An unsupervised approach for fault diagnosis of power transformers

Electrical utilities apply condition monitoring on power transformers (PTs) to prevent unplanned outages and detect incipient faults. This monitoring is often done using dissolved gas analysis (DGA) coupled with engineering methods to interpret the data, however the obtained results lack accuracy and reproducibility. In order to improve accuracy, various advanced analytical methods have been proposed in the literature. Nonetheless, these methods are often hard to interpret by the decision-maker and require a substantial amount of failure records to be trained. In the context of the PTs, failure data quality is recurrently questionable, and failure records are scarce when compared to nonfailure records. This work tackles these challenges by proposing a novel unsupervised methodology for diagnosing PT condition. Differently from the supervised approaches in the literature, our method does not require the labeling of DGA records and incorporates a visual representation of the results in a 2D scatter plot to assist in interpretation. A modified clustering technique is used to classify the condition of different PTs using historical DGA data. Finally, well-known engineering methods are applied to interpret each of the obtained clusters. The approach was validated using data from two different real-world data sets provided by a generation company and a distribution system operator. The results highlight the advantages of the proposed approach and outperformed engineering methods (from IEC and IEEE standards) and companies legacy method. The approach was also validated on the public IEC TC10 database, showing the capability to achieve comparable accuracy with supervised learning methods from the literature. As a result of the methodology performance, both companies are currently using it in their daily DGA diagnosis.     

Predicting hydrogen adsorption and desorption rates in cylindrical metal hydride beds: Empirical correlations and machine learning

A major limitation in chemisorptive hydrogen storage in metal hydrides is the long time required for the adsorption reaction during charging. This study investigates how the shape and material of the reaction chamber influences the adsorption and desorption rates. Numerical simulations of hydrogen storage in a cylindrical reaction chamber filled with LaNi₅ hydride are conducted for a range of chamber thermal conductivities and aspect ratios. The results show that adsorption and desorption processes are limited by thermal diffusion in the hydride bed and storage chamber. A storage efficiency is proposed based on an ideal isothermal process and used to evaluate the impact of chamber thermal conductivity and aspect ratio on the adsorption and desorption rates. Empirical correlations are proposed for predicting the adsorption and desorption efficiency of cylindrical LaNi₅ hydride beds. Finally, a machine-learning based data model for predicting storage efficiency in metal hydride chambers is presented. Comparison against the empirical correlations highlights that the machine learning-based data model can predict the storage efficiency more accurately.     

Deep learning based pipeline with multichannel inputs for patent classification

This work introduces a deep learning pipeline for automatic patent classification with multichannel inputs based on LSTM and word vector embeddings. Sophisticated text mining methods are used to extract the most important segments from patent texts, and a domain-specific pre-trained word embeddings model for the patent domain is developed; it was trained on a very large dataset of more than five million patents. The deep learning pipeline is using multiple parallel LSTM networks that read the source patent document using different input dimensions namely embeddings of different segments of patent texts, and sparse linear input of different metadata. Classifying patents into corresponding technical fields is selected as a use case. In this use case, a series of patent classification experiments are conducted on different patent datasets, and the experimental results indicate that using the segments of patent texts as well as the metadata as multichannel inputs for a deep neural network model, achieves better performance than one input channel.     

The impact of hydrogen refueling station subsidy strategy on China's hydrogen fuel cell vehicle market diffusion

Lack of hydrogen refueling stations (HRSs) has hindered the diffusion of hydrogen fuel cell vehicles (HFCVs) in the Chinese transport market. By combining the agent-based model (ABM) and the experience weighted attraction (EWA) learning algorithm, this paper explores the impact of government subsidy strategy for HRSs on the market diffusion of HFCVs. The actions of the parties (government, HRS planning department and consumers) and their interactions are taken into account. The new model suggests dynamic subsidy mode based on EWA algorithm yields better results than static subsidy mode: HFCV purchases, HRS construction effort, total number of HRSs and expected HRS planning department profits all outperform static data by around 27%. In addition, choosing an appropriate initial subsidy strategy can increase the sales of HFCVs by nearly 40%. Early investment from government to establish initial HRSs can also increase market diffusion efficiency by more than 76.7%.     

Predicting elastic modulus of porous La0.6Sr0.4Co0.2Fe0.8O3-δ cathodes from microstructures via FEM and deep learning

In this work, a deep learning accelerated homogenization framework is developed for prediction of elastic modulus of porous materials directly from their inner microstructures. The finite element method (FEM) and the homogenization theory are used to obtain the macroscopic properties of materials based on their microstructures. Based on a large dataset consisting of various microstructures and corresponding elastic properties via FEM, a deep convolutional neural network (CNN) is trained to capture the nonlinear functional relationship between the microstructure features and their macroscopic elastic properties. The deep learning model is finally well validated against extra new samples with excellent predictive performances. This demonstrates that the CNN deep learning model can be trusted as a surrogate model for the FEM based homogenization method, with the computation time being reduced by several orders of magnitude. The proposed deep learning framework is highly extendable for prediction of various macroscopic properties from microstructures.     

A Hybrid Virtual Cloud Learning Model during the COVID-19 Pandemic

The COVID-19 pandemic has affected the educational systems worldwide, leading to the near-total closures of schools, universities, and colleges. Universities need to adapt to changes to face this crisis without negatively affecting students’ performance. Accordingly, the purpose of this study is to identify and help solve to critical challenges and factors that influence the e-learning system for Computer Maintenance courses during the COVID-19 pandemic. The paper examines the effect of a hybrid modeling approach that uses Cloud Computing Services (CCS) and Virtual Reality (VR) in a Virtual Cloud Learning Environment (VCLE) system. The VCLE system provides students with various utilities and educational services such as presentation slides/text, data sharing, assignments, quizzes/tests, and chatrooms. In addition, learning through VR enables the students to simulate physical presence, and they respond well to VR environments that are closer to reality as they feel that they are an integral part of the environment. Also, the research presents a rubric assessment that the students can use to reflect on the skills they used during the course. The research findings offer useful suggestions for enabling students to become acquainted with the proposed system’s usage, especially during the COVID-19 pandemic, and for improving student achievement more than the traditional methods of learning.     

Soft sensing of water depth in combined sewers using LSTM neural networks with missing observations

Information and communication technologies combined with in-situ sensors are increasingly being used in the management of urban drainage systems. The large amount of data collected in these systems can be used to train a data-driven soft sensor, which can supplement the physical sensor. Artificial Neural Networks have long been used for time series forecasting given their ability to recognize patterns in the data. Long Short-Term Memory (LSTM) neural networks are equipped with memory gates to help them learn time dependencies in a data series and have been proven to outperform other type of networks in predicting water levels in urban drainage systems. When used for soft sensing, neural networks typically receive antecedent observations as input, as these are good predictors of the current value. However, the antecedent observations may be missing due to transmission errors or deemed anomalous due to errors that are not easily explained. This study quantifies and compares the predictive accuracy of LSTM networks in scenarios of limited or missing antecedent observations. We applied these scenarios to an 11-month observation series from a combined sewer overflow chamber in Copenhagen, Denmark. We observed that i) LSTM predictions generally displayed large variability across training runs, which may be reduced by improving the selection of hyperparameters (non-trainable parameters); ii) when the most recent observations were known, adding information on the past did not improve the prediction accuracy; iii) when gaps were introduced in the antecedent water depth observations, LSTM networks were capable of compensating for the missing information with the other available input features (time of the day and rainfall intensity); iv) LSTM networks trained without antecedent water depth observations yielded larger prediction errors, but still comparable with other scenarios and captured both dry and wet weather behaviors. Therefore, we concluded that LSTM neural network may be trained to act as soft sensors in urban drainage systems even when observations from the physical sensors are missing.     

Using deep learning to combine static and dynamic power analyses of cryptographic circuits

Side-channel attacks have shown to be efficient tools in breaking cryptographic hardware. Many conventional algorithms have been proposed to perform side-channel attacks exploiting the dynamic power leakage. In recent years, with the development of processing technology, static power has emerged as a new potential source for side-channel leakage. Both types of power leakage have their advantages and disadvantages. In this work, we propose to use the deep neural network technique to combine the benefits of both static and dynamic power. This approach replaces the classifier in template attacks with our proposed long short-term memory network schemes. Hence, instead of deriving a specific probability density model for one particular type of power leakage, we gain the ability of combining different leakage sources using a structural algorithm. In this paper, we propose three schemes to combine the static and dynamic power leakage. The performance of these schemes is compared using simulated test circuits designed with a 45-nm library.     

退火炉内罩中气体流动及钢卷传热的数值模拟

摘要：轧制力预报一直是热连轧过程控制模型的核心，浅层神经网络对复杂函数的表示能力有限，而深度学习模型通过学习一种深层非线性网络结构，实现复杂函数逼近。利用深度学习框架TensorFlow，构建了一种深度前馈神经网络轧制力模型，采用BP算法计算网络损失函数的梯度，运用融入Mini batch策略的Adam优化算法进行参数寻优，采用Early stopping、参数惩罚和Dropout正则化策略提高模型的泛化能力。基于上述建模策略，针对宝钢1880热连轧精轧机组的大量轧制历史数据进行了建模实验，对比分析了4种不同结构的前馈网络预测精度。结果表明，相比于传统SIMS轧制力模型，深度神经网络可实现轧制力的高精度预测，针对所有机架的预测精度平均提升21.11%。     

The cumulative power of incremental innovation and the role of project sequence management

In innovation and project management studies incremental development projects are perceived as theoreticlly and organisationally uninteresting. By means of a longitudinal study of product improvement projects at an automobile firm, this paper challenges such views and shows how the cumulative impact of the studied sequence resulted in a competitive repositioning of the company's product portfolio during a financially difficult period. Project managers achieved this by transcending the separation between exploration and exploitation projects; they not only adhered to time, cost and quality goals but also tried out new ways of testing and experimenting with controversial technical ideas. The paper analyzes the intensive inter project learning that generated these ambidextrous capabilities and emphasizes that practices at the project-level need to be buttressed by expanded management learning and capability development also at the sequence level.