A nonlinear control scheme is employed in this paper for controlling the speed of DC motors which are driven by DC-DC buck converters. The hybrid robust adaptive backstepping sliding mode control theory is used to design the proposed controller. To handle the sudden load variation effects, the load torque is considered as unknown terms which will estimate using an adaptation law. Exterior turbulences are also considered in the system's dynamical model to illustrate the robustness of the proposed hybrid control scheme. The overall system's stability is established using the Lyapunov theory. At last, simulation results are presented to exhibit the usefulness of the proposed control scheme in the presence of external disturbances and compared with an existing controller. Simulation results reveal that the proposed controller outperforms the existing controller in terms of overshoot and settling time.
This paper introduces a new static output feedback controller design method for continuous-time Takagi-Sugeno fuzzy systems, which can represent a wide class of nonlinear systems. Although many papers discuss state feedback controller design for fuzzy sytems, a fewer number of papers consider an output feedback controller design due to its difficulty. In this paper, our target is to propose an output feedback controller design method. In our control design, the integral structure of the membership functions, which are the same properties as the original membership functions, is utilized. A similar approach for state feedback controller design methods has been in the literature. Our proposed output feedback controller is designed based on a new fuzzy Lyapunov function which includes the double integral of the original membership functions of the system, and it shows a wider stabilizing area than the existing methods. At the end of the paper, a numerical example is provided to illustrate the newly proposed controller design method and to asymptotically stabilize the system.
A directional wideband microstrip line fed rectangular patch antenna has been proposed for the 28 GHz 5G applications. Initially, a conventional rectangular microstrip patch antenna has been designed thereafter to enhance the performances as well as to tune in the desired operating frequency range (24 GHz-30 GHz) a cutting edge with partial ground plane technique is used. At the intermediate stage, a star-shaped parasitic element has been used for the gain enhancement of the antenna. However, the high gain of the antenna is retained by optimizing the dimensions and position of the cutting edge as well as the ground plane without introducing any parasitic element on the patch which is the simplicity of the proposed antenna. The proposed antenna operates at the 28.201 GHz band ranging from 26.984 GHz to 29.551 GHz. It has an average radiation efficiency of 84.5%. As the proposed antenna possesses a smaller size ($\mathbf{20}\times \mathbf{17}\times \mathbf{1.575}\ \mathbf{mm}^{\mathbf{3}}$) and wide bandwidth of 2.567 GHz with a high gain and directivity over the entire operating band, it can be considered as a potential candidate for the 28 GHz millimetre wave (mm-wave) applications.
Intracranial Hemorrhage is a common brain injury that leads to a high mortality rate without prompt recognition. To address these issues, computer-aid diagnosis tools are rapidly being developed along with neural-network-based techniques to provide fast, reliable analysis and achieve accurate diagnosis decisions based on medical images. One of the most interesting applications in computer-aid diagnosis is Image Registration due to its practical features in clinical diagnosis and treatment planning. In this study, we present the non-rigid image registration for the 3D Computed Tomography image dataset of the Intracranial Hemorrhage Brain. By utilizing the affine transformation and a neural network model, we aim to predict the deformation vector field, map the real-world-collected dataset to the reference space and overcome the shifting data problem between the data analysis experiment on standard and real-world medical image analysis. Our test results gave that good registration performance is obtained in a very short time by using a neural network model, and the affine transformation significantly improves the real-world image registration. In addition, according to the distance from the hematoma area change ratio to the brain area change ratio, the characteristics of the major structure are determined to be preserved.
The ever-increasing demand of smart sensors for Internet of things applications, drove the change in outlook towards smart sensor system design. This paper focuses on smart sensing approach using single pristine tin-oxide (SnO2) sensing film based sensor. It is a novel attempt to not only identify but also quantify the bio-marker gases like ammonia (NH3), carbon monoxide (CO) and hydrogen sulphide (H2S) using single sensor with intelligent pattern recognition algorithms and its readout circuit. The presented work uses temperature modulated gas sensor response obtained at different concentrations for mentioned gases with in-house developed sensor for gas discrimination and quantification. This work uses wavelet features with Independent Component Analysis (ICA) based feature extraction technique and proposed Adaptive Gradient Boosting classifier for gas sensor response pattern recognition. The proposed approach provides gas detection accuracy 93.28%. Further, our approach uses maximum likelihood fit to quantify concentration of the predicted gas and reported the average MSE for all three gases is 0.11, which makes the proposed work smart to mitigate the cross-selectivity of MOS based smart sensor design in presence of drift and noise.
The Human-like ability to recognize emotion from speech has been an interesting field of research for quite a while now. In contrast, the emotion recognition from the sound of the crowd is a relatively new domain. Crowds express emotion as a collective group where the individual sounds combine together to make up emotions like cheering, booing, clapping, etc. As a result, recognizing emotion from crowd sound is very different from recognizing emotion from an individual's speech. Moreover, the lack of any large and diverse dataset makes it harder to perform machine learning analysis in this domain. In this paper, we present a relatively large and diverse dataset of the emotional sound of crowds collected from 70 different large crowd events. We collected data for 3 different types of emotion and organized the dataset into 5 different folds each containing a unique set of events. The diversity and organization will ensure the reliability of a machine learning model trained on this dataset. We also discuss the effectiveness of 34 different features and 2 analysis techniques on the proposed dataset. The dataset has been made publicly available for the community.
The main challenge of processing Hyperspectral Image is its high dimensionality. Most of the machine learning classification algorithm's accuracy diminishes as the dimensionality of feature increases. As a result, for working with Hyperspectral image classification, complex feature reduction techniques are performed. The proposed architecture of CNN hierarchically constructs high-level features by seeking low dimensional representation of Hyperspectral interpretation in an automated way, instead of working with a full spectral band or complex handcrafted features. Principal Component Analysis (PCA) is used to remove highly correlated spectral bands before feeding into the network. A combination of 3D and 2D convNet layers are used to preserve both the spectral and spatial information of the Hyperspectral data. A logistic regression layer is responsible for the classification task. The overall classification accuracy of the demonstrated approach is 99.43% which is much better than other conventional machine learning and deep learning methods.
Shipping containers provide numerous benefits to global transportation. They are used to transport cargo from more than 30,000 cargo ships sailing across the world. The shipping containers provide the best protection of goods. This is because once all the goods are loaded into the container, it is sealed completely. The objective of the container seal is to minimize the risk of someone accessing the container and taking cargo out and avoid someone putting illegal stuff into the container such as drugs, weapons of mass destruction. To this end, shipping container terminals are required to inspect security seals when containers pass the gate of intermodal terminals. The existing detection mechanism is based on the human visual system which is time-consuming and hazardous. In this paper, a deep learning-based framework is proposed to automate shipping container security seal detection. The proposed method consists of three components including, handlers and cam keepers detection, handlers and cam keepers super-resolution regions, and security seal classification. For handlers and cam keepers detection you only look once (Yolov5) is employed to detect them with high performance. Following that, the laplacian pyramid super-resolution network (LapSRN) image super-resolution technique is used to convert low-resolution handlers and cam keepers regions to high-resolution sub-images. Finally, EfficientNetB0 is employed to classify the super-resolution sub-images based on two categories, seal or no-seal. The proposed whole security seal detection system is trained end-to-end that can localize and recognize the regions containing security seals with high performance.
Automatic pain intensity estimation is emerging as a crucial requirement in painful analysis and has showed a high value in health care applications. The correctness and computational complexity are the two required keys for a pain assessment application. Inspired by the fact that facial expression of pain can be described by the pain-related Facial Action Unit (AU) combinations, we propose a novel three stages training for pain intensity estimation approach. Based on the Inception Resnet architecture, we firstly learn to predict AUs from the combination of DISFA and UNBC McMaster databases using Heatmap regression. Secondly, we learn to predict pain level as linear regression. Lastly, we use a LSTM model to exploit the temporal relation between video frames. With this approach, we take advantage of two databases for regularization to prevent overfitting, while still maintaining the efficiency in term of complexity of the Inception Resnet architecture. Experiments on the UNBC McMaster database show that our approach yields a promising result compared to the state-of-the-art in automatic pain intensity estimation. The code for testing and the models are available to download from https://github.com/glmanhtu/3stages
The electrocardiogram (ECG) is one of the simplest and oldest tools to assess the heart condition of cardiac patients. Heart diseases have emerged as one of the leading causes of death all over the world. According to the world health organization (WHO), millions of people are dying every year from heart-related diseases. A classification model that can early detect arrhythmia will be able to reduce this number by manyfold. Many researchers are working in this area and proposed many deep learning and machine Learning based models for arrhythmia classification. These models have high accuracy but require a machine with high computational power. Hence, these models are not sustainable options for the practical field. In this paper, we have proposed a 1D Convolutional Neural Network (CNN) model with high accuracy and low computational complexity. Our proposed methodology is appraised on the MIT-BIH arrhythmia dataset. We achieved overall 98.25% accuracy into five classes with an f1 score of 98.24%, precision 97.58%, and recall 96.79% which is better than previous results classifying arrhythmia. We can claim that our proposed method is better than most other existing models because of the higher accuracy with a simple architecture that can be run on an edge device with relatively low hardware configuration.
Skin cancer has become a severe problem for medical diagnosis. The adoption of Artificial Intelligence (AI) in pharmaceutical diagnosis is constantly improving. Lately, AI-based computer-aided diagnostics explications for the diagnosis of skin disease have been of prominent concern. Notwithstanding its importance, skin lesion segmentation inhabits an unresolved difficulty of variability in shade, texture, patterns, and obscure borders. Melanoma, precisely called malignant melanoma, is the deadliest kind of skin cancer. It is much more spread to other body organs if it remains undiagnosed and is not treated early. Hence, early screening has the utmost importance in enhancing the cure probability. The proposed algorithms significantly impacted skin cancer classification until today, but they decayed classification rates while using Non-Dermoscopic Digital Images. This paper presents a new approach of classifying skin lesions from non-dermoscopic digital images using Convolutional Neural Network and Neutrosophic Logic Support Vector Machine (CNN-NSVM) combined approach. CNN extracts the features from the images, and the neutrosophic logic employed with SVM classifies the skin lesion types. The proposed methodology is evaluated on the well-known malignant lesion image dataset from the digital image archive of the Department of Dermatology of the University Medical Center Groningen (UMCG). The obtained accuracy of the proposed algorithm is 91% which outperformed MED-NODE by 10% on a similar dataset.
Bangla Speech-To-Text (STT) conversion is a technology that provides a means of converting spoken Bangla language to a written Bangla text form. The standard of speech recognition in different languages is rising step by step however Bangla speech recognition has drawn exceptionally little attention. Building up an STT framework is a bulky procedure and it requires a few stages. Deep neural network-based architecture replaces the stages with neural network components which makes the task simpler and removes the dependency on hand-engineered rules. CNN-RNN networks with CTC criterion are utilized in this undertaking to construct a Bangla STT system that generates text from speech. The architecture is trained solely on speech samples and text transcripts. We used 215.53 hours of speech data set for training, which includes a wide variety of speech samples collected by people of different ages and genders. This paper shows a comparison between genuine text transcripts with produced text transcripts for a similar sound example. Comparison of results between implemented architecture and already existing Bangla STT has also been presented in this paper. The paper is concluded with a discussion about the word error rate and implementation challenges. The key contribution of this paper is to develop a gender and speaker-independent continuous speech-to-text conversion system for the Bangla language using deep learning.
The unacknowledged usage of Intellectual Property (IP), for example, academic or scientific writings, is considered plagiarism. People get involved in this unethical practice in order to achieve rewards or society's attention effortlessly. For example, students often copy & paste from other's documents to achieve better marks. However, it is crucial to detect plagiarism to reward the actual owners of IP like academic or scientific writings. The existing studies are better at identifying external online sources of plagiarism for a given document. However, busy academics need to identify plagiarism among a set of offline writings or documents more often. To this end, we develop a tool named AcPgChecker to assist busy academicians. Initially, this tool measure similarity among a set of documents using a well-known information retrieval technique, Cosine Similarity. Then, it compares the measured similarities with a predefined threshold to detect plagiarism. The main attractions of this tool are: it is open source and freely available for anyone whereas equivalent existing tools are very expensive.
Adaptive Network-based Fuzzy Inference System (ANFIS) is a promising model of explainable neural networks but rejection of illegal noise effects is an important issue in real application. In this paper, a novel approach for introducing noise clustering concepts into fuzzy $c$ -means-based ANFIS is proposed for robust modeling. In the premise part, noise fuzzy clustering is performed in the input data space for estimating fuzzy membership functions removing noise inputs. Then, in the consequence part, rule-wise robust regression models are estimated by removing noise outputs. As a result, the proposed hybrid robust ANFIS model simultaneously considers two types of noise generation schemes of the input-level and the output-level. The characteristics of the proposed method are demonstrated through numerical experiments such that input-level noise are rejected by degrading premise fuzzy memberships of noise objects so that their ANFIS outputs have small absolute values while output-level noise observations are rejected through robust regression.
Training supervised deep learning approaches requires a huge amount of labeled data. In the case of object segmentation on images, the creation of labeled data is expensive. Thus, creating a simple, fast and intuitive labeling process is of interest. Such an approach for object segmentation is to outline objects on a touch interface. However, outlines drawn on a touch device are too rough to be used as labels directly. To improve this, interactive segmentation algorithms can be applied. The results can be improved by utilizing prior information about the outline. In this paper, we examine what kind of information about user-drawn outlines can be used to improve the results of interactive segmentation algorithms on the use-case of segmenting stains on images of laundry. To do this, a user study has been conducted in which participants had to draw outlines on ten different images with varying amounts of stains. Our findings suggest, that 1) outlines often contain multiple stains, 2) outlines often cut through stains and 3) the shape of an outline is seldom similar to the shape of the contained stains. Thus, many assumptions integrated into established interactive segmentation algorithms do not hold which prevents their usage in a fast and intuitive labeling tool.
Real-time character animation for gaming and film industries is challenging and achieving production-ready quality requires a huge amount of time and resources. Animation through marker-based motion capture is quite a tiresome process that requires costly motion-capture suits, multiple cameras, and a large database. In this paper, we propose a model that aims to generate real-time character animation for biped locomotion in Unity ML(Machine Learning) agents using RL(Reinforcement learning) and IL(Imitation learning) algorithms. We first evaluate the training process with solely the state-of-the-art RL algorithm, PPO(Proximal Policy Optimization). Then we analyze the combination of IL algorithms BC(Behavioral Cloning) and GAIL(Generative Adversarial Imitation Learning) in conjunction with PPO. We further discuss the comparison between the two training results and show that our model can generate animations in real-time avoiding all the tedious work and large databases. We demonstrate that our approach is effortlessly easy to implement while maintaining the quality of the animation.
A nonlinear hybrid control approach is proposed in this paper to control the speed of a DC-DC buck converter driven DC motor. An adaptive backstepping sliding mode control theory is used to design the proposed controller. To handle the sudden load variation effects, the load torque is considered as unknown terms which will be estimated using an adaptation law. The theoretical stability of the system is ensured by using the Lyapunov theory. Finally, a simulation study is conducted to demonstrate the effectiveness of the proposed control scheme, and the performance is compared with an existing controller. Simulation results reveal that the proposed controller outperforms the existing controller in terms of overshoot and settling time.
Every year, the number of motor dysfunction patients has been rising. These patients require physical therapy and continuous observation and assessment of their exercises by a professional therapist. This process can take a longer time, leading to a staff shortage and increasing financial costs. Thus, a reliable rehabilitation framework is necessary to assess these exercises as precisely as possib...
We introduce research on social augmented reality with an emphasis on social face-to-face interaction. In social interaction interactants employ knowledge about their conversational partner and have their verbal interaction supported by nonverbal interaction cues. We survey the issues that arise when we pursue social interaction in augmented reality. Since handhelds and bulky head-mounted devices hardly allow unobtrusive interaction we pay extensive attention to developments in the field of smart glasses and smart contact lenses. The focus is on the use of smart augmented reality glasses during social interactions. Acceptance issues and disruption of social interaction due to the use of these devices are also touched upon.
Parkinson's Disease is caused by a decline in the production of dopamine due to the degeneration of brain cells. Dopamine is responsible for the communication between parts of the brain associated with the control and fluency of body movements. Hence, the disease manifests with a spectrum of movement disorders as well as non-motor features. It is now revealed that the non-motor symptoms may show many years prior to the onset of motor symptoms. Therefore, early and accurate diagnosis is crucial to stop or slow down the progression of the disease in its tracks. In this context, ensemble machine learning (ML) algorithms like boosting algorithms can play a significant role in detecting Parkinson's Disease at an early stage. In this paper, four boosting algorithms are studied and implemented in UCI Parkinson's Disease dataset. After rigorous simulation, the ML models exhibited satisfactory results in terms of different performance parameters like accuracy, precision, recall, F1-Score., AUC, Youden, specificity and error rate. However, the performances of the model are improved by tuning the hyperparameters with GridSearchCV. Hence, a detailed comparative analysis is portrayed where Light GBM displayed the highest accuracy of 93.39% after hyperparameter tuning. However, XGBoost and Gradient Boosting algorithm also depicted accuracies more than 90% but AdaBoost demonstrated maximum 87.22% accuracy with hyperparameter tuning.
