To address diagnostic delays and high false-alarm rates in multi-sensor systems, this paper proposes MIC-GAEAN, a lightweight network for anomaly detection. The framework constructs a global sensor correlation network using the Maximal Information Coefficient (MIC) to capture complex dependencies, which is dynamically refined via a lightweight attention mechanism. Notably, training excludes the target sensor's own historical data to prevent information leakage and enhance generalization. An adaptive MLP then performs point-level anomaly detection through single-step prediction. A novel Correlation Compensation Mechanism further uses healthy sensor data to set theoretical norms, distinguishing faults from normal variations and reducing false alarms. Validated on industrial data, MIC-GAEAN demonstrates high real-time accuracy, efficiency, and suitability for data-scarce, long-interval monitoring, offering a reliable solution for industrial system reliability.
Machine diagnostics is essential in maintaining high operational efficiency and minimizing machinery downtime. One way to achieve this goal is by implementing automated diagnostic systems. This paper presents a multi-sensor mapping system developed on a quadruped robotic platform. The robot combines diagnostic data from thermal and acoustic sensors with LiDAR data to create a 3D map of the working environment. Anomalous readings of temperature and sound pressure levels are identified and indicated on a 3D map with distinct color values through data fusion. Validation tests performed in a laboratory environment confirmed the feasibility of this approach by correctly identifying simulated cases of malfunctioning machinery. The system robustness was evaluated through repeatability tests. The proposed multi-modal mapping solution enables reliable monitoring of the technical state of machinery in complex environments.
Precise evaluation of vehicle CO2 emissions under real-world driving conditions is essential for meeting stringent regulations. While laboratory procedures such as WLTC ensure repeatability, they fail to reflect transient conditions typical of everyday driving, and PEMS-based RDE measurements remain costly. This study proposes a data-driven approach for reconstructing instantaneous CO2 concentration using onboard engine parameters-engine speed, exhaust gas temperature and exhaust mass flow rate-without direct CO2 sensing. Linear, nonlinear and ensemble machine learning models were evaluated using an RDE dataset of 5200 synchronized observations collected on a mixed urban-rural route. Ensemble methods, particularly Random Forest (R2 = 0.715, RMSE = 15,307 ppm) and XGBoost (R2 = 0.669), achieved the highest accuracy and reproduced steady-state conditions and rapid CO2 drops during fuel-cut events. The results confirm that reliable CO2 estimation can be achieved using a minimal OBD-based input set, enabling costeffective emission monitoring and real-time onboard applications.
Aviation piston engines (APEs), the primary propulsion systems in general aviation, are critical to flight safety. To understand the accident causation mechanism of APEs, this study proposes a data-driven methodology that integrates complex network (CN) theory with a three-step criticality analysis (TSCA). Maintenance records provide the data source for causal chain extraction and CN construction. In TSCA process, topology analysis using ICW-TOPSIS quantitatively identifies critical risk factors. Considering the uniqueness of component factors in physical systems, a hybrid metric integrating both objective and subjective dimensions is designed for accurate component evaluation. For critical risk path analysis, we develop an efficient path searching algorithm initiating from these prioritized components. The CN-TSCA findings enable the formulation of multi-phase safety control strategies. Leveraging historical maintenance records, this method effectively identifies risks in complex physical systems and provides a systematic framework for safety enhancement and preventive strategies.
This research addresses the inefficiencies in traditional forecasting methods for the intermittent and erratic demand for aircraft components, which typically results to either high inventory costs or expensive aircraft-on-ground situations from stockouts. The main objective is to build an enhanced forecasting model that uses installed base information to improve demand forecasting accuracy for aircraft components. An extensive literature review followed by a Delphi method study is used to identify the significant contextual factors impacting the demand and integrated into an Artificial Neural Network (ANN) forecasting model. The model's performance is evaluated using Mean Squared Error and benchmarked against the Croston method. This research provides efficient forecasting practices to the airline industry, probably reducing the operating expenses and improving the service quality.
Reliability of biomedical sensors is crucial for operational continuity and clinical safety, especially for critical patient monitoring systems. A Bayesian Deep Network (BDN) based model for predicting biomedical sensor failure probabilities is described in this work. The model analyzes various operational variables: sensor output signal, temperature, humidity, vibration, power consumption, and gives fault estimations in probabilities at every time interval. Differing from classical deterministic deep networks, the BDN considers weights and activation functions of the network as probabilistic variables, thus enabling the quantification of prediction confidence through epistemic uncertainty estimation via Monte Carlo sampling. Compared to standard Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) frameworks, the proposed method demonstrates 12% greater accuracy in positive predictions and 18% less false alarm rate. This suggests potential of Bayesian deep learning to enhance reliability for predictive maintenance of biomedical devices.
In the CAM machining of complex surfaces such as open blisk blades, unsmooth tool paths often occur at the common boundaries of compound surfaces. Conventional parameter alignment cannot guarantee C1 continuity in the Euclidean parameter domain. To generate boundary-consistent spiral tool paths on compound or trimmed surfaces, a method based on Coons reparameterization is proposed for computing the control lines required to generate spiral tool paths. On the unfolded Coons compound surfaces, spiral tool paths are generated directly by connecting the start and end points, avoiding the need to compute 3D points one by one. However, this may still fail to ensure C1 continuity at surface boundaries. Therefore, a 2D compound parameter domain is reconstructed via Coons mapping, which maintains the geometric features of the part and ensures C1-continuous tool paths at common boundaries. Simulation and machining of an open blisk blade verify the method's effectiveness, showing improved contour accuracy and an 18% reduction in surface roughness after optimization.
This paper proposes an adaptive subregion-based active Kriging (AS-AK) surrogate modeling approach. Firstly, an adaptive subregion decomposition strategy is developed to partition the candidate sample space into multiple concentric subregions, significantly enhancing the efficiency and accuracy of sampling. Subsequently, an active Kriging surrogate model is constructed, where the surrogate model is sequentially updated by iteratively selecting critical samples within each subregion to precisely approximate the highly nonlinear limit state function. Moreover, a collaborative multi-output surrogate modeling framework is further established to systematically handle correlations among multiple failure modes. Four benchmark numerical examples and an engineering application involving an aeroengine rigid-flexible coupling system illustrate that the proposed AS-AK method significantly outperforms existing reliability methods in both computational efficiency and accuracy.
During the cutting process of continuous miners, track plates are prone to deformation and damage, adversely affecting operational reliability and production efficiency. This study focuses on the EML340 continuous miner and establishes a simulation model based on the Discrete Element Method-Multi-Flexible Body Dynamics (DEM-MFBD) bidirectional coupling technology. Simulation results indicate that the load is mainly concentrated on the tight side of the track, with increasing intensity near the drive sprocket. The original track plate exhibits a maximum equivalent stress of 758.2 MPa and a safety factor of only 1.22. To address this issue, a multi-objective genetic algorithm implemented on the ISIGHT platform was employed to optimize the track plate structure, aiming to reduce weight and improve the safety factor. The optimized track plate weighs 25.42 kg, achieving a 9.2% reduction in weight compared to the original design, while the safety factor increases from 1.22 to 1.43. Industrial tests demonstrate that the optimized track plate operates without failure during service.
Light-emitting diodes (LEDs) have become indispensable in modern applications owing to their high energy efficiency, long lifespan, and robustness compared to conventional light sources. Given these attributes, the reliability of LEDs has become a crucial factor, directly influencing the ability of systems and devices to perform their intended functions over time. However, variations in materials, structures, and manufacturing processes introduce heterogeneity in their degradation behaviour, even under identical operating conditions. In applications demanding brightness stability, colour rendering, and reliability prediction, degradation homogeneity is crucial, making the analysis of LED heterogeneity essential. This article investigates such heterogeneity using feature extraction methods, kernel density estimation, and divergence measures based on degradation data obtained from optimized step-stress accelerated tests. The proposed approach is used to quantify and evaluate LED degradation variability and has clear implications for reliability assessment and predictive modelling.
Ground vibrations induced by tunnel blasting can severely impact nearby infrastructure. Therefore, accurate prediction of peak particle velocity (PPV) is essential for ensuring structural safety and engineering sustainability. This study proposes a PPV prediction model based on the Least Squares Support Vector Machine (LSSVM), optimised by a novel Adaptive Stagnation Whale Optimisation Algorithm (ASWOA). To address the limitations of the conventional WOA, a regionally dynamic threshold adjustment strategy based on stagnation counter is proposed. recording the number of consecutive iterations without improvement, and calculate the dynamic threshold by combining the decay coefficient to control the rate of change, thereby adaptively adjusts the trigger probability of spiral updates, improving global search capability. Compared with others models, the proposed method not only improves prediction accuracy but also ensure higher reliability in vibration prediction. Moreover, it provides an efficient tool for vibration control in tunnel blasting under complex geological conditions.
This paper addresses the challenges of weak parameter correlation, low detection accuracy, and poor interpretability in anomaly detection for rotating machinery under multi-condition operating scenarios. An explainable adaptive anomaly detection method is proposed. First, sensitivity and correlation analyses are employed to optimize the input parameters, and a spatial memory matrix is constructed by integrating an improved K-Nearest Neighbors algorithm with K-means clustering. Second, a multi-parameter anomaly detection model based on multivariate state estimation technique and sequential probability ratio test is developed to enable adaptive diagnosis of equipment operating conditions. Finally, error statistics are used to model the contribution trajectories of anomalous parameters, combined with a cumulative anomaly contribution rate metric to enhance the interpretability of anomaly localization. Experimental results show that the proposed method attains an average accuracy of 97.47% on multi-condition datasets, underscoring its wide applicability in industrial equipment monitoring.
For practical industrial production scenarios where collected vibration signals are easily interfered by environmental noise and bearing operating conditions are complex and variable, this paper proposes a Double Branch Lightweight Convolutional Neural Network (DBLCNN). The model adopts a dual-branch architecture: the onedimensional branch enhances feature extraction capability under low signal-to-noise ratio conditions, while the two-dimensional branch improves feature representation while significantly reducing the number of parameters. The complementary fault features extracted by the dual branches effectively enhance the accuracy of fault diagnosis. Under varying operating conditions, the model achieves an average accuracy of 95.58%; with the addition of 0 dB Gaussian white noise, its average accuracy under varying conditions remains at 90.17%. This study demonstrates that, even based on raw vibration signals without cumbersome preprocessing, the model can achieve excellent diagnostic performance in noisy environments and under variable operating conditions.
The key nodes in Transportation systems can improve the transportation system's performance efficiently and quickly when the maintenance resources are limited. A gated attention multi-channel graph convolutional network (KeyGAM-GCN) is proposed to identify the key nodes for complex transportation networks, which is an intelligent data-driven unsupervised key nodes identification method. In KeyGAM-GCN, a multi-channel graph convolutional network is developed to extract diverse topological and attribute features from transportation networks. A gated attention mechanism can fuse features by adaptively balancing the importance of different feature channels. To validate the effectiveness, experiments on 10 real-world transportation datasets are performed by comparing KeyGAM-GCN with several baselines in multiple metrics. The susceptible-infected-recovered-susceptible model is used to generate the nodes lables for evaluating the performance of the proposed method. The results show that KeyGAM-GCN can provide guidance for preventive maintenance for transportation systems.
This study presents a hybrid diagnostic approach combining the Continuous Wavelet Transform (CWT) and Convolutional Neural Networks (CNN) for assessing screw wear in a single-screw extruder operating under controlled conditions. Electrical current signals from the drive motor were analyzed to identify changes associated with the degradation of working components. CWT scalograms were used as time-frequency inputs for a CNN classifier, achieving a classification accuracy of 92.3% in distinguishing between new and worn screw states. Principal Component Analysis (PCA) confirmed clear separability of operating conditions, with the first two components explaining over 99% of the total variance. The results indicate that electrical signals contain diagnostically relevant information and that their combined analysis using CWT and CNN enables automated, non-invasive condition assessment with potential applicability in predictive maintenance systems without additional sensors.
This study investigates the mechanical safety of personal electric kick-scooters, specifically the vibration-induced failure of handlebar safety lock. Utilizing a dual-methodology approach, pre-experimental modal analysis via laser scanning was combined with real-world field measurements on asphalt and brick pavements using pneumatic and airless tires. Modal analysis identified a primary resonance frequency at 20 Hz, where airless tires exhibited significantly lower damping coefficients compared to pneumatic tires. Research results quantified a critical vibration that limit correlates with the unintended disengagement of the folding mechanism. On brick surfaces, airless tires produced vibrations 40% higher than pneumatic tires. These findings demonstrate that the reduction in damping provided by aftermarket airless tires directly compromises structural reliability, necessitating secondary locking redundancies for safe urban operation.
To address insufficient fault discriminability from excessive feature redundancy in rotating machinery analysis, an AHC-SCLS-driven clustering quality enhancement method for hierarchical sample entropy (HSE) is proposed. Hierarchical entropy decomposition first decouples multi-scale entropy features across frequency bands. Agglomerative hierarchical clustering (AHC) then constructs a hierarchical feature tree and reduces dimensionality via redundant attribute merging. A dual-criterion SCLS framework is integrated, where the average Silhouette Coefficient (SC) selects optimal cluster number and the Laplacian Score (LS) screens the most discriminative feature per cluster to form a refined subset. Experiments on the Ottawa University bearing and laboratory gearbox datasets validate the superiority of AHC-SCLS-optimized features. The PSO-SVM classifier achieves 99.29% accuracy on both datasets, other classifiers maintain 95.04%-97.87% accuracy, and the method outperforms mRMR, PCA and other traditional approaches across all classifiers.
Accurate prediction of bearing remaining useful life (RUL) is essential for reliable rotating machinery. However, multi-sensor degradation signals exhibit diverse temporal and spectral patterns that are often insufficiently captured by feature extractors using uniform processing, limiting their ability to model signal-specific degradation behavior and affecting prediction accuracy. This study proposes a frequency-adaptive feature extraction framework for bearing RUL prediction. The framework includes a Temporal Feature Extraction Network (TFEN) that employs dilated convolutions with adaptive configurations to capture degradation dynamics across multiple temporal scales, and a Transformer-based Spatial Feature Extraction Network (SFEN) to model inter-sensor dependencies. By aligning feature extraction with the dominant frequency characteristics of each sensor channel, the proposed method improves the representation of degradation features. Experiments on two bearing datasets demonstrate its effectiveness, showing consistently enhanced prediction accuracy relative to existing models.
This study highlights the problem of modernisation using a practical example of a rail vehicle pivot and constitutes an analysis of the impact of preventive renewals on the reliability of the upgraded structural node. The damage that occurs to the rubber vibration damper and, as a consequence, to the pivot itself, is a dependent failure. The developed model takes into account the random nature of the values of the operating times to failure of the pivot and the vibration damper cooperating with it, as well as the damage relationship between them. The model also includes preventive renewals of the vibration damper performed at fixed intervals of the vehicle's mileage, as well as information on its technical condition derived from inspections performed at shorter mileage intervals. As the analysis of the risk of damage to the pivot system has shown, changing the preventive renewal period of the vibration damper makes it possible to seek an acceptable risk value for the operator and to achieve the required safety integrity level (SIL).
In machining processes, monitoring the condition of cutting tools is crucial to ensure high productivity and consistent product quality. Tool Condition Monitoring (TCM) systems increasingly rely on machine learning techniques to analyze the large volume of sensor data generated during machining operations. In this study, real-time cutting force signals were acquired from milling experiments conducted under multiple tool wear conditions. Statistical feature extraction was performed on the force signals, followed by feature selection using decision tree-based importance analysis. A K-Nearest Neighbors (KNN) classifier was employed for tool condition classification, and hyperparameter tuning was carried out to enhance model performance. A comparative analysis of force signals revealed that the augmented force data in the X-direction achieved superior classification performance, with a maximum accuracy of 96%, compared to 78% for the Y-direction force signals. Data augmentation significantly reduced Type II error, which decreased from 3.04% to 0.14% for the X-direction force data. Hyperparameter tuning further improved model generalization, resulting in a testing accuracy of 95% and a training accuracy of 98% for the tuned KNN model. To enhance transparency and trustworthiness, a model-agnostic white-box interpretability framework was integrated with the KNN classifier, providing both global feature influence analysis and local, instance-level explanations of classification decisions. The proposed approach enables clear identification of dominant force features influencing tool wear classification and supports informed decision-making for tool maintenance. The results demonstrate that combining KNN classification with a model-agnostic interpretability layer offers an effective and transparent solution for force-based Tool Condition Monitoring in milling operations.
