.png){kind=icon}
.png){kind=icon}
Return
An Incremental Contrastive Learning Method for Compound Fault Diagnosis of Rolling Bearings
.png){kind=icon}
PRE
AI
J
Y
R
X
L
A
DOI：10.1109/jiot.2026.3703119.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Compound faults, which arise from the interaction of multiple simultaneous failures, pose significant risks to the maintenance of industrial machinery in the industrial Internet of Things (IIoT), leading to complex and unpredictable system failures. Traditional models tend to misclassify emerging compound faults as known categories due to a bias toward seen data. In addition, the delayed emergence of compound faults relative to single faults hampers prompt sample collection. Furthermore, compound fault datasets typically exhibit a long-tailed distribution. Driven by these challenges, we propose an incremental contrastive learning (CL) model based on cross-modal contrastive embedding (ICLCFD) to achieve an incremental fault diagnosis from single faults to compound faults. First, we employ a symmetrized dot pattern (SDP) image transformation to convert vibration signals into visual representations. We extract high-dimensional visual features from these SDP images using a multiscale residual convolutional neural network (MS-ResCNN) and generate low-dimensional semantic features based on vibration signals to obtain richer feature information. Subsequently, a novelty detection mechanism is developed using the predicted number of fault sources as a count-based indicator to identify newly emerging faults. Furthermore, we incorporate a difficulty-aware class rebalancing sampling strategy to prioritize hard-to-diagnose samples during incremental updates, mitigating the excessive impact of head-class samples on the diagnosis results. Experiments on three open-source bearing datasets [Case Western Reserve University (CWRU), Paderborn University (PU), and Xi’an Jiaotong University (XJTU-SY)] demonstrate that ICLCFD achieves a state-of-the-art (SOTA) accuracy of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$96.58 {\,}\pm {\,}0.28\%$ </tex-math></inline-formula>, outperforming existing methods by approximately 2%. The results confirm its effectiveness in handling unknown and compound faults in dynamic IoT maintenance pipelines.
Keywords:
Bearings
compound fault diagnosis
contrastive learning (CL)
generalized zero-shot learning (GZSL)
incremental learning
Journal
IF:
8.9
Papers:
1.4W
Citations:
7.8W
Organization
Cited Papers