A mobile robot positioning method based on inertial measurement unit and deep learning

Abstract

This invention discloses a mobile robot positioning method based on an inertial measurement unit (IMU) and deep learning. The method includes inputting raw IMU data into a trained deep learning model to obtain IMU error calibration parameters and noise parameters of an invariant Kalman filter. The IMU error calibration model is then used to obtain calibrated acceleration and angular velocity. These calibrated acceleration and angular velocity are then input into the invariant Kalman filter to obtain the mobile robot's positioning results. This method can adapt to the motion characteristics of different mobile robots, achieving higher-precision mobile robot positioning results. It also helps reduce the cost of mobile robot positioning and is less susceptible to environmental interference.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile robots, and in particular relates to a mobile robot positioning method based on an inertial measurement unit and deep learning. Background Art

[0002] As mobile robots are increasingly used in production and life, Simultaneous Localization and Mapping (SLAM) is regarded as a key technology for autonomous mobile robots in the field of robot positioning.

[0003] However, common SLAM systems typically use lidar or cameras as their primary positioning sensors, using costly computations to produce low-frequency (typically 10 Hz) high-precision positioning results. This significantly increases the cost of mobile robot positioning. Furthermore, lidar and cameras are also highly susceptible to environmental interference, resulting in positioning anomalies when dealing with varying light intensities and similar environmental structures. Furthermore, when a mobile robot is moving at high speeds, low-frequency positioning information cannot meet the required positioning accuracy.

[0004] While inertial measurement units (IMUs) can acquire a mobile robot's acceleration and angular velocity through their internal gyroscopes and accelerometers, and can be used to assist in positioning within the mobile robot, high-precision laser IMUs are often too expensive, and low-cost micro-electromechanical system inertial measurement units (MEMS-IMUs) inevitably contain a large amount of noise due to their electrical characteristics, making them incapable of being used by traditional algorithms for accurate positioning of mobile robots. Furthermore, with recent breakthroughs in artificial intelligence and deep learning technologies, deep learning-based perception technologies are playing an increasingly important role in the perception of unmanned driving environments. Deep neural networks have excellent fitting capabilities and have been applied to various mobile robot fields, including target tracking, semantic segmentation, and path planning.

[0005] Therefore, with the continuous development of mobile robot technology, there is an urgent need for a mobile robot positioning method that is low-cost, high-frequency, high-precision, and not susceptible to external interference. Summary of the Invention

[0006] The purpose of the present invention is to address the above-mentioned problems and propose a mobile robot positioning method based on inertial measurement unit and deep learning, which can adapt to the motion characteristics of different mobile robots, obtain more accurate mobile robot positioning results, and help reduce the positioning cost of mobile robots and is not easily affected by environmental interference.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] The present invention proposes a mobile robot positioning method based on an inertial measurement unit and deep learning. The mobile robot includes an inertial measurement unit, and the method includes the following steps:

[0009] S1. Obtain the original acceleration and original angular velocity of the mobile robot using an inertial measurement unit;

[0010] S2. Establish a deep learning model and train it. The deep learning model includes a dilated convolutional neural network and a linear fully connected network connected in sequence. The dilated convolutional neural network includes a dilated layer, a first convolutional module, a second convolutional layer module, a third convolutional module, a fourth convolutional module, and a fifth convolutional module connected in sequence. Each convolutional module includes a convolutional layer. The linear fully connected network includes a fully connected layer and a Tanh activation function layer connected in sequence.

[0011] S3. Input the original acceleration and original angular velocity into the trained deep learning model to obtain the error calibration parameters of the inertial measurement unit and the noise parameters of the invariant Kalman filter. The error calibration parameters of the inertial measurement unit include and The noise parameters of the invariant Kalman filter include r lat and r up ,in, They represent the proportional coefficients of the angular velocities of the x-axis, y-axis, and z-axis, respectively. In turn, they represent the proportional coefficients of the accelerations of the x-axis, y-axis, and z-axis. In turn, they represent the offset coefficients of the angular velocity of the x-axis, y-axis, and z-axis. Represents the offset coefficients of the x-axis, y-axis, and z-axis accelerations, r lat represents the velocity noise covariance in the y-axis direction, r up represents the velocity noise covariance in the z-axis direction;

[0012] S4. Obtain calibrated acceleration and angular velocity using the error calibration model of the inertial measurement unit. The error calibration model is established as follows:

[0013]

[0014] in,

[0015]

[0016]

[0017] Where, Represents the original IMU data at the nth moment, ω n represents the original angular velocity at the nth moment, a n represents the original acceleration at the nth moment, Represents the IMU data after calibration at the nth moment, represents the angular velocity after calibration at the nth moment, Indicates the acceleration after calibration at the nth moment, A n Represents the proportional coefficient matrix of the error calibration model at the nth moment, bias n represents the offset coefficient vector of the error calibration model at the nth moment;

[0018] S5. Input the calibrated acceleration and angular velocity into the invariant Kalman filter to obtain the positioning result of the mobile robot.

[0019] Preferably, the first convolution module, the second convolution layer module, the third convolution module and the fourth convolution module also include a BN layer, a GeLU activation function layer and a Dropout layer, and the convolution layer, the BN layer, the GeLU activation function layer and the Dropout layer in each convolution module are connected in sequence.

[0020] Preferably, the input dimension of the convolution layer of the first convolution module is 6, the output dimension is 16, the convolution kernel size is 7, and the expansion coefficient is 1; the input dimension of the convolution layer of the second convolution layer module is 16, the output dimension is 32, the convolution kernel size is 7, and the expansion coefficient is 4; the input dimension of the convolution layer of the third convolution module is 32, the output dimension is 64, the convolution kernel size is 7, and the expansion coefficient is 16; the input dimension of the convolution layer of the fourth convolution module is 64, the output dimension is 128, the convolution kernel size is 7, and the expansion coefficient is 64; the input dimension of the convolution layer of the fifth convolution module is 128, the output dimension is 14, the convolution kernel size is 7, and the expansion coefficient is 1.

[0021] Preferably, the calibrated acceleration and angular velocity are input into an invariant Kalman filter to obtain the positioning result of the mobile robot, as follows:

[0022] S51, the state variable x at the nth moment in the unchanged Kalman filter n The definition is as follows:

[0023]

[0024] Where, Represents the rotation matrix from the world coordinate system to the IMU coordinate system at the nth moment, Indicates the velocity of the inertial measurement unit in the world coordinate system at the nth moment, Indicates the position of the inertial measurement unit in the world coordinate system at the nth moment, represents the gyroscope bias of the inertial measurement unit at the nth moment, represents the accelerometer bias of the inertial measurement unit at the nth moment, Represents the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system at the nth moment, Represents the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system at the nth moment;

[0025] S52. Establish a nonlinear state prediction equation:

[0026]

[0027] Where, Represents the state quantity predicted at the nth moment, x n-1 is the state variable at the n-1th moment, and f(·) is the dynamic model of the mobile robot, which is expressed as follows:

[0028]

[0029] in,

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036]

[0037] Where, Represents the rotation matrix from the world coordinate system to the IMU coordinate system predicted at the nth moment, It represents the velocity of the inertial measurement unit in the world coordinate system predicted at the nth moment, represents the position of the inertial measurement unit in the world coordinate system predicted at the nth moment, represents the gyroscope bias of the inertial measurement unit predicted at time n, represents the accelerometer bias of the inertial measurement unit predicted at time n, Represents the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system predicted at the nth moment, It represents the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system predicted at the nth moment, dt is the time interval, and g is the gravitational acceleration of the current area;

[0038] S53. Establish the covariance equation:

[0039]

[0040] in,

[0041]

[0042]

[0043]

[0044] Where, is the predicted system state covariance matrix, F n For the nonlinear state prediction equation, n The Jacobian matrix, G n For the nonlinear state prediction equation The Jacobian matrix of the mobile robot, Q is the process noise, σ ω is the standard deviation of angular velocity, σ a is the standard deviation of acceleration, is the standard deviation of the angular velocity offset, is the standard deviation of the acceleration offset, is the standard deviation of the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system, is the standard deviation of the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system, P n-1 is the system state covariance matrix at the n-1th moment, 0 a×b Represents a row and b column 0 matrix, I d×d represents the identity matrix of dimension d;

[0045] S54. Establish the system observation equation:

[0046]

[0047] Where y n is the system observation value at the nth moment, h(x n ) is the state variable x at the nth moment n The system observation equation is is the observed value in the horizontal direction at the nth moment, is the vertical observation value at the nth moment, h lat (x n ) is the system observation equation in the horizontal direction at the nth moment, h up (x n ) is the system observation equation in the vertical direction at the nth moment, is the horizontal observation noise at the nth moment, is the vertical observation noise at the nth moment, n nrepresents the observation noise at the nth moment, the horizontal direction represents the y-axis direction of the mobile robot carrier coordinate system, and the vertical direction represents the z-axis direction of the mobile robot carrier coordinate system;

[0048] And satisfy the following formula:

[0049]

[0050] Where, is the speed of the mobile robot at the nth moment, is the forward speed of the mobile robot at the nth moment, is the horizontal speed of the mobile robot at the nth moment, is the vertical velocity of the mobile robot at the nth moment, and the forward direction represents the x-axis direction of the mobile robot carrier coordinate system;

[0051] S55. Establish the Kalman gain equation:

[0052]

[0053] in,

[0054]

[0055] Where K n is the Kalman gain coefficient at the nth moment, H n is h(x n ) is the Jacobian matrix, R is the observation noise covariance matrix;

[0056] S56. Establish nonlinear state estimation equation:

[0057]

[0058] in,

[0059]

[0060] Where z n is the zero-speed pseudo-observation, is the state quantity predicted for the nth moment The system observation equation of

[0061] S57. Estimated covariance update equation:

[0062]

[0063] Where, P n is the system state covariance matrix at the nth moment, I 21×21 is the identity matrix with dimension 21;

[0064] S58, the output result x of step S56 n As the positioning result of the mobile robot at the nth moment.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows: this application adopts a mobile robot positioning method based on inertial measurement unit and deep learning, and obtains the error calibration model parameters of IMU and the noise parameters of invariant Kalman filter through deep learning model training. It can adapt to the motion characteristics of different mobile robots, and perform real-time estimation of IMU noise and error, thereby obtaining higher-precision mobile robot positioning results, which helps to reduce the positioning cost of mobile robots, is not easily affected by environmental interference, and can cope with accurate positioning under different light intensities and similar environmental structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of the mobile robot positioning method based on inertial measurement unit and deep learning of the present invention;

[0067] Figure 2 Schematic diagram of the structure of the deep learning model of the present invention;

[0068] Figure 3 This is the flow chart of the invariant Kalman filter of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0070] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein in the specification of this application are for the purpose of describing specific embodiments only and are not intended to limit this application.

[0071] like Figure 1-3 As shown, a mobile robot positioning method based on an inertial measurement unit and deep learning, the mobile robot includes an inertial measurement unit, and the method includes the following steps:

[0072] S1. Use the inertial measurement unit to obtain the original acceleration and original angular velocity of the mobile robot.

[0073] The Inertial Measurement Unit (IMU) can obtain the motion acceleration and angular velocity of a mobile robot through its internal gyroscope and accelerometer. The raw acceleration and angular velocity information measured by the IMU is a set of 6-dimensional data, including the acceleration and angular velocity of the x, y, and z axes.

[0074] S2. Establish a deep learning model and train it. The deep learning model includes a dilated convolutional neural network and a linear fully connected network connected in sequence. The dilated convolutional neural network includes a dilated layer, a first convolutional module, a second convolutional layer module, a third convolutional module, a fourth convolutional module and a fifth convolutional module connected in sequence. Each convolutional module includes a convolutional layer. The linear fully connected network includes a fully connected layer and a Tanh activation function layer connected in sequence.

[0075] In one embodiment, the first convolution module, the second convolution layer module, the third convolution module and the fourth convolution module also include a BN layer, a GeLU activation function layer and a Dropout layer, and the convolution layer, the BN layer, the GeLU activation function layer and the Dropout layer in each convolution module are connected in sequence.

[0076] In one embodiment, the input dimension of the convolution layer of the first convolution module is 6, the output dimension is 16, the convolution kernel size is 7, and the expansion coefficient is 1; the input dimension of the convolution layer of the second convolution layer module is 16, the output dimension is 32, the convolution kernel size is 7, and the expansion coefficient is 4; the input dimension of the convolution layer of the third convolution module is 32, the output dimension is 64, the convolution kernel size is 7, and the expansion coefficient is 16; the input dimension of the convolution layer of the fourth convolution module is 64, the output dimension is 128, the convolution kernel size is 7, and the expansion coefficient is 64; the input dimension of the convolution layer of the fifth convolution module is 128, the output dimension is 14, the convolution kernel size is 7, and the expansion coefficient is 1.

[0077] Specifically, the deep learning model used is composed of a dilated convolutional neural network and a linear fully connected network, such as Figure 2 shown.

[0078] Among them, the dilated convolutional neural network first uses an expansion layer (ReplicationPad1d, Pad) to copy the boundary data and expand the data length by 510 units. It then contains 5 convolution modules, which are respectively recorded as the first convolution module, the second convolution layer module, the third convolution module, the fourth convolution module and the fifth convolution module. Each convolution module includes a convolution layer (CNN).

[0079] The convolution layer of the first convolution module has an input dimension of 6, an output dimension of 16, a convolution kernel size of 7, and a dilation coefficient of 1. It is followed by a Batch Normalization (BN) layer with the same dimension as the output dimension of the previous convolution layer, which is used to normalize the feature quantities obtained by convolution. It is followed by a GeLU activation function (GeLU) to make the training have a certain degree of randomness and obtain more robust results. Then, through the Dropout layer (Dropout), some parameters are set to 0 to reduce overfitting.

[0080] The convolution layer of the second convolution module has an input dimension of 16, an output dimension of 32, a convolution kernel size of 7, and a dilation coefficient of 4. It is followed by a Batch Normalization (BN) layer with the same dimension as the output of the previous convolution layer, which is used to normalize the features obtained by convolution. This is followed by a GeLU activation function, which makes the training more random and can produce more robust results. Then, through the Dropout layer, some parameters are set to 0 to reduce overfitting.

[0081] The convolution layer of the third convolution module has an input dimension of 32, an output dimension of 64, a convolution kernel size of 7, and a dilation coefficient of 16. It is followed by a Batch Normalization (BN) layer with the same dimension as the output of the previous convolution layer, which is used to normalize the features obtained by convolution. This is followed by a GeLU activation function, which makes the training more random and can produce more robust results. Then, through the Dropout layer, some parameters are set to 0 to reduce overfitting.

[0082] The convolution layer of the fourth convolution module has an input dimension of 64, an output dimension of 128, a convolution kernel size of 7, and a dilation coefficient of 64. It is followed by a Batch Normalization (BN) layer with the same dimension as the output of the previous convolution layer, which is used to normalize the features obtained by convolution. This is followed by a GeLU activation function, which makes the training more random and can produce more robust results. Then, through the Dropout layer, some parameters are set to 0 to reduce overfitting.

[0083] The input dimension of the convolution layer of the fifth convolution module is 128, the output dimension is 14, the convolution kernel size is 7, and the expansion coefficient is 1, resulting in 14-dimensional feature data.

[0084] The output end of the dilated convolutional neural network is connected to a linear fully connected network, which includes a fully connected layer (FC) and a Tanh activation function layer (Tanh). The activation function limits the parameter range and reduces the occurrence of overfitting during training.

[0085] Finally, 14-dimensional parameter data is obtained, of which the first 6 dimensions are the scale parameters of the IMU noise model, including the scale coefficients of the x-axis, y-axis, and z-axis angular velocity. The proportional coefficients of the accelerations on the x-axis, y-axis, and z-axis The 7th to 12th dimensions are the offset parameters of the IMU noise model, including the offset coefficients of the x-axis, y-axis, and z-axis angular velocity. Offset coefficients of x-axis, y-axis, and z-axis acceleration The last two dimensions are the observation noise parameters of the Invariant Kalman Filter (IEKF), including the horizontal velocity noise covariance r lat and vertical velocity noise covariance r up .

[0086] S3. Input the original acceleration and original angular velocity into the trained deep learning model to obtain the error calibration parameters of the inertial measurement unit and the noise parameters of the invariant Kalman filter. The error calibration parameters of the inertial measurement unit include and The noise parameters of the invariant Kalman filter include r lat and r up ,in, They represent the proportional coefficients of the angular velocities of the x-axis, y-axis, and z-axis, respectively. In turn, they represent the proportional coefficients of the accelerations of the x-axis, y-axis, and z-axis. In turn, they represent the offset coefficients of the angular velocity of the x-axis, y-axis, and z-axis. Represents the offset coefficients of the x-axis, y-axis, and z-axis accelerations, r lat represents the velocity noise covariance in the y-axis direction, r up represents the velocity noise covariance in the z-axis direction;

[0087] S4. Obtain calibrated acceleration and angular velocity using the error calibration model of the inertial measurement unit. The error calibration model is established as follows:

[0088]

[0089] in,

[0090]

[0091]

[0092] Where, Represents the original IMU data at the nth moment, ω n represents the original angular velocity at the nth moment, a n represents the original acceleration at the nth moment, Represents the IMU data after calibration at the nth moment, represents the angular velocity after calibration at the nth moment, Indicates the acceleration after calibration at the nth moment, A n Represents the proportional coefficient matrix of the error calibration model at the nth moment, bias n represents the offset coefficient vector of the error calibration model at the nth moment;

[0093] S5. Input the calibrated acceleration and angular velocity into the invariant Kalman filter to obtain the positioning result of the mobile robot.

[0094] In one embodiment, the calibrated acceleration and angular velocity are input into an invariant Kalman filter to obtain the positioning result of the mobile robot, as follows:

[0095] S51, the state variable x at the nth moment in the unchanged Kalman filter n The definition is as follows:

[0096]

[0097] Where, Represents the rotation matrix from the world coordinate system to the IMU coordinate system at the nth moment, Indicates the velocity of the inertial measurement unit in the world coordinate system at the nth moment, Indicates the position of the inertial measurement unit in the world coordinate system at the nth moment, represents the gyroscope bias of the inertial measurement unit at the nth moment, represents the accelerometer bias of the inertial measurement unit at the nth moment, Represents the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system at the nth moment, Represents the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system at the nth moment.

[0098] S52. Establish a nonlinear state prediction equation:

[0099]

[0100] Where, Represents the state quantity predicted at the nth moment, x n-1 is the state variable at the n-1th moment, and f(·) is the dynamic model of the mobile robot, which is expressed as follows:

[0101]

[0102] in,

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110] Where, Represents the rotation matrix from the world coordinate system to the IMU coordinate system predicted at the nth moment, It represents the velocity of the inertial measurement unit in the world coordinate system predicted at the nth moment, represents the position of the inertial measurement unit in the world coordinate system predicted at the nth moment, represents the gyroscope bias of the inertial measurement unit predicted at time n, represents the accelerometer bias of the inertial measurement unit predicted at time n, Represents the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system predicted at the nth moment, It represents the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system predicted at the nth moment, dt is the time interval, and g is the gravitational acceleration of the current area.

[0111] S53. Establish the covariance equation:

[0112]

[0113] in,

[0114]

[0115]

[0116]

[0117] Where, is the predicted system state covariance matrix, F n For the nonlinear state prediction equation, n The Jacobian matrix, G n For the nonlinear state prediction equation The Jacobian matrix of the mobile robot, Q is the process noise, σ ω is the standard deviation of angular velocity, σ a is the standard deviation of acceleration, is the standard deviation of the angular velocity offset, is the standard deviation of the acceleration offset, is the standard deviation of the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system, is the standard deviation of the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system, P n-1 is the system state covariance matrix at the n-1th moment, 0 a×b Represents a row and b column 0 matrix, I d×d Represents the identity matrix of dimension d.

[0118] S54. Establish the system observation equation:

[0119]

[0120] Where y n is the system observation value at the nth moment, h(x n ) is the state variable x at the nth moment n The system observation equation is is the observed value in the horizontal direction at the nth moment, is the vertical observation value at the nth moment, h lat (x n ) is the system observation equation in the horizontal direction at the nth moment, h up (x n ) is the system observation equation in the vertical direction at the nth moment, is the horizontal observation noise at the nth moment, is the vertical observation noise at the nth moment, n n represents the observation noise at the nth moment, the horizontal direction represents the y-axis direction of the mobile robot carrier coordinate system, and the vertical direction represents the z-axis direction of the mobile robot carrier coordinate system;

[0121] And satisfy the following formula:

[0122]

[0123] Where, is the speed of the mobile robot at the nth moment, is the forward speed of the mobile robot at the nth moment, is the horizontal speed of the mobile robot at the nth moment, is the vertical velocity of the mobile robot at the nth moment, and the forward direction represents the x-axis direction of the mobile robot carrier coordinate system.

[0124] S55. Establish the Kalman gain equation:

[0125]

[0126] in,

[0127]

[0128] Where K n is the Kalman gain coefficient at the nth moment, H n is h(x n ), and R is the observation noise covariance matrix.

[0129] S56. Establish nonlinear state estimation equation:

[0130]

[0131] in,

[0132]

[0133] Where z n is the zero-speed pseudo-observation, is the state quantity predicted for the nth moment The observation equation of the system.

[0134] S57. Estimated covariance update equation:

[0135]

[0136] Where, P n is the system state covariance matrix at the nth moment, I 21×21 is the identity matrix of dimension 21.

[0137] S58, the output result x of step S56 n As the positioning result of the mobile robot at the nth moment.

[0138] Among them, if the mobile robot system adopts a wheeled mobile robot, based on the kinematic characteristics of the wheeled mobile robot, when the mobile robot is operating normally without slipping, its y-direction velocity in the mobile robot carrier coordinate system is and z-direction velocity Should be 0. Therefore, this application uses the y-direction velocity of the mobile robot in the mobile robot carrier coordinate system as and z-direction velocity As a zero-speed pseudo-observable.

[0139] This application adopts a mobile robot positioning method based on inertial measurement unit and deep learning. The error calibration model parameters of the IMU and the noise parameters of the invariant Kalman filter are obtained through deep learning model training. It can adapt to the motion characteristics of different mobile robots and perform real-time estimation of IMU noise and error, thereby obtaining higher-precision mobile robot positioning results, which helps to reduce the positioning cost of the mobile robot, is not easily affected by environmental interference, and can cope with accurate positioning under different light intensities and similar environmental structures.

[0140] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0141] The above-described embodiments merely represent specific and detailed examples of the present application and should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A mobile robot positioning method based on an inertial measurement unit and deep learning, wherein the mobile robot includes an inertial measurement unit, characterized in that: The mobile robot positioning method based on inertial measurement unit and deep learning includes the following steps: S1. Obtain the original acceleration and original angular velocity of the mobile robot using an inertial measurement unit; S2. Establish a deep learning model and train it, wherein the deep learning model includes a dilated convolutional neural network and a linear fully connected network connected in sequence, wherein the dilated convolutional neural network includes a dilated layer, a first convolutional module, a second convolutional layer module, a third convolutional module, a fourth convolutional module, and a fifth convolutional module connected in sequence, each of the convolutional modules includes a convolutional layer, and the linear fully connected network includes a fully connected layer and a Tanh activation function layer connected in sequence; S3. Input the original acceleration and original angular velocity into the trained deep learning model to obtain the error calibration parameters of the inertial measurement unit and the noise parameters of the invariant Kalman filter. The error calibration parameters of the inertial measurement unit include and The noise parameters of the invariant Kalman filter include r lat and r up ,in, They represent the proportional coefficients of the angular velocities of the x-axis, y-axis, and z-axis, respectively. In turn, they represent the proportional coefficients of the accelerations of the x-axis, y-axis, and z-axis. In turn, they represent the offset coefficients of the angular velocity of the x-axis, y-axis, and z-axis. Represents the offset coefficients of the x-axis, y-axis, and z-axis accelerations, r lat represents the velocity noise covariance in the y-axis direction, r up represents the velocity noise covariance in the z-axis direction; S4. Obtain calibrated acceleration and angular velocity using an error calibration model of the inertial measurement unit. The error calibration model is established as follows: in, Where, Represents the original IMU data at the nth moment, ω n represents the original angular velocity at the nth moment, a n represents the original acceleration at the nth moment, Represents the IMU data after calibration at the nth moment, represents the angular velocity after calibration at the nth moment, Indicates the acceleration after calibration at the nth moment, A n Represents the proportional coefficient matrix of the error calibration model at the nth moment, bias n represents the offset coefficient vector of the error calibration model at the nth moment; S5. Input the calibrated acceleration and angular velocity into the invariant Kalman filter to obtain the positioning result of the mobile robot.

2. The mobile robot positioning method based on inertial measurement unit and deep learning according to claim 1, characterized in that: The first convolution module, the second convolution layer module, the third convolution module and the fourth convolution module also include a BN layer, a GeLU activation function layer and a Dropout layer, and the convolution layer, the BN layer, the GeLU activation function layer and the Dropout layer in each of the convolution modules are connected in sequence.

3. The mobile robot positioning method based on inertial measurement unit and deep learning according to any one of claims 1 or 2, characterized in that: The input dimension of the convolution layer of the first convolution module is 6, the output dimension is 16, the convolution kernel size is 7, and the expansion coefficient is 1; the input dimension of the convolution layer of the second convolution layer module is 16, the output dimension is 32, the convolution kernel size is 7, and the expansion coefficient is 4; the input dimension of the convolution layer of the third convolution module is 32, the output dimension is 64, the convolution kernel size is 7, and the expansion coefficient is 16; the input dimension of the convolution layer of the fourth convolution module is 64, the output dimension is 128, the convolution kernel size is 7, and the expansion coefficient is 64; the input dimension of the convolution layer of the fifth convolution module is 128, the output dimension is 14, the convolution kernel size is 7, and the expansion coefficient is 1.

4. The mobile robot positioning method based on inertial measurement unit and deep learning according to claim 1, characterized in that: The calibrated acceleration and angular velocity are input into the invariant Kalman filter to obtain the positioning result of the mobile robot, as follows: S51, the state variable x at the nth moment in the unchanged Kalman filter n The definition is as follows: Where, Represents the rotation matrix from the world coordinate system to the IMU coordinate system at the nth moment, Indicates the velocity of the inertial measurement unit in the world coordinate system at the nth moment, Indicates the position of the inertial measurement unit in the world coordinate system at the nth moment, represents the gyroscope bias of the inertial measurement unit at the nth moment, represents the accelerometer bias of the inertial measurement unit at the nth moment, Represents the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system at the nth moment, Represents the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system at the nth moment; S52. Establish a nonlinear state prediction equation: Where, Represents the state quantity predicted at the nth moment, x n-1 is the state variable at the n-1th moment, and f(·) is the dynamic model of the mobile robot, which is expressed as follows: in, Where, Represents the rotation matrix from the world coordinate system to the IMU coordinate system predicted at the nth moment, It represents the velocity of the inertial measurement unit in the world coordinate system predicted at the nth moment, represents the position of the inertial measurement unit in the world coordinate system predicted at the nth moment, represents the gyroscope bias of the inertial measurement unit predicted at time n, represents the accelerometer bias of the inertial measurement unit predicted at time n, Represents the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system predicted at the nth moment, It represents the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system predicted at the nth moment, dt is the time interval, and g is the gravitational acceleration of the current area; S53. Establish the covariance equation: in, Where, is the predicted system state covariance matrix, F n For the nonlinear state prediction equation, n The Jacobian matrix, G n For the nonlinear state prediction equation The Jacobian matrix of the mobile robot, Q is the process noise, σ ω is the standard deviation of angular velocity, σ a is the standard deviation of acceleration, is the standard deviation of the angular velocity offset, is the standard deviation of the acceleration offset, is the standard deviation of the rotation matrix from the mobile robot carrier coordinate system to the IMU coordinate system, is the standard deviation of the displacement vector from the mobile robot carrier coordinate system to the IMU coordinate system, P n-1 is the system state covariance matrix at the n-1th moment, 0 a×b Represents a row and b column 0 matrix, I d×d represents the identity matrix of dimension d; S54. Establish the system observation equation: Where y n is the system observation value at the nth moment, h(x n ) is the state variable x at the nth moment n The system observation equation is is the observed value in the horizontal direction at the nth moment, is the vertical observation value at the nth moment, h lat (x n ) is the system observation equation in the horizontal direction at the nth moment, h up (x n ) is the system observation equation in the vertical direction at the nth moment, is the horizontal observation noise at the nth moment, is the vertical observation noise at the nth moment, n n represents the observation noise at the nth moment, the horizontal direction represents the y-axis direction of the mobile robot carrier coordinate system, and the vertical direction represents the z-axis direction of the mobile robot carrier coordinate system; And satisfy the following formula: Where, is the speed of the mobile robot at the nth moment, is the forward speed of the mobile robot at the nth moment, is the horizontal speed of the mobile robot at the nth moment, is the vertical velocity of the mobile robot at the nth moment, and the forward direction represents the x-axis direction of the mobile robot carrier coordinate system; S55. Establish the Kalman gain equation: in, Where K n is the Kalman gain coefficient at the nth moment, H n is h(x n ) is the Jacobian matrix, R is the observation noise covariance matrix; S56. Establish nonlinear state estimation equation: in, Where z n is the zero-speed pseudo-observation, is the state quantity predicted for the nth moment The system observation equation of S57. Estimated covariance update equation: Where, P n is the system state covariance matrix at the nth moment, I 21×21 is the identity matrix with dimension 21; S58, the output result x of step S56 n As the positioning result of the mobile robot at the nth moment.

Images