Rotating machinery fault detection is critically important for industrial systems. However, current methods relying on single-modal signals often yield incomplete feature representation and limited robustness. Denoising diffusion probabilistic models (DDPMs) offer promise for anomaly detection due to their strong generative capabilities and ability to train solely on healthy data. To overcome these limitations, this paper introduces MFE-DDPM, a novel rotating machinery anomaly detection method integrating multimodal feature fusion with an attention-enhanced DDPM. Our approach significantly enhances detection accuracy and robustness. First, vibration and acoustic signals are converted into time-frequency spectrograms via time-frequency analysis and a hyperparameter-free encoding method, while temperature data are transformed into grayscale images. These modalities are fused at the data level by mapping them to the R, G, and B channels of an RGB image, respectively. Subsequently, the attention-enhanced DDPM detects anomalies based on the reconstruction error of samples generated from the fused input. Validation on the Ottawa Multimodal Bearing Dataset shows that MFE-DDPM achieves a highest anomaly detection accuracy of 99.50%, with an average accuracy of 98.65% over ten repeated trials, outperforming both conventional single- and dual-modal methods and multimodal approaches without attention mechanisms.

Fault identification plays a pivotal role in condition-based maintenance of rotating machinery, with identification accuracy highly dependent on the quality of extracted features. Multiscale permutation entropy (PE) methods have emerged as promising feature extraction tools due to the fast computation of PE and informative scalability of multiscale procedures. However, PE is unresponsive to amplitude variation due to the binary orbit similarity state, and the multiscale procedure suffers from scale information loss or even scale absence, all of which decrease the identification accuracy. To address these issues, this article proposes a novel approach termed the scale-compensation community distance entropy (SCDE) method for fault identification. On one hand, the community distance-based orbit similarity value is put forward to diversify orbit similarity states, achieving a dual-characteristic perception of both frequency and amplitude changes. On the other hand, the scale-compensation procedure is proposed to enrich overall and detailed information on continuous scales. The efficiency and superiority of SCDE are rigorously demonstrated using simulation data and experimental datasets.

Gas storage facilities have long faced bottlenecks in the core control systems of high-power compressors, where operational safety, efficiency, and automation under complex conditions remain key challenges. Traditional single-sensor monitoring offers limited perception and low data utilization, falling short of ensuring reliable operation and intelligent decision-making. With AI engineering advancing toward multi modal and heterogeneous industrial applications, research has increasingly focused on multi-sensor fusion and intelligent diagnostics. Although multi-sensor networks provide complementary information and enhanced perception, issues like data heterogeneity, class imbalance, and cross-domain distribution differences continue to constrain diagnostic performance. To address these issues, a Dual-branch Heterogeneous Synergistic Network (DHSNet) is proposed to achieve cross-modal fusion of vibration signals and infrared thermal imaging signals. The framework incorporates a cross domain adaptation strategy to enhance domain-invariant feature learning and a dynamic focal loss function to adaptively adjust class weights based on real-time output indicators, mitigating the impact of sample imbalance. Experimental results demonstrate that the proposed method achieves an average diagnostic accuracy of 93.27% in cross-load transfer tasks, outperforming other methods used in imbalanced scenarios by over 4%. Besides, the proposed method maintains robust diagnostic performance exceeding 90% accuracy across most load transfer scenarios, even under moderately imbalanced data conditions. Feature contribution analysis further validates the effectiveness of multi modal synergy, and revealing the complementary mechanism of multi-source data. This study enhances Prognostics and Health Management (PHM) systems' engineering applicability and perception capabilities in multi-sensor industrial environments, providing reliable multi-modal diagnostics to advance intelligent maintenance.

Fault detection in mechatronic transmissions is particularly challenging due to the nonstationary nature of monitoring signals arising from complex operating conditions, coupled with the high-safety requirements that limit the availability of fault data. Sparse linear parameter varying autoregressive moving average (Spa LPV-ARMA) model is a powerful tool for dealing with nonstationary time series, and good fitting results can be achieved through the basis function expansion, where parameters of the model are associated with additional variables. However, current research on Spa LPV-ARMA model only considers single basis function, overlooking the potential complementarity of multiple basis functions. This article proposes a novel enhanced Spa LPV-ARMA model with ensemble basis for mechatronic transmission fault detection. The proposed model incorporates the concept of ensemble learning by combining models with different basis functions, and a stepwise approach is utilized to select the models to be combined. The rational choice of the combination scale allows the ensemble model to have fewer parameters with higher accuracy. Simulation and experimental studies in mechatronic transmission are conducted, verifying that the proposed ensemble basis Spa LPV-ARMA model exhibits higher modeling accuracy and fault detection performance.

Accurate prediction of machinery’sremaining useful life (RUL) is essential for preventing catastrophic breakdowns and supporting predictive maintenance. Although RUL prediction has been extensively studied, most literature develops on unimodal data, which providesa limited and often biased perspective. Multimodal monitoring, which collects multiple sensor data, enables a more comprehensive understanding of degradation processes. While promising, significant challenges are encountered in existing methods: 1) point yet deterministic predictions are predominantly produced which, while potentially erroneous, tend to exhibit overconfidence, thereby lacking the dynamic uncertainty informing; 2) the processing of heterogeneous data and the achievement of physically interpretable fusion remain challenging; and 3) anomalies in the operation process are not appropriately identified. To address these issues, a new multimodal uncertainty-aware RUL prediction framework is proposed, grounded in stochastic modeling. Fractional stochastic differential equation-controlled subnets process each modality independently, wherein layer-wise transformations are modeled as state evolution in stochastic dynamical systems, allowing modality-specific uncertainty to be quantified without requiring parameter priors. A Lagrange multiplier-based fusion module is subsequently employed to perform explicit uncertainty-based fusion, enabling an interpretable and synergistic integration. Validation on harmonic drive reducers for robots demonstrates the superiority of the proposed framework, achieving an average improvement of 26.6% in RMSE and a 16.6% reduction in MAPE compared to state-of-the-art benchmarks. Furthermore, the method significantly reduces prediction uncertainty variance by 21.3%, offering more reliable insights into system degradation.

Surface electromyography (sEMG) is extensively employed to characterize human physiological signals within Internet of Things (IoT) systems, serving as a critical component in Human-Computer Interaction (HCI) and various other applications. Although neural networks have been widely applied to intelligent recognition of sEMG signals, existing methods often face significant challenges in accuracy degradation and computationally intensive processing when handling multisubject signals. To address these issues, this paper proposes a robust surface electromyography (sEMG) intelligent recognition method based on OmniScale XceptionTime-Enhanced Deep Belief Network (OmniXceptionDBN). The method first processes raw signals using Singular Spectrum Analysis (SSA) and Fast Fourier Transform (FFT), then integrates XceptionTime, OmniScaleCNN, and Deep Belief Networks (DBN) to construct the OmniXceptionDBN algorithm for sEMG recognition. The designed integrated network for sEMG signals (i.e., the OmniXceptionDBN algorithm) achieves recognition accuracies of 97.2% for single-subject and 85.9% for multi-subject recognition scenarios without requiring dataset-specific optimizations. Our approach effectively resolves the accuracy degradation when processing across individuals and the high computational complexity inherent in traditional methods, providing an efficient solution for intelligent sEMG recognition.

Scene reconstruction has attracted widespread attention due to its extensive applications in intelligent devices. Recently, 3D Gaussian Splatting (3DGS) has gained recognition as a prominent technique owing to its impressive pixel-level rendering quality and high reconstruction speed. However, 3DGS scenes constructed with only photometric supervision are highly prone to RGB overfitting. This problem neglects the depth of objects and the surface normal in the reconstructed scene, which leads to the lack of geometric consistency. To address these issues, this study introduces a coupled optimization of depth and normal within the 3D Gaussian Splatting (CODN-GS) framework. Firstly, a normal–depth–normal transformation is applied to ensure accurate capture of geometric features in the reconstructed scenes. Secondly, a robust monocular depth supervision model generates depth maps that are refined via global and local adjustments, which serve as guidance for the model to accurately learn scene geometry. Thirdly, a normal supervision model is incorporated as a complement to depth supervision, jointly optimizing the overall scene geometry. Finally, comprehensive experiments on the Replica, MipNerf360, and ScanNet datasets demonstrate that CODN-GS reduces RMSE-D and RMSE-N by at least 9% and 13%. These results confirm that the proposed method outperforms state-of-the-art methods in both depth and normal accuracy.

Zero-sample diagnostic methods have gained recognition in addressing the scarcity of gearbox fault samples, thereby being regarded as a promising technique to guarantee gearbox safety. However, historical zero-sample approaches typically neglect the use of multimodal noncontact sensing data and rarely consider the interpretability of the diagnostic process. This oversight limits their application in industrial environments that require high reliability or operate under extreme conditions. Therefore, this article presents a composite neuro-fuzzy system-guided cross-modal zero-sample diagnostic framework, termed FCZD-IA, which employs infrared thermography and acoustic data to monitor gearbox conditions. Specifically, FCZD-IA uses a proposed composite neural system as a decision-maker in the diagnostic task, while integrating a deep backbone network to discriminatively learn high-level fault features from multimodal data. Moreover, a specific training strategy is designed to guide the learning process of the FCZD-IA to promote robust and interpretable zero-sample diagnostics. Comprehensive experimental results validate the effectiveness of the proposed framework and its superiority over other competitive methods.

Bearing health monitoring (BHM) is vital in preventing unforeseen machinery shutdowns caused by frequent bearing failures. Within the BHM process, constructing health indicators takes center stage, serving the dual purpose of detecting incipient faults and assessing the monotonous degradation trend for predicting residual useful life. In terms of detecting incipient faults, permutation entropy (PE) serves as a promising tool due to its simplicity and rapid computation. However, when it comes to assessing irreversible degradation, PE often exhibits notable fluctuations and nonmonotonicity even after signal denoising processes. This issue arises from PE's vulnerability to impulsive noise and its invariance to monotonic signal transformations. To tackle this challenge, the article introduces a novel approach termed condition-adaptive permutation entropy (CAPE) for BHM. CAPE begins with a condition-based signal processing method to mitigate the influence of impulsive noise, followed by an amplitude-aware algorithm to break PE's invariance to monotonic signal processing. Moreover, CAPE adaptively selects fault-relevant permutation patterns to enhance its monotonicity. The effectiveness, superiority, and applicability of CAPE are rigorously demonstrated using simulation data and two experimental datasets.

In industrial scenarios, mechanical faults are episodic and uncertain. Thus the monitoring data collected is usually extremely imbalanced, resulting in intelligent diagnostic models that suffer from majority-class dominance, minority-class overfitting, and poor generalization performance. Therefore, a knowledge distillation-guided cost-sensitive ensemble learning framework is proposed. It effectively combines ensemble learning and cost-sensitive learning to fully extract the multiscale features, effectively leverage the critical multi-depth features, and emphasize classifying the most confusing classes. Specifically, multiple-scale feature extraction and multi-order fusion are first employed to fully utilize the fault information. Afterward, the complementary diagnostic knowledge at different depths of the network is embedded into a novel ensemble learning process for better integration decisions. Then an improved knowledge distillation method achieves the mutual transfer and sublimation of excellent diagnostic knowledge while focusing on the most confusing fault classes to achieve the effective representation of various types of faults. Finally, a cost-sensitive strategy is applied to further increase attention to minority classes. The experimental results for various complex data imbalance scenarios, including extreme imbalance, step imbalance, continuous imbalance, interclass imbalance, and intra-class imbalance, all indicate that the proposed method can achieve state-of-the-art performance and provide a promising solution for the practical industrial application of intelligent diagnostic methods.

The Industrial Internet of Things (IIoT) greatly facilitates prognostics and health management of complex industrial systems, wherein the vast amount of real-time data from the IIoT improves intelligent predictive maintenance of industrial systems. When processing industrial IoT data across devices, traditional subdomain adaptation-based methods ignore the local similarities across domains. Also, if fault classes are used to define subdomains, these methods may not be applicable when the target domain is unlabeled or has limited labels. To address the above challenges, a Graph-embedded Subdomain Adaptation Network (GSAN)-based approach is proposed to predict the remaining useful life under different machines in IIoT. Specifically, a manifold subdomain representation is established by manifold learning and local manifold discrepancies between each pair of manifold subdomains with the highest similarity are minimized. To maintain a divisible margin for each manifold, a self-supervised intra-manifold regularization module is developed. An extensive evaluation of six transfer scenarios is performed, and the experimental results show that GSAN can achieve more significant outcomes. This can provide some guidance for future work on prognostics across devices and subdomains.

Convolutional neural network (CNN)-based fault detection approaches based on multisource signals have attracted increasing interest from the research community and industrial practices, thanks to the powerful feature representation capability of CNN and the rapid development of sensor technology. Various strategies have been applied in existing CNN-based diagnostic models to learn features from 1-D real-valued multivariate data. However, the distribution gap and the intrinsic correlations among multisource mechanical signals during the learning process have been rarely considered, which may lead to suboptimal fault identification results. To tackle this issue, this article proposes a cross-modal fusion convolutional neural network (CMFCNN) for mechanical fault diagnosis, which performs modality-specific and cross-modal feature representation on multisource data. Specifically, CMFCNN adopts two parallel modality-specific networks and a cross-modal knowledge-sharing network to fully explore independent and shared features from the multisource mechanical signals. To achieve effective feature propagation and fusion, a cross-modal fusion module is introduced to integrate cross-modal features and pass the fused information to the next layer. Moreover, to alleviate overfitting and achieve a better diagnostic performance of the framework, an online soft-label training algorithm is adopted in the CMFCNN training phase. Extensive experimental results on the cylindrical rolling bearing dataset and the planetary gearbox dataset validate that the proposed CMFCNN outperforms seven state-of-the-art methods significantly, especially under strong noise conditions.

The prognosis of bearings is vital for condition-based maintenance of rotating machinery. This article proposes a systematic prognostic scheme for rolling element bearings. The proposed scheme infers the degradation progression by developing a novel health indicator (HI). This novel HI, derived from the spectral correlation, Wasserstein distance, and linear rectification, can reflect the changes in the probability distribution of all cyclic power-spectra over time. In other words, any form of variation in modulation characteristics can be revealed through the proposed novel indicator, even for the weak information buried by the internal or external noise. Furthermore, the developed HI can eliminate random fluctuations that often impair the remaining useful life (RUL) prediction accuracy. Then, a 3${\boldsymbol{\sigma }}$criterion-based technique is introduced to divide health stages. After that, the gated recurrent unit network is employed to predict the RUL of the bearing system, integrated with the Bayesian optimization algorithm to tune the optimal hyperparameters adaptively. This renders the establishment of an intelligent prognosis model with high prediction accuracy and generalization ability. Finally, experimental validations are conducted using the run-to-failure datasets of bearings. The obtained results demonstrate that the proposed HI has better monotonicity, and the proposed prognostic scheme can predict the RUL with high accuracy.