Joint anomaly detection from multi-joint signals is a key component in maintaining motion-control quality in industrial robots. In practical applications, this task requires models that achieve strong detection performance while remaining lightweight. However, existing methods often struggle to achieve a satisfactory balance between these two objectives. To address this issue, this paper proposes the Joint-aware Multi-Scale Attention Network (JMSANet), a lightweight architecture for joint-level anomaly detection. Specifically, JMSANet includes: (1) a joint position encoding (JPE) front-end that embeds joint identity into the stacked multi-joint input and performs pointwise projection to reduce inter-joint feature confusion; (2) a lightweight joint-aware hierarchical downsampling (LJHD) backbone built upon depthwise separable lightweight convolution blocks, which extracts multi-scale temporal representations via stage-wise downsampling with a compact footprint; and (3) a multi-scale aggregation (MSA) head that fuses multi-scale features through a lightweight multi-scale attention mechanism over global average pooling, enabling lightweight cross-scale integration. Experimental results on in-situ vibration data collected from an industrial robot show that JMSANet achieves competitive Macro-Precision, Macro-Recall, and Macro-F1 while maintaining a lightweight model footprint. Sensitivity analysis and ablation studies further verify the contributions of JPE, the hierarchical backbone, and the MSA head, supporting the effectiveness of the proposed design for the lightweight industrial robot anomaly detection task. The code is publicly available at https://github.com/Nashizumi45/JMSANet-A-Lightweight-Network-for-Industrial-Robot-Multi-Joint-Anomaly-Detection.
DzedzickisASubačiūtė-ŽemaitienėJŠutinysE, et al. Advanced applications of industrial robotics: new trends and possibilities. Appl Sci2021; 12(1): 135.
2.
BilalHObaidatMSAslamMS, et al. Online fault diagnosis of industrial robot using IORT and hybrid deep learning techniques: an experimental approach. IEEE Internet Things J2024; 11(19): 31422–31437. https://doi.org/10.1109/JIOT.2024.3418352
3.
LiCLiuQYuQ, et al. Few-shot diffusion domain generalization for diagnosing joint reducer faults in industrial robots. IEEE Trans Instrum Meas2025; 74: 1–10. https://doi.org/10.1109/TIM.2025.3576027
4.
LongJMouJZhangL, et al. Attitude data-based deep hybrid learning architecture for intelligent fault diagnosis of multi-joint industrial robots. J Manuf Syst2021; 61: 736–745. https://doi.org/10.1016/j.jmsy.2020.08.010
5.
HuXJiangCHuangY, et al. Smnet: a novel compositional generalization model for industrial robot multi-joint fault diagnosis. IEEE Internet Things J2026; 13(7): 13005–13018. https://doi.org/10.1109/JIOT.2026.3652582
6.
LeiYLiuHLiN, et al. Condition monitoring and fault diagnosis of industrial robots: a review. Sci China Technol Sci2025; 68(1): 1110301. https://doi.org/10.1007/s11431-024-2810-2
7.
HanTXieWPeiZ. Semi-supervised adversarial discriminative learning approach for intelligent fault diagnosis of wind turbine. Inform Sci2023; 648: 119496. https://doi.org/10.1016/j.ins.2023.119496
8.
HanTWangXGuoJ, et al. Health-aware joint learning of scale distribution and compact representation for unsupervised anomaly detection in photovoltaic systems. IEEE Trans Instrum Meas2025; 74: 3538811. https://doi.org/10.1109/TIM.2025.3571122
9.
WangYZhangMTangX, et al. A kmap optimized VMD-SVM model for milling chatter detection with an industrial robot. J Intell Manuf2022; 33(5): 1483–1502. https://doi.org/10.1007/s10845-021-01736-9
10.
FanJGuoYYinX. Instantaneous angular speed-based fault detection for planet and cycloidal gears in rotary vector reducers. IEEE Trans Instrum Meas2025; 74: 1–10. https://doi.org/10.1109/TIM.2025.3542103
11.
LuKChenCWangT, et al. Fault diagnosis of industrial robot based on dual-module attention convolutional neural network. Auton Intell Syst2022; 2(1): 12. https://doi.org/10.1007/s43684-022-00031-5
12.
WangTZhangLWangX. Fault detection for motor drive control system of industrial robots using CNN-LSTM-based observers. CES Trans Electr Mach Syst2023; 7(2): 144–152. https://doi.org/10.30941/CESTEMS.2023.00014
13.
ZhiZLiuLLiuD, et al. Fault detection of the harmonic reducer based on CNN-LSTM with a novel denoising algorithm. IEEE Sens J2021; 22(3): 2572–2581. https://doi.org/10.1109/JSEN.2021.3137992
14.
WangZWuDZhuJ. Transformer-based sequential contrastive learning for industrial robot time series anomaly detection. In: 2025 10th International conference on computer and information processing technology (ISCIPT), Fushun, China, 2–14 September 2025, pp. 169–174. IEEE.
15.
JiangCHuangYHuX, et al. JFDNet: a robust industrial robot multi-joint fault detection network under sensor fault and environmental noise scenarios. IEEE Sens J2026; 26(8): 12398–12407. https://doi.org/10.1109/JSEN.2026.3673312
16.
YanSShaoHWangJ, et al. Liconvformer: a lightweight fault diagnosis framework using separable multiscale convolution and broadcast self-attention. Expert Syst Appl2024; 237: 121338. https://doi.org/10.1016/j.eswa.2023.121338
17.
CholletF. Xception: deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition, Honolulu, HI, USA, 21–26 July 2017, pp. 1251–1258. IEEE.
18.
LoshchilovIHutterF. Fixing weight decay regularization in adam. arXiv: 1711.05101, 2017. ICLR 2019.
19.
HuXZhangKHuangY, et al. Rrdnet: robust fault diagnosis of railway switch machines under imperfect sensing. IEEE Trans Instrum Meas2026; 75: 2507215. https://doi.org/10.1109/TIM.2026.3676178
20.
ShaoHMingYLiuY, et al. Small sample gearbox fault diagnosis based on improved deep forest in noisy environments. Nondestr Test Eval2024; 40: 1–22. https://doi.org/10.1080/10589759.2024.2404489
21.
YanSShaoHZhuoL, et al. Incremental few-shot fault diagnosis with cosine-represent real-time evolved network of transmission systems. IEEE Trans Transp Electr2025; 11(4): 9544–9554. https://doi.org/10.1109/TTE.2025.3563630
22.
LiCXieWZhengB, et al. An enhanced CLKAN-RF framework for robust anomaly detection in unmanned aerial vehicle sensor data. Knowl Based Syst2025; 319: 113690. https://doi.org/10.1016/j.knosys.2025.113690
23.
PengYShaoHXiaoY, et al. Dual-stage interpretable domain generalization fault diagnosis: integrating prior knowledge and gradient-weighted class activation mapping. Eng Appl Artif Intell2026; 166: 113655. https://doi.org/10.1016/j.engappai.2025.113655
24.
ZhangGKongXWangQ, et al. Balancing discrepancy and consistency: adversarial single domain generalization in fault diagnosis. IEEE Trans Ind Inform2025; 21(10): 7923–7934. https://doi.org/10.1109/TII.2025.3584526
25.
HuXZhangHCaoJ, et al. PSRONet: a deep reinforcement learning-based sensor configuration framework in railway point machines fault diagnosis. IEEE Trans Instrum Meas2026; 75: 2500513. https://doi.org/10.1109/TIM.2026.3654732
26.
ZhangXYanSWangC, et al. Generalizable fault diagnosis under distribution shifts induced by unseen working conditions via synthetic and adversarial sample learning. IEEE Trans Reliab2025; 74(4): 5517–5530. https://doi.org/10.1109/TR.2025.3616336
27.
YuYGuoLGaoH, et al. PCWGAN-GP: a new method for imbalanced fault diagnosis of machines. IEEE Trans Instrum Meas2022; 71: 1–11. https://doi.org/10.1109/TIM.2022.3180431
28.
LiXLiSZhangG, et al. A novel interpretable dynamic weighted domain adaptation network for cross-domain fault diagnosis of bearings under time-varying speeds. Eng Appl Artif Intell2026; 170: 114173. https://doi.org/10.1016/j.engappai.2026.114173
29.
KrizhevskyASutskeverIHintonGE. Imagenet classification with deep convolutional neural networks. Adv Neural Inform Process Syst2012; 25: 1097–1105.
30.
SimonyanKZissermanA. Very deep convolutional networks for large-scale image recognition. Comput Vis Patt Recogn2014. ICLR 2015. https://doi.org/10.48550/arXiv.1409.1556
SandlerMHowardAZhuM, et al. Mobilenetv2: inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition, Salt Lake City, UT, USA, 18–23 June 2018, pp. 4510–4520. IEEE.
33.
ZhangXZhouXLinM, et al. Shufflenet: an extremely efficient convolutional neural network for mobile devices. In: 2018 IEEE/CVF conference on computer vision and pattern recognition, alt Lake City, UT, USA, 18–23 June 2018, pp. 6848–6856. IEEE.
34.
ChenJhong KaoSHeH, et al. Run, don’t walk: chasing higher flops for faster neural networks. In: 2023 IEEE/CVF conference on computer vision and pattern recognition (CVPR), Vancouver, BC, Canada, 17–24 June 2023, pp. 12021–12031. IEEE.
35.
HowardASandlerMChuG, et al. Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision, Seoul, Korea, 27 October2 November 2019, pp. 1314–1324. IEEE.
36.
IandolaFNHanSMoskewiczMW, et al. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <0.5 mb model size. arXiv: 1412.3555, 2016.
BaiSKolterJZKoltunV. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv: 1803.01271, 2018. https://doi.org/10.48550/arXiv.1803.01271
39.
FangHDengJBaiY, et al. CLFormer: a lightweight transformer based on convolutional embedding and linear self-attention with strong robustness for bearing fault diagnosis under limited sample conditions. IEEE Trans Instrum Meas2021; 71: 1–8. https://doi.org/10.1109/TIM.2021.3132327
40.
Ismail-FawazADevanneMBerrettiS, et al. Lite: Light inception with boosting techniques for time series classification. In: 2023 IEEE 10th international conference on data science and advanced analytics (DSAA), Thessaloniki, Greece, 9–13 October 2023,, pp. 1–10. IEEE.
41.
ZengAChenMZhangL, et al. Are transformers effective for time series forecasting? In: Proceedings of the AAAI conference on artificial intelligence, Washington, DC, USA, 7–14 February 2023, vol. 37, pp. 11121–11128. AAAI Press.
42.
LiuYHuTZhangH, et al. itransformer: inverted transformers are effective for time series forecasting. arXiv:2310.06625, 2023. ICLR 2024.
43.
ParkJKangS. Paano: patch-based representation learning for time-series anomaly detection. In: The fourteenth international conference on learning representations, Rio de Janeiro, Brazil, 23–27 April 2026. OpenReview.
YiKZhangQFanW, et al. Frequency-domain MLPS are more effective learners in time series forecasting. Adv Neural Inform Process Syst2023; 36: 76656–76679.
