1 State Key Laboratory of Mechanical System and Vibration, School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
2 National Key Laboratory of Advanced Micro and Nano Manufacture Technology, School of Integrated Circuits, Shanghai Jiao Tong University, Shanghai 200240, China
Ultra-flexible electrodes are nanometer-thick polymer-based brain-computer interface devices implanted via a “sewing machine” paradigm: microneedles rapidly inject the electrode into the target brain region under visual guidance and immediately withdraw, leaving only the independently floating electrode for long-term neural recording. Visual positioning errors, fragile structure, and long-distance transport can easily cause abnormalities such as hooking failure, fracture, and detachment. However, existing implantation systems pay little attention to real-time automatic identification of abnormal events, which undermines the reliability of automated implantation. To solve the problem, a force-vision fusion-based method was proposed for abnormal events identification, which leveraged the complementary characteristics of high-frequency, low-signal-to-noise-ratio (SNR) force signals and low-frequency, high-SNR visual signals. Specifically, it first identified abnormal events in real time by evaluating the change rate of force and the absolute deviation of the force signal from its initial value. Then, it identified the type of abnormal events with a vision-language model, which was fine-tuned with the low-rank adaptation method and structured textual prompt. Finally, the operation was adjusted according to the identified event type. Experimental results show that the proposed method accurately identifies both the timing and type of abnormal events. It not only prevents unnecessary tissue damage caused by inserting a microneedle without an attached microelectrode, but also enables re-hooking and insertion after hooking failure, thereby avoiding microelectrode waste and shortening implantation time.
LIU Chao, HAN Bo, CHEN Han-wei, YANG Kai-jie, WANG Long-chun, LIU Jing-quan, SHENG Xin-jun.
Force-vision Fusion-based Recognition Method for Abnormal Events in Ultra-flexible Microelectrode Implantation[J].
Science Technology and Engineering, 2026, 26(19): 8039-8048 DOI:10.12404/j.issn.1671-1815.2508478
在基于视觉的异常识别实验中,用于微调视觉语言模型的训练数据集包含304张图像,其中正常植入图像247张,钩取失败图像25张,电极断裂或脱落图像32张。用于评估的测试数据集包含256张图像,其中正常植入138张,钩取失败27张,电极断裂或脱落91张。需要注意的是,异常状况发生前的图像仍属于正常类别,因此训练和测试集中正常植入图像数量偏多符合真实数据分布。评价指标包括三部分:“是否存在严重运动模糊”的预测准确率、“是否异常”的预测准确率,以及异常类型的预测准确率(区分钩取失败和电极断裂或脱落)。由于操作区域通常不是正方形,且改变图像高宽比会导致微针和电极形变,因此图像在预处理时进行了等比例缩放,并限制其像素面积不超过262 144(512×512)。训练轮次设置为20,此时训练损失已收敛。考虑到边缘部署的推理时间和显存限制,模型结构与初始权重使用Qwen3-VL-2B-Instruct。由于模型权重较大,批大小与梯度累积步数均设置为4,相当于批训练大小为16的训练效果。优化器采用AdamW,学习率为2×10-4,权重衰减为0.005,其余设置沿用Qwen3-VL默认参数。训练和推理均使用一张RTX PRO 6000显卡完成。
RasiahN P, MaheshwaryR, KwonC S, et al. Complications of deep brain stimulation for Parkinson disease and relationship between micro-electrode tracks and hemorrhage: systematic review and meta-analysis[J]. World Neurosurgery, 2023, 171: 8-23.
[2]
CurrieA D, WongJ K, OkunM S. A review of temporal interference, nanoparticles, ultrasound, gene therapy, and designer receptors for Parkinson disease[J]. npj Parkinson’s Disease, 2024, 10(1): 195.
[3]
HansonT L, Diaz-BotiaC A, KharaziaV, et al. The “sewing machine” for minimally invasive neural recording[J/OL]. BioRxiv (2023-03-15)[2025-12-09]. https://doi.org110.1101/578542.
[4]
MuskE. An integrated brain-machine interface platform with thousands of channels[J]. Journal of Medical Internet Research, 2019, 21(10): 16194.
[5]
WeiC, WangY, PeiW, et al. Distributed implantation of a flexible microelectrode array for neural recording[J]. Microsystems & Nanoengineering, 2022, 8(1): 50.
[6]
LiangW, QinF, HanX, et al. Keypoint localization based on convolutional neural network for robotic implantation of flexible micro-electrodes[C]// 2022 IEEE 18th International Conference on Automation Science and Engineering (CASE). New York: IEEE, 2022: 822-827.
[7]
QinF, XuD, ZhangD, et al. Automated hooking of biomedical microelectrode guided by intelligent microscopic vision[J]. IEEE/ASME Transactions on Mechatronics, 2023, 28(5): 2786-2798.
[8]
AnY, YangJ, HeB, et al. A microscopic vision-based robotic system for floating electrode assembly[J]. IEEE/ASME Transactions on Mechatronics, 2024, 29(5): 3810-3820.
[9]
HanB, ChenH, LiuC, et al. An intelligent microscope vision system for hooking multi-thread flexible microelectrode based on few-shot segmentation[J]. IEEE Transactions on Automation Science and Engineering, 2025, 22: 11112-11123.
[10]
HanB, ChenH, LiuC, et al. Automated Insertion of flexible microelectrodes guided by vessel-aware microscopic vision[J]. IEEE/ASME Transactions on Mechatronics, 2025, 31(2): 1829-1840.
[11]
ShiB, WangF, HuoZ, et al. Contact force sensing and control for inserting operation during precise assembly using a micromanipulator integrated with force sensors[J]. IEEE Transactions on Automation Science and Engineering, 2022, 20(3): 2147-2155.
[12]
ZhangY, ZhaoA, WangC, et al. Robotic cell micromanipulation for posture adjustment of zebrafish embryonic cell[C]// Journal of Physics: Conference Series. IOP Publishing, 2025, 3101(1): 012014.
[13]
ZhouY, YangH, WangX, et al. A mosquito mouthpart-like bionic neural probe[J]. Microsystems & Nanoengineering, 2023, 9(1): 88.
[14]
HeB, GuoY, YangG. Integrated piezoelectric vibration and in situ force sensing for low-trauma tissue penetration[J]. Cyborg and Bionic Systems, 2025, 6: 0417.
[15]
ChenH, HanB, LiuC, et al. Delicate operation of a microneedle-forceps mechanism for ultra-flexible probe implantation[J]. IEEE Robotics and Automation Letters, 2025, 10(8): 8427-8434.
WangShuai, HeWenchun, WangFudi, et al. Research on question answering system based on large language model integrating knowledge graph and vector retrieval[J]. Science Technology and Engineering, 2024, 24(32): 13902-13910.