Generalization and the Backward Propagation Neural Network,

The internal stress difference between soft-ductile aluminum alloy substrate and hard-brittle Ni–W alloy coating will cause stress concentration, thus leading to the problem of poor bonding force. Herein, this work prepared the Ni–W graded coating on aluminum alloy matrix by the pulse electrodeposition method in order to solve the mechanical mismatch problem between substrate and coatings. More importantly, a backward propagation (BP) neural network was applied to efficiently optimize the pulse electrodeposition process of Ni–W graded coating. The SEM, EDS, XRD, Vickers hardness tester and Weighing scales are used to analyze the micromorphology, chemical element, phase composition, and micro hardness as well as oxidation weight increase, respectively. The results show that the optimal process conditions with BP neural network are as follows: the bath temperature is 30 °C, current density is 15 mA/cm2 and duty cycle is 0.3. The predicted value of the model agrees well with the experimental value curve, the relative error is minor. The maximum error is less than 3%, and the correlation coefficient is 0.9996. The Ni–W graded coating prepared by BP neural network shows good bonding with the substrate which has flat and smooth interface. The thickness of the coating is about 136 μm, which slows down the oxidation of the substrate and plays an effective role in protecting the substrate.

UAVs (unmanned aerial vehicles) have been widely used in many fields, where they need to be detected and controlled. Small-sample UAV recognition requires an effective detecting and recognition method. When identifying a UAV target using the backward propagation (BP) neural network, fully connected neurons of BP neural network and the high-dimensional input features will generate too many weights for training, induce complex network structure, and poor recognition performance. In this paper, a novel recognition method based on non-negative matrix factorization (NMF) with sparseness constraint feature dimension reduction and BP neural network is proposed for the above difficulties. The Edgeboxes are used for candidate regions and Log-Gabor features are extracted in candidate target regions. In order to avoid the complexity of the matrix operation with the high-dimensional Log-Gabor features, preprocessing for feature reduction by downsampling is adopted, which makes the NMF fast and the feature discriminative. The classifier is trained by neural network with the feature of dimension reduction. The experimental results show that the method is better than the traditional methods of dimension reduction, such as PCA (principal component analysis), FLD (Fisher linear discrimination), LPP (locality preserving projection), and KLPP (kernel locality preserving projection), and can identify the UAV target quickly and accurately.

An innovative efficient and fast neural networks in which hidden neurons are constructed based on Fourier series expansions (FSNN), half-range cosine (FCSNN) and sine expansions (FSSNN) are proposed and tested for linear and nonlinear regulation problems. The results of numerical examples using FSNN are compared with those obtained from traditional linear regression (LP), nonlinear regression (NLP), backward propagation neural networks (BPANN) and radial basis function neural networks (RBFNN). The results obtained from FSNN agree well with those obtained from LP, NLP, BPANN and RBFNN and show global approximation features to the fitting data. Only a few hidden neurons are required to obtain very good and fast convergence of regression as compared with BPANN and RBFNN.

Conventional retrieval and neural network methods are used simultaneously to retrieve sea surface wind speed (SSWS) from HH-polarized Sentinel-1 (S1) SAR images. The Polarization Ratio (PR) models combined with the CMOD5.N Geophysical Model Function (GMF) is used for SSWS retrieval from the HH-polarized SAR data. We compared different PR models developed based on previous C-band SAR data in HH-polarization for their applications to the S1 SAR data. The recently proposed CMODH, i.e., retrieving SSWS directly from the HH-polarized S1 data is also validated. The results indicate that the CMODH model performs better than results achieved using the PR models. We proposed a neural network method based on the backward propagation (BP) neural network to retrieve SSWS from the S1 HH-polarized data. The SSWS retrieved using the BP neural network model agrees better with the buoy measurements and ASCAT dataset than the results achieved using the conventional methods. Compared to the buoy measurements, the bias, root mean square error (RMSE) and scatter index (SI) of wind speed retrieved by the BP neural network model are 0.10 m/s, 1.38 m/s and 19.85%, respectively, while compared to the ASCAT dataset the three parameters of training set are −0.01 m/s, 1.33 m/s and 15.10%, respectively. It is suggested that the BP neural network model has a potential application in retrieving SSWS from Sentinel-1 images acquired at HH-polarization.

Advanced control systems are increasingly employed for intelligent factories. Fuzzy logic control (FLC) and backward propagation neural network (BPNN) control are investigated in this paper to realize position control for a linear switched reluctance motor (LSRM) against its nonlinear characteristics. Principles for FLC and BPNN control are introduced elaborately. Simulation results via BPNN show that dynamic position errors for the LSRM can be limited to 0.1 mm. Experimental results on FLC suggest that point-to-point position tracking for the motor can achieve 0.01 mm, constraining dynamic position error in 0.1 mm. By experiments, FLC for the LSRM performs better than traditional proportional-integral-derivative (PID) control, proving the effectiveness of the alleviation of the nonlinearity for the LSRM.

With the rapid development of cryptography and network communication, random number is becoming more and more important in secure data communication. The nonlinearity of backward propagation neural network (BPNN) is used to improve the traditional random number generator (RNG). SHA-2 (512) hash function can ensure the unpredictability of the produced random numbers. So, a novel and secure RNG architecture is proposed in the presented paper, which is BPNN based on SHA-2 (512) hash function. The quality of random number generated by this proposed architecture can well satisfy the security of cryptographic system according to results of test suites standardized by the U.S. The proposed architecture can be used to improve performances such as power consumption, flexibility, cost and area in network security and security for cryptographic systems.

To improve the low-speed dynamic performance of induction motor in direct torque control (DTC), a novel method of stator resistance identification based on wavelet network (WN) is presented and the determination of wavelet network structure is discussed. The inputs of the WN are the current error and the change in the current error and the output of the WN is the stator resistance error. The improved least squares algorithm (LSA) is used to fulfill the network structure and parameter identification. By the use of wavelet transform that accurately localizes the characteristics of a signal both in the time and frequency domains, the occurring instants of the stator resistance change can be identified by the multi-scale representation of the signal. Once the instants are detected, the accurate stator flux vector and electromagnetic torque are acquired by the parameter estimator, which makes the DTC applicable in the low region, optimizing the inverter control strategy. By detailed comparison between the wavelet and the typical backward-propagation (BP) neural network, the simulation results show that the proposed method can efficiently reduce the torque ripple and current ripple, superior to the BP neural network.

This paper presents an approach to digit recognition using single layer neural network classifier with Principal Component Analysis (PCA). The handwritten digit recognition is an important area of research as there are so many applications which are using handwritten recognition and it can also be applied to new application. There are many algorithms applied to this computer vision problem and many more algorithms are continuously developed on this to make the handwritten recognition classify digits more accurately with less computation involved. The proposed model in this paper aims to reduce the features to reduce computation requirements and successfully classify the digit into 10 categories (0 to 9). The system designed consists of backward propagation (BP) neural network and is trained and tested on the MNIST dataset of handwritten digit. The proposed system was able to obtain 98.39% accuracy on the MNIST 10,000 test dataset. The Principal Component Analysis (PCA) is used for feature extraction to curtail the computational and training time and at the same time produce high accuracy. It was clearly observed that the training time is reduced by up to 80% depending on the number of principal component selected. We will consider not only the accuracy, but also the training time, recognition time and memory requirements for entire process. Further, we identified the digits which were misclassified by the algorithm. Finally, we generate our own test dataset and predict the labels using this system.

Clostridium difficile infection (CDI) is now the most common cause of healthcare-associated infections, with increasing prevalence, severity, and mortality of nosocomial and community-acquired CDI which makes up approximately one third of all CDI. There are also increased rates of asymptomatic colonization particularly in high-risk patients. C difficile is a known collagenase-producing bacteria which may contribute to anastomotic leak (AL). Machine learning-augmented multivariable regression and propensity score (PS)-modified analysis was performed in this nationally representative case-control study of CDI and anastomotic leak, mortality, and length of stay for colectomy patients using the ACS-NSQIP database. Among 46735 colectomy patients meeting study criteria, mean age was 61.7years (SD 14.38), 52.2% were woman, 72.5% were Caucasian, 1.5% developed CDI, 3.1% developed anastomotic leak, and 1.6% died. In machine learning (backward propagation neural network)-augmented multivariable regression, CDI significantly increases anastomotic leak (OR 2.39, 95% CI 1.70-3.36; P < .001), which is similar to the neural network results. Having CDI increased the independent likelihood of anastomotic leak by 3.8% to 6.8% overall, and in dose-dependent fashion with increasing ASA class to 4.3%, 5.7%, 7.6%, and 10.0%, respectively, for ASA class I to IV. In doubly robust augmented inverse probability weighted PS analysis, CDI significantly increases the likelihood of AL by 4.58% (95% CI 2.10-7.06; P < .001). This is the first known nationally representative study on CDI and AL, mortality, and length of stay among colectomy patients. Using advanced machine learning and PS analysis, we provide evidence that suggests CDI increases AL in a dose-dependent manner with increasing ASA Class.

A methodology is presented for on-line damage identification of discrete structural systems. The damage characteristic of the system first can be detected and then identified from the change of its dynamic properties through a backward propagation neural network. The neural network is constructed by three multilayer subnets that perform the tasks of input pattern generation, damage location identification, and damage severity determination, respectively. The methodology is demonstrated on two spring-mass systems

There are many constraints in a semiconductor wafer fabrication facility (fab), especially time-constraints on processes. This paper gives a dispatching rule considering time-constraints (DRTC) that makes use of such information as due date of a job, workload of a machine, occupation time of a job on a machine, and time constraint of a process. The relative importance of each category is determined by weighting parameters, which are set by learning simulation samples with a backward propagation neural network (BPNN). A real fab simulation model is used to demonstrate the proposed method. The simulation results indicate that DRTC is superior to common rules (such as FIFO, EDD, CR and SRPT ) and the existing dispatching rule used in the fab (called as PRIOR) with more movements of WIP and higher time-constraints satisfaction proportion.

Mapping the transmission risk of Zika virus using machine learning models

The authors report on the application of the three-layer, backward error propagation neural network to the problem of reading sequenced DNA autoradiograms. The network is used for band identification by extracting two features: band intensity level and band intensity gradient. A training set of 16000 12*12 gray scale patterns is generated. Trained with these patterns, the network successfully learned to identify the degree of presence and absence of these two low level features. >

The training process of a deep neural network commonly consists of three phases: forward propagation, backward propagation, and weight update. In this paper, we propose a hardware architecture to accelerate the backward propagation. Our approach applies to neural networks that use rectified linear unit. Considering that the backward propagation results in a zero activation gradient when the corresponding activation is zero, we can safely skip the gradient calculation. Based on this observation, we design an efficient hardware accelerator for training deep neural networks by selectively computing gradients. We show the effectiveness of our approach through experiments with various network models.

The paper presents an adaptive speed controller for permanent magnet DC motor using an artificial neural network (ANN). The development of an efficient training algorithm is one of the key problems in designing such ANN. The output error vector of the neural network is usually used in training, instead of the actual process output error. Since the desired control action is usually unknown, the output error of the ANN controller is also unknown. A simple on-line training algorithm, which enables the neural network to be trained with the actual output error of the controlled drive plant is used. The direction of the controlled plant output response is the only a priori knowledge needed. The ANN based controller is effective, robust, and results in high performance permanent magnet DC motor drives. >