DE102020210902A1 - Computer-implemented method for calibrating sensor data of a rotation angle sensor, training method of a machine learning algorithm and magnetic resonance gyroscope system - Google Patents

Abstract

The invention relates to a computer-implemented method for calibrating sensor data from a rotation angle sensor (10) during measurement operation. The method uses a machine learning algorithm (A1) which is set up to determine a predefined target function (F) which represents a calibrated sensor signal from a rotation angle sensor (10). The invention also relates to a training method for the machine learning algorithm (A1), a magnetic resonance gyroscope system (1), a computer program and a computer-readable data carrier.

Description

The invention relates to a computer-implemented method for calibrating sensor data from a rotation angle sensor during measurement operation.

The invention also relates to a computer-implemented method for providing a trained machine learning algorithm for calibrating sensor data of a rotation angle sensor of a magnetic resonance gyroscope system during measurement operation.

The invention further relates to a nuclear magnetic resonance gyroscope system.

The machine learning algorithm is an algorithm trained or optimized using a machine learning method, e.g. a gradient method.

Precise sensor data is essential for the autonomous navigation of aircraft and vehicles. For this purpose, high-precision yaw rate sensors based on optical resonators are usually installed in aircraft, since a large pack size and high costs are justifiable there. High-precision on-board sensors are also required for autonomous driving and flying as well as for applications that have a poor connection to GPS, radar and similar systems, such as underwater navigation.

However, the size and required volume play an important role here. There are also gyroscopes that evaluate nuclear magnetic resonance signals from atomic nuclei with a non-vanishing magnetic moment. These show increased drift stability and increased accuracy compared to the MEMS yaw rate sensors currently used in the automotive industry.

Optically pumped angular rate sensors are based on a vapor cell. The resulting spin-larmor precession co larmor is read out in a vapor cell. An outer rotation co rot represents an additional rotation, which can be determined by reading out the rotation frequency ω_mess as follows: $$\mathrm{\omega }\_\text{mess}=\mathrm{\omega }\_\text{larmor}\pm \mathrm{\omega }\_\text{Red}.$$

However, these gyroscopes are also susceptible to certain types of noise that are correlated over time, which in particular can lead to the sensor oscillating.

So far, attempts have been made to suppress such temporal effects through shielding, construction or regulation. This causes costs and limits the size.

Models with which a correction is attempted are available for a few relationships. However, these often require an additional measurement unit or the addition of a measurement variable, e.g. by another isotope whose spin is measured. Disturbance variables such as laser frequency fluctuations, laser intensity fluctuations, laser width fluctuations, laser polarization fluctuations, temporal magnetic field fluctuations, spatial magnetic field fluctuations, field frequency fluctuations, vibrations, temperature fluctuations, temperature gradients and fluctuations in the supply current and supply voltage of the sensor are due to specific influences on a measured frequency and a measured amplitude, as well as due to their specific effects on the just mentioned properties of pump and detection laser, characterizable and separable from the measurement signal.

However, it turns out that these quantities can only be modeled to a small extent due to the non-linear and unknown interactions and relationships.

EP 22 82 167 A1 discloses a nuclear magnetic resonance gyro system consisting of a gyro cell sealed to confine an alkali metal vapor and a gyromagnetic isotope. The system further includes a magnetic field generator configured to generate a substantially uniform magnetic field provided in the gyro cell to cause gyromagnetic isotope precedence. The system further includes a rotation angle sensor configured to measure a rotation angle about an axis of the system based on the measured precession angle of the gyromagnetic isotope.

The invention is therefore based on the object of providing an improved method and system for calibrating sensor data from a rotation angle sensor.

The object is achieved with a computer-implemented method for calibrating sensor data of a rotation angle sensor during measurement with the features of patent claim 1 .

Furthermore, the object is achieved with a computer-implemented method for providing a trained machine learning algorithm for calibrating sensor data of a rotation angle sensor during measurement operation with the features of patent claim 8 .

In addition, the task is solved with a nuclear magnetic resonance gyroscope system with the features of patent claim 12 .

Disclosure of Invention

The present invention creates a computer-implemented method for calibrating sensor data, in particular rotation angle data, of a rotation angle sensor during measurement operation.

The method includes receiving a data set of sensor data comprising sensor data that is variable over time of a sensor signal from the angle of rotation sensor, the sensor data having a measurement component and a noise component based on a plurality of disturbance variables.

The method also includes determining at least one numerical value of a target function for the data set of sensor data using a machine learning algorithm, which is applied to the data set of sensor data, wherein the target function represents calibrated sensor data of the rotation angle sensor that corresponds to the measurement portion of the sensor data of the rotation angle sensor.

In addition, the method includes outputting a second data set comprising the at least one numerical value of the target function.

The present invention also provides a computer-implemented method for providing a trained machine learning algorithm for calibrating sensor data, in particular rotation angle data, of a rotation angle sensor during measurement operation.

The method includes receiving a first data set of input training data comprising time-variable first sensor data, in particular actual sensor data, of the rotation angle sensor and time-variable second sensor data, in particular actual sensor data, of at least one further sensor, the first sensor data having a measurement portion and a have a noise component based on a plurality of disturbance variables.

The method includes receiving a second data set of output training data, in particular setpoint sensor data, of the rotation angle sensor, which is applied to the rotation angle sensor via a test station.

The method also includes training a machine learning algorithm to determine the numerical value of a target function for the first data set of sensor data, the target function representing the measured portion of the first sensor data of the rotation angle sensor, calibrated first sensor data of the rotation angle sensor.

The present invention also provides a magnetic resonance gyroscope system. The system includes a vapor cell that includes an alkali metal and a noble gas isotope. The system further includes a magnetic field source configured to generate a magnetic field that is aligned with an axis of the magnetic resonance gyroscope system and that passes through the vapor cell. to precess the alkali metal and noble gas isotope.

The system also has a pump laser that is set up to generate an optical pump beam that spin-polarizes the alkali metal. In addition, the system includes a probe laser configured to generate an optical probe beam.

In addition, the system includes a rotation angle sensor configured to monitor the optical probe beam and to detect a rotation of the nuclear magnetic resonance gyroscope system about a sensitive axis based on a modulation of the optical probe beam in response to a precession of the noble gas isotope caused by the spin polarization of the Alkali metal originates to measure.

The system also has a computing device which is set up to execute the computer-implemented method according to the invention for calibrating sensor data, in particular rotation angle data, of a rotation angle sensor of a magnetic resonance gyroscope system in measurement operation.

The present invention also creates a computer program with program code in order to carry out the method according to the invention when the computer program is run on a computer.

The present invention also creates a computer-readable data carrier with program code of a computer program in order to carry out the method according to the invention when the computer program is run on a computer.

One idea of the present invention is to free the actual sensor signal from non-linear and/or unknown noise sources by using the trained machine learning algorithm to determine an essentially noise-free target value of the sensor data or the sensor signal of the rotation angle sensor.

The disturbance variables superimposed on the sensor signal have hitherto unknown interactions, with the present invention being able to free the sensor signal of the angle of rotation sensor, which changes over time, from the aforementioned non-linear or non-trivial influences and thus the quality of the sensor signal can be increased .

The method according to the invention is an efficient calibration method for sensor data from sensors that is suitable for industrial production. Thus, for example, production fluctuations can be compensated for by individual training of the specific target function.

A higher accuracy and drift stability of the sensor can thus be made possible due to the individual calibration, in particular in connection with temporal interactions of noise sources. The self-learning system can thus be used to correct the non-trivial influences of disturbance variables that are temporally correlated with one another.

Advantageous embodiments and developments result from the dependent claims and from the description with reference to the figures.

According to a preferred development, it is provided that the machine learning algorithm, in particular an artificial neural network, is set up to determine a numerical value of a prediction uncertainty of the numerical value of the target function representing the calibrated sensor data of the rotation angle sensor.

The algorithm is thus advantageously able to determine and output this additional information, which in turn can serve as a basis for decision-making for subsequent method steps.

According to a further preferred development, it is provided that if the numerical value of the prediction uncertainty of the numerical value of the target function representing the calibrated sensor data of the rotation angle sensor exceeds a predetermined threshold value, the sensor data of the rotation angle sensor, instead of the second data set output by the machine learning algorithm, comprising the at least one numerical value of the objective function can be used.

Due to the presence of the prediction uncertainty data as a basis for decision-making, a prediction reliability of the method according to the invention can thus be increased.

According to a further preferred development, it is provided that the data set of sensor data is received by the machine learning algorithm and, in parallel thereto, by a filter, with the filter at least partially freeing the sensor data from the noise component and a third data set comprising at least one numerical value from the measurement component outputs calibrated sensor data of the rotation angle sensor corresponding to the sensor data of the rotation angle sensor.

The method according to the invention thus advantageously has an additional mechanism in the form of the filter, which is set up to free the sensor signal received from the angle-of-rotation sensor from its noise component.

According to a further preferred development, it is provided that the second data set output by the machine learning algorithm is calculated linearly with the third data set output by the filter using a numerical value of the prediction uncertainty of the numerical value of the target function representing the calibrated sensor data of the rotation angle sensor to form a fourth data set will.

As a result, by providing the redundant mechanisms or function modules, an average increase in the reliability of the prediction can advantageously be made possible.

wobei µ der numerische Wert der Vorhersageunsicherheit des die kalibrierten Sensordaten des Drehwinkelsensors repräsentierenden numerischen Werts der Zielfunktion ist.According to a further preferred development, it is provided that the linear calculation of the second data set with the third data set is given by the following equation: $$\left(1-\mu \right)\text{numeric value}\left(\text{W2}\right)+\mu \text{numeric value}\left(\text{w1}\right)$$

where μ is the numerical value of the prediction uncertainty of the numerical value of the target function representing the calibrated sensor data of the rotation angle sensor.

In this way, a statistical reduction in the prediction uncertainty can be made possible in an advantageous manner.

According to a further preferred development, it is provided that if the numerical value of the prediction uncertainty exceeds a predetermined threshold value, only the third data set output by the filter is used, and if the numerical value of the prediction uncertainty falls below the predetermined threshold value, the second data set output by the machine learning algorithm is used.

It can thus be advantageously always ensured that the method always uses a calibration result that is as precise as possible.

According to a further preferred development, it is provided that the noise component of the first data set is generated by applying the plurality of disturbance variables to the angle of rotation sensor via the test station, with the disturbance variables being superimposed on the first sensor data.

It is thus advantageously possible to superimpose any number of disturbance variables on the sensor signal and thus train the algorithm using a large data set to identify non-linear correlations between the sensor signal, the disturbance variables and possibly a further plurality of available sensor signals.

die verrauschten Drehwinkelsensor-Daten s̃_t zum Zeitpunkt t ermittelt, wobei s _t die vom Algorithmus maschinellen Lernens entrauschten Drehwinkelsensor-Daten, $n_t^{i}i\in 1\dots N$

1 ... N Daten weiterer Sensoren, θ die Gewichte des Algorithmus maschinellen Lernens und T vorherige Schritte sind, die dem Algorithmus maschinellen Lernens rekurrent zugeführt werden.According to a further preferred development, it is provided that the target function for the first data set is a function of ${\tilde{s}}_t,n_{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="DE102020210902A1\_0012.png"\; state="\{\{state\}\}">t</figure-callout>}}^1,...,n_t^N;\theta ,s_{t-T:t-1},n_{t-T:t-1}^1,...,n_{t-T:t-1}^N$

determines the noisy rotation angle sensor data s̃ _t at time t, where s _t the rotation angle sensor data denoised by the machine learning algorithm, $n_t^{i}i\in 1...N$

1...N data of other sensors, θ are the weights of the machine learning algorithm, and T are previous steps fed recurrently to the machine learning algorithm.

The algorithm can thus advantageously be trained in such a way that it achieves an optimization with each training step or carries out an improved determination of the target function.

According to a further preferred development, it is provided that the at least one further sensor is a Hall sensor, a magnetometer, a temperature sensor, an acceleration sensor and/or a current sensor, and the plurality of disturbance variables are laser frequency fluctuations, laser intensity fluctuations, laser width fluctuations, laser polarization fluctuations, magnetic field fluctuations over time , spatial magnetic field fluctuations, field frequency fluctuations, vibrations, temperature fluctuations, temperature gradients, fluctuations in a supply current and/or fluctuations in a supply voltage of the angle of rotation sensor.

The disturbance variables therefore include a large number of factors which can influence the sensor signal from the angle-of-rotation sensor.

The configurations and developments described can be combined with one another as desired.

Further possible configurations, developments and implementations of the invention also include combinations of features of the invention described above or below with regard to the exemplary embodiments that are not explicitly mentioned.

character list

The accompanying drawings are provided to provide a further understanding of embodiments of the invention. They illustrate embodiments and, together with the description, serve to explain principles and concepts of the invention.

Other embodiments and many of the foregoing advantages will become apparent by reference to the drawings. The illustrated elements of the drawings are not necessarily shown to scale with respect to one another.

• 1 eine schematische Darstellung eines Kernspinresonanz-Gyroskopsystems gemäß einer bevorzugten Ausführungsform der Erfindung;
• 2 ein Ablaufdiagramm eines computerimplementierten Verfahrens zum Kalibrieren von Sensordaten eines Drehwinkelsensors im Messbetrieb gemäß einer ersten Ausführungsform der Erfindung;
• 3a ein Ablaufdiagramm des computerimplementierten Verfahrens zum Kalibrieren von Sensordaten des Drehwinkelsensors im Messbetrieb gemäß einer zweiten Ausführungsform der Erfindung;
• 3b ein Ablaufdiagramm des computerimplementierten Verfahrens zum Kalibrieren von Sensordaten des Drehwinkelsensors im Messbetrieb gemäß einer dritten Ausführungsform der Erfindung;
• 4 ein Ablaufdiagramm des computerimplementierten Verfahrens zum Kalibrieren von Sensordaten des Drehwinkelsensors im Messbetrieb gemäß der ersten bis dritten Ausführungsform der Erfindung; und
• 5 ein computerimplementiertes Verfahren zum Bereitstellen eines trainierten Algorithmus maschinellen Lernens zum Kalibrieren von Sensordaten eines Drehwinkelsensors im Messbetrieb gemäß einer bevorzugten Ausführungsform der Erfindung.

Show it:

• 1 a schematic representation of a magnetic resonance gyroscope system according to a preferred embodiment of the invention;
• 2 a flowchart of a computer-implemented method for calibrating sensor data of a rotation angle sensor during measurement according to a first embodiment of the invention;
• 3a a flowchart of the computer-implemented method for calibrating sensor data of the angle of rotation sensor in measurement mode according to a second embodiment of the invention;
• 3b a flowchart of the computer-implemented method for calibrating sensor data of the angle of rotation sensor in measurement mode according to a third embodiment of the invention;
• 4 a flowchart of the computer-implemented method for calibrating sensor data of the rotation angle sensor in measurement mode according to the first to third embodiment of the invention; and
• 5 a computer-implemented method for providing a trained algorithm machine learning for calibrating sensor data of a rotation angle sensor in measurement mode according to a preferred embodiment of the invention.

In the figures of the drawings, the same reference symbols designate the same or functionally identical elements, parts or components, unless otherwise stated.

1 Figure 12 shows a schematic representation of a magnetic resonance gyroscope system according to a preferred embodiment of the invention.

The nuclear magnetic resonance gyroscope system 1 has a vapor cell 18 comprising an alkali metal 18a and a noble gas isotope 18b.

The nuclear magnetic resonance gyroscope system 1 further comprises a magnetic field source 20 which is arranged to generate a magnetic field which is aligned with an axis of the nuclear magnetic resonance gyroscope system 1 and which passes through the vapor cell 18 to the alkali metal 18a and the noble gas isotope 18b precess.

The magnetic field source 20 generates one or more magnetic fields in one or more directions, e.g., a static magnetic field and at least one alternating magnetic field.

In addition, the nuclear magnetic resonance gyroscope system 1 has a pump laser 22 and first optics 23 . The pump laser 22 is set up to generate an optical pump beam through the first optics 23, which spin-polarizes the alkali metal 18a.

Furthermore, a probe laser 25 is provided, as well as second optics 27. The probe laser 25 is set up to use the second optics 27 to generate a preferably linearly polarized optical probe beam.

The magnetic resonance gyroscope system 1 further comprises a rotation angle sensor 10 configured to monitor the optical probe beam and a rotation of the magnetic resonance gyroscope system 1 about a sensitive axis based on a modulation of the optical probe beam in response to a precession of the noble gas isotope 18b resulting from the spin polarization of the alkali metal 14 to be measured.

For this purpose, the rotation angle sensor 10 has a photodetector 10a and a rotation detection component 10b.

In addition, a computing device 24 is provided, which is set up to execute the computer-implemented method according to the invention for calibrating sensor data, in particular rotation angle data, of a rotation angle sensor 10 of a magnetic resonance gyroscope system 1 in measurement operation.

2 shows a flow chart of a computer-implemented method for calibrating sensor data of a rotation angle sensor in measurement mode according to a first embodiment of the invention.

The method uses a machine learning algorithm A1, which is applied to a data set DS1 of sensor data comprising first sensor data D1, which is variable over time, of a sensor signal from the rotation angle sensor 10 of the magnetic resonance gyroscope system 1.

The sensor data D1 have a measurement portion MA and a noise portion RA based on a plurality of disturbance variables 14 . The machine learning algorithm A1 approximates or determines at least one numerical value W1 of a target function F for the data set DS1 of sensor data. The target function F represents calibrated sensor data of the rotation angle sensor 10 corresponding to the measurement portion MA of the sensor data D1 of the rotation angle sensor 10.

The machine learning algorithm A1 is also set up to output a second data set DS2 comprising the at least one numerical value W1 of the target function F.

The machine learning algorithm A1 is preferably implemented by an artificial neural network.

The machine learning algorithm A1 is set up to determine a numerical value μ of a prediction uncertainty of the numerical value W1 of the target function F representing the calibrated sensor data D1 of the rotation angle sensor 10 .

If the numerical value µ of the prediction uncertainty of the numerical value W1 of the objective function F representing the calibrated sensor data D1 of the rotation angle sensor 10 exceeds a predetermined threshold value SW, the nuclear magnetic resonance gyroscope system 1 uses the time-varying sensor data D1 of the sensor signal of the rotation angle sensor 10 instead of the machine generated by the algorithm Learning A1 output second data set DS2, which uses the at least one numerical value W1 of the target function F.

Since the sensor signal or the first data set DS1 based on it from the angle of rotation sensor 10 is sensor data that is variable over time, the machine learning algorithm A1 receives sensor data continuously or according to a sampling frequency of the angle of rotation sensor 10 continuously or at specified intervals during measurement operation and thus outputs it also according to the predetermined intervals or continuously numerical values of the target function in order to constantly calibrate or remove noise from the sensor data of the rotation angle sensor 10 during measurement operation.

3a shows a flow chart of the computer-implemented method for calibrating the rotation angle sensor during measurement operation according to a second embodiment of the invention.

The first data set DS1 of sensor data is received by the machine learning algorithm A1 and, in parallel thereto, by a filter A2. The filter is a log-in filter.

The filter A2 removes the noise component RA from the sensor data D1 at least in part and outputs a third data set DS3 comprising at least one numerical value W2 of the calibrated sensor data D1 of the rotation angle sensor 10 corresponding to the measurement component MA of the sensor data D1 of the rotation angle sensor 10.

The second data set DS2 output by the machine learning algorithm A1 is combined with the third data set DS3 output by the filter A2 using a numerical value μ of the prediction uncertainty of the numerical value W1 of the target function F representing the calibrated sensor data D1 of the angle of rotation sensor 10 to form a fourth data set DS4 calculated linearly FU.

wobei µ der numerische Wert der Vorhersageunsicherheit des die kalibrierten Sensordaten D1 des Drehwinkelsensors 10 repräsentierenden numerischen Werts W1 der Zielfunktion F ist.The linear calculation of the second data record DS2 with the third data record DS3 is given by the following equation: $$\left(1-\mu \right)\text{numeric value}\left(\text{W2}\right)+\mu \text{numeric value}\left(\text{w1}\right)$$

where μ is the numerical value of the prediction uncertainty of the numerical value W1 of the target function F representing the calibrated sensor data D1 of the rotation angle sensor 10 .

If the numerical value μ of the prediction uncertainty exceeds a predetermined threshold value SW, the magnetic resonance gyroscope system 1 only uses the third data set DS3 output by the filter A2.

If the numerical value μ of the prediction uncertainty falls below the predetermined threshold value SW, the magnetic resonance gyroscope system 1 uses the second data set DS2 output by the machine learning algorithm A1.

3b shows a flow chart of the computer-implemented method for calibrating sensor data of the angle of rotation sensor during measurement operation according to a third embodiment of the invention.

If the numerical value μ of the prediction uncertainty exceeds a predetermined threshold value SW, the magnetic resonance gyroscope system 1 only uses the third data set DS3 output by the filter A2.

If the numerical value μ of the prediction uncertainty falls below the predetermined threshold value SW, the magnetic resonance gyroscope system 1 uses the second data set DS2 output by the machine learning algorithm A1.

4 shows a flow chart of the computer-implemented method for calibrating sensor data of the rotation angle sensor in measurement mode according to the first to third specific embodiments of the invention.

The method includes receiving S1 a data set DS1 of sensor data comprising time-variable sensor data D1 of a sensor signal of the rotation angle sensor 10 of the nuclear magnetic resonance gyroscope system 1, the sensor data D1 having a measurement portion MA and a noise portion RA based on a plurality of disturbance variables 14.

The method also includes determining S2 at least one numerical value W1 of a target function F for the data set DS1 of sensor data using a machine learning algorithm A1, which is applied to the data set DS1 of sensor data, with the target function F corresponding to the measurement portion MA of the sensor data D1 of the rotation angle sensor 10 corresponding, calibrated sensor data D1 of the rotation angle sensor 10 represents.

In addition, the method includes an output S3 of a second data set DS2 comprising the at least one numerical value W1 of the target function F.

5 FIG. 1 shows a computer-implemented method for providing a trained machine learning algorithm for calibrating sensor data of a rotation angle sensor in measurement mode according to a preferred embodiment of the invention.

The method includes receiving S1' a first data set DS1 of input training data comprising time-variable first sensor data D1, in particular actual sensor data, of the rotation angle sensor 10 and time-variable second sensor data D2, in particular actual sensor data, of at least one further sensor 12 of the magnetic resonance Gyroscope system 1, the first sensor data D1 having a measurement component MA and a noise component RA based on a plurality of disturbance variables 14 .

Alternatively, the first sensor data D1 and the second sensor data D2 can be contained in two different data sets, for example, and can be received separately by the machine learning algorithm A1.

The method also includes receiving S2 ′ of a second data set DS2 of output training data, in particular target sensor data, of rotation angle sensor 10 , which is applied to rotation angle sensor 10 via a test station 16 .

The method also includes training S3' of a machine learning algorithm A1 to determine the numerical value W1 of a target function F for the first data set DS1 of sensor data, the target function F corresponding to the measured portion MA of the first sensor data D1 of the rotation angle sensor 10, calibrated first sensor data D1 of the rotation angle sensor 10 represents.

The machine learning algorithm A1 is trained S3′ using a gradient method.

The noise component RA of the first data set DS1 is generated by applying the plurality of disturbance variables 14 to the angle of rotation sensor 10 via the test station 16 . The disturbance variables 14 are superimposed on the first sensor data D1.

wobei s̃_t verrauschte Drehwinkelsensor-Daten zum Zeitpunkt t, s _t die vom Algorithmus maschinellen Lernens A1 entrauschten Drehwinkelsensor-Daten, nt i ∈ 1 ... N Daten weiterer Sensoren, θ die Gewichte des Algorithmus maschinellen Lernens A1 und T vorherige Schritte sind, die dem Algorithmus maschinellen Lernens A1 rekurrent zugeführt werden.The target function F for the first data set DS1 is given by the following equation: $$f\left({\tilde{s}}_t,n_{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="DE102020210902A1\_0012.png"\; state="\{\{state\}\}">t</figure-callout>}}^1,...,n_t^N;\theta ,s_{t-T:t-1},n_{t-T:t-1}^1,...,n_{t-T:t-1}^N\right)={\overline{s}}_t.$$

where s̃ _t noisy rotation angle sensor data at time t, s _t is the rotation angle sensor data denoised by machine learning algorithm A1, n i ∈ 1...N data of other sensors, θ is the weights of machine learning algorithm A1, and T is previous steps fed recurrently to machine learning algorithm A1.

The at least one additional sensor 12 of the nuclear magnetic resonance gyroscope system 1 is a Hall sensor. Alternatively, the at least one further sensor 12 of the nuclear magnetic resonance gyroscope system 1 can be formed by a magnetometer, a temperature sensor, an acceleration sensor and/or a current sensor.

The plurality of disturbance variables 14 are laser frequency fluctuations, laser intensity fluctuations, laser width fluctuations, laser polarization fluctuations, temporal magnetic field fluctuations, spatial magnetic field fluctuations, field frequency fluctuations, vibrations, temperature fluctuations, temperature gradients, fluctuations in a supply current and/or fluctuations in a supply voltage of the angle of rotation sensor 10.

The nuclear magnetic resonance gyroscope system 1 also includes a computer program with program code in order to carry out the method according to the invention when the computer program is executed on a computer or a computing device, in particular the nuclear magnetic resonance gyroscope system 1 .

The nuclear spin resonance gyroscope system 1 also includes a computer-readable data carrier with program code of a computer program to carry out the method according to the invention when the computer program is executed on a computer or a computing device, in particular the nuclear spin resonance gyroscope system.

The method according to the invention for calibrating sensor data, in particular rotation angle data, of a rotation angle sensor 10 during measurement operation is described above in connection with a nuclear magnetic resonance gyroscope system 1 by way of example. However, there are numerous alternative possible uses of the method in other technical areas in which rotation angle sensors are used.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• EP 2282167 A1 [0012]

Claims (14)

Computer-implemented method for calibrating sensor data, in particular rotation angle data, of a rotation angle sensor (10) in measurement mode, with the steps: Receiving (S1) a data set (DS1) of sensor data comprising sensor data (D1) variable over time of a sensor signal from the angle of rotation sensor (10), the sensor data (D1) having a measurement component (MA) and a noise component ( RA); Determining (S2) at least one numerical value (W1) of a target function (F) for the data set (DS1) of sensor data using a machine learning algorithm (A1), which is applied to the data set (DS1) of sensor data, the target function (F ) represents calibrated sensor data (D1) of the angle of rotation sensor (10) corresponding to the measured portion (MA) of the sensor data (D1) of the angle of rotation sensor (10); and Outputting (S3) a second data set (DS2) comprising the at least one numerical value (W1) of the target function (F). Computer-implemented method claim 1 , wherein the machine learning algorithm (A1), in particular an artificial neural network, is set up to calculate a numerical value (µ) of a prediction uncertainty of the numerical value (W1) of the target function (F ) to determine. Computer-implemented method claim 2 , wherein if the numerical value (µ) of the prediction uncertainty of the numerical value (W1) of the target function (F) representing the calibrated sensor data (D1) of the rotation angle sensor (10) exceeds a predetermined threshold value (SW), the sensor data (D1) of the rotation angle sensor ( 10) instead of the second data set (DS2) output by the machine learning algorithm (A1), comprising the at least one numerical value (W1) of the target function (F). Computer-implemented method claim 2 , the data set (DS1) of sensor data (D1) being received by the machine learning algorithm (A1) and in parallel thereto by a filter (A2), the filter (A2) separating the sensor data (D1) at least in part from the noise component (RA ) and outputs a third data set (DS3) comprising at least one numerical value (W2) from the measured portion (MA) of the sensor data (D1) of the rotation angle sensor (10) corresponding to calibrated sensor data (D1) of the rotation angle sensor (10). Computer-implemented method claim 4 , wherein the second data set (DS2) output by the machine learning algorithm (A1) matches the third data set (DS3) output by the filter (A2) using a numerical value (µ) of the prediction uncertainty of the calibrated sensor data (D1) of the rotation angle sensor (10) representing numerical value (W1) of the target function (F) to a fourth data set (DS4) is calculated linearly. Computer-implemented method claim 5 , where the linear calculation of the second data set (DS2) with the third data set (DS3) is given by the following equation: $$\left(1-\mu \right)\text{numeric value}\left(\text{W2}\right)+\mu \text{numeric value}\left(\text{w1}\right)$$

where μ is the numerical value of the prediction uncertainty of the numerical value (W1) of the target function (F) representing the calibrated sensor data (D1) of the rotation angle sensor (10). Computer-implemented method claim 5 , wherein if the numerical value (µ) of the prediction uncertainty exceeds a predetermined threshold value (SW), only the third data set (DS3) output by the filter (A2) is used, and wherein if the numerical value (µ) of the prediction uncertainty exceeds the predetermined threshold value (SW) fails, the second data set (DS2) output by the machine learning algorithm (A1) is used. Computer-implemented method for providing a trained machine learning algorithm for calibrating sensor data, in particular rotation angle data, of a rotation angle sensor (10) in measuring operation, with the steps: receiving (S1') a first data set (DS1) of input training data comprising first sensor data ( D1), in particular actual sensor data, of the rotation angle sensor (10) and temporally variable second sensor data (D2), in particular actual sensor data, of at least one further sensor (12), the first sensor data (D1) having a measurement portion (MA) and a having a noise component (RA) based on a plurality of disturbance variables (14); receiving (S2') a second data set (DS2) of output training data, in particular target sensor data, of the angle of rotation sensor (10), which are applied to the angle of rotation sensor (10) via a test station (16); and training (S3') a machine learning algorithm (A1) to determine the numerical value (W1) of a target function (F) for the first data set (DS1) of sensor data, the target function (F) the measurement portion (MA) of the first sensor data (D1) of the rotation angle sensor (10) corresponding, calibrated first sensor data (D1) of the rotation angle sensor (10). Computer-implemented method claim 8 , wherein the noise component (RA) of the first data set (DS1) is generated by applying the plurality of disturbance variables (14) via the test station (16) to the angle of rotation sensor (10), the disturbance variables (14) being based on the first sensor data (D1) be superimposed. Computer-implemented method claim 8 or 9 , where the objective function (F) for the first data set (DS1) depends on ${\tilde{s}}_t,n_t^1,...,n_t^N;\theta ,s_{t-T:t-1},n_{t-T:t-1}^1,...,n_{t-T:t-1}^N$

the noisy rotation angle sensor data s̃ _t determined at time t, where s̃ _t the rotation angle sensor data de-noised by the machine learning algorithm (A1), $n_t^{i}i\in 1...N$

Data from other sensors, θ the weights of the machine learning algorithm (A1), and T are previous steps fed recurrently to the machine learning algorithm (A1). Computer-implemented method according to one of Claims 8 until 10 , wherein the at least one further sensor (12) is a Hall sensor, a magnetometer, a temperature sensor, an acceleration sensor and/or a current sensor, and wherein the plurality of disturbance variables (14) are laser frequency fluctuations, laser intensity fluctuations, laser width fluctuations, laser polarization fluctuations, temporal magnetic field fluctuations, spatial magnetic field fluctuations, field frequency fluctuations, vibrations, temperature fluctuations, temperature gradients, fluctuations in a supply current and/or fluctuations in a supply voltage of the rotation angle sensor (10). A nuclear magnetic resonance gyroscope system (1) comprising: a vapor cell (18) comprising an alkali metal (18a) and a noble gas isotope (18b); a magnetic field source (20) arranged to generate a magnetic field which is aligned with an axis of the magnetic resonance gyroscope system (1) and which passes through the vapor cell (18) to form the alkali metal (18a) and the noble gas isotope (18b ) to precess; a pump laser (22) configured to generate an optical pump beam that spin-polarizes the alkali metal (18a); a probe laser (25) arranged to generate an optical probe beam; a rotation angle sensor (10) arranged to monitor the optical probe beam and a rotation of the nuclear magnetic resonance gyroscope system (1) about a sensitive axis based on a modulation of the optical probe beam in response to a precession of the noble gas isotope (18b), the to measure from the spin polarization of the alkali metal (14); and a computing device (24) which is set up to carry out the method according to one of Claims 1 until 7 to execute. Computer program with program code to implement the method according to one of Claims 1 until 7 to be performed when the computer program is run on a computer. Computer-readable data carrier with program code of a computer program for the method according to one of Claims 1 until 7 to be performed when the computer program is run on a computer.

Images