Dynamic Convolutional Neural Network Research Articles

A lightweight dynamic dual-damped wavelet-based convolutional neural network for interpretable bearing fault diagnosis

Wavelet-based convolutional neural networks (CNNS) have attracted widespread attention because they can improve the interpretability of intelligent fault diagnosis methods. However, the fault feature representation capability of typical wavelet-based convolution kernel frameworks must be strengthened to improve the diagnostic accuracy of complex faults. In the meantime, the large number of network parameters leads to high computational costs. To address these issues, a lightweight wavelet-based dynamic CNN, which comprises a dual-damping wavelet-based dynamic CNN (DWDC) block and a discrete wavelet transformation (DWT) enhancement (DWTE) block, is put forward. In the DWDC block, a wavelet convolution layer is initially designed, where a dual-damping wavelet is used as the kernel function to improve the match of the convolution kernel with fault impulses. Subsequently, a dynamic convolution layer with multiple parallel small-size convolutional kernels is designed to screen the fault features instead of a multilayer network structure, thereby greatly reducing the number of network parameters. Finally, the DWTE block is constructed by combining the DWT and residual dense block, and it can mine more fault information from the previously extracted features. The experiments on the variable speed bearing dataset, locomotive bearing dataset with constant speed and the Case Western Reserve University dataset prove that the proposed approach outperforms five classical CNN models and six advanced wavelet-based CNN models. In addition, it can effectively solve the issue of data imbalance because of its powerful feature extraction capability.

A novel lightweight dynamic focusing convolutional neural network LAND-FCNN for EEG emotion recognition

The inefficiency of model inference and the limited sample data are significant issues in electroencephalogram (EEG) emotion recognition models based on neural networks. Therefore, we propose a novel lightweight adaptive dynamic focusing convolutional neural network (LAND-FCNN). We use Partial Convolution (PConv) and Batch interactive attention in constructing the Backbone network. PConv reduces the computational cost by selectively extracting features from a subset of input channels, utilizing feature graph redundancy. Batch interactive attention establishes connections between various types of samples to address sample scarcity without relying on data augmentation. Then Backbone serves as the encoder for the Glance and Focus Dynamic Inference Networks (GFnet). The patch proposal network in GFnet adaptively captures task-relevant areas of each sample, allowing encoders to process only those areas. This increases inference speed and ensures reliable results with minimal computational effort. Extensive validation of diverse datasets with varying distributions confirms the efficacy and precision of LAND-FCNN.

Real-time Classification of Brain States in Functional MRI Using Dynamic Connectivity Patterns and Machine Learning

Understanding real-time brain states facilitates deeper insights into cognitive processes, emotional responses, and neurological phenomena. It provides researchers with a dynamic view of brain function, aiding in the development of novel therapies, advancing neuroscience, and fostering innovations in brain-computer interfaces and artificial intelligence. The research aims to achieve real-time classification of brain states using dynamic connectivity patterns and Convolutional Neural Network (CNN) algorithms. It focuses on how demographic variables, such as brain volume and medication usage, influence classification accuracy. The study employs Python programming with FSL and Nilearn for preprocessing tasks like motion correction and dynamic connectivity pattern extraction. Feature selection and data partitioning are managed using Scikit-learn, ensuring standardized feature values and handling missing data. The CNN architecture is customized to handle spatial and temporal features in functional MRI (fMRI) data, with convolutional layers extracting spatial features representing local connectivity patterns. Dynamic connectivity matrices visualization aids in understanding brain network reconfigurations over time. This approach offers real-time insights into cognitive processes and neurological disorders, guiding personalized interventions for brain health. In result, connectivity matrices and 3D brain network visualizations are pivotal for unraveling brain state dynamics. Connectivity matrices unveil intricate interactions among brain regions across resting, task execution, and stress response states, offering quantitative insights into functional connectivity patterns. Meanwhile, 3D visualizations provide spatial representations, showcasing complex interplays and architectural changes across states.

Real-time Classification of Brain States in Functional MRI Using Dynamic Connectivity Patterns and Machine Learning

Understanding real-time brain states facilitates deeper insights into cognitive processes, emotional responses, and neurological phenomena. It provides researchers with a dynamic view of brain function, aiding in the development of novel therapies, advancing neuroscience, and fostering innovations in brain-computer interfaces and artificial intelligence. The research aims to achieve real-time classification of brain states using dynamic connectivity patterns and Convolutional Neural Network (CNN) algorithms. It focuses on how demographic variables, such as brain volume and medication usage, influence classification accuracy. The study employs Python programming with FSL and Nilearn for preprocessing tasks like motion correction and dynamic connectivity pattern extraction. Feature selection and data partitioning are managed using Scikit-learn, ensuring standardized feature values and handling missing data. The CNN architecture is customized to handle spatial and temporal features in functional MRI (fMRI) data, with convolutional layers extracting spatial features representing local connectivity patterns. Dynamic connectivity matrices visualization aids in understanding brain network reconfigurations over time. This approach offers real-time insights into cognitive processes and neurological disorders, guiding personalized interventions for brain health. In result, connectivity matrices and 3D brain network visualizations are pivotal for unraveling brain state dynamics. Connectivity matrices unveil intricate interactions among brain regions across resting, task execution, and stress response states, offering quantitative insights into functional connectivity patterns. Meanwhile, 3D visualizations provide spatial representations, showcasing complex interplays and architectural changes across states.

A Novel Money Laundering Prediction Model Based on a Dynamic Graph Convolutional Neural Network and Long Short-Term Memory

Money laundering is an illicit activity that seeks to conceal the nature and origins of criminal proceeds, posing a substantial threat to the national economy, the political order, and social stability. To scientifically and reasonably predict money laundering risks, this paper focuses on the “layering” stage of the money laundering process in the field of supervised learning for money laundering fraud prediction. A money laundering and fraud prediction model based on deep learning, referred to as MDGC-LSTM, is proposed. The model combines the use of a dynamic graph convolutional network (MDGC) and a long short-term memory (LSTM) network to efficiently identify illegal money laundering activities within financial transactions. MDGC-LSTM constructs dynamic graph snapshots with symmetrical spatiotemporal structures based on transaction information, representing transaction nodes and currency flows as graph nodes and edges, respectively, and effectively captures the relationships between temporal and spatial structures, thus achieving the dynamic prediction of fraudulent transactions. The experimental results demonstrate that compared with traditional algorithms and other deep learning models, MDGC-LSTM achieves significant advantages in comprehensive spatiotemporal feature modeling. Specifically, based on the Elliptic dataset, MDGC-LSTM improves the Macro-F1 score by 0.25 compared to that of the anti-money laundering fraud prediction model currently considered optimal.

SANe: Space adaptation network for temporal knowledge graph completion

Temporal Knowledge Graphs (TKGs) model time-dependent facts as relations between entities at specific timestamps, making them well-suited for real-world scenarios. However, TKGs are susceptible to incompleteness, necessitating Temporal Knowledge Graph Completion (TKGC) to predict missing facts. Prior methods often struggle to effectively handle two critical properties of TKGs, time-variability and time-stability, simultaneously, which hinders their performance. In this paper, we propose Space Adaptation Network (SANe), a novel approach for TKGC. SANe adapts facts at different timestamps to distinct latent spaces, effectively addressing time-variability. Our model introduces Parameter Generation Network to produce separate neural networks for each snapshot, which are then encoded into different latent spaces. A dynamic convolutional neural network processes entities and relations, utilizing different learned parameters generated by parameter generation network with respect to timestamps. By handling different temporal snapshots separately, TKGC is transformed into static KGC, enabling the modeling of time-variability. Dynamic convolutional neural network efficiently learns collective knowledge over large periods and supplements more specific knowledge gradually in smaller periods, facilitating time-stability. To strike a balance between learning time-variability and time-stability, we introduce a time-aware parameter generator to produce parameters hierarchically based on year, month, and day timestamps. Long-term knowledge is effectively shared across adjacent snapshots within the same year or month, while short-term knowledge within a day is preserved in specific parameters. However, in unbalanced TKGs, where many facts occur in small intervals, the large number of parameters generated by time-aware parameter generator may remain underutilized. To address this, we propose Adaptive Parameter Generation with a partition tree, ensuring parameter load balancing while maintaining time-stability. We conduct extensive experiments on five benchmark datasets, demonstrating the superiority of SANe over existing methods for TKGC, achieving state-of-the-art performance. Our contributions include pioneering TKGC from the perspective of space adaptation, achieving a balance between time-variability and time-stability through latent space overlap constraints, and substantiating the effectiveness of our model through comprehensive experiments on rich temporal datasets.

Earthquake monitoring using deep learning with a case study of the Kahramanmaras Turkey earthquake aftershock sequence

Abstract. Seismic phase picking and magnitude estimation are fundamental aspects of earthquake monitoring and seismic event analysis. Accurate phase picking allows for precise characterization of seismic wave arrivals, contributing to a better understanding of earthquake events. Likewise, accurate magnitude estimation provides crucial information about an earthquake's size and potential impact. Together, these components enhance our ability to monitor seismic activity effectively. In this study, we explore the application of deep-learning techniques for earthquake detection and magnitude estimation using continuous seismic recordings. Our approach introduces DynaPicker, which leverages dynamic convolutional neural networks to detect seismic body-wave phases in continuous seismic data. We demonstrate the effectiveness of DynaPicker using various open-source seismic datasets, including both window-format and continuous recordings. We evaluate its performance in seismic phase identification and arrival-time picking, as well as its robustness in classifying seismic phases using low-magnitude seismic data in the presence of noise. Furthermore, we integrate the phase arrival-time information into a previously published deep-learning model for magnitude estimation. We apply this workflow to continuous recordings of aftershock sequences following the Turkey earthquake. The results of this case study showcase the reliability of our approach in earthquake detection, phase picking, and magnitude estimation, contributing valuable insights to seismic event analysis.

A new paradigm in electron microscopy: Automated microstructure analysis utilizing a dynamic segmentation convolutional neutral network

Over the past half century, the transmission electron microscope enabled insight into the fundamental arrangements and structures of materials. State-of-the-art electron microscopes can acquire large image datasets across multiple imaging modalities. However, the manual annotation process for feature or defect quantification may not be feasible with the modern microscope. Convolutional neural networks emerged to characterize individual microstructural features from an image in a cost-effective, consistent manner. However, many of these neural network approaches rely on thousands to hundreds of thousands of manual annotations of each feature type across hundreds of images to train the network for adequate performance. This work focused on the development and application of a pixel-wise defect detection machine-learning dynamic segmentation convolutional neural network with associated automated acquisition and postprocessing to identify microstructural features rapidly and quantitatively from a small initial dataset incorporating multiple imaging modes. The approach was demonstrated for characterization of superalloy 718 from both single image acquisition on multiple detectors to in-situ evolution captured with a single detector on a standard desktop computer to demonstrate the low barrier to entry required for widespread adoption. Pixel-by-pixel class identification was excellent with strong identification of chemically distinct phases, structurally distinct phases, and defect structures, thus demonstrating the new paradigm of machine learning-assisted characterization.

A workload identification method of industrial robot combining dynamic model and convolutional neural network

The motion joint system of industrial robot has obvious nonlinear characteristics, high dimensional running data and limited number of experimental samples. If mature neural network algorithm is used for identifying the workload of robots, it is easy to have over-learning problems, which seriously restricts the generalization ability of load identification method. In this paper, a load identification method combining the dynamic model of industrial robots with the neural network data model is proposed. A UR5 robot is used for carrying out multiple dynamic load testings to verify the effectiveness of the proposed identification method. Firstly, the workload identification by CNN algorithm is given, and the influence of parameters of prediction model and different neutral network are analyzed. Then the classical dynamic model of industrial robots with multi-degrees of freedom is established. The identified workload by dynamic model is also analyzed. At last, the deterministic information such as velocity and displacement is extracted from the calculation results of the dynamic model as the initial anchoring value. Then convolutional neural network is applied for compensating the residual highly nonlinear information. An improved mixing combination method is also proposed. This method can effectively deal with the interaction of different types of information in the data, and preliminarily cooperate the dynamic model and convolutional neural network. This provides a basic framework and method for solving the problem of parameter identification in multi-degree-of-freedom systems with small samples.

A new multiregional carbon emissions forecasting model based on a multivariable information fusion mechanism and hybrid spatiotemporal graph convolution network

Developing scientific and effective carbon emissions reduction policies relies heavily on precise carbon emission trend prediction. The existing complex spatiotemporal correlation and diverse range of influencing factors associated with multi-regional carbon emissions pose significant challenges to accurately modeling these trends. Under this constraint, this study is inspired by graph learning to establish a hybrid dynamic and static graph-based regional carbon emission network framework, which introduces a novel research standpoint for investigating short-term carbon emissions prediction (CEP). Specifically, a parallel framework of attribute-augmented dynamic multi-modal graph convolutional neural networks (ADMGCN) and temporal convolutional networks with adaptive fusion multi-scale receptive fields (AFMRFTCN) is proposed. The proposed model is evaluated against nineteen state-of-the-art models using daily carbon emission data from 30 regions in China, demonstrating its effectiveness in accurately predicting the trends of multi-regional carbon emissions. Conclusions are drawn as follows: First, especially in regions with marked periodicity, compared with the best baseline model, the mean absolute percentage error (MAPE) of our model is reduced by 20.19%. Second, incorporating graph convolutional neural networks (GCNs) with dynamic and static graphs is advantageous in extracting the spatial features of China's carbon emission network, which are influenced by geographical, economic, and industrial factors. Third, the parallel ADMGCN-AFMRFTCNs framework effectively captures the influence of external information on carbon emissions while mitigating the issue of low prediction accuracy resulting from univariate information. Fourth, the analysis reveals significant differences in the short-term (30-day) growth rate of carbon emissions among different regions. For example, Henan exhibits the highest growth rate (37.38%), while Guizhou has the lowest growth rate (−7.46%). It is valuable for policymakers and stakeholders seeking to identify regions with distinct emission patterns and prioritize mitigation efforts accordingly.

Dynamic Convolutional Neural Networks as Efficient Pre-Trained Audio Models

Dynamic Convolutional Neural Networks as Efficient Pre-Trained Audio Models

Ultra-short-term prediction of regional photovoltaic power based on dynamic graph convolutional neural network

Ultra-short-term prediction of regional photovoltaic power based on dynamic graph convolutional neural network

Dynamic Generative Residual Graph Convolutional Neural Networks for Electricity Theft Detection

Dynamic Generative Residual Graph Convolutional Neural Networks for Electricity Theft Detection

Estimating tree species composition from airborne laser scanning data using point-based deep learning models

Airborne laser scanning (ALS) data is increasingly being used for accurate estimations of forest inventory attributes, such as tree height and volume. However, determining tree species information, an important attribute required by inventories for a range of forest management applications, is challenging to extract from ALS data. This is due to complex species assemblages and vertical structures in certain forest environments, and a lack of spectral information for most ALS sensors. Our objective was to estimate tree species compositions in a Canadian boreal forest environment using ALS data and point-based deep learning techniques. To accomplish this, we utilised existing polygonal forest resource information to develop a large comprehensive dataset of tree species compositions across a 630,000 ha forest management area. We then used an adapted PointAugment generative adversarial network (GAN) coupled with the dynamic graph convolutional neural network to estimate tree species proportions. This innovative deep learning approach resulted in an overall R2adj of 0.61 for all tree species proportions. A weighted F1 score of 63% was observed when classifying the leading species of a plot, 66% for leading genus, and 85% when distinguishing between coniferous and deciduous dominated plots. Furthermore, the PointAugment GAN successfully generated augmented point clouds of forest plots which can alleviate issues associated with limited training samples for use in deep learning. This approach demonstrates the capability of point-based deep learning techniques to accurately estimate tree species compositions from ALS point clouds within an expansive forested region, characterized by a complex management history and a diversity of tree species. The source code along with the pretrained models are available at https://github.com/Brent-Murray/point-dl/tree/main/Pytorch/models/PointAugment.

Dynamic decomposition graph convolutional neural network for SSVEP-based brain–computer interface

The SSVEP-based paradigm serves as a prevalent approach in the realm of brain–computer interface (BCI). However, the processing of multi-channel electroencephalogram (EEG) data introduces challenges due to its non-Euclidean characteristic, necessitating methodologies that account for inter-channel topological relations. In this paper, we introduce the Dynamic Decomposition Graph Convolutional Neural Network (DDGCNN) designed for the classification of SSVEP EEG signals. Our approach incorporates layerwise dynamic graphs to address the oversmoothing issue in Graph Convolutional Networks (GCNs), employing a dense connection mechanism to mitigate the gradient vanishing problem. Furthermore, we enhance the traditional linear transformation inherent in GCNs with graph dynamic fusion, thereby elevating feature extraction and adaptive aggregation capabilities. Our experimental results demonstrate the effectiveness of proposed approach in learning and extracting features from EEG topological structure. The results shown that DDGCNN outperforms other state-of-the-art (SOTA) algorithms reported on two datasets (Dataset 1: 54 subjects, 4 targets, 2 sessions; Dataset 2: 35 subjects, 40 targets). Additionally, we showcase the implementation of DDGCNN in the context of synchronized BCI robotic fish control. This work represents a significant advancement in the field of EEG signal processing for SSVEP-based BCIs. Our proposed method processes SSVEP time domain signals directly as an end-to-end system, making it easy to deploy. The code is available at https://github.com/zshubin/DDGCNN.

DM-CNN: Dynamic Multi-scale Convolutional Neural Network with uncertainty quantification for medical image classification

Convolutional neural network (CNN) has promoted the development of diagnosis technology of medical images. However, the performance of CNN is limited by insufficient feature information and inaccurate attention weight. Previous works have improved the accuracy and speed of CNN but ignored the uncertainty of the prediction, that is to say, uncertainty of CNN has not received enough attention. Therefore, it is still a great challenge for extracting effective features and uncertainty quantification of medical deep learning models In order to solve the above problems, this paper proposes a novel convolutional neural network model named DM-CNN, which mainly contains the four proposed sub-modules : dynamic multi-scale feature fusion module (DMFF), hierarchical dynamic uncertainty quantifies attention (HDUQ-Attention) and multi-scale fusion pooling method (MF Pooling) and multi-objective loss (MO loss). DMFF select different convolution kernels according to the feature maps at different levels, extract different-scale feature information, and make the feature information of each layer have stronger representation ability for information fusion HDUQ-Attention includes a tuning block that adjust the attention weight according to the different information of each layer, and a Monte-Carlo (MC) dropout structure for quantifying uncertainty MF Pooling is a pooling method designed for multi-scale models, which can speed up the calculation and prevent overfitting while retaining the main important information Because the number of parameters in the backbone part of DM-CNN is different from other modules, MO loss is proposed, which has a fast optimization speed and good classification effect DM-CNN conducts experiments on publicly available datasets in four areas of medicine (Dermatology, Histopathology, Respirology, Ophthalmology), achieving state-of-the-art classification performance on all datasets. DM-CNN can not only maintain excellent performance, but also solve the problem of quantification of uncertainty, which is a very important task for the medical field. The code is available: https://github.com/QIANXIN22/DM-CNN.

Modelling tree biomass using direct and additive methods with point cloud deep learning in a temperate mixed forest

Airborne laser scanning (ALS) data has been widely used for total aboveground tree biomass (AGB) modelling, however, there is less research focusing on estimating specific tree biomass components (wood, branches, bark, and foliage). Knowledge about these biomass components is essential for carbon accounting, understanding forest nutrient cycling, and other applications. In this study, we compare additive AGB estimation (sum of estimated components) with direct AGB estimation using deep neural network (DNN) and random forest (RF) models. We utilise two point cloud DNNs: point-based Dynamic Graph Convolutional Neural Network (DGCNN) and Octree-based Convolutional Neural Network (OCNN). DNN and RF models were trained using a dataset comprised of 2336 sample plots from a mixed temperate forest in New Brunswick, Canada. Results indicate that additive AGB models perform similarly to direct models in terms of coefficient of determination (R2) and root-mean square error (RMSE), and reduced the mean absolute percentage error (MAPE) by 22% on average. Compared to RF, the DNNs provided a small improvement in performance, with OCNN explaining 5% more variation in the data (R2 = 0.76) and reducing MAPE by 20% on average. Overall, this study showcases the effectiveness of additive tree AGB models and highlights the potential of DNNs for enhanced AGB estimation. To further improve DNN performance, we recommend using larger training datasets, implementing hyperparameter optimization, and incorporating additional data such as multispectral imagery.

Detection of Fittings Based on the Dynamic Graph CNN and U-Net Embedded with Bi-Level Routing Attention

Accurate detection of power fittings is crucial for identifying defects or faults in these components, which is essential for assessing the safety and stability of the power system. However, the accuracy of fittings detection is affected by a complex background, small target sizes, and overlapping fittings in the images. To address these challenges, a fittings detection method based on the dynamic graph convolutional neural network (DGCNN) and U-shaped network (U-Net) is proposed, which combines three-dimensional detection with two-dimensional object detection. Firstly, the bi-level routing attention mechanism is incorporated into the lightweight U-Net network to enhance feature extraction for detecting the fittings boundary. Secondly, pseudo-point cloud data are synthesized by transforming the depth map generated by the Lite-Mono algorithm and its corresponding RGB fittings image. The DGCNN algorithm is then employed to extract obscured fittings features, contributing to the final refinement of the results. This process helps alleviate the issue of occlusions among targets and further enhances the precision of fittings detection. Finally, the proposed method is evaluated using a custom dataset of fittings, and comparative studies are conducted. The experimental results illustrate the promising potential of the proposed approach in enhancing features and extracting information from fittings images.

Novel Feature-Disentangled Autoencoder Integrating Residual Network for Industrial Soft Sensor

In order to overcome the low robustness and weak generalization in existing deep autoencoder (AE) for soft sensor modeling, a novel feature-disentangled autoencoder (FDAE) integrating residual network (Resnet) (FDAE-Resnet) is proposed in this paper. Different from the traditional deep AE that only can learn entangled features, the FDAE can obtain disentangled multi-source features including trend features, periodic features and spatial features by a new trend-periodic long short-term memory (TPLSTM) and a novel dynamic self-attention convolutional neural network (DSACNN). The trend and periodic signals decomposed from input variables are fed into the TPLSTM to learn trend and periodic features in time and frequency domain, respectively. Then, the DSACNN is utilized to capture dynamic spatial features in spatial domain by adding a new attention mechanism. Moreover, disentangled multi-source features are obtained by concatenating trend features, periodic features and spatial features together. Finally, the Resnet is utilized to build the soft sensor model by establishing the relationship between disentangled multi-sources features and outputs. To illustrate the effectiveness and superiority of the proposed method, the FDAE-Resnet is applied in the actual polypropylene process industry for melt index modeling. The experiment results show that compared with other state-of-the-art methods, the FDAE-Resnet can reduce the root mean square error by 26.2% and the mean absolute percentage error by 38.2% on average in the changed working conditions, respectively.

A deep learning model for efficient end-to-end stratification of thrombotic risk in left atrial appendage

Clot formation in the left atrial appendage (LAA) poses a high risk of ischemic strokes and systemic embolism to patients with atrial fibrillation (AF), the most common type of sustained heart arrhythmia that affects more than 35 million people worldwide. Hemodynamic metrics evaluated using computational fluid dynamics (CFD) have been employed to assess the risk of thrombosis in LAA, but its utilization in clinical settings is limited due to their cumbersome operations and high computational cost. To this end, we propose UDGCNN (U: U-net, DGCNN: Dynamic Graph Convolutional Neural Network) that can utilize data from the point cloud of patient-specific LAA geometries as inputs and predict multiple hemodynamic indexes for systematic assessment of the thrombotic risk. The novelty of the proposed model lies in introducing edge convolution layers to the PointNet structure to improve the model’s capacity to learn local features, which is essential for training and predicting complex patient-specific geometries. Moreover, the network structure of UDGCNN is optimized by integrating the Encoder-Decoder structure and skip-connection to extract hierarchical features from the point cloud. The accuracy and efficiency of the UDGCNN is examined by training and testing on 371 LAA geometries from patients with AF, the largest patient-specific dataset in the literature. Using mean absolute error and model inference time as metrics, we demonstrate that UDGCNN can provide an assessment of multiple hemodynamic metrics in 3D patient-specific geometries with prediction error ∼30% lower than those of the state-of-art PointNet model, whereas the inference time is 500-fold shorter compared to computational time CFD simulation. It is noted that UDGCNN is a general computational framework that can be extended to study various cardiovascular diseases. In summary, this study presents a new computational tool that enables end-to-end stratification of thrombotic risk in LAA based on bioimaging, thereby advancing the current screening approach in clinical practice.