用Matlab中的Neural Network Toolbox仿真赤道东太平洋SST的预报模型

基于NCEP/NCAR再分析资料和COADS海洋资料中的全球月平均海平面气压场、850hPa纬向风场及海表温度场 ,利用Matlab中的NeuralNetworkToolbox仿真环境和BP模型改进算法比较准确地仿真和反演出了南方涛动指数、赤道纬向风指数和滞后的赤道东太平洋海温之间的动力结构和预报模型。该模型具有很好的拟合精度和可行的预报效果 ,可在一定时效内预测赤道东太平洋月平均海温的变化趋势。由于所建系统是具有直接因果关系的预报模型 ,因此不仅可直接用于预测 。

Output-Feedback Based Simplified Optimized Backstepping Control for Strict-Feedback Systems with Input and State Constraints

In this paper,an adaptive neural-network(NN)output feedback optimal control problem is studied for a class of strict-feedback nonlinear systems with unknown internal dynamics,input saturation and state constraints.Neural networks are used to approximate unknown internal dynamics and an adaptive NN state observer is developed to estimate immeasurable states.Under the framework of the backstepping design,by employing the actor-critic architecture and constructing the tan-type Barrier Lyapunov function(BLF),the virtual and actual optimal controllers are developed.In order to accomplish optimal control effectively,a simplified reinforcement learning(RL)algorithm is designed by deriving the updating laws from the negative gradient of a simple positive function,instead of employing existing optimal control methods.In addition,to ensure that all the signals in the closed-loop system are bounded and the output can follow the reference signal within a bounded error,all state variables are confined within their compact sets all times.Finally,a simulation example is given to illustrate the effectiveness of the proposed control strategy.

WEIGHTED PSEUDO ALMOST-PERIODIC SOLUTIONS OF SHUNTING INHIBITORY CELLULAR NEURAL NETWORKS WITH MIXED DELAYS

In this paper, we prove the existence and the global exponential stability of the unique weighted pseudo almost-periodic solution of shunting inhibitory cellular neural networks with mixed time-varying delays comprising different discrete and distributed time delays. Some sufficient conditions are given for the existence and the global exponential stability of the weighted pseudo almost-periodic solution by employing fixed point theorem and differential inequality techniques. The results of this paper complement the previously known ones. Finally, an illustrative example is given to demonstrate the effectiveness of our results.

Fault Diagnosis for a Diesel Valve Train Based on Time-Freq uency Analysis and Probabilistic Neural Networks

The cone-shaped kernel distributions of vibration acceleration signals, whichwere acquired from the cylinder head in eight different states of a valve train, were calculatedand displayed in grey images. Probabilistic Neural Networks ( PAW) was used to classify the imagesdirectly after the images were normalized. By this way, the problem of fault diagnosis for a valvetrain was transferred to the classification of time-frequency images. As there is no need to extractfeatures from time-frequency images before classification, the fault diagnosis process is highlysimplified. The experimental results show that the vibration signals can be classified accurately bythe proposed methods.

Global Exponential Stability Analysis of a Class of Dynamical Neural Networks

The problem of the global exponential stability of a class of Hopfield neural networks is considered. Based on nonnegative matrix theory, a sufficient condition for the existence, uniqueness and global exponential stability of the equilibrium point is presented. And the upper bound for the degree of exponential stability is given. Moreover, a simulation is given to show the effectiveness of the result.

Optimization of the End Effect of Hilbert-Huang transform(HHT)

In fault diagnosis of rotating machinery, Hil- bert-Huang transform （HHT） is often used to extract the fault characteristic signal and analyze decomposition results in time-frequency domain. However, end effect occurs in HHT, which leads to a series of problems such as modal aliasing and false IMF （Intrinsic Mode Func- tion）. To counter such problems in HHT, a new method is put forward to process signal by combining the general- ized regression neural network （GRNN） with the bound- ary local characteristic-scale continuation （BLCC）. Firstly, the improved EMD （Empirical Mode Decompo- sition） method is used to inhibit the end effect problem that appeared in conventional EMD. Secondly, the gen- erated IMF components are used in HHT. Simulation and measurement experiment for the cases of time domain, frequency domain and related parameters of Hilbert- Huang spectrum show that the method described here can restrain the end effect compared with the results obtained through mirror continuation, as the absolute percentage of the maximum mean of the beginning end point offset and the terminal point offset are reduced from 30.113% and 27.603% to 0.510% and 6.039% respectively, thus reducing the modal aliasing, and eliminating the false IMF components of HHT. The proposed method caneffectively inhibit end effect, reduce modal aliasing and false IMF components, and show the real structure of signal components accuratelX.

Study on Process Parameters Optimization of Sheet Metal Forming Based on PFEA/ANN/GA

Sheet metal forming is widely applied to automobile, aviation, space flight, ship, instrument, and appliance industries.In this paper, based on analyzing the shortcoming of general finite element analysis (FEA), the conception of parametric finite element analysis (PFEA) is presented. The parametric finite element analysis, artificial neural networks(ANN) and genetic algorithm (GA) are combined to research thoroughly on the problems of process parametersoptimization of sheet metal forming. The author programs the optimization scheme and applies it in a research ofoptimization problem of inside square hole flanging technological parameters. The optimization result coincides wellwith the result of experiment. The research shows that the optimization scheme offers a good new way in die designand sheet metal forming field.

Off-Line Signature Recognition Based on Angle Features and Artificial Neural Network Algorithm

Handwritten signature recognition is presented based on an angle feature vector by using the artificial neural network （ANN） in this research. Each signature image will be represented by an angle vector. The feature vector will constitute the input to the ANN. The collection of signature images is divided into two sets. One set will be used for training the ANN in a supervised fashion. The other set which is never seen by the ANN will be used for testing. After training, the ANN will be tested by recognizing the signatures. When a signature is classified correctly, it is considered correct recognition, otherwise it is a failure. The achieved recognition rate of this system is 94%.

Shallow Convolutional Neural Networks for Acoustic Scene Classification

Recently, deep neural networks, which include convolutional neural networks（CNNs）, have been widely applied to acoustic scene classification（ASC）. Motivated by the fact that some simplified CNNs have shown improvements over deep CNNs, such as Visual Geometry Group Net（VGG-Net）, we have figured out how to simplify the VGG-Net style architecture to a shallow CNN with improved performance. Max pooling and batch normalization are also applied for better accuracy. With a series of controlled tests on detection and classification of acoustic scenes and events（DCASE） 2016 data sets, our shallow CNN achieves 6.7% improvement, and reduces time complexity to 5%, compared with the VGG-Net style CNN.

On-Line Fast Motor Fault Diagnostics Based on Fuzzy Neural Networks

An on-line method was developed to improve diagnostic accuracy and speed for analyzing run- ning motors on site. On-line pre-measured data was used as the basis for constructing the membership functions used in a fuzzy neural network （FNN） as well as for network training to reduce the effects of various static factors, such as unbalanced input power and asymmetrical motor alignment, to increase accuracy. The preprocessed data and fuzzy logic were used to find the nonlinear mapping relationships between the data and the conclusions, The FNN was then constructed to carry motor fault diagnostics, which gives fast accurate diagnostics. The on-line fast motor fault diagnostics clearly indicate the fault type, location, and severity in running motors. This approach can also be extended to other applications.

Exponential distance distribution of connected neurons in simulations of two-dimensional in vitro neural network development

The distribution of the geometric distances of connected neurons is a practical factor underlying neural networks in the brain. It can affect the brain＇s dynamic properties at the ground level. Karbowski derived a power-law decay distribution that has not yet been verified by experiment. In this work, we check its validity using simulations with a phenomenological model. Based on the in vitro two- dimensional development of neural networks in culture vessels by Ito, we match the synapse number saturation time to obtain suitable parameters for the development process, then determine the distri-bution of distances between connected neurons under such conditions. Our simulations obtain a clear exponential distribution instead of a power-law one, which indicates that Karbowski＇s conclusion is invalid, at least for the case of in vitro neural network development in two-dimensional culture vessels.

Ultimate Strength Prediction of Carbon/Epoxy Tensile Specimens from Acoustic Emission Data

The objective of this paper was to predict the residual strength of post impacted carbon/epoxy composite laminates using an online acoustic emission （AE） monitoring and artificial neural networks （ANN）. The laminates were made from eight-layered carbon （in woven mat form） with epoxy as the binding medium by hand lay-up technique and cured at a pressure of 100 kg/cm2 under room temperature using a 30 ton capacity compression molding machine for 24 h. 21 tensile specimens （ASTM D3039 standard） were cut from the cross ply laminates. 16 specimens were subjected to impact load from three different heights using a Fractovis Plus drop impact tester. Both impacted and non-impacted specimens were subjected to uniaxial tension under the acoustic emission monitoring using a 100 kN FIE servo hydraulic universal testing machine. The dominant AE parameters such as counts, energy, duration, rise time and amplitude are recorded during monitoring. Cumulative counts corresponding to the amplitude ranges obtained during the tensile testing are used to train the network. This network can be used to predict the failure load of a similar specimen subjected to uniaxial tension under acoustic emission monitoring for certain percentage of the average failure load.

A Comparison of Three Kinds of Multimodel Ensemble Forecast Techniques Based on the TIGGE Data

Based on the ensemble mean outputs of the ensemble forecasts from the ECMWF （European Centre for Medium-Range Weather Forecasts）, JMA （Japan Meteorological Agency）, NCEP （National Centers for Environmental Prediction）, and UKMO （United Kingdom Met Office） in THORPEX （The Observing System Research and Predictability Experiment） Interactive Grand Global Ensemble （TIGGE） datasets, for the Northern Hemisphere （10~ 87.5~N, 0~ 360~） from i June 2007 to 31 August 2007, this study carried out multimodel ensemble forecasts of surface temperature and 500-hPa geopotential height, temperature and winds up to 168 h by using the bias-removed ensemble mean （BREM）, the multiple linear regression based superensemble （LRSUP）, and the neural network based superensemble （NNSUP） techniques for the forecast period from 8 to 31 August 2007. A running training period is used for BREM and LRSUP ensemble forecast techniques. It is found that BREM and LRSUP, at each grid point, have different optimal lengths of the training period. In general, the optimal training period for BREM is less than 30 days in most areas, while for LRSUP it is about 45 days.

Artificial Neural Network （ANN） and Regression Tree （CART） applications for the indirect estimation of unsaturated soil shear strength parameters

The shear strength parameters of soil （cohesion and angle of internal friction） are quite essential in solving many civil engineering problems. In order to determine these parameters, laboratory tests are used. The main objective of this work is to evaluate the potential of Artificial Neural Network （ANN） and Regression Tree （CART） techniques for the indirect estimation of these parameters. Four different models, considering different combinations of 6 inputs, such as gravel %, sand %, silt %, clay %, dry density, and plasticity index, were investigated to evaluate the degree of their effects on the prediction of shear parameters. A performance evaluation was carried out using Correlation Coefficient and Root Mean Squared Error measures. It was observed that for the prediction of friction angle, the performance of both the techniques is about the same. However, for the prediction of cohesion, the ANN technique performs better than the CART technique. It was further observed that the model considering all of the 6 input soil parameters is the most appropriate model for the prediction of shear parameters. Also, connection weight and bias analyses of the best neural network （i.e., 6/2/2） were attempted using Connec- tion Weight, Garson, and proposed Weight-bias approaches to characterize the influence of input variables on shear strength parameters. It was observed that the Connection Weight Approach provides the best overall methodology for accurately quantifying variable importance, and should be favored over the other approaches examined in this study.

Prediction of selected biodiesel fuel properties using artificial neural network

Biodiesel is an alternative fuel to replace fossil- based diesel fuel. It has fuel properties similar to diesel which are generally determined experimentally. The experimental determination of various properties of biodiesel is costly, time consuming and a tedious process. To solve these problems, artificial neural network （ANN） has been considered as a vital tool for estimating the fuel properties of biodiesel, especially from its fatty acid （FA） composition. In this study, four ANNs have been designed and trained to predict the cetane number （CN）, flash point （FP）, kinematic viscosity （KV） and density of biodiesel using ANN with logsig and purelin transfer functions in the hidden layer of all the networks. The five most prevalent FAs from 55 feedstocks found in the literature utilized as the input parameters for the model are palmitic, stearic, oleic, linoleic and linolenie acids except for density network with a sixth parameter （temperature）. Other FAs that are present in the biodiesels have been considered based on the number of carbon atom chains and the level of saturation. From this study, the prediction accuracy and the average absolute deviation of the networks are CN （96.69%; 1.637%）, KV （95.80%; 1.638%）, FP （99.07%; 0.997%） and density （99.40%; 0.101%）. These values are reasonably better compared to previous studies on empirical correlations and ANN predictions of these fuel properties found in literature. Hence, the present study demonstrates the ability of ANN model to predict fuel properties of biodiesel with high accuracy.

Practical compensation for nonlinear dynamic thrust measurement system

The real dynamic thrust measurement system usually tends to be nonlinear due to the complex characteristics of the rig, pipes connection, etc. For a real dynamic measuring system, the nonlinearity must be eliminated by some adequate methods. In this paper, a nonlinear model of dynamic thrust measurement system is established by using radial basis function neural network （RBF-NN）, where a novel multi-step force generator is designed to stimulate the nonlinearity of the system, and a practical compensation method for the measurement system using left inverse model is proposed. Left inverse model can be considered as a perfect dynamic compensation of the dynamic thrust measurement system, and in practice, it can be approximated by RBF-NN based on least mean square （LMS） algorithms. Different weights are set for producing the multi-step force, which is the ideal input signal of the nonlinear dynamic thrust measurement system. The validity of the compensation method depends on the engine＇s performance and the tolerance error 0.5%, which is commonly demanded in engineering. Results from simulations and experiments show that the practical compensation using left inverse model based on RBF-NN in dynamic thrust measuring system can yield high tracking accuracy than the conventional methods.