Multilayer Perceptron Feed-forward Neural Network Research Articles

Solving Riccati-Type Nonlinear Differential Equations with Novel Artificial Neural Networks

In this study we investigate neural network solutions to nonlinear differential equations of Ricatti-type. We employ a feed-forward Multilayer Perceptron Neural Network (MLPNN), but avoid the standard back-propagation algorithm for updating the intrinsic weights. Our objective is to minimize an error, which is a function of the network parameters i.e., the weights and biases. Once the weights of the neural network are obtained by our systematic procedure, we need not adjust all the parameters in the network, as postulated by many researchers before us, in order to achieve convergence. We only need to fine-tune our biases which are fixed to lie in a certain given range, and convergence to a solution with an acceptable minimum error is achieved. This greatly reduces the computational complexity of the given problem. We provide two important ODE examples, the first is a Ricatti type differential equation to which the procedure is applied, and this gave us perfect agreement with the exact solution. The second example however provided us with only an acceptable approximation to the exact solution. Our novel artificial neural networks procedure has demonstrated quite clearly the function approximation capabilities of ANN in the solution of nonlinear differential equations of Ricatti type.

A Novel Method for Solving Ordinary Differential Equations with Artificial Neural Networks

This research work investigates the use of Artificial Neural Network (ANN) based on models for solving first and second order linear constant coefficient ordinary differential equations with initial conditions. In particular, we employ a feed-forward Multilayer Perceptron Neural Network (MLPNN), but bypass the standard back-propagation algorithm for updating the intrinsic weights. A trial solution of the differential equation is written as a sum of two parts. The first part satisfies the initial or boundary conditions and contains no adjustable parameters. The second part involves a feed-forward neural network to be trained to satisfy the differential equation. Numerous works have appeared in recent times regarding the solution of differential equations using ANN, however majority of these employed a single hidden layer perceptron model, incorporating a back-propagation algorithm for weight updation. For the homogeneous case, we assume a solution in exponential form and compute a polynomial approximation using statistical regression. From here we pick the unknown coefficients as the weights from input layer to hidden layer of the associated neural network trial solution. To get the weights from hidden layer to the output layer, we form algebraic equations incorporating the default sign of the differential equations. We then apply the Gaussian Radial Basis function (GRBF) approximation model to achieve our objective. The weights obtained in this manner need not be adjusted. We proceed to develop a Neural Network algorithm using MathCAD software, which enables us to slightly adjust the intrinsic biases. We compare the convergence and the accuracy of our results with analytic solutions, as well as well-known numerical methods and obtain satisfactory results for our example ODE problems.

Radiomics of high-resolution computed tomography for the differentiation between cholesteatoma and middle ear inflammation: effects of post-reconstruction methods in a dual-center study

ObjectivesTo evaluate the performance of radiomic features extracted from high-resolution computed tomography (HRCT) for the differentiation between cholesteatoma and middle ear inflammation (MEI), and to investigate the impact of post-reconstruction harmonization and data resampling.MethodsOne hundred patients were included in this retrospective dual-center study: 48 with histology-proven cholesteatoma (center A: 23; center B: 25) and 52 with MEI (A: 27; B: 25). Radiomic features (co-occurrence and run-length matrix, absolute gradient, autoregressive model, Haar wavelet transform) were extracted from manually defined 2D-ROIs. The ten best features for lesion differentiation were selected using probability of error and average correlation coefficients. A multi-layer perceptron feed-forward artificial neural network (MLP-ANN) was used for radiomics-based classification, with histopathology serving as the reference standard (70% of cases for training, 30% for validation). The analysis was performed five times each on (a) unmodified data and on data that were (b) resampled to the same matrix size, and (c) corrected for acquisition protocol differences using ComBat harmonization.ResultsUsing unmodified data, the MLP-ANN classification yielded an overall median area under the receiver operating characteristic curve (AUC) of 0.78 (0.72–0.84). Using original data from center A and resampled data from center B, an overall median AUC of 0.88 (0.82–0.99) was yielded, while using ComBat harmonized data, an overall median AUC of 0.89 (0.79–0.92) was revealed.ConclusionRadiomic features extracted from HRCT differentiate between cholesteatoma and MEI. When using multi-centric data obtained with differences in CT acquisition parameters, data resampling and ComBat post-reconstruction harmonization clearly improve radiomics-based lesion classification.Key Points• Unenhanced high-resolution CT coupled with radiomics analysis may be useful for the differentiation between cholesteatoma and middle ear inflammation.• Pooling of data extracted from inhomogeneous CT datasets does not appear meaningful without further post-processing.• When using multi-centric CT data obtained with differences in acquisition parameters, post-reconstruction harmonization and data resampling clearly improve radiomics-based soft-tissue differentiation.

Indirect tool monitoring in drilling based on gap sensor signal and multilayer perceptron feed forward neural network

In this study, an indirect tool monitoring was developed based on the installation of a gap sensor in measuring the signal related to the tool behavior during the drilling process. Eleven types of twist drills with different tool conditions were utilized to differentiate the sensorial signals based on the tool states. A statistical analysis was conducted in the signal processing, by extracting the gap sensor signal associates from each tool condition, using the skewness and kurtosis features. Multi-class classification was conducted using the multilayer perceptron (MLP) feed forward neural network (FF-NN) model to classify and predict the tool condition based on the skewness and kurtosis data. The architectures of the MLP FF-NN models were varied to optimize the classification accuracy. This study found that the tool condition was correlated to the displacement of the drill machine spindle because the runout occurred when the sensor signal displayed fluctuation and irregularity trends. The peak intensity of the gap sensor signals increased with increasing wear severity of the twist drill. An ideal MLP FF-NN structure was achieved when the classification performance was optimized to be consistent with the learning curve.

Prediction of maximum thermal crack width of RC abutments utilizing actual construction data and study on influential parameters using neural networks

In this study, an attempt was made to predict the maximum width of thermal cracking in RC abutments using field data. The data was obtained from Japanese Yamaguchi prefecture database which is a database of concrete structures in the prefecture constructed with appropriate concreting work. Reliable data was chosen carefully to avoid incorporation of possible human errors. Feed-forward multilayer perceptron neural network was used for prediction. k-fold cross validation was performed to avoid overfitting. By using the trained NN, several parametric studies were conducted to observe the influences of different parameters. Main idea of this research was to solve the complex problem of predicting the maximum crack width by using the parameters which are easier to obtain in the field. The results have shown the potential of predicting the crack width and observable influence of several parameters. The results will be helpful in proposing countermeasures to mitigate harmful thermal cracks.

Power Optimization Control Scheme for Doubly Fed Induction Generator Used in Wind Turbine Generators

Scientists and researchers are exploring different methods of generating and delivering electrical energy in an economical and reliable way, enabling them to generate electricity focusing on renewable energy resources. All of these possess the natural property of self-changing behavior, so the connection of these separate independent controllable units to the grid leads to uncertainties. This creates an imbalance in active power and reactive power. In order to control the active and reactive power in wind turbine generators with adjustable speed, various control strategies are used to allay voltage and current variations. This research work is focused on the design and implementation of effective control strategies for doubly fed induction generator (DFIG) to control its active and reactive power. A DFIG system with its control strategies is simulated on MATLAB software. To augment the transient stability of DFIG, the simulation results for the active and reactive power of conventional controllers are compared with three types of feed forward neural network controllers, i.e., probabilistic feedforward neural network (PFFNN), multi-layer perceptron feedforward neural network (MLPFFN) and radial basic function feedforward neural network (RBFFN) for optimum performance. Conclusive outcomes clearly manifest the superior robustness of the RBFNN controller over other controllers in terms of rise time, settling time and overshoot value.

Characterization of the Nonlinear Biaxial Mechanical Behavior of Human Ureter Using Constitutive Modeling and Artificial Neural Networks

Characterization of the mechanical properties of soft biological tissues is a fundamental issue in a variety of medical applications. As such, constitutive modeling of biological tissues that serves to establish a relationship between the kinematic variables has been used to formulate the tissue’s mechanical response under various loading conditions. However, the validation of the developed analytical and numerical models is accompanied by a length of computation time. Hence, the need for new advantageous methods like artificial intelligence (AI), aiming at minimizing the computation time for real-time applications such as in robotic-assisted surgery, sounds crucial. In this study, at first, the mechanical nonlinear characteristics of human ureter were obtained from planar biaxial test data, in which the examined specimens were simultaneously loaded along their circumferential and longitudinal directions. To do so, the biaxial stress-strain data was used to fit the well-known Fung and Holzapfel-Delfino hyperelastic functions using the genetic optimization algorithm. Next, the potential of Artificial Neural Networks (ANN), as an alternative method for prediction of the mechanical response of the tissue was evaluated such that, multilayer perceptron feedforward neural network with different architectures was designed and implemented and then, trained with the same experimental data. The results showed both approaches were practically able to predict the ureter nonlinearity and in particular, the ANN model can follow up the tissue nonlinearity during the entire loading phase in both low and high strain amplitudes (RMSE<0.02). Such results confirmed that neural networks can be a reliable alternative for modeling the nonlinear mechanical behavior of soft biological tissues.

Analysis of statistical coefficients and autoregressive parameters over intrinsic mode functions (IMFs) for epileptic seizure detection.

Epilepsy is a persistent neurological disorder impacting over 50 million people around the world. It is characterized by repeated seizures defined as brief episodes of involuntary movement that might entail the human body. Electroencephalography (EEG) signals are usually used for the detection of epileptic seizures. This paper introduces a new feature extraction method for the classification of seizure and seizure-free EEG time segments. The proposed method relies on the empirical mode decomposition (EMD), statistics and autoregressive (AR) parameters. The EMD method decomposes an EEG time segment into a finite set of intrinsic mode functions (IMFs) from which statistical coefficients and autoregressive parameters are computed. Nevertheless, the calculated features could be of high dimension as the number of IMFs increases, the Student's t-test and the Mann-Whitney U test were thus employed for features ranking in order to withdraw lower significant features. The obtained features have been used for the classification of seizure and seizure-free EEG signals by the application of a feed-forward multilayer perceptron neural network (MLPNN) classifier. Experimental results carried out on the EEG database provided by the University of Bonn, Germany, demonstrated the effectiveness of the proposed method which performance assessed by the classification accuracy (CA) is compared to other existing performances reported in the literature.

Non-Invasive Assessment of Breast Cancer Molecular Subtypes with Multiparametric Magnetic Resonance Imaging Radiomics.

We evaluated the performance of radiomics and artificial intelligence (AI) from multiparametric magnetic resonance imaging (MRI) for the assessment of breast cancer molecular subtypes. Ninety-one breast cancer patients who underwent 3T dynamic contrast-enhanced (DCE) MRI and diffusion-weighted imaging (DWI) with apparent diffusion coefficient (ADC) mapping were included retrospectively. Radiomic features were extracted from manually drawn regions of interest (n = 704 features per lesion) on initial DCE-MRI and ADC maps. The ten best features for subtype separation were selected using probability of error and average correlation coefficients. For pairwise comparisons with >20 patients in each group, a multi-layer perceptron feed-forward artificial neural network (MLP-ANN) was used (70% of cases for training, 30%, for validation, five times each). For all other separations, linear discriminant analysis (LDA) and leave-one-out cross-validation were applied. Histopathology served as the reference standard. MLP-ANN yielded an overall median area under the receiver-operating-characteristic curve (AUC) of 0.86 (0.77–0.92) for the separation of triple negative (TN) from other cancers. The separation of luminal A and TN cancers yielded an overall median AUC of 0.8 (0.75–0.83). Radiomics and AI from multiparametric MRI may aid in the non-invasive differentiation of TN and luminal A breast cancers from other subtypes.

Rainfall-river discharge modelling for flood forecasting using Artificial Neural Network (ANN)

This study is aimed at evaluating the applicability of Artificial Neural Network (ANN) model technique for river discharge forecasting. Feed-forward multilayer perceptron neural network trained with back-propagation algorithm was employed for model development. Hydro-meteorological data for the Imo River watershed, that was collected from the Anambra-Imo River Basin Development Authority, Owerri – Imo State, South-East, Nigeria, was used to train, validate and test the model. Coefficients of determination results are 0.91, 0.91 and 0.93 for training, validation and testing periodsrespectively. River discharge forecasts were fitted against actual discharge data for one to five lead days. Model results gave R2 values of 0.95, 0.95, 0.92, 0.96 and 0.94 for first, second, third, fourth, and fifth lead days of forecasts, respectively. It was generally observed that the R2 values decreased with increase in lead days for the model. Generally, this tech-nique proved to be effective in river discharge modelling for flood forecasting for shorter lead-day times, especially in areas with limited data sets.

Supervised machine learning for source allocation of per- and polyfluoroalkyl substances (PFAS) in environmental samples

Environmental contamination by per- and polyfluoroalkyl substances (PFAS) is widespread, because of both their decades of use, and their persistence in the environment. These factors can make identification of the source of contamination in samples a challenge, because in many cases contamination may originate from decades ago, or from a number of candidate sources. Forensic source allocation is important for delineating plumes, and may also be able to provide insights into environmental behaviors of specific PFAS components. This paper describes work conducted to explore the use of supervised machine learning classifiers for allocating the source of PFAS contamination based on patterns identified in component concentrations. A dataset containing PFAS component concentrations in 1197 environmental water samples was assembled based on data from sites from around the world. The dataset was split evenly into training and test datasets, and the 598-sample training dataset was used to train four machine learning classifiers, including three conventional machine learning classifiers (Extra Trees, Support-Vector Machines, K-Neighbors), and one multilayer perceptron feedforward deep neural network. Of the methods tested, the deep neural network and Extra Trees exhibited particularly high performance at classification of samples from a range of sources. The fact that the methods function on completely different principles and yet provide similar predictions supports the hypothesis that patterns exist in PFAS water sample data that can allow forensic source allocation. The results of the work support the idea that supervised machine learning may have substantial promise as a tool for forensic source allocation.

Optimization of solidification in die casting using numerical simulations and machine learning

In this paper, we demonstrate the combination of machine learning and three dimensional numerical simulations for multi-objective optimization of low pressure die casting. The cooling of molten metal inside the mold is achieved typically by passing water through the cooling lines in the die. Depending on the cooling line location, coolant flow rate and die geometry, nonuniform temperatures are imposed on the molten metal at the mold wall. This boundary condition along with the initial molten metal temperature affect the product quality quantified in terms of micro-structure parameters and yield strength. A finite volume based numerical solver is used to determine the temperature-time history and correlate the inputs to outputs. The objective of this research is to develop and demonstrate a procedure to obtain the initial and wall temperatures so as to optimize the product quality. The non-dominated sorting genetic algorithm (NSGA-II) is used for multi-objective optimization in this work. The number of function evaluations required for NSGA-II can be of the order of millions and hence, the finite volume solver cannot be used directly for optimization. Therefore, a multilayer perceptron feed-forward neural network is first trained using the results from the numerical solution of the fluid flow and energy equations and is subsequently used as a surrogate model. As an assessment, simplified versions of the actual problem are designed to first verify results of the genetic algorithm. An innovative local sensitivity based approach is then used to rank the final Pareto optimal solutions and select a single best design.

Prediction and optimization of slaughter weight in meat-type quails using artificial neural network modeling

Carcass yield of meat-type quails is strongly correlated with the weight of the birds at slaughter (slaughter weight [SW]; body weight at 45 D of age). Moreover, prediction of superior animals for SW at the earlier stages of the rearing period is favorable for producers. Therefore, the aim of the present study was to predict and optimize SW of Japanese quails based on their early growth performances, sex, and egg weight as predictors through artificial neural network (ANN) modeling. To construct the ANN model a feed-forward multilayer perceptron neural network structure was used. Moreover, sensitivity analysis was used to arrange the predictors in the ANN model(s) according to their predictive importance too. In addition, the optimization process was conducted to determine the optimum values for the input variables to yield maximum SW. The best-fitted network on input data to predict SW in Japanese quails was determined with 7 neurons in the input layer, 11 neurons in the hidden layer, and one neuron in the output layer. The coefficient of determination (R2) was 0.9404, 0.9359, and 0.9223 for training, validation, and testing phases, respectively. For the corresponding phases, SEM were also 51.8854, 52.2764, and 55.2572, respectively. According to sensitivity analysis, the most important input variable for prediction of SW was body weight at 20 D of age (BW20), whereas the less important input variables were weight of the birds at hatch and body weight at 5 D of age. The results of the neural network optimization indicated that all the input variables, except for BW20, were very similar but slightly higher than mean values (μ for each input variable). The results of this study suggest that the ANN provides a practical approach to predict the final body weight (SW) of Japanese quails based on early performances. Moreover, phenotypic selection for higher values of early growth traits did not ensure the achievement of maximum SW, except for BW20.

Optimal robust control of vehicle lateral stability using damped least-square backpropagation training of neural networks

Chassis control systems play a significant role in achieving the desired vehicle performance and stability during various severe maneuvers. A probabilistic estimation approach by hybridization of optimal robust control and a damped least-square backpropagation based neural networks (NN) is proposed to design a control system for dealing with unknown nonlinear dynamics of a passenger car. To this end, a four-wheel active steering (4WAS) model is employed and a multilayer perceptron (ML) feed-forward backpropagation neural network (FFBPNN) model is developed as an approximator. The optimal robust control is employed to regulate the yaw rate and side-slip angle of the vehicle to follow the desired vehicle response. The developed FFBPNN model is trained to distinguish the nonlinear dynamics of the vehicle and the corresponding optimal feedback gain during a wide range of operating conditions via the state variables. The robustness of the controller is evaluated using Lyapunov stability method. The performance of the proposed controller is analyzed considering the open-loop and closed-loop responses of the nonlinear vehicle model and a sliding mode controller to track the desired yaw rate and side-slip angle responses. The results obtained during severe maneuvers suggest that the proposed control method can substantially enhance the handling and stability performances of the vehicle.

Transient Thermography for Flaw Detection in Friction Stir Welding: A Machine Learning Approach

A systematic computational method to simulate and detect subsurface flaws, through nondestructive transient thermography, in aluminum (AL) sheets and friction stir (FS) welded sheets is proposed in this article. The proposed method relies on feature extraction methods and a data-driven machine learning modeling structure. Here, we propose the use of a multilayer perceptron feed-forward neural network with feature extraction methods to improve the flaw-probing depth of transient thermography inspection. Furthermore, for the first time, we propose thermographic signal linear modelling (TSLM), a hyper-parameter-free feature extraction technique for transient thermography. The new feature extraction and modeling framework was tested with out-of-sample experimental transient thermography data, and results show effectiveness in subsurface flaw detection of up to 2.3 mm deep in AL sheets [99.8% true positive rate (TPR) and 92.1% true negative rate (TNR)] and up to 2.2 mm deep in FS welds (97.2% TPR and 87.8% TNR).

AGRICULTURAL LAND CHANGE DETECTING AND FORECASTING USING COMBINATION OF FEEDFORWARD MULTILAYER NEURAL NETWORK, CELLULAR AUTOMATA AND MARKOV CHAIN MODELS

Abstract. This paper proposed a methodology for finding changes in agricultural land of Tehran during past years and simulating these changes for future years. The proposed method utilized the spatial GIS-based techniques and Landsat satellite imagery to predict agricultural land map for the future of Tehran. Therefore, a method for finding and predicting changes based on combining the feedforward multilayer perceptron neural network (MLP), cellular automata (CA), and Markov chain model were applied. In this regard, the Landsat images of 2002, 2008, and 2014 were classified by a binary support vector machine classifier into two classes of agricultural and non-agricultural. Then, the potential transition maps were generated by the neural network MLP and extensible areas were obtained by the Markov chain model. Finally, the results of these two steps were combined with the MOLA method and the 2020 and 2025 agricultural maps were predicted. The proposed method obtained the Kappa factor of 89.92% that indicates the high ability of the neural network and the CA–Markov for finding the changes and prediction in the city of Tehran.

Evaluation of phase equilibrium conditions of clathrate hydrates using connectionist modeling strategies

The potentials of gas hydrates to store large quantities of gases such as H2 and CH4 make them suitable candidates for energy storage applications. The presence of insoluble liquid hydrocarbons can lead to the formation of four-phase hydrates, and also affects the hydrate formation conditions. In this study, we propose a robust multilayer perceptron feed-forward artificial neural network (MLP-FF-ANN) and least squares support vector machines (LSSVM kernel RBF type) to forecast the four-phase equilibrium conditions of structure II (sII) gas hydrates. A total of 474 equilibrium data points for four-phase hydrate conditions are collected from the literature that are used to develop the models. The optimal ANN model structure contains a single hidden layer with 12 tangent-Sigmoid neurons, which is trained with the back-propagation algorithm. The magnitudes of average absolute relative deviation percent (AARD%) while employing the ANN and LSSVM models are 0.08 and 0.21, respectively. Although for the systems consisting of C2H6, N2, H2, and CO2, both methods feature a similar performance in estimating the equilibrium hydrate dissociation conditions, the ANN outperforms the LSSVM in systems including fluoromethane (R-41), difluoromethane (R-32), Kr, and H2S. Moreover, the performances of the ANN and LSSVM are compared to two thermodynamic (analytical) models and the ANN model is found the most accurate. This performance superiority brings potentials in demanding applications such as adaptive control, online state estimation, and process optimization in energy storage systems.

Design of a flattening filter using Fiber Bragg Gratings for EDFA gain equalization: an artificial neural network application

&#x0D; &#x0D; &#x0D; This paper presents a proposal for the non-uniform gain compensation of an Erbium-doped fiber optic amplifier (EDFA) in a Wavelength Division Multiplexed (WDM) system using Fiber Bragg Gratings (FBG). In this proposal, the multilayer perceptron feed-forward artificial neural network with backpropagation was trained under the secant method (one-step secant) and was selected according to mean square error measurement. The proposal optimizes FBG parameters such as center frequency, rejection level and length in order to determine a filtering response based on a reduced number of FBGS that will be used to flatten the non-linear response of the amplifier gain and avoid the per-carrier treatment of a standard flattening filter. While an artificial neural network with a 7-10-6 structure demonstrated the feasibility of equalizing the gain of an EDFA using as few as three FBGS, a 25-18-12 structure improved the results when the configuration consisted of an FBG array of six resonances that provided similar results to that featured by the standard gain-flattening filter. The proposal was evaluated in an amplified WDM system of eight optical carriers located between 195-196.4 THz.&#x0D; &#x0D; &#x0D;

Integrated Computational Alloy Design of Nickel-Base Superalloys

The nickel-base superalloys are unsurpassed for hot sections of the gas turbine engines whose efficiency can be improved by developing new superalloys with enhanced properties. The artificial neural networks (ANN) method is utilized in alloy design of single-crystal nickel-base superalloys. The ANN method is especially powerful in applications where developing a physically sound model is challenging such as alloy design with multiple components. A design space of 45,000 alloy compositions is generated, and then a multilayer perceptron (MLP) feed-forward (FF) artificial neural network is used to predict the density and creep lives of these alloys. The ANN method is coupled with physics-based (PHACOMP) calculations to predict and then correlate the creep rupture life to the composition, precipitate volume fraction, topologically closed packed phases, temperature, and stress. Then, an experimental alloy composition is selected from the design space, and its solidification behavior is investigated by a thermodynamics (CALPHAD)-based software. Moreover, single-crystal rods of the experimental alloy are grown and machined into creep samples, which are creep tested at three different stress–temperature couples. A good match between the experimental results and the ANN predictions is displayed in scatter plots and in Larson–Miller plots for the experimental alloy and for selected, well-known commercial single-crystal alloys. As the correlation of the microstructure to mechanical properties is still in its infancy by thermodynamics and mechanics-based software, an integrated ANN modeling is shown to be a powerful tool for finding a composition and establishing relationships between the microstructure and properties of alloys. This, of course, can help reduce the material development cycle time as aimed by the Integrated Computational Material Engineering (ICME).

Seasonal prediction of particulate matter over the steel city of India using neural network models

The particulate matter (PM) concentration forecast is an important component to evaluate the air quality over any region of the world. The results are vital when it comes to health concern issues related to air pollution in developing countries like India and China. The present study is focused on the prediction of future concentrations of respirable suspended particulate matter (RSPM, also called PM2.5) and suspended particulate matter (SPM, also called PM10) using the artificial neural networks (ANN) over tropical inland industrial site, Rourkela. Rourkela, popularly known as steel city of India is situated at the heart of a rich mineral belt and is endowed with many small, medium and large scale industries. This study aims to develop a seasonal ANN modelling approach for prediction of RSPM and SPM for the year 2013 by using the datasets of meteorology and RSPM and SPM concentrations covered from 2009 to 2013. Four neural network models namely, wavelet-based multi-layer perceptron feed forward neural network (WMLPNN), wavelet-based recurrent neural network, multi-layer perceptron feed forward neural network (MLPNN) and recurrent neural network (RNN) are used. The networks are trained using daily data from 2009 to 2012 on a seasonal basis, and daily predictions are performed for 2013 using seasonal based trained model. Five meteorological variables [temperature (T), relative humidity (RH), boundary layer height (BLH), surface pressure (SP), wind direction (WD) and wind speed (WS)] along with the Daubechies wavelet decomposed coefficients are considered as predictor variables in the models. A lagged scheme is introduced in wind speed input vector and networks are trained and tested up to 3 days lagged wind speed variable which greatly improved the prediction accuracy. WMLPNN is found to be outperforming all other tested schemes by providing promising results among all the tested models with lag in pre-monsoon, monsoon and winter seasons and without lag in wind speed for the post-monsoon season for both RSPM and SPM, elucidating the importance of predictor wind speed in the models. The values of R2 varied from 0.6 to 0.91 for RSPM and 0.7 to 0.98 for SPM and RMSE values varied from 0.09 to 0.14 for RSPM and 0.03 to 0.13 for SPM for WMLPNN model. The proposed WMLPNN model in the present study appeared to be reliable and promising for the prediction of PM2.5 and PM10 by providing aid in passing the alerts and notices for the betterment of living beings.