Traditional Neural Network Approach Research Articles

A Lightweight and Efficient Method of Structural Damage Detection Using Stochastic Configuration Network.

With the advancement of neural networks, more and more neural networks are being applied to structural health monitoring systems (SHMSs). When an SHMS requires the integration of numerous neural networks, high-performance and low-latency networks are favored. This paper focuses on damage detection based on vibration signals. In contrast to traditional neural network approaches, this study utilizes a stochastic configuration network (SCN). An SCN is an incrementally learning network that randomly configures appropriate neurons based on data and errors. It is an emerging neural network that does not require predefined network structures and is not based on gradient descent. While SCNs dynamically define the network structure, they essentially function as fully connected neural networks that fail to capture the temporal properties of monitoring data effectively. Moreover, they suffer from inference time and computational cost issues. To enable faster and more accurate operation within the monitoring system, this paper introduces a stochastic convolutional feature extraction approach that does not rely on backpropagation. Additionally, a random node deletion algorithm is proposed to automatically prune redundant neurons in SCNs, addressing the issue of network node redundancy. Experimental results demonstrate that the feature extraction method improves accuracy by 30% compared to the original SCN, and the random node deletion algorithm removes approximately 10% of neurons.

The babble of articulatory learning networks

Articulatory learning models utilize traditional neural network approaches to model speech. However,they differ in their inclusion of an intermediary step in which the model is trained to produce sound utilizing a physical model of the human vocal tract. This style of training is theoretically able to represent the impact of the physical nature of the human vocal tract on language in a way that is more directly translatable to the physical world. This project seeks to create an articulatory learning model which is trained to recognize and repeat input sounds utilizing the University of Wisconsin’s X-Ray Microbeam Database. We attempt this utilizing an inverse model which seeks to predict the vocal tract configuration of a given speech signal, and a forward model which predicts the acoustic form produced by vocal tract configurations. Further, we analyze the performance of the model across various training lengths to investigate the comparison to child language learning that is commonly made when implementing models of this type. Should this analysis of articulatory learners be valid, we look to find commonalities in the phonetic patterning of articulatory learning neural networks and those observed in the phonetic acquisition stages of child language development.

A Learning-Based Physical Model of Charge Transport in Room-Temperature Semiconductor Detectors

Room-temperature semiconductor radiation detectors (RTSDs) such as CdTe are becoming popular in computed tomography (CT) imaging. These detectors are often pixelated, requiring cumbersome postinteraction 3-D event reconstruction, which can benefit from detailed material characterization at the micron level. Transport properties and material defects with respect to electrons and holes are to be characterized, which is a labor-intensive process. Current state-of-the-art characterization is done either as a whole or at most pixel-by-pixel over the detector material. In this article, we propose a novel learning-based physical model to infer material properties at the microscopic level for RTSD. Our approach uses a novel physics-inspired learning model based on physical transport of charges with trapping centers for electrons and holes in the detector. The proposed model learns these material properties from known or measured input charges to the detector along with known or measured output signals and distributed charges in the bulk of the RTSD. The actual physical detector is divided into voxels in space and takes into account different material properties (such as drift, trapping, detrapping, and recombination) in each voxel as learnable model parameters. The model is based on a physics-inspired recurrent neural network model instead of traditional convolutional or fully connected networks. The advantage of our approach is the one-to-one relationship between the actual physical parameters of the voxels and learnable weights in the model, far fewer trainable parameters compared to traditional neural network approaches and less training time. The performance of our model has been evaluated on cadmium zinc telluride (CdZnTe), with voxels of three sizes, 25, 50, and <inline-formula> <tex-math notation="LaTeX">$100~ \mu{\mathrm {m}}$ </tex-math></inline-formula>, for single charge input as well as multiple charge inputs at different voxel positions. Our learning-based model provides material properties with higher spatial resolution and performs well in all scenarios and matches the actual physical parameters better than state-of-the-art classical approaches.

Variational inference in Bayesian neural network for well-log prediction

We have introduced a Bayesian neural network in quantitative log prediction studies with the goal of improving the petrophysical characterization and quantifying the uncertainty of model predictions. Neural network (NN) methods are gaining popularity in the petrophysics and geophysics communities; however, uncertainty quantification in model predictions is often neglected in the available literature, where the prediction is frequently performed in a deterministic setting. Determination of the uncertainty of the petrophysical model requires the estimation of the posterior distribution of the neural parameters that is generally mathematically intractable; for this reason, we adopt a variational approach to approximate the posterior model of the Bayesian network. To represent the uncertainty, we randomly draw samples from the posterior distribution of neural parameters to predict the model variables given the input data, leading to a learned-ensemble predictor. The proposed approach combines the ability of the NN of extracting hidden relations within the data set that physical relations cannot describe and the probabilistic framework for uncertainty quantification. We apply the proposed method to a well-log data set from the Volve Field, offshore Norway, to predict well logs in intervals where the data are incomplete or missing due to operational issues in the drilling procedure. The proposed approach is validated in intervals where the true data are available but not included in the training process. In the proposed application, the correlation coefficient between predictions and true data is greater than 0.9. In terms of accuracy, the results are comparable to those obtained using a traditional NN approach; however, the proposed method also provides a quantification of the uncertainty in the results, which offers additional information on the confidence in the predictions.

Defect Detection in Reinforced Concrete Using Random Neural Architectures

AbstractDetecting defects within reinforced concrete is vital to the safety and durability of our built infrastructure upon which we heavily rely. In this work a non‐invasive technique, ElectroMagnetic Anomaly Detection (EMAD), is used which provides information into the electromagnetic properties of the reinforcing steel and for which data analysis is currently performed visually: an undesirable process. This article investigates the first use of two neural network approaches to automate the analysis of this data: Echo State Networks (ESNs) and Extreme Learning Machines (ELMs) where fast and efficient training procedures allow networks to be trained and evaluated in less time than traditional neural network approaches. Data collected from real‐world concrete structures have been analyzed using these two approaches as well as using a simple threshold measure and a standard recurrent neural network. The ELM approach offers a significant improvement in performance for a single tendon‐reinforced structure, while two ESN architectures provided best performance for a mesh‐reinforced concrete structure.

Three-step approach for wafer sawing lane inspection

Wafer sawing performance must be closely monitored to ensure a satisfactory integrated circuits manufacturing yield. The inspection must allow the GO/NG decision to be fast and reliable, while also assuring that the training of the inspector is simple and not time consuming. The traditional neural-network approach to inspect images, while simple to implement, presents some disadvantages, including training efficiency and model effectiveness. Based on contour detection of the sawing lane, this work proposes a novel method combined with cross-center localization of sawing lanes, detection of sawing track, and four signatures to detect the abnormality of sawing effectively and timely. Our method does not need pretraining but runs faster and provides a better method with more effectiveness, higher flexibility, and immediate feedback to the sawing operation. An experiment using real data collected from an international semiconductor package factory is conducted to validate the performance of the proposed framework. The accurate acceptance rate and the accurate rejection rate are both 100%, while the false acceptance rate and false rejection rate are both zero as well. The results demonstrate that the proposed method is sound and useful for sawing inspection in industries.

A Nonlinear Artificial Intelligence Ensemble Prediction Model for Typhoon Intensity

Abstract A new nonlinear artificial intelligence ensemble prediction (NAIEP) model has been developed for predicting typhoon intensity based on multiple neural networks with the same expected output and using an evolutionary genetic algorithm (GA). The model is validated with short-range forecasts of typhoon intensity in the South China Sea (SCS); results show that the NAIEP model is clearly better than the climatology and persistence (CLIPER) model for 24-h forecasts of typhoon intensity. Using identical predictors and sample cases, predictions of the genetic neural network (GNN) ensemble prediction (GNNEP) model are compared with the single-GNN prediction model, and it has been proven theoretically that the former is more accurate. Computation and analysis of the generalization capacity of GNNEP also demonstrate that the prediction of the ensemble model integrates predictions of its optimized ensemble members, so the generalization capacity of the ensemble prediction model is also enhanced. This model better addresses the “overfitting” problem that generally exists in the traditional neural network approach to practical weather prediction.

A spatiotemporal neural network for recognizing partially occluded objects

In this paper, a spatiotemporal neural network for partially occluded object recognition is presented. The system consists of two major components: a feature extraction process and a spatiotemporal modular neural network. The former is made up of a sequence of preprocessing techniques including thresholding, boundary extraction, Gaussian filtering, and a split-and-merge algorithm to generate features that will represent the objects to be recognized. These acquired features are invariant to rotation, translation, and scaling and can serve as input to the spatiotemporal network that utilizes the concept of tap delay to account for spatial correlation between consecutive input features. A shape perceiver is designed into the network to extract continued parts of an object as well as to enable the inclusion of each object's unique characteristics into the system. Traditional neural network approaches for recognizing partially occluded objects have encountered significant problems because of the incomplete boundaries of the objects. In our approach, by creatively installing tap delays, the system can escape this limitation. Experimental results show that the proposed system can produce satisfactory results in efficiently and effectively recognizing partially occluded objects. Furthermore, intrinsic to this system is the ease by which it can be realized through parallel implementation, thus minimizing the tremendous time spent in matching object contours stored in a model database, as is the case in conventional recognition systems.

Digraph visualization using a neural algorithm with a heuristic activation scheme

This paper presents an activation scheme for use with Hopfield neural network algorithms that guarantees a valid solution for a particular category of problems. The technique monitors the appropriate neurons and heuristically controls their activation function. As a result it has been possible to eliminate several constraint terms from the energy function that normally would have been required to drive the network toward a valid solution. This saves time and eliminates the need for empirically determining a larger number of constants. This technique has been applied to the combinatorial optimization problem called hierarchical digraph visualization that arises in many application areas where it is necessary to visually realize the relationship between entities in complex systems. Results are presented that compare this new approach with a more traditional neural network approach as well as heuristic approaches, performance improvement in terms of the solution quality as well as execution time relative to both alternative techniques was achieved.

Neural network approach for minimizing the makespan of the general job-shop

A neural network approach is proposed to minimize the makespan of the job-shop scheduling, which is a combinatorial optimization problem. Our approach is based on the Hopfield interconnected neural networks model. In contrast to the traditional neural network approach based on the Hopfield model, our model changes the threshold values at each transition of neurons in order to make a non-delay schedule in addition to incorporating job- and shop-related constraints. As the modification may lead to non-optimal solutions, we increase the temperature of the network according to the Boltzmann machine mechanism and obtain other schedules until no better solution can be obtained within the specified number of tests. From the numerical experiments, 10 out of 15 problems are solved optimally, and remaining five problems are solved near-optimally within a reasonable computing time.