Surjya Kanta Pal

Wire Arc Additive Manufacturing (WAAM) involves strongly coupled, nonlinear and time-varying thermal–geometric dynamics, making autonomous control of layer geometry and surface quality a persistent challenge. Existing data-driven and supervised learning approaches primarily address forward prediction and not well suited for real-time, multi-objective decision-making under evolving process conditions. This paper presents a model-based Markov Decision Process (MDP) framework for closed-loop, real-time optimization of WAAM process parameters, explicitly targeting the control of layer-geometry accuracy and reduction of surface roughness and interlayer heat-input variation. The proposed method formulates current and travel speed selection as sequential decision-making actions within an MDP, supported by Generalized Additive Models (GAMs) to predict layer height, width, and roughness for unseen parameter combinations. A novel multi-criteria reward function is designed to jointly penalize geometric deviation, surface undulations along both layer height and width, and abrupt interlayer heat-input changes, enabling an adaptive trade-off that static or single-objective models cannot capture. MDP is employed to compute optimal policies within the inter-pass cooling time, making the approach suitable for real-time deployment without GPU acceleration. Experimental validation on thin-walled SS308L structures demonstrates that the proposed MDP-based control strategy significantly outperforms open-loop deposition. Compared to conventional parameter selection, layer height and width accuracy improved by 2.12% and 10%, respectively, while average surface roughness was reduced by 10.74% along the width and 25.2% along the height. Statistical analysis confirms tighter confidence intervals for both geometry and roughness, indicating improved process stability. The results establish the proposed framework as a computationally efficient, autonomous decision-making controller for WAAM, advancing automation capabilities for high-precision metal additive manufacturing.

This article presents a deep learning system designed to detect objects on a shop floor, classify them, estimate their pose, and direct a robot to pick them up without collision in real time. This system consists of specific algorithms devised for each of these tasks and has been implemented and tested in an industrial robot. Data collection involved capturing frames of several texture-less cuboid-shaped objects using a low-cost web camera. A convolutional neural network analyzes these frames to detect the object, which is then passed to the image processing module for pose estimation, followed by a sequence scheduling algorithm to determine its grasping. The system achieves 95% accuracy for object detection and 100% for object grasping. The practicality of the solution is confirmed through testing under various conditions such as fluctuating lighting, different colors, undefined orientation, and non-uniform spacing between the objects. The explainability of the model is demonstrated by successfully testing it in these situations. Results show the solution is agnostic to object orientation and background. FurtherFurthermore, it is computationally efficient, accurate, and cost-effective, as it utilizes a webcam, minimizing the overall cost. Additionally, the efficacy of the object detection model is found to be superior when compared to state-of-the-art algorithms. Received: 22 May 2024 | Revised: 26 May 2025 | Accepted: 24 September 2025 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in GitHub at https://github.com/AshishChouhan85/Dataset. Author Contribution Statement Ashish Chouhan: Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft. Shivam Shandilya: Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft. Pravanjan Nayak: Methodology, Validation, Data curation. Debasish Mishra: Methodology, Formal analysis, Investigation, Writing – review & editing, Visualization. Surjya Kanta Pal: Conceptualization, Resources, Writing – review & editing, Supervision.

Aluminum matrix composites have gained attention for high‐performance engineering applications, especially in transport industries for bearing liners, cylinder liners and rings, brake and clutch pads due to their wear resistance and excellent strength‐to‐weight ratio. The lightweight feature of the composites makes it ideal for improving fuel economy, reducing emissions, and optimizing structural integrity. In this study, carbon fiber (CF)/aluminium (Al) composites are fabricated through upward friction stir processing (UFSP), which can achieve uniform fiber dispersion and interfacial bonding between fiber and matrix and ensure uniform properties throughout the processed region. The study reveals a 14% increase in tensile strength, 36% in case of hardness, and 44% in case of impact resistance, with respect to the base Al matrix. Scanning electron microscopy (SEM) and electron backscatter diffraction fractography reveal uniformity in the dispersion of the fiber network within the matrix, which is essential for matrix‐to‐fiber load transfer and resistance to crack initiation. This study provides an insight into upscaling the UFSP technique to fabricate industrially deployable products.

Effective and reliable health monitoring is critical for ensuring the operational safety, efficiency, and longevity of industrial machinery. Multi-layer perceptron (MLP)-based neural networks, referred to as “Universal Approximators,” often face challenges in handling complex, non-linear data due to high computational costs, limited continual learning, and reliance on fixed, non-learnable activation functions. To address these challenges, this article evaluates the potential of Kolmogorov-Arnold networks (KANs) to enhance efficiency and generalization in machinery anomaly detection using a triaxial vibration data across four different health states. Unlike MLPs, KAN employs spline-based, trainable activations on edges, enhancing flexibility and learning efficiency. This architecture improves classification accuracy, outperforming MLP-based neural networks, and conventional autoencoders by 5.2%, 10.64%, and 6.54% on X , Y , and Z -axis data, respectively. Beyond accuracy, KAN demonstrates faster training, superior generalization, improved continual learning, and resilience to catastrophic forgetting. This study establishes KAN as a promising framework for next-generation intelligent machine health monitoring systems for real-world engineering applications such as predictive maintenance, condition-based monitoring, and autonomous fault diagnosis in rotating equipment, spindles in machining operation, wind turbine gearboxes, and industrial fans used in steel and power plants.