Small Amount Of Data Research Articles

Reducing the effect of face orientation using FaceMesh landmarks in drowsiness estimation based on facial thermal images

Abstract In this study, facial skin temperature distribution (FSTD) is focused on as a new driver monitoring index. FSTD is an autonomic index that can be measured remotely. Studies have been conducted to estimate drowsiness based on FSTD using modelng methods such as CNN, a type of deep learning, and sparse modeling, which can be trained with a small amount of data. These studies, however, only evaluated front-facing facial thermal images. FaceMesh is a model that extracts 478 3D facial feature landmarks from a 2D face image. In contrast to conventional models that extract only 68 facial feature landmarks, FaceMesh can extract facial feature landmarks for the entire face, including the cheeks, forehead, and other areas of the face that are in the blind spots. This study aims to improve the accuracy of drowsiness estimation by applying FaceMesh to automatically detect tilted faces and not including tilted images in the training data. As a result, the method proposed in this study improved drowsiness estimation accuracy by about 6% compared to the old method, which did not take face orientation into account.

BERT-Based Model and LLMs-Generated Synthetic Data for Conflict Sentiment Identification in Aspect-Based Sentiment Analysis

Aim/Purpose: Most research on aspect-based sentiment analysis (ABSA) classifies sentiment polarity into two classes (positive and negative) or three classes (positive, negative, and neutral) and does not include conflict sentiment. This study investigates the four-class sentiment classification (positive, negative, neutral, and conflict) and proposes a BERT-based model for identifying conflicting sentiment in ABSA. Furthermore, we employ the open-source large language models (LLMs) created by Meta, Llama 3, for generating synthetic data to support research on four-class sentiment classification in ABSA. Background: Public opinions and experiences on product reviews, social events, political movements, etc. can be used for exploring customer behavior, predicting customer preferences, understanding public sentiment, etc., so it becomes an important component in the decision-making process. Providing an accurate opinion will enable an individual, business, or organization to have an in-formed judgement before making a decision. Methodology: This study utilizes a methodology that includes generating synthetic data to augment the original datasets, designing the input representation, detecting aspect categories, performing a multi-label sentiment classification, and rep-resenting sentiment in a four-class sentiment classification. Contribution: This study provides an investigation on the four-class sentiment classification (positive, negative, neutral, and conflict) and proposes a BERT-based method to identify aspects with conflicting sentiment in ABSA. Moreover, it also evaluates Llama 3 for generating synthetic data to address the issues related to the data scarcity and imbalance datasets in the research of four-class senti-ment classification in ABSA. Findings: The investigation of the four-class sentiment classification task in ABSA demonstrates that identifying conflict sentiment is challenging for several reasons. Among them are (1) the lack of a public dataset for this research; (2) the small amount of data with conflict labels in the available dataset resulting in an imbalanced dataset; (3) conflict sentiment is a complex sentiment con-taining both positive and negative sentiments; and (4) conflicting sentiments are usually expressed in long and complicated sentences and involve implicit aspects. Our solution to these challenges involved generating synthetic data using Llama 3 and designing a BERT-based model on multi-label aspects for identifying aspect with conflict sentiment. Recommendations for Practitioners: Most existing ABSA models with four-class sentiment classification are con-ducted for the product reviews (mostly in the restaurant domain) and in high-resource languages (mainly in English). Therefore, users may need to make some adjustments to different domains and languages. Recommendation for Researchers: Due to the limited datasets available for research in aspect-based sentiment analysis with four-class sentiment classification, the development of a dataset to support this research is urgently needed. Impact on Society: By providing more accurate sentiment through aspect-based sentiment analy-sis, this study can better help people, organizations, or companies in get a view or an opinion about any product, service, or candidate in an electoral vote. Future Research: Future research on aspect-based sentiment analysis could utilize the large language models (LLMs) for conducting ABSA tasks including aspect term extraction, aspect category detection, and its sentiment polarities. We could also focus on evaluating the model for cross-domain and cross-language AB-SA system.

Texts of Different Emotional Classes and Their Topic Modeling

The article is devoted to studying verbalization specifics of various emotional states in the texts in the Russian language with the purpose to confirm or refute the hypothesis that texts of different emotional classes reflect the denotative situation not identically, which is reflected in thematic specifics and lexical content. The research material consisted of eight corpus texts in the Russian language, which were extracted from the public pages of the social network VKontakte. The texts were selected according to emotional hashtags that corresponded to eight basic emotions, according to H. Lцvheim’s model: anger, surprise, shame, enjoyment, disgust, distress, excitement, fear. The correspondence of emotion and hashtag was established in a preliminary psycholinguistic experiment. While analyzing the text collection, we used the method of computer thematic modeling to identify statistically non-random groups of words (topics). We applied the BERTopic neural network model to the collected data. As a result of the analysis, it was found that texts of 8 emotional classes contain an uneven number of topics, despite the fact that their number does not correlate directly with the amount of data: with a relatively small amount of data, there may be many topics, but in a voluminous corpus – few. The sets of words (tokens) that make up each non-random group (topic) differ in each subcorpora, reflecting the specifics of the denotative situation, which is formed under the influence of the emotional state of the speaker. The idea of diverse thematic “granularity” of texts of different emotional classes is theoretically justified.

Data enhanced deep transfer learning for health state evaluation of cutting tools

ABSTRACT Health state estimation of cutting tools is a key component for maintaining the reliability and stability of the manufacturing process. Although deep learning (DL) based tool wear indirect monitoring methods have been presented, most of them require a substantial amount of data to avoid high misclassification rates. To this end, this article proposes a novel estimation approach based on data-enhanced deep transfer learning (TL) methodology, which integrates multivariate data enhancements, deep convolution network, and generative adversarial learning for state assessment of the tool wear under small sample size. The proposed method includes three innovative points: 1) a multi-step data enhancement method is employed to augment the wavelet scalogram data as model input; 2) a novel training process is adopted to obtain optimal hyperparameters of TL-AlexNet and TL-VGGNet-16 model; 3) a deep transfer learning model for the enhanced data is presented to adaptively estimate the health state of the cutting tools. The proposed approach only requires a small quantity of data to achieve satisfactory estimation performance compared with other DL methods through experimental validation, and the average prediction accuracy for each stage is greater than 95%, which demonstrates the effectiveness and superiority of the presented model.

BERTugues: A Novel BERT Transformer Model Pre-trained for Brazilian Portuguese

Large Language Models (LLMs) are trained for English or multilingual versions, with superior performance in English. This disparity occurs because, in the training of multilingual models, only a relatively small amount of data is added for each additional language. Consequently, while these models can function in Portuguese, their performance is suboptimal. The first BERT model (Bidirectional Encoder Representations from Transformers) specifically trained for Brazilian Portuguese was BERTimbau in 2020, which enhanced performance across various text-related tasks. We followed the training approach of BERT/BERTimbau for BERTugues, while implementing several improvements. These included removing rarely used characters in Portuguese from the tokenizer, such as oriental characters, resulting in the addition of over 7,000 new tokens. As a result, the average length of sentence representations was reduced from 3.8 words with more than one token to 3.0, which positively impacted embedding performance by improving metrics relevant to classification problems. Two additional enhancements involved embedding emojis as tokens - an essential step for capturing conversational nuances - and filtering low-quality texts from the training dataset. These modifications improved performance across various tasks, raising the average F1 score from 64.8 % in BERTimbau to 67.9 % in BERTugues.

Computer Vision Technology for Short Fiber Segmentation and Measurement in Scanning Electron Microscopy Images

Computer vision technology for the automatic recognition and geometric characterization of carbon and glass fibers in scanning electron microscopy images is proposed. The proposed pipeline, combining the SAM model and DeepLabV3+, provides the generalizability and accuracy of the foundational SAM model and the ability to quickly train on a small amount of data via the DeepLabV3+ model. The pipeline was trained several times more rapidly with lower requirements for computing resources than fine-tuning the SAM model, with comparable inference time. On the basis of the pipeline, an end-to-end technology for processing images of electron microscopic fibers was developed, the input of which is images with metadata and the output of which is statistics on the distribution of the geometric characteristics of the fibers. This innovation is of great practical importance for modeling the physical characteristics of materials. This paper proposes a few-shot training procedure for the DeepLabV3+/SAM pipeline, combining the training of the DeepLabV3+ model weights and the SAM model parameters. It allows effective training of the pipeline using only 37 real labeled images. The pipeline was then adapted to a new type of fiber and background using 15 additional real labeled images. This article also proposes a method for generating synthetic data for training neural network models, which improves the quality of segmentation by the IoU and PixAcc metrics from 0.943 and 0.949 to 0.953 and 0.959, i.e., by 1% on average. The developed pipeline significantly reduces the time required to evaluate fiber length in scanning electron microscope images.

Utilizing physics‐augmented neural networks to predict the material behavior according to Yeoh's law

AbstractThis article discusses physics‐augmented neural network approaches in the field of hyperelastic material modeling. Physical conditions such as objectivity, material symmetry, or a stress– and energy‐free reference configuration are considered in the construction of the neural networks. In addition, a new approach for stress normalization is proposed. The neural network is used to learn the behavior of Yeoh's constitutive model with sparse data. Finally, the trained networks are incorporated into a three‐dimensional finite element framework and compared with the classical material model in terms of accuracy. The paper demonstrates the ability of physics‐augmented neural networks to model hyperelastic materials using a small amount of data that could be generated by experiments. Compared to the classical constitutive laws of Yeoh's model, our trained material showed no material instabilities that could occur due to poorly chosen material parameters.

Full-scale load testing for evaluating crack performance in precast prestressed concrete communication towers with annular sections

Tall prestressed concrete poles with annular cross-sections are increasingly used as communication towers. The annular section subjected to bending moments has a smaller extreme tension area, resulting in higher stress increments in the materials within this region. As slender concrete structures, their bending behavior and crack resistance performance need to be critically examined. However, large-scale tower structures have rarely been experimentally investigated. Moreover, current code-specified crack width calculation methods are designed for square sections, so their applicability to concrete members with annular sections requires further investigation. To address these issues, this study conducted full-scale static loading tests on two 30-meter-high tower structures made of prestressed high-strength concrete and evaluated the accuracy of code methods for estimating the maximum crack width. During the static loading tests, displacement, cracking moments, crack widths, and crack spacings were meticulously recorded, and a theoretical methodology for calculating the stress in the extreme tension bars was developed. Four calculation methods specified in structural codes were compared: the Concrete Design Code of China (GB50010), Eurocode 2, ACI 224R, and the Concrete Design Guideline of Japan (JGC2007). The experimental results showed that diameter significantly affects the stiffness and displacement of the structure, with reloading stiffness of the full-scale tower reducing by up to 11 % after exceeding the serviceability limit state. Regarding crack width prediction, substantial deviations were observed among the results of the four methods. On average, GB50010 underestimated crack widths, while JSCE, Eurocode 2, and ACI 224R overestimated them. Eurocode 2 and ACI 224R provided relatively conservative and accurate predictions. These findings highlight the need for improved and tailored code provisions for annular sections. It is also noted that the comparison was limited to crack widths <0.1 mm and a relatively small amount of data. Future research should focus on developing prediction methods specifically for prestressed concrete tower structures with annular sections.

Unsteady aerodynamic modeling and flight trajectory simulation of dual-spin projectile based on DNN and transfer-learning

To evaluate flight performance and aerodynamic characteristics of a dual-spin projectile, the coupled computational fluid dynamics and rigid body dynamics (CFD/RBD) method is commonly used, which can simultaneously solve the flight mechanics and flow field. However, the efficiency is compromised by the large number of CFD calculations required. This paper develops an unsteady aerodynamic modeling method that combines deep neural networks and transfer learning, which can accurately predict unsteady aerodynamics of dual-spin projectiles under varying initial conditions. Considering the influence of flight state and aerodynamic data from short-term historical moments, we integrate them as input features of the aerodynamic model to reduce the impact of long-term historical data. To enhance the model generalization under varying initial conditions, we fine-tune the built aerodynamic model using small amounts of data under new conditions by transfer learning. The proposed method is validated through interpolated and extrapolated prediction cases, respectively. The results indicate that the proposed method can achieve better accuracy and generalizability than long short-term memory neural networks and autoregressive moving average method in unsteady aerodynamic modeling of the dual-spin projectile. By coupling the flight dynamics equations with the aerodynamic model in the time domain, the flight simulation only takes a few seconds, which can reduce computing time by three orders of magnitude compared to the coupled CFD/RBD method.

Few-shot learning for inference in medical imaging with subspace feature representations.

Unlike in the field of visual scene recognition, where tremendous advances have taken place due to the availability of very large datasets to train deep neural networks, inference from medical images is often hampered by the fact that only small amounts of data may be available. When working with very small dataset problems, of the order of a few hundred items of data, the power of deep learning may still be exploited by using a pre-trained model as a feature extractor and carrying out classic pattern recognition techniques in this feature space, the so-called few-shot learning problem. However, medical images are highly complex and variable, making it difficult for few-shot learning to fully capture and model these features. To address these issues, we focus on the intrinsic characteristics of the data. We find that, in regimes where the dimension of the feature space is comparable to or even larger than the number of images in the data, dimensionality reduction is a necessity and is often achieved by principal component analysis or singular value decomposition (PCA/SVD). In this paper, noting the inappropriateness of using SVD for this setting we explore two alternatives based on discriminant analysis (DA) and non-negative matrix factorization (NMF). Using 14 different datasets spanning 11 distinct disease types we demonstrate that at low dimensions, discriminant subspaces achieve significant improvements over SVD-based subspaces and the original feature space. We also show that at modest dimensions, NMF is a competitive alternative to SVD in this setting. The implementation of the proposed method is accessible via the following link.

Joint Mixing Data Augmentation for Skeleton-based Action Recognition

Skeleton-based action recognition is beneficial for understanding human behavior in videos, and thus has received much attention in recent years as an important research area in action recognition. Current research focuses on designing more advanced algorithms to better extract spatio-temporal information from skeleton data. However, due to the small amount of data in the existing skeleton dataset and the lack of effective data augmentation methods, it is easy to lead to overfitting in model training. To address this challenge, we propose a mix-based data augmentation method, Joint Mixing Data Augmentation (JMDA), which can generally improve the effectiveness and robustness of various skeleton-based action recognition algorithms. In terms of spatial information, we introduce SpatialMix (SM), a method that projects the original 3D skeleton discrete information into a 2D space. Then, SM mixes the projected spatial information between two random samples during the training process to achieve the spatial-based mixing data augmentation. Concerning temporal information, we propose TemporalMix (TM). Leveraging the temporal continuity in skeleton data, we perform a temporal resize operation on the original skeleton data, and then merge two random samples during training to achieve the temporal-based mixed data augmentation. Additionally, we analyze the Feature Mismatch (FM) problem caused by introducing mix-based data augmentation into skeleton data. Then we propose a new data preprocessing method called Feature Alignment (FA) to effectively address this problem and improve model performance. Moreover, we propose a novel training pipeline, Joint Training Strategy (JTS), which combines multiple mix-based data augmentation methods for further improvement of model performance. Specifically, our proposed JMDA is plug-and-play and widely applicable to skeleton-based action recognition models. At the same time, the application of JMDA does not increase the model parameters and there is almost no additional training cost. We conduct extensive experiments on NTU RGB+D 60 and NTU RGB+D 120 datasets to demonstrate the effectiveness and robustness of the proposed JMDA on several mainstream skeleton-based action recognition algorithms.

Acoustic field visualization and source localization via physics-informed learning of sparse data with adaptive sampling

Traditional methods for acoustic field visualization require considerable effort for capturing large amounts of acoustic data to achieve a high resolution field map, highly limiting their widespread use. In this study, we propose an approach for acoustic field visualization based on physics-informed neural networks (PINNs) by using a small amount of data, subsequently realizing accurate acoustic source localization. First, we present a PINN model integrated with an acoustic Helmholtz equation and adaptive sampling, the performance of which is testified via numerical simulations. The “no mesh” character of PINN enables achieving high resolution acoustic field visualization without requiring the capture of numerous data in advance. Furthermore, we experimentally validate the performance of the proposed method, which demonstrates that the acoustic sources can be precisely localized with sparse field data acquisition within a small area. This work would find potential applications in various acoustics, such as acoustic communication, biomedical imaging, and virtual reality.

Fast Global and Local Semi-Supervised Learning via Matrix Factorization

Matrix factorization has demonstrated outstanding performance in machine learning. Recently, graph-based matrix factorization has gained widespread attention. However, graph-based methods are only suitable for handling small amounts of data. This paper proposes a fast semi-supervised learning method using only matrix factorization, which considers both global and local information. By introducing bipartite graphs into symmetric matrix factorization, the technique can handle large datasets effectively. It is worth noting that by utilizing tag information, the proposed symmetric matrix factorization becomes convex and unconstrained, i.e., the non-convex problem minx(1−x2)2 is transformed into a convex problem. This allows it to be optimized quickly using state-of-the-art unconstrained optimization algorithms. The computational complexity of the proposed method is O(nmd), which is much lower than that of the original symmetric matrix factorization, which is O(n2d), and even lower than that of other anchor-based methods, which is O(nmd+m2n+m3), where n represents the number of samples, d represents the number of features, and m≪n represents the number of anchors. The experimental results on multiple public datasets indicate that the proposed method achieves higher performance in less time.

Physics informed data-driven near-wall modelling for lattice Boltzmann simulation of high Reynolds number turbulent flows

Data-driven approaches offer novel opportunities for improving the performance of turbulent flow simulations, which are critical to wide-ranging applications from wind farms and aerodynamic designs to weather and climate forecasting. However, current methods for these simulations often require large amounts of data and computational resources. While data-driven methods have been extensively applied to the continuum Navier-Stokes equations, limited work has been done to integrate these methods with the highly scalable lattice Boltzmann method. Here, we present a physics-informed neural network framework for improving lattice Boltzmann-based simulations of near-wall turbulent flow. Using a small amount of data and integrating physical constraints, our model accurately predicts flow behaviour at a wide range of friction Reynolds numbers up to 1.0 × 106. In contradistinction with other models that use direct numerical simulation datasets, this approach reduces data requirements by three orders of magnitude and allows for sparse grid configurations. Our work broadens the scope of lattice Boltzmann applications, enabling efficient large-scale simulations of turbulent flow in diverse contexts.

Decision Transformer-Based Efficient Data Offloading in LEO-IoT.

Recently, the Internet of Things (IoT) has witnessed rapid development. However, the scarcity of computing resources on the ground has constrained the application scenarios of IoT. Low Earth Orbit (LEO) satellites have drawn people's attention due to their broader coverage and shorter transmission delay. They are capable of offloading more IoT computing tasks to mobile edge computing (MEC) servers with lower latency in order to address the issue of scarce computing resources on the ground. Nevertheless, it is highly challenging to share bandwidth and power resources among multiple IoT devices and LEO satellites. In this paper, we explore the efficient data offloading mechanism in the LEO satellite-based IoT (LEO-IoT), where LEO satellites forward data from the terrestrial to the MEC servers. Specifically, by optimally selecting the forwarding LEO satellite for each IoT task and allocating communication resources, we aim to minimize the data offloading latency and energy consumption. Particularly, we employ the state-of-the-art Decision Transformer (DT) to solve this optimization problem. We initially obtain a pre-trained DT through training on a specific task. Subsequently, the pre-trained DT is fine-tuned by acquiring a small quantity of data under the new task, enabling it to converge rapidly, with less training time and superior performance. Numerical simulation results demonstrate that in contrast to the classical reinforcement learning approach (Proximal Policy Optimization), the convergence speed of DT can be increased by up to three times, and the performance can be improved by up to 30%.

SONAR: An Adaptive Control Architecture for Social Norm Aware Robots

AbstractRecent advances in robotics and artificial intelligence have made it necessary or desired for humans to get involved in interactions with social robots. A key factor for the human acceptance of these robots is their awareness of environmental and social norms. In this paper, we introduce SONAR (for SOcial Norm Aware Robots), a novel robot-agnostic control architecture aimed at enabling social agents to autonomously recognize, act upon, and learn over time social norms during interactions with humans. SONAR integrates several state-of-the-art theories and technologies, including the belief-desire-intention (BDI) model of reasoning and decision making for rational agents, fuzzy logic theory, and large language models, to support adaptive and norm-aware autonomous decision making. We demonstrate the feasibility and applicability of SONAR via real-life experiments involving human-robot interactions (HRI) using a Nao robot for scenarios of casual conversations between the robot and each participant. The results of our experiments show that our SONAR implementation can effectively and efficiently be used in HRI to provide the robot with environmental and social and norm awareness. Compared to a robot with no explicit social and norm awareness, introducing social and norm awareness via SONAR results in interactions that are perceived as more positive and enjoyable by humans, as well as in higher perceived trust in the social robot. Moreover, we investigate, via computer-based simulations, the extent to which SONAR can be used to learn and adapt to the social norms of different societies. The results of these simulations illustrate that SONAR can successfully learn adequate behaviors in a society from a relatively small amount of data. We publicly release the source code of SONAR, along with data and experiments logs.

Machine Learning in Small and Medium-Sized Enterprises, Methodology for the Estimation of the Production Time

Data mining (DM) and machine learning (ML) are widely used in production planning and scheduling. Their application to production time estimation leads to improved planning and scheduling accuracy, resulting in increased overall efficiency. Small and medium-sized enterprises (SMEs) often have a small amount of data, which results in the limited adoption of DM and ML. Instead, production time estimation is still performed using rough approximations, which are inaccurate and non-reproducible. Therefore, this article proposes an ML methodology for production time estimation. It is adapted to the needs of SMEs and is applied with limited data. The methodology is based on the categorization of four job types (from A to D), the partitioning of data according to the limit theorem of data convergence, and the definition of risk based on metrics of probability and statistics. ML was applied by WEKA Workbench (Waikato Environment for Knowledge Analysis). It is also integrated into the Cross Industry Standard Process for DM. The methodology was implemented on data from a medium-sized company, Schoepstal Maschinenbau GmbH, for job types A and B to estimate machine/job cycle time, manufacturing cycle time, and lead time. Different accuracies were obtained for individual estimation models, confirming the strong dependence of the models on data quality. Suitable models were found for the implementation of the estimation of the manufacturing cycle time and the machine/job cycle time. The modeling of lead time estimation was unsuccessful. This was due to the weak dependence between the learning values and the values of the selected model attributes. The implementation of the methodology for job types C and D is the subject of further research.

Sports action recognition algorithm based on multi-modal data recognition

The recognition of sports action is an important research subject, which is conducive to the improvement of athletes’ own level. To improve the accuracy of multi-modal data action recognition, based on the Transformer module, this study introduces a multi-head attention mechanism, fuses multi-modal data, and constructs a multi-stream structured object relationship inference network. Based on PointNet++ network and combining five different data fusion frameworks, a motion recognition model that integrates RGB data and 3D skeleton point cloud is constructed. The results showed that the Top-1 accuracy of multi-stream structured object relationship inference network was 42.5% and 42.7%, respectively, which was better than other algorithms. The accuracy of the multi-modal fusion model was improved by 15.6% and 5.1% compared with the single mode, and by 5.4% and 2.6% compared with the dual mode, which showed its superiority in the action recognition task. This showed that the fusion of multi-modal data can provide more abundant information, so as to improve the accuracy of action recognition. The accuracy of the action recognition model combining RGB data and 3D skeleton point cloud was 84.3%, 87.5%, 90.2%, 90.6% and 91.2% after the combination of different strategies, which effectively compensated for the problem of missing information in 3D skeleton point cloud and significantly improved the accuracy of action recognition. With a small amount of data, the Top-1 accuracy of the multi-stream structured object relationship inference network in this study was superior to other algorithms, showing its advantages in dealing with complex action recognition tasks. In addition, the action recognition model that fuses RGB data and 3D skeleton point cloud also achieved higher accuracy, which is better than other algorithms. This study can meet the needs of motion recognition in different scenarios and has certain reference value.

Morphologic and hemodynamic analysis of intracranial mirror aneurysms

Abstract Intracranial mirror aneurysms are pathologic dilatations of cerebral arteries that occur at identical locations in both hemispheres. They do not always have identical sizes, shapes, or rupture states, despite identical patient-specific risk factors. As a subgroup of multiple aneurysms and thus small amounts of data, clinical research about their causes and development has not been conclusive. Using eight patients with 16 mirror aneurysms with different rupture states, we investigated differences between the hemispheres. We virtually performed stitching: detaching the aneurysm domes to place them on the opposite parent vessel to assess the influence of aneurysm morphology and upstream parent vessel shape on hemodynamics. In the unaltered datasets, we found that blood flow within aneurysms in the left hemisphere has higher wall shear stress and velocity. After stitching, we found that both upstream parent vessel as well as aneurysm morphology influence hemodynamics. Morphologic parameters like size, undulation index, and aneurysm angle had a higher influence. In a short rupture risk analysis we found a correlation between rupture and aspect ratio and non-sphericity index. Overall, we found possible indicators for mirror aneurysm analysis that illustrated the importance of morphological analysis, though more datasets are necessary for conclusive statements about special treatment for mirror aneurysms.

Bearing fault detection system based on a deep diffusion model

Bearings are crucial components of modern high-precision machinery and rotating machines. Excellent bearing failure detection systems are vital for ensuring that machines operate precisely. Advances in artificial neural networks (ANNs) and increases in computer processing speed have led to the application of many ANN models in various fields, including bearing failure detection, with excellent outcomes being achieved. However, to construct an ANN model that can precisely detect bearing failures, large quantities of data must be collected on various types of bearing failures. Thus, considerable time must be spent in data collection before rotating machines are operated on the production line, which increases costs for manufacturers. To overcome this problem, the present study used a diffusion model for data augmentation to improve the accuracy of an ANN model trained on a small quantity of bearing sound data. This study performed time-delay mapping to preprocess the data and convert them into a two-dimensional time-delay mapping diagram to reduce the dimensionality of the data features, a novel approach in the field of bearing failure detection. Finally, this study used a convolutional neural network model, which exhibited the optimal classification performance for time-delay mapping diagrams, for bearing failure detection. By comparing the results obtained from augmented and raw data, this study confirmed that using a diffusion model to augment data can improve the generalization ability of bearing failure detection models trained on a small quantity of data.