Physical Sciences: An Inspiration to the Neural Network Training

The goal of the paper is to create a training model based on real raw noisy data and train a neural network to determine the behavior of the fuel level, namely, to determine the time and volume of vehicle refueling, fuel consumption / excessive consumption / drainage.Various algorithms and data processing methods are used in fuel control and metering systems to get rid of noise. In some systems, primary filtering is used by excluding readings that are out of range, sharp jumps and deviations, and averaging over a sliding window. Research is being carried out on the use of more complex filters than simple averaging – by example, the Kalman filter for data processing.When measuring the fuel level using various fuel level sensor the data is influenced by many external factors that can interfere with the measurement and distort the real fuel level. Since these interferences are random and have a different structure, it is very difficult to completely remove them using classical noise suppression algorithms. Therefore, we use artificial intelligence, namely a neural network, to find patterns, detect noise and correct distorted data. To correct distorted data, you first need to determine which data is distorted, classify the data.In the course of the work, the raw data on the fuel level were transformed for use in the neural network training model. To describe the behavior of the fuel level, we use 4 possible classes: fuel consumption is observed, the vehicle is refueled, the fuel level does not change (the vehicle is idle), the data is distorted by noise. Also, in the process of work, additional tools of the DeepLearning4 library were used to load data training and training a neural network. A multilayer neural network model is used, namely a three-layer neural network, as well as used various training parameters provided by the DeepLearning4j library, which were obtained because of experiments.After training the neural network was used on test data, because of which the Confusion Matrix and Evaluation Metrics were obtained.In conclusion, finding a good model takes a lot of ideas and a lot of experimentation, also need to correctly process and transform the raw data to get the correct data for training. So far, a neural network has been trained to determine the state of the fuel level at a point in time and classify the behavior into four main labels (classes). Although we have not reduced the error in determining the behavior of the fuel level to zero, we have saved the states of the neural network, and in the future we will be able to retrain and evolve our neural network to obtain better results.

Download Full-text

TRAINING NEURAL NETWORK FOR TAXI PASSENGER DEMAND FORECASTING USING GRAPHICS PROCESSING UNITS

Ukrainian Journal of Information Technology ◽

10.23939/ujit2020.02.029 ◽

2020 ◽

Vol 2 (1) ◽

pp. 29-36

Author(s):

M. I. Zghoba ◽

◽

Yu. I. Hrytsiuk ◽

Keyword(s):

Neural Network ◽

Graphics Processing Units ◽

Neural Network Training ◽

Training Time ◽

Passenger Demand ◽

Network Training ◽

The Neural Network ◽

Input Dataset ◽

Speed Up ◽

Graphics Processing

The peculiarities of neural network training for forecasting taxi passenger demand using graphics processing units are considered, which allowed to speed up the training procedure for different sets of input data, hardware configurations, and its power. It has been found that taxi services are becoming more accessible to a wide range of people. The most important task for any transportation company and taxi driver is to minimize the waiting time for new orders and to minimize the distance from drivers to passengers on order receiving. Understanding and assessing the geographical passenger demand that depends on many factors is crucial to achieve this goal. This paper describes an example of neural network training for predicting taxi passenger demand. It shows the importance of a large input dataset for the accuracy of the neural network. Since the training of a neural network is a lengthy process, parallel training was used to speed up the training. The neural network for forecasting taxi passenger demand was trained using different hardware configurations, such as one CPU, one GPU, and two GPUs. The training times of one epoch were compared along with these configurations. The impact of different hardware configurations on training time was analyzed in this work. The network was trained using a dataset containing 4.5 million trips within one city. The results of this study show that the training with GPU accelerators doesn't necessarily improve the training time. The training time depends on many factors, such as input dataset size, splitting of the entire dataset into smaller subsets, as well as hardware and power characteristics.

Download Full-text

Power Quality Disturbance Identification and Optimization Based on Machine Learning

Distributed Generation & Alternative Energy Journal ◽

10.13052/dgaej2156-3306.3723 ◽

2021 ◽

Author(s):

Fei Long ◽

Fen Liu ◽

Xiangli Peng ◽

Zheng Yu ◽

Huan Xu ◽

...

Keyword(s):

Neural Network ◽

Power Quality ◽

Classification Method ◽

Neural Network Training ◽

Training Set ◽

Network Training ◽

The Neural Network ◽

Recognition Ability ◽

Power Quality Disturbance ◽

Depth Learning

In order to improve the electrical quality disturbance recognition ability of the neural network, this paper studies a depth learning-based power quality disturbance recognition and classification method: constructing a power quality perturbation model, generating training set; construct depth neural network; profit training set to depth neural network training; verify the performance of the depth neural network; the results show that the training set is randomly added 20DB-50DB noise, even in the most serious 20dB noise conditions, it can reach more than 99% identification, this is a tradition. The method is impossible to implement. Conclusion: the deepest learning-based power quality disturbance identification and classification method overcomes the disadvantage of the selection steps of artificial characteristics, poor robustness, which is beneficial to more accurately and quickly discover the category of power quality issues.

Download Full-text

Automatic building footprint extraction from UAV images using neural networks

Geodetski vestnik ◽

10.15292/geodetski-vestnik.2020.04.545-561 ◽

2020 ◽

Vol 64 (04) ◽

pp. 545-561

Author(s):

Zoran Kokeza ◽

Miroslav Vujasinović ◽

Miro Govedarica ◽

Brankica Milojević ◽

Gordana Jakovljević

Keyword(s):

Neural Network ◽

Surface Model ◽

Neural Network Training ◽

Implementation Evaluation ◽

Network Training ◽

Area Of Interest ◽

Cadastral Maps ◽

The Neural Network ◽

Building Footprint ◽

Uav Images

Up-to-date cadastral maps are crucial for urban planning. Creating those maps with the classical geodetic methods is expensive and time-consuming. Emerge of Unmanned Aerial Vehicles (UAV) made a possibility for quick acquisition of data with much more details than it was possible before. The topic of the research refers to the challenges of automatic extraction of building footprints on high-resolution orthophotos. The objectives of this study were as follows: (1) to test the possibility of using different publicly available datasets (Tanzania, AIRS and Inria) for neural network training and then test the generalisation capability of the model on the Area Of Interest (AOI); (2) to evaluate the effect of the normalised digital surface model (nDSM) on the results of neural network training and implementation. Evaluation of the results shown that the models trained on the Tanzania (IoU 36.4%), AIRS (IoU 64.4%) and Inria (IoU 7.4%) datasets doesn't satisfy the requested accuracy to update cadastral maps in study area. Much better results are achieved in the second part of the study, where the training of the neural network was done on tiles (256x256) of the orthophoto of AOI created from data acquired using UAV. A combination of RGB orthophoto with nDSM resulted in a 2% increase of IoU, achieving the final IoU of over 90%.

Download Full-text

Neural Network-WPCA Based Method for Multi-Objective Optimal Redundancy Allocation

International Journal of Reliability Quality and Safety Engineering ◽

10.1142/s0218539317500243 ◽

2017 ◽

Vol 24 (05) ◽

pp. 1750024

Author(s):

Yuanchen Fang ◽

Huyang Xu ◽

Nasser Fard

Keyword(s):

Neural Network ◽

Design Method ◽

Principal Component ◽

Estimation Accuracy ◽

Neural Network Training ◽

Performance Measurements ◽

Redundancy Allocation ◽

Training Performance ◽

Network Training ◽

The Neural Network

For systems with multiple redundancies, reliability evaluation in the redundancy allocation problem (RAP) constitutes a computational complexity. It has been demonstrated that neural network training provides an efficient approach to estimate the complex system reliability function. When executing the neural network algorithm, there are many parameters that need to be determined for improving the training performance. Therefore, robust experimental design method can be used to determine the neural network parameters. The traditional robust design methods are intended for a single response variable. However, the application of neural network method includes more than one measurement, such as estimation accuracy and time efficiency. In this paper, utility function is first estimated by neural network training, in which the algorithm parameters are determined by weighted principal component (WPCA)-based multi-response optimization which simultaneously optimizes more than one training performance measurements. Moreover, it is always desirable to simultaneously optimize several objectives in designing a system, such as reliability, cost, etc. Therefore, continuous WPCA-based multi-response design is then applied to obtain the best design of redundancies in RAP, which simultaneously optimize multiple objectives by taking into account the correlations between them.

Download Full-text

Impact of Asymmetric Weight Update on Neural Network Training With Tiki-Taka Algorithm

Frontiers in Neuroscience ◽

10.3389/fnins.2021.767953 ◽

2022 ◽

Vol 15 ◽

Author(s):

Chaeun Lee ◽

Kyungmi Noh ◽

Wonjae Ji ◽

Tayfun Gokmen ◽

Seyoung Kim

Keyword(s):

Neural Network ◽

Power Efficiency ◽

Calibration Method ◽

Neural Network Training ◽

Systematic Analysis ◽

Training Performance ◽

Actual Weight ◽

Network Training ◽

The Neural Network ◽

The Impact

Recent progress in novel non-volatile memory-based synaptic device technologies and their feasibility for matrix-vector multiplication (MVM) has ignited active research on implementing analog neural network training accelerators with resistive crosspoint arrays. While significant performance boost as well as area- and power-efficiency is theoretically predicted, the realization of such analog accelerators is largely limited by non-ideal switching characteristics of crosspoint elements. One of the most performance-limiting non-idealities is the conductance update asymmetry which is known to distort the actual weight change values away from the calculation by error back-propagation and, therefore, significantly deteriorates the neural network training performance. To address this issue by an algorithmic remedy, Tiki-Taka algorithm was proposed and shown to be effective for neural network training with asymmetric devices. However, a systematic analysis to reveal the required asymmetry specification to guarantee the neural network performance has been unexplored. Here, we quantitatively analyze the impact of update asymmetry on the neural network training performance when trained with Tiki-Taka algorithm by exploring the space of asymmetry and hyper-parameters and measuring the classification accuracy. We discover that the update asymmetry level of the auxiliary array affects the way the optimizer takes the importance of previous gradients, whereas that of main array affects the frequency of accepting those gradients. We propose a novel calibration method to find the optimal operating point in terms of device and network parameters. By searching over the hyper-parameter space of Tiki-Taka algorithm using interpolation and Gaussian filtering, we find the optimal hyper-parameters efficiently and reveal the optimal range of asymmetry, namely the asymmetry specification. Finally, we show that the analysis and calibration method be applicable to spiking neural networks.

Download Full-text

High-resolution soil moisture retrieval using a Neural Network approach from Sentinel-1 SAR data

10.5194/egusphere-egu2020-9995 ◽

2020 ◽

Author(s):

Qi Gao ◽

Maria Jose Escorihuela ◽

Nemesio Rodriguez-Fernandez ◽

Olivier Merlin ◽

Mehrez Zribi

Keyword(s):

Neural Network ◽

Soil Moisture ◽

High Resolution ◽

Neural Network Training ◽

Neural Network Approach ◽

Auxiliary Data ◽

Network Training ◽

The Neural Network ◽

Sar Data

High-resolution soil moisture product is important for agriculture-related managements including irrigation. We have investigated the Change Detection (CD) method using Sentinel-1 data for 100 m resolution soil moisture retrieval and got a Root Mean Square Error (RMSE) about 0.6 m3/m3. However, the result of this approach is not accurate enough for high-density crops like corn. Another approach needs to be studied to get better accuracy over all types of crops. The artificial neural network (NN) technique, which involves nonlinear parameterized mapping from an input vector to an output vector, is an appropriate tool for retrieving geophysical parameters from remote sensing data. Many studies have explored the NN approach for processing remotely sensed data, including retrieving soil moisture, however, only a few studies [Notarnicola et al., 2010; Paloscia et al., 2013, etc.] had investigated NN for soil moisture estimation over vegetation-covered areas, especially in a large scale.The objective of this study is to develop an approach based on neural networks to estimate soil moisture at high resolution over vegetation-covered areas from Sentinel-1 C-band SAR data. The quality of the output results depends directly on the quality of the input data used to train the NN and the reference data for the training, therefore, we performed our study over Catalonia, where we have many auxiliary data. The study is performed using both VV and VH polarization over the whole Catalonia. Apart from Sentinel-1 SAR data, auxiliary data including Sentinel-2 NDVI, SMAP soil moisture, CCI (ESA Climate Change Initiative) land cover, SIGPAC (Sistema de Informaci&#243;n Geogr&#225;fica de Parcelas Agr&#237;colas) land cover, irrigation index and crop type information from SIGPAC, and DEM (Digital elevation model) are also used for approach development. DISPATCH (Disaggregation based on Physical and Theoretical scale Change) soil moisture product at 1 km resolution is considered as the target in the Neural Network training, adding great value to our study. To prepare the Neural Network training, all data sets are co-registered at 1 km resolution within the same size and resampled for the same dates within one year (2017). Two indexes describing the normalized backscatter difference and soil moisture are introduced as equation (1) and (2):<table><tbody><tr><td>Index1 = (&#963;0i - &#963;0min) / (&#963;0max - &#963;0min)</td> <td>(1)</td> </tr><tr><td>Index2 = SMmin + (SMmax - SMmin) * Index1</td> <td>(2)</td> </tr></tbody></table>&#160;Different parameters were tested to train the Neural Network approach, the preliminary results show a correlation value compared with DISPATCH product about 0.71 over croplands, 0.73 over irrigated fields, and 0.65 over forests, considering Index1, Index2 and SMAP soil moisture. Works are still on-going to try to improve the results by better analyzing the SAR data performance over different fields and conditions. The final goal of the study is to produce 100 m resolution soil moisture product. After 1 km resolution study, we will apply the approach at 100m resolution, and the in-situ soil moisture will be used for validation.This work is inscribed within the Water4Ever project, which is funded by the European Commission under the framework of the ERA-NET COFUND WATERWORKS 2015 Programme.&#160;

Download Full-text

OPTIMIZATION OF THE INPUT DATA VECTOR TO IMPROVE THE NEURAL NETWORK TRAINING FOR OPC

10.29003/m2493.mmmsec-2021/137-140 ◽

2021 ◽

Author(s):

Georgiy Teplov ◽

Almira Galeeva ◽

Aleksey Kuzovkov

Keyword(s):

Neural Network ◽

Data Structure ◽

Input Data ◽

Input Vector ◽

Neural Network Training ◽

Network Training ◽

The Neural Network ◽

Data Vector ◽

The Impact

This work explored the impact of input data structure to improve the neural network training. The impact of two variants of the input data vector on the training accuracy of the network was studied. The first version of the input vector included the intensity of the exposure radiation map. The second version of the input vector included the intensity of the exposure radiation map and IC topology.

Download Full-text

Polarization-Based Haze Removal Using Self-Supervised Network

Frontiers in Physics ◽

10.3389/fphy.2021.789232 ◽

2022 ◽

Vol 9 ◽

Author(s):

Yingjie Shi ◽

Enlai Guo ◽

Lianfa Bai ◽

Jing Han

Keyword(s):

Neural Network ◽

Gradient Descent ◽

Transmission Model ◽

Neural Network Training ◽

Original Image ◽

Atmospheric Transmission ◽

Atmospheric Scattering ◽

Haze Removal ◽

Network Training ◽

The Neural Network

Atmospheric scattering caused by suspended particles in the air severely degrades the scene radiance. This paper proposes a method to remove haze by using a neural network that combines scene polarization information. The neural network is self-supervised and online globally optimization can be achieved by using the atmospheric transmission model and gradient descent. Therefore, the proposed method does not require any haze-free image as the constraint for neural network training. The proposed approach is far superior to supervised algorithms in the performance of dehazing and is highly robust to the scene. It is proved that this method can significantly improve the contrast of the original image, and the detailed information of the scene can be effectively enhanced.

Download Full-text