Cairo L. Nascimento

An NN-based approach for tuning servocontrollers

Neural networks (NN) are used in this paper to tune PI controllers for unknown plants, which may be nonlinear or open-loop unstable. A simple algorithm, which requires only knowledge of the plant output response direction, is used for training an NN controller, by employing the error between the reference and the plant output. Once this controller achieves good performance, its input-output behavior is approximated by a controller with PI structure, thereby enabling the computation of proportional and integral gains. These gains are familiar to process engineers and can be directly inserted into most existing softwares for process control in industry. Computer simulations on an unstable nonlinear plant and experimental results on a thermal plant are presented to illustrate the usefulness of the proposed approach.

Comparative case study of life usage and data-driven prognostics techniques using aircraft fault messages

Abstract Prognostics are a key activity in repair and maintenance operations. A recent approach to condition-based maintenance is the data-driven approach. This approach has been mostly based on past failure time measures, and sensed measurements of component degradation to derive estimates of the remaining useful life of equipment. An alternative source of data, rarely used in these models, is the stream of automatic messages derived from diagnostics systems, which consist of fault codes indicating abnormal events or deviations from optimal operation. Despite the richness and concise nature of these messages, their difficult interpretation poses significant challenges to its use in prognostics. This paper aims to show that data-driven prognostics based on this type of messages can be better suited to maintenance than time-based approaches. We illustrate this comparison with an industrial case study involving the removal times of a bleed valve from the aircraft air management system. Our experimental results reveal a significant accuracy improvement over the contrasting time-based models. We also establish the contribution to this improvement of the data-driven methods and message-related predictors.

Forecasting fault events for predictive maintenance using data-driven techniques and ARMA modeling

Abstract Presently, time-based airline maintenance scheduling does not take fault predictions into account, but happens at fixed time-intervals. This may result in unnecessary maintenance interventions and also in situations where components are not taken out of service despite exceeding their designed risk of failure. To address this issue we propose a framework that can predict when a component/system will be at risk of failure in the future, and therefore, advise when maintenance actions should be taken. In order to facilitate such prediction, we employ an auto-regressive moving average (ARMA) model along with data-driven techniques, and compare the performance of multiple data-driven techniques. The ARMA model adds a new feature that is used within the data-driven model to give the final prediction. The novelty of our work is the integration of the ARMA methodology with data-driven techniques to predict fault events. This study reports on a real industrial case of unscheduled removals of a critical valve of the aircraft engine. Our results suggest that the support vector regression model can outperform the life usage model on the evaluation measures of sample standard deviation, median error, median absolute error, and percentage error. The generalized linear model provides an effective approach for predictive maintenance with comparable results to the baseline. The remaining data-driven models have a lower overall performance.

FastSLAM filter implementation for indoor autonomous robot

In this paper, we present a FastSLAM particle filter algorithm used to efficiently map large indoor environments features. The proposed filter uses an unknown data association to match the extracted environment characteristics, such as walls and doors. Data association (DA) is chosen due to two reasons: 1) permit to rearrange the filter particles in the prediction phase of the filter, and 2) enable to incorporate the extracted features in the map of each particle. Indoor SLAM experiments were conducted in a long corridor composed by several wooden walls. These provisional walls were used to create a more challenging environment. From the map obtained by the mapping process, the robot is capable of navigating through the environment using the set of 22 predefined poses. The SLAM filter measurements are compared with their actual measured values.

Design of model predictive control via learning automata for a single UAV load transportation

In recent years, autonomous aerial robots have been successfully used to perform the construction of structures composed by parts that have similar dimensions and inertial moments. However, these proposed control systems are not able to accurately control the UAVs during the handling and transporting loads with various weights and balance features. In this paper, we investigate a robust and innovative control strategy for UAV load transportation system that can deal with the load characteristics and disturbances such as ground effect and control noise. Taking into account the nonlinear and under-actuated features of the quadrotor, a Learning Automata (LA) methodology is applied to tune the Nonlinear Model Predictive Controllers (NMPCs) in the various contexts of operation. Specifically, it applies LA to select the weighting parameters of the objective function in order to minimize tracking error provided by the plant. Simulation results demonstrate the learned weighting parameters can be efficiently employed to obtain NMPC controllers for tracking optimized trajectories to deal with different load conditions.

Iterative Decentralized Planning for Collective Construction Tasks with Quadrotors

This paper describes an iterative decentralized planning and learning method, based on stochastic learning automata theory and heuristic search techniques, to generate construction and motion strategies to build different types of three-dimensional structures using multiple quadrotors. This architecture is proposed to simultaneously solve three main problems: 1) the iterative generation of feasible construction and motion plans for each quadrotor; 2) the optimization with constraints on power and assembly while taking into account the dynamic nature of the environment, and 3) the planning of the translational speeds and selection of breakpoints for each vehicle. The quadrotors learn the optimal action policy to construct the structures while avoiding collisions during the loading and unloading procedures. In order to demonstrate the generality of the solution, simulated trials of the proposed autonomous construction system are presented where different three-dimensional structures are built.

A New Particle Weighting Strategy for Robot Mapping FastSLAM.

Nowadays, FastSLAM filters are the most widely used methods to solve the Simultaneous Localization and Mapping (SLAM) problem. In general, these approaches can use complex matrix formulation for computing the particle weighting procedure, during the execution of the SLAM algorithm. In this paper, we describe a new particle weight strategy for the FastSLAM filter, which can maintain the generation of particles in its most simplified form. Thus, this approach tries to estimate the robot poses and build the environment map using a simple geometric formulation for executing the particle weighting procedure. This method is capable of reducing the processing time and keeping the accuracy of the robot pose. Both simulation and experimental results demonstrate the feasibility of the proposed approach at enabling a robotic vehicle to accomplish the mapping of an unknown environment and also navigate through it.

Flexible Link Control with Three Different Structures Using Feedback-Error-Learning

This paper discusses three structures for neural control of a flexible link using the Feedback-ErrorLearning technique. This technique aims to acquire the inverse dynamic model of the plant and uses a neural network acting as an adaptive controller to improve the performance of a conventional non-adaptive feedback controller. The non-collocated control of a flexible link is characterized as a non-minimum phase system, which is difficult to be controlled by most control techniques. Three different neural approaches are used in this paper to overcome this difficulty. The first and second structures use a virtual redefined output as one of the inputs for the neural network and feedback controllers, while the third employs a delayed reference input signal in the feedback path and a tapped-delay line to process the reference input before presenting it to the neural network.