基于故障语义生成和广义零样本学习的复合故障诊断模型
董靖川 , 李佳明 , 王太勇 , 赵锡博 , 张正日 , 陈楷文
天津大学学报(自然科学与工程技术版) ›› 2026, Vol. 59 ›› Issue (11) : 1127 -1137.
基于故障语义生成和广义零样本学习的复合故障诊断模型
Compound Fault Diagnosis Model Based on Fault Semantics Generation and Generalized Zero-Shot Learning
滚动轴承复合故障诊断是故障诊断领域的一大挑战,主要是由于复合故障具有复杂的耦合特性,且在实际工业环境中难以获取和标记复合故障数据.现有的零样本复合故障诊断方法普遍缺乏同时识别单一故障和复合故障的能力.针对这一问题,提出了一种基于故障语义生成和广义零样本学习的复合故障诊断(FSG-GZSL)模型,以实现不同工作条件下的滚动轴承单一故障与复合故障的识别.该模型主要由复合故障预诊断模块、语义生成模块和复合故障诊断模块组成.首先,构建双通道域对抗神经网络作为复合故障预诊断模块,将时频图作为模块的输入,通过设定阈值条件来实现单一故障的识别并分离出复合故障.其次,设计了一种基于时频域统计特征的故障语义构建方法,用以提取样本的故障语义特征,并构建变分自编码器-生成式对抗网络(VAE-GAN)模型作为语义生成模块,通过学习单故障语义特征及其之间的耦合关系,生成伪复合故障语义.最后,利用 K-means 聚类模型作为复合故障诊断模块,先将伪复合故障语义输入形成聚类中心,通过比较样本故障语义与聚类中心的距离来实现复合故障的分类.本文在两个真实轴承故障数据集上进行了实验验证,分别达到了 96.02%和 97.41%的准确率,结果表明该方法在诊断效果上优于其他对比方法,且展现出较强的鲁棒性.
Diagnosing compound faults of rolling bearings remains a major challenge in fault diagnosis. This is mainly due to the complex coupling characteristics of compound faults and the difficulty in collecting and labeling compound fault data under real industrial conditions. Existing zero-shot compound fault diagnosis methods generally lack the ability to identify single and compound fault categories. To address this issue,a compound fault diagnosis model based on fault semantics generation and generalized zero-shot learning(FSG-GZSL) was proposed. This model aimed to recognize single and compound rolling bearing faults under various working conditions. The proposed model consisted of three modules:a pre-diagnosis module,a semantics generation module,and a compound fault diagnosis module. First,a dual-channel domain-adversarial neural network was developed as the pre-diagnosis module. It takes time-frequency images as input and identifies single faults by applying a threshold while separating compound faults. Next,a fault semantics construction method based on time-frequency statistical features was designed to extract semantic features from samples. A variational autoencoder-generative adversarial network(VAE-GAN) model was built as the semantics generation module,which would learn the semantic features of single faults and their coupling relationships to generate pseudo compound fault semantics. Finally,a K-means clustering model was employed as the diagnosis module. Pseudo compound fault semantics were used to form cluster centers,and compound faults were classified by comparing distances between the sample semantics and cluster centers. Experiments were conducted on two real-world bearing fault datasets and achieved accuracies of 96.02% and 97.41%,respectively. These results demonstrate that the proposed method outperforms other competing approaches in diagnostic accuracy and exhibits strong robustness.
| [1] |
王太勇, 王廷虎, 王 鹏, |
| [2] |
|
| [3] |
王国锋, 张旭东, 汪 菲, |
| [4] |
|
| [5] |
林诗麒, 陈智丽, 李宇鹏, |
| [6] |
|
| [7] |
|
| [8] |
程建刚, 毕凤荣, 张立鹏, |
| [9] |
|
| [10] |
周 旷, 覃文博, 孙天宇. 基于可信多源领域自适应的跨域滚动轴承故障诊断[J]. 控制与决策, 2025, 40(7): 2251-2260. |
| [11] |
|
| [12] |
|
| [13] |
王 鹏, 李丹青, 王 恒. 基于改进交替迁移学习的滚动轴承故障诊断算法[J]. 振动与冲击, 2024, 43(5): 239-249. |
| [14] |
|
| [15] |
魏文军, 张轩铭, 杨立本. 基于模糊聚类和改进 Densenet 网络的小样本轴承故障诊断[J]. 哈尔滨工业大学学报, 2024, 56(3): 154-163. |
| [16] |
|
| [17] |
刘 臻, 彭珍瑞, 王圣杰. 小样本下 SE—ResNet 与元迁移学习的变工况轴承故障诊断[J]. 振动工程学报, 2025, 38(6): 1199-1211. |
| [18] |
|
| [19] |
邹 松, 董绍江, 夏宗佑, |
| [20] |
|
| [21] |
|
| [22] |
李耀华, 赵 佳, 张鑫杰. 基于语义融合零样本学习的轴承复合故障诊断模型[J]. 振动与冲击, 2025, 44(10): 278-286. |
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
国家重点研发计划资助项目(2022YFB3303601)
天津市科学技术局资助项目(23YDPYGX00210)
/
| 〈 |
|
〉 |