Reliable anomaly detection and isolation are essential in safety-critical monitoring. Conventional principal component analysis (PCA) methods treat data from different sensors uniformly, which limits their ability to capture localized anomalies. This study proposes an Attention-PCA (A-PCA) framework that integrates spatial and temporal attention with PCA, enabling adaptive reweighting of sensors before projection. The PCA residuals are monitored using a multivariate exponentially weighted moving average (MEWMA) chart, with control limits estimated via kernel density estimation (KDE) to improve robustness under non-Gaussian conditions. Furthermore, anomaly isolation is enhanced through local interpretable model-agnostic explanations (LIME) applied in the PCA residual subspace, which provide clear sensor-level attribution and highlight variable-specific contributions to abnormal conditions. The approach is validated on a wheelchair posture dataset with three seat pressure sensors and one backrest ultrasonic sensor, showing effectiveness in detecting and isolating unbalanced seating conditions. Results demonstrate that A-PCA-MEWMA outperforms conventional PCA- and A-PCA-based monitoring schemes that rely on Hotelling’s T2 statistic and the squared prediction error (SPE). With both parametric and KDE thresholds, the proposed method achieves higher accuracy, reduced false alarms, and reliable isolation of right, left-, and forward-leaning abnormal postures. These findings confirm the effectiveness of combining attention mechanisms with explainable AI for practical and interpretable anomaly detection in human-centered monitoring systems.
This work presents a learning-based motion control framework for automating high-frequency (HF) brazing of copper tube joints, a repetitive yet unstructured task still dominated by manual operation on refrigerator manufacturing lines. We develop a digital-twin training stack using PyBullet and an RB5850 manipulator model, and formulate approach-braze-retreat behaviors as a continuous Cartesian control problem. A PPO actor-critic with convolutional visual encoders consumes RGB inputs (3x240x480) and outputs fine-grained end-effector increments (0.01 m resolution). To stabilize updates, we employ clipping-based surrogate objectives and separated learning rates for actor and critic, together with a replay buffer and mini-batch training. Simulation accelerates data collection and improves sample efficiency, yielding stable policy improvement within a few hundred epochs in simulation with monotonically increasing episodic rewards. We describe system integration toward real-world deployment, including induction-heating end-effector design and multimodal sensing (thermal/RGB/distance) for reward shaping and quality assurance. Ongoing work targets transfer from sim-to-real, closed-loop temperature/position control during brazing, and quantitative weld-quality evaluation under fixed-posture testbeds. The results indicate that PPO with a task-aware digital twin is a promising path to robust, generalizable HF brazing automation across product variants while reducing dependence on expert operators
This paper presents a motion planning approach designed to achieve energy efficiency in redundant robotic systems. The strategy proposed in this work is based on the optimization of the position of one of the redundant joints of the robot in each of the way points that define the task, while maintaining a fixed execution time. The configuration of the robot is then obtained through inverse kinematics, and the final trajectory is computed by interpolating the optimal joint positions of the robot in consecutive way points. Numerical and experimental results validate the proposed approach using a manipulator with seven degrees of freedom, and demonstrating the feasibility of the strategy in enhancing energy efficiency in redundant robots.
Accurate power measurement is essential in precision mechatronic devices, especially in embedded platforms where energy efficiency and reliability are critical. However, sensors like the INA219 are susceptible to inductive electromagnetic interference (EMI), often producing noisy power readings near motors or switching converters. This work presents a lightweight RMS-based filtering technique that estimates effective power via numerical integration over fixed sampling windows. Designed for real-time microcontroller implementation, the method avoids hardware filters and minimizes computational load. Validation is conducted on a solar-powered environmental monitoring robotic platform (The RaccoonBot) operating under strong EMI. Results show improved signal stability and fidelity compared to unfiltered and moving-average methods. The proposed approach enables reliable power sensing through embedded computation alone, offering a scalable solution for EMI-exposed, energy-constrained environments.
This work aims to present a new mapping methodology for AGV (Automated Guided Vehicle) systems. The goal is to achieve vehicle positioning through a PID (Proportional, Integral, and Derivative) controller, without requiring physical interaction with the path to be followed. The article provides a broad overview of data processing and applying techniques related to achieving accurate positioning without external feedback. This article presents satisfactory results, even though some sensor limitations must be considered.
This study is intended to design a functional prosthetic hand to increase mobility in individuals with wrist disarticulation amputation. Prosthetic hands play an important role in restoring critical hand functions, however, many existing models remain expensive, complex, and inaccessible, particularly in areas where medical resources are scarce. With the introduction of modern control systems designed to mimic natural hand movements, the prosthetic hand industry has undergone significant advancements. This study provides the design and concept of a functional prosthesis that gives priority to grip strength, cost-effectiveness, and durability. The structure of the prosthetic hand includes five functional fingers, each with three degrees of freedom, except for the thumb which has two. The structure allows the hands to perform a wide range of grips and finger postures that would mimic the natural hand movements more closely.
Intelligent unmanned aerial vehicles (UAVs) can effortlessly execute large-scale and complex missions due to their maneuverability and autonomy. Networks of heterogenous UAVs carrying various equipment and resources offer more opportunities to execute the tasks that single UAV may fail to do it alone as multiple UAVs can form coalitions and cooperatively share their resources and complete the missions. In this paper, two novel algorithms have been proposed to tackle the challenges that engulf the problem of distributed task allocation and coalition formation with multiple vehicles. A dynamic game-theory-based algorithm named GHOST where the UAVs autonomously act as rational players and move according to their preferences to choose the members and sort the tasks for their coalitions. In addition to an evolutionary algorithm with 3-generations (3GEA) used for planning the coordinates of UAVs. This algorithm makes use of an archive of previous best solutions and a shallow FNN (feedforward neural network) trained with multiple supervised algorithms to improve the convergence and diversity of solutions. The comparative analyses with more than 20 state-of-the-art clustering and evolutionary algorithms proved that the proposed algorithms could achieve optimal coalition structures and complete missions with a success rate of 85-100%.
PROFINET is a widely used industrial Ethernet standard which enables real-time communication between controllers and devices in industrial environments. This paper introduces PyPND-Device, a Python-based PROFINET IO Device implementation designed to run on low-cost hardware platforms such as Raspberry Pi. By leveraging the IO Base Interface of the PROFINET Driver and CPython bindings, PyPND Device enables real-time cyclic communication with PROFINET controllers. The proposed solution demonstrates how developers and testers can achieve a complete hardware-free PROFINET test environment, facilitating rapid development cycles and cost-effective testing. Experimental evaluations validate the feasibility of using Raspberry Pi for real-time PROFINET Device operation while maintaining strict timing requirements.
This work aims at the validation of the PIDA control algorithm performance. Control system performance analysis is conducted for the high-order integrating system consisting of three cascaded interconnected tanks. The analysis is done using MATLAB / Simulink environment. Control system based on PIDA algorithm is applied to the identified model of real industrial plant. Its identification is carried out using correlation analysis and pseudo random binary signal (PRBS) based experiment. The analysis addresses process variable filtering, Smith predictor, and compensation for valve nonlinearities. Moreover, the original PIDA algorithm is modified according to industrial requirements. The research focuses not only on control performance analysis, but also on proper actuators utilization in the energy-aware control. Results prove that properly deigned and tuned PIDA algorithm, especially using well designed acceleration term accompanied with proper filtering, allows to obtain significant control quality improvement and well decoupled disturbances. Proposed PIDA configuration enables to reach high control performance and energy-aware utilization of actuators.
This study proposes a control architecture that enables autonomous operation of an electric bike by adaptively switching between stabilization and trajectory tracking using sliding mode control. Stabilization control stabilizes the camber angle using a state-space model derived from an inverted pendulum representation. For trajectory tracking, a path-following method based on a kinematic model is used to control accurately the steering angle and vehicle path. The system monitors the estimated camber angle via an inertial measurement unit and switches to stabilization control when a predefined threshold is exceeded, preventing instability. The proposed method was validated through simulations and experiments demonstrating that it ensures high stability and accurate path tracking while avoiding interference between the two control modes.
Modeling and controlling actuators driven by smart materials, such as shape-memory alloys (SMAs), is challenging due to their strong nonlinearities, hysteresis characteristics, and Multiphysics coupling, which limits the effectiveness of classical model-based control. This paper presents and evaluates a data-driven Q-learning controller for SMA-based actuators that learns a state-feedback policy without requiring an explicit model of the plant. The target system is a single-input, single-output, 3D-printed compliant actuator driven by SMA wires, tasked with real-time trajectory tracking. The approach bypasses system identification and controller tuning in favor of policy learning via the action-value Q-function. In real-time simulation, the learned policy reliably tracks diverse reference signals (constant, sinusoidal, square). The mean absolute tracking error ranges from 0.8 to 1.9 percent, indicating good practical accuracy even though perfect tracking is not achieved—an acceptable trade-off for many smart-material applications.
Conflict Based Search(CBS) is one of the best-known methods to solve MAPF optimally. Its performance, however, degrades in large scale instances because it involves sequential high-level search and uninformed conflict resolution. In this work, we present the DCPCBS (Dynamically Conflict Prioritized CBS) algorithm with three enhancements to CBS that are intended to enhance scalability and runtime performance without compromising on completeness and optimality. First, we augment CBS with earliest-k conflict detection to allow the algorithm to detect and consider several early conflicts per node and branch on the most important one. Second, we add a conflict prioritization component that gives a dynamic criticality score to each conflict based on frequency of occurrence, proximity to agent goals in space, and proximity to temporal objectives, and direct the search towards the most influential conflicts. Third, we introduce parallel constraint tree (CT) expansion via multi-core processing and adaptive batching, enabling simultaneous resolution of multiple CT nodes to speed up search. All these improvements result in dramatic runtime and node expansions reductions without any degradation of solution quality. We experimentally validate our approach on challenging grid environments with high-density agent populations and show improved scalability.
In this study, the systematic investigation of layer size and neuron distribution in machine learning configurations for control theory was conducted. The analysis reveals that the changes in layer configuration and neuron distribution have an impact on system identification by the ML and control signal tracking performance. The obtained results demonstrate that equal-split and heavy-split structures in different multi-layer networks outperform alternative configurations, as shown by improvements in RMSE, MAE, and R2 metrics. These findings highlight the importance of carefully planning architectural choices in control schemes that are enhanced by machine learning. The proposed approach provides practical design guidelines for integrating neural networks in robotic control applications. Also, the tracking performance of the proposed ML-integrated control schemes was demonstrated in the numerical simulations.
We present an embedded, closed-loop controller that prevents defects in fused-deposition modeling (FDM) by combining a probabilistic defect estimator with a vision verifier and a constraint-handling model predictive controller (MPC). A 10-D multi-sensor feature vector drives a logistic Defect-Probability Estimator (DPE) refreshed every 200 ms; when the probability exceeds a threshold, a lightweight YOLOv4-tiny Vision Anomaly Verifier (VAV) refines high-risk events, and an OSQP-based MPC computes bounded adjustments to nozzle/bed temperatures, speed, flow, and fan duty. We model thermo-extrusion dynamics as a disturbed linear system and enforce tightened constraints to guarantee recursive feasibility under bounded disturbances. On a Raspberry Pi 5, the full loop meets a 200 ms budget with mean latency 132 ms (6 ms CI), 99th percentile 160 ms, and worst case 183 ms. When evaluated on 1000 PLA prints on a Prusa i3 MK3S, defects dropped from 200 to 80 jobs (20% to 8 %), and flawless completions rose from 650 to 880 (65% to 88 %). The predictive accuracy reached 91.2% with monthly adaptive retraining. The material savings are about 1 kg per 100 jobs (up to 2 kg; approximately 1-2 spools), which can recoup the $120-$150 sensor cost within 3-12 months depending on failure rates. The results indicate that multi-sensor prediction with risk-gated, CNN-assisted robust MPC enables reliable, low-cost quality control for FDM.
In industrial parks implementing a Virtual Utility Plant (VUP), shared facilities (e.g., steam and compressed air) enable collaborative utilization and improved energy efficiency. Accurate forecasting of utility demand-supply patterns is critical for routing and source-selection algorithms that optimize piping paths and supply allocation. While long-short-term memory (LSTM) models captured compressed air patterns at one site, they failed to model steam dynamics at another under varying operations, motivating more robust approaches. We propose a stacked-ensemble framework with LightGBM (LGBM), XGBoost, and CatBoost as base learners and compare it against Transformer baselines. The proposed method improves R2 from 0.81 to 0.98 (+0.17; +21.0% relative) and reduces CVRMSE from 24.8% to 7.23% (-17.57 percentage points; -70.9% relative) over a Transformer baseline. Case studies indicate reduced energy losses and actionable insights for routing and energy optimization in VUPs, advancing intelligent automation and sustainability goals.
This paper proposes a novel control law design method to address the control problem of automatic transition to hover of helicopters in low-speed forward flight state. Due to the complex dynamic characteristics of low-speed forward flight and the need for coordinated changes in altitude and speed to ensure a smooth transition, traditional methods have limited performance under these conditions. This article achieves coordinated control of height and speed by designing height commands and longitudinal ground speed commands, and introducing normal velocity feedback to enhance damping. Utilize Stateflow for multi state logic management and combine Simulink to complete system simulation. The results show that the designed system can effectively achieve a smooth transition from flying forward at low speeds to hovering, and both vertical and longitudinal accelerations meet the requirements of the index (<= 0.5 m/s (2)), verifying the effectiveness and reliability of the control strategy under low speed conditions.
Rising fuel costs and increasing CO2 emissions are driving the development of sustainable propulsion solutions for ships. In this context, it becomes increasingly interesting to explore alternative solutions to assist maritime propulsion and reduce fuel consumption.This study presents a path tracking control strategy for a four-lines tethered kite. The nonlinear dynamics are linearized around a desired trajectory using non-parametric closed-loop identification, and later producing a parametric linear model.A PID controller with notch and low-pass filtering (PID(NF)) is designed based on this linear model, ensuring effective trajectory tracking while satisfying stability margins and sensitivity requirements. Time-domain simulations confirm the controller’s effectiveness in dynamic flight.
The underlying model of a robot for the inverse kinematics is unable to represent arbitrary complexities. In contrast to this, more complex Machine Learning based models show capabilities to enhance the accuracy of a robot. This research has explored a combination of virtual model development and the transfer of said model into the physical world to enhance a robot’s absolute accuracy. It shows that the model can be developed on a simulated robot with no end effector considered and applied in the physical world with success. On a robot with an end effector, it reduces the absolute error radius by 33%. The additional transformation of the end effector and its complexity can be represented by the model without further development needs.
Implementing end-of-life (EoL) management strategies and moving towards sustainable development contribute to mitigating resource depletion and energy consumption. Within this context, scrap material sorting is a crucial step that organizes waste materials for further processing, such as recycling. This research proposes a sustainability-aware computer vision approach for scrap material recognition in automated sorting. First, this study presents a new dataset that includes images of various materials (e.g., aluminum, copper, etc.). To enhance the dataset, multiple data augmentation and preprocessing techniques, such as noise addition and rotation, are applied. Next, it fine-tunes the pre-trained YOLOv5, YOLOv8, YOLOv11, YOLOv12, and RetinaNet models using transfer learning. To promote green machine learning (ML) and align with sustainability, this study validates the models by using metrics related to sustainability factors (e.g., energy consumption and carbon emission) in addition to typical technical metrics, such as mAP@50, Precision, Recall, and F1 Score. Finally, a Pareto analysis is conducted to identify optimal models that balance technical performance and environmental impacts.
This paper presents a novel anomaly detection approach for automated CNC machining processes based on a Dynamic-Input LSTM (DI-LSTM) model. Conventional anomaly detection methods using single LSTM or autoencoder frameworks are often limited in capturing complex temporal dependencies in sensor data. To overcome this limitation, we propose a DI-LSTM architecture that allows flexible handling of multiple and variable input signals, enabling the model to capture diverse temporal features. The proposed model was applied to load rate and spindle speed signals collected from CNC machine sensors, and its performance was evaluated against conventional LSTM, LSTM-AE, and autoencoder methods. Experimental results demonstrate that the DI-LSTM model improves detection accuracy and stability, thereby contributing to enhanced reliability in automated machining processes. This study suggests that advanced machine learning techniques have the potential to support intelligent automation and predictive maintenance in manufacturing systems.
