Five-degree-of-freedom electromagnetic positioning system and method

Abstract

The invention discloses a five-degree-of-freedom electromagnetic positioning system and method, and the system comprises a transmitting unit which comprises a plurality of single-axis transmitting coils and is used for generating a magnetic field in a three-dimensional space; the receiving unit is used for collecting field intensity information in a three-dimensional space; the control and data processing unit is used for inputting the field intensity information collected by the receiving unit into a pose initial value prediction network and outputting initial pose information of the receiving unit; inputting the initial attitude information into a coil magnetic field modeling network to obtain theoretical field intensity; and the iterative optimization module acquires a field intensity difference according to the theoretical field intensity and the acquired field intensity information, updates the attitude information according to the field intensity difference, and returns to execute the step of acquiring the theoretical field intensity until a preset iteration termination condition is reached, thereby acquiring the final attitude information. The method can reduce the complexity of a traditional electromagnetic positioning algorithm, improves the positioning precision and speed, can be widely applied to the fields of auxiliary medical treatment, man-machine interaction and motion tracking, and achieves a multi-degree-of-freedom electromagnetic positioning function.

Description

Technical Field

[0001] The present invention relates to the fields of auxiliary medical treatment, human-computer interaction and motion tracking, and in particular to a five-degree-of-freedom electromagnetic positioning system and method. Background Art

[0002] In the fields of auxiliary medical treatment, human-computer interaction and motion tracking, the current mainstream positioning methods mainly include mechanical, acoustic, optical and electromagnetic, among which electromagnetic positioning technology is often used in the field of motion tracking. Compared with optical and mechanical methods, electromagnetic positioning has the advantage of not being obstructed by line of sight and is not easily affected by environmental factors such as temperature and humidity. Another advantage is that the positioning information is rich, that is, not only the spatial position but also the posture information can be obtained. Generally speaking, five-degree-of-freedom electromagnetic positioning refers to the use of electromagnetic signals to determine the three-dimensional coordinates and direction (without spin angle) of the target object. Its device generally includes a signal transmitting unit (such as a transmitting coil), a signal receiving unit (such as an induction coil), and a control and data processing unit.

[0003] Electromagnetic positioning algorithm refers to the method of calculating the position and direction of the target object in space by measuring and analyzing specific parameters in the electromagnetic field (such as the amplitude of the received signal, etc.). Most electromagnetic positioning algorithms need to rely on some kind of magnetic field model. The most widely used one is the magnetic dipole model, which abstracts the magnetic field source as a magnetic dipole to calculate the magnetic field distribution around it. The model calculation is simple and the positioning speed is fast, but the positioning accuracy of the model has a certain error compared with the real physics, especially in the near field. In addition, there is also a real mathematical model of the transmitting coil obtained through complex derivation, and the induced electromotive force of the receiving coil is expressed as a closed solution through numerical integration. The advantage of this method is that the magnetic field model of the transmitting coil is more accurate, and the disadvantage is that the numerical integration process takes a long time and the positioning speed is slow.

[0004] In addition, in the process of using a certain magnetic field model to calculate the posture, the iterative optimization algorithm performs iterative optimization through the gap between the theoretical value and the actual measured value, thereby approximating the true posture. The first step of the iterative optimization algorithm is to determine the initial posture (i.e., the initial value of the iteration). Currently, the global search algorithm (such as the PSO particle swarm algorithm) is mainly used to find the initial value, but this type of method has a slow search speed and may fall into a local optimum, making the above iterative optimization algorithm unable to converge, which means that the selection of the initial value of the iteration is particularly critical. Summary of the invention

[0005] In order to solve at least one of the technical problems existing in the prior art to a certain extent, the object of the present invention is to provide a five-degree-of-freedom electromagnetic positioning system and method based on neural network modeling.

[0006] The first technical solution adopted by the present invention is:

[0007] A five-degree-of-freedom electromagnetic positioning system, comprising:

[0008] A transmitting unit, comprising a plurality of single-axis transmitting coils of the same structure, for generating a magnetic field in a three-dimensional space;

[0009] A receiving unit, used for collecting field strength information in three-dimensional space;

[0010] The control and data processing unit includes a posture initial value prediction network, a coil magnetic field modeling network and an iterative optimization module; the field strength information collected by the receiving unit is input into the posture initial value prediction network, and the initial posture information of the receiving unit is output; the initial posture information is input into the coil magnetic field modeling network to obtain the theoretical field strength; the iterative optimization module obtains the field strength difference according to the theoretical field strength and the collected field strength information, updates the posture information according to the field strength difference, and returns to execute the step of obtaining the theoretical field strength until the preset iteration termination condition is reached to obtain the final posture information.

[0011] Furthermore, the structure of the coil magnetic field modeling network is a single-layer or multi-layer fully connected structure, the input is the coordinate information of the single-axis coil, and the output is the field intensity component;

[0012] The coil magnetic field modeling network is constructed in the following way:

[0013] A two-dimensional magnetic field model is established based on the fitting neural network, and then a three-dimensional magnetic field model is established through the spatial symmetry of the magnetic field; that is, the coil magnetic field modeling is used to realize the mapping between the two-dimensional space coordinates and the two-dimensional space magnetic field, and then the rotational symmetry of the three-dimensional magnetic field of the single-axis coil is used to map the two-dimensional space coordinates and the two-dimensional space magnetic field to the three-dimensional space.

[0014] Furthermore, mapping the two-dimensional space coordinates and the two-dimensional space magnetic field to the three-dimensional space includes:

[0015] Assume the two-dimensional space coordinates (y, z), the three-dimensional space coordinates (X, Y, Z), and the two-dimensional space magnetic field (B y ,B z ), three-dimensional magnetic field (B X ,B Y ,B Z );

[0016] From (X,Y,Z) we get (y,z): z=Z;

[0017] According to the neural network, (y,z) is obtained (B y ,B z ), further by (B y ,B z ) to get (B X ,B Y ,BZ ): B Z =B z .

[0018] Furthermore, the coil magnetic field modeling network is a single-plane coil magnetic field modeling fitting neural network, and its training data is located in any one-quarter quadrant of any plane (such as the yOz plane) where the center axis of the single-axis coil is located; the one-quarter quadrant magnetic field is mapped to the entire two-dimensional space by utilizing the plane symmetry of the magnetic field;

[0019] Since the rate of change of the magnetic field in the near field is large and the rate of change in the far field is small, the quarter quadrant is divided into several areas for training respectively. The sampling method of the training data is from dense to sparse from near to far, and several sub-neural networks are obtained. Then, the several sub-neural networks are combined into a single plane coil magnetic field modeling and fitting neural network.

[0020] Furthermore, in the yOz plane, the magnetic field is divided into only the component B along the y axis. y and the component B along the z-axis z , the magnitude of the magnetic field component depends only on the positions y and z. At the same time, B y The positive and negative symmetry is about the y-axis, and the positive and negative symmetry is about the z-axis. z The relationship is mirror symmetric about the y-axis and mirror symmetric about the z-axis. The formula is as follows:

[0021] B y (y,z)=-B y (y,-z)=-B y (-y,z)

[0022] B z (y,z)=B z (y,-z)=B z (-y,z)

[0023] That is, it is only necessary to model the regional magnetic field in the positive quarter quadrant to obtain the magnetic field of the entire plane.

[0024] Furthermore, the structure of the posture initial value prediction network is a single-layer or multi-layer fully connected structure, and the input is the field strength B (B1, B1, ... B n ), the output is the coordinate information P (x, y, z) corresponding to the receiving unit;

[0025] The initial pose prediction network is constructed in the following way:

[0026] A mapping relationship between the field strength information of the receiving unit and the five-degree-of-freedom posture of the transmitting unit is established based on the fitting neural network.

[0027] Furthermore, the training data of the posture initial value prediction network consists of multiple coordinate information in a fixed direction (such as vertically downward) and a fixed area, and the actual field strength of multiple transmitting coils corresponding to each coordinate information.

[0028] Furthermore, the coil on the receiving unit is a single-axis coil or a dual-axis coil;

[0029] When the coil on the receiving unit is a dual-axis coil, six-degree-of-freedom electromagnetic positioning can be achieved.

[0030] Furthermore, the number of the single-axis transmitting coils is greater than or equal to 5;

[0031] When the number of single-axis transmitting coils is 9, the 9 single-axis transmitting coils are distributed in a 3×3 array, and the 3 coils in each row or column are arranged in an orthogonal manner;

[0032] When the number of the single-axis transmitting coils is 8, two adjacent single-axis transmitting coils among the 8 single-axis transmitting coils are arranged in an orthogonal manner.

[0033] The second technical solution adopted by the present invention is:

[0034] A five-degree-of-freedom electromagnetic positioning method comprises the following steps:

[0035] Collect field strength information in three-dimensional space through a receiving unit;

[0036] The field strength information collected by the receiving unit is input into the initial posture prediction network to obtain the initial posture information of the receiving unit;

[0037] Input the initial posture information into the coil magnetic field modeling network to obtain the theoretical field strength;

[0038] The iterative optimization algorithm obtains the field strength difference based on the theoretical field strength and the collected field strength information, and updates the posture information based on the field strength difference;

[0039] The updated posture information is re-input into the coil magnetic field modeling network to obtain a new theoretical field strength until the preset iteration termination condition is reached to obtain the final posture information.

[0040] The present invention has the following beneficial effects:

[0041] (1) The present invention realizes the rapid and accurate modeling of magnetic field and the rapid and accurate prediction of initial value. By combining two neural networks and iterative optimization algorithm, a complete five-degree-of-freedom electromagnetic positioning algorithm is formed, which takes into account both positioning accuracy and speed.

[0042] (2) The present invention utilizes the rotational symmetry of the magnetic field of a single-axis coil to simplify the three-dimensional magnetic field into a planar magnetic field, thereby using a single-plane magnetic field modeling neural network to model the magnetic field in three-dimensional space. This not only reduces the computational complexity, increases the modeling speed, but also improves the positioning accuracy.

[0043] (3) The present invention uses field strength information and posture information to train a neural network to predict the initial value, which is faster and more accurate than the traditional global search algorithm and avoids falling into the local optimum.

[0044] (4) The electromagnetic positioning algorithm described in the present invention has strong scalability. Six-degree-of-freedom electromagnetic positioning can be achieved by simply replacing the single-axis coil of the receiving unit with a dual-axis coil. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the embodiments of the present invention or the drawings of related technical solutions in the prior art are introduced below. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 It is a basic flow chart of the five-degree-of-freedom electromagnetic positioning algorithm of the present invention;

[0047] Figure 2 is a yOz plane magnetic field vector distribution diagram of the transmitting coil of Example 1 of the present invention;

[0048] Figure 3 is a sampling point diagram within a positive quarter quadrant of Embodiment 1 of the present invention;

[0049] Figure 4 is a structural diagram of a neural network for single-plane magnetic field modeling according to Embodiment 1 of the present invention;

[0050] FIG5 is a diagram showing the effect of a single-plane magnetic field modeling neural network training according to Example 1 of the present invention (divided into three regions);

[0051] Figure 6 is a flow chart of a three-dimensional magnetic field modeling algorithm according to Embodiment 1 of the present invention;

[0052] Figure 7 is a configuration diagram of a transmitting unit according to Embodiment 1 of the present invention;

[0053] Figure 8 1 is a structural diagram of an iterative initial value prediction neural network according to Embodiment 1 of the present invention;

[0054] Fig. 9This is a diagram showing the training effect of an iterative initial value prediction neural network according to Embodiment 1 of the present invention;

[0055] Fig.10 This is a diagram showing the prediction effect of the initial value prediction neural network in 100 positions and postures according to Example 1 of the present invention;

[0056] Fig.11 is a flow chart of the LM algorithm based on neural network magnetic field modeling in Example 1 of the present invention;

[0057] Fig.12 This is a diagram of the continuous positioning effect of the LM algorithm based on neural network magnetic field modeling in Example 1 of the present invention in 400 positions;

[0058] Fig.13 is a configuration diagram of a transmitting unit according to Embodiment 2 of the present invention;

[0059] Fig.14 This is a diagram showing the training effect of an iterative initial value prediction neural network according to Embodiment 2 of the present invention;

[0060] Fig.15 This is a diagram showing the prediction effect of the initial value prediction neural network in 100 positions according to Example 2 of the present invention;

[0061] Fig.16 This is a diagram of the continuous positioning effect of the LM algorithm based on neural network magnetic field modeling in Example 2 of the present invention in 400 positions;

[0062] Fig.17 is a structural diagram of a single-plane magnetic field modeling neural network according to Embodiment 3 of the present invention;

[0063] FIG18 is a diagram showing the effect of the single-plane magnetic field modeling neural network training according to Example 3 of the present invention (divided into three regions);

[0064] Fig.19 4 is a structural diagram of an iterative initial value prediction neural network according to Embodiment 4 of the present invention;

[0065] Fig. 20 This is a diagram showing the training effect of an iterative initial value prediction neural network according to Embodiment 5 of the present invention;

[0066] Fig.21 It is a structural schematic diagram of the five-degree-of-freedom electromagnetic positioning system of the present invention. DETAILED DESCRIPTION

[0067] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limitations of the present invention. For the step numbers in the following embodiments, they are only provided for the convenience of explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0068] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., and orientations or positional relationships indicated are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0069] In the description of the present invention, "several" means one or more, "more" means more than two, "greater than", "less than", "exceed" etc. are understood as not including the number itself, and "above", "below", "within" etc. are understood as including the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing the technical features, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0070] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0071] like Figure 1 and Fig.21 As shown, the present invention provides a five-degree-of-freedom electromagnetic positioning system, comprising:

[0072] A transmitting unit, comprising a plurality of single-axis transmitting coils of the same structure, for generating a magnetic field in a three-dimensional space;

[0073] A receiving unit, used for collecting field strength information in three-dimensional space;

[0074] The control and data processing unit includes a posture initial value prediction network, a coil magnetic field modeling network and an iterative optimization module; the field strength information collected by the receiving unit is input into the posture initial value prediction network, and the initial posture information of the receiving unit is output; the initial posture information is input into the coil magnetic field modeling network to obtain the theoretical field strength; the iterative optimization module obtains the field strength difference according to the theoretical field strength and the collected field strength information, updates the posture information according to the field strength difference, and returns to execute the step of obtaining the theoretical field strength until the preset iteration termination condition is reached to obtain the final posture information.

[0075] The control and data processing unit in the system of this embodiment adopts two different neural networks (i.e., coil magnetic field modeling network and posture initial value prediction network), which correspond to the coil magnetic field modeling method based on neural network and the posture initial value prediction method based on neural network respectively. When the system is actually positioned, the first batch of field strength information collected by the sensor (i.e., receiving unit) is transmitted to the initial value prediction method, which predicts an initial value of posture and provides it to the iterative optimization algorithm (such as LM algorithm, gradient descent method, genetic algorithm, etc.); during each iteration, the optimization algorithm transmits the posture information to the coil magnetic field modeling method, which outputs the theoretical field strength corresponding to the posture, and the optimization algorithm analyzes the field strength difference between the theoretical field strength and the measured field strength, which is a vector (such as a gradient vector), and updates the posture information according to the field strength difference to gradually approach the real posture, thereby realizing the optimization of the posture.

[0076] As an optional implementation, the coil magnetic field modeling method uses a fitting neural network to help establish a two-dimensional magnetic field model, and then establishes a three-dimensional magnetic field model through the spatial symmetry of the magnetic field. The neural network is a single-plane coil magnetic field modeling fitting neural network, and its training data is located in any one of the planes (such as the yOz plane) where the axis of the single-axis coil is located. The one-quarter quadrant magnetic field can be mapped to the entire two-dimensional space using the plane symmetry of the magnetic field. Since the rate of change of the magnetic field in the near field is large and the rate of change in the far field is small, the one-quarter quadrant is divided into several areas for training respectively, and the sampling method of the training data is from near to far from dense to sparse, and several sub-neural networks are obtained, which are combined into a single-plane coil magnetic field modeling fitting neural network. The single-plane coil magnetic field modeling fitting neural network can realize the mapping between two-dimensional space coordinates and two-dimensional space magnetic field. Using the rotational symmetry of the three-dimensional magnetic field of the single-axis coil, this method maps the two-dimensional space coordinates and the two-dimensional space magnetic field to three-dimensional space.

[0077] As an optional implementation, the pose initial value prediction method uses a fitting neural network to help establish a mapping relationship between the field strength received by the induction coil and the five-degree-of-freedom pose of the induction coil. The core of the pose initial value prediction method lies in the neural network, which establishes a mapping relationship between the received field strength and coordinate information. The coordinate information and the posture information in a fixed direction are combined to realize the prediction of the pose initial value. The training data of the pose initial value prediction fitting neural network consists of multiple coordinate information in a fixed direction (such as vertically downward) and a fixed area, and the actual field strength of multiple transmitting coils corresponding to each coordinate information.

[0078] It should be noted that the iterative optimization algorithm in the present invention does not specify a certain type, and all optimization algorithms that can minimize the error between the target pose and the predicted pose are applicable to the present invention.

[0079] The present invention will be further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0080] Example 1

[0081] (1) Coil magnetic field modeling method

[0082] For a cylindrical transmitting coil, the magnitude of the field strength vector B around it is related to the number of turns n of the transmitting coil, the coil current I, the inner and outer diameters R of the coil, the coil height L, and the distance r from the field point (x0, y0, z0) to the center of the coil. It can be calculated through complex numerical integration, as shown in the following formula:

[0083]

[0084] In the formula, B x , B y , B z They correspond to the components of the field strength vector on the X, Y, and Z axes respectively, μ0 is the magnetic permeability of vacuum, R-in represents the inner diameter of the coil, R-out represents the outer diameter of the coil, and θ is the angle between the projection of the line from the field point to the center of the coil on the xOy plane and the x-axis.

[0085] Furthermore, considering the rotational symmetry of the cylindrical coil, the field strength it generates is also symmetrical, that is, the magnetic field in any plane where the central axis of the coil is located is exactly the same. Based on this, our solution is to establish a magnetic field model in the yOz plane to obtain a magnetic field model in the entire space. Figure 2 As shown, in the yOz plane, the magnetic field can only be divided into the component B along the y axis y and the component B along the z-axis z , the magnitude of the magnetic field component is only related to y and z, and at the same time, B y The positive and negative symmetry is about the y-axis, and the positive and negative symmetry is about the z-axis. zThe relationship is mirror symmetric about the y-axis and mirror symmetric about the z-axis, as shown in the following formula:

[0086] B y (y,z)=-B y (y,-z)=-B y (-y,z)

[0087] B z (y,z)=B z (y,-z)=B z (-y,z)

[0088] Therefore, we only need to model the regional magnetic field in the positive quarter quadrant to obtain the magnetic field of the entire plane.

[0089] like Figure 3 As shown, in this example, considering that the magnetic field has a larger rate of change in the near field and a smaller rate of change in the far field, a sampling method from near to far and from dense to sparse is adopted. Within the positive quarter quadrant, sampling and training are performed in three areas.

[0090] like Figure 4 As shown, the single plane magnetic field modeling neural network described in this example has a training input of normalized coordinates (y, z) and a training output of normalized numerical integrated field intensity (B y ,B z ), a single-plane magnetic field modeling fully connected neural network was established, the structure included 1 hidden layer, the number of neurons n = 200, the hidden layer transfer function was sigmoid function, and the output layer transfer function was linear function (purelin).

[0091] As shown in Figure 5, the training effect of the single-plane magnetic field modeling neural network described in this example is that the mean square error (MSE) reaches 10 -9 The correlation coefficient R between the predicted output and the target output reaches 1. Among them, Figure 5(a) shows the training effect of area 1, Figure 5(b) shows the training effect of area 2, and Figure 5(c) shows the training effect of area 3.

[0092] like Figure 6 As shown, the single-plane magnetic field modeling neural network described in this example can map the magnetic field vector of a single plane to the magnetic field vector of a three-dimensional space through the rotation transformation and spatial relationship of the rectangular coordinate system, thereby realizing magnetic field modeling of a complete space, as shown below:

[0093] From (X,Y,Z) we get (y,z):

[0094] According to the neural network, (y,z) is obtained (B y ,B z ),

[0095] Further by (B y ,B z ) to get (B X ,B Y ,B Z ): B Z =B z .

[0096] Among them, (y, z) is the two-dimensional space coordinate, (X, Y, Z) is the three-dimensional space coordinate, (B y ,B z ) is the two-dimensional magnetic field, (B X ,B Y ,B Z ) is the three-dimensional magnetic field.

[0097] (2) Initial pose prediction method

[0098] like Figure 7 As shown, in this example, regarding the configuration of the transmitting unit, we use an array distribution of 9 (≥5) single-axis transmitting coils. Each transmitting coil has the same specifications, and the three coils in each row / column are arranged in an orthogonal manner, which can ensure that a three-axis mutually orthogonal magnetic field is generated.

[0099] After the receiving unit obtains a set of field strength information (B), it needs to predict a posture P0 through B as the initial value of the optimization algorithm. Therefore, the core of the posture initial value prediction method adopted in this embodiment is to establish a neural network that can predict posture. During training, due to multiple transmitting coils and multiple posture directions, the data set will be very large.

[0100] like Figure 8 As shown, the initial pose prediction neural network of this example is based on Figure 7 The transmitter unit configuration shown in the figure selects vertical downward (0,0,-1) as a fixed direction and generates multiple coordinate points P(x,y,z) and corresponding field strengths B(B1,B2,…B9) as training data. With field strengths B(B1,B2,…B9) as training inputs and three-dimensional coordinates P(x,y,z) as training outputs, a fitting neural network for pose prediction is established, which contains two hidden layers, with n=50 neurons in each layer. The transfer function of the hidden layer is a hyperbolic tangent function (tan-sigmoid), and the transfer function of the output layer is a linear function (purelin).

[0101] like Fig. 9 As shown in the figure, the training effect of the initial pose prediction neural network described in this example is as high as 10 -2 The correlation coefficient R between the predicted output and the target output reaches 1.

[0102] like Fig.10 As shown, in this example, the performance of the initial value prediction neural network is tested. 100 postures are preset in the training space (all directions are vertically downward), and the numerical integrated field strength is used as the measured signal of the induction coil (under ideal conditions, measurement errors are not considered). The initial value prediction neural network is input for prediction. The results show that the cumulative time for 100 postures is about 0.5s, the average prediction time for each posture is 0.005s, and the average absolute value errors of the x, y, and z coordinates are approximately 0.14mm, 0.17mm, and 0.19mm.

[0103] like Fig.11 As shown, in this example, the LM algorithm (Levenberg-Marquardt) is selected as the optimization algorithm, the initial value prediction neural network and the magnetic field modeling neural network are assembled, and the performance of the complete 5-DOF electromagnetic positioning algorithm is tested based on the transmitting unit configuration described in this example.

[0104] like Fig.12 As shown in the figure, a smooth motion trajectory of the induction coil containing 400 postures is preset in a 1000mm side length cube space (both position and direction change), and the numerical integral field strength is used as the measured signal of the induction coil (under ideal conditions, measurement error is not considered). The results show that the cumulative time of 400 postures is about 16s (including initial value prediction and continuous positioning), the average prediction time of each posture is 0.04s, the positioning speed reaches 25Hz, the average absolute value error of the x, y, and z coordinates is about 0.02mm, 0.05mm, and 0.01mm, and the average angle error is 0.01 degrees after the direction vector is converted into an angle. The results show that the five-degree-of-freedom electromagnetic positioning algorithm described in this example achieves high-precision and high-speed continuous positioning.

[0105] Example 2

[0106] The difference between this embodiment and embodiment 1 lies in the configuration of the transmitting unit, in order to verify the universality of the algorithm of the present invention under different transmitting unit configurations.

[0107] like Figure 4 As shown in Example 1, the single-plane magnetic field modeling neural network described in this example has a training input of the normalized coordinates (y, z) and a training output of the normalized numerical integral field intensity (B y ,B z ), a single-plane magnetic field modeling fully connected neural network was established, the structure included 1 hidden layer, the number of neurons n = 200, the hidden layer transfer function was sigmoid function, and the output layer transfer function was linear function (purelin).

[0108] like Figure 6As shown, similar to Example 1, the single-plane magnetic field modeling neural network described in this example can map the magnetic field vector of a single plane to the magnetic field vector of a three-dimensional space through the rotation transformation and spatial relationship of the rectangular coordinate system, thereby realizing magnetic field modeling of the complete space.

[0109] like Fig.13 As shown, in this example, regarding the configuration of the transmitting unit, we use an array of 8 (≥5) single-axis transmitting coils. Each transmitting coil has the same specifications, and two adjacent coils are placed in an orthogonal manner to ensure that a three-axis mutually orthogonal magnetic field is generated.

[0110] like Fig.14 As shown, the initial pose prediction neural network described in this example is based on Fig.13 The transmitter unit configuration shown in the figure selects vertical downward (0,0,-1) as a fixed direction, and generates multiple coordinate points P (x,y,z) and corresponding field strengths B (B1,B2,…B8) as training data. With field strengths B (B1,B2,…B8) as training inputs and three-dimensional coordinates P (x,y,z) as training outputs, a fitting neural network for posture prediction is established, which includes two hidden layers, with n=50 neurons in each layer. The transfer function of the hidden layer is a hyperbolic tangent function (tan-sigmoid), and the transfer function of the output layer is a linear function (purelin). The training effect of the initial posture prediction neural network described in this example has a mean square error (mse) of 10 -2 The correlation coefficient R between the predicted output and the target output reaches 1.

[0111] like Fig.15 As shown, in this example, the performance of the initial value prediction neural network is tested. 100 postures are preset in the training space (all directions are vertically downward), and the numerical integrated field strength is used as the measured signal of the induction coil (under ideal conditions, measurement errors are not considered). The initial value prediction neural network is input for prediction. The results show that the cumulative time for 100 postures is about 0.5s, the average prediction time for each posture is 0.005s, and the average absolute value errors of the x, y, and z coordinates are approximately 0.2mm, 0.5mm, and 0.1mm.

[0112] like Fig.11 As shown, similar to Example 1, in this example, the LM algorithm (Levenberg-Marquardt) is selected as the optimization algorithm, the initial value prediction neural network and the magnetic field modeling neural network are assembled, and the performance of the complete 5-DOF electromagnetic positioning algorithm is tested based on the transmitting unit configuration described in this example.

[0113] like Fig.16As shown in the figure, a smooth motion trajectory of the induction coil containing 400 postures is preset in a 1000mm side length cube space (both position and direction change), and the numerical integral field strength is used as the measured signal of the induction coil (under ideal conditions, measurement error is not considered). The cumulative time of 400 postures is about 16s (including initial value prediction and continuous positioning), the average prediction time of each posture is 0.04s, the positioning speed reaches 25Hz, the average absolute value error of the x, y, and z coordinates is about 0.02mm, 0.02mm, and 0.01mm, and the average angle error is 0.01 degrees after the direction vector is converted into an angle. The results show that the five-degree-of-freedom electromagnetic positioning algorithm described in this example achieves high-precision and high-speed continuous positioning.

[0114] By comparing Example 1 and Example 2, it can be verified that the algorithm described in the present invention is universal under different transmitting unit configurations. As long as the transmitting unit configuration satisfies the distribution of multiple (≥5) identical single-axis transmitting coil arrays and can generate mutually orthogonal three-axis magnetic fields, a complete five-degree-of-freedom electromagnetic positioning algorithm can be implemented.

[0115] Example 3

[0116] The difference between this example and Example 1 lies in the structure of the single-plane magnetic field modeling neural network, in order to verify the versatility of the coil magnetic field modeling method based on neural network described in the present invention under different neural network structures.

[0117] like Fig.17 As shown, similar to Example 1, in this example, considering that the magnetic field has a larger rate of change in the near field and a smaller rate of change in the far field, a sampling method from near to far and from dense to sparse is adopted. In the positive quarter quadrant, three regions are respectively sampled and trained for single-plane magnetic field modeling neural network. The training input is the normalized coordinate (y, z), and the training output is the normalized numerical integral field intensity (B y ,B z ), a single-plane magnetic field modeling fully connected neural network was established. The difference is that the network structure contains 2 hidden layers, each layer has n=40 neurons, the hidden layer transfer function is the sigmoid function, and the output layer transfer function is a linear function (purelin).

[0118] As shown in Figure 18, the training effect of the single-plane magnetic field modeling neural network described in this example is that the mean square error (MSE) reaches 10 -12 The correlation coefficient R between the predicted output and the target output reaches 1. Among them, Figure 18(a) shows the training effect of area 1, Figure 18(b) shows the training effect of area 2, and Figure 18(c) shows the training effect of area 3.

[0119] By comparing Example 1 and Example 3, the versatility of the coil magnetic field modeling method based on neural network in different neural network structures can be verified.

[0120] Example 4

[0121] The difference between this example and Example 1 lies in the structure of the initial posture prediction neural network, in order to verify the universality of the neural network-based posture initial value prediction method described in the present invention under different neural network structures.

[0122] like Fig.19 As shown, the initial posture prediction neural network described in this example is similar to that in Example 1, both based on Figure 7 The transmitter unit configuration shown in the figure selects vertical downward (0,0,-1) as a fixed direction and generates multiple coordinate points P(x,y,z) and corresponding field strengths B(B1,B2,…B9) as training data. With field strengths B(B1,B2,…B9) as training inputs and three-dimensional coordinates P(x,y,z) as training outputs, a fitting neural network for pose prediction is established. The difference is that the network structure contains three hidden layers, with n=30 neurons in each layer, the hidden layer transfer function is a sigmoid function, and the output layer transfer function is a linear function (purelin).

[0123] like Fig. 20 As shown in the figure, the training effect of the initial pose prediction neural network described in this example is as high as 10 -7 The correlation coefficient R between the predicted output and the target output reaches 1.

[0124] By comparing Example 1 and Example 4, the versatility of the neural network-based initial pose prediction method of the present invention under different neural network structures can be verified.

[0125] In summary, the present invention has at least the following advantages and beneficial effects compared to the prior art:

[0126] (1) The present invention can quickly and accurately establish a magnetic field model and quickly and accurately predict the initial value. By combining two neural networks and an iterative optimization algorithm, a complete five-degree-of-freedom electromagnetic positioning system is formed, which takes into account both positioning accuracy and speed.

[0127] (2) The present invention uses the rotational symmetry of the single-axis coil magnetic field to simplify the three-dimensional magnetic field into a two-dimensional planar magnetic field, thereby using a planar magnetic field neural network to model the three-dimensional magnetic field. Compared with traditional numerical integration, the computational complexity is reduced and the modeling speed is improved; compared with the traditional magnetic dipole model, the positioning accuracy is improved.

[0128] (3) The present invention trains a neural network for predicting initial values ​​through field strength information and posture information. It is faster and more accurate than the traditional global search algorithm and can avoid falling into the dilemma of local optimality.

[0129] (4) The electromagnetic positioning algorithm described in the present invention has strong scalability. It only needs to replace the single-axis coil of the receiving unit with a dual-axis coil to achieve 6-DOF electromagnetic positioning.

[0130] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

[0131] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the described embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A five-degree-of-freedom electromagnetic positioning system, characterized in that: include: A transmitting unit, comprising a plurality of single-axis transmitting coils of the same structure, for generating a magnetic field in a three-dimensional space; A receiving unit, used for collecting field strength information in three-dimensional space; Control and data processing unit, including posture initial value prediction network, coil magnetic field modeling network and iterative optimization module; Inputting the field strength information collected by the receiving unit into the initial posture prediction network, and outputting the initial posture information of the receiving unit; Inputting the initial posture information into the coil magnetic field modeling network to obtain the theoretical field strength; The iterative optimization module obtains the field strength difference based on the theoretical field strength and the collected field strength information, updates the posture information based on the field strength difference, and returns to execute the step of obtaining the theoretical field strength until the preset iteration termination condition is reached to obtain the final posture information.

2. A five-degree-of-freedom electromagnetic positioning system according to claim 1, characterized in that: The structure of the coil magnetic field modeling network is a single-layer or multi-layer fully connected structure, the input is the coordinate information of the single-axis coil, and the output is the field intensity component; The coil magnetic field modeling network is constructed in the following way: A two-dimensional magnetic field model is established based on the fitting neural network, and then a three-dimensional magnetic field model is established through the spatial symmetry of the magnetic field; that is, the coil magnetic field modeling is used to realize the mapping between the two-dimensional space coordinates and the two-dimensional space magnetic field, and then the rotational symmetry of the three-dimensional magnetic field of the single-axis coil is used to map the two-dimensional space coordinates and the two-dimensional space magnetic field to the three-dimensional space.

3. A five-degree-of-freedom electromagnetic positioning system according to claim 2, characterized in that: Mapping the two-dimensional space coordinates and the two-dimensional space magnetic field to the three-dimensional space includes: Assume the two-dimensional space coordinates (y, z), the three-dimensional space coordinates (X, Y, Z), and the two-dimensional space magnetic field (B y ,B z ), three-dimensional magnetic field (B X ,B Y ,B Z ); From (X,Y,Z) we get (y,z): According to the neural network, (y,z) is obtained (B y ,B z ), further by (B y ,B z ) to get (B X ,B Y ,B Z ): B Z =B z 。 4. A five-degree-of-freedom electromagnetic positioning system according to claim 1 or 2, characterized in that: The coil magnetic field modeling network is a single-plane coil magnetic field modeling fitting neural network, and its training data is located in any one-quarter quadrant of any plane where the central axis of the single-axis coil is located; the one-quarter quadrant magnetic field is mapped to the entire two-dimensional space by utilizing the plane symmetry of the magnetic field; Since the rate of change of the magnetic field in the near field is large and the rate of change in the far field is small, the quarter quadrant is divided into several areas for training respectively. The sampling method of the training data is from dense to sparse from near to far, and several sub-neural networks are obtained. Then, the several sub-neural networks are combined into a single plane coil magnetic field modeling and fitting neural network.

5. A five-degree-of-freedom electromagnetic positioning system according to claim 4, characterized in that: In the yOz plane, the magnetic field is divided into only the component B along the y axis y and the component B along the z-axis z , the magnitude of the magnetic field component depends only on the positions y and z. At the same time, B y The positive and negative symmetry is about the y-axis, and the positive and negative symmetry is about the z-axis. z The relationship is mirror symmetric about the y-axis and mirror symmetric about the z-axis. The formula is as follows: B y (y,z)=-B y (y,-z)=-B y (-y,z) B z (y,z)=B z (y,-z)=B z (-y,z) That is, it is only necessary to model the regional magnetic field in the positive quarter quadrant to obtain the magnetic field of the entire plane.

6. The five-degree-of-freedom electromagnetic positioning system according to claim 1, characterized in that: The structure of the initial value prediction network of the posture is a single-layer or multi-layer fully connected structure, and the input is the field strength B (B1, B1, ... B1) obtained by the receiving unit about multiple transmitting coils n ), the output is the coordinate information P (x, y, z) corresponding to the receiving unit; The initial pose prediction network is constructed in the following way: A mapping relationship between the field strength information of the receiving unit and the five-degree-of-freedom posture of the transmitting unit is established based on the fitting neural network.

7. A five-degree-of-freedom electromagnetic positioning system according to claim 1 or 6, characterized in that: The training data of the posture initial value prediction network consists of multiple coordinate information in a fixed direction and a fixed area, and the field strength of multiple transmitting coils corresponding to each coordinate information.

8. The five-degree-of-freedom electromagnetic positioning system according to claim 1, characterized in that: The coil on the receiving unit is a single-axis coil or a double-axis coil; When the coil on the receiving unit is a dual-axis coil, six-degree-of-freedom electromagnetic positioning can be achieved.

9. The five-degree-of-freedom electromagnetic positioning system according to claim 1, characterized in that: The number of the single-axis transmitting coils is greater than or equal to 5; When the number of single-axis transmitting coils is 9, the 9 single-axis transmitting coils are distributed in a 3×3 array, and the 3 coils in each row or column are arranged in an orthogonal manner; When the number of the single-axis transmitting coils is 8, two adjacent single-axis transmitting coils among the 8 single-axis transmitting coils are arranged in an orthogonal manner.

10. A five-degree-of-freedom electromagnetic positioning method, characterized in that: The following steps are involved: Collect field strength information in three-dimensional space through a receiving unit; The field strength information collected by the receiving unit is input into the initial posture prediction network to obtain the initial posture information of the receiving unit; Input the initial posture information into the coil magnetic field modeling network to obtain the theoretical field strength; The iterative optimization algorithm obtains the field strength difference based on the theoretical field strength and the collected field strength information, and updates the posture information based on the field strength difference; The updated posture information is re-input into the coil magnetic field modeling network to obtain a new theoretical field strength until the preset iteration termination condition is reached to obtain the final posture information.

Images