In this study, the design and optimization of high-speed charging systems for supercapacitor banks are investigated to meet the fast charging needs of electric vehicles. The limitations of current energy storage technologies and the requirements for fast charging of electric vehicles are addressed. While examining the advantages and energy storage potentials of supercapacitors, appropriate capacity, and voltage levels are determined in the design process to increase the charging speed. Additionally, power electronic components and control strategies are discussed to enhance the efficiency of the charging system. These components and strategies aim to optimize the fast charging process by enabling supercapacitors to be charged with maximum efficiency. The effectiveness and optimization potential of the design are evaluated and discussed through simulations conducted using the Matlab/Simulink software program. This article emphasizes the significance of supercapacitor-based high-speed charging systems in meeting the fast charging needs of electric vehicles and provides a foundation for future advancements. With the progress of electric vehicle technologies, supercapacitor-based charging systems aim to enable faster and more efficient charging, thereby further popularizing the use of electric vehicles.
This article proposes two methods based on deep learning for estimating time-to-failure (TTF) of an industrial system using its degradation image. This provides an effective tool for predictive maintenance practitioners toward digitization of maintenance processes in Industry 4.0 transformation. Both methods utilize the long short-term memory (LSTM) networks for capturing temporal information. First methodology consists of two convolutional layers preceding a single LSTM layer to extract compact information from the individual images and rescue LSTM network from curse of dimensionality. Then, an LSTM layer estimates the TTF value from the extracted features. In the second approach, the dimension of the individual images are decreased by a fully connected neural network, which is trained as an autoencoder. A separate LSTM network is trained and run over this lower dimensional space. The strength of suggested architectures is shown using simulation data and a dataset of infrared image streams collected from rotating machinery. The performance comparison of proposed methods and other methods is also provided.
Bu çalışmada, konuşma işaretlerini sıkıştırmak için derin öğrenme tabanlı oto kodlayıcı ve artık vektör nicemlemesini temel alan sıkıştırma yöntemi önerilmiştir. Önerilen sıkıştırma yönteminde, öncelikle giriş konuşma işaretini daha düşük boyutlu bir uzaya atayan oto kodlayıcı kullanılmakta ve ardından oto kodlayıcı çıkışı, artık vektör nicemlemesi ile daha da sıkıştırılmaktadır. Sıkıştırma yöntemi, birbirine paralel çalışan iki farklı kod çözücü yapısı ve iki kod kitapçığı sayesinde farklı oranlarda sıkıştırma oranı sunmaktadır. Yöntemin başarımı konuşma kalitesini algısal değerlendirme metriği kullanılarak TIMIT veri kümesi ile test edilmiştir. Önerilen konuşma sıkıştırma yöntemi, 1.25 ve 2.5 kbps iletim hızları için sırasıyla 1.665 ve 1.985 konuşma kalitesini algısal değerlendirme skorları elde etmiştir.
Deep learning has been studied extensively for driver drowsiness detection using video data. However, since the proposed deep learning methods are computationally cumbersome, the commercial driver drowsiness detection methods are still using hand-crafted features such as lane deviation and percentage of eye closure. This study investigates a deep learning model that provides a fair drowsiness detection performance with a lightweight architecture. In the proposed method, Dlib library was used to detect the driver's face in individual frames of video data. The detected faces are fed into a pre-defined convolutional neural network architecture. Then, a long short-term memory network was used to capture the temporal information between the frame sequences to assess the state of drowsiness. The proposed model achieves a detection accuracy of 80% in a popular benchmark dataset. It was also verified that the model could be implemented on a commercial and inexpensive development board with a frame rate of 5 frames per second.
This study proposes a novel long short-term memory based sequence-to-sequence autoencoder model to compress ECG signals. The efficiency of this new method is illustrated on MIT-BIH Arrhythmia dataset. In the conducted experiments, the proposed architecture achieves %21.14 mean-independent percentage mean square difference (MPRD) with a constant compression ratio value of 33 : 1.
Industrial systems with multiple subsystems are monitored via various sensors to control the ongoing process. If the number of monitoring signals collected from these sensors is high and the number of faulty samples is low, then the machine learning methods may fail to provide effective solutions for fault detection and root cause identification. This paper proposes an efficient feature selection model based on the regularized LSTM neural networks, and fault detection and classification using an ensemble of binary LSTM classifiers. The model is verified in PHME Data Challenge 2021 which provides quality-control-pipeline monitoring data.
In the new data intensive world, predictive maintenance has become a central issue for the modern industrial plants. Monitoring of electric machinery is one of the most important challenges in predictive maintenance. Adaptive manufacturing processes/plants may be possible through the monitored conditions. In this respect, several attempts have been made to utilize deep learning algorithms for rotating machinery fault detection and diagnosis. Among them, deep autoencoders are very popular, because of their denoising effect. They are also implemented in electric machinery fault diagnostics in order to obtain lower order representation of signals. However, none of these efforts regard the autoencoders as compression units. Bearing in mind that spectra of vibration and current signals that are collected from electric machinery are critical instruments for detection and diagnosis of their faults, we propose that deep stacked autoencoder can be utilized as spectrum compression units. The performance of the proposed strategy are assessed using a bearing data set in three ways: (1)Rule-based classifiers are implemented on raw and compressed-decompressed spectrum and their performance are compared. (2) It is shown that the several machine learning classifiers such as support vector machines, artificial neural networks and k-nearest neighbour classifiers on compressed-decompressed spectrum achieves the performance of them on raw data. (3) A multi-layer perceptron (MLP) classifier is implemented on the low dimensional representation and it is demonstrated that the strategy of employing the same autoencoder as pretraining of feature extraction module cannot outperform the performance of this MLP classifier.
