Please wait a minute...
, No 9

■ Cover page(PDF 3152 KB) ■  Table of Content, September 2022(PDF 34 KB)

  • Text Independent Data-Level Fusion Network for Multimodal Sentiment Analysis
    Sachin Aggarwal and Smriti Sehgal
    2022, 18(9): 605-612.  doi:10.23940/ijpe.22.09.p1.605612
    Abstract    PDF (194KB)   
    References | Related Articles
    To understand human intentions behind a text accurately and as well as to reduce the tensions and misunderstandings caused by sarcasm and ambiguity, multimodal signals that include visual and audio signals should be utilized. This work is focused on creating a robust model to predict emotion depicted in a video by analyzing the visual and acoustic modalities. This paper has used a data level fusion of the features extracted by different feature extractors used for different modalities of the multimodal dataset. There are two reasons to use data-level fusion as it helps to reduce the computation time which is a major issue in most model-level fusion networks. Also, it will allow us to use the same model for all modalities as the data is already fused and there is no need to create a different model for each modality and then fuse them. Also, to check the effectiveness of the proposal, this work uses four different algorithms, which are BP Neural Network, Deep Neural Network RNN, and XGboost, to see the improvement in the accuracy and the other effectiveness evaluation parameters used in this paper. The datasets which are used for this study are CMU-MOSEI and CMU-MOSI. These datasets were labeled with the help of human agents where a video is classified to be depicting negative emotion or positive emotions. Some of the performance evaluation parameters used in this work are Cross-validation, Error rate, F1 score, Recall, and Precision. After analyzing the results, it was observed that the error rate of each improved model vs their base model was reduced by 8-11% with this approach.
    In-Wheel Motor Design with Thermal and Mechanical Model Analysis for Electric Bikes
    Sandeep Kumar Chawrasia, Aakash Das, and Chandan Kumar Chanda
    2022, 18(9): 613-625.  doi:10.23940/ijpe.22.09.p2.613625
    Abstract    PDF (988KB)   
    References | Related Articles
    The effect of CO2 emission on the environment and climate change and the scarcity of fossil fuels is a global issue nowadays. To resolve this issue, the world is trying to switch from fuel vehicles to electric vehicles. But it is taking too long due to many limitations of electric vehicles. So, to resolve the limitations, various research projects are currently being conducted, on different parts of the electric vehicle. For best performance and high efficiency and torque, BLDC motors are used. The outer rotor BLDC motors provide very high efficiency. Still, the bikes available in the market have quite low speed or are very costly. So, it is important to increase the speed at a cheaper price. And for that, the In-Wheel motor is the best option. The In-Wheel bike with a motor is significantly different from a petrol (or center motor drive) motorbike as the engine (or center motor) is separate from the wheel as any heating of the engine/ motor would not heat the wheels but in the case of the In-Wheel motor, heating of the motor would produce a similar type of heating in the wheels. If the temperatures and the total magnetic flux density of the In-Wheel motor go beyond normal values, then there would be a chance that the electric bike will be heated at a very high temperature and may catch fire too. Also, the pressure on the wheel will affect the performance of the motor. So, it is essential to properly analyze and calculate the temperatures and pressures at different points of the motor. Hence, in this work, a novel approach of the E-Magnetic model, Lab model, Thermal model, and Mechanical model are analyzed properly using Ansys Motor-CAD. The Thermal model is used for getting the best temperatures of various parts of the In-Wheel motor. And the mechanical model is used to get the best pressure at the different points of the motor.
    Naive Bayes and Neural Network Techniques for Marathi Poem Classification into Nine Rasa using Feature Selection
    Rushali A. Deshmukh
    2022, 18(9): 626-636.  doi:10.23940/ijpe.22.09.p3.626636
    Abstract    PDF (769KB)   
    References | Related Articles
    In this approach, we considered classifying poems written in Marathi, one of the popular Indian languages, into nine categories. Using this classification, a person who is unaware of the Marathi Language can come to know what kind of emotion the poem shows. Here we considered tf-idf to represent the features of the poem. We have used Univariate feature selection(Chi2), Tree-based models, L1-based feature selection, and Recursive Feature Elimination to select top-ranked features. For nine categories of poems - Fear, Joy, Love(Prem), Sadness, Vir(Courage), Wonder, Anger, Depression, and Peace - the Naive Bayes classifier achieves maximum accuracy of 85% with Chi2 feature selection. Then we considered six categories of poems for classification. Experimentation is done using Naive Bayes and Neural Network machine learning algorithms. Among all feature selection methods, with Chi2 feature selection, the highest accuracy achieved is 97%.
    Evaluation of Mechanical Behaviour of 3D Printed Structures using FDM Process
    A. Lakshumu Naidu, M. Jaya Krishna, and V. Ram Babu
    2022, 18(9): 637-643.  doi:10.23940/ijpe.22.09.p4.637643
    Abstract    PDF (388KB)   
    References | Related Articles
    3D printing plays a significant role in advanced manufacturing because it is adaptable. It is a type of additive manufacturing technique that prints an object layer by layer to create a 3D model. Fused Deposition Modelling (FDM) is a 3D printing technology procedure in which the material is heated and extruded layer by layer through a nozzle. When compared to other traditional processes such as injection moulding, 3D printing using the FDM process offers greater versatility. This research provides a methodology for evaluating the mechanical characteristics of various structured FDM polymers. Parametric research is provided to evaluate production aspects such as infill structure, which affect the mechanical performance of PLA and ABS-based specimens. However, understanding the mechanical characteristics of 3D printed structures is inadequate. Experiments were carried out in order to get a better understanding of the design and analysis of various types of 3D printed structures manufactured using the Fused Deposition Modelling (FDM) technology. The specimen is designed in Solid Works modelling programme and saved as a Stereolithography file (stl ). QIDI slicing software is used, with the same infill density and orientation. The specimen is sliced with a 40% infill density and a layer height of 0.2 mm. The G-Code file is being created. After slicing, it is produced using an Xmax wol 3D printer using the generated G-Code file. A compression test was done on a square cross section specimen. The test is carried out utilizing a Universal Testing Machine (UTM). Trihexagon, Cubic, Lines, Cubic Subdivision, Gyroid, Triangles, Octet, concentric, cross 3D, and grid are the infill structures employed. This was done to determine which material and infill structure is more durable and has higher compressive strength.
    Double Deep Q Network with Huber Reward Function for Cart-Pole Balancing Problem
    Shaili Mishra and Anuja Arora
    2022, 18(9): 644-653.  doi:10.23940/ijpe.22.09.p5.644653
    Abstract    PDF (535KB)   
    References | Related Articles
    The emergence of reinforcement learning defines a new research direction in control theory where feedback influences the system behavior in order to achieve the desired output. To date, this research work has focused on the cart pole balancing problem using deep reinforcement learning (Deep RL) algorithms. Deep RL is a comprehensive learning framework to study the interplay in environmental input parameters and corresponding output as feedback and further decision making to design a new parameter set to get better output validated in terms of an achieved reward. In this research paper, deep Q network (DQN) and Double deep Q network (DDQN) have been applied to the cart pole balancing problem and reward is measured using a novel loss function - Huber Loss. Comparison results of DQN with MSE and Huber show the fast convergence performance of the Huber loss function. Thereafter, DQN and Double DQN performance is validated by Huber loss itself. Performance outcome shows that DDQN reduced Huber loss and also converged much faster than DQN.
    Comparison of Health Index and Degree of Polymerization based on 2-Furfuraldehyde Compound for Transformer Condition Estimation
    M. S. Yahaya, N. Azis, H. Zainuddin, M. F. Basar, and M. H. Harun
    2022, 18(9): 654-659.  doi:10.23940/ijpe.22.09.p6.654659
    Abstract    PDF (214KB)   
    References | Related Articles
    This paper presents a study on the comparison of Health Index (HI) and the Degree of Polymerization (DP) models based on 2-furfuraldehyde (2-FAL) compound to determine the transformer condition. In total, 30 transformers ranging in age from 13 to 25 years were analyzed. First, the Transformer Health Index (THI) was computed based on yearly individual oil condition monitoring data that consisted of oil quality, dissolved gases, and furanic compounds. Next, the DP was computed based on 2-FAL using five different furfural analysis methods for DP models that are tavailable in literature. Next, the DP of each transformer was plotted against its 2-FAL values. Then, the percentage of DP named as Health Index DP (HIDP) for each transformer was established based on the maximum estimated DP of each model. Finally, the comparison between THI and HIDP were performed. It was found that there is an identical trend between THI and HIDP. However, the average error between THI and HIDP for each transformer is ranging from 8.28% to 37.18% and the total average error is 17.97%.
    Human Activity Recognition using Ensemble Convolutional Neural Networks and Long Short-Term Memory
    Sonika Jindal, Monika Sachdeva, and Alok Kumar Singh Kushwaha
    2022, 18(9): 660-667.  doi:10.23940/ijpe.22.09.p7.660667
    Abstract    PDF (384KB)   
    References | Related Articles
    Recent advances in artificial intelligence have transformed the world into a place where things can be recognized, the surroundings can be learned, and future sequences can be predicted. The advent of advanced technologies has resulted in improving the system and reducing the cost of monitoring systems. This study proposes an advanced ensemble approach of convolutional neural networks and long short-term memory (CNN-LSTM) for human activity recognition. The proposed approach evaluates the spatio-temporal features and recognizes the activities with enhanced accuracy. The method determines the activities by utilizing the RGB, skeleton, and depth-based attributes available in the dataset of UTD-MHAD. The experiments are conducted for the hand/arm-based 21 activities for which videos were captured with the help of depth and inertial sensors. The result evaluations are conducted with the measures of accuracy, precision, recall, and f-measure. These evaluations indicate the superior performance of the proposed ensemble approach compared to state-of-art techniques.
    Portable Learning Approach towards Capturing Social Intimidating Activities using Big Data and Deep Learning Technologies
    Mansi Mahendru and Sanjay Kumar Dubey
    2022, 18(9): 668-678.  doi:10.23940/ijpe.22.09.p8.668678
    Abstract    PDF (288KB)   
    References | Related Articles
    With the rise in the usage of different social media platforms, social intimidation has increasingly spread into these forums as it has given us endless chances to post anything for anyone. Previous studies have confirmed that exposure to this online social intimidation can have very serious offline consequences. With the growth of these multimodal social media platforms, there is an urgent requirement of some device methods for social intimidation detection and prevention. However, most of the prior research has focused on only textual posts for one or two topics of intimidation, namely sexism and racism. The principal objective of this research is to recognize social intimidation from multimodal posts such as text, memes, videos and audio and to target various social media networks such as Instagram, Twitter, and Facebook for several topics of harassment, namely religion based, personal attack, racism, sexism, physical appearance, etc. Previous research has stopped at detection, but this research has taken one step ahead to test the severity based on hate prediction score. The research study is performed using a combination of big data technology, namely Apache Spark, and several deep learning methods which are described below. The system is validated on five public datasets i.e., MLMA Hate Speech Dataset, MMHS150K Dataset, Hateful Memes Dataset, Instagram, Vine Dataset and measured on the basis of precision, recall and f1-score. Performance of the system has been inspected individually for every category of post under three subsections. The results attained specify that the proposed approach gives more feasible solution for social intimidation detection and its severity in online social networking platforms.
ISSN 0973-1318