Self-constraint vehicle inertial navigation method based on Transformer-Bayesian optimization motion classifier

Abstract

The invention provides a self-constraint vehicle inertial navigation method based on a Transform-Bayesian optimization motion classifier. The method comprises the following steps: constructing a deep learning network taking a Transform encoder as a core structure; taking inertial measurement data of the vehicle as input to obtain confidence of the vehicle belonging to each motion mode; determining a motion mode corresponding to the inertial measurement data as motion constraint information by adopting a motion mode classification rule based on a confidence threshold; wherein a threshold value in the motion mode classification rule is obtained through Bayesian optimization search to enable the cost of various recognition errors of sample data to be minimum; and fusing inertial measurement data with the motion constraint information by using invariant extended Kalman filtering to obtain vehicle navigation positioning information. The method can realize inertial data information enhancement, and has the advantages of low cost, high reliability, high precision and the like.

Description

Technical Field

[0001] The present invention relates to the technical fields of inertial navigation and pattern recognition, and particularly to a self-constrained vehicle inertial navigation method based on a Transformer-Bayesian optimized motion classifier. Background Art

[0002] Inertial navigation has significant advantages in environments such as satellite denial, but there is an error accumulation problem in inertial navigation solution, resulting in a decrease in its long-term navigation accuracy. Common error compensation methods use the Kalman filter algorithm to fuse external reference information and inertial navigation information, estimate the inertial navigation system error and feedback compensation. On this basis, integrated navigation algorithms that combine inertial navigation information with visual information, odometer information, and radar information have been developed. These methods have achieved certain results in specific application scenarios, but they cannot fully utilize the advantages of inertial navigation, restricting its autonomy, concealment, and reliability. Summary of the Invention

[0003] In view of this, the present invention provides a self-constrained vehicle inertial navigation method based on a Transformer-Bayesian optimized motion classifier. This method does not rely on external reference information and only uses inertial data and the kinematic constraints of the vehicle itself. Taking static and straight uniform motion as state constraints, it suppresses the errors of inertial navigation and has autonomy, concealment, and reliability.

[0004] To solve the above technical problems, the present invention is implemented as follows.

[0005] A self-constrained vehicle inertial navigation method based on a Transformer-Bayesian optimized motion classifier includes:

[0006] Step S1: Construct a deep learning network with a Transformer encoder as the core structure; take the inertial measurement data of the vehicle as the input to obtain the confidence of the vehicle belonging to each motion mode; the 3 motion modes include a static state, a straight-line motion state, and other motion states;

[0007] Step S2: According to the confidence obtained in step S1, adopt a motion mode classification rule based on a confidence threshold to determine the motion mode corresponding to the inertial measurement data; wherein, the threshold in the motion mode classification rule is obtained by Bayesian optimization to minimize the cost of various recognition errors of the sample data;

[0008] Step S3: Use the unscented extended Kalman filter for vehicle navigation and positioning; during the filtering process, when the motion mode is in the stationary state and the straight-line motion state, extract the motion features corresponding to the motion mode; use the extracted motion features as motion constraint information to simplify the vehicle state transition equation to constrain the accumulation of relevant motion features during the inertial navigation solution process; at the same time, the extracted motion features provide observations for the unscented extended Kalman filter to constrain the error accumulation.

[0009] Preferably, in step S1, the deep learning network with the Transformer encoder as the core structure includes an input layer, an initialization embedding layer, a Transformer encoder layer, a Dropout layer, a feed-forward neural network layer, and an output layer connected in sequence.

[0010] Preferably, in step S2, the method of using the motion mode classification rule based on the confidence threshold to determine the motion mode corresponding to the inertial measurement data includes:

[0011] Set the thresholds T0 and T1 for the stationary state and the straight-line motion state. The confidence levels of the stationary state, the straight-line motion state, and other motion states obtained in step S1 are score0, score1, and score2 respectively;

[0012] Case 1: If the confidence level score0 of the stationary state is greater than or equal to the stationary state threshold T0, then determine that the motion mode is the stationary state;

[0013] Case 2: If the confidence level score1 of the straight-line motion state is greater than or equal to the straight-line motion state threshold T1, then determine that the motion mode is the straight-line motion state;

[0014] If it is other cases, then determine that the motion mode is other motion states.

[0015] Preferably, the threshold optimization process in the motion mode classification rule is as follows:

[0016] Define that the stationary state provides strong constraints and the straight-line motion state provides weak constraints; the situation where a motion mode providing strong constraints is misidentified as a motion mode providing weak constraints is called a false alarm, and vice versa is a miss; a false alarm corresponds to a high cost, a miss corresponds to a low cost, and a correct identification corresponds to a zero cost;

[0017] Based on the cost corresponding to the prediction result of each sample in the test set, calculate the cost-sensitive error rate; use the Bayesian optimization algorithm to perform Gaussian process regression modeling on the cost-sensitive error rate, and use the expected acquisition function to evaluate the potential value of the sampled points to guide the search process and explore the optimal threshold combination T0 and T1 that minimizes the cost-sensitive error rate.

[0018] Preferably, define the cost of predicting the i-th type of motion state sample as the j-th type of motion state sample as cost ij , where i, j = 0, 1, 2; 0, 1, and 2 represent the stationary state, linear motion state, and other motion states respectively;

[0019] Then the cost of correct recognition cost ii = 0, and the costs corresponding to false alarms cost 10 , cost 20 and cost 21 are high costs, and the costs corresponding to missed detections cost 01 , cost 02 , cost 12 are low costs;

[0020] The cost-sensitive error rate is the mean of the costs corresponding to the prediction results of all samples in the test set.

[0021] Preferably, in step S2, after determining the motion mode corresponding to the inertial measurement data, further perform a holding mechanism judgment: at time t, if and only if the motion modes corresponding to the existing continuous M inertial measurement data are the same, then update the final motion state s t at time t to and use it as motion constraint information; otherwise, keep the final motion state s t-1 at the previous moment as motion constraint information.

[0022] Preferably, M = 5.

[0023] Preferably, the step S3 is:

[0024] Define the system state as an augmented Lie group state, and use an invariant extended Kalman filter for vehicle navigation and positioning; define the system observation value as that is, the vehicle speed in the body coordinate system, the constant bias of the gyroscope and the constant bias of the accelerometer R k 、v k are the vehicle motion attitude and vehicle speed in the navigation coordinate system;

[0025] When the motion mode is the stationary state, the vehicle speed v k in the navigation coordinate system = 0, the angular velocity ω b in the body coordinate system = 0, the acceleration α b in the body coordinate system = 0, which is used as prior information to simplify the vehicle state transfer equation to constrain the accumulation of speed and position errors during the inertial navigation solution process; at the same time, the speed v b in the body coordinate system = 0, and the angular velocity and acceleration output by the inertial device Provide observations, and the observed values are used to correct the constant bias of the inertial device and constrain the cumulative error;

[0026] When the motion mode is in a straight-line motion state, the three-axis attitude of the vehicle is a fixed value, which simplifies the vehicle state transfer equation as prior information to constrain the accumulation of attitude errors during inertial navigation solution; the velocity v in the body coordinate system b The lateral velocity component v in lat and the vertical velocity component v up are 0, providing partial observations for the velocity in the body coordinate system. By setting some rows of the observation matrix to zero, the velocity error is constrained.

[0027] Beneficial effects:

[0028] (1) The present invention adopts a motion mode recognition network based on Transformer, utilizes its modeling ability for long-sequence dependence relationships, improves the confidence of motion mode recognition, and has strong robustness.

[0029] (2) The present invention designs a motion mode recognition decision maker based on Bayesian optimization, fully considers the costs of different wrong decision results, optimizes the reliability of motion mode classification, and combines motion prior information to provide more accurate constraint support for subsequent inertial navigation information fusion.

[0030] (3) The present invention adopts motion constraint / inertial navigation fusion based on invariant extended Kalman filter, overcomes the limitation of classical Kalman filter that depends on system states for linearization in nonlinear systems, and improves the filtering accuracy and system robustness. Description of the drawings

[0031] Figure 1 is a framework diagram of the self-constrained vehicle inertial navigation method based on the Transformer-Bayesian optimization motion classifier provided by the present invention. Detailed implementation manners

[0032] The following combines the drawings and gives examples to describe the present invention in detail.

[0033] The present invention first generates the confidence of each motion mode to which the vehicle belongs through a motion mode recognition network based on Transformer, constructs a decision-making strategy based on the confidence of motion mode recognition, further uses a motion mode decision maker based on Bayesian optimization to extract the motion constraint information of the vehicle in the stationary and straight-line motion states, and finally uses invariant extended Kalman filter to fuse the motion constraint information and inertial measurement data to achieve accurate vehicle inertial autonomous navigation.

[0034] Figure 1The flowchart of the self-constrained vehicle inertial navigation method based on the Transformer-Bayesian optimization motion classifier of the present invention is shown. As Figure 1 shown, the method includes the following steps:

[0035] Step S1: Construct a deep learning network with a Transformer encoder as the core structure; take the inertial measurement data of the vehicle as the input to obtain the confidence of the vehicle belonging to each motion mode.

[0036] In this step, a deep learning network with a Transformer encoder as the core structure is constructed, including an input layer, an initial embedding layer, a Transformer encoder layer, a Dropout layer, a feed-forward neural network layer, and an output layer, and a motion mode recognition network with high recognition accuracy is trained using the vehicle inertial measurement data and the corresponding true motion mode labels. Taking the vehicle angular velocity and acceleration as the input, obtain the probability values of the vehicle belonging to each motion mode, that is, the confidence.

[0037] The 3 motion modes include a stationary state, a straight-line motion state, and other motion states. The probability values of the deep learning network outputting the vehicle belonging to the 3 motion modes include score0, score1, and score2.

[0038] Step S2: According to the confidence obtained in step S1, adopt a motion mode decision strategy based on a threshold to determine the motion mode corresponding to the inertial measurement data as the motion constraint information for step S3.

[0039] In this step, a motion mode classification rule based on a confidence threshold is constructed:

[0040]

[0041] where, is the predicted label of the i-th sample output by the deep learning network, and y i is the true label of which motion mode the i-th sample data belongs to. 0, 1, and 2 are class numbers, representing the stationary state, the straight-line motion state, and other motion states respectively. T0 and T1 are the thresholds for the stationary state and the straight-line motion state respectively. score0(i) and score1(i) are the confidence of the i-th sample data belonging to the stationary state and the confidence of belonging to the straight-line motion state obtained by inputting the deep learning network respectively.

[0042] Secondly, assign unequal weights to different recognition errors, measure the adaptability of motion mode recognition to subsequent navigation tasks using the cost-sensitive error rate, and further select Bayesian optimization to efficiently determine the confidence threshold that minimizes the cost-sensitive error rate to complete the preliminary decision of the motion mode.

[0043] The costs of different recognition errors are as follows:

[0044] Table 1 Cost Matrix

[0045]

[0046] Among them, cost ij represents the cost of predicting the i-th type of sample as the j-th type of sample. The stationary state provides strong constraints, the linear motion state provides weak constraints, and other motion states provide zero constraints. The situation where a motion pattern with strong constraints is recognized as a motion pattern with weak constraints is called a false alarm, and vice versa is a missed detection. The cost of correct recognition cost ii = 0, the costs corresponding to false alarms cost 10 , cost 20 , and cost 21 are high costs, and the costs corresponding to missed detections cost 01 , cost 02 , cost 12 are low costs;

[0047] The cost-sensitive error rate is:

[0048]

[0049] Among them, T = (T0, T1), N is the number of samples in the test set, y i is the true label of the i-th sample, is the predicted label of the i-th sample, is the cost of recognizing a sample with the true label y i as the predicted label , and the set of all costs is cost.

[0050] Adopt the Bayesian optimization algorithm. By performing Gaussian process regression modeling on the objective function cost-sensitive error rate and using the expected acquisition function to evaluate the potential value of the sampled points, guide the search process, and efficiently explore the optimal threshold combination that minimizes the cost-sensitive error rate, and find the T that minimizes the cost mean, which is the optimal threshold combination T*:

[0051]

[0052] The expected acquisition function is:

[0053]

[0054] Among them, f(x) is the objective function, f(x * ) is the current optimal objective function value, and ξ is the adjustment parameter.

[0055] Finally, according to the prior knowledge that the motion mode does not mutate, a state retention mechanism is formulated, and based on this mechanism, it is finally determined whether to update the final motion state:

[0056]

[0057] where s t is the final predicted value of the vehicle motion state; M is the continuous motion threshold, preferably M = 5; is the Kronecker function.

[0058] The meaning of this formula is that when and only when the prediction results are consistent for M consecutive times, the final motion state at time t is updated to (that is, ), otherwise the final motion state at the previous moment is maintained (that is, s t = s t-1 ), and the final motion mode recognition decision is completed.

[0059] Step S3: Motion constraint / INS fusion based on the invariant extended Kalman filter.

[0060] The stationary state is defined as:

[0061] v = 0 (1)

[0062] where v is the vehicle speed in the navigation coordinate system.

[0063] In the stationary state, the vehicle provides motion constraints:

[0064] v b = [v for , v lat , v up = 0 (2)

[0065] α b = 0 (3)

[0066] ω b = 0 (4)

[0067] where in this paper, the superscript b represents the data in the body coordinate system. v b , α b , ω b are the vehicle speed, acceleration, and angular velocity in the body coordinate system b, and v for , v lat , v up are the components of the vehicle speed v b , which are the forward speed component, lateral speed component, and vertical speed component, respectively.

[0068] The straight-line motion state is defined as:

[0069] yaw = const (5)

[0070] In the linear motion state, the vehicle provides motion constraints:

[0071] [v lat ,v up = 0 (6)

[0072] [roll, pitch, yaw] = const (7)

[0073] where yaw, roll, and pitch are the heading angle, roll angle, and pitch angle of the vehicle respectively, and const represents a constant value.

[0074] The inertial measurement data is fused with the motion constraint information output by the decision-making device by using the invariant extended Kalman filter. By processing the non-linear error in the Lie group space, the linearization limitation of the classical Kalman filter is overcome, and the vehicle navigation information is obtained.

[0075] In the stationary state, the motion characteristics (1)(3)(4) of the vehicle are used as prior information, which not only simplifies the vehicle state transfer equation, but also effectively constrains the accumulation of velocity and position errors in the inertial navigation solution process. In addition, the motion characteristics (2)(3)(4) provide observations for the velocity in the body coordinate system and the output of the inertial devices, which is beneficial to correcting the constant bias of the inertial devices and improving the measurement accuracy of the inertial devices, and constrains the accumulated error from the source.

[0076] In the linear motion state, the motion characteristic (7) of the vehicle also serves as prior information to constrain the accumulation of attitude errors. The motion characteristic (6) provides partial observations for the velocity in the body coordinate system and constrains the velocity error.

[0077] The specific steps for vehicle navigation and positioning using the invariant extended Kalman filter are as follows:

[0078] (1) Define the system state as the augmented Lie group state:

[0079]

[0080] where R k is the vehicle motion attitude in the navigation coordinate system, v k is the vehicle velocity in the navigation coordinate system, p k is the vehicle position in the navigation coordinate system, is the constant bias of the angular velocity output by the gyroscope, is the constant bias of the acceleration output by the accelerometer. χ k is the Lie group part, and b k is the non-Lie group part. In this paper, the subscript k represents the k-th iteration.

[0081] The corresponding system state estimation error is:

[0082]

[0083] where ξ k is the estimation error of the vehicle attitude, velocity, and position in the form of Lie algebra, is the estimation error of the constant bias of the inertial device, represents the constant estimated value of the inertial device.

[0084] (2) State propagation

[0085] When the vehicle motion state is identified as other motion states, the state prior estimate is:

[0086]

[0087] When the vehicle motion state is identified as the stationary state, the angular velocity a k and the acceleration ω k are set to 0; when the vehicle motion state is identified as the linear motion state, the angular velocity ω k is set to 0, and the motion constraint is used as a prior knowledge to constrain the error of the inertial measurement data.

[0088] Linearize the system state propagation equation, and the state transition matrix is:

[0089]

[0090] where I is the identity matrix, 0 is the zero matrix, and the subscript (3,3) represents the matrix size; (·) × represents constructing a three-dimensional vector into an anti-symmetric matrix.

[0091] Correspondingly, when the vehicle motion state is identified as the stationary state, is set to 0; for the linear motion state, the first 3 rows are set to 0.

[0092] Update the prior estimate covariance:

[0093]

[0094] where G k is the process noise propagation matrix, and Q k is the process noise covariance matrix.

[0095] (3) Define the system observation value as That is, the vehicle velocity in the body frame , the constant bias of the gyroscope and the constant bias of the accelerometer Then the observation matrix is:

[0096]

[0097] Among them, g is gravity.

[0098] When the vehicle motion state is recognized as a stationary state, the motion characteristics (2), (3), and (4) provide observations for the velocity in the body coordinate system and the outputs of inertial devices, and the observed values are

[0099] When the vehicle motion state is recognized as a straight-line motion state and other straight-line motion states, the motion characteristic (6) provides partial observations for the velocity in the body coordinate system, that is, only [v lat , v up = 0, the vehicle non-holonomic constraint. Accordingly, set the second to ninth rows of H k to 0.

[0100] Calculate the Kalman gain: N k+1 is the observation noise covariance matrix, and "+" represents the data after posterior estimation.

[0101] The posterior estimation of the system state is:

[0102]

[0103] Among them, represents the estimation of the Lie group part χ k and represents the observation estimation.

[0104] Update the posterior estimation covariance:

[0105]

[0106] Thus, the self-constrained vehicle inertial navigation based on the Transformer-Bayesian optimization motion classifier is completed.

[0107] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in this description can be different and are not limited. Therefore, those skilled in the art of the present invention can modify or equivalently replace the technical solutions recorded in the foregoing embodiments; and these modifications and replacements do not depart from the spirit and technical solutions of the present invention, and shall all fall within the protection scope of the present invention.

Claims

1. A self-constrained vehicle inertial navigation method based on Transformer-Bayesian optimization motion classifier, characterized in that: include: Step S1: Construct a deep learning network with Transformer encoder as the core structure; Using the inertial measurement data of the vehicle as input, obtaining the confidence that the vehicle belongs to each motion mode; the three motion modes include static state, linear motion state and other motion states; Step S2: according to the confidence obtained in step S1, a motion mode classification rule based on a confidence threshold is used to determine the motion mode corresponding to the inertial measurement data; wherein the threshold in the motion mode classification rule is obtained by Bayesian optimization to find the minimum cost of various recognition errors of the sample data; Step S3: using the invariant extended Kalman filter to perform vehicle navigation and positioning; During the filtering process, when the motion mode is in a stationary state and a linear motion state, the motion features corresponding to the motion mode are extracted; the extracted motion features are used as motion constraint information to simplify the vehicle state transfer equation to constrain the accumulation of related motion features during the inertial navigation solution; at the same time, the extracted motion features provide observations for the invariant extended Kalman filter to constrain error accumulation.

2. The method according to claim 1, characterized in that In step S1, the deep learning network with the Transformer encoder as the core structure includes an input layer, an initialization embedding layer, a Transformer encoder layer, a Dropout layer, a feedforward neural network layer and an output layer connected in sequence.

3. The method according to claim 1, characterized in that In step S2, the motion mode classification rule based on the confidence threshold is used to determine the motion mode corresponding to the inertial measurement data, including: Thresholds T0 and T1 of the static state and the linear motion state are set, and the confidences of the static state, the linear motion state and the other motion states obtained in step S1 are score0, score1 and score2 respectively; Case 1: If the static state confidence score0 is greater than or equal to the static state threshold T0, the motion mode is determined to be static; Case 2: If the linear motion state confidence score1 is greater than or equal to the linear motion state threshold T1, the motion mode is determined to be a linear motion state; If it is other cases, the motion mode is determined to be other motion states.

4. The method according to claim 3, characterized in that The threshold optimization process in the motion mode classification rule is as follows: It is defined that the static state provides a strong constraint, and the linear motion state provides a weak constraint; the situation where a motion pattern that provides a strong constraint is identified as a motion pattern that provides a weak constraint is called a false alarm, and vice versa is a missed detection; a false alarm corresponds to a high cost, a missed detection corresponds to a low cost, and a correct identification corresponds to zero cost; The cost-sensitive error rate is calculated based on the corresponding cost of the prediction result of each sample in the test set. The Bayesian optimization algorithm is used to guide the search process by performing Gaussian process regression modeling on the cost-sensitive error rate and using the expected acquisition function to evaluate the potential value of the sampled points, thereby exploring the optimal threshold combination T0 and T1 that minimizes the cost-sensitive error rate.

5. The method according to claim 3, characterized in that Define the cost of predicting the i-th motion state sample as the j-th motion state sample as cost ij , i, j = 0, 1, 2; 0, 1, 2 represent the static state, linear motion state and other motion states respectively; The cost of correct identification is ii =0, the cost corresponding to the false alarm 10 、cost 20 and cost 21 is a high cost, the cost corresponding to missed detection 01 、cost 02 、cost 12 For low cost; The cost-sensitive error rate is the mean of the costs corresponding to the prediction results of all samples in the test set.

6. The method according to claim 1, characterized in that In step S2, after determining the motion mode corresponding to the inertial measurement data, further determination of the holding mechanism is performed: at time t, if and only if there are M consecutive inertial measurement data corresponding to the same motion mode, the final motion state s at time t is set to t Updated to The value of is used as motion constraint information; otherwise, the final motion state s of the previous moment is maintained. t-1 , as motion constraint information.

7. The method according to claim 6, characterized in that M＝5。 8. The method according to claim 1, characterized in that The step S3 is: The system state is defined as the augmented Lie group state, and the invariant extended Kalman filter is used for vehicle navigation and positioning; the system observation value is defined as That is, the vehicle speed in the carrier coordinate system Gyroscope constant bias and accelerometer constant bias R k 、v k is the vehicle motion posture and vehicle speed in the navigation coordinate system; When the motion mode is stationary, the vehicle speed v in the navigation coordinate system k =0, angular velocity ω in the carrier coordinate system b = 0, acceleration α in the carrier coordinate system b = 0, which simplifies the vehicle state transfer equation as a priori information to constrain the accumulation of speed and position errors in the inertial navigation solution process; at the same time, the speed v in the carrier coordinate system b =0, the angular velocity output by the inertial device and acceleration Provide observations, the observed values ​​are To correct the constant deviation of inertial devices and constrain the cumulative error; When the motion mode is in linear motion, the three-axis attitude of the vehicle is a constant, which is used as a priori information to simplify the vehicle state transfer equation to constrain the accumulation of attitude errors during the inertial navigation solution process; the velocity v in the carrier coordinate system b The lateral velocity component v in lat , celestial velocity component v up It is 0, which provides partial observation for the velocity in the carrier coordinate system, and sets some rows of the observation matrix to zero to realize the constraint on the velocity error.

Images