A Hybrid Approach for Predicting Probability of Default in Peer-to-Peer (P2P) Lending Platforms Using Mixture-of-Experts Neural Network

Peer-to-peer (P2P) lending offers an alternative way to access credit. Unlike established lending institutions with proven credit risk management practices, P2P platforms rely on numerous independent variables to evaluate loan applicants’ creditworthiness. This study aims to estimate default probabilities using a mixture-of-experts neural network in P2P lending. The approach involves coupling unsupervised clustering to capture essential data properties with a classification algorithm based on the mixture-of-experts structure. This classic design enhances model capacity without significant computational overhead. The model was tested using P2P data from Lending Club, comparing it to other methods like Logistic Regression, AdaBoost, Gradient Boosting, Decision Tree, Support Vector Machine, and Random Forest. The hybrid model demonstrated superior performance, with a Mean Squared Error reduction of at least 25%.

Application of SOM neural network in clustering

The Self-Organizing Map (SOM) is an unsupervised neural network algorithm that projects high-dimensional data onto a two-dimensional map. The projection preserves the topology of the data so that similar data items will be mapped to nearby locations on the map. One of the SOM neural network’s applications is clustering of animals due their features. In this paper we produce an experiment to analyze the SOM in clustering different species of animals.

Fuzzy Cluster Neural Network Based on Wavelet Transform and Its Vibration Application

This paper advances a new approach based on wavelet and wavelet packet transforms in tandem with a fuzzy cluster neural network,abbreviated WPFCNN.Wavelets and wavelet packets decompose a vibration signal into different bands at different levels and provides multiresolution or multiscale views of a signal which is stationary or nonstationary. Fuzzy mathematics processes uncertain problems in engineering and converts the attributes extracted by wavelet packets to fuzzy membership degree.To achieve self-organizing classification,the MAXNET neural network is employed.WPFCNN integrates the advantages of wavelet packets and fuzzy cluster with MAXNET.The approach is adopted to process and classify vibration signal of a NH_3 compressor in a petrochemical plant.The results indicate that it is a useful and effective intelligence classification in the field of condition monitoring and fault diagnosis.

Hopfield Neural Network Approach to Clustering in Mobile Radio Networks

ＨｏｐｆｉｅｌｄＮｅｕｒａｌＮｅｔｗｏｒｋＡｐｐｒｏａｃｈｔｏＣｌｕｓｔｅｒｉｎｇｉｎＭｏｂｉｌｅＲａｄｉｏＮｅｔｗｏｒｋｓＪｉａｎｇＹａｎ；ＬｉＣｈｅｎｇｓｈｕ（ＮｏｒｔｈｅｒｎＪｉａｏｔｏｎｇＵｎｉｖｅｒｓｉｔｙ，Ｂｅｉｊｉｎｇ１０００４４）Ａｂ...

Prediction of Solar Radiation Using Data Clustering and Time-Delay Neural Network

In this paper, a combination of data clustering and artificial intelligence techniques are used to predict incoming solar radiation on a daily basis. The data clustering technique known as Perceptually Important Points is proposed, where time-series data is grouped into clusters separated by key characteristic points, which are later used as training data for an artificial neural network. The type of network used is known as a Focused Time-Delay Neural Network, and an analysis of the data is performed using the Mean Absolute Percentage Error scheme.

Simultaneous Forecast for Three Speciations of Heavy Metal Elements Using Fuzzy Cluster-Artificial Neural Network

The three speciations（water extract, adsorption and organic speciations） of Cu, Zn, Fe and Mn in geo-chemical samples were determined by fuzzy cluster-artificial neural network（FC-ANN） method coupled with atomic absorption spectrometry. A back-propagation artificial neural network with one input node and three export nodes was constructed, which could forecaste three speciations of heavy metals simultaneously. In the learning sample set, the three speciations of each element were allowed to change in a wide concentration range and the accuracy of the analysis was apparently increased via the learning sample set optimized with the help of the fuzzy cluster analysis. The average relative errors of the three speciations of Cu, Zn, Fe or Mn from 100 geo-chemical samples were less than 5%. The relative standard deviations of the three speciations of each of four heavy metals were 0.008%―4.43%.

Fault Diagnostics on Steam Boilers and Forecasting System Based on Hybrid Fuzzy Clustering and Artificial Neural Networks in Early Detection of Chamber Slagging/Fouling

The slagging/fouling due to the accession of fireside deposits on the steam boilers decreases boiler efficiency and availability which leads to unexpected shut-downs. Since it is inevitably associated with the three major factors namely the fuel characteristics, boiler operating conditions and ash behavior, this serious slagging/fouling may be reduced by varying the above three factors. The research develops a generic slagging/fouling prediction tool based on hybrid fuzzy clustering and Artificial Neural Networks (FCANN). The FCANN model presents a good accuracy of 99.85% which makes this model fast in response and easy to be updated with lesser time when compared to single ANN. The comparison between predictions and observations is found to be satisfactory with less input parameters. This should be capable of giving relatively quick responses while being easily implemented for various furnace types.

Method of neural network modulation recognition based on clustering and Polak-Ribiere algorithm

To improve the recognition rate of signal modulation recognition methods based on the clustering algorithm under the low SNR, a modulation recognition method is proposed. The characteristic parameter of the signal is extracted by using a clustering algorithm, the neural network is trained by using the algorithm of variable gradient correction （Polak-Ribiere） so as to enhance the rate of convergence, improve the performance of recognition under the low SNR and realize modulation recognition of the signal based on the modulation system of the constellation diagram. Simulation results show that the recognition rate based on this algorithm is enhanced over 30% compared with the methods that adopt clustering algorithm or neural network based on the back propagation algorithm alone under the low SNR. The recognition rate can reach 90% when the SNR is 4 dB, and the method is easy to be achieved so that it has a broad application prospect in the modulating recognition.

Wear State Recognition of Drills Based on K-means Cluster and Radial Basis Function Neural Network

Drill wear not only affects the surface smoothness of the hole, but also influences the life of the drill. Drill wear state recognition is important in the manufacturing process, which consists of two steps： first, decomposing cutting torque components from the original signals by wavelet packet decomposition （WPD）; second, extracting wavelet coefficients of different wear states （i.e., slight, normal, or severe wear） with signal features adapting to Welch spectrum. Finally, monitoring and recognition of the feature vectors of cutting torque signal are performed by using the K-means cluster and radial basis function neural network （RBFNN）. The experiments on different tool wears of the multivariable features reveal that the results of monitoring and recognition are significant and effective.

聚焦热度变化、主题动态与情感趋势的微博舆情演化研究

[研究目的]系统探讨微博舆情事件的演化特征,以提出针对性的对策建议,避免网络舆情扩散所可能引发的不利影响。[研究方法]为实现该目的,提出了基于CNN-BiLSTM-Attention的微博舆情多维特征演化分析框架,以深入剖析网络舆情的形成机制,进而优化对网络舆情的应对和处理策略。[研究结论]根据选取的事件从新浪微博获取数据,基于TF-IDF模型和K-Means聚类算法对微博舆情事件进行了维度划分,通过组合模型CNN-BiLSTM-Attention进行情感分类,并验证其准确性。最后,根据维度划分和情感分类的结果,结合舆情生命周期理论,从舆情热度、主题和情感三个方面研究了微博舆情事件的演化情况,并从生命周期和主题情感两方面得出网络舆情应对策略。

基于残差密集块的激光遥感图像中目标检测方法

为了提高对目标检测的效果,提出基于残差密集块的激光遥感图像中目标检测方法。首先,设计基于残差密集块的卷积神经网络,在设计ReLU激活函数并完成网络训练后,基于含噪激光遥感图像的初步特征提取结果,利用单个卷积展开卷积映射处理,抽取出潜在干净图像。然后,通过聚类处理的方式,得到激光遥感图像中车辆目标的显著图,再利用大律法,通过建立的特征比例关系的方式检测出其中的目标信息。实验结果表明,应用该方法有效滤除激光遥感图像中的噪声,并精准检测出激光遥感图像中的车辆目标。相比于3种传统方法,该方法检测结果均值误差的最小值仅为0.0156,说明该方法有效实现了设计预期。

基于遗传优化聚类的GRU无损电力监测数据压缩

针对电力调度中心监测数据记录体量大、存储困难的问题,提出基于遗传优化K-means聚类的门控循环单元神经网络无损数据压缩方法。首先,搭建分布式集群,将多维原始电力数据聚类成相似性较高的数据块,并利用遗传算法对聚类进行寻优,提高数据聚类的效果;再通过门控循环单元神经网络训练数据编码的概率分布模型,结合算术编码对数据进行编码压缩;最后,以多个电力数据集为算例进行分析。经验证本文所提的压缩算法能实现数据的高比例压缩、优化集群性能。

基于K-means聚类和特征空间增强的噪声标签深度学习算法

深度学习中神经网络的性能依赖于高质量的样本,然而噪声标签会降低网络的分类准确率。为降低噪声标签对网络性能的影响,噪声标签学习算法被提出。该算法首先将训练样本集划分成干净样本集和噪声样本集,然后使用半监督学习算法对噪声样本集赋予伪标签。然而,错误的伪标签以及训练样本数量不足的问题仍然限制着噪声标签学习算法性能的提升。为解决上述问题,提出基于K-means聚类和特征空间增强的噪声标签深度学习算法。首先,该算法利用K-means聚类算法对干净样本集进行标签聚类,并根据噪声样本集与聚类中心的距离大小筛选出难以分类的噪声样本,以提高训练样本的质量;其次,使用mixup算法扩充干净样本集和噪声样本集,以增加训练样本的数量;最后,采用特征空间增强算法抑制mixup算法新生成的噪声样本,从而提高网络的分类准确率。并在CIFAR10、CIFAR100、MNIST和ANIMAL-10共4个数据集上试验验证了该算法的有效性。

电动自行车穿行对公交乘客进出站行为的安全影响分析

为深刻认知公交乘客在横穿非机动车道进出公交站台过程中与纵向穿行的电动自行车形成的交叉型碰撞冲突的特征并探究冲突严重程度的影响因素,提出电动自行车穿行对公交乘客进出站行为的安全影响分析方法。首先,依据航拍获取的电动自行车和公交乘客运行轨迹,将公交乘客-电动自行车冲突划分为公交乘客避险、电动自行车避险、双方均避险等3类;然后在交叉型冲突特征分析的基础上,提出相对碰撞时间(TTR)指标,联合后侵入时间(PET)、安全减速度(DST)构建基于改进K-means聚类的冲突严重程度评价模型;最后分别从公交乘客和电动自行车视角提取5个冲突严重程度影响因素,建立基于BP神经网络的冲突严重等级预测模型,并对模型有效性进行检验。结果表明:速度差和相对速度对冲突严重等级的影响最大;电动自行车主动避险对冲突双方的运动状态影响最小,相对碰撞能量最低;所提冲突严重等级预测模型的计算结果与实际冲突等级具有较高符合度,能够在一定程度预测冲突严重程度。

NHNet——新型层次化遥感图像语义分割网络

深度学习分割方法是遥感图像分割领域的热点之一,主流的深度学习方法有卷积神经网络、transformer神经网络及两者的结合。特征提取是图像分割的重要环节,除了用卷积等方式提取特征,最近的研究聚焦于一些新的特征提取范式,如图卷积、小波变换等。本文利用聚类算法的区域构建属性,将改进的聚类算法用于骨干特征提取模块,同时使用卷积和视觉transformer作为辅助模块,以获取更丰富的特征表述;在模块基础上,提出了一种新型层次化遥感图像语义分割网络(NHNet);评估了NHNet语义分割的性能,并在LoveDA遥感数据集上与其他方法进行比较。结果表明,基于多特征提取的NHNet获得了竞争性的性能表现,平均交并比为49.64%,F_(1)分数为65.7%。同时,消融实验证明辅助模块提高了聚类算法分割的精确性,给NHNet分别提升了1.03%和2.41%的平均交并比。

基于视觉图像与激光点云融合的交通标志快速识别方法

交通标志对车辆交通起到重要作用和意义,而智能交通中交通标志识别由于标志特征提取效果差,导致识别率低、识别时间长,因此,提出一种新的基于视觉图像与激光点云融合的交通标志快速识别方法。采用双边滤波方法预处理原始激光点云数据;通过归一化处理得到视觉图像激光点云融合的目标空间激光点云位置测距数值;通过测距值获取目标图像位置,归一化处理交通标志视觉图像,引入k均值聚类算法二聚类处理图像,采用制作的切割模板切割图像感兴趣区域,提取交通标志图像的深度特征,结合卷积神经网络二次过滤特征,重新标定二次过滤后的特征,最终利用卷积神经网络模型实现交通标志快速识别;经实验对比证明,采用所提方法提取各个类型交通标志特征的提取效果较好,并且识别率达到89.74%,识别时间仅为13.1 s,干扰下识别时间最高仅为15.1 s,验证了该方法可以快速且准确识别各个类型的交通标志。

基于集群辨识和卷积神经网络-双向长短期记忆-时序模式注意力机制的区域级短期负荷预测

为了解决区域级短期电力负荷预测时输入特征过多和负荷时序性较强的问题,提出一种基于集群辨识和卷积神经网络(convolutional neural networks,CNN)-双向长短期记忆网络(bi-directional long short-term memory,BiLSTM)-时序模式注意力机制(temporal pattern attention,TPA)的预测方法。首先,将用电模式和天气作为影响因素,基于二阶聚类算法对区域内的负荷节点进行集群辨识,再从每个集群中挑选代表特征作为深度学习模型的输入,这样既能减少输入特征维度,降低计算复杂度,又能综合考虑预测区域的整体特征,提升预测精度。然后,针对区域电力负荷时序性的特点,用CNN-BiLSTM-TPA模型完成训练和预测,该模型能提取输入数据的双向信息生成隐状态矩阵,并对隐状态矩阵的重要特征加权,从多时间步上捕获双向时序信息用于预测。最后,在美国加利福尼亚州实例上分析验证了所提方法的有效性。

基于BIRCH聚类和递归神经网络的高铁强风预警算法

针对高铁临近风速预测需要克服数据的无周期性规律以及随机性较强的问题,构建了一种基于BIRCH聚类和LSTM递归神经网络算法的临近风速预测预警系统。该系统先做历史数据的交叉验证,然后用BIRCH进行在线增量聚类,最后根据聚类结果选取最接近当前预测时间序列的数据做LSTM的滚动训练并进行预测后得出预报预警结果,因此具有无需依赖数值预报产品以及随机数据适应性强的特点。实验证明,该系统的两种算法同时并行化在线运转,运行效率较高,预测效果较好,是解决强风预警问题的一种新方法。

挥发窑鼓风管的关键点识别及其摆放位置监测

窑头鼓风管的摆放位置是影响氧化锌挥发窑燃烧状态的重要操作参数之一,现仍由人工看火来手动调节,同时现场没有为挥发窑的运行优化记录完善的鼓风状态数据,也难以及时发现鼓风管被窑内渣块击中等安全隐患.针对上述问题,本文提出一种基于关键点识别的鼓风管位置监测方法.首先,文章对从窑头看火口采集的火焰视频数据集设计一种邻域关键点辅助的数据扩充方法,并构建级联金字塔网络(CPN)来预测鼓风管管口中心点的位置;然后,本文提出一种基于多帧图像的聚类分析算法来消除因烟尘遮挡所产生的异常点,并采用一种量化指标来实现对挥发窑鼓风管摆放位置的实时感知与记录;最后,本文基于现场采集的火焰视频数据进行了对比实验,结果表明所提出的关键点检测模型精度高、鲁棒性强,且鼓风管位置的量化准确率高达92.3%.

基于深度学习的集群系统故障预测方法

在面对集群系统故障预测时,长时间序列预测中存在因关键特征信息丢失而导致梯度消失或爆炸问题,从而影响了故障预测模型的准确性。基于此,提出一种新的基于深度学习的集群系统故障预测方法。该方法采用双向门控循环网络(bidirectional gate recurrent unit, BiGRU)来捕捉局部时序特征,同时采用Transformer来提高全局特征提取能力。通过BiGRU层中双向的信息传递获得集群系统日志上时序特征的动态变化,以获取集群事件中的潜在因果关系和局部时间特征,使用Transformer层并行处理BiGRU层输出的时间序列,得到全局的时间依赖性,继而由全连接神经网络层得到预测结果。通过由Blue Gene/L系统产生的真实日志所构建的公共数据集来验证方法的有效性,结果表明,所提方法优于对比方法,其最佳正确率和F1值分别达到91.69%和92.74%。