毕业生知识库

随着经济社会发展、交通运输规模膨胀、人口老龄化加剧等多种因素的交叉影响，创伤性骨折频繁发生，严重影响人类的生命和健康。下肢承担人体自重和实现基本运动，是骨折的常发和易发部位。下肢骨折的治疗包括复位手术以及手术后的康复。由于医师经验及术中设备的限制，传统骨折手术存在创伤大、手术风险高、难普及等不足，易发生复位不精准、二次感染等风险。同时，骨折复位与骨折康复两个过程严重脱节，且骨折康复过程缺乏有效的康复策略和量化的康复评价，常造成骨折康复盲目化、周期长甚至骨折术后不愈合等问题。集成有骨折精准复位功能和量化康复功能的机器人系统是应对上述问题的有效解决方案，已经成为学术界和工业界的研究热点。目前，骨折术后康复过程仍面临诸多挑战，如：主动康复训练中的意图识别、量化康复中评价指标的设定。本文在国家重点研发计划项目支持下，在合作团队研发的骨折康复机器人系统的基础上，根据实际临床需求，以术后主动康复、量化康复为切入点，研究基于脑电信号的运动意图识别算法、基于肌电信号的意图识别算法、骨折术后康复效果的量化评价等问题，主要研究内容和贡献如下：

分别基于脑电信号的时空特征和时间、空间、频率特征组合创新构建了意图识别模型，基于公开数据集的比较实验证明，所建立的模型可以获得比已有方法更高的识别率。一方面，将同一采样时刻所有通道的脑电信号投影到二维平面，在保留空间特征的同时将脑电信号转换为图像序列；然后采用卷积长短期记忆网络进行时空编码，生成时空特征；最后，采用卷积神经网络进一步提取特征，并采用softmax进行分类，得到基于脑电信号时空特征的意图识别模型。另一方面，利用复数Morlet波卷积将多个试次(Trial)的脑电信号转换成反映时间和频率特征的时频表示，设计注意力卷积网络提取各通道特征并赋予关键通道较高的权重，将权重和提取的特征进行聚合，并完成分类。最后，基于公开数据集的结果表明，本文的算法能够获得比已有算法更高的识别率。

针对脑电信号难以识别下肢多类运动意图、难以满足下肢骨折术后康复需求的问题，为下肢骨折患者设计了四类下肢动作，构建了基于表面肌电信号的下肢运动意图识别模型。针对四类下肢动作，采集四位下肢长骨骨折患者小腿关键骨骼肌的表面肌电信号。从数据分割长度、单特征选择、多个特征组合以及不同分类器等四个角度进行比较，设计最佳方案。结果表明：当窗口长度为200ms，选择过零点、波形长度、小波变换三种特征构成特征向量，采用BP神经网络作为分类器，可以获得最高的运动意图识别率。针对四位患者的下肢运动意图识别率分别为95.27%、96.28%、99.73%、98.16%。

提出基于表面肌电信号的康复阶段客观划分与下肢肌力评价方法，满足下肢骨折术后康复评价需求。针对康复阶段划分问题，设计50ms的数据窗口，优选平均绝对值、均方根、小波变换三种特征构成特征向量，使用支持向量机作为分类器，采用所设计的康复阶段划分模型对四位患者进行实验，可获得91.81\%的分类准确率。应用表面肌电指示下肢肌力状况，选用数据窗口为25ms，分别提取表面肌电信号的平均绝对值和均方根特征，比较健侧和患侧的肌力。结果显示：大部分患者患侧肌力明显小于健侧肌力。进而，基于该方法追踪患者患侧肌力随康复过程推进的变化情况。结果表明：患者经过一个月的康复训练，患侧腓骨长肌的力量有一定增长，朝着积极的方向发展。

With the cross-impact of various factors such as economic and social development, increasing population ageing, and expansion of transportation scale, traumatic fracture diseases, which occur frequently, are increasingly becoming a prominent problem affecting human life and health. The lower limbs are the common and prone parts of fractures because they bear the body's own weight and are responsible for basic movements. The treatment of lower limb fractures consists of reduction surgery and postoperative rehabilitation. Limited by the surgeon's experience and intraoperative equipment, traditional fracture surgery, with drawbacks of large trauma, high surgical risks and hard to popularization, is prone to lots of risks such as inaccurate reduction and secondary infection. At the same time, the two processes of fracture reduction and fracture rehabilitation are severely disconnected. The fracture rehabilitation process lacks effective rehabilitation strategies and quantitative assessment, which constantly causes severe problems, like aimless and long periods of rehabilitation, even nonunion after fractures. The robotic system integrated with accurate fracture reduction function and quantitative assessment function is an effective solution to the above problems, which has become a research hotspot in both academia and industry. At present, the rehabilitation process after fracture still faces a large number of challenges, such as the recognition of motion intentions in active rehabilitation training and the determination of criteria in quantitative rehabilitation. Therefore, supported by the National Key Research and Development project, this paper focuses on three key problems, including motion intention recognition algorithm based on EEG data, motion intention recognition algorithm based on sEMG data and the quantitative assessment of rehabilitation training based on the requirements of the robotic system and actual clinical needs. The main research contents and contributions of this paper are as follows:

Two intention recognition models based on the spatio-temporal and the spatio-temporal-spectral features of EEG data are innovatively constructed, separately. Extensive evaluations based on public dataset show that the proposed models outperform existing methods. On the one hand, EEG data are converted into image sequences while preserving the spatial features by projecting EEG data of all channels recorded at the same time onto a 2-D surface. Next, the Convolutional Long Short-Term Memory network is used to encode spatio-temporal features and generate a better representation from the obtained image sequence. The convolutional layer is used to further extract features, and softmax is used for classification. On the other hand, the complex Morlet wave convolution is used to convert the EEG signals of multiple trials into a time-frequency representation that reflects the time and spectral features. Next, the convolutional neural network combined with the attention mechanism is designed to extract the features of each channel and give the key channels higher attention weights. Finally, the attention weight and extracted features are aggregated for classification. Extensive evaluations based on public dataset show that the proposed models outperform existing methods.

Due to EEG data are difficult to recognize the multiple types of lower limb movement intentions, four types of lower limb movements are designed for patients with lower limb fractures. A lower limb motion intention recognition model based on sEMG data is proposed. For four types of lower limb movements, the sEMG data of key skeletal muscles of four patients with lower limb fractures are collected. Experiments are carried out from the four perspectives, such as data segmentation length, single feature selection, multiple feature combination, and different kinds of classifier. When window length is 200ms, and zero-crossing, waveform length, wavelet transform are combined to form the feature vector, using BP neural network as the classifier, the highest motion intention recognition rate can be obtained. The recognition rates of lower limb movement intention for four patients are 95.27%, 96.28%, 99.73%, 98.16%, respectively.

An sEMG-based method for objective division of rehabilitation stages and lower limb muscle strength assessment is proposed to meet the needs of rehabilitation assessment after lower limb fractures. To address the rehabilitation stage division problem, 50ms data window is designed. Mean absolute value, root mean square and wavelet transform are chosen to form feature vectors. The final classification accuracy of the support vector machine algorithm is 91.81\%. sEMG data are used to indicate the muscle strength of the lower limbs, and the data window is 25ms. The mean absolute value and root mean square features are extracted to compare the muscle strength of the healthy side and the affected side. The result shows that the muscle strength of the affected side in most patients is significantly weaker than that of the healthy side. Furthermore, the patient's muscle strength of the affected side is tracked along with the progress of the rehabilitation process based on this method. The result shows that the strength of the peroneus longus muscle on the affected side increases to a certain extent and develops in a positive direction after one month of rehabilitation training.