Large-space magnetic control capsule robot positioning system and method based on CRLocNet

Abstract

The invention discloses a CRLocNet-based large-space magnetic control capsule robot positioning system and method. The CRLocNet-based large-space magnetic control capsule robot positioning system comprises an external driving module, a capsule robot, a mobile sensor array module, a neural network-based signal processing module and a control module. The problems that a traditional positioning method is limited in positioning space and large in edge position positioning error are solved, the precision of the CRLocNet-based positioning method is not affected by an initial value, the error is not accumulated along with time, and the data correction sub-network is added to effectively correct the error between measurement data and a mathematical model. Physical information is introduced into a solution network to enhance interpretability and stability of a position solution sub-network, network weight and threshold are optimized through an improved northern eagle optimization algorithm, and high-precision large-space positioning of the magnetic control capsule robot is achieved.

Description

Technical Field

[0001] The present invention relates to a capsule robot positioning system, specifically a large-space magnetically controlled capsule robot positioning system and method based on CRLocNet, belonging to the technical field of ingestible micro-robot positioning. Background Art

[0002] With the development of robot technology in the field of biomedicine, more and more medical robots are used for disease diagnosis and treatment. A capsule robot is a swallowable micro-robot with a camera installed on its head, which can capture gastrointestinal tract images and transmit them to an external display device for physicians to perform gastrointestinal examinations and treatments. Compared with traditional wired gastroscopes, it has the advantages of being painless and non-invasive. With the innovation and development of related technologies, the driving and control technologies of capsule robots have become increasingly mature, and the magnetic field-based driving method has become the mainstream method in this research field.

[0003] Positioning, as one of the important functions of capsule robots, is the basis for key technologies such as closed-loop control and drug delivery treatment. In the prior art, such as a capsule robot positioning method disclosed in CN114947692A, this positioning method first establishes a coordinate system; then respectively establishes a magnetic dipole model and an integral model of a cylindrical permanent magnet; then obtains the data of the magnetic sensor array, uses the magnetic dipole model and the integral model to establish positioning equations respectively, and respectively defines the objective functions to be optimized; then uses the particle swarm optimization algorithm and the LM algorithm to optimize the objective functions of the magnetic dipole model and the integral model respectively to obtain the positioning result of the permanent magnet; finally, the three-dimensional pose of the capsule robot is more intuitively displayed in the upper computer. By combining the particle swarm optimization algorithm and the LM algorithm, the situation of solution failure caused by inappropriate initial values of the LM algorithm is avoided; the problem of large errors of the magnetic dipole model at close range is solved. The existing positioning technology places the permanent magnet inside the capsule robot, detects the spatial magnetic field through an external magnetic sensor, and obtains the real-time pose of the capsule robot, improving the efficiency and accuracy of the examination. Due to its special advantages of being unaffected by human tissues and harmless to the human body, the magnetic field-based positioning method has been widely used in the field of medical devices. However, due to the small magnetic field intensity generated by the permanent magnet inside the capsule, the space for accurate positioning by the magnetic sensor is limited, which easily leads to an increase in positioning errors in the edge area; the traditional solution method for magnetic positioning problems is based on the LM algorithm. As a common method for solving non-linear problems, the LM algorithm has been widely used in magnetic positioning. However, this method has problems that the positioning accuracy is greatly affected by the initial value and the positioning error accumulates over time. Therefore, there is still a need to design a large-space and high-precision magnetically controlled capsule robot positioning scheme for the above problems. Summary of the Invention

[0004] The object of the present invention is to provide a large-space magnetically controlled capsule robot positioning system and method based on CRLocNet to solve at least one of the above technical problems. By using a mobile sensor array, the effective positioning space of the capsule robot is increased, and a new method for precise positioning is provided for magnetic positioning by building a CRLocNet neural network.

[0005] The present invention realizes the above object through the following technical solutions: A large-space magnetically controlled capsule robot positioning system based on CRLocNet includes an external driving module, a capsule robot, a mobile sensor array module, a neural network-based signal processing module and a control module. The external driving module includes a robotic arm and an external driving permanent magnet, and the external driving permanent magnet is connected to the driving end of the robotic arm. The mobile sensor array module includes a magnetic sensor array and an orthogonal slide, and the magnetic sensor array is fixed on the mounting platform of the orthogonal slide;

[0006] The control module includes a microprocessor, a robotic arm control system and a mobile sensor array control system. The robotic arm control system receives the real-time position information of the capsule robot output by the neural network-based signal processing module; the magnetic sensor array moves synchronously with the external driving permanent magnet, and the external driving permanent magnet is located directly above the motion space of the capsule robot. The neural network-based signal processing module is embedded with a magnetic positioning model and a capsule robot positioning neural network CRLocNet including a data correction sub-network and a position solution operator network.

[0007] As a further solution of the present invention: the robotic arm is a six-degree-of-freedom robotic arm, the external driving permanent magnet has a cylindrical shape, and the capsule robot contains a small cylindrical permanent magnet inside;

[0008] The magnetic sensor array is arranged in a square matrix by n×n magnetic sensors. The orthogonal slide includes two mutually orthogonal linear modules, which are respectively driven by stepping motors. The magnetic sensor array moves in the x-y plane 100 mm below the capsule robot, and the moving range of the magnetic sensor array is larger than the x-y plane motion range of the capsule robot.

[0009] A large-space magnetically controlled capsule robot positioning method based on CRLocNet includes a robot positioning system. The positioning method includes the following steps:

[0010] Step 1: Collect data and establish a network training data set;

[0011] Step 2: Establish a data correction sub-network, and use the model data with added random noise for pre-training to realize the preliminary processing of the collected data;

[0012] Step 3: Establish a position solution operator network, and use the model-generated data for pre-training to realize its function of solving the magnetic dipole inverse model;

[0013] Step 4: Based on the collected data, establish the CRLocNet model and train it to achieve the function of calculating the position of the capsule robot;

[0014] Step 5: Improve the Northern Goshawk optimization algorithm to optimize the weights and thresholds of CRLocNet;

[0015] Step 6: Import the trained model into the host computer, collect magnetic field information in real time and calculate the pose of the capsule robot.

[0016] As a further solution of the present invention: In step 1, the established data set includes:

[0017] In the motion space of the capsule robot, a large number of sampling point information including spatial positions and attitude angles are randomly generated, where 20% is used as the measurement site set D m , and the rest is used as the simulation site D e ;

[0018] The external drive module drives the capsule robot to move to the measurement sites in sequence and collects the magnetic field intensity information through the sensor array;

[0019] The simulation software generates the magnetic field intensity data of all sampling points based on the magnetic dipole model at each sensor position as the data set of CRLocNet;

[0020] The generation of the simulation site measurement data is based on the magnetic dipole model:

[0021]

[0022] In the formula, P = (a, b, c) is the central position of the magnet inside the capsule robot, P s = (x i , y i , z i ) is an arbitrary sampling point in space, where i = 1, 2, N, B = (B x , B y , B z ) is the magnetic field intensity at this point, and the vector of the spatial distance and direction between this point and the permanent magnet can be expressed as P l = P s - P, R l is the modulus of P l ; μ0 is the magnetic permeability of vacuum, μ r is the relative magnetic permeability of air, M T is the dipole moment intensity of the magnet, and H0 = (m, n, p) is the magnetization direction of the permanent magnet at P.

[0023] As a further solution of the present invention: In step two, the data correction sub-network is a fully connected CNN network with c hidden layers, which is used to correct the real measurement data and remove the noise and errors generated by the real measurement data for various reasons. The number of nodes in the initial hidden layer is determined by the following formula:

[0024]

[0025] In the formula, N s is the number of samples in the training set, N i is the number of neurons in the input layer, N o is the number of neurons in the output layer, C a is an independent variable that can take any value. Usually, C a ∈(2, 10);

[0026] The pre-training data set is the simulated site D e with magnetic field intensity data generated based on the magnetic dipole model. The added noise conforms to the Gaussian distribution:

[0027]

[0028] As a further solution of the present invention: In step three, the position resolver network adopts the DenseNet network architecture, which includes an initial convolutional layer, Dense Block 1, Transition Layer 1, Dense Block 2, Transition Layer 2, global average pooling layer, and two fully connected layers. The two fully connected layers are used for the hidden layer and the output layer respectively. Dense Block 1 contains 3 convolutional layers, Dense Block 2 contains 2 convolutional layers, and the transition layer contains a 1×1 convolutional layer and a pooling layer; the pre-training data is the simulated site D e with magnetic field intensity data generated according to the magnetic dipole model.

[0029] As a further solution of the present invention: In step four, the capsule robot positioning network CRLocNet includes a data correction sub-network and a position resolver network. Its input data size is 3×n×n, where n is the number of rows and columns of the sensor square array, and 3 represents the number of channels, corresponding to the components of the magnetic field intensity measured by the sensor array (B dx , B dy , B dz ). The magnetic field intensity values measured by the sensor array are converted into an input in the same form as a color pixel image, and the association features between the position of the capsule robot and the magnetic sensors at different coordinates are extracted;

[0030] The training data set of the capsule robot positioning network CRLocNet is the real data measured at the measurement site D m ;

[0031] The pose loss function of the capsule robot positioning network CRLocNet is as follows:

[0032] L = L position + λ1L direction + λ2L magnetic

[0033] Among them, L position is the position error:

[0034] L position = ||p pred - p true ||2

[0035] L direction is the position error:

[0036] L direction = ||θ pred - θ true |||2

[0037] L magnetic is the magnetic field error, which represents the physical consistency check of the magnetic field. The magnetic field intensity is deduced based on the position predicted by the neural network, and the difference between the comparison and the measured value is used to ensure the consistency between the neural network output and the physical model:

[0038]

[0039] In the formula, p pred , θ pred and B pred are the position, angle, and the calculated magnetic field intensity predicted by the neural network respectively. p true , θ true and B measured are the real position of the capsule robot and the real measured value of the sensor respectively. λ1 and λ2 are the proportionality factors for balancing the weights, and ||·||2 is the 2-norm.

[0040] As a further solution of the present invention: In step five, the improved northern goshawk optimization algorithm is used to optimize the weights and thresholds of CRLocNet. The learning rate, DropoutRate, and Batch Size are selected as the parameters to be optimized. The iterative optimization goal of the improved northern goshawk optimization algorithm is to make the error between the predicted position and the actual position as small as possible. Therefore, the objective function is defined as:

[0041]

[0042] In the formula, E pred and E true are the positions of the capsule robot predicted by the neural network and its real position.

[0043] The Northern Goshawk Optimization Algorithm simulates the behavior of the northern goshawk during hunting, which is specifically described as follows:

[0044] First, initialize the northern goshawk population, which can be represented by the following population matrix:

[0045]

[0046] where X is the northern goshawk population matrix, and X i is the position of the i-th northern goshawk, and x i,j is the position of the i-th northern goshawk in the j-th dimension. N is the population size of the northern goshawks;

[0047] Subsequently, enter the first stage of hunting. Randomly select a prey in the space and then quickly attack. The purpose is to globally search the search space to determine the optimal area. This behavior can be described by the formula:

[0048] P i = X k , i = 1, 2,..., N, k = 1, 2,..., i - 1, i + 1,..., N

[0049]

[0050] where P i is the prey position of the i-th northern goshawk, and F Pi is the objective function value of the prey position of the i-th northern goshawk; k is a random integer from 1 to N, is the new position of the i-th northern goshawk, is the corresponding fitness value; r is a random number in the range of 0 to 1, and the value of I is 1 or 2;

[0051] After the northern goshawk first attacks the prey, the prey will try to escape. Therefore, a new round of pursuit is required; the extremely high pursuit speed of the northern goshawk allows them to chase and capture the prey almost in any situation. Assume that the attack range of the new round of pursuit is a circle with a radius of R. In the second stage:

[0052]

[0053] where t is the current iteration number, T is the maximum iteration number, is the new position of the i-th northern goshawk, is the new position of the i-th northern goshawk in the j-th dimension, is the new position of the i-th northern goshawk in the j-th dimension after the second stage update, is its corresponding objective function value.

[0054] As a further solution of the present invention: For the improved Northern Goshawk optimization algorithm, in the initialization stage, the Kent chaos mapping strategy is adopted, and the specific expression is as follows:

[0055]

[0056] In the formula, the control parameter a ∈ (0, 1), and the generated initial chaos sequence s i+1 ∈ (0, 1), which is used to initialize the position of the Northern Goshawk population, as shown in the following formula:

[0057] x ij = lb + s k (ub - lb).

[0058] In the formula, ub is the upper boundary of the individual position, lb is the lower boundary of the individual position, and s k is the Kent chaos mapping factor.

[0059] As a further solution of the present invention: For the improved Northern Goshawk optimization algorithm, in the first hunting stage, the elite guidance strategy is adopted, and the Northern Goshawk with the best position is used for position update; at the same time, the subtraction optimizer algorithm is introduced to update the position of the Northern Goshawk, and the positions of the global Northern Goshawks are comprehensively updated:

[0060]

[0061] In the formula, x best is the optimal Northern Goshawk position at the current moment, and v- represents a special subtraction operation:

[0062] X i - v X k = sign(F(X i ) - F(X k ))(X i - v * X k )

[0063] In the formula, X i and X k are the objective function values of the i-th Northern Goshawk and the k-th Northern Goshawk respectively, v is a vector with a dimension of m, v i ∈ [1, 2], and sign is the signum function;

[0064] For the improved Northern Goshawk optimization algorithm, in the second hunting stage, a dynamic update strategy based on the upper and lower limit positions is introduced to gradually narrow the target range:

[0065]

[0066] Wherein, ub is the upper boundary of the individual position, lb is the lower boundary of the individual position, and t is the current iteration number.

[0067] The beneficial effects of the present invention are as follows:

[0068] 1) By designing a mobile magnetic sensor array module, the problem of limited positioning space of the fixed sensor array and large positioning error at the edge position is solved, and high-precision positioning of the magnetic control capsule robot in a large space is realized;

[0069] 2) A neural network CRLocNet for positioning the magnetic control capsule robot is built, which improves the defects that the positioning accuracy of the traditional position calculation method is greatly affected by the initial value and the positioning error accumulates over time, and the positioning accuracy and speed meet the clinical requirements;

[0070] 3) The present invention designs a sub-network structure with data correction function, effectively corrects the error between the measurement data and the magnetic dipole model, introduces physical information into the neural network, and improves the stability and interpretability of the solution network;

[0071] 4) The initialization population of the northern goshawk optimization algorithm is improved based on the Kent chaotic mapping, and the first hunting stage and the second hunting stage of the algorithm are improved based on the subtraction optimizer and the dynamic position update strategy, effectively improving the global search ability and convergence rate of the algorithm. By optimizing the hyperparameters of the CRLocNet network through the improved northern goshawk optimization algorithm, the performance of the network model is effectively improved. Description of the Drawings

[0072] Figure 1 It is a schematic structural diagram of the capsule robot positioning system of the present invention;

[0073] Figure 2 It is a control flow chart of the capsule robot positioning system of the present invention;

[0074] Figure 3 It is a network framework diagram of the capsule robot positioning neural network CRLocNet of the present invention;

[0075] Figure 4 It is an algorithm flow chart of the improved northern goshawk optimization algorithm for the capsule robot of the present invention;

[0076] Figure 5 It is a convergence curve diagram of test function 1 in the improved northern goshawk optimization algorithm and multiple other intelligent optimization algorithms in the embodiment of the present invention;

[0077] Figure 6 It is a convergence curve diagram of test function 2 in the improved northern goshawk optimization algorithm and multiple other intelligent optimization algorithms in the embodiment of the present invention.

[0078] In the figure: 101, robotic arm; 102, external driving permanent magnet; 103, capsule robot; 104, magnetic sensor array; 105, orthogonal slide. Specific implementation mode

[0079] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0080] Embodiment 1, as Figures 1 to 2 shown, a large-space magnetic control capsule robot positioning system based on CRLocNet includes an external driving module, a capsule robot 103, a mobile sensor array module, a signal processing module based on a neural network, and a control module. The external driving module includes a robotic arm 101 and an external driving permanent magnet 102. The external driving permanent magnet 102 is connected to the driving end of the robotic arm 101. The mobile sensor array module includes a magnetic sensor array 104 and an orthogonal slide 105. The magnetic sensor array 104 is fixed on the mounting platform of the orthogonal slide 105;

[0081] The control module includes a microprocessor, a robotic arm control system, and a mobile sensor array control system. The robotic arm control system can adjust the movement of the external driving permanent magnet 102 by receiving the real-time position information of the capsule robot 103 output by the signal processing module based on the neural network, so as to drive the capsule robot 103 to move to the next target position; the magnetic sensor array 104 moves synchronously with the external driving permanent magnet 102, and the external driving permanent magnet 102 is located directly above the movement space of the capsule robot 103 to ensure that the capsule robot 103 and the external driving permanent magnet 102 are within the effective positioning space of the magnetic sensor. The signal processing module based on the neural network is embedded with a magnetic positioning model and a capsule robot positioning neural network CRLocNet including a data correction sub-network and a position resolution sub-network.

[0082] Embodiment 2, in addition to including all the technical features in Embodiment 1, this embodiment further includes: the robotic arm 101 is a six-degree-of-freedom robotic arm, the external driving permanent magnet 102 has a cylindrical shape, and the capsule robot 103 contains a small cylindrical permanent magnet inside;

[0083] The magnetic sensor array 104 consists of an n×n square arrangement of magnetic sensors. The orthogonal sliding stage 105 includes two mutually orthogonal linear modules, which are respectively driven by stepping motors. The magnetic sensor array 104 moves within the x-y plane 100 mm below the capsule robot, and the moving range of the magnetic sensor array 104 is larger than the x-y plane motion range of the capsule robot 103.

[0084] Example 3, as Figure 3 and Figure 4 shown, a large-space magnetic control capsule robot positioning method based on CRLocNet includes a robot positioning system. This positioning method includes the following steps:

[0085] Step 1: Collect data and establish a network training dataset;

[0086] Step 2: Establish a data correction sub-network, pre-train it using model data with added random noise, and achieve preliminary processing of the collected data;

[0087] Step 3: Establish a position solution operator network, pre-train it using model-generated data, and achieve its function of solving the magnetic dipole inverse model;

[0088] Step 4: According to the collected data, establish a CRLocNet model and train it to achieve the function of capsule robot position calculation;

[0089] Step 5: Improve the Northern Goshawk optimization algorithm to optimize the weights and thresholds of CRLocNet;

[0090] Step 6: Import the trained model into the host computer, collect magnetic field information in real time, and calculate the pose of the capsule robot.

[0091] Example 4, in addition to including all the technical features in Example 1, this example also includes:

[0092] In Step 1, the established dataset includes:

[0093] In the motion space of the capsule robot, a large number of sampling point information including spatial positions and attitude angles are randomly generated, where 20% is used as the measurement site set D m , and the rest is used as the simulation site D e ; The external drive module drives the capsule robot to move to the measurement sites in sequence and collects magnetic field intensity information through the sensor array; The simulation software generates the magnetic field intensity data of all sampling points based on the magnetic dipole model at each sensor position as the dataset of CRLocNet;

[0094] The generation of simulation site measurement data is based on the magnetic dipole model:

[0095]

[0096] Wherein, P = (a, b, c) is the central position of the internal magnet of the capsule robot, and P s = (x i , y i , z i ) is an arbitrary sampling point in space, where i = 1, 2, N, B = (B x , B y , B z ) is the magnetic field strength at this point. The vector of the spatial distance and direction between this point and the permanent magnet can be expressed as P l = P s - P, and R l is the modulus of P l ; μ0 is the magnetic permeability of vacuum, μ r is the relative magnetic permeability of air, M T is the dipole moment strength of the magnet, and H0 = (m, n, p) is the magnetization direction of the permanent magnet at P.

[0097] In step two, the data correction sub-network is a fully connected CNN network with c hidden layers, which is used to correct the real measurement data and remove the noise and errors generated by the real measurement data for various reasons. The number of nodes in the initial hidden layer is determined by the following formula:

[0098]

[0099] Wherein, N s is the number of samples in the training set, N i is the number of neurons in the input layer, N o is the number of neurons in the output layer, and C a is an independent variable that can take any value. Usually, C a ∈(2, 10);

[0100] The pre-training data set is the magnetic field strength data generated based on the magnetic dipole model for the simulation site D e with random noise added. The added noise conforms to the Gaussian distribution:

[0101]

[0102] In step 3, the position resolver network adopts the DenseNet network architecture, which includes an initial convolutional layer, Dense Block 1, Transition Layer 1, Dense Block 2, Transition Layer 2, a global average pooling layer, and two fully connected layers. The two fully connected layers are used for the hidden layer and the output layer respectively. Dense Block 1 contains 3 convolutional layers, Dense Block 2 contains 2 convolutional layers, and the transition layer contains a 1×1 convolutional layer and a pooling layer. The pre-training data is the simulation site D e Magnetic field intensity data generated according to the magnetic dipole model.

[0103] In step 4, the capsule robot positioning network CRLocNet includes a data correction sub-network and a position resolver network. The size of its input data is 3×n×n, where n is the number of rows and columns of the sensor square matrix, and 3 represents the number of channels, corresponding to the components of the magnetic field intensity measured by the sensor array (B dx , B dy , B dz ). The magnetic field intensity values measured by the sensor array are converted into inputs in the same form as a color pixel image, and the correlation features between the position of the capsule robot and different coordinate magnetic sensors are extracted;

[0104] The training data set of the capsule robot positioning network CRLocNet is the real data measured at the measurement site D m ;

[0105] The pose loss function of the capsule robot positioning network CRLocNet is:

[0106] L = L position + λ1L direction + λ2L magnetic

[0107] where L position is the position error:

[0108] L position = ||p pred - p true ||2

[0109] L direction is the position error:

[0110] L direction = ||θ pred - θ true |||2

[0111] L magneticis the magnetic field error, which represents the physical consistency check of the magnetic field. The magnetic field intensity is deduced based on the position predicted by the neural network, and the difference between the comparison and the measured value is used to ensure the consistency between the neural network output and the physical model:

[0112]

[0113] where p pred , θ pred and B pred are the position, angle predicted by the neural network and the magnetic field intensity calculated therefrom, respectively. p true , θ true and B measured are the true position of the capsule robot and the true measured value of the sensor, respectively. λ1 and λ2 are the proportionality factors for balancing the weights, and ||·||2 is the 2-norm.

[0114] In step five, the improved northern goshawk optimization algorithm is used to optimize the weights and thresholds of CRLocNet. The learning rate, DropoutRate, and Batch Size are selected as the parameters to be optimized. In order to achieve the best optimization effect, the iterative optimization objective of the improved northern goshawk optimization algorithm is designed to make the error between the predicted position and the actual position as small as possible. Therefore, the objective function is defined as:

[0115]

[0116] where E pred and E true are the positions of the capsule robot predicted by the neural network and its true position.

[0117] [[ID=3�]]The northern goshawk optimization algorithm simulates the behavior of the northern goshawk during hunting, which is specifically described as:

[0118] First, the northern goshawk population is initialized, which can be represented by the following population matrix:

[0119]

[0120] where X is the northern goshawk population matrix, X i is the position of the i-th northern goshawk, and x i,j is the position of the i-th northern goshawk in the j-th dimension. N is the population size of the northern goshawks;

[0121] Subsequently, enter the first stage of hunting. Randomly select a prey in the space and then quickly attack. The purpose is to globally search the search space to determine the optimal area. This behavior can be described by the formula as:

[0122] P i = X k, where \(i = 1, 2, \cdots, N\) and \(k = 1, 2, \cdots, i - 1, i + 1, \cdots, N\)

[0123]

[0124] In the formula, \(P\) i is the prey position of the \(i\)-th northern goshawk, and \(F\) Pi is the objective function value of the prey position of the \(i\)-th northern goshawk; \(k\) is a random integer from 1 to \(N\), is the new position of the \(i\)-th northern goshawk, and \(f\) is the corresponding fitness value; \(r\) is a random number in the range from 0 to 1, and the value of \(I\) is 1 or 2;

[0125] After the northern goshawk first attacks the prey, the prey will try to escape, so a new round of pursuit is needed; the extremely high pursuit speed of the northern goshawk allows them to chase and capture the prey almost in any situation. Assume that the attack range of the new round of pursuit is a circle with a radius of \(R\). In the second stage:

[0126]

[0127] In the formula, \(t\) is the current iteration number, and \(T\) is the maximum iteration number, is the new position of the \(i\)-th northern goshawk, is the new position of the \(j\)-th dimension of the \(i\)-th northern goshawk, is the new position of the \(j\)-th dimension of the \(i\)-th northern goshawk after the second-stage update, and \(f\) is its corresponding objective function value.

[0128] Furthermore, for the improved northern goshawk optimization algorithm, in the initialization stage, the Kent chaos mapping strategy is adopted, and the specific expression is as follows:

[0129]

[0130] In the formula, the control parameter \(a\in(0, 1)\), and the generated initial chaos sequence \(s\) i+1 \(\in(0, 1)\) is used to initialize the position of the northern goshawk population, as shown in the following formula:

[0131] \(x\) ij \(= lb + s\) k (ub - lb).

[0132] In the formula, \(ub\) is the upper boundary of the individual position, \(lb\) is the lower boundary of the individual position, and \(s\) k is the Kent chaos mapping factor.

[0133] For the improved Northern Goshawk optimization algorithm, in the first hunting stage, an elite-guided strategy is adopted, and the Northern Goshawk at the best position is used for position update; at the same time, the Subtraction Optimizer algorithm is introduced to update the position of the Northern Goshawk, and the positions of the global Northern Goshawks are comprehensively updated:

[0134]

[0135] In the formula, x best is the optimal position of the Northern Goshawk at the current moment, and v- represents a special subtraction operation:

[0136] X i - v X k = sign(F(X i ) - F(X k ))(X i - v * X k )

[0137] In the formula, X i and X k are the objective function values of the i-th Northern Goshawk and the k-th Northern Goshawk respectively, v is a vector with dimension m, v i ∈[1, 2], and sign is the signum function;

[0138] For the improved Northern Goshawk optimization algorithm, in the second hunting stage, a dynamic update strategy based on upper and lower limit positions is introduced to gradually narrow the target range:

[0139]

[0140] In the formula, ub is the upper boundary of the individual position, lb is the lower boundary of the individual position, and t is the current iteration number.

[0141] Example 5, A positioning method for a large-space magnetically controlled capsule robot based on CRLocNet. The commonly used test functions selected in this example are as follows:

[0142]

[0143] The improved Northern Goshawk optimization algorithm (INGO) is compared with the basic Northern Goshawk optimization algorithm (NGO), the Dung Beetle Optimization algorithm (DBO), the Golden Jackal Optimization algorithm (GJO), the Subtraction Averaging Optimization algorithm (SAO), and the Sparrow Search algorithm (SSA) to verify the performance of the improved Northern Goshawk optimization algorithm. To ensure the fairness of the test, the population size of each algorithm is set to 30, and the maximum number of iterations is set to 500.

[0144] As Figure 5As shown, it can be seen that the improved Northern Goshawk Optimization Algorithm has greater advantages in terms of convergence speed and optimization ability compared to other algorithms.

[0145] Example 6, A positioning method for large-space magnetically controlled capsule robots based on CRLocNet. The commonly used test functions selected in this example are as follows:

[0146]

[0147] Compare the improved Northern Goshawk Optimization Algorithm (INGO) with the basic Northern Goshawk Optimization Algorithm (NGO), Dung Beetle Optimization Algorithm (DBO), Golden Jackal Optimization Algorithm (GJO), Subtraction Mean Optimization Algorithm (SAO), and Sparrow Search Algorithm (SSA) to verify the performance of the improved Northern Goshawk Optimization Algorithm. To ensure the fairness of the test, the population size of each algorithm is set to 30, and the maximum number of iterations is set to 500.

[0148] As Figure 6 shown, it can be seen that the improved Northern Goshawk Optimization Algorithm has greater advantages in terms of convergence speed and optimization ability compared to other algorithms.

[0149] The capsule robot 103 enters the gastrointestinal tract through the oral cavity and upper digestive tract of the subject, and its movement is controlled by an external drive module to complete corresponding tasks such as imaging, drug delivery, and sampling. The mobile sensor module and the external drive module are located on both sides of the subject. The structural diagram given in the present invention is in the form of the subject lying down for examination. During the implementation process, the sensor array and the driving permanent magnet can be placed on the front and back sides of the subject respectively to also achieve standing examination and positioning.

[0150] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

[0151] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A large-space magnetically controlled capsule robot positioning system based on CRLocNet, comprising an external driving module, a capsule robot (103), a mobile sensor array module, a signal processing module based on a neural network, and a control module, characterized in that: The external driving module includes a robotic arm (101) and an external driving permanent magnet (102). The external driving permanent magnet (102) is connected to the driving end of the robotic arm (101). The mobile sensor array module includes a magnetic sensor array (104) and an orthogonal slide table (105). The magnetic sensor array (104) is fixed on the mounting platform of the orthogonal slide table (105). The control module includes a microprocessor, a robotic arm control system, and a mobile sensor array control system. The robotic arm control system receives the real-time position information of the capsule robot (103) output by the signal processing module based on a neural network. The magnetic sensor array (104) moves synchronously with the external driving permanent magnet (102), and the external driving permanent magnet (102) is located directly above the motion space of the capsule robot (103). The signal processing module based on a neural network is embedded with a magnetic positioning model and a capsule robot positioning neural network CRLocNet including a data correction sub-network and a position solver network.

2. The large-space magnetically controlled capsule robot positioning system according to claim 1, wherein: The robotic arm (101) is a six-degree-of-freedom robotic arm. The external driving permanent magnet (102) has a cylindrical shape. The capsule robot (103) contains a small cylindrical permanent magnet inside. The magnetic sensor array (104) is arranged in a square matrix by n×n magnetic sensors. The orthogonal slide table (105) includes two mutually orthogonal linear modules, which are respectively driven by stepping motors. The magnetic sensor array (104) moves in the x-y plane 100 mm below the capsule robot, and the moving range of the magnetic sensor array (104) is larger than the x-y plane motion range of the capsule robot (103).

3. A positioning method for a large-space magnetically controlled capsule robot based on CRLocNet, comprising the large-space magnetically controlled capsule robot positioning system according to any one of claims 1 to 2, characterized in that, The large-space magnetic control capsule robot positioning method includes the following steps: Step 1: Collect data and establish a network training data set. Step 2: Establish a data correction sub-network and pre-train it with model data added with random noise to achieve preliminary processing of the collected data. Step 3: Establish a position solver network and pre-train it with model-generated data to achieve its function of solving the magnetic dipole inverse model. Step 4: Establish a CRLocNet model based on the collected data and train it to achieve the capsule robot position calculation function. Step 5: Improve the Northern Goshawk optimization algorithm to optimize the weights and thresholds of CRLocNet. Step 6: Import the trained model into the host computer, collect magnetic field information in real time, and calculate the pose of the capsule robot.

4. The large-space magnetic control capsule robot positioning method according to claim 3, characterized in that: In the above Step 1, the established data set includes: In the motion space of the capsule robot, a large number of sampling point information including spatial positions and attitude angles are randomly generated, where 20% is used as the measurement site set D m , and the rest are used as simulation sites D e ; The external driving module drives the capsule robot to move to the measurement sites in sequence, and the magnetic field intensity information is collected through the sensor array. The simulation software generates the magnetic field intensity data of all sampling points based on the magnetic dipole model at each sensor position as the data set of CRLocNet. The generation of the simulation site measurement data is based on the magnetic dipole model: Wherein, P = (a, b, c) is the central position of the internal magnet of the capsule robot, and P s = (x i , y i , z i ) is an arbitrary sampling point in space, where i = 1, 2, N, B = (B x , B y , B z ) is the magnetic field strength at this point, and the vector of the spatial distance and direction between this point and the permanent magnet can be expressed as P l = P s - P, and R l is the modulus of P l ; μ0 is the magnetic permeability of vacuum, μ r is the relative magnetic permeability of air, M T is the dipole moment strength of the magnet, and H0 = (m, n, p) is the magnetization direction of the permanent magnet at P.

5. The large-space magnetically controlled capsule robot positioning method according to claim 3, characterized in that: In the above Step 2, the data correction sub-network is a fully connected CNN network with c hidden layers, which is used to correct the real measurement data and remove the noise and errors generated by the real measurement data for various reasons. The number of initial hidden layer nodes is determined by the following formula: Where N s is the number of training set samples, N i is the number of neurons in the input layer, N o is the number of neurons in the output layer, C a is an independent variable that can take any value. Usually, C a ∈(2, 10); The pre-training dataset is the simulated site D with added random noise e Based on the magnetic field intensity data generated by the magnetic dipole model, the added noise conforms to the Gaussian distribution:

6. The large-space magnetically controlled capsule robot positioning method according to claim 3, characterized in that: In the third step, the position resolver network adopts a DenseNet network architecture, including an initial convolutional layer, Dense Block 1, Transition Layer 1, Dense Block 2, Transition Layer 2, a global average pooling layer, and two fully connected layers, where the two fully connected layers are used for the hidden layer and the output layer respectively. Dense Block 1 contains 3 convolutional layers, Dense Block 2 contains 2 convolutional layers, and the transition layer contains a 1×1 convolutional layer and a pooling layer; the pre-training data is the simulation site D e Magnetic field intensity data generated according to the magnetic dipole model.

7. The large-space magnetically controlled capsule robot positioning method according to claim 3, wherein: In the fourth step, the capsule robot positioning network CRLocNet includes a data correction sub-network and a position solution operator network. Its input data size is 3×n×n, where n is the number of rows and columns of the sensor square array, and 3 represents the number of channels, corresponding to the components of the magnetic field intensity measured by the sensor array (B dx , B dy , B dz ). The magnetic field intensity values measured by the sensor array are converted into inputs in the same form as a color pixel image, and the correlation features between the position of the capsule robot and different coordinate magnetic sensors are extracted; The training dataset of the capsule robot localization network CRLocNet is the real data measured at the measurement site D m measured data; The pose loss function of the capsule robot positioning network CRLocNet is as follows: L = L position + λ1L direction + λ2L magnetic where L position is the position error: L position = ||p pred -p true ||² L direction is the position error: L direction = || θ pred - θ true || 2 L magnetic is the magnetic field error, representing the physical consistency check of the magnetic field. The magnetic field intensity is deduced based on the position predicted by the neural network, and the consistency between the neural network output and the physical model is ensured by comparing the difference with the measured value: where p pred , θ pred and B pred are the position, angle, and the calculated magnetic field strength predicted by the neural network respectively, p true , θ true and B measured are the true position of the capsule robot and the true measurement values of the sensor respectively, λ1 and λ2 are the proportionality factors for balancing the weights, and ||·||2 is the 2-norm.

8. The large-space magnetic control capsule robot positioning method according to claim 3, characterized in that: In step 5, the improved northern goshawk optimization algorithm is used to optimize the weights and thresholds of CRLocNet. The learning rate, DropoutRate, and Batch Size are selected as the parameters to be optimized. The iterative optimization objective of the improved northern goshawk optimization algorithm is to minimize the error between the predicted position and the actual position. Therefore, the objective function is defined as: where, E pred and E true are the positions of the capsule robot predicted by the neural network and its true position; The northern goshawk optimization algorithm simulates the behavior of the northern goshawk during hunting, which is specifically described as follows: First, the northern goshawk population is initialized, which is represented by the following population matrix: In the formula, X is the population matrix of northern goshawks, and X i is the position of the i-th northern goshawk, and x i,j is the position of the i-th northern goshawk in the j-th dimension, and N is the population size of northern goshawks; Subsequently, enter the first stage of hunting. Randomly select prey in the space and then quickly attack, aiming to globally search the search space and determine the optimal area. This behavior is described by the formula as: P i = X k , i = 1, 2, ..., N, k = 1, 2, ..., i - 1, i + 1, ..., N where P i is the prey position of the i-th northern goshawk, and F Pi is the objective function value of the prey position of the i-th northern goshawk; k is a random integer from 1 to N, and X i new,P1 is the new position of the i-th northern goshawk, and F i new,P1 is the corresponding fitness value; r is a random number in the range of 0 to 1, and the value of I is 1 or 2; After the northern goshawk first attacks the prey, the prey will try to escape. Therefore, a new round of pursuit needs to be carried out. Assume that the attack range of the new round of pursuit is a circle with a radius of R. In the second stage: where \(t\) is the current iteration number, \(T\) is the maximum iteration number, \(X\) i new,P1 is the new position of the \(i\)-th Northern Goshawk, \(x\) i new,P2 is the new position of the \(j\)-th dimension of the \(i\)-th Northern Goshawk, \(x\) i,j P2 is the new position of the \(j\)-th dimension of the \(i\)-th Northern Goshawk after the second-stage update, \(F\) i new,P2 is its corresponding objective function value.

9. The positioning method of the large-space magnetically controlled capsule robot according to claim 8, wherein: For the improved northern goshawk optimization algorithm, in the initialization stage, the Kent chaos mapping strategy is adopted, and the specific expression is as follows: where the control parameter \(a\in(0,1)\), and the generated initial chaotic sequence \(s\) i+1 \(\in(0,1)\) is used to initialize the positions of the northern goshawk population, as shown in the following formula: x ij = lb + s k (ub - lb) where \(u_b\) is the upper bound of the individual position, \(l_b\) is the lower bound of the individual position, and \(s\) k is the Kent chaos mapping factor.

10. The large-space magnetic control capsule robot positioning method according to claim 8, characterized in that: For the improved northern goshawk optimization algorithm, in the first hunting stage, the elite guidance strategy is adopted, and the northern goshawk at the best position is used to update the position; at the same time, the subtraction optimizer algorithm is introduced to update the position of the northern goshawk, and the position of the global northern goshawk is comprehensively updated: where x best is the optimal position of the northern goshawk at the current moment, v - represents a special subtraction operation: X i-v X k = sign(F(X i )) - F(X k ))(X i - v * X k ) where X i and X k are the objective function values of the i-th northern goshawk and the k-th northern goshawk respectively, v is a vector of dimension m, v i ∈ [1, 2], and sign is the signum function; For the improved northern goshawk optimization algorithm, in the second hunting stage, a dynamic update strategy based on the upper and lower limit positions is introduced to gradually narrow the target range: In the formula, ub is the upper boundary of the individual position, lb is the lower boundary of the individual position, and t is the current iteration number.

Images