Original Sample Set Research Articles

Envelope rotation forest: A novel ensemble learning method for classification

Ensemble learning is an important field and research hotpot of machine learning. However, the existing ensemble learning methods construct subsets by directly sampling the same original sample set. Therefore, the subsets suffer from the low quality of original sample set, such as low separability, low diversity, high sensitivity to noisy samples. As a result, the improvement effect of existing ensemble methods is limited. To address these problems, a new ensemble learning method - envelope rotation forest was proposed. Firstly, envelope sample transformation mechanism (ESTM) is designed to transform the original sample set to multiple layers of envelope sample sets, thereby improving the quality of the sample subsets. The ESTM consists of multilayer iterative mean clustering (MIMC) and interlayer consistency mechanism (ICM). Secondly, the two-dimensional base classifier matrix generation and fusion mechanism (2D_BCMG&FM) is designed to enhance classification performance. The final classification results are obtained by fusing the two-dimensional classification labels. For verification, the proposed method is tested on 26 datasets and is compared with representative methods. It is observed that the accuracy is improved by up to 18.77%, which highlights the significance of considering envelope transformation on original samples for subsequent subset construction.Data and code can be found at: https://github.com/GRabase/ERF

A novel semi-supervised learning rolling bearing fault diagnosis method based on SNNGAN

In practical industrial environments, rotating machinery typically operates under normal conditions. As a result, the signals collected are primarily normal signals. This imbalance in the sample data diminishes the effectiveness of fault diagnosis. To address this issue, this paper produces a novel semi-supervised fault diagnosis approach based on a Siamese neural network combined with a generative adversarial network (SNNGAN) to enhance classification accuracy. Firstly, vibration signals collected are subjected to continuous wavelet transformation to obtain time–frequency representations, which are utilized for pre-training convolutional encoders in the generator and discriminator. Subsequently, a cosine similarity algorithm is employed to ensure the quality of generated samples. For generated data, set a similarity threshold. Those surpassing the threshold are assigned their corresponding labels and added to the original sample set. Otherwise, those falling below the threshold are transformed back into vibration vectors through an inverse transform and then serve as input to create new samples. Finally, fault diagnosis experiments are conducted on the newly balanced data set. In four imbalanced data experiments, the results demonstrate that SNNGAN outperforms other methods in average accuracy, G-mean, and F1 score, with average accuracy values of 0.919, 0.948, 0.927, and 0.953 for the respective datasets. Therefore, SNNGAN exhibits outstanding fault diagnosis performance under conditions of data imbalance.

Learning from feature and label spaces’ bias for uncertainty-adaptive facial emotion recognition

Developing an accurate deep model for facial emotion recognition is a long-term challenge. It is because the uncertainty of emotions, stemming from the ambiguity of different emotional categories and the difference of subjective annotations, can ruin the ability of model to achieve the desired optimization. This paper constructs two distinct datasets, namely original sample set and ambiguous sample set, to explore an effective ambiguous knowledge transfer method to realize the adaptive awareness of uncertainty in facial emotion recognition. The original sample set is the weakly-augmented data with relatively low uncertainty, as most emotions are clean in reality. Meanwhile, the ambiguous sample set is strongly-augmented data that introduces feature and label bias with regard to emotion, which are with relatively high uncertainty. The proposed framework consists of two sub-nets, which are trained using the original set and the ambiguous set respectively. To achieve uncertainty-adaptive learning for two sub-nets, we introduce two modules. One is the cross-space attention consistency learning module that performs attention coupling across original and ambiguous feature spaces, achieving uncertainty-aware representation learning in feature granularity. The other is the soft-label learning module that models and utilizes uncertainty in label granularity, through aligning the posterior distributions between original label space and ambiguous label space. Experimental studies on public datasets indicate that our method is competitive with the state-of-the-art.

Research on the method of diesel particulate filters carbon load recognition based on deep learning

Because the carbon load inside a diesel particulate filters (DPF) affects the DPF regeneration, and the carbon load recognition is significant for the particulate matter (PM) emission control. It is necessary to investigate an on-board DPF carbon load recognition method because the carbon load cannot be directly measured by sensors. Aiming to build a DPF carbon load prediction model adopting the deep learning method, this paper proposes a DPF carbon load identification model based on different experimental parameters using a layered one dimension convolutional neural network (1D-CNN) method. To improve data validity, this paper adopts two data-processing methods. The data pre-processing adopts data splicing method to complete the construction of the original sample set, and the data after-processing uses wavelet packet transform method to establish the feature sample sets. The model adopts the optimal feature dataset constructed by three input parameters, i.e., temperature difference, pressure difference, and exhaust mass flow, and has both high training accuracy and test accuracy above 90 %. The pressure difference is the most important influencing input parameter, and the three-parameter sample set (ΔT + ΔP + Q) has great recognition accuracy and good model stability with the high training accuracy and test accuracy as well as less iteration.

Generation of high-dimensional vibration signal and its application in fault diagnosis

Imperfect data, such as data scarcity and imbalance, have a negative impact on intelligent fault diagnosis. Generative adversarial networks (GANs) have proven to be a potential method for augmenting data. However, the overall quality of data generated by GANs still requires improvement. In this study, a multiscale feature-fusion GAN was proposed for generating high-dimensional signals. A strategy that generates signals with dimensionality higher than the desired dimensionality was proposed. The generated signals were resampled to improve the phase diversity of the generated signals. To enhance the feature extraction capability of the GAN for high-dimensional signals, a multiscale feature extraction structure was designed. The integration of multiscale feature extraction and fusion was achieved without a significant increase in the computational burden. To enhance the amplitude diversity of the generated signals, a reconstruction network was designed that directly constrained the spatial distribution of the generated signals. Experimental results show that the model has advantages in terms of the similarity and diversity of the generated signals. The effectiveness of the model in fault diagnosis was verified using two motor datasets, where the fault diagnosis model using the sample set augmented by the proposed model obtained a 36.32% improvement in accuracy compared with using the original sample set. In addition, in comparison experiments, the model achieved a higher diagnostic accuracy improvement of 28.87% compared to four other published models.

Machine learning classifier-based dynamic surrogate model for structural reliability analysis

The performance functions of large-scale complex structures are often implicit and nonlinear, which leads to problems such as high computational costs and low computational accuracy in reliability calculation. In this regard, a dynamic machine learning classifier surrogate model based on Monte Carlo simulation (DMLC-MCS) is proposed. The training sample points are generated by the Markov chain Monte Carlo (MCMC) method and numerical analysis to create the training sample dataset, and the surrogate model based on machine learning classifiers (MLCs) are used to reconstruct the limit state function (LSF). Then, samples are extracted by MCS technique, and the LSF values are predicted by the trained surrogate model. An iterative process is proposed around the most probable point (MPP), and the failure probability obtained by the MCS technique is taken as the convergence condition. If the convergence condition is not satisfied, the MPP information is added to the original sample set to refine the surrogate model. Compared with the traditional reliability method, the proposed method significantly reduces the computational cost on the premise of ensuring high accuracy. In addition, the method is easy to combine with numerical analysis and is proven to be applicable for reliability analysis of real-world complex engineering problems.

Critical assessment of chemometric models employed for varietal authentication of wine based on UHPLC-HRMS data

The use of metabolic fingerprinting combined with advanced chemometric tools for wine authentication has increased in recent years. Although numerous studies, showing different authentication strategies, have been published, rarely any attention has been paid to the stability of used classification models over a longer time period.Here, we present a reliable and robust metabolic fingerprinting-based multiclass strategy for varietal authentication of wine. Analysis was conducted using ultra-high-performance liquid chromatography coupled to high-resolution tandem mass spectrometry. Two sets of commercial wine samples, one for the creation of classification models (201 wines, five red and five white grape varieties) and one for the verification of their validity over a longer time period (138 wines, three white varieties), were analysed. The generated data from the first sample set were subjected to orthogonal partial least squares discriminant analysis (OPLS-DA). The resulting models were validated and used to build decision trees, which enabled the classification of wine samples according to the grape variety. The individual classification rates of the OPLS-DA models were 90–100%. Overall classification rates of the decision trees were 94 and 96% for red and white wines, respectively. In case of the white wine decision tree, verification of its validity over a longer time period was performed using an additional sample set, analysed four months after the original sample set. From the additional sample set, 87% of samples were correctly classified, thus, the stability of the OPLS-DA classification models over a longer time period was verified. In addition, 25 varietal markers of significant statistical importance, mostly flavonoids, phenolic acids and their derivatives, were tentatively identified.

Ecofriendly shiitake authentication using bulk and amino acid-specific stable isotope models

Accurate authentication of ecofriendly shiitake, such as organic and pesticide-free shiitake, is required to improve food safety and to increase reliability of the national agrofood certification system; however, this is a challenging task. Therefore, this study examined the feasibility of bulk and compound-specific isotope analyses to discriminate ecofriendly shiitake against conventional counterparts. Using the compound-specific isotope model, the classification accuracy was greater (100%) than that of the bulk isotope model (74.5%) for each original sample set. In the compound-specific model, a cutting score of −4.42 discriminated organic shiitake from pesticide-free shiitake and a cutting score of 4.87 discriminated organic shiitake from conventional shiitake. The isotope fractionation trend was less influenced by shiitake type and the amino acid synthetic pathway. Thus, the compound-specific isotope model of amino acids may be a good complementary authentication tool to overcome the limitations of bulk stable isotopes or a pesticide residue test.

A spatial distance-based spatial clustering algorithm for sparse image data

By allocating each object to one of the predefined categories, image classification deeply understands the attributes and features of the data on each object in the scene, and further mines the potential features and internal connections of the data, supporting the subsequent application decision-making with necessary structured data. One of the key challenges to image classification is how to accurately classify sparse data, when there is an imbalance between different categories of data, i.e., how to identify small objects in images. Recognizing a person in satellite images is such a challenging task. These objects are sparse either globally or in each recognizable local segment. Therefore, they are often overlooked by the classifier, or removed as noises. During deep learning, feature sparsity means the samples contain too much useless information, which suppresses the generalization and accuracy of the model. To solve the problem, this paper presents a spatial distance-based spatial clustering algorithm for sparse image data (SDBSCA-SID). Firstly, the imaging range of the image sensor constitutes a two-dimensional (2D) constraint space. Under the constraint, spatial clustering was carried out based on the features of each sample to aggregate dense data into primary categories, and aggregate sparse data and noises into secondary categories. Referring to the 2D constrained space, multiple spatial classification surfaces were constructed to aggregate the sparse data to the two sides of these surfaces as much as possible. If the error is minimized, then the sparse data belong to these classification surfaces. To shorten the convergence time of the clustering algorithm on imbalanced data, the original sample set was cut into slices, and assigned to several calculation units for separate clustering. Next, the same-class clusters were merged through reduction. Finally, the obtained class labels were compared with the preset class labels, wrapping up the semantic segmentation of images. The stability and accuracy of our algorithm were demonstrated through tests on image samples.

How Does Sample Labeling and Distribution Affect the Accuracy and Efficiency of a Deep Learning Model for Individual Tree-Crown Detection and Delineation

Monitoring and assessing vegetation using deep learning approaches has shown promise in forestry applications. Sample labeling to represent forest complexity is the main limitation for deep learning approaches for remote sensing vegetation classification applications, and few studies have focused on the impact of sample labeling methods on model performance and model training efficiency. This study is the first-of-its-kind that uses Mask region-based convolutional neural networks (Mask R-CNN) to evaluate the influence of sample labeling methods (including sample size and sample distribution) on individual tree-crown detection and delineation. A flight was conducted over a plantation with Fokienia hodginsii as the main tree species using a Phantom4-Multispectral (P4M) to obtain UAV imagery, and a total of 2061 manually and accurately delineated tree crowns were used for training and validating (1689) and testing (372). First, the model performance of three pre-trained backbones (ResNet-34, ResNet-50, and ResNet-101) was evaluated. Second, random deleting and clumped deleting methods were used to repeatedly delete 10% from the original sample set to reduce the training and validation set, to simulate two different sample distributions (the random sample set and the clumped sample set). Both RGB image and Multi-band images derived from UAV flights were used to evaluate model performance. Each model’s average per-epoch training time was calculated to evaluate the model training efficiency. The results showed that ResNet-50 yielded a more robust network than ResNet-34 and ResNet-101 when the same parameters were used for Mask R-CNN. The sample size determined the influence of sample labeling methods on the model performance. Random sample labeling had lower requirements for sample size compared to clumped sample labeling, and unlabeled trees in random sample labeling had no impact on model training. Additionally, the model with clumped samples provides a shorter average per-epoch training time than the model with random samples. This study demonstrates that random sample labeling can greatly reduce the requirement of sample size, and it is not necessary to accurately label each sample in the image during the sample labeling process.

Difference-value background based on the subset of the category in hyperspectral anomaly detection

Hyperspectral anomaly detection (HAD) is an important unsupervised target detection technique. Background modelling techniques occupy an important component in HAD. The background modelling process is easily disturbed by the anomalous target, resulting in a large error between the estimated value and the real value. To solve this problem, this paper proposes a new background difference value method. Firstly, the hyperspectral datasets are divided into several subsets by endmembers, and then the subsets with fewer samples are removed from the original sample set to a new background set. Finally, the background model parameters are estimated by the new background set. Experimental results show that the proposed algorithm has better detection performance in complex scenes.

Compressed Sensing-Based Multitarget CFAR Detection Algorithm for FMCW Radar

Constant false alarm rate (CFAR) detection algorithms, which are widely used in frequency-modulated continuous wave (FMCW) radar systems, achieve target detection by employing a threshold determined on the basis of a predicted background level. However, in multitarget scenarios, the multitarget shadowing effect can lead to inaccurate prediction of the background level and improper setting of the threshold, which then results in severely degraded CFAR performance. To combat this multitarget shadowing effect, a novel CFAR algorithm based on sparsity adaptive correlation maximization (SACM-CFAR) is proposed in this work. The proposed SACM-CFAR algorithm realizes target detection by utilizing the correlation between linear measurements of the radar intermediate frequency (IF) signal and the sensing matrix. To achieve a desired false alarm rate, the proposed algorithm determines the threshold by estimating the distributed parameters of the reduced sample set obtained by removing the detected targets from the original sample set. Both simulation results and field test results verify that the proposed algorithm outperforms conventional algorithms in multitarget scenarios.

High Wind Speed Inversion Model of CYGNSS Sea Surface Data Based on Machine Learning

In response to the deficiency of the detection capability of traditional remote sensing means (scatterometer, microwave radiometer, etc.) for high wind speed above 25 m/s, this paper proposes a GNSS-R technique combined with a machine learning method to invert high wind speed at sea surface. The L1-level satellite-based data from the Cyclone Global Navigation Satellite System (CYGNSS), together with the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Centers for Environmental Prediction (NCEP) data, constitute the original sample set, which is processed and trained with Support Vector Regression (SVR), the combination of Principal Component Analysis (PCA) and SVR (PCA-SVR), and Convolutional Neural Network (CNN) methods, respectively, to finally construct a sea surface high wind speed inversion model. The three models for high wind speed inversion are certified by the test data collected during Typhoon Bavi in 2020. The results show that all three machine learning models can be used for high wind speed inversion on sea surface, among which the CNN method has the highest inversion accuracy with a mean absolute error of 2.71 m/s and a root mean square error of 3.80 m/s. The experimental results largely meet the operational requirements for high wind speed inversion accuracy.

ПРАКТИЧЕСКИЙ ПОДХОД К ОПРЕДЕЛЕНИЮ ДОСТАТОЧНОСТИ ЭКСПЕРИМЕНТАЛЬНОЙ ВЫБОРКИ ВЕКТОРНЫХ СИГНАЛОВ ГАЗОАНАЛИТИЧЕСКОЙ МУЛЬТИСЕНСОРНОЙ ЛИНЕЙКИ ДЛЯ ОБУЧЕНИЯ ИСКУССТВЕННОЙ НЕЙРОННОЙ СЕТИ

An algorithm for finding a sufficient amount of a representative set of sample data for training artificial neural networks using the numerical parameters of the original sample set is proposed. The algorithm is carried out on the example of analyzing a vector signal generated by a gas analytical multisensor micro-arrays based on a thin SnO2 film when calibrated to the effect of CO, isopropanol, and ethanol in a mixture with air. As a result of the operation of the algorithm, the minimum required volume of the training sample of artificial neural networks was found, which allows achieving a high (more than 99 %) recognition quality. The results show the efficiency of the proposed algorithm.

Novel adaptive sample space expansion approach of NIR model for in-situ measurement of gasoline octane number in online gasoline blending processes

For real-time measurement of gasoline octane number in online gasoline blending processes by near infrared (NIR) spectroscopy, this paper presents an adaptive sample space expansion approach (ASSEA) to improve the real-time measurement accuracy and to shorten the time for establishing the model. To improve the adaptability of the model, a comprehensive index that combines the characteristics of the query samples and the distance between the spectra is adopted to select the guide spectra. A sample construction method is proposed to supplement the original sample set with insufficient sample size. The new spectra are constructed with the labeled spectra based on the guide spectra. Subsequently, the concentration information of the constructed spectra is obtained on the basis of multi-way partial least square method. The proposed ASSEA method is compared with six representative methods. The prediction performance of ASSEA method (RMSE is 0.2442, and R2 is 0.8767) is significantly better than other methods. And compared with the traditional PLS method, RMSE and R2 are reduced and increased by 33.4% and 45.9%, respectively. Therefore, its efficiency is validated through an industrial gasoline blending process.

Semantic segmentation sample augmentation based on simulated scene generation—case study on dock extraction from high spatial resolution imagery

ABSTRACT Deep learning-based semantic segmentation methods, such as fully convolutional networks (FCNs), are state-of-the-art techniques for object extraction from high spatial resolution images. However, collecting massive scene-formed training samples typically required in FCNs is time-consuming and labour-intensive. A suit of automatic sample augmentation schemes based on simulated scene generation is proposed in this study to reduce the manual workload. Proposed schemes include style transfer, target embedding, and mixed modes by utilizing techniques, such as texture transfer, image inpainting, and region-line primitive association framework, which automatically expand the sample set on the basis of a small number of real samples. Dock extraction experiments using UNet are conducted on China’s GaoFen-2 imagery with expanded sample sets. Results showed that the proposed schemes can successfully generate sufficient simulation samples, increase sample diversity, and subsequently improve semantic segmentation accuracy. Compared with the results that use the original real sample set, measures of F 1-score (F 1) and intersection over union (IoU) of dock extraction accuracy demonstrate a maximum improvement of 20.53% and 23.01%, respectively, after sample augmentation.

Identification of NTRK gene fusions in lung adenocarcinomas in the Chinese population.

The molecular profile of neurotrophic tyrosine kinase receptor (NTRK) gene fusions in lung adenocarcinoma (LUAD) is not fully understood. Next‐generation sequencing (NGS) and pan‐tyrosine kinase receptor (TRK) immunohistochemistry (IHC) are powerful tools for NTRK fusion detection. In this study, a total of 4,619 LUAD formalin‐fixed, paraffin‐embedded tissues were collected from patients who underwent biopsy or resection at the Shanghai Chest Hospital during 2017–2019. All specimens were screened for NTRK1 rearrangements using DNA‐based NGS. Thereafter, the cases with NTRK1 rearrangements and cases negative for common driver mutations were analyzed for NTRK1/2/3 fusions using total nucleic acid (TNA)‐based NGS and pan‐TRK IHC. Overall, four NTRK1/2 fusion events were identified, representing 0.087% of the original sample set. At the DNA level, seven NTRK1 rearrangements were identified, while only two TPM3‐NTRK1 fusions were confirmed on TNA‐based NGS as functional. In addition, two NTRK2 fusions (SQSTM1‐NTRK2 and KIF5B‐NTRK2) were identified by TNA‐based NGS in 350 ‘pan‐negative’ cases. Two patients harboring NTRK1/2 fusions were diagnosed with invasive adenocarcinoma, while the other two were diagnosed with adenocarcinoma in situ and minimally invasive adenocarcinoma. All four samples with NTRK fusions were positive for the expression of pan‐TRK. The two samples with NTRK2 fusions showed cytoplasmic staining alone, while the other two samples with NTRK1 fusions exhibited both cytoplasmic and membranous staining. In summary, functional NTRK fusions are found in early‐stage LUAD; however, they are extremely rare. According to this study's results, they are independent oncogenic drivers, mutually exclusive with other driver mutations. We demonstrated that NTRK rearrangement analysis using a DNA‐based approach should be verified with an RNA‐based assay.

Regional Characterization Study of Fatty Acids and Tocopherol in Organic Milk as a Tool for Potential Geographical Identification.

Reliable geographical identification can protect producers of excellent agroproducts, and also provide reliable purchasing information to satisfy consumers. Therefore, this study aimed to evaluate the regional and monthly variation in fatty acid (FA) and tocopherol (TOC) in organic milk (OM) and develop a geographical discriminant model of OM in Korea. In this study, OM had α-TOC and showed a regional or monthly difference of 3–5%. Moreover, C16:0, C18:1 n9 cis + trans, C18:0, and C14:0 were the predominant FAs in OM, and OM mostly had higher ∑UFA, including nutritionally desirable FAs; but lower ∑SFA among four regions or in April and August (p < 0.001). The model prepared using stepwise discriminant analysis showed a classification accuracy of 100% for original and cross-validated sample sets. Our results have characterized regional and monthly nutritional variations of OM, thereby potentially suggesting the applicability of a reliable Korean geographical identification labeling system using nutrient compositional analysis of OM.

Conditional Generative Adversarial Network-Based Training Sample Set Improvement Model for the Semantic Segmentation of High-Resolution Remote Sensing Images

To achieve high segmentation quality, deep semantic segmentation neural networks (DSSNNs) need to be trained on diverse direction, location, and neighboring category combinations for each pixel in the input-output image patches. Achieving this goal requires a large sample set. However, in many practical application scenarios, a very large training sample set is too expensive to achieve, or insufficient remote sensing image data are available. These limitations directly affect the quality of the results of DSSNNs. To address the above-mentioned problem, this article proposes a conditional generative adversarial network (CGAN)-based training sample set improvement model (CGAN-TSIM) for the semantic segmentation of high-resolution remote sensing images. In CGAN-TSIM, the generator model of the CGAN can generate a sample image when a ground-truth image is an input as a “condition.” A condition generation mechanism is designed to create ground-truth images, and these ground-truth conditions are used to drive the CGAN to generate samples containing more diverse object combinations, directions, and locations. These generated images can be added to the original training sample set to improve their spatial information diversity. Rather than simply relying on passively finding samples that contain diverse spatial information, CGAN-TSIM extracts high-level spatial information from the original training images and actively generates new sample images. Experiments show that the samples generated by CGAN-TSIM can improve the quality of the sample set. Compared with other traditional methods, CGAN-TSIM enables better classification accuracy.

Short-Term Traffic Flow Intensity Prediction Based on CHS-LSTM

Short-term traffic flow prediction is an important basis of intelligent transportation systems. Its accuracy directly affects the performance of traffic control and induction. To improve prediction accuracy, in this paper, a traffic flow prediction model is proposed by combining hierarchical agglomerative clustering (HAC), standardized Euclidean distance (SED), and a long short-term memory network (LSTM). The proposed model is called the CHS-LSTM model. In this model, HAC is used to carry out cluster analysis on the original traffic flow data sample set of target detection sections. Based on the results of cluster analysis, traffic flow data are divided into several categories. For different categories of traffic flow, SED is used to calculate the spatial correlation of the road network where the target sections are located, and the K sections most relevant to the target detection sections are used in the construction of an input data matrix for LSTM. The prediction result with the minimum root-mean-square error is regarded as the final prediction result. The electronic toll collection data of Taiwan are used as the fundamental data in this paper to verify the performance and effectiveness of the CHS-LSTM model. The experimental results indicate that the CHS-LSTM model can effectively improve the prediction accuracy of LSTM. Moreover, compared with well-known models generally used for predicting the intensity of traffic flow, our proposed model also shows its superiority.