CN-121996903-A - Training method and positioning method for deep learning model

CN121996903ACN 121996903 ACN121996903 ACN 121996903ACN-121996903-A

Abstract

The application discloses a training method of a deep learning model, which comprises the steps of obtaining an observed pseudo range and a theoretical pseudo range of a target GNSS satellite at a target historical moment, and determining a target measurement noise value as a label (label) of a training sample, namely a true value, based on the observed pseudo range and the theoretical pseudo range. And then training the deep learning model by using the training sample and taking the target measured noise value as a loss true value until the predicted measured noise value of the target GNSS satellite output by the deep learning model meets the preset condition with the loss value output by the loss function corresponding to the target measured noise value, thereby completing the training of the deep learning model. Because the theoretical pseudo range (pseudo range true value) is calculated based on the positioning result that the positioning precision meets the preset precision requirement, the measurement noise true value is constructed based on the observed pseudo range of the target GNSS satellite at the target historical moment and the theoretical pseudo range, reliable true value data can be provided for training the deep learning model, and the accuracy of the measurement noise output by the trained deep learning model is ensured.

Inventors

GAO XILE
FANG XING
LUO LEIGANG
YUAN HELIANG
HE YULI
YANG FAN

Assignees

北京高德云图科技有限公司

Dates

Publication Date: 20260508
Application Date: 20241101

Claims (10)

1. A method of training a deep learning model, the method comprising: acquiring an observed pseudo range and a theoretical pseudo range of a target GNSS satellite at a target historical moment; Determining a target measurement noise value according to the observed pseudo range and the theoretical pseudo range of the target GNSS satellite; training a deep learning model by using the target measurement noise value as a loss true value until the predicted measurement noise value of the target GNSS satellite output by the deep learning model and the loss value output by a loss function corresponding to the target measurement noise value meet a preset condition, wherein the training sample is a set of any satellite observation at any historical moment, the input of each training sample at least comprises target GNSS satellite characteristics and all GNSS satellite characteristics, and the output of each training sample is the measurement noise value of the target GNSS satellite.
2. The method of claim 1, wherein the observed pseudoranges of the target GNSS satellite at the target historic time are from a target device, the method further comprising: based on the model parameters of the target equipment, acquiring the model characteristics of the target equipment; The input of each training sample further comprises the model features respectively spliced with the target GNSS satellite features and all GNSS satellite features.
3. The method according to claim 2, wherein the deep learning model comprises a embedding layer, and the obtaining the model characteristics of the target device is specifically: And taking the model parameters of the target equipment as the input of the embedding layers, and extracting the model characteristics of the target equipment through the embedding layers.
4. The method according to claim 3, wherein training the deep learning model with the target measured noise as a loss truth value by using a training sample comprises: Inputting the target GNSS satellite characteristics and the model characteristics into a first full-connection layer of the deep learning model to obtain observation characteristics of the target GNSS satellite; inputting all GNSS satellite characteristics and model characteristics into a convolution layer and a pooling layer of the deep learning model to obtain environmental characteristics; And inputting the observation characteristics and the environmental characteristics of the target GNSS satellite into a splicing layer of the deep learning model, and obtaining a predicted measurement noise value of the target GNSS satellite through a second full-connection layer connected with the splicing layer in the deep learning model.
5. The method of any of claims 1-4, wherein determining the target measurement noise value based on the observed pseudorange and the theoretical pseudorange of the target GNSS satellite comprises: And taking the square of the difference between the observed pseudo range and the theoretical pseudo range of the target GNSS satellite as a target measurement noise value.
6. The method according to any one of claim 1 to 4, wherein, The target GNSS satellite characteristics comprise actual observation data of a target GNSS satellite at the target historical moment; the all GNSS satellite characteristics comprise actual observation data of all GNSS satellites at any historical time, the all GNSS satellites comprise the target GNSS satellite, and the any historical time comprises the target historical time.
7. The method of claim 6, wherein the actual observed data comprises at least any one of or a combination of the following: Observations related to satellite positioning quality, observations related to atmospheric error estimation quality, observation type, check relation between different observations, observations related to residuals.
8. The method according to any one of claims 1-4, further comprising: And obtaining a theoretical pseudo range of the target GNSS satellite at the target historical moment based on the historical positioning result corresponding to the target historical moment, wherein the positioning precision of the historical positioning result meets a preset precision requirement.
9. A method of positioning, the method comprising: Acquiring GNSS current measurement data observed by target equipment, wherein the GNSS current measurement data comprises target GNSS satellite characteristics observed by target GNSS satellites at the current moment and all GNSS satellite characteristics observed by all GNSS satellites at the current moment; Inputting at least the GNSS current measurement data into a deep learning model trained by the method of any one of claims 1 to 8, to obtain a measurement noise value of the target GNSS satellite; and updating the measurement noise matrix in the Kalman filtering algorithm based on the measurement noise value to obtain a positioning result of the target equipment.
10. The method according to claim 9, wherein the method further comprises: Obtaining model parameters of the target equipment; The step of inputting at least the GNSS current measurement data into a deep learning model obtained by training by the method of any one of claims 1 to 8 to obtain a measurement noise value of the target GNSS satellite, specifically includes: Inputting the model parameters and the GNSS current measurement data into the deep learning model, wherein embedding layers of the deep learning model obtain model characteristics of the target equipment based on the model parameters; Inputting the target GNSS satellite characteristics and the model characteristics into a first full-connection layer of the deep learning model to obtain observation characteristics of the target GNSS satellite; inputting all GNSS satellite characteristics and model characteristics into a convolution layer and a pooling layer of the deep learning model to obtain environmental characteristics; And inputting the observation characteristics and the environmental characteristics of the target GNSS satellite into a splicing layer of the deep learning model, and obtaining a measured noise value of the target GNSS satellite through a second full-connection layer connected with the splicing layer in the deep learning model.

Description

Training method and positioning method for deep learning model Technical Field The application relates to the field of positioning, in particular to a training method and a positioning method of a deep learning model. Background When positioning is performed according to observations of the global navigation satellite system (Global Navigation SATELLITE SYSTEM, GNSS), a positioning algorithm based on Kalman filtering needs to obtain measurement noise of each observation to form a measurement noise matrix for updating the filtering algorithm. Currently, the mainstream measurement noise estimation method is obtained through nonlinear transformation according to fields such as signal to noise ratio in GNSS observation. On the one hand, the definition of the signal to noise ratio has model and manufacturer differences, and on the other hand, the current measurement noise is difficult to accurately measure by a single characteristic. The present inventors have found that academia attempts to predict the measured noise of GNSS satellites in real time using a trained measured noise model to dynamically update the measured noise matrix. However, because of the difficulty in acquiring the measurement noise true value data, the accuracy of the measurement noise predicted by the existing measurement noise model is low, which affects the positioning accuracy and cannot be landed in industry. Disclosure of Invention In view of the above, the present application provides a training method and a positioning method for a deep learning model, which can provide an optimal estimation of measurement noise for a positioning algorithm based on kalman filtering, and ensure positioning accuracy. In order to solve the problems, the technical scheme provided by the application is as follows: In a first aspect of the present application, there is provided a training method of a deep learning model, the method comprising: acquiring an observed pseudo range and a theoretical pseudo range of a target GNSS satellite at a target historical moment; Determining a target measurement noise value according to the observed pseudo range and the theoretical pseudo range of the target GNSS satellite; And training the deep learning model by using the training sample and taking the target measured noise value as a loss true value until the predicted measured noise value of the target GNSS satellite output by the deep learning model and the loss value output by the loss function corresponding to the target measured noise value meet the preset condition. The training samples are any satellite observation set at any historical moment, and the input of each training sample at least comprises target GNSS satellite characteristics and all GNSS satellite characteristics, and the output of each training sample is the measured noise value of the target GNSS satellite. In a second aspect of the present application, there is provided a positioning method comprising: Acquiring GNSS current measurement data observed by target equipment, wherein the GNSS current measurement data comprises target GNSS satellite characteristics observed by target GNSS satellites at the current moment and all GNSS satellite characteristics observed by all GNSS satellites at the current moment; inputting at least the GNSS current measurement data into a deep learning model to obtain a measurement noise value of the target GNSS satellite, wherein the deep learning model is obtained through training by the method of the first aspect; And updating the measurement noise matrix in the Kalman filtering algorithm based on the measurement noise value to obtain a positioning result of the target equipment. In a third aspect of the present application, there is provided a training apparatus for a deep learning model, the apparatus comprising: the acquisition unit acquires an observed pseudo range and a theoretical pseudo range of the target GNSS satellite at the target historical moment; The determining unit is used for determining a target measurement noise value according to the observed pseudo range and the theoretical pseudo range of the target GNSS satellite; The training unit is used for training the deep learning model by using the training sample and taking the target measured noise value as a loss true value until the predicted measured noise value of the target GNSS satellite output by the deep learning model and the loss value output by the loss function corresponding to the target measured noise value meet the preset condition. The training samples are any satellite observation set at any historical moment, and the input of each training sample at least comprises target GNSS satellite characteristics and all GNSS satellite characteristics, and the output of each training sample is the measured noise value of the target GNSS satellite. In a fourth aspect of the application, there is provided a positioning device comprising: The acquisition unit is used for acquiring GNSS current mea