Backward Propagation Research Articles

Deep learning-based vibration stress and fatigue-life prediction of a battery-pack system

The primary concerns of the automotive industry are structural integrity and battery-pack system (BPS) reliability. To ascertain the appropriate thickness of critical BPS components (e.g., the long bracket, crossbeam, and bottom shell), engineers are required to perform several simulations of finite element (FE) analysis. This procedure is laborious and requires a significant amount of time. This study introduces a very effective approach for predicting vibration-induced stress and fatigue life, intending to enhance dependability in the design process. Firstly, a three-layer lithium battery model is utilized to investigate the impact of vibration on the maximal Mises stress at various states of charge. This information is used to model the BPS, including the batteries. The nonlinear FE model of the BPS is validated using the modal test results. Secondly, a sensitivity analysis is performed to ascertain the impact of component thickness on the maximum Mises stress. For vibration simulations, multiple design variables are selected to collect data. Next, the nonlinear relationship between inputs (thicknesses of critical components) and outputs (maximum Mises stress and minimum fatigue life) is described using a deep learning (DL) modeling framework with forward and backward propagation. Finally, the accuracy of the DL model is assessed by measuring error functions and comparing its performance to six commonly employed methods. Furthermore, the inclusion of Gaussian noise is employed to assess the model’s robustness and ability to generalize. Additionally, the establishment of fatigue-life and stress boundaries serves to offer designers valuable insights. The results indicate that the proposed method for predicting vibration stress and fatigue life is highly efficient and cost-effective, making it useful for the design of a robust and reliable BPS.

Monochromatic wave equation reflection traveltime inversion

Wave equation reflection traveltime inversion (WRTI) is a fundamental method for interrogating the deep subsurface macrovelocity model. The time-domain WRTI calculates the gradient through the crosscorrelation of the source background and receiver scattered wavefields and the crosscorrelation of the source scattered and receiver background wavefields. During the backward propagation of the adjoint source, access to the source background and scattered snapshots of the entire propagation time is required. Because the source and receiver wavefields are extrapolated in opposite temporal directions, these source background and scattered wavefields are either stored in memory or reconstructed by additional wavefield extrapolation. This leads to a heavy memory burden and/or high computational cost. To address these challenges, we develop a new frequency-domain WRTI that is based on a monofrequency. By using the wavenumber coverage and gradient analyses, we prove that the monochromatic WRTI can produce comparable wavenumbers with time-domain and multifrequency-based WRTI. Our method directly stores monochromatic wavefields in memory because only one frequency is used for the gradient calculation. Therefore, our method significantly outperforms the conventional time-domain implementation in terms of memory requirement and computational cost. The accuracy of the monochromatic WRTI method is validated by gradient and result comparisons of several models with different complexities. Finally, we determine the effectiveness of our method by applying it to offshore streamer data.

Acoustic absorption of3Dprinted glycol‐modified polyethylene terephthalate composites with organically modified montmorillonite and short carbon fibers: Experimentation andANNbased predictive strategy

AbstractIn this article, the acoustic properties of 3D printed glycol‐modified polyethylene terephthalate (PETG) reinforced with organically modified montmorillonite (OMMT) nanoclay/short carbon fiber (SCF) nanocomposites are experimentally investigated using ASTM E1050‐08 standard. To this end, the sound absorption coefficient (SAC) of different PETG composites was calculated with the aid of two microphone impedance tube, working in the frequency range of 50–6300 Hz. The effect of different weight percentages (wt%) of OMMT nanoclay, SCFs, and 3D printing infill density on the acoustic behavior of PETG nanocomposites is studied. The experimental results reveal that higher wt% of OMMT nanoclay and SCFs has a beneficial effect on sound absorption. Further, the trend of variation of SAC is justified with morphological studies. Also, an artificial neural network (ANN) based prediction methodology to predict SAC is developed using the datasets obtained from the experimentation. Levenberg–Marquardt backward propagation algorithm with 20 neurons trains the ANN model. Using the trained ANN model, the acoustic properties of PETG/OMMT/SCF nanocomposites with different operating frequencies, infill density and wt% of reinforcements are predicted with less than 5% average error. This can be beneficial in eliminating the fabrication and experimentation costs incurred while assessing the acoustic properties of the PETG composites.HighlightsPETG/OMMT nanoclay/SCF composite filaments are fabricated.Acoustic absorption of 3D printed PETG composites are experimentally studied.ANN based predictive tool is developed using experimental data.ANN tool predicts the sound absorption and reduces experimentation cost.Microstructural studies justify the trend of sound absorption.

Forward and backward propagation guided-mode waves of electric dipole resonances in an h-BN metasurface

The generation of a backward propagation (BP) and forward propagation (FP) guided mode in a double ridge h-BN metasurface (DRM) is investigated. h-BN is a hyperbolic metamaterial with two Reststrahlen frequency bands (RBs). The incident wave can excite guided modes in DRM, where the Poynting vector parallel component of the FP mode is identical and the BP mode is opposite to that of the incident wave. Based on the optimized structure, the frequency range of the BP mode for TE waves was found near the type II hyperbolic band (HB II), while for TM waves, it was found in the gap between the range of HB I and HB II in h-BN. To comprehensively understand the physics underlying BP and FP modes, we present the electric and magnetic field intensities in DRM, the electric field profile of beam steering, and the radiated powers of multipole resonances. The electric dipole (ED) moment contributes most significantly to the FP and BP modes, with its power being much greater than that of other multipolar moment modes. A proportion of BP mode power and incident power decreases with increasing incident angle of TE waves. There are two peaks of the proportion with TM waves. The maximum proportion is near 75% ofTE waves and 16% ofTM waves. The DRM has shown promising potential in the field of sensors based on BP mode, with a sensitivity of 3.9675 µm/RIU of TE waves and 5.1479 µm/RIU of TM waves. These findings suggest that DRMs hold significant promise for the development of optical metasurfaces, optical switches, and high-performance sensors.

An Integrated Method for Ship Heading Control Using Motion Model Prediction and Fractional Order Proportion Integration Differentiation Controller

Due to the influence of the natural environment, it is very challenging to control the movement of ships to navigate safely and avoid potential risks induced by external environmental factors, especially for the development of autonomous ships in inland or restricted waterways. In this research, we propose an integrated approach for ship heading control that improves the timeliness and robustness of navigation. Recursive least squares and backward propagation neural networks are utilized to identify the ship motion model parameters under the influence of external factors and predict their development in real time. A particle swarm optimization-integrated Fractional Order Proportion Integration Differentiation (FOPID) controller is then designed based on the dynamically identified motion model to achieve accurate heading control for ships navigating in restricted waterways. A case study was conducted based on the Korea Venture Large Crude Carrier 2 (KVLCC2) model to verify the effectiveness, and a comparison between the conventional FOPID controller and the improved FOPID controller was also conducted. The results indicate that the proposed identification–prediction–optimization FOPID controller has faster speed on stabilization and has higher robustness against external influences, which could provide added value for the development of a motion controller for the autonomous ship for inland and restricted waterway navigation.

Statistical Compact Modeling With Artificial Neural Networks

This work proposes a statistical modeling approach for the artificial neural network (ANN) based compact model (CM). The method of retaining part of the network features of the nominal device and further finetuning the network parameters (variational neurons) is found to accurately reproduce the static variation. A mapping from process variation to network parameters is derived by combining the proposed variational neuron selection algorithm and the backward propagation of variance (BPV) method. In addition, a secondary classification of the selected variational neurons is applied to model the fabrication-induced correlation between n-and p-type devices. The NN-based statistical modeling approach has been well implemented and verified on the GAA simulation data and the 16nm node foundry FinFET, which indicates its great potential in modeling emerging and advanced device technology.

Lightweight spiking neural network training based on spike timing dependent backpropagation

Spiking neural networks are energy efficient and biological interpretability, communicating through sparse, asynchronous spikes, which makes them suitable for neuromorphic hardware. However, due to the nature of binary weights and spike trains in time-coded binarized spiking neural networks, their forward propagation may cause neurons to not fire spikes, and their backward propagation has non-differentiable problems. Moreover, the current use of deep and complex network structures generates a large number of redundant parameters. Therefore, we need effective methods to improve energy efficiency without reducing accuracy. We propose a dynamic threshold model that can reduce the number of dead neurons. We combine the backpropagation algorithm and the spike timing dependent plasticity algorithm to avoid the non-differentiable problem. We propose a neuron pruning strategy based on adaptive firing time threshold. This pruning strategy prunes 267 neurons in a network of 600 neurons, reducing the network size and obtaining a more compact network structure. The energy efficiency is improved by 0.55×, while the classification accuracy is lost by 1.1%.

Enabling Binary Neural Network Training on the Edge

The ever-growing computational demands of increasingly complex machine learning models frequently necessitate the use of powerful cloud-based infrastructure for their training. Binary neural networks are known to be promising candidates for on-device inference due to their extreme compute and memory savings over higher-precision alternatives. However, their existing training methods require the concurrent storage of high-precision activations for all layers, generally making learning on memory-constrained devices infeasible. In this article, we demonstrate that the backward propagation operations needed for binary neural network training are strongly robust to quantization, thereby making on-the-edge learning with modern models a practical proposition. We introduce a low-cost binary neural network training strategy exhibiting sizable memory footprint reductions while inducing little to no accuracy loss vs Courbariaux & Bengio’s standard approach. These decreases are primarily enabled through the retention of activations exclusively in binary format. Against the latter algorithm, our drop-in replacement sees memory requirement reductions of 3–5×, while reaching similar test accuracy (± 2 pp) in comparable time, across a range of small-scale models trained to classify popular datasets. We also demonstrate from-scratch ImageNet training of binarized ResNet-18, achieving a 3.78× memory reduction. Our work is open-source, and includes the Raspberry Pi-targeted prototype we used to verify our modeled memory decreases and capture the associated energy drops. Such savings will allow for unnecessary cloud offloading to be avoided, reducing latency, increasing energy efficiency, and safeguarding end-user privacy.

Model Construction and Analysis of Deep Learning-based Cybersecurity Awareness Enhancement for College Students

Abstract This paper constructs a network security intelligence analysis model based on deep learning methods. Firstly, the weights and thresholds of network packets are modeled using the convolutional neural network algorithm to extract the main information features. Then, the backward propagation algorithm is used for layer-by-layer propagation, combined with an unsupervised autoencoder to achieve the network parameter update. The results show that the model can recognize a variety of network viruses, with an average detection rate of 97%, and the error rate is kept around 0.5%. The network security intelligence analysis model is based on the deep learning method to analyze and warn about network intrusion data, effectively improving college students’ awareness about network security.

Optimizing wavefront coding for extended depth of field: a synchronous algorithm for optical element and decoding optimization.

Wavefront coding (WFC) combines phase mask design and image restoration algorithm to extend the depth of field (DOF) for various applications. However, discrete design limits finding globally optimal solutions, increasing the complexity of system design, and affecting the accuracy and robustness of image restoration. An end-to-end imaging system design has emerged to break through these limitations by integrating optical design and image processing algorithms. In this study, we propose an algorithm that synchronously optimizes the optical elements and decoding algorithm in WFC using ray-tracing simulation. We also derive formulas for the optical layer's forward and backward propagation for joint optimization of the optical layer and decoding algorithm. Experimental verification demonstrates the algorithm's effectiveness in optimizing the WFC system and offers improved performance under a unified design framework.

Nonlinear Acoustic Holography with Adaptive Sampling.

Accurate and efficient numerical simulation of highly nonlinear ultrasound propagation is essential for a wide range of therapeutic and physical ultrasound applications. However, due to large domain sizes and the generation of higher harmonics, such simulations are computationally challenging, particularly in three-dimensional problems with shock waves. Current numerical methods are based on computationally inefficient uniform meshes that require the highest harmonic to be resolved across the entire domain. To address this challenge, we present an adaptive numerical algorithm for computationally efficient nonlinear acoustic holography. At each propagation step, the algorithm monitors the harmonic content of the acoustic signal and adjusts its discretization parameters accordingly. This enables efficient local resolution of higher harmonics in areas of high nonlinearity while avoiding unnecessary resolution elsewhere. Furthermore, the algorithm actively adapts to the signal's nonlinearity level, eliminating the need for prior reference simulations or information about the spatial distribution of the harmonic content of the acoustic field. The proposed algorithm incorporates an upsampling process in the frequency domain to accommodate the generation of higher harmonics in forward propagation and a downsampling process when higher harmonics are decimated in backward propagation. The efficiency of the algorithm was evaluated for highly nonlinear three-dimensional problems, demonstrating a significant reduction in computational cost with a nearly 50-fold speedup over a uniform mesh implementation. Our findings enable a more rapid and efficient approach to modeling nonlinear high-intensity focused ultrasound wave propagation.

A study on the prediction of development strategies of art education in universities based on deep learning model

Abstract In this paper, we propose a deep learning restoration method for fine art painting images based on a contextual encoder and select a convolutional self-encoder-based image restoration model for workflow analysis, including forward inference and backward network propagation. We design and develop a software system for predicting the grades of college art and painting students and apply the neural network prediction model to this system to realize the grade prediction of college art students. The accuracy of the GDPN model is 0.922, the precision is 0.904, and the recall is 0.966, which can consider both the temporal and overall information in the click behavior and achieve a better prediction effect.

Improved Performance of Optical Fiber Core: Role of Plasmonic Metals Activation

Optical fiber core with customized characteristics became demanding for diverse high performance applications. Based on this idea, the optical fiber core was activated using various plasmonic metals (beryllium, chromium, and nickel) to improve its refractive index, sensitivity and bandwidth. The influence of various wavelengths and core radii on three modes (LP01, LP11 and LP21) propagation was determined using finite element analysis (FEM). The COMSOL MULTIPHYSICS software was used for the computation. The fiber core radii of the plasmonic metal activated and wavelengths were varied to control the forward and backward energy propagation as well as the modal dispersion relation. Quantities like effective refractive index, attenuation, propagation constant and diffusion coefficient for the three modes as a function of wavelengths and fiber core radii were calculated, which showed maximum values at shorter wavelengths. Irrespective of the type of metal activation in the fiber core, the refractive index of LP01 mode for the core of radius 200 nm was more significantly affected compared to others. Regardless of different metals inclusion, the dispersion relation (refractive index versus frequency) for all modes was strongly positive, showing increasing values for radius in the order of 200, 400, 600 nm. Plasmonic metals Cr and Ni displayed best effect, while Be required high values of V to get LP01 in a narrow range and other modes appeared in a larger range than V. Present results may be useful for the development of high performance optical fiber core.

The construction of a teacher training system for artificial intelligence professionals in universities oriented to RNN models

Abstract In this paper, a custom recurrent neural network is constructed based on RNN. The internal structure of the GRU recurrent neural is mainly analyzed, and the activation function is used to ensure the smooth flow of information in the backward propagation of the neural network. By analyzing the recurrent neural in the processing sequence as well as correlation data, the time-based backpropagation algorithm is constructed. Take advantage of Highway connection in backward propagation to improve the computational speed of the improved recurrent neural network. The moving average method is used to deal with abnormal data and combined with the SGD algorithm to avoid the problem of using all training samples in one iteration so as to establish the optimal model of teacher training time series. The results show that: 48% of the training target factors account for teacher training mainly focused on professional knowledge and 35% on practical ability.

A Multilayer Framework for Online Metric Learning.

Online metric learning (OML) has been widely applied in classification and retrieval. It can automatically learn a suitable metric from data by restricting similar instances to be separated from dissimilar instances with a given margin. However, the existing OML algorithms have limited performance in real-world classifications, especially, when data distributions are complex. To this end, this article proposes a multilayer framework for OML to capture the nonlinear similarities among instances. Different from the traditional OML, which can only learn one metric space, the proposed multilayer OML (MLOML) takes an OML algorithm as a metric layer and learns multiple hierarchical metric spaces, where each metric layer follows a nonlinear layer for the complicated data distribution. Moreover, the forward propagation (FP) strategy and backward propagation (BP) strategy are employed to train the hierarchical metric layers. To build a metric layer of the proposed MLOML, a new Mahalanobis-based OML (MOML) algorithm is presented based on the passive-aggressive strategy and one-pass triplet construction strategy. Furthermore, in a progressively and nonlinearly learning way, MLOML has a stronger learning ability than traditional OML in the case of limited available training data. To make the learning process more explainable and theoretically guaranteed, theoretical analysis is provided. The proposed MLOML enjoys several nice properties, indeed learns a metric progressively, and performs better on the benchmark datasets. Extensive experiments with different settings have been conducted to verify these properties of the proposed MLOML.

UAV Hyperspectral Data Combined with Machine Learning for Winter Wheat Canopy SPAD Values Estimation

Chlorophyll is an important indicator for monitoring crop growth and is vital for agricultural management. Therefore, rapid and accurate estimation of chlorophyll content is important for decision support in precision agriculture to accurately monitor the SPAD (Soil and Plant Analyzer Development) values of winter wheat. This study used winter wheat to obtain canopy reflectance based on UAV hyperspectral data and to calculate different vegetation indices and red-edge parameters. The best-performing vegetation indices and red-edge parameters were selected by Pearson correlation analysis and multiple stepwise regression (MSR). SPAD values were estimated using a combination of vegetation indices, vegetation indices and red-edge parameters as model factors, two types of machine learning (ML), a support vector machine (SVM), and a backward propagation neural network (BPNN), and partial least squares regression (PLSR) for four growth stages of winter wheat, and validated using independent samples. The results show that for the same data source, the best vegetation indices or red-edge parameters for estimating SPAD values differed at different growth stages and that combining vegetation indices with red-edge parameters gave better estimates than using only vegetation indices as an input factor for estimating SPAD values. There is no significant difference between PLSR, SVM, and BPNN methods in estimating SPAD values, with better stability of the estimated models using machine learning methods. Different growth stages have a large impact on winter wheat SPAD values estimates, with the accuracy of the four growth stage models increasing in the following order: booting < heading < filling < flowering. This study shows that using a combination of vegetation indices and red-edge parameters can improve SPAD values estimates compared to using vegetation indices alone. In the future, the choice of appropriate factors and methods will need to be considered when constructing models to estimate crop SPAD values.

Theoretical study of a Janus layered photonic structures based on improved particle swarm algorithm for detection of serum creatinine and glucose concentration

An improved particle swarm optimization (IPSO) algorithm is proposed to perform parameter optimization of the sensor which is realized by Janus layered photonic structures (LPS) based on the optical Tamm state (OTS) and defect layer, which can be used to detect serum creatinine (SCR) solution concentration (CSCR ) and glucose solution concentration (CG ) by adjusting the thickness and refractive index (RI) of the medium in the structure to control the peak and frequency changes of the transmission and absorption peaks. Due to the characteristic of Janus, the light incident from different directions can produce different effects. The OTS based on metal properties has sensitive and narrow absorption peaks when the light comes in forward propagation, its absorption characteristics allowing accurate measurement of CSCR with a high sensitivity (S) of 172 THz/RIU and a relevant quality factor (Q) of 651 and a lower detection limit (DL) of 2.3163 × 10−4. It has a significant Q of 4.5 × 104, DL of 4.825 × 10−5, and S of 11.7857 THz/RIU with its transmission characteristics which can be used to detect CG when the light comes in backward propagation. It can be said that the IPSO has a certain application prospect for the design of optical sensing.

A Method for Backward Failure Propagation in Conceptual System Design

The increased complexity of modern system designs and demands for quicker time to market have made safety-related verification and validation of such systems more challenging. Incorporating safety and risk considerations at the early stages of design is one way to acquire a more robust initial design for novel systems. Inductive fault analysis has its significance at final stages of design, e.g., verification and validation. However, to preclude certain system failure states—especially for the systems with high failure consequences, a designer would innately prefer to trace back and remedy the causes of failure, as compared to a more cumbersome activity of identifying the faults individually and sifting the combinations that lead to the failure of interest. The work presented in this paper is aimed at the development of a backward failure propagation methodology for analyzing the origins of functional failures in a conceptual design of systems including but not limited to nuclear, mechanical, aerospace, process, electrical/electronics, telecommunication, automotive, etc. This method allows the designer to achieve a robust early design based on the analyses of the system’s functional dependencies before proceeding to the detailed design and testing stages. The insights provided by the analysis at the conceptual design stage also reduce redesign efforts, testing costs, and project delays. The proposed method is a functional analysis approach that extends the Integrated System Failure Analysis for backward failure propagation. When provided with an abstract system configuration, a system’s functional model, and a system’s behavioral model, it utilizes a known functional state (typically a failure) to acquire system component modes and the states of other functions. The method includes inversion of the functional failure logic and component behavioral rules using propositional logic and deductive analysis to assess valid states of a system that satisfy the given initial conditions. To test the method’s scalability, we applied the proposed method to a simplified representation of the secondary loop of a typical pressurized water reactor.

Ion-acoustic waves in weakly ionized plasmas with charge-exchange collisions

Ion-neutral charge-exchange collisions in plasmas of laboratory, space, and astrophysical origins are fundamental to understanding wave dissipation and wave generation phenomena. This paper implements a charge-exchange collision operator in the Boltzmann–Poisson system equations for a weakly ionized plasma. When considering an electric field perturbation, the governing kinetic equations provide significant results concerning the plasma conductivity and the dielectric function, appearing in simple, sensible forms. The present analysis reveals a backward wave propagation phenomenon at maximum conductivity when the wavenumber of the plasma wave is smaller than the reciprocal of the ion-neutral collisions mean free path. In addition, it is shown that ion-neutral coupling resulting from charge-exchange collisions enhances ion-acoustic waves below and beyond the ion plasma frequency and leads to the onset of a fundamental instability that overcomes Landau damping under certain circumstances. The collisionless model is recovered as a limiting case, i.e., in the asymptotic limit of a long mean free path.

Physics-informed deep learning for traffic state estimation based on the traffic flow model and computational graph method

Traffic state estimation (TSE) is a critical task for intelligent transportation systems. However, it is extremely challenging because the traffic data quality is often affected by the installation position of devices, data collection frequency, interference during the transmission process, etc., thus causing the problem of data sparsity or data missing. To address the issue of traffic state estimation under the scenario of data sparsity, we propose a TSE model that combines the computational graph with physics-informed deep learning (PIDL) methods. Firstly, we apply the computational graph method to determine the parameters of the traffic fundamental diagram. These parameters are embedded into the computational graph framework, and their values are determined through the forward propagation of variables and the backward propagation of errors. Next, we employ the PIDL method to realize TSE (taking the LWR model based on the Greenshields fundamental diagram as an example). The PIDL leverages the advantages of data-driven and model-driven approaches to achieve accurate traffic state estimation. Case studies are conducted using the NGSIM dataset under two sparse data scenarios: loop detectors and probe vehicles. Experimental results demonstrate that PIDL can accurately reconstruct the traffic state of the entire road segment based on partially observed data. Furthermore, compared to pure deep learning methods and other baseline models, PIDL performs better in situations with sparse data, thereby proving the feasibility of integrating domain knowledge with deep learning frameworks. This paper fully acknowledges the issue of data sparsity in TSE and effectively addresses it by applying the PIDL method to achieve precise TSE, which holds significant implications for the control and management of real traffic flow.