Inertial navigation error suppression method and system based on LSTM-MMAE hybrid architecture, electronic equipment and storage medium

Abstract

The invention provides an inertial navigation error suppression method based on an LSTM-MMAE hybrid architecture, and the method comprises the steps: carrying out the data preprocessing of original IMU data, and obtaining a standardized time sequence; capturing a long-term time sequence dependency relationship of the standardized time sequence through a gating mechanism to obtain a speed increment predicted value; the noise covariance matrix of the speed increment predicted value and the observation noise covariance matrix of the theoretical trajectory speed are adjusted and fused through a plurality of parallel filtering models with different parameters, and optimal fusion estimation coping with different operation tasks is obtained; and carrying out speed prediction on the optimal fusion estimation to obtain optimal fusion speed estimation, and carrying out iteration on the optimal fusion speed estimation to obtain a target motion track. According to the method, through fusion of the LSTM neural network and multi-model adaptive estimation, IMU error accumulation is dynamically inhibited in a GPS failure scene, and the positioning precision and the system reliability are improved.

Description

Technical Field

[0001] The present invention relates to the field of inertial navigation and positioning technology, and in particular to an inertial navigation error suppression method, system, electronic device and storage medium based on an LSTM-MMAE hybrid architecture. Background Art

[0002] Currently, GPS / IMU fusion measurement systems are widely used in positioning and navigation. However, their architectural flaw, which strongly relies on a continuous GPS signal, poses significant risks in signal-restricted environments. When GPS fails due to environmental obstruction or hardware failure, the system degenerates into a pure IMU integration mode. Furthermore, due to the IMU's noise drift characteristics, position errors accumulate over time, causing positioning accuracy to plummet within minutes, severely limiting system reliability. Therefore, it is crucial to design an inertial navigation error suppression method, system, electronic device, and storage medium based on a LSTM-MMAE hybrid architecture. Summary of the Invention

[0003] The purpose of the present invention is to provide an inertial navigation error suppression method, system, electronic device and storage medium based on LSTM-MMAE hybrid architecture, which improves positioning accuracy and system reliability by integrating LSTM neural network and multi-model adaptive estimation.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] An inertial navigation error suppression method based on LSTM-MMAE hybrid architecture includes the following steps:

[0006] Perform data preprocessing on the raw IMU data to obtain a standardized time series; data preprocessing includes: noise filtering, time alignment and normalization;

[0007] The long-term temporal dependencies of the standardized time series are captured through the gating mechanism to obtain the speed increment prediction value;

[0008] The noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity are adjusted and fused through multiple parallel filtering models with different parameters to obtain the optimal fusion estimate.

[0009] The optimal fusion estimate is subjected to velocity prediction to obtain the optimal fusion velocity estimate, and the optimal fusion velocity estimate is iterated to obtain the target motion trajectory.

[0010] Optionally, a gating mechanism is used to capture the long-term temporal dependencies of the normalized temporal sequence to obtain a speed increment prediction value, including:

[0011] The standardized time series is selectively discarded through the sigmoid activation function and the hidden state of the previous moment to obtain the forgotten output;

[0012] The content information of the forgotten output is filtered and updated into the state cell to obtain the candidate cell state vector;

[0013] The state information that needs to be hidden in the candidate cell state vector is determined by the output parameters, and the speed increment prediction value is obtained.

[0014] Optionally, the noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity are adjusted and fused respectively through multiple parallel filtering models with different parameters to obtain an optimal fusion estimate, including:

[0015] Each filtering model is equipped with different noise covariance matrices and observation noise covariance matrices;

[0016] The velocity increment prediction value and the theoretical trajectory reference value are used as input data of the filtering model;

[0017] Update the filter model based on the input data to obtain the posterior state estimate and prediction error covariance matrix;

[0018] The matching degree between the current observation data and the filtering model is determined based on the likelihood calculated from the prediction error covariance matrix;

[0019] Determine the probability of belonging to the current filter model based on the model probability calculated based on the posterior state estimate;

[0020] The posterior state estimates are weightedly fused according to the matching degree and probability to obtain the optimal fusion estimate.

[0021] Optionally, the filter model is updated and predicted based on the input data to obtain a posterior state estimate and a prediction error covariance matrix, including:

[0022] Predict the current state and error covariance based on the input data and the state estimate at the previous moment;

[0023] Calculate the residual and Kalman gain of the current observation data;

[0024] The current state and error covariance are corrected by the residual and Kalman gain respectively to obtain the posterior state estimate and prediction error covariance matrix.

[0025] An inertial navigation error suppression system based on LSTM-MMAE hybrid architecture, including:

[0026] The data preprocessing module is used to perform data preprocessing operations on the original IMU data to obtain a standardized time series;

[0027] The feature extraction module is used to capture the long-term temporal dependencies of the normalized time series through a gating mechanism to obtain the speed increment prediction value;

[0028] The fusion module is used to adjust and fuse the noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity through multiple parallel filtering models with different parameters to obtain the optimal fusion estimate;

[0029] The recursive integration output module predicts the speed of the optimal fusion estimate to obtain the optimal fusion speed estimate, and iterates the optimal fusion speed estimate to obtain the target motion trajectory.

[0030] Optionally, the data preprocessing module includes: a noise filtering unit, a time alignment unit and a normalization layer;

[0031] The noise filtering unit is used to filter and denoise the original IMU data to generate denoised data;

[0032] The time alignment unit is used to perform time alignment processing on the denoised data to generate aligned data;

[0033] The normalization layer is used to normalize the aligned data and generate a standardized time series.

[0034] Optionally, the feature extraction module is an LSTM network, including: a forget gate, an input gate, and an output gate;

[0035] The forget gate selectively discards the standardized time series through the sigmoid activation function and the hidden state of the previous moment to obtain the forget output;

[0036] The input gate is used to filter the content information of the forgotten output and update it into the state cell to obtain the candidate cell state vector;

[0037] The output gate is used to determine the state information that needs to be hidden in the candidate cell state vector through the output parameters to obtain the speed increment prediction value.

[0038] Optionally, the fusion module is an MMAE model, including: multiple parallel Kalman filters and a dynamic weight allocation unit; the Kalman filters are all connected to the dynamic weight allocation unit.

[0039] An electronic device, comprising:

[0040] at least one processor;

[0041] a memory communicatively coupled to the at least one processor;

[0042] The memory stores instructions that can be executed by at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned error suppression method.

[0043] A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the error suppression method.

[0044] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the inertial navigation error suppression method based on the LSTM-MMAE hybrid architecture provided by the present invention comprises: performing data preprocessing on the original IMU data to obtain a standardized time series sequence; capturing the long-term time series dependency of the standardized time series sequence through a gating mechanism to obtain a velocity increment prediction value; adjusting and fusing the noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity through multiple parallel filtering models with different parameters to obtain an optimal fusion estimate; performing velocity prediction on the optimal fusion estimate to obtain an optimal fusion velocity estimate, and iterating the optimal fusion velocity estimate to obtain the target motion trajectory. By fusing the LSTM neural network with multi-model adaptive estimation, this method dynamically suppresses IMU error accumulation in GPS failure scenarios, thereby improving positioning accuracy and system reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a diagram of the inertial navigation error suppression method of the present invention;

[0047] Figure 2 Schematic diagram of the LSTM network structure of an embodiment of the present invention;

[0048] Figure 3 Schematic diagram of the MMAE network structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1 As shown, the present invention provides an inertial navigation error suppression method based on LSTM-MMAE hybrid architecture, comprising the following steps:

[0052] Step 100: Perform data preprocessing on the raw IMU data to obtain a standardized time series; data preprocessing includes: noise filtering, time alignment and normalization;

[0053] Step 200: Capture the long-term temporal dependency of the standardized temporal sequence through a gating mechanism to obtain a speed increment prediction value;

[0054] Specifically, the gating mechanism is composed of a forget gate, an input gate, and an output gate. First, the forget gate outputs the parameter f t Determine how many cells C in the previous state of the normalized time series to discard t-1 The information is expressed as:

[0055] f t =σ(W f ·[h t-1 ,X t ]+b f );

[0056] Where σ is the sigmoid activation function, W f is the weight matrix of the forget gate, h t-1 is the hidden state of the previous step, X t is the current input, b f Is the bias term of the forget gate. Then the output parameter i is input through the gate t Determine how much of the current output information of the forget gate needs to be updated to the state cell C t In the expression:

[0057] i t =σ(W i ·[h t-1 ,X t ]+b i );

[0058]

[0059] Where W i is the weight matrix of the input gate, b i is the bias term of the input gate, W C is the weight matrix of candidate state cells, b C is the bias term of the candidate state cell, is the candidate cell state vector. Finally, the output gate outputs the parameter o t Decide on the state of cell C t How much information is output as the hidden state and the output speed increment prediction value is expressed as:

[0060] o t =σ(W o ·[h t-1 ,X t ]+b o );

[0061] h t =o t *tanh(C t );

[0062] Where W o is the weight matrix of the output gate, b o is the bias term of the output gate, the symbol * represents the element-level Hadamard product, h t is the current hidden state. The update equation of the state cell is:

[0063] Step 300: Using multiple parallel filtering models with different parameters, the noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity are adjusted and fused to obtain the optimal fusion estimate. The specific steps include:

[0064] Each filtering model is equipped with different noise covariance matrices and observation noise covariance matrices;

[0065] The velocity increment prediction value and the theoretical trajectory reference value are used as input data of the filtering model;

[0066] Update the filter model based on the input data to obtain the posterior state estimate and prediction error covariance matrix. Specifically, the current state and error covariance are predicted based on the input data and the state estimate at the previous moment. The residual and Kalman gain of the current observation data are calculated.

[0067] The current state and error covariance are corrected by the residual and Kalman gain respectively to obtain the posterior state estimate and prediction error covariance matrix.

[0068] The matching degree between the current observation data and the filtering model is determined based on the likelihood calculated from the prediction error covariance matrix;

[0069] Determine the probability of belonging to the current filter model based on the model probability calculated based on the posterior state estimate;

[0070] The posterior state estimates are weightedly fused according to the matching degree and probability to obtain the optimal fusion estimate for different operation tasks.

[0071] It should be noted that the traditional filtering model includes three steps: state prediction, measurement update, and state update. In the prediction step, the state and error covariance of the current moment are predicted based on the state estimation and control input of the previous moment. State prediction It is based on the state estimation of the previous moment to predict the state of the current moment, and the prediction error covariance matrix P k|k-1 It represents the uncertainty of the prediction, which is expressed as:

[0072]

[0073] Among them A k-1 is the state transfer matrix from the previous moment to the current moment, B k-1 is the control input matrix, u k-1 is the control input vector, Q k-1 is the process noise covariance matrix.

[0074] In the measurement update step, the residual and Kalman gain are calculated using the observation data at the current moment. The Kalman gain matrix K determines the weight of the observation data in the update, and the observation error covariance matrix S k Represents the uncertainty of the observation process, and the residual y can represent the difference between the observed data and the predicted observed data. The expression is:

[0075]

[0076] Among them H k is the observation matrix, R k is the observation noise covariance matrix, z k is the observed data, i.e. the velocity of the theoretical trajectory, is the optimized IMU velocity increment, y k It is the speed after fusion optimization.

[0077] In the state update step, the predicted state estimate and covariance are modified based on the current observation data. It is the model output at the current moment, the prediction error covariance matrix P k|k It is the input of the next round of state prediction, and its expression is:

[0078]

[0079] P k|k =(IK k H k )P k|k-1 ;

[0080] where I is the identity matrix.

[0081] Compared with the traditional filtering model, the embodiment of the present invention adds three key steps: likelihood calculation, model probability update and state fusion. k |M (i) ) indicates that in model M (i) Next, the residual y k The probability density of occurrence, the higher the likelihood, the closer the current observation data is to the model M (i) The higher the matching degree, the calculation formula is:

[0082]

[0083] where y k represents the residual of the sub-model, M (i) represents the i-th model, represents the observation error covariance matrix of the i-th model, m is the observation vector z k dimension.

[0084] Model probability Indicates that at time k, the i-th model M (i) The posterior probability indicates that the current system is in model M (i) The probability is calculated as follows:

[0085]

[0086] Where N is the total number of models.

[0087] State estimation vector Represents the comprehensive state estimation vector at time k, which is the output of the system at the current moment, P k is the comprehensive error covariance matrix at time k, which represents the uncertainty of the comprehensive state and is used as the input of the next round of the system. The calculation formula is:

[0088]

[0089] in Represents the i-th model M (i) The posterior state estimate at time k is the output of the sub-model, is the i-th model M (i) The posterior error covariance matrix at time k.

[0090] Step 400: Perform velocity prediction on the optimal fusion estimate to obtain the optimal fusion velocity estimate, and iterate the optimal fusion velocity estimate to obtain the target motion trajectory.

[0091] Specifically, for different running tasks, the model parameters are adaptively adjusted, and then multiple filter models with different parameters are predicted in parallel to estimate the parameters with the best effect, thereby obtaining the optimal fusion speed estimation. The speed results are then iteratively fused to obtain the target motion trajectory.

[0092] The present invention also provides an inertial navigation error suppression system based on LSTM-MMAE hybrid architecture, comprising:

[0093] The data preprocessing module is used to perform data preprocessing operations on the original IMU data to obtain a standardized time series;

[0094] The feature extraction module is used to capture the long-term temporal dependencies of the normalized time series through a gating mechanism to obtain the speed increment prediction value;

[0095] The fusion module is used to adjust and fuse the noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity through multiple parallel filtering models with different parameters to obtain the optimal fusion estimate;

[0096] The recursive integration output module is used to perform velocity prediction on the optimal fusion estimate to obtain the optimal fusion velocity estimate, and iterate the optimal fusion velocity estimate to obtain the target motion trajectory.

[0097] Specifically, the data preprocessing module includes: a noise filtering unit, a time alignment unit and a normalization layer;

[0098] The noise filtering unit is used to filter and denoise the original IMU data to generate denoised data;

[0099] The time alignment unit is used to perform time alignment processing on the denoised data to generate aligned data;

[0100] The normalization layer is used to normalize the aligned data and generate a standardized time series.

[0101] like Figure 2 As shown, the feature extraction module is an LSTM network, including: forget gate, input gate and output gate;

[0102] The forget gate selectively discards the standardized time series through the sigmoid activation function and the hidden state of the previous moment to obtain the forget output;

[0103] The input gate is used to filter the content information of the forgotten output and update it into the state cell to obtain the candidate cell state vector;

[0104] The output gate is used to determine the state information that needs to be hidden in the candidate cell state vector through the output parameters to obtain the speed increment prediction value.

[0105] like Figure 3 As shown, the fusion module is an MMAE model, including: multiple parallel Kalman filters and dynamic weight allocation units; the Kalman filters are all connected to the dynamic weight allocation units.

[0106] An embodiment of the present invention further provides an electronic device, characterized by comprising:

[0107] at least one processor;

[0108] a memory communicatively coupled to the at least one processor;

[0109] The memory stores instructions that can be executed by at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned error suppression method.

[0110] An embodiment of the present invention further provides a non-transitory computer-readable storage medium storing computer instructions, where the computer instructions are used to enable a computer to execute the above-mentioned error suppression method.

[0111] The beneficial effects of the present invention are as follows:

[0112] 1) LSTM prediction replaces integral operations, reducing algorithm complexity, improving computational efficiency and accuracy, and significantly suppressing errors;

[0113] 2) By dynamically switching models using MMAE, data accuracy remains consistent across multiple motion modes;

[0114] 3) The multi-model adaptive estimation model (MMAE) dynamically adjusts the process noise covariance matrix and the observation noise covariance matrix of each sub-model, achieving optimal model switching under different motion modes, improving the motion diversity of the feed cabin, and enhancing the accuracy and stability of trajectory estimation;

[0115] 4) No additional sensor expansion is required and it is compatible with existing GPS / IMU systems, solving the problems of IMU data accumulation error and insufficient real-time performance of theoretical trajectory data in scenarios where GPS is unavailable.

[0116] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0117] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for suppressing inertial navigation errors based on LSTM-MMAE hybrid architecture, characterized in that: The steps include: Performing data preprocessing on the raw IMU data to obtain a standardized time series; the data preprocessing includes: noise filtering, time alignment and normalization; Capturing the long-term temporal dependency of the standardized temporal sequence through a gating mechanism to obtain a speed increment prediction value; The noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity are adjusted and fused respectively through multiple parallel filtering models with different parameters to obtain an optimal fusion estimate; The optimal fusion estimate is subjected to speed prediction to obtain an optimal fusion speed estimate, and the optimal fusion speed estimate is iterated to obtain a target motion trajectory.

2. The inertial navigation error suppression method based on the LSTM-MMAE hybrid architecture according to claim 1 is characterized in that The long-term temporal dependency of the standardized time series is captured by a gating mechanism to obtain a speed increment prediction value, including: The standardized time series is selectively discarded through a sigmoid activation function and a hidden state at a previous moment to obtain a forgotten output; Filtering the content information of the forgotten output and updating it into the state cell to obtain a candidate cell state vector; The state information to be hidden in the candidate cell state vector is determined by outputting parameters to obtain the speed increment prediction value.

3. The inertial navigation error suppression method based on the LSTM-MMAE hybrid architecture according to claim 1 is characterized in that The noise covariance matrix of the velocity increment prediction value and the observed noise covariance matrix of the theoretical trajectory velocity are adjusted and fused respectively through multiple parallel filtering models with different parameters to obtain an optimal fusion estimate, including: Each of the filtering models is equipped with a different noise covariance matrix and observation noise covariance matrix; Using the speed increment prediction value and the theoretical trajectory reference value as input data of the filtering model; Performing update prediction on the filtering model according to the input data to obtain a posterior state estimate and a prediction error covariance matrix; Determining the matching degree between the current observation data and the filtering model according to the likelihood calculated by the prediction error covariance matrix; Determine the probability of belonging to the current filtering model according to the model probability calculated by the posterior state estimation; The posterior state estimates are weightedly fused according to the matching degree and the probability to obtain the optimal fused estimate.

4. The inertial navigation error suppression method based on the LSTM-MMAE hybrid architecture according to claim 3 is characterized in that The filter model is updated and predicted according to the input data to obtain a posterior state estimate and a prediction error covariance matrix, including: Predicting the current state and error covariance based on the input data and the state estimate at the previous moment; Calculating the residual and Kalman gain of the current observation data; The current state and the error covariance are respectively corrected by the residual and the Kalman gain to obtain the posterior state estimate and the prediction error covariance matrix.

5. An inertial navigation error suppression system based on LSTM-MMAE hybrid architecture, characterized in that: include: The data preprocessing module is used to perform data preprocessing operations on the original IMU data to obtain a standardized time series; a feature extraction module, configured to capture the long-term temporal dependency of the standardized temporal sequence through a gating mechanism to obtain a speed increment prediction value; A fusion module is used to adjust and fuse the noise covariance matrix of the velocity increment prediction value and the observation noise covariance matrix of the theoretical trajectory velocity through multiple parallel filtering models with different parameters to obtain an optimal fusion estimate; The recursive integration output module is used to perform speed prediction on the optimal fusion estimate to obtain an optimal fusion speed estimate, and iterate the optimal fusion speed estimate to obtain a target motion trajectory.

6. The inertial navigation error suppression system based on the LSTM-MMAE hybrid architecture according to claim 5, characterized in that: The data preprocessing module includes: a noise filtering unit, a time alignment unit and a normalization layer; The noise filtering unit is used to perform filtering and denoising processing on the original IMU data to generate denoised data; The time alignment unit is used to perform time alignment processing on the denoised data to generate aligned data; The normalization layer is used to perform normalization processing on the aligned data to generate the standardized time series.

7. The inertial navigation error suppression system based on the LSTM-MMAE hybrid architecture according to claim 5, characterized in that: The feature extraction module is an LSTM network, including a forget gate, an input gate, and an output gate; The forget gate selectively discards the standardized time series through the sigmoid activation function and the hidden state at the previous moment to obtain a forget output; The input gate is used to filter the content information of the forgotten output and then update it into the state cell to obtain a candidate cell state vector; The output gate is used to determine the state information that needs to be hidden in the candidate cell state vector through output parameters to obtain the speed increment prediction value.

8. The inertial navigation error suppression system based on the LSTM-MMAE hybrid architecture according to claim 5, characterized in that: The fusion module is an MMAE model, including: Multiple Kalman filters and dynamic weight allocation units are arranged in parallel; the Kalman filters are all connected to the dynamic weight allocation units.

9. An electronic device, characterized in that: include: at least one processor; a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the error suppression method according to any one of claims 1 to 4.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to cause a computer to execute the error suppression method according to any one of claims 1 to 4.

Images