Industrial Internet and edge computing provide massive computing resources to legacy industrial automation systems. It also poses critical challenges in handling massive customization requirements. Due to high reconfiguration delays, existing automation systems cannot adapt to those changes in real-time. This paper proposes a virtualized industrial edge runtime for software-defined automation systems with a lightweight architecture, rapid startup, fast application launch, and dynamic resource allocation. The proposed runtime system enables dynamic resource reconfiguration with a low memory footprint to address high deployment latency. The experimental results demonstrate that the proposed method can significantly reduce response time and memory usage, thereby meeting real-time constraints for software-defined automation systems.
This research is on the sparse identification of nonlinear dynamic systems (SINDy) for the modeling of a diesel engine air-path system and its use in the nonlinear model predictive control (NLMPC) of the air-path system. The proposed modeling strategy involves a binary optimization scheme based on genetic algorithm (GA) to promote model accuracy and sparsity. Prior studies have demonstrated the applicability of SINDy-based modeling using conventional sequentially thresholded least-squares (STLSQ) method for a diesel engine air-path system. This air-path system is a nonlinear multi-input multi-output (MIMO) system of high degree. However, the modeling accuracy achieved by conventional STLSQ-based approach has been limited, particularly when the library is added with nonlinear terms of higher order. Therefore, this research proposes a new data-driven model discovery approach based on GA-based binary optimization and STLSQ with multiple thresholds, referred to as Binary-STLSQ-MT, to overcome the above limitations. As a result, Binary-STLSQ-MT successfully identified models using the library that include higher-order terms, for which conventional STLSQ tends to suffer from numerical instability or divergence. It also achieved a higher coefficient of determination than the conventional approach, with approximately 2% improvement for y1. Furthermore, when applied to NLMPC, the model identified by Binary-STLSQ-MT achieved superior control performance, reducing the NLMPC objective function values by approximately 57% compared with the conventional STLSQ-based model. These findings indicate that the proposed modeling approach substantially extends the practical applicability of SINDy to realistic nonlinear MIMO control problems.
This article introduces a fully soft-switched non-isolated high step-down/up bidirectional converter. The proposed structure utilizes two magnetic cores, one is the high voltage side inductor and the other is a multi-purpose core providing both high conversion ratio as well as soft switching condition. Soft switching is facilitated by the energy stored in the leakage inductance, enabling zero voltage switching for both main switches. The converter has the merits of wide input voltage range and reduced switch current stress due to extended duty cycle at high step-down state. Additionally, employing one of two main switches as a synchronous rectifier in each direction reduces conduction loss and enhances overall efficiency. Though utilizing four switches in its structure, the converter needs only three gate signals in each direction. All these features result in a highly efficient, easily controlled, with enhanced step-down/up conversion ratio. A prototype of the proposed converter operating at 100 W and 100 kHz with 400 V high side and 24 V low side voltages is implemented and the experimental results exhibit 97.84% / 97.5% efficiency in high step down/up directions.
This paper introduces and validates the concept of DC low-voltage ride-through (DC-LVRT) for critical DC loads through the utilization of a non-inverting buck-boost converter (NIBBC), acting as a front-end converter in the DC power chain for powering Polymer Electrolyte Membrane (PEM) hydrogen electrolyzers (HEs). The objective of the control strategy for the NIBBC is to maintain a stable DC bus output even during severe input voltage disturbances, preventing protective shutdowns. The control scheme employs dual voltage-loop and single current-loop PI controllers, and a new feedforward mode-transition compensator designed to inject precise and fast duty support during sudden input voltage drops. A system-oriented model of the front-end and downstream conversion stages is developed as an equivalent load impedance for controller design. An extensive stability and robustness analysis of the proposed strategy is presented and experimentally validated on a Silicon Carbide (SiC)-based prototype. The system was subjected to three disturbance profiles, operating the NIBBC in buck, buck-boost, and buck-boost marginal modes. Experimental results demonstrate that the control system maintains the output voltage within 20% of its 100 V reference, with recovery within 40 ms for all cases. The feedforward controller reduces voltage dips by up to 18%. A detailed hardware versus simulation analysis quantifies the impact of real-world imperfections on control performance, underscoring the importance of such considerations for reliable DC-LVRT implementation.
This paper presents new pulse-width modulation (PWM) strategies for symmetrical six-phase inverters feeding single-neutral-point loads that eliminate or significantly reduce common-mode voltage (CMV) and suppress auxiliary-plane voltages. The proposed methods are derived using a constraint-driven switching-state synthesis approach and control modulation to a reduced set of switching states that inherently produce zero CMV or low CMV levels. By appropriate vector selection and sequencing, the x - y subspace voltages are eliminated and the average voltage in the 0+ - 0- plane is forced to zero, thereby reducing zero-sequence current ripple. Closed-form analytical expressions are derived for the duty cycles, enabling offline computation and low real-time computational burden. No virtual vectors are generated, and no online optimization is required. Symmetric switching sequences are used to limit switching losses and dead-time effects. The modulation range of the proposed methods is analyzed, and experimental results are introduced to validate the performance of the proposed PWMs. A comprehensive comparison with conventional PWM methods and a recently published model predictive control (MPC)-based PWM strategy is also presented. Comparative evaluation with existing PWM techniques demonstrates near-complete CMV suppression for one proposed strategy and approximately 67.7% CMV reduction for the other, while maintaining current quality and loss levels close to the reference PWM. The results further show that the proposed methods require substantially lower computational effort than the MPC-based approach while maintaining competitive performance. Additional simulation results using a symmetrical six-phase permanent magnet synchronous motor drive demonstrate the applicability of the proposed methods to electric-drive systems, confirming suitability for real-time implementation.
Reliable fault detection and diagnosis (FDD) in wind turbine systems remains a challenging task because of the continuously changing operating conditions encountered in practical applications. Variations in wind speed, turbulence, mechanical loading, and measurement noise introduce significant uncertainties that often reduce the effectiveness of conventional data-driven diagnostic methods. To address these challenges, this paper proposes a robust FDD framework designed for wind turbines operating under dynamic and uncertain environments. All results presented in this work are obtained from simulation studies conducted in a MATLAB/Simulink environment. The simulation model captures realistic operating conditions by incorporating variable wind speeds, turbulence, and load-dependent dynamics. Model parameters are selected from physically meaningful ranges reported in the wind energy literature, improving the realism of the simulated operating scenarios while avoiding arbitrary parameter tuning. To enhance robustness under uncertain conditions, Digital Twin-based data augmentation and small measurement perturbations are introduced during the training stage. Fault-related information is characterized using a hybrid feature extraction strategy that combines time-domain, frequency-domain, and statistical descriptors. The extracted features are subsequently processed using multiple classifiers whose outputs are integrated through early-, late-, and decision-level fusion strategies with attention-based weighting. Simulation results obtained from multiple wind turbine fault scenarios demonstrate that the proposed framework consistently provides accurate and robust fault diagnosis, outperforming several conventional methods. Although the achieved classification performance is very high, the results are interpreted with appropriate caution because controlled simulation environments and well-separated fault characteristics can contribute to optimistic performance estimates. To minimize the risk of data leakage, strict scenario-level separation between training and testing datasets is maintained throughout the evaluation.
Accurate control-oriented composite hydrodynamic parameter knowledge is essential for high-performance control of uncrewed surface vessels (USVs), yet conventional identification methods rely on persistent excitation (PE), which is often impractical in real operations. This paper proposes an integrated framework combining online parameter estimation with prescribed-time line-of-sight (PT-LOS) cascade guidance control. The estimation scheme uses Dynamic Regressor Extension and Mixing (DREM) to decouple a 14-parameter composite dynamic model into independent scalar regressions, followed by a prescribed-time least-squares (PT-LS) estimator that guarantees practical prescribed-time convergence under bounded disturbances and exact prescribed-time convergence in the disturbance-free case of the estimation errors at a user-defined time. The approach requires only finite excitation from standard maneuvers, eliminating the need for PE. Leveraging these estimates, a prescribed-time cascade guidance controller is developed to achieve precise velocity tracking, with adaptive compensation for unknown disturbances. The framework ensures practical prescribed-time stability (PPTS) of both the composite parameter estimation errors and the tracking errors, independent of initial conditions. Theoretical analysis and experimental results validate the approach, demonstrating accurate composite parameter identification and high-precision tracking under realistic operating conditions.
Autonomous driving systems demand strict real-time properties and safety for practical realization, and development is progressing in heterogeneous environments where the industry standard AUTOSAR Adaptive Platform (AUTOSAR AP) and Robot Operating System 2 (ROS 2) coexist; however, end-to-end latency analysis in such mixed environments remains a challenge. A framework called “CART” was proposed in a previous study to integrate trace data from both platforms; however, its evaluation was limited to functional verification with small-scale applications, and its effectiveness in an environment with complexity and high load close to actual autonomous driving systems was unverified. Therefore, this paper proposes the “Evaluation Platform for Tracing of Autonomous Driving System Combined AUTOSAR AP and ROS 2,” which extends CART and integrates CARLA, a high-fidelity simulator, and a practical autonomous driving software stack (Autoware). This platform targets the sequence from the reception of sensor data by the autonomous driving system, through processing via ROS 2 and AUTOSAR AP, to the emission of control commands immediately before they are forwarded to the simulator, enabling the tracking of overall system behavior and the identification of bottlenecks in complex mixed environments. The results indicate that integrated tracing and latency analysis using CART can be applied to a mixed AUTOSAR AP and ROS 2 autonomous driving stack executed in a cloud-based simulator-driven evaluation environment. Furthermore, quantitative evaluation confirmed that the CPU and memory overhead from the ara::log-based tracing method adopted by CART remains at a low level comparable to the ROS 2 standard tracing tool, affirming its viability for large-scale systems. The proposed platform supports detailed performance evaluation and reduces manual effort by enabling unified end-to-end latency analysis across the heterogeneous AUTOSAR AP and ROS 2 autonomous-driving stack.
Recent reliability studies from 2025 indicate that power semiconductor modules remain the most critical failure point in inverters, accounting for approximately 38 % of total system failures. The increase in power density and reduction in system volume are common requirements, especially in the automotive sector, targeting 200 kW/L by 2030, which increases thermal stress and requires advanced non-intrusive SOH monitoring. Conventional electrical-based indicators are often limited by high switching noise and inherent temperature dependence, which can lead to inaccurate aging estimation if proper thermal compensation is not applied, and by the complexity of on-board conditioning circuits. This paper proposes a novel method for IGBT aging characterization based only on junction (Tj) and case (Tc) temperature monitoring. The proposed approach consists of simulating the power losses and thermal behavior of a power transistor under various predefined aging conditions. These simulations provide the Tj and Tc for each aging state, which are then correlated with conventional aging indicators such as the on-state conduction resistance and the junction-to case thermal resistance. This unique approach enables decoupling of conduction-related degradation mechanisms (Mcon) from thermal-path degradation mechanisms (Mth). Experimental validation of power modules demonstrates that the proposed method accurately tracks SOH using conventional indicators, such as conduction and thermal resistances, with errors of 0.1–1.5% over most of the device's operational life.
This paper presents a physics-informed representation approach for supervised classification of predefined stator inter-turn short-circuit (ITSC) severity classes in three-phase induction motors. The term physics-informed is used here in a representation-level sense: physical knowledge is introduced through deterministic input construction, not through physics-constrained neural-network losses or architecture modifications. The proposed formulation constructs voltage-current trajectories from the measured phase currents ($v$-$i_{\phi }$) and from current components obtained with the Conservative Power Theory (CPT). CPT provides a time-domain, projection-based decomposition of the stator-current vector into active ($\boldsymbol{i}_{\mathrm{a}}$), reactive ($\boldsymbol{i}_{\mathrm{r}}$), unbalanced ($\boldsymbol{i}_{\mathrm{u}}$), and void ($\boldsymbol{i}_{\mathrm{v}}$) components in $L^{2}[0,T]^{3}$. Pairing phase voltages with each CPT component yields component-wise trajectory families whose geometry is associated, respectively, with average power transfer, magnetic energy exchange, inter-phase asymmetry, and residual distortion. At the terminal-variable level, ITSC progression is interpreted as a redistribution of current energy from the balanced active-reactive subspace toward imbalance and distortion subspaces, quantified through $\mathcal {F}_{\text{ITSC}}$ and expressed in trajectory deformation, slope variation, phase asymmetry, and loop thickening. The approach is experimentally assessed on a 1 HP induction motor under torque variation, supply-voltage imbalance, and multiple ITSC severity levels. Five CPT-based trajectory representations, namely $v$-$i_{\mathrm{a}}$, $v$-$i_{\mathrm{r}}$, $v$-$i_{\mathrm{u}}$, $v$-$i_{\mathrm{v}}$, and the hybrid reactive-void representation $v$-$i_{\text{rv}}$, are compared with the conventional raw $v$-$i_{\phi }$ representation using a manually parameterized convolutional neural network (CNN) and a pretrained DenseNet-169 model. In the evaluated dataset, the raw $v$-$i_{\phi }$ trajectory produced the largest observed classification accuracies, with repeated-run mean values of 99.77% for the manual CNN and 98.93% for DenseNet-169 . CPT-derived trajectories, particularly $v$-$i_{\mathrm{r}}$ and $v$-$i_{\text{rv}}$, yielded comparable classification performance while decomposing the diagnostic patterns into physically interpretable energetic components. The evaluated learning task is therefore closed-set severity classification under controlled operating conditions, with physical attribution supplied by the CPT-based trajectory construction. The results position CPT-derived trajectories as complementary representations: the raw $v$-$i_{\phi }$ trajectory concentrates the largest amount of discriminative information in this dataset, whereas CPT-derived trajectories support interpretation of voltage-current deformation through component-wise energetic attribution.
This paper presents a boost high power factor rectifier integrated with a two-switch forward converter operating in continuous conduction mode (CCM) at both the input and output. The use of overlapping PWM modulation is proposed to reduce the low-frequency ripple in the output voltage, enabling a significant reduction in the required capacitance and eliminating the need for electrolytic capacitors while maintaining high power factor operation. An auxiliary demagnetization circuit is introduced to ensure complete transformer demagnetization at the zero-crossing of the AC input voltage, where the duty cycle becomes large. The analysis of the performance improvement using Silicon Carbide (SiC) synchronous rectification at the converter output is also presented. A 500 W converter was implemented to validate the proposed solution, achieving a capacitance-to-power ratio of 0.164 $\mu$F/W, an efficiency of 91.4%, a THD of 4.778%, a power factor of 0.9953, and a fast dynamic response of 2 ms at the converter output.
Stability and transparency are intrinsically coupled and must be addressed simultaneously in delayed bilateral teleoperation systems, where there is a fundamental trade-off between ensuring closed-loop stability and achieving accurate force–position coordination. This challenge becomes particularly critical when interacting with remote environments under time-varying communication delays, since the human operator introduces nonlinear, time-varying, and often unpredictable dynamics into the closed-loop system. This paper presents an adaptive neural-network-based compensation strategy embedded within a model-based control framework to enhance dual coordination in bilateral teleoperation. By combining classical control design with online neural network adaptation, the proposed controller compensates for parametric uncertainties, unmodeled dynamics, interaction forces, and communication delays without requiring explicit models of the human operator or the remote environment. The adaptive structure enables real-time learning of unknown nonlinearities while preserving the stability guarantees provided by the underlying control architecture. A theoretical analysis of the closed-loop teleoperation system is developed, demonstrating stability in the presence of human-applied forces, environment interaction forces, and time-varying communication delays. Numerical simulations conducted on two-degree-of-freedom manipulators validate the feasibility and practical viability of the proposed approach. Different explicit models of the human operator are considered in simulation to assess the sensitivity to operator dynamics. The results show that a bounded dual coordination of force and position is achieved under delayed communication regardless of the assumed operator model, supporting the potential application of the method in real-world teleoperation scenarios.
This paper presents a non-isolated, single-stage, bidirectional on-board charger (OBC) for electric vehicles based on a differential Y-rectifier with one phase-inductor per phase, enabling seamless operation on single- and three-phase grids. A straightforward, unchanged modulation and current-control strategy is retained across 1ϕ/3ϕ operation, yielding mode-invariant device stresses and efficiency, and supporting equal rated power in both configurations. The topology features a markedly reduced component count and a common reference coincident with the HV battery, simplifying gate-drive and sensing circuitry, and lowering isolation requirements in the control path. In addition, the converter integrates a buck-type power pulsation buffer (PPB) that compensates the 2·fac ripple otherwise imposed on the battery, thereby allowing a compact energy buffer and high power density. Grid-compliant operation is ensured under Ubatt>Ûac (practically ∼400 V), targeting modern 800-V battery platforms. The proposed converter is tested by means of a hardware prototype rated at 11 kW that operates over 85–256 Vrms, 40–65 Hz, with battery voltages ≥ 400 V (tested up to 850 V). The design achieves 10.2 kW/L power density and peak efficiencies > 97%, with PF ≥ 0.99 and THDi ≤ 5%. Experimental results validate operation in both grid configurations and confirm invariant device RMS/peak stresses, identical control, and equivalent efficiency between 1ϕ and 3ϕ operation, establishing a cost-effective, high-power-density alternative to two-stage OBCs.
Data centers are emerging as one of the fastest-growing electricity consumers worldwide due to the rapid expansion of cloud computing, artificial intelligence (AI), and digital services. The large-scale integration of AI data centers into electric power systems introduces significant challenges for grid planning, operation, stability, power quality, and compliance with evolving grid codes. Modern data centers are characterized by power-electronic-dominated infrastructures, including high-density computing platforms, advanced cooling systems, on-site renewable energy and energy storage resources, all of which exhibit dynamic behaviors distinct from conventional passive loads. Consequently, their increasing integration necessitates the development and application of appropriate modeling frameworks to accurately assess grid impacts and enable effective control and coordination strategies. This papers covers grid-integrated data centers, with a focus on their electrical and cooling architectures, and associated modeling approaches. In addition, modeling of critical data center components, interconnection requirements, and key integration challenges are examined. Finally, emerging opportunities for data centers to provide frequency regulation and flexibility services are discussed, outlining future research directions toward reliable, efficient, and grid-interactive data center integration.
This article presents the design, implementation, and experimental validation of an Internet-of-Things (IoT)-based structural health monitoring (SHM) system for suspension bridges using synchronized vibration measurements. The proposed framework integrates distributed micro-electro-mechanical systems (MEMSs)-based sensing nodes, global positioning system (GPS)-assisted time synchronization, low-power wireless communication, and cloud-based modal analysis to support continuous monitoring of structural conditions. The main contribution of this work is the end-to-end integration of synchronized low-power IoT sensing, multitier wireless communication, and cloud-based vibration analysis within a unified SHM framework, together with experimental validation on a laboratory-scale suspension bridge. In the proposed system, each leaf node incorporates a high-accuracy digital accelerometer, ESP-NOW communication, and a duty-cycled sleep mode to reduce power consumption, while a central gateway aggregates the acquired data and transmits them to a remote server through a cellular network. Natural-frequency analysis, mode-shape evaluation, and the modal assurance criterion (MAC) are employed to assess changes in structural dynamic behavior. The system was evaluated using a laboratory-scale suspension bridge instrumented with five synchronized IoT sensing nodes and subjected to controlled shake-table excitations. Four hanger- and main-cable-damage scenarios were introduced to evaluate the overall operation of the proposed framework. The results demonstrate successful synchronized vibration acquisition, reliable wireless data transmission, and the ability to capture damage-sensitive modal changes, including natural-frequency shifts, mode-shape distortions, and MAC variations under different damage conditions. Communication experiments further show that ESP-NOW reduces energy consumption compared with transmission control protocol/internet protocol (TCP/IP) over Wi-Fi, extending node battery lifetime by approximately 40% through sleep-mode operation. Overall, the proposed SHM framework provides a cost-effective, energy-efficient, and synchronized IoT solution for vibration-based monitoring of suspension bridges under laboratory conditions. Future work will focus on field validation on real bridge structures and long-term operation under practical environmental conditions.
This article presents a short-term (1-h ahead) physics-guided multirepresentation feature fusion framework for photovoltaic (PV) power prediction with uncertainty quantification based on quantile regression for interval forecasting. The proposed hybrid methodology integrates feature-representation fusion, deep learning architectures, including long short-term memory (LSTM) and gated recurrent unit (GRU), and machine learning models, such as support vector regression (SVR) and random forest (RF), along with dimensionality reduction techniques, namely principal component analysis (PCA) and autoencoder (AE), and a fully defined physics-based digital twin (DT) model with explicit irradiance-to-power equations and parameter calibration. The DT residual, computed causally using only training data, is incorporated as a physics-informed feature to capture unmodeled nonlinearities. PIs are learned using the pinball (quantile) loss function, replacing heuristic assumptions, and feature-representation fusion reduces uncertainty while enhancing robustness. Validation on real meteorological datasets from Izki and Manah stations (2017-2023) shows that conventional feature sets achieve root mean squared error (RMSE) values between 0.14-0.17 with coefficient of determination (R-2 ) ranging from 0.47-0.61. Incorporating DT residuals significantly improves performance, while maintaining realistic generalization performance, with LSTM achieving RMSE approximate to 0.0105 and R-2 up to 0.9979. SVR and RF models also benefit, achieving RMSE <= 0.0525 and R-2 >= 0.9467. Comprehensive evaluation including ensemble comparisons, ablation studies, and uncertainty metrics, namely prediction interval coverage probability and mean prediction interval width, confirms high coverage and sharpness, while feature-representation fusion ensures consistent, low-uncertainty predictions. Further experiments on a physically inspired synthetic dataset with explicitly defined generation process and temporal resolution of 1000 samples demonstrate similar trends. Using DT residual features, LSTM and GRU achieve R-2 > 0.998 with RMSE approximate to 0.042, while RF reaches R-2 = 0.897 and a relatively higher RMSE approximate to 0.326. PCA and AE representations improve computational efficiency but provide comparatively lower predictive accuracy. Direct comparison between individual models and the proposed ensemble highlights the robustness and stability of the fusion strategy, particularly in high-noise conditions. In summary, the framework offers a reproducible, physically grounded, and uncertainty-aware solution for next-generation solar energy forecasting, explicitly quantifying prediction uncertainty and leveraging physics-informed residuals to enhance predictive reliability and interpretability across both real and synthetic scenarios.
Control software design in Cyber-Physical Systems plays a vital role in ensuring maintainability and adaptability in response to evolving process requirements. Variability-aware designs facilitate the efficient integration of new functionalities and depend on the selected design elements and on the way these elements interact. Designers can enhance flexibility by applying design patterns and adopting alternative design elements when an existing design requires substantial modification effort. In IEC 61499, using discrete interfaces as a design element in Function Block (FB) implementations often incurs high effort due to numerous discrete connection instances and limited encapsulation. Complexity is further increased by implicit one-to-many connections among subsystems, which negatively affect system understandability and maintainability. This paper analyzes two baseline implementations that follow pure Service-Oriented Architecture (SOA) interaction patterns: choreography and orchestration, implemented in IEC 61499 using FBs with discrete interfaces. We then restructure these baseline implementations using design patterns and adapter interfaces, resulting in two new implementations with adapter interfaces and a hybrid interaction structure. We evaluate modification effort across three variability cases to quantify the changes required in the original and redesigned implementations. The results demonstrate a measurable reduction in modification effort and improvements in control software maintainability. Based on these findings, we derive guidelines to support practitioners in selecting suitable designs and interface strategies for a specific variability context in control software design.
Device fingerprinting enables various asset management scenarios. This work presents a comprehensive evaluation of machine learning (ML) and deep learning (DL) approaches for fingerprinting IP-connected devices using the publicly available network mapper (Nmap) operating system (OS) database. Focusing on the classification of three critical device characteristics, i.e., operating system family, vendor, and device type, fingerprinting is treated as a multiclass classification problem. The performance of eleven algorithms, including six classical ML models and five deep neural network (DNN) architectures, is evaluated. The experiments explore the impact of data imbalance, employing synthetic minority oversampling technique and adaptive synthetic for synthetic data augmentation. The empirical results show that tree-based ML models, such as decision trees, random forests, and extreme gradient boosting, achieve high accuracy at significantly lower training and inference costs than DNNs. Although DNNs can also yield competitive area under the curve (AUC) scores, their performance benefits are outweighed by training requirements, making them less suitable for most practical applications. This study concludes that traditional ML approaches remain preferable for OS, vendor, and device type classification in IP-connected device fingerprinting scenarios, and it highlights key tradeoffs in performance, complexity, and computational overhead. The results are of interest to industry stakeholders, including researchers, security practitioners, and system designers who integrate device identification into their asset management, security, and critical infrastructure systems.
The increasing scale and complexity of photovoltaic (PV) power plants require reliable and cost-effective monitoring strategies capable of identifying operational anomalies under diverse climatic conditions. However, a challenge in this context lies in the strong dependence of many existing approaches on high-quality local solar irradiance measurements, which are often unavailable, incomplete, or economically unfeasible in real-world PV installations. This article proposes an integrated methodology for PV power prediction and anomaly detection using satellite-based solar data combined with statistical thresholding techniques. Satellite-derived meteorological datasets are used to train supervised regression models to estimate PV power, which is subsequently employed as an external and physically consistent reference for anomaly detection. Monthly characteristic curves are constructed from the predicted power using statistical measures, enabling the definition of adaptive operating thresholds that account for seasonal and intraday variability. Anomalies are identified based on both statistical deviations from these dynamic limits and physically inconsistent behaviors between irradiance and measured power, such as power drops under increasing irradiance conditions. The methodology is evaluated using two independent satellite datasets with different temporal and spatial characteristics, allowing a comparative analysis of linear and tree-based machine learning models in terms of prediction accuracy, robustness, and explainability. In this way, results demonstrate that satellite-based power prediction provides a stable and scalable baseline for anomaly detection, particularly suitable for PV plants with limited instrumentation or without local irradiance sensors. The proposed approach effectively captures seasonal patterns of operational stress and distinguishes different types of anomalous behavior, offering a practical and interpretable solution for large-scale PV monitoring and performance assessment.
Accurate fault diagnosis of electrical machines is essential for operational reliability and safety. Handcrafted features remain attractive in practice because of their interpretability and computational efficiency. However, this traditional approach relies on individual feature vectors and is limited in preserving inherent temporal dynamics. To address this issue, a label-consistent input structuring based on the second stage sliding window (SSSW) method is proposed. This approach retains the benefits of handcrafted features and arranges feature vectors into temporally coherent sequences while ensuring label consistency. A hyperparameter optimization scheme is incorporated with a long short-term memory classifier to reduce manual tuning. Performance evaluations demonstrate the robustness of the proposed SSSW method across diverse operating conditions and input signal configurations. These include single-phase currents, multi-phase currents, vibration, and fused current-vibration signals. Notably, the proposed approach achieves classification accuracy of up to 100% and demonstrates stable learning behavior. Experimental verification on a laboratory scale permanent magnet synchronous motor testbed further validates the proposed method under realistic measurement noise and interference conditions. Finally, combining handcrafted features with the proposed SSSW input structuring method provides a practical, scalable solution for reliable electrical machine fault diagnosis.
