Aug 2026· Structural Health Monitoring· 0 citations· 20 references
TL;DR
A semantic knowledge transfer framework to diagnose compound faults using only single-fault data for training, providing a robust solution for mechanical fault diagnosis under zero-sample scenarios is proposed.
Abstract
Compound fault diagnosis of wind turbine gearboxes (WTGs) has received extensive attention. Existing deep learning-based methods typically require sufficient compound fault samples for model training. However, collecting such samples is extremely difficult and often impractical in real-world industrial scenarios. Inspired by zero-shot learning, this paper proposes a semantic knowledge transfer framework to diagnose compound faults using only single-fault data for training. Within this framework, a semantic knowledge library is first constructed to encode human expert intelligence into high-fidelity knowledge vectors, establishing a shared semantic space for all fault classes. To ensure signal representations align with these expert semantics, a time-frequency informative perceptron is introduced to capture comprehensive fault signatures by simultaneously capturing discriminative features from both time and frequency domains. Finally, imbalance-robust knowledge learners are designed to bridge the gap between physical features and knowledge labels while mitigating the inherent class imbalance effects. The proposed framework is validated on a self-built WTG compound fault test platform. Experimental results showcase its exceptional effectiveness and superiority in recognizing unseen compound faults, providing a robust solution for mechanical fault diagnosis under zero-sample scenarios.
Inspired by few-shot learning, a support–query task construction strategy is introduced, reformulating classification as conditional task reasoning and a global coverage and local balance task sampling strategy is designed to enhance task diversity and mitigate sample imbalance.
Cross-condition fault diagnosis remains a fundamental challenge in intelligent operation and maintenance of hydropower equipment, where deep learning methods suffer 20%−30% accuracy degradation under operating condition drifts. Large language models (LLMs), through massive multi-domain pre-training, offer cross-domain generalization that may help overcome this limitation. However, applying LLMs to industrial vibration signals faces several difficulties: a modality gap between continuous physical time series and discrete semantic sequences, multi-fault class aliasing in the semantic space, and limited sensitivity to transient fault pulses. This paper proposes FD-SE-LLM, a Semantic-Enhanced Large Language Model Framework for Fault Diagnosis targeting hydropower carbon brush bearings. The framework includes Time-domain Semantic Cross-Correlation (TSCC) to project periodic and transient features into the LLM semantic space, a Fault-specific Multi-class Conditional Variational Autoencoder (FM-CVAE) to disentangle fault class-specific representations from environmental noise, and an embedded Bearing Time-Adapter to enhance transient pulse sensitivity. Experiments on hydropower, CWRU, and JNU datasets show that FD-SE-LLM achieves 91.5% average cross-condition accuracy, outperforming deep learning baselines by 6.8−15.1 pp and existing LLM-based approaches by 3.9−6.3 pp.
Unknown authors· Journal of Dynamics Monitori...· 0 citations
The diagnosis of compound faults in bearings is a challenging issue in real industrial scenarios, as these faults have a significant impact on production and safety. Moreover, the data related to these compound faults are difficult to collect and label, making data-driven fault diagnosis methods inapplicable. To address the scarcity of compound-fault data and the absence of labeled samples in bearing systems, this article proposes a zero-shot diagnosis method based on semantic embedding. By training solely on single-fault data, this approach effectively overcomes the shortage of compound fault samples. First, a binary semantic representation is constructed for each single fault using Top-
k
masking, and a neural semantic composer (NSC) is pretrained to generate realistic compound-fault prototypes. Second, a multisensor robust convolutional neural network is designed for feature extraction. Finally, a synthetic compound-fault training strategy is introduced to enhance model generalization. By aligning the compound semantic targets generated from feature mixing and NSC, this strategy enables the model to learn the patterns in unseen compound samples. The necessity of each module is demonstrated through ablation experiments. The effectiveness and generalization ability of the method are verified using the HDU and HUST datasets. Under various operating conditions, the proposed method demonstrates optimal or nearly optimal accuracy. The code will be made publicly available on
https://github.com/su-yibei/zero-shot-fault-diagnosis
.
Sitong Jiao, Shijie Ning, Xiaomin Zhu et al.· Structural Health Monitoring· 0 citations
This paper proposes a knowledge graph-guided condition-aware meta-learning (KG-CAML) method for few-shot fault diagnosis under sensor data variations induced by fluctuating raw materials and changing operational loads. Within a meta-learning framework, KG-CAML integrates data-driven sensor correlations with expert process knowledge by constructing a KG that captures dependency structures among equipment and measurement variables. A relational graph convolutional network is employed as the encoder to extract consistent feature representations. On this basis, the correlation alignment metric is introduced to quantify the distributional similarity and alignment loss between training tasks and test conditions. Furthermore, to mitigate knowledge forgetting during meta-learning, a task memory updater is designed to dynamically update and retain task representations across stages, thereby continuously accumulating and transferring critical diagnostic knowledge. Experimental results on Tennessee Eastman processes and Wastewater Treatment Plant demonstrate that the diagnostic performance of KG-CAML outperforms state-of-the-art methods under multiple working conditions and few-shot scenarios significantly.
Ke Wu, Yuan Xu, Yi Luo et al.· Measurement science and tech...· 0 citations
Long-term operation of industrial equipment results in the continuous accumulation of monitoring data for emerging fault types. This phenomenon leads to catastrophic forgetting in intelligent diagnostic models and degrades overall performance. A semantic-guided orthogonal subspace network (SOSN) is proposed for class-incremental continual intelligent fault diagnosis. The architecture utilizes lightweight adapters to construct task-specific orthogonal subspaces within a frozen pre-trained backbone. These subspaces ensure that gradient updates for novel tasks remain theoretically isolated from historical parameters. To reduce the dependency on historical raw data, a semantic-guided prototype synthesis mechanism is developed. This mechanism reconstructs historical class centers by leveraging feature similarity across different task subspaces. Extensive evaluations are performed on rolling bearing datasets and a joint heterogeneous dataset comprising wheelset bearings and elevator door systems. Experimental results demonstrate that SOSN significantly outperforms mainstream class-incremental learning approaches in diagnostic accuracy. The framework effectively balances representational plasticity and diagnostic stability across diverse mechanical systems.
He Ren, Yang Liu, Gongning Li et al.· Journal of Vibration and Con...· 0 citations
Fault diagnosis for complex industrial equipment plays a crucial role in safeguarding production safety and advancing the capabilities of intelligent operation and maintenance. Current deep learning approaches have demonstrated promising accuracy in fault classification tasks; however, their signal representations alone cannot provide a transparent interface for embedded large language models. To tackle the aforementioned challenges, we propose PGA-LLM, a novel fault diagnosis framework for industrial equipment that leverages large language models via probability-guided alignment. First, a variational autoencoder (VAE)-based signal encoder embedded with reconstruction constraints is established. Joint reconstruction and classification objectives balance discriminative representation learning and signal reconstruction. Second, the probability-guided alignment (PGA) module combines fault-class probability guidance with a residual feature path; a learned gate fuses both paths before continuous soft-prompt projection. Furthermore, a progressive three-stage training scheme is adopted, encompassing encoder pre-training, interface optimization, and low-rank adaptation (LoRA) of Qwen2.5-1.5B. Extensive experiments are carried out on four standard datasets, CWRU, Gear, Mixed, and MBHM, and the Stage 2 signal-side output achieves classification accuracies of 97.1%, 99.0%, 93.4%, and 96.3%, respectively. The report-generation branch provides a schema-constrained signal-to-language interface for maintenance-oriented reporting.
Tao Wang, Yanqiang Di, Shao-Chong Feng et al.· Technologies· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
