Geomagnetic road positioning method and system based on deep learning

Abstract

The invention provides a geomagnetic road positioning method and system based on deep learning, and belongs to the field of geomagnetic road positioning. The problem that when large-scale data and complex environments are processed through traditional geomagnetic positioning precision, positioning precision and robustness are insufficient is solved. According to the invention, by increasing the input dimension, 9-dimensional data covering the triaxial magnetic field intensity, the total magnetic field amount, the total magnetic field gradient, the course information and the triaxial geomagnetic rotation vector, the geomagnetic environment characteristics are comprehensively captured; the wavelet transform technology is introduced to pre-process the data, noise and burrs in geomagnetic waveforms are effectively filtered out, and the data quality is improved; and deep features in the geomagnetic data are extracted by using the CNN, and time sequence information is captured through a Bi-LSTM-LSTM framework, so that high-precision longitude and latitude positioning is realized. The method gives full play to the learning advantages of a deep learning model, and remarkably improves the robustness and accuracy of the system through the training of large-scale geomagnetic data.

Description

Technical Field

[0001] The present invention relates to the technical field of geomagnetic road positioning, and in particular to a geomagnetic road positioning method and system based on deep learning. Background Art

[0002] Geomagnetic navigation is a method of positioning using the Earth's magnetic field. The Earth's magnetic field has different strengths and directions in different geographical locations, and this spatial variation makes geomagnetic navigation possible. Due to the stability and global coverage of the geomagnetic field, geomagnetic navigation is particularly suitable for environments that cannot be covered by satellite signals, such as indoors, underground, and underwater. Traditional geomagnetic navigation mainly relies on the matching of geomagnetic maps with real-time geomagnetic data, and determines the position by comparing the correlation between the measured geomagnetic data and the known geomagnetic map. Therefore, the accuracy of traditional geomagnetic positioning is largely limited to the accuracy and resolution of the geomagnetic map and the interference of the local environment, and the accuracy and robustness of the algorithm are insufficient when processing large-scale data and complex environments. Summary of the invention

[0003] The technical problems to be solved by the present invention are:

[0004] In order to solve the problem that the accuracy of traditional geomagnetic positioning is limited by the accuracy and resolution of geomagnetic maps and the interference of local environment, and the positioning accuracy and robustness are insufficient when processing large-scale data and complex environments.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] The present invention provides a geomagnetic road positioning method based on deep learning, comprising the following steps:

[0007] S100, collecting original geomagnetic data in the offline training stage, collecting three-axis magnetic field intensity, total magnetic field, total magnetic field gradient, heading information and three-axis geomagnetic rotation vector data for the road to be located, and dividing the original geomagnetic data into a training set and a test set;

[0008] S200, performing data preprocessing on the training set in the original geomagnetic data collected in step S100, including downsampling the high-frequency magnetic field data, denoising by wavelet transform, dividing the geomagnetic sequence into a plurality of continuous subsequences by using a sliding window, and performing normalization processing, and finally forming a training data set;

[0009] S300, constructing a CNN-Bi-LSTM-LSTM deep learning model, and performing model training using the training set data set obtained in step S200 to obtain a trained deep learning model;

[0010] S400. In the online stage, the test set in step S100 is input into the deep learning model trained in step S300, the geomagnetic data of the test set is processed, and finally the predicted location information is output.

[0011] Further, in step S100, the total magnetic field, the total magnetic field gradient and the heading angle are calculated by formula (1):

[0012] F=[m x m y m z m a m g θαβγ] T

[0013]

[0014] Among them, the component of the geomagnetic field in three-dimensional space is m x 、m y and m z , respectively represent the components of the geomagnetic field in the north, east and vertically downward directions; m a and m g They represent the total magnetic field and the total magnetic field gradient respectively; θ is the heading angle, which represents the deflection angle of the current object relative to the north direction; α, β, and γ represent the geomagnetic rotation vector respectively.

[0015] Furthermore, in step S200, the collected high-frequency magnetic field is downsampled to obtain magnetic field data with equal time intervals and moderate frequencies.

[0016] Furthermore, in step S200, a sliding window of uniform size is used to divide the geomagnetic sequence into a plurality of continuous subsequences, and the plurality of continuous subsequences are a series of magnetic field sequences with equal time span.

[0017] Furthermore, in step S200, during normalization processing, the original geomagnetic data and the latitude and longitude data of the location in S100 are normalized to the range of [0,1] according to formula (2).

[0018]

[0019] Among them, min(s) and max(s) are the minimum and maximum values ​​of the geomagnetic sequence s, respectively.

[0020] Furthermore, in step S300, abnormal features are extracted from the multidimensional geomagnetic sequence of step S200 through the CNN layer, and the extracted features are input into the Bi-LSTM layer to predict the user's location through historical geomagnetic measurement data.

[0021] Furthermore, the last LSTM layer is used to extract enhanced features.

[0022] A geomagnetic road positioning system based on deep learning, the system has a program module corresponding to the above steps, and executes the steps in the above geomagnetic road positioning method based on deep learning when running.

[0023] A computer-readable storage medium stores a computer program, wherein the computer program is configured to implement the steps of a geomagnetic road positioning method based on deep learning when called by a processor.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention provides a geomagnetic road positioning method and system based on deep learning, which can comprehensively capture the network architecture of global magnetic field, local magnetic field and precise location characteristics, not only covering the traditional basic information such as axial magnetic field strength and total magnetic field, but also innovatively introducing the total magnetic field gradient, heading information and three-axis geomagnetic rotation vector data, a total of nine-dimensional data, which greatly enriches the dimension of the geomagnetic fingerprint library, so that the model can more comprehensively capture the subtle changes in the road environment; this innovation not only effectively reduces the common mismatching risk in traditional geomagnetic fingerprint matching, but also realizes the seamless conversion from magnetic field time series data to spatial features and then to precise location information, significantly improving the robustness and accuracy of geomagnetic positioning;

[0026] Compared with the traditional LSTM and CNN-LSTM algorithms, the single architecture of the traditional LSTM or CNN-LSTM is abandoned, and a more efficient network structure suitable for the characteristics of this task is adopted. The positioning accuracy of the present invention is greatly improved, and the error rate is greatly reduced. Downsampling, wavelet denoising, sliding window, CNN and other technologies are adopted to effectively reduce noise interference, improve the purity and continuity of data, and lay a solid foundation for subsequent positioning calculations, especially in complex road conditions such as turns. Its trajectory tracking is smoother and fits the actual path. This is because the outdoor road environment is complex and changeable, including different road materials, traffic flow and weather conditions. Vehicle geomagnetic positioning technology can adapt to these complex environments, provide stable and reliable positioning services, and show excellent robustness and accuracy; the efficiency of data collection is higher and the time cost is lower;

[0027] By reconstructing the network structure and optimizing the geomagnetic signal feature extraction and position mapping mechanism, it not only opens up new ideas for vehicle positioning technology, but also provides strong technical support for applications in smart cities, intelligent transportation systems and other fields. With the continuous optimization of algorithms and the popularization of smart devices, the application prospects of this invention will be broader and can cover the entire urban transportation network, which is expected to bring revolutionary changes to vehicle navigation, traffic management, urban planning and other aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of a geomagnetic road positioning method based on deep learning in an embodiment of the present invention;

[0029] Figure 2 It is a framework diagram of a geomagnetic road positioning method based on deep learning in an embodiment of the present invention;

[0030] Figure 3 1 is a network structure diagram of a one-dimensional convolutional neural network in an embodiment of the present invention;

[0031] Figure 4 This is a structural diagram of a basic unit of an LSTM network in an embodiment of the present invention;

[0032] Figure 5 Schematic diagram of a Bi-LSTM network in an embodiment of the present invention;

[0033] Figure 6 A comparison diagram of longitude and latitude trajectories of different models in an embodiment of the present invention;

[0034] Figure 7 This is a physical picture of the vehicle used for collecting original geomagnetic data in the comparative experiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0036] Specific implementation plan 1: Combine Figures 1 to 5 As shown, the present invention provides a geomagnetic road positioning method based on deep learning, comprising the following steps:

[0037] S100, determining the original geomagnetic data to be collected,

[0038] The algorithm of the present invention is divided into two stages: offline training and online positioning. In order to extract more geomagnetic features for positioning, multiple geomagnetic data are used as input. Therefore, in the offline training stage, the vehicle collects a large amount of three-axis magnetic field intensity, total magnetic field, total magnetic field gradient, heading information and three-axis geomagnetic rotation vector data of the predicted positioning area along the road, wherein the total magnetic field, total magnetic field gradient and heading angle are calculated by formula (1).

[0039] F=[m x m y m z m a m g θαβγ] T

[0040]

[0041] Among them, the component of the geomagnetic field in three-dimensional space is m x 、m y and m z , respectively represent the components of the geomagnetic field in the north, east and vertically downward directions; m a and m g They represent the total magnetic field and the total magnetic field gradient respectively; θ is the heading angle (i.e., azimuth), which indicates the deflection angle of the current object relative to the north direction, usually defined as clockwise; α, β, and γ are the geomagnetic rotation vectors respectively;

[0042] S200, preprocessing the original geomagnetic data obtained in step S100,

[0043] The present invention can use a smart phone or a device with a magnetic sensor to collect data. In order to achieve more accurate positioning, the collected data needs to be preprocessed, including:

[0044] S210, considering that the sampling frequency of the magnetic sensor is relatively high, down-sampling the collected high-frequency magnetic field to obtain magnetic field data with equal time intervals and moderate frequencies;

[0045] S220, when collecting data, the slight shaking / movement of the user's mobile phone may cause noise. To prevent the noise from adversely affecting the positioning prediction, the wavelet transform denoising technology is used to remove the noise data that reduces the positioning performance;

[0046] S230, since it is very cumbersome and time-consuming to extract features from a long input sequence, a sliding window of uniform size is used to divide the geomagnetic sequence into several continuous subsequences to ensure that the number of magnetic field data points contained in each magnetic field sequence is equal, that is, a series of magnetic field sequences with equal time spans are generated;

[0047] S240, in order to avoid the influence of too large difference in absolute value of magnetic field intensity on training, the original geomagnetic data and longitude and latitude data of the position corresponding to each point in each sequence in step S100 are normalized to the range of [0,1] according to formula (2), and a large number of complete magnetic field-position sequences with equal time span are constructed to form a training data set for training CNN-LSTM-Bi-LSTM;

[0048]

[0049] Among them, min(s) and max(s) are the minimum and maximum values ​​of the geomagnetic sequence s, respectively;

[0050] S300: construct a CNN-LSTM-Bi-LSTM deep learning model, and use the training set data obtained in step S240 to perform model training.

[0051] Combination Figure 3 The figure shows the structure of an existing one-dimensional convolutional neural network (CNN). CNN can automatically learn and extract useful features from raw data without manually designing feature extractors, which makes CNN more flexible and effective in dealing with complex problems.

[0052] Deep learning methods achieve more accurate positioning by learning more extensive abstract and high-level features from geomagnetic sequences; Long Short-Term Memory (LSTM) is an improved recurrent neural network (RNN) with better long-term memory, which can effectively convey and represent long sequence information without causing long-term useful information to be forgotten; LSTM network mainly consists of 4 parts, namely forget gate, input gate, candidate memory and output gate. This design ensures memory update, combined with Figure 4 The figure shows the structure of the basic unit of the LSTM network. In geomagnetic positioning, the LSTM network can be used to model the spatiotemporal characteristics and dynamic changes of the geomagnetic field, that is, to predict the user's location based on historical geomagnetic measurement data.

[0053] In the LSTM network, by inputting x into the forget gate ft t With the state memory unit S t-1 、Intermediate output h t-1 Together determine the forgotten part of the state memory unit; x in the input gate it t After the changes of σ function and tanh function respectively, the vector retained in the state memory unit is jointly determined; the intermediate state unit outputs h t The updated candidate memory cell S t With output o t Jointly determine, the calculation formula is shown in formula (3);

[0054]

[0055] Among them, f t is the output vector of the sigmoid neural layer, W f 、b f is the parameter of the sigmoid function at this stage, h t-1 is the output at the previous moment, x t is a new input; i t Indicates the value to be updated in the unit state, W i 、b i Represents the parameters of the sigmoid function at this stage; Indicates the new data that needs to be added to the cell state, W S 、b S Represents the parameters of the tanh function; S t-1 is the state of the unit at the previous moment, S t is the new unit status after updating; o t Indicates the judgment condition of the output gate, W o 、b o Indicates the parameters of the sigmoid function at this stage, h t Indicates the final output.

[0056] Based on the conventional LSTM network, the Bi-LSTM network introduces a bidirectional processing mechanism, which includes two independent LSTM layers, one for processing the forward sequence (from front to back) and the other for processing the reverse sequence (from back to front). Finally, the output of the Bi-LSTM network is concatenated from the hidden states in two directions. The above processing allows the model to simultaneously obtain the geomagnetic information before and after the current time. The principle diagram is shown in the figure below. Figure 5 As shown;

[0057] In the model proposed in the present invention, abnormal features are extracted from the multidimensional geomagnetic sequence through the CNN layer, and the extracted features are input into the Bi-LSTM layer, and the user's location is predicted through historical geomagnetic measurement data; in order to extract more subsequence context information, the last LSTM layer is applied to extract enhanced features to achieve more accurate prediction; finally, two fully connected layers are applied to output the predicted longitude and latitude respectively; in the model training stage, the preprocessed feature data and location data are input into the hierarchical deep learning network model for training; after completing the above training stage, a trained model can be obtained;

[0058] S400, in the online stage, the input is a magnetic field sequence with the same time span as the offline stage sequence. The trained model is used to process and analyze the geomagnetic data of the test set, and the predicted position information is output, which can finally be compared with the actual position output by GPS.

[0059] Specific implementation scheme 2: The present invention provides a geomagnetic road positioning system based on deep learning, which has a program module corresponding to the above steps and executes the steps in the above-mentioned geomagnetic road positioning method based on deep learning during operation.

[0060] The other combinations and connection relationships of this embodiment are the same as those of the first embodiment.

[0061] Specific implementation scheme three: The present invention provides a computer-readable storage medium, characterized in that: the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of a geomagnetic road positioning method based on deep learning when called by a processor.

[0062] The other combinations and connection relationships of this embodiment are the same as those of the first embodiment.

[0063] Comparative experiment

[0064] Experimental scenario: Place the smartphone on Figure 7 In the car shown, the relative posture and position of the mobile phone on the car are kept unchanged during driving. The data collection site is a circle of roads around the Harbin Institute of Technology (Weihai) Library (total length is 525m). The magnetic collection equipment and position collection equipment are IQOO Neo8 model mobile phones. The collection frequency is 10HZ. The collection method is to drive the test vehicle along the road in the test site for many times. The route goes around the library four times in total, of which the data collected in three circles are divided into training sets, and the data in one circle are divided into test sets. The present invention compares the path obtained by training with the path obtained by the actual GPS. The matching accuracy is within 5m.

[0065] Model parameters: This model sets the feature dimension to 64 and the LSTM hidden layer to 3. In addition, the batch size is set to 64 during training, the sequence sliding window step is 40, and the training is repeated for 500 rounds. The Adam optimizer is used as the parameter trainer, and the learning rate is 0.0001. The model is implemented using the PyTorch library, and the detailed configuration is shown in Table 1.

[0066] Table 1 Experimental configuration

[0067]

[0068] Comparison and evaluation of algorithm performance in actual environment:

[0069] The present invention uses a smart phone to process the geomagnetic data collected from the actual environment, inputs it into the improved neural network model for training, and compares the traditional LSTM algorithm, the CNN-LSTM algorithm and the algorithm in this paper with the actual GPS path. The obtained path is as follows Figure 6 shown.

[0070] Table 2 Comparison of positioning performance of different model algorithms

[0071]

[0072] It can be seen from Table 2 that the positioning accuracy of the algorithm of the present invention is the highest compared with the LSTM and CNN-LSTM algorithms, and can basically keep consistent with the actual trajectory. The positioning accuracy of the LSTM and CNN-LSTM algorithms has declined significantly, among which the LSTM algorithm has the largest error and has a divergent trend. From a numerical point of view, the mean square error of the distance of the algorithm of the present invention is improved by 82.17% and 43.95% respectively compared with the LSTM and CNN-LSTM algorithms. Although the training time of the algorithm of the present invention is the longest, considering its significantly improved positioning accuracy, this training time is acceptable. From Figure 6 It can be seen that at the corners, the trajectory trend of the present invention is smoother and most consistent with the actual trajectory, indicating that under the same parameters, the algorithm of the present invention has better robustness and higher accuracy.

[0073] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A geomagnetic road positioning method based on deep learning, characterized in that: The following steps are involved: S100, collecting original geomagnetic data in the offline training stage, collecting three-axis magnetic field intensity, total magnetic field, total magnetic field gradient, heading information and three-axis geomagnetic rotation vector data for the road to be located, and dividing the original geomagnetic data into a training set and a test set; S200, performing data preprocessing on the training set in the original geomagnetic data collected in step S100, including downsampling the high-frequency magnetic field data, denoising by wavelet transform, dividing the geomagnetic sequence into a plurality of continuous subsequences by using a sliding window, and performing normalization processing, and finally forming a training data set; S300, constructing a CNN-Bi-LSTM-LSTM deep learning model, and performing model training using the training set data set obtained in step S200 to obtain a trained deep learning model; S400. In the online stage, the test set in step S100 is input into the deep learning model trained in step S300, the geomagnetic data of the test set is processed, and finally the predicted location information is output.

2. The method for geomagnetic road positioning based on deep learning according to claim 1, characterized in that: In step S100, the total magnetic field, the total magnetic field gradient and the heading angle are calculated by formula (1): Among them, the component of the geomagnetic field in three-dimensional space is m x 、m y and m z , respectively represent the components of the geomagnetic field in the north, east and vertically downward directions; m a and m g They represent the total magnetic field and the total magnetic field gradient respectively; θ is the heading angle, which represents the deflection angle of the current object relative to the north direction; α, β, and γ represent the geomagnetic rotation vector respectively.

3. The method for geomagnetic road positioning based on deep learning according to claim 1, characterized in that: In step S200, the collected high-frequency magnetic field is downsampled to obtain magnetic field data with equal time intervals and moderate frequencies.

4. The method for geomagnetic road positioning based on deep learning according to claim 3, characterized in that: In step S200, a sliding window of uniform size is used to divide the geomagnetic sequence into a plurality of continuous subsequences, wherein the plurality of continuous subsequences are a series of magnetic field sequences with equal time span.

5. The method for geomagnetic road positioning based on deep learning according to claim 4, characterized in that: In step S200, during normalization processing, the original geomagnetic data and the latitude and longitude data of the location in S100 are normalized to the range of [0,1] according to formula (2). Among them, min(s) and max(s) are the minimum and maximum values ​​of the geomagnetic sequence s, respectively.

6. The method for geomagnetic road positioning based on deep learning according to claim 1, characterized in that: In step S300, abnormal features are extracted from the multidimensional geomagnetic sequence of step S200 through the CNN layer, and the extracted features are input into the Bi-LSTM layer to predict the user's location through historical geomagnetic measurement data.

7. The method for geomagnetic road positioning based on deep learning according to claim 6, characterized in that: The final LSTM layer is used to extract enhanced features.

8. A geomagnetic road positioning system based on deep learning, characterized in that: The system has a program module corresponding to the steps of any one of claims 1 to 7 above, and executes the steps in the above-mentioned deep learning-based geomagnetic road positioning method when running.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the deep learning-based geomagnetic road positioning method described in any one of claims 1 to 7 when called by a processor.

Images