Multi-model Fusion Method Research Articles

Fine-grained photometric classification using multi-model fusion method with redshift estimation

The modern sky surveys accelerates astronomical data collection. We proposed a multi-model fusion method aimed at comprehensive and fine-grained astronomical source classification. This method incorporates a redshift estimation model using the mixture density network into a source classification model. Based on 1.2 million sources from the SDSS and the ALLWISE, we performed three-class experiments for stars, quasars, and galaxies, four-class experiments to further classify galaxies into normal and emission-line galaxies (NGs; ELGs), and seven-class experiments where ELG were refined into active galactic nuclei (AGNs), broad-line galaxies (BLs), star-forming galaxies (SFs), and starburst galaxies (SBs). In all experiments, our proposed method is superior to direct classification. In three- and four-class, we obtains 0.77% and 1.14% improvement in accuracy, demonstrating the effectiveness of adding redshift estimation. Meanwhile, three machine learning algorithms were stacked into one by us to finish fine-grained classification, which achieved an accuracy of 78.5%, with F1 scores of 99.2% for stars, 97% for quasars, 64.3% for NGs, 60.8% for AGNs, 68.3% for BLs, 87.2% for SBs, and 71.3% for SFs. The NMAD and R2 for the redshift estimation part of our method are 0.18 and 0.916, while it has only 2.65% outliers. The method we proposed further mines the information contained in the photometry to achieve comprehensive and fine-grained classification, which will be beneficial for immediate analysis in large-scale surveys. Besides, this method can leverage feature importance to stimulate new insights for astronomers.

A modeling method for the control circuit of a double-fed induction generator in a wind power system based on multimodel fusion

Abstract The current modeling techniques of a Double-fed induction generator (DFIG) for wind power systems have shortcomings in the idealization of equipment internal parameters and unitization of system coupling. In this paper, a multi-model fusion method of doubly-fed induction wind turbine modeling for wind power systems is proposed. Firstly, the 1.5 MW DFIG body is constructed according to the parameters of the physical prototype. Then, to realize the system’s control of DFIG, the DIFG external circuit and control circuit of the wind power system are built to realize multi-system coupling simulation modeling. Through simulation analysis, the effectiveness of the multi-model coupling simulation system and the availability of the method are proved.

State of Health Estimation for Lithium-Ion Batteries Under Arbitrary Usage Using Data-Driven Multimodel Fusion

Accurately estimating the state of health (SoH) of batteries is indispensable for the safety, reliability, and optimal energy and power management of electric vehicles. However, from a data-driven perspective, complications, such as dynamic vehicle operating conditions, stochastic user behaviors, and cell-to-cell variations, make the estimation task challenging. This work develops a data-driven multi-model fusion method for SoH estimation under arbitrary usage profiles. All possible operating conditions are categorized into six scenarios. For each scenario, an appropriate feature set is extracted to indicate the SoH. Based on the obtained features, four machine learning algorithms are applied individually to train SoH estimation models using time-series data. In addition to the estimates at the current time step, a histogram data-based and online adaptive model is taken from previous work for predicting the next-step SoH. Then, a Kalman filter is applied to systematically fuse the results of all the estimation and prediction models. Experimental data collected from different types of batteries operated under diverse profiles verify the effectiveness and practicability of the developed method, as well as its superiority over individual models.

Mapping the Planting Area of Winter Wheat at 10-m Resolution Using Sentinel-2 Data and Multimodel Fusion Method

Mapping the Planting Area of Winter Wheat at 10-m Resolution Using Sentinel-2 Data and Multimodel Fusion Method

Automatic Chinese Food recognition based on a stacking fusion model.

With commercialization of deep learning models, daily precision dietary record based on images from smartphones becomes possible. This study took advantage of Deep-learning techniques on visual recognition tasks and proposed a big-data-driven Deep-learning model regressing from food images. We established the largest data set of Chinese dishes to date, named CNFOOD-241. It contained more than 190,000 images with 241 categories, covering Staple food, meat, vegetarian diet, mixed meat and vegetables, soups, dessert category. This study also compares the prediction results of three popular deep learning models on this dataset, ResNeXt101_32x32d achieving up to 82.05% for top-1 accuracy and 97.13% for top-5 accuracy. Besides, this paper uses a multi-model fusion method based on stacking in the field of food recognition for the first time. We built a meta-learner after the base model to integrate the three base models of different architectures to improve robustness. The accuracy achieves 82.88% for top-1 accuracy.Clinical Relevance-This study proves that the application of artificial intelligence technology in the identification of Chinese dishes is feasible, which can play a positive role in people who need to control their diet, such as diabetes and obesity.

A feature reuse based multi-model fusion method for state of health estimation of lithium-ion batteries

Precise estimation of lithium-ion batteries’s state of health (SOH) is crucial for their safe and stable operation. The existing SOH estimation methods mainly focus on a single model with multi-features. Hence, the potential of each feature cannot be fully exploited. To improve the accuracy of prediction model, a feature reuse(FR) based multi-model fusion method is proposed in this paper. Firstly, four features are extracted by analyzing the battery aging process from multiple perspectives such as sample entropy and time(micro/macro), particularly a new feature is proposed to reflect changes in battery energy. The four health factors are as first-level input and second-level feature reuse. Secondly, the preliminary SOH predictions are generated respectively by using k-nearest neighbor regression, random forest regression, linear regression, and extremely randomized tree regression models with different feature inputs. Finally, the initial SOH predictions were fused with the input features from the first stage using a Bayesian linear regression algorithm. Inspired by the advantages of feature reuse method and multi-model fusion, the feature reuse method is applied for fusing Bayesian regression multi-model. To verify the effectiveness of the proposed model, comparative experiments are carried on the Oxford and CALCE battery degradation dataset. Comparing with the single model estimation and methods without feature reuse, the proposed method has better accuracy and stronger robustness in SOH estimation.

Active Learning Integrated Portable Skin Lesion Detection System Based on Multimodel Fusion

In skin lesion detection systems, a large number of labeled images are usually required to achieve high segmentation accuracy, which hinders the effectiveness and timeliness of disease diagnosis. To this end, a high-precision skin lesion detection system is proposed in this paper. The hardware part of the system adopts a modular design, fully considering ergonomics, portability, and miniaturization. In the software part, an active learning ensemble with a multi-model fusion method (ALEM) is proposed to achieve efficient and accurate skin lesion region segmentation. The core idea of ALEM is to use multiple uncertainty strategies of active learning to obtain the most uncertain pixels to be marked in the skin lesion image when marking image pixels. The experiment shows that the average Dice coefficient and average Jaccard index of ALEM on ISIC-2016 are 82.81% and 92.4% respectively and that on ISIC-2017 are 87.51% and 79.26% respectively. It is worth noting that ALEM still outperforms in tests with only 80% of the training data and no more than 15% pixel annotation per image on average. Our system achieves an average AUC of 91.01% on the ISIC2017 and is tested for effectiveness on real skin. The skin lesion detection system developed in this paper is expected to bring convenience to doctors and patients and speed up the diagnosis of diseases.

An Effective Multimodel Fusion Method for SAR and Optical Remote Sensing Images

An Effective Multimodel Fusion Method for SAR and Optical Remote Sensing Images

Practical battery State of Health estimation using data-driven multi-model fusion

Due to dynamic vehicle operating conditions, random user behaviors, and cell-to-cell variations, accurately estimating the battery state of health (SoH) is challenging. This paper proposes a data-driven multi-model fusion method for battery capacity estimation under arbitrary usage profiles. Six feasible and mutually excluded scenarios are meticulously categorized to cover all operating conditions. Four machine learning (ML) algorithms are individually trained using time-series data to estimate the current time step battery capacity. Additionally, a prediction model based on the histogram data is adopted from previous work to predict the next step capacity value. Then, a Kalman filter (KF) is applied to fuse all the estimation and prediction results systematically. The developed method has been demonstrated on cells operated under diverse profiles to verify its effectiveness and practicability.

Analysis of Interaction Grouping Modeling Fusion Group Behavior Recognition Algorithm

In order to make full use of the effective information in the video, this paper proposes a multi-model interactive video behavior recognition method. In order to solve the problems of incomplete human target detection and redundant feature extraction, YOLO_V4 is used to detect the human body and remove the redundant background information. Then, it is proposed to introduce the channel attention model SE-NET into the Inception_V3 network, so as to strengthen the extraction of key features and make the network pay more attention to the details of key features. Finally, the feature information is sent to LSTM network with memory function for action recognition and classification. The multi-model mutual fusion algorithm proposed in this paper is tested and verified on an internationally published UT-Interaction data set. The experimental results show that the accuracy of interactive behavior recognition is improved, and the improved accuracy is 85.1%, which indicates that the multi-model fusion method has higher accuracy.

Contactless blood oxygen estimation from face videos: A multi-model fusion method based on deep learning

Blood Oxygen (SpO2), a key indicator of respiratory function, has received increasing attention during the COVID-19 pandemic. Clinical results show that patients with COVID-19 likely have distinct lower SpO2 before the onset of significant symptoms. Aiming at the shortcomings of current methods for monitoring SpO2 by face videos, this paper proposes a novel multi-model fusion method based on deep learning for SpO2 estimation. The method includes the feature extraction network named Residuals and Coordinate Attention (RCA) and the multi-model fusion SpO2 estimation module. The RCA network uses the residual block cascade and coordinate attention mechanism to focus on the correlation between feature channels and the location information of feature space. The multi-model fusion module includes the Color Channel Model (CCM) and the Network-Based Model(NBM). To fully use the color feature information in face videos, an image generator is constructed in the CCM to calculate SpO2 by reconstructing the red and blue channel signals. Besides, to reduce the disturbance of other physiological signals, a novel two-part loss function is designed in the NBM. Given the complementarity of the features and models that CCM and NBM focus on, a Multi-Model Fusion Model(MMFM) is constructed. The experimental results on the PURE and VIPL-HR datasets show that three models meet the clinical requirement(the mean absolute error ⩽ 2%) and demonstrate that the multi-model fusion can fully exploit the SpO2 features of face videos and improve the SpO2 estimation performance. Our research achievements will facilitate applications in remote medicine and home health.

A Multimodel Fusion Method for Cardiovascular Disease Detection Using ECG.

Objective. Electrocardiogram (ECG) is an important diagnostic tool that has been the subject of much research in recent years. Owing to a lack of well-labeled ECG record databases, most of this work has focused on heartbeat arrhythmia detection based on ECG signal quality. Approach. A record quality filter was designed to judge ECG signal quality, and a random forest method, a multilayer perceptron, and a residual neural network (RESNET)-based convolutional neural network were implemented to provide baselines for ECG record classification according to three different principles. A new multimodel method was constructed by fusing the random forest and RESNET approaches. Main Results. Owing to its ability to combine discriminative human-crafted features with RESNET deep features, the proposed new method showed over 88% classification accuracy and yielded the best results in comparison with alternative methods. Significance. A new multimodel fusion method was presented for abnormal cardiovascular detection based on ECG data. The experimental results show that separable convolution and multiscale convolution are vital for ECG record classification and are effective for use with one-dimensional ECG sequences.

A multi-feature-based multi-model fusion method for state of health estimation of lithium-ion batteries

Battery state-of-health (SOH) estimation is a critical concern of the battery management system, which significantly affects the safe and stable operation of electric vehicles. The existing SOH estimation methods mainly focus on a single model with single-source features. Hence, the generalizability of these methods is limited. In this paper, a multi-feature-based multi-model fusion method is proposed for the SOH estimation of lithium-ion batteries. Firstly, the key factors of the battery aging process are analyzed from multiple sources such as voltage, temperature, and incremental capacity curves. Seven health factors are extracted as first-level input. Secondly, the preliminary SOH predictions are generated by using multiple linear regression, support vector regression, and Gaussian process regression models, respectively. Finally, a random forest model is used to fuse the preliminary SOH predictions. To improve the model prediction accuracy, the proposed model extracts features from different sources to fully describe the battery aging process. Inspired by the advantages of multi-model fusion, the random forest regressor method is applied for fusing the multi-model. To verify the effectiveness of the proposed model, comparative experiments are carried on the Oxford battery degradation dataset. Comparing with single feature or single model estimation methods, the results demonstrate the proposed method has better accuracy and stronger robustness in SOH estimation. • A novel SOH estimation method is trained with a two-layer structure. • Seven health factors of batteries are extracted from multiple sources. • The preliminary SOH predictions are generated by using MLR, SVR, and GPR. • A random forest model is used to fuse the preliminary SOH predictions.

Multi-model fusion short-term power load forecasting based on improved WOA optimization.

The high accuracy of short-term power load forecasting has a pivotal role in helping power companies to construct reasonable production scheduling plans and avoid resource waste. In this paper, a multi-model short-term power load prediction method based on Variational mode decomposition (VMD)-improved whale optimization algorithm (IWOA)-wavelet temporal convolutional network (WTCN)-bidirectional gated recurrent unit (BiGRU)-attention and CatBoost model fusion is proposed. First, VMD was employed to decompose the load data into different intrinsic mode functions. Second, a WTCN was utilized to extract the load data features, and multi-dimensional feature factors were integrated into the BiGRU network for model training. Moreover, an attention mechanism was added to enhance the influence degree of important information. The WTCN-BiGRU-attention model is improved by the WOA algorithm to optimize the hyperparameters of the network. Finally, the model was fused with the predicted data of the CatBoost network by the mean absolute percentage error-reciprocal weight (MAPE-RW) algorithm to construct the best fusion model. Compared with other forecasting models, the proposed multi-model fusion method has higher accuracy in short-term power load forecasting using the public data set for an Australian region.

Multi-step short-term wind speed prediction based on integrated multi-model fusion

Wind power generation is uncontrollable and non-adjustable green energy. The accurate prediction of wind speed is of great significance to the operation and maintenance of wind farms and the safety of the power grid. A novel wind speed prediction approach based on mode decomposition combined with integrated multi-model fusion is proposed. The purpose of the integrated multi-model fusion method is to improve the accuracy of wind speed prediction. Two decomposition algorithms, empirical mode decomposition and local mean decomposition are used to decompose the wind speed into several components respectively. After decomposition, the odd and even sequences are extracted to form two new sequences. According to the characteristics of the components obtained from the decomposition of odd and even sequences, stochastic configuration network and support vector machine are selected as the prediction models. Meanwhile, the particle swarm optimization algorithm is introduced to optimize the parameters of support vector machine. After all decomposition components of the odd and even sequence are predicted by stochastic configuration networks and support vector machine, the predicted values are superimposed to obtain the final predicted value of the odd and even sequence. Finally, the predicted values of odd and even series are integrated to get the final predicted value of wind speed. By comparing the actual and predicted values of short-term wind speed, RMSE, R2, MAE, MAPE, Pearson’s test, prediction error distribution box-plot, and Taylor diagram are used to judge the performance indicators of the prediction model. The case study results show that, compared with the state-of-the-art prediction models, RMSE decreases by about 0.1, MAPE decreases by about 0.05%, MAE decreases by about 0.2, and R2 increases by about 0.2. The other performance indicators of the proposed approach are also better than other models, which prove that the integrated multi-model fusion model can improve the accuracy and effectiveness of wind speed prediction. The proposed prediction approach has high prediction accuracy, can correctly reflect the short-term wind speed law, and has a good application prospect.

MMFGRN: a multi-source multi-model fusion method for gene regulatory network reconstruction.

Lots of biological processes are controlled by gene regulatory networks (GRNs), such as growth and differentiation of cells, occurrence and development of the diseases. Therefore, it is important to persistently concentrate on the research of GRN. The determination of the gene-gene relationships from gene expression data is a complex issue. Since it is difficult to efficiently obtain the regularity behind the gene-gene relationship by only relying on biochemical experimental methods, thus various computational methods have been used to construct GRNs, and some achievements have been made. In this paper, we propose a novel method MMFGRN (for "Multi-source Multi-model Fusion for Gene Regulatory Network reconstruction") to reconstruct the GRN. In order to make full use of the limited datasets and explore the potential regulatory relationships contained in different data types, we construct the MMFGRN model from three perspectives: single time series data model, single steady-data model and time series and steady-data joint model. And, we utilize the weighted fusion strategy to get the final global regulatory link ranking. Finally, MMFGRN model yields the best performance on the DREAM4 InSilico_Size10 data, outperforming other popular inference algorithms, with an overall area under receiver operating characteristic score of 0.909 and area under precision-recall (AUPR) curves score of 0.770 on the 10-gene network. Additionally, as the network scale increases, our method also has certain advantages with an overall AUPR score of 0.335 on the DREAM4 InSilico_Size100 data. These results demonstrate the good robustness of MMFGRN on different scales of networks. At the same time, the integration strategy proposed in this paper provides a new idea for the reconstruction of the biological network model without prior knowledge, which can help researchers to decipher the elusive mechanism of life.

An effective multi-model fusion method for EEG-based sleep stage classification

Stage 1 (S1) and REM sleep are the two key stages in EEG-based sleep stage classification, which are of great significance to the study of neurocognitive ability and sleep diseases. Recently, various methods have been widely studied, and achieved good classification performance, however, most existing studies have a common problem of the low detection rate of S1 and REM sleep. In this paper, we focus on improving the detection performance of S1 and REM sleep and present an effective multi-model fusion method by using hybrid-channel EEG signals, which consists of two parts: the detection of merged stage of S1 and REM sleep and the classification between these two stages. First, we detect S1 and REM sleep by distinguishing the merged stage from other sleep stages using C-SVM model and single-channel EEG signals. To overcome the influence caused by class imbalance, a one-class OC-SVM model of the merged stage is established to correct S1 and REM sleep from the misclassified negative samples. Then, through analyzing the EEG characteristic between S1 and REM sleep and extracting the classification features of multiple sub-bands, we classify S1 and REM sleep using two-channel EEG signals. Finally, the proposed method is tested and analyzed for commonly used dataset of Sleep-EDFX. The results show that this method can effectively detect S1 and REM sleep and promote the application of sleep quality evaluation, fatigue detection, sleep disease diagnosis.

Testing the Feasibility of a Multi-Model Fusion Method for Monitoring the Action of Rehabilitating Stroke Patients in Care Management

Post-stroke care encounters challenges, including high cost, lack of professionals, and insufficient rehabilitation state evaluation. Computer technology can alleviate these issues, as it allows health care professionals (HCP) to quantify the workload and thus enhance rehabilitation care quality. In this paper, a novel multi-model fusion method, terms as pose dual-stream network (PDSN), is devised, aiming to test the feasibility of monitoring the training actions of rehabilitating stroke patients in care management. In particular, this deep-learning-based algorithm combines human pose estimation and dual-stream networks in an innovative way. We utilize an improved OpenPose to estimate human pose from videos obtained by the low-cost monocular camera. In dual-stream networks, the spatial and motion streams are flexibly integrated. The spatial stream network combines the Gated Recurrent Unit (GRU) and attention mechanism to extract spatiotemporal data, while the motion stream network is composed of improved multi-layer 1D Convolutional Neural Networks (CNN), which enhanced by causal and dilated convolution skillfully. Additionally, an adaptive weight fusion strategy is used to fuse the two networks for the final action classification. Results show high accuracy on two public datasets and a dataset created by us, which validate the superiority and feasibility of our method.

Joint dimensionality reduction and metric learning for image set classification

Compared with the traditional classification task based on a single image, an image set contains more complementary information, which is of great benefit to correctly classify a query subject. Thus, image set classification has attracted much attention from researchers. However, the main challenge is how to effectively represent an image set to fully exploit the latent discriminative feature. Unlike in previous works where an image set was represented by a single or a hybrid mode, in this paper, we propose a novel multi-model fusion method across the Euclidean space to the Riemannian manifold to jointly accomplish dimensionality reduction and metric learning. To achieve the goal of our framework, we first introduce three distance metric learning models, namely, Euclidean-Euclidean, Riemannian-Riemannian and Euclidean-Riemannian to better exploit the complementary information of an image set. Then, we aim to simultaneously learn two mappings performing dimensionality reduction and a metric matrix by integrating the two heterogeneous spaces (i.e., the Euclidean space and the Riemannian manifold space) into the common induced Mahalanobis space in which the within-class data sets are close and the between-class data sets are separated. This strategy can effectively handle the severe drawback of not considering the distance metric learning when performing dimensionality reduction in the existing set based methods. Furthermore, to learn a complete Mahalanobis metric, we adopt the L2,1 regularized metric matrix for optimal feature selection and classification. The results of extensive experiments on face recognition, object classification, gesture recognition and handwritten classification demonstrated well the effectiveness of the proposed method compared with other image set based algorithms.

Multimodel fusion method via sparse representation at pixel-level and feature-level

In current fusion methods based on sparse representation (SR) and different frequency, the SR is usually applied to fusion of the low-frequency components. In contrast, the direct fusion is usually adopted for high-frequency components due to their significant diversity. However, the effect of the direct fusion is degraded by the redundant information resulting from the correlation between original signals. A multimodel fusion framework is proposed by applying the SR to low-frequency fusion at pixel-level and high-frequency fusion at feature-level, respectively. First, the multimodal images are decomposed into high-frequency and low-frequency components by nonsubsampled contourlet transform (NSCT). Second, the universal high-frequency dictionary is constructed by using the fast independent component analysis (ICA) of the source high-frequency and its subband images. They represent the general feature part and unique feature part for the high-frequency signals, respectively. The universal low-frequency dictionary is constructed by using the original low-frequency signals. Third, the direct fusion of the high-frequency is converted into sparse coefficients fusion in fast ICA domain. Moreover, the multiple directive contrasts by modifying sum-modified Laplacian are taken as the fusion rule. The low-frequency signals are fused by using an activity measure based on weights. Finally, the fused image is obtained by inverse NSCT on the merged components. The experiments are conducted on three types of image pairs, and the results demonstrate that the proposed method outperforms seven state-of-the-art methods, in terms of four subjective and objective evaluations.