Previous Time Step Research Articles

Data-driven prediction of unsteady pressure distributions based on deep learning

In the present work, an efficient Reduced-Order Model is developed for the prediction of motion-induced unsteady pressure distributions. The model is trained on the basis of synthetic data generated by full-order Computational Fluid Dynamics (CFD) simulations. The nonlinear identification task is to predict a snapshot representing the pressure distribution for the current time step based on respective snapshots of previous time steps and applied excitation. Once a Reduced-Order Model is conditioned on training data, it can predict sequences of the pressure distribution in a recurrent manner based on the excitation signal. Hence, it is able to capture the motion-induced nonlinear unsteady aerodynamics for a given configuration at fixed free-stream conditions. In this way, computationally extensive CFD simulations can be substituted by the application of the more efficient Reduced-Order Model. The nonlinear behavior of the aerodynamic system is captured based on a deep convolutional neural network. The performance of the Reduced-Order Model is demonstrated based on the LANN (Lockheed-Georgia, Air Force Flight Dynamics Laboratory, NASA-Langley and NLR) wing performing high-amplitude pitching motion in transonic flow. The unsteady aerodynamics of the considered test case is dominated by nonlinear effects due to complex moving shock structures both on the upper and lower surface of the wing. The Reduced-Order Model yields a superior prediction accuracy at a speed-up of more than three orders of magnitude compared to the employed CFD method.

A method for capacity estimation of lithium-ion batteries based on adaptive time-shifting broad learning system

Accurate capacity estimation of lithium-ion batteries can improve the safety and reliability of equipment. Deep learning provides a new approach for capacity estimation. However, it is still difficult to adapt to the changes in different stages of capacity degradation due to many and complex parameters in deep structure network and the time-invariant model parameters. Considering that Broad Learning System is a neural network independent of deep structure, which has a simple single hidden layer structure and fewer parameters, this paper proposes an Adaptive Time-shifting Broad Learning System (ATBLS) capacity estimation model. For input layer, the input data of each time step is combined with the output of previous time step to get a new input, so as to provide local information to hidden unit of the network and provide guidance for local parameters. For hidden layer, the hidden unit of each time step is weighted-fused with the hidden unit of previous time step to update the parameters, so as to reflect the long-term dynamics of sequence. The experiment are conducted on three different data sets. And the effectiveness of ATBLS is verified by comparing with other methods.

Real-time core temperature prediction of prismatic automotive lithium-ion battery cells based on artificial neural networks

For modeling of the non-linear heat generation and thermal effects in Li-ion batteries, artificial neural networks are a great solution to represent the thermal behavior for the battery management system in an electric vehicle. Several studies have proven the high accuracy with large benefits in computing-time and complexity compared to detailed electrochemical–thermal models. Commonly used feedforward networks need to prove suitability for dynamic applications but are always limited due to missing information about the previous time steps or need external sensor information. In this work, a novel Nonlinear AutoRegressive with eXogenous (NARX)-network is developed and parameterized for a large 25Ah prismatic cell. The NARX is compared to a feedforward using the same general structure and input data in terms of training, validation behavior, long-term prediction and dynamic driving application. Both ANN approaches prove to be adequate for the temperature prediction with an accuracy within 1K during long-term prediction of 10h. Additionally, in a BEV application with real-time requirements the thermal models predicting the dynamic temperature behavior with high precision and robustness without even a temperature input in case of the NARX-approach.

Improved thermal multiple-relaxation-time lattice Boltzmann model for liquid-vapor phase change

In this paper, an improved thermal multiple-relaxation-time (MRT) lattice Boltzmann (LB) model is proposed for simulating liquid-vapor phase change. A temperature equation is first derived for liquid-vapor phase change, where the latent heat of vaporization is decoupled with the equation of state. Therefore, the latent heat of vaporization can be arbitrarily specified in practice, which significantly improves the flexibility of the present LB model for liquid-vapor phase change. The Laplacian term of temperature is avoided in the proposed temperature equation and the gradient term of temperature is calculated through a local scheme. To solve the temperature equation accurately and efficiently, an improved MRT LB equation with nondiagonal relaxation matrix is developed. The implicit calculation of the temperature, caused by the source term and encountered in previous works, is avoided by approximating the source term with its value at the previous time step. As demonstrated by numerical tests, the results by the present LB model agree well with analytical results, experimental results, or the results by the finite difference method where the fourth-order Runge-Kutta method is employed to implement the discretization of time.

Sequential sparse Bayesian learning for beamforming

A sequential Bayesian method for beamforming is presented for the estimation of the directions of arrivals (DOAs) of source signals which are varying over time. The sparse Bayesian learning (SBL) uses prior information of unknown source amplitudes as a multi-variate Gaussian with zero-mean and time-varying variance parameters. For sequential processing, we utilize the unknown variance as statistical information across time and propose a sequential SBL-based method to improve time-varying DOA estimation performance. The suggested method has two steps. A prediction step calculates the prior distribution of the current variance parameter from the variance parameter estimated at the previous time step, and an update step incorporates the current measurements. This SBL approach with sequential processing provides high-resolution for time-varying DOA tracking. We evaluate the proposed method using simulated data and real data from the SWellEx-96 experiment.

CSMA/CA channel hopping in IoT environment toward intelligent multi-user diversity

In massive IoT (Internet of Things) environment, which constitutes a goal of the 5G standard, the communication channel of an MTC (machine-type-communication) device with narrowband transmission has problems that are deep fading and interference with the transmission performance. For improved communications in IoT environment, researchers aimed to increase the transmission throughput by using larger contention windows between channel hops when the probability of successful data communication drops during the previous time step to prevent the rapid increase in the number of connected devices. Hence, researchers propose a model to change the interframe spacing according to the number of contention windows for hopping and data prioritization by quickly determining network traffic variations in a massive network environment. Specifically, researchers propose CSMA/CA (carrier-sense multiple access with collision avoidance) channel hopping to prevent deep fading and obtain multiuser diversity in multichannel environments as well as improve the channel quality. Researchers confirmed that the proposed CSMA/CA channel hopping provides a data transfer rate about twice that of conventional TDMA (time-division multiple access) channel hopping for 60 connected devices. As a result of analyzing data transfer rates in channel hopping, it was shown that the collision probability is not high when the number of nodes is small, so the number of channels and contention window size are not significantly affected, but as the number of nodes increases, the number of channels and contention window size are greatly affected.

Predicting transient diesel engine NOx emissions using time-series data preprocessing with deep-learning models

Deep-learning models were developed and evaluated for predicting the engine-out emission of NOx—one of the main pollutants emitted from diesel engines—under transient conditions, that is, the Worldwide Harmonized Light Vehicles Test Procedure (WLTP). Phenomena of transient conditions are difficult to predict accurately via the conventional modeling approaches. Two algorithms—the deep neural network (DNN) and long short-term memory (LSTM)—were evaluated regarding the accuracy and calculation time. Training was performed using measured data, and the results indicated that the LSTM model ( R2 = 0.9777, RMSE = 20.6 ppm) was more accurate than the DNN model ( R2 = 0.9671, RMSE = 25.5 ppm). However, the DNN model had a significantly higher computation speed (0.36 s) than the LSTM model (1381.0 s). Data preprocessing was performed to insert time-related information into the data; the DNN model trained with the measured data lacked this feature. Data were preprocessed by the calculation of the weighted average of previous timestep data to the current-timestep data. The weighted average was calculated according to various ratios, for example, 7:3, 6:4, 5:5, 4:6, and 3:7. By applying a 7:3 weighted average for training the DNN model, the accuracy of the DNN model was achieved to an R2 value of 0.9741, and RMSE 22.8 ppm (only R2 value 0.0036 smaller and RMSE 2.2 ppm larger than the LSTM model) without sacrificing the calculation speed. The results of this study suggest that data preprocessing of the DNN model is an effective method for achieving accuracy as high as that of the LSTM model. The developed DNN model for the NOx emission prediction can be used as a virtual sensor for real-time prediction owing to its accuracy and computation time.

Detecting dynamic community by fusing network embedding and nonnegative matrix factorization

Many systems in nature and society, ranging from types of cancers to engineering, can be modeled as temporal networks. Identification of dynamic communities is fundamental for investigating the mechanisms of systems. Compared with community detection in static networks, dynamic community detection is challenging because it simultaneously leverages clustering accuracy with clustering drift. Current methods employ the temporal smoothness framework to balance these two items, but the inaccuracy and complexity of these methods have been criticized. To overcome these issues, a novel algorithm called NE2NMF for dynamic community detection is proposed by fusing Network Embedding based Evolutionary Nonnegative Matrix Factorization. Specifically, we use third-order smoothness to characterize clustering drift by fusing snapshots of previous, current, and subsequent time steps, thereby improving the accuracy of the characterization of dynamic communities. Moreover, to reduce time complexity, we decompose the network embedding matrix, rather than the adjacency matrix of each snapshot on the basis of the equivalence between evolutionary non-negative matrix factorization (ENMF) and network embedding. Results demonstrate that NE2NMF outperforms state-of-the-art approaches in terms of accuracy and robustness.

Ion temperature gradient control using reinforcement learning technique

Plasma with an internal transport barrier (ITB) is desirable for a steady-state tokamak reactor because of its high confinement quality and high bootstrap current fraction. However, the local pressure gradient tends to be steep and the plasma often becomes unstable. In this study, an ion temperature gradient control system based on neutral beam injection (NBI) is developed using the reinforcement learning technique. The response characteristics of an ion temperature gradient to NBI are non-linear and sensitive to experimental conditions, which makes it difficult to develop a robust control system. Our control system is trained for plasmas with a wide range of ITB strengths. Using the reinforcement learning technique, the system acquires a robust control feature through several thousand iterations of trial and error in an integrated transport simulation hosted by TOPICS. The control system is composed of neural networks (NNs) whose input variables are the ion temperature gradient, the current NBI power, and the NBI powers for several previous control time steps. The trained system can determine a control output which is suitable for the response characteristics inferred from the input variables. The trained control system is tested in the TOPICS simulation using plasma models based on two experimental plasmas of JT-60U with different ITB strengths. It is shown that the ion temperature gradient can be appropriately controlled for both plasmas, which supports the expectation that this system is applicable to real experiments.

Sequential sparse Bayesian learning for time-varying direction of arrival.

This paper presents methods for the estimation of the time-varying directions of arrival (DOAs) of signals emitted by moving sources. Following the sparse Bayesian learning (SBL) framework, prior information of unknown source amplitudes is modeled as a multi-variate Gaussian distribution with zero-mean and time-varying variance parameters. For sequential estimation of the unknown variance, we present two sequential SBL-based methods that propagate statistical information across time to improve DOA estimation performance. The first method heuristically calculates the parameters of an inverse-gamma hyperprior based on the source signal estimate from the previous time step. In addition, a second sequential SBL method is proposed, which performs a prediction step to calculate the prior distribution of the current variance parameter from the variance parameter estimated at the previous time step. The SBL-based sequential processing provides high-resolution DOA tracking capabilities. Performance improvements are demonstrated by using simulated data as well as real data from the SWellEx-96 experiment.

Approaches for Efficiently Detecting Frontier Cells in Robotics Exploration.

Many robot exploration algorithms that are used to explore office, home, or outdoor environments, rely on the concept of frontier cells. Frontier cells define the border between known and unknown space. Frontier-based exploration is the process of repeatedly detecting frontiers and moving towards them, until there are no more frontiers and therefore no more unknown regions. The faster frontier cells can be detected, the more efficient exploration becomes. This paper proposes several algorithms for detecting frontiers. The first is called Naïve Active Area (NaïveAA) frontier detection and achieves frontier detection in constant time by only evaluating the cells in the active area defined by scans taken. The second algorithm is called Expanding-Wavefront Frontier Detection (EWFD) and uses frontiers from the previous timestep as a starting point for searching for frontiers in newly discovered space. The third approach is called Frontier-Tracing Frontier Detection (FTFD) and also uses the frontiers from the previous timestep as well as the endpoints of the scan, to determine the frontiers at the current timestep. Algorithms are compared to state-of-the-art algorithms such as Naïve, WFD, and WFD-INC. NaïveAA is shown to operate in constant time and therefore is suitable as a basic benchmark for frontier detection algorithms. EWFD and FTFD are found to be significantly faster than other algorithms.

Deep residual U-net convolution neural networks with autoregressive strategy for fluid flow predictions in large-scale geosystems

The inherent complexity of the fluid flow in subsurface systems brings potential inevitable uncertainty in their characterization. Computationally intensive high-dimensional inversion problems often emerge in solving the fluid flow problems of various scenarios, which required to be probed. To improve the efficiency of solving such problems, surrogate strategies are widely used to quantify the uncertainty of underground multiphase flow models. In this paper, a deep learning surrogate model is developed for predicting the time-dependent dynamic multiphase flow in a two-dimensional (2D) channelized geological system. The surrogate model is combined with a residual U-net and an autoregressive strategy, which considers the output at the previous time step as input and predict the output at the current time step. The residual U-net has a symmetric network structure similar to U-net and contains extra residual units. The rich skip connections in the network can promote information dissemination and achieve better prediction performance with fewer parameters. We demonstrated the performance of the autoregressive residual U-net (AR-Runet) for predicting the migration of solute transport in heterogeneous 2D binary model. The result shows the AR-Runet surrogate model can provide an accurate approximation of saturation and pressure fields at different times. We also have demonstrated that with the autoregressive strategy this network can achieve similar predict results with relatively less training data. The performance of the AR-Runet network is also compared with the autoregressive Dense net (AR-Dense). The findings indicate that the AR-Runet can provide effective measures for developing surrogate model and uncertainty analysis in dynamic multiphase flow predictions of subsurface systems.

A Method of the Lagrange Polynomial Interpolation with Weighting Factors for Reducing Computational Time

This paper proposes a novel extrapolation method known as the Lagrange polynomial interpolation with weighting factors (LPI-WF) to determine initial guess values in each time step for solving equation systems of transient problems using iterative methods. The LPI-WF method is developed from the Lagrange polynomial interpolation to determine the direct extrapolation values and multiply them with the weighting factors. The weighting factors are calculated by considering the number of the previous time steps involved in the extrapolation and the duration between the present time step and the previous time steps. Thus, the LPI-WF method is proper for use with highorder temporal schemes. The key advantages of the LPI-WF method are that the computational time required to achieve the steady-state condition of the transient problems is reduced and the computation codes with the LPI-WF method is more stable than without the LPI-WF method at high time step values. A performance test of the LPI-WF method is carried out by comparing the computational time based on the “lid-driven cavity flow” problem for Reynolds numbers of 1000 and 5000. The test result shows that the computational time of the problem when the LPI-WF method is adopted can be reduced up to 10.46 % compared to the conventional method.

Two second-order and linear numerical schemes for the multi-dimensional nonlinear time-fractional Schrödinger equation

This paper presents two second-order and linear finite element schemes for the multi-dimensional nonlinear time-fractional Schrodinger equation. In the first numerical scheme, we adopt the L2-1σ formula to approximate the Caputo derivative. However, this scheme requires storing the numerical solution at all previous time steps. In order to overcome this drawback, we develop the $\mathcal {F}L2$ -1σ formula to construct the second numerical scheme, which reduces the computational storage and cost. We prove that both the L2-1σ and $\mathcal {F}L2$ -1σ formulas satisfy the three assumptions of the generalized discrete fractional Gronwall inequality. Furthermore, combining with the temporal-spatial error splitting argument, we rigorously prove the unconditional stability and optimal error estimates of these two numerical schemes, which do not require any time-step restrictions dependent on the spatial mesh size. Numerical examples in two and three dimensions are given to illustrate our theoretical results and show that the second scheme based on $\mathcal {F}L2$ -1σ formula can reduce CPU time significantly compared with the first scheme based on L2-1σ formula.

Mathematical modeling of dynamic related processes based on the finite element techniques

In the process of investigation the functioning of information systems with the help of mathematical modeling, problems can arise that are solved using finite element techniques. For example, In the present work we solved the axially symmetric dynamic problem of related phenomena under microscale thermal loading. The statement of the problem includes: Cauchy relations, equations of motion, heat conduction equation, initial conditions, thermal and mechanical boundary conditions. The nonlinear behavior of a material is described by the unified model of flow. The problem is solved numerically by the time step integration method, iterative method and finite element method. Equations of the evolution for the inelastic flow model are integrated by the second-order Euler implicit method with the use of the rule of «middle point». The system of nonlinear transcendental equations obtained in each time step is solved by the method of simple iteration. To accelerate convergence, we apply the Stephensen – Aitken procedure. The equations of motion are integrated by the Newmark method, whereas the heat-conduction equation is integrated by the first-order implicit method. The problem of heat-conduction is linearized by finding the temperature-dependent thermal characteristics according to the temperature distribution in the previous time step or previous iteration. The main results obtained in the work are the following: quantitative estimations of temperature effects of thermostructural-mechanical coupling, which are caused by volumetric thermoelastic effect, dissipation of mechanical energy and hidden heat are obtained.

Convergence analysis of Fourier pseudo-spectral schemes for three-dimensional incompressible Navier-Stokes equations

<p style='text-indent:20px;'>The stability and convergence of the Fourier pseudo-spectral method are analyzed for the three dimensional incompressible Navier-Stokes equation, coupled with a variety of time-stepping methods, of up to fourth order temporal accuracy. An aliasing error control technique is applied in the error estimate for the nonlinear convection term, while an a-priori assumption for the numerical solution at the previous time steps will also play an important role in the analysis. In addition, a few multi-step temporal discretization is applied to achieve higher order temporal accuracy, while the numerical stability is preserved. These semi-implicit numerical schemes use a combination of explicit Adams-Bashforth extrapolation for the nonlinear convection term, as well as the pressure gradient term, and implicit Adams-Moulton interpolation for the viscous diffusion term, up to the fourth order accuracy in time. Optimal rate convergence analysis and error estimates are established in details. It is proved that, the Fourier pseudo-spectral method coupled with the carefully designed time-discretization is stable provided only that the time-step and spatial grid-size are bounded by two constants over a finite time. Some numerical results are also presented to verify the established convergence rates of the proposed schemes.</p>

Time-Varying DOA Tracking Algorithm Based on Generalized Labeled Multi-Bernoulli

Direction of arrival (DOA) tracking for multi-sources is a hot issue in array signal processing. To deal with the problem that sources DOA and their number are time-varying, a DOA tracking algorithm based on Generalized Labeled Multi-Bernoulli (GLMB) filter is proposed. Since the measurement value has only one set of data, the measurement association mapping (MAM) does not match, which leads to deviations in the GLMB filter update step. In this regard, we used the estimated sources number of the previous time step as the measurement number of the current time step, and successfully achieved MAM matching. Subsequently, particle filtering is used to approximate the posterior distribution of DOA, where the particle likelihood function can be calculated by the multi-signal classification (MUSIC) spatial spectrum function. In addition, by exponentially weighting the likelihood function, the number of particles in the high likelihood region of the posterior distribution increases, which makes the GLMB filter pruning and merging operations more effective. Simulation results show that the method is better than the probability hypothesis density DOA (PHD-DOA) algorithm in tracking state sources and estimating the number of targets.

KeyMemoryRNN: A Flexible Prediction Framework for Spatiotemporal Prediction Networks

Most previous recurrent neural networks for spatiotemporal prediction have difficulty in learning the long-term spatiotemporal correlations and capturing skip-frame correlations. The reason is that the recurrent neural networks update the memory states only using information from the previous time step node and the networks tend to suffer from gradient propagation difficulties. We propose a new framework, KeyMemoryRNN, which has two contributions. The first is that we propose the KeyTranslate Module to extract the most effective historical memory state named keyword state, and we propose the KeyMemory-LSTM which uses the keyword state to update the hidden state to capture the skip-frame correlation. In particular, KeyMemoryLSTM has two training stages. In the second stage, KeyMemoryLSTM adaptively skips the update of sometime step nodes to build a shorter memory information flow to alleviate the difficulty of gradient propagation to learn the long-term spatiotemporal correlations. The second is that both KeyTranslate Module and KeyMemoryLSTM are flexible additional modules, so we can apply them to most RNN-based prediction networks to build KeyMemoryRNN with different base network. The KeyMemoryRNN achieves the state-of-the-art on three spatiotemporal prediction tasks, and we provide ablation studies and memory analysis to verify the effectiveness of KeyMemoryRNN.

A Predictive Energy Management Strategy for Multi-mode Plug-in Hybrid Electric Vehicle based on Long short-term Memory Neural Network

Abstract The plug-in hybrid electric vehicles (PHEV) provide a promising solution to increasingly severe tailpipe pollution and make possible the high-efficiency utilization of energy. To strengthen the overall fuel economy of a multi-mode PHEV, this paper proposes a real-time predictive energy management strategy (EMS) based on the Long Short-term Memory (LSTM) neural network. In order to forecast the short-term vehicle velocity with speed of previous time steps, a LSTM network is built and trained using speed profile of multiple representative driving cycles; a model predictive control architecture solved by dynamic programming (DP) in prediction horizon is also structured to realize successive online optimization of power allocation between the internal combustion engine and electric motors. The simulation of proposed strategy, convention adaptive equivalent consumption minimization strategy (A-ECMS) and the offline global optimization are carried out with UDDS and HWFET cycles to investigate the effectiveness and adaptivity of the proposed strategy, the result shows that the near-optimality of fuel economy is realized, and the fuel economy is elevated compared with the A-ECMS, therefor the potential for practical application of the proposed strategy is proved.

Non-Autoregressive vs Autoregressive Neural Networks for System Identification

The application of neural networks to non-linear dynamic system identification tasks has a long history, which consists mostly of autoregressive approaches. Autoregression, the usage of the model outputs of previous time steps, is a method of transferring a system state between time steps, which is not necessary for modeling dynamic systems with modern neural network structures, such as gated recurrent units (GRUs) and Temporal Convolutional Networks (TCNs). We compare the accuracy and execution performance of autoregressive and non-autoregressive implementations of a GRU and TCN on the simulation task of three publicly available system identification benchmarks. Our results show, that the non-autoregressive neural networks are significantly faster and at least as accurate as their autoregressive counterparts. Comparisons with other state-of-the-art black-box system identification methods show, that our implementation of the non-autoregressive GRU is the best performing neural network-based system identification method, and in the benchmarks without extrapolation, the best performing black-box method.