based on the AlexNet architecture, which achieves exceptional classification accuracy. This integrated approach, combining segmentation, classification, and thermal analysis provides a reliable and cost-effective system for early breast cancer detection. Present study demonstrates that thermography, supported by computational models and artificial intelligence offers a promising supplementary tool for breast cancer screening. It bridges the gap between non-invasive imaging and precise tumor localization, contributing to the early detection and treatment of breast cancer.Item Investigation on Wire Electro Discharge Machining Characteristics of TiNiCu Shape Memory Alloys(National Institute of Technology Karnataka, Surathkal, 2020) Roy, Abhinaba.; S, Narendranath.Shape memory alloys are well known across academia and industries due to their unique functional capabilities, such as shape memory effect and superelasticity besides other useful properties. They are also known for their toughness, resistance to corrosion, improved fatigue life and damping capabilities. Shape memory effect is exhibited by these group of alloys due to reverse martensitic phase transformation which transforms de-twinned martensites back to twinned martensites. This phase transformation of shape memory alloys occurs without any change in state of the material, which contextually known as diffusionless transformation. Superelasticity, on the other hand is exhibited by these alloys, when the alloy is handled at an operating temperature higher than its austenitic temperature. Ni rich NiTi shape memory alloy for example can be processed to be superelastic at room temperature. These incredible qualities qualify shape memory alloys as potential materials for smart applications such as sensors and actuators. A vast majority of these alloys exhibit shape memory effect due to thermal load and some of them are also influenced by a magnetic field. Thermally induced shape memory alloys have formed wide applicability due to ease of use and economic factor. Among these alloys, TiNi based shape memory alloys are most widely researched and put into applications compared to Cu-based or Fe-based alloys. Phase transformation temperature of TiNi based shape memory alloys lie within a nominal operating temperature range (60⁰C-100⁰C) which makes them more suitable for sensing and actuating applications. However, with addition of a ternary element, phase transformation temperature of these alloys can be tailored to specific needs. Addition of Cu as ternary element in TiNi binary alloy system was found to reduce its phase transformation temperature and narrow transformation hysteresis. Cu addition also facilitates thermal conductivity making it more sensitive to change in thermal flux. Therefore, TiNiCu ternary shape memory alloys could be used for much sensitive applications. Major challenge these alloys impose is poor machinability with conventional machining techniques. High tool wear, poor machined surface quality and additional post-machining processes compromise finish quality, accuracy of the end product and increase the cost involved. This is where non-conventional machining techniques proved as an added advantage to process these functional alloys and soon became a more popular choice over conventional machining techniques. Non-conventional machining process like laser beammachining (LBM), water jet machining (WJM), electrochemical machining (ECM) and electrodischarge machining (EDM) result to better machining characteristics compared to conventional machining techniques. due to non-contact nature of the tool-workpiece interface. However, thick recast layer, oxidation, burr formation are some of machining defects that non-conventional machining techniques exhibit. Wire electrodischarge machining (WEDM) is a variant of traditional electrodischarge machine (EDM) where machining is carried out using an wire electrode. Sparking between wire electrode and workpiece results in removal of workpiece material through local melting. Advantage of WEDM over EDM is that through CNC any desired profile can be cut imposing minimum damage to workpiece material. Sensors and actuators incorporating shape memory effect are generally micro shaped components which undergoes microscopic shape change. Major aim of this study is to investigate WEDM characteristics of various homologous TiNiCu shape memory alloys and to optimize machining responses so as to produce components without compromising accuracy and quality. Six different TiNiCu shape memory alloys were vacuum melted and characterized in terms of microstructure, phases present, phase transformation temperatures and microhardness. Optical microscope with image analyzer, X-ray diffractrometer, differential scanning calorimeter and microhardness tester were used to perform aforementioned characterization. Further, to determine the quality of machining, the following output responses namely material removal rate (MRR), surface roughness (SR), kerf width (KW), recast layer thickness (RLT), machined surface microhardness (MH) and machined surface morphology were studied and reported. Ti50Ni25Cu25 exhibited least thermal hysteresis (~6⁰C) which indicates its suitability as ideal material for sensor and actuator applications. Due to varying thermal conductivity of vacuum melted homologous TiNiCu shape memory alloys, variation in WEDM responses were observed. Thereafter, prediction of WEDM responses was carried out using Artificial Neural Network (ANN) and optimization of WEDM responses was performed using Genetic Algorithm (GA). After a thorough investigation, WEDM process parameters to machine homologous TiNiCu shape memory alloys were reported and discussed in detail.Item Experimental Investigation on Estimation and Prediction of Sound in Percussive Drilling(National Institute of Technology Karnataka, Surathkal, 2013) Kivade, Sangshetty; Murthy, Ch. S. N.; Vardhan, HarshaThis research work was taken up with the objectives of developing general prediction models for the determination of uni-axial compressive strength (UCS), abrasivity, tensile strength (TS) and Schmidt rebound number (SRN) for sedimentary and igneous rocks using penetration rate and sound level produced during percussive drilling. To carry out this investigation fabricated pneumatic drill set-up on the laboratory scale was used. In the present work shale, dolomite, sand stone, lime stone and hematite were the sedimentary rocks, whereas dolerite, soda granite, black granite, basalt and gabbros were the igneous rocks used in this investigation. For all the above mentioned rocks their mechanical properties were determined as per the suggested methods of International Society of Rock Mechanics (ISRM). The laboratory investigation on all the sedimentary and igneous rocks using the drill set-up was carried out to find the penetration rate (mm/s) and sound level (dB (A)) produced by varying air pressure from 392 to 588 kPa, thrust from 100 to 1000 N and with varying drill bits and types (integral chisel drill bit: 30, 34 and 40 mm diameter, threaded (R22) cross drill bit: 35 and 38 mm diameter). The data generated in the laboratory investigation was utilized for the development of regression models for predicting rock properties like, UCS, abrasivity, TS, and SRN using air pressure, thrust, bit diameter, penetration rate and sound level. Further, regression models were also developed for predicting penetration rate and sound level using air pressure, thrust, bit diameter and rock properties as input parameters. In a similar way, i.e. utilizing the same input parameters for determining the rock properties and predicting the sound level and penetration rate, Artificial Neural Network (ANN) models were developed. A comparison was made between the results obtained using various regression models developed and the ANN models. Results of this investigation indicate that ANN models are superior over regression models.Item Development of a Hybrid Recurrent Neural Network Based Intelligent Decision Support System with Reverse Mapping for CNC Machining(National Institute of Technology Karnataka, Surathkal, 2018) Malghan, Rashmi Laxmikant; S. Rao, Shrikantha; D’Souza, R. J.The growth of consumer demands for better quality metal cutting related products has motivated the metal cutting industry to continuously enhance quality control of metal cutting processes. Of the several processes, the face-milling is one of the most fundamental metal removal operations used. It is affected by machining process parameters like cutting force, ambient conditions, coolant type, tool parameters and material properties. Nowadays, diverse types of materials have been used based on the condition requirements like strength, weight, corrosion resistance, etc. Metal reinforced composites have tailorable properties which widen their applications. Machining of composite materials is difficult to carry out due to the anisotropic and non-homogeneous structure of composites and the high abrasiveness of their reinforcing constituents. In this study on milling of AA6061 and AA6061-4.5%Cu-5%SiCp composite, formation of unwanted scratches on the surface of the material were witnessed due to presence of hard particles, resulting in increased surface roughness. Design of experiment is used to analyse the machining process parameters. Taguchi orthogonal array design is used to analyse the levels of the experiment. The Analysis of Variance (ANOVA) is also used to evaluate the contribution of process parameters on milling process output variables for both alloys and composites. The mathematical models for cutting force, surface roughness and power consumption are developed using response surface methodology(RSM). Under utilization of machine capacity limits the efficient use of machines and is presently continually being run at sub-optimal conditions. In this study, a novel technique is introduced wherein the desired depth of cut is achieved with lesser number of passes, lesser time and also by consuming lesser power. Planning a strategy for better machine utilization based on power constraint in machining using PID logic. Further, prediction of responses of milling process are carried out using artificial neural network (ANN) with feed forward architecture using error back propagation learning algorithm. A reverse mapping neural network (NN) has been implemented as a novel architecture, which can derive the input responses, based on the desired system outputs. Reverse mapping approach can be treated as advisory system in absence of human experts, can predict the settings of various process parameters in a milling process to achieve the desired responses as per the requirements of end user. Further this model can be implemented to adjust the process parameters in on-line control of the milling quality. The validity of themodels is established. The ANN models formulated for cutting force, surface roughness and power consumption are found to predict the corresponding responses quite accurately, within the acceptable limits of prediction errors. To explore the dynamic learning capacity of Elman Simple Recurrent Neural Network as advancement over ANN model, the corresponding RNN model was developed. The convergence problem of RNN model was overcome by an innovative way by using Hybrid Recurrent Neural Network (HRNN). The biases and weights are borrowed in a HRNN model with feedback connections, from a partially trained ANN model having similar architecture. The HRNN formulated using this methodology is able to predict the relationship between input and output data and a good correlation is achieved. With reduced learning time, it is observed that an HRNN modelled from a partially trained ANN has equivalent prediction capability and is superior to ANN in terms of computational time. It is noteworthy that, prediction helps the investigator to determine the outputs as well as inputs, but since it fails to estimate the global extreme values of the response responsible for the best product quality (minimum defects). Identifying the extreme values for the conflicting outputs poses difficulties. Traditional methods (DOE, RSM, Grey Relational Analysis and Classical engineering approach) might fail to determine the global optimum values as searches are carried out in single direction. Evolutionary algorithms (Particle Swarm Optimization(PSO)) through their heuristic search mechanisms determine the global solutions at many distinct locations in multi-dimensional space, simultaneously. The lower and upper levels of machining parameters were opted as constraints. The optimized results were cross verified with experiments and found to have good agreement with the experimental values. PSO outperforms Grey Relational Analysis and RSM thus can be utilized as a tool to optimize and predict results during machining of AA6061 and AA6061-4.5%CU-5%SiCp. Graphical user interface (GUI) has been designed using available API libraries which include two main modules, namely, Prediction (both forward and reverse mapping) and Optimization. Each model has the sub components for prediction of cutting force, surface roughness and power consumption. There is provision to obtain outputs by manually feeding the inputs as well for plotting bar graphs by varying one parameter at a time, keeping others constant.