Inertial initial alignment method and system for deep learning

Abstract

The invention provides a deep learning inertial initial alignment method and system, and the method comprises the steps: inputting the angular rate and specific force information of an inertial device into a trained deep learning network, obtaining a speed increment, and obtaining the speed under a satellite denial condition; the deep learning network is a TCN network with a cross spiral structure; and on the basis of the speed output by the deep learning network, through the constructed initial attitude matrix model, an initial attitude is solved by adopting an MKF algorithm, and inertial initial alignment is completed. According to the method, the feature extraction and information association capability in the TCN residual module is fully mined, the coupling relationship between inertial device data and speed increment is deeply explored, and the speed estimation precision is effectively improved, so that an effective parameter information reference is provided for denial inertial initial alignment, and the inertial navigation initial alignment precision is improved. The method is suitable for initial alignment of the inertial navigation system under the satellite denial condition.

Description

Technical Field

[0001] The present invention belongs to the field of inertial initial alignment of an inertial navigation system during satellite denial, and particularly relates to a deep learning-based inertial initial alignment method and system. Background Art

[0002] The inertial navigation system (INS) has the advantages of strong autonomy, good concealment, small volume, etc., and is widely used in the fields of weapons, industry, vehicle-mounted, aerospace, aviation, navigation, etc. The core principle of the INS is to sense the acceleration and angular velocity information of the carrier through an inertial measurement unit (IMU), and to continuously estimate the position, velocity, and attitude through integral operations. Its performance depends to a large extent on the accuracy of the initial alignment. The alignment time and alignment accuracy of the initial alignment directly affect the starting time of the INS operation and the accuracy of subsequent navigation solutions. This technology has become a difficult problem that needs to be solved urgently in inertial navigation technology.

[0003] At present, there are mainly three types of methods for initial alignment according to the stage: rough alignment, fine alignment, and rough alignment during movement. The rough alignment directly uses external reference information to solve the initial attitude of the inertial navigation system, and has the advantage of rapidity. The fine alignment constructs an initial attitude error equation of the inertial navigation system and performs filtering fusion with external reference information to accurately obtain the initial attitude. The rough alignment during movement constructs a Wahba problem solving framework with the help of external reference information during the movement of the carrier to quickly determine the initial attitude. The essential idea of these three types of methods is to rely on external references to calibrate the initial attitude of the inertial navigation system. In particular, the initial alignment method based on satellite navigation has become the mainstream initial alignment framework. However, in the case of satellite denial conditions such as indoor, underground, canyon, or complex urban environments, the lack of effective external reference information not only limits the application of the inertial navigation system in complex environments, but also poses a severe challenge to its combat capabilities in the military field. Therefore, how to achieve high-precision initial alignment under satellite denial conditions has become an urgent problem to be solved.

[0004] In recent years, the rapid development of artificial intelligence (AI) technology has brought new opportunities to the field of inertial navigation. As an important branch of artificial intelligence, deep learning has made breakthrough progress in the fields of image recognition, natural language processing, speech recognition, etc. Its powerful feature extraction and pattern recognition capabilities provide new ideas for solving complex nonlinear problems. In the field of inertial navigation, researchers have begun to explore the application of deep learning technology in tasks such as initial alignment, attitude estimation, and error compensation, and have achieved a series of encouraging results. Therefore, this patent designs a deep learning-based inertial initial alignment method to effectively solve the problem of inertial initial alignment of an inertial navigation system under satellite denial conditions. Summary of the Invention

[0005] To overcome the deficiencies of the prior art, the present invention proposes a deep learning-based inertial initial alignment method and system, which effectively improves the speed estimation accuracy, thereby providing effective parameter information reference for the denied inertial initial alignment and improving the inertial navigation system initial alignment accuracy.

[0006] The technical solution for achieving the object of the present invention is as follows:

[0007] An inertial initial alignment method based on deep learning, comprising the steps of:

[0008] Input the angular rate and specific force information of the inertial device into the trained deep learning network to obtain the speed increment, and then obtain the speed under satellite denial conditions; the deep learning network is a TCN network with a cross-helix structure;

[0009] Based on the obtained speed under satellite denial conditions, use the MKF algorithm to solve the parameters of the constructed initial attitude matrix model to obtain the initial attitude and complete the initial alignment of the inertial navigation system.

[0010] Further, the deep learning network is a TCN network with a cross-helix structure, and the residual module is connected to the TCN module in reverse. The output of the residual module is directly connected to the causal convolution layer, the output of the Dropout layer is directly connected to the convolution layer, and a skip connection is made to the next residual module.

[0011] Further, the causal convolution layer uses the Padding(·) function to pad zeros at the end of the time series and then performs convolution. Then the causal convolution at the subsequence x j is:

[0012]

[0013] where the filter function is F = (f1, f2,..., f N ), the time series X = (x1, x2,..., x j ), j is the number of sequences, x j is the j-th sequence, f n is the n-th filter function, and N is the number of filters.

[0014] Further, the deep learning network increases the receptive field of the neural network by adding a dilation rate to the convolution kernel. For the sequence and the filter function, the dilated convolution G on the sequence element s is:

[0015]

[0016] where * represents the convolution operation, c is the convolution kernel size, d is the dilation coefficient, f(i) is the i-th convolution kernel, and s - d·i is the past direction.

[0017] Further, the connection method of the residual module is as follows:

[0018] g = Activation(X + C(X))

[0019] In the formula, C(X) is convolution, and Activation(·) is an activation function.

[0020] Further, the constructed initial attitude matrix model is as follows:

[0021]

[0022] Among them, the parameters α and β are:

[0023]

[0024] Among them, b(t) is the vehicle coordinate system at time t, n(0) is the navigation coordinate system at time 0, b(0) is the vehicle coordinate system at time 0, ω ie is the earth's angular rotation rate, f b is the three-axis specific force measured by the accelerometer, g n is the local gravitational acceleration, v n is the velocity.

[0025] Further, the recurrence solution formulas for the parameters α and β are:

[0026]

[0027] An inertial initial alignment system based on deep learning includes:

[0028] A velocity estimation unit that inputs the angular rate and specific force information of inertial devices into a trained deep learning network to obtain a velocity increment, and further obtains the velocity under satellite denial conditions; the deep learning network is a TCN network with a cross-helix structure;

[0029] An initial attitude solution unit that, based on the obtained velocity under satellite denial conditions, uses the MKF algorithm to solve the parameters of the constructed initial attitude matrix model to obtain the initial attitude and complete the initial alignment of the inertial navigation system.

[0030] An inertial initial alignment device includes: a memory, a processor, and a computer program stored on the memory. When the processor executes the computer program, the steps of the inertial initial alignment method are implemented.

[0031] A computer storage medium stores an executable program, and when the executable program is executed by a processor, the steps of the inertial initial alignment method are implemented.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: First, the present invention constructs an inertial initial alignment framework during rejection, designs a data-driven speed estimator, and based on the inertial navigation solution principle, uses inertial device data as the network input and speed increment as the network output, and proposes a new type of cross-time convolutional network (XTCN). It fully exploits the feature extraction and information association capabilities in the TCN residual module, deeply explores the coupling relationship between inertial device data and speed increment, effectively improves the speed estimation accuracy, thereby providing effective parameter information reference for rejection inertial initial alignment and improving the inertial navigation initial alignment accuracy. This method is applicable to the initial alignment of inertial navigation systems under satellite rejection conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 FIG. is a flowchart of inertial initial alignment under satellite rejection conditions.

[0034] Figure 2 FIG. is a schematic diagram of an inertial initial alignment framework under satellite rejection conditions.

[0035] Figure 3 FIG. is a schematic diagram of the XTCN structure.

[0036] Figure 4 FIG. is a schematic diagram of an extended XTCN structure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] Next, the embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0038] Combined with Figure 1 and Figure 2 , a deep learning-based inertial initial alignment method, the derivation and implementation process is as follows:

[0039] The key problem of inertial navigation initial alignment is to determine the attitude matrix from the vehicle coordinate system to the navigation coordinate system The alignment process starts at time 0. According to the chain rule, the attitude matrix can be decomposed into:

[0040]

[0041] where the navigation coordinate system at time t is n(t), and the vehicle coordinate system is b(t). The navigation coordinate system at time 0 is n(0), and the vehicle coordinate system is b(0).

[0042] According to the velocity differential equation of SINS (strapdown inertial navigation system):

[0043]

[0044] where R e is the mean radius of the earth, ω ieis the angular rate of the Earth's rotation, ω en is the position rate; [Lλh] T is the position, representing latitude, longitude and altitude respectively; is the velocity, representing east, north and vertical velocities respectively. f b is the three-axis specific force measured by the accelerometer, g n is the local gravitational acceleration.

[0045] Simplifying Equation (2) gives:

[0046]

[0047] In the formula,

[0048] The solution to Equation (3) is the Wahba problem, which is solved using the MKF algorithm.

[0049] The specific recursive solutions for α and β are as follows:

[0050]

[0051] Generally, the velocity v in Equation (4) n can be obtained from GNSS (Global Navigation Satellite System). However, when satellite signals are denied, how to provide a velocity reference for inertial initial alignment and achieve effective alignment of the inertial navigation system becomes a difficult problem that urgently needs to be solved. With the rapid development of deep learning technology, "data" has become an increasingly important resource. Therefore, a data-driven online velocity estimation method can be constructed by combining the coupling relationship between inertial devices (gyroscopes and accelerometers) and velocity in Equation (2). Discretizing Equation (2) gives:

[0052]

[0053] In the formula, F represents to the mapping relationship of. represent the output values of the gyroscope and accelerometer at time t respectively; represents the velocity increment.

[0054] Combining Equation (4) and Equation (5), it can be seen that the solution of α k is only related to inertial devices (gyroscopes and accelerometers), and the solution of β k is related to velocity and contains the velocity increment In the harmful acceleration parts of equations (4) and (5), the gravitational acceleration is the main factor and can be regarded as a fixed constant. Therefore, a non-linear mapping relationship can be constructed between the angular rate and specific force information output by inertial devices (gyroscopes and accelerometers) and the velocity increment. Through historical data, a deep learning method is used to construct an end-to-end learning model from inertial devices to velocity increment to achieve online estimation of velocity under satellite denial conditions and provide a velocity reference for inertial initial alignment.

[0055] TCN is a new deep learning network structure specifically designed for time series prediction. The convolutional layer of TCN combines two structures: dilated convolution and causal convolution, and a residual module is designed. To fully exploit the feature extraction and information mining capabilities in the TCN residual module, inspired by the double helix molecular structure of deoxyribonucleic acid (DNA) in biology, a TCN with a cross-helix structure (XTCN) is constructed. As Figure 3 shown, based on TCN, XTCN uses the residual module to connect to the TCN module in reverse, directly connects the output of the residual module to the causal convolutional layer, and directly connects the output of the Dropout layer to the convolutional layer, with a skip connection to the next residual module. The skip connection in XTCN allows the gradient to be directly passed to the shallow layer through the identity mapping, bypassing the non-linear transformation path, effectively alleviating the vanishing gradient problem, making it easier for the network to learn the residual mapping rather than directly learning the output. In addition, cross-layer reuse of features is also achieved, enhancing the model's ability to express the data output by inertial devices, extracting robust features from the noisy inertial device data, and deeply mining the coupling relationship between inertial device data and velocity increment.

[0056] The reverse connection idea of XTCN can be continuously passed on. As Figure 4 , on the basis of the XTCN structure in Figure 3 , the network structure can be continuously extended according to actual needs, giving full play to the advantages of residual connections. By increasing the complexity of the network structure, it can be adapted to complex application scenarios and improve the generalization ability of the network.

[0057] In XTCN, causal convolution is a special convolution operation. When performing convolution operations, only the features of the previous moments are considered, and the information of the subsequent moments will not be leaked, ensuring the causal relationship of time series data and avoiding the influence of future information on the current prediction. Let the filter function be F = (f1, f2,..., f N ), N be the number of filters, the time series X = (x1, x2,..., x j ), j be the number of sequences. To ensure that the sequence can be divisible by the convolution kernel size, the Padding(·) function is used to pad 0 at the end of the time series and then perform convolution. Then in the subsequence x jThe causal convolution at [location] is as follows:

[0058]

[0059] In the formula, x j is the j-th sequence. f n is the n-th filter function.

[0060] In XTCN, dilated convolution is a convolution operation that enlarges the receptive field. By adding a dilation rate to the convolution kernel, dilated convolution can increase the receptive field of the neural network without increasing the number of parameters, thereby better capturing long-term dependent temporal relationships and improving the accuracy of time series prediction. For the sequence X = (x1, x2,..., x J ) and the filter function F = (f1, f2,..., f N ), the dilated convolution G on the sequence element s is:

[0061]

[0062] In the formula, * represents the convolution operation. c is the convolution kernel size. d is the dilation coefficient. f(i) is the i-th convolution kernel. s - d·i is the past direction.

[0063] In XTCN, based on the residual module, the TCN network structure is connected through a spiral structure, and the residual connection method is as follows:

[0064] g = Activation(X + C(X)) (8)

[0065] In the formula, C(X) is the convolution. Activation(·) is the activation function.

[0066] This embodiment also provides an inertial initial alignment system based on deep learning, including:

[0067] A speed estimation unit that inputs the angular rate and specific force information of the inertial device into the trained deep learning network to obtain the speed increment, and further obtains the speed under satellite denial conditions; the deep learning network is a TCN network with a cross-spiral structure;

[0068] An initial attitude solving unit that, based on the speed obtained under satellite denial conditions, uses the MKF algorithm to solve the parameters of the constructed initial attitude matrix model to obtain the initial attitude and complete the initial alignment of the inertial navigation system.

[0069] This embodiment also provides an inertial initial alignment device, including: a memory, a processor, and a computer program stored on the memory. When the processor executes the computer program, the steps of the inertial initial alignment method are implemented.

[0070] This embodiment also provides a computer storage medium, which stores an executable program. When the executable program is executed by a processor, the steps of the inertial initial alignment method are implemented.

[0071] In summary, the present invention constructs a non-linear mapping relationship between the angular rate and specific force information output by inertial devices (gyroscopes and accelerometers) and the velocity increment; through historical data, a deep learning method is used to construct an end-to-end learning model from inertial devices to velocity increment to realize the online estimation of velocity under satellite denial conditions and provide a velocity reference for inertial initial alignment.

[0072] The above embodiments only represent the implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A deep learning-based inertial initial alignment method, characterized in that: Includes steps: Input the angular rate and specific force information of the inertial device into the trained deep learning network to obtain the velocity increment, and then obtain the velocity under satellite denial conditions; Based on the acquired velocity under satellite denial conditions, the MKF algorithm is used to solve the constructed initial attitude matrix model parameters, obtain the initial attitude, and complete the initial alignment of the inertial navigation system.

2. The inertial initial alignment method based on deep learning according to claim 1, characterized in that: The deep learning network is a TCN network with a cross-spiral structure, which uses a residual module to reversely connect the TCN module, directly connects the output of the residual module to the causal convolution layer, directly connects the output of the Dropout layer to the convolution layer, and jumps to the next layer of residual modules.

3. The inertial initial alignment method based on deep learning according to claim 2, characterized in that: The causal convolution layer uses the Padding(·) function to add 0 to the end of the time series and then performs convolution. j The causal convolution at is: Where, the filter function is F = (f1, f2, ..., f N ), time series X = (x1, x2, ..., x j ), j is the number of sequences, x j is the jth sequence, f n is the nth filter function, and N is the number of filters.

4. The inertial initial alignment method based on deep learning according to claim 3, characterized in that: The deep learning network increases the receptive field of the neural network by adding a dilation rate to the convolution kernel. For the sequence and filter function, the dilated convolution G on the sequence element s is: Where * represents the convolution operation, c is the convolution kernel size, d is the dilation coefficient, f(i) is the i-th convolution kernel, and sd·i is the past direction.

5. The inertial initial alignment method based on deep learning according to claim 3, characterized in that: The connection method of the residual module is: g = Activation(X + C(X)) Where C(X) is the convolution and Activation(·) is the activation function.

6. The inertial initial alignment method based on deep learning according to claim 1, characterized in that: The constructed initial posture matrix model is: Among them, the parameters α and β are: Where b(t) is the carrier coordinate system at time t, n(0) is the navigation coordinate system at time 0, b(0) is the carrier coordinate system at time 0, ω ie is the Earth's rotation angular velocity, ω en is the position velocity, f b is the triaxial specific force measured by the accelerometer, g n is the local gravitational acceleration, v n For speed.

7. The inertial initial alignment method based on deep learning according to claim 6, characterized in that: The recursive solution formula for parameters α and β is:

8. An inertial initial alignment system based on deep learning for implementing any of the methods described in claims 1-7, characterized in that: include: The velocity estimation unit inputs the angular rate and specific force information of the inertial device into the trained deep learning network to obtain the velocity increment, and then obtain the velocity under the satellite denial condition; the deep learning network is a TCN network with a cross-helical structure; The initial attitude solving unit uses the MKF algorithm to solve the constructed initial attitude matrix model parameters based on the acquired velocity under satellite denial conditions, obtains the initial attitude, and completes the initial alignment of the inertial navigation system.

9. An inertial initial alignment device, characterized in that: include: A memory, a processor and a computer program stored in the memory, wherein the processor implements the steps of the inertial initial alignment method according to any one of claims 1 to 7 when executing the computer program.

10. A computer storage medium, characterized in that: The computer storage medium stores an executable program, and the executable program is executed by a processor to implement the steps of the inertial initial alignment method according to any one of claims 1 to 7.

Images