Search

Joint human activity localization and recognition has broad application prospects in human-computer interaction, virtual reality, smart healthcare system, security monitoring and robotics. Ultra-wideband (UWB) is an emerging technology adopted in real-time location system (RTLS) and has shown satisfactory performance in the task of human activity localization. However, few studies have been carried out to simultaneously recognize human activities based on UWB RTLS, which limits the use of UWB RTLS in many applications. In this study, we develop a RTLS based on UWB for the joint task of activity localization and recognition. A compressed sensing-based activity recognition approach is proposed for the task of activity recognition and several machine learning methods are designed to further improve the activity localization accuracy for the task of activity localization. The experimental results show that our UWB RTLS achieves good performance in this joint task.

High-precision indoor positioning has nowadays emerged as a critical function for many applications. However, many existing indoor positioning systems either fail to achieve a high positioning accuracy or are very easily affected by indoor environments composed of many obstacles, preventing them from satisfying many application requirements. Ultra wideband (UWB) has recently drawn extensive attention in the field of indoor positioning due to its great ability to achieve high ranging and localization accuracy while minimizing the effect of multipath interference. Meanwhile, advanced artificial intelligence and signal processing techniques have been explored to improve the precision and performance of indoor positioning system. In this paper, a high-precision UWB indoor positioning system integrating artificial intelligence and signal processing techniques is designed. And field tests are also conducted to validate the design of the system.

The application of compressed sensing and machine learning in power quality disturbance (PQD) classification has drawn more and more attention. This paper presents an adaptive compressed sensing and machine learning (ACSML) method to classify both single PQD and combined PQDs with the consideration of the correlation and the sparsity properties in the PQD signals. This method first uses random projection to reduce the dimensionality of the PQD signals. Meanwhile, a simplified near neighbor algorithm is proposed to reduce the size of the required PQD signal training dataset. The PQD classification problem is finally solved using an adaptive compressed sensing classification algorithm. Experiment results show that the proposed ACSML method achieves higher classification accuracy and faster classification speed in classifying PQDs than the other existing compressed sensing based PQD classification methods.

Indoor positioning system plays an important role in smart grid. Although GPS is the predominant outdoor positioning technology, it is unsuitable to be used in many fields of smart grid for three main reasons: first, signals sent from GPS could easily get blocked by solid materials such as metal or brick; second, the complex electromagnetic interference induced by electrical circuits greatly affects GPS signals; third, GPS can only achieve meter-level real time positioning accuracy, which is far from sufficient for many requirements of smart grid applications. Some other indoor positioning technologies, such as Bluetooth, Wi-Fi, ultrasound, infrared and RFID, fail in either the positioning accuracy, the positioning range, or the positioning speed required in many smart grid applications. Therefore, this paper proposes a real time indoor positioning system for smart gird based on a more promising technology, ultra-wideband (UWB). UWB is suitable for real-time localization in smart grid because UWB has short radio frequency pulse duration and wide bandwidth, which can minimize the effects of multipath interference and allow for high-resolution ranging and easier material penetration. In addition, since high-accuracy position information is required in many smart grid fields, a comprehensive framework integrating several artificial intelligence techniques, including outlier detection, line-of-sight/non-line-of-sight classification, filter design, range measurement correction and maximum likelihood localization estimation, is also proposed to further improve the positioning accuracy. At last, the performance of this system is verified through a series of experiments.

Compressive sensing related methods have achieved success in power system fault classification. However, those methods, either only focus on sparsity without considering the correlation within power system fault signals, or emphasize too much on the correlation but neglect the sparsity property of power system fault signals, resulting in the decrease of classification precision and the increase of the number of training samples needed. In addition, those methods do not explore other machine learning techniques at the same time to further improve the classification accuracy and speed. To overcome the drawbacks and limitations of those methods, this paper proposes an adaptive compressive sensing and machine learning classification framework to quickly and accurately classify power system faults considering both the correlation and the sparsity properties inherited in power system fault signals. First, random projection is applied to reduce the dimensionality of the fault signal. Second, a k-nearest neighbor based algorithm is designed to reduce the number of fault signals needed in the training dataset. Finally, the classification of power system faults is formulated as solving an adaptive L1-norm and L2-norm combined objective function using an alternating direction optimization method. Simulation results and analyses show that the proposed framework has better performance than the existing compressive sensing based power system faulty classification methods in both accuracy and speed.

Depression has been a major concern for the employers due to its high prevalence and cost. In United States, 6.8% of the work population had depression and it is accountable for over $100 billion workplace loss every year. Early detection and timely intervention is the key to address this challenge. In this paper, we proposed to leverage a large claim dataset to build depression risk prediction model. This approach is easy to implement and scale because the claim data is accessible for employers at no additional cost. We derived a subset of over 7.2 million patients in a 3-year period, with 2,099 variables including the diagnosis, procedure, medication, and health service provider information. We trained two models (Least Absolute Shrinkage and Selection Operator (LASSO) and random forest *RF)) with first two years data to predict the onset of depression in the third year. Both models achieved the best comparable performance so far according to our knowledge. The important variables identified by our algorithms are mental health related medications and diagnoses (other than depression), gender (female), and other chronic conditions such as back pain. We further validated these variables by confirming in clinical literature. We concluded that the claim data based depression prediction is a viable and low cost approach, especially for self-insured employers.

A key aspect of the mimetic defense system is the scheduling of heterogeneous executions to ensure that the system is dynamic. The existing scheduling strategy only performs random scheduling to satisfy heterogeneity and service quality. Combined with the historical information of the arbiter in the mimicry defense system, this paper proposed a random seed scheduling algorithm based on historical confidence for maximum heterogeneity and quality of service. First, the seed performer is randomly selected in the historical confidence executor that satisfies the threshold, and then the scheduling scheme is determined according to the maximum heterogeneity type and the quality of service indicator. The simulation results showed that the scheduling period of the algorithm is lower than the random scheduling algorithm. However, the scheduling effect is higher than the random scheduling algorithm, which achieved a good dynamic balance between dynamics, security, and quality of service.

We investigate facial expression recognition using state-of-the-art classification models. Recently, CNNs have been extensively used for face recognition. However, CNNs have not been thoroughly evaluated for facial expression recognition. In this paper, we train and test a CNN model for facial expression recognition. The performance of the CNN model is used as benchmark for evaluating other pre-trained deep CNN models. We evaluate the performance of Inception and VGG which are pre-trained for object recognition, and compare these with VGG-Face which is pre-trained for face recognition. All experiments are performed on publicly available face databases, namely, CK+, JAFFE and FACES.

We investigate facial expression recognition based on geometric features, and image appearance, using a range of classifier models. First, we evaluate expression recognition of face images based on geometric features, namely, facial landmarks based on the Active Appearance Model, and Action Units based on the Facial Action Coding System. A generalized linear model and a neural network are used to classify face images based on these geometric features. Second, the classification achieved by facial landmarks and Action Units is compared with the classification achieved by convolutional neural network (CNN) which extracts image features from raw pixels. Finally, we use transfer learning technique to evaluate classification using a pre-trained model. All experiments are performed on the state-of-the-art CK+ face database.