Search

Structure Learning for Bayesian network (BN) is an important problem with extensive research. It plays central roles in a wide variety of applications in Alibaba Group. However, existing structure learning algorithms suffer from considerable limitations in real world applications due to their low efficiency and poor scalability. To resolve this, we propose a new structure learning algorithm LEAST, which comprehensively fulfills our business requirements as it attains high accuracy, efficiency and scalability at the same time. The core idea of LEAST is to formulate the structure learning into a continuous constrained optimization problem, with a novel differentiable constraint function measuring the acyclicity of the resulting graph. Unlike with existing work, our constraint function is built on the spectral radius of the graph and could be evaluated in near linear time w.r.t. the graph node size. Based on it, LEAST can be efficiently implemented with low storage overhead. According to our benchmark evaluation, LEAST runs 1 to 2 orders of magnitude faster than state of the art method with comparable accuracy, and it is able to scale on BNs with up to hundreds of thousands of variables. In our production environment, LEAST is deployed and serves for more than 20 applications with thousands of executions per day. We describe a concrete scenario in a ticket booking service in Alibaba, where LEAST is applied to build a near real-time automatic anomaly detection and root error cause analysis system. We also show that LEAST unlocks the possibility of applying BN structure learning in new areas, such as large-scale gene expression data analysis and explainable recommendation system.

The new distributed propulsion configuration (DPC) can effectively improve aircraft fuel efficiency and reduce pollution, but also brings in cross-coupling effects between the aircraft aerodynamics and propulsion system. To tackle this problem, an integrated flight/propulsion optimal control is required, which is a constrained optimization problem without an explicit objective function, and most traditional gradient-based optimization algorithms cannot be applied to this problem. In this paper, a novel Bayesian optimization algorithm based on particle swarm optimization (BO-PSO) is proposed to solve this complex optimization problem. Bayesian optimization samples the objective function and fits a Gaussian process model to the sample set, then optimizes objective function via a proxy optimization of an acquisition function which is closely related to the Gaussian process model fitted. The parallel computing ability of particle swarm optimization is further explored to accelerate the convergence speed of the algorithm. Simulation results on the SAX-40 aircraft demonstrate that the proposed BO-PSO algorithm is effective for the integrated flight/propulsion optimal control and the accuracy and convergence speed of the BO-PSO algorithm are effectively improved compared to the pure Bayesian Optimsiation algorithm.

Micro Aerial Vehicles(MAVs) have been seen rapid progress in the indoor entertaining, security monitoring, as well as search and rescue activities. Localization capability is one of the basis of MAVs autonomous movement and path planning. This paper proposes a novel indoor localization algorithm using a prior map for the mini MAVs with a sparse sonar array. Four mutually perpendicular installed ultrasonic sensors provide distances of each directions. An improved multiple rays model is constructed to approximate the detection range of the ultrasonic sensor. Then a maximum a posteriori (MAP) estimation of the MAV's location is accomplished using a Gaussian distribution prior and a Gaussian likelihood function. The proposed algorithm is valid on a test platform and the results indicate good localization performance and robustness against compass measurement noise.

Aircraft loss-of-control (LOC) is the major contributing factor to fatal accidents and is characterised by the manoeuvring of aircraft beyond the allowable flight envelopes. This paper proposes an online learning and inference based method for aircraft flight envelope estimation in order to prevent aircraft LOC. The lift and drag coefficients of the aircraft are identified online using an extended Kalman filter and the aircraft flight dynamics. A Gaussian process regression model then learns and infers the up-to-date form of the lift curve from both prior knowledge and the identification data. The extremum of the inferred lift curve, including the maximum lift coefficient and the critical angle of attack are used to compute the flight envelope estimate of the aircraft. Numerical simulation on the NASA generic transportation model (GTM) shows that the proposed method can effectively estimate the aircraft lift and drag coefficients, and by using the extremum on the up-to-date lift curve inferred, return the flight envelope under a wingtip impairment condition.

Particulate matter pollution becomes an increasingly important problem in developing countries and poses a hazard to human health especially in indoor environment. At the same time, modern buildings are equipped with advanced sensing and control technologies, offering control and optimization capabilities over a large range of parameters. This paper proposes an open platform of a WiFi-enabled indoor air quality monitoring and control system, which could be incorporated into such a ‘smart building’ structure. The complete software and hardware design of this system is presented, along with a series of control experiments. The proposed system operates over an existing WiFi wireless network utilizing the MQTT protocol. It is capable of monitoring the indoor air quality as well as controlling an air purifier to regulate the particulate matters concentration. Experiment results under a real world office environment demonstrate the effectiveness of the proposed design.

The objective of this research was to use the image data captured by the Kinect to control a robotic vehicle. The strategy is to collect a series of depth map in real-time of the robots environment, which contains range information of obstacles. The depth sensor in the Kinect can been used to produce 3D image and the movement decision is made by processing the data from sensor Kinect to identify regions which may contain nearly obstacles which should not be hit. Using the depth image data and 3D data (XYZ) obtained by the Kinect, the laptop can navigate the vehicle SRV-1 to avoid obstacles and build a 3D global map in the indoor environment. As the result of this research, the mapping for 2D and 3D data is combined with robot movement and can be operate in real-time effectively. The collision avoidance system is able to detect obstacles from depth information and navigate SRV-1 robot to choose a collision free path. The robot platform with sensor Kinect is powerful to acquired range data used in terms of navigation and mapping technique.

Aircraft centre of gravity (C.G.) is important for aircraft safety and performance. This paper proposes the use of Gaussian process regression (GPR) models for the estimation of the C.G. location of fixed-wing aircraft. The major benefit of using a GPR model is that it is a data-based approach explicitly tackling uncertainties caused by the quality and quantity of the data as well as sensor measurement noise. The proposed method consists of two steps: the estimation of the fuel tank's C.G. using the GPR model trained with fuel weight property data, and the computation of aircraft C.G. by the C.G. equation. A numerical case study of a transport aircraft shows that the proposed method achieves small mean squared error and gives good estimate of the aircraft C.G. under simulated flight scenarios.

State estimation of nonlinear dynamic systems is an important problem in practice. This paper proposes a recursive state estimation method for nonlinear dynamic systems using Gaussian processes (GP) and pre-computed local linear models. Gaussian processes exhibit remarkable learning, or nonlinear regression capabilities from measurement data. The incorporation of pre-computed local linear models reduces the amount of data required and improves the regression performance of the GP. Based on such an improved GP model for nonlinear dynamic systems, a recursive Bayesian filtering method is implemented for the estimation of the unknown states of the system. Simulations on an aircraft benchmark model demonstrate that the proposed method is capable of estimating the unknown angle of attack of the aircraft, and the existence of the local linear models significantly improves the estimation performance of the filter. This method is especially useful in a control systems design context in which local linearisations of nonlinear dynamic systems are usually readily available.

Large scale variations in the mass or center of gravity (C.G.) pose serious problems for the control of aircraft in situations as heavy load airdrop on cargo aircraft, the airliners which employ the active C.G. control technology, aircraft with structural damages during the flight. To solve the problem of attitude control of such aircraft, a general model for the aircraft is established and a disturbance rejection attitude controller is proposed. The proposed model comprises of a 6-DOF dynamics model and the mass and C.G. variation model, which can handle various mass or C.G. variations and can reflect the dynamic couplings introduced by such variations. For the attitude control of such aircraft, based on the framework of observation and compensation, an explicit model following controller with the extension of extended state observer (ESO) is established. Real-time simulations demonstrate that the proposed model can accurately reflect the aircrafts dynamic responses during the variations, and with the compensation from the ESO, the model following controller can improve the performance and meet the requirements of precision attitude control.

We review the full spectrum of solutions to the Mathieu differential equation y'' + [a - 2q cos(2z)]y = 0, and we describe a numerical algorithm which allows a flexible approach to the computation of all the Mathieu functions. We use an elegant and compact matrix notation which can be readily implemented on any computing platform. We give some explicit examples (written in the programming language Scilab) that provide a ready-to-use package for solving the Mathieu differential equation and related applications in several fields.