Real Training Data Research Articles

Electric vehicle battery state-of-charge estimation based on optimized deep learning strategy with varying temperature at different C Rate

The estimation of the state-of-charge (SOC) in battery technology is a vital task for the battery management system (BMS). In this study, a modeling framework is presented for SOC estimation that makes use of methodologies that involve machine learning (ML). Aside from providing the user with real-time feedback on the battery's remaining capacity, precise knowledge of SOC imposes further control over the charging/discharging process, hence elongating the battery's useful lifespan. This decreases the possibility of over-voltage. As a result, the proposed work creates a foundation for the development of realistic training data at different temperature and different C rate. This analysis demonstrates how the discharge voltage curves change at various temperatures and various C Rate. Change in voltage, internal resistance and SOC is analyzed at various temperature and various C Rate. This paves the way for a more exact SOC estimation that is driven by data-driven methods. According to the findings of this study, the energy storage capacity is greater when the C rate of charging is lower, and it is possible to deliver more energy when the C rate of discharging is lower. According to the findings of this study, the discharge of the battery is linear for the SoC values that are less than 90%. The error in the SoC estimation is brought down to 0.835% by using this approach.

Virtual Footwear Try-on in Augmented Reality using Deep Learning Models

Abstract Customization is an increasing trend in fashion product industry to reflect individual lifestyles. Previous studies have examined the idea of virtual footwear try-on in augmented reality (AR) using a depth camera. However, the depth camera restricts the deployment of this technology in practice. This research proposes to estimate the 6-DoF pose of a human foot from a color image using deep learning models to solve the problem. We construct a training dataset consisting of synthetic and real foot images that are automatically annotated. Three convolutional neural network models (DOPE, DOPE2, and YOLO6d) are trained with the dataset to predict the foot pose in real-time. The model performances are evaluated using metrics for accuracy, computational efficiency, and training time. A prototyping system implementing the best model demonstrates the feasibility of virtual footwear try-on using a RGB camera. Test results also indicate the necessity of real training data to bridge the reality gap in estimating the human foot pose.

Generative adversarial network based synthetic data training model for lightweight convolutional neural networks.

Inadequate training data is a significant challenge for deep learning techniques, particularly in applications where data is difficult to get, and publicly available datasets are uncommon owing to ethical and privacy concerns. Various approaches, such as data augmentation and transfer learning, are employed to address this problem, which help to some extent in removing this limitation. However, after a certain amount of data augmentation, the quality of the generated data stalls, and transfer learning suffers from the issue of negative transfer. This paper proposes a novel generative adversarial network-based synthetic data training (GAN-ST) model to generate synthetic data for training a lightweight convolutional neural network (CNN). An enhanced generator is proposed to quickly saturate and cover the colour space of the training distribution. The GAN-ST model is based on Deep Convolutional Generative Adversarial Network(s) (DCGAN) and Conditional Generative Adversarial Network(s) (CGAN) models, which consist of an enhanced generator. The study evaluates the accuracy of a CNN model on the MNIST and CIFAR 10 datasets using both original and synthetic data. The results revealed an impressive classifier accuracy on the MNIST dataset, achieving an accuracy of 99.38% on GAN-ST-generated synthetic training data, which is only 0.05% lower than the performance on original data-based training. The classifier performance on the CIFAR dataset is also remarkable, achieving an accuracy of 90.23%. The performance of CNN trained using GAN-ST-based synthetic data is notable, with the most considerable improvement of 0.66% and 7.06%, over a single GAN-based synthetic data training for the MNIST and CIFAR datasets, respectively. By training two GANs independently, the GAN-ST model covers different parts of the original data distribution, resulting in a more diverse and realistic training data set for the classifier. This diverse set of synthetic data, when used to train a CNN, shows better generalization to new data, leading to improved classification accuracy.

Using Synthetic Tree Data in Deep Learning-Based Tree Segmentation Using LiDAR Point Clouds

Deep learning, neural networks and other data-driven processing techniques are increasingly used in the analysis of LiDAR point cloud data in forest environments due to the benefits offered in accuracy and adaptability to new environments. One of the downsides of these techniques in practical applications is the requirement for manually annotated data necessary for training neural networks, which can be time consuming and costly to attain. We develop an approach to training neural networks for forest tree stem segmentation from point clouds that uses synthetic data from a custom tree simulator, which can generate large quantities of training examples without manual human effort. Our tree simulator captures the geometric characteristics of tree stems and foliage, from which automatically-labelled synthetic point clouds can be generated for training a semantic segmentation algorithm based on the PointNet++ architecture. Using evaluations on real aerial and terrestrial LiDAR point clouds from a range of different forest sites, we demonstrate our synthetic data-trained models can out-perform, or provide comparable performance with models trained on real data from other sites or when available real training data is limited (increases in IoU from 1–7%). Our simulation code is open-source and made available to the research community.

GeoSynth: A Photorealistic Synthetic Indoor Dataset for Scene Understanding.

Deep learning has revolutionized many scene perception tasks over the past decade. Some of these improvements can be attributed to the development of large labeled datasets. The creation of such datasets can be an expensive, time-consuming, and imperfect process. To address these issues, we introduce GeoSynth, a diverse photorealistic synthetic dataset for indoor scene understanding tasks. Each GeoSynth exemplar contains rich labels including segmentation, geometry, camera parameters, surface material, lighting, and more. We demonstrate that supplementing real training data with GeoSynth can significantly improve network performance on perception tasks, like semantic segmentation. A subset of our dataset will be made publicly available at https://github.com/geomagical/GeoSynth.

Self-Supervised Category-Level 6D Object Pose Estimation With Optical Flow Consistency

Category-level 6D object pose estimation aims at determining the pose of an object of a given category. Most current state-of-the-art methods require a significant amount of real training data to supervise their models. Moreover, annotating the 6D pose is very time consuming, error-prone, and it does not scale well to a large amount of object classes. Therefore, a handful of methods have recently been proposed to use unlabelled data to establish weak supervision. In this letter we propose a self-supervised method that leverages the 2D optical flow as a proxy for supervising the 6D pose. To this purpose, we estimate the 2D optical flow between consecutive frames based on the pose estimation. Then, we harness an off-the-shelf optical flow method to enable weak supervision using a 2D-3D optical flow based consistency loss. Experiments show that our approach for self-supervised learning yields state-of-the-art performance on the NOCS benchmark, and it reaches comparable results with some fully-supervised approaches.

Redefining the digital triplet for surrogate system integration

Predictive Maintenance (PdM) requires methods and tools to convey process information to maintenance planners allowing for data-driven repair decisions. Cyber-Physical Systems (CPS) and Digital Twins (DT) are current tools that transform data for informed decision making; however, the successful deployment of these tools is hampered by missing or low levels of training data for machine specific events such as failure. This paper proposes a standardized framework for adapting data from offline environments to train online systems without real world failure training data. This novel process, the Surrogate Digital Triplet (SDTr) framework, incorporates a third system, the surrogate triplet, to transfer data between the lab (offline) and production (online) environment. SDTr standardizes the data, information, and knowledge interfaces between systems to pass offline learning to the real world in a traceable manner.

Object Detector Differences when Using Synthetic and Real Training Data

To train well-performing generalizing neural networks, sufficiently large and diverse datasets are needed. Collecting data while adhering to privacy legislation becomes increasingly difficult and annotating these large datasets is both a resource-heavy and time-consuming task. An approach to overcome these difficulties is to use synthetic data since it is inherently scalable and can be automatically annotated. However, how training on synthetic data affects the layers of a neural network is still unclear. In this paper, we train the YOLOv3 object detector on real and synthetic images from city environments. We perform a similarity analysis using Centered Kernel Alignment (CKA) to explore the effects of training on synthetic data on a layer-wise basis. The analysis captures the architecture of the detector while showing both different and similar patterns between different models. With this similarity analysis, we want to give insights on how training synthetic data affects each layer and to give a better understanding of the inner workings of complex neural networks. The results show that the largest similarity between a detector trained on real data and a detector trained on synthetic data was in the early layers, and the largest difference was in the head part. The results also show that no major difference in performance or similarity could be seen between frozen and unfrozen backbone.

An optimized patch-point based approach for seismic fault interpretation using CNN

The interpretation of fault is essential for the oil and gas industries. This paper proposes an optimized patch-point-based approach for interpreting faults in a seismic data set using a convolutional neural network (CNN). We extract small patches of data for training and identify the fault patches. Next, we separately train seismic data points that are previously labeled as fault or non-fault. The strategy is to apply patch classification followed by analyzing fault patchs’ points to get the fault's location. We consider a mixture of synthetic and real data for training and as well as for testing. This method has used only the seismic amplitude values and has not considered any seismic attribute. We do normalization and quantization of seismic data to act as input to the CNN network, and the results show good accuracy when applied to synthetic and real data.

Single Shot X-ray Speckle Tracking Phase Contrast Imaging with a Low Brilliance Lab Source

In this paper, we present a single-shot X-ray speckle tracking (XST) phase-contrast imaging (PCI) method for a non-coherent polychromatic laboratory source. XST typically requires a coherent X-ray source, such as a synchrotron, to produce speckle. For a laboratory source without sufficient coherence, the absorption pattern of a random mask, such as sandpaper, have been proposed. However, to track the very small phase shifts in the absorption speckle below the size of the pixels of a standard flat panel detector in the range of 100μm, hundreds of detector images are taken, each moved by a linear stage by a fraction of a pixel to determine the phase shift. As a result, both acquisition time and radiation dose are high. Moreover, precise linear stages are required for sub-pixel movement of the random mask during acquisition of the projection images. In contrast, in this paper we propose a neural network called PCINet to acquire the differential phase image with only one reference image without the sample and one sample image. The proposed system does not require precise linear stages because only a single acquisition is needed. Due to the incoherent source, an absorption-based pseudo-speckle is generated as a reference image. In order to increase the contrast of the pseudo-speckle, a novel, cheap and easy to manufacture tungsten random spatial distrubution of particles is proposed as an alternative to sandpaper. A method for generating realistic training data for the neural network using real X-ray images in a laboratory setup was developed. The training data of the neural network was generated as close as possible to the real data by constructing the input image pairs pixel by pixel from multiple real X-ray images of the pseudo-speckle. The experimental resukts show that it is possible to obtain the differential phase image with only one reference image and one sample image with superimposed absorption speckles in the case of an incoherent laboratory setup.

Synthesizing 3D Gait Data with Personalized Walking Style and Appearance

Extracting gait biometrics from videos has been receiving rocketing attention given its applications, such as person re-identification. Although deep learning arises as a promising solution to improve the accuracy of most gait recognition algorithms, the lack of enough training data becomes a bottleneck. One of the solutions to address data deficiency is to generate synthetic data. However, gait data synthesis is particularly challenging as the inter-subject and intra-subject variations of walking style need to be carefully balanced. In this paper, we propose a complete 3D framework to synthesize unlimited, realistic, and diverse motion data. In addition to walking speed and lighting conditions, we emphasize two key factors: 3D gait motion style and character appearance. Benefiting from its 3D nature, our system can provide various gait-related data, such as accelerometer data and depth map, not limited to silhouettes. We conducted various experiments using the off-the-shelf gait recognition algorithm and draw the following conclusions: (1) the real-to-virtual gap can be closed when adding a small portion of real-world data to a synthetically trained recognizer; (2) the amount of real training data needed to train competitive gait recognition systems can be reduced significantly; (3) the rich variations in gait data are helpful for investigating algorithm performance under different conditions. The synthetic data generator, as well as all experiments, will be made publicly available.

A Data Driven Framework for QoE-Aware Intelligent EN-DC Activation

In emerging 5G networks, User Equipment camps traditionally on 4G network. Later, if the user requests a 5G service, it can simultaneously camp on 4G and 5G using EUTRAN New-Radio Dual-Connectivity (EN-DC) approach. In EN-DC, poor radio-conditions in either 4G or 5G network can be detrimental to user Quality-of-Experience (QoE). Although operators want to maximize EN-DC activation to fully utilize the 5G network, sub-optimal parameter configuration to turn on ENDC can compromise key-performance-indicators due to excessive radio-link-failures (RLFs) or voice-muting. While the need to maximize the EN-DC activation is obvious for maximizing the 5G network's utility, RLF and mute avoidance are vital to maintain the QoE requirements. To achieve aforementioned tradeoff, this paper presents the first solution to optimally configure the EN-DC activation parameters. We collect two datasets from real network to develop machine-learning-models to predict RLF and muting, respectively. We also investigate and compare the potential of various under-sampling, oversampling, and synthetic data generation techniques including Tomek-Links and Generative Adversarial Networks for their potential to address the data imbalance problem inherent in the real network training data. Leveraging these models, we formulate and solve two QoE-aware optimization problems that can maximize EN-DC activation while minimizing RLF or voice-muting. System-level simulation-based results show that compared to state-of-the-art solution that does not take into account RLF or voice-muting risk in EN-DC activation, the proposed solution can intelligently determine ENDC activation criteria that minimize the risk of RLF and voice muting while giving the operator's desired level of priority to maximize 5G network utilization.

Single Image Reflection Removal Based on Residual Attention Mechanism

Affected by shooting angle and light intensity, shooting through transparent media may cause light reflections in an image and influence picture quality, which has a negative effect on the research of computer vision tasks. In this paper, we propose a Residual Attention Based Reflection Removal Network (RABRRN) to tackle the issue of single image reflection removal. We hold that reflection removal is essentially an image separation problem sensitive to both spatial and channel features. Therefore, we integrate spatial attention and channel attention into the model to enhance spatial and channel feature representation. For a more feasible solution to solve the problem of gradient disappearance in the iterative training of deep neural networks, the attention module is combined with a residual network to design a residual attention module so that the performance of reflection removal can be ameliorated. In addition, we establish a reflection image dataset named the SCAU Reflection Image Dataset (SCAU-RID), providing sufficient real training data. The experimental results show that the proposed method achieves a PSNR of 23.787 dB and an SSIM value of 0.885 from four benchmark datasets. Compared with the other most advanced methods, our method has only 18.524M parameters, but it obtains the best results from test datasets.

Generating Paired Seismic Training Data with Cycle-Consistent Adversarial Networks

Deep-learning-based seismic data interpretation has received extensive attention and focus in recent years. Research has shown that training data play a key role in the process of intelligent seismic interpretation. At present, the main methods used to obtain training data are synthesizing seismic data and manually labeling the real data. However, synthetic data have certain feature differences from real data, and the manual labeling of data is time-consuming and subjective. These factors limit the application of deep learning algorithms in seismic data interpretation. To obtain realistic seismic training data, we propose label-to-data networks based on cycle-consistent adversarial networks in this work. These networks take random labels and unlabeled real seismic data as input and generate synthetic seismic data that match the random labels and have similar features to the real seismic data. Quantitative analysis of the generated data demonstrate the effectiveness of the proposed methods. Meanwhile, test results on different data indicate that the generated data are reliable and can be applied for seismic fault detection.

Style transfer-enabled Sim2Real framework for efficient learning of robotic transesophageal echocardiography using simulation

Recently, various machine learning techniques have been proposed to realize automatic ultrasound (US) image analysis for robotic US acquisition tasks. However, obtaining large amounts of real US images for training is usually expensive or even infeasible in some clinical applications, such as transesophageal echocardiography (TEE). An alternative is to build a simulator to generate synthetic US data for training, but the differences between simulated and real US images may result in poor model performance. This work presents a novel Sim2Real framework to learn robotic US image analysis based on simulated data to address the challenges of limited real US data for robotic TEE. A style transfer model is proposed based on unsupervised contrastive learning to convert real US images into the simulation style. Moreover, a task-relevant model is designed to combine CNNs with vision transformers and trained only on the simulated data to generate task-dependent predictions. We demonstrate the effectiveness of our method in an image regression task to predict the probe position based on US images in robotic TEE. Our results show that using only simulated US data and a small amount of unlabelled real data for training, our method can achieve comparable performance to semi-supervised and fully supervised learning methods. Moreover, the effectiveness of our previously proposed CT-based US image simulation method is also indirectly confirmed.

Training and Tuning of Neuro - Fuzzy Control Laws for the Machining of Prosthetics

Adaptive machining is a process of performing real-time parameter updates based on changes in external operating conditions. With the advancements being made in the area of customised implants, this study is conducted to determine the utilisation of Artificial Neural Network (ANN) within a Computer Numerical Control (CNN) manufacturing cell to perform real-time tool offset adjustment in the manufacture of prosthetic knees. This study integrates smart sensor technology in the CNC machining cell, enabling the acquisition of force data for a critical tool used in the machining of a Tibial component. The collected time series data is used to compare performance between a random forest classification algorithm and Bi-directional Long Term Short Memory (LSTM) neural networks. Pre-processing improvements are presented for the random forest algorithm using a time series data conversion step. In the case of the Bi-directional LSTM model, 2D time series data is converted to a 3D array using a novel projection technique. The result of both techniques is then converted to real world tool offset value by implementing a fuzzy logic controller. Results are presented with recommendations to the move from synthetic to real training data for future deployments.

SAR-TSCC: A Novel Approach for Long Time Series SAR Image Change Detection and Pattern Analysis

Change detection has played an increasingly important role in multi-temporal remote sensing applications recently. Long time series analysis is providing new information of land cover changes and improving the quality and accuracy of the change information being derived from remote sensing. The purpose of this study is to dig for more change temporal information and change pattern information from synthetic aperture radar (SAR) image time series (ITS), which is of great significance for monitoring urban area changes, conducting land-use surveys, and renovating illegal constructions. In the study, a novel unified framework for long Time Series SAR image Change detection and Change pattern mining (SAR-TSCC) was proposed for land cover change mapping. To obtain the most notable change time rapidly, a fast SAR ITS change point search method based on pruned exact linear time (SAR-PELT) algorithm was adopted. Meanwhile, the deep time series classification network, named SAR-TST, was implemented to recognize the change patterns, which is based on time series transformer (TST) architecture. Considering the lack of real training data, a novel synthetic data generation method is developed. The combination of the synthetic and real data enhanced the generalization of the classifiers. The proposed framework was used for monitoring a large urbanization area in the northwest of Hong Kong, China. The Cosmo Skymed (CSK) time series data acquired from 2013 to 2020 were exploited for land cover change analysis. Experiment results showed that our approach achieved state-of-the-art performance, as the time accuracy reached 86% and the classification accuracy on the four main change patterns (impulse, step, cycle, and complex) is over 99%. In particular, the proposed SAR-TST model showed remarkable advantages in the presence of insufficient real data.

Critical evaluation of the use of artificial data for machine learning based de novo peptide identification

Proteins are essential components of all living cells and so the study of their in situ expression, proteomics, has wide reaching applications. Peptide identification in proteomics typically relies on matching high resolution tandem mass spectra to a protein database but can also be performed de novo. While artificial spectra have been successfully incorporated into database search pipelines to increase peptide identification rates, little work has been done to investigate the utility of artificial spectra in the context of de novo peptide identification. Here, we perform a critical analysis of the use of artificial data for the training and evaluation of de novo peptide identification algorithms. First, we classify the different fragment ion types present in real spectra and then estimate the number of spurious matches using random peptides. We then categorise the different types of noise present in real spectra. Finally, we transfer this knowledge to artificial data and test the performance of a state-of-the-art de novo peptide identification algorithm trained using artificial spectra with and without relevant noise addition. Noise supplementation increased artificial training data performance from 30% to 77% of real training data peptide recall. While real data performance was not fully replicated, this work provides the first steps towards an artificial spectrum framework for the training and evaluation of de novo peptide identification algorithms. Further enhanced artificial spectra may allow for more in depth analysis of de novo algorithms as well as alleviating the reliance on database searches for training data.

Continuous Learning Method of Radar HRRP Based on CVAE-GAN

To improve the catastrophic forgetting that existed in the high-resolution range profile (HRRP) target recognition model of offline training, this paper proposed a continuous learning method of radar HRRP based on conditional variational auto-encoding and generative adversarial network (CVAE-GAN). Firstly, this method generates data through CVAE to simulate the real training data, and replays it in the series of subsequent tasks. By using generators, the proposed method does not need to save the original HRRP data and can ensure the privacy of the training dataset. Secondly, the proposed method takes the categorical data label as a generating condition, solves the unbalanced categories of the generated data samples, and improves the accuracy of the radar HRRP target recognition. Finally, by combining the attention mechanism and GAN network of the transformer, the generative capacity of the CVAE generator is enhanced effectively. In this paper, the continuous learning method is verified by a validation framework with three task settings. The experimental results show that the proposed method has a better continuous learning ability and better engineering practicality in various tasks compared with the previous regularization methods which can guarantee the privacy of the training dataset.

Joint physics-based and data-driven time-lapse seismic inversion: Mitigating data scarcity

In carbon capture and sequestration, developing rapid and effective imaging techniques is crucial for real-time monitoring of the spatial and temporal dynamics of [Formula: see text] propagation during and after injection. With continuing improvements in computational power and data storage, data-driven techniques based on machine learning (ML) have been effectively applied to seismic inverse problems. In particular, ML helps alleviate the ill-posedness and high computational cost of full-waveform inversion (FWI). However, such data-driven inversion techniques require massive high-quality training data sets to ensure prediction accuracy, which hinders their application to time-lapse monitoring of [Formula: see text] sequestration. We develop an efficient “hybrid” time-lapse workflow that combines physics-based FWI and data-driven ML inversion. The scarcity of the available training data is addressed by developing a new data-generation technique with physics constraints. The method is validated using a synthetic [Formula: see text]-sequestration model based on the Kimberlina storage reservoir in California. Our approach is shown to synthesize a large volume of high-quality, physically realistic training data, which is critically important in accurately characterizing the [Formula: see text] movement in the reservoir. The developed hybrid methodology can also simultaneously predict the variations in velocity and saturation and achieve high spatial resolution in the presence of realistic noise in the data.