Inverse Neural Network Research Articles

Real-time online inversion of GA-PSO-BP flux leakage defects based on information fusion: Numerical simulation and experimental research

Currently, defect inversion is always a difficult problem in magnetic flux leakage (MFL) detection, and all kinds of algorithms cannot solve this problem effectively. Back propagation (BP) neural network is widely used in the reconstruction of MFL. However, BP neural network has problems such as slow training speed, low recognition accuracy, and easy to fall into local minima. In this study, an information fusion method combining fuzzy set theory and neural network is studied to eliminate outliers caused by vibration, so as to improve the reliability of data in vibration environment. And an improved particle swarm optimization (PSO) algorithm (GA-PSO-BP) is applied to invert the defect size from the flux leakage signal. The traditional particle swarm optimization method has strong dependence on the initial value and can only obtain the local optimal solution. In this paper, the crossover and mutation operations in the genetic algorithm are used to improve the population diversity and global search ability of the particle swarm algorithm, and the weights and thresholds of the network are adjusted to make the predicted output continuously approach the expected output. The defect size is estimated using an inversion technique based on GA-PSO-BP. The results show that the inversion method based on GA-PSO-BP can improve the average error accuracy of defect depth by 5.83% and the average error accuracy of defect length by 4.87%. Thus, the inversion method based on GA-PSO-BP is superior to the BP neural network inversion technology. Besides, the improved algorithm can improve the defect inversion speed and meet the requirements of real-time online detection in a vibration environment.

Estimation of Soil Organic Carbon Using Vis-NIR Spectral Data and Spectral Feature Bands Selection in Southern Xinjiang, China.

Soil organic carbon (SOC) plays an important role in the global carbon cycle and soil fertility supply. Rapid and accurate estimation of SOC content could provide critical information for crop production, soil management and soil carbon pool regulation. Many researchers have confirmed the feasibility and great potential of visible and near-infrared (Vis-NIR) spectroscopy in evaluating SOC content rapidly and accurately. Here, to evaluate the feasibility of different spectral bands variable selection methods for SOC prediction, we collected a total of 330 surface soil samples from the cotton field in the Alar Reclamation area in the southern part of Xinjiang, which is located in the arid region of northwest China. Then, we estimated the SOC content using laboratory Vis-NIR spectral. The Particle Swarm optimization (PSO), Competitive adaptive reweighted sampling (CARS) and Ant colony optimization (ACO) were adopted to select SOC feature bands. The partial least squares regression (PLSR), random forest (RF) and convolutional neural network (CNN) inversion models were constructed by using full-bands (400–2400 nm) spectra (R) and feature bands, respectively. And we also analyzed the effects of spectral feature band selection methods and modeling methods on the prediction accuracy of SOC. The results indicated that: (1) There are significant differences in the feature bands selected using different methods. The feature bands selected methods substantially reduced the spectral variable dimensionality and model complexity. The models built by the feature bands selected by CARS, PSO and ACO methods showed the different potential of improvement in model accuracy compared with the full-band models. (2) The CNN model had the best performance for predicting SOC. The R2 of the optimal CNN model is 0.90 in the validation, which was improved by 0.05 and 0.04 in comparison with the PLSR and RF model, respectively. (3) The highest prediction accuracy was archived by the CNN model using the feature bands selected by CARS (validation set R2 = 0.90, RMSE = 0.97 g kg−1, RPD = 3.18, RPIQ = 3.11). This study indicated that using the CARS method to select spectral feature bands, combined with the CNN modeling method can well predict SOC content with higher accuracy.

Physics Embedded Deep Neural Network for Solving Full-Wave Inverse Scattering Problems

In this work, we design an iterative deep neural network to solve full-wave inverse scattering problems (ISPs) in the 2-D case. Forward modeling neural networks that predict the scattered field are embedded in an inversion neural network. In an iterative manner, the inversion network predicts the model update from the residual between the simulated data and the observed data. The proposed inversion network can achieve super-resolution reconstruction meanwhile keeping the simulated data of reconstructed models well consistent with the observed data. We validate this method with both synthetic and experimental data inversion. Results show that the inversion network can predict models with high accuracy, efficiency, and good generalization ability. By combining deep learning and physical simulation together, the proposed method provides a way for real-time imaging with high reliability and accuracy.

Multi-Parameter Inversion of AIEM by Using Bi-Directional Deep Neural Network

A novel multi-parameter inversion method is proposed for the Advanced Integral Equation Model (AIEM) by using bi-directional deep neural network. There is a very complex nonlinear relationship between the surface parameters (dielectric constant and roughness) and radar backscattering coefficient. The traditional inverse neural network, which is constructed by using the backscattering coefficients as the input and the surface parameters as the output, leads to bad convergence and wrong results. This is because many sets of surface parameters can get the same backscattering coefficient. Therefore, the proposed bi-directional deep neural network starts with building an AIEM-based forward deep neural network (AIEM-FDNN), whose inputs are the surface parameters and outputs are the backscattering coefficients. In this way, the weights and biases of the forward deep neural network can be optimized and predicted, which can be used for the backward deep neural network (AIEM-BDNN). Then, the multi-parameters are updated by minimizing the loss between the output backscattering coefficients with the measured ones. By inserting a sigmoid function between the input and the first hidden layer, the input multi-parameters can be efficiently approximated and continuously updated. As a result, both the forward and backward deep neural networks can be built with these weights and biases. By sharing the weights and biases of the forward network, the training of the inverse network is avoided. The bi-directional deep neural network can not only predict the backscattering coefficient but can also inverse the surface parameters. Numerical results are given to demonstrate that the RMSE of the backscattering coefficients calculated by the proposed bi-directional neural network can be reduced to 0.1%. The accuracy of the inversion parameters, including the real and imaginary parts of the dielectric constant, the root mean square height and the correlation length, can be improved to 97.56%, 91.14%, 99.04% and 98.45%, respectively. At the same time, the bi-directional neural network also has good accuracy for the inversion of the POLARSCAT measured data.

An Efficient Knowledge-Based Artificial Neural Network for the Design of Circularly Polarized 3-D-Printed Lens Antenna

An efficient knowledge-based artificial neural network (KBANN) is proposed, and it is used for the design of circularly polarized (CP) lens antenna in this article. In this KBANN, forward neural network (FNN) and inverse neural network (INN) are included. In this model, INN is the major component to predict the antenna structure parameters. As multiple performance indices are required, INN requires a large number of training samples to deduce complex mapping relationship. To solve this problem, FNN is introduced to provide prior knowledge for INN. FNN generates a huge training dataset for INN training, and then the trained INN can directly output the geometric parameters by feeding the target electromagnetic responses as input. This article solves the problem of multiple performance indices in antenna design, and a CP lens antenna with wideband, good axial ratio, and high gain is designed and fabricated to verify the effectiveness of the KBANN model.

Multi-parameter methane measurement using near-infrared tunable diode laser absorption spectroscopy based on back propagation neural network

In order to achieve high-precision and continuous measurement of methane (CH4) isotope abundance (δ13CH4), a near-infrared CH4 sensor system was developed based on back propagation (BP) neural network using tunable diode laser absorption spectroscopy (TDLAS). A pair of CH4 isotope absorption lines with similar low-state energy levels near 1658 nm were selected for minimizing temperature dependence. A LabVIEW-based signal processing platform was realized for real-time signal processing. Wavelet denoising (WD) algorithm was employed to pre-process the absorption signal for increasing signal-to-noise ratio (SNR), and the BP neural network algorithm was adopted to invert δ13CH4. The BP neural network algorithm exhibits potential for multi-parameter analysis, including gas temperature and pressure. Related experiments were carried out to prove the high-precision and reliability of the neural network inversion method. Through experimental comparison, the measurement precision of δ13CH4 is enhanced by ∼ 11 times using the neural network inversion method compared to the maximum absorbance extracting method.

Inversion using adaptive physics‐based neural network: Application to magnetotelluric inversion

ABSTRACTA new trend to solve geophysical problems aims to combine the advantages of deterministic inversion with neural network inversion. The neural networks applied to geophysical inversion have had limited success due to the need for extensive training of datasets and the lack of generalizability to out‐of‐sample scenarios. Deterministic regularized inversion often requires a good starting model to avoid possible local minima in highly nonlinear problems. We have developed a physics‐based neural network procedure that combines the advantages of deterministic and neural network inversions in a coupled inversion scheme. The new inversion algorithm is formulated as a constrained minimization problem with a composite objective function composed of data misfit, neural network training function and a coupling model objective function that links the two inversion schemes through a reference model. We investigate two strategies to update the reference model in the coupling model objective function using either a fully trained network or an adaptively trained network that keeps evolving and learning during the inversion. First, we propose a procedure to generate the training dataset required for the fully trained network to improve the network generalization. Second, we extend the analysis of the adaptively trained network inversion without preparing a training dataset, so it is suitable for most exploration scenarios. Our approach enables the network to learn only the relevant model distribution. Predictions from the network are used to constrain the deterministic inversion using a new predicted reference model. We demonstrate the convergence of our procedure in recovering high‐resolution resistivity models when applied to synthetic magnetotellurics data. We compare the recovered resistivity model from a deterministic inversion, a neural network inversion and physics‐based neural network inversions using fully and adaptively trained networks, respectively. The results indicate that the adaptive physics‐based neural network is superior as it does not require preparing a training dataset but can still recover reliable resistivity models.

Neural Network Application in Dispersion Curve Inversion of Seabed Geoacoustic Parameters

This paper estimates seabed shear-wave velocities and the thickness of the surface sediments using Scholte wave dispersion curves extracted from data. Common surface wave dispersion curve inversion methods are divided into: local linearization methods and global optimization methods. These are model-driven, the inversion process takes a lot of time, and it is easy to get a local optimum, and the inversion results are inaccurate. Aiming at the shortcomings of the existing surface wave dispersion inversion methods, this paper introduces an inversion method based on neural network, and fits the Scholte wave dispersion curve to obtain the geoacoustic parameters of the surface sediments. Neural network inversion is data-driven, and the model is extracted from the data, which can improve the speed and accuracy of surface wave inversion. By simulating the shallow sea model, better results are obtained, and at the same time, the experimental data is used for calculation, the inversion results are close to the traditional methods, and the inversion speed is improved.

Research on Orbital Angular Momentum Multiplexing Communication System Based on Neural Network Inversion of Phase

An adaptive optical wavefront recovery method based on a residual attention network is proposed for the degradation of an Orbital Angular Momentum multiplexing communication system performance caused by atmospheric turbulence in free-space optical communication. To prevent the degeneration phenomenon of neural networks, the residual network is used as the backbone network, and a multi-scale residual hybrid attention network is constructed. Distributed feature extraction by convolutional kernels at different scales is used to enhance the network’s ability to represent light intensity image features. The attention mechanism is used to improve the recognition rate of the network for broken light spot features. The network loss function is designed by combining realistic evaluation indexes so as to obtain Zernike coefficients that match the actual wavefront aberration. Simulation experiments are carried out for different atmospheric turbulence intensity conditions, and the results show that the residual attention network can reconstruct the turbulent phase quickly and accurately. The peaks to valleys of the recovered residual aberrations were between 0.1 and 0.3 rad, and the root means square was between 0.02 and 0.12 rad. The results obtained by the residual attention network are better than those of the conventional network at different SNRs.

Inverse Design of Distributed Bragg Reflectors Using Deep Learning

Distributed Bragg Reflectors are optical structures capable of manipulating light behaviour, which are formed by stacking layers of thin-film materials. The inverse design of such structures is desirable, but not straightforward using conventional numerical methods. This study explores the application of Deep Learning to the design of a six-layer system, through the implementation of a Tandem Neural Network. The challenge is split into three sections: the generation of training data using the Transfer Matrix method, the design of a Simulation Neural Network (SNN) which maps structural geometry to spectral output, and finally an Inverse Design Neural Network (IDNN) which predicts the geometry required to produce target spectra. The latter enables the designer to develop custom multilayer systems with desired reflection properties. The SNN achieved an average accuracy of 97% across the dataset, with the IDNN achieving 94%. By using this inverse design method, custom-made reflectors can be manufactured in milliseconds, significantly reducing the cost of generating photonic devices and thin-film optics.

A Novel Approach for Identifying Hyper-Elastic Material Parameters of Cartilage based on FEM and Neural Networks

Cartilage damage and degeneration may lead to osteoarthritis for both animals and humans. Quantitative studies on the nonlinear hyper-elastic behavior of cartilages are essential to evaluate cartilage tissue deterioration. However, direct identification of the material behavior is not feasible. This paper presents a procedure to characterize the nonlinear mechanical behavior of the cartilage tissue by an inverse method using measurable structural quantities. First, a two-way neural network (NN) is established, which uses the fully trained forward problem neural network instead of the forward problem solver to generate training samples for inverse problem neural network. Moreover, based on the experimental data of the kangaroo shoulder joint, a nonlinear finite element (FE) model is then created to produce a dataset for training the forward network. Furthermore, intensive studies are conducted to examine the performance of our two-way NN method for the prediction of cartilage hyper-elastic material parameters by comparison with the direct inverse NN method. When only the direct inverse problem neural network is used for training, all samples are from FE simulations and the simulation time is 50.7 h, and the prediction time is tens of seconds. Besides, our two-way neural network calls the trained forward NN to collect training samples, and all the samples can be obtained in seconds, with which the simulation time is only 78 s. The predicted results are in good agreement with the experimental data, and the comparison shows that our two-way NN is an efficient and proficient method to predict the parameters for other biological soft tissues.

Impact of material homogeneity assumption on cortical stiffness estimates by MR elastography.

Inversion algorithms used to convert acquired MR elastography wave data into material property estimates often assume that the underlying materials are locally homogeneous. Here we evaluate the impact of that assumption on stiffness estimates in gray-matter regions of interest in brain MR elastography. We describe an updated neural network inversion framework using finite-difference model-derived data to train convolutional neural network inversion algorithms. Neural network inversions trained on homogeneous simulations (homogeneous learned inversions [HLIs]) or inhomogeneous simulations (inhomogeneous learned inversions [ILIs]) are generated with a variety of kernel sizes. These inversions are evaluated in a brain MR elastography simulation experiment and in vivo in a test-retest repeatability experiment including 10 healthy volunteers. In simulation and in vivo, HLI and ILI with small kernels produce similar results. As kernel size increases, the assumption of homogeneity has a larger effect, and HLI and ILI stiffness estimates show larger differences. At each inversion's optimal kernel size in simulation (7 × 7 × 7 for HLI, 11 × 11 × 11 for ILI), ILI is more sensitive to true changes in stiffness in gray-matter regions of interest in simulation. In vivo, there is no difference in the region-level repeatability of stiffness estimates between the inversions, although ILI appears to better maintain the stiffness map structure as kernel size increases, while decreasing the spatial variance in stiffness estimates. This study suggests that inhomogeneous inversions provide small but significant benefits even when large stiffness gradients are absent.

On direct weight inverse approach for identifying composite parameters based on two-way TrumpetNets

Reliable composite material parameter identification is of practical engineering significance in view of the extensive applications of composite materials in high-tech industries, such as aerospace, automobile industry and medical fields, etc. Efficient and easy-to-use computational tools are needed, and a new theoretical framework of two-way TrumpetNets have been proposed recently. It can solve both the forward and inverse problems effectively. The striking beauty of this theory is that the solution of the inverse problem is expressed by direct weight inversion (DWI) approach in an explicit formula. This study extends and implements the explicit inverse solution using two-way TrumpetNets for identification of material parameters of composite laminates. We successfully inversely identify up to seven parameters for composite laminates, which is essentially a 7-Dimensional inverse problem. It is found that by application of the two-way TrumpetNets, regularization parameters are not required in the DWI formula due to the intrinsic regularization by the least square formulation. The influence of different network structures, i.e., the number of hidden layers and the number of each layer, on the inverse result is explored. For similar composite material parameter identification, the computational efficiency is significantly improved in view of that, for inverse identification of composite parameters, it takes only 20 min computational time by two-way TrumpetNets. While this computational time is one sixth of that by the traditional inverse neural network using the same desktop computer, the computational accuracy is not compromised.

Fuzzy deep wavelet neural network with hybrid learning algorithm: Application to electrical resistivity imaging inversion

Electrical resistivity imaging (ERI) is a non-invasive imaging technique for measuring resistivity, and the inversion problem of ERI is non-linear and non-convex. Traditional fuzzy neural network based on gradient descent is known to be inept for its low accuracy and does not ensure global convergence. In order to solve above problems, we present a fuzzy deep wavelet neural network (FDWNN) inversion method trained by an accelerated hybrid learning algorithm to invert resistivity data of ERI. Firstly, a novel FDWNN model, which integrates the fuzzy clustering-based premise part with the deep WNN-based consequent part, is applied to improve the prediction accuracy and enhance the interpretability of ERI inversion. Secondly, an adaptive shuffled frog leaping algorithm (ASFLA) is introduced to balance the exploration and exploitation during the search process intelligently. In the proposed ASFLA, an adaptive mutation rule is applied to improve the local search and a differential leaping strategy is presented to enhance the global search. Finally, an accelerated hybrid learning algorithm integrating the ASFLA and a weight decay backpropagation (wdBP) method is designed, which keeps the advantages of the SFLA in finding global optimal values, while speeds up the convergence and improves the generalization through wdBP simultaneously. Moreover, five experiments are introduced to evaluate the feasibility and applicability of the FDWNN algorithm by comparison with other contenders.

Hybrid Memetic Pretrained Factor Analysis-Based Deep Belief Networks for Transient Electromagnetic Inversion

As a trenchless detection approach, the transient electromagnetic method (TEM) can effectively detect well-conducted geoelectric structures, such as groundwater structures. The nonlinear TEM inversion process represented by neural network (NN) inversion does not rely on the initial model, which helps it efficiently and accurately obtain geoelectric structures. In this article, a factor analysis (FA)-based deep belief network (DBN) inversion framework built on hybrid memetic pretrained (HMP), HMP-FADBN, is proposed. We construct a DBN with restricted Boltzmann machines (RBMs) and backpropagation NNs (BPNNs) as the base fitting skeleton. FA is used to reduce the dimensionality of the feature space of the DBN. The hybrid memetic (HM) whale optimization algorithm (WOA) pretrains the network parameters and uses the memetic strategy to adjust the coordinated development system of social and individual cognition to enhance the network training effect. Numerical examples show that the prediction accuracy of the proposed HMP-FADBN for TEM geoelectric models is improved from more than 10% to approximately 2%. Moreover, 5%, 10%, and 15% noise tests show that the trained NN has good generalization and denoising abilities, and the prediction accuracy is less than 3% (affected by the maximum noise). Finally, the developed method is successfully applied to a landslide TEM survey, and the predicted quasi-2D geoelectric structure is consistent with the original geological structure.

Mining of Weak Fault Information Adaptively Based on DNN Inversion Estimation for Fault Diagnosis of Rotating Machinery

Fault feature extraction plays a significant role in bearing fault diagnosis, especially in incipient fault period. Recently, deep neural network has been favored by many researchers due to its excellent hierarchical feature extraction capabilities. The existing diagnosis methods based on deep neural network mostly take original condition monitoring data as input and further convert fault diagnosis into pattern recognition issues. Although it improves the level of intelligent diagnosis, it is still confronted with practical problems. Since deep neural network includes many non-linear mapping layers, the extracted fault-related features show a high degree of abstraction, which reduces the practical level of this technology. Aimed at this problem, a mining method of understandably weak fault information is proposed based on deep neural network inversion estimation. From perspective of neuron response degree, this method mines the most sensitive input pattern that maximizes the neuron’s activation value of network output layer in the original input feature space. It can intuitively mine the weak fault information from the original signal. Two bearing experiments verify the effectiveness and reliability of the proposed method.

Linearized inverse Schrödinger potential problem with partial data and its deep neural network inversion

<p style='text-indent:20px;'>We study the linearized inverse Schrödinger potential problem with (many) partial boundary data. By fixing specific partial boundary these measurements are realized by the linearized local Dirichlet-to-Neumann map. When the wavenumber is assumed to be large, we verify a Hölder type increasing stability by constructing the complex exponential solutions in a reflection form. Meanwhile, the linearized inverse Schrödinger potential problem admits an integral equation where the unknown potential function is indirectly contained there. Such a formulation allows us to adopt a deep neural network inversion algorithm. Numerical examples show that one can reconstruct the unknown potential function stably within the partial boundary data setting.</p>

Tree Internal Defected Imaging Using Model-Driven Deep Learning Network

The health of trees has become an important issue in forestry. How to detect the health of trees quickly and accurately has become a key area of research for scholars in the world. In this paper, a living tree internal defect detection model is established and analyzed using model-driven theory, where the theoretical fundamentals and implementations of the algorithm are clarified. The location information of the defects inside the trees is obtained by setting a relative permittivity matrix. The data-driven inversion algorithm is realized using a model-driven algorithm that is used to optimize the deep convolutional neural network, which combines the advantages of model-driven algorithms and data-driven algorithms. The results of the comparison inversion algorithms, the BP neural network inversion algorithm, and the model-driven deep learning network inversion algorithm, are analyzed through simulations. The results shown that the model-driven deep learning network inversion algorithm maintains a detection accuracy of more than 90% for single defects or homogeneous double defects, while it can still have a detection accuracy of 78.3% for heterogeneous multiple defects. In the simulations, the single defect detection time of the model-driven deep learning network inversion algorithm is kept within 0.1 s. Additionally, the proposed method overcomes the high nonlinearity and ill-posedness electromagnetic inverse scattering and reduces the time cost and computational complexity of detecting internal defects in trees. The results show that resolution and accuracy are improved in the inversion image for detecting the internal defects of trees.

A Contribution to the Modelling of Fouling Resistance in Heat Exchanger-Condenser by Direct and Inverse Artificial Neural Network

A Contribution to the Modelling of Fouling Resistance in Heat Exchanger-Condenser by Direct and Inverse Artificial Neural Network

Imaging subsurface orebodies with airborne electromagnetic data using a recurrent neural network

Conventional interpretation of airborne electromagnetic data has been conducted by solving the inverse problem. However, with recent advances in machine learning (ML) techniques, a 1D deep neural network inversion that predicts a 1D resistivity model using multifrequency vertical magnetic fields and sensor height information at one location has been applied. Nevertheless, since the final interpretation of this 1D approach relies on connecting 1D resistivity models, 1D ML interpretation has low accuracy for the estimation of an isolated anomaly, as in conventional 1D inversion. Thus, we have developed a 2D interpretation technique that can overcome the limitations of 1D interpretation and consider spatial continuity by using a recurrent neural network (RNN). We generate various 2D resistivity models, calculate the ratio of primary and induced secondary magnetic fields of the vertical direction in ppm scale using a vertical magnetic dipole source, and then train the RNN using the resistivity models and the corresponding electromagnetic (EM) responses. To verify the validity of 2D RNN inversion, we apply the trained RNN to synthetic and field data. Through application of the field data, we demonstrate that the design of the training data set is crucial to improve prediction performance in a 2D RNN inversion. In addition, we investigate changes in the RNN inversion results of field data dependent on the data preprocessing. We demonstrate that using two types of data, logarithmic transformed data and linear-scale data, having different patterns of input information can enhance the prediction performance of the EM inversion results.