Train Configuration Research Articles

RepDI: A light-weight CPU network for apple leaf disease identification

Apple disease is one of the major factors affecting apple production, and the visual diagnosis of apple leaves is an efficient disease identification solution. In this paper, we propose an efficient lightweight model based on structural reparameterization for apple leaf disease identification, called RepDI for short. To achieve faster inference on the CPU devices, we introduce depth-wise separable convolution and structural reparameterization technology in RepDI, which has different structures during training and inference. In addition, to better capture diseased leaves and disease regions in complex contexts, we propose the parallel dilated attention mechanism module and embed it into RepDI. Experiments show that RepDI can achieve state-of-the-art performance in disease identification task, compared to most lightweight models. Meanwhile, RepDI achieves the fastest inference speed on our desktop CPU, which is an important factor in practical applications. Furthermore, we collect and annotate a novel dataset for apple leaf diseases from real scenarios, called Real-ALD, which is more challenging than previous datasets. And RepDI achieves a top-1 accuracy of 98.92 in the Real-ALD dataset under a limited training configuration. Our code is released to contribute to the plant protection community and we will further explore the potential of RepDI for down-stream detection, segmentation tasks.

Multiobjective evolutionary pruning of Deep Neural Networks with Transfer Learning for improving their performance and robustness

Evolutionary Computation algorithms have been used to solve optimization problems in relation with architectural, hyper-parameter or training configuration, forging the field known today as Neural Architecture Search. These algorithms have been combined with other techniques such as the pruning of Neural Networks, which reduces the complexity of the network, and the Transfer Learning, which lets the import of knowledge from another problem related to the one at hand. The usage of several criteria to evaluate the quality of the evolutionary proposals is also a common case, in which the performance and complexity of the network are the most used criteria. This work proposes MO-EvoPruneDeepTL, a multi-objective evolutionary pruning algorithm. MO-EvoPruneDeepTL uses Transfer Learning to adapt the last layers of Deep Neural Networks, by replacing them with sparse layers evolved by a genetic algorithm, which guides the evolution based in the performance, complexity and robustness of the network, being the robustness a great quality indicator for the evolved models. We carry out different experiments with several datasets to assess the benefits of our proposal. Results show that our proposal achieves promising results in all the objectives, and direct relation are presented among them. The experiments also show that the most influential neurons help us explain which parts of the input images are the most relevant for the prediction of the pruned neural network. Lastly, by virtue of the diversity within the Pareto front of pruning patterns produced by the proposal, it is shown that an ensemble of differently pruned models improves the overall performance and robustness of the trained networks.

Model reference adaptive system speed estimator based on type-1 and type-2 fuzzy logic sensorless control of electrical vehicle with electrical differential

Introduction. In this paper, a new approach for estimating the speed of in-wheel electric vehicles with two independent rear drives is presented. Currently, the variable-speed induction motor replaces the DC motor drive in a wide range of applications, including electric vehicles where quick dynamic response is required. This is now possible as a result of significant improvements in the dynamic performance of electrical drives brought about by technological advancements and development in the fields of power commutation devices, digital signal processing, and, more recently, intelligent control systems. The system’s reliability and robustness are improved, and the cost, size, and upkeep requirements of the induction motor drive are reduced through control strategies without a speed sensor. Successful uses of the induction motor without a sensor have been made for medium- and high-speed operations. However, low speed instability and instability under various charge perturbation conditions continue to be serious issues in this field of study and have not yet been effectively resolved. Some application such as traction drives and cranes are required to maintain the desired level of torque down to low speed levels with uncertain load torque disturbance conditions. Speed and torque control is more important particularly in motor-in-wheel traction drive train configuration where vehicle wheel rim is directly connected to the motor shaft to control the speed and torque. Novelty of the proposed work is to improve the dynamic performance of conventional controller used of model reference adaptive system speed observer using both type-1 and type-2 fuzzy logic controllers. Purpose. In proposed scheme, the performance of the engine is being controlled, fuzzy logic controller is controlling the estimate rotor speed, and results are then compared using type-1 and type-2. Method. For a two-wheeled motorized electric vehicle, a high-performance sensorless wheel motor drive based on both type-2 and type-1 fuzzy logic controllers of the model reference adaptive control system is developed. Results. Proved that, using fuzzy logic type-2 controller the sensorless speed control of the electrical differential of electric vehicle EV observer, much better results are achieved. Practical value. The main possibility of realizing reliable and efficient electric propulsion systems based on intelligent observers (type-2 fuzzy logic) is demonstrated. The research methodology has been designed to facilitate the future experimental implementation on a digital signal processor.

Pedestrian Detection System using YOLOv5 for Advanced Driver Assistance System (ADAS)

The technology in transportation is continuously developing due to reaching the self-driving vehicle. The need of detecting the situation around vehicles is a must to prevent accidents. It is not only limited to the conventional vehicle in which accident commonly happens, but also to the autonomous vehicle. In this paper, we proposed a detection system for recognizing pedestrians using a camera and minicomputer. The approach of pedestrian detection is applied using object detection method (YOLOv5) which is based on the Convolutional Neural Network. The model that we proposed in this paper is trained using numerous epochs to find the optimum training configuration for detecting pedestrians. The lowest value of object and bounding box loss is found when it is trained using 2000 epochs, but it needs at least 3 hours to build the model. Meanwhile, the optimum model’s configuration is trained using 1000 epochs which has the biggest object (1.49 points) and moderate bounding box (1.5 points) loss reduction compared to the other number of epochs. This proposed system is implemented using Raspberry Pi4 and a monocular camera and it is only able to detect objects for 0.9 frames for each second. As further development, an advanced computing device is needed due to reach real-time pedestrian detection. 

ChampKit: A framework for rapid evaluation of deep neural networks for patch-based histopathology classification

Background and ObjectiveHistopathology is the gold standard for diagnosis of many cancers. Recent advances in computer vision, specifically deep learning, have facilitated the analysis of histopathology images for many tasks, including the detection of immune cells and microsatellite instability. However, it remains difficult to identify optimal models and training configurations for different histopathology classification tasks due to the abundance of available architectures and the lack of systematic evaluations. Our objective in this work is to present a software tool that addresses this need and enables robust, systematic evaluation of neural network models for patch classification in histology in a light-weight, easy-to-use package for both algorithm developers and biomedical researchers. MethodsHere we present ChampKit (Comprehensive Histopathology Assessment of Model Predictions toolKit): an extensible, fully reproducible evaluation toolkit that is a one-stop-shop to train and evaluate deep neural networks for patch classification. ChampKit curates a broad range of public datasets. It enables training and evaluation of models supported by timm directly from the command line, without the need for users to write any code. External models are enabled through a straightforward API and minimal coding. As a result, Champkit facilitates the evaluation of existing and new models and deep learning architectures on pathology datasets, making it more accessible to the broader scientific community. To demonstrate the utility of ChampKit, we establish baseline performance for a subset of possible models that could be employed with ChampKit, focusing on several popular deep learning models, namely ResNet18, ResNet50, and R26-ViT, a hybrid vision transformer. In addition, we compare each model trained either from random weight initialization or with transfer learning from ImageNet pretrained models. For ResNet18, we also consider transfer learning from a self-supervised pretrained model. ResultsThe main result of this paper is the ChampKit software. Using ChampKit, we were able to systemically evaluate multiple neural networks across six datasets. We observed mixed results when evaluating the benefits of pretraining versus random intialization, with no clear benefit except in the low data regime, where transfer learning was found to be beneficial. Surprisingly, we found that transfer learning from self-supervised weights rarely improved performance, which is counter to other areas of computer vision. ConclusionsChoosing the right model for a given digital pathology dataset is nontrivial. ChampKit provides a valuable tool to fill this gap by enabling the evaluation of hundreds of existing (or user-defined) deep learning models across a variety of pathology tasks. Source code and data for the tool are freely accessible at https://github.com/SBU-BMI/champkit.

EVALUATING BERT AND PARSBERT FOR ANALYZING PERSIAN ADVERTISEMENT DATA

This paper discusses the impact of the Internet on modern trading and the importance of data generated from these transactions for organizations to improve their marketing efforts. The paper uses the example of Divar, an online marketplace for buying and selling products and services in Iran, and presents a competition to predict the percentage of a car sales ad that would be published on the Divar website. Since the dataset provides a rich source of Persian text data, the authors use the Hazm library, a Python library designed for processing Persian text, and two state-of-the-art language models, mBERT and ParsBERT, to analyze it. The paper's primary objective is to compare the performance of mBERT and ParsBERT on the Divar dataset. The authors provide some background on data mining, Persian language, and the two language models, examine the dataset's composition and statistical features, and provide details on their fine-tuning and training configurations for both approaches. They present the results of their analysis and highlight the strengths and weaknesses of the two language models when applied to Persian text data. The paper offers valuable insights into the challenges and opportunities of working with low-resource languages such as Persian and the potential of advanced language models like BERT for analyzing such data. The paper also explains the data mining process, including steps such as data cleaning and normalization techniques. Finally, the paper discusses the types of machine learning problems, such as supervised, unsupervised, and reinforcement learning, and the pattern evaluation techniques, such as confusion matrix. Overall, the paper provides an informative overview of the use of language models and data mining techniques for analyzing text data in low-resource languages, using the example of the Divar dataset.

Artificial neural network-based multiple-input multiple-output metamodel for prediction of design parameters for a high-speed rail viaduct

The prediction of the design parameters of short to medium-span supported bridges in critical locations (such as canal/road crossings) under the action of high-speed trains has been investigated in this article. An artificial neural network (ANN)-based MIMO (multiple-input multiple-output) metamodels is proposed in conjunction with the semi-analytical framework of simply-supported bridges. Three cases, namely single moving load, series of moving loads at equal spacing (HSLM-B), and as per conventional train configuration (HSLM-A) recommended in Eurocode1: EN 1991-2 (2003), are considered. The prime novelty of the article is to develop a dimensionless semi-analytical framework to train and validate a MIMO metamodel implementing ANN for predicting the multiple dynamic responses of bridges under high-speed loads. The dependency of the maximum dynamic responses, that is, displacement, shear force, and bending moment, on the governing parameters (structural and loading) have been elucidated using Pearson's correlation matrix for the three different train configurations. Further, the robustness and efficiency of the best-fitted metamodels have been compared, and a user interface has been developed for ease of implementation. This platform evaluates the responses such as displacement, shear force, bending moment, and structural safety confirming the standards of Eurocode EN 1990:2002 + A1:2005 (E).

An efficient method for predicting wheel-rail forces in coupled nonlinear train-track-bridge system using artificial neural networks

Wheel-rail forces are an essential indicator for vehicle safety evaluation. The calculation of wheel-rail forces for the coupled nonlinear train-track-bridge system using the direct numerical integration method is time-consuming, hindering the timely safety and reliability assessment of train operations. In this paper, an efficient method based on the nonlinear autoregressive with exogenous input artificial neural network (NARX-ANN) was proposed to predict the nonlinear wheel-rail force of trains on bridges caused by track irregularity. Firstly, samples of track irregularity time history were randomly generated using the stochastic harmonic function method. The vertical and lateral wheel-rail forces of the coupled nonlinear train-track-bridge system were calculated by numerical integration. Secondly, the NARX-ANN model was established and trained with a small number of randomly selected wheel-rail force samples. Finally, the wheel-rail forces obtained by numerical integration were regarded as the ground truth to verify the prediction accuracy of the NARX-ANN model. The influence of training configuration, initial output, and the number of time delays on prediction accuracy was systematically analyzed. The results showed that the NARX-ANN model could accurately predict the time history of wheel-rail forces. In addition, its computational efficiency is 19.35 times higher than that of the numerical integration method. It is hoped that this study can guide the stochastic analysis of the train–track–bridge system.

Experimental study on wheel-rail rolling contact fatigue damage starting from surface defects under various operational conditions

Surface defects in different railway lines experience different wheel-rail rolling contact actions owing to various railway line characteristics and vehicle operation parameters. This study focused on the evolution of defective wheel-rail rolling contact fatigue (RCF) behaviour under different operation conditions. Conical defects were created using an indentation apparatus. Wheel-rail RCF tests were performed using a twin-disc testing machine. Various operational conditions were simulated considering different slip ratios and wheel linear velocities on uphill and flat lines, as well as different vehicle axle loads and train configurations. The results indicate that surface damage, crack depths, and wear debris size increased only with decreasing wheel linear velocities, whereas the crack angles increased only with decreasing slip ratio. Meanwhile, both the wear rate and oxidation degree of wear debris were simultaneously affected by the slip ratio and wheel linear velocities. Under rolling conditions on flat lines, vehicle axle loads had a negligible influence on the wheel-rail RCF behaviour.

Change of Direction Ability as a Sensitive Marker of Adaptation to Different Training Configurations, and Different Populations: Results from Four Experiments.

This article includes four separate experiments. In the first experiment male beach handball players (n = 24) were randomly assigned to regular training (n = 12) or plyometric and sprint training (n = 12). In the second experiment, male players were assigned to a handball practice only (n = 12), a plyometric training (n = 12), or an eccentric-overload (e.g., versa-pulley machine) training group (n = 12). In the third experiment, participants were assigned to padel training (n = 12) or specific on-court neuromuscular technical actions (n = 12). In the fourth experiment, females between 50-59 years (n = 25), 60-64 years (n = 25), and 65-70 years (n = 25) completed 10 weeks of bench stepping training involving jumps, and were compared to age-matched controls (n = 45). The COD ability was assessed with the 10-m COD ability test (experiments one and two), the COD ability test with 90º and 180º turns (third experiment), and with the timed-up-and-go test (fourth experiment). In experiment one, greater COD improvement was noted in the experimental group compared to the control group (p < 0.05). In experiment two, both intervention groups similarly improved COD when compared to the control group (p < 0.05). In experiment three, although no significant group-time interactions were observed for COD, the experimental group improved all COD measures pre-post-training (p < 0.05), with a larger effect size for COD with 180° turn to the right compared to the control group (effect size = 0.8 vs. 0.3). In experiment four, the three training groups improved COD ability compared to the control group (p < 0.05). In conclusion, COD is a sensitive marker of adaptation to different training configurations in these diverse groups.

Wireless Federated Learning With Hybrid Local and Centralized Training: A Latency Minimization Design

Wireless federated learning (FL) is a collaborative machine learning (ML) framework in which wireless client-devices independently train their ML models and send the locally trained models to the FL server for aggregation. In this paper, we consider the coexistence of privacy-sensitive client-devices and privacy-insensitive yet computing-resource constrained client-devices, and propose an FL framework with a hybrid centralized training and local training. Specifically, the privacy-sensitive client-devices perform local ML model training and send their local models to the FL server. Each privacy-insensitive client-device can have two options, i.e., (i) conducting a local training and then sending its local model to the FL server, and (ii) directly sending its local data to the FL server for the centralized training. The FL server, after collecting the data from the privacy-insensitive client-devices (which choose to upload the local data), conducts a centralized training with the received datasets. The global model is then generated by aggregating (i) the local models uploaded by the client-devices and (ii) the model trained by the FL server centrally. Focusing on this hybrid FL framework, we firstly analyze its convergence feature with respect to the client-devices' selections of local training or centralized training. We then formulate a joint optimization of client-devices' selections of the local training or centralized training, the FL training configuration (i.e., the number of the local iterations and the number of the global iterations), and the bandwidth allocations to the client-devices, with the objective of minimizing the overall latency for reaching the FL convergence. Despite the non-convexity of the joint optimization problem, we identify its layered structure and propose an efficient algorithm to solve it. Numerical results demonstrate the advantage of our proposed FL framework with the hybrid local and centralized training as well as our proposed algorithm, in comparison with several benchmark FL schemes and algorithms.

Deep Learning for Predictive Analytics in Reversible Steganography

Deep learning is regarded as a promising solution for reversible steganography. There is an accelerating trend of representing a reversible steo-system by monolithic neural networks, which bypass intermediate operations in traditional pipelines of reversible steganography. This end-to-end paradigm, however, suffers from imperfect reversibility. By contrast, the modular paradigm that incorporates neural networks into modules of traditional pipelines can stably guarantee reversibility with mathematical explainability. Prediction-error modulation is a well-established reversible steganography pipeline for digital images. It consists of a predictive analytics module and a reversible coding module. Given that reversibility is governed independently by the coding module, we narrow our focus to the incorporation of neural networks into the analytics module, which serves the purpose of predicting pixel intensities and a pivotal role in determining capacity and imperceptibility. The objective of this study is to evaluate the impacts of different training configurations upon predictive accuracy of neural networks and provide practical insights. In particular, we investigate how different initialisation strategies for input images may affect the learning process and how different training strategies for dual-layer prediction respond to the problem of distributional shift. Furthermore, we compare steganographic performance of various model architectures with different loss functions.

ANN crowds in early-stage design: an investigation of influence of small training sets on prediction

The objective of this paper is to present and evaluate an approach to early-stage design product performance predicting that employs the “wisdom of the crowd” of artificial neural networks (ANN). The wisdom of the crowd has been used to gather insights, ideas, and problems by using responses from large numbers of people. In this work, 189 different ANN architectures will be used instead of human agents, thus creating an ANN Crowd. This research explores the question “should every agent in the crowd has the same voice?” To answer this, this research aims to broadly understand the sensitivity of influencing factors (training set selection, nodes, layers, and architecture) in the performance of the ANN Crowd. This will then support refinement to achieve generalizability for the ANN Crowds. The specific focus of this paper is to explore the effect of the training set selection on prediction accuracy. To do this, four different scenarios are explored (2 inputs x 2 outputs). ANN Crowds are created for these scenarios with different product training configurations to test how the training set influences the predictions (k-fold = 10). Results show that training set selection has a statistical and practical influence on prediction accuracy. Analyzing the presence of edge cases in training sets and the effects of training targets suggests further influence of training set selection. Future work includes a deeper exploration of the influence of edge cases within the training sets. Building on this foundation, the members of the ANN Crowd (different architectures) will be studied for their influence on prediction performance.

SMSG: Profiling-Free Parallelism Modeling for Distributed Training of DNN

The increasing size of deep neural networks (DNNs) raises a high demand for distributed training. An expert could find good hybrid parallelism strategies, but designing suitable strategies is time and labor-consuming. Therefore, automating parallelism strategy generation is crucial and desirable for DNN designers. Some automatic searching approaches have recently been studied to free the experts from the heavy parallel strategy conception. However, these approaches all rely on a numerical cost model, which requires heavy profiling results that lack portability. These profiling-based approaches cannot lighten the strategy generation work due to the non-reusable profiling value. Our intuition is that there is no need to estimate the actual execution time of the distributed training but to compare the relative cost of different strategies. We propose SMSG (Symbolic Modeling for Strategy Generation), which analyses the cost based on the communication and computation semantics. With SMSG, the parallel cost analyses are decoupled from hardware characteristics. SMSG defines cost functions for each kind of operator to quantitatively evaluate the amount of data for computation and communication, which eliminates the heavy profiling tasks. Besides, SMSG introduces how to apply functional transformation by using the Third Homomorphism theorem to control the high searching complexity. Our experiments show that SMSG can find good hybrid parallelism strategies to generate an efficient training performance similar to the state of the art. Moreover, SMSG covers a wide variety of DNN models with good scalability. SMSG provides good portability when changing training configurations that a profiling-based approach cannot.

Evaluating deep learning for predicting epigenomic profiles.

Deep learning has been successful at predicting epigenomic profiles from DNA sequences. Most approaches frame this task as a binary classification relying on peak callers to define functional activity. Recently, quantitative models have emerged to directly predict the experimental coverage values as a regression. As new models continue to emerge with different architectures and training configurations, a major bottleneck is forming due to the lack of ability to fairly assess the novelty of proposed models and their utility for downstream biological discovery. Here we introduce a unified evaluation framework and use it to compare various binary and quantitative models trained to predict chromatin accessibility data. We highlight various modeling choices that affect generalization performance, including a downstream application of predicting variant effects. In addition, we introduce a robustness metric that can be used to enhance model selection and improve variant effect predictions. Our empirical study largely supports that quantitative modeling of epigenomic profiles leads to better generalizability and interpretability.

A generalizable, uncertainty-aware neural network potential for GeSbTe with Monte Carlo dropout

A Bayesian neural network potential (NNP) achieved with the Monte Carlo dropout approximation method is developed for GeSbTe alloys. The Bayesian NNP is shown to be more generalizable than its classical counterpart, yielding reasonable predictions on structures that are not directly in the training configurations, and is able to output uncertainty estimates for the predictions. Its application to a molecular dynamics (MD) simulation is also presented, and the validity of the obtained trajectory is evaluated by comparing it to Density Functional Theory (DFT).

Evaluating the Robustness of Click Models to Policy Distributional Shift

Many click models have been proposed to interpret logs of natural interactions with search engines and extract unbiased information for evaluation or learning. The experimental set-up used to evaluate them typically involves measuring two metrics, namely the test perplexity for click prediction and nDCG for relevance estimation. In both cases, the data used for training and testing is assumed to be collected using the same ranking policy. We question this assumption. Important downstream tasks based on click models involve evaluating a different policy than the training policy, i.e., click models need to operate under policy distributional shift . We show that click models are sensitive to it. This can severely hinder their performance on the targeted task: conventional evaluation metrics cannot guarantee that a click model will perform equally well under distributional shift. In order to more reliably predict click model performance under policy distributional shift, we propose a new evaluation protocol. It allows us to compare the relative robustness of six types of click models under various shifts, training configurations and downstream tasks. We obtain insights into the factors that worsen the sensitivity to policy distributional shift, and formulate guidelines to mitigate the risks of deploying policies based on click models.

Few-Shot Classification with Dual-Model Deep Feature Extraction and Similarity Measurement

From traditional machine learning to the latest deep learning classifiers, most models require a large amount of labeled data to perform optimal training and obtain the best performance. Yet, when limited training samples are available or when accompanied by noisy labels, severe degradation in accuracy can arise. The proposed work mainly focusses on these practical issues. Herein, standard datasets, i.e., Mini-ImageNet, CIFAR-FS, and CUB 200, are considered, which also have similar issues. The main goal is to utilize a few labeled data in the training stage, extracting image features and then performing feature similarity analysis across all samples. The highlighted aspects of the proposed method are as follows. (1) The main self-supervised learning strategies and augmentation techniques are exploited to obtain the best pretrained model. (2) An improved dual-model mechanism is proposed to train the support and query datasets with multiple training configurations. As examined in the experiments, the dual-model approach obtains superior performance of few-shot classification compared with all of the state-of-the-art methods.

A systematic review and meta-analysis demonstrating Klotho as an emerging exerkine

Klotho is an anti-aging protein with several therapeutic roles in the pathophysiology of different organs, such as the skeletal muscle and kidneys. Available evidence suggests that exercise increases Klotho levels, regardless of the condition or intervention, shedding some light on this anti-aging protein as an emergent and promising exerkine. Development of a systematic review and meta-analysis in order to verify the role of different exercise training protocols on the levels of circulating soluble Klotho (S-Klotho) protein. A systematic search of the Cochrane Central Register of Controlled Trials (CENTRAL), MEDLINE through PubMed, EMBASE, CINAHL, CT.gov, and PEDro. Randomized and quasi-randomized controlled trials that investigated effects of exercise training on S-Klotho levels. We included 12 reports in the analysis, comprising 621 participants with age ranging from 30 to 65 years old. Klotho concentration increased significantly after chronic exercise training (minimum of 12 weeks) (Hedge’ g [95%CI] 1.3 [0.69–1.90]; P < 0.0001). Moreover, exercise training increases S-Klotho values regardless of the health condition of the individual or the exercise intervention, with the exception of combined aerobic + resistance training. Furthermore, protocol duration and volume seem to influence S-Klotho concentration, since the effect of the meta-analysis changes when subgrouping these variables. Altogether, circulating S-Klotho protein is altered after chronic exercise training and it might be considered an exerkine. However, this effect may be influenced by different training configurations, including protocol duration, volume, and intensity.

Applying a hybrid model considering the interaction between train and power system for energy consumption surveys

AbstractPower supply system plays an important role in the operation of urban rail systems. It has to provide high‐quality energy for trains continuously. Train's energy consumption is varied directly with operation conditions such as train's headway, train configuration, line characteristic and many other factors. Therefore, using a powerful simulation tool for estimating energy consumption considering these factors seems inevitable. This paper tries to develop a hybrid model in which the train model is based on acquired data from real measurement and an adapted network electrical model. Train empirical model is based on neural network on the basis of online data measured in line one of Shiraz metro and used for training of model. Then benefiting from optimization algorithms, a novel model of energy supply system of trains, including Traction Substations (TSS) and Overhead Catenary System (OCS), is proposed. Finally, the behaviour of trains during operation in 15‐min headway was simulated and the results under same conditions were compared with practical measurements. Currently, this model helps to determine power system behaviour during different operation conditions, predicting energy consumption for different headways, Train's driver behaviour during acceleration, deceleration and braking and also operation risk analysis during traction substations outages.