基于矩神经网络下的不变分布

给出了更新过程的一种扩散逼近.然后讨论了在更新输入的情况下,应用IF模型对神经元系统构建矩神经网络,包括神经元发放脉冲(Spike)的一阶和二阶统计量.在讨论了矩神经网络的不变分布的同时,通过模拟给出了不同情形下的不动点和边界曲线.

Prediction of Subcellular Localization of Eukaryotic Proteins Using Position-Specific Profiles and Neural Network with Weighted Inputs

Subcellular location is one of the key biological characteristics of proteins. Position-specific profiles （PSP） have been introduced as important characteristics of proteins in this article. In this study, to obtain position-specific profiles, the Position Specific lterative-Basic Local Alignment Search Tool （PSI-BLAST） has been used to search for protein sequences in a database. Position-specific scoring matrices are extracted from the profiles as one class of characteristics. Four-part amino acid compositions and lst-7th order dipeptide compositions have also been calculated as the other two classes of characteristics. Therefore, twelve characteristic vectors are extracted from each of the protein sequences. Next, the characteristic vectors are weighed by a simple weighing function and inputted into a BP neural network predictor named PSP-Weighted Neural Network （PSP-WNN）. The Levenberg-Marquardt algorithm is employed to adjust the weight matrices and thresholds during the network training instead of the error back propagation algorithm. With a jackknife test on the RH2427 dataset, PSP-WNN has achieved a higher overall prediction accuracy of 88.4% rather than the prediction results by the general BP neural network, Markov model, and fuzzy k-nearest neighbors algorithm on this dataset. In addition, the prediction performance of PSP-WNN has been evaluated with a five-fold cross validation test on the PK7579 dataset and the prediction results have been consistently better than those of the previous method on the basis of several support vector machines, using compositions of both amino acids and amino acid pairs. These results indicate that PSP-WNN is a powerful tool for subcellular localization prediction. At the end of the article, influences on prediction accuracy using different weighting proportions among three characteristic vector categories have been discussed. An appropriate proportion is considered by increasing the prediction accuracy.

Exponential stability criteria on neural networks with continuously distributed delays

The exponential stability of a class of neural networks with continuously distributed delays is investigated by employing a novel Lyapunov-Krasovskii functional. Through introducing some free-weighting matrices and the equivalent descriptor form, a delay-dependent stability criterion is established for the addressed systems. The condition is expressed in terms of a linear matrix inequality （LMI）, and it can be checked by resorting to the LMI in the Matlab toolbox. In addition, the proposed stability criteria do not require the monotonicity of the activation functions and the derivative of a time-varying delay being less than 1, which generalize and improve earlier methods. Finally, numerical examples are given to show the effectiveness of the obtained methods.

QUANTITATIVE APPROACH TO REAL TIME STARTUP TIME DETERMINATION OF DOWNSTREAM PHASES IN CONCURRENT ENGINEERING

Implementation of simultaneous execution phases in the concurrent engineering (CE) needs careful planning when the downstream phase could be activated as the upstream phase developed to a certain point. The determination of startup time of overlapping jobs in CE has long been a disturbance in manufacturing industry implementing CE programs. A novel model based on both fuzzy logic and neural network is proposed to mathematically formulate the inter-connective information between the two coupled phases in CE projects, and to determine the startup time of downstream phases in real time. The information transferring between the two coupled phases is quantified by using the negative Shannon entropy. Based on this algorithm, a PDM-based framework is proposed to narrow the gap between pro-duct design and manufacture, in which five modules are built to monitor, reshuffle and implement the simultaneous executing processes. Finally, an example is given to illustrate applications of the algorithm in the real world.

Absolute Exponential Stability of Generalized Dynamical Neural Networks

This paper investigates the absolute exponential stability of generalized neural networks with a general class of partially Lipschitz continuous and monotone increasing activation functions. The main obtained result is that if the interconnection matrix T of the neural system satisfies that - T is an H matrix with nonnegative diagonal elements, then the neural system is absolutely exponentially stable(AEST). The Hopfield network, Cellular neural network and Bidirectional associative memory network are special cases of the network model considered in this paper. So this work gives some improvements to the previous ones.

A Low Resolution Image Restoration Method based on BP Neural Network

In order to investigate the restoration of low resolution images, the linear and nonlinear interpolation methods were applied for the interpolation of the com- mon information matrix obtained from a series of pictures, getting the restructuring matrix. The characteristic block with the best restoration effect was determined by analyzing the pixel difference of the common information of each image at the same position. Then the characteristic blocks and their original blocks were used to build and train neural network. Finally, images were restored by the neural network and the differences between pictures were reduced. Experimental results showed that this method could significantly improve the efficiency and precision of algorithm.

Orthogonal moment based texture segmentation

Texture segmentation is a necessary step to identify the surface or an object in an image. We present a Legendre moment based segmentation algorithm. The Legendre moments in small local windows of the image are computed first and a nonlinear transducer is used to map the moments to texture features and these features are used to construct feature vectors used as input data. Then an RBF neural network is used to perform segmentation. A k-mean algorithm is used to train the hidden layers of the RBF neural network. The training of the output layer is the supervised algorithm based on LMS. The algorithm has been successfully used to segment a number of gray level texture images. Compared with the geometric moment-based texture segmentation, we can reduce the error rates using orthogonal moments.

A novel fully-integrated miniature six-axis force/torque sensor

This paper presents a new designed miniature six DOF （degree of freedom） force/torque sensor. This sensor is fully integrated with a micro DSP （digital signal processor）, so all the signal conditioning, A/D, decoupling, digital-signals serial output are performed in the sensor. Some experimental results are presented to demonstrate the capability of the proposed design. Finally, a neural network was used for decoupling the interacting signals, compared with the conventional method using the inverse matrix, this new method is more accurate.

An intelligent method for contact fatigue reliability analysis of spur gear under EHL

To complete the contact fatigue reliability analysis of spur gear under elastohydrodynamic lubrication(EHL) efficiently and accurately, an intelligent method is proposed. Oil film pressure is approximated using quadratic polynomial with intercrossing term and then mapped into the Hertz contact zone. Considering the randomness of the EHL, material properties and fatigue strength correction factors, the probabilistic reliability analysis model is established using artificial neural network(ANN). Genetic algorithm(GA) is employed to search the minimum reliability index and the design point by introducing an adjusting factor in penalty function. Reliability sensitivity analysis is completed based on the advanced first order second moment(AFOSM). Numerical example shows that the established probabilistic reliability analysis model could correctly reflect the effect of EHL on contact fatigue of spur gear, and the proposed intelligent method has an excellent global search capability as well as a highly efficient computing performance compared with the traditional Monte Carlo method(MCM).

Rotation, Scaling and Translation Invariant Blind Image Watermarking Scheme Utilizing Zernike Moments

A novel adaptive blind image watermarking scheme resistant to Rotation, scaling and translation (RST) attacks is proposed in this paper. Based on fuzzy clustering theory and Human visual system (HVS) model, the spread spectrum watermark is adaptively embedded in Discrete wavelet transform (DWT) domain. In order to register RST transform parameters, a hierarchical neural network is utilized to learn image geometric pattern represented by low order Zernike moments. Watermark extraction is carried out after watermarked image has been synchronized without original image. It only needs a trained neural network.Experiments show that it can embed more robust watermark under certain visual distance, effectively resist Joint photographic experts group (JPEG) compression, noise and RST attacks.

Automated visual inspection of surface defects based on compound moment invariants and support vector machine

The traditional inspection methods are mostly based on manual inspection which is very likely to make erroneous judgments due to personal subjectivity or eye fatigue, and can＇t satisfy the accuracy. To overcome these difficulties, we develop a machine vision inspection system. We first compare several kinds of methods for feature extraction and classification, and then present a real-time automated visual inspection system for copper strips surface （CSS） defects based on compound moment invariants and support vector machine （SVM）. The proposed method first processes images collected by hardware system, and then extracts feature characteristics based on grayscale characteristics and morphologic characteristics （Hu and Zernike compound moment invariants）. Finally, we use SVM to classify the CSS defects. Furthermore, performance comparisons among SVM, back propagation （BP） and radial basis function （RBF） neural networks have been involved. Experimental results show that the proposed approach achieves an accuracy of 95.8% in detecting CSS defects.

A new neural network model for the feedback stabilization of nonlinear systems

A new neural network model termed ‘standard neural network model’ (SNNM) is presented, and a state-feedback control law is then designed for the SNNM to stabilize the closed-loop system. The control design constraints are shown to be a set of linear matrix inequalities (LMIs), which can be easily solved by the MATLAB LMI Control Toolbox to determine the control law. Most recurrent neural networks (including the chaotic neural network) and nonlinear systems modeled by neural networks or Takagi and Sugeno (T-S) fuzzy models can be transformed into the SNNMs to be stabilization controllers synthesized in the framework of a unified SNNM. Finally, three numerical examples are provided to illustrate the design developed in this paper.

EOM-CCSD-Based Neural Network Diabatic Potential Energy Matrix for ^(1)πσ^(*)-Mediated Photodissociation of Thiophenol

A new diabatic potential energy matrix(PEM)of the coupled~^(1)ππ^(*)and~1πσ*states for the~1πσ*-mediated photodissociation of thiophenol was constructed using a neural network(NN)approach.The diabatization of the PEM was specifically achieved by our recent method[Chin.J.Chem.Phys.34,825(2021)],which was based on adiabatic energies without the associated costly derivative couplings.The equation of motion coupled cluster with single and double excitations(EOM-CCSD)method was employed to compute adiabatic energies of two excited states in this work due to its high accuracy,simplicity,and efficiency.The PEM includes three dimensionalities,namely the S-H stretch,C-S-H bend,and C-C-S-H torsional coordinates.The root mean square errors of the NN fitting for the S1 and S2 states are 0.89 and 1.33 me V,respectively,suggesting the high accuracy of the NN method as expected.The calculated lifetimes of the S1 vibronic 00 and31 states are found to be in reasonably good agreement with available theoretical and experimental results,which validates the new EOM-CCSD-based PEM fitted by the NN approach.The combination of the diabatization scheme solely based on the adiabatic energies and the use of EOM-CCSD method makes the construction of reliable diabatic PEM quite simple and efficient.

Seismic Health Monitoring of Foundations Using Artificial Neural Networks

Damage identification plays an important role in structural health monitoring systems. Despite variety in damage identification methods, little attention has been paid to the seismic damage identification of foundations. When shear walls serve as the lateral load resistance system of structures, foundations may subject to the high level of concentrated moment and shear forces. Consequently, they can experience severe damage. Since such damage is often internal and not visible, visual inspections cannot identify the location and the severity of damage. Therefore, a robust method is required for damage localization and quantification of foundations. According to the concept of performance-based seismic design of structures, the seismic behavior of foundations is considered as Force-Controlled. Therefore, for damage identification of foundation, internal forces should be estimated during ground motions. In this study, for real-time seismic damage detection of foundations, a method based on artificial neural networks was proposed. A feed-forward multilayer neural network with one hidden layer was selected to map input samples to output parameters. The lateral displacements of stories were considered as the input parameters of the neural network while moment and shear force demands at critical points of foundations were taken into account as the output parameters. In order to prepare well-distributed data sets for training the neural network, several nonlinear time history analyses were carried out. The proposed method was tested on the foundation of a five-story concrete shear wall building. The obtained results revealed that the proposed method was successfully estimated moment and shear force demands at the critical points of the foundation.

CONSTRUCTION OF THE ENCRYPTION MATRIX BASED ON CHEBYSHEV CHAOTIC NEURAL NETWORKS

The paper proposes a novel algorithm to get the encryption matrix. Firstly, a chaotic sequence generated by Chebyshev chaotic neural networks is converted into a series of low-order integer matrices from which available encryption matrices are selected. Then, a higher order encryption matrix relating real world application is constructed by means of tensor production method based on selected encryption matrices. The results show that the proposed algorithm can produce a "one-time pad cipher" encryption matrix with high security; and the encryption results have good chaos and auto-correlation with the natural frequency of the plaintext being hidden and homogenized.

Dynamics Analysis for Generic Projection Continuous-Time RNNs with Bounded Matrices

The dynamics anMysis of recurrent neural networks （RNNs） is a first and necessary step for any practical applications of them. In the present paper, the easily verified theorem is found to ascertain the asymptotical stability for generic RNN model with projection mapping under the critical condition that a discriminant matrix defined by the networks is semi-positive definite. The results given here not only improve deeply upon the existing relevant critical as well as non-critical dynamics conclusions in literature, but also can be used in the practical application of RNNs directly.

A forward-inverse dynamics modeling framework for human musculoskeletal multibody system

Multibody musculoskeletal modeling of human gait has been proved helpful in investigating the pathology of musculoskeletal disorders.However,conventional inverse dynamics methods rely on external force sensors and cannot capture the nonlinear muscle behaviors.Meanwhile,the forward dynamics approach is computationally demanding and only suited for relatively simple tasks.This study proposed an integrated simulation methodology to fulfill the requirements of estimating foot-ground reaction force,tendon elasticity,and muscle recruitment optimization.A hybrid motion capture system,which combines the marker-based infrared device and markerless tracking through deep convolutional neural networks,was developed to track lower limb movements.The foot-ground reaction forces were determined by a contact model for soft materials,and its parameters were estimated using a two-step optimization method.The muscle recruitment problem was first resolved via a static optimization algorithm,and the obtained muscle activations were used as initial values for further simulation.A torque tracking procedure was then performed by minimizing the errors of joint torques calculated by musculotendon equilibrium equations and inverse dynamics.The proposed approach was validated against the electromyography measurements of a healthy subject during gait.The simulation framework provides a robust way of predicting joint torques,musculotendon forces,and muscle activations,which can be beneficial for understanding the biomechanics of normal and pathological gait.

SZ-Derivations,PZ-Derivations and S-Derivations of a Matrix Algebra over Commutative Rings

Let R be a commutative ring with identity, Nn(R) the matrix algebra consisting of all n × n strictly upper triangular matrices over R with the usual product operation. An R-linear map φ : Nn(R) → Nn(R) is said to be an SZ-derivation of Nn(R) if x2 = 0 implies that φ(x)x+xφ(x) = 0. It is said to be an S-derivation of Nn(R) if φ(x2) = φ(x)x+xφ(x) for any x ∈ Nn(R). It is said to be a PZ-derivation of Nn(R) if xy = 0 implies that φ(x)y+xφ(y) = 0. In this paper, by constructing several types of standard SZ-derivations of Nn(R), we first characterize all SZ-derivations of Nn(R). Then, as its application, we determine all S-derivations and PZ- derivations of Nn(R), respectively.