Over the last years, the computational power of mobile devices such as smartphones and tablets has grown dramatically, reaching the level of desktop computers available not long ago. While standard smartphone apps are no longer a problem for them, there is still a group of tasks that can easily challenge even high-end devices, namely running artificial intelligence algorithms. In this paper, we present a study of the current state of deep learning in the Android ecosystem and describe available frameworks, programming models and the limitations of running AI on smartphones. We give an overview of the hardware acceleration resources available on four main mobile chipset platforms: Qualcomm, HiSilicon, MediaTek and Samsung. Additionally, we present the realworld performance results of different mobile SoCs collected with AI Benchmark 1 that are covering all main existing hardware configurations. * We also thank Przemyslaw Szczepaniak (pszczepaniak@google.com), Google Inc., for writing and editing sections 2.7, 3.1 and 3.2. 1

No abstract

Rear-end collision crash is one of the most common accidents on the road. Accurate driving style recognition considering rear-end collision risk is crucial to design useful driver assistance systems and vehicle control systems. The purpose of this study is to develop a driving style recognition method based on vehicle trajectory data extracted from the surveillance video. First, three rear-end collision surrogates, Inversed Time to Collision (ITTC), Time-Headway (THW), and Modified Margin to Collision (MMTC), are selected to evaluate the collision risk level of vehicle trajectory for each driver. The driving style of each driver in training data is labelled based on their collision risk level using K-mean algorithm. Then, the driving style recognition model’s inputs are extracted from vehicle trajectory features, including acceleration, relative speed, and relative distance, using Discrete Fourier Transform (DFT), Discrete Wavelet Transform (DWT), and statistical method to facilitate the driving style recognition. Finally, Supporting Vector Machine (SVM) is applied to recognize driving style based on the labelled data. The performance of Random Forest (RF), K-Nearest Neighbor (KNN), and Multi-Layer Perceptron (MLP) is also compared with SVM. The results show that SVM overperforms others with 91.7% accuracy with DWT feature extraction method.

No abstract

No abstract

No abstract

Dissolved gas analysis (DGA) is widely used to detect the incipient fault of power transformers. However, the accuracy is greatly limited by selection of DGA features and performance of fault diagnostic model. This paper proposed a fault diagnostic method integrating feature selection and diagnostic model optimization. Firstly, this paper set up three feature sets with eight basic DGA gases, 28 DGA gas ratios and 36 hybrid DGA features, respectively. Then, to eliminate the interference of weak-relevant and irrelevant features, the genetic-algorithm-SVM-feature-screen (GA-SVM-FS) model was built to screen out the optimal hybrid DGA features subset (OHFS) from three feature sets. Next, using the OHFS as the input, the support vector machine (SVM) multi-classifier optimized by ISGOSVM (SVM classifier optimized by improved social group optimization) was built to diagnose fault types of transformers. Finally, the performance of OHFS and ISGOSVM diagnostic model was tested and compared with traditional DGA features and diagnostic models, respectively. The results show that the OHFS screened out is comprised of 14 features, including 12 gas ratios and two gases. The accuracy of OHFS is 3–30% higher than traditional DGA features, and the accuracy of ISGOSVM can increase by 3% to 14% compared with the SGOSVM (SVM classifier optimized by social group optimization), GASVM (SVM classifier optimized by genetic algorithm optimization), PSOSVM (SVM classifier optimized by particle swarm optimization), and SVM diagnostic models. The proposed approach integrating the OHFS with ISGOSVM achieves the highest accuracy of fault diagnose (92.86%).

No abstract