Method and arrangement for the neural modelling of a dynamic system with non-linear stochastic behavior

• US 6,272,480 B1
• Filed: 10/19/1998
• Issued: 08/07/2001
• Est. Priority Date: 10/17/1997
• Status: Expired due to Fees

• Current Assignee: Siemens AG
• Original Assignee: Siemens AG
• Inventors: Briegel, Thomas, Tresp, Volker
• Primary Examiner(s): Powell, Mark R.
• Assistant Examiner(s): STARKS, WILBERT L
• Application Number:
  US09/175,068
• Time in Patent Office: 1,023 Days
• Field of Search: 706/25, 706/44, 700/29
• US Class Current: 706/44
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs

Abstract

In a method and arrangement for the neural modelling of a dynamic system with non-linear stochastic behavior wherein only a few measured values of the influencing variable are available and the remaining values of the time series are modelled, a combination of a non-linear computerized recurrent neural predictive network and a linear error model are employed to produce a prediction with the application of maximum likelihood adaption rules. The computerized recurrent neural network can be trained with the assistance of the real-time recurrent learning rule, and the linear error model is trained with the assistance of the error model adaption rule that is implemented on the basis of forward-backward Kalman equations. This model is utilized in order to predict values of the glucose-insulin metabolism of a diabetes patient.

Claims

1. A method for neural modelling of a dynamic system with non-linear stochastic behavior, the method comprising the steps of:

modelling a system behavior of a dynamic system as a time series of at least one influencing variable for a prediction;

forming the influencing variable as an additive combination of a deterministic output quantity of a computerized recurrent neural network and a linearly modelled system error; and

adapting the computerized recurrent neural network at a first point in time with an error model adaption error formed as a difference between the influencing variable of the system measured at the first point in time and the linearly modelled system error.

2. The method of claim 1, further comprising the step of modelling the system error with a Kalman filter, modelling a missing measured value in the time series with a variance of infinity, and modelling a measured value in the time series with a variance of zero.

3. The method of claim 2, further comprising the step of employing the following system equations:

4. The method of claim 1, further comprising the step of training the computerized recurrent neural network with a real-time learning rule.

5. The method of claim 1, further comprising the step of employing a linear error model adapted with forward/backward Kalman filter equations.

6. The method of claim 1, further comprising the step of modelling the glucose/insulin metabolism of a diabetes patient as said dynamic system, and modelling at least a glucose level of the blood as said influencing variable.

7. The method of claim 6, further comprising the step of supplying respective time series to the computerized recurrent neural network for training each one of insulin dose administered, quantity of food ingested, sport activity endured, current blood glucose level, and preceding estimated blood glucose level.

8. A method for neural modelling of a dynamic system with non-linear stochastic behavior, the method comprising the steps of:

modelling a system behavior as a time series of at least one influencing variable for a prediction;

forming the influencing variable as an additive combination of a deterministic output quantity of a computerized recurrent neural network and a linearly modelled system error; and

adapting a model of the system error with a neural model adaption error formed as a difference between the influencing variable supplied to the network at a point in time as an input quantity and a corresponding output quantity output by the adapted network.

9. The method of claim 8, further comprising the step of modelling the system error with a Kalman filter, modelling a missing measured value in the time series with a variance of infinity, and modelling a measured value in the time series with a variance of zero.

10. A method according to claim 9 employing the following system equations:

11. The method of claim 8, further comprising the step of training the computerized recurrent neural network with a real-time learning rule.

12. The method of claim 8, further comprising the step of employing a linear error model adapted with forward/backward Kalman filter equations.

13. The method of claim 8, further comprising the step of modelling the glucose/insulin metabolism of a diabetes patient as said dynamic system, and modelling at least a glucose level of the blood as said influencing variable.

14. The method of claim 13, further comprising the step of supplying respective time series to the computerized recurrent neural network for training each one of insulin dose administered, quantity of food ingested, sport activity endured, current blood glucose level, and preceding estimated blood glucose level.

15. A system for neural modelling of a dynamic environment with non-linear stochastic behavior, comprising:

a computerized recurrent neural network for predicting at least one influencing variable of the dynamic environment;

a first subsystem for error modelling of a system error of the dynamic environment;

a second subsystem for forming an adaption error for a first point in time as a difference from the system error and the influencing variable at the first point in time; and

a third subsystem for adapting the computerized recurrent neural network with the adaption error, whereby the computerized neural network is trained with a system behavior of said dynamic system.

16. The system of claim 15, wherein the computerized recurrent neural network comprises a multi-layer perceptron having a hidden layer with three neurons.

17. The system of claim 15, wherein the computerized neural network further comprises a subsystem for modeling a blood glucose level as said influencing variable.