Bearings are critical components in machinery, and accurately predicting their remaining useful life (RUL) is essential for effective predictive maintenance. Traditional RUL prediction methods often rely on manual feature extraction and expert knowledge, which face specific challenges such as handling non-stationary data and avoiding overfitting due to the inclusion of numerous irrelevant features. This paper presents an approach that leverages Continuous Wavelet Transform (CWT) for feature extraction, a Channel-Temporal Mixed MLP (CT-MLP) layer for capturing intricate dependencies, and a dynamic attention mechanism to adjust its focus based on the temporal importance of features within the time series. The dynamic attention mechanism integrates multi-head attention with innovative enhancements, making it particularly effective for datasets exhibiting non-stationary behaviour. An experimental study using the XJTU-SY rolling bearings dataset and the PRONOSTIA bearing dataset revealed that the proposed deep learning algorithm significantly outperforms other state-of-the-art algorithms in terms of RMSE and MAE, demonstrating its robustness and accuracy.

The explainability of recommendation systems refers to the ability to explain the logic that guides the system’s decision to endorse or exclude an item. In industrial-grade recommendation systems, the high complexity of features, the presence of embedding layers, the existence of adversarial samples and the requirements for explanation accuracy and efficiency pose significant challenges to current explanation methods. This paper proposes a novel framework AdvLIME (Adversarial Local Interpretable Model-agnostic Explanation) that leverages Generative Adversarial Networks (GANs) with Embedding Constraints to enhance explainability. This method utilizes adversarial samples as references to explain recommendation decisions, generating these samples in accordance with realistic distributions and ensuring they meet the structural constraints of the embedding module. AdvLIME requires no modifications to the existing model architecture and needs only a single training session for global explanation, making it ideal for industrial applications. This work contributes two significant advancements. First, it develops a model-independent global explanation method via adversarial generation. Second, it introduces a model discrimination method to guarantee that the generated samples adhere to the embedding constraints. We evaluate the AdvLIME framework on the Behavior Sequence Transformer (BST) model using the MovieLens 20 M dataset. The experimental results show that AdvLIME outperforms traditional methods such as LIME and DLIME, reducing the approximation error of real samples by 50% and demonstrating improved stability and accuracy.

The original online version of this article was revised: the statement of Data availability and Competing interests have been added.

A correction to this article is available online at https://doi.org/10.1007/s43684-025-00110-3.

Transportation of heavy loads is often performed by multi-axle multi-steered heavy duty vehicles In this article a novel nonlinear optimal control method is applied to the kinematic model of the five-axle and three-steering coupled vehicle system. First, it is proven that the dynamic model of this articulated multi-vehicle system is differentially flat. Next. the state-space model of the five-axle and three-steering vehicle system undergoes approximate linearization around a temporary operating point that is recomputed at each time-step of the control method. The linearization is based on Taylor series expansion and on the associated Jacobian matrices. For the linearized state-space model of the five-axle and three-steering vehicle system a stabilizing optimal (H-infinity) feedback controller is designed. This controller stands for the solution of the nonlinear optimal control problem under model uncertainty and external perturbations. To compute the controller’s feedback gains an algebraic Riccati equation is repetitively solved at each iteration of the control algorithm. The stability properties of the control method are proven through Lyapunov analysis. The proposed nonlinear optimal control approach achieves fast and accurate tracking of setpoints under moderate variations of the control inputs and minimal dispersion of energy by the propulsion and steering system of the five-axle and three-steering vehicle system.

Water pipeline leaks pose significant risks to urban infrastructure, leading to water wastage and potential structural damage. Existing leak detection methods often face challenges, such as heavily relying on the manual selection of frequency bands or complex feature extraction, which can be both labour-intensive and less effective. To address these limitations, this paper introduces a Frequency-Informed Transformer model, which integrates the Fast Fourier Transform and self-attention mechanisms to enhance water pipe leak detection accuracy. Experimental results show that FiT achieves 99.9% accuracy in leak detection and 98.7% in leak type classification, surpassing other models in both accuracy and processing speed, with an efficient response time of 0.25 seconds. By significantly simplifying key features and frequency band selection and improving accuracy and response time, the proposed method offers a potential solution for real-time water leak detection, enabling timely interventions and more effective pipeline safety management.

The field of human motion data capture and fusion has a broad range of potential applications and market opportunities. The capture of human motion data for wearable sensors is less costly and more convenient than other methods, but it also suffers from poor data capture accuracy and high latency. Consequently, in order to overcome the limitations of existing wearable sensors in data capture and fusion, the study initially constructed a model of the human joint and bone by combining the quaternion method and root bone human forward kinematics through mathematical modeling. Subsequently, the sensor data calibration was optimized, and the Madgwick algorithm was introduced to address the resulting issues. Finally, a novel human joint motion data capture and fusion model was proposed. The experimental results indicated that the maximum mean error and root mean square error of yaw angle of this new model were 1.21° and 1.17°, respectively. The mean error and root mean square error of pitch angle were maximum 1.24° and 1.19°, respectively. The maximum knee joint and elbow joint data capture errors were 3.8 and 6.1, respectively. The suggested approach, which offers a new path for technological advancement in this area, greatly enhances the precision and dependability of human motion capture, which has a broad variety of application possibilities.

Modern industries dependent on reliable asset operation under constrained resources employ intelligent maintenance methods to maximize efficiency. However, classical maintenance methods rely on assumed lifetime distributions and suffer from estimation errors and computational complexity. The advent of Industry 4.0 has increased the use of sensors for monitoring systems, while deep learning (DL) models have allowed for accurate system health predictions, enabling data-driven maintenance planning. Most intelligent maintenance literature has used DL models solely for remaining useful life (RUL) point predictions, and a substantial gap exists in further using predictions to inform maintenance plan optimization. The few existing studies that have attempted to bridge this gap suffer from having used simple system configurations and non-scalable models. Hence, this paper develops a hybrid DL model using Monte Carlo dropout to generate RUL predictions which are used to construct empirical system reliability functions used for the optimization of the selective maintenance problem (SMP). The proposed framework is used to plan maintenance for a mission-oriented series k-out-of-n:G system. Numerical experiments compare the framework’s performance against prior SMP methods and highlight its strengths. When minimizing cost, maintenance plans are frequently produced that result in mission survival while avoiding unnecessary repairs. The proposed method is usable in large-scale, complex scenarios and various industrial contexts. The method finds exact solutions while avoiding the need for computationally-intensive parametric reliability functions.

Most machine learning-based remaining useful life (RUL) prediction methods only yield point predictions, and their “black-box” nature results in low interpretability. Stochastic process-based modeling can predict RUL probability density function (PDF), yet it often suffers from inaccurate modeling and failure to fully utilize historical degradation data of the same equipment type. To overcome these limitations, this paper integrates the two approaches and proposes an Attention-Gaussian-LSTM-Wiener (AG-LSTM-Wiener)-based RUL prediction method, enabling dynamic weighted fusion of predicted PDFs. An AG-LSTM-Wiener model with a two-branch structure is constructed. Health indicator (HI) is fed into the corresponding branch models to generate two different PDF curves. Decision blocks are employed to estimate RUL, from which weights are derived to achieve dynamic weighted fusion of the PDFs. Experiments on the CMPASS turbofan engine degradation dataset validate the proposed method’s effectiveness. Results demonstrate that the proposed method not only prevents PDF curve distortion but also improves the prediction accuracy compared with other methods. With the root mean squared error (RMSE) and Score reduced by 32.8% and 46.1% on average, and the mean squared error of PDF (

This research addresses the diagnosis of broken rotor bar faults in three-phase induction motors, focusing on steady-state conditions under different load levels and fault severity. Although numerous techniques exist, there is still a significant gap in comprehensive comparative evaluations that rigorously assess the interaction between signal processing, feature selection, and pattern classifiers, particularly concerning their robustness to noise and multiple performance criteria. An experimental investigation was carried out with electrical current and mechanical vibration signals, several signal preprocessing techniques, two feature selection strategies, Correlation-Based Feature Selection (CFS) and Wrapper, and a wide range of pattern classifiers, Decision Tree (DT), Naive Bayes (NB), Artificial Neural Network (ANN), and Support Vector Machine (SVM). The performance of the configurations was quantified by a multicriteria indicator, complemented by a dedicated robustness assessment by introducing white noise into the input signals. The most significant results reveal that vibration signals exhibit superior diagnostic robustness compared to electrical current signals, especially under noisy conditions. Furthermore, Wrapper-based feature selection consistently outperforms CFS, and configurations combining Wrapper with DT or NB classifiers emerge as the most suitable for detecting and diagnosing broken bars. Furthermore, the Wrapper-DT configuration efficiently classified defects even with the inclusion of 40% noise. This work provides data-driven insights into robust configurations for broken bar diagnosis, guiding the development of more reliable predictive maintenance systems, emphasizing signal modality, robust feature selection, and real-time applications.

Object detection serves as a challenging yet crucial task in computer vision. Despite significant advancements, modern detectors remain struggling with task alignment between localization and classification. In this paper, Global Collaborative Learning (GCL) is introduced to address these challenges from often-overlooked perspectives. First, the essence of GCL is reflected in the label assignment of the detector. Adjusting the loss function to transform samples with strong localization yet weak classification into high-quality samples in both tasks, provides more effective training signals, enabling the model to capture key consistent features. Second, the spirit of GCL is embodied in the head design. By enabling global feature interaction within the decoupled head, the approach ensures that final predictions are made more comprehensively and robustly, thereby preventing the two independent branches from converging into suboptimal solutions for their respective tasks. Extensive experiments on the challenging MS COCO and CrowdHuman datasets demonstrate that the proposed GCL method substantially enhances performance and generalization capabilities.

Significant progress has been made in distributed unmanned aerial vehicle (UAV) swarm exploration. In complex scenarios, existing methods typically rely on shared trajectory information for collision avoidance, but communication timeliness issues may result in outdated trajectories being referenced when making collision avoidance decisions, preventing timely responses to the motion changes of other UAVs, thus elevating the collision risk. To address this issue, this paper proposes a new distributed UAV swarm exploration framework. First, we introduce an improved global exploration strategy that combines the exploration task requirements with the surrounding obstacle distribution to plan an efficient and safe coverage path. Secondly, we design a collision risk prediction method based on relative distance and relative velocity, which effectively assists UAVs in making timely collision avoidance decisions. Lastly, we propose a multi-objective local trajectory optimization function that considers the positions of UAVs and static obstacles, thereby planning safe flight trajectories. Extensive simulations and real-world experiments demonstrate that this framework enables safe and efficient exploration in complex environments.

With the growing penetration of renewable energy, the impact of renewable uncertainties on power system secure operation is of increasing concern. Based on a recently developed linear power flow model, we formulate a chance-constrained optimal power flow (CC-OPF) in transmission networks that provides a concise way to regulate the security regarding both power and voltage behaviors under renewable uncertainties, the latter of which fails to be captured by the conventional DC power flow model. The formulated CC-OPF finds an optimal operating point for the forecasted scenario and the corresponding generation participation scheme for balancing power fluctuations such that the expectation of generation cost is minimized and the probabilities of line overloading and voltage violations are sufficiently low. The problem under the Gaussian distribution of renewable fluctuations is reformulated into a deterministic problem in the form of second-order cone programming, which can be solved efficiently. The proposed approach is also extended to the non-Gaussian uncertainty case by making use of the linear additivity of probability terms in the Gaussian mixture model. The obtained results are verified via numerical experiments on several IEEE test systems.

Vision Transformers (ViTs) have achieved state-of-the-art performance on various computer vision tasks. However these models are memory-consuming and computation-intensive, making their deployment and efficient inference on edge devices challenging. Model quantization is a promising approach to reduce model complexity. Prior works have explored tailored quantization algorithms for ViTs but unfortunately retained floating-point (FP) scaling factors, which not only yield non-negligible re-quantization overhead, but also hinder the quantized models to perform efficient integer-only inference. In this paper, we propose H-ViT, a dedicated post-training quantization scheme (e.g., symmetric uniform quantization and layer-wise quantization for both weights and part of activations) to effectively quantize ViTs with fewer Power-of-Two (PoT) scaling factors, thus minimizing the re-quantization overhead and memory consumption. In addition, observing serious inter-channel variation in LayerNorm inputs and outputs, we propose Power-of-Two quantization (PTQ), a systematic method to reducing the performance degradation without hyper-parameters. Extensive experiments are conducted on multiple vision tasks with different model variants, proving that H-ViT offers comparable(or even slightly higher) INT8 quantization performance with PoT scaling factors when compared to the counterpart with floating-point scaling factors. For instance, we reach 78.43 top-1 accuracy with DeiT-S on ImageNet, 51.6 box AP and 44.8 mask AP with Cascade Mask R-CNN (Swin-B) on COCO.

As autonomous driving technology advances from assisted to higher levels of autonomy, the complexity of operational environments and the uncertainty of driving tasks continue to increase, posing significant challenges to system safety. The key to ensuring safety lies in conducting comprehensive and rational risk assessments to identify potential hazards and inform policy optimization. Consequently, risk assessment has emerged as a critical component for ensuring the safe operation of higher-level autonomous driving systems. This review focuses on research into risk assessment for autonomous driving. It systematically surveys the state-of-the-art literature from three key perspectives: risk sources, assessment methodologies, data foundations, and system architectures. For each perspective, the paper provides an in-depth analysis of representative technical approaches, modeling principles, and typical application scenarios, while summarizing their research characteristics and applicable boundaries. Finally, this paper synthesizes the three fundamental challenges that persist in current research and further explores future directions and development opportunities. It provides a theoretical foundation and methodological references for the development of autonomous driving systems that exhibit high safety and reliability.

AutoML systems seek to assist Artificial Intelligence users in finding the best configurations for machine learning models. Following this line, recently the area of Automated Reinforcement Learning (AutoRL) has become increasingly relevant, given the growing increase in applications for reinforcement learning algorithms. However, the literature still lacks specific AutoRL systems for combinatorial optimization, especially for the Sequential Ordering Problem (SOP). Therefore, this paper aims to present a new AutoRL approach for SOP. For this, two new methods are proposed using hyperparameter optimization and metalearning: AutoRL-SOP and AutoRL-SOP-MtL. The proposed AutoRL techniques enable the combined tuning of three SARSA hyperparameters, being ϵ-greedy policy, learning rate, and discount factor. Furthermore, the new metalearning approach enables the transfer of hyperparameters between two combinatorial optimization domains: TSP (source) and SOP (target). The results show that the application of metalearning generates a reduction in computational cost in hyperparameter optimization. Furthermore, the proposed AutoRL methods achieved the best solutions in 23 out of 28 simulated TSPLIB instances compared to recent literature studies.

In many industrial settings, fleets of assets are required to operate through alternating missions and breaks. Fleet Selective Maintenance (FSM) is widely used in such contexts to improve the fleet performance. However, existing FSM models assume that upcoming missions are identical and require only a single system configuration for completion. Additionally, these models typically assume that all missions must be completed, overlooking resource constraints that may prevent readying all systems within the available break duration. This makes mission prioritization and assignment a necessary consideration for the decision-maker. This work proposes a novel FSM model that jointly optimizes system to mission assignment, component and maintenance level selection, and repair task allocation. The proposed framework integrates analytical models for standard components and Deep Neural Networks (DNNs) for sensor-monitored ones, enabling a hybrid reliability assessment approach that better reflects real-world multi-component systems. To account for uncertainties in maintenance and break durations, a chance-constrained optimization model is developed to ensure that maintenance is completed within the available break duration with a specified confidence level. The optimization model is reformulated using two well-known techniques: Sample Average Approximation (SAA) and Conditional Value-at-Risk (CVaR) approximation. A case study of military aircraft fleet maintenance is investigated to demonstrate the accuracy and added value of the proposed approach.

Machine learning, a revolutionary and advanced technology, has been widely applied in the field of stock trading. However, training an autonomous trading strategy which can effectively balance risk and Return On Investment without human supervision in the stock market with high uncertainty is still a bottleneck. This paper constructs a Bayesian-inferenced Gated Recurrent Unit architecture to support long-term stock price prediction based on characteristics of the stock information learned from historical data, augmented with memory of recent up- and-down fluctuations occur in the data of short-term stock movement. The Gated Recurrent Unit architecture incorporates uncertainty estimation into the prediction process, which take care of decision-making in an ever-changing dynamic environment. Three trading strategies were implemented in this model; namely, a Price Model Strategy, a Probabilistic Model Strategy, and a Bayesian Gated Recurrent Unit Strategy, each leveraging the respective model’s outputs to optimize trading decisions. The experimental results show that, compared with the standard Gated Recurrent Unit models, the modified model exhibits a huge tremendous/dramatic advantage in managing volatility and improving return on investment Return On Investment. The results and findings underscore the significant potential of combining Bayesian inference with machine learning to operate effectively in chaotic decision-making environments.

Autonomous driving systems (ADS) are at the forefront of technological innovation, promising enhanced safety, efficiency, and convenience in transportation. This study investigates the potential of end-to-end reinforcement learning (RL) architectures for ADS, specifically focusing on a Go-To-Point task involving lane-keeping and navigation through basic urban environments. The study uses the Proximal Policy Optimization (PPO) algorithm within the CARLA simulation environment. Traditional modular systems, which separate driving tasks into perception, decision-making, and control, provide interpretability and reliability in controlled scenarios but struggle with adaptability to dynamic, real-world conditions. In contrast, end-to-end systems offer a more integrated approach, potentially enhancing flexibility and decision-making cohesion.

This research introduces CARLA-GymDrive, a novel framework integrating the CARLA simulator with the Gymnasium API, enabling seamless RL experimentation with both discrete and continuous action spaces. Through a two-phase training regimen, the study evaluates the efficacy of PPO in an end-to-end ADS focused on basic tasks like lane-keeping and waypoint navigation. A comparative analysis with modular architectures is also provided. The findings highlight the strengths of PPO in managing continuous control tasks, achieving smoother and more adaptable driving behaviors than value-based algorithms like Deep Q-Networks. However, challenges remain in generalization and computational demands, with end-to-end systems requiring extensive training time.

While the study underscores the potential of end-to-end architectures, it also identifies limitations in scalability and real-world applicability, suggesting that modular systems may currently be more feasible for practical ADS deployment. Nonetheless, the CARLA-GymDrive framework and the insights gained from PPO-based ADS contribute significantly to the field, laying a foundation for future advancements in AD.