Under variable conditions, different fault types in complex mechanical systems exhibit distinct cross-component causal coupling patterns, leading to fluctuations in monitoring signals and unstable feature representations. To address these issues, this study proposes a novel mechanical fault diagnosis method, termed causal coupling modeling and multi-order spectral representation alignment. First, a neural Granger mechanism is used to quantify inter-channel causal strengths, while category-level priors and sample-level dynamic information are jointly incorporated to construct an interpretable cross-component dynamic causal graph. Second, a multi-order spectral attention representation module is developed to achieve joint graph and feature mapping of raw time-domain signals and adaptive fusion of multi-scale features, thereby enhancing fault-related feature representation under variable conditions. Finally, a primary-auxiliary view collaborative alignment strategy is introduced to promote cross-view information complementarity and graph-level representation consistency, thus alleviating feature shifts caused by condition variations. Experimental results demonstrate that the proposed method effectively captures cross-component causal coupling patterns, improves fault discriminability, and achieves superior diagnostic performance with improved stability under the evaluated varying-condition scenarios and enhanced interpretability.
Artificial Intelligence (AI) facilitates adaptive, real-time energy management in residential microgrids operating under uncertain and dynamic conditions. Traditional rule-based and static optimisation methods are often inadequate for addressing fluctuations in electricity prices, household demand, renewable generation, and electric vehicle (EV) availability. To address these challenges, this study presents a hierarchical real-time energy management framework that integrates Model Predictive Control (MPC) with Deep Double Q-Network (DDQN) reinforcement learning. The proposed framework accommodates microgrids with photovoltaic (PV) generation, hybrid energy storage, and bidirectional Vehicle-to-Home (V2H) and Home-to-Vehicle (H2V) energy exchange. The predictive scheduling layer generates short-term operating plans, while the reinforcement learning layer refines control actions in real time at device, household, and supervisory levels. Simulation results demonstrate that the framework reduces computational burden by 31.4%, grid dependency by 21%, and peak demand by 14–19%, while increasing renewable energy utilisation by 18–25%. The approach also achieves approximately 96% convergence stability and a real-time response time below 0.42 s. These findings indicate that the proposed AI-based framework delivers efficient, robust, and practical real-time coordination for residential microgrids, even in the presence of renewable intermittency and uncertain EV availability.
The utilization of automated event detection methods enhances surveilling the operation of technical systems via sensor data. Particularly within the automotive domain, there exists an exigency for methodologies that discern events embedded within sensor data. In active vehicle safety systems, diverse driving scenarios are simulated, encompassing scrutiny of vehicle dynamics evaluated by external sensors. Therefore, automated detection methods expedite evaluation by eliminating the need for human-based assessments. This study investigates an event detection method to identify activation intervals of the Lane Departure Warning function within sensor data. The proposed model segments the sensor-driven data based on predefined window and step sizes, subsequently generating scalograms from temporal segments. Scalograms exhibiting significant overlap with the target event are labeled as target scalograms. The model comprises an autoencoder network integrated with a density estimation function. The autoencoder undergoes training to reconstruct target scalograms. The encoder module serves as a mechanism for extracting features from training data, which act as inputs for the density function. The algorithm's effectiveness is determined by evaluating its ability to distinguish the target scalograms from those generated from all temporal segments of unseen sensor data. The algorithm, evaluated on unseen datasets, achieves a precision rate of 94% in pinpointing the event's location, with a time interval offset averaging 0.5 s. This study demonstrates the potential of employing one-class classifiers—traditionally used for anomaly detection—for event detection within sensor data of active vehicle safety experiments. It further highlights that their performance can be enhanced by integrating multiple complementary models.
This study employs machine learning (ML) techniques to investigate the compressive strength, flexural strength, and thermal conductivity of aerogel-incorporated concrete (AIC). A database comprising 388 sets of experimental parameters and corresponding performance results was established and applied to train and evaluate six supervised ML algorithms. The results demonstrate that the Light Gradient Boosting Machine (LightGBM) achieved the best predictive performance for compressive strength and thermal conductivity, whereas the Extreme Gradient Boosting (XGBoost) performed best for flexural strength. All optimal models achieved coefficients of determination above 0.95 on the test set. To improve model transparency, the Shapley Additive Explanations analysis was introduced to interpret the optimal models. Parameter analysis was further conducted to reveal the relative contribution and directional effect of each input feature on the predicted properties. The results indicate that aerogel content and water-to-binder ratio are the dominant factors governing the three performance indicators, whereas the effects of mineral admixture content, curing age, and testing condition vary across different target properties. Based on these findings, design recommendations were proposed to enhance compressive strength, improve flexural strength, and reduce thermal conductivity, thereby providing a data-driven basis for the mix design of AIC. Furthermore, an interactive software platform integrating prediction and interpretability functions was developed to enable rapid performance assessment. Finally, a multi-objective optimization framework was developed to simultaneously maximize compressive strength and flexural strength and minimize thermal conductivity. The resulting trade-off solutions were systematically evaluated to identify representative mix designs corresponding to different performance priorities.
Gas leak monitoring is a knowledge-intensive engineering task critical for maintaining mechanical integrity and service safety. In practical gas leakage scenarios, signals are weak and noisy, making accurate monitoring highly challenging. Existing methods cannot sufficiently capture the multi-channel spatiotemporal features and multi-view coupling relationships of signals under real operating conditions, resulting in low monitoring accuracy. Therefore, a multi-view spatiotemporal perception network (MSPN) based on multi-channel acoustic emission signals is proposed for adaptive gas leak monitoring. Specifically, a multi-channel screening strategy is first designed to identify reliable signal channels. Subsequently, the symmetric dot pattern (SDP) method is used to transform signals into image representations, thereby enhancing feature learnability. Next, a spatially enhanced Mamba (SEMamba) and a temporally enhanced bidirectional gated recurrent unit (TEBiGRU) are designed to form a multi-view feature extraction framework. The spatial features and temporal evolution patterns of multi-channel signals are captured by this framework. Finally, an adaptive optimization mechanism is designed to learn the optimal fusion weights of multi-view features during model training, enabling more effective integration of complementary spatial and temporal information for gas leakage monitoring. Experimental results on two measured gas leakage datasets demonstrate that the proposed method achieves high monitoring accuracy, outperforming fourteen state-of-the-art diagnostic approaches. The framework not only provides accurate and reliable signal-based diagnostics but also enables mechanical system health monitoring in industrial pipeline operations.
