Patel College of Science & Technology

Theft of energy is a significant issue worldwide. It not only strains distribution networks but also results in huge losses for power companies. For example, in India, developing countries, non-technical losses due to theft contribute to $20-30 \%$ of electricity distribution losses. Manual inspections and rule-based systems, which are the traditional theft detection methods, are slow, expensive, and ineffective against the new theft detection methods. With the integration of digitalized power systems and smart meters, new possibilities for applying machine learning to identify theft in real-time and with greater precision arise. In this work, a framework based on Decision Trees (DTs) to detect electricity theft using smart meters is discussed. Analysed data consists of 20,000 simulated meter readings with attributes: Voltage, Current, Power, Energy, Power Factor, Load Consumption, and Loss. Preprocessing stages included missing value treatment, normalization, and, again, feature extraction. A 70:30 split was used to create training and testing subsets. A Decision Tree Classifier, using the entropy criterion, was developed in Google Colab using the Scikit-learn and Pandas Python libraries. Exceptional results were achieved, with 99.96% accuracy along with almost flawless precision, recall, and F1-score. The confusion matrix only indicated two misclassifications during the test, which illustrates both the proposed method's effectiveness and the validation of the test. This framework surpasses comparable accuracy, interpretability, and computational efficiency in traditional approaches, as well as in the most advanced studies deployed in machine learning. For these reasons, it becomes extremely efficient and reasonably expected in contemporary smart grid implementations.

With the evolution in wireless communication, there are many security threats over the internet. The intrusion detection system (IDS) helps to find the attacks on the system and the intruders are detected. Previously various machine learning (ML) techniques are applied on the IDS and tried to improve the results on the detection of intruders and to increase the accuracy of the IDS. This paper has proposed an approach to develop efficient IDS by using the principal component analysis (PCA) and the random forest classification algorithm. Where the PCA will help to organise the dataset by reducing the dimensionality of the dataset and the random forest will help in classification. Results obtained states that the proposed approach works more efficiently in terms of accuracy as compared to other techniques like SVM, Naive Bayes, and Decision Tree. The results obtained by proposed method are having the values for performance time (min) is 3.24 minutes, Accuracy rate (%) is 96.78 %, and the Error rate (%) is 0.21 %.

Glaucoma is a disease that relates to the vision of the human eye. This disease is considered as the irreversible disease that results in the vision deterioration. Much deep learning (DL) models have been developed for the proper detection of glaucoma so far. So this paper presents architecture for the proper glaucoma detection based on the deep learning by making use of the convolutional neural network (CNN). The differentiation between the patterns formed for glaucoma and non-glaucoma can find out with the use of the CNN. The CNN provides a hierarchical structure of the images for differentiation. Proposed work can be evaluated with a total of six layers. Here the dropout mechanism is also used for achieving the adequate performance in the glaucoma detection. The datasets used for the experiments are the SCES and ORIGA. The analysis is performed for both the dataset and the obtained values are. 822 and. 882 for the ORIGA and SCES dataset respectively.

Computer network traffic control is a torrid research topic nowadays, as this task helps in various applications like anomaly detection, congestion control and bandwidth control. Different machine learning techniques are used for this purpose earlier, such as autoregressive integrated moving averages (ARIMA), recurrent neural network (RNN), etc. Here a framework on long short term neural network is proposed for network traffic prediction. The proposed framework makes use of real network traces from TIER-1 ISP. These traces are used to make the predictions from the proposed framework that uses Long Short Term Model (LSTM). The aim is to generate the predictions at very short time scales (<; 30seconds). As there is diversity in the network traffic, a feature-based clustering framework is employed to work as the preprocessing stage to cluster similar time series together. The results state that LSTM can be used for the prediction of network traffic with low errors.

Object detection has become an important task for various purposes in our daily lives. Machine learning techniques have been used for this task from earlier but they are used for the classification of image-based species to extract the feature set. This task of deciding the feature set helps to decide the desired object detection. To overcome the object classification problern, this paper proposes a transfer learning-based deep learning method. The different convolutional neural networks (CNN) are studied in this work. Here for the improvement in the result, the majority voting scheme is used. The overall work is carried out on the CUB 200-2011 dataset. The results obtained have shown incredible improvement in the accuracy of the proposed work when compared to the different CNN models.