Intelligent fault diagnosis has become an indispensable technique for ensuring machinery reliability. However, existing methods suffer significant performance decline in real-world scenarios where models are tested under unseen working conditions, while domain adaptation approaches are limited to their reliance on target domain samples. Moreover, most existing studies rely on single-modal sensing signals, overlooking the complementary nature of multi-modal information for improving model generalization. To address these limitations, this paper proposes a multi-modal cross-domain mixed fusion model with dual disentanglement for fault diagnosis. A dual disentanglement framework is developed to decouple modality-invariant and modality-specific features, as well as domain-invariant and domain-specific representations, enabling both comprehensive multi-modal representation learning and robust domain generalization. A cross-domain mixed fusion strategy is designed to randomly mix modality information across domains for modality and domain diversity augmentation. Furthermore, a triple-modal fusion mechanism is introduced to adaptively integrate multi-modal heterogeneous information. Extensive experiments are conducted on induction motor fault diagnosis under both unseen constant and time-varying working conditions. The results demonstrate that the proposed method consistently outperforms advanced methods and comprehensive ablation studies further verify the effectiveness of each proposed component and multi-modal fusion.
As modern machinery moves toward higher levels of automation and complexity, the demand for equipment reliability has become increasingly stringent. Unexpected failures may result in substantial economic losses, unplanned downtime, and even catastrophic accidents [1]. To mitigate these risks, industries are deploying large-scale sensor networks to continuously monitor diverse signals such as vibration, current, and acoustic [2]. By *Corresponding authors. Email addresses: huang.yixiang@sjtu.edu.cn (Y. Huang); qinchengjin@sjtu.edu.cn (C. Qin) leveraging these monitoring data, fault diagnosis can be performed to detect anomalies at an early stage and enable timely maintenance interventions [3].
Recently, intelligent data-driven fault diagnosis methods based on deep learning have attracted substantial attention due to their strong ability to automatically learn discriminative representations from raw sensor measurements. Compared with traditional physics-based approaches and conventional machine-learning methods that rely heavily on hand-crafted feature engineering, deep learning methods eliminate the need for complex prior expert knowledge and manual feature extraction, thereby demonstrating highly promising diagnostic performance [4,5]. For example, vibration signals of rotating machinery can be directly fed into deep learning models, allowing them to automatically learn relevant features for fault diagnosis. Borghesani et al. [6] leveraged 1D convolutional neural network (CNN) to accomplish bearing fault diagnosis and attempted to explain the vibration feature extraction process. Liu et al. [7] designed a residual network with multiscale kernels for motor fault diagnosis using vibration signals. However, raw one-dimensional signals often exhibit complex temporal characteristics and non-stationary behaviors, which may limit the ability of conventional 1D models to fully capture informative patterns. To better represent the underlying time-frequency structures of vibration signals, many studies transform 1D vibration signals into 2D time-frequency representations with short-time fourier transform (STFT) [8] or wavelet transform [9], and then employ 2D networks to extract more expressive features from these transformed images. Nevertheless, vibration signals may exhibit limited sensitivity to certain electrical faults in electromechanical machines such as motors [10]. Consequently, motor current signals have also been widely adopted for fault diagnosis [11]. For example, Jimenez-Guarneros et al. [12] designed a lightweight 1D CNN to diagnose mechanical and electrical faults of induction motor with current signals. Furthermore, acoustic signals also contain fault-related information and have therefore been explored for machinery fault diagnosis. Zhang et al. [13] employed acoustic signals together with a graph convolutional network (GCN) to diagnose bearing faults, while Xiao et al. [14] utilized a denoising autoencoder to achieve motor fault diagnosis based on acoustic measurements. These studies collectively demonstrate the effectiveness of deep learning-based diagnostic methods using acoustic signals.
Despite the remarkable progress of intelligent fault diagnosis methods, their generalization ability remains a major obstacle to practical applications as machines typically operate under variable working conditions. Domain shifts induced by changes in operating speed and load can significantly degrade the performance of models trained on data from source conditions. To address this issue, domain adaptation (DA) techniques have been extensively investigated to improve the robustness of diagnostic models under distribution shifts. By aligning feature distributions from source domains to target domains, these methods aim to mitigate the adverse effects of condition variability and enable more reliable cross-condition fault diagnosis [15]. Consequently, DA based on discrepancy minimization [16] and domain adversarial learning [17] and subdomain adaptation (SDA) [18] methods have demonstrated effectiveness in machinery fault diagnosis across various working conditions.
Nevertheless, a major limitation hindering the practical applications of DA methods is their reliance on target domain data for distribution alignment. In real industrial scenarios, this requirement is often difficult to satisfy as machines frequently operate under newly emerging or previously unseen working conditions for which no samples have been collected in advance. To overcome this constraint, domain generalization (DG) has emerged as an attractive alternative, aiming to learn models that can generalize to unseen target conditions without accessing any target-domain samples during training [19]. Instead of explicitly aligning source and target distributions, DG seeks to extract domain-invariant and discriminative features from multiple source domains, thereby enhancing the model’s robustness to distribution shifts and demonstrating promising
This content is AI-processed based on open access ArXiv data.