Lower Training Time Research Articles

Physics-informed neural networks for weakly compressible flows using Galerkin–Boltzmann formulation

In this work, we study the Galerkin–Boltzmann formulation within a physics-informed neural network (PINN) framework to solve flow problems in weakly compressible regimes. The Galerkin–Boltzmann equations are discretized with second-order Hermite polynomials in microscopic velocity space, which leads to a first-order conservation law with six equations. Reducing the output dimension makes this equation system particularly well suited for PINNs compared with the widely used D2Q9 lattice Boltzmann velocity space discretizations. We created two distinct neural networks to overcome the scale disparity between the equilibrium and non-equilibrium states in collision terms of the equations. We test the accuracy and performance of the formulation with benchmark problems and solutions for forward and inverse problems with limited data. We compared our approach with the incompressible Navier–Stokes equation and the D2Q9 formulation. We show that the Galerkin–Boltzmann formulation results in similar L2 errors in velocity predictions in a comparable training time with the Navier–Stokes equation and lower training time than the D2Q9 formulation. We also solve forward and inverse problems for a flow over a square, try to capture an accurate boundary layer, and infer the relaxation time parameter using available data from a high-fidelity solver. Our findings show the potential of utilizing the Galerkin–Boltzmann formulation in PINN for weakly compressible flow problems.

A deep learning deviation-based scheme to defend against false data injection attacks in power distribution systems

Defending against false data injection attacks (FDIAs) in cyber-physical power systems is crucial. Detection in power distribution systems is complex due to load variations, uncertainties, and fewer meters. Defense strategies include model-driven and data-driven approaches, but model-based methods can trigger false alarms due to threshold setting issues. The current research proposes a novel data-driven method to address threshold setting issues in detecting and localizing FDIAs in power distribution systems. First, a dataset is created by recording estimated measurement values using an unscented Kalman filter and weighted least squares across various attack scenarios. These estimated measurements are then fed into a deep artificial neural network (ANN) for binary classification to detect attacks. The output, along with the estimated measurements, is used by another ANN to localize the corrupted meter zone. This deep learning-based approach improves threshold setting over the common chi-square method. Results show that the proposed deep learning method for FDIA detection and localization outperforms a recently proposed ensemble of shallow models. The area under the curve value increases by about 5% with lower training time. The approach is also effective against previously unseen attack strategies and different feeder topologies.

An Efficient Visual Tracking System Based on Extreme Learning Machine in the Defence and Military Sector

Visual tracking is the capacity to estimate or forecast a target object’s location in each frame of a video after specifying its starting position. Visual tracking is of essential relevance in defence and military operations. The military can use it to improve situational awareness, improve precise targeting, acquire intelligence in real-time, and efficiently respond to a variety of threats and circumstances. In the past, object tracking systems have relied mostly on algorithms based on deep learning techniques and these tracking algorithms are lacking in both accuracy and speed. In this research, an Extreme Learning Machine-based visual tracking system has been proposed that incorporates properties like high accuracy, low training time, and less network computing complexity as compared to existing deep learning-based tracking algorithms. The Haar wavelet transform is utilized in the recommended technique for feature learning, while the extreme learning machine is utilised for classification and recognition. A benchmark dataset object tracking benchmark-2013 has been used to carry out the experiments. The experiment values indicated that the proposed technique has accomplished enhanced performance over another tracking model. Additionally, we tested the proposed method’s accuracy and robustness regarding certain visual characteristics: Illumination variation, occlusion, deformation, out-of-plane rotation, background clutters, and in-plane rotation. The findings of the simulation revealed that the objects in videos have been 84% accurately tracked by the suggested method.

Improved semantic-guided network for skeleton-based action recognition

A fundamental issue in skeleton-based action recognition is the extraction of useful features from skeleton joints. Unfortunately, the current state-of-the-art models for this task have a tendency to be overly complex and parameterized, which results in low model training and inference time efficiency for large-scale datasets. In this work, we develop a simple but yet an efficient baseline for skeleton-based Human Action Recognition (HAR). The architecture is based on adaptive GCNs (Graph Convolutional Networks) to capture the complex interconnections within skeletal structures automatically without the need of a predefined topology. The GCNs are followed and empowered with an attention mechanism to learn more informative representations. This paper reports interesting accuracy on a large-scale dataset NTU-RGB+D 60, 89.7% and 95.0% on respectively Cross-Subject, and Cross-View benchmarks. On NTU-RGB+D 120, 84.6% and 85.8% over Cross-Subject and Cross-Setup settings, respectively. This work provides an improvement of the existing model SGN (Semantic-Guided Neural Networks) when extracting more discriminant spatial and temporal features.

Geometric deep learning for enhanced quantitative analysis of microstructures in X-ray computed tomography data

Quantitative microstructural analysis of XCT 3D images is key for quality assurance of materials and components. In this paper we implement a Graph Convolutional Neural Network (GCNN) architecture to segment a complex Al-Si Metal Matrix composite XCT volume (3D image). We train the model on a synthetic dataset and we assess its performance on both synthetic and experimental, manually-labeled, datasets. Our simple GCNN shows a comparable performance, measured via the Dice score, to more standard machine learning methods, but uses a greatly reduced number of parameters (less than 1/10 of parameters), features low training time, and needs little hardware resources. Our GCNN thus achieves a cost-effective reliable segmentation.

A geometric algebra-based approach for myoelectric pattern recognition control and faster prosthesis recalibration

ObjectiveAlthough many advancements have been made on myoelectric pattern-recognition, the control of poly-articulated upper-limb prostheses remains insufficiently robust. Electrode-shift, sweat or fatigue degrade the performance of classifiers over time, resulting in unfruitful device usage and frequent re-calibration. To tackle this issue, here we introduce two models − µP6 and µP8 − that combine Geometric Algebra with nearest-neighbor classification. We aim at reducing both the necessary training data and training time and, unlike most current state-of-the-art algorithms, we exploit an alternative geometric representation (and visualization) of the EMG signal as different polygons for different types of gestures, facilitating the explanation of the decision-making process to a layman. Moreover, we explore four abstention strategies to reduce the number of misclassifications. MethodsWe perform an offline analysis on two datasets, alongside two other standard models: nonlinear logistic regression (NLR) and linear discriminant analysis (LDA). ResultsEven with few training data, the proposed algorithms achieve high F1-scores (>0.95), significantly higher or non-significantly different from the values obtained with NLR and LDA, while maintaining relatively low abstentions rates and training times (<2 ms). ConclusionThe proposed algorithms allow to reduce the amount of training data and training times without compromising recognition rates. SignificanceThe proposed algorithms may contribute for a faster prosthesis re-calibration procedure while allowing to re-gain high recognition rates. Furthermore, the decision-making process is explainable and interpretable, potentially improving user trust and acceptance.

Improving the Concrete Crack Detection Process via a Hybrid Visual Transformer Algorithm.

Inspections of concrete bridges across the United States represent a significant commitment of resources, given their biannual mandate for many structures. With a notable number of aging bridges, there is an imperative need to enhance the efficiency of these inspections. This study harnessed the power of computer vision to streamline the inspection process. Our experiment examined the efficacy of a state-of-the-art Visual Transformer (ViT) model combined with distinct image enhancement detector algorithms. We benchmarked against a deep learning Convolutional Neural Network (CNN) model. These models were applied to over 20,000 high-quality images from the Concrete Images for Classification dataset. Traditional crack detection methods often fall short due to their heavy reliance on time and resources. This research pioneers bridge inspection by integrating ViT with diverse image enhancement detectors, significantly improving concrete crack detection accuracy. Notably, a custom-built CNN achieves over 99% accuracy with substantially lower training time than ViT, making it an efficient solution for enhancing safety and resource conservation in infrastructure management. These advancements enhance safety by enabling reliable detection and timely maintenance, but they also align with Industry 4.0 objectives, automating manual inspections, reducing costs, and advancing technological integration in public infrastructure management.

Faulty cell prediction accuracy comparison of machine learning algorithms using temperature sensor placement optimization approach in immersion cooled Li-ion battery modules

Immersion cooling is a promising thermal management system for LiBs where the cells are submerged in a thermally conductive coolant, thus improving heat dissipation and prolonging battery life. The present study investigates an aligned arranged 4S4P immersion-cooled LiB module to develop a best-fit machine learning (ML) model that could predict the fault position in the module depending on the inputs from the temperature sensors that are optimized under different operating and fault conditions. The Pearson Correlation Coefficient (PCC) feature selection approach is used to optimize the sensors, while the data is generated from numerical simulations on a 4S4P immersion-cooled LiB module. The model validation through internal experimental trials is performed on a 2S2P battery module. The four ML classification algorithms, which include the K Nearest Neighbors, Random Forest (RF), Extreme Gradient, and Long Short Term Memory (LSTM), are trained on the optimized sensors data and are further tested internally and externally to assess their performance on unseen data within and outside the training range. A 5-fold cross-validation process is also implemented, and following a comprehensive comparison of the predictions based on the accuracies and model elapsed time, the best-fit model is identified. The results conclude that while the LSTM model slightly outweighs the other models with an accuracy of 98.18% for the specific external test cases, the RF model is chosen as the best-fit model with a prediction accuracy of 99.7% based on the error metrics and low training time for the internal testing. The outcomes of this present work contribute to the early identification of battery module failures, enhancing safety and reducing costs.

Classification feature selection and dimensionality reduction based on logical binary sine-cosine function arithmetic optimization algorithm

Arithmetic optimization algorithm (AOA) is a meta-heuristic algorithm inspired by mathematical operations. AOA has been diffusely used for optimization issues on continuous domains, but few scholars have studied discrete optimization problems. In this paper, we proposed Binary AOA (BAOA) based on two strategies to handle the feature selection problem. The first strategy used S-shaped and V-shaped shift functions to map continuous variables to discrete variables. The second strategy was to combine four logical operations (AND, OR, XOR, XNOR) on the basis of the transfer function, and constructed a parameter model based on the sine and cosine function. An enhanced logic binary sine–cosine function arithmetic optimization algorithm (LBSCAOA) was proposed to realize the position update of variables. Its purpose was to improve the algorithm's global search capabilities and local exploitation capabilities. In the simulation experiments, 20 datasets were selected to testify the capability of the proposed algorithm. Since KNN had the advantages of easy understanding and low training time complexity, this classifier was selected for evaluation. The performance of the improved algorithm was comprehensively evaluated by comparing the average classification accuracy, the average number of selected features, the average fitness value and the average running time. Simulation results showed LBSCAOA with V-Shaped “V4” stood out among many improved algorithms. On the other hand, LBSCAOA with V-Shaped “V4” was used as a representative to compare with other typical feature selection algorithms to verify its competitivenes.

Exploring deep echo state networks for image classification: a multi-reservoir approach

Echo state networks (ESNs) belong to the class of recurrent neural networks and have demonstrated robust performance in time series prediction tasks. In this study, we investigate the capability of different ESN architectures to capture spatial relationships in images without transforming them into temporal sequences. We begin with three pre-existing ESN-based architectures and enhance their design by incorporating multiple output layers, customising them for a classification task. Our investigation involves an examination of the behaviour of these modified networks, coupled with a comprehensive performance comparison against the baseline vanilla ESN architecture. Our experiments on the MNIST data set reveal that a network with multiple independent reservoirs working in parallel outperforms other ESN-based architectures for this task, achieving a classification accuracy of 98.43%. This improvement on the classical ESN architecture is accompanied by reduced training times. While the accuracy of ESN-based architectures lags behind that of convolutional neural network-based architectures, the significantly lower training times of ESNs with multiple reservoirs operating in parallel make them a compelling choice for learning spatial relationships in scenarios prioritising energy efficiency and rapid training. This multi-reservoir ESN architecture overcomes standard ESN limitations regarding memory requirements and training times for large networks, providing more accurate predictions than other ESN-based models. These findings contribute to a deeper understanding of the potential of ESNs as a tool for image classification.

Multimodal subspace identification for modeling discrete-continuous spiking and field potential population activity

Objective. Learning dynamical latent state models for multimodal spiking and field potential activity can reveal their collective low-dimensional dynamics and enable better decoding of behavior through multimodal fusion. Toward this goal, developing unsupervised learning methods that are computationally efficient is important, especially for real-time learning applications such as brain–machine interfaces (BMIs). However, efficient learning remains elusive for multimodal spike-field data due to their heterogeneous discrete-continuous distributions and different timescales. Approach. Here, we develop a multiscale subspace identification (multiscale SID) algorithm that enables computationally efficient learning for modeling and dimensionality reduction for multimodal discrete-continuous spike-field data. We describe the spike-field activity as combined Poisson and Gaussian observations, for which we derive a new analytical SID method. Importantly, we also introduce a novel constrained optimization approach to learn valid noise statistics, which is critical for multimodal statistical inference of the latent state, neural activity, and behavior. We validate the method using numerical simulations and with spiking and local field potential population activity recorded during a naturalistic reach and grasp behavior. Main results. We find that multiscale SID accurately learned dynamical models of spike-field signals and extracted low-dimensional dynamics from these multimodal signals. Further, it fused multimodal information, thus better identifying the dynamical modes and predicting behavior compared to using a single modality. Finally, compared to existing multiscale expectation-maximization learning for Poisson–Gaussian observations, multiscale SID had a much lower training time while being better in identifying the dynamical modes and having a better or similar accuracy in predicting neural activity and behavior. Significance. Overall, multiscale SID is an accurate learning method that is particularly beneficial when efficient learning is of interest, such as for online adaptive BMIs to track non-stationary dynamics or for reducing offline training time in neuroscience investigations.

Unveiling the Impact of Extreme Learning Machine in the Defence and Military Sector

Among the most well-known machine learning algorithms, Extreme Learning Machine (ELM) has seen widespread use across a variety of fields, including the defence and military industries. For problems like sluggish technique and iteratively altering the hidden layer’s parameters to optimise the efficiency of the gradient descent approach, a cutting-edge machine learning algorithm known as the ELM has been developed. Depending on the specific objective and circumstance, the Extreme Learning Machine (ELM) may be more appropriate than Deep Neural Network (DNN) techniques. The models constructed in this manner perform quite well in generalisation. The following three goals are emphasised in its unconventional structure: 1) A great degree of accuracy in learning 2) less need for direct human involvement 3) an extremely rapid rate of learning, and moreover, it provides an optimal response for the whole world. As a result of its quick training, flexibility, and resilience, the Extreme Learning Machine (ELM) has several uses in this field, including target detection and tracking, image and signal processing, cybersecurity and intrusion detection, decision support systems, pattern recognition and classification, etc. According to our findings, the ELM approach was used with low training time and the testing accuracy is excellent. Also, this study presents the contribution of the revolutionary machine learning algorithm ELM to the defence and military sectors.

Optimizing echo state networks for continuous gesture recognition in mobile devices: A comparative study

Continuous gesture recognition can be used to enhance human-computer interaction. This can be accomplished by capturing human movement with the use of the Inertial Measurement Units in smartphones and using machine learning algorithms to predict the intended gestures. Echo State Networks (ESNs) consist of a fixed internal reservoir that is able to generate rich and diverse nonlinear dynamics in response to input signals that capture temporal dependencies within the signal. This makes ESNs well-suited for time series prediction tasks, such as continuous gesture recognition. However, their application has not been rigorously explored, with regard to gesture recognition. In this study, we sought to enhance the efficacy of ESN models in continuous gesture recognition by exploring diverse model structures, fine-tuning hyperparameters, and experimenting with various training approaches. We used three different training schemes that used the Leave-one-out Cross-validation (LOOCV) protocol to investigate the performance in real-world scenarios with different levels of data availability: Leaving out data from one user to use for testing (F1-score: 0.89), leaving out a fraction of data from all users to use in testing (F1-score: 0.96), and training and testing using LOOCV on a single user (F1-score: 0.99). The obtained results outperformed the Long Short-Term Memory (LSTM) performance from past research (F1-score: 0.87) while maintaining a low training time of approximately 13 seconds compared to 63 seconds for the LSTM model. Additionally, we further explored the performance of the ESN models through behaviour space analysis using memory capacity, Kernel Rank, and Generalization Rank. Our results demonstrate that ESNs can be optimized to achieve high performance on gesture recognition in mobile devices on multiple levels of data availability. These findings highlight the practical ability of ESNs to enhance human-computer interaction.

Cross-domain prediction approach of human lower limb voluntary movement intention for exoskeleton robot based on EEG signals.

Exoskeleton robot control should ideally be based on human voluntary movement intention. The readiness potential (RP) component of the motion-related cortical potential is observed before movement in the electroencephalogram and can be used for intention prediction. However, its single-trial features are weak and highly variable, and existing methods cannot achieve high cross-temporal and cross-subject accuracies in practical online applications. Therefore, this work aimed to combine a deep convolutional neural network (CNN) framework with a transfer learning (TL) strategy to predict the lower limb voluntary movement intention, thereby improving the accuracy while enhancing the model generalization capability; this would also provide sufficient processing time for the response of the exoskeleton robotic system and help realize robot control based on the intention of the human body. The signal characteristics of the RP for lower limb movement were analyzed, and a parameter TL strategy based on CNN was proposed to predict the intention of voluntary lower limb movements. We recruited 10 subjects for offline and online experiments. Multivariate empirical-mode decomposition was used to remove the artifacts, and the moment of onset of voluntary movement was labeled using lower limb electromyography signals during network training. The RP features can be observed from multiple data overlays before the onset of voluntary lower limb movements, and these features have long latency periods. The offline experimental results showed that the average movement intention prediction accuracy was 95.23% ± 1.25% for the right leg and 91.21% ± 1.48% for the left leg, which showed good cross-temporal and cross-subject generalization while greatly reducing the training time. Online movement intention prediction can predict results about 483.9 ± 11.9ms before movement onset with an average accuracy of 82.75%. The proposed method has a higher prediction accuracy with a lower training time, has good generalization performance for cross-temporal and cross-subject aspects, and is well-prioritized in terms of the temporal responses; these features are expected to lay the foundation for further investigations on exoskeleton robot control.

The SSHVEP Paradigm-Based Brain Controlled Method for Grasping Robot Using MVMD Combined CNN Model.

In recent years, the steady-state visual evoked potentials (SSVEP) based brain control method has been employed to help people with disabilities because of its advantages of high information transmission rate and low training time. However, the existing SSVEP brain control methods cannot adapt to dynamic or unstructured environments. Moreover, the recognition accuracy from the conventional decoding algorithm still needs to improve. To address the above problems, this study proposed a steady-state hybrid visual evoked potentials (SSHVEP) paradigm using the grasping targets in their environment to improve the connection between the subjects' and their dynamic environments. Moreover, a novel EEG decoding method, using the multivariate variational mode decomposition (MVMD) algorithm for adaptive sub-band division and convolutional neural network (CNN) for target recognition, was applied to improve the decoding accuracy of the SSHVEPs. 18 subjects participated in the offline and online experiments. The offline accuracy across 18 subjects by the 9-target SSHVEP paradigm was up to 95.41 ± 2.70 %, which is a 5.80% improvement compared to the conventional algorithm. To further validate the performance of the proposed method, the brain-controlled grasping robot system using the SSHVEP paradigm was built. The average accuracy reached 93.21 ± 10.18 % for the online experiment. All the experimental results demonstrated the effectiveness of the brain-computer interaction method based on the SSHVEP paradigm and the MVMD combined CNN algorithm studied in this paper.

A scalable blockchain-enabled federated learning architecture for edge computing.

Various deep learning techniques, including blockchain-based approaches, have been explored to unlock the potential of edge data processing and resultant intelligence. However, existing studies often overlook the resource requirements of blockchain consensus processing in typical Internet of Things (IoT) edge network settings. This paper presents our FLCoin approach. Specifically, we propose a novel committee-based method for consensus processing in which committee members are elected via the FL process. Additionally, we employed a two-layer blockchain architecture for federated learning (FL) processing to facilitate the seamless integration of blockchain and FL techniques. Our analysis reveals that the communication overhead remains stable as the network size increases, ensuring the scalability of our blockchain-based FL system. To assess the performance of the proposed method, experiments were conducted using the MNIST dataset to train a standard five-layer CNN model. Our evaluation demonstrated the efficiency of FLCoin. With an increasing number of nodes participating in the model training, the consensus latency remained below 3 s, resulting in a low total training time. Notably, compared with a blockchain-based FL system utilizing PBFT as the consensus protocol, our approach achieved a 90% improvement in communication overhead and a 35% reduction in training time cost. Our approach ensures an efficient and scalable solution, enabling the integration of blockchain and FL into IoT edge networks. The proposed architecture provides a solid foundation for building intelligent IoT services.

OFIDS : Online Learning-Enabled and Fingerprint-Based Intrusion Detection System in Controller Area Networks

As a widely used industrial field bus, the controller area network (CAN) lacks security mechanisms (e.g., encryption and authentication). With the development of network connectivity and the increase in external communication, CAN networks are vulnerable to security attacks (e.g., masquerade). Attackers can intrude on CAN networks to control CAN devices, causing security threats. A fingerprint-based intrusion detection system (IDS) in CAN networks can detect masquerade attacks by scanning the unique clock signals of CAN devices. However, most state-of-the-art fingerprint-based IDSs commonly use an analog-to-digital converter module with a low frequency of 60 MHz to sample CAN signals, lowering the accuracy of the trained fingerprint-based IDS model. In addition, almost all fingerprint-based IDSs are trained offline and then detected online in CAN networks, ignoring that system clock signals of hardware change over time, resulting in degraded detection performance. An intrusion detection system based on physical fingerprints can achieve effective identification of masquerade attacks. This paper proposes an online learning-enabled and fingerprint-based IDS (OFIDS) in CAN networks to increase the sampling frequency, simplify the sampling circuit, shorten the delay time, improve the detection response time, and increase the detection accuracy. OFIDS uses a high-speed comparator (i.e., TLV3501) and field-programmable gate arrays (FPGA; i.e., Xilinx ZYNQ-7010) to sample the CAN_High signal, achieving a low sampling delay time of 4.5 ns and a high sampling frequency of 1 GHz. The self-adaptability of the backpropagation neural network is taken advantage of and used to train the OFIDS model with a detection accuracy of 99.9992%. The OFIDS model is deployed to a CAN network prototype with five CAN devices (i.e., two Arduino UNO boards and three STM32 microcontrollers), where online learning is conducted with a low training time of less than 2 min to update the OFIDS model. Experimental results show that OFIDS can achieve at least 99.99% detection accuracy within 0.18 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\mu$</tex-math></inline-formula> s in a CAN network prototype and can achieve 98% detection accuracy in a real vehicle.

Can pretrained networks be used in fault diagnosis of monoblock centrifugal pump?

Monoblock centrifugal pumps are used daily and play an essential role in agriculture, wastewater treatment, food, paper, gas and oil industries. Extended and endless use of monoblock centrifugal pump can affect the desired pump characteristics, compromising the efficiency of the pump. The condition of a pump is judged based on the high vibration, unusual noise, leakage, etc. Maintaining pump health is essential to increase the productivity in industry. Countless experiments were carried out by researchers to maintain the efficiency of the pump. The main shortcomings of traditional fault diagnosis are the need for a high degree of human intellect and professional knowledge. Such drawbacks have made the researchers provide autonomous and intelligent diagnostic tools. The objective of this article is to investigate the possibility of use of deep learning technique to identify the type of fault present in the monoblock centrifugal pump. First, vibration signals from the monoblock centrifugal pump were acquired and plotted as images that serve as an input for deep learning (DL) algorithms. The deep learning system can analyse the status of the monoblock centrifugal pump and learn from the vibration signal plots. In this paper, pre-trained networks such as ResNet-50, AlexNet, GoogleNet and VGG-16 were used to diagnose the fault in the monoblock centrifugal pump. The outcome of the hyperparameters namely learning rate (LR), batch size (BS), split ratio (SR) and solver were considered and the best performing network was recommended for the monoblock centrifugal pump fault diagnostics. Findings: A higher classification accuracy of 100% with lower training time of 177 s was obtained while using ResNet-50.

Self-Attention-Based Deep Convolution LSTM Framework for Sensor-Based Badminton Activity Recognition.

Sensor-based human activity recognition aims to classify human activities or behaviors according to the data from wearable or embedded sensors, leading to a new direction in the field of Artificial Intelligence. When the activities become high-level and sophisticated, such as in the multiple technical skills of playing badminton, it is usually a challenging task due to the difficulty of feature extraction from the sensor data. As a kind of end-to-end approach, deep neural networks have the capacity of automatic feature learning and extracting. However, most current studies on sensor-based badminton activity recognition adopt CNN-based architectures, which lack the ability of capturing temporal information and global signal comprehension. To overcome these shortcomings, we propose a deep learning framework which combines the convolutional layers, LSTM structure, and self-attention mechanism together. Specifically, this framework can automatically extract the local features of the sensor signals in time domain, take the LSTM structure for processing the badminton activity data, and focus attention on the information that is essential to the badminton activity recognition task. It is demonstrated by the experimental results on an actual badminton single sensor dataset that our proposed framework has obtained a badminton activity recognition (37 classes) accuracy of 97.83%, which outperforms the existing methods, and also has the advantages of lower training time and faster convergence.

PLAME: Piecewise-Linear Approximate Measure for Additive Kernel SVM

Additive Kernel SVM has been extensively used in many applications, including human activity detection and pedestrian detection. Since training an additive kernel SVM model is very time-consuming, which is not scalable to large-scale datasets, many efficient solutions have been developed in the past few years. However, most of the existing methods normally fail to achieve one of these three important conditions which are (1) low classification error, (2) low memory space, and (3) low training time. In order to simultaneously fulfill these three conditions, we develop the new piecewise-linear approximate measure (PLAME) for additive kernels. By incorporating PLAME with the well-known dual coordinate descent method, we theoretically show that this approach can achieve the above three conditions. Experimental results on twelve real datasets show that our approach can achieve the best trade-off between the accuracy, memory space, and training time compared with different types of state-of-the-art methods.