Binoy. B. Nair

Tungsten inert gas (TIG) welding of AA5083 aluminum alloy is sensitive to heat input and shielding conditions, which can lead to defects such as porosity and burn-through that compromise weld quality. Conventional non-destructive testing methods are generally performed after welding and therefore provide limited capability for real-time process monitoring. This study presents an acoustic emission (AE)-based frequency-domain framework for real-time classification of weld conditions during TIG welding of AA5083. Acoustic signals were acquired at a sampling rate of 10 kHz and transformed into frequency-domain representations using the Fast Fourier Transform (FFT). Decision Tree (DT), Support Vector Machine (SVM), Artificial Neural Network (ANN), and ensemble-learning classifiers were evaluated using the resulting spectral features. To address the high dimensionality of the FFT representation, Principal Component Analysis (PCA) was incorporated into the modeling workflow, with the retained components determined from the model-development data. Five-fold cross-validation and grid-search-based hyperparameter tuning were used during model development, while an independent test set was reserved for final evaluation. The best-performing classifiers achieved a classification accuracy of approximately 0.99 for distinguishing good-weld, porosity, and burn-through conditions. Weld-condition labels were independently validated using visual inspection and radiographic testing. FFT and Short-Time Fourier Transform (STFT) analyses were used to characterize global and time-dependent spectral behavior, respectively, while SHapley Additive exPlanations (SHAP) were employed to identify frequency regions contributing to model predictions. A comparative analysis with time-domain statistical features further demonstrated the stronger discriminative capability of the FFT-derived representation under the investigated conditions. The proposed framework combines high classification performance with interpretable frequency-domain information and provides a basis for in-process weld-condition monitoring and quality control of TIG-welded AA5083.

Dashboard cameras are becoming increasingly prevalent in vehicles, leading to a significant demand for reliable methods to extract essential metadata, such as timestamps, geolocation, and speed, from recorded footage. This metadata is essential for contextualizing events captured by these cameras, facilitating tasks such as accident reconstruction, security monitoring, and forensic analysis. However, many modern low-cost dashboard cameras overlay text directly onto video rather than logging metadata separately. This study aims to develop a solution that is independent of specific devices for extracting metadata from dash- board camera footage. Optical Character Recognition (OCR) technology presents a promising avenue due to its hardware-agnostic capabilities. The research investigates the application of pre-trained OCR models, including Tesseract, KerasOCR, and EasyOCR, to extract overlaid data from images and videos cap- tured by dashboard cameras, while also assessing the limitations and potential failure scenarios of these models. The findings indicate that a combination of traditional OCR algorithms and preprocessing techniques achieves an average Character Recognition Rate (CRR) of 50.6% and a Character Error Rate (CER) of 49.4%. To address the shortcomings of traditional OCR methods, a novel Transformer-based Optical Character Recognition (TrOCR) approach is pro- posed. Extensive training and validation of the TrOCR model, utilizing a mixed 1dataset of real and synthetic footage, significantly enhance character recognition accuracy to 84%, with a corresponding reduction in character error rate to 16%. Furthermore, the incorporation of post-processing techniques results in excep- tional accuracy of 97% and a negligible character error rate of only 0.08%. This research introduces a device-independent approach to extract metadata from dashboard camera footage, demonstrating the efficacy of the TrOCR method and underscoring the benefits of post-processing techniques for live data extraction.

This article presents experimental fault diagnosis research of engines through the integration of bidirectional transfer learning, federated learning, and a new hybrid architecture of Transformer and Deep Neural Networks (DNNs) for out-of-distribution (OOD) testing. For the research purpose, two unique engine datasets were chosen: one for standard car engines and generators, and the other for engines of remote-controlled model cars and drones. For the initial step, bidirectional transfer learning through the employment of traditional DNNs was utilized to eliminate the domain difference among these heterogeneous engine datasets so that fault diagnosis can be effectively achieved for small and large engines. The second step was to develop a federated learning system to facilitate collaborative DNN model training on decentralized engine data to ensure data privacy with overall enhanced generalization. Lastly, the research proposes a federated transfer learning approach that employs Transformer-DNN hybrid architecture. The Transformer component, with positional encoding and multi-head attention mechanisms, was tailored to extract sequential dependencies in engine operational data, whereas the DNN part was dedicated to discriminative feature extraction for classification. Outstanding results show that transfer learning and federated transfer learning are highly effective in fault diagnosis of small engines; however, OOD testing based on the hybrid architecture substantially enhances accuracy in fault detection for large engines. Comparative experiments show significant improvements in classification accuracy and robustness in comparison to baseline DNN models. These results form a convincing basis for the development of scalable, privacy-preserving predictive maintenance systems that are flexible for deployment on a range of engine applications within both automotive and unmanned systems environments.

Fault diagnostics in internal combustion engines (ICEs) is vital for optimal operation and avoiding costly breakdowns. This paper reviews methodologies for ICE fault detection, including model-based and data-driven approaches. The former uses physical models of engine components to diagnose defects, while the latter employs statistical analysis of sensor data to identify patterns indicating faults. Various methods for ICE fault identification, such as vibration analysis, thermography, acoustic analysis, and optical approaches, are reviewed. This paper also explores the latest approaches for detecting ICE faults. It highlights the challenges in the diagnostic process and ways to enhance result accuracy and reliability. This paper concludes with a review of the progress in fault identification in ICE components and prospects, highlighted by an experimental investigation using 16 machine learning algorithms with seven feature selection techniques under three load conditions to detect faults in a four-cylinder ICE. Additionally, this study incorporates advanced deep learning techniques, including a deep neural network (DNN), a one-dimensional convolutional neural network (1D-CNN), Transformer and a hybrid Transformer and DNN model which demonstrate superior performance in fault detection compared to traditional machine learning methods.

Deep learning-based models are ideally suited for accurately predicting electrical load in a smart grid. However, the computational overhead in training models and identifying optimal training hyperparameters are challenging problems. Two important challenges are addressed in this study. Firstly, a novel technique is proposed to accelerate the model training process by selecting optimal mini-batch size. The second challenge encountered is maintaining and operationalizing individual consumption forecasting models as they are not easily scalable. A methodology is proposed to generate deep learning-based models capable of forecasting energy consumption simultaneously for large user groups. The proposed methodology was empirically validated on data from 4710 individual users in ISSDA CER Dataset (approximately 25,632 datapoints/user) and 11 Feeder dataset data from NYISO (approximately 106163 datapoints/feeder). Eight deep learning algorithms trained using eleven mini-batch combinations, for two forecasting horizons resulted in 176 workflows. Training on 167 sample CER ISSDA user data resulted in 24,288 models and 75,360 individual models for the optimal mini batch. Similarly, 1,936 models were generated for the 11 NYISO Feeder datasets. A clustering approach is employed to group users with similar consumption patterns resulting in 107 models. Effectiveness of models is validated with eight deep learning algorithms trained for two forecasting horizons. Results demonstrate a 93.74% improvement on an average in convergence compared to conventional mini-batch sizes, achieving one among top three accuracy across various prediction horizons and user categories. This underscores the importance of tailored mini-batch selection methods for effective energy forecasting and management in Smart Grid systems.

Malware detection is an ever-evolving area given that the strides in the detection capabilities being matched by radical attempts to bypass the detection. As the sophistication of malware continues to increase, the demand for innovative approaches to improve detection capabilities become paramount. Machine learning/Deep learning models are being increasingly used for Malware Detection, however one of the most important and frequently overlooked aspects of building such models is feature encoding. This research paper explores the importance of feature encoding to improve the efficiency of threat detection and proposes a novel entropy-based encoding scheme for the categorical features present in the data extracted from malicious inputs. The KDDCUP99, UNSW-NB15 and CIC-Evasive-PDF-Mal2022 datasets have been used to evaluate the effectiveness of the proposed encoding scheme. The results of the proposed encoding scheme are validated against seven other encoding schemes to ascertain the credibility and usability of the proposed scheme. The efficiency of the proposed system evaluated by applying different encoded versions of the datasets to train various machine learning models and determining the classification performance of the models on each dataset. The machine learning models trained with the proposed encoding scheme produced stable classification results and outperformed other encoding schemes when dimensionality reduction was applied on the data. The ensemble classifier trained using the proposed scheme was able to classify the data with an F1 score of 99.99% when the dimension-reduced entropy-encoded KDD Cup99 dataset was used to build the model. On the CIC-Evasive PDF-Mal2022 dataset, the entropy encoding has exhibited a slightly improved classification parameters with the ensemble methods yielding a peak F1 score of 99.27%. We have also determined the feature importance values of the features present in the datasets to study the change in the contribution levels of the features when multiple categorical encoding schemes are applied upon the data.