Overcoming the inefficiency in the procurement of railway wagon components is a crucial step in promoting the modernization of railway enterprises' governance capabilities. This study formulates components purchasing for railway wagon as a two-stage Markov Decision Process (MDP) that jointly optimizes planned purchase quantities and supplier selection. Inputs to the MDP (demand signals and inventory) are treated as pre-processed operational forecasts from the enterprise data platform; the paper therefore focuses on the design of state/action spaces, reward construction, and the value-iteration solution for producing actionable purchasing decision. Using real business records from a large railway equipment group, simulation results demonstrate that the MDP-based decision reduces total purchasing cost, shortens purchasing lead time, and improves supplier quality metrics compared with legacy manual strategies.
In intelligent welding, achieving the automation of weld quality inspection is a significant challenge, and weld seam marking is of crucial importance. For this purpose, a method based on edge detection using binary image preprocessing was developed on the MATLAB platform. Compared with the traditional multi-sensor fusion approach, this method does not require complex sensor integration, simplifying the implementation process. Compared with neural network methods, it is more flexible and simpler. The method first preprocesses the image into a binary image and then compares the weld seam feature marking with the Roberts, Prewitt, Sobel, and Canny operators. The results show that the Canny operator demonstrates a significant performance advantage in the comparison of four indicators: point sharpness, entropy, average gradient, and Quality Assessment of Blended Features. Its performance is 3 to 25 times that of other operators, and it performs best in weld seam feature texture detection, showing high robustness.
This study proposes a wave-induced pneumatic energy converter that transforms vertical wave motion into airflow to produce electrical power. A laboratory-scale prototype was fabricated and tested by means of a Scotch-yoke mechanism that replicates wave-induced displacement. Owing to its compact architecture and reduced component count, the device can be manufactured at relatively low cost. Under regular-wave conditions (height = 0.5 m; period = 1 s), the mean outlet air velocity reached 30.72 ms-1 and the peak pneumatic power was 43.34 W. The measured energy-conversion efficiency was 25.71%, corresponding to a volumetric energy density of approximately 875 Wm-3. Although the experiments were restricted to laboratory scale, the results confirm proof-of-concept. The observed performance and characteristics indicate potential for deployment in near-shore and off-shore environments; nevertheless, validation for full-scale sea conditions remains outside the scope of the present work.
The global information perception of power equipment is the key to supporting the efficient and stable operation of the new power system. This paper adopts the digital twin technology and constructs a new framework for panoramic perception of transformer vibration status. To address the difficulties in obtaining power defect samples, the dominance of normal samples, and the reliance on large-scale data of multi-modal large models, a power defect data enhancement method based on diffusion models is proposed. Under the premise of ensuring the rationality of the generated image structure, this method utilizes the trained multi-modal large model Qwen-VL-Max to extract the high-order semantic information of real power scene images and combines the prompt engineering technology to generate synthetic images with power defect features and high quality. Moreover, for the common problem of data sparsity in multi-modal recommendation systems, a multi-modal fusion recommendation algorithm based on collaborative self-supervised learning is proposed. This algorithm effectively enhances the representation ability of multi-modal data through the joint learning of the deep features of the data, thereby alleviating the problem of performance decline in recommendations caused by data sparsity. Experimental results show that, compared with the current mainstream multi-modal recommendation algorithms, this algorithm has significant improvements in multiple recommendation evaluation indicators.
The financial sector is undergoing a profound transformation with the integration of artificial intelligence (AI) and cloud computing technologies. A notable advancement is the deployment of a deep learning classification system that integrates Bidirectional Gated Recurrent Units (BiGRU) with the Fruit Fly Optimization Algorithm (FOA) to enhance complex banking operations. The BiGRU model efficiently analyzes financial transactions, customer profiles, and risk patterns by processing sequential data with long-term dependencies. FOA, inspired by the foraging behavior of fruit flies, optimizes the network's performance and computational efficiency. A cloud-based implementation of the BiGRU-FOA framework ensures scalability, real-time processing, and seamless integration with existing banking infrastructure. Experimental results demonstrate that BiGRU-FOA outperforms traditional machine learning techniques and standalone deep learning models in financial dataset classification, achieving superior accuracy, precision, and recall. This model enhances fraud detection, customer segmentation, and credit risk assessment, paving the way for more efficient and intelligent banking operations. By leveraging this advanced AI-driven framework, banks can improve decision-making processes, enhance operational efficiency, and offer personalized financial services. This research highlights the potential of deep learning and optimization technologies in revolutionizing the banking sector, enabling a more secure, efficient, and customer-centric approach.
Accurate supply chain demand forecasting is critical for inventory optimization and risk reduction. This study proposes three enhanced forecasting models, Dynamic Weight Fusion-MHA model (DWFMHA_UDAF). These models are built upon Bidirectional Long Short-Term Memory (BiLSTM) and Nonlinear Autoregressive models with exogenous inputs (NARX) to address the limitations of fixed-weight neural-ensemble architectures such as UDAF. By introducing Dynamic Weight Fusion (DWF) and multi-head attention (MHA), the proposed architectures adaptively reflect temporal demand shifts. The proposed models are evaluated through a focused methodological validation using a highly volatile time-series (Store 1) from the Rossmann Store Sales dataset. This specific store was selected as a complex testbed to rigorously assess the adaptive capability of the DWF mechanism under dynamic demand shifts and exogenous influences. Empirical validation on the Rossmann Store Sales dataset, benchmarking ten forecasting models, showed that DWF_UDAF achieved the lowest Mean Absolute Error (MAE = 0.147). In contrast, the statistical ARIMAX baseline recorded the lowest Root Mean Squared Error (RMSE = 0.206). Statistical tests, specifically one-way ANOVA and Tukey's HSD, confirmed that DWF_UDAF outperformed both fixed-weight and attention-based ensemble architectures in MAE (p < 0.001). In contrast, MHA-based models exhibited degraded accuracy. This study attributes this performance drop to the structural mismatch between the high data requirements of Multi-Head Attention and the limited sample size of a single store, which led to overfitting and attention weight collapse. This work provides both theoretical and practical evidence that adaptive-fusion mechanisms outperform static structures in complex supply chain forecasting scenarios.
This paper presents research on the optimisation of a vibratory transport on an assembly line. The study covers a special case of vibratory transport characterised by uninterrupted contact between the transported body and the vibratory trough's surface, and the interaction is modelled using Coulomb's model with stiction. In the first phase of the research, an analysis of external forces acting on the body during relative motion is conducted, and the conditions for non-hopping vibratory regimes are defined. The differential equations of relative motion are derived, and the precise moments at which relative sliding on the vibratory conveyor occurs are analytically determined. The proposed theoretical model of vibratory motion is validated with computer simulations performed in SolidWorks Motion Analysis. The obtained simulation results confirm the analytical predictions and are discussed in detail at the end of the paper.
Social media has become an important channel for expressing emotional experiences and potential psychological distress, making automated psychological state recognition a key technical challenge for early risk warning systems. Psychological signals in text are distributed across multiple linguistic levels, ranging from character-level expressive variations to word-level semantics and sentence-level psychological structure, which limits the effectiveness of single-scale models. This paper proposes a multi-scale deep learning architecture for psychological state recognition from social media text. The approach integrates character-level and word-level representations, multi-scale convolutional modules for local semantic extraction, attention-based global semantic modeling, and cross-scale feature fusion. By jointly capturing fine-grained linguistic cues and global psychological context, the proposed model enhances the discriminative power of psychological representations. Experiments conducted on a multi-class mental health text dataset demonstrate that the proposed method consistently outperforms traditional machine learning models, conventional deep learning architectures, and attention-enhanced baselines in terms of accuracy, precision, recall, and F1-score. Furthermore, the model outputs are transformed into temporal risk signals, enabling the identification of weak, accumulating, and accelerating psychological risk patterns. The results indicate that multi-scale text modeling provides an effective technical solution for psychological state recognition and establishes a practical basis for the development of early psychological risk warning systems.
Person re-identification (ReID) in large-scale surveillance requires methods that are both accurate and efficient. While deep hashing enables compact binary representations, it often suffers from accuracy degradation due to the domain gap between raw features and hash codes. This paper proposes a unified, open-source framework for fast person ReID that introduces a cross-domain loss function to explicitly bridge the feature and hash spaces. Our model-agnostic training strategy integrates seamlessly with existing architectures such as ResNet and OSNet. Experiments on Market1501 and CUHK03 demonstrate that the proposed framework outperforms state-of-the-art deep hashing and fast ReID methods, achieving up to 8.61% higher mean Average Precision (mAP). Extensive ablation studies validate the contribution of the cross-domain loss, and evaluations across multiple backbones confirm the framework's versatility. The results show that our approach not only improves accuracy but also provides a strong, reproducible baseline for efficient person re-identification.
This research presents a meticulously engineered dual-port, dual-radial implantable MIMO (Multiple Input Multiple Output) antenna specifically designed to operate within the 1.39-1.42 GHz frequency band of the Wireless Medical Telemetry System (WMTS). The proposed antenna, with compact dimensions of 19 & times; 19 & times; 1.27 mm(3), utilizes Rogers RO6010 as both substrate and superstrate, featuring a dielectric constant (epsilon(r)) of 10.2, a thickness of 0.635 mm, and a loss tangent of 0.0023. The design achieves high isolation (> 20 dB) between antenna elements, ensuring minimal mutual coupling, enhanced signal integrity, and reduced interference in implantable biomedical communication environments. The antenna demonstrates excellent performance metrics including a low Envelope Correlation Coefficient (ECC < 0.1), signifying robust signal decorrelation and high data reliability. Furthermore, the diversity gain (DG) approaches 10 dB, improving reception diversity and mitigating multipath and fading issues. Despite its compact and low-power design, the antenna achieves a peak gain of-28 dB and a measured channel capacity (CC) of 8 bps/Hz, indicating suitability for high-data-rate applications. Fabrication and measurement results exhibit strong agreement with simulated data, confirming the proposed antenna's practical feasibility for next-generation biomedical implant communication systems.
The widespread adoption of renewable energy within integrated energy systems introduces significant uncertainty, posing challenges to their stable and economical operation. Mitigating the costly degradation of electrochemical energy storage equipment while effectively managing fluctuations on both the generation and load sides has become a critical issue. Against this backdrop, this paper proposes a scheduling framework integrating interval-type robust optimisation with a two-layer model predictive control approach. During the day-ahead planning phase, an interval-type robust scheduling model is constructed. This model employs k-means clustering to partition uncertainty sets for renewable generation and load demand, thereby circumventing the excessive conservatism inherent in traditional robust optimisation. For the intraday phase, a two-layer model predictive control framework is established to smooth the rolling optimisation time scale differences across electricity, heat, gas, and hydrogen loads. Concurrently, a storage lifetime cost term is incorporated into the optimisation objective to extend the energy storage system's lifespan. This is combined with an electricity price-based demand response model to enhance scheduling flexibility. Simulation results demonstrate the proposed method's multifaceted advantages. Compared to conventional robust optimisation, the interval-based robust optimisation model maintains high user satisfaction (96.43%) while reducing average daily system operating costs by approximately 8.24% and lowering the monthly capacity degradation rate of the energy storage system from 1.831% to 1.617%. The intraday robust model predictive control (MPC) demonstrated robust performance under extreme scenarios, with total cost increases of merely 2.91%, significantly lower than the 16.23% observed in conventional MPC. The proposed scheduling strategy, integrating interval robust optimisation with model predictive control, offers an effective solution for addressing robustness, economic efficiency, and equipment lifespan management within integrated energy systems under renewable energy and load uncertainty.
This study investigates the influence of arc length correction and dynamic correction on the geometry and stability of bead-on-plate welds produced using the MIG Pulse Synergic process. S235 carbon steel was used as the base material, with Inconel 718 as the filler metal. Welds were evaluated through visual inspection, 3D scanning, and statistical analysis. Results showed that arc length correction had a significant impact on bead width, height, weld toe angle, and heat input, while dynamic correction had a more limited effect, primarily influencing process stability and bead width at extreme settings. The findings provide guidance for optimizing welding parameters to improve WAAM process.
This article takes false information in social media as a research case to explore the influence mechanism and action path of users' information verification behavior, aiming to enhance the public's ability to distinguish online information and promote the self-purification of the online environment. This study, based on the Fine Processing Possibility Model (ELM), combined with questionnaire surveys and structural equation models, constructs a model of the influence mechanism of social media users' false information verification behavior, and conducts empirical tests on the relevant influencing factors. This research can provide a reference for the management of social media platforms and the prevention of false information by users. Using only one modal discriminator cannot fully exploit the invariance between modalities, thus limiting the accuracy of cross-modal retrieval. This study innovatively constructs a cross-modal feature alignment framework based on dual-channel all-modal encoders and adversarial learning. This model adopts a parallel architecture design, featuring two feature encoding channels: visual and text. It maps heterogeneous data to a unified representation space through a deep neural network. During the model optimization process, we embed a classification and discrimination module in the hidden layer of the feature extraction network and adopt an adversarial training strategy for multi-objective optimization, so that the obtained shared feature space simultaneously possesses: 1) cross-modal semantic consistency; 2) Modal invariant feature expression ability. The experimental results show that an individual's emotional tendency and risk perception level can significantly affect their information discrimination behavior patterns. These two psychological factors constitute the core dimensions of users' information verification decisions.
In the study, the influence of rake angle, clearance angle, inclination angle, approach angle and nose angle on the dimensional deviation and roughness of the machined surface during dry turning of Ti6Al4V alloys was investigated. The experimental investigation utilized a custom design of experiments. The results indicated that the dimensional deviation varied between 0.053 mm and 0.081 mm, while the surface roughness ranged from 1.892 mm to 2.141 mm, depending on the different combinations of insert angles. An analysis of the results showed significant effects of both the main and quadratic interactions of the input parameters on the output parameters. Notably, altering the angles of the cutting insert affected dimensional deviation by 52.83% and surface roughness by 13.16%. Additionally, a strong correlation was found between the output parameters. Confirmation experiments were conducted to validate the developed predictive equations, and the percentage errors demonstrated the accuracy of the modelling process.
This study aims to evaluate the impact of artificial intelligence (AI) integration on the performance and governance efficiency of financial institutions. To address potential endogeneity concerns arising from reverse causality and omitted variable bias, we employ System Generalized Method of Moments (System GMM) estimator, complemented by Fixed Effects (FE) and Random Effects (RE) models for robustness checks. Our findings indicate that AI integration significantly enhances return on assets (ROA), operational efficiency, risk-adjusted returns, and customer satisfaction while reducing compliance costs and regulatory breaches. However, challenges such as algorithmic bias and workforce displacement must be addressed. In conclusion, AI offers substantial benefits to financial institutions, but ethical considerations and strategic workforce planning are essential for sustainable integration. These insights provide valuable guidance for financial institutions and policymakers aiming to harness AI's potential while mitigating associated risks.
Intrusion Detection Systems (IDSs) in Internet of Things (IoT) environments face persistent challenges, including class imbalance in network traffic data and the limited interpretability of black-box machine learning models. This paper proposes a novel, interpretable framework that effectively addresses both concerns. We introduce a Diffusion Model-based Synthetic Data Generator (DM-SDG) coupled with Prototype-Based Undersampling (PBUS) to mitigate class imbalance issues without compromising data integrity. For enhanced feature selection and dimensionality reduction, a dual-stage feature refinement strategy is employed using Self-Supervised Feature Filtering (SSFF) and SHAP-Guided Recursive Pruning (SGRP). Our classification stage incorporates Graph Attention Networks (GATs) and Transformer-based Intrusion Detection Systems (T-IDS), which provide improved context-awareness and sequence modeling in dynamic IoT environments. To enhance transparency and model trustworthiness, we integrate three explainability mechanisms: Counterfactual Explanations (CE), SHAP Interaction Values, and Explainable Concept Activation Vectors (ECAVs), enabling both global and local interpretation of detection decisions. The proposed solution is evaluated on benchmark datasets including CICIDS2018, CIC-ToN-IoT, and NF-UNSW-NB15-v2. Experimental results demonstrate accuracy improvements ranging from 0.5% to 2.4%, along with consistent F1-score and MCC gains of 1.5-3.5% over leading baselines such as CTGAN-ENN. Our framework achieves a balanced trade-off between detection accuracy, computational efficiency, and explainability, making it highly suitable for deployment in real-time IoT security infrastructures.
Load frequency control is crucial for keeping the stability of power systems by balancing energy supply and demand in electrical grids, essential for an accurate and fast-acting controller to address sizeable parametric uncertainties. In particular, the growing population of electric vehicles (EVs) requires advanced control and optimization strategies to ensure secure and stable operation. In this paper, the effect of the INFO algorithm on load frequency control is investigated as a new method in the system with renewable energy sources and EV Battery. The novelty of the proposed study lies in the INFO-based optimal tuning of the controller for a PV-thermal system with integrated EV battery and its detailed comparison with several traditional and recent optimization algorithms such as genetic algorithm (GA), firefly algorithm (FA), chess algorithm (CA), flood algorithm (FLA), sinhcosh algorithm (SCHO), artificial rabbits optimization (ARO) and black widow optimization algorithm (BWOA). The obtained results show that the controller optimized with the INFO algorithm provides a robust dynamic response with fast settling time and the lowest overshoot and undershoot values. Compared to the results of other controllers, an average improvement of approximately 30% in system frequency overshoot and approximately 40% in settling time was achieved.
Hyperspectral (HS) pansharpening aims to generate high-spatial-resolution hyperspectral (HRHS) images by fusing panchromatic (PAN) images with low-spatial-resolution hyperspectral (LRHS) images. However, many existing HS pansharpening methods fail to capture global dependencies between cross-modal features, leading to spectral and spatial distortions.To address this issue, we propose a hyperspectral pansharpening network based on information cross embedding (HPN-ICE). The model progressively fuses HS and PAN image features through two modules: the global feature fusion module (GFFM) and the multi-directional feature enhancement module (MFEM). In GFFM, a feature embedding fusion module (FEFM) is firstly designed based on the information cross embedding, which efficiently fuses spectral and spatial features by establishing cross dependencies between two modal features. Then, a frequency-domain channel attention module (FCAM) is constructed to enhance the global spectral information in the frequency domain. MFEM is constructed to enhance the local details of fused features in multi-dimensional directions. Extensive experiments conducted on three widely used datasets demonstrate that HPN-ICE achieves significant improvements in both spatial and spectral quality metrics over some state-of-the-art (SOTA) methods. The code will be released on GitHub.
The rapid growth of Internet of Vehicles (IoV) platforms has raised significant concerns about data security and storage efficiency. This study proposes a novel approach that is integrating cloud computing, machine learning, and advanced database management to enhance IoV data security and storage. MongoDB is utilized for data storage. The logistic regression and Support Vector Machine (SVM) algorithms are combined for security detection. By this way, our method demonstrates improved performance over traditional approaches. Experimental results showed a 4.60% increase in data recognition precision, a 4.66% higher recall, and a 2.55% improvement in the F1 score compared to existing models. The proposed system also exhibits enhanced data storage efficiency and robust security detection capabilities. These findings demonstrate our method has significant potential for improving IoV platform security and data management in real-world applications.
This paper presents a systematic implementation and analysis of a multi-stage BIOS boot process for the D1-H RISC-V application processor, addressing the critical challenges of limited on-chip storage and complex memory management requirements in modern embedded systems. We propose a three-stage boot architecture integrating on-chip BROM firmware, Secondary Program Loader (SPL), and main program execution, alongside an efficient storage allocation strategy utilizing external Nand Flash and DRAM. Our implementation demonstrates significant technical innovations in three key areas: (1) a modular storage structure design that optimizes memory utilization across different boot stages, achieving efficient code migration between Nand Flash (128 MB) and DRAM (512 MB); (2) an adaptive boot process that enables flexible configuration for various startup scenarios, supporting both development and production environments; and (3) a novel engineering framework that enhances code portability and maintainability. Performance analysis reveals that our implementation achieves a boot time of 10 ms for the complete startup sequence, with memory utilization efficiency of 15% compared to conventional approaches. The system successfully manages code migration between storage media with a transfer rate of 100 MB/s, demonstrating reliable operation across multiple test scenarios. We validate our design through comprehensive testing on the ADL-D1-H platform, showing successful integration with development tools and supporting direct program downloads through serial ports. This work provides practical insights for BIOS design in RISC-V systems and establishes a replicable framework for implementing efficient boot processes in resource-constrained embedded environments. The proposed solution eliminates the need for external download devices and enables direct serial port programming, significantly simplifying development, research, and remote update processes.
