基于神经网络和模糊逻辑的平台罗经故障检测

在基于神经网络的平台罗经故障检测中,为了提高故障检测灵敏度,根据船载平台罗经故障检测的特点,提出了以模糊逻辑和指数加权平均处理估计误差的故障检测方法,并用实船航行数据仿真.该方法对未知输入等干扰不敏感而对故障敏感,且可根据故障的大小自动调节检测时间的长短.对不易检测的小故障,自动延长检测时间以利用更多的信息从而提高检测的正确率;对手较大的故障,自动缩短检测时间从而减少检测延时和累积误差.     

基于相对动态误差的轴承故障特征参数提取北大核心CSCD

轴承故障信号中存在有关故障的异常信息,对维护机械安全有着重大意义。轴承故障信号经小波包分解之后,故障的异常信息主要体现在分解频段的动态误差上,而各个频段的动态误差一般由标准差能量熵和标准差均值来描述。为了凸显轴承故障的区分特征,通过轴承故障尺寸去刻度动态误差,利用相应的轴承故障特征参数提取相对动态误差,是有效的处理方法。基于此思路,本文针对小波包分解后不同频段分量的标准差,计算其能量熵以及均值。然后把对应频段的标准差能量熵和标准差均值相加作为特征参数,在同一尺度下定性分析。同时把轴承信号不同频段的特征参数相加后的数值与轴承故障尺寸相比,通过产生的相对动态误差进行定量分析,最终实现对轴承故障的有效区分。实验结果表明,本文所提方法对轴承故障有很好的区分效果。     

基于IPSO-LSTM神经网络的光伏发电系统故障检测

光伏发电系统的故障检测对光伏发电系统的安全运行至关重要。为提高光伏发电系统的故障检测效率，提出一种基于改进粒子群算法优化长短期记忆(IPSO-LSTM)神经网络的故障检测方法。首先通过构建改进的粒子群算法优化双层LSTM网络，对光伏发电系统的发电功率进行实时预测；然后，将LSTM网络预测的发电功率和系统实际的发电功率的误差作为残差值，当残差值大于设定的故障检测阈值时，可以确定系统发生故障。试验结果表明：改进粒子群算法优化的LSTM神经网络比传统的LSTM网络的故障检测性能更优越。     

改进的径向基函数神经网络及其应用

本文基于不同扩展中心的径向基函数神经网络（RBF）,提出了一个性能函数用于表示每个输入神经元对于网络输出精度影响的重要性,并据此对输入神经元进行自适应删减,将此方法应用于时间序列预测中,可自适应地选取最佳嵌入维数和时间延迟。对于股指预测的实验表明改进的RBF神经网络比未改进的RBF神经网络有更高的预测精度。     

平台罗经系统误差自补偿技术研究

利用计算机技术实补偿平台罗经系统的误差是本文的主要目的。首先建立了加速度计,陀螺,平台罗经的误差模型,主要讨论了怎样从测量误差,安装误差及陀螺非常值漂移误差中解算出平台罗经的系统误差,从而对加速度计的输出和平台罗经的姿态进行补偿,最后对本文提出的方法进行了仿真,给出了仿真曲经,结果表明,该补偿方法具有一定的效果。     

基于QPSO改进LSTM发动机怠速预测的FPID控制

以北京现代伊兰特G4GD发动机为试验台,将电控系统故障作为实验变量,测得规定时间内双传感器组合发生故障时的发动机怠速,并选原车ECU较难控制的6种组合怠速故障进行分析。基于量子粒子群算法（QPSO）对长短时记忆神经网络（LSTM）隐含层节点、训练次数与学习率进行寻优预测,将预测结果与多种神经网络进行对比,并通过均方根误差（RMSE）评价指标进行判断。使用Origin数据拟合将预测输出结果进行数值拟合,之后输入Matlab中使用Simulink搭建控制单元模型,由模糊常量-积分-微分（FPID）控制器对输出结果进行怠速控制。结果表明：基于量子粒子群算法改进的长短时记忆神经网络预测效果最好;模糊常量-积分-微分控制器对怠速的控制可有效缩短电子控制单元（ECU）的控制时间,无超调,且可有效调节至规定怠速。     

混沌时间序列预测方法

针对混沌时间序列预测问题,深入分析了基于径向基神经网络与基于记忆库两种预测方法,进而应用两种方法分别对Logistic序列、Kent序列进行预测,仿真结果表明:输入节点个数严重影响RBF神经网络预测性能,样本容量的增大与预测精度提高只是在一定范围内呈正比;选取系统的关联维数作为预测阶数时,基于记忆库预测方法的平均预测误差较小,且其预测误差随样本容量的增多而减小.     

蓝宝石晶体高温传感器的信号解调及标定

为了使用蓝宝石晶体高温传感器对工业烟气温度进行长期在线监测,设计了蓝宝石晶体高温传感器的解调系统,并对其进行了标定。对蓝宝石晶体高温传感器的白光偏振干涉测温原理进行了理论分析,并利用离散腔长变换(DGT)解调算法对光程差信息进行解调。在此基础上建立了一套以热电偶为参照的标定系统,使用S型高温热电偶采集温度数据,得到了光程差-温度样本。分别利用二次多项式拟合法与BP神经网络法对传感器的输出曲线进行了拟合与泛化,并进行了对比。实验结果表明:在800~1 300℃温度范围内,与二次多项式拟合方法相比,BP神经网络的拟合精度较高,拟合残差均值达到0.33℃;泛化能力强,多次泛化结果误差均值为0.56℃,均方误差为0.55℃。最终使用BP神经网络方法对传感器进行标定,使得传感解调系统满足了工业测高温的精度要求。     

基于深度神经网络(DNN)的压电陶瓷前馈补偿研究

针对压电陶瓷固有的迟滞非线性,设计了一种基于深度神经网络(DNN)的前馈补偿控制系统。该系统包含1个输入层、7个隐藏层和1个输出层。实验结果表明,开环情况下压电陶瓷的位移线性误差达8.91μm。施加神经网络前馈补偿后,压电陶瓷的最大位移误差降低到80 nm,稳态误差为±20 nm。进一步测试表明,在10～100 Hz输入频率下系统最大误差小于100 nm,均方根误差为0.01μm,验证了深度神经网络能够准确补偿压电陶瓷动态迟滞非线性,具有较好的频率泛化能力。     

基于情感神经网络的风电功率预测

风力发电功率预测对于风能并网具有重要意义.采用一种可用于复杂系统和模式建模的新型神经网络——情感神经网络,对风力发电功率进行预测.为防止ENN在训练时陷入局部最优解,提出采用遗传算法对其进行训练.采用预测误差的均方根和标准差衡量预测准确性、稳定性,对ENN性能进行了检验.结果表明,相比于人工神经网络、支持向量机和自滑动回归模型,ENN能够获得更高的预测准确率和预测可靠性.     

小波神经网络在短时交通流量预测中的应用

随着交通基础设施建设和智能运输系统的发展,交通规划和交通诱导成为交通领域的研究热点,对交通规划和交通诱导而言,准确的交通流量预测是其实现的前提和关键。短时交通流量预测是一个时间序列预测问题,文中应用小波神经网络对短时交通流量进行了预测。首先,对神经网络、小波分析等相关理论进行了简要介绍。在此基础上,采用5-7-1小波神经网络结构,以Morlet小波基函数作为隐含层节点的传递函数,将车流量数据输入该模型中,以训练小波神经网络,并用训练好的神经网络来预测短时交通流量。从预测结果来看,小波神经网络的预测结果较为准确,网络预测值接近期望值,效果较好。     

Futuristic speed prediction using auto‐regression and neural networks for mobile ad hoc networks

In this paper, we propose a speed prediction model using auto‐regressive integrated moving average (ARIMA) and neural networks for estimating the futuristic speed of the nodes in mobile ad hoc networks (MANETs). The speed prediction promotes the route discovery process for the selection of moderate mobility nodes to provide reliable routing. The ARIMA is a time‐series forecasting approach, which uses autocorrelations to predict the future speed of nodes. In the paper, the ARIMA model and recurrent neural network (RNN) trains the random waypoint mobility (RWM) dataset to forecast the mobility of the nodes. The proposed ARIMA model designs the prediction models through varying the delay terms and changing the numbers of hidden neuron in RNN. The Akaike information criterion (AIC), Bayesian information criterion (BIC), auto‐correlation function (ACF), and partial auto‐correlation function (PACF) parameters evaluate the predicted mobility dataset to estimate the model quality and reliability. The different scenarios of changing node speed evaluate the performance of prediction models. Performance results indicate that the ARIMA forecasted speed values almost match with the RWM observed speed values than RNN values. The graphs exhibit that the ARIMA predicted mobility values have lower error metrics such as mean square error (MSE), root MSE (RMSE), and mean absolute error (MAE) than RNN predictions. It yields higher futuristic speed prediction precision rate of 17% to 24% throughout the time series as compared with RNN. Further, the proposed model extensively compares with the existing works.     

Prediction of cyclosporine dosage in patients after kidney transplantation using neural networks

This paper proposes the use of neural networks for individualizing the dosage of cyclosporine A (CyA) in patients who have undergone kidney transplantation. Since the dosing of CyA usually requires intensive therapeutic drug monitoring, the accurate prediction of CyA blood concentrations would decrease the monitoring frequency and, thus, improve clinical outcomes. Thirty-two patients and different factors were studied to obtain the models. Three kinds of networks (multilayer perceptron, finite impulse response (FIR) network, and Elman recurrent network) and the formation of neural-network ensembles are used in a scheme of two chained models where the blood concentration predicted by the first model constitutes an input to the dosage prediction model. This approach is designed to aid in the process of clinical decision making. The FIR network, yielding root-mean-square errors (RMSEs) of 52.80 ng/mL and mean errors (MEs) of 0.18 ng/mL in validation (10 patients) showed the best blood concentration predictions and a committee of trained networks improved the results (RMSE = 46.97 ng/mL, ME = 0.091 ng/mL). The Elman network was the selected model for dosage prediction (RMSE = 0.27 mg/Kg/d, ME = 0.07 mg/Kg/d). However, in both cases, no statistical differences on the accuracy of neural methods were found. The models' robustness is also analyzed by evaluating their performance when noise is introduced at input nodes, and it results in a helpful test for models' selection. We conclude that neural networks can be used to predict both dose and blood concentrations of cyclosporine in steady-state. This novel approach has produced accurate and validated models to be used as decision-aid tools.     

人体头发中重金属元素铬含量的高光谱检测

头发中的重金属含量可以反映出人体健康的变化。提出了运用高光谱数据检测头发中重金属元素铬含量的方法。对头发的透射率波长曲线进行了包络线消除、吸收特征参量化等处理。以化学检测的铬含量作为标准数据,化学检测精度可达90%以上。然后训练人工神经网络,通过调节网络的隐含层层数、结点数和激活函数来优化模型。实验计算表明,隐含层层数为1,结点数为7或9的人工神经网络的预测效果较好。利用统计实验结果对人工神经网络的内部精度和外部精度进行评价。人体头发中铬的敏感波段为1380 nm～1550 nm、1880 nm～2100 nm、2120 nm～2210 nm;训练后的神经网络预测的均方根误差为13%,精度达87%。实验结果表明,应用高光谱技术可以快速无损地检测人体头发中的重金属元素铬的含量。     

A Flexible Analytical Markov Model for the IEEE 802.15.4 Unslotted Mechanism in Single-Hop Hierarchical Wireless Networks with Hidden Nodes

In this paper, an analytical Markov model (AMM) for the IEEE 802.15.4 unslotted CSMA/CA mechanism in single-hop hierarchical wireless networks (HWN) with hidden nodes is presented. The proposed AMM is flexible enough to operate both in saturation and non-saturation regions. Its prediction accuracy was compared with that of two well known models in the literature as well as with simulation results obtained by using the network simulator (NS)-2. The comparison has indicated that when $r$ , which denotes the maximum frame retransmission attempts, is set to zero and one-tier single-hop HWN with no hidden nodes is considered, the proposed AMM predicts the simulation results with a very small mean absolute percentage error (MAPE). Similarly, when a single-hop HWN with two tiers and hidden nodes is considered, then only one of the two well known models is applicable and the proposed AMM predictions present the smallest MAPE over the simulation results. Lastly, in the case of four-tier single-hop HWN with hidden nodes, only the proposed AMM is applicable and it predicts the simulation results at a respectable level of accuracy. When $r>0$ and the unslotted mechanism continues by incrementing the number of retransmission attempt by one upon channel access failure events, only the proposed AMM can imitate the behavior of the NS-2 IEEE 802.15.4 module.     

量子衍生神经网络的设计与实现

当使用神经网络解决问题时, 得到的结果与神经网络的逼近能力有很大关系。如何提高神经网络的逼近能力目前还没有较为理想的解决方法。本文提出了一种利用多位量子受控非门来构造神经网络模型的新方法。该模型为三层结构,隐层为量子神经元,输出层为普通神经元。量子神经元由量子旋转门和多位受控非门组成,利用多位受控非门中目标量子位的输出向输入端的反馈,实现对输入序列的整体记忆,利用多位受控非门的受控关系获得量子神经元的输出。基于量子计算原理设计了该模型的L M学习算法。该模型可从宽度和深度两方面获取输入序列的特征。纸牌预测的实验结果表明,当输入节点数和序列长度比较接近时,该模型对训练集的识别率比普通神经网络有大约8%的提高,从而揭示了量子计算机制对提高网络逼近能力的有效性。      

HBP: improvement in BP algorithm for an adaptive MLP decision feedback equalizer

Though the decision feedback equalizer (DFE) with multilayer perceptron (MLP) structure can be trained effectively by the backpropagation (BP) algorithm, it is always accompanied by the problem of local minimum. In order to solve some problems of the local minimum in the BP algorithm and to improve the performance of the BP algorithm under the same MLP structure, we combine the hierarchical approach and the BP algorithm to implement the MLP DFE, and we call the new scheme hierarchical BP (HBP) algorithm. Based on the hierarchical approach, from the input layer to the output layer of the MLP, every two layers of neural nodes (with one hidden layer) will be trained with an individual BP algorithm. Therefore, the entire MLP can be trained by several independent BP algorithms, unlike the standard BP algorithm, which utilizes only one BP algorithm to train the whole MLP structure. The results of performance evaluation indicate that the HBP algorithm not only strongly reduces the mean squared error but also yields a much lower bit-error rate than the standard BP algorithm does for equal computational cost and conditions.     

Multilayer feedforward neural networks with single powers-of-twoweights

A new algorithm for designing multilayer feedforward neural networks with single powers-of-two weights is presented. By applying this algorithm, the digital hardware implementation of such networks becomes easier as a result of the elimination of multipliers. This proposed algorithm consists of two stages. First, the network is trained by using the standard backpropagation algorithm. Weights are then quantized to single powers-of-two values, and weights and slopes of activation functions are adjusted adaptively to reduce the sum of squared output errors to a specified level. Simulation results indicate that the multilayer feedforward neural networks with single powers-of-two weights obtained using the proposed algorithm have generalization performance similar to that of the original networks with continuous weights