Early fault detection of degraded gears at different speeds is both essential and challenging. Adaptive blind deconvolution methods have shown considerable promise for extracting fault characteristics from complex vibration signals. Their performance depends on accurate cyclic frequency estimation and optimal filter length selection. However, this estimation often fails due to gear meshing shock interference and early weak fault characteristics. Additionally, determining the filter length relies on additional metrics with inefficient search strategies, thereby limiting the overall reliability and efficiency. To address these issues, an adaptive blind deconvolution via convolutional neural network (ABDCNN) is proposed. First, we employ an envelope harmonic product spectrum guided by gear frequency-domain features to reduce interference from noise and meshing shocks, enabling precise estimation of the target cyclic frequency. Then, an attention mechanism is integrated into the convolutional neural network to jointly optimize filter coefficients and length estimation, thereby improving computational efficiency. Simulations and gear contact fatigue experiments demonstrate that ABDCNN enables more efficient detection of early faults across different speeds while maintaining strong interpretability.

Abstract Sample imbalance across operating conditions severely constrains data-driven bearing fault detection performance. This challenge is more acute due to the complete absence of fault samples under target operating conditions. To address this limitation, this paper proposes a framework integrating Variational Mode Decomposition (VMD), nonlinear dynamic modeling, and physics guidance. First, a two-degree-of-freedom nonlinear system is established to simulate physically realistic dynamic behaviors. Second, VMD decouples fault feature components from source-domain fault signals. These components are superimposed with target-domain normal signals and dynamic responses to form hybrid samples. Then, a diffusion-based Wasserstein Generative Adversarial Network (WGAN) architecture is adopted. Simultaneously, auxiliary classifiers for operating conditions and fault types are incorporated as physics-regularized losses to enforce physical consistency. The fault feature decoupling component based on VMD and the generation framework jointly mitigate the domain gap between dynamically generated response signals and measured signals. Experimental results indicate generated samples from the proposed method achieve superior scores across multiple evaluation metrics. And classification networks trained on generated samples achieve diagnostic accuracy exceeding 90% in multiple cross-condition tasks, which confirms the superiority of the developed model.

Abstract Accurate online remaining useful life (RUL) estimation for lithium-ion batteries depends on reliable capacity-degradation modelling from measurable signals. This paper proposes an improved whale optimisation algorithm-optimised physics-informed neural network (IWOA-PINN) to address noisy incremental-capacity curves, degradation-inconsistent estimates, and hyperparameter sensitivity. A second-order RC equivalent-circuit model reconstructs the Q – V relationship, from which peak height, peak voltage, time at peak voltage, and time to reach the maximum surface temperature (MATT) are extracted. A lightweight capacity regressor is regularised by a training-only PH- Q prior and a soft adjacent-cycle monotonicity penalty, while IWOA selects the hidden-layer size, training epochs, learning rate, and two loss weights. RUL is mapped from the estimated capacity trajectory and the end-of-life threshold. Experiments on a public Xi’an Jiaotong University dataset and a laboratory-built platform show capacity-estimation RMSE around 0.01 Ah and mean RUL errors of 2.0–3.0 cycles under the tested fixed and random-walk partial-discharge protocols. The reported RUL errors reflect rolling end-of-life localisation as cycle-wise measurements become available, rather than single-origin open-loop forecasting. Under a unified comparison protocol, IWOA-PINN achieves the lowest mean RMSE values on B12 and B22, at 0.0084 and 0.0062 Ah, respectively. Its mean RUL errors are 3.0 and 1.1 cycles, with the former comparable to the strongest baseline and the latter the lowest among the compared models.

Data-driven fault diagnosis is an important approach for gearbox health management. In practical industrial scenarios, however, fault samples are usually scarce while normal samples are relatively abundant, resulting in severe class imbalance and degraded diagnostic performance. To alleviate this problem, a gated attention-enhanced improved conditional diffusion model is proposed for generating high-quality 1D vibration fault signals under small-sample conditions. Fault-type labels are embedded into the diffusion process to enable class-conditional generation of fault samples. Meanwhile, a gated attention-enhanced 1D U-Net denoising network is constructed to improve the extraction of long-range dependencies and salient fault features from vibration signals. In addition, a joint loss function incorporating spectral skewness is introduced to enhance time-domain reconstruction accuracy and frequency-domain fidelity simultaneously. Experiments on two planetary gearbox datasets show that the proposed method achieves better performance than representative generative models in terms of distribution similarity and waveform consistency. The generated samples are further used to balance the training set, leading to improved downstream diagnostic performance across ResNet-, ShuffleNetV2-, and MobileNetV3-based classifiers.

Intelligent fault diagnosis methods based on deep learning have made significant progress in the field of planetary gearbox fault diagnosis. However, in practical industrial scenarios, existing diagnostic models face challenges of excessive complexity that hinder their deployment on resource-constrained hardware platforms for real-time equipment monitoring. Although lightweight models typically employ fewer parameters and simplified architectures, they demonstrate insufficient diagnostic accuracy when confronted with intense noise in engineering practice. To address these issues, this paper proposes EPA-RepViT, a lightweight model with enhanced noise robustness. Firstly, an efficient parallel attention module is designed, combining channel attention and spatial attention mechanisms to enhance the model’s ability to perceive important channels and regions. Secondly, a concise stem module is developed by decoupling the dual-branch downsampling from channel expansion, thereby improving the stem’s representational capacity while increasing the model’s inference speed. Finally, the RepViT model is adapted into a one-dimensional form and pruned, and integrated with the aforementioned modules to achieve end-to-end intelligent fault diagnosis for planetary gearboxes. Experimental results on two planetary gearbox fault datasets demonstrate that the proposed model achieves superior diagnostic accuracy and operational efficiency compared to representative lightweight models, along with enhanced noise robustness.

Traditional fault diagnosis methods based on supervised learning often suffer performance degradation when data distributions shift, limiting their effectiveness in real-world industrial applications. In contrast, unsupervised transfer learning can leverage knowledge from a labeled source domain to address similar tasks in an unlabeled target domain. To tackle performance drops caused by distribution shifts under varying gear loads, this paper proposes a domain-adversarial fault diagnosis method integrating multi-band filtering and multi-scale residual learning (MBF-MSRDA). Specifically, multi-band filtering is used to enhance fault features across frequency bands, while a residual network extracts multi-scale features. To reduce domain discrepancies, multi-layer maximum mean discrepancy (MMD) is applied for effective feature alignment. Experimental results validate that the proposed method achieves high fault classification accuracy under diverse load transfer conditions, demonstrating strong generalization and domain adaptability.

Abstract Accurate remaining useful life (RUL) prediction of rotating machinery is critical for industrial reliability and maintenance safety. In practical prognostic scenarios, data scarcity and operating-condition variations often challenge model generalization. Recent large language model (LLM)-based time-series methods offer a promising solution by exploiting transferable sequence representations. However, most existing LLM-based RUL prediction methods still rely mainly on implicit parametric knowledge, input reprogramming, or generic adapter structures, which may cause unreliable extrapolation and insufficient use of traceable historical degradation experience. To address these limitations, this paper proposes RAF-LLM, a retrieval-augmented LLM framework for rotating machinery RUL prediction. RAF-LLM constructs a degradation-specific external knowledge base from historical run-to-failure trajectories and introduces a learnable dual-branch retriever with prediction-oriented alignment. The retrieved historical degradation trajectories serve as non-parametric memory to support RUL regression, enabling predictions to be grounded in explicit degradation precedents rather than relying solely on pretrained LLM parameters. Meanwhile, continuous embedding projection and gated residual fusion are employed as supporting modules to align degradation-sensitive physical features with the LLM semantic space and integrate retrieved historical knowledge with current observations. A retrieval-alignment objective further guides the retriever to select candidates that are not only similar in feature space but also useful for RUL estimation. Repeated experiments on two public run-to-failure datasets show that RAF-LLM achieves competitive average prediction performance against representative baselines. The results provide empirical support for retrieval-assisted RUL prediction and case-based traceability under the evaluated settings.

Gearboxes are crucial for reliable power transmission in rotating machinery. Single-modal diagnostic methods limit feature expression, while multimodal methods, while improving accuracy, are still limited by insufficient cross-modal feature mining and inadequate representation of multi-scale fault features. To address this issue, this paper proposes a multimodal time-frequency feature fusion network (MFFNet). This network efficiently extracts fault signatures by collaboratively analyzing short-time Fourier transform (STFT) time-frequency graphs and Markov transition field (MTF) state transition features. First, feature mode decomposition (FMD) based on finite impulse response (FIR) filters is used to adaptively extract fault-sensitive modes from the raw vibration signal. Then, a dual-channel time-frequency representation is constructed using the STFT and MTF to encode the time-frequency and state transition features of the time series. To enhance feature integration, a feature fusion module (FFM) combining a cross-modal self-attention mechanism and an adaptive weighted fusion strategy was designed, as well as a multi-scale fusion module (MFM) for cross-resolution feature alignment. This enhanced the model’s ability to model fault information in complex non-stationary vibration signals and ensured the integrity of fault feature extraction. Finally, an adaptive decision-level fusion strategy was employed to integrate multi-branch diagnosis results. Experimental results demonstrate that MFFNet outperforms the comparison models on the planetary gearbox vibration dataset, achieving a fault diagnosis accuracy of 99.50

Heterogeneous gear fault vibration signals are often influenced by the coupling of multi-level modulation sources, which severely hinders the separation of informative features and obstructs accurate tracking of their intrinsic temporal evolution. Moreover, current research in fault diagnosis lacks effective co-modeling strategies that can jointly characterize the multi-source coupling mechanisms and their dynamic progression, making it difficult to decouple fault-relevant patterns from interfering components in such modulated signals.To address this issue, we propose a physics-guided spatiotemporal decoupling framework, which integrates spatiotemporal modeling with the underlying physical modulation mechanisms. First, inspired by the kinematic characteristics of gear systems, a heuristic selective decision solver is designed to extract modulation patterns dominated by rotational periodicity, thereby achieving effective isolation of multi-source interference. Next, the extracted patterns are transformed into physically meaningful enhanced representations via a spatiotemporal encoder, which simultaneously preserves the local structural details and temporal evolution of fault features. Furthermore, a spatiotemporal coupling capture network is developed, incorporating a physically salient attention mechanism to adaptively emphasize critical fault-related components, significantly improving the discriminative capability of the learned features. Finally, a neural operator-based classifier is employed to accomplish robust fault recognition. Experimental results demonstrate that the proposed method achieves high diagnostic accuracy and robustness across a variety of complex operating conditions, enabling efficient and reliable fault diagnosis for gear transmission systems.

Abstract As a critical component of rotating machinery, bearings directly affect the operational safety of the system. However, in real industrial scenarios, the imbalanced distribution of fault samples impairs the accuracy of diagnostic systems. Existing fault diagnosis methods still suffer from limited representation capability in fusing high-dimensional heterogeneous features with fault information, while language model-based diagnostic approaches face challenges such as insufficient interpretability. To address these issues, this paper proposes a frequency-aware multi-view feature fusion fault diagnosis framework, termed frequency-aware and multi-view fusion bearing fault diagnosis framework (FMFDF). First, raw vibration signals are converted into statistical semantic prompts and time–frequency spectrograms to construct multi-view inputs consisting of text and images. Then, a fault frequency feature extraction network (FF-FEN) is designed to embed FF information into the feature extraction process, where frequency-constrained convolutions enhance the feature representation of FF bands. Subsequently, a multi-head attention mechanism is employed to align and fuse the textual semantic features with the frequency-aware time–frequency features, followed by fault classification using a LoRA-fine-tuned BERT and classification network. The proposed method is validated on the case western reserve university and Jiangnan university bearing datasets. Experimental results demonstrate that FMFDF achieves favourable diagnostic stability under imbalanced sample conditions and offers a certain degree of interpretability.

Root cracks induced by gear tooth bending fatigue represent a typical high-cycle fatigue failure mode. Previous studies have rarely addressed the effects of manufacturing and assembly errors on root crack propagation behaviour and fatigue life. Therefore, this study first employs single-tooth bending tests to simulate high-cycle loading conditions, utilizing synchronized high-speed imaging and vibration monitoring to characterize crack initiation and growth processes. Through computational analysis of maximum bending stresses and experimental investigations, a root crack simulation model under meshing conditions was established. Model validity was verified via experimental comparisons, enabling comprehensive analysis of bending fatigue failure mechanisms. Finally, the validated model investigates the influence of surface roughness and centre distance variations on root crack propagation and fatigue failure. The findings reveal that load variations do not alter crack paths but reduce fatigue life with increasing load. Elevated surface roughness promotes deeper crack penetration into the gear substrate, with optimized roughness effectively enhancing bending fatigue resistance. Centre distance does not influence crack propagation path, but appropriately improving assembly precision increases bending fatigue life.

Previous deep learning-based methods for gear fault diagnosis and assessment, due to their black-box nature, result in the extracted signal features and diagnostic results lacking practical physical significance and interpretability. This study combines deep learning methods with the dynamic characteristics of gears and proposes a gear fault evaluation method named convolutional-neural-network-based inverse physics-informed neural network (CNN-IPINN). First, a neural network loss function is designed based on the gear dynamics equation to construct an IPINN for solving inverse problems in gear dynamics. This design enables the neural network to extract the actual time-varying meshing stiffness (TVMS) from gear vibration signals, which serves as the core basis for gear fault diagnosis and assessment, thereby enhancing the interpretability of the network. Considering the issues of low accuracy, model complexity, and slow operation speed in traditional physics-informed neural networks (PINNs), as well as the spatial correlation of gear vibration signals, this study introduces CNN as the backbone network of PINN to construct CNN-IPINN for extracting the TVMS of gears. Finally, based on the actual gear experimental dataset, diagnoses and evaluations are performed on gear faults involving varying degrees of wear, pitting, and crack damage. This approach achieved highly accurate and interpretable gear fault diagnosis, thereby demonstrating broad prospects in engineering applications.

Pedestrian detection based on deep networks has become a research hotspot in the field of computer vision. With the rapid development of the Internet of Things (IoT) and autonomous driving technology, the deployment of pedestrian detection models on mobile devices places higher demands on the accuracy and real-time performance of detection. In addition, fully integrating multimodal information can further improve the robustness of the model. To this end, this article proposes a novel multimodal fusion YOLOv5 network for pedestrian detection. Specifically, to improve the performance of multi-scale pedestrian detection, we enhance contextual awareness abilities by embedding the multi-head self-attention (MSA) mechanism and graph convolution operations in the existing YOLOv5 framework. In addition, we can fully explore the real-time advantages of the YOLOv5 framework in pedestrian detection tasks. To improve multimodal information fusion, we introduce the joint cross-attention fusion mechanism to enhance knowledge interaction between different modalities. To validate the effectiveness of the proposed model, we conduct a large number of experiments on two multimodal pedestrian detection datasets. All the results confirm that our proposed model obtains the highest performance in terms of multi-scale pedestrian detection. Moreover, compared to other multimodal deep models, our proposed model still shows superior performance.

To address the challenges of difficult feature extraction from gearbox fault signals and low accuracy in fault diagnosis, this paper presents a novel diagnostic model that integrates Variational Mode Decomposition (VMD) optimized by the Chaotic Hippopotamus Optimization algorithm (CHO) with a Bidirectional Gated Recurrent Unit (BiGRU) neural network. Firstly, CHO is used to optimize the number of modal decomposition and the penalty factor of VMD, using a composite index that combines the ratio of permutation entropy (PE) to mutual information entropy (MIE) as the fitness function. Then, the optimized VMD method decomposes the signal, and nine statistical features are extracted from the optimal Intrinsic Mode Function (IMF)components to form feature vectors. Finally, the feature vectors are input into the BiGRU for fault recognition. Experimental results demonstrate that the proposed model integrating CHO-VMD-BiGRU, can accurately extract fault features and outperform other algorithms, showing promising practical application value.

Gear wear is an inevitable consequence of friction and load during operation. However, nonlinear and non-stationary nature of the vibration signals under variable speed conditions and their complex interaction with gear wear make it extremely challenging to extract related features from them. Since the gear wear area visually indicates the severity under varying speed conditions, we propose a vibration-based method for monitoring the area to quantify wear severity. First, non-stationary vibration signals are resampled angular domain using the multisynchronous extraction transform to mitigate the impact rotational speed variations. Then, vibration components related to gear wear are extracted using the adaptive local synchronization fitting technique. Subsequently, a novel indicator, termed the cumulative comprehensive energy ratio, is constructed by integrating the multidimensional characteristics to monitor the wear area and visually assess wear severity. Finally, the relationships between the constructed indicator and the percentage of gear area under six different speed conditions are investigated on the CL-100 gear contact test rig. The results demonstrate that the proposed method effectively quantifies wear under variable speed conditions.

The vibration of helicopter spiral bevel gears has a complex signal pattern that is nonlinear, non-stationary, and with strong background noise, which makes it difficult to meet the requirements of intelligent fault diagnosis by simply using a physical model, signal processing, or machine learning. A framework that integrates vibration mechanism models and deep learning mechanisms based on higher-order spectrum is proposed to accomplish helicopter spiral bevel gear fault diagnosis using the optimal symplectic geometry mode bi-spectrum mutual information maximization encoder (OSBMME). Firstly, recursive symplectic geometry mode decomposition is employed to extract the fault-sensitive optimal symplectic geometry mode (OSGM) from complex vibration signals by combining it with the mechanism model to eliminate the effects of interfering signals and out-of-band noise. After that, the OSGM is transformed into a bispectral contour spectrogram with higher order statistics, maintaining the signal's phase information and enhancing noise immunity. Finally, a deep mutual information maximization encoder network based on a combination of unsupervised and supervised methods extracts the fault information from the spectrogram and completes the fault diagnosis of helicopter spiral bevel gears utilizing a comparative learning mechanism. The performance verification is carried out on the helicopter main reducer vibration experimental platform through normal as well as injected four fault modes, and the results of the ablation and comparison experiments show that OSBMME can effectively extract high-quality expression of sensitive information from nonlinear, non-stationary, and strongly noisy vibration signals, realize higher accuracy of helicopter spiral bevel gear fault diagnosis, and have better robustness under fluctuating working conditions.

Lithium-ion batteries (LIB) are widely used in aviation applications. However, inconsistencies in cell performance within battery packs (BPs) can significantly affect energy efficiency and reduce lifespan, potentially compromising flight safety. To address this critical issue, this paper proposes a novel energy balancing method that integrates inductors and a DC/DC converter to optimize energy distribution within aviation LIB, thereby enhancing pack consistency. In this approach, inductors are used to implement "peak shaving" by transferring excess energy from high-voltage cells, while the DC/DC converter achieves "valley filling" by providing energy to low-voltage cells, thus stabilizing the overall cell voltages. The balancing operation is driven by the state of charge (SOC) of the cells as the primary control parameter, enabling precise and efficient energy transfer among cells in a 28 V aviation battery pack (ABP). Comprehensive simulation and experimental validation were conducted to assess the effectiveness of the proposed balancing method. Simulation and experimental results demonstrate that the proposed method reduces the time required to balance 1 % SOC by approximately 20 s, representing a 10 % improvement in balancing speed. The proposed approach achieves faster balancing speed and improved balancing performance compared to conventional balancing methods. This work offers a practical solution for enhancing the reliability and lifespan of aviation LIB, contributing to safer and more efficient aviation power systems.

To effectively extract early fault features in rotating machinery signals contaminated by intense noise, this article proposes a novel fault detection method based on feature mode decomposition (FMD) and a maximum correlated kurtosis combined L-Kurtosis deconvolution (MCK-LKD) algorithm. The method innovatively introduces a combined kurtosis indicator named CK-LK and establishes the MCK-LKD algorithm by integrating the statistical characteristics of correlated kurtosis (CK) and L-Kurtosis (LK), significantly improving the detection capability for fault impact components. Furthermore, a two-stage processing framework FMD-MCK-LKD is constructed, effectively combining the noise reduction advantages of FMD with the fault enhancement capabilities of the MCK-LKD algorithm. The proposed method is tested on both simulated and experimental signals, with comparative analysis conducted against other mainstream deconvolution algorithms. The results demonstrate that the signal decomposition and reconstruction method based on FMD, combined with the MCK-LKD algorithm, exhibits significant advantages and superior performance in early fault detection of rotating machinery.

Deep learning-based models have achieved promising performance for video behavior anomaly detection, but these models require high computational complexity for processing video data. Moreover, the performance is severely restricted by different challenges, i.e., lighting conditions, background noise and occlusion. The human skeleton has a compact spatial structure and rich semantic information, which can be more robust to video data defection. Although the human skeleton has good real-time performance, the recognition accuracy still needs further improvement. To this end, we propose a novel discrete variational feature transfer learning (DVFTL) framework in which the spatiotemporal graph-embedded Transformer module is designed to construct feature extraction backbones. Specifically, we extract human skeleton information from video frames, and combine graph convolution with the Transformer encoder to explore the local and global dependencies of joint points in the human skeleton. To analyze abnormal behavior uncertainty, we constructed the extraction network based on the discrete variational reconstruction mechanism. Moreover, to further improve the skeleton detection performance, we introduce the distillation transfer learning mechanism from the video variational network to the skeleton variational network. we conducted extensive comparative experiments on two publicly available datasets. The experimental results show that the skeleton-based variational network achieves high performance. Moreover, the introduction of the transfer learning mechanism can further improve the performance of our proposed model.

Download This Paper Open PDF in Browser Add Paper to My Library Share: Permalink Using these links will ensure access to this page indefinitely Copy URL Copy DOI