Wind turbine blades suffer from diverse and continually evolving damage, yet the most critical failure modes occur infrequently and are hard to document, leaving conventional supervised methods constrained by their training distribution and costly to retrain as new damage types emerge. To address this, we propose a training-free, knowledge-grounded inspection framework that integrates Retrieval-Augmented Generation (RAG) with Vision-Language Models (VLM). A multimodal knowledge base is constructed, comprising technical documentation, representative reference images, and domain-specific guidelines. A hybrid text-image retriever with keyword-aware reranking assembles the most relevant context to condition the VLM at inference, injecting domain knowledge without task-specific training. We evaluate the framework on 110 labeled blade images covering four mechanism-based damage categories: healthy state, surface damage, environmental damage, and structural damage. The Full RAG framework achieves an overall accuracy of 94.55%, with a macro-averaged precision of 0.9786, recall of 0.8750, and F1 score of 0.9055. Ablation studies demonstrate that both the hybrid text-image retrieval mechanism and the keyword-aware reranking step are individually necessary for robust multi-class damage identification, with single-modality retrieval leading to systematic failure on specific damage categories. Comparison against AnomalyCLIP confirms that structured domain knowledge integration is the key differentiating factor over general-purpose zero-shot vision models. Comparison against YOLO-based detectors further highlights the generalization limitations of supervised pipelines across imaging conditions and damage types. The proposed framework requires no task-specific model adaptation, supports on-the-fly knowledge base updates, and produces structured diagnostic reports with traceable reasoning, offering a data-efficient and interpretable solution for industrial blade inspection.

Multiphoton photoreduction enables high-fidelity fabrication of complex 3D microstructures, yet reliable process-structure-property (PSP) prediction remains difficult because the available data are sparse, heterogeneous, and interaction-dominated. In this regime, conventional feature-vector models are statistically underdetermined, making them prone to spurious correlations, poor regime transfer, and unstable post hoc explanations, whereas mechanistic pipelines depend on calibrated submodels that are rarely available during early process development. We present PSP-HDC, a graph-structured hyperdimensional computing framework that encodes a directed PSP graph as an internal prior for representation, inference, and explanation. A trainable scalar-to-hypervector encoder learns parameter-specific embeddings on a fixed hyperdimensional basis to accommodate heterogeneous scales and noise. Sample representations are then composed through graph-aligned binding and bundling along directed PSP dependencies, and prediction is performed by associative-memory retrieval against class prototypes. Because the same prototype memories support both decision making and attribution, PSP-HDC provides intrinsic explanations at the parameter, group, and within-group levels, while memory alignment and separation quantify prototype formation during training. On sheet-resistance regime prediction for the 3D platform, PSP-HDC achieves an accuracy of 0.910 +/- 0.077 over 1000 random splits and 0.896 under process-fold generalization, outperforming strong baselines.

Robotic additive manufacturing unlocks large build volumes, non-planar deposition, and even cooperative multi-arm fabrication capabilities beyond gantry-style systems. This flexibility arises from high degrees of freedom serial kinematics that let the end-effector reorient along complex toolpaths. However, unoptimized orientation choices can amplify configuration-dependent joint jerk and tracking error, eroding print accuracy. Although conventional solutions, such as configuration-space orientation search and jerk-constrained time-optimal planning can mitigate this issue, their iterative solvers scale poorly on long toolpaths limiting large-scale deployment. We introduce a learning-in-the-loop orientation-planning framework that minimizes joint-space jerk. A data-driven jerk surrogate predicts joint jerk from local toolpath features in a low-dimensional parameterization, enabling a rapid evaluation optimizer and avoiding exhaustive search. Our experimental results show that across three representative parts the planner produced reduced geometric deviation by up to 76.72% while lowering simulated joint jerk by up to 19.9%.

Selecting an encoding configuration for Hyperdimensional Computing (HDC) determines whether a model fits the memory budget of an Internet of Things (IoT) device, yet the knowledge needed to make that choice is dispersed across thousands of papers in a form no optimizer can query. The configuration space is large and each evaluation is expensive, so exhaustive search is infeasible; sequential optimizers are the standard remedy, but they begin from an uninformed prior and spend early evaluations exploring blindly. We present an agentic Retrieval-Augmented Generation (RAG) recommendation system that converts the HDC literature into an informed starting point for this search. A corpus of 1,518 HDC papers is indexed multimodally. A vision-language model (VLM) describes 1,755 figure regions extracted from source PDFs, making plotted evidence such as accuracy against dimension curves retrievable alongside prose and tables. Given a task profile stating the dataset characteristics, target hardware, objectives, and memory budget, the agent retrieves evidence under an allocation that reserves context for figures, and returns an encoding configuration with supporting citations. That configuration seeds the first trial of a constrained multi-objective search over accuracy and inference memory, which then refines it, and feasibility is enforced by external measurement rather than by the model. On network intrusion detection (CICIDS2017) and bearing fault diagnosis (CWRU), the literature-grounded start raised the hypervolume of the recovered Pareto front by 17.4% and 19.1% over an identically budgeted search from a random draw, at no additional evaluation cost. A natural extension is to let the system perform the data preprocessing that the practitioner now supplies, so that a deployment can be described end to end from raw sensor streams.

High-precision CNC machining of free-form aerospace components requires bounded compensations informed by inspection, simulation, and process knowledge. Off-the-shelf large language model (LLM) assistants can generate text, but they do not reliably execute risk-constrained multi-step numerical workflows or provide auditable provenance for high-stakes decisions. We present multi-agent knowledge analysis (MAKA), a human-in-the-loop decision-support architecture that separates intent routing, tools-only quantitative analysis, knowledge graph retrieval, and critic-based verification that enforces physical plausibility, safety bounds, and provenance completeness before recommendations are surfaced for human approval. MAKA is instantiated on a Ti-6Al-4V rotor blade machining testbed by fusing virtual-machining path-tracking error fields, cutting-force and deflection simulations, and scan-based 3D inspection deviation maps from 16 blades. The analysis decomposes deviation into an evidence-linked pathing component, a drift-based wear proxy capturing systematic evolution across parts, a residual systematic compliance term, and a variability proxy for instability-aware escalation. In a three-level tool-orchestration benchmark (single-step through ≥3-step stateful sequences), MAKA improves successful tool execution by up to 87.5 percentage points relative to an unstructured single-model interaction pattern with identical tool access. Digital twin what-if studies show MAKA can coordinate traceable compensation candidates that reduce predicted surface deviation from order 10^-2in to approximately ± 10^-3in over most of the blade within the simulation environment, providing a pre-deployment verification signal for risk-aware human decision-making.

Robotic laser profiling is widely used for dimensional verification and surface inspection, yet measurement fidelity is often dominated by sensor configuration rather than robot motion. Industrial profilers expose multiple coupled parameters, including sampling frequency, measurement range, exposure time, receiver dynamic range, and illumination, that are still tuned by trial-and-error; mismatches can cause saturation, clipping, or missing returns that cannot be recovered downstream. We formulate instruction-conditioned sensing parameter recommendation; given a pre-scan RGB observation and a natural-language inspection instruction, infer a discrete configuration over key parameters of a robot-mounted profiler. To benchmark this problem, we develop Instruct-Obs2Param, a real-world multimodal dataset linking inspection intents and multi-view pose and illumination variation across 16 objects to canonical parameter regimes. We then propose ScanHD, a hyperdimensional computing framework that binds instruction and observation into a task-aware code and performs parameter-wise associative reasoning with compact memories, matching discrete scanner regimes while yielding stable, interpretable, low-latency decisions. On Instruct-Obs2Param, ScanHD achieves 92.7

Prognostics & Health Management (PHM) systems collect high-rate, multi-sensor signals that must traverse bandwidth-limited networks and run on resource-constrained edge devices. Uniform, ratio-driven signal compression often erases weak yet diagnostic structures and degrades decision quality under operating drift. Many learned codecs store dense floating-point latent codes that are ill-suited to edge budgets, require per-sensor or per-fault specialization, and rarely embed physics. Under drift, these methods smear the sidebands and harmonics that carry critical evidence of degradation in vibration and acoustic signals. We introduce the Edge-aware Physicsinformed Concurrent Multiplexed Transformer (EPicMT), a signal-processing framework, which couples selective transmission with physics-guided signal reconstruction within a single architecture spanning compression, reconstruction, and decision tasks. EPicMT implements contentaware token selection (sampling) to control the number of transmitted coefficients, and structure-aware, non-uniform quantization that assigns finer steps to harmonic/envelope bands and coarser steps to broadband background, achieving bit-budgeted signal compression rather than mere dimensionality reduction. On the decoder, a physics-guided Transformer with time-frequency priors, multi-scale envelope consistency and a physically-constrained diffuser, regularizes reconstruction so that spectral energy, phase relations, and envelope trends are preserved. Concurrent reflects the joint optimization of diagnostic and prognostic objectives within the same network, while multiplexed denotes parameter sharing across tasks through a common backbone and token interface, enabling unified training and signal-adaptive bit allocation under varying operating conditions. Case studies on an in-house gearbox testbed, the CWRU bearing dataset, and NASA CMAPSS demonstrate high-fidelity signal reconstructions at compression ratios up to 64:1 while maintaining or improving fault-classification accuracy and remaining useful life prediction relative to representative autoencoder and Transformer baselines. The learned gating reduces transmitted bits and on-device compute without hand-tuned heuristics, and an embedded deployment confirms feasibility on resource-constrained hardware.

Precision process planning in Computer Numerical Control (CNC) machining demands rapid, context-aware decisions on tool selection, feed-speed pairs, and multi-axis routing, placing immense cognitive and procedural burdens on engineers from design specification through final part inspection. Conventional rule-based computer-aided process planning and knowledge-engineering shells freeze domain know-how into static tables, which become limited when dealing with unseen topologies, novel material states, shifting cost-quality-sustainability weightings, or shop-floor constraints such as tool unavailability and energy caps. Large language models (LLMs) promise flexible, instruction-driven reasoning for tasks but they routinely hallucinate numeric values and provide no provenance. We present Augmented Retrieval Knowledge Network Enhanced Search Synthesis (ARKNESS), the end-to-end framework that fuses zero-shot Knowledge Graph (KG) construction with retrieval-augmented generation to deliver verifiable, numerically exact answers for CNC process planning. ARKNESS (1) automatically distills heterogeneous machining documents, G-code annotations, and vendor datasheets into augmented triple, multi-relational graphs without manual labeling, and (2) couples any on-prem LLM with a retriever that injects the minimal, evidence-linked subgraph needed to answer a query. Benchmarked on 155 industry-curated questions spanning tool sizing and feed-speed optimization, a lightweight 3B-parameter Llama-3 augmented by ARKNESS matches GPT-4o accuracy while achieving a +25 percentage point gain in multiple-choice accuracy, +22.4 pp in F1, and 8.1x ROUGE-L on open-ended responses.

Industrial anomaly detection (IAD) systems must spot subtle defects with only a handful of reference images while production lines, lighting, and materials change daily, a setting where conventional CNN-based or hand-tuned one-class models quickly break down. Multimodal Large Language Models (MLLMs) offer the potential to overcome these limitations by effectively combining visual and textual data processing capabilities. However, they often lack essential domain-specific knowledge, such as industry-specific defect tolerances and reasoning steps, limiting their effectiveness in targeted IAD tasks. We introduce MoXpert, a multi-expert framework designed specifically to enhance MLLM performance for IADs. MoXpert leverages a gated Mixture-of-Experts structure to dynamically select and integrate expert module outputs tailored to the specific queries and characteristics of inspected objects. It consists of four interconnected expert modules: Reference Extractor, providing a contextual baseline through retrieval of similar normal images; Knowledge Guide, offering detailed, domain-specific insights; Reasoning Expert, facilitating structured and logical reasoning for complex queries; and Decision Maker, synthesizing inputs from all modules to deliver accurate and contextually aware responses. Evaluations conducted on the comprehensive MMAD benchmark demonstrate that MoXpert significantly enhances adaptability, precision, and robustness, effectively meeting the practical demands of real-world industrial anomaly detection. The code of the proposed method can be found at https://github.com/Ed1sonChen/MoXpert.

Energy-aware path planning is central to robotic manufacturing, as it demands accurate, low-latency predictions of both trajectory feasibility and energy consumption. Physics-based estimators are accurate but slow and platform-specific, while existing learned surrogates are fast yet often mis-score infeasible paths and transfer poorly across robot models. We present Kinematics-Guided Multi-Task (KG-MT), a kinematics-aware architecture that jointly performs prediction of path feasibility and energy consumption during robotic additive manufacturing processes. By injecting inverse kinematics as a shared computational foundation into a shared backbone, KG-MT internalizes reachability, joint and velocity limits, as well as local Jacobian conditioning, yielding features that benefit both tasks. We tune hyperparameters via Bayesian optimization and study two adaptation regimes, architecture-only and architecture-with-weights transfer, to reduce target-data needs and training time. Comprehensive evaluations under homogeneous and heterogeneous robot scenarios show that the proposed model not only outperforms traditional two-stage pipelines but also drastically reduces computation time while maintaining high prediction accuracy. In cross-robot transfer tests, KG-MT achieves 97.98 % feasibility accuracy with 2.60 % energy MAE in the homogeneous transfer setting and 96.82 % accuracy with 3.79 % MAE in the heterogeneous setting. Critically, for real-world additive manufacturing applications, KG-MT performs 312 times faster than analytical simulations and 2.2 times faster than cascaded neural network surrogates. KG-MT provides a practical foundation for cross-platform, energy-aware planning in robotic manufacturing, supporting path optimization, robot placement, and sustainable operations, and is readily extensible to additional objectives (e.g., jerk or thermal constraints) without re-architecting the model.

Wind turbine blades are a load-bearing element whose undetected damage propagates rapidly. Reliable, automated inspection is therefore a high-leverage problem. Existing supervised deep-learning approaches, however, are constrained by the scarcity of pixel-annotated damage data and produce models that cannot adapt to new defect types without retraining. Vision-language models (VLMs) offer a path around the data bottleneck through their generalist visual reasoning, but two obstacles remain: they lack domain-specific knowledge, and they cannot produce the pixel-level spatial outputs that maintenance planning requires. This work introduces a unified few-shot framework that resolves both obstacles within a single inference pass. We establish that VLM cross-modal attention, generated as a byproduct of caption synthesis, encodes spatially structured task-relevant information, and we introduce a knowledge-guided IoU-based mechanism that converts this passive signal into pixel-level segmentation masks without any segmentation training. Domain expertise is injected through a threetier hierarchical knowledge base, defect taxonomy, visual exemplars, and spatial patterns, that mirrors how expert inspectors progressively refine their assessment. On the public DTU dataset, the framework achieves 98.3% classification accuracy and 0.862 Dice coefficient using only 15 training samples per class, outperforming the strongest supervised baseline by 13.3%; cross-domain validation on the NEU surface defect benchmark confirms the framework transfers to a fundamentally different industrial inspection setting. The significance is twofold: methodologically, we show that attention as a side product of generation can be repurposed as an active grounding signal; practically, we demonstrate that few-shot VLM-based inspection is viable for industrial deployment where labeled data is scarce.

Smart manufacturing can significantly improve efficiency and reduce energy consumption, yet the energy demands of AI models may offset these gains. This study utilizes in-situ sensing-based prediction of geometric quality in smart machining to compare the energy consumption, accuracy, and speed of common AI models. HyperDimensional Computing (HDC) is introduced as an alternative, achieving accuracy comparable to conventional models while drastically reducing energy consumption, 200× for training and 175 to 1000× for inference. Furthermore, HDC reduces training times by 200× and inference times by 300 to 600×, showcasing its potential for energy-efficient smart manufacturing.

Robotic inspection relies on accurate sensing to acquire high-fidelity geometric measurements for defect detection and metrology. While prior work has focused on robot motion and viewpoint planning, how to configure sensing parameters remains largely underexplored, despite their decisive impact on measurement quality. We propose SenseHD, a robotic sensing system that formulates scanner configuration as an instruction-conditioned sensing decision. Instead of predicting precise parameter values, SenseHD treats sensing parameters as discrete sensing actions and selects stable sensing regimes through hyperdimensional associative memory. Experiments on a real robotic inspection platform demonstrate that SenseHD robustly selects appropriate sensing configurations and significantly improves inspection reliability, while remaining lightweight and efficient compared to baseline methods.

The escalating demand for enhanced efficiency, customization, and flexibility in manufacturing drives the development of multifunctional systems, such as 5-axis computer numerical control (CNC) machines that combine milling, drilling, and turning operations. Integrating in-situ sensing data (e.g., load, torque, axis position, and power) with explainable artificial intelligence (XAI) offers a pathway to decipher complex process dynamics and ensure part quality. However, conventional XAI methods require computationally intensive post hoc analyses that introduce latency, obstructing real-time monitoring in dynamic manufacturing systems, while traditional learning models exhibit inadequate generalization across heterogeneous, interdependent manufacturing tasks, leading to fragmented insights and scalability constraints. This paper introduces MuItiHD, a graph-based hyperdimensional computing framework that intrinsically integrates explainability, multi-task learning, and computational efficiency to overcome these limitations. By encoding multi-channel time series data into a structured graph, MuItiHD captures interdependencies among signals using hyperdimensional representations, enabling computationally efficient parallel processing and rapid inference. Its cognitive operations, such as bundling and binding, enhance multi-task learning by modeling diverse tasks simultaneously while also quantifying signal significance and certainty to deliver explainable predictions of geometric quality metrics (e.g., surface roughness and dimensional accuracy). A real-world case study on a 5-axis CNC process interprets the geometric quality of three tasks: left 25.4 mm counterbore diameter, right 25.4 mm counterbore diameter, and a 2.54 mm radius. The framework demonstrated superior performance, achieving an F1-score of 95.4% and outperforming alternative machine learning approaches in both quality prediction and training efficiency by up to 36x, establishing it as a robust solution for advanced manufacturing systems.

Additive manufacturing (AM) technologies have undergone significant advancements through the integration of cooperative robotics additive manufacturing (C-RAAM) platforms. By deploying AM processes on the end effectors of multiple robotic arms, not only are traditional constraints such as limited build volumes circumvented, but systems also achieve accelerated fabrication speeds, cooperative sensing capabilities, and in-situ multi-material deposition. Despite advancements, challenges remain, particularly regarding defect generation including voids, cracks, and residual stress. Various factors contribute to these issues, including toolpath planning (i.e., slicing strategies), part decomposition for cooperative printing, and motion planning (i.e., path and trajectory planning). This review first examines the critical aspects of system control for C-RAAM systems consisting of slicing and motion planning. The methods for the mitigation of defects through the adjustment of these aspects and the process parameters of AM methods are then described in the context of how they modify the AM process: pre-process, inter-layer (i.e., during layer pauses), and mid-layer (i.e., during material deposition). The application of advanced sensing technologies, including high-resolution cameras, laser scanners, and thermal imaging, for capturing of micro, meso, and macro-scale defects is explored. The role of digital twins is analyzed, emphasizing their capability to simulate and predict manufacturing outcomes, enabling preemptive adjustments to prevent defects. Finally, the outlook and future opportunities for developing next-generation C-RAAM systems are outlined.

Smart manufacturing requires on-device intelligence that meets strict latency and energy budgets. HyperDimensional Computing (HDC) offers a lightweight alternative by encoding data as high-dimensional hypervectors and computing with simple operations. Prior studies often assume that the qualitative relation between HDC hyperparameters and performance is stable across applications. Our analysis of two representative tasks, signal-based quality monitoring in Computer Numerical Control (CNC) machining and image-based defect detection in Laser Powder Bed Fusion (LPBF), shows that this assumption does not hold. We map how encoder type, projection variance, hypervector dimensionality, and data regime shape accuracy, inference latency, training time, and training energy. A formal complexity model explains predictable trends in encoding and similarity computation and reveals nonmonotonic interactions with retraining that preclude a closed-form optimum. Empirically, signals favor nonlinear Random Fourier Features with more exclusive encodings and saturate in accuracy beyond moderate dimensionality. Images favor linear Random Projection, achieve high accuracy with small dimensionality, and depend more on sample count than on dimensionality. Guided by these insights, we tune HDC under multiobjective constraints that reflect edge deployment and obtain models that match or exceed the accuracy of state-of-the-art deep learning and Transformer models while delivering at least 6x faster inference and more than 40x lower training energy. These results demonstrate that domain-aware HDC encoding is necessary and that tuned HDC offers a practical, scalable path to real-time industrial AI on constrained hardware. Future work will enable adaptive encoder and hyperparameter selection, expand evaluation to additional manufacturing modalities, and validate on low-power accelerators.

We introduce ScanBot, a benchmark for instruction-conditioned, high-precision surface scanning with robot-mounted industrial laser profilers. Unlike existing robot learning datasets that emphasize coarse behaviors such as grasping, navigation, or dialogue, ScanBot targets sensing-centric tasks where sub-millimeter motion continuity, strict stand-off control, and stable scanner settings are essential for acquiring usable geometry. The dataset contains scanning trajectories over twenty objects, including electronic components and structured 3D-printed parts, and spans six task types that range from broad inspection to fine-grained detail scanning and geometry-critical operations, including metrology and registration. Each episode is specified by natural language instructions and paired with synchronized first-person RGB-D, third-person video, laser height profiles, robot joint and pose traces, and scanner-parameter logs. These requirements expose a gap: despite recent progress, learning-based models often fail to produce stable and feasible scan motions under fine-grained instructions and real laser-profiling constraints. To reflect how industrial scanning is actually done, we evaluate methods through a two-stage pipeline. Stage I asks the model to "set up the sensor" by recommending scanner parameters, while Stage II asks it to "move like a scanner" by producing smooth, feasible trajectories that maintain stand-off and cover the intended region under precision demands.

Decentralized federated learning faces privacy risks because model updates can leak data through inference attacks and membership inference, a concern that grows over many client exchanges. Differential privacy offers principled protection by injecting calibrated noise so confidential information remains secure on resource-limited IoT devices. Yet without transparency, black-box training cannot track noise already injected by previous clients and rounds, which forces worst-case additions and harms accuracy. We propose PrivateDFL, an explainable framework that joins hyperdimensional computing with differential privacy and keeps an auditable account of cumulative noise so each client adds only the difference between the required noise and what has already been accumulated. We evaluate on MNIST, ISOLET, and UCI-HAR to span image, signal, and tabular modalities, and we benchmark against transformer-based and deep learning-based baselines trained centrally with Differentially Private Stochastic Gradient Descent (DP-SGD) and Renyi Differential Privacy (RDP). PrivateDFL delivers higher accuracy, lower latency, and lower energy across IID and non-IID partitions while preserving formal (epsilon, delta) guarantees and operating without a central server. For example, under non-IID partitions, PrivateDFL achieves 24.42

Wind turbine blade inspection using drone-based imaging has emerged as a highly promising solution for scalable, low-cost monitoring of large wind farms. However, the majority of existing visual inspection methods still rely heavily on either handcrafted features or large-scale labeled datasets to train traditional deep learning models. These approaches face significant challenges in terms of adaptability, requiring time-consuming retraining or fine-tuning whenever new defect patterns, lighting conditions, or inspection angles arise. In this work, we propose a novel few-shot visual reasoning pipeline based on visual -text Retrieval-Augmented Generation (RAG) integrated with a pre-trained Vision-Language Model (VLM), designed to reduce reliance on manual labeling and enhance adaptability across blade inspection scenarios. Unlike conventional pipelines, our system does not require task-specific fine-tuning. Instead, it performs in-context few-shot reasoning by retrieving relevant visual-textual examples from a structured knowledge base and prompting the language model to reason over these alongside the current images description. To demonstrate its practical potential, we construct a domain-specific knowledge base including structured textual files and optional curated image-caption examples. We show that our visual-text RAG-VLM framework is able to reason about damage type and severity in a flexible, interpretable, and data-efficient manner. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)