AI-assisted GNSS / INS deep coupling carrier phase prediction method and device
Through the AI-assisted GNSS/INS deep-coupled carrier phase prediction method, the weighted compensation parameters of future time are estimated using the combined model of CNN, GRU and attention mechanism layer, which solves the problems of carrier phase loss and re-acquisition in complex urban environments, and improves the accuracy of carrier phase prediction and the continuity of positioning.
Patent Information
- Application Number
- CN202510521669.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-24
- Publication Date
- 2025-08-01
AI Technical Summary
In complex urban environments, GNSS signal occlusion causes frequent carrier phase loss and re-acquisition, affecting the high-precision positioning continuity and availability of GNSS/INS deep coupling systems. The existing carrier phase prediction methods are simple to process before the prediction starts and are prone to errors.
The AI-assisted GNSS/INS deep-coupled carrier phase prediction method is used to establish an information transmission model between phase prediction and tracking loops, and use a combined model of convolutional neural network (CNN), gated cyclic unit (GRU) and attention mechanism layer to estimate the weighted compensation parameters of future time to reduce prediction errors.
It effectively reduces the carrier phase prediction error, improves the accuracy of carrier phase prediction and the continuity of positioning, and performs well in static and dynamic conditions.
Smart Images
Figure CN120403618A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of integrated navigation, and particularly relates to a deep-coupled carrier phase prediction method and device for a Global Navigation Satellite System (GNSS) / Inertial Navigation System (INS) assisted by artificial intelligence (AI). Background Art
[0002] GNSS / INS integrated navigation has first evolved into a loosely coupled system that fuses data at the positioning level and a tightly coupled system that fuses data at the observable level according to the data fusion method. However, the loose and tight coupling algorithms cannot change the baseband signal processing logic inside the GNSS receiver and cannot improve the quality of GNSS observables. Therefore, the GNSS / INS deep-coupled system came into being, enabling the information of each sensor to be deeply fused and mutually supplemented at the receiver signal processing level.
[0003] However, with the modern development of cities, in the face of complex urban environments mainly including tree-lined roads, high-rise buildings, viaducts, etc., even when using the GNSS / INS deep-coupled system, problems such as inaccurate positioning still occur due to frequent GNSS signal blockages. When the signal is blocked, GNSS observables such as pseudorange and carrier phase will be severely affected. Since positioning based on pseudorange can only achieve meter-level accuracy, real-time kinematics (RTK) and precise point positioning (PPP) technologies based on pseudorange and carrier phase can achieve centimeter-level positioning. And the carrier phase is more vulnerable than other observables. Therefore, maintaining the quality of the carrier phase is both the key and the difficulty in the positioning process. In complex urban environments, the carrier phase frequently encounters loss of lock and reacquisition, which makes it difficult for the GNSS receiver to continuously output high-quality carrier phase observations. During the process of the carrier phase changing from discontinuous to continuous, RTK and PPP need to re-fix the ambiguity to restore the continuously available high-precision positioning performance.
[0004] Therefore, when it is difficult to obtain the carrier phase of a certain satellite channel due to signal interference, the signal of other channels should be used to assist in predicting the carrier phase of this channel. In this way, when the signal is restored, the error of the carrier phase is small, and re-fixing the ambiguity can be avoided. Most of the existing carrier phase prediction methods focus on the processing after the start of carrier phase prediction, and the steps before entering the prediction are relatively simple, which may result in obvious errors. Summary of the Invention
[0005] To address this issue, the present invention adopts a prediction method that takes into account the processing of cache information before the start of prediction. At the same time, the present invention analyzes that there is a certain gap between the calculation time of the compensation parameter utilized by the original prediction theory and the end time of prediction, resulting in a certain error in the calculated compensation parameter. A new prediction mechanism is proposed: using AI to estimate the weighted compensation parameter containing future time information, which can reduce the error caused by the compensation parameter.
[0006] Object of the present invention: To overcome the problem that in a complex urban environment, the carrier phase frequently encounters loss of lock and reacquisition, and RTK and PPP need to re-fix the ambiguity, which seriously affects the continuity and availability of high-precision positioning. The present invention provides an AI-assisted GNSS / INS deep-coupled carrier phase prediction method and device, which can utilize the information of normal channels to predict the divergence of the carrier phase of occluded channels, thereby avoiding re-fixing the cycle ambiguity.
[0007] The content of the present invention is as follows: An AI-assisted GNSS / INS deep-coupled carrier phase prediction method, and the implementation steps are as follows: Step 1, establish an information transfer model between the phase prediction and the tracking loop. After the carrier phase starts to be predicted, use the information of multiple normal channels to obtain the correlation relationship between the predicted Doppler and the compensation parameter; Step 2, establish an AI estimation weighted compensation parameter model. When predicting the carrier phase, use multiple compensation parameters continuously calculated within a period of time to predict the compensation parameter for a certain future time, calculate the weighted compensation parameter, and use the weighted compensation parameter to obtain the predicted Doppler, thereby obtaining the carrier phase; Step 3, establish a carrier phase prediction system process model. According to the characteristics of the carrier phase error sequence in the loop, determine whether to retain the cache information of the error, and control the opening and closing of the carrier phase prediction during the alternation of satellite signal interruption and recovery; Step 4, obtain the data of GNSS and INS, and perform GNSS / INS deep-coupled integrated navigation to obtain the motion Doppler of each satellite channel. Combine the loop Doppler of the normal satellite channels, and use the models established in Steps 1, 2, and 3 to predict the carrier phase to obtain the carrier phase prediction result.
[0008] Furthermore, the specific implementation manner of Step 1 is as follows: Step 1.1, construct a reference relationship between the predicted Doppler and the carrier phase; Step 1.2, use a single normal satellite channel to obtain an expression for the single-satellite predicted Doppler: Step 1.3, determine the satellite motion Doppler update frequency and the receiver motion Doppler update frequency, so that the satellite motion Doppler update frequency and the receiver motion Doppler update frequency are synchronously updated with the loop; Step 1.4: Obtain the correlation relationship expression of multi-satellite predicted Doppler and compensation parameters using multiple normal satellite channels; the multi-satellite predicted Doppler in Step 1.4 The relationship with the compensation parameter is as follows:
[0009] Among them, the superscript j represents the satellite number, is the motion Doppler of the open-loop prediction channel, is the clock drift Doppler of the receiver, is the estimation of other errors in the prediction channel, is in other normal channels Take the average value, is the other error of the normal channel, is at the start of the prediction, and The difference of, is called the compensation parameter.
[0010] Furthermore, in Step 2, the AI estimation weighted compensation parameter model is a combined model of a convolutional neural network CNN, a gated recurrent unit GRU, and an attention mechanism layer Attention. Among them, CNN is used to extract features from the input time series data. GRU receives the features extracted by CNN as input and processes the long-term dependencies in the sequence data. The attention mechanism is applied to the output sequence of GRU. Through the attention mechanism, important time steps in the GRU output sequence are focused on, and different weights are assigned to different time steps, so as to better capture the key information in the time series and then generate a context vector containing key information; finally, the compensation parameter at the future moment is obtained through the output layer.
[0011] Furthermore, CNN includes two convolutional layers and two pooling layers. The first convolutional layer sets 16 convolutional kernels with a size of 3, a stride of 1, uses same padding, and uses the ReLU activation function to extract the local relationship between adjacent time points and introduce non-linearity; the subsequent first pooling layer uses max pooling, with a window size and stride of 2 to reduce the data volume; the second convolutional layer increases the number of convolutional kernels to 32, with the same size of 3, a stride of 1, same padding, and the ReLU activation function to learn more complex feature combinations; the second pooling layer is set the same as the first pooling layer to further reduce the sequence length; GRU receives the feature sequence output by CNN, sets 64 units, and is used to capture the long-term dependencies in the time series, and the output of each time step contains the sequence information before that time point; The attention mechanism layer weights the GRU output sequence, uses a learnable vector as the query, takes the 64-dimensional feature vectors at each time step in the GRU output sequence as the key and value respectively, calculates the attention weights by computing the similarity between the query and the key, and sums the weighted values to generate a context vector containing key information; The output layer is a fully connected layer with 1 neuron.
[0012] Further, starting from the current moment, one time point is taken every certain interval going forward, a total of k time points are obtained, and then the compensation parameters at these k moments are calculated. These compensation parameters form a time series, and inputting them into the AI estimation weighted compensation parameter model can obtain the compensation parameters at future moments.
[0013] Further, at the beginning of the prediction, estimate the weighted compensation parameters containing future time information , which are weighted by the compensation parameters at future moments and the calculated compensation parameters at the current moment to obtain, and the formula is expressed as:
[0014] is the weighting coefficient, and is used for phase prediction; thus, the relationship formula between the predicted Doppler and the carrier phase is obtained:
[0015] where the superscript j represents the satellite number, is the motion Doppler of the open-loop prediction channel, is the clock drift Doppler of the receiver, is the in other normal channels, taking the average value, is the other error of the normal channel.
[0016] Further, the specific implementation method of step three is as follows: Step 3.1, use the first preset interval time to periodically extract the carrier phase error sequence in the loop and detect whether its error fluctuation is normal. In the time period included in this error sequence, the earliest time point is the detection start point, and the latest time point is the detection end point. The detection end point of the previous detection is also the detection start point of the next detection. In one detection, if the phase error fluctuation is normal, delete the cache of the phase error sequence in the previous detection.
[0017] Step 3.2, if in a single detection, abnormal phase error fluctuation appears for the first time, set the detection result to warning, do not delete the cached information, and perform detection at the second preset interval time, where the second preset interval time is less than the first preset interval time. If it is determined that the carrier phase error fluctuation is abnormal, retain the cached data and enter the phase prediction preparation stage. When the receiver signal is blocked, enter the phase prediction stage; Step 3.3, perform a reverse search on the time axis to determine the time point when the carrier phase error is less than the error threshold. If the first time monitoring point does not meet the carrier phase error requirement, the detection result of the first time monitoring point is rejection, and continue the reverse search on the time axis. If the second time monitoring point meets the carrier phase error requirement, the detection result of the second time monitoring point is allowed, and continue to detect the third time monitoring point. If both the second time monitoring point and the third time monitoring point meet the carrier phase error requirement, the detection result of the third time monitoring point is allowed; Step 3.4, for any time monitoring point with a detection result of allowed, perform a forward search on the time axis to the next time point, and start carrier phase prediction until the carrier-to-noise ratio of the satellite signal reflects that the signal has returned to normal, then restore the open loop of the prediction channel to the closed loop, and the carrier phase prediction ends.
[0018] Further, the specific implementation method for determining whether the carrier phase error is within the normal range is as follows: Let the time between two detections be T , the number of phase values be N , the phase value sequence be , the i -th phase value in the sequence be , the set threshold be , the Allan variance block time be t , the detection method can be specifically divided into: (1) calculating the proportion of the phase error within the threshold during this period r , (2) calculating the Allan variance of the phase error at different block times t ; expressed by the formula:
[0019] When r and both meet the set conditions, it is determined that the phase error during this period is normal and the signal is basically not interfered.
[0020] Further, the GNSS data includes position, timestamp, and satellite signal strength information; the INS data includes speed, angular velocity, and acceleration information.
[0021] The present invention also provides an AI-assisted GNSS / INS deep coupling carrier phase prediction device, including: A processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an AI-assisted GNSS / INS deep coupling carrier phase prediction method as described in the above technical solution.
[0022] Aiming at the problem of discontinuous carrier phase caused by frequent signal interruptions in complex environments, the present invention provides an AI-assisted carrier phase prediction method for a deep coupling system at the baseband signal processing level. It uses AI to estimate weighted compensation parameters more suitable for prediction, including the entire complete process of carrier phase prediction starting from whether to discard cached data, which can reduce the prediction error of the carrier phase and has better prediction effects under both static and dynamic conditions. This method has been verified to be feasible through actual verification and has good practicability. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 It is a flowchart of the present invention.
[0024] Figure 2 It is a diagram of information transfer between the phase prediction and the tracking loop.
[0025] Figure 3 It is a diagram of the CNN+GRU+Attention neural network architecture.
[0026] Figure 4 It is a flowchart of the carrier phase prediction system.
[0027] Figure 5 It is a curve of carrier phase error divergence at 30 s.
[0028] Figure 6 It is a curve of carrier phase error divergence at 60 s. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0029] The following further describes the present invention in conjunction with the drawings and specific embodiments: The present invention focuses on the field of deep coupling of combined navigation. In view of the problem that frequent signal interruptions in complex environments cause discontinuous carrier phase, an AI-assisted carrier phase prediction method applied to the baseband signal processing level in a deep coupling system is provided, covering the system process of carrier phase prediction starting from whether to discard cached data. In the original prediction theory, the calculation time of the compensation parameters and the time interval between the end of the prediction are relatively large, resulting in a large error in the compensation parameters, which in turn causes a large prediction error. Therefore, the present invention proposes to use AI to estimate weighted compensation parameters containing future time information, which effectively improves this situation and can reduce the prediction error of the carrier phase. It has better prediction effects under both static and dynamic conditions. The present invention is a complete prediction method that has been verified to be feasible in practice and has good practicality.
[0030] As attached Figure 1 Shown is the flow chart of the present invention, as shown in the attached Figure 2 The figure shows the information transmission between the phase prediction and tracking loops. The specific steps of the present invention are as follows: Step 1: Establish an information transfer model between the phase prediction and tracking loops. This model focuses on how to use the information of each normal channel (including the movement speed of the satellite and receiver, and the loop tracking Doppler) to obtain the correlation between the predicted Doppler and the compensation parameters after the carrier phase prediction begins. The details are as follows: 1.1, Construct the relationship between the predicted Doppler and carrier phase:
[0031] in, , The time and The corresponding carrier phase, is the predicted Doppler frequency, is the estimated error of the carrier phase.
[0032] 1.2, derive the expression for obtaining the predicted Doppler using a single normal satellite channel: The Doppler measurement value of the receiver is mainly composed of four components: (1) the movement of the satellite and the receiver; (2) the clock drift of the satellite and the receiver; (3) the rate of change of the propagation path delay; and (4) the receiver noise. The receiver noise here includes all errors caused by the receiver, including the thermal noise of electronic devices such as antennas and amplifiers, signal quantization errors, cross-correlation interference between satellite signals, etc. The Doppler error caused by satellite clock drift, the rate of change of propagation path delay and receiver noise is small and difficult to predict when predicting the carrier phase, and can be uniformly recorded as other errors. Therefore, set the satellite j The corresponding channel is the open-loop prediction channel, which predicts Doppler The expression is:
[0033] where the superscript j represents the satellite number, is the motion Doppler of the open-loop prediction channel. is the clock drift Doppler of the receiver, is the other error of this channel.
[0034] In the normal channel, the following relationship exists:
[0035] where, is the loop Doppler obtained by normal loop tracking, is the motion Doppler of this normal channel, is the other error of this normal channel.
[0036] The clock drift of the open-loop tracking channel can be predicted using the clock drift obtained from the normal tracking channel. When the satellite signal is interrupted, the open-loop tracking channel uses the predicted Doppler to update the numerically controlled oscillator (NCO) of the carrier tracking loop and outputs continuous carrier phase observations. For the open-loop tracking channel i , if can be accurately estimated, the carrier phase prediction error will be in a relatively low range. First, consider the Doppler caused by relative motion, which is generated by the motion of both the inertial navigation and the receiver and can be written as:
[0037] where, represents the Doppler caused by satellite motion, represents the Doppler caused by receiver motion, Meanwhile, for the motion Doppler associated with satellite i , the following relationship exists:
[0038] where, p and v are the position and measured velocity of the receiver. and are the position and velocity of satellite i . is the carrier wavelength. represents the unit direction vector between the receiver and satellite numbered i , which is calculated from the satellite position and the receiver position and is used to project the receiver motion and satellite motion onto the line of sight between the two.
[0039] From equations (4), (5), and (6), for the channeli , there is
[0040] Designating the in Equation (2) as , a form of predicted Doppler can be obtained. From Equations (2), (3), and (7), we can get
[0041] This is the expression for obtaining the predicted Doppler using a single normal satellite channel. However, since only the satellites in the normal channels are used, the existing information is not fully utilized. Moreover, even for normal channels, there are still differences in the i calculated by different channels. Therefore, in the subsequent steps, we will consider how to integrate the information of each normal channel to obtain other error estimates of the predicted channel . .
[0042] 1.3. Increase the and update frequency so that it is updated synchronously with the loop.
[0043] Use k to represent the th satellite velocity update time, which coincides with the loop update time and no recursion is required. Then the change in satellite Doppler k can be estimated from the Doppler at the k th satellite velocity update and the Doppler at the -1th satellite velocity update. This relationship can be written as:
[0044] where is the loop update frequency and is the frequency at which the receiver calculates the satellite velocity update.
[0045] Use k to represent the Doppler obtained by the k th recursion during the loop update while the satellite velocity remains unchanged between the n th and the +1th satellite velocity update times. Then can be expressed as:
[0046] The change in receiver Doppler <000023)]] can be obtained from the acceleration k at the th receiver velocity update provided by the inertial navigationIt is obtained that this relationship can be expressed as:
[0047] where is the update frequency of the inertial navigation for the receiver speed, is the time between two inertial navigation updates.
[0048] Similarly, use to represent the k th and the k +1th satellite speed update moments. The Doppler obtained by the n th recursion during the loop update while the receiver speed is not updated is and can be written as:
[0049] Through the above method, and have obtained relatively accurate values at each loop update time.
[0050] 1.4, Derive the expression for obtaining the predicted Doppler using multiple normal satellite channels: The Doppler measured by the phase-locked loop minus the estimated motion Doppler can obtain the Doppler of the clock drift and other errors of each channel. They have the same trend because each channel has a consistent clock drift Doppler, but due to different other errors in each channel, the total Doppler is not exactly the same. The reasons for the difference are the satellite speed calculation error and the different ionospheric and tropospheric delays of different satellites. If directly using in the normal tracking channels to assist the open-loop tracking channel, the divergence degree of the carrier phase error may be relatively large in a short time.
[0051] Although there are certain differences in the DC bias, the parts that fluctuate rapidly with time have strong consistency. can be divided into a slowly varying bias part and a rapidly varying jitter part, denoted as and respectively. is the 100-epoch sliding average of with an epoch interval of 20 ms, is expressed as
[0052] changes rapidly and irregularly, and the fluctuations of dynamic data are more intense than those of static data. Generally speaking, the gap between different channels is very small, and due to the The change trends are almost the same. As can be seen from Equation (13), for each channel also has the same change trend.
[0053] Taking the average value of in other normal channels can better approximate the change trend in the open-loop channel . However, and may still have an error close to a constant value. Therefore, at the beginning of the prediction, the difference between can be recorded . This is called the compensation parameter. Then, the estimation of other errors in the predicted channel can be expressed as:
[0054] Specifying in Equation (2) as can obtain another form of the predicted Doppler :
[0055] Compared with Equation (8), this form makes more full use of the Doppler information of each normal channel and can achieve higher accuracy in prediction.
[0056] Step 2: Establish an AI model for predicting the future-time compensation parameter. In Step 1, a carrier phase prediction method was obtained through theoretical derivation of formulas. However, in practice, since the calculated compensation parameter will be different at different time points, and the farther the time for calculating is from the end time of the prediction, the greater the possible error. Compared with calculating the current-time during prediction, this paper proposes a new prediction mechanism: at the beginning of the prediction, estimate the weighted compensation parameter containing future-time information. The weighted compensation parameter at the future time is weighted with the calculated current-time compensation parameter . The formula is expressed as:
[0057] is the weighting coefficient, and is used for phase prediction.
[0058] Then the predicted Doppler can be expressed as
[0059] Because at the future time Closer to the predicted end time, so the resulting error should be smaller. Since the formula derivation method uses the cached information of the baseband, it can only calculate the current, that is, at the start of the prediction . Therefore, to estimate the at future times, it is necessary to use artificial intelligence.
[0060] Using formula derivation can obtain not only the compensation parameters at the current time, but also the compensation parameters at past times. Starting from the current time, this invention takes a time point every 200 ms backwards, a total of 50 time points are obtained, and the compensation parameters at 50 moments can be calculated. These compensation parameters form a time series. By predicting the 10th point after this sequence, that is, the compensation parameter at 2 s, and then calculating the weighted compensation parameter. At the position where the signal is not interrupted, the true compensation parameter for the next 2 s can be obtained through calculation, and these true values are used to train the neural network. When performing carrier phase prediction, artificial intelligence is used to estimate the compensation parameter for the next 2 s, calculate the weighted compensation parameter, and replace the compensation parameter at the current time for phase prediction. After verification, the neural network composed of a Convolution Neural Network (CNN), a Gated Recurrent Unit (GRU), and an Attention mechanism has a good estimation effect on this task, and this neural network can be abbreviated as CNN+GRU+Attention.
[0061] CNN is a deep learning model commonly used to process data with a grid structure. Its core components are the convolutional layer and the pooling layer. In the combined model, CNN is mainly used for feature extraction of the input data. For time series data, local features and patterns of the data are extracted through convolutional operations. These features can capture short-term dependencies and local variations in the data. The convolutional layer performs sliding convolutional operations on the input data using multiple convolutional kernels to extract local features. These convolutional kernels can automatically learn different feature patterns, such as edges, textures, etc. The pooling layer is used to downsample the feature map, reducing the data dimension while retaining important feature information, enhancing the robustness of the model and its invariance to changes such as translation and scaling. GRU is a variant of the Recurrent Neural Network (RNN) used to process sequential data. It controls the flow and forgetting of information by introducing update gates and reset gates. GRU receives the features extracted by CNN as input and processes the long-term dependencies in the sequential data. Since GRU can retain historical information, it can further integrate and process the local features extracted by CNN, capturing the changing trends and patterns of the data in the time dimension. The update gate determines how much information in the current state comes from the previous state and how much comes from the current input. The reset gate determines how to combine the previous state with the current input. This enables GRU to better capture long-term dependencies in long sequences while avoiding the gradient vanishing or explosion problems that may occur in traditional RNNs. The purpose of Attention is to enable the model to selectively focus on different parts of the input sequence. It determines the importance weights of each element in the input by calculating attention scores. These scores are usually calculated based on the relevance of the input elements to the current task. Then, the input is weighted and summed according to these weights, enabling the model to focus on more critical information and improving the model's performance and ability to capture important features. The attention mechanism is applied to the output sequence of GRU. Through the attention mechanism, the model can automatically focus on important time steps in the GRU output sequence, assign different weights to different time steps, and thus better capture the key information in the sequence. Finally, the context vector can be fused with other information for subsequent prediction tasks. As attached Figure 3 Shown is the architecture diagram of the CNN+GRU+Attention neural network. The input is a sequence of compensation parameters calculated for recent times, and the output is the compensation parameters for future times.
[0062] Specifically, in the neural network of the present invention for time series prediction to estimate the compensation parameters for the next 2s, the CNN+GRU+Attention structure is designed as follows: Convolutional Neural Network (CNN): The CNN part aims to extract local features from the input time series of compensation parameters. The input is a time series composed of compensation parameters at 50 time points taken every 200 ms from the current time point backwards, with a dimension of (batch_size, 50, 1).
[0063] The CNN contains two convolutional layers and two pooling layers. The first convolutional layer is set with 16 convolutional kernels of size 3, a stride of 1, using'same' padding, and the ReLU activation function to extract the local relationships between adjacent time points and introduce non-linearity. The subsequent first pooling layer uses max pooling with a window size and stride of 2 to reduce the data volume. The second convolutional layer increases the number of convolutional kernels to 32, still with a size of 3, a stride of 1,'same' padding, and the ReLU activation function to learn more complex feature combinations. The second pooling layer has the same settings as the first pooling layer to further reduce the sequence length. After two convolutional and pooling operations, the sequence length is reduced from 50 to 12, and the feature dimension increases to 32.
[0064] Gated Recurrent Unit (GRU): The GRU receives the feature sequence with a dimension of (batch_size, 12, 32) output by the CNN, and is set with 64 units to capture the long-term dependencies in the time series. Its output dimension is (batch_size, 12, 64), and the output at each time step contains the sequence information before that time point.
[0065] Attention Mechanism: The attention mechanism weights the GRU output sequence, focusing on the important time steps related to the prediction of the compensation parameters in the next 2s. A 64-dimensional learnable vector is used as the query, and the 64-dimensional feature vectors at each time step in the GRU output sequence are used as the key and value respectively. The attention weights are obtained by calculating the similarity between the query and the key, and the weighted sum of the values is used to generate a context vector containing the key information.
[0066] Output Layer: The context vector obtained by the attention mechanism is input into the fully connected layer. Since the goal is to predict a single compensation parameter value in the next 2s, the fully connected layer is set with 1 neuron.
[0067] This CNN + GRU + Attention architecture leverages the CNN's local feature extraction capabilities, the GRU's long-term dependency processing capabilities, and the attention mechanism's information-focusing capabilities to accurately predict compensation parameters for the next two seconds. During training, the known true compensation parameters for the next two seconds are used as labels, and the backpropagation algorithm updates the model parameters to improve prediction accuracy.
[0068] Step 3: Establish a carrier phase prediction system process model, determine whether to retain the error cache information based on the characteristics of the carrier phase error sequence in the loop, and control the on and off of the carrier phase prediction in the alternation of satellite signal interruption and recovery. Figure 4 The flowchart of the carrier phase prediction system is shown below. The details are as follows: 3.1. Use the first preset interval time to periodically extract the carrier phase error sequence in the loop and detect whether its error fluctuation is normal. In the time period contained in the error sequence, the earliest time point is the detection start point, and the latest time point is the detection end point. The detection end point of the previous detection is also the detection start point of the next detection. In one detection, if the phase error fluctuation is normal, the cache of the phase error sequence in the previous detection is deleted. After two discarded cache tests, the results are both allowed, that is, the phase errors are within the normal range, and the data of these two cache units can be deleted.
[0069] 3.2. If the third discarded cache test reveals an abnormal phase error trend during this period, the test result is a warning. The cached data from this test to the previous one is retained, and the test interval is shortened before another test is performed. If the anomaly persists during the fourth test, the cached data is retained and the phase prediction preparation phase begins. This phase prediction phase begins until the receiver detects signal obstruction by calculating the satellite signal's carrier-to-noise ratio.
[0070] 3.3. By searching backward on the time axis, we look for a time point where the phase error is less than the threshold. The first time point does not meet the phase error requirement, and the detection result is rejected, so we continue to search backward on the time axis. The second time point meets the phase error requirement, and the detection result is allowed, but we cannot start phase prediction directly from this point, because it is possible that an abnormal change in phase error occurred before this point, but the phase error detected near this point happens to be smaller, so we need to detect another time point in the past. If the phase errors between these two time points are within the normal range, we can confirm that the latter time point is available.
[0071] 3.4. When the detection result of the previous time point is confirmed to be available, search forward to the next time point on the time axis, and then start phase prediction according to Equation (1) until the calculated carrier-to-noise ratio of the satellite signal reflects that the signal returns to normal, then restore the open loop of the prediction channel to a closed loop, and the carrier phase prediction ends.
[0072] During the whole process, it is necessary to judge whether the phase error within the time interval is normal for many times. The judgment method is: Let the time between two detections be T , the number of phase values be N , the phase value sequence be , the i -th phase value of the sequence be , the set threshold be , the Allan variance block time be t . The detection method proposed in this study can be specifically divided into: (1) calculating the proportion of the phase error within the threshold during this period r , (2) calculating the Allan variance of the phase error at different block times t . It is expressed by the formula:
[0073] The above algorithm is called the phase error detection algorithm. When r the proportion is close to 100% and is within the normal variance interval stored by the system using the previous variance, this algorithm judges that the phase error during this period is normal and the signal is basically not interfered.
[0074] Step 4: Obtain GNSS and INS data. GNSS data includes information such as position, timestamp, and satellite signal strength; INS data includes information such as speed, angular velocity, and acceleration. And perform GNSS / INS deep coupling integrated navigation to obtain the motion Doppler of each satellite channel, and compare and verify this value with the receiver speed information output by the on-vehicle SPAN-CPT6 device to ensure the accuracy of the integrated navigation. Then, combined with the loop Doppler of the normal satellite channel, use the models established in Steps 1, 2, and 3 to perform carrier phase prediction. Specifically, Step 1 includes obtaining the value of the compensation parameter using the satellite speed, receiver speed, and loop Doppler, and obtaining the relationship between the predicted Doppler and the compensation parameter. Step 2 includes estimating the weighted compensation parameter containing future time information using multiple continuously calculated compensation parameters within a period of time to make the compensation parameter at the start of prediction more accurate. Substituting the new compensation parameter into the formula can obtain the predicted Doppler, and thus obtain the carrier phase. Step 3 includes the system process of carrier phase prediction, involving whether to retain cache information and how to judge the interruption and recovery of the signal.
[0075] The following is an illustration of the comparison effect between the present invention and the existing non-AI prediction method, i.e., the method described in the invention titled "A Combined Navigation Deep-Coupled Carrier Phase Prediction Method and System", as Figure 5 shown is the carrier phase error divergence curve at 30 s. From Figure 5 it can be seen that for non-AI prediction ( Figure 5 (a) in Figure 5 ), at 30 s, the root mean square of the carrier phase error is 0.203 cycles. For AI prediction ( Figure 6 (b) in Figure 6 ), at 30 s, the root mean square of the carrier phase error is 0.138 cycles. Compared with non-AI prediction, the root mean square of the error is reduced by 31.7%. As Figure 6 shown is the carrier phase error divergence curve at 60 s. For non-AI prediction ( (a) in
[0076] ), at 60 s, the root mean square of the carrier phase error is 0.414 cycles. For AI prediction ( (a) in ), at 60 s, the root mean square of the carrier phase error is 0.311 cycles. Compared with non-AI prediction, the root mean square of the error is reduced by 24.9%. When performing carrier phase prediction, the smaller the divergence of the carrier phase error in the prediction channel, the closer the carrier phase prediction is to the true value, thereby making the positioning more continuous and robust. Therefore, the method of this patent can achieve better results in carrier phase prediction. On the other hand, the embodiment of the present invention also provides an AI-assisted GNSS / INS deep-coupled carrier phase prediction device, including: a processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an AI-assisted GNSS / INS deep-coupled carrier phase prediction method as described in the above technical solution.
[0077] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. An AI-assisted GNSS / INS deep coupling carrier phase prediction method, characterized in that It includes the following steps: Step 1: Establish an information transfer model between the phase prediction and the tracking loop. After the carrier phase starts to be predicted, the correlation relationship between the predicted Doppler and the compensation parameters is obtained by using the information of multiple normal channels. Step 2: Establish an AI estimation weighted compensation parameter model. When predicting the carrier phase, use multiple compensation parameters calculated continuously within a period of time to predict the compensation parameters for a certain future time, calculate the weighted compensation parameters, and obtain the predicted Doppler using the weighted compensation parameters, thereby obtaining the carrier phase. Step 3: Establish a carrier phase prediction system process model. According to the characteristics of the carrier phase error sequence in the loop, determine whether to retain the cached information of the error, and control the opening and closing of the carrier phase prediction during the alternation of satellite signal interruption and recovery. Step 4: Obtain the data of GNSS and INS, perform GNSS / INS deep coupling integrated navigation, obtain the motion Doppler of each satellite channel, combine the loop Doppler of the normal satellite channels, and use the models established in Steps 1, 2, and 3 to predict the carrier phase to obtain the carrier phase prediction result.
2. An AI-assisted GNSS / INS deep coupling carrier phase prediction method according to claim 1, characterized in that: The specific implementation method of Step 1 is as follows: Step 1.1: Construct a reference relationship formula between the predicted Doppler and the carrier phase. Step 1.2: Use a single normal satellite channel to obtain an expression for the single-satellite predicted Doppler. Step 1.3: Determine the satellite motion Doppler update frequency and the receiver motion Doppler update frequency, so that the satellite motion Doppler update frequency, the receiver motion Doppler update frequency, and the loop are updated synchronously. Step 1.4: Use multiple normal satellite channels to obtain the correlation relationship expression between the multi-satellite predicted Doppler and the compensation parameters. Multi-satellite predicted Doppler in Step 1.4 The relationship with the compensation parameter is as follows: where the superscript j represents the satellite number, is the motion Doppler of the open-loop prediction channel, is the clock drift Doppler of the receiver, is the estimation of other errors in the prediction channel, in other normal channels takes the average value, is the other error of the normal channel, at the start of the prediction, and the difference, is called the compensation parameter.
3. An AI-assisted GNSS / INS deep coupling carrier phase prediction method according to claim 1, wherein: The AI estimation weighted compensation parameter model in Step 2 is a combined model of a convolutional neural network CNN, a gated recurrent unit GRU, and an attention mechanism layer Attention. Among them, CNN is used to extract features from the input time series data. GRU receives the features extracted by CNN as input and processes the long-term dependencies in the sequence data. The attention mechanism is applied to the output sequence of GRU. Through the attention mechanism, important time steps in the GRU output sequence are focused on, and different weights are assigned to different time steps to better capture the key information in the time series and then generate a context vector containing key information. Finally, the compensation parameters for future moments are obtained through the output layer.
4. An AI-assisted GNSS / INS deep-coupled carrier phase prediction method according to claim 1, characterized in that: CNN includes two convolutional layers and two pooling layers. The first convolutional layer is set with 16 convolutional kernels of size 3, a stride of 1, and same padding, and uses the ReLU activation function to extract the local relationship between adjacent time points and introduce non-linearity. Subsequently, the first pooling layer uses max pooling, with a window size and a stride of 2 to reduce the data volume. The second convolutional layer increases the number of convolutional kernels to 32, still of size 3, a stride of 1, same padding, and the ReLU activation function to learn more complex feature combinations. The second pooling layer is set the same as the first pooling layer to further reduce the sequence length. The GRU receives the feature sequence output by the CNN, sets 64 units to capture long-term dependencies in the time series, and the output at each time step contains the sequence information before that time point. The attention mechanism layer weights the output sequence of the GRU. The learnable vector is used as the query, and the 64-dimensional feature vectors at each time step in the GRU output sequence are used as the key and value respectively. The attention weights are obtained by calculating the similarity between the query and the key, and the weighted sum of the values is used to generate a context vector containing key information. The output layer is a fully connected layer with 1 neuron.
5. An AI-assisted GNSS / INS deep coupling carrier phase prediction method according to claim 3, characterized in that: Starting from the current moment, one time point is taken every certain interval forward, and a total of k time points are obtained. Then, the compensation parameters at these k moments are calculated, and these compensation parameters form a time series, which is input into the AI estimation weighted compensation parameter model to obtain the compensation parameters at future moments.
6. An AI-assisted GNSS / INS deep-coupled carrier phase prediction method according to claim 1 or 3, characterized in that: At the beginning of prediction, a weighted compensation parameter containing future time information is estimated , which is obtained by weighting the compensation parameter at the future moment and the calculated compensation parameter at the current moment , and is expressed by the formula as follows: is a weighting coefficient, and use to perform phase prediction; thus, the relationship between the predicted Doppler and the carrier phase is obtained: where the superscript j represents the satellite number, is the motion Doppler of the open-loop prediction channel, is the clock drift Doppler of the receiver, is in other normal channels taking the average value, is other errors of the normal channel.
7. An AI-assisted GNSS / INS deep coupling carrier phase prediction method according to claim 1, characterized in that: The specific implementation method of step three is as follows: Step 3.1: Use the first preset interval time to periodically extract the carrier phase error sequence in the loop and detect whether its error fluctuation is normal. In the time period included in this error sequence, the earliest time point is the detection start point, and the latest time point is the detection end point. The detection end point of the previous detection is also the detection start point of the next detection. In one detection, if the phase error fluctuation is normal, the cache of the phase error sequence in the previous detection is deleted. Step 3.2: If, for the first time in one detection, the phase error fluctuation is abnormal, the detection result is set to warning, and the cache information is not deleted. Detection is performed at the second preset interval time, and the second preset interval time is less than the first preset interval time. If it is determined that the carrier phase error fluctuation is abnormal, the cache data is retained, and the phase prediction preparation stage is entered. When the receiver signal is blocked, the phase prediction stage is entered. Step 3.3: Perform a reverse search on the time axis to determine the time points where the carrier phase error is less than the error threshold. If the first time monitoring point does not meet the carrier phase error requirement, the detection result of the first time monitoring point is rejected, and continue to perform a reverse search on the time axis. If the second time monitoring point meets the carrier phase error requirement, the detection result of the second time monitoring point is allowed, and continue to detect the third time monitoring point. If both the second time monitoring point and the third time monitoring point meet the carrier phase error requirement, the detection result of the third time monitoring point is allowed. Step 3.4: For any time monitoring point with an allowed detection result, perform a forward search on the time axis to the next time point and start carrier phase prediction until the carrier-to-noise ratio of the satellite signal reflects that the signal has returned to normal, then restore the open loop of the prediction channel to the closed loop, and the carrier phase prediction ends.
8. An AI-assisted GNSS / INS deep coupling carrier phase prediction method according to claim 7, characterized in that: The specific implementation method for determining whether the carrier phase error fluctuation is normal is as follows: Let the time between two detections be T , the number of phase values be N , the phase value sequence be , the i -th phase value of the sequence be , the set threshold be , the Allan variance block time be t , the detection method can be specifically divided into: (1) calculating the proportion of the phase error within the threshold during this period r , (2) calculating the Allan variance t of the phase error at different block times ; expressed by the formula as: When r and both meet the set conditions, it is determined that the phase error during this period is normal and the signal is basically not interfered with.
9. An AI-assisted GNSS / INS deep coupling carrier phase prediction method according to claim 1, characterized in that: GNSS data includes position, timestamp, and satellite signal strength information; INS data includes speed, angular velocity, and acceleration information.
10. An AI-assisted GNSS / INS deep-coupled carrier phase prediction device, characterized in that, Including: A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an AI-assisted GNSS / INS deep coupling carrier phase prediction method according to any one of claims 1-9.