Closed-loop cortical control of direction using support vector machines.

Motor neuroprosthetics research has focused on reproducing natural limb motions by correlating firing rates of cortical neurons to continuous movement parameters. We propose an alternative system where specific spatial-temporal spike patterns, emerging in tasks, allow detection of classes of behavior with the aid of sophisticated nonlinear classification algorithms. Specifically, we attempt to examine ensemble activity from motor cortical neurons, not to reproduce the action this neural activity normally precedes, but rather to predict an output supervisory command to potentially control a vehicle. To demonstrate the principle, this design approach was implemented in a discrete directional task taking a small number of motor cortical signals (8-10 single units) fed into a support vector machine (SVM) to produce the commands Left and Right. In this study, rats were placed in a conditioning chamber performing a binary paddle pressing task mimicking the control of a wheelchair turning left or right. Four animal subjects (male Sprague-Dawley rats) were able to use such a brain-machine interface (BMI) with an average accuracy of 78% on their first day of exposure. Additionally, one animal continued to use the interface for three consecutive days with an average accuracy over 90%.     

Acquisition of optimal attribute subset through genetic algorithm using GNP‐based class association rule mining

Attribute selection is a technique to prune less relevant information and discover high‐quality knowledge. It is especially useful for the classification of a large database, because the preprocessing of data increases the possibility that predictor attributes given to the mining algorithm become more relevant to the class attribute. In this paper, a method to acquire the optimal attribute subset for the genetic network programming (GNP) based class association rule mining has been proposed, and this attribute selection process using genetic algorithm (GA) leads to a higher accuracy for classification. Class association rule mining through GNP is conducted with a small subset of data rather than the original large number of attributes; thus simple but important rules are obtained for classification while the local optimal problem is avoided. Simulation results with educational data show that the classification accuracy is largely improved from 52.73 to 74.54%, when classification is made using the optimal attribute subset. © 2014 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.     

PD recognition with knowledge-based preprocessing and neuralnetworks

Partial discharge (PD) patterns are an important tool for the diagnosis of HV insulation systems. Human experts can discover possible insulation defects in various representations of the PD data. One of the most widely used representations is phase-resolved PD (PRPD) patterns. We present a method for the automated recognition of PRPD patterns using a neural network (NN) for the actual classification task. At the core of our method lies a preprocessing scheme that extracts relevant features from the raw PRPD data in a knowledge-based way, i.e. according to physical properties of PD gained from PD modeling. This allows a very small NN to be used for classification. In addition to the classification of single-type patterns (one defect) we present a method to separate superimposed patterns stemming from multiple defects. High recognition rates are achieved with a large number of single patterns generated by stochastic PD simulations. Our network architecture compares favorably with a more traditional network architecture used previously for PRPD classification. These results are confirmed by classification of patterns measured in laboratory experiments and power stations     

SIMONE: a realistic neural network simulator to reproduce MEA-based recordings.

Contemporary multielectrode arrays (MEAs) used to record extracellular activity from neural tissues can deliver data at rates on the order of 100 Mbps. Such rates require efficient data compression and/or preprocessing algorithms implemented on an application specific integrated circuit (ASIC) close to the MEA. We present SIMONE (Statistical sIMulation Of Neuronal networks Engine), a versatile simulation tool whose parameters can be either fixed or defined by a probability distribution. We validated our tool by simulating data recorded from the first olfactory relay of an insect. Different key aspects make this tool suitable for testing the robustness and accuracy of neural signal processing algorithms (such as the detection, alignment, and classification of spikes). For instance, most of the parameters can be defined by a probabilistic distribution, then tens of simulations may be obtained from the same scenario. This is especially useful when validating the robustness of the processing algorithm. Moreover, the number of active cells and the exact firing activity of each one of them is perfectly known, which provides an easy way to test accuracy.     

Evolutionary selection of interesting class association rules using genetic relation algorithm

During the last years, several association rule‐based classification methods have been proposed, these algorithms may quickly generate accurate rules. However, the generated rules are often very large in terms of the number of rules and usually complex and hardly understandable for users. Among all the rules generated by the algorithms, only some of them are likely to be of any interest to the domain expert analyzing the data. Most of the rules are either redundant, irrelevant or obvious. In this paper, a new method for selecting the interesting class association rules is proposed by an evolutionary method named genetic relation algorithm. The algorithm evaluates the relevance and interestingness of the discovered association rules by the relationships between the rules in each generation using a specific measure of distance among them giving a reduced set of rules as the result in the final generation. This small rule set has the following properties: (i) accurate as it has at least the same classification accuracy as the complete association rule set, (ii) interesting because of the diversity of rules and (iii) comprehensible because it is more understandable for the users as the number of attributes involved in the rules is also small. The efficiency of the proposed method is compared with other conventional methods including genetic network programming‐based mining using ten databases and the experimental results show that it outperforms others keeping a good balance between the classification accuracy and the comprehensibility of the rules. © 2011 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.     

基于BP神经网络的运动想象脑电信号分类研究

运动想象脑机接口因具有更大的自主性、灵活性,在脑机互联领域得到了广泛应用,相比较其它范式分类准确率偏低,限制了其发展。本文利用时频图谱、脑地形图两种特征分析方法对上肢运动想象脑电信号进行了特征分析,并采用滤波器组共空间模式(filter bank co-space,FBCSP)特征提取算法对上肢运动想象信号数据进行了特征提取,再将提取结果分别利用支持向量机（support vector machine,SVM）算法、K-最近邻(K-Nearest Neighbor)算法、反向传播（back propagation,BP）神经网络三种分类算法进行分类,研究结果发现SVM算法、KNN算法、BP神经网络算法应用在上肢运动想象脑机接口系统的平均分类准确率分别为76.45%、74.55%、81.70%,BP神经网络算法相比SVM算法、KNN算法在上肢运动想象任务的分类准确率上分别高出了5.25%、7.15%,并且t检验后得到分类准确率均具有极显著的统计学差异,并利用ROC曲线和AUC值检测了分类器效果,BP神经网络的AUC值相比SVM算法、KNN算法也分别提升了0.1226、0.1285,表明BP神经网络分类算法相比较SVM算法、KNN算法更适用于上肢运动想象脑机接口系统,提高了系统的分类准确率,推动了上肢运动想象脑电信号实际应用的发展进程。     

基于MGS-LGBM算法的电力系统暂态稳定评估

多粒度级联轻型梯度提升机(MGS-LGBM)具有超参数设置简单、模型泛化能力强、分类准确率高、训练评估快等特点。为提高电力系统暂态稳定评估的准确性和快速性,将MGS-LGBM引入电力系统暂态评估中。首先通过时域仿真提取原始数据,构造能够反映系统稳定情况的23维特征量,输入MGS-LGBM模型中,稳定结果作为输出量,利用模型中的多粒度扫描和级联结构对样本特征和结果进行高效并行训练。通过新英格兰10机39节点系统仿真验证MGS-LGBM算法,通过与其它机器学习算法比较,算法在提高暂态评估准确率的同时兼顾快速性,且在含有无关特征和训练集较少的情况下仍能保持较好的评估性能。     

基于模糊多目标遗传优化算法的节假日电力负荷预测

多目标遗传优化算法的一个优点就是可在一次迭代计算中寻找到问题的多个非劣最优解。该文应用多目标遗传算法和关联规则算法提出一个基于模糊规则的电力负荷模式分类系统。在此分类系统中采用多目标遗传优化算法从众多模糊分类规则中自动挑选出具有较好识别性能和可解释性的模糊规则，并利用模糊关联规则挖掘通过启发式规则选择改善遗传算法的搜索性能。经仿真试验表明此分类系统具有较好的分类性能，可为节假日负荷预测提供更为充分的历史数据，从而改善其负荷预测性能。     

Single limb performance following contralateral bimanual limb training.

Recent studies on intermanual transfer of reaching movements suggest that this transfer is conducted over an "extrinsic" coordinate system. We hypothesize that training reaching movements in a force field with both hands at the same time, in the same position (bimanual grip) will be more beneficial in promoting transfer of the learned skill to the dominant hand than training the unimpaired limb on the same movements in the same force field since the representation of the movement should be invariant of the limb. However, unlike intermanual transfer, bimanual transfer has the potential to involve infinite number of actuator combinations, or joint configurations, interfering with consistent transfer. The efficacy of this method of transfer has implications for people with hemiparesis since the less-affected arm could potentially "instruct" the more-affected arm how to move. Here, we report on an experiment that evaluates and compares the skill transfer between limbs in a reaching task: 1) intermanual transfer (from the nondominant to the dominant hand) and 2) bimanual transfer (from a bimanual grip to the dominant hand) with healthy subjects. We used two methods from which to judge the transfer: performance in the presence of the force field or by errors made during "catch trials" when the forces were unexpectedly removed as subjects changed hands (known as after effects of adaptation). We found only a small amount of transfer (20% of that seen in the practiced limb) with both types of training, and surprisingly there was no significant difference in the movement accuracy between these two training methods. Moreover, the direction of the after effects supports the assertion that the nervous system generalizes these movements in an extrinsic coordinate system. Accordingly, the limb must experience the dynamics singularly in order to develop an internal model.     

Question Classification for Chinese Cuisine Question Answering System

This paper presents an integrated method on question classification for Chinese cuisine question answering (QA) system. First, we exploit the domain knowledge to enrich question preprocessing, then classification features are extracted by means of domain attributes and the rule-based classifier is constructed. Support vector machine (SVM) classifier is used for secondary classification to the questions which cannot be matched with rules. A prototype system based on the proposed method has been constructed and an experiment on 453 natural language questions collected from Internet has been carried out. It achieved an accuracy of 96.22%. Result shows that a small number of linguistically motivated domain features can efficiently classify questions of Chinese cuisine QA system. Copyright © 2009 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.     

A virtual reality environment for designing and fitting neural prosthetic limbs.

Building and testing novel prosthetic limbs and control algorithms for functional electrical stimulation (FES) is expensive and risky. Here, we describe a virtual reality environment (VRE) to facilitate and accelerate the development of novel systems. In the VRE, subjects/patients can operate a simulated limb to interact with virtual objects. Realistic models of all relevant musculoskeletal and mechatronic components allow the development of entire prosthetic systems in VR before introducing them to the patient. The system is used both by engineers as a development tool and by clinicians to fit prosthetic devices to patients.     

基于GA和ELM的电能质量扰动识别特征选择方法

电力系统中海量暂态扰动的分析与治理需要以高效准确的扰动分类为基础。现有扰动识别方法缺少合理的特征选择环节,分类器过于复杂,不能满足高效分类的需要。提出一种新的电能质量扰动特征选择方法。首先,对原始信号使用S变换进行预处理,提取具有代表性的25种扰动信号特征构建原始特征集合;然后,根据极限学习机识别准确率构造用于扰动特征选择的遗传算法适应度函数;最后,用遗传算法来进行迭代运算,确定最优特征集合。实验证明,新方法能够有效去除冗余特征,在保证分类准确率前提下,有效降低分类器复杂度,提高分类效率。     

基于阴影增强和注意力机制的高光谱图像分类

基于深度学习网络的高光谱图像分类能够有效地提取图像中的特征信息，促进遥感图像中丰富信息的挖掘与利用。然而，现有方法性能仍然受限于阴影信息不能充分提取、特征不能有效利用。针对阴影区域信息提取，动态随机共振能够利用噪声增强信号，提高信息的表达能力；针对特征利用，在卷积神经网络中嵌入注意力机制，能够在其提取的高层特征的基础上，从空间维度和通道维度进一步提取融合，筛选出对当前任务目标更为关键的特征，提升网络分类性能。实验结果表明：通过在含有阴影区域的真实高光谱图像数据集Hydice上仿真，动态随机共振能够有效增强信号进而将分类精度从96.48%提升到97.14%,卷积注意块的加入使分类精度提升了0.408 4%。进一步与其他分类方法在Hydice、Indian Pines、Pavia University进行实验对比验证，本文方法分类精度分别达到了97.436 1%、99.219 5%和99.929 9%,对不同数据集的分类都具有良好的表现，相较于其他方法具有明显优势，证明了该方法的有效性和良好的分类性能，在高光谱图像分类领域具有广阔的应用前景。     

一种基于图卷积神经网络的加密流量分类方法

深度学习算法被广泛应用于网络流量分类领域并取得较好效果。然而,对抗攻击的出现给其安全性带来了严重威胁,使得当前主流的基于卷积神经网络模型的分类算法的精度严重下降。针对此,本文提出了一种抗流量分类中灰度图对抗攻击的加密流量分类方法。所提方法通过提取数据包负载长度、包序列、方向、簇等流量交互信息构建拓扑图,将加密流量分类问题转化为图分类问题。接着,本文使用基于图卷积神经网络的分类方法进行特征的学习分类,图卷积神经网络模型可以自动从输入的拓扑图中提取特征,将特征映射到嵌入空间中的不同表示来区分不同的图结构。实验结果表明,本文所提方法不仅能够避免对抗攻击,且在公开数据集上的分类性能也较现有典型方法提高了5%以上。     

基于融合度量策略的小样本金属缺陷识别研究

针对目前金属表面缺陷分类，数据稀缺且标注步骤繁琐昂贵的问题，将小样本度量学习引入金属表面缺陷分类中，提出了一种小样本分布度量网络模型FDM-FSL:用信息更加丰富的细节描述子来表征图像特征，并通过空间注意力机制筛选获得更具判别力的描述子信息，最后引入融合KL散度和EMD距离的图像到类的度量方式以考虑查询集和支持集类别的分布一致性。实验结果表明，提出的网络模型在小样本数据集MiniImageNet上拥有更加优良的度量能力，5类5样本下平均识别精度相较经典的RelationNet、CovaMNet、DN4算法识别准确率提高了6.34%、5.78%、1.25%。在金属缺陷数据集NEU-DET上5类5样本平均识别准确率分别提高了2.87%、3.34%、2.5%。     

A Hybrid Support Vector Fuzzy Inference System for the Classification of Leakage Current Waveforms Portraying Discharges

Several techniques have been applied on leakage current waveforms in order to extract information regarding electrical activity on high-voltage insulators. However, a fully representative value is yet to be defined. In this article, a hybrid support vector fuzzy inference system is introduced as a classification tool. The system incorporates fuzzy logic, genetic algorithms, and support vector machines. Apart from the classification accuracy achieved, the system also produces a set of fuzzy rules under which the classification is made, allowing a further insight of the process. A comparison is made to other classification tools previously applied on the same data set.     

Support Vector Machine for Classification of Voltage Disturbances

The support vector machine (SVM) is a powerful method for statistical classification of data used in a number of different applications. However, the usefulness of the method in a commercial available system is very much dependent on whether the SVM classifier can be pretrained from a factory since it is not realistic that the SVM classifier must be trained by the customers themselves before it can be used. This paper proposes a novel SVM classification system for voltage disturbances. The performance of the proposed SVM classifier is investigated when the voltage disturbance data used for training and testing originated from different sources. The data used in the experiments were obtained from both real disturbances recorded in two different power networks and from synthetic data. The experimental results shown high accuracy in classification with training data from one power network and unseen testing data from another. High accuracy was also achieved when the SVM classifier was trained on data from a real power network and test data originated from synthetic data. A lower accuracy resulted when the SVM classifier was trained on synthetic data and test data originated from the power network.     

利用加速度信号时频域特征的枪击识别研究

当前枪支射弹可靠检测及精确计数是枪弹管控的难点之一。为提高基于加速度信号的射弹检测算法的精度和可靠性,提出一种新的射击信号时域特征提取方法—时域分段特征提取法,可避免时域特征过度依赖于加速度瞬时尖峰的问题。首先,提取了枪击加速度样本信号的时域和频域各类统计特征。然后,采用机器学习分类算法K近邻、逻辑回归、支持向量机以及决策树和随机森林进行枪击识别建模。最后,探索和比较各种单一特征对枪击事件识别模型性能的影响。实验结果表明,所提取的主波动域面积特征具有最优的区分度,能够在多数机器学习算法上达到99%以上的分类准确率。     

一种面向类不平衡加密流量的端到端分类模型

当前流量分类方法在面对类不平衡流量时,往往存在着在少数类上的分类效果不佳的问题。针对该问题,提出了一种面向类不平衡加密流量的端到端分类模型。所提模型在传统卷积神经网络模型的基础上添加了一个Inception模块进行特征融合,让模型能提取到更丰富的特征,弥补了少数类因样本数量少所带来的特征学习上的不足;同时引入一个通道-空间域注意力模块,对Inception模块所融合的特征根据重要程度赋予相应的权值,使模型更多地关注到更重要的特征,增强流量特征的表征能力。与此同时,为减少网络参数,采用卷积层加全局平均池化层的组合代替模型中的全连接层。实验结果表明,相较于当前典型流量分类模型,所提模型在数据集少数类上具有更优的分类性能,精确率、召回率和F1-Score均有显著提高,其中综合性能指标F1-Score在某些少数类上的提升达到了15%~18%。     

Uncalibrated Workpiece Positioning Method for Peg-in-hole Assembly Using an Industrial Robot

This paper proposes an uncalibrated workpiece positioning method for peg-in-hole assembly of a device using an industrial robot. Depth images are used to identify and locate the workpieces when a peg-in-hole assembly task is carried out by an industrial robot in a flexible production system. First, the depth image is thresholded according to the depth data of the workpiece surface so as to filter out the background interference. Second, a series of image processing and the feature recognition algorithms are executed to extract the outer contour features and locate the center point position. This image information, fed by the vision system, will drive the robot to achieve the positioning, approximately. Finally, the Hough circle detection algorithm is used to extract the features and the relevant parameters of the circular hole where the assembly would be done, on the color image, for accurate positioning. The experimental result shows that the positioning accuracy of this method is between 0.6-1.2 mm, in the used experimental system. The entire positioning process need not require complicated calibration, and the method is highly flexible. It is suitable for the automatic assembly tasks with multi-specification or in small batches, in a flexible production system.