Artificial Neural Network Configuration Research Articles

An artificial neural network to model response of a radiotherapy beam monitoring system.

The integral quality monitor (IQM) is a real-time radiotherapy beam monitoring system, which consists of a spatially sensitive large-area ion chamber, mounted at the collimator of the linear accelerator (linac), and a calculation algorithm to predict the detector signal for each beam segment. By comparing the measured and predicted signals the system validates the beam delivery. The current commercial version of IQM uses an analytic method to predict the signal, which requires a semi-empirical approach to determine and optimize various calculation parameters. The process of developing the calculation model is complex and time consuming, and moreover, the model cannot be easily generalized across various beam delivery platforms with different combinations of beam energy, beam flattening, beam shaping elements, and Linac models. Therefore, as an alternative solution, we investigated the feasibility of developing a machine learning (ML) method, using an artificial neural network (ANN), to predict the ion chamber signal. In developing an ANN, it is not necessary to explicitly account for each of the elements of beam interactions with various structures in the beam path to the ion chamber. The ANN was designed with multilayer perceptron (MLP). The input layer consisted of multiple features, derived from the geometrical characteristics of beam segments. Gradient descent error backpropagation technique was used to train the ANN. The combined training dataset included 270 rectangular fields, and 801 clinical IMRT fields delivered using 6 MV beams on Varian TrueBeamTM and Elekta InfinityTM . Each of 12 different ANN configurations (3 different sets of input features×4 different sets of number of hidden nodes) was simulated 10 times with randomly selected 80% of data for training and the remaining data for validation. Artificial neural networks with one hidden layer, consisting of 10 nodes, and 10 input features provided optimum results. Once the feature sets were extracted, the time required for the network training was on the order of a few minutes, and the time required to perform an output calculation per field was only fraction of a second. More than 95% of clinical intensity-modulated radiation therapy (IMRT) segments were calculated within±3.0% modeling error for Varian Truebeam (90% and ±3.3% for Elekta Infinity). A total of 3320 volumetric-modulated arc therapy (VMAT) segments from Truebeam were calculated using the ANN trained with IMRT fields. More than 95% of the cumulative VMAT beam segments were within 3.6% modeling error, similar to the performance for IMRT segments. In general the modeling error was found to be inversely proportional to the size and intensity of the beam segment. A prototype ANN has been developed for predicting the signals of the IQM system, with substantially less efforts compared to the analytic model. The performance of the ANN was found to be at least equivalent to that of the analytic method, in terms of average and maximum error, for 6 MV beams on both Varian TrueBeam and Elekta Infinity platforms.

Slow Degradation Fault Detection in a Harsh Environment

The ever increasing challenges posed by the science projects in astronomy have skyrocketed the complexity of the new generation telescopes. Due to the climate and sky requirements, these high-precision instruments are generally located in remote areas, suffering from the harsh environments around it. These modern telescopes not only produce massive amounts of scientific data, but they also generate an enormous amount of operational information. The Atacama Large Millimeter/submillimeter Array (ALMA) is one of these unique instruments, generating more than 50 Gb of operational data every day while functioning in conditions of extreme dryness and altitude. To maintain the array working under extreme conditions, the engineering teams must check over 130,000 monitoring points, combing through the massive datasets produced every day. To make this possible, predictive tools are needed to identify, hopefully beforehand, the occurrence of failures in all the different subsystems. This work presents a novel fault detection scheme for one of these subsystems, the Intermediate Frequency Processors (IFP). This subsystem is critical to process the information gathered by each antenna and communicate it, reliably, to the correlator for processing. Our approach is based on echo state networks, a configuration of artificial neural networks, used to learn and predict the signal patterns. These patterns are later compared to the actual signal, to identify failure modes. Additional preprocessing techniques were also added since the signal-to-noise ratio of the data used was very low. The proposed scheme was tested in over seven years of data from 132 IFPs at ALMA, showing an accuracy of over 70%. Furthermore, the detection was done several months earlier, on average, when compared to what human operators did. These results help the maintenance procedures, increasing reliability while reducing humans’ exposure to the harsh environment where the antennas are. Although applied to a specific fault, this technique is broad enough to be applied to other types of faults and settings.

Estimation of cerebral blood flow velocity during breath-hold challenge using artificial neural networks

The effect of untreated Obstructive Sleep Apnoea (OSA) on cerebral haemodynamics and CA impairment is an active field of research interest. A breath-hold challenge is usually used in clinical and research settings to simulate cardiovascular and cerebrovascular changes that mimic OSA events. This work utilises temporal arterial oxygen saturation (SpO2) and photoplethysmography (PPG) signals to estimate the temporal cerebral blood flow velocity (CBFv) waveform. Measurements of CBFv, SpO2, and PPG, were acquired concurrently from volunteers performing two different protocols of breath-hold challenge in the supine position. Past values of the SpO2 and PPG signals were used to estimate the current values of CBFv using different permutations and topologies of supervised learning with shallow artificial neural networks (ANNs). The measurements from one protocol were used to train the ANNs and find the optimum topologies, which in turn were tested using the other protocol. Data collected from 10 normotensive, healthy subjects (four females, age 28.5 ± 6.1 years, Body Mass Index (BMI) 24.0 ± 4.7 kg/m2) were used in this study. The results show that different subjects have different optimum topologies for ANNs, thus indicating the effects of inter-subject variability on ANNs. Successfully reconstructed blind waveforms for the same subject group in the second protocol showed a reasonable accuracy of 60–80% estimation compared to the measured waveforms. HypothesisTemporal waveforms for SpO2 and PPG contain adequate information to estimate the temporal CBFv waveform using ANNs. MethodologyConcurrent measurements of SpO2 and PPG using pulse oximetry from the forehead and CBFv from the middle cerebral artery (MCA) using transcranial Doppler (TCD) were recorded from healthy, normotensive subjects performing a breath-hold challenge. The breath-hold challenge mimicked the cerebrovascular response to apnoea, and was recorded by measuring CBFv in MCA. Two protocols were used, each consisting of five breath-holding manoeuvres and differing in terms of the time between the five successive breath-holds. Using data from one protocol, several permutations of the temporal values of SpO2 and PPG signals were used as inputs to different ANN topologies, in order to train and find the optimum model. The optimum model was evaluated using the data from the other protocol as a blind dataset. ResultsUsing the first protocol for training, optimum ANN configurations were found to be different for each subject, and accuracy of 75–87% was achieved. When these optimum ANN models were tested using the second protocol as a blind dataset, the accuracy achieved was around 60–80%. ConclusionsA novel approach employing temporal records of SpO2 and PPG can be used to estimate the CBFv waveform using ANNs with acceptable accuracy. Increases in the size and diversity of the population dataset and the use of features extracted from SpO2 and PPG signals are needed for generalisation of the method and potential future clinical applications.

Implementation of artificial neural network model for continuous hydrogen production using confectionery wastewater

In the present study, an artificial neural network (ANN) was implemented to estimate the hydrogen production from confectionery wastewater. From the experimental investigation, it could be concluded that maximum COD removal efficiency of 99% and hydrogen production rate of 6570 mL/d was achieved at 7.00 kg COD/m3d and 24 h HRT. To validate this, a back propagation ANN configuration of 4-12-4-2 was opted. The modelling was performed using the input parameters like time, influent chemical oxygen demand (COD), effluent pH and volatile fatty acids (VFA). The correlation coefficient between the experimental and predicted hydrogen production rate was 0.996. The result of the tested data for hydrogen production rate was successful. The calculated average percentage error (APE) for hydrogen production rate was 0.0004. As the APE values were closer to zero, the trained ANN model fitted well with the experimental data.

Dst Index Forecast Based on Ground‐Level Data Aided by Bio‐Inspired Algorithms

AbstractIn this study, different hybridized techniques that combine an artificial neural network (ANN) with bio‐inspired optimization algorithms such as particle swarm optimization (PSO), genetic algorithm (GA), and a hybridized PSO+GA were applied to update the ANN connection weights and so forecast the disturbance storm time (Dst) index. The past values of Dst index time series were used as input parameters to forecast its variation from 1 to 6 hours ahead. The database collected considers 233,760 hourly data from 01 January 1990 to 31 August 2016, containing storms and quiet period, grouped into three data sets: learning set (116,880 hourly data points), validation set (58,440 data points), and testing set (58,440 data points). Several ANN configurations were studied and optimized during the training process by evaluating the root mean square error (RMSE) and the correlation coefficient (R). An analysis of the predictive capability of each method was made year per year, and according to the levels of the geomagnetic storm. Also, an additional test was applied to the proposed method using 17 intense geomagnetic storms reported during solar cycle 24, including the St. Patrick's Day storm of 2015. Results show that the hybridized ANN+PSO method can forecast the Dst index quite accurately from 1 to 3 h in advance (with RMSE≤5 nT and R≥0.9), while the ANN+PSO+GA method can forecast the Dst index quite accurately from 4 to 6 h ahead (RMSE≤7 nT and R≥0.8)

Artificial neural networks can be used for Ambrosia pollen emission parameterization in COSMO-ART

The emission parameterization is a crucial part of numerical pollen dispersion models. This paper shows that Artificial Neural Networks (ANNs) can substantially improve the performance of the Ambrosia pollen emission in numerical pollen dispersion models such as COSMO-ART. Based on simultaneous measurement of Ambrosia pollen concentrations and meteorological variables in the source area, ANNs were trained to predict the diurnal profile of pollen emission. Six different combinations of explanatory meteorological variables were trained with five different ANN configurations resulting in 30 candidate emission models. The best network configuration for each combination of explanatory variables were used as emission parameterization in the numerical pollen dispersion model COSMO-ART. In addition, two benchmarks were implemented: an emission parameterization based on sigmoid functions and an artificial neural network using only time as an explanatory variable. The Ambrosia pollen seasons of 2015 and 2016 were simulated using the two benchmarks and the six emission parameterizations. The modelled diurnal profile of emission fluxes at 15 different sites from Serbia, Hungary and France with strong local pollen sources were compared with observed concentrations. Artificial Neural Networks based emission parameterization substantially improved the performance of the Ambrosia pollen emission in COSMO-ART compared to the emission based on the sigmoid functions in all these three countries. However, a time-related explanatory variable must be used. This suggests that the ANN-based emission parameterizations can be used at distant locations as well. On the other hand, the use of meteorological related parameters did not increase the performance compared with the time-only benchmark.

Improving feature location accuracy via paragraph vector tuning

Abstract Context Feature location techniques are still not highly accurate despite advances in the field. Objective This paper aims at investigating the impact of applying different tunings to paragraph vector to the feature location problem. It evaluates the influence of different artificial neural network (ANN) configurations for learning rate and negative sampling loss function in paragraph vectors training. Method The suggested weight configuration relies on the search for an adequate ANN learning rate and an adequate calibration of negative sampling skip-gram mode of the Doc2vec (DV) algorithm. A dataset with 633 feature descriptions, extracted from six open-source Java projects, organized within method granularity, is used for the empirical assessment. Results Our results suggest that feature location techniques benefit from the use of paragraph vector with systematic tuning. We show that an adequate update policy for ANN weights can increase feature location accuracy. An adequate calibration for negative sampling also improved accuracy. We got it with no default values of negative sampling pointed by literature. Moreover, an ensemble of learning rate policies and the use of a tuned DV negative sampling option had overcome state-of-the-art approaches. Conclusions We show evidence of a relationship between hyper-parameter settings and accuracy gain. Modern paragraph vector approaches require adequate calibration to produce better results, and we have improved the accuracy of feature location process with proper tuning.

Subjective Air Traffic Complexity Estimation Using Artificial Neural Networks

Air traffic complexity is usually defined as difficulty of monitoring and managing a specific air traffic situation. Since it is a psychological construct, best measure of complexity is that given by air traffic controllers. However, there is a need to make a method for complexity estimation which can be used without constant controller input. So far, mostly linear models were used. Here, the possibility of using artificial neural networks for complexity estimation is explored. Genetic algorithm has been used to search for the best artificial neural network configuration. The conclusion is that the artificial neural networks perform as well as linear models and that the remaining error in complexity estimation can only be explained as inter-rater or intra-rater unreliability. One advantage of artificial neural networks in comparison to linear models is that the data do not have to be filtered based on the concept of operations (conventional vs. trajectory-based).

Gap Prediction in Hybrid Graphene-Hexagonal Boron Nitride Nanoflakes Using Artificial Neural Networks

The electronic properties of graphene nanoflakes (GNFs) with embedded hexagonal boron nitride (hBN) domains are investigated by combined ab initio density functional theory calculations and machine-learning techniques. The energy gaps of the quasi-0D graphene-based systems, defined as the differences between LUMO and HOMO energies, depend not only on the sizes of the hBN domains relative to the size of the pristine graphene nanoflake but also on the position of the hBN domain. The range of the energy gaps for different configurations increases as the hBN domains get larger. We develop two artificial neural network (ANN) models able to reproduce the gap energies with high accuracies and investigate the tunability of the energy gap, by considering a set of GNFs with embedded rectangular hBN domains. In one ANN model, the input is in one-to-one correspondence with the atoms in the GNF, while in the second model the inputs account for basic structures in the GNF, allowing potential use in upscaled systems. We perform a statistical analysis over different configurations of ANNs to optimize the network structure. The trained ANNs provide a correlation between the atomic system configuration and the magnitude of the energy gaps, which may be regarded as an efficient tool for optimizing the design of nanostructured graphene-based materials for specific electronic properties.

Short-term wind speed forecasting in Uruguay using computational intelligence

Short-term wind speed forecasting for Colonia Eulacio, Soriano Department, Uruguay, is performed by applying an artificial neural network (ANN) technique to the hourly time series representative of the site. To train the ANN and validate the technique, data for one year are collected by one tower, with anemometers installed at heights of 101.8, 81.8, 25.7, and 10.0 m. Different ANN configurations are applied for each site and height; then, a quantitative analysis is conducted, and the statistical results are evaluated to select the configuration that best predicts the real data. This method has lower computational costs than other techniques, such as numerical modelling. For integrating wind power into existing grid systems, accurate short-term wind speed forecasting is fundamental. Therefore, the proposed short-term wind speed forecasting method is an important scientific contribution for reliable large-scale wind power forecasting and integration in Uruguay. The results of the short-term wind speed forecasting showed good accuracy at all the anemometer heights tested, suggesting that the method is a powerful tool that can help the Administración Nacional de Usinas y Transmissiones Eléctricas manage the national energy supply.

Predicting subjective measures of walkability index from objective measures using artificial neural networks

Increasing urbanization has been one of the most significant concerns of urban managers. The role of non-motorized transportation in sustainable urban development is vitally important for reducing overweight and obesity among citizens. These issues have led to numerous studies of the association between environmental characteristics, walkability index and levels of human health. The encouragement of public walking and cycling requires measures of walkability indices. One of the common challenges in measuring a walkability index is the complexity of the connection between the subjective indices resulting from public opinion and objective measures of geographic data. The scientific novelty of this paper lies in two aspects: First we developed and evaluated several artificial neural network (ANN) configurations for predicting subjective measures of walkability index from objective measures. Second, we introduced an index for two distinctive modes of walkability: daily shopping and recreation purposes that ranges from 1 to 10. The parameters of land-use diversity, population density, intersection density, network density, access to public transportation, green spaces and commercial places were utilized to calculate the objective value of the walkability index. The determination of subjective value of the walkability index was achieved using fieldwork reports. The resulting index was tested in districts 1 and 3 of Region 18 in the city of Tehran. The quantities used to evaluate the results included RMSE, MAE, MBE, and R. Network training was performed using the Levenberg-Marquardt algorithm. A 10-fold cross validation was used to evaluate and compare the performance of different network configurations. Our findings indicated that the best walkability index for purchasing can be estimated using Levenberg-Marquardt algorithm with one hidden layer and seven neurons. This configuration resulted in a correlation coefficient and an RMSE of 93.79% and 0.1368 respectively. To predict the walkability index for recreational purpose, the best result was obtained using Levenberg-Marquardt algorithm, representing a combination of one layer with four neurons for which the correlation coefficient and RMSE are 90.71% and 0.1602 respectively.

Determination of the Boyle temperature of pure gases using artificial neural networks

The Boyle temperature represents the temperature at which a real gas behaves more like an ideal gas. It is an useful thermodynamic parameter but it is difficult to measure experimentally. Thus, in this work we provide the Boyle temperatures of 16 pure gases, namely: argon, carbon dioxide, carbon monoxide, deuterium, helium-3, helium-4, hydrogen, krypton, methane, neon, nitric oxide, nitrogen, oxygen, tetrafluoromethane, tritium and xenon. The Boyle temperatures of the gases were determined by setting to zero their second virial coefficient (B) functions. Starting from 1663 available experimental data of B, an artificial neural network (ANN) was trained to correlate the second virial coefficient of the gases to two parameters: the reduced temperature and the critical molar volume. In order to obtain results as accurate as possible, different ANN configurations were tested, with variable numbers of neurons/hidden layers and up to 10000 iterations. It was found that the best configuration consists of 2 hidden layers with 24 neurons each. The ANN root-mean-square error was compared to those of other correlations available in literature and resulted to be very low, being equal to 2.74 cm3/mol. The temperatures determined by setting the ANN mapping of B to zero are comparable to those obtained from accurate curve fittings and, by virtue of their low deviations, can be adopted for calculation and as reference for experimental measures.

FORECASTING VALUES OF CHROMATICITY OF DRINKING AND SOURCE WATERS USING ARIMA MODEL AND NEURAL NETWORK

Aim. In the present investigation artificial neural network (ANN) and ARIMA-model are compared for forecasting of data of colour of water.Methods. Data corresponds to the colour of water of groundwater and drinking water of water intake of south-east region of the Republic of Belarus. The definition of colour was carried out for the period from 2009 to 2017. twice a day, the time series of values included 5215 values. The parameters of the models were estimated by 85% of the time series values, and the remaining 15% of the values (the test period) compared the forecast values with the actual ones. Optimal configurations of ARIMA-models were determined from the results of comparing the averaged values of the root mean squared errors (RMSE); optimal configurations of ANN were determined from the results of comparing the averaged values of RMSE and correlation coefficients (CC) on the test periods.Results. Comparison of forecasting methods was carried out on the basis of the averaged values of mean absolute error and mean relative error on the test periods. It was revealed that ANN allows to obtain the predicted values of colour of water more accurate than ARIMA-model.Main conclusions. Software implementation of ANN in the MATLAB environment empowers with sufficient accuracy get forecast values of groundwater and drinking water for 100 values.

Determination of the effects of operating conditions on the output power of the inverter and the power quality using an artificial neural network

In this study, the effects of operating conditions of photovoltaic (PV) panels in a power plant installed in Burdur province in Turkey with 8 MW capacity, the output power of the inverter and the power quality in the grid were examined experimentally in January and July. In addition, an artificial neural network (ANN) was utilized to estimate the power at the output of the inverter under different operating conditions and optimal operating conditions were attempted to be determined. In the ANN configuration, the output power generated in PV panels (PDC), radiation intensity, relative humidity and temperature measurements were used as input data, output power (PAC) at the junction of the inverter, power factor (PF) and the frequency values were used as output data. In order to ensure the integration of the energy obtained from the solar power plant by using the obtained results, it was tried to estimate the levels of the factors affecting the inverter output power and efficiency in terms of the proper operating conditions and efficiency. The success level of the ANN results was observed to be above 99% with the experimental results.

Performance comparison of neural network training algorithms in the modeling properties of steel fiber reinforced concrete

Our study is aimed at modeling the effect of three contributory factors, namely aspect ratio, water cement ratio and cement content on the water intake/absorption, compressive strength, flexural strength, split tensile strength and slump properties of steel fiber reinforced concrete. Artificial neural network (ANN) as a multilayer perceptron normal feed forward network was integrated to develop a predictive model for the aforementioned properties. Five training algorithms belonging to three classes: gradient descent, Levenberg Marquardt (quasi Newton) and genetic algorithm (GA). The ANN configuration consists of the input layer with three nodes, a single hidden layer of ten nodes of the output layer with five nodes. The study also compared the performance of all algorithms with regards to their predicting abilities. The ANN training was done by splitting the experimental data into the training and testing set. The divergence of the RMSE between the output and target values of the test set was monitored and used as a criterion to stop training. Although the convergence speed of GA was far higher than all other training algorithm, it performed better in predicting the water intake/absorption, split tensile strength and slump properties. However, incremental back propagation (IBP) and batch back propagation (BBP) outperformed GA in predicting the compressive strength and flexural strength respectively. The overall performance of the training algorithm was assessed using the coefficient of determination and the absolute fraction of variance obtained for the test data set and GA was found to have the highest value of 0.94 and 0.92 respectively. In determining the properties fiber reinforced concrete according to GA–ANN implementation, the water/cement ratio played slightly more dominant role than the aspect ratio and this was followed by cement content.

A cognitive approach for laser milled PMMA surface characteristics forecasting

In the present work, different Artificial Neural Networks (ANN) architectures were developed and applied to predict the surface characteristics (roughness and depth) of laser milled pockets, performed on poly-methyl-methacrylate (PMMA) sheets. The experimental data were obtained by adopting a 30 W CO2 laser source, fixing the average power at the maximum value and changing the wave mode (continuous or pulsed mode), the scan speed and the etching distance in large range. The depth and the roughness parameters (Ra and Rt), of machined surfaces were acquired by a 3D Surface Profiling System and adopted for the ANN training together with the process parameters. In order to allow network convergence, ANN training was executed by applying a random variable noise to the input data (Rn). The Mean Absolute Percentage Error (MAPE) was adopted to evaluate the ability of ANNs in surface characteristics forecasting. The results show a strong influence of the adopted ANN configuration on the forecasting ability. Nevertheless, a careful selection of the network architecture allows forecasting the roughness with a MAPE lower than 7%.

Artificial Neural Networks to Assess Emotional States from Brain-Computer Interface

Estimation of human emotions plays an important role in the development of modern brain-computer interface devices like the Emotiv EPOC+ headset. In this paper, we present an experiment to assess the classification accuracy of the emotional states provided by the headset’s application programming interface (API). In this experiment, several sets of images selected from the International Affective Picture System (IAPS) dataset are shown to sixteen participants wearing the headset. Firstly, the participants’ responses in form of a self-assessment manikin questionnaire to the emotions elicited are compared with the validated IAPS predefined valence, arousal and dominance values. After statistically demonstrating that the responses are highly correlated with the IAPS values, several artificial neural networks (ANNs) based on the multilayer perceptron architecture are tested to calculate the classification accuracy of the Emotiv EPOC+ API emotional outcomes. The best result is obtained for an ANN configuration with three hidden layers, and 30, 8 and 3 neurons for layers 1, 2 and 3, respectively. This configuration offers 85% classification accuracy, which means that the emotional estimation provided by the headset can be used with high confidence in real-time applications that are based on users’ emotional states. Thus the emotional states given by the headset’s API may be used with no further processing of the electroencephalogram signals acquired from the scalp, which would add a level of difficulty.

In pursuit of the best artificial neural network configuration for the prediction of output parameters of corrugated plate heat exchanger

Artificial Neural Network (ANN) is very effective tool for modeling non-linear multivariable relationships. In this study, the aim is to develop ANN models to predict the overall heat transfer coefficient, effectiveness factor, friction factor of cold and hot fluids: output parameters of the plate heat exchanger. The experimental data of two non-Newtonian fluids: Carboxymethyl cellulose (CMC) and Xanthan gum (XG) as cold fluid in different concentrations varying from 0.1 to 0.6%w/w and water as hot fluid is taken for the analysis of ANN. Four-input parameters namely cold fluid Reynolds number and Prandtl number, hot fluid Reynolds number and Prandtl number and five-input parameters namely concentration of the cold fluid, cold fluid Reynolds number and Prandtl number, hot fluid Reynolds number and Prandtl number are employed to develop two different models for each of the fluids. It is observed that the configuration 4-5-5-5-4 and 5-5-5-5-4 is best for XG with maximum absolute relative error as 6.792 and 4.03, respectively. Similarly, configuration 4-6-4 and 5-5-4 is best for CMC with a maximum absolute relative error of 5.85 and 5.392, respectively.

Characterization of a triple concentric-tube heat exchanger with corrugated tubes using Artificial Neural Networks (ANN)

This work presents a model of Artificial Neural Networks (ANNs) capable of accurately predicting the heat transfer rate and pressure drop in a triple concentric-tube heat exchanger (TTHX) with corrugated and non-corrugated inner tubes and involving a fluid typically used in the food industry. Pitch and depth are varied in the case of corrugated tubes. The ANN model was developed and validated using a huge databank including 181 experimental datasets. The best training algorithm is the Bayesian regulation. A back-propagation algorithm, which is considered to be the most common learning method for ANNs, is used in the training and testing of the network. Different network configurations were tested, and the optimum ANN configuration consisted of a network with two hidden layers with 15 and 21 nodes in the first and second layer, respectively. The ANNs results were found to be in good agreement with the experimental data, with the absolute average relative deviation (AARD) being below 1.91% for the heat transfer coefficient and below 3.82% for the pressure drop, respectively. The simplicity of the developed ANN model and its low levels of error for a huge experimental databank are some of the key features of the model.

A non-negative artificial neural network configuration for noise predictions

Artificial Neural Networks (ANN) provide an accurate relationship between user-defined inputs and outputs for complex problems. Their basic units are called neurons. The computation that takes place in each neuron is a weighted summation of its inputs. These results are then implemented as an input to an activation or transfer function, which in turn defines the final neuron’s output. The transfer function configuration is very important when defining the final purpose of the ANN. There are many standard activation functions like pure linear, tangent sigmoid, logistic sigmoid, among others. In this study, an ANN capable of predicting tire pavement interaction noise (TPIN) is constructed. Its outputs correspond to sound pressure in narrowband. Thus, the main requirement is that the outputs must be positive. Several neuron transfer functions were investigated for both hidden and output layers. Finally, a customized hybrid linear-sigmoid transfer function was selected. The final ANN configuration is able to capture typical TPIN spectral content within 400 Hz–1600 Hz. In addition, the error is significantly reduced if compared to those obtained after implementing standard transfer functions.