滚动轴承作为连接机械设备转动与静止部分的关键组件，其运行状态直接影响到设备整机安全，因此对滚动轴承进行精准、有效的故障诊断尤为重要。滚动轴承故障诊断的关键在于故障信息提取，所提取的特征信息对故障诊断结果起着决定性作用。传统故障诊断方式通过人工提取故障特征，多依赖于专家经验。随着深度学习等人工智能技术不断发展，基于深度学习的智能诊断方法逐渐应用于滚动轴承故障诊断领域。本文以滚动轴承为研究对象，结合时频分析和深度学习对滚动轴承振动信号进行分析和处理，在深入研究振动信号特征提取的基础上实现滚动轴承智能故障诊断。

滚动轴承振动信号时频图像生成方法研究。基于滚动轴承振动信号非线性、非平稳特点，采用时间-频率分析方法研究短时傅里叶变换（Short-time Fourier Transform，STFT）、连续小波变换（Continuous Wavelet Transform，CWT）和经验模态分解-伪魏格纳分布（Empirical Mode Decomposition&Pseudo-Wigner-ville Distribution，EMDPWVD）三种方法。利用这些时频分析方法提取故障特征，验证方法对故障特征表达的有效性。时频分析为基于Transformer的滚动轴承故障诊断提供图像化特征数据集。

（2）基于Vision Transformer（ViT）滚动轴承故障诊断模型研究。针对单一和复合故障时频图像数据集，基于接受词向量输入格式的编码-解码结构Transformer网络，提出一种接受图像输入格式且具备位置编码的ViT模型。首先将输入图像按照预定尺寸切分为图像块序列并加入位置编码保持序列的相对位置关系，再将其输入ViT模型中利用自注意力机制完成特征提取。研究模型内部敏感参数及网络层数设置的影响，得到三种时频图像对应的最优故障识别模型，在测试集上故障诊断准确率达到92.67%。

（3）基于池化层优化ViT滚动轴承故障诊断方法研究。由于直接转换所得ViT模型不包含卷积层和池化层，仅依靠自注意力机制完成特征提取，输入的序列在网络传输过程中均保持同一尺寸，特征提取能力受限。为探索空间特征尺寸变换对网络性能改进的有益作用，提出一种结合池化层的PiT模型（Pooling-based Vision Transformer，PiT）。实验证明，引入池化层后提升了原ViT模型的泛化能力，具有更好的故障识别能力，提升了模型的故障识别精度，同一样本输入下故障识别准确率较原ViT提升了3.70%。

本文以滚动轴承为研究对象，以振动时频图像为输入，通过深度学习算法完成了基于ViT模型的滚动轴承智能故障诊断，并在此基础上发展了改进的PiT模型，通过实验验证证明了所提出的时频图像与深度学习模型结合可以实现滚动轴承故障智能精密诊断。

As a key component connecting the rotating and stationary parts of mechanical equipment, the running status of rolling bearings directly affects the safety of the entire equipment. Therefore, accurate and effective fault diagnosis of rolling bearings is particularly important. The key to fault diagnosis of rolling bearings lies in the extraction of fault information, and the extracted feature information plays a decisive role in the fault diagnosis results. Traditional fault diagnosis methods rely heavily on expert experience by manually extracting fault features. With the continuous development of artificial intelligence technologies such as deep learning, intelligent diagnosis methods based on deep learning are gradually applied in the field of rolling bearing fault diagnosis. This article takes rolling bearings as the research object, combines time-frequency analysis and deep learning to analyze and process vibration signals of rolling bearings. Based on in-depth research on feature extraction of vibration signals, intelligent fault diagnosis of rolling bearings is achieved.

(1) Research on time-frequency image generation method for vibration signals of rolling bearings. Based on the nonlinear and non-stationary characteristics of rolling bearing vibration signals, time-frequency analysis method is used to study three methods: Short-time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), and Empirical Mode Decomposition&Pseudo-Wigner-ville Distribution (EMDPWVD). Use these time-frequency analysis methods to extract fault features and verify the effectiveness of the method in expressing fault features. Time-frequency analysis provides a graphical feature dataset for Transformer based fault diagnosis of rolling bearings.

(2) Research on the Fault Diagnosis Model of Rolling Bearings Based on Vision Transformer (ViT). A ViT model with position encoding is proposed for single and composite fault time-frequency image datasets based on an encoding decoding structure Transformer network that accepts word vector input formats. Firstly, the input image is segmented into a sequence of image blocks according to a predetermined size and position encoding is added to maintain the relative position relationship of the sequence. Then, it is input into the ViT model to complete feature extraction using a self attention mechanism. By studying the influence of sensitive parameters and network layer settings within the model, the optimal fault recognition model corresponding to three time-frequency images was obtained, with a fault diagnosis accuracy of 92.67% on the test set.

(3) Research on Fault Diagnosis Method for ViT Rolling Bearings Based on Pooling Layer Optimization. Due to the fact that the directly converted ViT model does not include convolutional and pooling layers, and only relies on self attention mechanism to complete feature extraction, the input sequence remains the same size during network transmission, resulting in limited feature extraction capability. To explore the beneficial effects of spatial feature size transformation on network performance improvement, a Pooling-based Vision Transformer (PiT) model is proposed, which combines pooling layers. Experiments have shown that the introduction of pooling layer enhances the generalization ability of the original ViT model, has better fault recognition ability, and improves the fault recognition accuracy of the model. Under the same sample input, the fault recognition accuracy is improved by 3.70% compared to the original ViT.

This article takes rolling bearings as the research object, takes vibration time-frequency images as input, and completes intelligent fault diagnosis of rolling bearings based on the ViT model through deep learning algorithm. Based on this, an improved PiT model is developed. Experimental verification shows that the combination of the proposed time-frequency images and deep learning models can achieve intelligent and precise diagnosis of rolling bearing faults.

[1]雷亚国，贾峰，周昕，等. 基于深度学习理论的机械装备大数据健康监测方法[J]. 机械工程学报，2015，51(21): 49-56.

[2]何正嘉，曹宏瑞，訾艳阳，等. 机械设备运行可靠性评估的发展与思考[J]. 机械工程学报，2014，50(2): 171-186.

[3]屈梁生，何正嘉. 机械故障诊断学[M].上海：上海科学技术出版社，1986.

[4]王国彪，何正嘉，陈雪峰，等. 机械故障诊断基础研究“何去何从”[J]. 机械工程学报，2013，49(1): 63-72.

[5]国家自然科学基金委员会工程与材料科学部. 机械工程学科发展战略报告：2011～2020[M]. 北京：科学出版社，2010.

[6]罗嘉宁. 实验与仿真数据驱动的滚动轴承故障严重性评估研究[D]. 哈尔滨：哈尔滨工业大学，2019.

[7]雷亚国，贾峰，孔德同，等. 大数据下机械智能故障诊断的机遇与挑战[J]. 机械工程学报，2018，54(05): 94-104.

[8]陈志强，陈旭东，José Valente de Olivira，等. 深度学习在设备故障预测与健康管理中的应用[J]. 仪器仪表学报，2019，40(09): 206-226.

[9]张雪英，栾忠权，刘秀丽. 基于深度学习的滚动轴承故障诊断研究综述[J]. 设备管理与维修，2017(18): 130-133.

[10]陈是扦，彭志科，周鹏. 信号分解及其在机械故障诊断中的应用研究综述[J]. 机械工程学报，2020，56(17): 91-107.

[11]李舜酩，郭海东，李殿荣. 振动信号处理方法综述[J]. 仪器仪表学报，2013，34(08): 1907-1915.

[12]李恒，张氢，秦仙蓉，等. 基于短时傅里叶变换和卷积神经网络的轴承故障诊断方法[J]. 振动与冲击，2018，37(19): 124-131.

[13]张东，冯志鹏. 迭代广义短时Fourier变换在行星齿轮箱故障诊断中的应用[J]. 工程科学学报，2017，39(04): 604-610.

[14]Wang Y, Deng S E, Wang H D. Fault Diagnosis Method for Rolling Bearings Based on SG Filter and STFT with Combining Window Functions[J]. Bearing, 2011.

[15]Morlet J, Arens G, Fourgeau E. et al. Wave propagation and sampling theory-Part 1: Complex signal scattering in multilayered media. Geophysics, 1982,47(2):203~221.

[16]袁建虎，韩涛，唐建，等. 基于小波时频图和CNN的滚动轴承智能故障诊断方法[J]. 机械设计与研究，2017，33(02): 93-97.

[17]周建华，郑攀，王帅星，等. 基于小波时频图和卷积神经网络的行星齿轮箱故障诊断方法[J]. 机械传动，2022，46(01): 156-163.

[18]王亚萍，许迪，葛江华，等. 基于SPWVD时频图纹理特征的滚动轴承故障诊断[J].振动. 测试与诊断，2017，37(01): 115-119+203.

[19]闫维明，马裕超，何浩祥，等. 双线性时频分布交叉项提取及损伤识别应用[J]. 振动. 测试与诊断，2014，34(06): 1014-1021+1166.

[20]Shiau Y H, Wu M C. Detecting characteristics of information masked by a laser-triggered microwave system via Hilbert-Huang transform[J]. Optics Communications, 2010, 283(9): 1909-1916

[21] Yang W X, Tavner P J. Empirical mode decomposition,an adaptive approach for interpreting shaft vibratory signals of large rotating machinery[J]. Journal of Sound and Vibration, 2009, 321(3-5): 1144-1170

[22]Fan H W, Shao S J, Zhang X H, et al. Intelligent Fault Diagnosis of Rolling Bearing Using FCM Clustering of EMD-PWVD Vibration Images[J]. IEEE Access, 2020, 8: 145194-145206.

[23]Mu Weijie, Jiang Haibo. Internal Combustion Engine Fault Diagnosis Method Based MICEEMD-PWVD[J]. IOP Conference Series: Earth and Environmental Science, 2021, 632(3).

[24]樊红卫，张旭辉，曹现刚，等. 智慧矿山背景下我国煤矿机械故障诊断研究现状与展望[J]. 振动与冲击，2020，39(24): 194-204.

[25]郭一帆，唐家银. 基于机器学习算法的寿命预测与故障诊断技术的发展综述[J]. 计算机测量与控制，2019，27(03): 7-13.

[26]肖乾浩. 基于机器学习理论的机械故障诊断方法综述[J]. 现代制造工程，2021，No.490(07): 148-161.

[27]翟嘉琪，杨希祥，程玉强，等. 机器学习在故障检测与诊断领域应用综述[J]. 计算机测量与控制，2021，29(03): 1-9.

[28]李巍华，单外平，曾雪琼. 基于深度信念网络的轴承故障分类识别[J]. 振动工程学报，2016，29(02): 340-347.

[29]Yang Yu, Yu Dejie, Cheng Junsheng. A roller bearing fault diagnosis method based on EMD energy entropy and ANN[J]. Journal of Sound and Vibration, 2005, 294(1).

[30]宋丹丹，魏域琴，范启富. 基于SVM和KPCA的滚动轴承故障诊断[J]. 化工自动化及仪表，2019，46(12): 988.

[31]黄海松，魏建安，任竹鹏，等. 基于失衡样本特性过采样算法与SVM的滚动轴承故障诊断[J]. 振动与冲击，2020，39(10): 65-74+132.

[32]陶骏杰. 基于小波分析和支持向量机的滚动轴承故障诊断研究[D]. 北京：北京化工大学，2019.

[33]石志炜，张丽萍，钟成豪，等. 改进PSO优化SVM的滚动轴承故障诊断[J]. 福州大学学报(自然科学版)，2020，48(03): 333-340.

[34]Meng Li. The Application of PCA and SVM in Rolling Bearing Fault Diagnosis[J]. Advanced Materials Research, 2012, 1601(430-432).

[35]刘鲲鹏，苏涛，赵磊，等. 滚动轴承故障模式识别方法研究现状分析[J]. 内燃机与配件，2018，No.275(23): 62-63.

[36]欧阳承达，赵红梅，张军. 基于模糊聚类法的滚动轴承故障诊断研究[J]. 自动化应用，2020(06): 29-31.

[37]朱永泽，陈志刚. 基于图拉普拉斯矩阵和改进K均值聚类的滚动轴承故障诊断[J]. 电子测量技术，2021，44(10): 45-51.

[38]尹宝才,王文通,王立春.深度学习研究综述[J].北京工业大学学报,2015,41(01):48-59.

[39]王猛，王光君，张雷，等. 基于时频分析的风力发电机轴承剩余寿命预测[J]. 山东电力技术，2022，49(12): 32-38+100.

[40]童靳于，罗金，郑近德. 基于增强深度自编码网络的滚动轴承故障诊断方法[J]. 中国机械工程，2021，32(21): 2617-2624.

[41]李巍华，单外平，曾雪琼. 基于深度信念网络的轴承故障分类识别[J]. 振动工程学报，2016，29(02): 340-347.

[42]Tao J, Liu Y L, Yang D, et al. Fault diagnosis of rolling bearing using deep belief networks[C]. 2015 International Symposium on Material, Energy and Environment Engineering. [S.l.]: Atlantis Press, 2015.

[43]郗涛，葛增元，王莉静. 基于参数优化的VMD与DBN的滚动轴承故障诊断[J]. 组合机床与自动化加工技术，2022(12): 57-61..

[44] Y LeCun, B Boser, J Denker, et al. Backpropagation applied to handwritten zip code recognition[J]. Neural Computation, 1989, 541-551.

[45]许理，李戈，余亮，等. 基于小波包与CNN的滚动轴承故障诊断[J]. 四川理工学院学报(自然科学版)，2018，31(03): 54-59.

[46]Fan Hongwei, Xue Ceyi, Zhang Xuhui, et al. Vibration Images-Driven Fault Diagnosis Based on CNN and Transfer Learning of Rolling Bearing under Strong Noise[J]. Shock and Vibration, 2021, 2021(2):1-16.

[47]杨有帅. 基于Transformer的图像特征提取方法研究[D]. 电子科技大学，2022. DOI:10.27005/d.cnki.gdzku.2022.003530.

[48]Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]. Neural Inf. Process. Syst. 2017, 30, 5998–6008.

[49]Wang Q, Li B, Xiao T, et al. Learning Deep Transformer Models for Machine Translation[C]. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019, 1810-1822.

[50]Dai ZH, Yang ZL, Yang YM, et al. Transformer-XL: Attentive Language Models beyond a Fixed-length Context[J]. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019, 2978-2988.

[51]刘文婷，卢新明. 基于计算机视觉的Transformer研究进展[J]. 计算机工程与应用，2022，58(06): 1-16.

[52]Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv: 2010. 11929, 2020.

[53]赵爽，魏国辉，赵文华，等. 基于Vision Transformer的小儿肺炎辅助诊断研究[J/OL]. 激光与光电子学进展：1-16[2023-04-05]. http://kns.cnki.net/kcms/detail/31.1690.tn.20220713.1915.575.html.

[54]王贵参，杨承林，蒲佳佳，等. 基于改进的Vision Transformer杂草生长周期识别方法研究[J]. 长春工业大学学报，2022，43(06): 712-718.

[55]Bazi Yakoub, Bashmal Laila, Rahhal Mohamad M. et al. Vision Transformers for Remote Sensing Image Classification[J]. Remote Sensing, 2021, 13(3).

[56]Wu H, Xiao B, Codella N, et al. Cvt: Introducing convolutions to vision transformers [C]. Proceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, 2021: 22-31.

[57]Heo B, Yun S, Han D, et al. Rethinking spatial dimensions of vision transformers [C]. Proceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, 2021: 11936-11945.

[58]何翔. 滚动轴承故障机理及智能化检测技术研究[D]. 成都：西南交通大学，2017.

[59]孔亚林. 基于振动信号的滚动轴承故障诊断方法研究[D]. 大连：大连理工大学，2006.

[60]宋琼. 红外图像质量增强及小目标检测算法研究[D]. 武汉：华中科技大学，2019.

[61]胡安洲. 主客观一致的图像感知质量评价方法研究[D]. 合肥：中国科学技术大学，2014.

[62]任欢，王旭光. 注意力机制综述[J]. 计算机应用，2021，41(S1): 1-6.

[63]高耀东，侯凌燕，杨大利. 基于多标签学习的卷积神经网络的图像标注方法[J]. 计算机应用，2017，37(01): 228-232.

[64]肖瑛，冯长建. 组合窗函数的短时傅里叶变换时频表示方法[J]. 探测与控制学报，2010，32(03): 43-47.

[65]郝欢，王华力，魏勤. 经验模态分解理论及其应用[J]. 高技术通讯，2016，26(01): 67-80.