State-of The-art Approaches Research Articles

DetSegDiff: A joint periodontal landmark detection and segmentation in intraoral ultrasound using edge-enhanced diffusion-based network

Individuals with malocclusion require an orthodontic diagnosis and treatment plan based on the severity of their condition. Assessing and monitoring changes in periodontal structures before, during, and after orthodontic procedures is crucial, and intraoral ultrasound (US) imaging has been shown a promising diagnostic tool in imaging periodontium. However, accurately delineating and analyzing periodontal structures in US videos is a challenging task for clinicians, as it is time-consuming and subject to interpretation errors. This paper introduces DetSegDiff, an edge-enhanced diffusion-based network developed to simultaneously detect the cementoenamel junction (CEJ) and segment alveolar bone structure in intraoral US videos. An edge feature encoder is designed to enhance edge and texture information for precise delineation of periodontal structures. Additionally, we employed the spatial squeeze-attention module (SSAM) to extract more representative features to perform both detection and segmentation tasks at global and local levels. This study used 169 videos from 17 orthodontic patients for training purposes and was subsequently tested on 41 videos from 4 additional patients. The proposed method achieved a mean distance difference of 0.17 ± 0.19 mm for the CEJ and an average Dice score of 90.1% for alveolar bone structure. As there is a lack of multi-task benchmark networks, thorough experiments were undertaken to assess and benchmark the proposed method against state-of-the-art (SOTA) detection and segmentation individual networks. The experimental results demonstrated that DetSegDiff outperformed SOTA approaches, confirming the feasibility of using automated diagnostic systems for orthodontists.

Goal-conflict identification based on local search and fast boundary-condition verification based on incremental satisfiability filter

Identifying boundary conditions (BCs) is of fundamental importance for goal-conflict analysis. BCs are able to capture particular combinations of circumstances that make some special conflicts, namely goal divergences, in which the goals of the requirement cannot be satisfied as a whole. Unfortunately, existing approaches only handle small-scale cases within a reasonable running time and do not consider how to search more general BCs because of the search space explosion and the high computational cost of verifying BCs. In this paper, we observed that the BCs identified by the state-of-the-art (SOTA) approach share similarities in structure. It inspires us to develop a BC identification approach, named LOGION, based on local search, to better exploit the similarity in structures. Specifically, we utilize the structural similarity of BCs to construct the neighborhood relation of our local search algorithm. Besides, we design strengthening and weakening operators of formulae to capture the semantics of more general BCs in local searching. In order to boost the BC verification, we propose a lasso-based incremental satisfiability filter (LISF). The primary intuition behind LISF is using trace checking as a light-weight pre-checking, and if it is not successful, to call the satisfiability checking. We conduct extensive experiments to evaluate LOGION, comparing against SOTA approaches. Experimental results show that LOGION produces about 1 order of magnitude more general BCs than the SOTA approaches. It is also shown that LISF significantly boosts the speed, especially for large-scale cases (up to about 1 order of magnitude faster on average).

Capacitive sensing of frost growth dynamics on aluminum surfaces with different wettabilities

Frost formation is a critical issue in many industries including air conditioning and refrigeration, aviation, transportation, and cryogenics. Frost buildup on heat exchanger surfaces reduces their efficiency by acting as a thermal resistance and by reducing or blocking the airflow. Defrosting is therefore necessary to maintain system efficiency. However, defrost initiation and cycle duration are critical parameters that govern system performance. Excessively early or late defrosting can result in energy loss. Frost detection is thus essential to carry out defrosting efficiently. To know when to initiate defrost, both the density and thickness of the frost must be known as both affect the frost thermal resistance. However, state of the art approaches are not sensitive to both frost density and thickness. Here, a capacitive sensing method is proposed that is sensitive to both the density and thickness of the frost. The method can achieve precise frost detection, offering insights into the ideal time to initiate defrosting by being sensitive to factors that affect frost specific thermal resistance. Experiments are carried out on plain (hydrophilic), superhydrophobic, and superhydrophilic aluminum surfaces at three different surface temperatures (−6 °C, −10 °C, −16 °C) to investigate the sensing performance. The capacitance measured increases with time as frost forms and grows. This change in capacitance, along with mass and thickness data, was used as a tool to characterize frost growth. The results indicate that decreasing wettability (superhydrophilic to plain (hydrophilic) to superhydrophobic) results in a decreasing trend in mass accumulation, frost density, and change in capacitance and an increasing trend in frost thicknesses and thermal resistance. It was also demonstrated that increasing surface temperature results in decreasing changes in capacitance (measurement sensitivity), mass accumulation, frost thicknesses, and thermal resistance, with an increasing trend in frost density. A strong correlation between frost mass and change in capacitance was observed regardless of surface temperature or wettability. This rendered implications for the best time of defrosting initiation. The proposed approach offers high accuracy, sensitivity, and convenience as a non-invasive method to study dynamic frost growth with insights into more optimal defrosting cycles, contributing to the development of more efficient energy systems.

A Novel Hierarchical Vision Transformer and Wavelet Time-Frequency Based on Multi-Source Information Fusion for Intelligent Fault Diagnosis.

Deep learning (DL) has been widely used to promote the development of intelligent fault diagnosis, bringing significant performance improvement. However, most of the existing methods cannot capture the temporal information and global features of mechanical equipment to collect sufficient fault information, resulting in performance collapse. Meanwhile, due to the complex and harsh operating environment, it is difficult to extract fault features stably and extensively using single-source fault diagnosis methods. Therefore, a novel hierarchical vision transformer (NHVT) and wavelet time-frequency architecture combined with a multi-source information fusion (MSIF) strategy has been suggested in this paper to boost stable performance by extracting and integrating rich features. The goal is to improve the end-to-end fault diagnosis performance of mechanical components. First, multi-source signals are transformed into two-dimensional time and frequency diagrams. Then, a novel hierarchical vision transformer is introduced to improve the nonlinear representation of feature maps to enrich fault features. Next, multi-source information diagrams are fused into the proposed NHVT to produce more comprehensive presentations. Finally, we employed two different multi-source datasets to verify the superiority of the proposed NHVT. Then, NHVT outperformed the state-of-the-art approach (SOTA) on the multi-source dataset of mechanical components, and the experimental results show that it is able to extract useful features from multi-source information.

E2ETCA: End-to-end training of CNN and attention ensembles for rice disease diagnosis1

Rice is one of the most important crops worldwide. Diseases of the rice plant can drastically reduce crop yield and even lead to complete loss of production. Early diagnosis can reduce the severity and help efforts to establish effective treatment and reduce the usage of pesticides. Traditional machine learning approaches have already been employed for automatic diagnosis. However, they heavily rely on manual preprocessing of images and handcrafted features, which is challenging, time-consuming, and may require domain expertise. Recently, a single end-to-end deep learning (DL)-based approach was employed to diagnose rice diseases. However, it is not highly robust, nor is it generalizable to every dataset. Hence, we propose a novel end-to-end training of convolutional neural network (CNN) and attention (E2ETCA) ensemble framework that fuses the features of two CNN-based state-of-the-art (SOTA) models along with those of an attention-based vision transformer model. These fused features are utilized for diagnosis by the addition of an extra fully connected layer with softmax. The whole procedure is performed end-to-end, which is very important for real-world applications. Additionally, we feed the extracted features into a traditional machine learning approach support vector machine for classification and further analysis. To verify the effectiveness of our proposed E2ETCA framework, we demonstrate it on three publicly available datasets: the Mendeley Rice Leaf Disease Image Samples dataset, the Kaggle Rice Diseases Image dataset, the Bangladesh Rice Research Institute dataset, and a combination of these three datasets. On the basis of various evaluation metrics (accuracy, precision, recall, and F1-score), our proposed E2ETCA framework exhibits superior performance to existing SOTA approaches for rice disease diagnosis, which can also be generalizable in similar other domains.

Learnable convolutional attention network for knowledge graph completion

Recently, graph convolutional networks (GCNs) and graph attention networks (GATs) have been used extensively in knowledge graph completion (KGC), which aims to solve the incompleteness of knowledge graphs (KGs). However, both GCNs and GATs have limitations in the KGC task, and the best method is analyzing the neighbors of each entity (pre-validating), while this process is prohibitively expensive. Furthermore, relations in KGs have specific semantics and should be considered when aggregating neighbor information (message passing). To address the above limitations, we propose a learnable convolutional attention network for knowledge graph completion named LCA-KGC. LCA-KGC introduces a knowledge graph convolutional attention network using a convolution operation before the attention mechanism to ensure structural information acquisition and avoid redundant information stacking. Furthermore, to complete the autonomous switching of GNNs types and eliminate the necessity of pre-validating the local structure of KGs, LCA-KGC designs a learnable knowledge graph convolutional attention network by comprising three types of GNNs in one learnable formulation. Moreover, a learnable message function is proposed to emphasize relational semantics when aggregating neighbor information. Extensive experiments on standard KG datasets validate the effectiveness of the proposed innovations, and LCA-KGC achieves state-of-the-art (SOTA) performance compared to existing approaches (e.g., compared to SOTA approaches, LCA-KGC improves MRR from 0.360 to 0.372 on FB15k-237 dataset, and Hits@3 from 0.561 to 0.581 on YAGO3-10 dataset).

H3E: Learning air combat with a three-level hierarchical framework embedding expert knowledge

Learning to perform air combat autonomously has been a long-standing challenge. The design of intelligent game strategies was difficult due to complex dynamic constraints and long decision-making process. Most existing approaches depend heavily on simplified aircraft models or complex hand-crafted rules. Different from previous works, this paper presents H3E, a novel and efficient Three-level Hierarchical decision framework embedding Expert knowledge, which gives birth to various strategies for high-fidelity one-on-one beyond-visual-range (BVR) air combat game. Inspired by the way pilots make decisions, we build a hierarchical framework to divide the air combat into several sub systems in which each level can perform its own task with much smaller exploration space. In addition, to make full use of the expert knowledge while minimizing its limitation, a novel “Rule-Imitation-Reinforcement” (RIR) training paradigm with an adjustable expert-guidance (AEG) loss function is established, which can increase the exploration efficiency as well as the game win rate. Finally, this work is evaluated in the Intelligent Air Game Simulator (IAGSim), a high-fidelity air combat simulation platform, through a series of games against the state-of-the-art (SOTA) methods. The learning process and game results verify the superior performance of our framework in terms of the exploration efficiency (higher rewards with the same training samples) and win rate (at least 72.9%) compared with the existing SOTA approaches.

Low SLL enhanced gain proximity coupled linear antenna array for 5G wireless communications

This manuscript presents a proximity-fed linear seven-element array antenna with significantly suppressed cross-polarization level for fifth generation (5G) wireless communications. The proposed structure uses a Chebyshev form of power distribution to couple the non-uniform radiating elements. Generally, a conventional series-fed array with identical and similar number of elements provides a minimum sidelobe level (SLL) of 13.5 dB and a gain up to 14.51 dBi (maximum). However, the proposed array provides SLL lower than −21.11 dB and a gain of 15.54 dBi at 28 GHz. Moreover, measured cross-pol discrimination is observed to be greater than 36 dB in elevation plane and in the azimuth plane as well. The proposed antenna array includes seven rectangular patches of varying widths on the upper substrate and is fed through proximity coupling using slit-loaded segmented feed lines on the lower substrate. The developed array antenna exhibits a 5.82 % fractional bandwidth from 26.52 to 28.11 GHz with a radiation efficiency of 91 %. Experimental and simulation results of vital antenna parameters are in close agreement. The overall radiation behavior of the proposed antenna array is found to be superior when compared with the other state of the art approaches, as shown in a comprehensive comparison table.

LDCNet: Lightweight dynamic convolution network for laparoscopic procedures image segmentation

Medical image segmentation is fundamental for modern healthcare systems, especially for reducing the risk of surgery and treatment planning. Transanal total mesorectal excision (TaTME) has emerged as a recent focal point in laparoscopic research, representing a pivotal modality in the therapeutic arsenal for the treatment of colon & rectum cancers. Real-time instance segmentation of surgical imagery during TaTME procedures can serve as an invaluable tool in assisting surgeons, ultimately reducing surgical risks. The dynamic variations in size and shape of anatomical structures within intraoperative images pose a formidable challenge, rendering the precise instance segmentation of TaTME images a task of considerable complexity. Deep learning has exhibited its efficacy in Medical image segmentation. However, existing models have encountered challenges in concurrently achieving a satisfactory level of accuracy while maintaining manageable computational complexity in the context of TaTME data. To address this conundrum, we propose a lightweight dynamic convolution Network (LDCNet) that has the same superior segmentation performance as the state-of-the-art (SOTA) medical image segmentation network while running at the speed of the lightweight convolutional neural network. Experimental results demonstrate the promising performance of LDCNet, which consistently exceeds previous SOTA approaches. Codes are available at github.com/yinyiyang416/LDCNet.

3D Human Pose Estimation from multi-view thermal vision sensors

Human Pose Estimation from images allows the recognition of key daily activity patterns in Smart Environments. Current State-of-the-art (SOTA) 3D pose estimators are built on visible spectrum images, which can lead to privacy concerns in Ambient Assisted Living solutions. Thermal Vision sensors are being deployed in these environments, as they preserve privacy and operate in low brightness conditions. Furthermore, multi-view setups provide the most accurate 3D pose estimation, as the occlusion problem is overcome by having images from different perspectives. Nevertheless, no solutions in the literature use thermal vision sensors following a multi-view scheme.In this work, a multi-view setup consisting of low-cost devices is deployed in the Smart Home of the University of Almería. Thermal and visible images are paired using homography, and SOTA solutions such as YOLOv3 and Blazepose are used to annotate the bounding box and 2D pose in the thermal images. ThermalYOLO is built by fine-tuning YOLOv3 and outperforms YOLOv3 by 5% in bounding box recognition and by 1% in IoU value. Furthermore, InceptionResNetV2 is found as the most appropriate architecture for 2D pose estimation.Finally, a 3D pose estimator was built comparing input approaches and convolutional architectures. Results show that the most appropriate architecture is having three single-channel thermal images processed by independent convolutional backbones (ResNet50 in this case). After these, the output is fused with the 2D poses. The resulting convolutional neural network shows excellent behaviour when having occlusions, outperforming single-view SOTA approaches in the visible spectrum.

Towards energy balancing optimization in wireless sensor networks: A novel quantum inspired genetic algorithm based sinks deployment approach

A Wireless Sensor Network (WSN) with multiple sinks alleviates the bottleneck problem of imbalanced energy consumption and leads to improved scalability and convenience in data collection for diverse range of large-scale applications. However, introducing more than one sink also brings a new set of challenges to overcome. To minimize the uneven energy consumption among sensors for longevity of the WSN, it is required to find optimal locations of the sinks that will permit most of the sensors to forward their data using minimum hops and to determine the optimal route for each sink with consistent route length. Routing in large networks is an NP-Hard problem. Previous schemes have solved this problem using multiple sinks deployment either in planned ways or random manner without simultaneously considering the possibilities of network topological changes, edge cases of sensors, and problem of uneven path length. In this context, this paper proposes a quantum inspired genetic algorithm based multiple sinks deployment approach (Q-GEMS) for energy consumption balancing among sensors to extend the network lifetime. Specially, the exiting approaches have focused on either minimizing energy consumption or reducing the path length. In addition to addressing these constraints, the proposed Q-GEMS is designed to be resilient which adapts to any changes in the planned topologies of deployed sinks in anticipation of sensor failures and external influences. Also to achieve greater energy consumption balance among sensors by equalizing the path length and bring down the cost of WSN by minimizing the number of sinks, novel heuristics have been proposed for positioning the sinks, binding the sensors to sinks, evaluating the quality of solutions and updating the Q-bit population. The experimental results show that the proposed Q-GEMS performs comparatively better than state of the art approaches and achieving network lifetime 5.1%, 21%, and 24% longer than that of EEMS, EESS, and DPSO+PSO respectively.

A Novel Artificial-Intelligence-Based Approach for Classification of Parkinson's Disease Using Complex and Large Vocal Features.

Parkinson's disease (PD) affects a large proportion of elderly people. Symptoms include tremors, slow movement, rigid muscles, and trouble speaking. With the aging of the developed world's population, this number is expected to rise. The early detection of PD and avoiding its severe consequences require a precise and efficient system. Our goal is to create an accurate AI model that can identify PD using human voices. We developed a transformer-based method for detecting PD by retrieving dysphonia measures from a subject's voice recording. It is uncommon to use a neural network (NN)-based solution for tabular vocal characteristics, but it has several advantages over a tree-based approach, including compatibility with continuous learning and the network's potential to be linked with an image/voice encoder for a more accurate multi modal solution, shifting SOTA approach from tree-based to a neural network (NN) is crucial for advancing research in multimodal solutions. Our method outperforms the state of the art (SOTA), namely Gradient-Boosted Decision Trees (GBDTs), by at least 1% AUC, and the precision and recall scores are also improved. We additionally offered an XgBoost-based feature-selection method and a fully connected NN layer technique for including continuous dysphonia measures, in addition to the solution network. We also discussed numerous important discoveries relating to our suggested solution and deep learning (DL) and its application to dysphonia measures, such as how a transformer-based network is more resilient to increased depth compared to a simple MLP network. The performance of the proposed approach and conventional machine learning techniques such as MLP, SVM, and Random Forest (RF) have also been compared. A detailed performance comparison matrix has been added to this article, along with the proposed solution's space and time complexity.

A Closer Look at Debiased Temporal Sentence Grounding in Videos: Dataset, Metric, and Approach

Temporal Sentence Grounding in Videos (TSGV) , which aims to ground a natural language sentence that indicates complex human activities in an untrimmed video, has drawn widespread attention over the past few years. However, recent studies have found that current benchmark datasets may have obvious moment annotation biases, enabling several simple baselines even without training to achieve state-of-the-art (SOTA) performance. In this paper, we take a closer look at existing evaluation protocols for TSGV, and find that both the prevailing dataset splits and evaluation metrics are the devils that lead to untrustworthy benchmarking. Therefore, we propose to re-organize the two widely-used datasets, making the ground-truth moment distributions different in the training and test splits, i.e., out-of-distribution (OOD) test. Meanwhile, we introduce a new evaluation metric “dR@ n ,IoU= m ” that discounts the basic recall scores especially with small IoU thresholds, so as to alleviate the inflating evaluation caused by biased datasets with a large proportion of long ground-truth moments. New benchmarking results indicate that our proposed evaluation protocols can better monitor the research progress in TSGV. Furthermore, we propose a novel causality-based Multi-branch Deconfounding Debiasing (MDD) framework for unbiased moment prediction. Specifically, we design a multi-branch deconfounder to eliminate the effects caused by multiple confounders with causal intervention. In order to help the model better align the semantics between sentence queries and video moments, we enhance the representations during feature encoding. Specifically, for textual information, the query is parsed into several verb-centered phrases to obtain a more fine-grained textual feature. For visual information, the positional information has been decomposed from the moment features to enhance the representations of moments with diverse locations. Extensive experiments demonstrate that our proposed approach can achieve competitive results among existing SOTA approaches and outperform the base model with great gains.

Self-knowledge distillation for surgical phase recognition.

Advances in surgical phase recognition are generally led by training deeper networks. Rather than going further with a more complex solution, we believe that current models can be exploited better. We propose a self-knowledge distillation framework that can be integrated into current state-of-the-art (SOTA) models without requiring any extra complexity to the models or annotations. Knowledge distillation is a framework for network regularization where knowledge is distilled from a teacher network to a student network. In self-knowledge distillation, the student model becomes the teacher such that the network learns from itself. Most phase recognition models follow an encoder-decoder framework. Our framework utilizes self-knowledge distillation in both stages. The teacher model guides the training process of the student model to extract enhanced feature representations from the encoder and build a more robust temporal decoder to tackle the over-segmentation problem. We validate our proposed framework on the public dataset Cholec80. Our framework is embedded on top of four popular SOTA approaches and consistently improves their performance. Specifically, our best GRU model boosts performance by [Formula: see text] accuracy and [Formula: see text] F1-score over the same baseline model. We embed a self-knowledge distillation framework for the first time in the surgical phase recognition training pipeline. Experimental results demonstrate that our simple yet powerful framework can improve performance of existing phase recognition models. Moreover, our extensive experiments show that even with 75% of the training set we still achieve performance on par with the same baseline model trained on the full set.

Emotion recognition at a distance: The robustness of machine learning based on hand-crafted facial features vs deep learning models

Emotion estimation from face expression analysis is nowadays a widely-explored computer vision task. In turn, the classification of expressions relies on relevant facial features and their dynamics. Despite the promising accuracy results achieved in controlled and favorable conditions, the processing of faces acquired at a distance, entailing low-quality images, still suffers from a significant performance decrease. In particular, most approaches and related computational models become extremely unstable in the case of the very small amount of useful pixels that is typical in these conditions. Therefore, their behavior should be investigated more carefully. On the other hand, real-time emotion recognition at a distance may play a critical role in smart video surveillance, especially when controlling particular kinds of events, e.g., political meetings, to try to prevent adverse actions.This work compares facial expression recognition at a distance by: 1) a deep learning architecture based on state-of-the-art (SOTA) proposals, which exploits the whole images to autonomously learn the relevant embeddings; 2) a machine learning approach that relies on hand-crafted features, namely the facial landmarks preliminarily extracted using the popular Mediapipe framework. Instead of using either the complete sequence of frames or only the final still image of the expression, like current SOTA approaches, the two proposed methods are designed to use rich temporal information to identify three different stages of emotion. Expressions are time-split accordingly into four phases to better exploit their temporal-dependent dynamics. Experiments were conducted on the popular Extended Cohn-Kanade dataset (CK+). It was chosen for its wide use in related literature, and because it includes videos of facial expressions and not only still images. The results show that the approach relying on machine learning via hand-crafted features is more suitable for classifying the initial phases of the expression and does not decay in terms of accuracy when images are at a distance (only 0.08% of decay). On the contrary, deep learning not only has difficulties classifying the initial phases of the expressions but also suffers from relevant performance decay when considering images at a distance (52.68% accuracy decay).

3D-Sensitive Encoding of Pharmacophore Features.

In the presence of structural data, one sometimes need to compare 3D ligands. We design an overlay-free method to rank order 3D molecules in the pharmacophore feature space. The proposed encoding includes only two fittable parameters, is sparse, and is not too high dimensional. At the cost of an additional parameter, to delineate the binding site from a protein-ligand complex, the method can compare binding sites. The method was benchmarked on the LIT-PCBA data set for ligand-based virtual screening experiments and the sc-PDB and a Vertex data set when comparing binding sites. In similarity searches, the proposed method outperforms an open-source software doing optimal superposition of ligand-based pharmacophores and RDKit's 3D pharmacophore fingerprints. When comparing binding sites, the method is competitive with state of the art approaches. On a single CPU core, up to 374,000 ligand/s or 132,000 binding site/s can be rank ordered. The "AutoCorrelation of Pharmacophore Features" open-source software is released at https://github.com/tsudalab/ACP4.

Persistent homology of coarse-grained state-space networks.

This work is dedicated to the topological analysis of complex transitional networks for dynamic state detection. Transitional networks are formed from time series data and they leverage graph theory tools to reveal information about the underlying dynamic system. However, traditional tools can fail to summarize the complex topology present in such graphs. In this work, we leverage persistent homology from topological data analysis to study the structure of these networks. We contrast dynamic state detection from time series using a coarse-grained state-space network (CGSSN) and topological data analysis (TDA) to two state of the art approaches: ordinal partition networks (OPNs) combined with TDA and the standard application of persistent homology to the time-delay embedding of the signal. We show that the CGSSN captures rich information about the dynamic state of the underlying dynamical system as evidenced by a significant improvement in dynamic state detection and noise robustness in comparison to OPNs. We also show that because the computational time of CGSSN is not linearly dependent on the signal's length, it is more computationally efficient than applying TDA to the time-delay embedding of the time series.

SleepPPG-Net: A Deep Learning Algorithm for Robust Sleep Staging From Continuous Photoplethysmography.

Sleep staging is an essential component in the diagnosis of sleep disorders and management of sleep health. Sleep is traditionally measured in a clinical setting and requires a labor-intensive labeling process. We hypothesize that it is possible to perform automated robust 4-class sleep staging using the raw photoplethysmography (PPG) time series and modern advances in deep learning (DL). We used two publicly available sleep databases that included raw PPG recordings, totalling 2,374 patients and 23,055 hours of continuous data. We developed SleepPPG-Net, a DL model for 4-class sleep staging from the raw PPG time series. SleepPPG-Net was trained end-to-end and consists of a residual convolutional network for automatic feature extraction and a temporal convolutional network to capture long-range contextual information. We benchmarked the performance of SleepPPG-Net against models based on the best-reported state-of-the-art (SOTA) algorithms. When benchmarked on a held-out test set, SleepPPG-Net obtained a median Cohen's Kappa ( κ) score of 0.75 against 0.69 for the best SOTA approach. SleepPPG-Net showed good generalization performance to an external database, obtaining a κ score of 0.74 after transfer learning. Overall, SleepPPG-Net provides new SOTA performance. In addition, performance is high enough to open the path to the development of wearables that meet the requirements for usage in clinical applications such as the diagnosis and monitoring of obstructive sleep apnea.

Self-adaptive attention fusion for multimodal aspect-based sentiment analysis.

Multimodal aspect term extraction (MATE) and multimodal aspect-oriented sentiment classification (MASC) are two crucial subtasks in multimodal sentiment analysis. The use of pretrained generative models has attracted increasing attention in aspect-based sentiment analysis (ABSA). However, the inherent semantic gap between textual and visual modalities poses a challenge in transferring text-based generative pretraining models to image-text multimodal sentiment analysis tasks. To tackle this issue, this paper proposes a self-adaptive cross-modal attention fusion architecture for joint multimodal aspect-based sentiment analysis (JMABSA), which is a generative model based on an image-text selective fusion mechanism that aims to bridge the semantic gap between text and image representations and adaptively transfer a textual-based pretraining model to the multimodal JMASA task. We conducted extensive experiments on two benchmark datasets, and the experimental results show that our model significantly outperforms other state of the art approaches by a significant margin.

Added value of thermodynamic white-box data for HCCI combustion prediction

Homogeneous charge compression ignition (HCCI) is a promising combustion process to reducing both greenhouse gas and pollutant emissions in the transportation sector. The combustion starts when the thermodynamic state of the cylinder charge reaches the auto-ignition properties of the fuel, which is challenging to model. In addition, the strong cyclic coupling due to required exhaust gas recirculation leads to a significant dependence on the combustion of the previous cycle.This work focuses on identifying accurate models to predict the combustion process for potential use in an MPC. In this work HCCI is achieved using negative valve overlap (NVO) on a fully flexible valve train. The fluctuating composition and temperature of the cylinder charge is not directly measurable and not taken into account in state of the art approaches. It is assumed that it has a decisive influence on the start of combustion. Hence, in this work, the thermodynamic state is modeled by a detailed physical white-box model, which is used as an additional input for the combustion prediction. To investigate the added value of the thermodynamic state data, an Artificial neural network (ANN) is trained to predict combustion both with and without the thermodynamic state data from the white-box model. The root-mean-square error of the ANN including thermodynamic data is reduced by 12.7%. Thus, by adding the thermodynamic white-box data, the HCCI combustion prediction can be significantly improved.