Self-adaptive integrated navigation method, system and terminal applied to obstacle area

Abstract

The invention discloses a self-adaptive integrated navigation method, system and terminal applied to an obstacle area, and the method comprises the following steps: acquiring a specific force and an angular rate output by an inertial sensor, inputting the specific force and the angular rate into a trained long-short-term memory network, and outputting a predicted displacement increment; acquiring current data output by an inertial navigation system, and judging whether the predicted displacement increment obeys target distribution or not according to the current data and the predicted displacement increment; when the predicted displacement increment does not obey target distribution, calculating an observation noise variance matrix according to an attenuation factor, and generating a target displacement increment according to the observation noise variance matrix and the predicted displacement increment; and generating a vehicle position according to the target displacement increment and the current data. According to the invention, accurate navigation can still be carried out when a global navigation satellite system signal is completely unavailable.

Description

Technical Field

[0001] The present invention relates to the technical field of obstacle area navigation, and in particular to an adaptive combined navigation method, system and terminal applied in obstacle areas. Background Art

[0002] Multi-sensor integrated navigation technology is widely used in aerospace, autonomous driving, railway detection and other navigation and positioning fields, among which the combination of global navigation satellite system (GNSS) and inertial navigation system (INS) is the most common. However, inertial sensors are susceptible to accelerometer and gyroscope noise, causing position and attitude estimation errors to deviate rapidly over time.

[0003] At present, when the global navigation satellite system signal is completely unavailable, relying solely on inertial sensors will cause errors to deviate rapidly, so various machine learning methods can be used to predict GNSS pseudo observations. Among them, with the assistance of the LSTM (Long Short-Term Memory Network) method, GNSS and INS integrated navigation can use historical navigation data without building a physical motion model, process data in highly nonlinear systems, and have the potential for real-time processing.

[0004] However, the current training and prediction of LSTM models are based on empirical parameters, resulting in their actual performance being largely constrained by the selection of hyperparameters, which makes it impossible for GNSS and INS integrated navigation to perform more accurate navigation with the assistance of the LSTM method.

[0005] Therefore, the prior art still needs to be improved and developed. Summary of the invention

[0006] The main purpose of the present invention is to provide an adaptive combined navigation method, system, terminal and computer-readable storage medium for application in obstacle areas, aiming to solve the problem that the LSTM model is trained and predicted based on empirical parameters in the prior art, resulting in its actual performance being largely subject to the selection of hyperparameters, thereby making it impossible for the GNSS and INS integrated navigation to perform navigation more accurately with the assistance of the LSTM (Long Short-Term Memory Network) method.

[0007] To achieve the above object, the present invention provides an adaptive integrated navigation method applied in an obstacle area, and the adaptive integrated navigation method applied in an obstacle area comprises the following steps:

[0008] Obtaining the specific force and angular rate output by the inertial sensor, inputting the specific force and the angular rate into the trained long short-term memory network, and outputting the predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on a particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain the trained long short-term memory network;

[0009] Acquire current data output by the inertial navigation system, and determine whether the predicted displacement increment complies with the target distribution according to the current data and the predicted displacement increment;

[0010] When the predicted displacement increment does not obey the target distribution, calculating the observation noise variance matrix according to the attenuation factor, and generating the target displacement increment according to the observation noise variance matrix and the predicted displacement increment;

[0011] A vehicle position is generated according to the target displacement increment and the current data.

[0012] Optionally, the training process of the long short-term memory network specifically includes:

[0013] Construct a training data set based on the specific force and angular rate pre-acquired by the inertial sensor and the corresponding GPS displacement increment;

[0014] According to the specific force and angular rate pre-acquired by the inertial sensor, the particle swarm optimization algorithm is used to optimize the hyperparameters of the long short-term memory network to obtain the optimized hyperparameters;

[0015] According to the optimized hyperparameters, the long short-term memory network is trained using the training data set to obtain a trained long short-term memory network.

[0016] Optionally, the method of optimizing the hyperparameters of the long short-term memory network using a particle swarm optimization algorithm based on the specific force and angular rate pre-acquired by the inertial sensor to obtain the optimized hyperparameters specifically includes:

[0017] The population is constructed according to the hyperparameters of the long short-term memory network and the particle swarm optimization algorithm, and the population is iteratively optimized according to the specific force and angular rate pre-acquired by the inertial sensor, and the optimal solution corresponding to the hyperparameters is found in each iterative optimization process;

[0018] When the number of iterations reaches a preset number, the optimal solution corresponding to the hyperparameter obtained in this iteration is used as the optimized hyperparameter.

[0019] Optionally, the iterative optimization of the population according to the specific force and angular rate pre-acquired by the inertial sensor, and finding the optimal solution corresponding to the hyperparameters in each iterative optimization process, specifically includes:

[0020] Iteratively optimizing the population according to the specific force and angular rate pre-acquired by the inertial sensor, wherein the hyperparameters include the number of neurons and the learning rate;

[0021] In each iteration, the number of neurons and particles corresponding to the learning rate are updated according to the update formula to obtain multiple local optimal solutions;

[0022] For each local optimal solution, a long short-term memory network is used to process the pre-acquired specific force and angular velocity to obtain a training output, and a particle swarm optimization algorithm is used to process the pre-acquired specific force and angular velocity to obtain a predicted output. A fitness function is calculated based on the training output and the predicted output, and an optimal solution for the number of neurons and the learning rate is selected from multiple local optimal solutions based on the fitness function.

[0023] Optionally, the acquiring current data output by the inertial navigation system and judging whether the predicted displacement increment obeys the target distribution according to the current data and the predicted displacement increment specifically includes:

[0024] Acquire current data output by the inertial navigation system, and calculate the prediction residual and standard error according to the current data output by the inertial navigation system and the predicted displacement increment;

[0025] A test statistic is constructed according to the prediction residual and the standard error, and whether the predicted displacement increment obeys a standard normal distribution is determined according to the test statistic.

[0026] Optionally, when the predicted displacement increment does not obey the target distribution, calculating the observation noise variance matrix according to the attenuation factor, and generating the target displacement increment according to the observation noise variance matrix and the predicted displacement increment specifically includes:

[0027] When the predicted displacement increment does not obey the target distribution, calculating the observation noise variance matrix according to the attenuation factor;

[0028] Generate state parameters and corresponding observation noise variance matrix according to the predicted displacement increment and the current data;

[0029] A gain matrix is ​​calculated according to the state parameters and the corresponding observation noise variance matrix, and a target displacement increment is generated according to the gain matrix, the state parameters and the predicted displacement increment.

[0030] Optionally, when the predicted displacement increment does not obey the target distribution, calculating the observation noise variance matrix according to the attenuation factor specifically includes:

[0031] When the predicted displacement increment does not obey the target distribution, constructing the covariance of the prediction residual;

[0032] An adaptive factor is constructed according to the attenuation factor, and an observation noise variance matrix is ​​calculated according to the adaptive factor and the covariance of the prediction residual.

[0033] In addition, to achieve the above-mentioned purpose, the present invention further provides an adaptive integrated navigation system applied in an obstacle area, wherein the adaptive integrated navigation system applied in an obstacle area comprises:

[0034] A network output module is used to obtain the specific force and angular rate output by the inertial sensor, input the specific force and the angular rate into the trained long short-term memory network, and output the predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on the particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain the trained long short-term memory network;

[0035] A judgment module, used for acquiring current data output by the inertial navigation system, and judging whether the predicted displacement increment complies with the target distribution according to the current data and the predicted displacement increment;

[0036] A target displacement increment generation module, used for calculating an observation noise variance matrix according to an attenuation factor when the predicted displacement increment does not obey a target distribution, and generating a target displacement increment according to the observation noise variance matrix and the predicted displacement increment;

[0037] The vehicle position generating module is used to generate the vehicle position according to the target displacement increment and the current data.

[0038] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an adaptive combined navigation program applied in an obstacle area, which is stored in the memory and can be run on the processor, and when the adaptive combined navigation program applied in the obstacle area is executed by the processor, the steps of the adaptive combined navigation method applied in the obstacle area as described above are implemented.

[0039] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an adaptive combined navigation program applied in an obstacle area, and when the adaptive combined navigation program applied in the obstacle area is executed by a processor, the steps of the adaptive combined navigation method applied in the obstacle area as described above are implemented.

[0040] In the present invention, the specific force and angular rate output by the inertial sensor are obtained, the specific force and the angular rate are input into the trained long short-term memory network, and the predicted displacement increment is output, wherein the hyperparameters of the long short-term memory network are optimized based on the particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain the trained long short-term memory network; the current data output by the inertial navigation system is obtained, and the predicted displacement increment is judged whether the predicted displacement increment obeys the target distribution according to the current data and the predicted displacement increment; when the predicted displacement increment does not obey the target distribution, the observation noise variance matrix is ​​calculated according to the attenuation factor, and the target displacement increment is generated according to the observation noise variance matrix and the predicted displacement increment; the vehicle position is generated according to the target displacement increment and the current data. In the present invention, the particle swarm optimization algorithm is used to determine the global optimal values ​​of the two hyperparameters of the number of neurons and the learning rate in the LSTM model, aiming to improve the training and prediction performance, and the attenuation factor is introduced to dynamically adjust the observation noise covariance matrix to reduce the influence of potential observation anomalies on the filter estimation, so that the present invention can still accurately navigate when the global navigation satellite system signal is completely unavailable. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of a preferred embodiment of the adaptive combined navigation method applied in an obstacle area of ​​the present invention;

[0042] Figure 2 It is a schematic diagram of the network architecture of the adaptive combined navigation method applied in the obstacle area of ​​the present invention;

[0043] Figure 3 It is a schematic diagram of the trajectory of a vehicle in a geodetic coordinate system in an adaptive combined navigation method in an obstacle area according to the present invention;

[0044] Figure 4 It is a schematic diagram of position increment, attitude increment and specific force increment during vehicle movement in the adaptive combined navigation method applied in an obstacle area of ​​the present invention;

[0045] Figure 5 It is a schematic diagram of a parameter optimization fitness curve in the adaptive combined navigation method applied in an obstacle area of ​​the present invention;

[0046] Figure 6 It is a schematic diagram of optimizing the number of neurons and learning rate parameters in the LSTM model in the adaptive combined navigation method applied in the obstacle area of ​​the present invention;

[0047] Figure 7 It is a schematic diagram of RMSE of LSTM model training based on PSO parameter iterative optimization in the adaptive integrated navigation method applied in the obstacle area of ​​the present invention;

[0048] Figure 8 It is a schematic diagram of GPS position increment and latitude, longitude and altitude errors predicted based on PSO-LSTM in the adaptive integrated navigation method applied in the obstacle area of ​​the present invention;

[0049] Fig. 9 It is a schematic diagram of comparison of test results based on PSO-LSTM inference in the adaptive combined navigation method applied in the obstacle area of ​​the present invention;

[0050] Fig.10 It is a schematic diagram of a carrier three-dimensional position error based on EKF, attenuated adaptive EKF and inertial dead reckoning in an adaptive combined navigation method applied in an obstacle area according to the present invention;

[0051] Fig.11 It is a schematic diagram of the improvement of positioning accuracy of AKF relative to EKF when the global positioning system is interrupted once every 100 seconds in the adaptive combined navigation method applied to the obstacle area of ​​the present invention;

[0052] Fig.12 It is a structural diagram of a preferred embodiment of the adaptive integrated navigation system applied in an obstacle area of ​​the present invention;

[0053] Fig.13 It is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention clearer and more specific, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0055] Multi-sensor integrated navigation technology is widely used in aerospace, autonomous driving, railway detection and other navigation and positioning fields, among which the combination of global navigation satellite system (GNSS) and inertial navigation system (INS) is the most common. However, inertial sensors are susceptible to accelerometer and gyroscope noise, causing position and attitude estimation errors to deviate rapidly over time.

[0056] At present, when the GNSS signal is completely unavailable, relying solely on inertial sensors will cause errors to deviate rapidly, so various machine learning methods can be used to predict GNSS pseudo observations. Among them, with the assistance of the LSTM method, GNSS and INS integrated navigation can make better use of historical navigation data without the need to establish a physical motion model, can process data in highly nonlinear systems, and has the potential for real-time processing.

[0057] However, the current training and prediction of LSTM models are based on empirical parameters, resulting in its actual performance being largely constrained by the selection of hyperparameters, which makes it impossible for GNSS and INS integrated navigation to perform more accurate navigation with the assistance of the LSTM (Long Short-Term Memory Network) method.

[0058] In response to one or more of the above problems, the present application obtains the specific force and angular rate output by the inertial sensor, inputs the specific force and the angular rate into a trained long short-term memory network, and outputs a predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on a particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain a trained long short-term memory network; current data output by the inertial navigation system is obtained, and whether the predicted displacement increment obeys a target distribution is determined based on the current data and the predicted displacement increment; when the predicted displacement increment does not obey the target distribution, an observation noise variance matrix is ​​calculated based on an attenuation factor, and a target displacement increment is generated based on the observation noise variance matrix and the predicted displacement increment; and a vehicle position is generated based on the target displacement increment and the current data.

[0059] The adaptive integrated navigation method applied in an obstacle area according to a preferred embodiment of the present invention is as follows: Figure 1 As shown, the adaptive integrated navigation method applied in the obstacle area includes the following steps:

[0060] Step S10, obtaining the specific force and angular rate output by the inertial sensor, inputting the specific force and the angular rate into the trained long short-term memory network, and outputting the predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on the particle swarm optimization algorithm to obtain the optimized hyperparameters, and based on the optimized hyperparameters, the long short-term memory network is trained to obtain the trained long short-term memory network.

[0061] Specifically, in the present application, when GPS is available, the corresponding navigation can be realized by the GPS / INS loosely coupled state model and observation model, wherein the corresponding loosely coupled GPS / INS integrated state space model and observation model are established in the "north-east-down" coordinate system, and the position and velocity difference between the GPS and INS inertial reckoning is used as the observation input. The corresponding discretized loosely coupled state and observation model can be expressed as follows:

[0062]

[0063] Among them, X k,k-1 Represents 9 state parameters, including position error, velocity error, attitude error, gyroscope drift, accelerometer bias, lever arm error between INS and GPS antenna, and time synchronization error; is the state transition matrix; Z k Represents the position difference between GPS and INS; H k is the time observation matrix; W k-1,k-1 and V k represents the state noise vector and the observation noise vector, and the corresponding covariance matrix can be written as Q k-1,k-1 and R k . The corresponding filter is EKF.

[0064] Therefore, when GPS is available, its output is usually used as the main reference; however, when GPS is not available, data fusion between the two sensors cannot be performed, but the corresponding processing can be achieved through the present invention.

[0065] In the present invention, a trained long short-term memory network is used instead of GPS to achieve corresponding position prediction.

[0066] Furthermore, the training process of the long short-term memory network specifically includes:

[0067] Construct a training data set based on the specific force and angular rate pre-acquired by the inertial sensor and the corresponding GPS displacement increment;

[0068] According to the specific force and angular rate pre-acquired by the inertial sensor, the particle swarm optimization algorithm is used to optimize the hyperparameters of the long short-term memory network to obtain the optimized hyperparameters;

[0069] According to the optimized hyperparameters, the long short-term memory network is trained using the training data set to obtain a trained long short-term memory network.

[0070] Specifically, the LSTM model has a simple structure and few parameters. During training, the LSTM network uses the back-propagation algorithm to adjust weights while minimizing the loss function through gradient descent. During prediction, the network uses its memory cells to retain key historical information, thereby effectively capturing long-term dependencies in time series data. This ability makes LSTM particularly effective in processing sequence data. The LSTM network can be represented by memory cells connected at adjacent time steps, and its basic results are as follows Figure 2 As shown. Figure 2 As can be seen, each rectangular block represents a storage block, which is represented by a forget gate f t An input gate i t , output gate O t and a state storage block C t-1 Composition, and g t To modify C t-1 ;X t and h t-1Represents the input at the current time. In LSTM, the forget gate determines the part of the state storage block that needs to be forgotten. The vector that needs to be retained in the state storage block is determined by the sigmoid function (δ) of the input gate and the tanh activation function. The final output h t is the updated state memory block C t And the output gate gets the final output.

[0071] In the LSTM network, the learning rate controls the step size of the weight update during the gradient descent process. Setting it too high or too low will lead to unstable training; the number of neurons in the hidden layer directly affects the representation ability of the network. The particle swarm optimization algorithm has advantages in convergence speed, adaptability and computational complexity. Therefore, in the present invention, before using the LSTM network to predict the position and velocity increments during GPS interruption, PSO (particle swarm optimization algorithm) is first applied to optimize these two hyperparameters, thereby improving the training efficiency and prediction performance of the model.

[0072] After the optimization is completed, LSTM is trained based on the optimized parameters.

[0073] Furthermore, the hyperparameters of the long short-term memory network are optimized using a particle swarm optimization algorithm based on the specific force and angular rate pre-acquired by the inertial sensor to obtain the optimized hyperparameters, which specifically include:

[0074] The population is constructed according to the hyperparameters of the long short-term memory network and the particle swarm optimization algorithm, and the population is iteratively optimized according to the specific force and angular rate pre-acquired by the inertial sensor, and the optimal solution corresponding to the hyperparameters is found in each iterative optimization process;

[0075] When the number of iterations reaches a preset number, the optimal solution corresponding to the hyperparameter obtained in this iteration is used as the optimized hyperparameter.

[0076] Specifically, in this application, the number of neurons and the learning rate are represented by symbols α and ε, respectively. The initial dimension of the population and the number of particles in each population are represented by D and N, respectively; the initial position information x i and speed information v i The two parameters (α, ε) corresponding to each particle are randomly obtained within a predetermined range, where i = 1, 2, ..., N. In each iteration, the particle updates its velocity and position by the following formula:

[0077]

[0078] in represents d particles in iterations i to k (d = 1, ..., D), c1 and c2 represent individual learning factors and group learning factors respectively; r1 and r2 represent random numbers in the interval [0, 1], and r1 and r2 increase the randomness of the search; represents d particles in iterations i to k-1 (d = 1, ..., D); it should be noted that the position and velocity of the particles should be limited to a certain range to avoid falling into the local optimum or missing the global optimal solution; and w represents the inertia weight, which is used to adjust the search range of the solution space; Represents the global optimal solution.

[0079] In order to select the global optimal solution (α, ε) of two hyperparameters from the particle swarm, the fitness function value of each particle must be calculated in each iteration. The fitness function F fitness The construction is as follows:

[0080]

[0081] Where Num represents the number of training samples, y j is the training output of the network, Represents the predicted output. Therefore, in each iteration, the optimal solution is found for the particles in the population based on the calculated fitness function value.

[0082] Furthermore, the iterative optimization of the population according to the specific force and angular rate pre-acquired by the inertial sensor and finding the optimal solution corresponding to the hyperparameters in each iterative optimization process specifically includes:

[0083] Iteratively optimizing the population according to the specific force and angular rate pre-acquired by the inertial sensor, wherein the hyperparameters include the number of neurons and the learning rate;

[0084] In each iteration, the number of neurons and particles corresponding to the learning rate are updated according to the update formula to obtain multiple local optimal solutions;

[0085] For each local optimal solution, a long short-term memory network is used to process the pre-acquired specific force and angular velocity to obtain a training output, and a particle swarm optimization algorithm is used to process the pre-acquired specific force and angular velocity to obtain a predicted output. A fitness function is calculated based on the training output and the predicted output, and an optimal solution for the number of neurons and the learning rate is selected from multiple local optimal solutions based on the fitness function.

[0086] Specifically, in the present application, when iterating, the long short-term memory network processes the pre-acquired specific force and angular velocity to obtain the training output and the particle swarm optimization algorithm processes the pre-acquired specific force and angular velocity to obtain the predicted output to calculate the corresponding fitness function, thereby selecting the optimal solution for the number of neurons and the learning rate from the multiple local optimal solutions generated.

[0087] Step S20: acquiring current data output by the inertial navigation system, and judging whether the predicted displacement increment complies with the target distribution according to the current data and the predicted displacement increment.

[0088] Specifically, in this application, considering that the GPS pseudo-position predicted by the LSTM network based on PSO may contain outliers, when performing filtering estimation, relying solely on the prediction residual may not be sufficient to fully detect anomalies in the observation model, which may lead to unreliable estimation results. Therefore, it is necessary to construct a test statistic based on the prediction residual information to identify observation anomalies.

[0089] Further, the acquiring current data output by the inertial navigation system and judging whether the predicted displacement increment obeys the target distribution according to the current data and the predicted displacement increment specifically include:

[0090] Acquire current data output by the inertial navigation system, and calculate the prediction residual and standard error according to the current data output by the inertial navigation system and the predicted displacement increment;

[0091] A test statistic is constructed according to the prediction residual and the standard error, and whether the predicted displacement increment obeys a standard normal distribution is determined according to the test statistic.

[0092] Specifically, in this application, the test statistic follows a standard normal distribution and can be expressed as:

[0093]

[0094] in represents the prediction residual, and the corresponding standard error is It can be calculated by the following formula:

[0095]

[0096]

[0097] in, is the one-step predicted value of the state. represents the diagonal elements of the observation noise covariance matrix, represents the one-step prediction value of the state covariance, and i is the row element of the observation matrix. A test statistic is constructed based on the obtained prediction residual and the standard error. If the test statistic exceeds the threshold, it does not obey the normal distribution at this time, indicating that there is an observation abnormality.

[0098] Step S30: When the predicted displacement increment does not conform to the target distribution, calculate the observation noise variance matrix according to the attenuation factor, and generate a target displacement increment based on the observation noise variance matrix and the predicted displacement increment.

[0099] For the phenomenon of abnormal observations, it is necessary to introduce a gradual adaptation factor to dynamically adjust the observation noise covariance matrix to reduce the impact of abnormal observations.

[0100] Further, when the predicted displacement increment does not conform to the target distribution, calculating the observation noise variance matrix according to the attenuation factor, and generating a target displacement increment based on the observation noise variance matrix and the predicted displacement increment specifically includes:

[0101] When the predicted displacement increment does not conform to the target distribution, calculate the observation noise variance matrix according to the attenuation factor;

[0102] Generate state parameters and the corresponding observation noise variance matrix based on the predicted displacement increment and the current data;

[0103] Calculate the gain matrix according to the state parameters and the corresponding observation noise variance matrix, and generate a target displacement increment based on the gain matrix, the state parameters, and the predicted displacement increment.

[0104] Specifically, as shown in the following formula, omitting the subscript i, the covariance of the prediction residual can be expressed as:

[0105]

[0106] where R k is the observation noise variance matrix. The variance of the prediction residual represents the ensemble average of the random sequence and can be replaced by the time average in the discrete equation. Then the formula corresponding to the covariance of the prediction residual can be rewritten as:

[0107]

[0108] Since the observation noise is mainly caused by external factors, it is prone to significant changes. Fixed observation noise cannot effectively adapt to the dynamic system, and it is necessary to perform adaptive adjustment based on historical and current observation information. The gradual adaptation factor can make full use of historical observation noise information and prediction residual information. Therefore, introduce an attenuation factor b (0 < b < 1) and express the adaptation factor as:

[0109]

[0110] Then the formula corresponding to the covariance of the prediction residual after rewriting can be written as:

[0111]

[0112] It can be seen that the adaptive factor increases the contribution of the current observation information to the state estimation.

[0113] Then, according to the predicted displacement increment and the current data, the state parameters and the corresponding observation noise variance matrix are generated, as shown in the following formula:

[0114]

[0115] The gain matrix is ​​calculated by the state parameters and the corresponding observation noise variance matrix, and the target displacement increment is generated according to the gain matrix, the state parameters and the predicted displacement increment, which is specifically expressed as follows:

[0116]

[0117] Where K k represents the gain matrix, P k represents the estimated state covariance matrix, and is the estimated value after filtering, i.e., the target displacement increment.

[0118] Furthermore, when the predicted displacement increment does not obey the target distribution, the observation noise variance matrix is ​​calculated according to the attenuation factor, specifically including:

[0119] When the predicted displacement increment does not obey the target distribution, constructing the covariance of the prediction residual;

[0120] An adaptive factor is constructed according to the attenuation factor, and an observation noise variance matrix is ​​calculated according to the adaptive factor and the covariance of the prediction residual.

[0121] Specifically, in this application, when the predicted displacement increment does not obey the target distribution, the covariance of the predicted residual is constructed, that is, the corresponding Then, γ is constructed by the attenuation factor b. k , to calculate the corresponding observation noise variance matrix.

[0122] Step S40: Generate a vehicle position according to the target displacement increment and the current data.

[0123] Specifically, after obtaining the standard displacement increment and the current data, the corresponding vehicle position is generated through EKF.

[0124] Furthermore, the present invention has been experimented accordingly. The corresponding GPS and INS data acquisition system includes a Leica 1200 GPS receiver and a SPAN-CPT integrated navigation and positioning system. The GPS sampling rate is 10 Hz, while the operating frequency of the inertial sensor is 100 Hz. The total duration of the data acquisition is about 1810 seconds. The differential positioning between the rover and the static base station provides a reference position. The trajectory of the aircraft and its starting and ending points are as follows Figure 3 As shown. Table 1 lists the bias and random walk parameters of the accelerometer and gyroscope. The initial position error of the inertial navigation system is set to 1 meter in the north and east directions, and 1.5 meters in the south direction. The initial velocity error is defined as 0.5 m / s, while the initial attitude errors of pitch, roll, and heading are set to 1°, 1°, and 3°, respectively. The standard deviation of the GPS positioning error in the two horizontal directions is set to 0.5 meters and 1 meter in the downlink direction. Using the parameters in Table 1 and the above settings, the initial state covariance and observation noise covariance of the extended Kalman filter are established.

[0125] Table 1. Bias and random walk parameters of inertial sensors

[0126]

[0127] The mechanization of inertial navigation is to calculate the vehicle's speed and displacement increment by integrating the output of the accelerometer and gyroscope. The specific force and angular velocity output by the inertial sensor are nonlinearly related to the displacement increment. Therefore, the vehicle's motion characteristics, such as straight-line motion, turning, climbing, and acceleration, are also reflected in the sensor's output information. Figure 4 Shows along Figure 3 The displacement increments of the vehicle in the longitude, latitude and altitude directions and the sensor output information for the vehicle traveling along the trajectory shown. It can be seen that when the vehicle turns, the heading rate will change significantly, and when accelerating or decelerating in a straight line, the specific force output will change significantly.

[0128] In order to verify that the proposed method can effectively predict the displacement increment of the vehicle during the GPS signal interruption, a simulation was performed in which the GPS was available from 1 to 1200 seconds and unavailable from 1201 to 1810 seconds. According to the available time period, the specific force and angular rate output by the inertial sensor were used as the input of LSTM training, and the corresponding reference displacement increment was used as the training output. First, the PSO algorithm was used to iteratively optimize the number of neurons and the learning rate parameters to obtain the global optimal value. Figure 5 The iterative results of the fitting function values ​​in the latitude, longitude and altitude directions are shown. It can be seen that after six PSO optimization iterations, the LSTM training adaptation values ​​in the three directions are all consistent. At this time, the global optimal values ​​of the corresponding number of neurons and learning rate are used for the training and prediction of the LSTM model. These values ​​are shown in Table 2.

[0129] Table 2. Bias and random walk parameters of inertial sensors

[0130]

[0131]

[0132] Figure 6 The optimization results of the two hyperparameters, the number of neurons and the learning rate, during each PSO iteration of the LSTM model are shown. This shows that after six iterations, the global optimal positions and optimal values ​​of the two hyperparameters have been determined. This also shows that simply relying on empirical values ​​to replace the hyperparameters of the LSTM model does not ensure effective training and prediction performance. Therefore, using the PSO algorithm for global hyperparameter optimization can improve the effect of LSTM.

[0133] When training an LSTM model based on the optimized hyperparameters of the number of neurons and the learning rate, an appropriate number of iterations must be set to ensure that the model has enough opportunities to learn the data features while avoiding overfitting. Figure 7 As shown in Figure 1, as the number of iterations increases, the root mean square error of the LSTM model training output tends to stabilize. After about 22 iterations, the training root mean square error between adjacent iterations is less than 1 mm.

[0134] To further verify the accuracy of the method for predicting vehicle displacement increments, the specific force and angular velocity outputs of the first 1200 seconds were used as the input and output of the training set. The reference true value of the training set displacement increment is marked as 1#, while the predicted increment of the training set is marked as 3#. The predicted increment of the subsequent 610 seconds of the test set is denoted as 4#, and the corresponding reference true value is denoted as 2#. The displacement increments of the training set and the test set and their corresponding predicted increments are plotted on Figure 8 As can be seen from the figure, whether it is a training set or a test set, the displacement increment predicted by the PSO-LSTM model (i.e., the corresponding method of the present invention) is very close to the actual increment. Figure 8 The corresponding error sequence is plotted in (b), where the errors of the training set and the test set are concentrated between -0.5 meters and 0.5 meters, and only the prediction error during turning fluctuates greatly. In the test set, the displacement increment error in the longitude direction is relatively large in certain periods, close to 1 meter, and the height error is even close to 2 meters. Table 2 also lists the root mean square values ​​of the incremental prediction errors in the longitude, latitude and altitude directions of the training set and the test set. The effective values ​​of the training set are between 0.122 and 0.151 meters, while the effective values ​​of the test set are between 0.154 and 0.186 meters. This shows that when the GPS signal is not available, the method proposed in the present invention can effectively capture the motion characteristics of the vehicle, and the predicted displacement increment is within the decimeter range.

[0135] When GPS signals are unavailable, relying solely on inertial sensors to estimate vehicle position results in errors that accumulate rapidly over time. Similarly, using the proposed method to predict the vehicle displacement increments based on the specific forces and angular velocities output by the inertial sensors also results in error accumulation. Therefore, the position calculated when the GPS signal was interrupted based on the predicted increments cannot be directly fused with the position derived from inertial navigation. Therefore, during the EKF measurement update process, the adaptive factors are calculated based on the predicted residuals. The method of the present invention allows for dynamic adjustment of observation noise using historical observation information when GPS signals are available, thereby reducing or mitigating the impact of accumulated or significant errors on state estimates.

[0136] Fig. 9 The traditional EKF ( Fig. 9 (a)) and the decay adaptive EKF ( Fig. 9 (b)) The chi-square statistic. The degrees of freedom are set to 3 and the significance levels are set to 0.1 and 0.01. At a significance level of 10%, the vehicle position calculated based on the predicted displacement increment is closer to the actual vehicle position. Otherwise, the observation noise needs to be adjusted before filtering estimation. Fig. 9 As shown, the vehicle positions calculated based on the predicted increments do have outliers, which cannot be fully verified by the chi-square test.

[0137] Fig.10 The figure shows the vehicle 3D position error calculated using EKF, adaptive EKF (AKF), and pure inertial navigation (SINS) during the period when GPS signal is unavailable. It can be seen from the figure that although both the inertial navigation position estimate and the vehicle pseudo-position based on the predicted displacement increment will have error accumulation, the error divergence rate of the former is significantly higher than that of the latter. When the GPS unavailability time exceeds 100 seconds, the position error of the inertial navigation exceeds several hundred meters, which cannot be fully displayed in the figure due to the large error. However, the fusion positioning using the predicted vehicle pseudo-position can still correct the position error of the inertial navigation well. As the interruption time increases, the comprehensive positioning error also shows an increasing trend. In addition, it can be seen from the figure that after introducing the adaptive factor, the positioning error is reduced compared with the scheme using only EKF. This shows that the dynamic adjustment of observation noise reduces the impact of observation anomalies, thereby improving positioning accuracy.

[0138] In order to further analyze the degree of improvement in positioning accuracy of the AKF scheme compared with the EKF, the position errors in the north, east and down directions are recorded every 100 seconds, and the root mean square error (RMSE) of the two methods is calculated. The statistical results and improvement percentages are shown in Tables 3 and Fig.11. It can be seen that as the GPS unavailable time increases, the positioning errors in the three directions increase to varying degrees. The positioning error of the AKF scheme is lower than that of the EKF scheme. Among the six statistical results, the maximum improvement in the north direction is about 23.6%, the maximum improvement in the east direction is about 18.3%, and the maximum improvement in the south direction is about 22.7%. The average effective values ​​of the positioning errors of the EKF scheme in each period are 7.163 meters, 11.020 meters and 20.178 meters, respectively, while the errors of the AKF scheme are 5.664 meters, 9.227 meters and 16.294 meters, respectively. These results further demonstrate the effectiveness of the attenuated adaptive filtering proposed in the present invention in comprehensive positioning, and improve the reliability of machine learning-assisted positioning during GPS unavailability.

[0139] Table 3. Effective values ​​of positioning errors in the north, east and south directions every 100 seconds (unit: meters)

[0140]

[0141] The present invention obtains the specific force and angular rate output by the inertial sensor, inputs the specific force and the angular rate into the trained long short-term memory network, and outputs the predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on the particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain the trained long short-term memory network; the current data output by the inertial navigation system is obtained, and the predicted displacement increment is judged whether the predicted displacement increment obeys the target distribution according to the current data and the predicted displacement increment; when the predicted displacement increment does not obey the target distribution, the observation noise variance matrix is ​​calculated according to the attenuation factor, and the target displacement increment is generated according to the observation noise variance matrix and the predicted displacement increment; the vehicle position is generated according to the target displacement increment and the current data. In the present invention, the particle swarm optimization algorithm is used to determine the global optimal values ​​of the two hyperparameters of the number of neurons and the learning rate in the LSTM model, aiming to improve the training and prediction performance, and the attenuation factor is introduced to dynamically adjust the observation noise covariance matrix to reduce the influence of potential observation anomalies on the filter estimation, so that the present invention can still accurately navigate when the global navigation satellite system signal is completely unavailable.

[0142] Furthermore, if Fig.12 As shown, based on the above-mentioned adaptive integrated navigation method applied in an obstacle area, the present invention also provides an adaptive integrated navigation system applied in an obstacle area, wherein the adaptive integrated navigation system applied in an obstacle area includes:

[0143] The network output module 121 is used to obtain the specific force and angular rate output by the inertial sensor, input the specific force and the angular rate into the trained long short-term memory network, and output the predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on the particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain the trained long short-term memory network;

[0144] A judgment module 122 is used to obtain current data output by the inertial navigation system, and judge whether the predicted displacement increment complies with the target distribution according to the current data and the predicted displacement increment;

[0145] A target displacement increment generation module 123, configured to calculate an observation noise variance matrix according to an attenuation factor when the predicted displacement increment does not conform to a target distribution, and to generate a target displacement increment according to the observation noise variance matrix and the predicted displacement increment;

[0146] The vehicle position generating module 124 is used to generate the vehicle position according to the target displacement increment and the current data.

[0147] Furthermore, if Fig.13 As shown, based on the above-mentioned adaptive combined navigation method and system applied in obstacle areas, the present invention also provides a terminal accordingly, and the terminal includes a processor 10, a memory 20 and a display 30. Fig.13 Only some components of the terminal are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0148] The memory 20 may be an internal storage unit of the terminal in some embodiments, such as a hard disk or memory of the terminal. The memory 20 may also be an external storage device of the terminal in other embodiments, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Further, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed in the terminal, such as the program code of the installation terminal, etc. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, the memory 20 stores an adaptive combined navigation program 40 applied in an obstacle area, and the adaptive combined navigation program 40 applied in the obstacle area can be executed by the processor 10, thereby realizing the adaptive combined navigation method applied in the obstacle area in the present application.

[0149] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the adaptive combined navigation method applied in an obstacle area.

[0150] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 30 is used to display information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0151] In one embodiment, when the processor 10 executes the adaptive combined navigation program 40 for use in an obstacle area in the memory 20 , the steps of the above adaptive combined navigation method for use in an obstacle area are implemented.

[0152] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an adaptive combined navigation program applied in an obstacle area, and when the adaptive combined navigation program applied in the obstacle area is executed by a processor, the steps of the adaptive combined navigation method applied in the obstacle area as described above are implemented.

[0153] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, method, article or terminal including the element.

[0154] Of course, those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the program can be stored in a computer-readable storage medium that can be read by a computer, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0155] It should be understood that the application of the present invention is not limited to the above examples. For ordinary technicians in this field, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. An adaptive integrated navigation method applied in an obstacle area, characterized in that: The adaptive integrated navigation method applied in the obstacle area comprises: Obtaining the specific force and angular rate output by the inertial sensor, inputting the specific force and the angular rate into the trained long short-term memory network, and outputting the predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on a particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain the trained long short-term memory network; Acquire current data output by an inertial navigation system, and determine whether the predicted displacement increment complies with a target distribution according to the current data and the predicted displacement increment; When the predicted displacement increment does not obey the target distribution, calculating the observation noise variance matrix according to the attenuation factor, and generating the target displacement increment according to the observation noise variance matrix and the predicted displacement increment; A vehicle position is generated according to the target displacement increment and the current data.

2. The adaptive integrated navigation method applied in obstacle areas according to claim 1, characterized in that: The training process of the long short-term memory network specifically includes: Construct a training data set based on the specific force and angular rate pre-acquired by the inertial sensor and the corresponding GPS displacement increment; According to the specific force and angular rate pre-acquired by the inertial sensor, the particle swarm optimization algorithm is used to optimize the hyperparameters of the long short-term memory network to obtain the optimized hyperparameters; According to the optimized hyperparameters, the long short-term memory network is trained using the training data set to obtain a trained long short-term memory network.

3. The adaptive integrated navigation method applied in obstacle areas according to claim 2 is characterized in that: The method uses a particle swarm optimization algorithm to optimize the hyperparameters of the long short-term memory network according to the specific force and angular rate pre-acquired by the inertial sensor to obtain the optimized hyperparameters, which specifically include: The population is constructed according to the hyperparameters of the long short-term memory network and the particle swarm optimization algorithm, and the population is iteratively optimized according to the specific force and angular rate pre-acquired by the inertial sensor, and the optimal solution corresponding to the hyperparameters is found in each iterative optimization process; When the number of iterations reaches a preset number, the optimal solution corresponding to the hyperparameter obtained in this iteration is used as the optimized hyperparameter.

4. The adaptive integrated navigation method applied in obstacle areas according to claim 3 is characterized in that: The iterative optimization of the population according to the specific force and angular rate pre-acquired by the inertial sensor and finding the optimal solution corresponding to the hyperparameters in each iterative optimization process specifically includes: Iteratively optimizing the population according to the specific force and angular rate pre-acquired by the inertial sensor, wherein the hyperparameters include the number of neurons and the learning rate; In each iteration, the number of neurons and particles corresponding to the learning rate are updated according to the update formula to obtain multiple local optimal solutions; For each local optimal solution, a long short-term memory network is used to process the pre-acquired specific force and angular velocity to obtain a training output, and a particle swarm optimization algorithm is used to process the pre-acquired specific force and angular velocity to obtain a predicted output. A fitness function is calculated based on the training output and the predicted output, and an optimal solution for the number of neurons and the learning rate is selected from multiple local optimal solutions based on the fitness function.

5. The adaptive integrated navigation method applied in obstacle areas according to claim 1, characterized in that: The acquiring current data output by the inertial navigation system and judging whether the predicted displacement increment complies with the target distribution according to the current data and the predicted displacement increment specifically includes: Acquire current data output by the inertial navigation system, and calculate the prediction residual and standard error according to the current data output by the inertial navigation system and the predicted displacement increment; A test statistic is constructed according to the prediction residual and the standard error, and whether the predicted displacement increment obeys a standard normal distribution is determined according to the test statistic.

6. The adaptive integrated navigation method applied in obstacle areas according to claim 1, characterized in that: When the predicted displacement increment does not obey the target distribution, calculating the observation noise variance matrix according to the attenuation factor, and generating the target displacement increment according to the observation noise variance matrix and the predicted displacement increment specifically includes: When the predicted displacement increment does not obey the target distribution, calculating the observation noise variance matrix according to the attenuation factor; Generate state parameters and corresponding observation noise variance matrix according to the predicted displacement increment and the current data; A gain matrix is ​​calculated according to the state parameters and the corresponding observation noise variance matrix, and a target displacement increment is generated according to the gain matrix, the state parameters and the predicted displacement increment.

7. The adaptive integrated navigation method applied in obstacle areas according to claim 6, characterized in that: When the predicted displacement increment does not obey the target distribution, the observation noise variance matrix is ​​calculated according to the attenuation factor, specifically including: When the predicted displacement increment does not obey the target distribution, constructing the covariance of the prediction residual; An adaptive factor is constructed according to the attenuation factor, and an observation noise variance matrix is ​​calculated according to the adaptive factor and the covariance of the prediction residual.

8. An adaptive integrated navigation system used in an obstacle area, characterized in that: The adaptive integrated navigation system applied in the obstacle area comprises: A network output module is used to obtain the specific force and angular rate output by the inertial sensor, input the specific force and the angular rate into the trained long short-term memory network, and output the predicted displacement increment, wherein the hyperparameters of the long short-term memory network are optimized based on the particle swarm optimization algorithm to obtain the optimized hyperparameters, and the long short-term memory network is trained based on the optimized hyperparameters to obtain the trained long short-term memory network; A judgment module, used for acquiring current data output by the inertial navigation system, and judging whether the predicted displacement increment complies with the target distribution according to the current data and the predicted displacement increment; A target displacement increment generation module, used for calculating an observation noise variance matrix according to an attenuation factor when the predicted displacement increment does not obey a target distribution, and generating a target displacement increment according to the observation noise variance matrix and the predicted displacement increment; The vehicle position generating module is used to generate the vehicle position according to the target displacement increment and the current data.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and an adaptive combined navigation program applied in an obstacle area that is stored in the memory and can be run on the processor. When the adaptive combined navigation program applied in the obstacle area is executed by the processor, the steps of the adaptive combined navigation method applied in the obstacle area as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an adaptive combined navigation program applied in an obstacle area, and when the adaptive combined navigation program applied in an obstacle area is executed by a processor, the steps of the adaptive combined navigation method applied in an obstacle area as described in any one of claims 1-7 are implemented.

Images