A neural network assisted vehicle adaptive autonomous navigation method

Abstract

The application provides a neural network assisted vehicle adaptive autonomous navigation method, which comprises the following steps: designing an improved LSTM model for multi-parameter parallel prediction; collecting inertial measurement unit data and performing inertial navigation calculation; performing inertial / satellite integrated navigation calculation when satellite navigation is effective; training the designed parallel LSTM model based on the integrated navigation result; predicting vehicle speed and attitude information based on the trained machine learning model when satellite navigation is ineffective; and using variational Bayes based on the prediction result of the machine learning model to perform adaptive integrated navigation when satellite navigation is ineffective. The application overcomes the problem of insufficient cumulative accuracy of vehicle navigation error when only relying on inertial navigation without reference information, and can realize full autonomous and high-precision navigation and positioning of the vehicle, which has important practical application value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of inertial navigation, and particularly relates to a neural network assisted vehicle adaptive autonomous navigation method. BACKGROUND

[0002] High-precision navigation and positioning information is a prerequisite for vehicle precise control, reasonable path planning and safe automatic driving. At present, vehicle navigation and positioning mainly adopts inertial / satellite integrated navigation. However, the precision and reliability of satellite navigation directly depend on satellite signal conditions, and cannot be used in scenes without satellite signals such as underground parking lots and tunnels, and are not stable enough in environments such as building shielding. At this time, the error of inertial navigation based on integral solution accumulates with time. Especially in the vehicle scene, due to the volume and cost restrictions, a small low-cost inertial navigation system is used, and the sensor precision is limited, and the error accumulation characteristics are particularly obvious. Therefore, it is urgent to design a vehicle full autonomous navigation method that does not depend on external reference information such as satellite navigation. SUMMARY

[0003] The present application aims to at least solve one of the technical problems existing in the prior art.

[0004] To this end, the present application provides a neural network assisted vehicle adaptive autonomous navigation method.

[0005] The technical solution of the present application is as follows:

[0006] According to an aspect, a neural network assisted vehicle adaptive autonomous navigation method is provided, which includes:

[0007] Step one, design an improved LSTM model for multi-parameter parallel prediction, including:

[0008] The improved LSTM model includes four sub-models:

[0009] 1) Sub-model 1: horizontal velocity prediction model

[0010] The horizontal velocity prediction model predicts eastward and northward velocities, and the neural network framework based thereon is LSTM, and the model input and output are designed as follows:

[0011] Model input: gyro output excluding harmful components such as earth rotation angular velocity Attitude quaternion Atti calculated by pure inertial navigation insQ , navigation system acceleration f excluding gravity acceleration components n ; wherein the navigation system acceleration f n is calculated by combining the original accelerometer output with the attitude calculated by pure inertial navigation;

[0012] Model output: eastward velocity and northward velocity;

[0013] 2) Submodel 2: Skyward velocity prediction model

[0014] The model framework and model input of the skyward velocity prediction model are the same as those of submodel 1, and the model output is the skyward velocity;

[0015] 3) Submodel 3: Horizontal attitude prediction model

[0016] The horizontal attitude prediction model predicts the pitch angle and roll angle, and the neural network framework and model input are the same as those of submodel 1, and the model output is the pitch angle and roll angle;

[0017] 4) Submodel 4: Heading prediction model

[0018] The heading prediction model predicts the heading angle in quaternion form, and the neural network framework and model input are the same as those of submodel 1, and the model output is the attitude in quaternion form;

[0019] Step two, collect inertial measurement unit data and perform inertial navigation solution;

[0020] Step three, acquire satellite navigation data and perform integrated navigation solution when GNSS is effective, including:

[0021] Perform multiple GNSS-effective car tests, and perform inertial / satellite integrated navigation based on the inertial measurement unit data and GNSS data collected in step two to calculate the reference attitude quaternion Atti referQ and the reference navigation system velocity, which are used to train the improved LSTM model;

[0022] Step four, train the improved LSTM model designed in step one based on the inertial navigation solution results in step two and the integrated navigation results in step three;

[0023] Step five, based on the improved LSTM model trained in step four, predict vehicle attitude and velocity information using inertial navigation data in the case of GNSS invalidity;

[0024] Step six, use variational Bayesian filtering based on the prediction results in step five for adaptive integrated navigation.

[0025] Further, when designing the submodel 4, a quaternion form loss function is designed as follows:

[0026]

[0027] where Atti predQ = [q 0p q 1p q 2p q 3p ] Tand Atti referQ = [q 0t q 1t q 2t q 3t ] T are quaternion forms of predicted and reference attitude respectively; AQ = [Aq0Aq1Aq2Aq3] T is the difference between the predicted attitude Atti predQ and the reference attitude Atti referQ , and the calculation formula is:

[0028]

[0029] The reference attitude here is calculated by inertial / satellite integrated navigation.

[0030] Further, the step three specifically includes:

[0031] Collecting inertial measurement unit data, performing inertial navigation solution, obtaining gyro output excluding harmful components Attitude quaternion Atti insQ of pure inertial navigation solution; and based on Atti insQ and accelerometer output, calculating navigation system acceleration f n excluding gravity acceleration component, for input of parallel LSTM model.

[0032] Further, the step six specifically includes:

[0033] Based on the inertial measurement unit data collected in step two, using variational Bayesian filtering for integrated navigation with the prediction result of step five as reference information, the step execution process is as follows:

[0034] 6.1 Establish an integrated navigation state space model, including state equation and measurement equation;

[0035] 6.2 Initialize the parameters of the integrated navigation algorithm;

[0036] 6.3 Perform time update;

[0037] 6.4 Perform measurement update.

[0038] Further, in step 6.1:

[0039] The state equation is X k = Φ k,k-1 X k-1 + w k-1 , and the state quantity X includes north-east misalignment angle North-east velocity error δv n = [δV n δV udelta V e ] T latitude error altitude error δh, longitude error δλ, body-vehicle gyro drift ε b = [ε x ε y ε z ] T body-vehicle accelerometer bias body-vehicle lever arm error L out time delay T d ;

[0040] The measurement equation is Z k = H k X k + V k , where Z k is the difference between the improved LSTM model predicted attitude and velocity information and the inertial navigation calculation results.

[0041] Further, in step 6.2:

[0042] Initialize the parameters of the combined navigation algorithm based on the variational Bayesian filter, including the state quantity estimation result m k , the estimation mean square error P k , the system noise variance matrix Q k , the initial value of the shape parameter of the inverse Gamma distribution α0, and the initial value of the inverse scale parameter of the inverse Gamma distribution β0.

[0043] Further, in step 6.3, the time update is performed according to the following formula:

[0044]

[0045]

[0046]

[0047]

[0048] where Φ k / k-1 is calculated according to the inertial navigation error equation, ρ i ∈(0.9,1], i=1,…,d is the attenuation coefficient, and are the shape parameter and inverse scale parameter of the inverse Gamma distribution, respectively, and d represents the dimension of the measurement.

[0049] Further, in step 6.4, the measurement update is performed in an iterative manner:

[0050] First, initialize the parameters: Then perform the iterative calculation as follows:

[0051] for n=0:N-1

[0052]

[0053]

[0054]

[0055]

[0056] end

[0057] where N represents the number of iterations, and the output is obtained after the iteration is completed. is the calculated time-varying measurement noise variance.

[0058] Further, the autonomous navigation method further comprises:

[0059] When there is auxiliary information such as satellite navigation, steps 2 to 4 are performed to update the machine learning model online; when there is no auxiliary information such as satellite navigation, steps 2, 3, 5, 6.3 and 6.4 in step 6 are performed until the test ends and the navigation is completed.

[0060] According to another aspect, a computer device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the above method when executing the computer program.

[0061] The above technical solution is aimed at the problem that only inertial navigation cannot meet the high-precision navigation and positioning requirements of a vehicle when reference information such as satellite navigation is invalid, and proposes a neural network assisted vehicle adaptive autonomous navigation method, which trains an improved LSTM model for multi-parameter parallel prediction when satellite navigation is valid, to realize high-precision prediction of motion parameters of different orders of magnitude. When satellite navigation is invalid, the trained model is used to predict the vehicle speed and attitude, and the adaptive integrated navigation is assisted by the variational Bayesian filter.

[0062] The scheme proposed by the present application does not rely on reference information other than inertial navigation, can realize full autonomous navigation of a vehicle, that is, can enable a vehicle to realize full autonomous and high-precision navigation only by using the inertial measurement unit installed on the vehicle itself in the absence of any reference information, and significantly reduces the time accumulation error of inertial navigation, can be applied to satellite navigation invalid scenarios such as tunnels and buildings, and has important practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0063] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0064] Figure 1 is a flow chart of an embodiment of the method of the application. DETAILED DESCRIPTION

[0065] It should be noted that the embodiments and features of the application herein disclosed can be combined with each other without departing from the scope of the application. The following describes the technical solutions in the embodiments of the application clearly and completely with reference to the drawings of the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, rather than all the embodiments of the application. The description of the at least one example embodiment is actually only illustrative, but not intended to limit the application and its application or use in any way. Based on the embodiments of the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0066] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it will be further understood that the terms "comprise" and / or "include" when used in this specification, specify the presence of stated features, steps, operations, devices, components and / or combinations thereof.

[0067] Unless specifically stated otherwise, the relative arrangements of the components and steps illustrated in these embodiments and the numerical expressions and values set forth herein are not limiting. It should be understood that the various parts of the drawings are not necessarily drawn to scale, and that, for the purpose of convenience and clarity, not all components can be shown in a given figure. Techniques, methods, and devices known to those of ordinary skill can not be discussed in detail, but rather can be summarily described in order not to unnecessarily obscure aspects of the present application. In the description of the example embodiments above, any specific values should be interpreted as examples only and not as limiting. Other example embodiments of the example embodiments can have different values. It should be noted that like references herein represent like elements or parts throughout the various illustrative drawings and that the use of "for example" or "e.g." in connection with a particular item or use of an item does not exclude the use of that item or the use of a similar item in other embodiments.

[0068] Reference Figure 1The embodiment provides a neural network assisted vehicle adaptive autonomous navigation method.

[0069] Step one: design an improved LSTM model for multi-parameter parallel prediction, specifically including the following steps:

[0070] The improved LSTM model for multi-parameter parallel prediction includes four sub-models, and specific information is as follows:

[0071] 1) Sub-model 1: horizontal velocity prediction model

[0072] The horizontal velocity prediction model predicts eastward and northward velocities, and the neural network framework is LSTM, and the model input and output are as follows:

[0073] Model input: gyroscope output excluding harmful components such as the angular velocity of the earth rotation Attin calculated by pure inertial navigation insQ , navigation system acceleration f excluding the gravity acceleration component n . Wherein, the navigation system acceleration f n is calculated by combining the original accelerometer output and the attitude calculated by pure inertial navigation;

[0074] Model output: eastward velocity and northward velocity

[0075] 2) Sub-model 2: skyward velocity prediction model

[0076] The model framework and model input are the same as those of sub-model 1, and the model output is the skyward velocity;

[0077] 3) Sub-model 3: horizontal attitude prediction model

[0078] The horizontal attitude prediction model predicts the pitch angle and roll angle, and the neural network framework and model input are the same as those of sub-model 1, and the model output is the pitch angle and roll angle;

[0079] 4) Sub-model 4: heading prediction model

[0080] In order to overcome the problem that the heading angle is not derivable when switching between 0° and 360°, the heading angle is estimated in the form of a quaternion, the neural network framework and model input are the same as those of sub-model 1, and the model output is the attitude in the form of a quaternion; The loss function in the form of a quaternion is designed as follows:

[0081]

[0082] Wherein, Atti predQ =[q 0p q 1p q 2p q 3p ]T and Atti referQ = [q 0t q 1t q 2t q 3t ] T are quaternion forms of the predicted and reference attitudes respectively; AQ = [Aq0Aq1Aq2Aq3] T is the predicted attitude Atti predQ and the difference between the reference attitude Atti referQ is calculated as:

[0083]

[0084] The reference attitude here is calculated by inertial / satellite integrated navigation;

[0085] Since the variation of pitch and roll angles is small in magnitude, the accuracy is limited when combined with the heading angle for prediction, so the pitch and roll angles use the prediction results of sub-model 3, and sub-model 4 only uses the heading angle prediction result;

[0086] Step two: collect inertial measurement unit data and perform inertial navigation solution

[0087] Collect inertial measurement unit data, perform inertial navigation solution, and obtain gyro output excluding harmful components Attipure insQ is the attitude quaternion of pure inertial navigation solution, and based on Attipure insQ and accelerometer output, the navigation system acceleration f n excluding the gravity acceleration component is calculated, which is used as the input of the parallel LSTM model;

[0088] Step three: acquire satellite navigation data and perform integrated navigation solution when GNSS is effective

[0089] Multiple GNSS effective car tests are performed, and inertial / satellite integrated navigation is performed based on the collected inertial measurement unit data and GNSS data to calculate the reference attitude quaternion Atti referQ = [q 0t q 1t q 2t q 3t ] T and the reference navigation system velocity, which are used to train the parallel LSTM model;

[0090] Step four: train the parallel LSTM model designed in step one based on the inertial navigation results in step two and the integrated navigation results in step three;

[0091] Step five: Based on the parallel LSTM model trained in step four, the vehicle attitude and velocity information are predicted using inertial navigation data in the case of GNSS invalidity.

[0092] Step six: Adaptive integrated navigation is performed using variational Bayesian filtering based on the prediction results in step five, specifically

[0093] Based on the inertial measurement unit data collected in step two, the prediction results in step five are used as reference information, and variational Bayesian filtering is used for integrated navigation. The specific step execution process is as follows:

[0094] 6.1 Establishing the state space model of integrated navigation

[0095] The state space model used by the integrated navigation algorithm includes state equations and measurement equations:

[0096] The state equation is X k =Φ k,k-1 X k-1 +w k-1 . The state quantity X includes the north-east misalignment angle the north-east velocity error δv n =[δV n δV u δV e ] T , the latitude error the altitude error δh, the longitude error δλ, the gyro drift ε b =[ε x ε y ε z ] T , the accelerometer bias of the carrier system the carrier system arm error L out , the time delay T d , etc.

[0097] The measurement equation is Z k =H k X k +V k . The measurement quantity Z k is the difference between the predicted attitude and velocity information of the parallel LSTM model and the inertial navigation solution results.

[0098] 6.2 Initialization of integrated navigation algorithm parameters

[0099] The parameters of the integrated navigation algorithm based on variational Bayesian filtering are initialized, including the state quantity estimation result m k , the estimation mean square error P k , the system noise variance matrix Q k , the initial value of the shape parameter of the inverse Gamma distribution α0, and the initial value of the inverse scale parameter of the inverse Gamma distribution β0.

[0100] 6.3 Execution of time update

[0101] The time update is executed according to the following formula:

[0102]

[0103]

[0104]

[0105]

[0106] where Φ k / k-1 According to the inertial navigation error equation, ρ i ∈ (0.9, 1], i = 1, …, d is the attenuation coefficient, and are the shape parameter and inverse scale parameter of the inverse Gamma distribution respectively, and d represents the dimension of the measurement;

[0107] 6.4 Execution of measurement update

[0108] The measurement update is performed in an iterative manner: first, initialize the parameters: Then perform the iterative calculation as follows:

[0109] for n = 0: N-1

[0110]

[0111]

[0112]

[0113]

[0114] end

[0115] where N represents the number of iterations. After the iteration is completed, output where is the time-varying measurement noise variance calculated.

[0116] In addition, when there is auxiliary information such as satellite navigation, steps 2 to 4 are executed to update the machine learning model online; when there is no auxiliary information such as satellite navigation, steps 2, 3, 5 and 6.3 and 6.4 in steps 6 are executed until the experiment ends and the navigation is completed.

[0117] According to another embodiment, a computer device is provided, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the above method when executing the computer program.

[0118] It can be seen that the embodiment of the present application relates to a neural network assisted vehicle adaptive autonomous navigation method, when satellite navigation is effective, an improved LSTM model of multi-parameter parallel prediction is trained to realize high-precision prediction of motion parameters of different orders of magnitude; when satellite navigation is invalid, the improved LSTM model is used to predict the vehicle speed and attitude, to assist inertial navigation to perform integrated navigation, and to use variational Bayesian filtering to perform adaptive integrated navigation in view of the problem that the noise of the prediction result of the LSTM model is complex and time-varying. Thus, high-precision, fully autonomous navigation and positioning of the vehicle are realized without relying on reference information such as GNSS, which has important practical application value.

[0119] Features described and / or illustrated with respect to one embodiment can be used in the same or similar manner in one or more other embodiments and / or in combination with or in place of features in other embodiments.

[0120] It should be emphasized that the term "comprises / comprising" when used in this text refers to the presence of a feature, whole, step or component, but does not exclude the presence or addition of one or more other features, wholes, steps, components or combinations thereof.

[0121] The above method of the present application can be realized by hardware, or by hardware in combination with software. The present application relates to a computer readable program which, when executed by a logic component, enables the logic component to realize the above-described device or component, or to realize the above-described various methods or steps. The present application also relates to a storage medium for storing the above program, such as a hard disk, a magnetic disk, an optical disk, a DVD, a flash memory, etc.

[0122] Many of the features and advantages of the embodiments are apparent from the detailed description, and thus, it is intended that the appended claims cover all such features and advantages of the embodiments within their true spirit and scope. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the embodiments of the present application to the exact construction and operation described and illustrated, and accordingly, all suitable modifications and equivalents can be resorted to, falling within the scope of the embodiments.

[0123] The parts of the present application not described in detail are known to those skilled in the art.

Claims

1. A neural network-assisted adaptive autonomous navigation method for vehicles, characterized in that, The autonomous navigation method includes: Step 1: Design an improved LSTM model for multi-parameter parallel prediction, including: The improved LSTM model is designed to include four sub-models: 1) Sub-model 1: Horizontal velocity prediction model The horizontal velocity prediction model predicts eastward and northward velocities. The neural network framework used is LSTM, and the model's input and output are designed as follows: Model input: Gyroscope output after deducting the harmful component of Earth's rotation angular velocity. Attitude quaternion Atti calculated by pure inertial navigation insQ The acceleration of the navigation system, f, after deducting the gravitational acceleration component. n Among them, the acceleration f of the navigation system n Attitude calculation is performed by combining the raw accelerometer output with pure inertial navigation solutions. Model output: Eastward velocity and Northward velocity; 2) Sub-model 2: Celestial velocity prediction model The model framework and model input of the celestial velocity prediction model are the same as those of sub-model 1, and the model output is the celestial velocity. 3) Sub-model 3: Horizontal attitude prediction model The horizontal attitude prediction model predicts pitch and roll angles. The neural network framework and model input are the same as those of sub-model 1, and the model output is pitch and roll angles. 4) Sub-model 4: Heading prediction model The heading prediction model uses quaternions to predict the heading angle. The neural network framework and model input are the same as those of sub-model 1. The model output is the attitude in quaternion form. Step 2: Acquire data from the inertial measurement unit and perform inertial navigation calculations; Step 3: When GNSS is available, acquire satellite navigation data and perform integrated navigation calculations, including: Multiple GNSS-enabled vehicle tests were conducted. Based on the inertial measurement unit data and GNSS data acquired in step two, inertial / satellite integrated navigation was performed, and the reference attitude quaternion Atti was calculated. referQ The velocity of the reference navigation system is used to train the improved LSTM model; Step 4: Train the improved LSTM model designed in Step 1 based on the inertial navigation solution results from Step 2 and the combined navigation results from Step 3; Step 5: Based on the improved LSTM model trained in Step 4, predict vehicle attitude and velocity information using inertial navigation data when GNSS is ineffective; Step Six: Based on the prediction results from Step Five, use variational Bayesian filtering to perform adaptive integrated navigation; When designing sub-model 4, the accompanying quaternion form loss function is designed as follows: Among them, Atti predQ =[q 0p q 1p q 2p q 3p ] T and Atti referQ =[q 0t q 1t q 2t q 3t ] T These are the quaternion forms of the predicted attitude and the reference attitude, respectively; ΔQ = [Δq0Δq1Δq2Δq3] T Atti, the attitude prediction result predQ Atti with reference attitude referQ The difference between them is calculated using the following formula: The reference attitude here is calculated using inertial / satellite integrated navigation; Step six specifically includes: Based on the inertial measurement unit data acquired in step two, and using the prediction results from step five as reference information, variational Bayesian filtering is used for integrated navigation. The execution process is as follows: 6.1 Establish a state-space model for integrated navigation, including state equations and measurement equations; 6.2 Initialize the integrated navigation algorithm parameters; 6.3 Execution time update; 6.4 Perform measurement updates.

2. The neural network-assisted adaptive autonomous navigation method for vehicles according to claim 1, characterized in that, Step three specifically includes: Data is collected from the inertial measurement unit, inertial navigation calculations are performed, and the gyroscope output is obtained after deducting the harmful component of the Earth's rotation angular velocity. Attitude quaternion Atti calculated by pure inertial navigation insQ ; and based on Atti insQ The accelerometer output is used to calculate the navigation system acceleration f after deducting the gravitational acceleration component. n , used as input for parallel LSTM models.

3. The neural network-assisted adaptive autonomous navigation method for vehicles according to claim 1, characterized in that, In step 6.1: The state equation is X k =Φ k,k-1 X k-1 +w k-1 The state quantity X includes the north-to-east misalignment angle. North-East Velocity Error δv n =[δV n δV u δV e ] T Latitude error Altitude error δh, longitude error δλ, and gyroscope drift ε of the carrier system b =[ε x ε y ε z ] T , Load system accelerometer zero bias▽ b =[▽ x ▽ y ▽ z ] T , load system arm error L out Time delay T d ; The measurement equation is Z k =H k X k +V k Measurement Z k To improve the difference between the attitude and velocity information predicted by the LSTM model and the inertial navigation solution.

4. The neural network-assisted adaptive autonomous navigation method for vehicles according to claim 3, characterized in that, In step 6.2: Initialize the parameters of the integrated navigation algorithm based on variational Bayesian filtering, including the state quantity estimation result m. k Estimate the mean square error P k System noise variance matrix Q k Initial values ​​for the shape parameter α0 and the inverse scaling parameter β0 of the inverse Gamma distribution.

5. The neural network-assisted adaptive autonomous navigation method for vehicles according to claim 4, characterized in that, In step 6.3, the time update is performed according to the following formula: Where, Φ k / k-1 ρ is calculated based on the inertial navigation error equation. i ∈(0.9,1], i=1,…,d is the attenuation coefficient, and denoted as the shape parameter and inverse scaling parameter of the inverse Gamma distribution, respectively, and d represents the dimension of the measurement.

6. The neural network-assisted adaptive autonomous navigation method for vehicles according to claim 5, characterized in that, In step 6.4, the measurement update is performed iteratively: First, initialize the parameters: i = 1, ..., d, and then the iterative calculation is performed as follows: In the formula, N represents the number of iterations, and the output is after the iteration is complete. This is the time-varying measurement noise variance that is solved.

7. A neural network-assisted adaptive autonomous navigation method for vehicles according to any one of claims 1-6, characterized in that, The autonomous navigation method also includes: When satellite navigation assistance information is available, execute steps two through four to update the machine learning model online; when satellite navigation assistance information is unavailable, execute steps 6.3 and 6.4 in steps two, three, five, and six until navigation is completed at the end of the experiment.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method of any one of claims 1-7.

Images