Training Database Research Articles

Utilizing semantic-level computer vision for fracture trace characterization of hard rock pillars in underground space

This study utilizes a semantic-level computer vision-based detection to characterize fracture traces of hard rock pillars in underground space. The trace images captured by photogrammetry are used to establish the database for training two convolutional neural network (CNN)-based models, i.e., U-Net (University of Freiburg, Germany) and DeepLabV3+ (Google, USA) models. Chain code technology, polyline approximation algorithm, and the circular window scanning approach are combined to quantify the main characteristics of fracture traces on flat and uneven surfaces, including trace length, dip angle, density, and intensity. The extraction results indicate that the CNN-based models have better performances than the edge detection methods-based Canny and Sobel operators for extracting the trace and reducing noise, especially the DeepLabV3+ model. Furthermore, the quantization results further prove the reliability of extracting the fracture trace. As a result, a case study with two types of traces (i.e., on flat and uneven surfaces) demonstrates that the applied semantic-level computer vision detection is an accurate and efficient approach for characterizing the fracture trace of hard rock pillars.

Learning-based intelligent trajectory planning for auto navigation of magnetic robots.

Introduction: Electromagnetically controlled small-scale robots show great potential in precise diagnosis, targeted delivery, and minimally invasive surgery. The automatic navigation of such robots could reduce human intervention, as well as the risk and difficulty of surgery. However, it is challenging to build a precise kinematics model for automatic robotic control because the controlling process is affected by various delays and complex environments. Method: Here, we propose a learning-based intelligent trajectory planning strategy for automatic navigation of magnetic robots without kinematics modeling. The Long Short-Term Memory (LSTM) neural network is employed to establish a global mapping relationship between the current sequence in the electromagnetic actuation system and the trajectory coordinates. Result: We manually control the robot to move on a curved path 50 times to form the training database to train the LSTM network. The trained LSTM network is validated to output the current sequence for automatically controlling the magnetic robot to move on the same curved path and the tortuous and branched new paths in simulated vascular tracks. Discussion: The proposed trajectory planning strategy is expected to impact the clinical applications of robots.

A global land cover training dataset from 1984 to 2020

State-of-the-art cloud computing platforms such as Google Earth Engine (GEE) enable regional-to-global land cover and land cover change mapping with machine learning algorithms. However, collection of high-quality training data, which is necessary for accurate land cover mapping, remains costly and labor-intensive. To address this need, we created a global database of nearly 2 million training units spanning the period from 1984 to 2020 for seven primary and nine secondary land cover classes. Our training data collection approach leveraged GEE and machine learning algorithms to ensure data quality and biogeographic representation. We sampled the spectral-temporal feature space from Landsat imagery to efficiently allocate training data across global ecoregions and incorporated publicly available and collaborator-provided datasets to our database. To reflect the underlying regional class distribution and post-disturbance landscapes, we strategically augmented the database. We used a machine learning-based cross-validation procedure to remove potentially mis-labeled training units. Our training database is relevant for a wide array of studies such as land cover change, agriculture, forestry, hydrology, urban development, among many others.

Application of the Variational Principle to Create a Measurable Assessment of the Relevance of Objects Included in Training Databases

Application of the Variational Principle to Create a Measurable Assessment of the Relevance of Objects Included in Training Databases

Deep learning-based extraction and quantification of features in XCT images of steel corrosion in concrete

The corrosion of rebars in concrete has been regarded as one of the most predominant factors for the degradation of reinforced concrete (RC) structures. The accurate predictions of corrosion damages are the basis for the safety assessment of in-service RC structures and the enhancement of new structure designs. X-ray computed tomography (XCT), known for its non-destructive capabilities, has been widely utilized to understand the corrosion process inside the concrete, offering three-dimensional (3D) visualization and aiding in the identification of the parameters in the predictive model of steel corrosion. This study uses U-Net, a deep learning network, to detect various phases of steel corrosion evolution in concrete, including the identification of rebar, corrosion products, pores, cement mortar, and corrosion-induced cracking. A galvanostatic accelerated corrosion test was conducted to simulate rebar corrosion within the mortar. The XCT test was performed, producing a series of XCT images that revealed the corrosion-induced cracks in the concrete. This set of images formed a comprehensive training database for the U-Net model. The performance of the model was significantly improved by data augmentation within the training dataset, thereby enabling it to recognize corrosion-induced cracking and pores. The U-Net model proved effective in accurately segmenting the XCT images into different phases, demonstrating a high accuracy rate of 99.18% on the validation dataset. A mean Intersection over Union (IoU) of 82.1% and a mean F1 score of 89.1% are observed on the test dataset, showing superior segmentation performance on XCT images. This significant advancement has automated the process of visualizing and 3D reconstructing the steel corrosion process inside the concrete. Concurrently, it has enabled the quantification of the volumetric expansion coefficient of corrosion products. The findings reveal spatial and temporal variations in this coefficient due to the migration of the corrosion product through the cracks.

AI to Train AI: Using ChatGPT to Improve the Accuracy of a Therapeutic Dialogue System

In this work, we present the use of one artificial intelligence (AI) application (ChatGPT) to train another AI-based application. As the latter one, we show a dialogue system named Terabot, which was used in the therapy of psychiatric patients. Our study was motivated by the fact that for such a domain-specific system, it was difficult to acquire large real-life data samples to increase the training database: this would require recruiting more patients, which is both time-consuming and costly. To address this gap, we have employed a neural large language model: ChatGPT version 3.5, to generate data solely for training our dialogue system. During initial experiments, we identified intents that were most often misrecognized. Next, we fed ChatGPT with a series of prompts, which triggered the language model to generate numerous additional training entries, e.g., alternatives to the phrases that had been collected during initial experiments with healthy users. This way, we have enlarged the training dataset by 112%. In our case study, for testing, we used 2802 speech recordings originating from 32 psychiatric patients. As an evaluation metric, we used the accuracy of intent recognition. The speech samples were converted into text using automatic speech recognition (ASR). The analysis showed that the patients’ speech challenged the ASR module significantly, resulting in deteriorated speech recognition and, consequently, low accuracy of intent recognition. However, thanks to the augmentation of the training data with ChatGPT-generated data, the intent recognition accuracy increased by 13% relatively, reaching 86% in total. We also emulated the case of an error-free ASR and showed the impact of ASR misrecognitions on the intent recognition accuracy. Our study showcased the potential of using generative language models to develop other AI-based tools, such as dialogue systems.

Development of AI-Based Tools for Power Generation Prediction

This study presents a model for predicting photovoltaic power generation based on meteorological, temporal and geographical variables, without using irradiance values, which have traditionally posed challenges and difficulties for accurate predictions. Validation methods and evaluation metrics are used to analyse four different approaches that vary in the distribution of the training and test database, and whether or not location-independent modelling is performed. The coefficient of determination, R2, is used to measure the proportion of variation in photovoltaic power generation that can be explained by the model’s variables, while gCO2eq represents the amount of CO2 emissions equivalent to each unit of power generation. Both are used to compare model performance and environmental impact. The results show significant differences between the locations, with substantial improvements in some cases, while in others improvements are limited. The importance of customising the predictive model for each specific location is emphasised. Furthermore, it is concluded that environmental impact studies in model production are an additional step towards the creation of more sustainable and efficient models. Likewise, this research considers both the accuracy of solar energy predictions and the environmental impact of the computational resources used in the process, thereby promoting the responsible and sustainable progress of data science.

Precise strength prediction of endogenous promoters from Escherichia coli and J-series promoters by artificial intelligence

BackgroundPromoter strength plays a critical role in modulating protein expression in genetic engineering. However, there are only a few studies on the strength of promoters from the comprehensive genomic database of sigma factors. To circumvent the time and resource-intensive experimental approach, artificial intelligence (AI) is considered to construct a complete database of proposed promoters from Escherichia coli, and further utilizing prediction algorithms to evaluate the promoter strength and confirmed using intensity of green fluorescent protein (GFP). MethodsThe promoter database was constructed using partial information from Ecocyc, and predictive strength of the promoters was calculated via the phiSITE hunter tool. Among the 1744 promoter entries in the database were derived from E. coli MG1655, while total of 935 sigma factor 70 (σ70) promoters were identified. Then, the training database was applied to develop a precise tool for predicting promoter strength using machine learning and six deep learning models. The accuracy of predictions was confirmed through wet experiments conducted on endogenous and J-series promoters. Significant findingsBy employing a deep learning model, particularly the Convolutional Neural Network (CNN), the promoter prediction fitness of phiSITE, which relied on traditional alignment metrics, was approved. On the other hand, phiSITE demonstrated satisfied result in the fluorescence experiments using 7 endogenous promoters, achieving an R-squared (R2) at 0.93. When applied the same model to predict the strength of J-series promoters, the best R2 achieved 0.99. Thus, CNN model represents as an effective evaluation of AI-based promoter strength.

Revolutionising food advertising monitoring: a machine learning-based method for automated classification of food videos.

Food advertising is an important determinant of unhealthy eating. However, analysing a large number of advertisements (ads) to distinguish between food and non-food content is a challenging task. This study aims to develop a machine learning-based method to automatically identify and classify food and non-food ad videos. Methodological study to develop an algorithm model that prioritises both accuracy and efficiency in monitoring and classifying advertising videos. From a collection of Brazilian television (TV) ads data, we created a database and split it into three sub-databases (i.e. training, validation and test) by extracting frames from ads. Subsequently, the training database was classified using the EfficientNet neural network. The best models and data-balancing strategies were investigated using the validation database. Finally, the test database was used to apply the best model and strategy, and results were verified with field experts. The study used 2124 recorded Brazilian TV programming hours from 2018 to 2020. It included 703 food ads and over 20 000 non-food ads, following the protocol developed by the INFORMAS network for monitoring food marketing on TV. The results showed that the EfficientNet neural network associated with the balanced batches strategy achieved an overall accuracy of 90·5 % on the test database, which represents a reduction of 99·9 % of the time spent on identifying and classifying ads. The method studied represents a promising approach for differentiating food and non-food-related video within monitoring food marketing, which has significant practical implications for researchers, public health policymakers, and regulatory bodies.

Automatic segmentation of atrial fibrillation and flutter in single-lead electrocardiograms by self-supervised learning and Transformer architecture.

Automatic detection of atrial fibrillation and flutter (AF/AFL) is a significant concern in preventing stroke and mitigating hemodynamic instability. Herein, we developed a Transformer-based deep learning model for AF/AFL segmentation in single-lead electrocardiograms (ECGs) by self-supervised learning with masked signal modeling (MSM). We retrieved data from 11 open-source databases on PhysioNet; 7 of these databases included labeled ECGs, while the other 4 were without labels. Each database contained ECG recordings with durations of ≥30s. A total of 24 intradialytic ECGs with paroxysmal AF/AFL during 4h of hemodialysis sessions at Seoul National University Hospital were used for external validation. The model was pretrained by predicting masked areas of ECG signals and fine-tuned by predicting AF/AFL areas. Cross-database validation was used for evaluation, and the intersection over union (IOU) was used as a main performance metric in external database validation. In the 7 labeled databases, the areas marked as AF/AFL constituted 41.1% of the total ECG signals, ranging from 0.19% to 51.31%. In the evaluation per ECG segment, the model achieved IOU values of 0.9254 and 0.9477 for AF/AFL segmentation and other segmentation tasks, respectively. When applied to intradialytic ECGs with paroxysmal AF/AFL, the IOUs for the segmentation of AF/AFL and non-AF/AFL were 0.9896 and 0.9650, respectively. Model performance by different training procedure indicated that pretraining with MSM and the application of an appropriate masking ratio both contributed to the model performance. It also showed higher IOUs of AF/AFL labels than in previous studies when training and test databases were matched. The present model with self-supervised learning by MSM performs robustly in segmenting AF/AFL.

Digital geotechnics: from data-driven site characterisation towards digital transformation and intelligence in geotechnical engineering

ABSTRACT Geotechnical engineering is experiencing a paradigm shift towards digital transformation and intelligence, driven by Industry 4.0 and emerging digital technologies, such as machine learning. However, development and application of machine learning are relatively slow in geotechnical practice, because extensive training databases are a key to the success of machine learning, but geotechnical data are often small and ugly, leading to the difficulty in developing a suitable training database required for machine learning. In addition, convincing examples from real projects are rare that demonstrate the immediate added value of machine learning to geotechnical practices. To facilitate digital transformation and machine learning in geotechnical engineering, this study proposes to develop a project-specific training database that leverages on digital transformation of geotechnical workflow and reflects both project-specific data collected from various stages of the geotechnical workflow and domain knowledge in geotechnical practices, such as soil mechanics, numerical analysis principles, and prior engineering experience and judgment. A real ground improvement project is presented to illustrate the proposed method and demonstrate the added value of digital transformation and machine learning in geotechnical practices.

Automated adaptive chemistry for Large Eddy Simulations of turbulent reacting flows

Large Eddy Simulations (LES) of turbulent reacting flows carried out with detailed kinetic mechanisms have a key role for the discovery of the physical and chemical processes occurring in combustion systems, and are essential for the development of efficient, stable, and non-pollutant technologies. Nevertheless, these simulations require a large amount of computational resources, making their utilization for large-scale systems, such as industrial burners and gas turbines, impractical. In this work, we combine state-of-the-art machine learning algorithms and model reduction methods to deliver a fully automated strategy for performing LES with adaptive chemistry. This strategy is based on the Sample-Partitioning Adaptive Chemistry (SPARC) algorithmic procedure, which consists of four steps: the generation of a training dataset, its partitioning in clusters, the generation of a set of reduced chemical mechanisms specifically tailored to each cluster and, lastly, the numerical simulation of the case of interest with adaptive chemistry enabled by an on-the-fly classification of every grid point.The SPARC approach has already been demonstrated to substantially reduce the computational effort of reactive flows simulations. However a non-negligible level of user interventions is needed, upon which the method’s success critically depend. Therefore, with the goal of boosting the performance of this workflow and minimise the user-specified degrees of freedom, we plug in and exploit the Local Principal Component Analysis augmented with an automated Bayesian-optimised search for optimal clustering solutions, and the Computational Singular Perturbation method with an additional layer of automation based on the Tangential Stretching Rate for minimally-sized reduced mechanisms. We employ a cheap and easy-to-generate 1-dimensional-flames training database and we demonstrate the efficiency, accuracy and robustness of this strategy with an application to LES of the Adelaide Jet in Hot Coflow (AJHC) burner, a turbulent reacting flow exhibiting intense turbulence-chemistry interactions.

White dwarf Random Forest classification through Gaia spectral coefficients

Context. The third data release of Gaia has provided approximately 220 million low resolution spectra. Among these, about 100 000 correspond to white dwarfs. The magnitude of this quantity of data precludes the possibility of performing spectral analysis and type determination by human inspection. In order to tackle this issue, we explore the possibility of utilising a machine learning approach, based on a Random Forest algorithm. Aims. Our goal is to analyse the viability of the Random Forest algorithm for the spectral classification of the white dwarf population within 100 pc from the Sun, based on the Hermite coefficients of Gaia spectra. Methods. We utilised the assigned spectral type from the Montreal White Dwarf Database for training and testing our Random Forest algorithm. Once validated, our algorithm model was applied to the rest of the unclassified white dwarfs within 100 pc. First, we started by classifying the two major spectral type groups of white dwarfs: hydrogen-rich (DA) and hydrogen-deficient (non-DA). Next, we explored the possibility of classifying the various spectral subtypes, including the secondary spectral types in some cases. Results. Our Random Forest classification presented a very high recall (>80%) for DA and DB white dwarfs, and a very high precision (>90%) for DB, DQ, and DZ white dwarfs. As a result we have assigned a spectral type to 9446 previously unclassified white dwarfs: 4739 DAs, 76 DBs (60 of them DBAs), 4437 DCs, 132 DZs, and 62 DQs (nine of them DQpec). Conclusions. Despite the low resolution of Gaia spectra, the Random Forest algorithm applied to the Gaia spectral coefficients proves to be a highly valuable tool for spectral classification.

Pre-Fault Dynamic Security Assessment of Power Systems for Multiple Different Faults via Multi-Label Learning

With the advancement of machine learning techniques, data-driven power system dynamic security assessment (DSA) has received great research interests. Traditional methods usually apply one DSA model for one specific fault and cannot simultaneously address different multiple faults by one model. To solve this issue and further improve the DSA accuracy performance, this paper proposes a novel DSA method based on multi-label learning (MLL) with a training database for sufficient and incomplete/uneven coverage labels scenarios. By considering fault correlation between different faults, the proposed MLL-based DSA method can handle multiple faults simultaneously with the high stability assessment accuracy performance. Moreover, this paper provides the detailed optimization process and tight mathematical proof for the proposed MLL-based DSA method. Comprehensive simulation tests and comparisons were conducted on the New England 10-machine 39-bus testing system and the synthetic Illinois 49-machine 200-bus testing system.

Neural Network-Enhanced Fault Diagnosis of Robot Joints

Industrial robots play an indispensable role in flexible production lines, and the faults caused by degradation of equipment, motors, mechanical system joints, and even task diversity affect the efficiency of production lines and product quality. Aiming to achieve high-precision multiple size of fault diagnosis of robotic arms, this study presents a back propagation (BP) multiclassification neural network-based method for robotic arm fault diagnosis by taking feature fusion of position, attitude and acceleration of UR10 robotic arm end-effector to establish the database for neural network training. The new algorithm achieves an accuracy above 95% for fault diagnosis of each joint, and a diagnostic accuracy of up to 0.1 degree. It should be noted that the fault diagnosis algorithm can detect faults effectively in time, while avoiding complex mathematical operations.

XCNN-SC: Explainable CNN for SARS-CoV-2 variants classification and mutation detection

The COVID-19 pandemic spread rapidly all over the world in 2019, causing the deaths of millions of people. One of the main challenges in controlling the pandemic was the high rate of virus mutation, which evaded the immune system and reduced the vaccine's effectiveness. Due to the differences in symptoms, transmission and mortality rate, and effective prevention strategies of each variant, identifying and classifying different variants is a great necessity. In the present study, 82802 whole genomes of Alpha, Beta, Delta, Eta, Epsilon, Lota, and Omicron variants of SARS-CoV-2 were obtained from the NCBI database. The label encoding method was applied to convert the genomic sequences to numeric sequences. Then, a single-layer 1D-CNN model was used to classify seven variants of SARS-CoV-2 and achieved 99.78% accuracy, 97.98% precision, 97.66% sensitivity, 99.95% specificity, and 98.68% F1-score. The max pool layer feature maps of this network were investigated to discover the sequence discriminative regions in SARS-CoV-2 variants and their effect on biological functional differences. The feature map examination led to the detection of three mutations; one of them was a missense mutation, which has been reported previously. The rest of the detected mutations were silent mutations. In conclusion, the achieved results indicated that CNN models can extract effective features to discriminate several SARS-CoV-2 variants. Also, investigation of CNN feature maps led to an explainable CNN revealing the learned knowledge of the network from the training database. This knowledge helped us to detect mutations and their functional effects in different variants.

Detection and Mapping of Chestnut Using Deep Learning from High-Resolution UAV-Based RGB Imagery

The semantic segmentation method based on high-resolution RGB images obtained by unmanned aerial vehicle (UAV) provides a cost-effective way to improve the accuracy of detection and classification in forestry. Few studies have explored the impact of sample distribution patterns on deep learning model detection accuracy. The study was carried out using the data from the 4.78 km2 RGB image of a chestnut (Castanea mollissima Blume) plantation obtained by the DJI Phantom 4-RTK, and the model training was conducted with 18,144 samples of manually delineated chestnut tree clusters. The performance of four semantic segmentation models (U-Net, DeepLab V3, PSPNet, and DeepLab V3+) paired with backbones (ResNet-34, ResNet-50) was evaluated. Then, the influence of chestnut data from different planting patterns on the accuracy and generalization performance of deep learning models was examined. The results showed that the combination of DeepLab V3 with ResNet-34 backbone gives the best performance (F1 score = 86.41%), while the combination of DeepLab V3+ with ResNet-50 backbone performed the worst. The influence of different backbone networks on the detection performance of semantic segmentation models did not show a clear pattern. Additionally, different spatial distribution patterns of chestnut planting affected the classification accuracy. The model MIX, trained on comprehensive training data, achieves higher classification accuracies (F1 score = 86.13%) compared to the model trained on single training data (F1 score (DP) = 82.46%; F1 score (SP) = 83.81%). The model performance in complex scenario data training is superior to that of the model in simple scene data training. In conclusion, comprehensive training databases can improve the generalization performance of chestnut classification with different spatial distribution patterns. This study provides an effective method for detecting chestnut cover area based on semantic segmentation, allowing for better quantitative evaluation of its resource utilization and further development of inventories for other tree species.

Rapid discovery of promising materials via active learning with multi-objective optimization

Developing efficient methods to find materials that satisfy multiple properties simultaneously is an important task so that the material screening process can be accelerated with reduced resources. In this study, the performance of active learning based on multi-objective Bayesian optimization is demonstrated for two well-known inorganic material databases. First, this method is applied to two-dimensional materials to optimize their electronic and mechanical properties. This indicates that the hypervolume-based method is more effective than the other methods of exploitation, exploration, and random selection. All optimal Pareto front (PF) can be found within sampling only from 16% to 23% of the entire search space using the expected hypervolume improvement (EHVI). To ensure the generality of the developed platform, we implemented this method on a general inorganic material database. Again, the EHVI performed better than the other methods in all cases. In the most data-deficient case, where the initially known data comprise 0.1% of the total data, 61% of the data in the entire search space are required to find the optimal PF. This is 36%p less than that of the random selection or exploitation methods. However, exploitation and exploration methods do not always perform worse than the EHVI. These methods are more effective than the EHVI in searching for more than a certain number of PFs. The results clearly demonstrate the applicability and efficacy of the multi-objective active learning platform. We believe that even with a small number of initial training databases, this method can accelerate the material-screening process with significantly reduced resources.

Flow Prediction for Hydropower Generation using LMST Neural Networks

The objective of this investigation is to predict the necessary flow to satisfy the demand in order to guarantee the generation of energy in a hydropower, for which the artificial neural network LMST (Long Short-Term Memory) was used. In this context, in this article, historical data from the years 2010 to 2019 of flow and power of the plant was used, which were provided by it, in the first instance the data was purified to group them and graph the heat diagram with which the variables of interest were determined, then the training was carried out with which the network was structured with 10 neurons, 2 hidden layers, 2 freezing layers and 1 output layer. In addition, the RMSprop optimizer was selected with 10 numbers of delays with which an absolute error of 5.12% was obtained, finally tests were made comparing the results obtained in the training with data from one day after those that were used in the training and testing database. It is worth mentioning that by predicting the flow for the production of electrical energy, it contributes to the National Development Plan “Toda una Vida” of Ecuador, which is why it focuses on optimizing the use of water resources for required production levels.

Sistem Absensi Cerdas Menggunakan Open CV Berbasis Pengenalan Wajah

Abstract
 
 The face is an important part of the human body uniquely identify a person. Using facial characteristics as biometrics, facial recognition systems can be implemented. That the most demanding task in any organization is marking attendance. In the traditional attendance system, students are summoned byteachers and their presence or absence marked accordingly. However, this traditional technique is time consuming and boring. In this project, Open CV is based on face recognition approach has been proposed. This model integrates a camera captures the input image, an algorithm for detecting the face of a input images, encoding and identifying faces, tagging presence in a spreadsheet and convert it to a PDF file. That the training database is created by training the system with faces from authorized students. The cropped image is then saved as a database with their respective labels. The features are extracted using the LBPH algorithm.
 
 Keywords: LBPH, , Attendance, Biometric, Face Recognition