DETERMINING FUTURE SWITCHING BEHAVIOR OF A SYSTEM UNIT

Abstract

A computer-implemented method for configuring a system model and a computer-implemented method for configuring a sensor model. There is also described a computer-implemented method for determining future switching behavior of a system unit, with the following steps: a) receiving the configured system model; b) receiving the configured sensor model, c) the configured sensor model being a probability distribution regarding how the sensor unit will behave in the specific time period; d) establishing at least one random sample of behavior of a sensor unit by sampling from the probability distribution; and e) determining the future switching behavior of the system unit and/or at least one associated statistical value on the basis of the established random sample by means of the trained system model. There is also described a corresponding computer program product.

Claims

1-8. (canceled)

9. A computer-implemented method for configuring a system model, the system model being a machine learning model for determining a switching behavior of a system unit, the method comprising: a. providing at least one training data set with a plurality of known input elements of the system unit, in each case for a specific point in time or time period; wherein b. the plurality of known input elements of the system unit includes at least one sensor data set of a sensor unit; c. configuring the system model by a machine learning method using the at least one training data set; and d. providing the configured system model as output.

10. The computer-implemented method according to claim 9, wherein the machine learning method is a rule-based approach, selected from the group consisting of: neural network and decision tree.

11. A computer-implemented method for configuring a sensor model, the sensor model being a machine learning model for determining a behavior of a sensor unit, the method comprising: a. providing at least one training data set with a plurality of known input elements of a sensor unit, in each case for a specific point in time or time period; wherein b. the plurality of known input elements of the sensor unit includes at least one sensor data set for a specific point in time; c. configuring the sensor model by a machine learning method using the at least one training data set; wherein d. the configured sensor model is a probability distribution regarding how the sensor unit will behave in the specific time period; and e. providing the configured sensor model as output.

12. The computer-implemented method according to claim 11, wherein the machine learning method is a stochastic approach.

13. The computer-implemented method according to claim 12, wherein the machine learning method is a Poisson process.

14. A computer-implemented method for determining future switching behavior of a system unit, the method comprising: a. receiving a configured system model, the system model being a machine learning model for determining a switching behavior of a system unit and the system model being configured by a computer-implemented method which includes the following: a1. providing at least one training data set with a plurality of known input elements of the system unit, in each case for a specific point in time or time period; wherein a2. the plurality of known input elements of the system unit includes at least one sensor data set of a sensor unit; a3. configuring the system model by a machine learning method using the at least one training data set; b. receiving a configured sensor model, the sensor model being a machine learning model for determining a behavior of a sensor unit and the sensor model being configured by a computer-implemented method which includes the following: b1. providing at least one training data set with a plurality of known input elements of a sensor unit, in each case for a specific point in time or time period; wherein b2. the plurality of known input elements of the sensor unit includes at least one sensor data set for a specific point in time; b3. configuring the sensor model by a machine learning method using the at least one training data set; wherein b4. the configured sensor model is a probability distribution regarding how the sensor unit will behave in the specific time period; c. ascertaining at least one random sample of a behavior of a sensor unit by sampling from the probability distribution; and d. determining the future switching behavior of the system unit and/or at least one associated statistical value with the aid of the trained system model based on the ascertained random sample.

15. The computer-implemented method according to claim 14, wherein the statistical value is selected from the group consisting of median, mean, and variance.

16. The computer-implemented method according to claim 14, further comprising: carrying out a step selected from the group consisting of: outputting a future switching behavior of the system unit and/or at least one associated statistical value on a display unit; storing the future switching behavior of the system unit and/or at least one associated statistical value in a storage unit; and communicating the future switching behavior of the system unit and/or at least one associated statistical value to a computing unit.

17. A computer program product comprising a computer program with program code for carrying out the method according to claim 9 when the computer program is executed on a program-controlled device.

Description

4. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Presently preferred embodiments of the invention are described further in the following detailed description with reference to the following FIGURES.

[0057] FIG. 1 shows an exemplary detector trajectory.

5. DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0058] Preferred embodiments of the present invention are described below with reference to the FIGURE.

Method for Configuring a System Model (Deterministic Switching Behavior)

[0059] In accordance with one embodiment of the invention, the internal deterministic switching behavior of a light signaling system as system unit is approximated with the aid of a mathematical or logical function as decision tree. In the approximation it is assumed that the true or actual switching behavior of the light signaling system is unknown. However, the switching behavior follows fixed rules, for example as a schematic circuit diagram for the intelligent logic module from Siemens “LOGO”. What rules are possible is furthermore known. These permissible rules of control of the light signaling system are represented as a decision tree. The search for the internal deterministic switching behavior of the light signaling system is realized by generating decision trees that may explain the historical data as accurately as possible.

[0060] In this case, it is possible to use evolutionary or particle-based optimization methods for searching in these symbolic search spaces. The permissible rules are modeled as function blocks and are interconnected, tested and improved by the optimization methods. Both evolutionary and particle-based optimization methods consider generations of possible solutions. In this case, the best solutions of a parent generation preferably pass on information to the child generation thereof, such that the entire population becomes successful over time.

[0061] Given the training data set comprising input elements such as e.g. the sensor signals of the last t-T seconds, the instantaneous signal plan (as index) at the point in time t, the state of the SGR (“signal group”) of the last t-T seconds, the current state of all SGRs (“signal groups”) is calculated. In other words, the deterministic switching behavior of the light signaling system is represented given the context. Consequently, the input elements may also be referred to as context features.

Method for Configuring a Sensor Model (Stochastic Detector Behavior)

[0062] In contrast to the deterministic switching behavior, the behavior of the sensor units of a light signaling system is not predetermined by deterministic rules, but rather is dependent on environmental influences.

[0063] By way of example, it is not possible to predict when the pushbutton of a pedestrian crossing light will be actuated.

[0064] This is owing to the fact that it is not known when a pedestrian will arrive at the crossing light. However, historical data can be used to learn how often a pedestrian should be expected. By way of example, in the evening approximately 2 pedestrians per 5 minutes may be expected at a light signaling system.

[0065] Traffic-situational signals such as detectors and local public transport (public transport messages) are therefore modeled as a stochastic process, e.g. as a Poisson process. Given the training data set comprising input elements such as e.g. calendar data (time of day, weekday, holiday), detector states, public transport of the last T seconds, this process forecasts a probability distribution regarding detectors and public transport for a horizon of the coming H seconds (e.g. 5 minutes). In other words, the stochastic switching behavior of the detectors is determined given the context. Consequently, the input elements may also be referred to as context features.

[0066] In accordance with one embodiment of the invention, the probability distribution is modeled in the form of the parameters of a Poisson distribution. A Bayesian method, in particular a Gaussian process or a Bayesian neural network, is used as a regressor that forecasts the parameters of the Poisson distribution from the context. A stochastic process is modeled by way of the forecast of the parameters of a Poisson distribution with the aid of a Bayesian method. This stochastic process cannot explicitly predict the point in time of the next detector actuation in reality. Instead, it is possible to calculate statistics regarding the actuation (e.g. mean waiting time for the next actuation, variance or quantiles of this waiting time) or to take random samples. Such random samples are used for forecasting.

Method for Determining Future Switching Behavior of a System Unit (Forecast by Means of Rollouts)

[0067] Since the behavior of the detectors cannot be predicted exactly, an exact prediction of the switching behavior of the system unit in the future is not possible either. Instead, the distribution of possible switching behaviors of the light signaling system in the future is predicted. Statistics may be calculated from this distribution. The distribution of possible switching behaviors is generated by means of probabilistic rollouts that take account of the multiplicity of possible developments in the future. The rollouts are generated as follows:

[0068] 1. Sampling of detector trajectories from the stochastic detector model. An exemplary detector trajectory is shown in FIG. 1.

[0069] 2. Sampling of signal group trajectories/development of the crossing light model with detector trajectories from step (1).

[0070] 3. Calculation of relevant statistics from the signal group trajectories, e.g. median (forecast), variance (uncertainty). Optionally, the entire architecture, in particular the stochastic model, may additionally be trained directly with regard to the forecast quality. For this purpose, the error signal of the forecast (e.g. of the remaining time) is backpropagated through the symbolic model (step 1) and then further to the parameters of the stochastic model (step 2). Focusing of the learning may be achieved as a result: importance is attached particularly to the stochastic patterns of the DET/OEV signals during training, which are also important for the forecast of the remaining time.