IMPROVE THE ACCURACY OF NANO-SATELLITE'S CONTROL SYSTEMS Control systems are one of the main and complex systems of on- board microsatellite equipment. For control it is necessary to know first of all the parameters of the orientation of the object relative to the given coordinate system. Inaccuracy in determining the orientation parameters, in turn, leads to an incorrect execution of the control task. In the context of the wide application of control systems in spacecraft, the problem of increasing the accuracy for controlling the movement of nano-satellite’s is becoming increasingly important. To solve such problems, various methods and means are used, the traditional one of which is the Kalman filter. The Kalman filter is a recursive method for estimating parameters, proposed in the 60's. The last century, using for work mathematical model of the movement of the object and the current vector of measurements. Despite the great popularity of the filter, a number of serious problems arise in its application: • the problem of specifying the initial approximations of the orientation parameters, since for certain initial conditions the filter can not converge; • Linearization problem: a filter for its operation requires a linearized motion model; Therefore, only with slow movements (with frequent measurements), the filter can give a satisfactory estimate of the state vector; • tuning problem: covariance error matrices are used in the filter, the tuning of which strongly affects the convergence rate and the error of the estimated state vector. Each of these problems for each spacecraft is solved in its own way, depending on the tasks assigned, so there is no generally accepted methodology for working with the filter. The complexity of the method is that, for example, for a six-dimensional state vector, the number of parameters to be selected is 155, and manual selection takes a very long time. Structural diagram of the Kalman filter
Gk + 1, k is the perturbation matrix;
Fk + 1, k is the transition matrix of the state of the system; wk is the perturbation vector; Xk is the state vector of the dynamical system; Zk is a vector of measurement; Hk is the measurement matrix; νk is the measurement error vector. Other problems associated with the problems of the classical Kalman filter are associated with the use of a more accurate non- linearized model of errors, which takes into account: • errors in determining scale factors; • errors in setting the sensor sensitivity axes relative to the basic axes of the spacecraft; • Constant components of sensor drifts; • errors from unbalance of sensors; • errors in setting the sensitivity sensitivity axes relative to the axes of the instrument base; • errors in the displacement of the zeros of the velocity meters. For this model of errors, the Kalman filter's dimension greatly increases, which requires additional on-board computer resources. Another problem is that individual error model coefficients can be determined if the system is in steady state. When the system is operating in a dynamic mode, these coefficients depend on the flight parameters, and they are not known. As one of the best means of improving control accuracy, which does not require specifying a system model and significant computing power, was developed in 2011 by Sebastian Madgwick. The Madgwick filter uses quaternions, which allows you to use data from the accelerometer and magnetometer for analytical calculations, as well as to obtain an error in the direction of the gyroscope in the form of a quaternion derivative by optimization using the gradient descent method. The task of the Madgwick filter is to calculate a single orientation estimate by measuring the accelerometer, gyroscope and magnetometer. The structure diagram of the Madgwick filter
- rate of change in orientation;
- calculated based on accelerometer measurements
and magnetometer; β - error compensation for measuring the rate of change in orientation - orientation quaternion. Modules consisting of a three-axis gyro, accelerometer and magnetometer. The author of the paper cites the results of experiments showing that the level of accuracy of the developed filter exceeds that of the filter based on the Kalman method: - less than 0.6 rms deviation in the stationary state; - less than 0.8 rms deviation in the mobile state. The advantages of the Madgwick filter also include: - cheapness of computational resources - 277 simple arithmetic operations each filter update; - efficiency at low sampling rates (e.s 10 Hz). Another promising option is the possibility of introducing a fuzzy neural network into the control systems. The concept of fuzzy logic was introduced by the Azerbaijani professor Lutfi-Zade in 1965. In his work, the notion of a set was extended by the assumption that the membership of an element to a set can take any values in the interval [0 ... 1], and not just 0 or 1, such sets were called fuzzy. Due to their inherent properties, there is no need to compile a mathematical model and, consequently, there are no problems associated with increasing the dimensionality of systems and the inability to analytically record the model of errors. The use of fuzzy logic allows to ensure the formalization of a complex object at the level of relatively simple rules, and the use of neural networks ensures the adaptability of control systems. The joint use of the apparatus of fuzzy logic and artificial neural networks allows solving a large range of problems of automatic control: the tasks of synthesis, identification, adaptive and robust control, and others. Let's list the advantages of neural network systems over conventional ones, which are manifested only when solving difficultly-algorithmizable problems: 1) neural networks make decisions based on experience acquired by them independently; "Independently" in this case means that the creator of the system does not need to establish the relationships between the input data and the necessary solution, spending time on a variety of statistical processing, selecting a mathematical apparatus, creating and testing mathematical models; 2) the decision made by the neural network is not categorical; the network issues a solution along with a degree of confidence in it, which leaves the user able to critically evaluate the result; 3) the neural network allows modeling the decision-making situation; 4) neural networks issue a response quickly (fractions of a second), which allows them to be used in various dynamic systems that require immediate decision-making; 5) neural networks in the tasks of correction of the classification model, minimization of training parameters, etc. allow to simplify the process of creating expert systems, to determine the directions of scientific search; 6) neural networks are resistant to destruction and interference; 7) neural networks have a high level of parallelization of algorithms. Thank you for your attention!
(Lecture Notes in Mathematics 2034) Andrea Bonfiglioli, Roberta Fulci (Auth.)-Topics in Noncommutative Algebra_ the Theorem of Campbell, Baker, Hausdorff and Dynkin -Springer-Verlag Berlin Heidelberg