Structural Feature Extraction Research Articles

Dual-kernel driven convolutional sparse learning for bearing transient feature enhancement

The transient structure feature of bearing reflects the dynamic response of vibration, which is of vital importance for rotating machinery fault diagnosis. However, the measured signals are complicated and non-stationary, where the transient impacts are always submerged in heavy background noise. In this study, a novel dual-kernel driven convolutional sparse learning (DKCSL) is proposed to enhance the transient feature with the intrinsic structure mined. Considering the benefits of data-driven representation with time–frequency manifold (TFM) learning and model-driven representation with wavelet analysis, two time–frequency kernels with different principles are constructed for 2-dimensional convolutional sparse learning (CSL) in a fusion way of both data-driven and model-driven. From the perspective of image processing, the representation of desired transient structure features can be well sensed from time-frequency distribution (TFD) by those two dual-driven time–frequency kernels. In this manner, the desired defective intrinsic features can be mined, which is consistent with the physical dynamic response. In addition, image entropy is introduced to output the optimal TFM kernel and a series of inverse transform are applied to obtain the enhanced signal waveform. Simulated and experimental data are tested here to verify the creativity and effectiveness of the proposed DKCSL method, and a comparison with the traditional single TFM kernel-based sparse result and wavelet kernel-based sparse result are further provided, which deeply indicates the originality of our method that its potential in intrinsic structure feature extraction with dual-kernel sensing for bearing fault diagnosis.

GA-SCS: Graph-Augmented Source Code Summarization

Automatic source code summarization system aims to generate a valuable natural language description for a program, which can facilitate software development and maintenance, code categorization, and retrieval. However, previous sequence-based research did not consider the long-distance dependence and highly structured characteristics of source code simultaneously. In this article, we present a Transformer-based Graph-Augmented Source Code Summarization (GA-SCS), which can effectively incorporate inherent structural and textual features of source code to generate an effective code description. Specifically, we develop a graph-based structure feature extraction scheme leveraging abstract syntax tree and graph attention networks to mine global syntactic information. And then, to take full advantage of the lexical and syntactic information of code snippets, we extend the original attention to a syntax-informed self-attention mechanism in our encoder. In the training process, we also adopt a reinforcement learning strategy to enhance the readability and informativity of generated code summaries. We utilize the Java dataset and Python dataset to evaluate the performance of different models. Experimental results demonstrate that our GA-SCS model outperforms all competitive methods on BLEU, METEOR, ROUGE, and human evaluations.

A dual attention driven multiscale-multilevel feature fusion approach for hyperspectral image classification

ABSTRACT Deep learning has achieved promising results for hyperspectral image (HSI) classification in recent years due to its hierarchical structure and automatic feature extraction ability from raw data. The HSI has continuous spectral information, allowing for the precise identification of materials by capturing minute spectral differences. Convolutional neural networks (CNNs) have proven to be effective feature extractors for HSI classification. However, inherent network limitations prevent them from adequately mining and representing the sequence attributes of spectral signatures and learning critical and valuable features from both spectral and spatial dimensions simultaneously. This paper proposes a deep learning-based framework called a novel dual attention-based multiscale-multilevel ConvLSTM3D (DAMCL) to address these challenges. In this work, our contribution is threefold; firstly, a dual attention mechanism is proposed, effectively learning critical and valuable features from spectral and spatial dimensions. Secondly, multiscale ConvLSTM3D blocks can learn the discriminative features alongside handling long-range dependencies of spectral data. Thirdly, these features are combined by a multilevel feature fusion approach to maximize the impact of features learned at different levels. To assess the performance of the proposed method, extensive experiments are carried out on five different benchmark datasets containing complex and challenging land cover classes. The results confirm that the proposed method outperforms state-of-the-art techniques with a small number of training samples in terms of overall accuracy (OA), average accuracy (AA), and Kappa . The overall accuracy of 98.88%, 99.42%, 99.20%, 95.37%, and 92.57% is achieved over the Indian Pines, Salinas Valley, University of Pavia, Houston 2013, and Houston 2018 datasets, respectively.

A deep neural network model for coreference resolution in geological domain

Coreference resolution of geological entities is an important task in geological information mining. Although the existing generic coreference resolution models can handle geological texts, a dramatic decline in their performance can occur without sufficient domain knowledge. Due to the high diversity of geological terminology, coreference is intricately governed by the semantic and expressive structure of geological terms. In this paper, a framework CorefRoCNN based on RoBERTa and convolutional neural network (CNN) for end-to-end coreference resolution of geological entities is proposed. Firstly, the fine-tuned RoBERTa language model is used to transform words into dynamic vector representations with contextual semantic information. Second, a CNN-based multi-scale structure feature extraction module for geological terms is designed to capture the invariance of geological terms in length, internal structure, and distribution. Thirdly, we incorporate the structural feature and word embedding for further determinations of coreference relations. In addition, attention mechanisms are used to improve the ability of the model to capture valid information in geological texts with long sentence lengths. To validate the effectiveness of the model, we compared it with several state-of-the-art models on the constructed dataset. The results show that our model has the optimal performance with an average F1 value of 79.78%, which is a 1.22% improvement compared to the second-ranked method.

A Spatial Alignment Method for UAV LiDAR Strip Adjustment in Nonurban Scenes

LiDAR strip adjustment is a key prerequisite for subsequent applications based on point cloud data since it inevitably suffers from spatial discrepancies caused by laser ranging errors, mounting errors, etc. Most current LiDAR strip adjustment methods rely on the extraction of structural features which are often unsuitable for non-urban scenes. Alternative strip adjustment methods based on correspondence distance minimalization ignore spatial alignment. To overcome these limitations, this paper presents an accurate spatial alignment method for UAV LiDAR strip adjustment in non-urban scenes. Firstly, we construct a novel point cloud feature descriptor called Spherical Shell Point Feature (SSPF) to extract multi-dimensional non-structural features that are robust to non-urban point clouds. The constructed SSPF is then combined with point coordinates to generate embedded features, which simultaneously consider the point coordinates and spatial alignment. Finally, the embedded features are utilized by a two-stage matching method to match pair-wise points of two adjacent strips. The proposed method is validated on two non-urban datasets collected by two types of LiDARs, which reduces the digital surface model discrepancies by 0.252 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</i> and 0.221 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">m</i> , respectively, and proves its superiority compared to mainstream strip adjustment methods as well.

CPInformer for Efficient and Robust Compound-Protein Interaction Prediction.

Recently, deep learning has become the mainstream methodology for Compound-Protein Interaction (CPI) prediction. However, the existing compound-protein feature extraction methods have some issues that limit their performance. First, graph networks are widely used for structural compound feature extraction, but the chemical properties of a compound depend on functional groups rather than graphic structure. Besides, the existing methods lack capabilities in extracting rich and discriminative protein features. Last, the compound-protein features are usually simply combined for CPI prediction, without considering information redundancy and effective feature mining. To address the above issues, we propose a novel CPInformer method. Specifically, we extract heterogeneous compound features, including structural graph features and functional class fingerprints, to reduce prediction errors caused by similar structural compounds. Then, we combine local and global features using dense connections to obtain multi-scale protein features. Last, we apply ProbSparse self-attention to protein features, under the guidance of compound features, to eliminate information redundancy, and to improve the accuracy of CPInformer. More importantly, the proposed method identifies the activated local regions that link a CPI, providing a good visualisation for the CPI state. The results obtained on five benchmarks demonstrate the merits and superiority of CPInformer over the state-of-the-art approaches.

Drug-target interaction prediction using high order nonlinear features via neural factorization machines

Medication design and repositioning are sped up by the prediction of drug-target interactions (DTIs). Two main kinds of prediction methods are commonly used, which are based on chemical structure feature extraction and deep learning methods. However, on the one hand, the DTI prediction approaches based on chemical structural feature extraction may not wholly explore the possible network characteristics in the data. On the other hand, many deep learning methods call for numerous layers of neural networks to be layered to learn higher-order feature interaction information. To sum up, the existing computation methods often have the limitations of gradient disappearance and overfitting. This study presents a novel method (JRD-NFM) by calculating Jaccard similarities, getting an eigenvector through Restarted random walk (RWR), and generating low-dimensional feature vectors by Disposition Component Analysis (DCA). Besides decoding the topological features and similarity information of target and drug node, it can also get the context information of a single network. Considering the advantages of Neural Factorization Machines (NFM) in extracting high-order nonlinear features and processing sparse data, this study use NFM to classifier the collection data to integrate drug and target biochemical structure information. The experimental results demonstrate that JRD-NFM can outperform widely used deep learning methods and conventional chemical structure approaches. It could provide fresh ideas for combining molecular structure and heterogeneous network data to predict DTIs.

A Study on Structural and Textural Feature Extraction for Contactless Fingerprint Classification

Objectives: The Recent COVID 19 Pandemic has capped the efficacy of the mostly used existing touch-based fingerprint detection. Hence, the main objective of this study is to develop a lightweight, robust and efficient touchless fingerprint identification model. To cope up with touchless environment demands, an emphasis is made on improvement in local conditions, feature modalities as well as learning environment. Methods: Considering these objectives, the focus of this study is on two different methods for classification of contactless fingerprint. In the first method structural features like minutiae details are extracted followed by classification by SSIM. In the second method GLCM textural features were extracted followed by classification using Random Forest algorithm. In the proposed method, performance assessment is done by considering data samples of 1000 random users that are collected from different benchmark databases like Hong Kong Polytechnic University 3D-fngerprint images Database Version 2.0, Touchless Fingerprint Database of IIT Bombay, IIT Kanpur, IIT Jodhpur. Findings: Though, touchless fingerprint detection is considered as a viable alternative; yet, the real-time data complexities like non-linear textural patterns, dusts, non-uniform local conditions like illumination, contrast, orientation make it complex for realization. Moreover, the likelihood of ridge discontinuity and texture damages can majorly limit its efficacy. Novelty: The proposed model mainly focusses on reducing Equal Error Rate and improving the accuracy of contactless fingerprint classification by extracting textural features rather just sticking to conventional structural feature-minutiae. The Proposed method outperforms when compared with the existing state of the art methods by achieving an accuracy of 94.72%, precision of 98.84%, recall of 97.716%, FMeasure 0.9827 and a reduced EER of about 0.084. The key novelty of this approach was that it doesn’t require any surface 3D reconstruction, rather it employed different mathematical approaches to retrieve surface normal and minutiae information. Keywords: SSIM; GLCM; Contactless Fingerprint; Minutiae; EER; Confusion Matrix

Cross-domain fault diagnosis of rotating machinery based on graph feature extraction

Transfer learning can realize cross-domain fault diagnosis of rotating machinery, where the model trained on many labeled samples collected in one working condition can be transferred to insufficient samples collected in the target working condition. Currently, the data features cannot be completely extracted by existing methods when the data distribution gap of the samples collected in different working conditions is quite large. In order to fully extract the data features of rotating machinery to achieve cross-domain fault diagnosis, this paper investigated a cross-domain fault diagnosis model of rotating machinery based on graph feature extraction. The proposed method can realize unsupervised fault diagnosis on rotating machinery running under different working conditions by extracting the numerical and structural features of source and target domains. First of all, data features with large data distribution gaps need to be fully extracted, so a convolutional network based on a deformable convolutional network (De-conv) is designed to extract the features with large differences in data distribution under various working conditions. Secondly, features are extracted based on a convolutional neural network for data values in existing domain adaptation (DA) methods while the structure features of machine monitoring data are ignored. Therefore, a composite spectral-based graph convolutional network is designed to extract structural features of data. Thirdly, fully extracted features are input into a universal DA network to achieve cross-domain fault diagnosis of unknown faults in rotating machinery under changing working conditions. Finally, a benchmarking data set and a data set collected from a practical experimental platform are used to verify the effectiveness of the proposed model, and the results show that it is more suitable for cross-domain fault diagnosis of rotating machinery than other comparison models.

Seismic P‐wave first‐arrival picking model based on EQK‐IncResNet

SummarySeismic P‐wave first arrival picking is one of the critical problems in seismic source research. At present, the picking of seismic P‐wave first arrival mainly relies on the combination of traditional methods and manual picking. However, traditional algorithms have poor picking performance and low manual picking efficiency, which cannot meet the needs of research and application. Based on the deep learning network InceptionResNet‐V2, this paper proposes an improved network (EQK‐IncResNet) that can be used for seismic P‐wave first arrival picking. Compared with traditional methods, this model does not need to set the threshold manually and only needs to input the three‐component data of the seismic waveform data to intelligently identify the first‐arrival of the seismic P‐wave. This model has a lightweight structure and excellent feature extraction capabilities, which can effectively identify low signal‐to‐noise ratio data. The experimental results show that within the error thresholds of 0.1, 0.3, and 0.5 s, the hit rate of this method is 74.45%, 96.79%, and 98.68%, respectively. The average picking error is 0.031 s, which is better than the traditional methods such as AR‐AIC + STA/LTA and the mainstream deep learning methods such as GRU.

GEOLOGICAL FEATURES IDENTIFIED FROM FIELD OBSERVATIONS AND REMOTE SENSING DATA ON THE UM TAGHIR AREA, EASTERN DESERT, EGYPT

The current study presents the integration between field observations and remotely sensed data for detection and extraction of geological structural features using Sentinel-2A and Aster DEM images. The area under investigation is represented by the Neoproterozoic East African Orogeny encompassing a part of the Arabian Nubian Shield. All the rock units studied belong to the Late Cryogenian-Ediacaran periods that are divided into two large terrains (continental arc terrain and oceanic arc terrain). The metagabbro and metavolcaniclastic rocks make up the oceanic terrain, while the gabbro, tonalite, granodiorite, dokhan volcanic, monzogranite and alkali feldspar granite comprise continental arc terrain. The Sentinel-2A remote sensing and ASTER DEM data have meaningful application in respect of geological interpretation. Lineament analysis is one of the most useful tools in geological mapping and mineral exploration. Several methods of processing and extracting lithological information and lineaments were applied to the Sentinel-2A and ASTER DEM data covering the present study. The methods include various image enhancements (FCC, MNF and PCA) and the application of directional filters (Sobel). The study results show that the area was subjected to stresses of various directions (WNW – ESE, NW-SE, NE-SW, N-S, NNE – SSW, and E-W). There occurred some important structure-related and mineralization events like migmatization (in granodiorites) and bearing mineralization (in gabbro), which are associated with major elements of evidence-based structural control of the area and with the proximity of the Quena Safaga shear zone-related mineralization.

An improved two-stream compression convolution network for rolling bearing fault diagnosis

The timely and accurate fault diagnosis of rolling bearings plays a vital role in ensuring the safe and reliable operation of many complex mechanical systems. However, most of the existing rolling bearing fault diagnosis models have complex structures, need a large number of samples, and cannot fully use the information contained in the signal. Based on these problems, an intelligent method for detecting and identifying rolling bearing faults is proposed based on an improved two-stream compression convolution network. The proposed method has a compact structure and powerful feature extraction capability, which consists of two modules. One adopts depthwise separable convolution and squeeze and excitation network, named DSCSE, which can fully extract the spatial features and greatly reduce the number of parameters. The other utilizes a one-dimensional convolutional neural network and spatial dropout mechanism, named 1DCNNSD, which can efficiently extract the temporal features and reduce model complexity. Meanwhile, to reduce the loss of information contained in the signal when the linear rectification unit is under negative input, the improved ReLU is designed. Numerous experiments demonstrate that the novel approach has higher accuracy, better generalization performance, and robustness than other methods under small samples.

NeurstrucEnergy: A bi-directional GNN model for energy prediction of neural networks in IoT

A significant demand rises for energy-efficient deep neural networks to support power-limited embedding devices with successful deep learning applications in IoT and edge computing fields. An accurate energy prediction approach is critical to provide measurement and lead optimization direction. However, the current energy prediction approaches lack accuracy and generalization ability due to the lack of research on the neural network structure and the excessive reliance on customized training dataset. This paper presents a novel energy prediction model, NeurstrucEnergy. NeurstrucEnergy treats neural networks as directed graphs and applies a bi-directional graph neural network training on a randomly generated dataset to extract structural features for energy prediction. NeurstrucEnergy has advantages over linear approaches because the bi-directional graph neural network collects structural features from each layer's parents and children. Experimental results show that NeurstrucEnergy establishes state-of-the-art results with mean absolute percentage error of 2.60%. We also evaluate NeurstrucEnergy in a randomly generated dataset, achieving the mean absolute percentage error of 4.83% over 10 typical convolutional neural networks in recent years and 7 efficient convolutional neural networks created by neural architecture search. Our code is available at https://github.com/NEUSoftGreenAI/NeurstrucEnergy.git.

Community detection based on competitive walking network embedding method

Currently, much of the information of the real world is network-structured, and extracting hidden information from network-structured data helps to understand the corresponding systems, but can also be a challenging problem. In recent years, network embedding has been an effective way to extract network information, which represents nodes in complex networks as low-dimensional space vectors, while preserving the properties of the network. Community attributes are an important property of networks, and in most network embedding algorithms, the community structure is usually ignored or cannot be explicitly preserved. In this paper, we propose a new network embedding framework that explicitly considers community structure feature extraction. The framework, called competitive walking network embedding (CWNE), extracts sample sequences by competitive walking and obtains node representation vectors by skip-gram training. Competitive walking allows the extracted sample sequences to be concentrated within the same community, effectively preserving the community structure features of the network. The results of testing the proposed method on artificial and real-world networks show that our model is more effective in detecting community structure in networks. In addition, visualization experiments show that the results of CWNE show that nodes from the same community are more tightly distributed in low-dimensional space.

Improving Network Representation Learning via Dynamic Random Walk, Self-Attention and Vertex Attributes-Driven Laplacian Space Optimization.

Network data analysis is a crucial method for mining complicated object interactions. In recent years, random walk and neural-language-model-based network representation learning (NRL) approaches have been widely used for network data analysis. However, these NRL approaches suffer from the following deficiencies: firstly, because the random walk procedure is based on symmetric node similarity and fixed probability distribution, the sampled vertices’ sequences may lose local community structure information; secondly, because the feature extraction capacity of the shallow neural language model is limited, they can only extract the local structural features of networks; and thirdly, these approaches require specially designed mechanisms for different downstream tasks to integrate vertex attributes of various types. We conducted an in-depth investigation to address the aforementioned issues and propose a novel general NRL framework called dynamic structure and vertex attribute fusion network embedding, which firstly defines an asymmetric similarity and h-hop dynamic random walk strategy to guide the random walk process to preserve the network’s local community structure in walked vertex sequences. Next, we train a self-attention-based sequence prediction model on the walked vertex sequences to simultaneously learn the vertices’ local and global structural features. Finally, we introduce an attributes-driven Laplacian space optimization to converge the process of structural feature extraction and attribute feature extraction. The proposed approach is exhaustively evaluated by means of node visualization and classification on multiple benchmark datasets, and achieves superior results compared to baseline approaches.

Seizure detection algorithm based on improved functional brain network structure feature extraction

Epilepsy is one of the most common neurological disorders. Accurate detection of epileptic seizures is essential for treatment. A seizure detection method with the structure of functional brain network, time, and frequency multi-domain features is presented in this paper. Pearson correlation coefficient (PCC) and mutual information (MI) can characterize the correlation between channels of EEG, and be used to construct PCC and MI combined functional brain networks (PMNet). Then the complex network theory is used to calculate the structure features of PMNet to explore whether it reflects seizure or not. After extraction of the multi-domain features, principal component analysis (PCA) is used to eliminate irrelevant feature information before feeding the features into the support vector machine (SVM) classification method. The method is tested on two open datasets (CHB-MIT and SIENA) and verified on the West syndrome (WS) dataset from the Children’s Hospital, Zhejiang University School of Medicine. The results demonstrate that the method based on the structure of functional brain networks and other multi-domain EEG features is competitive with other comparison methods. Furthermore, base on the method with brain network features, this paper has developed a simple automatic seizure detection system in hospital applications, which can reduce the workload of EEG technicians in reading EEG and marking seizures.

Transportation Mode Detection Combining CNN and Vision Transformer with Sensors Recalibration Using Smartphone Built-In Sensors.

Transportation Mode Detection (TMD) is an important task for the Intelligent Transportation System (ITS) and Lifelog. TMD, using smartphone built-in sensors, can be a low-cost and effective solution. In recent years, many studies have focused on TMD, yet they support a limited number of modes and do not consider similar transportation modes and holding places, limiting further applications. In this paper, we propose a new network framework to realize TMD, which combines structural and spatial interaction features, and considers the weights of multiple sensors’ contributions, enabling the recognition of eight transportation modes with four similar transportation modes and four holding places. First, raw data is segmented and transformed into a spectrum image and then ResNet and Vision Transformers (Vit) are used to extract structural and spatial interaction features, respectively. To consider the contribution of different sensors, the weights of each sensor are recalibrated using an ECA module. Finally, Multi-Layer Perceptron (MLP) is introduced to fuse these two different kinds of features. The performance of the proposed method is evaluated on the public Sussex-Huawei Locomotion-Transportation (SHL) dataset, and is found to outperform the baselines by at least 10%.

An ensemble-based approach using structural feature extraction method with class imbalance handling technique for drug-target interaction prediction

An ensemble-based approach using structural feature extraction method with class imbalance handling technique for drug-target interaction prediction

Towards using visual, semantic and structural features to improve code readability classification

Context:Code readability, which correlates strongly with software quality, plays a critical role in software maintenance and evolvement. Although existing deep learning-based code readability models have reached a rather high classification accuracy, only structural features are utilized which inevitably limits their model performance. Objective:To address this problem, we propose to extract readability-related features from visual, semantic, and structural aspects from source code in an attempt to further improve code readability classification. Method:First, we convert a code snippet into a RGB matrix (for visual feature extraction), a token sequence (for semantic feature extraction) and a character matrix (for structural feature extraction). Then, we input them into a hybrid neural network that is composed of BERT, CNN, and BiLSTM for feature extraction. Finally, the extracted features are concatenated and input into a classifier to make a code readability classification. Result:A series of experiments are conducted to evaluate our method. The results show that the average accuracy could reach 85.3%, which outperforms all existing models. Conclusion:As an innovative work of extracting readability-related features automatically from visual, semantic, and structural aspects, our method is proved to be effective for the task of code readability classification.

Adaptive spatiotemporal graph convolutional network with intermediate aggregation of multi-stream skeleton features for action recognition

Video-based action recognition is a challenging problem due to the rapid and uncertain changes in human actions. Recent studies show that incorporating video and human body skeleton helps improve action recognition performance. These methods generally use graph convolutional networks (GCNs) to extract structural features of the human body joints from skeleton data. Yet, most GCN-based methods have some limitations in skeleton-based action recognition. (1) The graph structure of the human body joints is time-invariant, making it difficult to represent the changing relationship between joints across actions. (2) Methods relying on single-stream models only utilize limited information of skeleton data, such as joints or bones, and fail to consider coherent features of movements. (3) Methods relying on multi-stream models have considerable parameters and are inefficient for real-life applications. To address these problems, we propose an adaptive spatiotemporal graph convolutional network with intermediate aggregation of multi-stream skeleton features for action recognition. First, our method learns an adaptive graph structure representing the changing relationship between joints. Secondly, we facilitate a multi-stream model to extract various features from the skeleton, including joint-stream, bone-stream, and motion-stream. Moreover, an intermediate aggregation strategy is employed to aggregate these features and to reduce the parameters of this model. The proposed method has been validated on various benchmarks and a real-world abnormal action dataset. Extensive experimental results show that our method achieves excellent performance in skeleton-based action recognition.