Search

Purpose: To develop a deep-learning system for automatic osteoporotic vertebral compression fractures (OVCF) detection at the thoracolumbar junction using low-dose computed tomography (CT) images. Materials and methods: 500 individuals were enrolled in this retrospective study, including 270 normal and 230 OVCF cases. The cases were divided into the training, validation, and test sets in the ratio of 6:2:2. First, a localization model using Faster R-CNN was trained to identify and locate the target thoracolumbar junction, then a 3D AnatomyNet model was trained to finely segment the target vertebrae in the localized image. Finally, 3D DenseNet model was applied for detecting OVCF on target vertebrae. Manual annotation by experienced radiologists and a clinically made diagnosis of OVCF were used as the gold standards. The performance of the detecting system was evaluated through the area under the curve (AUC) for receiver operating characteristic (ROC) analysis, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV). Results Our automated segmentation method achieved a mean Dice coefficient of 0.95 for vertebral bodies (T12-L2) segmentation on the testing dataset, with dice coefficients greater than 0.9 accounting for 96.6%. For the diagnostic performance of our system for OVCF, the AUC, sensitivity, specificity, PPV and NPV for the four-fold cross-validation on the testing dataset were 98.1%, 95.7%, 92.6%, 91.7% and 96.2%, respectively. Conclusions A deep-learning system has been developed to automatically segment vertebral bodies and accurately detect OVCF using low-dose CT.

This article mainly focuses on the most common types of high-speed railways malfunctions in overhead contact systems, namely, unstressed droppers, foreign-body invasions, and pole number-plate malfunctions, to establish a deep-network detection model. By fusing the feature maps of the shallow and deep layers in the pretraining network, global and local features of the malfunction area are combined to enhance the network's ability of identifying small objects. Further, in order to share the fully connected layers of the pretraining network and reduce the complexity of the model, Tucker tensor decomposition is used to extract features from the fused-feature map. The operation greatly reduces training time. Through the detection of images collected on the Lanxin railway line, experiments result show that the proposed multiview Faster R-CNN based on tensor decomposition had lower miss probability and higher detection accuracy for the three types faults. Compared with object-detection methods YOLOv3, SSD, and the original Faster R-CNN, the average miss probability of the improved Faster R-CNN model in this paper is decreased by 37.83%, 51.27%, and 43.79%, respectively, and average detection accuracy is increased by 3.6%, 9.75%, and 5.9%, respectively.

An overhead contact system (OCS) is key to providing power to high-speed railways. OCS detection is an important measure to ensure the safe operation of a high-speed railway. At present, OCS anomaly detection mainly relies on the manual analysis of the images regularly collected by the 4C system, which is very inefficient and can easily miss anomalies. Although some classification and object detection methods based on deep learning can be used for OCS anomaly detection, the effective training of deep networks can be difficult to support due to the small number of anomaly OCS image samples. Considering that most OCS faults are abnormal fasteners, we propose an abnormal detection method based on normal images, called the nested residual encoder-decoder network (NRE-Net). This network consists of two nested encoder-decoder networks, where the encoder is the shared part, and a residual structure is added to the encoding and decoding branches to enhance the feature expression ability. The experimental results show that the method can greatly improve the accuracy of anomaly detection for the CIFAR-10 dataset and OCS fastener dataset. Compared with the previous state-of-the-art approaches, the $F_{1}$ score of the proposed method for the two classes fastener in the OCS fastener dataset has increased by 10.8% and 11.9%, respectively.

The traditional information search technology performs full-text indexing on the data in the Internet, searches for information by means of keyword matching index, and returns information to the user. This retrieval method does not accurately understand the user’s needs, and returns relevant links rather than the information the user needs. The user needs to find relevant information from the linked documents. In order to improve the shortcomings of the above traditional search technology, this paper is based on the knowledge map of multi-data source intelligent information distributed search technology, through data acquisition in the Internet, complete the transformation of data to knowledge to form a knowledge network and provide information retrieval. This paper studies the construction of knowledge maps for application domain representation, proposes a semantic similarity model with constraints and implicit feedback correction mechanism, and explores the realization of intelligent information search under certain conditions. Through the analysis of prototype experimental data in the application field, the accuracy of information search based on knowledge map can reach 90%, which has strong practicability.

The power sector continues to accumulate a large amount of data resources, including relevant standard specifications, technical documents, management documents, fault resolution records. How to quickly query and intelligently search these documents is of great value for grid dispatching and fault recovery. The domain search system of traditional power grid is based on keywords, and has the problems of low precision and recall rate. It cannot understand the business language and cannot support semantic reasoning. This paper designs and implements a method based on knowledge mapping to construct the power domain semantic search system. Semantic knowledge extraction of unstructured data is carried out by intelligent domain segmentation technology, organized and stored as knowledge mapping, and semantic search for support reasoning is realized based on knowledge mapping. The process of constructing domain semantic search system is introduced. Experiments show that the accuracy rate and recall rate of the method have been greatly improved.

Droppers are key components of high-speed railway overhead catenary systems, which are exposed to the external environment and are prone to breakage faults due to the impact of wind force and the pantograph on moving trains day after day. How to identify dropper breakage or relaxation faults through acceleration signals installed in the carrier cable and contact wire is a challenging problem. In this study, the experimental section of the Lanzhou-Xinjiang high-speed railway was simulated on the basis of the bow-network dynamic simulation model, in which the overhead catenary system was subjected to the force of pulsating wind alone or pulsating wind and the pantograph at the same time. In the experiment, we collected 10 channel signals from five acceleration sensors when two droppers were normal or broken. We established a 1D-CNN model of four categories and then determined the hyperparameters of the deep network structure and the important parameters of the network optimization scheme through the Bayesian optimization algorithm. Furthermore, we selected the lowest sensor number to identify dropper fracture faults by a large number of experiments according to mechanics principles. The experimental results show that the proposed methods have a higher identification accuracy rate, recall rate, and robustness than the traditional artificial feature extraction approaches. Therefore, the detection methods proposed provide an effective way to identify high-speed railway catenary dropper faults on the basis of acceleration sensors.

Anomaly detection is an important branch of computer vision. At present, a variety of deep learning models are applied to anomaly detection. However, the lack of abnormal samples makes supervised learning difficult to implement. In this paper, we mainly study abnormal detection tasks based on unsupervised learning and propose a Fully-Nested Encoder-decoder Framework. The main part of the proposed generating model consists of a generator and a discriminator, which are adversarially trained based on normal data samples. In order to improve the image reconstruction capability of the generator, we design a Fully-Nested Residual Encoder-decoder Network, which is used to encode and decode the images. In addition, we add residual structure into both encoder and decoder, which reduces the risk of overfitting and enhances the feature expression ability. In the test phase, a distance measurement model is used to determine whether the test sample is abnormal. The experimental results on the CIFAR-10 dataset demonstrate the excellent performance of our method. Compared with the existing models, our method achieves the state-of-the-art result.

To find a way of the interpretability of deep learning, in this paper, a features back-tracking (FBT) approach based on a sparse deep learning architecture is proposed. Firstly, for a deep belief network (DBN), both the Kullback-Leibler divergence of the hidden neurons and the L1 norm penalty on the connection weights are introduced. Thus, the sparse response mechanism as well as the sparse connection of the brain neurons can be simulated directly. That means the DBN can learn a sparse framework and an effective sparse data representation. On this basis, the feature back-tracking technique is put forward. For both the single nucleotide polymorphisms (SNPs) data and MNIST data, FBT has quite well performance on searching for the risk loci on the genes as well as the important sites of the digit data. It reveals that the proposed FBT method can pick out the essential features by deep learning architecture with quite high classification accuracy and data storage ability. Utilizing the sparse layer-wise feature learning to achieve key features from the original data, is an effective attempt to explore the profound mechanism of human brain and interpretability of deep learning.

The business knowledge of intelligent transportation inspection is scattered in the business system, database and expert experience, and it is impossible to visualize the entire business logic of transportation inspection. Aiming at the characteristics of complex and wide variety of operational inspection business relationships, an intelligent operational inspection knowledge mapping is constructed. Through the steps of operation inspection business ontology construction, knowledge acquisition, knowledge fusion, knowledge mapping display, knowledge mapping application and other steps, a logical knowledge mapping of operation inspection business is formed, which facilitates the recognition of the entire business relationship and assists the decision-making analysis of the operation inspection personnel.

In order to realize the explicit expression of the distribution network data relationship of the power grid enterprises and the efficient search of data resources, a method for modeling and intelligent searching of distribution network data resources based on knowledge mapping is proposed. This method firstly identifies the core ontology of the distribution network based on the ontology construction process and the distribution network business, then maps and extracts the identified ontology model and the relational database physical model to form a knowledge mapping of distribution network data resources. On this basis, combining the BiLSTM-CRF-based distribution network named entity recognition algorithm and the BERT-based Chinese semantic matching algorithm, a distribution network data resource search engine technology is proposed to realize the search text feature extraction and result matching. Finally, the application verification data of a power grid company shows that the distribution network knowledge mapping has a good effect on the expression and visualization of data resource relationships. At the same time, the accuracy rate of the proposed distribution network data resource search engine technology method reaches 72.2%.

The uniformly pseudo-projection-anti-monotone (UPPAM) neural network model, which can be considered as the unified continuous-time neural networks (CNNs), includes almost all of the known CNNs individuals. Recently, studies on the critical dynamics behaviors of CNNs have drawn special attentions due to its importance in both theory and applications. In this paper, we will present the analysis of the UPPAM network under the general critical conditions. It is shown that the UPPAM network possesses the global convergence and asymptotical stability under the general critical conditions if the network satisfies one quasi-symmetric requirement on the connective matrices, which is easy to be verified and applied. The general critical dynamics have rarely been studied before, and this work is an attempt to gain an meaningful assurance of general critical convergence and stability of CNNs. Since UPPAM network is the unified model for CNNs, the results obtained here can generalize and extend the existing critical conclusions for CNNs individuals, let alone those non-critical cases. Moreover, the easily verified conditions for general critical convergence and stability can further promote the applications of CNNs.

A simple europium complex, Eu(TTA)3AA (TTA=4,4,4-trifluoro-1-(2-thienyl)-1,3-butanedione, AA=acrylic acid) was synthesized by a simple solution method. Then, two kinds of rubber matrix, nitrile-butadiene rubber (NBR) and silicone rubber (SiR) were used as the substrate for Eu(TTA)3AA to prepare fluorescence composites. The neat Eu(TTA)3AA complex showed the ability of self-polymerization when it was heated at 145 °C. It was found that the fluorescence intensity of the neat Eu(TTA)3AA decreased over 70% when the polymerization time was over 25 min at 145 °C. The results also revealed that the polymerizated Eu(TTA)3AA could be dispersed in nano-scale in two matrices and the luminescent intensities decreased 52% in NBR matrix, and 95% in SiR matrix compared with two relative compounds without crosslinking. To optimize the luminescence intensity of the composites, the Eu(TTA)3AA polymerization kinetic process in matrix was investigated in detail by controlling the temperature, the crosslinking agent, etc. The results showed that the peroxide could accelerate Eu(TTA)3AA self-polymerization in the rubber matrix, and therefore improved the dispersion, but not be helpful for the luminescence intensity enhancement. In addition, the relatively higher luminescence intensity in Eu(TTA)3AA/NBR in comparison to that of Eu(TTA)3AA/SiR might contribute to the interaction between nitrile group (–CN) in NBR and Eu-complexes, suggesting that the luminescence intensity of the composites also depended on the matrix characteristics.

There have been numerous recurrent neural network models diversely developed for modeling or simulating the associative memory behavior of human beings in the past decades, and the existing results for each model individual are in certain sense redundant with similarity. By utilizing the innate character of general activation operators, i.e., the uniformly pseudo-projection-anti-monotone property, a unified continuous-time recurrent neural network model is introduced, which can jointly cover almost all of the known continuous-time recurrent neural network individuals. Under the critical condition which is the intrinsic bounded line of stability and instability, we develop some convergence and stability theory for the unified recurrent neural network model when the time is continuous. The study shows that the approach adopted in the present paper is powerful, particularly in the sense of unifying, simplifying and extending the currently existing various models and dynamics results of continuous-time RNNs.

The sparsity driven classification technologies have attracted much attention in recent years, due to their capability of providing more compressive representations and clear interpretation. Two most popular classification approaches are support vector machines (SVMs) and kernel logistic regression (KLR), each having its own advantages. The sparsification of SVM has been well studied, and many sparse versions of 2-norm SVM, such as 1-norm SVM (1-SVM), have been developed. But, the sparsification of KLR has been less studied. The existing sparsification of KLR is mainly based on L1 norm and L2 norm penalties, which leads to the sparse versions that yield solutions not so sparse as it should be. A very recent study on L1/2 regularization theory in compressive sensing shows that L1/2 sparse modeling can yield solutions more sparse than those of 1 norm and 2 norm, and, furthermore, the model can be efficiently solved by a simple iterative thresholding procedure. The objective function dealt with in L1/2 regularization theory is, however, of square form, the gradient of which is linear in its variables (such an objective function is the so-called linear gradient function). In this paper, through extending the linear gradient function of L1/2 regularization framework to the logistic function, we propose a novel sparse version of KLR, the 1/2 quasi-norm kernel logistic regression (1/2-KLR). The version integrates advantages of KLR and L1/2 regularization, and defines an efficient implementation scheme of sparse KLR. We suggest a fast iterative thresholding algorithm for 1/2-KLR and prove its convergence. We provide a series of simulations to demonstrate that 1/2-KLR can often obtain more sparse solutions than the existing sparsity driven versions of KLR, at the same or better accuracy level. The conclusion is also true even in comparison with sparse SVMs (1-SVM and 2-SVM). We show an exclusive advantage of 1/2-KLR that the regularization parameter in the algorithm can be adaptively set whenever the sparsity (correspondingly, the number of support vectors) is given, which suggests a methodology of comparing sparsity promotion capability of different sparsity driven classifiers. As an illustration of benefits of 1/2-KLR, we give two applications of 1/2-KLR in semi-supervised learning, showing that 1/2-KLR can be successfully applied to the classification tasks in which only a few data are labeled.

The vertical stripe defects on silicon steel surface seriously affect the appearance and electromagnetic properties of silicon steel products. Eliminating such defects is adifficult and urgent technical problem. This paper investigates the relationship between the defects and their influence factors by classification methods. However, when the common classification methods are used in the problem, we cannot obtain a classifier with high accuracy. Byanalysis of the data set, we find that it is imbalanced and inconsistent. Because the common classification methods are based on accuracy-maximization criterion, they are not applicable to imbalanced and inconsistent data set. Thus, we propose asupport-degree-maximization criterion and anovel cost-sensitive loss function and also establish an improved L1/2 regularization approach for solution of the problem. Moreover, by employing reweighted iteration gradient boosting algorithm, we obtain a linear classifier with a high support degree. Through analyzing the classifier, we formulate a rule under which the silicon steel vertical stripe defects do not occur in the existing production environment. By applying the proposed rule to 50TW600 silicon steel production, the vertical stripe defects of the silicon steel products have been greatly decreased.

The backpropogation (BP) neural networks have been widely applied in scientific research and engineering. The success of the application, however, relies upon the convergence of the training procedure involved in the neural network learning. We settle down the convergence analysis issue through proving two fundamental theorems on the convergence of the online BP training procedure. One theorem claims that under mild conditions, the gradient sequence of the error function will converge to zero (the weak convergence), and another theorem concludes the convergence of the weight sequence defined by the procedure to a fixed value at which the error function attains its minimum (the strong convergence). The weak convergence theorem sharpens and generalizes the existing convergence analysis conducted before, while the strong convergence theorem provides new analysis results on convergence of the online BP training procedure. The results obtained reveal that with any analytic sigmoid activation function, the online BP training procedure is always convergent, which then underlies successful application of the BP neural networks.

Support vector classification (SVC) is an efficient tool to solve classification which is the fundamental problem in data mining, but its efficiency is limited on the small or middle sized data sets. In this paper a new preprocessing algorithm - clarifier method (CM) is proposed to accelerate training of SVC especially when training data are of large mode. The core idea of CM is based on the fact that of all the training vectors of SVC, only support vectors have a most significant influence on the optimal result, meanwhile the other vectors can be not taken into account with little deviation of the final result. The approach can distinguish probable support vectors directly by the labels of training vectors in its neighborhood region. Then only these training vectors are preserved to form the training set of SVC. A series of artificial and real world simulation results show the efficiency and effectiveness of CM. The number of vectors in SVC training becomes much less and the training time is decreased without loss of the generalization capability of the SVC even in the case that noisy training samples are added in

A new pre-processing method for regression is developed. The core idea is using three rules to clarify regression raw data. The rules are realized through introducing a judge function on the regression datum whose value determines the importance of the datum. By applying the rules, a new pre-processing method for regression is developed. Performance of the new method on a series of simulations demonstrate that it not only significantly increases computational efficiency and robustness, but also preserves generalization capability of a regression method. Incorporated with any regression method, the developed method then can be efficiently applied to regression of large data sets.