Three-layer Feedforward Network Research Articles

An intelligent hybrid deep learning model for rolling bearing remaining useful life prediction

ABSTRACT Remaining Useful Life (RUL) prediction of rolling bearings is one of the intricate and important issues for equipment intelligent maintenance and health management. Various machine learning models and methods have been applied to rolling bearing RUL prediction. However, a single model cannot effectively extract state information and obtain accurate prediction results, and its generalisation is not stable under the condition of small sample data. Therefore this paper proposes an intelligent hybrid deep learning model for achieving accurate RUL prediction of rolling bearings. Firstly, the one-dimensional vibration signal is transformed into the corresponding two-dimensional time-frequency diagram via Continuous Wavelet Transform (CWT). Secondly, the diagram is input into a Multilayer Perceptron (MLP) consisting of a basic three-layer feed-forward network to obtain a one-dimensional feature vector. And lastly, the obtained feature vector is input into an integrated model based on Deep Autoregressive and Transformer to produce the probability distribution and obtain the prediction results of rolling bearing RUL. Extensive experiments on two rolling bearing datasets show that the proposed model outperforms six other comparative models in extracting bearing fault features and predicting bearing RUL, which demonstrates that the proposed model can effectively extract bearing fault features and accurately predict bearing remaining useful life.

A basic identification of late auditory evoked potentials at infrasound frequencies: Support via neural network–based signal processing

Concerns about potential health risks from infrasound exposure require objective investigation methods for the perception of infrasound by humans. Neural-network methods were applied to identify weak infrasound evoked potentials and to investigate the efficacy of these techniques. Two simple networks – a fully connected three-layer feed-forward network and a two-layer autoencoder – were designed for the analysis of electroencephalography (EEG) data recorded in response to infrasound stimuli presented. Late auditory evoked potentials (LAEPs) were detected and correlation factors between the target and the output of the networks were determined to assess the differentiation ability of the two networks. Both networks were able to differentiate infrasound evoked potentials from noise but showed different performance levels with respect to a differentiation threshold. The results can be applied to objectively identify perception events by infrasound in humans.

Numerical prediction of microstructure and hardness for low carbon steel wire Arc additive manufacturing components

The objective of this research is to define an optimized strategy for wire arc additive manufacturing (WAAM) process with a numerical tool. In this paper, a three-dimensional scaled steel plate model (120 mm × 24 mm × 6 mm) is built for numerical simulations to replace the expensive physical WAAM experiments. A series of thermo-metallo-mechanical analyses is conducted by using ABAQUS CAE user subroutines in which, thermal, metallurgical and mechanical models for gas metal arc welding (GMAW) process are implemented. Groups of variables in WAAM process namely wire melting current, voltage and torch moving speeds, as well as interval layer cooling coefficients, are chosen as input parameters of the artificial neural network (ANN). The two targeted values are hardness and ultimate strength of the steel plate model, which are obtained by fully coupled thermo-metallo-mechanical simulations or by quick results from commercial software JmatPro, and by real measurements on the test samples of WAAM experimental plates. In this primary study, Taguchi algorithm is used for data preparations of the numerical experiments. The tests are fulfilled with Neural Networks (NN) tool in MATLAB. A typical three-layer feedforward network is used for general works. The developed models show good prediction capability for hardness and ultimate strength. The ANN algorithm presents high potential in bidirectional modeling to develop the WAAM strategy tool. It is possible to realize the inverse modeling from properties obtained by fractions of phase volume to WAAM processing parameters.

Acid-activated Hibiscus sabdariffa seed pods biochar for the adsorption of Chloroquine phosphate: Prediction of adsorption efficiency via machine learning approach.

• Acid treated agrowaste was used for Chloroquine removal. • Maximum monolayer adsorption capacity for the system was 161.29 mg/g. • The mechanism of adsorption was predominantly chemisorption. • Freundlich and Pseudo Second Order models best described the adsorption system. • ANN model predicted the adsorption behaviour of CQ onto the prepared adsorbent. A constant increase in the release of pollutants into the environment poses a threat as this practice continually degrades water quality. One of the proffered solutions is the adsorption of pollutants onto adsorbents. The present study emphasized the performance of Hibiscus sabdariffa seed pods (HSP1) for removing Chloroquine phosphate (CQ) from aqueous media via adsorption. A machine learning tool (Artificial Neural Network, ANN) was used to predict the efficiency of CQ removal onto the prepared biochar. Different operational parameters were investigated for their influence and impact on CQ adsorption. HSP1 was also characterized using different physicochemical and spectrophotometric techniques. HSP1 remarkably displayed an outstanding performance for the elimination of CQ with a q max of 161.29 mg/g and percentage removal of 96.01%. The experimental data are well predicted by Pseudo Second Order (PSO) and Freundlich isotherm, indicating multilayer adsorption. The practicability and spontaneity of the adsorption process were confirmed through thermodynamic studies. A three-layer feedforward backpropagation network for the adsorption data was established using the Levenberg Marquardt training algorithm. The ANN model with structure 5-14-1, with Mean Square Error (8.01) and R 2 value (0.9823) was able to envisage good removal efficiency. It was concluded that the ANN model could predict the adsorption behavior of CQ on HSP1.

Fault prediction model on electrical power network using artificial neural network-based time series: A case study of Ayede-Eruwa/ Lanlate Feeder

Increase in size of electrical power network usually results in a rise in fault level and consequently in huge economic losses to energy providers and consumers in the distribution systems. Therefore, it is very important to be proactive in dealing with faults on distribution feeder systems not only to reduce financial havoc, but to save lives and improve the quality of life of the people. A case study of Ayede-Eruwa/Lanlate Oyo State, Nigeria 33kV line is considered. An Artificial Neural Network based Time Series (ANN-TS) fault predictive model is developed for forecasting of faults on the above chosen electrical power network. Daily forced outage readings of the substation’s feeders for three years were collected and modeled using a three-layer feed-forward network ANN-TS. The results in the frequency of fault prediction show that there is an overlap between the observed and predicted values. The annual Mean Average Percentage Error (MAPE) varies between 0.004% and 25%, and the feeders’ average MAPE ranges from 6% to 10%. The fault duration annual MAPE varies between 0.001% and 25.54% while the feeders’ average MAPE varies between 6% and 11%. The energy loss prediction follows the same trend with the annual MAPE alternating between 0.01% and 26.75%, andthe feeders’ average MAPE between 6% and 10%. The average overall MAPE of each feeder is between 6% and 10% which indicates that the developed model is about 90% to 94% accurate. Although, the model is designed for Ayede-Eruwa/Lanlate feeder, it could be utilized for effective prediction of faults in any power distribution network.

Layout Optimization of Urban Cultural Space Construction Based on Forward Three-Layer Neural Network Model.

The change of urban cultural space layout is a multi-variable, multi-objective, and restricted research process. The optimization of urban cultural space construction and layout is a multi-objective decision-making problem that needs to be solved urgently. Based on the forward three-layer neural network theory, this paper constructs an optimization model for the construction and layout of urban cultural space evaluation of the layout of cultural space. This paper first analyzes the feasibility of combining the forward three-layer neural network model with the optimization and adjustment of cultural space layout structure. Taking the three-layer feedforward network as an example, the structure optimization model based on the forward three-layer neural network is selected, and the established model is used to reflect the internal environment of the objective world. Structure and perform dynamic simulation. In the process of simulation modeling, from the aspects of system description, model structure, logical analysis, reasoning, and interpretation, two effective computer dynamic simulation methods, namely, forward three-layer neural network model and system dynamics SD model, were carried out for theoretical comparison and identification. The experimental results show that the feasibility and calculation error of the application of the optimization model are relatively good, reaching 0.897 and 6.21%, respectively. The number of newly added cultural spaces and the expansion speed show an increasing trend, expanding at an average annual speed of about 35 km2, effectively increasing the quality of regional planning and construction layout.

Polyaromatic hydrocarbons biodegradation using mix culture of microorganisms from sewage waste sludge: application of artificial neural network modelling.

In this study, we aimed to examine the tolerance of mixed culture of microorganisms isolated from sewage waste sludge to degrade high concentrations of polyaromatic hydrocarbons, naphthalene, and phenanthrene. The performance of the artificial neural network (ANN) model to predict and simulate the experimental biodegradation results was investigated. The mixed culture of microorganisms was isolated from sewage waste sludge and adopted to biodegrade naphthalene and phenanthrene at different concentrations (100-1000mg/L). Sewage waste sludge obtained from wastewater treatment plants. A three-layer feed-forward network with a sigmoid transfer function (logsig) at the hidden layer, a linear transfer function (purelin) at the output layer, and a backpropagation training algorithm was used to set the ANN model. The results of this study show that naphthalene at concentrations of 100, 300, 700, and 1000 mg/L was depleted after incubation with the mixed culture for 6, 8, 14, and 16 days, respectively. For phenanthrene, depletion of 100, 300, 600, and 1000 mg/L was achieved after 8, 11, 16, and 19 days of incubation, respectively. A high correlation coefficient of 99.5% between the predicted and the experimental results were obtained by using the AAN model. The results indicated that the mixed culture of microorganisms from sewage waste sludge could effectively consume naphthalene and phenanthrene as carbon and energy sources. Also, the ANN model could efficiently predict the experimental results for biodegradation treatment.

CLASSIFICATION OF BANANA RIPENING STAGES BY ARTIFICIAL NEURAL NETWORKS AS A FUNCTION OF PLANT PHYSICAL, PHYSICOCHEMICAL, AND BIOCHEMICAL PARAMETERS

Brazil is currently the 4th world’s largest banana producer, producing around 7 million tons. In this scenario, several studies have been developed with a large amount of data, such as climatic, morphological, and nutritional data, in an attempt to improve these numbers even further. This study aims to classify banana ripening stages by artificial neural networks (ANN) as a function of plant physical, physicochemical, and biochemical parameters. The used ANN consisted of a three-layer feedforward backpropagation network, with eight neurons in the input layer (physical, physicochemical, and biochemical parameters), ten neurons in the intermediate layer, and two neurons in the output layer (classification of banana ripening stages). The results showed three configurations. ANN presented an excellent result for the training phase, with 100% accuracy in the sample classification for the three configurations. The validation and testing phases, that is, the classification of samples that were not part of the training, showed 91.6% and 94.4% accuracy in the first and second configurations, respectively, and 89.5% accuracy in the third configuration.

Pattern Recognition for Data Retrieval using Artificial Neural Network

Data retrieval is an important aspect of data management. In this paper, we design an ANN to recognize the learned patterns. We use a three-layer feed-forward network for the training of patterns (bitmap data). We implement two kinds of recognition: forced recognition and custom-specific recognition. The ANN model developed recognizes the pattern even if there is variation in the applied test patterns from the learned/trained patterns. In particular, we discuss the fault tolerance offered by neural networks. The characteristic of fault tolerance depends upon the type of distribution taken from random numbers. The conventional network is based on the concept of memorization whereas the neural network is based on the concept of generalization.

Modeling of milk lactose removal by column adsorption using artificial neural networks: MLP and RBF

Artificial neural network (ANN) techniques are effective in modeling nonlinear processes, are simple to implement and require low computational time. In this work, the lactose adsorption process for continuous flow in a fixed-bed column with a molecularly imprinted polymer (MIP) adsorbent was modeled using an ANN technique. The neural models allowed predicting the relative lactose concentration (C/C0) from the interactions between the variables of contact time (min), temperature (?C), granulometry (mesh), bed height (cm) and flow rate (mL min-1). The ANN models were developed in MATLAB using multilayer perceptrons (MLP) and a radial basis function network (RBF). The MLP model was developed using a three-layer feed forward backpropagation network with 5, 8 and 4 neurons in the first, second and third layer, respectively. The function (RBF) network is also proposed and its performance is compared to a traditional network type. The best architecture configuration RBF model was developed using 5, 14 and 1 neurons in the first, second and third layer, respectively. The proposal of development of mathematical models applied to multi-component adsorption system for milk using these approaches is innovative. The resulting breakthrough curve models for lactose adsorption were in good agreement with the experimental results. Performance indices, such as R?, MSE, RMSE, SSE, MAE and RME were used to evaluate the reliabilities and accuracies of the models. A comparison between the ANN models shows the ability to predict the breakthrough curves of lactose removal in the milk adsorption process. Though, the MLP network model shows more accurately a higher correlation coefficient (R2 = 0.9751) and lower values for the obtained error indices. The accuracy of the model is confirmed by the comparison between the predicted and experimental data. The results showed that both neural models efficiently described the non-linear process of lactose adsorption in a fixed-bed column.

A versatile neuromorphic system based on simple neuron model

Brain-inspired neuromorphic computing has attracted much attention for its advanced computing concept. However, the massive hardware cost in fully-connected architectures makes it challenging to build a large-scale neuromorphic system. In this work, we report a compact, programmable, versatile, and scalable neuromorphic architecture. To demonstrate the concept of the neuromorphic architecture, a neuromorphic system consisting of four cores is implemented on an FPGA platform. On the one hand, the neuromorphic system is extremely compact and hardware-saving. The computing block based on a simple digital leaky Integrate-and-Fire (LIF) model only costs 69 logic elements (LEs); only one physical neuron is implemented in each core, and it can be reused as hundreds of virtual neurons by time-division-multiplexing; only four 9-bit synaptic weights are assigned to each neuron, which effectively alleviates the hardware explosion in fully-connected architecture. On the other hand, the neuromorphic system is programmable and versatile, and can perform different neural network computing. The neuromorphic system mapped with a three-layer feedforward network successfully recognizes the MNIST handwritten digits with an accuracy of 96.26%, and it also effectively realizes different convolution operations which are basic computing operations in convolutional neural networks. Last but not least, each neuromorphic core has its own router module, making it convenient to scale up.

Application of neural networks and fuzzy systems for the intelligent prediction of CO2-induced strength alteration of coal

CO2 sequestration and enhanced coal bed methane (ECBM) extraction necessitate CO2 injection into coal reservoirs that affect the coal strength properties and long-term integrity of the seam. Evaluation of CO2-induced coal strength alterations is essential to minimize the reservoir damage. Advanced soft computing models have become prevalent in rock mechanics field, as they are capable of learning trends from complex data sets, preserving the experience and using it for predictions. We present two models viz. artificial neural network (ANN) and adoptive neuro-fuzzy inference system (ANFIS) to predict the strength alterations of coal, under various CO2 saturation conditions. Model performances are compared with linear and non-linear multivariate regression analyses (L-MRA and NL-MRA). We consider three effective input parameters (i.e. coal type, CO2 saturation pressure and CO2 interaction time) and one output parameter (i.e. unconfined compressive strength (UCS)) in the models. ANN consists of a three-layer feed-forward back-propagation network with a 3-5-1 architecture and ANFIS consists of [4 4 4] Gaussian type membership functions. Model results confirm that ANFIS has the highest prediction capacity followed by ANN, with R2 equal to 0.9954 and 0.9933, respectively. Both L-MRA and NL-MRA prediction performances are not satisfactory, as R2 values are only 0.7854 and 0.7821 for two models, respectively. Thus, general statistical models like MRA fail to precisely predict the complex strength alterations. From the verified models, we show that well-trained ANN and ANFIS models can successfully fit and forecast the experimental data, and are able to predict the long-term CO2 saturation effect on coal strength.

A Core Method for the Weak Completion Semantics with Skeptical Abduction


 
 
 The Weak Completion Semantics is a novel cognitive theory which has been successfully applied to the suppression task, the selection task, syllogistic reasoning, the belief bias effect, spatial reasoning as well as reasoning with conditionals. It is based on logic programming with skeptical abduction. Each program admits a least model under the three-valued Lukasiewicz logic, which can be computed as the least fixed point of an appropriate semantic operator. The semantic operator can be represented by a three-layer feed-forward network using the core method. Its least fixed point is the unique stable state of a recursive network which is obtained from the three-layer feed-forward core by mapping the activation of the output layer back to the input layer. The recursive network is embedded into a novel network to compute skeptical abduction. This paper presents a fully connectionist realization of the Weak Completion Semantics.
 
 

Efficient coding matters in the organization of the early visual system

Individual areas in the brain are organized into a hierarchical network as a result of evolution. Previous work indicated that the receptive fields (RFs) of individual areas have been evolved to favor metabolically efficient neural codes. In this paper, we propose that not only the RFs of individual areas, but also the organization of adjacent neurons and the hierarchical structure composed of these areas have been evolved to support efficient coding. To verify this hypothesis, we introduce a feed-forward three-layer network to simulate the early stages of human visual system. We emphasize that the network is not a purely feed-forward one since it also includes intra-layer connections, which are essential but usually ignored in the literature. Simulation results strongly reveal that (1) the obtained RFs of the simulated retinal ganglion cells (RGCs) or neurons in the lateral geniculate nucleus (LGN) and V1 simple neurons are consistent to the neurophysiological data; (2) the responses of closer RGCs are more correlated, and V1 simple neurons with similar orientations prefer to cluster together; (3) the hierarchical organization of the early visual system is beneficial for saving energy, which accords with the requirement of metabolically efficient neural coding in the process of human brain evolution.

Higher Network Activity Induced by Tactile Compared to Electrical Stimulation of Leech Mechanoreceptors.

The tiny ensemble of neurons in the leech ganglion can discriminate the locations of touch stimuli on the skin as precisely as a human fingertip. The leech uses this ability to locally bend the body-wall away from the stimulus. It is assumed that a three-layered feedforward network of pressure mechanoreceptors, interneurons, and motor neurons controls this behavior. Most previous studies identified and characterized the local bend network based on electrical stimulation of a single pressure mechanoreceptor, which was sufficient to trigger the local bend response. Recent studies showed, however, that up to six mechanoreceptors of three types innervating the stimulated patch of skin carry information about both touch intensity and location simultaneously. Therefore, we hypothesized that interneurons involved in the local bend network might require the temporally concerted inputs from the population of mechanoreceptors representing tactile stimuli, to decode the tactile information and to provide appropriate synaptic inputs to the motor neurons. We examined the influence of current injection into a single mechanoreceptor on activity of postsynaptic interneurons in the network and compared it to responses of interneurons to skin stimulation with different pressure intensities. We used voltage-sensitive dye imaging to monitor the graded membrane potential changes of all visible cells on the ventral side of the ganglion. Our results showed that stimulation of a single mechanoreceptor activates several local bend interneurons, consistent with previous intracellular studies. Tactile skin stimulation, however, evoked a more pronounced, longer-lasting, stimulus intensity-dependent network dynamics involving more interneurons. We concluded that the underlying local bend network enables a non-linear processing of tactile information provided by population of mechanoreceptors. This task requires a more complex network structure than previously assumed, probably containing polysynaptic interneuron connections and feedback loops. This small, experimentally well-accessible neuronal system highlights the general importance of selecting adequate sensory stimulation to investigate the network dynamics in the context of natural behavior.

Thermal Performance of Flat Tube Louver Fin Heat Exchanger using Artificial Neural Network

In this paper, an application of artificial neural networks (ANNs) is presented to predict the outlet temperatures on both fluid side of flat tube louver fin type heat exchanger used in flat tube louver fin heat exchanger. A validated numerical code is developed using MATLAB to generate huge data sets which gives outlet temperature of both the fluids. Three-layer feed-forward back propagation neural network is developed to model the thermal performance of heat exchanger. The ANN is thereafter trained using the numerical data to predict the outlet temperatures in heat exchanger. ANN configurations are set by selecting the number of hidden layers and nodes to achieve error convergence. Different network configurations are also examined for searching a better network for prediction. Keywords: Artificial Neural Networks, louver fin, radiator, numerical code, compact heat exchanger Cite this Article Manish K. Rathod. Thermal Performance of Flat Tube Louver Fin Heat Exchanger using Artificial Neural Network. Journal of Thermal Engineering and Applications . 2017; 4(3): 6–12p.

Prediction of creep curve of HP40Nb steel using artificial neural network

Simulation of creep curves using data obtained from a limited number of short-time creep tests is helpful for predicting the long-time creep life of materials by extrapolation techniques. The present paper demonstrates the application of artificial neural network (ANN) for the prediction of creep curves for HP40Nb micro-alloyed steel. The network consists of stress, temperature and time as the input parameters and the creep strain as the output parameter. The data used are taken from accelerated creep tests carried out at constant temperatures in the range 650–1050 °C and constant stresses 47–120 MPa. The network was trained using a three-layer feed-forward back-propagation network, having a 15-neuron hidden layer, using the Levenberg–Marquardt optimization algorithm. After successful network training, the model was subjected to several tests to demonstrate consistent prediction capability. 98% of the creep strain data could be predicted within an error of ±10% deviation from the experimental values. An additional experiment carried out to check the prediction capability of the model confirms very good prediction capability, with a correlation coefficient of 0.994, by ANN modeling.

Analyzing the Characteristics of Ship-Generated Waves in an Inland-Restricted Channel Using an Artificial Neural Network (ANN)

Mao, L. and Chen, Y., 2018. Analyzing the characteristics of ship-generated waves in an inland-restricted channel using an artificial neural network (ANN).Many empirical formulas based on experimental data have been proposed using regression analysis to analyze characteristics of ship-generated waves for engineering applications. However, ship-generated waves have an inherent variability, which suggests that a traditional method such as regression analysis might be inadequate. This paper presents an alternative nonlinear method using an artificial neural network (ANN) for clarifying the characteristics of ship-generated waves in an inland-restricted channel. A three-layered feed-forward back-propagation network with six input variables, five hidden layer nodes, and one output variable is established to predict the maximum ship-generated wave height based on sets of data collected from field observations conducted in Xicheng Canal. Comparisons of the predicted results from the proposed model with observed data show that the maximum ship-generated wave height can be predicted more accurately using the ANN compared to the two empirical formulas. The proposed model was then used to analyze the quantitative relationship between maximum ship-generated wave height in Xicheng Canal and various impact factors. Maximum ship-generated wave heights for 500-ton and 800-ton ships are given for engineering design in Xicheng Canal. These results show that the ANN model is a useful tool for both the prediction and mechanistic analysis of ship-generated waves in an inland-restricted channel.

Encoding of Tactile Stimuli by Mechanoreceptors and Interneurons of the Medicinal Leech.

For many animals processing of tactile information is a crucial task in behavioral contexts like exploration, foraging, and stimulus avoidance. The leech, having infrequent access to food, developed an energy efficient reaction to tactile stimuli, avoiding unnecessary muscle movements: The local bend behavior moves only a small part of the body wall away from an object touching the skin, while the rest of the animal remains stationary. Amazingly, the precision of this localized behavioral response is similar to the spatial discrimination threshold of the human fingertip, although the leech skin is innervated by an order of magnitude fewer mechanoreceptors and each midbody ganglion contains only 400 individually identified neurons in total. Prior studies suggested that this behavior is controlled by a three-layered feed-forward network, consisting of four mechanoreceptors (P cells), approximately 20 interneurons and 10 individually characterized motor neurons, all of which encode tactile stimulus location by overlapping, symmetrical tuning curves. Additionally, encoding of mechanical force was attributed to three types of mechanoreceptors reacting to distinct intensity ranges: T cells for touch, P cells for pressure, and N cells for strong, noxious skin stimulation. In this study, we provide evidences that tactile stimulus encoding in the leech is more complex than previously thought. Combined electrophysiological, anatomical, and voltage sensitive dye approaches indicate that P and T cells both play a major role in tactile information processing resulting in local bending. Our results indicate that tactile encoding neither relies on distinct force intensity ranges of different cell types, nor location encoding is restricted to spike count tuning. Instead, we propose that P and T cells form a mixed type population, which simultaneously employs temporal response features and spike counts for multiplexed encoding of touch location and force intensity. This hypothesis is supported by our finding that previously identified local bend interneurons receive input from both P and T cells. Some of these interneurons seem to integrate mechanoreceptor inputs, while others appear to use temporal response cues, presumably acting as coincidence detectors. Further voltage sensitive dye studies can test these hypotheses how a tiny nervous system performs highly precise stimulus processing.

Modelling and Optimization of Homogenous Photo-Fenton Degradation of Rhodamine B by Response Surface Methodology and Artificial Neural Network

The predictive ability of response surface methodology (RSM) and artificial neural network (ANN) in the modelling of photo-Fenton degradation of Rhodamine B (Rh-B) was investigated in the present study. The dye degradation was studied with respect to four factors viz., initial concentration of dye, concentration of H2O2 and Fe2+ ions and process time. Central composite design (CCD) was used to evaluate the effect of four factors and a second order regression model was obtained. The optimum degradation of 99.84% Rh-B was obtained when 159 ppm dye, 239 ppm H2O2, 46 ppm Fe2+ were treated for 27 min. The independent variables were fed as inputs to ANN with the percentage dye degradation as outputs. For the optimum percentage dye degradation, a three-layered feed-forward network was trained by Levenberg-Marquardt (LM) algorithm and the optimized topology of 4:10:1 (input neurons: hidden neurons: output neurons) was developed. A high regression coefficient (R2 = 0.9861) suggested that the developed ANN model was more accurate and predicted in a better way than the regression model given by RSM (R2 = 0.9112).