DE102024002670B3 - Sensor arrangement and method for merging multiple sensor states - Google Patents

Abstract

The invention relates to a sensor arrangement (1) comprising at least:
- a plurality of sensors (S1 to Sm),
- at least one weighting module (G1 to Gm) to which sensor signals (Si(S1) to Si(Sm)) of the sensors (S1 to Sm) can be fed, and
- a sensor fusion module (4) connected downstream of the at least one weighting module (G1 to Gm), to which weighted sensor states (gSZ(S1) to gSZ(Sm)) can be supplied,
wherein each sensor (S1 to Sm) comprises an associated Kalman filter (KF1 to KFm) which is configured to cyclically determine either a sensor-based sensor state (sSZ(S1) to s(SZ(Sm)) or to exclusively determine, in particular to predict, a model-based sensor state (mSZ(S1) to mSZ(Sm)),
wherein the weighting module (G1 to Gm) is configured to weight the determined sensor states (sSZ(S1) to sSZ(Sm), mSZ(S1) to mSZ(Sm)) depending on their relationship to one another, and
wherein the sensor fusion module (4) is configured to fuse the weighted sensor states (gSZ(S1) to gSZ(Sm)) of all sensors (S1 to Sm) and to determine a sensor fusion value (SFW).

Description

The invention relates to a sensor arrangement according to the features of the preamble of claim 1 and a method for fusing several sensors of a sensor arrangement.

Sensor arrangements designed as multisensors, comprising a plurality of sensors for determining a state, for example, a physical quantity, a system, an environment, an object, or the like, are well known. Sensors are often subject to faults. Furthermore, in such a multisensor arrangement, individual sensors can fail or provide erroneous measurement data.

From the DE 10 2020 114 969 A1 A method and a device for determining an estimation model is described, which is designed to provide an estimated value of a wheel radius of the wheel for a value of a wheel speed of a wheel of a vehicle.

The US 2003/0014147 A1 is directed to improving the accuracy with which a stationary array sensor provides transverse measurements by providing offset compensation for the stationary array sensor using the output of a scanning sensor associated with the manufacturing process.

From the DE 196 33 884 A1 is a method for error-tolerant position determination of an object and from the DE 10 2021 121 715 A1 Anomaly detection in multidimensional sensor data is known.

The invention is based on the object of providing a sensor arrangement that determines a sufficiently accurate status even in the event of a fault. Furthermore, a method for operating the sensor arrangement is to be specified.

The first-mentioned object is achieved according to the invention by a sensor arrangement having the features of claim 1. The second-mentioned object is achieved according to the invention by a method having the features of claim 12.

Advantageous embodiments of the invention are the subject of the subclaims.

The sensor arrangement according to the invention comprises at least a plurality of sensors, at least one weighting module to which sensor signals from the sensors can be fed, and a sensor fusion module connected downstream of the at least one weighting module to which weighted sensor states can be fed, wherein each sensor comprises an associated Kalman filter which is configured to cyclically either determine, in particular estimate, a sensor-based sensor state or exclusively determine, in particular predict, a model-based sensor state, wherein the weighting module is configured to weight the determined sensor states depending on their relationship to one another, and wherein the sensor fusion module is configured to fuse the weighted sensor states of all sensors and determine a sensor fusion value.

• - entweder den Sensorzustand mittels eines Modells anhand der aktuellen Sensorsignale und anhand von Modelldaten oder nur anhand der aktuellen Sensorsignale zu schätzen
• - oder, wenn keine aktuellen Sensorsignale verfügbar sind, den Sensorzustand ausschließlich mittels des Modells anhand von Modelldaten vorherzusagen.

Preferably, the respective Kalman filter can be configured, for example, to process current sensor signals of the associated sensor and per cycle

• - either to estimate the sensor state using a model based on the current sensor signals and model data or only based on the current sensor signals
• - or, if no current sensor signals are available, to predict the sensor state using the model based on model data only.

Furthermore, the Kalman filter can be configured to determine the sensor states of the respective sensor or each sensor over a predetermined time range and/or over a predetermined number of cycles, in particular to determine, preferably estimate, its sensor-based sensor states, and/or to determine, in particular predict, exclusively its model-based sensor states.

In addition, the respective Kalman filter of the associated sensor can be deactivated if the specified time range and/or the specified number of cycles is exceeded and sensor values from this sensor are missing.

Furthermore, the respective Kalman filter of the associated sensor can be configured to feed the sensor-based sensor states determined for this sensor over the specified time range and/or over the specified number of cycles and/or the exclusively model-based sensor states to the associated weighting module for weighting and to weight them.

For example, the weighting module can be configured to combine all sensor-based sensor states determined over the specified time range and/or over the specified number of cycles and/or exclusively model-based sensor states of the respective sensor in a weighted manner.

In addition, the weighting module can be configured to compare the number of sensor-based sensor states determined over the specified time range and/or over the specified number of cycles and the determined number of model-based sensor states of the respective sensor and to weight them in such a way that, if the number of sensor-based sensor states is higher than the number of model-based sensor states of the respective sensor over the specified time range or over the specified number of cycles, a weighted sensor state of this sensor is weighted higher than if the number is reversed.

Optionally, both the determined number of sensor-based sensor states and the determined number of model-based sensor states of all Kalman filters as well as the determined weighted sensor states of all sensors can be fed to a normalization module which is configured to normalize or standardize the sensor states and/or weighted sensor states of all sensors.

In addition, the optionally unified sensor states and/or weighted sensor states of all sensors can be fed to the fusion module, which is configured to link the optionally unified sensor states and/or weighted sensor states with the respective dynamic weighting and, as a result, to determine the sensor fusion value for the considered predefined period of time or for the considered predefined cycles.

The sensors of the sensor arrangement, in particular designed as a multi-sensor arrangement, are, for example, microelectronic sensors that measure mechanical, physical, and/or chemical variables. The sensors can, for example, consist of at least the following sensors: environmental sensors, radar sensors, lidar sensors, camera sensors, speed sensors, acceleration sensors, yaw rate sensors, temperature sensors, inclination sensors, 3D sensors, 2D sensors, or the like.

The method according to the invention for operating the sensor arrangement described above, in particular for fusing the sensor signals of the multiple sensors of this sensor arrangement by means of dynamic weighting of determined sensor states, provides that each sensor comprises an associated Kalman filter which cyclically either determines, in particular estimates, a sensor-based sensor state or exclusively determines, in particular predicts, a model-based sensor state, wherein the determined sensor-based sensor states or the determined model-based sensor states are subsequently weighted depending on their relationship to one another and these weighted sensor states of all sensors are fused and a sensor fusion value is determined.

The previously described sensor arrangement and the described dynamic weighting method can be used, for example, in measuring systems with multiple sensors which record and transmit measured values in the second range, for example 3 seconds, for example in keyless go systems, security systems such as collision monitoring systems, temperature monitoring systems in electric vehicles or the like.

The dynamic weighting method enables a highly responsive and accurate sensor array that dynamically adapts to changes in the sensor behavior of the associated sensors during each cycle, for example, whether the sensors are actively sensing and transmitting sensor signals or not. Dynamic weighting (also called the dynamic weighting method) is particularly applicable to short-term, real-time applications that require rapid response to changing conditions, such as motion, temperature, or power tracking. By taking sensor malfunctions into account, it ensures that the final sensor fusion value is more accurately adjusted to the state to be determined based on the current sensor activity.

Embodiments of the invention are explained in more detail below with reference to a drawing.

• 1 schematisch ein Beispiel für eine Sensoranordnung mit einem zugehörigen Funktionsablaufplan,
• 2 schematisch ein Beispiel für die Sensoranordnung mit zugehörigen Kalman-Filtern und deren Funktionsfähigkeit in Abhängigkeit von Sensoreingangssignalen,
• 3 schematisch ein Diagramm zur Ermittlung eines Sensorfusionswertes über mehrere Messzyklen.

Showing:

• 1 schematically an example of a sensor arrangement with an associated functional flow chart,
• 2 schematically an example of the sensor arrangement with associated Kalman filters and their functionality depending on sensor input signals,
• 3 schematic diagram for determining a sensor fusion value over several measurement cycles.

Corresponding parts are provided with the same reference numerals in all figures.

1 shows schematically an example of a sensor arrangement 1 with an associated functional flow chart.

The sensor arrangement 1 comprises a plurality of sensors S1 to Sm.

The sensors Sm can, for example, be microelectronic sensors that measure mechanical, physical, and/or chemical variables. In particular, the sensors Sm are installed in a vehicle and serve to monitor vehicle functions and/or vehicle variables, such as vehicle speed, vehicle acceleration, vehicle environment, or the like. For example, the sensors Sm can be designed as at least environmental sensors, radar sensors, lidar sensors, camera sensors, speed sensors, acceleration sensors, yaw rate sensors, temperature sensors, inclination sensors, 3D sensors, 2D sensors, or the like.

The sensors Sm measure regularly, in particular in predetermined cycles Zn and/or over a predetermined time period T (shown in 2 ), physical variables such as vehicle speed, temperature, vehicle environment, or the like. The resulting sensor signals Si(S1) to Si(Sm) can be stored, in particular buffered, in associated memories 2.1, 2.2 to 2m, optionally with a time stamp. Instead of one memory 2.m per sensor Sm, a single memory 2.m with multiple memory areas can also be provided. This allows sensor signals Si(Sm) from the past to be read, analyzed, and/or re-evaluated.

The sensor arrangement 1 further comprises at least one weighting module G1 to Gm, to which the sensor signals Si(S1) to Si(Sm) of the sensors Sm can be fed. For example, an associated weighting module Gm is provided for each sensor Sm for determining weighted sensor states gSZ(S1) to gSZ(Sm) for the respective sensor Sm based on the supplied sensor signals Si(S1) to Si(Sm) of the respective sensor Sm.

The at least one weighting module Gm is followed by a sensor fusion module 4, to which at least the weighted sensor states gSZ(S1) to gSZ(Sm) of all sensors Sm can be fed.

Each sensor Sm comprises an associated Kalman filter KF1 to KFm, which is configured to cyclically either determine, in particular estimate, a sensor-based sensor state sSZ(S1) to sSZ(Sm), or exclusively determine, in particular predict, a model-based sensor state mSZ(S1) to mSZ(Sm).

The respective Kalman filter KFm can be, for example, a moving average filter, a sum filter, a linear Kalman filter, an unscented Kalman filter, an extended Kalman filter or the like.

The respective weighting module Gm is configured to weight the determined sensor states mSZ(Sm) and/or sSZ(Sm) depending on their relationship to each other and to determine weighted sensor states gSZ(S1) to gSZ(Sm) for each sensor Sm.

The sensor fusion module 4 is configured to fuse the weighted sensor states gSZ(S1) to gSZ(Sm) of all sensors Sm and to determine a sensor fusion value SFW and to store it in a memory 8.

Optionally, a normalization module 6 can be connected between the weighting module(s) Gm and the sensor fusion module 4. The normalization module 6 is configured, for example, to normalize the weighted sensor states gSZ(Sm), in particular to standardize them, as shown in the example according to 2 is described in more detail.

The sensor arrangement 1 according to the invention is based on a dynamic weighting method which adapts weighting factors GF1 to GFm (shortly called weights) in real time on the basis of the results, in particular of the supplied sensor signals Si(Sm), of the Kalman filter KFm, which can be influenced by sensor timeouts or signal losses of the respective sensor S1 to Sm and are taken into account by the Kalman filter KFm.

2 shows schematically an example of the sensor arrangement 1 with three sensors S1 to S3 with the associated Kalman filters KF1 to KF3 for each sensor S1 to S3 and their functionality as a function of the measured sensor signals Si(S1) to Si(S3), which can be or are supplied as sensor input signals to the weighting module G1 to Gm over the predetermined time range T, for example every 10 ms within a time range T of one day or one hour or the like, and/or in predetermined cycles Zn, for example every 10 ms or every 10 s.

The sensor arrangement 1 according to the invention enables a dynamic weighting of the measured sensor signals Si(S1) to Si(S3) without precise knowledge of the properties of the sensors S1 to S3.

For conventional sensor arrangements 1 it is essential to consider error margins, responses to influences such as environmental influences, and to know the variability of measurement accuracy depending on location and use.

The sensor arrangement 1 according to the invention is based on a new dynamic weighting method that does not require detailed information about sensor properties. Instead, the sensor arrangement 1 according to the invention uses the data inherent in the respective Kalman filter F1 to Fm to adapt to the unique properties of each sensor S1 to Sm.

For this reason, the sensor array 1 includes a corresponding Kalman filter KFm for each sensor Sm. In other words, each sensor Sm is paired with its own Kalman filter KFm.

This configuration improves the accuracy of the recorded values/measurement data of the sensor signals Si(Sm) with each new measurement.

Since each sensor Sm is coupled to or includes a specific Kalman filter KFm, the corresponding Kalman filter KFm adapts to the specific properties of the associated sensor Sm. Consequently, the behavior of the sensors Sm is encapsulated and represented by the Kalman filters KFm themselves. This allows the reliability of the sensors Sm to be derived directly from the Kalman filters KFm, which are then used directly to calculate the weights used in sensor fusion.

For example, the Kalman filter KFm can be implemented in an electronic unit of the respective electronic sensor Sm as a software module.

The dynamic weighting method does not require extensive knowledge of the sensors Sm used, thus creating a robust sensor array 1 without the need for additional information.

The sensor arrangement 1 is configured, in particular via its modules such as the Kalman filters KFm, the weighting module Gm, the optional normalization module 6 and the sensor fusion module 4, to interpolate the previously acquired sensor signals Si(Sm) of the respective sensor Sm and optionally the missing sensor information when a sensor Sm temporarily stops transmitting measurement data (= sensor signals Si(Sm)) but then receives them again. For this purpose, the Kalman filter KFm of the respective sensor Sm is configured accordingly.

For this purpose, the respective Kalman filter KFm comprises an associated model, in particular a mathematical model, which is configured to adapt to the availability of the respective sensor Sm, regardless of the physical quantity to be measured and/or other measurement parameters and/or sensor properties. For example, the respective Kalman filter KFm is configured to estimate/predict the respective system state of the respective sensor Sm based on measurement activity, in particular acquisition times/availability of the respective sensor Sm for measuring or non-acquisition times/unavailability of the respective sensor Sm for measuring, and to generate a model-based sensor state mSZ(Sm).

In prior art filters, if, for example, a sensor Sm no longer transmits measurement data and thus no longer transmits sensor signals Si(Sm) for an extended period of time, the accuracy of the filter in question decreased significantly, as this filter was dependent only on its internal mathematical model. These known models are inherently limited and cannot fully capture the complexity of the tracked parameters, especially when it comes to highly nonlinear behaviors such as human movement. Consequently, the fused values also deviated from the actual situation if the weights were not adjusted to account for this increasing inaccuracy.

The sensor arrangement 1 according to the invention is therefore based on a dynamic weighting method which reacts to these inaccuracies by adapting weighting factors GF1 to GFm (shortly called weights) in real time based on the results of the Kalman filters KFm, which can be influenced by sensor timeouts or signal losses.

The respective weighting module Gm automatically reduces the weighting factor GFm of the sensor(s) Sm with less reliable sensor signals Si(Sm) or no sensor signals Si(Sm), for example due to signal losses, outliers or the like, while increasing the weighting factor GFm of those sensors Sm that provide more consistent measurement data, thereby improving the resilience of the entire sensor arrangement 1 against sensor malfunctions or limitations.

The dynamic weighting method operates within the sensor array 1, which is designed as a sensor fusion system. The sensor array 1 is based on sensor-assigned Kalman filters KFm. Each sensor Sm integrated into the fusion process is assigned its own Kalman filter KFm. These Kalman filters KFm are capable of improving the accuracy of the Sensors Sm to refine the parameter values reported by using both sensor measurements and a predefined mathematical model.

When a sensor Sm does not provide any measurements, the implemented Kalman filter KFm relies exclusively on its mathematical model to make predictions and determines or generates exclusively model-based sensor states mSZ(Sm). Therefore, the output of the Kalman filter KFm is referred to as a "prediction" when it is based solely on the model.

However, if a sensor Sm provides measurements, the implemented Kalman filter KFm can rely on the sensor signals Si(Sm) and optionally on the mathematical model to determine or generate sensor-based sensor states sSZ(Sm). This output of the Kalman filter KFm, based on both the sensor signals Si(Sm) and the mathematical model, is referred to as an "estimate" because it combines both the model and the actual sensor measurements.

It should be noted that while the mathematical model(s) of the Kalman filter (KFm) are sufficiently accurate, they cannot perfectly replicate the subtleties of real-world parameters. There is always a discrepancy between model predictions and actual parameters. Sensor measurements are also inherently subject to error.

The provided Kalman filters KFm are designed to reduce these inaccuracies by combining the errors of the sensor measurements and the mathematical model to determine, in particular estimate, the sensor-based sensor state sSZ(Sm).

The sensor arrangement 1 thus makes it possible to differentiate for each sensor Sm between the sensor-based estimation and thus the combined model- and sensor-based estimation and the exclusively model-based prediction and preferably to combine these and to use them for the dynamic adaptation of the weighting factors GFm of the respective sensor Sm.

An estimate of the model- and sensor-based estimate of the sensor-based sensor state sSZ(Sm) generated by a well-calibrated Kalman filter KFm is generally more accurate than a single measurement or a prediction of the model-based sensor state mSZ(Sm) based solely on the mathematical model. Consequently, the estimate of the sensor-based sensor state sSZ(Sm) by the Kalman filter KFm is closer to the actual value ("ground truth") than the prediction of the model-based sensor state mSZ(Sm) by the Kalman filter KFm, which does not consider sensor data.

2 illustrates the estimation/prediction of the respective Kalman filter KFm of sensor arrangement 1.

As shown, each Kalman filter KFm generates an update A (= updated estimate) of the sensor-based sensor state sSZ(Sm) when it receives a measurement M from its corresponding sensor S1 to S3. If no measurement M is available, the Kalman filter KFm generates a prediction V of the exclusively model-based sensor state mSZ(Sm).

During each cycle Zn, each Kalman filter KFm provides either a prediction V or an update A of the estimate, depending on the availability of measurements M and thus of acquired sensor signals Si(Sm). Once all Kalman filters KFm have completed their calculations, their results are fed to the associated weighting module Gm and combined and weighted using the dynamic weighting method, specifically generating weighted sensor states gSZ(Sm), as shown in 1 This dynamic weighting method ensures that the sensor array 1 remains unaffected by the randomness of the sensor measurements or the asynchrony between the sensors Sm.

The dynamic weighting method takes advantage of the available sensor signals Si(Sm) by tracking the number of predictions V and updates A (= estimates) produced by each Kalman filter in the last cycles Zn.

In the 2 For example, in the system shown, the Kalman filter KF1 for sensor S1 has produced six updates A (= estimates) and five predictions V in the last eleven cycles Zn. In contrast, the Kalman filter KF3 for sensor S3 produced four updates A (= estimates) and seven predictions V in the same period. This information is analyzed and evaluated using the weighting module Gm. For example, the weighting module G3 evaluates the output of the Kalman filter KF1 for sensor S1 as likely more accurate and thus higher than that of the Kalman filter KF3 for sensor S3, since more exclusively model-based predictions V were executed over the same period T than for sensor S1 with more sensor-based updates A.

• Der Gewichtungsfaktor GFm hoch ist bei höherer Genauigkeit eines Sensors Sm über n-Zyklen Zn oder über einen Zeitbereich T, wenn die Anzahl von Aktualisierungen A (= Schätzungen) größer ist als die Anzahl von Vorhersagen V: A > V

For such dynamic weighting, the weighting modules Gm can, for example, comprise and implement a corresponding weighting function for generating the weighting factor GFm. For example, the weighting modules Gm can have implemented a weighting function such that

• The weighting factor GFm is high at higher accuracy of a sensor Sm over n cycles Zn or over a time range T, if the number of updates A (= estimates) is greater than the number of predictions V: A > V

This weighting function (= logic) can be extended to scenarios in which a sensor Sm is inactive for a longer time period T or exceeds a specified time period T or a specified number of cycles Zn.

If the sensor Sm times out, the associated Kalman filter KFm will not make any predictions V (as indicated by the dashed frame as an option in 2 shown) because it does not receive any new/current measurements M (sensor signals Si(Sm)) and thus cannot update itself with new measurements M. Consequently, the reliability of the mathematical predictions V decreases with less available information.

In this case, the sensor array 1, specifically the weighting module Gm, counts how many updates A (=estimates), how many predictions V, and how many "no outputs" (dashed border of the predictions V) were performed by each Kalman filter KFm. For example, if the first Kalman filter KF1 delivered seven updates A, two predictions V, and two cycles with "no results" in the last eleven cycles Zn, while the second Kalman filter KF2 delivered two updates A, five predictions V, and four "no results," the weighting algorithm of the weighting module Gm would assign a larger weighting factor GF1 to the first Kalman filter KF1. This is because the first Kalman filter KF1 contains more recent data.

The dynamic weighting method takes these considerations into account to determine the appropriate weight for each sensor Sm in the subsequent fusion process of the sensor fusion module 4.

At the end of each cycle Zn, when all Kalman filters KFm have completed their filtering functions and detections, the weighting module Gm counts and stores the number of predictions V, updates A (= estimates), and "no outputs" for the last "n" cycles Zn. To determine the corresponding weighting factors GFm for each sensor Sm and derive them from these detections, for example, the number of updates A and the number of predictions V can be combined with the predefined weighting factors GFm, which prioritize (favor) updates A, while "no outputs" are evaluated with a value of zero.

• - der Multiplikationsfaktor größer ist für A > V.

In other words, the respective weighting factor GFm of the associated sensor Sm is higher the higher the number of its updates A compared to its predictions V or no measurement results over the time range T under consideration or over the number of predefined cycles Zn under consideration. The respective weighting factor GFm is, for example, a multiplication factor, where:

• - the multiplication factor is larger for A > V.

The weighted sensor states gSZ(Sm) of the respective/all sensors Sm can then be summed using the weighting module Gm and divided by the total number of considered cycles Zn to obtain a non-normalized weighting factor nGFm (shown in 1 ) for each sensor Sm. This procedure is performed for all sensors Sm.

Subsequently, the unnormalized weighting factors nGFm are normalized to a scale between 0 and 1 using the normalization module 6 and output as a dynamic weight dG. A normalization function underlying the normalization module 6 is configured such that sensors Sm with a larger number of updates A (estimates) receive a higher dynamic weight dG than sensors Sm with more predictions V.

After normalizing or unifying the unnormalized weighting factors nGFm of all sensors Sm, the output of each Kalman filter KFm in the current cycle Zn (be it a prediction V or an update A) for sensor fusion is multiplied by the respective dynamic weight dG by the sensor fusion module 4 and all are added to obtain a final fused value, the sensor fusion value SFW for the respective cycle Zn or the cycles Zn to be considered.

1 illustrates the process by which the algorithm determines the fused sensor signals Si(Sm) in each cycle Zn using the dynamic weighting method. 1 it is evident that the number of Senso ren Sm has no influence on the functioning of the algorithm.

3 shows a comparison of the determined sensor fusion values SFW according to the invention with dynamic weighting dG and without and thus according to the state of the art. 3 shows a comparison between the adaptation of the dynamic weighting dG with the dynamic weighting method for sensor S1 with first solid line VL1 and for sensor S2 with first dashed line GL1 according to the invention and the adaptation according to the known transition weighting method for sensor S1 with second solid line VL2 and for sensor S2 with second dashed line GL2.

The well-known transition weighting method with a second solid line VL2 and a second dashed line GL2 does not distinguish between updates A (estimates), predictions V, and "no outputs"; it simply gradually reduces the weight of a sensor S1, S2 to zero when the respective sensor S1, S2 stops transmitting data for a certain period. The weight is then increased when the sensor S1, S2 transmits information again, with a corresponding reduction in the weight of the other sensor S2, S1.

In contrast, the dynamic weighting method, implemented in the sensor array 1, reacts to changes in real time at each cycle Zn and distinguishes between updates A (estimates), predictions V and “no outputs”, as described previously, this is shown in 3 based on the cyclic changes of the respective dynamic weighting dG of the first solid line VL1 for sensor S1 and the first dashed line GL1 for sensor S2.

The dynamic weighting method of the previously described sensor array 1 offers an alternative to conventional approaches when detailed knowledge about the reliability of the sensors Sm is lacking. It exploits the inherent capabilities of the implemented Kalman filters KFm for each sensor Sm to determine reliability weights without requiring additional information or introducing additional delays.

The dynamic weighting method results in a highly responsive sensor array 1 that adapts to changes in sensor behavior every cycle Zn, for example, whether the sensors Sm are actively acquiring and transmitting sensor signals Si(Sm). The dynamic weighting method is particularly valuable in short-term, real-time applications that require rapid response to changing conditions, such as motion, temperature, or current tracking. By accounting for sensor malfunctions, it ensures that the final sensor fusion value SFW is more accurately aligned with the ground truth and sensor activity.

The sensor arrangement 1 and the described dynamic weighting method can be used, for example, in measuring systems with several sensors Sm, which record and transmit measured values in the second range, for example 3 seconds, for example in keyless go systems, security systems, such as collision monitoring systems, temperature monitoring systems in electric vehicles or the like.

List of reference symbols

1

Sensor arrangement

2.1 to 2.m

memory

4

Sensor fusion module

6

Standardization module

8

memory

dG

dynamic weighting

GL1, GL2

dashed lines

gSZ(S1) to gSZ(Sm)

weighted sensor state

KF1 to KFm

Kalman filter

mSZ(S1) to mSZ(Sm)

model-based sensor state

nGFm

non-standardized weighting factor

sSZ(S1) to sSZ(Sm)

sensor-based sensor state

GF1 to GFm

Weighting factor

G1 to Gm

Weighting module

SFW

Sensor fusion value

S1 to Sm

sensor

Si(S1) to Si(Sm)

Sensor signal

VL1, VL2

Solid lines

Claims (12)

Sensor arrangement (1), characterized by - a plurality of sensors (S1 to Sm), - at least one weighting module (G1 to Gm) to which sensor signals (Si(S1) to Si(Sm)) of the sensors (S1 to Sm) can be fed, and - a sensor fusion module (4) connected downstream of the at least one weighting module (G1 to Gm), to which weighted sensor states (gSZ(S1) to gSZ(Sm)) can be supplied, wherein each sensor (S1 to Sm) comprises an associated Kalman filter (KF1 to KFm) which is configured to cyclically determine either a sensor-based sensor state (sSZ(S1) to s(SZ(Sm)) or to determine, in particular to predict, a model-based sensor state (mSZ(S1) to mSZ(Sm)), wherein the weighting module (G1 to Gm) is configured to weight the determined sensor states (sSZ(S1) to sSZ(Sm), mSZ(S1) to mSZ(Sm)) depending on their relationship to one another, and wherein the sensor fusion module (4) is configured to fuse the weighted sensor states (gSZ(S1) to gSZ(Sm)) of all sensors (S1 to Sm) and to generate a sensor fusion value (SFW) to be determined. Sensor arrangement (1) according to Claim 1 , characterized in that the respective Kalman filter (KF1 to KFm) is set up to process current sensor signals (Si(S1) to Si(Sm)) and for each cycle (Zn) - either to estimate the sensor-based sensor state (sSZ(S1) to sSZ(Sm)) as the sensor state by means of a model based on the current sensor signals (Si(S1) to Si(Sm)) and on the basis of model data or only based on the current sensor signals (Si(S1) to Si(Sm)) - or, if no current sensor signals (Si(S1) to Si(Sm)) are available, to predict the model-based sensor state (mSZ(S1) to mSZ(Sm)) as the sensor state exclusively by means of the model based on model data. Sensor arrangement (1) according to Claim 1 or 2 , characterized in that the Kalman filter (KF1 to KFm) is set up to determine the sensor states of the respective sensor (S1 to Sm) or each sensor (Sm) over a predetermined time range (T) and/or over a predetermined number of cycles (Zn), in particular to determine, in particular to estimate, its sensor-based sensor states (sSZ(S1) to sSZ(Sm)), and/or to determine, in particular to predict, exclusively its model-based sensor states (mSZ(S1) to mSZ(Sm)). Sensor arrangement (1) according to Claim 3 , characterized in that the respective Kalman filter (KF1 to KFm) of the associated sensor (S1 to Sm) can be deactivated if the predetermined time range (T) and/or the predetermined number of cycles (Zn) is exceeded and sensor values (Si(S1) to Si(Sm)) are missing from this sensor (S1 to Sm). Sensor arrangement (1) according to Claim 3 , characterized in that the respective Kalman filter (KF1 to KFm) of the associated sensor (S1 to Sm) is set up to supply the sensor-based sensor states (sSZ(S1) to sSZ(Sm)) determined over the predetermined time range (T) and/or over the predetermined number of cycles (Zn) for this sensor (S1 to Sm) and/or the exclusively model-based sensor states (mSZ(S1) to mSZ(Sm)) to the associated weighting module (G1 to Gm) for weighting. Sensor arrangement (1) according to Claim 3 or 5 , characterized in that the weighting module (G1 to Gm) is set up to combine all sensor-based sensor states (sSZ(S1) to sSZ(Sm)) and/or exclusively model-based sensor states (mSZ(S1) to mSZ(Sm)) of the respective sensor (S1 to Sm) determined over the predetermined time range (T) and/or over the predetermined number of cycles (Zn) in a weighted manner. Sensor arrangement (1) according to Claim 3 , 5 or 6 , characterized in that the weighting module (G1 to Gm) is set up to compare the number of sensor-based sensor states (sSZ(S1) to sSZ(Sm)) determined over the predetermined time range (T) or over the predetermined number of cycles (Zn) and the determined number of model-based sensor states (mSZ(S1) to mSZ(Sm)) of the respective sensor (S1 to Sm) and to weight them in such a way that, if the number of sensor-based sensor states (sSZ(S1) to sSZ(Sm)) is higher than the number of model-based sensor states (mSZ(S1) to mSZ(Sm)) of the respective sensor (S1 to Sm) over the predetermined time range (T) or over the predetermined number of cycles (Zn), a weighted sensor state (gSZ(S1) to gSZ(Sm)) of this sensor (S1 to Sm) is weighted higher than if the number is reversed. Sensor arrangement (1) according to Claim 7 , characterized in that both the determined number of sensor-based sensor states (sSZ(S1) to sSZ(Sm)) and the determined number of model-based sensor states (mSZ(S1) to mSZ(Sm)) of all Kalman filters (KF1 to KFm) as well as the determined weighted sensor states (gSZ(S1) to gSZ(Sm)) of all sensors (S1 to Sm) can be fed to a normalization module (6) which is set up to standardize the weighted sensor states (gSZ(S1) to gSZ(Sm)) of all sensors (S1 to Sm). Sensor arrangement (1) according to one of the preceding claims, characterized in that the weighted sensor states (gSZ(S1) to gSZ(Sm)) of all sensors (S1 to Sm) can be fed to the sensor fusion module (4), which is set up to link the weighted sensor states (gSZ(S1) to gSZ(Sm)) with the respective dynamic weighting (dG) and to determine the resulting sensor fusion value (SFW) for the considered specified period of time (T) or for the considered specified cycles (Zn). Sensor arrangement (1) according to one of the preceding claims, characterized in that the sensors (S1 to Sm) are microelectronic sensors (S1 to Sm) which measure mechanical, physical and/or chemical variables. Sensor arrangement (1) according to one of the preceding claims, characterized in that the sensors (S1 to Sm) are formed from at least the following sensors (S1 to Sm): environmental sensors, radar sensors, lidar sensors, camera sensors, speed sensors, acceleration sensors, yaw rate sensors, temperature sensors, inclination sensors, 3D sensors, 2D sensors or the like. Method for operating a sensor arrangement (1), in particular for fusing sensor signals (Si(S1) to Si(Sm)) of a plurality of sensors (S1 to Sm) of the sensor arrangement (1) by means of dynamic weighting (dG) of determined sensor states, wherein each sensor (S1 to Sm) comprises an associated Kalman filter (KF1 to KFm) which cyclically either determines, in particular estimates, a sensor-based sensor state (sSZ(S1) to sSZ(Sm)) or exclusively determines, in particular predicts, a model-based sensor state (mSZ(S1) to mSZ(Sm)), wherein subsequently the determined sensor-based sensor states (sSZ(S1) to sSZ(Sm)) or the determined model-based sensor states (mSZ(S1) to mSZ(Sm)) are weighted depending on their relationship to one another and the weighted sensor states (gSZ(S1) to gSZ(Sm)) of all sensors (S1 to Sm) are fused and a sensor fusion value (SFW) is determined.