A magnetic positioning and velocity measurement method for synthesizing cross-axis effects of an artificial intelligence network model

Abstract

The application discloses a magnetic positioning and speed measuring method for synthesizing and correcting cross-axis effects of an artificial intelligence network model, relates to the field of magnetic positioning and speed measuring, and aims to improve the positioning and linear speed measuring precision. In the first step, an artificial intelligence deep learning framework training network model is constructed, which comprises a cross-axis effect correction module, a three-axis tilt and sensitivity correction module and a sensor center correction module; the collected magnetic field data is subjected to the artificial intelligence deep learning framework training network model to obtain correction parameters; in the second step, the correction parameters are used to correct the magnetic field data, so that the corrected magnetic field data can be applied to positioning and speed measurement.

Description

TECHNICAL FIELD

[0001] The application relates to the field of magnetic positioning and speed measurement, in particular to a magnetic positioning and speed measurement method for comprehensively calibrating cross-axis effects by using an artificial intelligence network model. BACKGROUND

[0002] The magnetic positioning and speed measurement technology has high precision of sub-millimeter level and can be applied to precise positioning and speed measurement of an observed object. A three-axis anisotropic magnetoresistance sensor with high sensitivity, low power consumption, small size and low cost is commonly used. However, the three-axis measurement is sometimes affected by cross-axis effects: when the magnetic field perpendicular to the measurement axis is strong, the reading value of the measurement axis is disturbed, leading to measurement distortion. In addition, due to the influence of manufacturing process limitations or the influence of circuit board welding factors, the three-axis sensor has three-axis slight tilt and sensitivity difference, and the center position of the sensor is offset. These factors are sometimes confused with each other, which together causes measurement error and affects the positioning accuracy and linear speed measurement accuracy of the observed object.

[0003] In the prior art, some anisotropic magnetic field sensors offset the cross-axis effect by reversing the magnetic field coil through hardware, for example, CN109932669A, a digital compensation type AMR magnetic field measurement device, but the volume of the sensor is increased, and the cost of the three-axis sensor is also increased. The application trains a miniature artificial intelligence network model to systematically solve the cross-axis effect and other problems when calculating the positioning and linear speed without increasing the hardware cost. SUMMARY

[0004] The application aims to provide a magnetic positioning and speed measurement method for comprehensively calibrating cross-axis effects by using an artificial intelligence network model, which aims to solve the measurement error problem caused by cross-axis effects, three-axis tilt, sensor center position offset and the like in the prior art, and improve the positioning and linear speed measurement accuracy.

[0005] In order to solve the technical problems in the prior art, the application adopts the following technical solutions: first, obtaining correction parameters: a sensor array is provided, the sensor is a three-axis anisotropic magnetoresistance type sensor, magnetic field data output by each sensor at different positions and different orientations of a magnet is collected, an artificial intelligence deep learning framework training network model is constructed, including a cross-axis effect correction module, a three-axis tilt and sensitivity correction module and a sensor center correction module, the collected magnetic field data is used to train the artificial intelligence deep learning framework training network model to obtain correction parameters; second, applying the correction parameters to positioning and speed measurement: reading the magnetic field data output by each sensor, correcting the magnetic field data by using the correction parameters, calculating the relative position of the magnet and the sensor according to the corrected magnetic field data, and finally obtaining the speed of the magnet relative to the sensor according to the change rate of the relative position of the magnet and the sensor with time.

[0006] As a further scheme of the present application: the cross-axis effect correction module performs n times of iterative calculation on the magnetic field value of the i th sensor output ( V ix ,V iy ,V iz ) to obtain the corrected magnetic field value ( H ix ’(n) ,H iy ’(n) ,H iz ’(n) ), and the iterative formula is: wherein (Bx, By, Bz) are parameters that need to be trained and optimized; the initial value of iteration is H ikx ,H iky ,H ikz ; the iteration is stopped when the iteration is up to the N th time, and the corrected magnetic field value obtained in the N th iteration is recorded as (Bx, By, Bz). H ix ’(0) = 0, H iy ’(0) = 0, H iz ’(0) = 0; when the iteration is up to the N th time, the iteration is stopped when (Bx, By, Bz) is less than or equal to a set value, and the corrected magnetic field value obtained in the N th iteration is recorded as (Bx, By, Bz). , , H ix ’ , H iy ’ , H iz ’ .

[0007] As a further scheme of the present application: the three-axis tilt and sensitivity correction module uses a 3x3 linear transformation matrix A i3x3 to correct (Bx, By, Bz) to (Bx, By, Bz). H ix ’ , H iy ’ , H iz ’ H ix , H iy , H ​​iz ), the calculation formula is: .

[0008] As a further scheme of the application: the sensor center correction module adopts the formula: correct the sensor center position, wherein d i ( Δ ix ,Δ iy , Δ iz ) is the distance vector of the ith sensor to the magnet, x 、 y 、 z is the position of the magnet;( x isensor 、 y isensor 、 z isensor ) is the ideal position of the ith sensor center known;( Δ ix ,Δ iy ,Δ iz ) is the correction parameter of the ideal position of the sensor center, which is 3 parameters to be trained.

[0009] As a further scheme of the application: the formula is: d i ( Δ ix ,Δ iy ,Δ iz ) and the 3 components of the magnet towards u ( u x ,u y , u z ) are substituted into the dipole model to calculate the dipole model magnetic field vector H i ; wherein m 0 is the magnetic distance size constant of the magnet measured in advance; for the ith sensor, the artificial intelligence deep learning framework trains the network model with the measured magnetic field of the sensor after correction ( H ix , H iy , H iz ) and the dipole model magnetic field vectorH i The mean square error of the above formula is taken as the loss function Loss: The artificial intelligence deep learning framework trains the network model based on the Pytorch framework float64 precision, and uses the Adam function as the optimization function to obtain all the correction parameters H ikx ,H iky ,H ikz ), A i3x3 , ( Δ ix ,Δ iy ,Δ iz ).

[0010] As a further scheme of the present application: during training, first make all the correction parameters of the sensors the same, train the same correction parameters with the data collected by all the sensors, and the correction parameters obtained by training are regarded as the mean value of the correction parameters of all the sensors; then, each sensor takes the mean value as the initial value, adjusts and trains according to the data collected by the sensor, and obtains the correction parameter of each sensor.

[0011] As a further scheme of the present application: according to the formula: Given the initial guess value of the position ( x , y , z ) and orientation ( u x ,u y ,u z ) of the magnet, the magnetic field value at the position of each sensor is calculated by substituting the dipole formula, and the square difference between the magnetic field value obtained by correction of all the sensors ( H ix , H iy , H iz ) is calculated, and the final optimized value of ( x , y , z ) and ( u x ,u y ,u z ) is obtained by L-M iteration.

[0012] As a further scheme of the present application: the position ( x , y, z ) Numerical derivation of time t to obtain the linear speed of the magnet .

[0013] Advantages of the present application:

[0014] 1. Multiple correction to improve measurement accuracy: joint correction of cross-axis effect, three-axis tilt, and interference caused by sensor center offset.

[0015] 2. Support large-scale parallel computing: use deep learning framework to realize batch processing of multi-sensor data, meet the real-time requirements of positioning and speed measurement.

[0016] 3. No additional hardware: correction method through program calculation, no increase in sensor power consumption and volume, suitable for consumer-level devices.

[0017] 4. Strong adaptability: through training, it can adapt to individual differences of different models and different batches of sensors.

[0018] 5. Simple calibration process: only a good symmetry magnet and a fixed layer plate are needed, without complex external equipment.

[0019] 6. Ensure the stability of each sensor parameter training: during training, first make all sensor correction parameters the same, use data collected by all sensors to train the same correction parameters, and the trained correction parameters are considered as the mean of all sensor correction parameters. Then each sensor takes the mean as the initial value, and fine-tunes the training according to the data collected by the sensor to obtain the correction parameters of each sensor. This helps to ensure the stability of each sensor parameter training. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a schematic diagram of data collection of the magnet above multiple sensors.

[0021] Figure 2 is the network model of the i-th sensor correction.

[0022] Figure 3 is a method for calculating the relative position of the magnet according to the output of all sensors. DETAILED DESCRIPTION

[0023] The application relates to a magnetic positioning and speed measuring method for synthetically correcting cross-axis effect of an artificial intelligence network model. In order to solve the measurement error problems caused by cross-axis effect, three-axis tilt and sensor center position offset in the prior art, the following methods are adopted. One is a primary correction method: data acquisition is performed to obtain data required for training an artificial intelligence network model; a network model is constructed, which comprises correction of cross-axis effect, further correction of three-axis tilt and sensitivity, and correction of a sensor magnetic field measurement center; the network model is trained by using the collected data to obtain parameters of the above correction. Two is an application method for positioning and speed measurement: in actual measurement, the magnetic field value measured by the artificial intelligence network model is corrected to calculate the positioning and speed.

[0024] As shown in the accompanying drawings, Figure 1 In the primary correction method, data acquisition is performed on a plurality of three-axis anisotropic magnetoresistance type sensors 2 above a plurality of positions and different orientations of a symmetric magnet 1 (such as an axially magnetized cylinder), and the magnetic field readings of the sensors 2 are collected to form a group of data containing the positions ( x , y , z ) orientations ( u x ,u y ,u z ) and the readings of each sensor ( V ix ,V jy ,V kz ) of the magnet 1. In particular, after collecting data at each position, the N / S poles of the magnet are preferably adjusted, and the sensors are sampled again as another group of data. Preferably, the data of the sensors 2 whose magnetic field readings change by more than a set value after the N / S poles of the magnet are adjusted are sampled again to enhance the training of the correction of the cross-axis effect.

[0025] In the primary correction method of the embodiment, the magnet 1 used is a neodymium iron boron material axially magnetized symmetric cylinder with a diameter of 1 cm and a height of 1 cm. The data are collected by using a 4x4 sensor 2 array, and the sensor model is LIS3MDL. As shown in the accompanying drawings, Figure 1 Above the sensor 2 array, the layer plate for fixing the magnet 1 contains a large number of pre-processed and engraved slot holes that can accommodate the magnet 1. In order to prevent the cylindrical magnet 1 from falling through the slot hole when standing, a layer of layer plate without slot hole can be placed under the layer plate with slot hole to support the magnet 1. When the magnet is placed in the slot hole, the position ( x , y , z ) and orientation ( u x,u y ,u z ). With the help of a large number of slots in the layer board, the magnet 1 is placed at various positions and directions in space, and the magnetic field readings of the sensor 2 are collected to form a set of data of magnet positions and sensor readings.

[0026] In the initial correction method, an artificial intelligence network model is constructed as shown in the attached figure Figure 2 , including the following trainable correction modules:

[0027] The first is the cross-axis effect correction module, which is aimed at the magnetic field value output by the i th sensor (B i ) and the magnetic field value output by the j th sensor (B j ). V ix ,V iy ,V iz The correction calculation method of multiple iterations is as follows: Wherein (B i ) is the magnetic field value output by the i th sensor, (B j ) is the magnetic field value output by the j th sensor, and (B i ) is the magnetic field value output by the i th sensor. H ikx ,H iky ,H ikz ) is the parameter that needs to be trained and optimized. The initial value of iteration H ix ’(0) = 0, H iy ’(0) = 0, H iz ’(0) = 0. In practice, different iteration times need to be tried until the iteration H ix ’(n) , H iy ’(n) , H iz ’(n) relative H ix ’(n-1) , H iy ’(n-1) , H iz ’(n-1) change value within the set value range; or even if the iteration number is increased, the training effect cannot be improved. In the above iteration, H ix ’(n-1) , H iy ’(n-1) , H iz ’(n-1)The order is not symmetrical, which reflects the essence of the cross-axis effect of the sensor used, that is, the correlation between the axes.

[0028] The network model further includes a triaxial tilt and sensitivity correction module. The output of the i-th sensor cross-axis effect module ( H ix ’(n) , H iy ’(n) , H iz ’(n) ), abbreviated as ( H ix ’ , H iy ’ , H iz ’ After a 3x3 linear transformation matrix A i3x3 Output corrected ( H ix , H iy , H iz The calculation method is as follows: The network model includes the calculation of the magnetic field generated by magnet 1 on each sensor 2, which is obtained by the dipole model: in d i It is the distance vector from the i-th sensor 2 to the magnet 1, calculated using the sensor center correction method below; u It is a unit vector indicating the orientation of the known magnet 1, consisting of 3 components ( u x ,u y ,u z ); m 0 is the magnitude constant of the magnetic moment of magnet 1, which was measured in advance.

[0029] In the network model, further correction of the sensor center is added, as shown below: in x , y , z This is position 1 of magnet; x isensor , y isensor , z isensor ) is the known ideal location of the center of the i-th sensor 2; Δix ,Δ iy ,Δ iz ) is the correction of the ideal position of the sensor 2 center, and is 3 parameters to be trained.

[0030] The d i ( Δ ix ,Δ iy ,Δ iz ) and the 3 components of the magnet 1 towards u ( u x ,u y ,u z ) are substituted into the dipole model, and the calculated dipole model magnetic field vector H i . For the i-th sensor 2, the network model uses the corrected magnetic field measured by the sensor H ix , H iy , H iz ) and the mean square error of the dipole model magnetic field vector H i as the loss function (Loss). The network model is trained based on the Pytorch framework float64 precision, and Adam is used as the optimization function to obtain all the correction parameters H ikx , H iky ,H ikz ), A i3x3 , Δ ix ,Δ iy ,Δ iz .

[0031]

[0032] During training, the correction parameters of all sensors 2 are first made the same, and the same correction parameters are trained using data collected by all sensors 2. The correction parameters obtained by training are considered as the mean value of the correction parameters of all sensors 2. Then each sensor 2 takes this mean value as the initial value, and fine-tunes the training according to the data collected by this sensor 2 to obtain the correction parameters of each sensor 2. This helps to ensure the stability of the parameter training of each sensor 2.

[0033] After the above corrections, in the application methods of positioning and speed measurement, the outputs of all sensors 2 are known ( V ix , V iy ,V iz First, the position of magnet 1 needs to be calculated. x , y , z ). As attached Figure 3 The output of each sensor 2 passes through the cross-axis effect correction module and the triaxial tilt and sensitivity correction module to obtain the corrected magnetic field. H ix , H iy , H iz The position of magnet 1 (). x , y , z ) and orientation ( u x ,u y ,u z The magnetic field, after being optimized and corrected by all sensor 2 outputs, is obtained through an optimization algorithm. H ix , H iy , H iz ), and the position of magnet 1 ( x , y , z ) and orientation ( u x ,u y ,u z Substituting the values ​​into the dipole formula, the magnetic field of each sensor 2 is calculated to minimize the difference in squared values. A commonly used optimization algorithm is to, given the position of magnet 1 (…),… x , y , z ) and orientation ( u x ,u y ,u z The initial guess value of ) is obtained by LM iteration. x , y , z )and( u x ,uy ,u z The final optimized value.

[0034]

[0035] The magnetic field of each sensor 2 calculated by the dipole formula is as follows, showing only the position of magnet 1 ( x , y , z ) and orientation ( u x ,u y ,u z ) is the unknown quantity to be optimized, and the rest are known quantities, including ( Δ ix ,Δ iy ,Δ iz ) is a known quantity obtained through the above correction method.

[0036]

[0037] In positioning and speed measurement applications, magnet 1 can be replaced, as long as its volume is small enough that it can be approximated as a dipole for sensor 2. To measure the velocity of magnet 1 over a fixed distance, fixed front and rear boundaries are set, for example, the front boundary is... x = x0 The back boundary is x = x1 Monitor every moment. t Positioning of magnet 1 x t , y t , z t Record entry to the front boundary. x = x0 time t0 and positioned as ( x0 , y t1 , z t1 ), departure boundary time t1 and positioned as ( x1 , y t1 , z t1 Then the velocity of magnet 1 after passing through a fixed distance is:

[0038] In the positioning and velocity measurement method, the above method is essentially to measure the relative positioning and relative linear velocity of the magnet 1 to the sensor. It can be used for multiple sensor fixation, and the magnet 1 moves relative to the sensor; it can also be used for magnet 1 fixation, and multiple sensors as a whole are in motion, for example, multiple sensors are fixed on a moving object, there is a magnet in each interval on the ground, and the position and speed through the fixed distance of the moving object are measured.

Claims

1. A magnetic positioning and velocity measurement method for synthesizing and calibrating cross-axis effects of an artificial intelligence network model, characterized in that: The first step is to obtain the correction parameters: set up a sensor array, the sensor uses a three-axis anisotropic magnetoresistance type sensor, collect the magnetic field data output by each sensor at different positions and different orientations; build an artificial intelligence deep learning framework training network model, including a cross-axis effect correction module, a three-axis tilt and sensitivity correction module, and a sensor center correction module; use the collected magnetic field data to train the network model through the artificial intelligence deep learning framework, and obtain the correction parameters; the second step is to apply the correction parameters to positioning and speed measurement: read the magnetic field data output by each sensor, correct the magnetic field data using the correction parameters, and calculate the relative position of the magnet and the sensor according to the corrected magnetic field data, and finally obtain the linear velocity of the magnet relative to the sensor according to the rate of change of the relative position of the magnet and the sensor with time, wherein the cross-axis effect correction module performs n times of iteration calculation on the magnetic field value output by the i-th sensor to obtain the corrected magnetic field value ( V ix ,V iy ,V iz ) and the iteration formula is: H ix ’(n) ,H iy ’(n) ,H iz ’(n) H ikx ,H iky ,H ikz ) is a parameter that needs to be trained and optimized; the initial value of iteration H ix ’(0) =0, H iy ’(0) =0, H iz ’(0) =0; the corrected magnetic field value obtained by the n-th iteration is denoted as ( H ix ’ ,H iy ’ ,H iz ’ ).​​ The three-axis tilt and sensitivity correction module uses a 3x3 linear transformation matrix A i3x3 The( H ix ’ , H iy ’ , H iz ’ ) correction is( H ix , H iy , H iz ), and the calculation formula is: .

2. The magnetic positioning and velocity measurement method of synthesizing cross-axis effects of an artificial intelligence network model according to claim 1, characterized in that: The sensor center correction module adopts the formula: The sensor center position is corrected, wherein d i ( Δ ix ,Δ iy , Δ iz is a distance vector from the ith sensor to the magnet, x 、 y 、 z is a position where the magnet is located;( x isensor 、 y isensor 、 z isensor is an ideal position of the ith sensor center which is known;( Δ ix ,Δ iy ,Δ iz is a correction parameter of the sensor center ideal position, and is three parameters to be trained.

3. The magnetic positioning and velocity measurement method of synthesizing cross-axis effects of an artificial intelligence network model according to claim 2, characterized in that: Will d i ( Δ ix ,Δ iy ,Δ iz ) and the 3 components of the magnet towards u ( u x ,u y ,u z ) into the dipole model, the dipole model magnetic field vector H i ; where m 0 is the magnetic distance size constant of the magnet measured in advance; for the ith sensor, the mean square error of the corrected magnetic field H ix , H iy , H iz ) measured by the sensor and the dipole model magnetic field vector H i is taken as the loss function Loss: The artificial intelligence deep learning framework trains the network model based on the Pytorch framework float64 precision, and uses the Adam function as the optimization function to solve all the correction parameters H ikx ,H iky ,H ikz ), A i3x3 , Δ ix ,Δ iy ,Δ iz .

4. The magnetic positioning and velocity measurement method of synthesizing cross-axis effects of an artificial intelligence network model according to claim 3, characterized in that: During training, the correction parameters of all sensors are first made the same, the same correction parameters are trained using the data collected by all sensors, and the correction parameters obtained by training are regarded as the mean value of the correction parameters of all sensors; then each sensor takes the mean value as the initial value, and adjusts and trains according to the data collected by the sensor to obtain the correction parameters of each sensor.

5. The magnetic positioning and velocity measurement method of synthesizing cross-axis effects of an artificial intelligence network model according to claim 4, characterized in that: According to the formula: Given the position of the magnet ( x , y , z ) and orientation ( u x ,u y ,u z The initial guess value is substituted into the dipole formula to calculate the magnetic field value at the location of each sensor, and then compared with the magnetic field value obtained from the calibration of all sensors. H ix , H iy , H iz The difference of squares between the two is minimized, and the result is obtained by LM iteration. x , y , z )and( u x ,u y ,u z The final optimized value.

6. The magnetic positioning and velocity measurement method of synthesizing cross-axis effects of an artificial intelligence network model according to claim 5, characterized in that: The obtained position of the magnet ( x , y , z By numerically differentiating the magnet with respect to time t, the linear velocity of the magnet can be obtained. .

Images