BITS Faculty Publications

Permanent URI for this communityhttp://localhost:4000/handle/123456789/1867

Browse

Filters

2021 - 20234

Settings

Author: search.filters.author.Sangwan, Kuldip SinghSubject: search.filters.subject.Deep Learning

Search Results

Now showing 1 - 4 of 4
  • No Thumbnail Available
    Item
    An Image-Based Approach for Structural Damage Recognition and Segmentation Using Deep Transfer Learning
    (Springer, 2023-07) Sangwan, Kuldip Singh
    This research aims to determine the feasibility of using image-based deep learning techniques to inspect the damage and recognize its category in the building components. This analysis helps to determine the structure's health and its quantification in terms of damage by using image segmentation. The validation of the proposed approach is done by using PEER Hub ImageNet (Φ-Net), which is a benchmark dataset of structural images. Miniaturized VGG-16 CNN network and its customized version-based architectures have been tested on the dataset to find their adaptability to structural domain classification. To avoid overfitting in the classes with lesser samples, the transfer learning is applied using a feature extractor and fine-tuning strategies. Different experiments are designed to find the optimal model parameters and their scope for a particular image recognition task. To quantify the damage in recognition tasks such as images with cracks or spalling, pixel-based segmentation is implemented to highlight the regions where the damage occurred and its area in the region of interest. The accuracy scores of 97% for a binary class problem reveal the potential use of transfer learning-based deep learning models in structural damage recognition and segmentation even for a multiclass challenging scene.
  • No Thumbnail Available
    Item
    Deep Learning-Based Mosquito Species Detection Using Wingbeat Frequencies
    (Springer, 2022-02) Sangwan, Kuldip Singh
    The outbreak of mosquito-borne diseases such as malaria, dengue, chikungunya, Zika, yellow fever, and lymphatic filariasis has become a major threat to human existence. Hence, the elimination of harmful mosquito species has become a worldwide necessity. The techniques to reduce and eliminate these mosquito species require the monitoring of their populations in regions across the globe. This monitoring can be performed by automatic detection from the sounds of their wingbeats, which can be recorded in mosquito suction traps. In this paper, using the sounds emitted from their wingbeats, we explore the detection of the six most harmful mosquito species. From 279,566 wingbeat recordings in the wingbeat kaggle dataset, we balance the data across the six mosquito species using data augmentation techniques. With the use of state-of-the-art machine learning models, we achieve detection accuracies of up to 97%. These models can then be integrated with mosquito suction traps to form an efficient mosquito species detection system.
  • No Thumbnail Available
    Item
    A Computer Vision Based Approach forDriver Distraction Recognition Using Deep Learning and Genetic Algorithm Based Ensemble
    (Springer, 2021-10) Sangwan, Kuldip Singh
    As the proportion of road accidents increases each year, driver distraction continues to be an important risk component in road traffic injuries and deaths. The distractions caused by increasing use of mobile phones and other wireless devices pose a potential risk to road safety. Our current study aims to aid the already existing techniques in driver posture recognition by improving the performance in the driver distraction classification problem. We present an approach using a genetic algorithm-based ensemble of six independent deep neural architectures, namely, AlexNet, VGG-16, EfficientNet B0, Vanilla CNN, Modified DenseNet and InceptionV3 + BiLSTM. We test it on two comprehensive datasets, the AUC Distracted Driver Dataset, on which our technique achieves an accuracy of 96.37%, surpassing the previously obtained 95.98%, and on the State Farm Driver Distraction Dataset, on which we attain an accuracy of 99.75%. The 6-Model Ensemble gave an inference time of 0.024 s as measured on our machine with Ubuntu 20.04(64-bit) and GPU as GeForce GTX 1080 .
  • No Thumbnail Available
    Item
    Development of a Machine Learning based model for Damage Detection, Localization and Quantification to extend Structure Life
    (Elsevier, 2021) Sangwan, Kuldip Singh
    Structural Health Monitoring (SHM) has been researched for a long time and continues to be an active area of research. Initial work on SHM involved identification of hand-crafted features and predictive models relied on statistical methods. The recent improvements in computing capabilities, coupled with better integration of sensor data, has led to the emergence of more effective techniques in terms of scalability and predictive power. Machine learning offers a solution through automatic feature extraction algorithms, and scalable and noise robust models. Convolutional Neural Networks (CNN) have been used as state-of-art classifiers for images as well as for text. This paper proposes the use of the monitored structure’s transmissibility functions for the structure under observation, which can be fed into a novel composite architecture consisting of Deep CNN followed by multivariate linear regressors to detect, localize, and quantify the damage extent in a system. The proposed method was tested on the Los Alamos’ Eight degree-of-freedom (DOF) structure, and the Structural Beam Data from Laboratory of Mechanical Vibrations and Rotor Dynamics, University of Chile. This study on damage localization and quantification can be leveraged to comment on the safety and soundness of the structure under inspection and can help in making more informed inferences. It is expected that, in general, this will lead to extended structure life, which not only improves the resource utilization in terms of structure maintenance and its longevity but also decreases the carbon footprint and capital expenditure.