Patent Application
20250096713
The invention is based on the object of providing a method for operating an electrical drive system that permits the state of the electrical drive system to be monitored as simply and reliably as possible.
The method according to the invention is used to operate an electrical drive system and comprises the steps of: acquiring a number of input variables pertaining to the electrical drive system, for example between 2 and 10 input variables, determining at least one state variable from the acquired input variables by means of an observer, determining at least one disturbance variable from the acquired input variables by means of the observer, controlling the electrical drive system on the basis of the at least one determined state variable, inter alia, and monitoring the state of the electrical drive system by means of machine learning, the at least one disturbance variable forming input data for a machine learning model. Input variables pertaining to the observer may be, for example: an actual position of a load as measured by means of a motor sensor, a torque, or a motor current, a position of the load as measured by means of a load sensor, a target position of the load, etc.
In one embodiment, the electrical drive system comprises: an electric motor and a phase-rotation indicator or encoder mechanically coupled to the electric motor, and a mechanical load moved, or driven, by means of the electric motor, and a load sensor mechanically coupled to the mechanical load. Measured variables ascertained for the observer are selected from: a phase-rotation indicator position generated by means of the phase-rotation indicator, a load sensor position generated by means of the load sensor, and a torque generated by means of the electric motor.
In one embodiment, the state variable is a position and/or a speed of the mechanical load.
In one embodiment, the at least one disturbance variable represents a friction occurring in the electrical drive system.
In one embodiment, the at least one disturbance variable represents a torque generated by means of the electric motor.
In one embodiment, the observer is an extended Kalman filter.
In one embodiment, the state of the electrical drive system determined by means of machine learning is evaluated for the purpose of preventive maintenance of the electrical drive system.
According to the invention, machine learning is combined with physical model knowledge.
According to the invention, for example model-based sensor fusion of the load sensor with the phase-rotation indicator and the torque can be carried out in order to obtain a more accurate position and speed of the mechanical load. This can be accomplished by using an extended Kalman filter as an observer, for example. The extended Kalman filter is an observer that also estimates the disturbance variables, i.e. calculates, or assesses, errors in the measurements. The variances between the measurement and the calculation are plausibility checked.
The invention is based on the insight that the observed values, in particular the disturbance variables, are not only suitable for showing whether the observer is working. They also indicate whether a state of the electrical drive system changes. An increasing friction or other changes lead to variances between the values determined in the model and the values from the real measurements. The estimated disturbance variables in the observer therefore increase significantly.
According to the invention, the observer is used with a disturbance variable estimator to generate additional data that are better suited as input data for machine learning for the purpose of condition monitoring than data such as torque, speed, etc. The disturbance variables determined by means of the observer provide information about changes in reality compared with the model directly and immediately. It is possible to observe variables for which no or only poor measurements exist. The disturbance variables compress information considerably compared with measured values. The use of an observer in order to evaluate data is a way of making better statements with fewer measurements. Additionally, the evaluation of the disturbance variables is well suited to learning condition monitoring by means of machine learning without previously labeled measurements, for example labeled as “good” or “corrupt” measurement.
The use of an observer to estimate friction or other disturbance variables also permits a specific statement about the machine state to be made with only a few values, for example 2 values. This reduces the need to transmit a large amount of data to an evaluation unit.
In addition to reconstructed measurement and state variables, such as position and speed, the observer according to the invention also outputs disturbance variables. This is accomplished by calculating the disturbance variables in the same way as other states. By way of example, the disturbance variable may be a torque. The greater the variance of the response of the drive system compared with the drive system model, the greater the disturbance variable becomes. If the drive system responds in the same way as the associated drive system model, the disturbance variables disappear. By way of example, the torque is then 0 Nm.
As a result of the use of the disturbance variables in machine learning, the AI methods become simpler, for example a simple decision tree suffices, fewer data are needed and the methods are more easily transferrable to different drive systems. It is no longer necessary to train from zero. Only the limits for the decision tree need to be adapted. This can also be done by way of a model, in principle.
If an observer is already used for the drive system, for example, as in accordance with the invention, for sensor fusion, it is also a simple matter to use the observer for preventive maintenance (condition monitoring).
The invention permits optimized use of machine learning, for example for condition monitoring or preventive maintenance, with lower hardware costs. The observer provides a dual benefit in this case, specifically optimizing sensor signals and generating input data for the machine learning model.
The invention is described in detail hereinbelow with reference to the drawings.
The at least one disturbance variable S can represent a friction occurring in the electrical drive system 100 and/or a torque generated by means of the electric motor 1, for example.
In respect of the basic design and basic function of observers, or extended Kalman filters, reference will also be made to the relevant specialist literature.
Measured variables m1, . . . , mm ascertained for the observer 200 are for example selected from a phase-rotation indicator position D1 generated by means of the phase-rotation indicator 2, a load sensor position D2 generated by means of the load sensor 4, and a torque D3 generated by means of the electric motor 1, see also
The state variable(s) Z may be a position and/or a speed of the mechanical load 3.
The electrical drive system 100 is controlled on the basis of the at least one determined state variable Z.
| Number | Date | Country | Kind |
|---|---|---|---|
| 10 2023 125 209.7 | Sep 2023 | DE | national |
