DE19505506C2 - Method for using an observer model for torque estimation and prediction for an asynchronous machine - Google Patents

Description

The invention relates to a method according to the Oberbe handle of claim 1. Such a method is by the case by J. Böcker "Discrete-Time Model of an Induction Motor" ETEP Vol. 1, No. 2, March / April 1991, pp. 65-71 known.

The problem is that of an asynchronous machine inside torque applied to a drive (on the rotor shaft) can not be measured and therefore no feedback for the Torque control loop is available, for example the tension control of reels on; however, it is by no means limited to such an application.

Reel is the winding of tape-like dimensions terials into a fixed bundle for further processing or the transportation as well as the unwinding around the material a process feed. This technology is particularly used in paper, Plastic and metal industries used. Here are special demands placed on the belt tension to a firm Reel to guarantee a safe process flow or the product properties (surface quality, Ma material thickness). That is the task of the band pull regulation.

The belt tension required is provided by an electric drive upset. They are almost exclusively alien excitations Electricity machines in use because of the large collar diameter can also be operated in the field weakening range.

The technological conditions allow it in particular for Hot rolling does not measure the strip tension and for a direct one Use belt tension control. The present process can  are only partially regulated indirectly (which is possible) The most accurate model of all components involved available have to be. So far, the use of three-phase current has failed motors because the torque given to the shaft is not exactly correct can determine enough from the measurable sizes. The mechanical The moment of the reel drive must in the case considered here namely at least with an accuracy of 1% overall Speed range, including river weakening, can be specified.

For asynchronous machines in high-quality industrial applications the field-oriented regulation is the standard. Here is the Control of field and torque, similar to the DC machine with external excitation, through two separate, linear control loops set (cf. Blaschke, F: The method of field theories for controlling the induction machine, "Siemens Research and development reports ", 1/1972, pages 184 to 193).

For example, an asynchronous machine for a reel two problems arise. It is immediately apparent that because of the (almost always) missing Measuring device (torque measuring shaft) no control variables feedback can take place. So this is about a quasi-control, in which the initially mentioned observer term model forms an estimated torque value and this on the Torque setpoint is regulated. The more accurate this observer model, the more precise the control. Unfortunately, that is Relationship between current and torque in three-phase machines not as easy (proportional) as with direct current machine. There the armature current can be used as a very accurate measure of serve the generated moment. For the nonlinear system of Asynchronous machine should actually be a non-linear one Observers are designed. However, such approaches lead to complex non-linear control structures.

The second problem is that information about the state sizes of the rotor (field or current) are not measurable Rotor flow angle can only be obtained from these. Of the Rotor flux angle is used to transform the mathematical  Model in the coordinate system rotating with the rotor flux required for field-oriented control and back.

Observer models of the asynchronous machine, which make an estimate of the required states, are used to solve both problems. The following problems arise:

• - discretization in digital systems,
• - parameter deviations,
• - Effects not included in the mathematical model
• - (thermal changes, saturation and mechanical
• - friction).

These problems have hitherto been the use of asynchronous machines in torque controls with accuracy requirements of <1% control error prevented.

Particular attention is paid to the method according to the invention on the quality of modeling of the mechanical moment to a better control quality in the asynchronous machine (and thus in the spe zial application) to achieve the tension control.

The requirements for high-precision torque control, such as Example of the tension control on reels, however, require that be as much information as possible about the asynchronous machine is known and used. It is a matter of regulation technical view of a torque control. Controls are always only as precise as the process used for them models, as no information about the actual value of the size to be regulated is available. This must be as precise as possible to be modeled. In the case of indirect moments control of an asynchronous machine can do that Moment (as presupposed) cannot be measured, it will stand however, measurands such as rotor speed, stator current and Stator voltage of the asynchronous machine available, so that a Torque calculator, the torque to be applied by the drive tracks the process conditions, fed with this data can be.  

To verify and evaluate observer models, Output quantity, ie the estimated moment with that of the "real" Systems, i.e. the actual moment on the machine shaft be compared. You can do this with a torque measurement wave, if - as provided here - before commissioning such a torque measurement on the Ro Tor shaft of the asynchronous machine is (still) possible.

Are knowledge of the physical mode of action of a Process, you can do this through a physical motivated modeling in a mathematical model transfor lubricate. But in general it can be said that through a physika Modeling almost never provides a complete description of a process can be achieved. As simple examples the very often neglected phenomena like friction, Manipulated variable restrictions, distributed or time-variant para be called meters. The reason for this is in the good ver consistency of the resulting, mostly linear, model le to see.

In an asynchronous machine, effects such as satti are very often magnetic components, current displacement, asymmetry drive of the coils, thermal influences and not ideal positions links not taken into account. In many cases the math is matical description but also very complex or not at all only known. Modeling such effects is difficult because the time constants of their dynamics are sometimes around sizes orders up and down of those in the observer model reproduced system description differs. Control algorithms for systems with such properties are not easy to develop.

If you have the available information in a ma thematic structure, so it usually has to the necessary parameters for what is available in the specific case System can be adjusted. You get an image of the real one Systems in the form of a structurally incomplete system with  estimated parameters. In the procedure to be considered here However, knowledge of the real machine parameters is required set.

Both the neglect of certain dynami mentioned above effects as well as the imprecise knowledge of the selected Model parameters lead to highly precise dynamic requirements Compliance with the control error, such as in the indirect belt tension control, for an incorrect calculation the reference variable in the torque calculator.

The estimated parameters of the model are different from the real ones Process parameters, so causes the moment equation in loading observer model in the implemented equation system an error because of the wrong proportionality factor consisting of the Model parameters. Another source of error arises from the imprecise rotor flux position calculation.

By the article by Hong-Te Su, N. Bhat, PA Minderman and TJ McAvoy: Integrating Neural Networks with First Principles Models for Dynamic Modeling "IFAC Dynamics and Control of Chemical Reactors (DYCORD-92), Maryland, USA, 1992, pages 327 to 332, in particular FIG. 1, it is known to improve the estimated value of an observer model in dynamic chemical processes by additionally using a neural network.

Neural networks are based on learning and generalization of effects encoded in input and output data pairs are.

Contrary to the procedure for mathematical-physical Motivated modeling is used when teaching neural Less prior knowledge about the process to be modeled provided. The multilayer perceptron (MLP) and the radial Basic Function Networks (RBF), two very frequently used networks  types of work, can approximate the effects contained in the training data as precisely as desired.

With the aforementioned use of a neural network in a chemical process, the neural network as well as the observer model the input variable of the current chemical pro zeses entered. In addition, the observer error resulting from the difference between the output of the real process and the estimate of the observer model, which neural network fed. The correction quantity supplied by the neural network is then the Added estimate of the observer model.

Such a method may be suitable for a chemical process; however, it doesn't make sense for Indirect torque control can be used on asynchronous machines, as the latter are dynamic System with a very large entrance space, which makes it very expensive to create a neural network work out. In addition, the acquisition of the output of the real system (namely the Torque measurement on the rotor shaft) not possible with the asynchronous machine considered here, so that a simple transfer of the known method to the use of the input is given observer model by itself.

In general, the use of neural networks in field-oriented controlled converters is also necessary fed three-phase drives common (see: Godoy-Simoes, M; Bose-B-K .: Neural network based estimation of feedback signals for a vector controlled induction motor drive. Proceedings of 1994 IEEE Industry Applications Society Annual Meeting, vol. 1, Denver, CO, USA, 2-6 Oct. 1994, vol. 1 1994, pp. 471-479; R. Buhl and R. D. Lorenz, "Design and implementation of neural networks for digital regulation of inverter drives, "in Conf. Rec. IEEE Ind. Applicat. Soc. Annu. Meeting, Oct. 1991, pp. 415-421; K.P. Toh, E.P. Nowicki, and F. Ashrafzadeh, "a flux estimator for field oriented control of an induction motor using an artificial neural network, "in Conf. Rec. IEEE Ind. Applicat. Soc. Annu. Meeting, October 1994, pp. 585-592).

The neural networks are trained with measured values, or with values from parallel running models can be obtained. With the publication of Godoy-Simoes, M; Bose-B-K, a. a. It is known that parallel to the neural network there is a torque and flux  estimator (DSP estimator) is running. The neural network is u. a. due to torque deviation trained.

The use of a neural network instead of the observer model mentioned at the beginning also leads not because of the previously mentioned large entrance room for training the neural network to a sufficient limitation of the estimation error.

The invention has for its object the Qua in a method of the type mentioned to improve the quality of the torque estimate provided by the observer model.

This object is ge according to the invention by the features characterized in claim 1 solves.  

Advantageously, the entire system is therefore not affected by the replicated neural network, but only that not through that mathematical model captured effects. The a priori Structural knowledge is used to win the observer model and then the unstructured parts are replaced by the Mapped neural model. The neural model approximates therefore the torque error of the mathematical model based on of the observer error and those detectable during operation Measured variables of the system.

Because, by definition, the shaft torque of the Asynchronous machine can (still) be measured at this time the neural network based on the error deviation of the observer Estimated torque delivered from the actual model entering moment on simple (freely entered from the entrance room restricted) way.

Advantageous embodiments of the method according to the invention are marked in the subclaims.

The method according to the invention will now be described with reference to Exemplary embodiments illustrated in the drawing will. Show it

Fig. 1 is a schematic diagram of an arrangement for the implementing of the method according to the invention,

Fig. 2 is a schematic diagram for training the neural network before start of operation and

Fig. 3 is a block diagram for the application of the method according to the invention in an indirect manual control.

Referring to FIG. 1, an input (manipulated variable for the Asyn chronmaschine) and both an asynchronous machine (the nonlinear herein by their voltage System M _{P_x} and by their () mechanical system of equations (torque equation) M _{P_Y} sented repre is) and fed to an observer model ( which here mathematically simulates the voltage equation system _{P_math_beob.} and the mechanical equation system _{P_math_y} ).

A real measurable stator size x the asynchronous machine, e.g. B. the stator currents and the corresponding stator size calculated by the observer model (ie z. B. the estimated stator currents of the asynchronous machine are compared, and the resulting observer error ex is _input to a trained on this observer neural network _{P_neuronal} (here, for example, an RBF network).

With the rotational frequency ω (t) of the rotor of the asynchronous machine both the real machine and the observer model as finally also affects the neural network.

The torque error m _el learned from the neural network and formed from its input variable, the observer error ex, is added as a correction value to the part _{P_math_y of} the observer model emulating the mechanical equation system of the asynchronous _machine , and thus forms a highly precise estimated value _{el_beob_RBF} , on the while operation, the immeasurable shaft torque is regulated.

Fig. 2 shows in principle how the neural network described in its arrangement to Fig. 1, which, as already mentioned, should preferably be an RBF network (but it can also be a multilayer perceptron), is trained before the start of operation. Then the actual shaft torque y can (still) be detected (for example by temporarily providing the drive with a torque measuring shaft). It is compared with the mathematical simulation _{P_math_y of} the (non-linear) mechanical system (corresponding to the torque equation) based on the estimated value for the air gap torque for the shaft torque. The deviation em _el representing the observer error as a whole is in turn compared with the correction value m _el output by the neural network due to the observer error ex. The deviation e _modeling determined in each case serves the learning process of the neural network.

In Fig. 3 the integration of the method explained with reference to FIGS . 1 and 2 according to the invention in the process of an indirect strip tension control is shown in principle. A command variable, namely the required strip tension from a system superior to the reel and measured variables from the higher system (for example speed v of the strip to be reeled and radius r of the coil on the reel) are fed to a torque computer as well as from a process model _P one corresponding to that Fig. 1 explained method (ie including the neural network) determined control variable estimate (t). The torque calculator forms a reference variable r (t) for the control of an actuator (for example, a converter connected upstream of the asynchronous machine), which acts on the process (ie here the asynchronous machine) with the manipulated variable u (t). In real operation, faults not modeled in the observer (process) model also influence the process (asynchronous machine). Overall, a control variable y (t) results, which is adjusted to the reference variable r (t). To estimate the control variable estimate size (t), the process (asynchronous machine) traces the measured variables, which can also be detected during operation, by measurement, as well as the manipulated variable u (t), to the process model _P.

In summary, it can be said that the modeling of the torque of the asynchronous machine for the indirect belt tension control can be improved by the hybrid observer RBF network structure _{P_Beobachter_RBF} shown in FIG. 1. This also applies in particular to parameter changes. The advantage over the purely classical (physical-mathematical) modeling is that effects with unknown dynamics, which are represented in an input-output data set, can be recorded with the help of the neural RBF network. This can make a significant contribution to improving the control quality for indirect strip tension control.

Another advantage of the structure according to the invention is that them from a classic structure by means of an additional one neural network can be obtained. It is an addition previously successfully used methods possible without to have to forego the advantages of the previous solutions. For the control of an asynchronous machine, this means that the tried and tested for power electronic control Principle of field orientation can be used.

Claims (3)

1.Method for operating an asynchronous machine in which an observer model which thematically simulates the asynchronous machine ma and calculates the mechanical moment on the shaft of the asynchronous machine forms an estimated torque value from the electrical input variables supplied to the asynchronous machine, the deviation being at least one stator size calculated by the observer model Asynchronous machine is traced back to the viewer from the corresponding measured stand size actually occurring as an observer error, characterized

• 1. - that during operation, the observer error is entered into a neural network and the neural network then provides a correction value for the accompanying (shaft) torque estimate, which is calculated overall by the observer model,
• 2. - that the torque is measured before the start of operation and compared with the torque estimate by the observer, and
• 3. - that the neural network is taught in with the resulting error in the (shaft) torque estimate.

2. The method according to claim 1 marked by use with reel drives. 3. The method according to claim 1 characterized, that as a neural network a Radial Basis Function (RBF) network or a multilayer Perceptron is used.