A federated learning inertial navigation fusion positioning method

Abstract

A federated learning inertia navigation fusion positioning method uses a federated edge learning method to predict speed, obtains positioning information, and then uses single-anchor ranging to perform segmented fusion, comprising the following steps: inertial navigation data input; coordinate mapping conversion; federated edge model and training; coordinate mapping conversion; segmented fusion positioning.The application innovatively uses a federated learning data-driven inertia navigation method, which can solve the problems caused by centralized deep learning methods, and simultaneously adopts an edge single-anchor ranging segmented fusion positioning method, which can greatly improve the positioning accuracy of the system at low cost.

Description

Technical Field

[0001] This invention relates to the field of indoor positioning technology. Background Technology

[0002] Inertial navigation is a crucial component of indoor positioning, but widely used commercial inertial navigation devices suffer from significant deviations. Traditional dead reckoning methods exhibit substantial positioning errors over short periods, failing to meet indoor positioning requirements. Deep learning methods have been employed to improve inertial navigation positioning accuracy. Publication number CN107655472B utilizes a deep learning-based error compensation method for high-precision inertial navigation devices, improving navigation accuracy to some extent at the error level. Publication number CN108168548B invented a pedestrian inertial navigation system assisted by machine learning algorithms and models. By training sensor information, it enables real-time fault detection of the pedestrian navigation system, reducing the failure rate to some extent. Publication number CN112595313A inputs the position and velocity acquired by inertial navigation devices into a machine learning model to establish auxiliary navigation information, avoiding the divergence of positioning errors over long periods and improving positioning accuracy. However, these methods raise security and privacy concerns regarding inertial navigation data, and centralized data collection is difficult, resulting in low training efficiency. Summary of the Invention

[0003] This invention proposes for the first time an inertial navigation and positioning framework based on federated learning, and innovatively proposes a segmented fusion positioning method to solve the problem that a single anchor point cannot provide full coverage in large scenes.

[0004] This invention addresses the problems of security and privacy, difficulty in data acquisition, and low training efficiency in existing centralized deep learning-based inertial navigation methods. It proposes a federated learning-based inertial navigation method that solves these problems. Furthermore, it employs a segmented fusion positioning method at the edge ranging end, which can greatly improve the positioning accuracy of the system at a low cost.

[0005] The technical solution of this invention is as follows:

[0006] A federated habitual navigation fusion positioning method is characterized by using a federated edge learning method to predict velocity, obtain positioning information, and then using single-anchor-point ranging for segmented fusion.

[0007] A federated habitual navigation fusion positioning method, characterized by comprising:

[0008] Step 1: Input Inertial Navigation Data

[0009] Pedestrians use their mobile phones to acquire raw inertial navigation data, which includes X, Y, and Z axis data from gyroscopes and accelerometers.

[0010] Step 2, Coordinate Mapping and Transformation

[0011] The specific calculation method for transforming the vehicle coordinate system to HACF (Heading Unknown Coordinate System) is as follows:

[0012]

[0013] In the formula R represents the data in HACF and the carrier coordinate system, respectively. grv It is the game rotation vector obtained by the mobile phone. The data in the carrier coordinate system can be converted to HACF through the game rotation vector.

[0014] Step 3, Federated Edge Model and Training

[0015] The federated learning model system consists of a central server and N clients, where each client k (k = 1, 2, 3...N) maintains a local dataset D. k The size of the local dataset is |D k |=n k Meanwhile, the central server also holds a copy of the test dataset D. t This is used to test the results of the current model training; at the beginning of the t-th round, the central server will randomly select customers with a proportion of C, denoted as |S|, to participate in the model training process. The central server will then assign global model weights w to these customers. t Send to the selected client, upon receiving w t Afterwards, each selected client passes through Get the updated local model parameters Where f(D) k ,w t () is a loss function; it updates the local deep learning model. Uploaded to the central server, the central server uses it to update the global model parameters; after obtaining a new global model each time, it is processed through dataset D. t The training is completed when the accuracy meets the requirements or the number of rounds reaches a pre-set threshold.

[0016] Step 4: Coordinate Mapping and Transformation

[0017] The two-dimensional velocity vector in the unknown course coordinate system (HACF) in step 3 needs to be transformed to the world coordinate system in step 5;

[0018] Step 5: Segmented fusion positioning

[0019] After obtaining the optimal federated edge model through training in step 3, the input is transformed into three-axis data from the gyroscope and accelerometer in step 2. The output of the optimal federated edge model is the two-dimensional vector velocity v in a heading-unknown coordinate system. x and v yThis velocity vector is then transformed to the world coordinate system, and the velocity is integrated. Where k represents time, location information can be obtained. This location information is used for mathematical modeling in step 5. The modeling also requires distance measurement information from the edge, i.e., Z(k) = |P(k) - P BAS |

[0020] In step 3, FedAVG is used as the algorithm for training the federated learning model.

[0021] In step 3, equation (1) is used to represent the update of the global model parameters:

[0022]

[0023] In step 3, the deep learning model used for training is a 1D version of the ResNet-18 architecture, with a fully connected layer of 512 units added at the end to regress the 2D velocity vector:

[0024]

[0025] In step 4, the data in the world coordinate system and HACF are first converted to the polar coordinate system, and then the statistical average of the difference between the polar angles during this period is calculated, thereby converting the predicted values ​​in HACF to the world coordinate system.

[0026] The segmented fusion in step 5 combines the model output prediction results from federated learning in step 3 with the ranging results from the edge ranging end.

[0027] like Figure 4 The specific method is as follows:

[0028] Step 5.1 Obtain the edge ranging value of the UWB hardware;

[0029] Step 5.2 Input the prediction results output by the federated learning model in Step 3 and the ranging results from the edge ranging end in Step 5.1, and proceed to Step 5.3;

[0030] Step 5.3 Determine whether the pedestrian is within the range of the ranging value based on the ranging threshold M:

[0031] If the distance is outside the range (the distance measured is greater than M, where M is set to 20 meters), the location information is predicted by using the input inertial navigation data and the output of the two-dimensional velocity integral from the federated edge model.

[0032] If the distance is within the range of the measured value, then particle filtering is used to fuse the measured value and the predicted value.

[0033] In step 5.2, when using particle filter fusion, a mathematical model of the system needs to be established as shown in equation (3). The first three equations are state equations established by combining the two-dimensional velocity vector output by the federated edge model, and the last equation is the observation equation established based on UWB hardware ranging.

[0034]

[0035] In the formula, x(k) and y(k) are the cumulative distances along the x-axis and y-axis, respectively, representing the location information. P(k) and Z(k) are the pedestrian's position coordinates at time k and the distance measured from the UWB base station anchor point to the mobile phone, respectively. BAS This refers to the location of the anchor point for a single UWB base station. W k V k These are process noise and observation noise, respectively. First, initialize the particle set, {(x i ,y i Let |i=1,2,...,N}, where N is the number of particles, and x,y are the two-dimensional positions of the particles. Then, update the position P(k) to P(k+1) according to equation (3). For each particle, update the particle weight according to equations (3)(4)(5). Then, make resampling reduce particle degradation.

[0036]

[0037]

[0038] This invention innovatively uses a federated learning data-driven inertial navigation method, which can solve the problems caused by centralized deep learning methods. At the same time, it adopts a single-anchor-point ranging and segmented fusion positioning method at the edge, which can greatly improve the positioning accuracy of the system at low cost. Attached Figure Description

[0039] Figure 1 Coordinate system diagram;

[0040] Figure 2 Block diagram of the present invention;

[0041] Figure 3 Diagram of a federated learning framework driven by inertial navigation data;

[0042] Figure 4 Segmented fusion positioning flowchart;

[0043] Figure 5 Comparison chart of segmented fusion localization results in federated learning. Detailed Implementation

[0044] This invention innovatively proposes an inertial navigation fusion positioning method based on federated learning, which mainly solves the problems of privacy and security, difficulty in data collection, and low training efficiency in the data-driven process of inertial navigation. At the same time, it innovatively proposes a single anchor point segmented fusion positioning method to ensure positioning accuracy.

[0045] The coordinate system involved in the system is as follows: Figure 1 As shown.

[0046] The raw data acquired via the mobile phone in step 1 belongs to the carrier coordinate system. It needs to be transformed to the heading-agnostic coordinate system (HACF) required in step 3 before being input into the federated learning model. It's called a heading-agnostic coordinate system because the Z-axis is aligned with the direction of gravity, but the X-axis direction is uncertain. Finally, the federated edge model, after training, regresses to output a two-dimensional vector velocity v. x and v y This data is also in a course-unknown coordinate system (HACF). However, the fusion positioning in step 5 needs to be performed in the world coordinate system, so the data output by the model needs to be converted to the world coordinate system.

[0047] Figure 2 This is the framework design for the present invention. The specific implementation steps of the present invention are:

[0048] Step 1: Input inertial navigation data.

[0049] The inertial navigation data includes X, Y, and Z axis data from the gyroscope and accelerometer. This raw data is obtained through the pedestrian's mobile phone. The sampling frequency can be set according to the actual application. Here it is set to 200Hz.

[0050] Step 2: Coordinate mapping transformation.

[0051] Step 2 here involves transforming the carrier coordinate system to HACF. The specific calculation method is as follows:

[0052]

[0053] In the formula R represents the data in HACF and the carrier coordinate system, respectively. grv It is the game rotation vector obtained by the mobile phone. The data in the carrier coordinate system can be converted to HACF through the game rotation vector.

[0054] Step 3: Federation Edge Model and Training.

[0055] This invention employs FedAVG as the algorithm for training federated learning models. The federated learning model system consists of a central server and N clients, where each client k (k = 1, 2, 3...N) maintains a local dataset D. kThe size of the local dataset is |D k |=n k Meanwhile, the central server also holds a copy of the test dataset D. t This is used to test the results of the current model training. At the beginning of the t-th round, the central server will randomly select customers with a proportion of C, denoted as |S|, to participate in the model training process. The central server will then assign global model weights w... t Send to the selected client, upon receiving w t Afterwards, each selected client passes through Get the updated local model parameters Where f(D) k ,w t () is a loss function. It updates the local deep learning model. The data is uploaded to a central server, which uses it to update the global model parameters. Each time a new global model is obtained, it is processed through dataset D. t The training is completed when the accuracy meets the requirements or the number of rounds reaches a pre-set threshold.

[0056]

[0057] The deep learning model used for training in this invention is a 1D version of the ResNet-18 architecture, with a fully connected layer of 512 units added at the end to regress the 2D velocity vector.

[0058]

[0059] Figure 3 This is the model framework we are building. In the diagram, a represents sending the global model to the client, b represents updating the deep learning model using local data, c represents sending the local deep learning model to the central server, and d represents updating the global model.

[0060] Step 4: Coordinate mapping and transformation.

[0061] The two-dimensional velocity vector in the unknown heading coordinate system (HACF) of step 3 needs to be transformed to the world coordinate system of step 5. This invention utilizes the relatively stable characteristic of inertial navigation data in the initial time. First, the data in the world coordinate system and HACF are transformed to the polar coordinate system. Then, the statistical average of the difference in polar angles during this period is calculated, thereby transforming the predicted value in HACF to the world coordinate system.

[0062] Step 5: Segmented fusion positioning.

[0063] After obtaining the optimal federated edge model through training in step 3, the input is transformed into three-axis data from the gyroscope and accelerometer in step 2. The output of the optimal federated edge model is the two-dimensional vector velocity v in a heading-unknown coordinate system. x and v y This velocity vector is then transformed to the world coordinate system, and the velocity is integrated. Where k represents time, location information can be obtained. This location information is used for mathematical modeling in step 5. The modeling also requires distance measurement information from the edge, i.e., Z(k) = |P(k) - P BAS |

[0064] The segmented fusion in step 5 combines the model output prediction results from federated learning in step 3 with the ranging results from the edge ranging end. Figure 4 This is a flowchart of segmented fusion. This invention uses a single anchor point (e.g., Figure 5 Ranging is performed using base stations (in the network), which can greatly save hardware costs. The hardware for the ranging end can be selected according to the specific application scenario; here, we use ultra-wideband (UWB) as the edge ranging end. However, the coverage of a single anchor point indoors is limited. Therefore, this invention uses a segmented fusion positioning method to cover a larger indoor scene.

[0065] like Figure 4 The specific method is as follows:

[0066] Step 5.1 Obtain the edge ranging value of the UWB hardware;

[0067] Step 5.2 Input the prediction results output by the federated learning model in Step 3 and the ranging results from the edge ranging end in Step 5.1, and proceed to Step 5.3;

[0068] Step 5.3 Determine whether the pedestrian is within the range of the ranging value based on the ranging threshold M:

[0069] If the distance is outside the range (the distance measured is greater than M, where M is set to 20 meters), the location information is predicted by using the input inertial navigation data and the output of the two-dimensional velocity integral from the federated edge model.

[0070] If the distance is within the range of the measured value, then particle filtering is used to fuse the measured value and the predicted value.

[0071] In step 5.2, when using particle filter fusion, a mathematical model of the system needs to be established as shown in equation (3). The first three equations are state equations established by combining the two-dimensional velocity vector output by the federated edge model, and the last equation is the observation equation established based on UWB hardware ranging.

[0072]

[0073] In the formula, x(k) and y(k) are the cumulative distances along the x-axis and y-axis, respectively, representing the location information. P(k) and Z(k) are the pedestrian's position coordinates at time k and the distance measured from the UWB base station anchor point to the mobile phone, respectively. BAS This refers to the location of the anchor point for a single UWB base station. W k V k These are process noise and observation noise, respectively. First, initialize the particle set, {(x i ,y i Let |i=1,2,...,N}, where N is the number of particles, and x,y are the two-dimensional positions of the particles. Then, update the position P(k) to P(k+1) according to equation (3). For each particle, update the particle weight according to equations (3)(4)(5). Then, make resampling reduce particle degradation.

[0074]

[0075]

[0076] Figure 5 This is the result image after segmented fusion positioning. A total of 303 meters were walked indoors. Compared with the centralized method, the segmented fusion positioning method based on federated learning and navigation proposed in this invention improves positioning performance by 47%.

[0077] Technical effects:

[0078] This invention can solve the problems of cumbersome indoor UWB deployment, difficult maintenance, and high hardware costs. It is suitable for low-cost indoor positioning locations where the accuracy requirements are not particularly high. Walking around a square with a side length of 8m for 4 circles, a distance of 128 meters, the cumulative deviation is 1.8 meters, which can meet general indoor positioning needs.

Claims

1. A method for fedified habitual navigation fusion positioning, characterized in that, The method uses federated edge learning to predict velocity, obtain location information, and then uses single-anchor ranging for segmented fusion. Step 1: Input Inertial Navigation Data Pedestrians' mobile phones acquire raw inertial navigation data, which includes X, Y, and Z axis data from gyroscopes and accelerometers. Step 2, Coordinate Mapping and Transformation The carrier coordinate system is transformed to HACF (Heading Unknown Frame Coordinate System), and the calculation method is as follows: In the formula , These represent the data in HACF and the carrier coordinate system, respectively. It is the game rotation vector obtained by the mobile phone, which is converted to HACF through the data in the game rotation vector carrier coordinate system; Step 3, Federated Edge Model and Training The federated learning model system consists of a central server and N clients, where each client... Maintain a local dataset The size of the local dataset is Meanwhile, the central server also holds a copy of the test dataset. This is used to test the results of the current model training; at the beginning of the t-th round, the central server will randomly select customers with a proportion of C, denoted as |S|, to participate in the model training process. The central server will then adjust the global model weights. Send to the selected client, upon receiving Afterwards, each selected client passes through Get the updated local model parameters ,in It is a loss function; it updates the local deep learning model. Uploaded to the central server, which then uses it to update the global model parameters; After obtaining a new global model, it is used in the dataset. The training is completed when the accuracy meets the requirements or the number of rounds reaches a pre-set threshold. Step 4: Coordinate Mapping and Transformation The two-dimensional velocity vector in the unknown heading coordinate system HACF in step 3 needs to be transformed to the world coordinate system in step 5. Step 5: Segmented fusion positioning After obtaining the optimal federated edge model through training in step 3, the input is converted into three-axis data from the gyroscope and accelerometer in step 2. The output of the optimal federated edge model is the two-dimensional vector velocity in a heading-unknown coordinate system. and This velocity vector is then transformed to the world coordinate system, and the velocity is integrated. Where k represents time, the location information is obtained. This location information is used for mathematical modeling in step 5. The modeling also requires distance measurement information from the edge. Where P(k) and Z(k) are the pedestrian's position coordinates at time k and the distance measured from the UWB base station anchor point to the mobile phone, respectively. It is the location of the anchor point of a single UWB base station.

2. The federated habitual navigation fusion positioning method as described in claim 1, characterized in that, In step 3, FedAVG is used as the algorithm for training the federated learning model.

3. The federated habitual navigation fusion positioning method as described in claim 1, characterized in that, In step 3, equation (1) is used to represent the update of the global model parameters: 。 4. The federated habitual navigation fusion positioning method as described in claim 1, characterized in that, In step 3, the deep learning model used for training is a 1D version of the ResNet-18 architecture, with a fully connected layer of 512 units added at the end to regress the 2D velocity vector: （2）。 5. The federated habitual navigation fusion positioning method as described in claim 1, characterized in that, In step 4, the data in the world coordinate system and HACF are first converted to the polar coordinate system, and then the statistical average of the difference between the polar angles during this period is calculated, thereby converting the predicted values ​​in HACF to the world coordinate system.

6. The federated habitual navigation fusion positioning method as described in claim 1, characterized in that, The segmented fusion in step 5 is to fuse the model output prediction results of federated learning in step 3 and the ranging results of the edge ranging end. Step 5.1 Obtain the edge ranging value of the UWB hardware; Step 5.2 Input the prediction results output by the federated learning model in Step 3 and the ranging results from the edge ranging end in Step 5.1, and proceed to Step 5.3; Step 5.3 Determine whether the pedestrian is within the range of the ranging value based on the ranging threshold M: If it is not within the range, the location information is predicted by using the input inertial navigation data and the output of the two-dimensional velocity integral of the federated edge model. If the distance is within the range of the measured value, then particle filtering is used to fuse the measured value and the predicted value.

7. The federated habitual navigation fusion positioning method as described in claim 6, characterized in that, In step 5.2, when using particle filter fusion, a mathematical model of the system needs to be established as shown in equation (3). The first three equations are state equations established by combining the two-dimensional velocity vector output by the federated edge model, and the last equation is the observation equation established based on UWB hardware ranging: （3） In the formula, x(k) and y(k) are the cumulative distances along the x-axis and y-axis, respectively, representing the position information. , These are process noise and observation noise, respectively; first, initialize the particle set. N is the number of particles, x and y are the two-dimensional positions of the particles, and then the position P(k) is updated to P(k+1) according to equation (3). For each particle, the weight of the particle is updated according to equations (3), (4), and (5). Then, resampling is used to reduce particle degradation. （4） （5）。

Images