This paper systematically reviews the various research results of in-depth learning applied in the field of industrial fault diagnosis, in order to fill the gaps in the evolution route, key technology direction and practical application of existing reviews. During the combing process, it can be observed that the development of deep learning in this field has gradually extended from the construction of basic network architecture to the development of hybrid architecture. The current research direction is focused on the innovative development of technology directions such as attention mechanism, transfer learning and generative adversarial networks. In the actual landing process, the industrial scene puts forward multiple requirements for the model, including the difficulty of data acquisition, the difficulty of model logic interpretation, and the need for the calculation speed to match the pace of industrial production. This paper further prospects the future directions of automated machine learning, multimodal fusion, physical information fusion, etc., which can facilitate the follow-up researchers to carry out related work and promote the integration of intelligent diagnosis technology into the independent operation and maintenance mode.
深度学习的发展起始于早期的多层神经网络以及反向传播算法,然而在过去曾由于理论方面的瓶颈以及算力的限制而陷入发展的低谷期[14],一直到大约2006年的时候,借助无监督预训练[15]、线性整流函数(Rectified linear unit,ReLU)激活函数、图形处理器(Graphics processing unit,GPU)计算以及大规模数据等关键方面的突破才得以实现复兴,并且由此催生了以卷积神经网络和循环神经网络作为代表的深度架构[16]。在现阶段,深度学习凭借包含数亿参数的深层模型,在图像识别、语音处理以及自然语言理解等众多领域已经达到了近乎人类的性能水平,并且在工业界得到了广泛的应用,不过在无监督学习、复杂推理以及与强化学习相结合的通用人工智能等更具挑战性的领域,目前仍然处于探索的阶段。
在2006年左右,深度学习有过一次“复兴”,而这次复兴则很大程度上归功于无监督学习方法的突破。深度信念网络(Deep belief net,DBN)等无监督预训练方法通过逐层学习特征检测器,为深层网络的训练提供了有效的参数初始化策略。例如,文献[26]证明了无监督预训练在有限标注数据场景下的有效性,该方法特别适用于工业故障诊断中故障样本稀缺的实际情况,通过在大量无标签数据上预训练,模型能够学习到数据的基础分布特征。除此之外,生成对抗网络(Generative adversarial networks,GAN)和变分自编码器(Variational autoencoder,VAE)等生成模型为解决数据不平衡问题提供了新思路。一项研究提出的具有平衡串行CNN和Transformer的GAN(Generative adversarial networks with balancing serial CNN and transformer,BCTGAN)通过串联式CNN-Transformer结构生成高质量故障样本,在极端不平衡条件下仍能保持极高的诊断准确率[27]。另外一种VAE-CNN混合模型在低信噪比的噪声环境下仍表现优异[28]。
非接触式诊断是深度学习的一个重要应用方向。基于红外热像仪和迁移学习的诊断方法,仅需少量标记样本即可实现转子-轴承系统故障的准确识别,有效解决了振动传感器安装受限场景的监测难题[26]。这种方法的核心创新在于从无监督的卷积自编码器向有监督的增强卷积神经网络进行参数迁移,使得模型初始参数已经包含了从红外图像中提取通用特征的强大能力。另一项创新性工作设计了无监督域共享CNN架构,引入Cauchy核诱导的最大均值差异(Maximum mean discrepancy,MMD)替代传统的高斯核函数。这种改进不仅提升了跨工况特征适配的计算效率,还增强了在从稳态速度到时变速度等复杂迁移场景中的鲁棒性,为处理非平稳信号提供了新的技术路径[29]。
小样本学习的主要目的是,在每类故障只有几个标注样本的情形下,训练出具备强大泛化能力的诊断模型。主流方法有3大类。基于数据增强的方法,借助生成模型或者基于物理机理的仿真,比如机构特征生成模型(Mechanism character generative model,MCGM)合成虚拟故障样本,有效增加训练集。基于度量学习的方法,如原型网络、匹配网络,学习一个嵌入空间,让同类故障样本的特征相互接近,异类样本彼此远离,在推理时通过计算和少数支持样本的距离进行分类。基于元学习的方法在大量相关任务上展开元训练,使模型拥有快速适应新故障类别的能力。这些方法让模型在只有轴承、齿轮等旋转机械的少量实测故障数据时,依然能够实现较高的诊断精度,特别适用于新型设备或罕见故障的初期诊断阶段[30]。
迁移学习要把从一个相关领域(即源域,如某种稳定工况下的振动数据)学到的知识,迁移并适配到数据稀缺的新领域(即目标域,比如变工况下的数据),以此降低对目标域标注数据的依赖。依据源域与目标域的差异程度,迁移学习在故障诊断中呈现多种形式。特征迁移是较常见的方式,在源域上预训练特征提取器(如卷积自编码器[32]),再将其参数迁移到目标域模型作为初始化,使模型拥有从目标域少量数据中快速学习的能力。实例迁移方法是通过重加权源域样本,筛选出与目标域分布相似的样本参与训练。为应对更复杂的工况分布差异(即域偏移),域自适应方法被广泛研究,其核心是缩小源域与目标域在特征空间中的分布差异。例如,基于MMD的方法通过最小化域间距离来学习域不变特征[29],基于对抗学习的方法引入域判别器,驱动特征提取器产生混淆域判别器的特征,更精细的方法如分层交替迁移学习和基于局部中心矩差异(Local Central moment discrepancy,LCMD)的类别级对齐[33],致力于在更细粒度上实现特征分布的对齐,提高在变负载、变转速等复杂迁移场景下的诊断准确率和鲁棒性。
一方面,CNN的架构不断演进以契合工业信号特性:一维卷积神经网络可直接处理原始振动信号,达成端到端诊断,例如借助引入残差学习在强噪声环境下保持了较高的准确率[36],借助多通道输入模拟图像处理,可充分运用故障信息的空间相关性[37];二维卷积神经网络借助短时傅里叶变换[38]或者希尔伯特-黄变换[39]将时序信号转变为时频图像,利用CNN在图像识别方面的强大性能;多尺度卷积神经网络凭借并行运用不同尺寸卷积核来同时捕获故障的局部细节与全局趋势,比如基于选择性核块构造神经网络去噪的时间尺度自适应的卷积神经网络(Novel time-scale adaptive CNN with neural network denoiser constructed based on the selective kernel block,SKND-TSACNN)能自适应关注不同尺度特征[40],多尺度CNN与LSTM的混合模型[41]依靠特征融合实现了较好的噪声鲁棒性。
MinhN Q, KhiemN T, GiangV H. Fault classification and localization in power transmission line based on machine learning and combined CNN-LSTM models[J]. Energy Reports, 2024, 12: 5610-5622.
[2]
WangZ, LiuQ, ChenH, et al. A deformable CNN-DLSTM based transfer learning method for fault diagnosis of rolling bearing under multiple working conditions[J]. International Journal of Production Research, 2021, 59(16): 4811-4825.
[3]
YeL, MaX, WenC. Rotating machinery fault diagnosis method by combining time-frequency domain features and CNN knowledge transfer[J]. Sensors, 2021, 21(24): No.8168.
[4]
WangH, XuJ, YanR, et al. A new intelligent bearing fault diagnosis method using SDP representation and SE-CNN[J]. IEEE Transactions on Instrumentation and Measurement, 2020, 69(5): 2377-2389.
[5]
DibajA, EttefaghM M, HassannejadR, et al. A hybrid fine-tuned VMD and CNN scheme for untrained compound fault diagnosis of rotating machinery with unequal-severity faults[J]. Expert Systems with Applications, 2021, 167: No.114094.
[6]
BuC, LiuT, WangT, et al. A CNN-architecture-based photovoltaic cell fault classification method using thermographic images[J]. Energies, 2023, 16(9): No.3749.
[7]
PhamM T, KimJ M, KimC H. 2D CNN-based multi-output diagnosis for compound bearing faults under variable rotational speeds[J]. Machines, 2021, 9(9): No.199.
[8]
HuP, ZhaoC, HuangJ, et al. Intelligent and small samples gear fault detection based on wavelet analysis and improved CNN[J]. Processes, 2023, 11(10): No.2969.
[9]
SongX, CongY, SongY, et al. A bearing fault diagnosis model based on CNN with wide convolution kernels[J]. Journal of Ambient Intelligence and Humanized Computing, 2022, 13(8): 4041-4056.
[10]
ZhaoK, XiaoJ, LiC, et al. Fault diagnosis of rolling bearing using CNN and PCA fractal based feature extraction[J]. Measurement, 2023, 223: 113754.
[11]
WangX, ZhouP, PengX, et al. Fault location of transmission line based on CNN-LSTM double-ended combined model[J]. Energy Reports, 2022, 8: 781-791.
[12]
ZhangY, HeL, ChengG. MLPC-CNN: a multi-sensor vibration signal fault diagnosis method under less computing resources[J]. Measurement, 2022, 188: No.110407.
[13]
ChengY, WangD, ZhouP, et al. A survey of model compression and acceleration for deep neural networks[EB/OL]. [2025-12-19].
[14]
SchmidhuberJ. Deep learning in neural networks: an overview[J]. Neural Networks, 2015, 61: 85-117.
[15]
HintonG E, OsinderoS, TEHY W. A fast learning algorithm for deep belief nets[J]. Neural Computation, 2006, 18(7): 1527-1554.
[16]
KrizhevskyA, SutskeverI, HintonG E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.
[17]
ZachariadesC, XavierV. A review of artificial intelligence techniques in fault diagnosis of electric machines[J]. Sensors, 2025, 25: No.5128.
[18]
ChoiD J, HanJ H, ParkS U, et al. Comparative study of CNN and RNN for motor fault diagnosis using deep learning[C]∥2020 IEEE 7th International Conference on Industrial Engineering and Applications, Bangkok, 2020: 693-696.
[19]
HintonG, DengL, YuD, et al. Deep neural networks for acoustic modeling in speech recognition: the shared views of four research groups[J]. IEEE Signal Processing Magazine, 2012, 29(6): 82-97.
[20]
RumelhartD E, HintonG E, WilliamsR J. Learning representations by back-propagating errors[J]. Nature, 1986, 323: 533-536.
[21]
LecunY, BottouL, BengioY, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.
[22]
WangJ, JoshiG. Cooperative SGD: a unified framework for the design and analysis of local-update SGD algorithms[J]. JMLR, 2021, 22(213): 1-50.
[23]
ZhuY, LiG, TangS, et al. Acoustic signal-based fault detection of hydraulic piston pump using a particle swarm optimization enhancement CNN[J]. Applied Acoustics, 2022, 192: No.108718.
[24]
TangS, ZhuY, YuanS. Intelligent fault identification of hydraulic pump using deep adaptive normalized CNN and synchrosqueezed wavelet transform[J]. Reliability Engineering & System Safety, 2022, 224: No.108560.
[25]
ZhangC, ChenP, JiangF, et al. Fault diagnosis of nuclear power plant based on sparrow search algorithm optimized CNN-LSTM neural network[J]. Energies, 2023, 16(6): No.2934.
[26]
ZhiyiH, HaidongS, XiangZ, et al. An intelligent fault diagnosis method for rotor-bearing system using small labeled infrared thermal images and enhanced CNN transferred from CAE[J]. Advanced Engineering Informatics, 2020, 46: No.101150.
[27]
ChenH, WeiJ, HuangH, et al. Novel imbalanced fault diagnosis method based on generative adversarial networks with balancing serial CNN and Transformer (BCTGAN)[J]. Expert Systems with Applications, 2024, 258: No.125171.
[28]
WangY, LiD, LiL, et al. A novel deep learning framework for rolling bearing fault diagnosis enhancement using VAE-augmented CNN model[J]. Heliyon, 2024, 10(15): No.e35407.
[29]
CaoH, ShaoH, ZhongX, et al. Unsupervised domain-share CNN for machine fault transfer diagnosis from steady speeds to time-varying speeds[J]. Journal of Manufacturing Systems, 2022, 62: 186-198.
[30]
YuG, WuP, LvZ, et al. Few-shot fault diagnosis method of rotating machinery using novel MCGM based CNN[J]. IEEE Transactions on Industrial Informatics, 2023, 19(11): 10944-10955.
[31]
JiaW, DongZ, ShiH, et al. A zero-shot bearing fault diagnosis framework utilizing spatial relationships among primary features under variable working conditions[J]. Mechanical Systems and Signal Processing, 2025, 238: No.113103.
[32]
LvY, QiX, ZhengX, et al. Unsupervised quantitative judgment of furnace combustion state with CBAM-SCAE-based flame feature extraction[J]. Journal of the Energy Institute, 2024, 116: No.101733.
[33]
MengZ, CaoW, SunD, et al. Research on fault diagnosis method of MS-CNN rolling bearing based on local central moment discrepancy[J]. Advanced Engineering Informatics, 2022, 54: No.101797.
[34]
WangC, LiH, ZhangK, et al. Intelligent fault diagnosis of planetary gearbox based on adaptive normalized CNN under complex variable working conditions and data imbalance[J]. Measurement, 2021, 180: No.109565.
[35]
GuY, ZhangY, YangM, et al. Motor on-line fault diagnosis method research based on 1D-CNN and multi-sensor information[J]. Applied Sciences, 2023, 13(7): No.4192.
[36]
PengD, LiuZ, WangH, et al. A novel deeper one-dimensional CNN with residual learning for fault diagnosis of wheelset bearings in high-speed trains[J]. IEEE Access, 2019, 7: 10278-10293.
[37]
LiX, LiP, ZhangZ, et al. CNN-LSTM-based fault diagnosis and adaptive multichannel fusion calibration of filament current sensor for mass spectrometer[J]. IEEE Sensors Journal, 2024, 24(2): 2255-2269.
[38]
ChenJ, JiangJ, GuoX, et al. An efficient CNN with tunable input-size for bearing fault diagnosis[J]. International Journal of Computational Intelligence Systems, 2021, 14(1): No.625.
[39]
YuanZ, ZhangL, DuanL, et al. Intelligent fault diagnosis of rolling element bearings based on HHT and CNN[C]∥IEEE 2018 Prognostics and System Health Management Conference, Chongqing, China, 2018: 292-296.
[40]
YuZ, ZhangC, LiuJ, et al. SKND-TSACNN: a novel time-scale adaptive CNN framework for fault diagnosis of rotating machinery[J]. Knowledge-Based Systems, 2023, 275: No.110682.
[41]
ChenX, ZhangB, GaoD. Bearing fault diagnosis base on multi-scale CNN and LSTM model[J]. Journal of Intelligent Manufacturing, 2021, 32(4): 971-987.
[42]
LuoR, LiY, GuoH, et al. Cross-operating-condition fault diagnosis of a small module reactor based on CNN-LSTM transfer learning with limited data[J]. Energy, 2024, 313: No.133901.
[43]
KangJ, ZhuX, ShenL, et al. Fault diagnosis of a wave energy converter gearbox based on an Adam optimized CNN-LSTM algorithm[J]. Renewable Energy, 2024, 231: No.121022.
[44]
SnyderQ, JiangQ, TRIPPE. Integrating self-attention mechanisms in deep learning: a novel dual-head ensemble transformer with its application to bearing fault diagnosis[J]. Signal Processing, 2025, 227: No.109683.
[45]
ZhuQ X, QianY S, ZhangN, et al. Multi-scale Transformer-CNN domain adaptation network for complex processes fault diagnosis[J]. Journal of Process Control, 2023, 130: No.103069.
[46]
ChengD, LiuL, YuZ. CNN-based intelligent fault-tolerant control design for turbofan engines with actuator faults[J]. IEEE Access, 2021, 9: 28122-28139.
[47]
JiangT, GradusJ L, RoselliniA J. Supervised machine learning: a brief primer[J]. Behavior Therapy, 2020, 51(5): 675-687.
[48]
LiY F, LiangD M. Safe semi-supervised learning: a brief introduction[J]. Frontiers of Computer Science, 2019, 13(4): 669-676.
[49]
PanS J, YangQ. A survey on transfer learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2010, 22(10): 1345-1359.
[50]
冯志刚,王颖,王宇.基于多通道融合的齿轮箱故障诊断方法[J/OL]. [2025-12-19].
[51]
WangK, GouC, DuanY, et al. Generative adversarial networks: introduction and outlook[J]. IEEE/CAA Journal of Automatica Sinica, 2017, 4(4): 588-598.
[52]
PengD, WangH, DesmetW, et al. RMA-CNN: a residual mixed-domain attention CNN for bearings fault diagnosis and its time-frequency domain interpretability[J]. Journal of Dynamics, Monitoring and Diagnostics, 2023, 2: 115-132.
[53]
HuangY, ZhangJ, LiuR, et al. Improving accuracy and interpretability of CNN-based fault diagnosis through an attention mechanism[J]. Processes, 2023, 11(11): No. 3233.
[54]
ZhangJ, LiangJ, LiuJ. Fault diagnosis in reactor coolant pump with an automatic CNN-based mixed model[J]. Progress in Nuclear Energy, 2024, 175: No.105294.
[55]
HeC, GeD, YangM, et al. A data-driven adaptive fault diagnosis methodology for nuclear power systems based on NSGAII-CNN[J]. Annals of Nuclear Energy, 2021, 159: No. 108326.
[56]
DembczyńskiK, WaegemanW, ChengW, et al. On label dependence and loss minimization in multi-label classification[J]. Machine Learning, 2012, 88(1/2): 5-45.
[57]
NamJ, KimJ, MencíaE L, et al. Large-scale multi-label text classification-revisiting neural networks[J]. Lecture Notes in Computer Science, 2014, 8725: 437-452.
[58]
ZhangZ. Graph neural network-based bearing fault diagnosis using Granger causality test[J]. Expert Systems with Applications, 2024, 242: No.122827.
[59]
WeissK, KhoshgoftaarT M, WangD. A survey of transfer learning[J]. Journal of Big Data, 2016, 3(1): No.9.
[60]
CHENQ, ZHANGF, WANGY, et al. Bearing fault diagnosis based on efficient cross space multiscale CNN transformer parallelism[J]. Scientific Reports, 2025, 15(1): No.12344.
LiaoY L, SmidtT. EQUIFORMER: Equivariant graph attention transformer for 3D atomistic graphs[EB/OL]. [2025-12-19].
[63]
AliM, AliM, HussainM, et al. Generative adversarial networks (GANs) for medical image processing: recent advancements[J]. Archives of Computational Methods in Engineering, 2025, 32(2): 1185-1198.
[64]
BaratchiM, WangC, LimmerS, et al. Automated machine learning: past, present and future[J]. Artificial Intelligence Review, 2024, 57(5): No.122.
[65]
Alvarez-melisD, JaakkolaT S. On the robustness of interpretability methods[EB/OL]. [2025-12-19].
[66]
SiddiqueN I. Fault classification and location of a PMU-equipped active distribution network using deep convolution neural network (CNN)[J]. Electric Power Systems Research, 2024, 229: No. 110178.
[67]
HeX, ZhaoK, ChuX. AutoML: a survey of the state-of-the-art[J]. Knowledge-Based Systems, 2021, 212: No.106622.
[68]
ZhaoX, ZhuX, ZhangR, et al. Imbalanced fault diagnosis of a conditional variational auto-encoder with transfer and adversarial structures[J]. Advanced Engineering Informatics, 2025, 64: No. 103032.
[69]
ShangZ, WangZ, PanC, et al. Physics-informed auto-encoder based on digital twin for rolling bearing fault diagnosis under imbalanced sample conditions[J]. Engineering Applications of Artificial Intelligence, 2026, 163: No.113017.
[70]
LuZ, ChuZ, ZhuM, et al. Unsupervised feature learning using locality-preserved auto-encoder with complexity-invariant distance for intelligent fault diagnosis of machinery[J]. Applied Intelligence, 2025, 55(6): 379.
[71]
BayoudhK. A survey of multimodal hybrid deep learning for computer vision: architectures, applications, trends, and challenges[J]. Information Fusion, 2024, 105: No.102217.
[72]
YunT, LiJ, MaL, et al. Status, advancements and prospects of deep learning methods applied in forest studies[J]. International Journal of Applied Earth Observation and Geoinformation, 2024, 131: 103938.
[73]
ChenC, WangT, MaoD, et al. A multi-scale graph pyramid attention network with knowledge distillation towards edge computing robotic fault diagnosis[J]. Expert Systems with Applications, 2025, 260: No.125469.
[74]
YuZ, WangY, ZongX, et al. GRUDMU-DSCNN: an edge computing method for fault diagnosis with missing data[J]. Applied Intelligence, 2025, 55(2): No.140.
[75]
ZamanW, SiddiqueM F, UllahS, et al. Hybrid deep learning model for fault diagnosis in centrifugal pumps: a comparative study of VGG16, ResNet50, and wavelet coherence analysis[J]. Machines, 2024, 12(12): No.905.
[76]
WangX, JiangH, MuM, et al. A trackable multi-domain collaborative generative adversarial network for rotating machinery fault diagnosis[J]. Mechanical Systems and Signal Processing, 2025, 224: No.111950.
[77]
WangX, JiangH, ZengT, et al. An adaptive fused domain-cycling variational generative adversarial network for machine fault diagnosis under data scarcity[J]. Information Fusion, 2026, 126: No.103616.
[78]
FanC, ZhangY, MaH, et al. A novel lightweight DDPM-based data augmentation method for rotating machinery fault diagnosis with small sample[J]. Mechanical Systems and Signal Processing, 2025, 232: No.112741.