Software Dependability and Security
Software dependability and security are critical to assuring the resilience of complex systems. Despite decades of work in this area, software remains a weak link in system integrity, leading to failures that compromise safety and/or impose financial costs.
The challenge posed is at once of critical importance and immense. We believe progress is best made through a new approach that focuses on mitigating the types of software bugs that are most difficult to address with conventional methods, and the team we have assembled is singularly well qualified to pursue this path. To meet the challenge, the proposed program will carry out education in the area of software faults, failures and their mitigations in the development cycle and specifically during system operation. We plan to provide a course to graduate students, young researchers and software engineers studying or working in software engineering field that can play an important role in both the master program for electrical and computer engineering and the undergraduate program for interdisciplinary data science at Duke Kunshan University.
Environment Diversity -based Software Fault Tolerance and Its Applications
Modern life depends on devices and systems containing a moderate to significant amount of software whose dependability is critical to the reliability of the device or system. Software fault tolerance has hitherto been based on design diversity, with high implementation costs limiting the scope of application.
Affordable software fault tolerance using the newer notion of environmental diversity is studied in this project, which is predicated on the existence of elusive software faults known as environment dependent bugs with transient characteristics. Here, the environment for a software system is taken to mean the operating systems resources and other concurrently running applications. This project mainly focuses on the following four research aspects: environmental factor identification, key environmental control techniques, environmental diversity-based fault tolerance approaches, and applications to Android systems. Research on failure data analysis, experimental research, analytical modeling and optimization techniques for open source software will be employed. The fruits of the research will effectively contribute to reducing the cost of software fault tolerance, while reducing the impact of environment dependent bugs on software reliability/availability. It will also contribute to the emergence and development of environment dependent bugs related research in software engineering.
Server Failure Caused by Memory Uncorrectable Error
To manage the faults in a complex distributed system, the key challenges include The high dimensional features and small training set (faulty behaviors) make it difficult to model the healthy system using the derived models for diagnosis and detection.
The high-order correlation among the features of normal and abnormal behaviors of the distributed system makes it difficult to classify the two behaviors and to evaluate the potential impact of any failures to the distributed system.
Therefore, we address these challenges by proposing the following techniques:
We use multiple symbolization techniques to discretize and compress long-term time series; we then introduce an unsupervised learning approach by describing several symbol-based clustering and classification methods to analyze the health status of the system.
We propose a Generative-Adversarial-Network based model that can learn from the normal behaviors of the distributed system. The derived generators can capture the high-order correlation among these highly dimensional features in the distributed system, and the derived discriminators can be good at detecting anomalies. The multi-discriminators then form a classifier to isolate and identify the faults. The overall GAN model can test and evaluate the impact of any fault and its propagation. Based on our preliminary research, LSTM-based neural-networks are advocated in the generator/discriminator model that can memorize the complex change of states and the triggering of events in the system and capture their temporal and spatial correlations.
Wireless Physical Layer Algorithm Optimization Based on Machine Learning
5G communication systems generally adopt a massive MIMO system, which dramatically increases the number of variables for channel estimation and equalization algorithms and the complexity of the solution. This project proposes to use machine learning technology to realize and optimize the physical layer algorithms, and reduce the dependence of development communication systems on empirical knowledge. The artificial intelligence technology proposed in this project can significantly reduce the dependence on the channel models, greatly improve the system performance, and is robust under different channels, interference conditions and signal-to-noise ratios.
Automatic Speech Recognition and Speech Synthesis for General Application
In real life, documentary tools, such as shorthand and dictation, are necessary for maintaining an objective record of such things as conferences and court trials. It is important to have a comprehensive record so that conferences and court trials can be retrospectively accurate. Nevertheless, this kind of work is mechanical and tedious. However, a stenographer could be replaced by mature automatic speech recognition (ASR) technology. The recording work would be easily accomplished by ASR technology capable of providing an accurate and real-time record. Similarly, speech synthesis, a technology that synthesizes speech from text, performs well in reading and broadcast areas. Based on a deep neural network, we are cooperating with Kunshan local enterprise to apply ASR technology and speech synthesis to the government system, providing both speech-to-text and text-to-speech services. Wangjin Technology constantly provides live-action data for analyzing, and then we apply it language models to further lower the recognition error rate. Our cooperation aims to reduce the workload of enterprises and governments, and at the same time, to improve work efficiency.
Multimodal Behavior Signal Analysis
We aim to provide multimodal assistive techniques for diagnosis and treatment of behavioral and developmental disorders by focusing on three research topics: capturing behavioral signals, measuring behavioral variables, and predicting behavioral labels. We collaborate with the Child Developmental Behavior Center at the Third Afﬁliated Hospital of Sun Yat-sen University and analyze the audio-visual data collected in a real clinical setting from young children with autism spectrum disorders (ASD). Audio and video sensors can record the child-clinician interaction in a consistent fashion, and state-of-the-art signal processing, artiﬁcial intelligence methods can facilitate quantitative analyses and modeling of various psychology inspired codes and labels. We aim to provide an automated assessment and diagnosis framework of ASD to assist and augment, rather than supplant, an expert clinician.
Audio Speech and Language Processing
We perform research on various kinds of topics in audio, speech and language processing, such as speaker verification, spoken language identification, speaker diarization, speaker age and gender recognition, speaker emotion recognition, speech recognition, keyword spotting, voice conversion, speech synthesis, microphone array, etc. In 2018, we collaborated with multiple companies on several aforementioned topics. Speech signals not only contain lexicon information, they also deliver various kinds of paralinguistic speech attribute information, such as speaker, language, gender, age, emotion, channel, voicing, psychological states, etc. The core technical question of paralinguistic speech attribute recognition is utterance level supervised learning based on text independent or text dependent speech signal with flexible duration. We propose an end-to-end framework with new designs on variable length data loader, frontend convolutional network, encoding (or pooling) layer, loss function, data augmentation, transfer learning and multitask learning.
We aim to build a smart headlight system by realizing pedestrian and head detection, real time object tracking, and light interaction with pedestrians for night driving based on the monocular vision and an LED headlight system. For instance, once a pedestrian is automatically detected at night, our smart headlight system will project the LED headlight onto the pedestrian to highlight him/her. The project has investigated large-scale night driving data on real road scenes using deep learning technology. It will make a big breakthrough in the performance of pedestrian detection at night for autonomous driving systems.
Data Analytics for Smart Manufacturing
To ensure high quality and yield, today’s advanced manufacturing systems are equipped with thousands of sensors to continuously collect measurement data for process monitoring, defect diagnosis and yield learning. In particular, the recent adoption of Industry 4.0 has promoted a set of enabling technologies for low-cost data sensing, processing and storage of manufacturing process. While a large amount of data has been created by the manufacturing industry, statistical algorithms, methodologies and tools are immediately needed to process the complex, heterogeneous and high-dimensional data in order to address the issues posed by process complexity, process variability and capacity constraint. The objective of this project is to explore the enormous opportunities for data analytics in the manufacturing domain and provide data-driven solutions for manufacturing cost reduction.
Single-Channel Real-Time Drowsiness Detection Based on Electroencephalography
Drowsiness is characterized by a low level of consciousness and difficulty in staying awake. The risk of falling asleep at this stage is high, especially when engaged in monotonous tasks such as operating heavy machinery or driving a vehicle. Recently, the American Academy of Sleep Medicine (AASM) issued a statement emphasizing the dangers of driving when drowsy. One can quantify the direct impact of drowsy driving via the large number of accidents caused by a lack of alertness. The US National Highway Traffic Safety Administration estimates that 100,000 police-reported crashes are directly caused by drowsiness. In this work, we propose a novel single-channel, real-time drowsiness detection algorithm that is suitable for portable applications with low computational complexity. Our proposed method measures the average power of the EEG signal in multiple 1-second time windows over eight frequency bands: the delta band (1—3 Hz), the theta band (4—7 Hz), the low-alpha band (8—9 Hz), the high-alpha band (10—12 Hz), the low-beta band (13—17 Hz), the high-beta band (18—30 Hz), the low-gamma band (31—40 Hz) and the high-gamma band (41—50 Hz). Next, it extracts a set of important features by adopting a novel approach based on adaptive counters. These features are eventually used for accurate drowsiness detection.
Intelligent Business Based on Analytics
We closely collaborate with leading domestic industrial enterprises. By using their sales and logistics data, we provide customers with guidance on pricing and discounts on all categories of products. The project is combined with new retailing, using data-driven methodology for all aspects from production to sales, and providing advice on enterprise data management.
High Performance Iris Recognition
Iris recognition has emerged as one of the most accurate and reliable biometric approaches for human recognition. Based on computer vision and machine learning algorithms, this project is developing an automatic iris recognition algorithm that can be used over medium and long distances. The recognition algorithm also has high recognition accuracy in a non-ideal environment, and can accurately identify the user even in the case of the user wearing glasses, or if their illumination changes or signal noise. The recognition accuracy of the project algorithm on the international iris recognition database reaches the world's leading level.