Forward Artificial Neural Network Research Articles

A Bidirectional Deep Learning Approach for Designing MEMS Sensors

To achieve the desired characteristics for MEMS sensors, traditional design process obtains the geometrical parameters based on complex theoretical calculations and interactive finite element (FE) simulations, which are time-consuming and data-consuming. To solve the above problems, a data-driven bidirectional design approach based on deep learning (DL) method is introduced to improve the design efficiency of MEMS sensors in this work. By using the piezoresistive acceleration sensor as a design example, the forward artificial neural network (ANN) with the sensor geometrical parameters as the input and the sensor performance as the output is trained and realized by using 1000 groups of data collected through FE simulation. This forward ANN can accurately predict the sensor performance including the measurement range, sensitivity, and resonant frequency. In addition, the inverse ANN with the sensor performance as the input and the sensor geometrical parameters as the output is also achieved by using a tandem network. This inverse ANN can provide the geometrical parameters directly and instantly according to the target performance. Both the forward and inverse networks cost only about 6 ms for each task and the mean relative errors are less than 3%. The high efficiency and low relative error indicate that DL is a promising approach to improve the design efficiency for MEMS sensors.

An Artificial Neural Network Method for High-Accurate and High-Efficient MEMS Pressure Sensor Design

Microelectromechanical system (MEMS) devices have numerous advantages including small sizes, high performance, and easy integration capability and thus have been widely used in the Internet of Things (IoT). A typical MEMS device usually includes a set of performance parameters, and each parameter is sensitive to the device geometries but with different regularities and weights, thus resulting in the complexity of MEMS device design. The conventional design method is mainly based on iterative finite-element (FE) simulation and optimization, which is time-consuming and inefficient. To address the above issues, a bidirectional artificial neural network (ANN)-based method is explored and used as the design method by using an MEMS pressure sensor as a design example. First, a forward ANN with the geometries and performance as the input and output, respectively, is trained and constructed, which can accurately predict the performance. Then, an inverse ANN with the performance and geometries as the input and output, respectively, is also investigated. By means of a tandem network, the nonuniqueness issue of the inverse ANN caused by a one-to-many response from the input to the output can be well addressed. This tandem network can output the corresponding geometries instantly according to the target performance. This work demonstrates the great potential of the ANN as a new and facile strategy in MEMS device design.

Identification of delamination severity in a tapered FRP composite plate

The typical failure in fibre-reinforced polymer (FRP) laminated composites is delamination. It reduces structural stiffness and changes in dynamic responses, which can be successfully used for assessing structural health monitoring (SHM) of composite structures. SHM often requires the solution of an inverse problem. This study focuses on solving the inverse problem with minimal computational time to predict delamination size, in-plane location, and interface. For the inverse model, two computational intelligence algorithms, the surrogate-assisted real-coded genetic algorithm (SARGA) and an artificial neural network (ANN) were used. A surrogate model is a fast-executing model for predicting the shift in the natural frequency using a forward ANN. For ANN training, the database size was optimized. The layerwise theory was solved using the finite element method for a laminated tapered FRP plate with ply drop-off, and ‘constrained mode’ is used to model the delamination. The experimental modal analysis was conducted on a glass-fibre reinforced polymer tapered composite plate with one intact plate and three artificially delaminated plates. The algorithms are validated using numerically simulated and experimentally measured frequencies. The numerical validation showed a better prediction of delamination size, in-plane location, and interface when compared with the experimental validation. Nevertheless, SARGA can reasonably predict the size, in-plane (x) location and interface of the delamination with an average overlapping ratio of 72.89%. The comparison of algorithms demonstrates the application of SARGA with the frequency-based approach for delamination detection and real-time condition monitoring in real-world applications.

Electrochemical evaluation of an Acanthocereus tetragonus aqueous extract on aluminum in NaCl (0.6 M) and HCl (1 M) and its modelling using forward and inverse artificial neural networks

• Acanthocereus tetragonus aqueous extract was evaluated in acidic and neutral media as a corrosion inhibitor. • Forward ANNs were trained with two Aluminum electrochemical evaluations databases. • Larger databases were calculated from Forward ANNs. • Inverse ANNs (IANNs) were trained with the large databases. • Time and concentration of the Acanthocereus tetragonus extract were found from optimal IANNs. An innovative numerical method based on a machine learning approach is presented in order to model the electrochemical behaviour of an Acanthocereus tetragonus aqueous extract on Aluminum in acidic and neutral media. Experimental data of an electrochemical evaluation of Aluminum in HCl (1 M) and NaCl (0.6 M) were used to generate the training set for forward Artificial Neural Networks (ANN). Later, this nonlinear relationship is inverted and refined with the purpose of design and train an inverse ANN that solves the following inverse problem: to find the concentration of a green corrosion inhibitor and the exposure time as a function of pH values, real and imaginary impedance values, and a frequency range of measure between 10,000 and 0.01 Hz.

Survey on Sign Language Detection Application

Neural networks, as its name suggests, is a machine learning technique which is modeled after the brain structure. It comprises of a network of learning units called neurons. These neurons learn how to convert input signals (e.g. picture of a cat) into corresponding output signals (e.g. the label “cat”), forming the basis of automated recognition. A convolutional neural network (CNN, or ConvNet) is a type of feed¬forward artificial neural network in which the connectivity pattern between its neurons is inspired by the organization of the animal visual cortex.

MuscleNET: mapping electromyography to kinematic and dynamic biomechanical variables by machine learning

Objective. This paper proposes machine learning models for mapping surface electromyography (sEMG) signals to regression of joint angle, joint velocity, joint acceleration, joint torque, and activation torque. Approach. The regression models, collectively known as MuscleNET, take one of four forms: ANN (forward artificial neural network), RNN (recurrent neural network), CNN (convolutional neural network), and RCNN (recurrent convolutional neural network). Inspired by conventional biomechanical muscle models, delayed kinematic signals were used along with sEMG signals as the machine learning model’s input; specifically, the CNN and RCNN were modeled with novel configurations for these input conditions. The models’ inputs contain either raw or filtered sEMG signals, which allowed evaluation of the filtering capabilities of the models. The models were trained using human experimental data and evaluated with different individual data. Main results. Results were compared in terms of regression error (using the root-mean-square) and model computation delay. The results indicate that the RNN (with filtered sEMG signals) and RCNN (with raw sEMG signals) models, both with delayed kinematic data, can extract underlying motor control information (such as joint activation torque or joint angle) from sEMG signals in pick-and-place tasks. The CNNs and RCNNs were able to filter raw sEMG signals. Significance. All forms of MuscleNET were found to map sEMG signals within 2 ms, fast enough for real-time applications such as the control of exoskeletons or active prostheses. The RNN model with filtered sEMG and delayed kinematic signals is particularly appropriate for applications in musculoskeletal simulation and biomechatronic device control.

Welding parameters prediction for arbitrary layer height in robotic wire and arc additive manufacturing

In wire and arc additive manufacturing, the weld bead geometry determined the slicing layer height, which was decided by the welding parameters. Generally, the determination of the welding parameters relied on empirical and experimental data through the trial-and-error methods that incur considerable time and cost. To obtain the proper welding process parameters according to the desired single bead geometry and layer height, a full factorial experimental design matrix was applied to collect the original data of welding parameters and bead geometrical variables. A forward artificial neural network (FANN) was built to predict the bead geometry form the welding parameters. Then, a closed-loop iteration method combined a genetic algorithm (GA) and the FANN model (FANN-GA) was developed to search for the most optimal welding process parameters in accordance with the selected bead geometrical variables. The results confirmed that the FANN-GA model has a good performance on the backward prediction of the welding process parameters compared with the direct backward artificial neural network (BANN). Several groups of single layer multi-bead and multi-layer multi-bead experiment were performed to testify the proposed method, and the relative error between the desired and actual layer height was small. The proposed method makes it possible to fabricate the component with an arbitrary desired layer height, and could be used in the adaptive slicing additive manufacturing or surface coating.

ON-Line MRI Image Selection and Tumor Classification using Artificial Neural Network

When soft tissue planning is important, usually, the Magnetic Resonance Imaging (MRI) is a medical imaging technique of selection. In this work, we show a modern method for automated diagnosis depending on a magnetic resonance images classification of the MRI. The presented technique has two main stages; features extraction and classification. We obtained the features corresponding to MRI images implementing Discrete Wavelet Transformation (DWT), inverse and forward, and textural properties, like rotation invariant texture features based on Gabor filtering, and evaluate the meaning of every property in the classification. The classifier is according to Feed Forward Back Propagation Artificial Neural Network (FP-ANN) in the classification stage. The properties thereafter derived to be implemented to teach a neural network based binary classifier that will be automatically able to conclude whether the image is that of a pathological, suffering from brain lesion, or a normal brain. The proposed algorithm obtained the sensitivity of 97.50%, specificity of 82.86% and accuracy of 94.3% for clinical Brain MRI database. This outcome proofs that the presented algorithm is robust and effective compared with other recent techniques.

Analysis of Temperature Profiles in Longitudinal Fin Designs by a Novel Neuroevolutionary Approach

Real application problems in physics, engineering, economics, and other disciplines are often modeled as differential equations. Classical numerical techniques are computationally expensive when we require solutions to our mathematical problems with no prior information. Hence, researchers are more interested in developing numerical methods that can obtain better solutions with fewer efforts and computational time. Heuristic algorithms are considered suitable candidates for such type of problems. In this research, we have developed a new neuroevolutionary algorithm that combines the power of feed- forward artificial neural networks (ANNs) and a modern metaheuristic, the Symbiotic Organism Search (SOS) algorithm. With our new approach, we have analyzed the simultaneous surface convection and radiation during heat transfer in different models of fins/heat exchangers. Longitudinal fins are considered with concave parabolic, rectangular and trapezoidal shapes. We have analyzed our problem in two scenarios and six sub-cases. Our solutions are of high quality, with minimum residual errors in all cases. We have established the quality of our results by calculating values of different performance indicators like Root-mean-square error (RMSE), Absolute error (AE), Generational distance (GD), Mean absolute deviation (MAD), Nash-Sutcliffe efficiency (NSE), Error in Nash-Sutcliffe efficiency (ENSE). Statistical and graphical analysis of our results suggests that our approach is suitable for handling real application problems. We have compared our results with state-of-the- art results, and the outcome of our analysis points to the superiority of our approach.

A Non-Polynomial, Non-Sigmoidal, Bounded and Symmetric Activation Function for Feed – Forward Artificial Neural Networks

Feed-forward artificial neural networks are universal approximators of continuous functions. This property enables the use of these networks to solve learning tasks. Learning tasks in this paradigm are cast as function approximation problems. The universal approximation results for these networks require at least one hidden layer with non-linear nodes, and also require that the non-linearities be non-polynomial in nature. In this paper a non-polynomial and non-sigmoidal non-linear function is proposed as a suitable activation function for these networks. The usefulness of the proposed activation function is shown on 12 function approximation task. The obtained results demonstrate that the proposed activation function outperforms the logistic / log-sigmoid and the hyperbolic tangent activation functions.

Photocatalytic Removal of RhB by Ag and Mg Co-Doped ZnO Nanoparticles: Modeling of Operational Parameters Using ANN Based on RSM Data

In this study, using response surface methodology (RSM), we focused on optimizing Rhodamine B (RhB) removal via Ag, Mg co-doped ZnO nanoparticles. Statistical analysis and lack of fit indicated that the model was adequate for optimizing the process. The optimum conditions were 240 mg L–1 photocatalyst dosage, 5 mg L–1 RhB concentration and 10 min irradiation time. A mathematical equation was created between RhB removal percent and effective operational parameters. The results calculated through RSM equation were used for modeling the process via artificial neural networks (ANN). Operational parameters such as photocatalyst dosage, initial RhB concentration and irradiation time were selected as the input variables. A three-layered feed forward back propagation ANN model was developed to process modeling. Data used in RSM design was simulated for testing the accuracy of the model generated by ANN. The comparison of the ANN predicted data and the experimental data revealed that the proposed method is significantly more effective in terms of the process modeling. The results obtained via ANN modeling indicated that the photocatalyst dosage with the relative importance of 49.61% can be considered as the most effective parameter for RhB photocatalytic removal.

Modelling of Kerf Width in Plasma Jet Metal Cutting Process using ANN Approach

In this paper Artificial Neural Network (ANN) model was developed for prediction of kerf width in plasma jet metal cutting process. Process parameters whose influence was analyzed are cutting height, cutting speed and arc current. An L18 (21x37) Taguchi orthogonal array experiment was conducted on aluminium sheet of 3 mm thickness. Using the experimental data a feed – forward backpropagation artificial neural network model was developed. After the prediction accuracy of the developed model was verified, the model was used to generate plots that show influence of process parameters and their interactions on analzyed kerf width and to get conlusions about process parameters values that lead to minimal kerf width.

Classification of Ultrasoud Thyroid Nodule using Feed Forward Neural Network

Presently, one of the main issues to create challenges in medicine sciences by developing technology is the disease diagnosis with high accuracy. An automatic system is developed that classifies the thyroid nodule images using machine learning algorithms. Ultrasound imaging is the best way to prediction of which type of thyroid is there. Ultrasound thyroid Nodules images were distinguishing in two groups Benign (non-cancerous) and Malignant (cancerous). Artificial Neural Networks are considered as the best solutions to achieve this goal and involve in widespread researches to diagnose the diseases. In this paper, we consider a Feed Forward Multi-layer Perception Artificial Neural Networks using back propagation learning algorithm to classify Thyroid disease.

A method of parameterising a feed forward multi-layered perceptron artificial neural network, with reference to South African financial markets

No analytic procedures currently exist for determining optimal artificial neural network structures and parameters for any given application. Traditionally, when artificial neural networks have been applied to financial modelling problems, structure and parameter choices are often made a priori without sufficient consideration of the effect of such choices. A key aim of this study is to develop a general method that could be used to construct artificial neural networks by exploring the model structure and parameter space so that informed decisions could be made relating to the model design. In this study, a formal approach is followed to determine suitable structures and parameters for a Feed Forward Multi-layered Perceptron artificial neural network with a Resilient Propagation learning algorithm with a single hidden layer. This approach is demonstrated through the modelling of four South African economic variables, namely the average monthly returns on the money, bond and equity markets as well as monthly inflation. Artificial neural networks can be constructed on the aforementioned variables in isolation or, jointly, in an integrated model. The performance of a range of more traditional time series models is compared with that of the artificial neural network models. The results suggest that, on a statistical level, artificial neural networks perform as well as time series models at forecasting the returns for financial markets. Hybrid models, combining artificial neural networks with the time series models, are constructed, trained and tested for the money market and for the rate of inflation. They appear to add value to the time series models when forecasting inflation, but not for the money market.

Application of Artificial Neural Networks for Evaluating Pressure Filtration of Coal Refuse Slurries

ABSTRACTBench-scale pressure filtration testing was performed to evaluate the dewatering characteristics of two coal refuse slurries, which were collected from thickener underflow streams of two coal preparation plants. Pressure filtration provides an opportunity to produce drier solids (filter cake) that can be stacked or mixed with the coarse refuse and improved water (filtrate) recovery that can be reused in the plant. This paper examines the effects of some of the major influencing variables such as pressure, slurry pH, feed solids concentration, fines fraction of solids in the slurry, filtration time, and temperature on dewatering thickener underflow slurry. Experimental results indicated that the overall filtrate flux increased with increase in pressure and temperature while it decreased with increase in fines fraction, pH, filtration time, and solids concentration. A total of 82 experimental results were used to develop a feed forward back-propagating artificial neural network (ANN) model. The model had R2 values over 0.9 for both the training and the testing datasets, indicating the goodness of fit. Sensitivity analysis performed using the ANN model indicated that filtration time and pH were the most significant variables influencing filtrate flux.

Combined inverse-forward artificial neural networks for fast and accurate estimation of the diffusion coefficients of cartilage based on multi-physics models

Analytical and numerical methods have been used to extract essential engineering parameters such as elastic modulus, Poisson׳s ratio, permeability and diffusion coefficient from experimental data in various types of biological tissues. The major limitation associated with analytical techniques is that they are often only applicable to problems with simplified assumptions. Numerical multi-physics methods, on the other hand, enable minimizing the simplified assumptions but require substantial computational expertise, which is not always available. In this paper, we propose a novel approach that combines inverse and forward artificial neural networks (ANNs) which enables fast and accurate estimation of the diffusion coefficient of cartilage without any need for computational modeling. In this approach, an inverse ANN is trained using our multi-zone biphasic-solute finite-bath computational model of diffusion in cartilage to estimate the diffusion coefficient of the various zones of cartilage given the concentration-time curves. Robust estimation of the diffusion coefficients, however, requires introducing certain levels of stochastic variations during the training process. Determining the required level of stochastic variation is performed by coupling the inverse ANN with a forward ANN that receives the diffusion coefficient as input and returns the concentration-time curve as output. Combined together, forward-inverse ANNs enable computationally inexperienced users to obtain accurate and fast estimation of the diffusion coefficients of cartilage zones. The diffusion coefficients estimated using the proposed approach are compared with those determined using direct scanning of the parameter space as the optimization approach. It has been shown that both approaches yield comparable results.

Predicting Failure Strength of Randomly Oriented Short Glass Fiber-Epoxy Composite Specimen by Artificial Neural Network Using Acoustic Emission Parameters

Acoustic emission (AE) peak amplitude and cumulative energy emitted during 50% of failure of composite specimen was collected, analyzed, and utilized to predict the ultimate tensile strength (UTS) using artificial neural network (ANN) and the performance of various training algorithm on prediction was analyzed. AE data have been collected from finite numbers of randomly oriented short glass fiber-epoxy tensile specimens, while loading up to failure in a tensile testing machine. AE response from each of the specimen was classified and segregated by understanding the failure mechanism. A feed forward back-propagation type ANN was designed and the segregated data of amplitude hits and cumulative energy was processed using two separate networks to predict the UTS of corresponding specimens using it with appropriate parameters and the results were analyzed.

Application of Back Propagation Neural Network Model for Predicting Flank Wear of Yttria Based Zirconia Toughened Alumina (ZTA) Ceramic Inserts

A back propagation neural network model has been adopted for the flank wear prediction of zirconia toughened alumina (ZTA) insert in turning operation. The experiments are performed on AISI 4340 steel using developed yttria based ZTA inserts. These inserts are prepared through wet chemical co-precipitation route followed by powder metallurgy process. Machining conditions such as cutting speed, feed rate and depth of cut are selected as input to the neural network model and flank wear of the inserts corresponding to these conditions has been chosen as the output of the network. The experimentally measured values are used to train the feed forward back propagation artificial neural network for prediction of those conditions. The convergence of the mean square error both in training and testing come out very well. The performance of the trained neural network has been validated with experimental data. The results demonstrate that the machining model is suitable and the optimization strategy satisfies practical requirements.

Implementation of Switched Beam Smart Antenna Using Artificial Neural Network

The capacity enhancement promised by switched beam smart antenna is a function of efficient selection of the active beam at each point in time. This article presents the use of artificial neural network (ANN) to improve the performance of switched beam smart antenna. The proposed method is based on feed forward backpropagation ANN. In this design, samples of the calibration of the footprint of the base station is mapped to the radiation pattern of Butler matrix (BM) base station antenna arrays. A $$4\times 4$$4×4 BM is designed to implement the developed technique and a digitizer applied to obtain the ANN training data. The results are compared with the existing numerical method and negative selection algorithm based on received signal strength indicator. It is shown that the proposed method exhibits a better switching performance.

Prediction of resonant frequency of a circular patch frequency selective structure using artificial neural network

In this paper, efforts have been given to develop a method which relates resonant frequency with the periodicity of a frequency selective structure comprising of two dimensional arrays of circular patches. Firstly resonant frequencies for various periodicities of the frequency selective surfaces have been calculated by method of moment after selecting proper basis function. Results obtained by this method have been used to train a feed forward back propagation artificial neural network with three hidden layers. Genetic algorithm based back propagation neural network and gradient decent based artificial neural networks are used in this process. From the trained network, resonant frequencies for given periodicities (even excluding the training data set) may be readily available. Now for the same combination of periodicity resonant frequency has been again calculated by method of moment. Some of these results are compared with the measured data for validation. Each time it is found that the results are in good parity.