Deconvolutional Neural Network Research Articles

Spatial–temporal prediction model for unsteady near-wall flow around cylinder based on hybrid neural network

A hybrid neural network based on Densely Connected Convolutional Networks (DenseNet), Convolutional Long Short-Term Memory Neural Network (ConvLSTM), and Deconvolutional Neural Network (DeCNN) is employed to predict unsteady flow fields. The utilization of DenseNet makes the model more compact and makes the prediction of three-dimensional flow affordable. The ConvLSTM is implemented to predict multiple future time steps which improves prediction efficiency. The proposed model transforms the time sequences of velocity and pressure fields into uniform spatial–temporal topology as input and captures nonlinear feature information in the spatial–temporal domain. Numerical simulations are conducted for the flow around cylinder at different Reynolds numbers and the near-wall flow around cylinder with different gap ratios, and training samples for the neural network inputs are established. The predicted results are compared with the numerical simulation results, showing good agreement. From the prediction cycle, it can be seen that good prediction results can be maintained in the first three prediction cycles. The prediction results of the three-dimensional unsteady flow around a cylinder near a plane wall, exhibit remarkable accuracy, successfully capturing the evolution of turbulent vortex structures. This signifies that the prediction model is highly effective in capturing the spatial–temporal variations of complex unsteady flows.

Mean Squared Error May Lead You Astray When Optimizing Your Inverse Design Methods

Abstract When performing time-intensive optimization tasks, such as those in Topology or Shape Optimization, researchers have turned to machine-learned (ML) Inverse Design methods i.e., predicting the optimized geometry from input conditions to replace or warm start traditional optimizers. Such methods are often optimized to reduce the Mean Squared Error (MSE) or Binary Cross-Entropy between the output and a training dataset of optimized designs. While convenient, we show that this choice may be myopic. Specifically, we compare two methods of optimizing the hyper-parameters of easily reproducible Machine Learning models including random forest (RF), k-nearest neighbors (KNN), and Deconvolutional Neural Network (DeCNN) model for predicting the three optimal topology problems. We show that under both direct Inverse Design and when warm starting further Topology Optimization (TO), using MSE metrics to tune hyper-parameters produces less performant models than directly evaluating the objective function, though both produce designs that are almost one order of magnitude better than using the common uniform initialization. We also illustrate how warm starting impacts both the convergence time, the type of solutions obtained during optimization, and the final designs. Overall, our initial results portend that researchers may need to revisit common choices for evaluating ID methods that subtly trade-off factors in how an ID method will actually be used. We hope our open-source dataset and evaluation environment will spur additional research in those directions.

Performance prediction of co-rotating disk cavity with finned vortex reducer based on machine learning

Machine learning method was applied for rapid and precise prediction of flow and heat transfer characteristics within co-rotating disk cavity with a finned vortex reducer. A new data preprocessing scheme was developed to reduce modeling cost. By using this scheme, classical radial basis function neural network (RBFNN) shows better prediction performance compared with deconvolutional neural network. Furthermore, RBFNN has a simpler topological structure and fewer hyperparameters. By testing, the relative root mean square error for total pressure, total temperature, and swirl ratio is 0.51 %, 0.32 %, and 1.99 %, and corresponding coefficient of determination reaches 99.68 %, 96.55 %, and 99.16 %. The effects of input parameters on the outputs of RBFNN were analyzed, and a sensitivity analysis was conducted. Within the investigated parameter range, the total pressure loss increases as dimensionless mass flow rate, rotational Reynolds number, and radial position of the fin increases. The total temperature drop increases with the increase of dimensionless mass flow rate and rotational Reynolds number, but the decrease of radial position of the fin. Additionally, an increase in dimensionless mass flow rate or a decrease in rotational Reynolds number causes the increase of swirl ratio. This research provides an efficient modeling method for components in secondary air system in gas turbines.

Deconvolutional Neural Network for Generating Spray Trajectory of Shoe Soles

The footwear industry is moving towards automation and intellectualization. To overcome the drawbacks of the high-cost and low-efficiency traditional manual shoe sole gluing process, automatic methods were utilized for generating spray trajectories. Currently, most of the reported automatic methods for generating spray trajectories mainly rely on the outer contour bias method. However, the glue is only applied to the area near the edge/contour of shoe soles and the fixed offset distance in the outer contour bias method cannot adapt to the immense amount of shoe styles with high precision and achieve applicability for irregular and unique sole designs. An intuitive yet logical approach to fulfill the requirements is to utilize the deconvolutional neural network for generating shoe sole spray trajectories. In this work, we treated the glue trajectory prediction as an image-to-image prediction and established a novel deconvolutional neural network to generate shoe sole spray trajectories. The as-proposed deconvolutional neural network for generating spray trajectory offered significant advantages over the existing bias-based methods, including: (1) based on the novel deconvolutional neural network, the proposed method for generating shoe sole spray trajectories exhibits greater applicability to irregular shoe soles, which improves the spray accuracy without compromising the spray efficiency; (2) we discard all the pooling layers, which only consist of convolutional and deconvolutional layers, to preserve more spatial information and achieve higher spray accuracy through end-to-end mapping from shoe sole images to shoe sole spray trajectories, resulting in an improved spray accuracy without sacrificing spray efficiency. The Dice similarity coefficient and Hausdorff distance were used as the evaluation metrics to assess the performance of our approach. Our proposed method showed an ultra-high accuracy and precision with a Dice similarity coefficient over 99.25% and a Hausdorff distance less than 1.2 mm, which are ~10% higher than the spray accuracy of other reported traditional methods. Our findings would bring significant improvements to the field of automatic shoe sole spray trajectory generation, which has the potential to promote the utilization of intelligent technologies in the footwear industry.

Meshless optical mode solving using scalable deep deconvolutional neural network

Optical mode solving is of paramount importance in photonic design and discovery. In this paper we propose a deep deconvolutional neural network architecture for a meshless, and resolution scalable optical mode calculations. The solution is arbitrary in wavelengths and applicable for a wide range of photonic materials and dimensions. The deconvolutional model consists of two stages: the first stage projects the photonic geometrical parameters to a vector in a higher dimensional space, and the second stage deconvolves the vector into a mode image with the help of scaling blocks. Scaling block can be added or subtracted as per desired resolution in the final mode image, and it can be effectively trained using a transfer learning approach. Being a deep learning model, it is light, portable, and capable of rapidly disseminating edge computing ready solutions. Without the loss of generality, we illustrate the method for an optical channel waveguide, and readily generalizable for wide range photonic components including photonic crystals, optical cavities and metasurfaces.

Multi-Level Deep Generative Adversarial Networks for Brain Tumor Classification on Magnetic Resonance Images

The brain tumor is an abnormal and hysterical growth of brain tissues, and the leading cause of death affected patients worldwide. Even in this technology-based arena, brain tumor images with proper labeling and acquisition still have a problem with the accurate and reliable generation of realistic images of brain tumors that are completely different from the original ones. The artificially created medical image data would help improve the learning ability of physicians and other computer-aided systems for the generation of augmented data. To overcome the highlighted issue, a Generative Adversarial Network (GAN) deep learning technique in which two neural networks compete to become more accurate in creating artificially realistic data for MRI images. The GAN network contains mainly two parts known as generator and discriminator. Commonly, a generator is the convolutional neural network, and a discriminator is the deconvolutional neural network. In this research, the publicly accessible Contrast-Enhanced Magnetic Resonance Imaging (CE-MRI) dataset collected from 2005-to 2020 from different hospitals in China consists of four classes has been used. Our proposed method is simple and achieved an accuracy of 96%. We compare our technique results with the existing results, indicating that our proposed technique outperforms the best results associated with the existing methods.

Fast Prediction of Flow Field around Airfoils Based on Deep Convolutional Neural Network

We propose a steady-state aerodynamic data-driven method to predict the incompressible flow around airfoils of NACA (National Advisory Committee for Aeronautics) 0012-series. Using the Signed Distance Function (SDF) to parameterize the geometric and flow condition setups, the prediction core of the method is constructed essentially by a consecutive framework of a convolutional neural network (CNN) and a deconvolutional neural network (DCNN). Impact of training parameters on the behavior of the proposed CNN-DCNN model is studied, so that appropriate learning rate, mini-batch size, and random deactivation rate are specified. Tested by “unseen” airfoil geometries and far-field velocities, it is found that the prediction process is three orders of magnitudes faster than a corresponding Computational Fluid Dynamics (CFD) simulation, while relative errors are maintained lower than 1% on most of the sample points. The proposed model manages to capture the essential dynamics of the flow field, as its predictions correspond reasonably with the reconstructed field by proper orthogonal decomposition (POD). The performance and accuracy of the proposed model indicate that the deep learning-based approach has great potential as a robust predictive tool for aerodynamic design and optimization.

PREFACE: A SPECIAL SELECTION ON EMERGING TECHNIQUES FOR BIOMECHANICS–PART II

PREFACE: A SPECIAL SELECTION ON EMERGING TECHNIQUES FOR BIOMECHANICS–PART II

Towards End-to-End Image Compression and Analysis with Transformers

We propose an end-to-end image compression and analysis model with Transformers, targeting to the cloud-based image classification application. Instead of placing an existing Transformer-based image classification model directly after an image codec, we aim to redesign the Vision Transformer (ViT) model to perform image classification from the compressed features and facilitate image compression with the long-term information from the Transformer. Specifically, we first replace the patchify stem (i.e., image splitting and embedding) of the ViT model with a lightweight image encoder modelled by a convolutional neural network. The compressed features generated by the image encoder are injected convolutional inductive bias and are fed to the Transformer for image classification bypassing image reconstruction. Meanwhile, we propose a feature aggregation module to fuse the compressed features with the selected intermediate features of the Transformer, and feed the aggregated features to a deconvolutional neural network for image reconstruction. The aggregated features can obtain the long-term information from the self-attention mechanism of the Transformer and improve the compression performance. The rate-distortion-accuracy optimization problem is finally solved by a two-step training strategy. Experimental results demonstrate the effectiveness of the proposed model in both the image compression and the classification tasks.

Network Anomaly Traffic Detection Algorithm Based on RIC-SC-DeCN.

In the research of network abnormal traffic detection, in view of the characteristics of high dimensionality and redundancy in traffic data and the loss of original information caused by the pooling operation in the convolutional neural network, which leads to the problem of unsatisfactory detection effect, this paper proposes a network abnormal traffic detection algorithm based on RIC-SC-DeCN to improve the above problems. Firstly, a recursive information correlation (RIC) feature selection mechanism is proposed, which reduces data redundancy through the maximum information correlation feature selection algorithm and recursive feature elimination method. Secondly, a skip-connected deconvolutional neural network model (SC-DeCN) is proposed to reduce the information loss by reconstructing the input signal. Finally, the RIC mechanism and the SC-DeCN model are merged to form a network abnormal traffic detection algorithm based on RIC-SC-DeCN. The experimental results on the CIC-IDS-2017 dataset show that the RIC feature selection mechanism proposed in this paper has the highest accuracy when using MSCNN as the detection model compared to the other three, which can reach 96.22%. Compared with the other five models, the SC-DeCN model has the highest detection accuracy, while the model training time is moderate and can reach 96.55%. Compared with the SC-DeCN model, the RIC-SC-DeCN model reduces the overall training time by 45.50%, while the accuracy rate is increased to 97.68%. It shows that the algorithm proposed in this paper has a good detection effect in the detection of network abnormal traffic.

Data-Driven Modeling of Geometry-Adaptive Steady Heat Convection Based on Convolutional Neural Networks

Heat convection is one of the main mechanisms of heat transfer, and it involves both heat conduction and heat transportation by fluid flow; as a result, it usually requires numerical simulation for solving heat convection problems. Although the derivation of governing equations is not difficult, the solution process can be complicated and usually requires numerical discretization and iteration of differential equations. In this paper, based on neural networks, we developed a data-driven model for an extremely fast prediction of steady-state heat convection of a hot object with an arbitrary complex geometry in a two-dimensional space. According to the governing equations, the steady-state heat convection is dominated by convection and thermal diffusion terms; thus the distribution of the physical fields would exhibit stronger correlations between adjacent points. Therefore, the proposed neural network model uses convolutional neural network (CNN) layers as the encoder and deconvolutional neural network (DCNN) layers as the decoder. Compared with a fully connected (FC) network model, the CNN-based model is good for capturing and reconstructing the spatial relationships of low-rank feature spaces, such as edge intersections, parallelism, and symmetry. Furthermore, we applied the signed distance function (SDF) as the network input for representing the problem geometry, which contains more information compared with a binary image. For displaying the strong learning and generalization ability of the proposed network model, the training dataset only contains hot objects with simple geometries: triangles, quadrilaterals, pentagons, hexagons, and dodecagons, while the testing cases use arbitrary and complex geometries. According to the study, the trained network model can accurately predict the velocity and temperature field of the problems with complex geometries, which has never been seen by the network model during the model training; and the prediction speed is two orders faster than the CFD. The ability of accurate and extremely fast prediction of the network model suggests the potential of applying reduced-order network models to the applications of real-time control and fast optimization in the future.

Two-dimensional film-cooling effectiveness prediction based on deconvolution neural network

For film cooling in high-pressure turbines, it is vital to predict the temperature distribution and film cooling effectiveness on the blade surface downstream of the cooling hole. This temperature distribution and film cooling effectiveness depend on the interaction between the hot mainstream and the coolant jet. However, it is difficult to correlate accurately due to the complex mechanism. Based on deep learning techniques, a theoretic model using Deconvolutional Neural Network (Deconv NN) was developed to model the non-linear and high-dimensional mapping between coolant jet parameters and the surface temperature distribution on a flat plate. Computational Fluid Dynamics (CFD) was utilized to provide data for the training models. The input of the model includes blowing ratio, density ratio, hole inclination angle and hole diameters etc. With rigorous testing and validation, it is found that the predicted results are in good agreement with results from CFD. It is compared against the existing semi-empirical correlations and other machine learning techniques, such as support vector machine method. Dataset with different size is tested. The results suggest that the performance and robustness of Deconv NN is much better than other methods.

Three-dimensional tire-pavement contact stresses prediction by deep learning approach

ABSTRACT The demand for fast and accurate tire-pavement contact modelling is becoming increasingly prevalent with the advancement of pavement design and finite-element modelling. This paper presents a tool for fast and accurate prediction of non-uniform tire-pavement contact stresses utilising deep learning. Two truck tires, under various wheel loading, inflation pressure, and slip ratio conditions, were considered. The developed deep learning model, ContactNet, is a deconvolutional neural network consisting of two fully connected layers, one reshape layer, and five deconvolution layers with millions of neurons. Two validated finite-element truck tire models were used to generate a contact stresses database with 1800 simulated results. The database was then used for training and testing of the ContactNet. The ContactNet resulted in average errors of 0.80%, 0.77%, 0.90%, and 0.57% in predicting maximum vertical stress, effective contact area, maximum longitudinal stress, and maximum transverse stress. The mean absolute error of the ContactNet prediction is 0.91 kPa. This significantly outperformed four conventional machine-learning regression methods investigated in this study, including polynomial regression, k-nearest neighbours, multi-layer perceptron, and random forests.

Deblurring of Images using Novel Deconvolutional Neural Network (DNN) Algorithm to Enhance the Accuracy and Comparing with Richardson-Lucy Deconvolution Algorithm (RLD)

Aim- Machine learning techniques are rapidly used in the area of digital image processing research due to its impressive results in deconvolution and deblurring of images. The objective of this study is to evaluate the performance of Novel DNN algorithm in deblurring of images by comparing it with the RLD algorithm. Materials and Methods - Novel Deconvolutional Neural Network (DNN) and Richardson-Lucy Deconvolution (RLD) algorithms were implemented to deblur the input images upto 256 pixels range. These algorithms were implemented to enhance the accuracy rate of deblurred images using MATLAB Software and analyzed by collecting the dataset of 40 samples with 80% of pretest power. Results - From the MATLAB simulation result, DNN achieves image deblurring rate with 98% accuracy and RLD method achieves image deblurring rate with 92% accuracy. The significance value obtained as (P < 0.002). Conclusion - Novel DNN classifier appears to have better accuracy compared to RLD Classifier.

Time-variant prediction of flow over an airfoil using deep neural network

In this article, we propose an unsteady data-driven reduced order model (ROM) (surrogate model) for predicting the velocity field around an airfoil. The network model applies a convolutional neural network (CNN) as the encoder and a deconvolutional neural network (DCNN) as the decoder. The model constructs a mapping function between temporal evolution of the pressure signal on the airfoil surface and the surrounding velocity field. For improving the model performance, the input matrix is designed to further incorporate the information of the Reynolds number, the geometry of the airfoil, and the angle of attack. The DCNN works as the decoder for better reconstructing the spatial and temporal information of the features extracted by the CNN encoder. The training and testing datasets of flow fields under different conditions are obtained by solving the Navier–Stokes equations using the computational fluid dynamics method. After model training, the neural network based ROM shows accurate and dramatically fast predictions on the flow field of the testing dataset with extended angles of attack and Reynolds numbers. According to the current study, the neural network-based ROM has exhibited attractive potentials on ROM of the unsteady fluid dynamic problem, and the model can potentially serve on investigating flow control or optimization problems in the future.

Unsteady reduced-order model of flow over cylinders based on convolutional and deconvolutional neural network structure

In this paper, we propose a neural network based reduced-order model for predicting the unsteady flow field over single/multiple cylinders. The neural network model constructs a mapping function between the temporal evolution of the pressure signal on the cylinder surface and the surrounding velocity field, where Convolutional Neural Network (CNN) layers are used as the encoder and deconvolutional neural network layers are used as the decoder. Compared with the network model with a fully connected (FC) decoder, the model with the deconvolution connected (DC) decoder is good for capturing and reconstructing the spatial relationships of low-rank feature spaces, such as edge intersections, parallelism, and symmetry, while the fluid flow, which is described by Navier–Stokes equations containing convection and diffusion terms, displays outstanding features of locality. In this article, the performance of the network models with the FC decoder and the DC decoder is evaluated by studying the problem of flow over a single cylinder first, and then the complexity of the flow structure of the studied problems is enhanced by increasing the number of cylinders and the Reynolds number. The results indicate that both the CNN-FC decoder model and CNN-DC decoder model achieve fast and accurate prediction on the velocity field, and the CNN-DC decoder model gives more robust and precise performance for all studied problems.

Sleep stage classification for child patients using DeConvolutional Neural Network

Studies from the literature show that the prevalence of sleep disorder in children is far higher than that in adults. Although much research effort has been made on sleep stage classification for adults, children have significantly different characteristics of sleep stages. Therefore, there is an urgent need for sleep stage classification targeting children in particular. Our method focuses on two issues: The first is timestamp-based segmentation (TSS) to deal with the fine-grained annotation of sleep stage labels for each timestamp. Compared to this, popular sliding window approaches unnecessarily aggregate such labels into coarse-grained ones. We utilize DeConvolutional Neural Network (DCNN) that inversely maps features of a hidden layer back to the input space to predict the sleep stage label at each timestamp. Thus, our DCNN can yield better classification performances by considering labels at numerous timestamps. The second issue is the necessity of multiple channels. Different clinical signs, symptoms or other auxiliary examinations could be represented by different Polysomnography (PSG) recordings, so all of them should be analyzed comprehensively. We therefor exploit multivariate time-series of PSG recordings, including 6 electroencephalograms (EEGs) channels, 2 electrooculograms (EOGs) channels (left and right), 1 electromyogram (chin EMG) channel and two leg electromyogram channels. Our DCNN-based method is tested on our SDCP dataset collected from child patients aged from 5 to 10 years old. The results show that our method yields the overall classification accuracy of 84.27% and macro F1-score of 72.51% which are higher than those of existing sliding window-based methods. One of the biggest advantages of our DCNN-based method is that it processes raw PSG recordings and internally extracts features useful for accurate sleep stage classification. We examine whether this is applicable for sleep data of adult patients by testing our method on a well-known public dataset Sleep-EDFX. Our method achieves the average overall accuracy of 90.89% which is comparable to those of state-of-the-art methods without using any hand-crafted features. This result indicates the great potential of our method because it can be generally used for timestamp-level classification on multivariate time-series in various medical fields. Additionally, we provide source codes so that researchers can reproduce the results in this paper and extend our method.

Deconvolutional neural network for image super-resolution

This study builds a fully deconvolutional neural network (FDNN) and addresses the problem of single image super-resolution (SISR) by using the FDNN. Although SISR using deep neural networks has been a major research focus, the problem of reconstructing a high resolution (HR) image with an FDNN has received little attention. A few recent approaches toward SISR are to embed deconvolution operations into multilayer feedforward neural networks. This paper constructs a deep FDNN for SISR that possesses two remarkable advantages compared to existing SISR approaches. The first improves the network performance without increasing the depth of the network or embedding complex structures. The second replaces all convolution operations with deconvolution operations to implement an effective reconstruction. That is, the proposed FDNN only contains deconvolution layers and learns an end-to-end mapping from low resolution (LR) to HR images. Furthermore, to avoid the oversmoothness of the mean squared error loss, the trained image is treated as a probability distribution, and the Kullback–Leibler divergence is introduced into the final loss function to achieve enhanced recovery. Although the proposed FDNN only has 10 layers, it is successfully evaluated through extensive experiments. Compared with other state-of-the-art methods and deep convolution neural networks with 20 or 30 layers, the proposed FDNN achieves better performance for SISR.

Efficient Deconvolution Architecture for Heterogeneous Systems-on-Chip.

Today, convolutional and deconvolutional neural network models are exceptionally popular thanks to the impressive accuracies they have been proven in several computer-vision applications. To speed up the overall tasks of these neural networks, purpose-designed accelerators are highly desirable. Unfortunately, the high computational complexity and the huge memory demand make the design of efficient hardware architectures, as well as their deployment in resource- and power-constrained embedded systems, still quite challenging. This paper presents a novel purpose-designed hardware accelerator to perform 2D deconvolutions. The proposed structure applies a hardware-oriented computational approach that overcomes the issues of traditional deconvolution methods, and it is suitable for being implemented within any virtually system-on-chip based on field-programmable gate array devices. In fact, the novel accelerator is simply scalable to comply with resources available within both high- and low-end devices by adequately scaling the adopted parallelism. As an example, when exploited to accelerate the Deep Convolutional Generative Adversarial Network model, the novel accelerator, running as a standalone unit implemented within the Xilinx Zynq XC7Z020 System-on-Chip (SoC) device, performs up to 72 GOPs. Moreover, it dissipates less than 500 mW@200 MHz and occupies ~5.6%, ~4.1%, ~17%, and ~96%, respectively, of the look-up tables, flip-flops, random access memory, and digital signal processors available on-chip. When accommodated within the same device, the whole embedded system equipped with the novel accelerator performs up to 54 GOPs and dissipates less than 1.8 W@150 MHz. Thanks to the increased parallelism exploitable, more than 900 GOPs can be executed when the high-end Virtex-7 XC7VX690T device is used as the implementation platform. Moreover, in comparison with state-of-the-art competitors implemented within the Zynq XC7Z045 device, the system proposed here reaches a computational capability up to ~20% higher, and saves more than 60% and 80% of power consumption and logic resources requirement, respectively, using ~5.7× fewer on-chip memory resources.

Super-Resolution Ultrasound Lamb Wave NDE Imaging of Anisotropic Airplane Laminates via Deconvolutional Neural Network

A deconvolutional neural network (DNN) is integrated into the ultrasound (UT) pulse-echo Lamb wave nondestructive evaluation (NDE) imaging technique to achieve subwavelength super-resolution defect images for the application of anisotropic composite airplane-laminated structures. First, numerical simulation has been performed to simulate the multilayer velocity–frequency dispersion relation of the symmetric and antisymmetric Lamb wave modes. Then, a super-resolution DNN is developed with all Lamb wave modes as the convolutional kernels to obtain the subwavelength defects image of the laminated structures. After that, the effectiveness of the pulse-echo Lamb wave NDE imaging technique is verified by the experimental test of a standard aluminum metal plate with three calibration holes of 2/3/5 mm diameters at the UT center frequency of 2 MHz and a line defect. Finally, the experiment of the pulse-echo Lamb wave NDE imaging technique is carried out with an L-joint coupon sample made of anisotropic graphite–epoxy airplane materials. The comparison between the experimental result and the conventional pitch-catch C-scan shows that the Lamb wave NDE technique can reveal more details of the defects, indicating its promising application in anisotropic layers’ structures defects inspection.