This study investigates the surface improvement of Direct Metal Laser Sintering (DMLS) produced AlSi10Mg alloy using Abrasive Flow Machining (AFM). Due to the layer by layer manufacturing mechanism, DMLS components typically exhibit high surface roughness, partially fused particles, and irregular surface features that may adversely affect service performance. A full factorial experimental design was conducted by varying abrasive grain size (240 and 400 mesh), abrasive concentration (40% and 60%), and cycle number. Surface roughness, material removal rate (MRR), macro scale morphology, and scanning electron microscopy (SEM) analyses were used to evaluate the effectiveness of the AFM process. The results showed that AFM significantly improved the surface quality under all investigated conditions, achieving roughness reductions exceeding 90% under optimum processing parameters. Cycle number was identified as the dominant factor governing both roughness evolution and material removal behavior. SEM observations revealed the progressive removal of melt pool related asperities, adhered particles, and manufacturing induced surface irregularities, resulting in a more homogeneous surface morphology. The optimum experimental condition was obtained at 400 mesh abrasive size, 60% abrasive concentration, and approximately 20 cycles. Statistical analyses confirmed the significance of the observed trends, while numerical optimization predicted an optimum condition close to the experimentally determined processing window. The findings demonstrate that AFM is an effective post processing technique for improving the surface quality of DMLS produced AlSi10Mg components and provide practical guidance for finishing additively manufactured metallic parts with complex geometries.
This study is concerned with a resilient path-following control strategy for autonomous electric vehicles (AEVs) based on an adaptive event-triggered mechanism (AETM) under randomly occurring multiple cyber-attacks, including denial-of-service (DoS), deception attacks and actuator faults. Initially, to alleviate bandwidth consumption constraints, an adaptive communication scheduling mechanism is formulated, in which the triggering threshold is dynamically regulated in accordance with the evolving system performance requirements. Furthermore, unlike existing AETM-based path-following schemes for AEVs, a unified switched-system model is constructed to simultaneously encapsulate DoS attack scenarios and deception attacks models within the proposed adaptive triggering function. Moreover, in view of the unavoidable presence of actuator faults in practical control implementations, a resilient fault-tolerant event-triggered control scheme is formulated to enhance vehicle performance. Furthermore, by introducing a novel piecewise Lyapunov functional construction, the co-design condition is derived for the triggering parameters and the resilient path following controller, while guaranteeing global exponential stability of the resulting switched closed-loop AEV dynamics. Finally, a numerical case study is presented to validate the effectiveness of the proposed framework and analysis.
Since most distributed power sources in microgrids lack the necessary inertia, the frequency stability of the system deteriorates. By emulating the dynamic characteristics of traditional synchronous generators, virtual synchronous generator (VSG) technology provides essential inertia support for power systems, emerging as an effective solution. However, while providing essential inertia and damping, VSGs inevitably induce active power oscillations, a problem that is critically exacerbated when multiple VSGs operate in parallel. To address this issue, this paper proposes an active power-frequency oscillation suppression strategy based on a fractional-order control scheme for the parallel VSG system. By incorporating two fractional-order feedback paths into the VSG active-power control loop, the proposed strategy provides additional degrees of freedom for shaping the active-power and frequency responses. Consequently, the proposed strategy effectively attenuates active-power and frequency oscillations in the parallel VSG system. Subsequently, the closed-loop stability of the parallel VSG system under the proposed fractional-order strategy is established using a Lyapunov energy function. Finally, hardware experiments on a two-unit parallel VSG platform show that, under a resistive-load disturbance, the proposed strategy reduces the maximum active-power overshoot from 9.4 % to 2.6 % and shortens the settling times of the active-power and frequency responses from approximately 0.50 s to 0.21 s, relative to traditional VSG control. This corresponds to an approximately 58 % reduction in settling time, while the desired 2:1 steady-state active-power-sharing ratio is maintained. The improvement in active-power–frequency dynamics is also retained under inductive-load and nonlinear-load disturbances.
Electrodynamic levitation (EDL) is a promising technology for high-speed transportation, but its parametric design limits require robust physical validation. This paper presents the numerical modeling, physical prototyping, and experimental characterization of a novel 10-pole rotating Halbach-array disk. By evaluating rotor diameters between 100 and 200 mm across various air gaps, we quantify the resulting force profiles, Lift-to-Drag (L/D) ratios, and the underlying dynamic effects. Findings reveal that this specific configuration reliably generates sufficient lift for a 130 kg vehicle when optimized for minimal air gaps and high rotational speeds. Discrepancies between finite element analyses and physical tests are explicitly calculated, providing a realistic design framework that accounts for manufacturing tolerances and magnet arrangement deviations in future Hyperloop applications.
Maximizing power extraction while mitigating structural vibrations remains a major control challenge in floating offshore wind turbines (FOWTs) due to coupled wind–wave–platform dynamics. This study proposes a genetic algorithm–optimized artificial neural network (GA–ANN) control framework for the below-rated operating region of a 5 MW barge-type FOWT. The control strategy combines direct generator speed control for maximum power point tracking with an ANN-based torque controller whose parameters are optimized using genetic algorithms. The main contribution of this work is a comparative analysis of cost-function formulations used in the GA optimization of the neural controller. These formulations incorporate both generator speed regulation and tower-top dynamic response, enabling a systematic evaluation of their influence on the optimization process and the resulting control performance. In particular, additive and multiplicative cost-function structures are investigated to analyze how different aggregation strategies affect the trade-off between energy capture and structural vibration mitigation. Numerical simulations performed in OpenFAST demonstrate that the proposed GA–ANN control strategy outperforms a baseline variable-speed controller, delivering higher and less oscillatory electrical power while significantly reducing tower-top acceleration levels. The results highlight the critical role of cost-function design in evolutionary optimization of wind turbine controllers and demonstrate the potential of the proposed methodology to improve both performance and structural durability in floating offshore wind systems.
This study presents an experimental and finite element (FE) validation investigation of reinforced concrete (RC) beams under quasi-static moving loads, focusing on the applicability of Fiber Bragg Grating (FBG) sensors for low-amplitude strain monitoring. Two full-scale beam specimens with identical dimensions were examined: an unstrengthened control beam and a beam strengthened with an externally bonded carbon-fiber-reinforced polymer (CFRP) laminate. Moving loads of 100, 122, and 150 kg, all within the elastic response range of the beams, were applied using a custom cart system to simulate quasi-static moving-load conditions. FBG sensors installed at embedded and external locations were used to capture real-time strain responses. Collocated conventional electrical strain gauges were also installed; however, the strain amplitudes generated during testing were generally below 10 µε, and reliable strain–time histories could not be extracted from the electrical gauges. Therefore, the detailed strain-response analysis was based on the FBG measurements.The recorded FBG strain–time histories showed clear rise–peak–decay patterns corresponding to load movement along the span. External sensors measured higher strain amplitudes than embedded sensors, consistent with their proximity to the outer tensile fiber rather than the neutral axis. The CFRP-strengthened beam exhibited higher localized surface strain at the laminate level than the control beam, consistent with strain transfer to the bonded CFRP layer. A three-dimensional FE model developed in Abaqus reproduced the experimental strain–time response with good agreement, with peak-strain deviations generally within 15 %. Stress distributions in the concrete, steel reinforcement, and CFRP layer were consistent with the expected elastic flexural response under the applied loading. Within the tested low-amplitude strain range, the FBG system provided clear, high-resolution strain–time histories and supported FE correlation under quasi-static moving-load conditions. The validated FE model provides a basis for future parametric studies of FRP-strengthened RC members under broader loading, damage, and strengthening configurations than those examined in this elastic-range, single-specimen-pair study.
The transition toward sustainable energy systems demands electrical generators that are highly efficient, reliable, and durable. Among emerging topologies, Flux-Switching Generators (FSGs) offer considerable potential. Nevertheless, without a systematic design methodology, FSGs may suffer from excessive torque ripple and reduced efficiency. This study presents a comprehensive multi-objective optimization of an FSG using four meta-heuristic algorithms: Gray Wolf Optimizer (GWO), Particle Swarm Optimization (PSO), Non-dominated Sorting Genetic Algorithm II (NSGA-II), and Multi-Objective Differential Evolution (MODE). The objective is to minimize total loss, active material volume, and torque ripple, while treating the average electromagnetic torque magnitude as a competing objective.To ensure a fair comparison, each algorithm is executed in 10 independent runs with identical population sizes and iteration limits, and equal weighting is assigned to all objective functions. Because GWO and PSO minimize a scalarized weighted cost whereas NSGA-II and MODE rely on Pareto dominance, all four designs are compared through a common normalized cost. The optimal designs obtained from each method are further validated through Finite Element Analysis (FEA), which indicates that bulk saturation is markedly reduced with respect to the preliminary design, the cores operating predominantly at about 1.4–1.6 T, while high flux-density regions remain confined to localized tooth-tip and magnet-interface areas. Notably, although the optimized designs operate at a lower torque magnitude, they attain a higher power density and efficiency than the preliminary machine, indicating that the framework yields more compact and efficient generators. Statistical evaluation based on the Friedman test and Holm-corrected pairwise Wilcoxon tests indicates that the four algorithms are statistically equivalent in solution quality; nevertheless, GWO attains the best mean rank, the lowest mean and minimum cost, and roughly half the computation time of the other methods, and is therefore identified as the most practical and efficient approach for FSG optimization.
This work presents a LiNbO3 laterally excited bulk acoustic resonator (XBAR) with staggered diamond-hole interdigital transducers (IDTs) for suppressing higher-order A1-n spurious modes. By introducing multidirectional acoustic scattering and disrupting lateral periodicity, the proposed structure weakens the coherent buildup of lateral standing waves while preserving the target A1 mode. Comparative analysis of conventional, parallel, and staggered electrode configurations shows that the staggered layout provides a better balance between spurious suppression and main-mode preservation. Devices fabricated on 400 nm-thick Z-cut LiNbO3 thin films operate at 4.57 GHz, exhibiting a Bode-Qmax of 527 and an electromechanical coupling coefficient (kt2) of 22.2 %, corresponding to a figure of merit (FOM = kt2 × Q) of 117. Compared with the conventional design, the spurious mode level is reduced from 0.588 to 0.285. The proposed structure provides a practical route toward high-frequency, low-loss LiNbO3 acoustic resonators.
The increase in the need for electrical energy day by day causes the number of connections in the network to increase and planning becomes more difficult. The power flow analysis is an integral part of the work carried out on the power systems and it is can easily be performed thanks to the increased speed and capacity of computers. However, in a system with thousands of buses, presenting the results to make predictions and quickly determine the status of the system is also a major problem. The use of geographic information systems is recommended to provide situational awareness for power system reliability, especially due to blackouts that have occurred in many countries in recent years. Rather than proposing a new power flow algorithm, this study introduces an integrated GIS-based (ArcGIS) visualization framework that enables the geographical interpretation of power flow analysis results obtained with the Power System Analysis Toolbox (PSAT) software. In addition, a long-term planning study was carried out by using the real interconnected 66-bus system in Istanbul province of Turkey. This region was chosen because it consumes more electrical energy than many OECD countries due to its high population density and industrial load intensity. It also has three Bosphorus crossings due to its geographical features. It has been observed that the results obtained by using free software are very close to the results of commercial package program (PowerWorld Simulator) that can perform both analysis and visualization. The proposed workflow and framework supports multiple visualization layers and long-term planning scenarios, providing an effective decision-support environment for transmission planning and system operation.
Models that are both highly accurate and computationally efficient for edge deployment are required for IoT sensor streams for real-time patient health prediction. To predict catastrophic health events, this work presents TCB-Health, a hybrid Temporal Convolutional Network (TCN) and Bidirectional LSTM (Bi-LSTM) architecture that uses multivariate physiological inputs, including respiration, accelerometer data, ECG, and PPG. The large MIMIC-IV waveform database (>10,000 patients) was used for additional validation, and the WESAD wearable dataset (15 participants) was employed for the investigations. The proposed model surpasses the CNN-only, GRU, and solo TCN baselines, achieving classification accuracy of 98.67%, precision of 98.54%, recall of 98.82%, F1-score of 98.90%, and inference latency of less than 10 ms per data window. The method reduces computational costs by around 35% in FLOPs and 28% in parameters, enabling reliable real-time deployment on edge devices with constrained resources. These results demonstrate that TCB-Health, in IoT-enabled healthcare systems, provides a scalable and effective solution for remote patient monitoring, early detection of cardiac abnormalities, and real-time clinical decision support.
This paper presents a multi-objective framework for online KPI-based adaptive pulse-width modulation (PWM) selection in six-phase electric vehicle traction inverters. Unlike conventional modulation approaches that rely on fixed strategies or heuristic switching rules, the proposed method formulates PWM selection as an online supervisory decision process based on measured inverter performance indicators. At each control interval, multiple PWM candidates are evaluated using measured indicators, including common-mode voltage (CMV), current distortion, circulating currents, and switching losses. These objectives are combined into a normalized cost function to identify the most suitable modulation strategy under varying operating conditions. The proposed framework performs PWM selection directly from measured KPIs without requiring a detailed inverter or load model for the supervisory decision process. In addition, two new PWM schemes are developed to reduce CMV while maintaining acceptable current quality and switching performance. These schemes are integrated into the proposed framework and evaluated alongside conventional modulation techniques. The approach is validated through both simulation and experimental studies on a six-phase inverter prototype. Results demonstrate consistent convergence to suitable PWM strategies under steady-state conditions and effective adaptation under dynamic operation. The proposed PWM schemes achieve reductions in RMS CMV of up to 67% compared with the reference PWM strategy. Furthermore, Proposed PWM1 maintains THD below 0.83% across the investigated operating range and achieves up to 67% lower current distortion at a modulation index of 0.8 compared with a recent low-CMV model predictive control method, while requiring up to 72% lower switching losses and 53% lower execution time. The implementation results further confirm the practical feasibility of the proposed framework, with an execution time of approximately 21.44 μs on a dSPACE MicroLabBox platform. The proposed framework provides a scalable and practical solution for online adaptive PWM selection in multiphase traction systems.
In data-driven remaining useful life (RUL) prediction, practitioners face a recurring design choice: should engineering effort focus on health indicator (HI) construction or on model architecture? This paper provides quantitative tools to answer that question before model training begins, using only knowledge of the degradation physics. We develop an information-theoretic framework that yields three practical outputs. First, a single scalar diagnostic — the information non-uniformity index η, computable from the degradation model class — predicts whether a given domain is HI-limited (η>1: invest in HI design) or model-limited (η<0.3: invest in model architecture). Second, a closed-form formula determines the minimum HI dimensionality d∗ needed to capture a prescribed fraction of the available prognostic signal, preventing both under-dimensioned HIs (irreducible accuracy loss) and over-dimensioned ones (overfitting risk). Third, a bottleneck ratio ρHI quantifies the fraction of prognostic information lost at the HI compression step versus the modeling step, directly identifying the pipeline’s weakest link. These tools are derived from three theorems establishing the fundamental limits of HI-based prediction for exponential, power-law, and stretched-exponential degradation classes, with robustness extensions to non-Gaussian noise, correlated channels, and nonlinear HIs. Cross-domain comparison against published results from turbofan engines, lithium-ion batteries, and rolling element bearings is consistent with the framework’s predictions: domains flagged as HI-limited (batteries, bearings) are those where published studies report that HI design dominates model architecture. Because the per-domain η values are derived from assumed degradation-class parameters rather than computed from raw data, this cross-domain ordering serves as a consistency check; the primary causal evidence is a controlled experiment on three benchmark datasets (C-MAPSS FD001, NASA Battery, C-MAPSS FD004), which confirms that the effective optimal HI dimensionality increases with system complexity as predicted.
Multimodal medical image fusion aims to generate an informative fused image by integrating complementary information from different imaging modalities, thereby facilitating clinical observation, disease diagnosis, and treatment planning. However, brain MRI-SPECT fusion remains challenging because of the limited availability of paired training samples, the intrinsic heterogeneity between anatomical and functional modalities, and the difficulty of simultaneously preserving fine structural details and global functional responses. To address these issues, this paper proposes a local–global decoupled dual-network framework for small-sample MRI-SPECT fusion, termed LGD-Fusion. Specifically, the proposed method constructs a local branch and a global branch with different modeling responsibilities. The local branch employs rotation convolution and wavelet convolution to enhance direction-sensitive structural detail extraction and multi-scale frequency representation, while the global branch integrates large-kernel convolution and state-space modeling to capture long-range contextual dependencies and global semantic information. To further reduce redundant feature entanglement between branches, a shared-private decoupling module is introduced to explicitly distinguish branch-consistent information from branch-specific complementary information. In addition, a dual-head reconstruction strategy is designed to jointly perform base-intensity allocation and residual-detail compensation, enabling stable fusion reconstruction in an unsupervised manner. Experimental results on small-sample brain MRI-SPECT image pairs demonstrate that LGD-Fusion achieves favorable performance in structural preservation, edge-detail transfer, and visual quality, while effectively maintaining functional information under limited-data conditions. These results indicate that the proposed framework provides a promising solution for robust and interpretable multimodal medical image fusion when only limited paired samples are available.
The rapid evolution of sixth-generation (6G) wireless systems has intensified the need for adaptive modulation strategies capable of operating reliably and efficiently in highly dynamic terahertz (THz) communication environments. Conventional threshold-based and static optimization methods are often inadequate in such settings because they cannot fully capture the strong nonlinearity, severe attenuation, molecular absorption, and time-varying behaviour of THz channels. To address these challenges, this paper proposes an online meta-optimized reinforcement learning framework for link-level adaptive modulation in THz communication systems. The proposed Hybrid RL–PSO framework integrates Proximal Policy Optimization (PPO) with Particle Swarm Optimization (PSO), where PPO performs state-aware modulation selection and PSO is periodically invoked during training to refine reward-shaping coefficients and selected learning hyperparameters. This closed-loop interaction enables the learning process to remain responsive to evolving channel conditions while improving policy-update stability and modulation-decision quality. Comprehensive simulations are conducted using M-QAM transmission over a physics-based THz channel model across a wide range of signal-to-noise ratio conditions. The proposed framework is evaluated against Deep Q-Network, Advantage Actor–Critic, standalone PPO, PSO-only optimization, GA-based optimization, and conventional threshold-based adaptive modulation. The results demonstrate improved uncoded bit error rate, link-level spectral efficiency, energy efficiency, throughput–reliability balance, and convergence stability under common evaluation conditions. Average reward is used only as an internal indicator of learning progression and convergence rather than as a direct cross-method performance measure. Additional evaluations under dynamic channel variations, multi-user operation, and parameter perturbations further demonstrate the robustness and scalability of the proposed approach. Overall, the proposed online Hybrid RL–PSO framework provides an effective solution for intelligent link-level physical-layer adaptation in future 6G THz communication systems.
Automated grading of Diabetic Retinopathy (DR) remains a challenging computer vision task due to the need to jointly model fine-grained local lesions and global structural patterns in retinal images. Existing hybrid architectures combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) typically rely on static feature fusion strategies, which limit adaptability to varying image characteristics and often result in feature misalignment. To overcome this limitation, this study proposes Reti-TransNet, a dynamic hybrid framework that integrates an EfficientNet-B0 backbone with a Swin Transformer branch through an Adaptive Gated Fusion (AGF) mechanism. The AGF module formulates feature fusion as a control problem and employs a learnable gating network to dynamically recalibrate channel-wise contributions of local and global representations. In addition, a Nelder–Mead optimization scheme is applied to fine-tune decision thresholds, directly maximizing the Quadratic Weighted Kappa (QWK) metric to preserve ordinal consistency. Experimental results on the APTOS 2019 dataset show that the proposed approach achieves a QWK score of 0.905 (95% CI: 0.88–0.93) and a multiclass accuracy of 84.31% in the primary configuration (mean QWK of 0.897 ± 0.006 and mean accuracy of 79.94% ± 3.21% across five independent runs), outperforming representative static fusion baselines. Further evaluation on the external IDRiD dataset indicates favorable primary screening performance under domain shift, achieving an AUC of 0.963 for healthy class detection, although detailed multiclass generalization remains limited due to sensitivity drops in early-stage disease detection. The proposed dynamic fusion strategy offers a promising and resource-efficient approach for computer-aided DR diagnosis.
This study presents a spatially integrated techno-economic methodology for replacing fossil fuels with residual biomass in agro-industrial heat generation, using chili dehydration plants in Zacatecas, Mexico, as a case study. The approach combines Geographic Information Systems (GIS), biomass resource assessment, transport network optimization, and energy system analysis to determine the technical, economic, and environmental feasibility of biomass-based heat supply. Forty-three dehydration plants comprising 504 drying tunnels were geolocated, and their thermal energy demands quantified, resulting in an annual biomass requirement of 44,899 t of dry woodchips. A GIS-based inventory identified 186 biomass collection points, mainly concentrated in the state’s western forested areas. Spatial buffer and route-network analyses revealed that 46,023 tons of residual biomass per year are available within a 150 km service radius, sufficient to supply the entire network. A transport routing algorithm based on a distance-priority greedy heuristic was developed and benchmarked against the exact linear programming (LP) solution for the underlying capacitated transportation problem (186 origins, 43 destinations, 7,998 candidate routes), achieving an optimality gap of 1.79 %. The resulting logistics system has a total transportation cost of USD 839,335 per year and a total delivered biomass cost of approximately USD 1.89 million per year. The technical analysis indicates that replacing the existing fuel‑oil‑based boilers with biomass boilers and the existing LPG burners with biomass burners can reliably meet the plants’ thermal demand. From an environmental perspective, the transition to biomass would reduce greenhouse-gas emissions (CO2e) by 81.3 % relative to the baseline of 35,050 t CO2/year, depending on the extent of seasonal biomass substitution. The results indicate the potential of residual biomass to provide a cost-effective, low-carbon heat source for chili dehydration in Zacatecas, highlighting the value of GIS-based planning and optimized logistics for regional bioenergy deployment.
The proliferation of electric vehicles (EVs) and the growing adoption of vehicle-to-grid (V2G) technology have established onboard chargers (OBCs) as critical interfaces between the power grid and EV battery systems. Among the constituent power conversion stages, the isolated bidirectional DC–DC converter is pivotal, as it must simultaneously provide galvanic isolation, support bidirectional energy flow, and maintain high efficiency across a wide output voltage range encompassing both 400 V and 800 V battery architectures. This paper presents a systematic review of isolated bidirectional DC–DC converter topologies for OBC applications, covering Dual Active Bridge (DAB), Inductor-Inductor-Capacitor (LLC), Capacitor-Inductor-Inductor-Capacitor (CLLC), and reconfigurable and two-stage configurations. Each topology is analyzed with respect to its operating principles, soft-switching capability, control strategies, voltage gain flexibility, and suitability for V2G operation. A comparative assessment categorizes the design strategies employed to extend the output voltage range, including variable DC-link control, topology reconfiguration, two-stage cascaded architectures, and hybrid modulation schemes. The role of wide-bandgap (WBG) semiconductor technologies—specifically silicon carbide (SiC) and gallium nitride (GaN)—in enabling high-frequency, high-efficiency converter designs is also examined. Key design requirements, regulatory standards, and recent industry trends are addressed. The findings indicate that CLLC resonant converters and reconfigurable topologies are the most promising options for next-generation bidirectional on-board chargers (OBCs). They offer an excellent balance of efficiency, wide voltage adaptability, and compatibility with both 400 V and 800 V battery platforms.
Traditional powder metallurgy (PM) techniques typically produce functionally graded composite materials (FGCMs) with planar interfaces, which limit the design possibilities and mechanical performance of the resulting materials. In this study, a novel separator-assisted molding technique was developed to fabricate FGCMs with non-planar interfaces. This technique utilizes adjustable separators to control the precise volume of each graded layer. Utilizing this method, three-layer FGCMs consisting of 0, 30, and 60 wt% B4C, as well as four-layer FGCMs with 0, 20, 40, and 60 wt% B4C reinforcement, were successfully fabricated. The microstructure, density, hardness, and transverse rupture strength (TRS) of the produced FGCMs were systematically evaluated. Microstructural characterization via SEM revealed that the reinforcement was uniformly distributed within the low-wt.% B4C layers, showing no evidence of delamination or macroscopic porosity. Hardness data were statistically analyzed using Levene’s test for homogeneity of variance, followed by a one-way ANOVA. Tukey’s post hoc test was subsequently applied to detect significant differences between the groups (p < 0.05). The highest Brinell hardness was obtained from the four-layered FGCM containing 40 wt% B4C, reaching 161 HBW. TRS data were validated via Levene’s test and statistically analyzed using a two-way ANOVA to evaluate the effects of the number of layers and the tested surface type. The maximum TRS value was achieved by the four-layered FGCM at 628,56 MPa, specifically when the load was applied to the reinforced surface. Overall, the results demonstrate that the separator-assisted molding technique effectively produces FGCMs with complex interfaces and enhanced mechanical properties, holding great promise for advanced structural applications.
Grinding burn is a critical defect that compromises the integrity of high-precision components. While analytical models rely on rigorous physical formulations, they often lack the flexibility to adapt to real variability. In contrast, standard machine learning (ML) also fails by ignoring grinding process understanding. To bridge this gap, this study transfers expert knowledge to the model by extracting signal features from the distinct phases of the grinding cycle. Signals from Computer Numerical Control (CNC), accelerometers, and acoustic emissions were segmented into approach, roughing, and spark-out stages to enable independent phase learning. Additionally, a feature selection strategy was proposed, successfully reducing the 1425 initial features to a set of just 32. Machine learning models reached 96.34% accuracy in known scenarios, far exceeding the 76.72% analytical baseline. When tested on grinding conditions different from those used during training, the staged methodology achieved 80%–85% accuracy. This performance not only outperforms the analytical results, but also notably surpasses standard ML global models, which failed with accuracies below 60%. Embedding the physical structure of the process into machine learning provides the contextual awareness necessary for reliable industrial monitoring, enabling superior generalisation and robustness compared to both traditional analytical models and standard data-driven techniques.
Metal surfaces sustain scratches and pits during manufacturing processes, which reduce product quality and system safety. In recent years, machine vision methods have been widely adopted for product quality inspection due to their high efficiency and non-contact features. However, highly reflective metals exhibit uneven illumination due to shape variations, which obscure tiny defects. Existing detection solutions struggle to reliably capture fine defects under uneven illumination and deploy efficient models in dynamic manufacturing environments. These challenges define the core difficulty of highly reflective metal inspection, which this review addresses through a research roadmap. We present a systematic review and prospective view of optical imaging principles and methods for defect detection on highly reflective metal surfaces. We systematically reviewed 2D and 3D defect detection techniques for highly reflective metal surfaces. This review further analyzes data-driven solutions, including automated model design and lightweight architectures, to address deployment challenges such as parameter redundancy and limited generalization. Meanwhile, it evaluates the industrial applicability of large vision models in few-shot or zero-shot scenarios. Finally, we outline systematic optimization mechanisms in manufacturing scenarios to address the challenges of highly reflective metal surface inspection within complex industrial environments.
