Categorical Cross-entropy Loss Research Articles

DeepRing: Convolution Neural Network based on Blockchain Technology

Background: This paper addresses specific challenges in predictive modeling, namely transparency issues, susceptibility to data manipulation, and fairness concerns. To overcome these obstacles, the study introduces DeepRing, approach that combines Convolutional Neural Networks (CNNs) and blockchain technology. Objective: DeepRing aims to enhance prediction integrity, data security, and fairness, thereby improving the ethical considerations, reliability, and accountability of predictive models. Methods: involves iterative training of a CNN model on five diverse datasets, including CIFAR-10, Fashion-MNIST, MNIST, CIFAR-100, and a Hands dataset. The CNN architecture incorporates Conv2D layers, MaxPooling2D layers, and Dense layers. Training metrics such as accuracy and sparse categorical cross-entropy loss are monitored, with the Adam optimizer employed. While achieving high accuracy on Plam (0.5300), MNIST (0.9978) and Fashion MNIST (0.9673), DeepRing exhibits moderate performance on CIFAR-10 (0.9296) and lower accuracy on CIFAR-100 (0.5973). Results: demonstrate the effectiveness of DeepRing in improving accuracy and enhancing model performance across various datasets. However, further development and validation are essential for successful model implementation, further development and validation are essential for successful model implementation. Conclusions: Introduces DeepRing as an innovative solution to address key challenges in predictive modeling, specifically focusing on transparency issues, susceptibility to data manipulation, and fairness concerns. By combining Convolutional Neural Networks (CNNs) with blockchain technology, DeepRing aims to elevate prediction integrity, enhance data security, and promote fairness, thereby contributing to the improvement of ethical considerations, reliability, and accountability in predictive modelling.

Deep learning model for detection of hotspots using infrared thermographic images of electrical installations

Hotspots in electrical power equipment or installations are a major issue whenever it occurs within the power system. Factors responsible for this phenomenon are many, sometimes inter-related and other times they are isolated. Electrical hotspots caused by poor connections are common. Deep learning models have become popular for diagnosing anomalies in physical and biological systems, by the instrumentality of feature extraction of images in convolutional neural networks. In this work, a VGG-16 deep neural network model is applied for identifying electrical hotspots by means of transfer learning. This model was achieved by first augmenting the acquired infrared thermographic images, using the pre-trained ImageNet weights of the VGG-16 algorithm with additional global average pooling in place of conventional fully connected layers and a softmax layer at the output. With the categorical cross-entropy loss function, the model was implemented using the Adam optimizer at learning rate of 0.0001 as well as some variants of the Adam optimization algorithm. On evaluation, with a test IRT image dataset, and a comparison with similar works, the research showed that a better accuracy of 99.98% in identification of electrical hotspots was achieved. The model shows good score in performance metrics like accuracy, precision, recall, and F1-score. The obtained results proved the potential of deep learning using computer vision parameters for infrared thermographic identification of electrical hotspots in power system installations. Also, there is need for careful selection of the IR sensor’s thermal range during image acquisition, and suitable choice of color palette would make for easy hotspot isolation, reduce the pixel to pixel temperature differential across any of the images, and easily highlight the critical region of interest with high pixel values. However, it makes edge detection difficult for human visual perception which computer vision-based deep learning model could overcome.

Failure Diagnosis for Dental Air Turbine Handpiece with Payload Using Feature Engineering and Temporal Convolution Network.

The internal mechanisms of dental air turbine handpieces (DATHs) have become increasingly intricate over time. To enhance the operational reliability of dental procedures and guarantee patient safety, this study formulated temporal convolution network (TCN) prediction models with the functions of causality in time sequence, transmitting memory, learning, storing, and fast convergence for monitoring the health and diagnosing the rotor and collet failure of DATHs. A handpiece mimicking a dentist's hand load of 100 g was employed to repeatedly mill a glass porcelain block back and forth for cutting. An accelerometer was employed to capture vibration signals during free-running of unrestrained operation of the handpiece, aiming to discern the characteristic features of these vibrations. These data were then utilized to create a diagnostic health classification (DHC) for further developing a TCN, a 1D convolutional neural network (CNN), and long short-term memory (LSTM) prediction models. The three frameworks were used and compared for machine learning to establish DHC prediction models for the DATH. The experimental results indicate that, in terms of DHC predicted for the experimental dataset, the square categorical cross-entropy loss function error of the TCN framework was generally lower than that of the 1D CNN, which did not have a memory framework or the drawback of the vanishing gradient problem. In addition, the TCN framework outperformed the LSTM model, which required a longer history to provide sufficient diagnostic ability. Still, high accuracies were achieved both in the direction of feed-drive milling and in the gravity of the handpiece through vibration signals. In general, the failure classification prediction model could accurately predict the health and failure mode of the dental handpiece before the use of the DATH when an embedded sensor was available. Therefore, this model could prove to be a beneficial tool for predicting the deterioration patterns of real dental handpieces in their remaining useful life.

Remote Measurement of Tide and Surge Using a Deep Learning System with Surveillance Camera Images

The latest progress in deep learning approaches has garnered significant attention across a variety of research fields. These techniques have revolutionized the way marine parameters are measured, enabling automated and remote data collection. This work centers on employing a deep learning model for the automated evaluation of tide and surge, aiming to deliver accurate results through the analysis of surveillance camera images. A mode of deep learning based on the Inception v3 structure was applied to predict tide and storm surges from surveillance cameras located in two different coastal areas of Italy. This approach is particularly advantageous in situations where traditional tide sensors are inaccessible or distant from the measurement point, especially during extreme events that require accurate surge measurements. The conducted experiments illustrate that the algorithm efficiently measures tide and surge remotely, achieving an accuracy surpassing 90% and maintaining a loss value below 1, evaluated through Categorical Cross-Entropy Loss functions. The findings highlight its potential to bridge the gap in data collection in challenging coastal environments, providing valuable insights for coastal management and hazard assessments. This research contributes to the emerging field of remote sensing and machine learning applications in environmental monitoring, paving the way for enhanced understanding and decision-making in coastal regions.

Deep Learning Model for Colon Cancer Classification using InceptionV3

Colon cancer is a significant global public health concern, necessitating an accurate and timely diagnosis for effective treatment. Leveraging advancements in deep learning, this study proposes a novel approach to colon cancer classification using InceptionV3 convolutional neural network architecture. A dataset comprising 1600 colonoscopy images divided into colon_aca (adenocarcinoma) and colon_n (normal) classes was utilized. The model demonstrated promising performance, achieving a training accuracy of 98.86% and a validation accuracy of 99.74% after 100 epochs. This success was accomplished by employing a Stochastic Gradient Descent (SGD) optimizer with a learning rate of 0.0001 and momentum of 0.9, along with categorical cross-entropy loss. Our findings underscore the importance of deep learning models, specifically InceptionV3, in facilitating the precise classification of colon cancer, thus offering a valuable tool for assisting clinicians in early detection and treatment decision-making. Future research may explore the integration of additional clinical data and the evaluation of alternative deep learning architectures to further enhance diagnostic accuracy.

75 AI-Enhanced Hematocrit Analysis: Revolutionizing rapid anemia detection in small ruminants

Abstract In recent times, there has been a growing focus on the well-being and healthcare of small ruminants, particularly in relation to the issue of anemia due to infection with blood-feeding gastrointestinal nematodes, such as Haemonchus contortus. The objective of this study was to assess the hematocrit levels in blood samples obtained from small ruminants, specifically goats. Additionally, the study is an attempt to design and create a quick sensor for identifying anemia, which could be conveniently used on farms. A total of 75 mature, intact male Spanish goats were subjected to hematocrit analysis to ascertain their spectrum of hematocrit values and association with anemic conditions. Simultaneously, a unique sensor with user-friendly features was developed to promptly provide farmers with feedback on the anemic status of animals using AI-based machine learning algorithms, enabling timely intervention. The sensor utilized a semi-invasive approach with minimal blood sample requirement. 30 µL of blood was dropped on glycerol soaked Whatman filter paper No. 1, and images were taken at 90 sec, 150 sec, and 270 sec to capture the blood pattern on filter paper soaked in glycerol using a cell phone. These images were correlated with actual hematocrit value obtained from conventional hematocrit analysis. A RMSprop with adjusted learning rate convolution neural network (RMSprop-CNN) based images classification model was developed for the classification of different patterns at different levels of hematocrit. The model was trained on a total of 1,000 images with a training and testing split of 80:20 ratio. For the optimization process, the Adam optimizer with a learning rate of 0.001 is employed. The model is compiled with a categorical cross-entropy loss function, aiming to improve its accuracy metric over training iterations. The initial studies demonstrated a detection accuracy level of 70.31 % at 10 epochs in identifying different hematocrit levels (Level 1; healthy, to Level 5; severely anemic) but improved significantly up to 94.89 % by 100 epochs, resulting in a notable reduction in the time and level of knowledge previously necessary for conducting such evaluations. The present study not only provides valuable insights into the hematological characteristics of small ruminants, but also establishes a foundation for the development of a straightforward and efficient method for detecting anemia. Consequently, this advancement has the potential to enhance the overall care and well-being of animals within agricultural environments.

Deep learning for automatic volumetric segmentation of left ventricular myocardium and ischaemic scar from multi-slice late gadolinium enhancement cardiovascular magnetic resonance.

This study details application of deep learning for automatic volumetric segmentation of left ventricular (LV) myocardium and scar and automated quantification of myocardial ischaemic scar burden from late gadolinium enhancement cardiovascular magnetic resonance (LGE-CMR). We included 501 images and manual segmentations of short-axis LGE-CMR from over 20 multinational sites, from which 377 studies were used for training and 124 studies from unique participants for internal validation. A third test set of 52 images was used for external evaluation. Three models, U-Net, Cascaded U-Net, and U-Net++, were trained with a novel adaptive weighted categorical cross-entropy loss function. Model performance was evaluated using concordance correlation coefficients (CCCs) for LV mass and per cent myocardial scar burden. Cascaded U-Net was found to be the best model for the quantification of LV mass and scar percentage. The model exhibited a mean difference of -5 ± 23 g for LV mass, -0.4 ± 11.2 g for scar mass, and -0.8 ± 7% for per cent scar. CCC were 0.87, 0.77, and 0.78 for LV mass, scar mass, and per cent scar burden, respectively, in the internal validation set and 0.75, 0.71, and 0.69, respectively, in the external test set. For segmental scar mass, CCC was 0.74 for apical scar, 0.91 for mid-ventricular scar, and 0.73 for basal scar, demonstrating moderate to strong agreement. We successfully trained a convolutional neural network for volumetric segmentation and analysis of LV scar burden from LGE-CMR images in a large, multinational cohort of participants with ischaemic scar.

COVID-19 Detection on Chest X-Ray Images Using Modified VGG-19

A global pandemic infecting human respiratory system identified as COVID-19 emerged in 2019 has been a critical issue concerning worldwide, including Indonesia. The massive and relatively easy spread of the virus requires COVID-19 patients to be separated from other patients to avoid hospitals being the major source of the virus spread. The gold standard in COVID-19 testing known as rRt-PCR was established to help detecting the virus. This method, however, is found to be less time-efficient due to the sample examination taking 5-7 hours to conduct. Another COVID-19 testing method called Rapid Antigen Test is claimed to be more practical since the sample examination only takes 5-30 minutes to complete. One disadvantage of this testing method is its sensitivity, which is believed to be relatively low. Further supporting examinations are needed in diagnosing COVID-19 when using the Rapid Antigen Test, one of which is chest imaging. Most hospitals in Indonesia are type C hospitals, which means radiology specialists’ availability is limited and sometimes not available. Therefore, it is necessary to develop an automation system for detecting COVID-19 disease on chest X-ray images using the Convolutional Neural Network method to help doctors interpreting chest X-rays of patients effectively faster in determining whether the patient is potentially infected with COVID-19 or not. The algorithm proposed in this study was modified VGG-19, with ADAM optimization method and categorical cross-entropy loss function. This algorithm resulted in a model accuracy rate of 81%, a precision/recall model of 62%, and a specificity model of 87%. The area under the curve in the Receiver Operatic Characteristic Curve indicates that the accuracy level of the test is below the Excellent curve, suggesting that the model created successfully serves as a screening test tool.

Tomato leaf disease identification via two–stage transfer learning approach

In the last ten years, there has been an upsurge in focus on sustainable agribusiness as a response to the bio-hazards posed by the effects of climate change, severe weather events, population growth, increased demands for food security, and a lack of natural resources. Generally, tomato plants are very susceptible to a wide variety of diseases. So, when it comes to maintaining the quality of tomato crops, having a quick diagnosis that is also correct plays a significant role. Nowadays, deep learning (DL), notably convolutional neural networks (CNNs), have achieved exceptional results in numerous applications, including categorizing tomato plant diseases. This study aims to enhance the disease identification algorithm for tomatoes by using transfer learning to reduce the time needed for the model to be trained and improve its identification accuracy. The model is based on the VGGNet that has been pre-trained using ImageNet and two inception blocks. In addition, the improved categorical cross-entropy loss function for the multi-attribute identification task and two-stage transfer learning were included in the model training process. Our model obtains a better identification accuracy for tomato diseases in the test set than other state-of-the-art techniques, as shown by the experiments' outcomes. The findings indicate substantial values produced by the suggested approach, with 99.23% of accuracy. The suggested model is a beneficial tool for farmers in helping to detect and protect tomatoes from disease due to the considerably high success rate achieved by the model.

DRFL: Federated Learning in Diabetic Retinopathy Grading Using Fundus Images

Diabetic retinopathy (DR) is a complication of diabetic Mellitus, developing retinal lesions that impair vision. The DR detection in the early stages avoids permanent vision loss. The treatments provide relief, and the vision loss due to DR is irreversible. The manual grading of DR is time-consuming and prone to human errors. The other real-time problem is exchanging patient fundus image information with hospitals worldwide while upholding the organisations' privacy concerns. When training a deep learning (DL) network, two critical factors to keep in mind are creating a collaborative platform and protecting patient data privacy. Therefore, an automated DR detection technique is required while protecting patient data and privacy. In this work, we propose a novel DR severity grading technique based on Federated Learning (FL), a recent advancement in DL called DRFL. FL is a new research paradigm that allows DL models to be trained collectively without disclosing clinical information. In DRFL, we combined the Federated averaging (FedAvg) technique and the median of the categorical cross-entropy loss. Since in comparison to FedAvg, the median cross-entropy is better suited for either under-fitted or over-fitted clients. Also, we propose a novel central server that extracts multi-scale features from the fundus images to identify small lesions present in the fundus image. In this work, we consider five clients holding different preprocessed fundus images collected from publicly available databases such as MESSIDOR-2, IDRiD, Kaggle, and a local database collected from Silchar Medical College and Hospital. The proposed model obtained an accuracy of 98.6%, specificity of 99.3%, precision of 97.25%, and an F1 score of 97.5%, which are better results than other techniques.

Brain Tumor Segmentation Using Deep Learning on MRI Images.

Brain tumor (BT) diagnosis is a lengthy process, and great skill and expertise are required from radiologists. As the number of patients has expanded, so has the amount of data to be processed, making previous techniques both costly and ineffective. Many academics have examined a range of reliable and quick techniques for identifying and categorizing BTs. Recently, deep learning (DL) methods have gained popularity for creating computer algorithms that can quickly and reliably diagnose or segment BTs. To identify BTs in medical images, DL permits a pre-trained convolutional neural network (CNN) model. The suggested magnetic resonance imaging (MRI) images of BTs are included in the BT segmentation dataset, which was created as a benchmark for developing and evaluating algorithms for BT segmentation and diagnosis. There are 335 annotated MRI images in the collection. For the purpose of developing and testing BT segmentation and diagnosis algorithms, the brain tumor segmentation (BraTS) dataset was produced. A deep CNN was also utilized in the model-building process for segmenting BTs using the BraTS dataset. To train the model, a categorical cross-entropy loss function and an optimizer, such as Adam, were employed. Finally, the model's output successfully identified and segmented BTs in the dataset, attaining a validation accuracy of 98%.

Classification and regression-based convolutional neural network and long short-term memory configuration for bridge damage identification using long-term monitoring vibration data

Considerable attention has recently been focused on classification and regression-based convolutional neural network (CNN) and long short-term memory (LSTM) due to their excellent performance in capturing complex spatial and temporal information characteristics for structural damage identification. However, few studies have considered structural damage identification as a classification and regression problem. In addition, bridges in practical engineering are vulnerable to various environmental and vehicle loading conditions. Hence, this study proposed a new two-stage CNN–LSTM configuration for bridge damage identification using vibration data considering the influence of temperatures. First, a classification-based CNN–LSTM is designed to perform multiclass damage detection tasks, and then a regression-based CNN–LSTM is developed for damage localization and severity prediction tasks. The performance of the proposed damage identification method was evaluated through a simulation dataset of a concrete highway bridge model and a field experiment dataset of Z24-bridge (Switzerland). In addition, a set of statistical evaluation metrics such as sparse categorical cross-entropy loss, accuracy, confusion matrix, mean squared loss, mean absolute error, mean absolute percentage error, and coefficient of determination were used to compare the damage identification performance of the proposed CNN–LSTM configuration with a regular CNN model and conventional machine learning (ML) algorithms. Prediction results indicate that the proposed CNN–LSTM model outperforms the regular CNN model and conventional ML algorithms for bridge damage identification.

Flood Susceptibility Mapping Using Image-Based 2D-CNN Deep Learning: Overview and Case Study Application Using Multiparametric Spatial Data in Data-Scarce Urban Environments

This study presents results for urban flood susceptibility mapping (FSM) using image-based 2D-convolutional neural networks (2D-CNN). The model input multiparametric spatial data comprised of land-useland-cover (LULC), digital elevation model (DEM), and the topographic and hydrologic conditioning derivatives, precipitation, and soil types. The implemented dropout regularization 2D-CNN with ReLU activation function, categorical cross-entropy loss function, and AdaGrad optimizer produced the case study area FSM with overall accuracy (OA) of 82.5%. The image-based 2D-CNN outperformed the multilayer perceptron (MLP) neural network by 18.4% in terms of overall accuracy and with corresponding lower MAE and higher F1-measures of 10.9% and 0.989, as compared to 25.6% and 0.877, respectively, for MLP-ANN results. The accuracy of the 2D-CNN that produced FSM map and the model efficiency were evaluated using area under the ROC curve (AUC) with respective success and prediction rates of 0.827 and 0.809. Using image-based 2D-CNN, 27% of the 247.7 km2 of the studied area was mapped with a high risk of flooding, with MLP-ANN overestimating the degree of high flood risk by 4.7%. Based on the gain ratio index analysis of the flood conditioning factors (FCFs), the most significant FCFs were LULC (18.5%), precipitation (14.9%), proximity to river (13.3%), and elevation (12.4%). Soil types contributed 8.6%, slope 9.1%, and the DEM-derived hodological conditioning indicators contributed 23.2%. The study results demonstrate that in urban areas with scarce hydrological monitoring networks, the use of image-based 2D-CNN with multiparametric spatial data can produce high-quality flood susceptibility maps for flood management in urban environments.

Adversarial Robustness via Fisher-Rao Regularization.

Adversarial robustness has become a topic of growing interest in machine learning since it was observed that neural networks tend to be brittle. We propose an information-geometric formulation of adversarial defense and introduce Fire, a new Fisher-Rao regularization for the categorical cross-entropy loss, which is based on the geodesic distance between the softmax outputs corresponding to natural and perturbed input features. Based on the information-geometric properties of the class of softmax distributions, we derive an explicit characterization of the Fisher-Rao Distance (FRD) for the binary and multiclass cases, and draw some interesting properties as well as connections with standard regularization metrics. Furthermore, we verify on a simple linear and Gaussian model, that all Pareto-optimal points in the accuracy-robustness region can be reached by Fire while other state-of-the-art methods fail. Empirically, we evaluate the performance of various classifiers trained with the proposed loss on standard datasets, showing up to a simultaneous 1% of improvement in terms of clean and robust performances while reducing the training time by 20% over the best-performing methods.

HYBRID DEEP LEARNING WITH MULTIMODAL DATA AUGMENTATION FOR RIVERBANK TREE RING CLASSIFICATION

The purpose of this research was to develop a model for counting or classifying annual tree rings of lamphu trees along riverbanks. The study was based on image processing combined with deep learning and machine learning to develop the model. The image data augmentations were processed through image preprocessing with the region of interest, denoising, color model conversion, blurring, thresholding, and dilation. This process transforms one image’s data into seventy images as the dataset. Deep learning models developed with Visual Geometry Group-16-based convolutional neural networks applied different support vector machine classifiers: L2 norm, categorical hinge loss, and the Weston-Watkins. The results showed that the model developed with the categorical hinge loss-support vector machine classifier gave the highest model efficiency with an accuracy of 94.07% and a categorical cross-entropy loss of 5.48%. The models developed with the Weston-Watkins and L2 norm were followed by an accuracy of 93.70% and 93.26%, respectively. The model developed with the Softmax function as a classifier that gave the lowest order of accuracy of 92.96%. Keywords: Convolutional Neural Network, Data Augmentation, Deep Learning, Support Vector Machine, Tree Ring DOI： https://doi.org/10.35741/issn.0258-2724.58.3.27

A Lightweight Robust Deep Learning Model Gained High Accuracy in Classifying a Wide Range of Diabetic Retinopathy Images

Diabetic retinopathy (DR) is a common complication of diabetes mellitus, and retinal blood vessel damage can lead to vision loss and blindness if not recognized at an early stage. Manual DR detection using large fundus image data is time-consuming and error-prone. An effective automatic DR detection system can be significantly faster and potentially more accurate. This study aims to classify fundus images into five DR classes, using deep learning methods, with the highest possible accuracy and the lowest possible computational time. Three distinct DR datasets, APTOS, Messidor2, and IDRiD, are merged, resulting in 5,819 raw images. Before training the model, various image preprocessing techniques are applied to remove artifacts and noise from the images and improve their quality. Three augmentation techniques: geometric, photometric, and elastic deformation, are used to create a balanced dataset. A shallow convolutional neural network (CNN) is developed using three blocks of convolutional layers and maxpool layers with a categorical cross-entropy loss function, Adam optimizer, 0.0001 learning rate, and 64 batch size as a base model, and this is also employed to determine the best data augmentation method for further processing. A study to optimize the performance is then conducted by changing different components and hyperparameters of the base model, resulting in our proposed RetNet-10 model. Six cutting-edge models are employed for comparison. Our proposed RetNet-10 model performed the best, with a testing accuracy of 98.65%. MobileNetV2, VGG16, Xception, VGG19, InceptionV3 and ResNet50 achieved testing accuracies of 91.42%, 90.16%,89.57%, 88.21%, 87.68% and 87.23%, respectively. The model is also trained with several k values to assess its robustness. After image processing and data augmentation, using the combined dataset, and fine-tuning the base model, our proposed RetNet-10 model outperformed other automated methods for DR diagnosis.

Optical Character Recognition System for Digit Recognition Using Deep Learning

Because it is so difficult to distinguish handwritten digits, digit identification is one of the most critical applications in computer vision. This is one of the reasons why it is so tough. The field of handwritten character recognition is one in which a great deal of application of numerous deep learning models has occurred. The startling parallels that can be drawn between deep learning and the brain are primarily responsible for its meteoric rise in popularity. In this study, the Artificial Neural Network and the Convolutional Neural Network, two of the most used Deep Learning algorithms, were investigated with an eye toward the recognition process's feature extraction and classification phases. With the assistance of the categorical cross-entropy loss and the ADAM optimizer, the models were trained on the MNIST dataset. Backpropagation and gradient descent are the two methods utilized during the training process of neural networks that contain reLU activations and carry out automatic feature extraction. In computer vision, one of the most common and widely used classifiers is the Convolution Neural Network, sometimes referred to as ConvNets or Convolutional neural networks. This network is used for the recognition and categorization of images.

A Tagging Model using Segmentation Proposal Network

This paper presents a tagging model used the Segmentation map as reference regions. The suggested model leverages an encoder-decoder architecture combined with a proposal layer and dense layers for accurate object tagging and segmentation. The proposed model utilizes a pre-trained VGG16 encoder to extract high-level features from input images, followed by a decoder network that reconstructs the image. A proposal layer generates a binary map indicating the presence or absence of objects at each location in the image. The proposal layer is integrated with the decoder output and further refined by a convolutional layer to produce the final segmentation. Two dense layers are employed to predict object classes and bounding box coordinates. The model is trained using a custom loss function that combines categorical cross-entropy loss and means squared error loss. Experimental results demonstrate the effectiveness of the proposed model in achieving accurate object tagging and segmentation.

Real-time driving risk assessment using deep learning with XGBoost

Traffic crashes typically occur in a few seconds and real-time prediction can significantly benefit traffic safety management and the development of safety countermeasures. This paper presents a novel deep learning model for crash identification based on high-frequency, high-resolution continuous driving data. The method consists of feature engineering based on Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) and classification based on Extreme Gradient Boosting (XGBoost). The CNN-GRU architecture captures the time series characteristics of driving kinematics data. Compared to normal driving segments, safety-critical events (SCEs)—i.e., crashes and near-crashes (CNC)—are rare. The weighted categorical cross-entropy loss and oversampling methods are utilized to address this imbalance issue. An XGBoost classifier is utilized instead of the multi-layer perceptron (MLP) to achieve a high precision and recall rate. The proposed approach is applied to the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP 2 NDS) data with 1,820 crashes, 6,848 near-crashes, and 59,997 normal driving segments. The results show that in a 3-class classification system (crash, near-crash, normal driving segments), the accuracy for the overall model is 97.5%, and the precision and recall for crashes are 84.7%, and 71.3% respectively, which is substantially better than benchmarks models. Furthermore, the recall of the most severe crashes is 98.0%. The proposed crash identification approach provides an accurate, highly efficient, and scalable way to identify crashes based on high frequency, high-resolution continuous driving data and has broad application prospects in traffic safety applications.

Personalized Activity Recognition with Deep Triplet Embeddings.

A significant challenge for a supervised learning approach to inertial human activity recognition is the heterogeneity of data generated by individual users, resulting in very poor performance for some subjects. We present an approach to personalized activity recognition based on deep feature representation derived from a convolutional neural network (CNN). We experiment with both categorical cross-entropy loss and triplet loss for training, and describe a novel loss function based on subject triplets. We evaluate these methods on three publicly available inertial human activity recognition datasets (MHEALTH, WISDM, and SPAR) comparing classification accuracy, out-of-distribution activity detection, and generalization to new activity classes. The proposed triplet algorithm achieved an average 96.7% classification accuracy across tested datasets versus the 87.5% achieved by the baseline CNN algorithm. We demonstrate that personalized algorithms, and, in particular, the proposed novel triplet loss algorithms, are more robust to inter-subject variability and thus exhibit better performance on classification and out-of-distribution detection tasks.