A decision-level multimodal fusion deep learning strategy is proposed for the effective fault detection of rolling bearings based on long-term fault signals collected from multiple sensors. First, key features are extracted from the multimodal signal set using singular spectrum analysis (SSA), and these features are transformed into a composite dataset that combines short-time Fourier transform (STFT) images and time series data. Based on this, a recursive gated convolutional neural network (RGCNN) is designed to process the STFT image data, while a 1D convolutional neural network (1DCNN) is specifically optimized for training with time series data. Furthermore, decision-level multimodal feature fusion is achieved by applying a weighted average method to integrate the features from different deep learning models, aiming to obtain more comprehensive fault prediction results. The proposed method, multimodal fusion fault detection (MFFD), is validated on the Paderborn and Ottawa rolling bearing datasets, which include various typical faults. Experimental results demonstrate the effectiveness of the proposed approach. Compared to traditional single-modality deep learning models, the proposed method shows significant improvements in fault diagnosis accuracy and generalization capability.

Accurate engine fault classification with limited labeled data is critical for the safety and reliability of rotating machinery. This task is challenging because operating regimes are time-varying, and key variables must satisfy physical constraints, under which traditional feature classifier pipelines degrade and deep networks tend to overfit. We propose PhysGMM-MoE, a physics-aware Gaussian Mixture Model (GMM)-Mixture-of-Experts (MoE) framework for small-sample engine fault classification. At the data level, PhysGMM-MoE fits class-conditional, regime-aware GMMs and performs physically constrained, distance-based quality control to selectively augment minority classes while preserving engine operating semantics. At the model level, a heterogeneous pool of lightweight statistical experts and a lightweight Transformer-based deep expert (ECFT-Transformer) capture complementary neighborhood cues and high order multi-sensor correlations, and an L2-regularized logistic regression meta-learner fuses expert outputs via stacking. We evaluate fault classification on the 3500-DEFault diesel-engine dataset using the adopted eight-class cylinder-fault labeling (H, F1–F7) built from in-cylinder pressure statistics and torsional-vibration harmonics; although severity levels exist in the dataset, this study focuses on classification rather than severity estimation. With 40 training samples per class, PhysGMM-MoE achieves a mean accuracy of 0.9875, exceeding SMOTE+XGBoost by 0.0086, and attains the best macro precision/recall/F1 of 0.9878/0.9826/0.9889, demonstrating strong performance under the adopted small-sample setting.