Volume Of Training Data Research Articles

Towards Efficient Neural Decoder for Dexterous Finger Force Predictions.

Dexterous control of robot hands requires a robust neural-machine interface capable of accurately decoding multiple finger movements. Existing studies primarily focus on single-finger movement or rely heavily on multi-finger data for decoder training, which requires large datasets and high computation demand. In this study, we investigated the feasibility of using limited single-finger surface electromyogram (sEMG) data to train a neural decoder capable of predicting the forces of unseen multi-finger combinations. We developed a deep forest-based neural decoder to concurrently predict the extension and flexion forces of three fingers (index, middle, and ring-pinky). We trained the model using varying amounts of high-density EMG data in a limited condition (i.e., single-finger data). We showed that the deep forest decoder could achieve consistently commendable performance with 7.0% of force prediction errors and R2 value of 0.874, significantly surpassing the conventional EMG amplitude method and convolutional neural network approach. However, the deep forest decoder accuracy degraded when a smaller amount of data was used for training and when the testing data became noisy. The deep forest decoder shows accurate performance in multi-finger force prediction tasks. The efficiency aspect of the deep forest lies in the short training time and small volume of training data, which are two critical factors in current neural decoding applications. This study offers insights into efficient and accurate neural decoder training for advanced robotic hand control, which has the potential for real-life applications during human-machine interactions.

Sample Size Considerations for Fine-Tuning Large Language Models for Named Entity Recognition Tasks: Methodological Study.

Large language models (LLMs) have the potential to support promising new applications in health informatics. However, practical data on sample size considerations for fine-tuning LLMs to perform specific tasks in biomedical and health policy contexts are lacking. This study aims to evaluate sample size and sample selection techniques for fine-tuning LLMs to support improved named entity recognition (NER) for a custom data set of conflicts of interest disclosure statements. A random sample of 200 disclosure statements was prepared for annotation. All "PERSON" and "ORG" entities were identified by each of the 2 raters, and once appropriate agreement was established, the annotators independently annotated an additional 290 disclosure statements. From the 490 annotated documents, 2500 stratified random samples in different size ranges were drawn. The 2500 training set subsamples were used to fine-tune a selection of language models across 2 model architectures (Bidirectional Encoder Representations from Transformers [BERT] and Generative Pre-trained Transformer [GPT]) for improved NER, and multiple regression was used to assess the relationship between sample size (sentences), entity density (entities per sentence [EPS]), and trained model performance (F1-score). Additionally, single-predictor threshold regression models were used to evaluate the possibility of diminishing marginal returns from increased sample size or entity density. Fine-tuned models ranged in topline NER performance from F1-score=0.79 to F1-score=0.96 across architectures. Two-predictor multiple linear regression models were statistically significant with multiple R2 ranging from 0.6057 to 0.7896 (all P<.001). EPS and the number of sentences were significant predictors of F1-scores in all cases ( P<.001), except for the GPT-2_large model, where EPS was not a significant predictor (P=.184). Model thresholds indicate points of diminishing marginal return from increased training data set sample size measured by the number of sentences, with point estimates ranging from 439 sentences for RoBERTa_large to 527 sentences for GPT-2_large. Likewise, the threshold regression models indicate a diminishing marginal return for EPS with point estimates between 1.36 and 1.38. Relatively modest sample sizes can be used to fine-tune LLMs for NER tasks applied to biomedical text, and training data entity density should representatively approximate entity density in production data. Training data quality and a model architecture's intended use (text generation vs text processing or classification) may be as, or more, important as training data volume and model parameter size.

Cross-domain structural damage identification using transfer learning strategy

In damage identification of civil structures, training deep neural networks (DNNs) often requires a large volume of annotated training data. However, artificial damage to real-world structures is always forbidden, resulting in insufficient data for training. Transfer learning provides a solution that allows structural damage information in a data-rich domain to be transferred and shared as the prior knowledge to a data-scarce domain, thereby indirectly augmenting available training data in the latter domain. To this end, a multi-task transfer learning strategy has been adopted for the purpose of achieving cross-domain damage identification between a plane frame in the source domain and a three-dimensional frame in the target domain. The strategy can share the learning knowledge from the domain with sufficient training data to the target domain with insufficient data. Thereby, the DNN (named DNN#2) in the target domain can be trained on a small amount of training data. Owing to their ability in data feature extraction, stacked auto-encoders (SAEs) are used to construct the desirable DNN#1 and DNN#2 corresponding to different damage identification tasks, named Task#1 and Task#2, in the two domains. The two DNNs share the hidden layers in order to share damage feature information between the source and target domains. The training datasets of the two domains are first used to jointly pre-train the auto-encoders’ parameters in an unsupervised learning way. Afterwards, a supervised fine-tuning step is carried out to retraining the entire SAEs for better performance. By these means, Task#2 receives some prior knowledge from Task#1, and thus, it is accomplished on limited training data when Task#1 is synchronously achieved. The analysis results demonstrate that the proposed multi-task learning strategy requires only a single training session to simultaneously realize the cross-domain damage identification of two different steel frames.

Enforcing Traffic Safety: A Deep Learning Approach for Detecting Motorcyclists’ Helmet Violations Using YOLOv8 and Deep Convolutional Generative Adversarial Network-Generated Images

In this study, we introduce an innovative methodology for the detection of helmet usage violations among motorcyclists, integrating the YOLOv8 object detection algorithm with deep convolutional generative adversarial networks (DCGANs). The objective of this research is to enhance the precision of existing helmet violation detection techniques, which are typically reliant on manual inspection and susceptible to inaccuracies. The proposed methodology involves model training on an extensive dataset comprising both authentic and synthetic images, and demonstrates high accuracy in identifying helmet violations, including scenarios with multiple riders. Data augmentation, in conjunction with synthetic images produced by DCGANs, is utilized to expand the training data volume, particularly focusing on imbalanced classes, thereby facilitating superior model generalization to real-world circumstances. The stand-alone YOLOv8 model exhibited an F1 score of 0.91 for all classes at a confidence level of 0.617, whereas the DCGANs + YOLOv8 model demonstrated an F1 score of 0.96 for all classes at a reduced confidence level of 0.334. These findings highlight the potential of DCGANs in enhancing the accuracy of helmet rule violation detection, thus fostering safer motorcycling practices.

Hyperparameter tuning for deep learning semantic image segmentation of micro computed tomography scanned fiber-reinforced composites

Image segmentation with deep learning models has significantly improved the accuracy of the pixel-wise labeling of scientific imaging which is critical for many quantitative image analyses. This has been feasible through U-Net and related architecture convolutional neural network models. Although the adoption of these models has been widespread, their training data pool and hyperparameters have been mostly determined by educated guesses through trial and error. In this study, we present observations of how training data volume, data augmentation, and patch size affect deep learning performance within a limited data set. Here we study U-Net model training on four different samples of x-ray CT images of fiber-reinforced composites. Because the training process is not deterministic, we relied on seven-fold replication of each experimental condition to avoid under-sampling and observe model training variance. Unsurprisingly, we find greater training data volume strongly benefits individual models’ final accuracy and learning speed while depressing variance among replicates. Importantly, data augmentation has a profound benefit to model performance, especially in cases with a low abundance of ground truth, and we conclude that high coefficients of data augmentation should be used in scientific imaging semantic segmentation models. Future work to describe and measure image complexity is warranted and likely to ultimately guide researchers on the minimum required training data volume for particular scientific imaging deep learning tasks.

Comprehensive transferability assessment of short-term cross-building-energy prediction using deep adversarial network transfer learning

Data-driven models are widely used for building-energy predictions (BEPs). In practice, these models may fail when the available data on the target building is insufficient. Transfer learning (TL), where useful knowledge from information-rich source buildings with sufficient data is learned to enhance the prediction of information-poor buildings with data shortages, can address the problem of poor prediction performance caused by insufficient data. However, an important issue is finding suitable information-rich buildings to maximize the advantages of TL in cross- BEP. To address this issue, this study explored the selection of source buildings from three perspectives: building-energy data similarity between source and target buildings, building information characteristics, and the volume of training data. The impact of these three factors on the performance improvement of cross- BEP was assessed in a data-centric manner. Based on our previous studies, we selected a deep adversarial neural network (DANN) as the TL strategy for cross-BEP and systematically investigated the performance improvement and transferability of DANN from multiple perspectives of both post-hoc and ex-ante analysis. The Building Data Genome Project datasets were used for validation. Thirty-six buildings of six types and 180 source-target building pairs were considered. Our results demonstrated that DANN could effectively improve model performance by 40–90 % and 20%–80 % compared to non-optimized LSTM and parameter-optimized LSTM. When the same type and location source-target building pairs were only considered, the DTW index showed a relative strong negative linear correlation with the DANN prediction performance improvement, and the goodness of fitting is around 0.80. For building energy data within one year considered, DANN should be trained using no less than 6-month source domain data and no more than 4-week target domain data to improve transferability and reduce the cross-building energy prediction error.

Rapid 2D 23Na MRI of the calf using a denoising convolutional neural network

Purpose23Na MRI can be used to quantify in-vivo tissue sodium concentration (TSC), but the inherently low 23Na signal leads to long scan times and/or noisy or low-resolution images. Reconstruction algorithms such as compressed sensing (CS) have been proposed to mitigate low signal-to-noise ratio (SNR); although, these can result in unnatural images, suboptimal denoising and long processing times. Recently, machine learning has been increasingly used to denoise 1H MRI acquisitions; however, this approach typically requires large volumes of high-quality training data, which is not readily available for 23Na MRI. Here, we propose using 1H data to train a denoising convolutional neural network (CNN), which we subsequently demonstrate on prospective 23Na images of the calf. Methods1893 1H fat-saturated transverse slices of the knee from the open-source fastMRI dataset were used to train denoising CNNs for different levels of noise. Synthetic low SNR images were generated by adding gaussian noise to the high-quality 1H k-space data before reconstruction to create paired training data. For prospective testing, 23Na images of the calf were acquired in 10 healthy volunteers with a total of 150 averages over ten minutes, which were used as a reference throughout the study. From this data, images with fewer averages were retrospectively reconstructed using a non-uniform fast Fourier transform (NUFFT) as well as CS, with the NUFFT images subsequently denoised using the trained CNN. ResultsCNNs were successfully applied to 23Na images reconstructed with 50, 40 and 30 averages. Muscle and skin apparent TSC quantification from CNN-denoised images were equivalent to those from CS images, with <0.9 mM bias compared to reference values. Estimated SNR was significantly higher in CNN-denoised images compared to NUFFT, CS and reference images. Quantitative edge sharpness was equivalent for all images. For subjective image quality ranking, CNN-denoised images ranked equally best with reference images and significantly better than NUFFT and CS images. ConclusionDenoising CNNs trained on 1H data can be successfully applied to 23Na images of the calf; thus, allowing scan time to be reduced from ten minutes to two minutes with little impact on image quality or apparent TSC quantification accuracy.

Weight Averaging and re-adjustment ensemble for QRCD

Question Answering (QA) is a prominent task in the field of Natural Language Processing (NLP) with extensive applications. Recently, there has been a notable surge in research interest concerning the development of QA systems for the Holy Qur’an, an Islamic religious text. The Qur’an Reading Comprehension Dataset (QRCD) Malhas and Elsayed (2020) is a highly commendable effort in this respect. It stands as the first benchmark dataset specifically designed for a set of directly answerable questions from the Qur’an. Each question in the dataset is meticulously labeled with all potential answers sourced from the Holy Qur’an. From our perspective, the main challenge in QRCD stems from the limited volume of training data it offers. As a solution we propose an innovative approach to build a Deep Neural Network (DNN) ensemble, centered around Ara-Electra model (Antoun et al., 2021), that we called Weight Averaging and Re-adjustment (WAR) model. The model is constructed by computing running averages of all model states that evolve during a single training session and ensuring that model weights are readjusted prior to each training epoch, in order to hold it back from over fitting the training data. The scheme results in a single standalone model that exhibits the benefits of multi-model ensembles. It is distinguished from other ensembles proposed for QRCD that accumulates outputs from multiple expert models and employs classic techniques like hard voting or score averaging on output probabilities to build unified results. Each expert model costs individual training time and compute resources. The WAR model outperforms existing systems with improved generalization over unseen data. It achieves F1, partial Reciprocal Rank (pRR), and exact-match (EM) scores of 0.567, 0.60 and 0.29 respectively, exceeding best reported QRCD scores by 3%, 1.5% and 0.69% respectively. Notably, we are comparing our results with the top scores from different models, highlighting our model’s consistent performance across all three metrics.

ADAPTIVE DOMAIN-SPECIFIC NAMED ENTITY RECOGNITION METHOD WITH LIMITED DATA

The ever-evolving volume of digital information requires the development of innovative search strategies aimed at obtaining the necessary data efficiently and economically feasible. The urgency of the problem is emphasized by the growing complexity of information landscapes and the need for fast data extraction methodologies. In the field of natural language processing, named entity recognition (NER) is an essential task for extracting useful information from unstructured text input for further classification into predefined categories. Nevertheless, conventional methods frequently encounter difficulties when confronted with a limited amount of labeled data, posing challenges in real-world scenarios where obtaining substantial annotated datasets is problematic or costly. In order to address the problem of domain-specific NER with limited data, this work investigates NER techniques that can overcome these constraints by continuously learning from newly collected information on pre-trained models. Several techniques are also used for making the greatest use of the limited labeled data, such as using active learning, exploiting unlabeled data, and integrating domain knowledge. Using domain-specific datasets with different levels of annotation scarcity, the fine-tuning process of pre-trained models, such as transformer-based models (TRF) and Toc2Vec (token-to-vector) models is investigated. The results show that, in general, expanding the volume of training data enhances most models' performance for NER, particularly for models with sufficient learning ability. Depending on the model architecture and the complexity of the entity label being learned, the effect of more data on the model's performance can change. After increasing the training data by 20%, the LT2V model shows the most balanced growth in accuracy overall by 11% recognizing 73% of entities and processing speed. Meanwhile, with consistent processing speed and the greatest F1-score, the Transformer-based model (TRF) shows promise for effective learning with less data, achieving 74% successful prediction and a 7% increase in performance after expanding the training data to 81%. Our results pave the way for the creation of more resilient and efficient NER systems suited to specialized domains and further the field of domain-specific NER with sparse data. We also shed light on the relative merits of various NER models and training strategies, and offer perspectives for future research.

TiMix: Text-Aware Image Mixing for Effective Vision-Language Pre-training

Self-supervised Multi-modal Contrastive Learning (SMCL) remarkably advances modern Vision-Language Pre-training (VLP) models by aligning visual and linguistic modalities. Due to noises in web-harvested text-image pairs, however, scaling up training data volume in SMCL presents considerable obstacles in terms of computational cost and data inefficiency. To improve data efficiency in VLP, we propose Text-aware Image Mixing (TiMix), which integrates mix-based data augmentation techniques into SMCL, yielding significant performance improvements without significantly increasing computational overhead. We provide a theoretical analysis of TiMix from a mutual information (MI) perspective, showing that mixed data samples for cross-modal contrastive learning implicitly serve as a regularizer for the contrastive loss. The experimental results demonstrate that TiMix exhibits a comparable performance on downstream tasks, even with a reduced amount of training data and shorter training time, when benchmarked against existing methods. This work empirically and theoretically demonstrates the potential of data mixing for data-efficient and computationally viable VLP, benefiting broader VLP model adoption in practical scenarios. Our code is available on https://github.com/chaoyajiang/TiMiX/tree/main.

VIXEN: Visual Text Comparison Network for Image Difference Captioning

We present VIXEN - a technique that succinctly summarizes in text the visual differences between a pair of images in order to highlight any content manipulation present. Our proposed network linearly maps image features in a pairwise manner, constructing a soft prompt for a pretrained large language model. We address the challenge of low volume of training data and lack of manipulation variety in existing image difference captioning (IDC) datasets by training on synthetically manipulated images from the recent InstructPix2Pix dataset generated via prompt-to-prompt editing framework. We augment this dataset with change summaries produced via GPT-3. We show that VIXEN produces state-of-the-art, comprehensible difference captions for diverse image contents and edit types, offering a potential mitigation against misinformation disseminated via manipulated image content. Code and data are available at http://github.com/alexblck/vixen

Dynamic Industrial Optimization: A Framework Integrates Online Machine Learning for Processing Parameters Design

We develop the online process parameter design (OPPD) framework for efficiently handling streaming data collected from industrial automation equipment. This framework integrates online machine learning, concept drift detection and Bayesian optimization techniques. Initially, concept drift detection mitigates the impact of anomalous data on model updates. Data without concept drift are used for online model training and updating, enabling accurate predictions for the next processing cycle. Bayesian optimization is then employed for inverse optimization and process parameter design. Within OPPD, we introduce the online accelerated support vector regression (OASVR) algorithm for enhanced computational efficiency and model accuracy. OASVR simplifies support vector regression, boosting both speed and durability. Furthermore, we incorporate a dynamic window mechanism to regulate the training data volume for adapting to real-time demands posed by diverse online scenarios. Concept drift detection uses the EI-kMeans algorithm, and the Bayesian inverse design employs an upper confidence bound approach with an adaptive learning rate. Applied to single-crystal fabrication, the OPPD framework outperforms other models, with an RMSE of 0.12, meeting precision demands in production.

IDENTIFICATION OF ROOTKITS IN NETWORK TRAFFIC WITH USING THE BAGGING OF CLASSIFIERS

The paper proposes an approach to identify anomalies in network traffic based on the use of machine learning classifiers. The solution allows you to determine the resulting state class by averaging the votes of individual classifiers. The approach was evaluated on the NSL-KDD public dataset. A compar- ison of the performance of classifiers and their averaged evaluation using the Weka tool was performed. The NSL-KDD set has been optimized, with an emphasis on "rootkit" type attacks, as one of the most diffi- cult types of attacks to detect. Using the bagging-based approach implemented in the Weka application, it was possible to obtain accuracy results – 99.94%. During the experiment, a tendency of increasing accu- racy in the application of bagging on open data was revealed as the volume of training data increases. The proposed approach can be applied in the design of systems for detecting attacks and other abnormal states of information systems. The results of the accuracy of the average assessment require further re- search in order to improve the indicators. It is possible to modernize the approach of averaging the votes of classifiers by excluding/adding other classifiers, qualitative selection of attributes and their features, in- creasing the number of training samples for classification.

Fabric defect image generation method based on the dual-stage W-net generative adversarial network

Due to the intricate and diverse nature of textile defects, detecting them poses an exceptionally challenging task. In comparison with conventional defect detection methods, deep learning-based defect detection methods generally exhibit superior precision. However, utilizing deep learning for defect detection requires a substantial volume of training data, which can be particularly challenging to accumulate for textile flaws. To augment the fabric defect dataset and enhance fabric defect detection accuracy, we propose a fabric defect image generation method based on Pix2Pix generative adversarial network. This approach devises a novel dual-stage W-net generative adversarial network. By increasing the network depth, this model can effectively extract intricate textile image features, thereby enhancing its ability to expand information sharing capacity. The dual-stage W-net generative adversarial network allows generating desired defects on defect-free textile images. We conduct quality assessment of the generated fabric defect images resulting in peak signal-to-noise ratio and structural similarity values exceeding 30 and 0.930, respectively, and a learned perceptual image patch similarity value no greater than 0.085, demonstrating the effectiveness of fabric defect data augmentation. The effectiveness of dual-stage W-net generative adversarial network is established through multiple comparative experiments evaluating the generated images. By examining the detection performance before and after data augmentation, the results demonstrate that mean average precision improves by 6.13% and 14.57% on YOLO V5 and faster recurrent convolutional neural networks detection models, respectively.

Artificial Neural Networks for Long-Bone Fracture Detection Based on X-Ray Image Contour Features

Abstract: Two contour-based fracture detection strategies are presented in the ensuing research. The programs are designed to help medical practitioners identify human long bone fractures from X-ray pictures. The line-based fracture detection systems suggested in [1] served as the foundation for the creation of the contour-based fracture. Convolutional Neural Networks (CNN) are widely used in current Computer Aided Diagnosis (CAD) systems to classify broken X-ray images. Even while the current CAD systems achieve excellent accuracy, doing so comes at a cost: a large volume of training data. By using identified contours in X-ray pictures, the suggested techniques aim to achieve high classification accuracy with little training data. Two methods exist for contour-based fracture detection. The first is the enhanced CHFB scheme, while the second is the Standard Contour Histogram Feature-Based (CHFB). The modified CHFB scheme differs from the other one in that it removes the detected flesh contours surrounding the leg region.

Semantic representation and comparative analysis of physical activity sensor observations using MOX2-5 sensor in real and synthetic datasets: a proof-of-concept-study

The widespread use of devices like mobile phones and wearables allows for automatic monitoring of human daily activities, generating vast datasets that offer insights into long-term human behavior. A structured and controlled data collection process is essential to unlock the full potential of this information. While wearable sensors for physical activity monitoring have gained significant traction in healthcare, sports science, and fitness applications, securing diverse and comprehensive datasets for research and algorithm development poses a notable challenge. In this proof-of-concept study, we underscore the significance of semantic representation in enhancing data interoperability and facilitating advanced analytics for physical activity sensor observations. Our approach focuses on enhancing the usability of physical activity datasets by employing a medical-grade (CE certified) sensor to generate synthetic datasets. Additionally, we provide insights into ethical considerations related to synthetic datasets. The study conducts a comparative analysis between real and synthetic activity datasets, assessing their effectiveness in mitigating model bias and promoting fairness in predictive analysis. We have created an ontology for semantically representing observations from physical activity sensors and conducted predictive analysis on data collected using MOX2-5 activity sensors. Until now, there has been a lack of publicly available datasets for physical activity collected with MOX2-5 activity monitoring medical grade (CE certified) device. The MOX2-5 captures and transmits high-resolution data, including activity intensity, weight-bearing, sedentary, standing, low, moderate, and vigorous physical activity, as well as steps per minute. Our dataset consists of physical activity data collected from 16 adults (Male: 12; Female: 4) over a period of 30–45 days (approximately 1.5 months), yielding a relatively small volume of 539 records. To address this limitation, we employ various synthetic data generation methods, such as Gaussian Capula (GC), Conditional Tabular General Adversarial Network (CTGAN), and Tabular General Adversarial Network (TABGAN), to augment the dataset with synthetic data. For both the authentic and synthetic datasets, we have developed a Multilayer Perceptron (MLP) classification model for accurately classifying daily physical activity levels. The findings underscore the effectiveness of semantic ontology in semantic search, knowledge representation, data integration, reasoning, and capturing meaningful relationships between data. The analysis supports the hypothesis that the efficiency of predictive models improves as the volume of additional synthetic training data increases. Ontology and Generative AI hold the potential to expedite advancements in behavioral monitoring research. The data presented, encompassing both real MOX2-5 and its synthetic counterpart, serves as a valuable resource for developing robust methods in activity type classification. Furthermore, it opens avenues for exploration into research directions related to synthetic data, including model efficiency, detection of generated data, and considerations regarding data privacy.

Сегментация перекрывающихся изображений деревьев на цифровых снимках лесных массивов

The use of decision support systems based on computer vision and artificial intelligence significantly improves the working conditions for the operators of technological machines in the timber sector, whose work implies high intensity and psycho-emotional overload. By means of computer vision and artificial intelligence the operator can quickly and easily obtain the data on the state of the cutting area and adopt the optimal solution for holding the working operation. This facilitates his work and reduces the time spent searching and analyzing the data on the cutting area. Meanwhile, one of the key elements of such a system is a subsystem for automatic segmentation of objects in the photograph. We have explored the possibility of segmenting overlapping objects in the photographs of forest areas using a convolutional neural network based on the Mask R-CNN architecture. Unlike in most works on similar topics, the objects of this study are color photographs taken by an RGB camera rather than a lidar. This creates the prospect for reducing the cost of hardware and software systems used to support decision-making by the operators of logging machines. The images of the stems and crowns of coniferous and deciduous trees overlapping each other are the segmented objects under consideration. Using the GIMP graphic editor, we have manually marked the color photographs depicting a total of 134 trees of 4 different species: spruce, aspen, birch and pine. Utilizing the developed database, we have carried out an experiment to further train the Mask R-CNN convolutional neural network for segmentation of overlapping parts of the trees in the digital photographs of forest areas. The neural network has been pre-trained using the Microsoft COCO dataset containing more than 200,000 images of 80 different classes of objects such as people, cars, animals and various items. While training the neural network, the images supplied to its input were subjected to a series of various linear and nonlinear geometric transformations, which made it possible to increase the volume of training data by 11 times. As a result, the accuracy of segmentation of the images of the stems and crowns of coniferous and deciduous trees overlapping each other has reached 79 %, which allows the use of neural networks based on a similar architecture in decision support systems for logging machine operators.

A robust neural network model for fault detection in the presence of mislabelled data

AbstractSeveral data‐driven methodologies for process monitoring and detection of faults or abnormalities have been developed for the safety of processing systems. The effectiveness of data‐based models, however, is impacted by the volume and quality of training data. This work presents a robust neural network model for addressing the mislabelled and low‐quality data in detecting faults and process abnormalities. The approach is based on harnessing data quality features along with supervisory labels in the network training. The data quality has been computed using the Mahalanobis distances and trusted centres of each class of data such as normal and faulty data. The method has been examined for detecting abnormalities in two case studies; a continuous stirred tank heater problem for detecting leaks and the Tennessee Eastman chemical process for detecting step and sticking faults. The performance of the proposed robust artificial neural networks (ANN) model is evaluated in terms of accuracy, fault detection rate, false alarm rate, and classification index at varying extents of mislabelling, namely, 1%, 5%, and 10% mislabelled data. The proposed model demonstrates higher detection performance, especially at increased labels of mislabelled data where the performance of the conventional ANN is severely impacted. The proposed methodology can be advantageous in handling mislabelled and low‐quality data issues which are crucial in the data‐driven modelling of processing systems.

Особенности разработки полносвязных нейросетей для решения задачи оценивания липофильности органических соединений

The assessment of lipophilicity of small organic compounds plays a crucial role in the development and optimization of new drugs. Unfortunately, experimental methods require significant time and resources, including the use of laboratory equipment and reagents. Additionally, manual verification and data adjustment often increase the process's labor intensity. In contrast, computational methods like machine learning offer faster and less resource-intensive ways to assess lipophilicity, allowing for efficient processing of large data volumes and adaptation to complex relationships between molecular structure and lipophilicity. Developing neural network models for lipophilicity assessment is challenging due to insufficient experimental data and high computational costs with graph neural network models. This work presents an analysis of popular methods for describing chemical structures for building fully connected neural network models, less demanding in training data volume. Based on this analysis, features best describing organic compounds from an open lipophilicity dataset collected from the ChEMBL database are selected. The search for the optimal neural network model architecture for the chosen features is conducted.

Analysis of the relationship between training data volume and model quality for surrogate models in physical simulations

Typically, the quality of AI models is highly dependent on the amount and quality of the available training data. While for many applications of AI several million training datasets are available in accessible databases and can be easily extended, the generation of training data for surrogate models of physics simulations is computationally demanding. To identify the necessary amount of training data for a sufficient good AI model, we are evaluating the performance of a surrogate model for a thermomechanical production process with different sizes of artificially generated training data.