Fusion optimization method and device based on Kalman filtering and LSTM cascade, and integrated navigation method and system

Abstract

The invention discloses a fusion optimization method based on Kalman filtering and LSTM cascade. The fusion optimization method comprises three steps of dynamic state estimation, time sequence error modeling and closed-loop fusion optimization. IMU (Inertial Measurement Unit) data is used as input, dynamic state estimation is realized through Kalman filtering, modeling system errors are separated, an LSTM (Long Short Term Memory) network is used for carrying out time sequence modeling on a residual error sequence to capture nonlinear errors, and finally, a corrected state quantity is fed back to a Kalman filtering updating link through a closed-loop feedback mechanism. And collaborative optimization of error compensation and state estimation is realized. According to the method, IMU error accumulation is effectively inhibited through a dynamic-data dual-drive mechanism, the navigation precision and robustness in a complex scene are remarkably improved, and the requirements of high-dynamic applications such as intelligent driving and unmanned aerial vehicle navigation can be met. Finally, the optimized IMU data and the GNSS observation value are fused for integrated navigation, high-precision navigation solution is achieved through Kalman filtering, and indoor and outdoor seamless positioning and the all-attitude control requirement of a high-dynamic carrier are met.

Description

Technical Field

[0001] The present invention relates to the field of navigation technology, and more specifically, to a fusion optimization method and device based on Kalman filtering and LSTM cascade, and a combined navigation method and system. Background Art

[0002] With the rapid development of intelligent driving, drone navigation, and mobile robotics, high-precision, high-reliability integrated navigation systems have become a key technology for supporting complex scenarios. Among the many integrated navigation solutions, integrated navigation using inertial measurement units (IMUs) and global navigation satellite systems (GNSSs) is widely adopted due to their complementary nature. IMUs provide high-frequency autonomous motion information, while GNSSs provide absolute position corrections. However, traditional methods often rely on Kalman filters (KFs) or their extended forms for dynamic fusion. However, their performance is limited by linear assumptions and the accuracy of noise models, making them incapable of coping with IMU nonlinear errors and GNSS anomalies. In recent years, machine learning techniques have been introduced into the navigation field, but purely data-driven approaches lack prior knowledge of system dynamics and rely on high-quality training data, making them difficult to maintain robustness in dynamic and sudden scenarios.

[0003] An IMU-GNSS integrated navigation system primarily consists of an inertial measurement unit (IMU) and a global navigation satellite system (GNSS). The IMU typically includes a three-axis accelerometer and a three-axis gyroscope to measure the vehicle's acceleration and angular velocity. The GNSS receiver receives signals from multiple satellites and uses triangulation to calculate the vehicle's three-dimensional position and velocity. The integrated navigation system combines IMU and GNSS data through data fusion algorithms to improve navigation accuracy and reliability. Common data fusion algorithms include the Kalman filter and its extensions, such as the extended Kalman filter (EKF) and the unscented Kalman filter (UKF). Furthermore, adaptive filtering methods improve the robustness of filtering algorithms by adjusting the noise covariance matrix in real time. For example, the adaptive KF algorithm based on innovation covariance can adjust the filter gain when the GNSS signal is intermittent. Machine learning techniques, such as long-short-term memory (LSTM), have also been introduced in the navigation field to capture nonlinear error patterns in the data.

[0004] Traditional filtering methods perform poorly when the statistical characteristics of dynamic noise are unknown or sudden, have limited ability to model nonlinear errors, and suffer from high computational complexity. The isolated use of machine learning techniques and filtering algorithms results in low information utilization, a lack of prior knowledge of system dynamics, and insufficient physical interpretability. Existing research is often limited to improving single methods and lacks systematic validation of the "filtering + machine learning" fusion strategy, particularly inadequate robustness assessment in dynamic and sudden change scenarios. Summary of the Invention

[0005] To address the challenges of existing technologies, this paper proposes a fusion optimization method based on a cascaded Kalman filter and LSTM. Furthermore, it provides an IMU-GNSS integrated navigation framework using this cascaded Kalman filter and LSTM. This approach aims to effectively combine the advantages of model-driven and data-driven approaches, improving the stability and robustness of navigation systems in complex scenarios. The technical framework achieves this fusion through a cascaded closed-loop structure of a KF (model-driven) and an LSTM (data-driven). The LSTM models the nonlinear errors in the residuals and feeds the corrected state back to the KF update process, forming a closed-loop optimization process.

[0006] Traditional methods are mostly one-way filtering processes. Compared with existing adaptive methods, the present invention, in terms of scene adaptability, targets complex mountain scenes by capturing terrain-related nonlinear error patterns through LSTM, combined with KF's adaptive noise adjustment, effectively solving the problem that existing methods do not clearly address the IMU nonlinear drift in mountainous environments, and significantly improving stability in complex scenes. In the joint optimization of noise covariance adjustment, KF is used to separate the modelable system errors, and only the remaining noise is adaptively adjusted, reducing computational complexity. At the same time, the time series modeling capability of LSTM for residuals is utilized to dynamically correct the estimated deviation of the noise covariance, while existing methods mostly rely on statistical models or fixed thresholds. Specific improvements include: first, using the Kalman filter to estimate the IMU's gyroscope and accelerometer biases in real time, suppressing high-frequency noise through state equations and observation equations, and separating the modelable system errors (gyroscope zero bias, accelerometer zero bias), and using an adaptive noise covariance adjustment strategy to improve the robustness of the filtering algorithm in complex mountainous scenarios; using the time series modeling capability of the LSTM network to capture the nonlinear error pattern in the residual and achieve effective compensation for the IMU's nonlinear error; feeding the LSTM corrected state back to the update link of the Kalman filter to form a closed-loop optimization mechanism, making full use of the system's dynamic characteristics, improving information utilization, and taking into account both real-time and long-term stability.

[0007] In order to achieve the above objectives, the first aspect of the present invention provides a fusion optimization method based on Kalman filtering and LSTM cascade, comprising: Taking IMU data as input, Kalman filtering is used to perform dynamic state estimation and separate modelable system errors, outputting state estimates and residual sequences, where the residual sequence is a sequence consisting of the difference between the observed values ​​and the state estimates; The pre-built LSTM network model is used to obtain the residual correction value according to the input residual; The state estimation value output by the Kalman filter and the residual correction obtained by the LSTM network are fed back to the Kalman filter update stage to perform collaborative optimization of error compensation and state estimation to obtain the corrected IMU data.

[0008] In one embodiment, the IMU data is used as input, a Kalman filter is used to perform dynamic state estimation and separate modelable system errors, and the output state estimate and residual sequence are included: Using three-axis angular velocity data and three-axis acceleration data as input, the state prediction is performed through the prediction equation to separate the gyroscope bias and accelerometer bias errors; The state estimates and residual sequences are generated by adjusting the covariance matrix update.

[0009] In one embodiment, a pre-built LSTM network model is used to obtain a residual correction value based on the input residual, including: Build a two-layer LSTM network; The historical data set is input into the LSTM network model. The nonlinear error pattern related to the terrain is learned by LSTM. The model parameters are optimized using the mean square error loss function to obtain the trained LSTM network model: The residual is input into the trained LSTM network model. The input layer receives the residual sequence, and the output layer generates a nonlinear error compensation as the residual correction amount.

[0010] In one embodiment, the state estimation value output by the Kalman filter and the residual correction value obtained by the LSTM network are fed back to the Kalman filter update stage to perform collaborative optimization of error compensation and state estimation, including: The state estimate and the residual correction are used as inputs in the Kalman filter update phase to correct the state and obtain the corrected IMU state. Adopting adaptive noise covariance adjustment strategy to reset the covariance matrix; The corrected state and the reset covariance matrix are used as the initial values ​​of the next filtering moment to form a closed-loop optimization and obtain the corrected IMU data. In one embodiment, the state estimate and the residual correction are used as inputs in the Kalman filter update phase to correct the state, thereby obtaining a corrected IMU state, including:

[0011] in, is the corrected IMU state, is the state estimate, is the residual correction amount.

[0012] In one embodiment, the covariance matrix is ​​reset using an adaptive noise covariance adjustment strategy, including:

[0013] in, is the reset covariance matrix, is the reset factor.

[0014] Based on the same inventive concept, the second aspect of the present invention provides a combined navigation method based on Kalman filtering and LSTM cascade, which performs IMU-GNSS loose combined navigation on the corrected IMU data obtained in the first aspect and the positioning result of PPP to obtain a combined navigation result.

[0015] Based on the same inventive concept, the third aspect of the present invention provides a fusion optimization device based on Kalman filtering and LSTM cascade, comprising: The Kalman filter dynamic state estimation module is used to take IMU data as input, use Kalman filtering to perform dynamic state estimation and separate modelable system errors, and output state estimation values ​​and residual sequences, where the residual sequence is a sequence consisting of the difference between the observed value and the state estimation value; The LSTM time series error modeling module is used to obtain the residual correction value based on the input residual using the pre-built LSTM network model; The fusion optimization module is used to feed back the state estimation value output by the Kalman filter and the residual correction obtained by the LSTM network to the Kalman filter update stage, perform collaborative optimization of error compensation and state estimation, and obtain the corrected IMU data.

[0016] Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the fusion optimization method based on Kalman filtering and LSTM cascade described in the first aspect is implemented.

[0017] Based on the same inventive concept, the fifth aspect of the present invention provides a combined navigation system, including the third aspect of the fusion optimization device based on Kalman filtering and LSTM cascade, a PPP solution module and a fusion module, wherein the PPP solution module is used to receive the observation quantity and ephemeris information broadcast by the GNNS satellite, and perform PPP solution to output position information. The fusion module is used to correct the navigation parameters through Kalman filtering, and feed back the IMU data output by the fusion optimization device based on Kalman filtering and LSTM cascade as an error into the inertial navigation solution to obtain the positioning result.

[0018] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows: The present invention provides a fusion optimization method based on Kalman filtering and LSTM cascade, which mainly includes dynamic state estimation, multi-source error modeling, and closed-loop fusion optimization. Taking the three-axis angular velocity, acceleration and GNSS observation values ​​of the vehicle-mounted IMU as input, real-time posture estimation and linear error separation are achieved through Kalman filtering; an LSTM neural network model is constructed to learn the nonlinear error pattern from the time series residual to correct the original IMU data; a closed-loop feedback strategy is designed to reversely inject the corrected state quantity into the Kalman filter update link, and combined with covariance reset to suppress historical error accumulation. This method effectively suppresses IMU error accumulation through a dynamic-data dual-drive mechanism, significantly improves navigation accuracy and robustness in complex scenarios, and can meet the needs of high-dynamic applications such as intelligent driving and drone navigation.

[0019] Furthermore, a combined navigation method based on Kalman filtering and LSTM cascade is provided, which integrates the optimized IMU data with GNSS observation values ​​for combined navigation. High-precision navigation solution is achieved through Kalman filtering, supporting seamless indoor and outdoor positioning and full attitude control requirements of highly dynamic carriers. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] 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 or the description of the prior art. Obviously, the drawings described below are 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.

[0021] Figure 1 This is an overall flow chart of the fusion optimization method based on Kalman filtering and LSTM cascade in an embodiment of the present invention; Figure 2 Flowchart of Kalman filtering in an embodiment of the present invention; Figure 3 This is a structural principle diagram of the LSTM model in an embodiment of the present invention; Figure 4 This is a flowchart of loose combined navigation in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. 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 shall fall within the scope of protection of the present invention.

[0023] First, the relevant methods and technologies involved in the present invention are described: (1) Inertial Measurement Unit: A device used to measure the three-axis attitude angle (or angular rate) and acceleration of an object. Its core components include accelerometers and gyroscopes. The accelerometer is used to detect the acceleration signals of the object in the three independent axes in the carrier coordinate system, while the gyroscope is used to detect the angular velocity signals of the carrier relative to the navigation coordinate system. Through these two sensors, the IMU can measure the angular velocity and acceleration of the object in three-dimensional space, and calculate the object's attitude, velocity, displacement and other information based on this. Some advanced IMUs are also equipped with magnetometers to assist in calibrating directional drift, thereby improving measurement accuracy.

[0024] (2) Global Navigation Satellite System: It is a technology for positioning based on a satellite navigation system, mainly composed of the US GPS system, Russia's GLONASS system, China's BeiDou navigation system and Europe's Galileo system. Its working principle is based on triangulation, which determines the precise position of the receiver by receiving signals from at least four satellites with known positions and measuring the signal propagation time. The signal transmitted by the satellite contains time and position information. The receiver synchronizes time with the pseudo-random code in the satellite signal, measures the time delay and calculates the distance. This technology can accurately determine the position, speed and time information of the receiving device, thereby realizing positioning and navigation functions.

[0025] (3) LSTM neural network model: LSTM (Long Short-Term Memory) is a special recurrent neural network structure designed to solve the gradient vanishing and gradient exploding problems of traditional RNN when processing long sequence data. LSTM can effectively capture long-term dependencies through the design of carefully designed gating mechanisms and memory cells, and performs well in sequence modeling tasks. The core of the LSTM network is the LSTM unit, each of which contains a memory cell and three control gates: forget gate, input gate, and output gate. The forget gate determines what information to discard from the memory cell at the previous moment; the input gate determines what new information to obtain from the current input and the previous state and update the memory cell state; the output gate determines what state to output from the memory cell as the output at the current moment. Through the interaction of these gating mechanisms and memory cells, LSTM can selectively retain and forget information, thereby better capturing long-term dependencies. LSTM has been widely used in natural language processing, speech recognition, time series prediction and other fields, showing excellent performance.

[0026] (4) Machine learning: A branch of artificial intelligence (AI) that enables computer systems to automatically learn and improve from data without being explicitly programmed. Its core goal is to develop algorithms and models that enable computers to recognize patterns based on data, make decisions, and continuously improve their performance as they accumulate experience. The application range of machine learning is wide, covering a variety of fields, from image recognition and natural language processing to predictive analysis and recommendation systems.

[0027] (5) Kalman filter: The Kalman filter is an efficient recursive filter that can estimate the state of a dynamic system from a series of incomplete and noisy measurements. The key idea of ​​the Kalman filter is data fusion and iterative updating. By combining prior information and actual measurements, the system can better estimate the current state and gradually improve the accuracy of the estimate through iterative updates. It is a recursive process that only uses the current input measurement and the previously calculated state and its uncertainty matrix to operate, without requiring additional past information.

[0028] (6) Integrated navigation: It is a technology that improves positioning accuracy and reliability by integrating information from multiple navigation systems (such as inertial navigation system (INS), global navigation satellite system (GNSS), visual navigation, laser ranging, sonar detection, odometer, etc.). Its core principle is to utilize the complementarity of different navigation systems and integrate the measurement data of each system through data fusion algorithms (such as Kalman filter, particle filter, etc.) to obtain more accurate and reliable navigation information. Example 1 This embodiment discloses a fusion optimization method based on Kalman filtering and LSTM cascade, see Figure 1 ,include: S1: Taking IMU data as input, Kalman filtering is used to perform dynamic state estimation and separate modelable system errors, and output state estimation values ​​and residual sequences, where the residual sequence is a sequence consisting of the difference between the observed values ​​and the state estimation values.

[0029] Specifically, S1 is the Kalman filter dynamic state estimation step, and the input IMU data includes three-axis angular velocity data and three-axis acceleration data.

[0030] S2: Use the pre-built LSTM network model to obtain the residual correction value based on the input residual.

[0031] Specifically, S2 is the LSTM time series error modeling step, which uses the pre-built LSTM network model to capture the nonlinear error pattern related to the terrain and obtain the residual correction amount based on the input residual.

[0032] S3: Feed the state estimation value output by the Kalman filter and the residual correction obtained by the LSTM network back to the Kalman filter update stage to perform collaborative optimization of error compensation and state estimation to obtain the corrected IMU data.

[0033] Specifically, the S3 closed-loop fusion optimization step feeds back the LSTM corrected state to the update link of the Kalman filter, forming a closed-loop optimization mechanism that fully utilizes the system dynamics, improves information utilization, and takes into account both real-time and long-term stability.

[0034] The present invention discloses a combined navigation method based on Kalman filtering and LSTM cascade, belonging to the field of high-precision navigation and positioning technology. The method mainly carries out the work of dynamic state estimation, time series error modeling and closed-loop fusion optimization. Taking IMU data as input, dynamic state estimation is realized and modelable system errors (gyroscope bias, accelerometer bias) are separated through Kalman filtering. The residual sequence is time-series modeled using an LSTM network to capture nonlinear errors. Finally, the corrected state quantity is fed back to the Kalman filter update link through a closed-loop feedback mechanism to achieve coordinated optimization of error compensation and state estimation. This method effectively suppresses IMU error accumulation through a dynamic-data dual-drive mechanism, significantly improves navigation accuracy and robustness in complex scenarios, and can meet the needs of high-dynamic applications such as intelligent driving and drone navigation.

[0035] The main protection points include two aspects: (1) Closed-loop fusion architecture and state correction equation design: A bidirectional interactive closed-loop fusion mechanism is constructed to feed back the error compensation value output by the LSTM to the state update link of the Kalman filter in real time, and eliminate historical error interference through a covariance reset strategy. This mechanism breaks through the limitation of one-way error propagation in traditional methods and realizes the dynamic synergy of model-driven and data-driven. The protection point focuses on the design of the state correction equation and covariance reset of closed-loop fusion to ensure the stability of the system in dynamic mutation scenarios.

[0036] (2) Error source hierarchical decoupling method and its implementation path: A hierarchical decoupling method for error sources is proposed. The Kalman filter prioritizes linear noise and modelable errors, and the LSTM performs time series modeling on nonlinear residuals. The protection point covers the design of error separation and the implementation path of the data and model dual-driven architecture, effectively improving the compensation accuracy of complex errors. In the specific implementation, the Kalman filter is used to estimate and separate the system error and noise in real time, and the LSTM is used to capture the nonlinear error pattern. Finally, the corrected state is fed back to the update link of the Kalman filter to form a closed-loop optimization.

[0037] The prior art discloses CN117332813A, an IMU attitude estimation method based on KF and LSTM fusion (prior art 1) and CN118089710A, an inertial navigation performance maintenance method based on IMU and LSTM assistance (prior art 2). The following describes the differences between the present invention and the prior art from the aspects of inventive concept, technical route, specific implementation method, and technical effect.

[0038] (1) Inventive concept The core of this invention lies in integrating the advantages of Kalman filtering and LSTM to build a closed-loop fusion architecture, achieving dynamic collaboration between model-driven and data-driven approaches to improve the accuracy and robustness of integrated navigation systems in complex scenarios. Particularly focusing on the nonlinear drift of IMUs in complex mountainous scenarios, this method, through the cascaded closed-loop structure of KF and LSTM, forms a bidirectional interactive closed-loop fusion mechanism, overcoming the limitation of the one-way error propagation in traditional cascade methods.

[0039] Existing Technology 1: This technique primarily focuses on IMU pose estimation, leveraging an attention mechanism to redistribute the weights of the Kalman filter and LSTM network outputs to improve pose estimation accuracy and robustness. Its core approach is to fuse the outputs of the two methods through the attention mechanism, rather than building a closed-loop fusion architecture.

[0040] Existing Technology 2: This approach focuses on maintaining inertial navigation performance, using Kalman filtering for error correction and combining it with LSTM to build an inertial device error prediction model to address the error divergence problem of inertial navigation systems. This approach focuses on using LSTM to model and predict inertial device errors to compensate for them.

[0041] Error Correction in Prior Art 2 Error Correction Type: Existing Technology 2 primarily corrects inertial device errors within the inertial navigation system. Specifically, in a GNSS / INS integrated navigation system, Kalman filtered information is combined with the previous moment's inertial navigation information to calculate inertial device errors. This inertial device error includes errors in the gyroscope and accelerometer along three axes and is related to changes in the carrier's motion state.

[0042] Correction Purpose and Method: The goal is to address the problem of error divergence in the inertial navigation system and improve the accuracy and reliability of inertial navigation system positioning. This is achieved by building an LSTM-based inertial device error prediction model. This model is trained when satellite signals are valid. When satellite signals fail, the trained model is used to predict and compensate for inertial device errors, thereby correcting inertial device errors.

[0043] The difference between the present invention and prior art 2 Error correction scope and focus: Prior Art 2 focuses on correcting errors in inertial devices, specifically the measurement errors of accelerometers and gyroscopes. These errors are caused by factors such as the device's inherent characteristics and changes in the carrier's motion state, such as errors caused by external interference such as carrier vibration.

[0044] This invention not only focuses on inertial device errors but also corrects nonlinear errors in the integrated navigation system. After separating modelable system errors through Kalman filtering, the invention uses LSTM to model the nonlinear errors in the residual sequence. This nonlinear error may arise from the influence of various factors on the entire integrated navigation system, not just the inertial device itself, such as complex terrain and dynamic environment.

[0045] LSTM application and error prediction method Existing Technology 2: This model uses LSTM to build an inertial device error prediction model. The model takes as input information related to the carrier's motion state, such as the position, velocity, and attitude change of the inertial navigation system (INS), as well as actual inertial device measurements. It then predicts the accelerometer and gyroscope errors along three axes. Each axis's error prediction is implemented using a separate model (six models in total). These models are trained to predict inertial device errors, thereby compensating for the IMU's actual measurements.

[0046] The present invention employs a two-layer LSTM network, whose input is the residual sequence after Kalman filtering. The LSTM network learns the nonlinear error pattern in the residual, taking the residual as input and outputting a nonlinear error compensation value (residual correction value). This correction value is then fused with the state estimate of the Kalman filter and fed back into the Kalman filter update phase to achieve coordinated optimization of error compensation and state estimation. This approach utilizes the LSTM as a nonlinear modeling tool for the Kalman filter residual sequence, forming a closed-loop fusion optimization mechanism with the Kalman filter. This approach focuses on optimizing the state estimate of the entire navigation system, rather than simply predicting and compensating for inertial device measurement errors.

[0047] (2) Differences in technical routes and specific implementation methods The present invention: Closed-loop fusion architecture: A bidirectional interactive closed-loop fusion mechanism is constructed, feeding the error compensation values ​​output by the LSTM into the Kalman filter's state update process in real time. A covariance reset strategy is used to eliminate historical error interference. Specifically, the Kalman filter is used to estimate the IMU's gyroscope and accelerometer biases in real time, separating modelable system errors. An adaptive noise covariance adjustment strategy is used to improve the robustness of the filtering algorithm. An LSTM network is used to model nonlinear errors in the residuals, effectively compensating for IMU nonlinear errors. The corrected LSTM state is fed back into the Kalman filter's update process, forming a closed-loop optimization.

[0048] A hierarchical decoupling method for error sources: The Kalman filter prioritizes linear noise and modelable errors, while the LSTM performs time series modeling for nonlinear residuals. In its implementation, the Kalman filter estimates and separates system errors and noise in real time, while the LSTM captures nonlinear error patterns. The corrected state is then fed back to the Kalman filter's update phase.

[0049] Prior art 1: Model fusion: The outputs of the Kalman filter and LSTM network are weighted and redistributed through an attention mechanism to achieve accurate IMU attitude estimation. Specifically, the Kalman filter and LSTM network are used to predict the IMU attitude separately. The attention mechanism then calculates the fusion weights of the two methods to obtain the final attitude estimate.

[0050] Error handling method: The main focus is on how to fuse the outputs of the two methods through the attention mechanism to overcome the problem that traditional filtering has weak adaptability to disturbance interference. It does not explicitly involve the coordinated optimization of error compensation and state estimation through a closed-loop feedback mechanism.

[0051] Prior art 2: Model fusion method: Build an LSTM-based inertial device error prediction model and combine it with Kalman filtering for error correction. Specifically, Kalman filtering technology is used to correct inertial navigation errors, fusing the carrier's position and velocity information with information measured by the IMU to obtain the carrier's motion state information. An LSTM-based inertial device error prediction model is built based on inertial device errors. This model is trained when satellite signals are valid, and the trained model is used to predict and compensate for inertial device errors in the event of satellite failure.

[0052] Error handling method: focuses on using LSTM to model and predict inertial device errors to compensate for errors, involving the coordinated optimization of error compensation and state estimation through a closed-loop feedback mechanism.

[0053] (3) Technical effects The present invention: Improved stability in complex scenarios: Through the closed-loop fusion architecture and error source hierarchical decoupling method, the nonlinear drift problem of IMU in complex mountainous scenarios is effectively solved, and the stability and robustness of the integrated navigation system in complex scenarios are improved.

[0054] Improved error compensation accuracy: Through the synergy of Kalman filtering and LSTM, comprehensive processing of linear noise, modelable errors, and nonlinear residuals is achieved, effectively improving the accuracy of error compensation.

[0055] Reduced computational complexity: The KF algorithm separates the modelable system errors and only performs adaptive adjustments on the remaining noise, which reduces computational complexity.

[0056] Prior art 1: Improved attitude estimation accuracy: By introducing a neural network, the impact of external disturbances on attitude prediction accuracy is reduced to a certain extent, and the prediction deviation of the traditional Kalman filter in the nonlinear system is corrected, thereby improving the accuracy of attitude estimation.

[0057] Enhanced robustness: The robustness of pose estimation is improved by using the attention mechanism to fuse the outputs of the Kalman filter and the LSTM network. However, this work mainly focuses on the pose estimation field and does not explicitly address stability improvements in complex mountainous scenarios.

[0058] Prior art 2: Maintaining inertial navigation performance: By combining Kalman filtering with LSTM, effective prediction and compensation of inertial device errors are achieved, solving the problem of error divergence in the inertial navigation system and improving the accuracy and reliability of the inertial navigation system's positioning.

[0059] Maintaining performance during satellite failure: When satellite signals fail, the trained LSTM model is used to predict and compensate for inertial devices, ensuring the continuity and accuracy of the navigation system. However, the main focus is on maintaining inertial navigation performance, and the coordinated optimization of error compensation and state estimation through a closed-loop feedback mechanism is not explicitly addressed.

[0060] In one embodiment, S1 may be implemented as follows: S1.1: Using three-axis angular velocity data and three-axis acceleration data as input, the state prediction is performed through the prediction equation to separate the gyroscope bias and accelerometer bias errors. S1.2: Generate state estimates and residual sequences by adjusting the covariance matrix update.

[0061] Specifically, the Kalman filter is a linear estimation algorithm used to solve the linear problem of Gaussian white noise. By fusing the system model and the observation model, it obtains the past state estimate and the latest measurement value based on the minimum mean. The square root error criterion is used to weight the estimate to obtain the best estimate. The present invention uses the Kalman filter to filter the IMU raw data as input to obtain the state estimate and residual. The filtering process of the Kalman filter is divided into two steps: prediction and update. Its state vector is: (1) in, Indicates attitude angle; Indicates the gyroscope zero bias; Indicates the accelerometer bias.

[0062] The Kalman filter prediction equation is as follows: (2) (3) in, Indicates time A priori state estimate of ; is the state transition matrix; Indicates time The posterior state estimate at ; is the control input matrix, the control input Mapping into state space; Indicates time control input. It's time The a priori estimated covariance of ; is the time step The posterior estimated covariance of ; is the process noise covariance matrix. The first equation in this formula represents the previous state estimate, while the second equation calculates the covariance of the previous estimate.

[0063] The state update equation of the Kalman filter is as follows: (4) (5) (6) in, It's time Kalman gain; is the observation matrix; is the observation noise covariance matrix. It's time The posterior state estimate of ; It's time The measurement vector of . It's time The posterior estimated covariance of ; is the identity matrix. The first equation in the formula is the update of the Kalman gain, the second equation is the posterior state estimate, and the third equation is the final covariance calculation.

[0064] The original IMU data is processed by Kalman filtering to obtain the state estimate and residual as output. The residual equation is: (7) in, is the data residual; is the state estimate. The Kalman filter process is as follows Figure 2 shown.

[0065] In one embodiment, S2 may be implemented as follows: S2.1: Build a two-layer LSTM network; S2.2: Input the historical data set into the LSTM network model, use LSTM to learn the nonlinear error pattern related to the terrain, and use the mean square error loss function to optimize the model parameters to obtain a trained LSTM network model: S2.3: The residual is input into the trained LSTM network model. The input layer receives the residual sequence, and the output layer generates a nonlinear error compensation as the residual correction amount.

[0066] Specifically, the historical dataset covers scenes such as urban roads and tunnels.

[0067] LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN) that introduces a gating mechanism to solve the gradient vanishing or gradient exploding problems of standard RNN when processing long sequence data. The core structure of LSTM is the memory cell, which controls the flow of information through the forget gate, input gate, and output gate to effectively manage long-term dependencies. In time series correction, LSTM receives the residual sequence output by the Kalman filter. As input, the forecast residual correction LSTM uses historical residual information to capture potential temporal patterns and improve the accuracy of the Kalman filter.

[0068] The structure of LSTM is as follows Figure 3 The horizontal line Ct running through the cells in the figure represents the cell state of the LSTM. ■ represents the neural network layer, which is similar to the activation function (Neural Network Layer); ● represents the pointwise operation, such as the corresponding point multiplication, addition, and subtraction operations (Pointwise Operation Layer); Indicates data flow direction (Vector Transfer); Indicates vector concatenation; Indicates Vector Copy.

[0069] The gyroscope data information stored and involved in the calculation in the memory cells of the LSTM neural network can be gated to add key information and delete useless information. The operation of the gate switch can be achieved through the Sigmoid function. Among them, Sigmoid is a monotonic bounded function, and all values ​​are between 0 and 1. 0 means that the gate is closed and the gyroscope data information is not allowed to pass through. 1 means that the gate is open and all input gyroscope data information is allowed to pass through. As shown in the figure, the architecture of the LSTM neural network model predicting gyroscope data contains three gates to control the gyroscope data information involved in model learning. The LSTM neural network structure designed in this embodiment mainly includes the following five steps: Step 1: Data normalization; Step 2: Use the forget gate in the LSTM structure, also known as the forget gate, to discard useless residual data information through the Sigmoid function. According to the state vector and residual data information in the hidden layer, through the operation of the gate, a vector between 0 and 1 is output. The operation of the forget gate is calculated according to the following formula: (8) in, Is the output of the forget gate, ranging from 0 to 1; is the Sigmoid activation function; is the weight matrix of the forget gate; is the hidden state at the previous moment; is the residual input at the current moment; is the bias term of the forget gate.

[0070] Step 3: Determine what information the cell adds through the input gate. This step is divided into two steps. First, use the hidden layer state vector and , which information to update is determined by the operation of the output gate, and the update parameters are calculated according to the following formula: (9) Secondly, using the hidden state vector in the model and residual input information Through the operation of the tanh layer, the candidate memory unit state is obtained , Calculate according to the following formula: (10) in, is the output of the input gate, ranging from 0 to 1; is the weight matrix of the input gate; is the bias term of the input gate. is the candidate memory cell state; is the hyperbolic tangent activation function; is the weight matrix of the candidate memory cell state; is the bias term for the candidate memory cell state.

[0071] Step 4: Old cell information contained in the cell state Perform the update operation and regard the updated cell information as the new cell state information The specific update rule is to control the candidate cell state according to the forget gate The information stored in the input gate is used to add new key information to the new candidate cell information. middle.

[0072] (11) Step 5: After updating the cell state, the hidden layer state vector needs to be output and input information Determine the output information.

[0073] (12) (13) (14) in, Is the output of the output gate, and its value range is between 0 and 1; is the weight matrix of the output gate; is the bias term of the output gate; The hidden state at the current moment; ; is the residual correction; is the weight matrix of the fully connected layer; is the bias term of the fully connected layer.

[0074] After obtaining the trained LSTM model, the residual error can be input into the model to obtain the residual error correction.

[0075] In one embodiment, S3 may be implemented as follows: The state estimate and the residual correction are used as inputs in the Kalman filter update phase to correct the state and obtain the corrected IMU state. Adopting adaptive noise covariance adjustment strategy to reset the covariance matrix; The corrected state and the reset covariance matrix are used as the initial values ​​of the next filtering moment to form a closed-loop optimization and obtain the corrected IMU data. In the specific implementation process, the residual correction value obtained by the LSTM neural network and the state estimation value obtained by the Kalman filter can be used to obtain the corrected IMU state: (15) in, is the corrected IMU state; is the state estimate; is the residual correction. The corrected IMU state is fed back to the update phase of the Kalman filter and the covariance matrix is ​​reset: (16) in, is the reset covariance matrix; is the reset coefficient. The corrected state and covariance are used as the initial value of the next filtering moment, forming a closed-loop optimization.

[0076] The method of the present invention is described below by means of specific examples, and the specific implementation steps are as follows: Step 1: Multi-source data collection and preprocessing Hardware configuration: (1) Use the vehicle-mounted IMU module to synchronously collect the three-axis angular velocity at a sampling frequency of 200Hz and acceleration ( ).

[0077] (2) Use GNSS receiver to obtain PPP positioning data, including latitude and longitude , elevation and northeastern sky speed .

[0078] Step 2: Dynamic State Estimation (Kalman Filter) State vector definition: Construct a 9-dimensional error state vector, including attitude angles (heading angle, pitch angle, roll angle), gyroscope bias, and accelerometer bias.

[0079] Kalman filter process: Prediction stage: Based on the IMU dynamic model, the state transition matrix is ​​used to predict the prior state estimate and covariance matrix at the current moment.

[0080] Update phase: Calculate the Kalman gain and update the posterior state estimate and covariance.

[0081] Step 3: Timing Error Modeling (LSTM Network) Residual sequence construction: Extract the residual sequence after Kalman filter update (the difference between the observation value and the state estimate) and construct the input data according to the 1 second time window (100 sampling points).

[0082] Network structure and training: Build a two-layer LSTM network with 128 hidden units in each layer and the input layer receiving the residual sequence , the output layer generates nonlinear error compensation ( ).

[0083] The network training uses the mean square error loss function, and the parameters are optimized through historical data sets (covering scenarios such as urban roads and tunnels). The learning rate is set to 0.001 and the batch size is 64.

[0084] Step 4: Closed-loop fusion optimization and output The error compensation output by LSTM is fused with the Kalman filter state estimate to obtain the corrected IMU data. The corrected IMU state is fed back to the update stage of the Kalman filter, and the covariance matrix is ​​reset. The corrected state and covariance are used as the initial values ​​of the filter at the next moment to form a closed-loop optimization.

[0085] Step 2 focuses on preliminary dynamic state estimation and separation of modelable system errors. Taking IMU data (including triaxial angular velocity and triaxial acceleration) as input, it uses the Kalman filter algorithm to perform state prediction and separate modelable system errors such as gyroscope bias and accelerometer bias. By adjusting the covariance matrix, it ultimately outputs the initial state estimate and residual sequence. This process lays the foundation for the entire error processing process and provides essential information for subsequent in-depth error analysis and compensation.

[0086] Step 4, based on step 2, introduces the residual correction amount obtained by time series modeling of the residual sequence using the LSTM network, thereby achieving collaborative optimization of error compensation and state estimation. Specifically, step 4 combines the state estimate output by step 2 with the residual correction amount obtained by the LSTM network in step 3, and feeds it back to the update link of the Kalman filter. By adopting an adaptive noise covariance adjustment strategy to reset the covariance matrix, and using the corrected state and the reset covariance matrix as the initial values ​​of the filter at the next moment, a closed-loop optimization mechanism is formed. This process makes full use of the LSTM network's ability to learn nonlinear errors, further corrects the state estimate, effectively improves the adaptability and compensation effect of the Kalman filter for nonlinear errors in complex scenarios, and enhances the accuracy and reliability of the navigation system in dynamic environments.

[0087] The relationship between the two: The Kalman filter update phase in step 2 is the starting point and foundation of the entire error processing process. Its main task is to perform initial state estimation based on IMU data and isolate modelable system errors, providing raw estimates and residual information for subsequent processing. The Kalman filter update phase in step 4 is a deepening and improvement of step 2. It integrates the LSTM network's modeling results of nonlinear errors in the residual sequence and feeds the corrected information back to the Kalman filter to achieve coordinated optimization of error compensation and state estimation. The two do not exist in isolation, but are closely connected through the Kalman filter framework. Step 4 can be seen as an extension and enhancement of step 2. The two work together to enable the entire navigation system to achieve more accurate and reliable state estimation and error compensation in complex environments, effectively improving the performance of the navigation system.

[0088] In specific applications, the corrected IMU data and PPP positioning results are combined for IMU-GNSS loose combined navigation, and the final output is combined navigation. The result includes the optimized three-dimensional position (longitude, latitude, elevation), speed and attitude angle. The data update rate is 200Hz, which meets the real-time requirements of high-dynamic scenes.

[0089] Example 2 Based on the same inventive concept, this embodiment discloses a combined navigation method based on Kalman filtering and LSTM cascade, which performs IMU-GNSS loose combined navigation on the corrected IMU data obtained in Example 1 and the positioning result of PPP to obtain a combined navigation result.

[0090] Example 3 Based on the same inventive concept, this embodiment discloses a fusion optimization device based on Kalman filtering and LSTM cascade, comprising: The Kalman filter dynamic state estimation module is used to use IMU data as input, perform dynamic state estimation using Kalman filtering, separate modelable system errors, and output state estimation values ​​and residual sequences, where the residual sequence is a sequence consisting of the difference between the observed values ​​and the state estimation values; The LSTM time series error modeling module is used to obtain the residual correction value based on the input residual using the pre-built LSTM network model; The fusion optimization module is used to feed back the state estimation value output by the Kalman filter and the residual correction obtained by the LSTM network to the Kalman filter update stage, perform collaborative optimization of error compensation and state estimation, and obtain the corrected IMU data.

[0091] Since the device introduced in the second embodiment of the present invention is the device used to implement the fusion optimization method based on Kalman filtering and LSTM cascade in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the device, so it is not described here. All devices used in the method in the first embodiment of the present invention belong to the scope of protection of the present invention. Example 4 Based on the same inventive concept, this embodiment discloses a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the fusion optimization method based on Kalman filtering and LSTM cascade described in Example 1 is implemented.

[0092] Since the computer device described in Example 4 of the present invention is the computer device used to implement the fusion optimization method based on the Kalman filter and LSTM cascade in Example 1 of the present invention, those skilled in the art will be able to understand the specific structure and variations of the computer device based on the method described in Example 1 of the present invention, and therefore will not be described in detail here. All computer devices used in the method of Example 1 of the present invention fall within the scope of protection of the present invention.

[0093] Example 5 Based on the same inventive concept, this embodiment discloses a combined navigation system, including the fusion optimization device based on Kalman filtering and LSTM cascade, the PPP solution module and the fusion module in Example 3, wherein the PPP solution module is used to receive the observation quantity and ephemeris information broadcast by the GNNS satellite, and perform PPP solution to output position information. The fusion module is used to correct the navigation parameters through Kalman filtering, and feed back the IMU data output by the fusion optimization device based on Kalman filtering and LSTM cascade as an error into the inertial navigation solution to obtain the positioning result.

[0094] Specifically, the PPP / INS loose combination algorithm architecture is as follows Figure 4The algorithm consists of three parts: the first part mechanically organizes the specific force and angular velocity information provided by the inertial measurement unit and outputs navigation information such as speed and position; the second part receives observations and ephemeris information broadcast by GNNS satellites and performs PPP solution to output position information; the third part performs Kalman filtering, corrects navigation parameters through Kalman filtering, and feeds back errors related to the inertial sensor into the inertial navigation solution.

[0095] The PPP / INS loose combination model uses a 21-dimensional error state vector for Kalman filtering. Expressed as: (17) in, yes Position error vector of the system; yes Velocity error vector under the system; is the attitude error vector; is the gyroscope process noise vector; is the accelerometer process noise vector; is the gyroscope measurement noise vector; is the accelerometer measurement noise vector.

[0096] After the state vector is updated, the current error state quantity is obtained for: (18) in, is the discrete state transfer matrix; for The error state vector at time t; is the noise vector after conversion.

[0097] 3.2 Observation equation Because IMU mechanical arrangement typically determines the IMU's geometric center, while PPP calculations typically calculate the receiver antenna phase center, the antenna arm distance between the two must be considered. The PPP / INS loosely coupled model pre-measures the antenna arm distance and compensates for the antenna phase center.

[0098] The error observation equation is expressed as: (19) in, is the measurement error; is the error state vector; is the measurement error noise; represents the observation matrix, which is expressed as: (20) in, yes The identity matrix of yes The zero matrix of ; It is from Tie to The coordinate transformation matrix of the system; is the antenna arm vector, that is, the vector from the IMU measurement center to the PPP antenna phase center. The projection of the system.

[0099] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0100] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0101] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, the present invention is intended to include such changes and modifications to the embodiments of the present invention if they fall within the scope of the claims and their equivalents.

Claims

1. A fusion optimization method based on Kalman filtering and LSTM cascade, characterized in that: include: Taking IMU data as input, Kalman filtering is used to perform dynamic state estimation and separate modelable system errors, outputting state estimates and residual sequences, where the residual sequence is a sequence consisting of the difference between the observed values ​​and the state estimates; The pre-built LSTM network model is used to obtain the residual correction value according to the input residual; The state estimation value output by the Kalman filter and the residual correction obtained by the LSTM network are fed back to the Kalman filter update stage to perform collaborative optimization of error compensation and state estimation to obtain the corrected IMU data.

2. The fusion optimization method based on Kalman filtering and LSTM cascade according to claim 1 is characterized in that: Taking IMU data as input, Kalman filtering is used to perform dynamic state estimation and separate modelable system errors, outputting state estimates and residual sequences, including: Using three-axis angular velocity data and three-axis acceleration data as input, the state prediction is performed through the prediction equation to separate the gyroscope bias and accelerometer bias errors; The state estimates and residual sequences are generated by adjusting the covariance matrix update.

3. The fusion optimization method based on Kalman filtering and LSTM cascade according to claim 1 is characterized in that: The pre-built LSTM network model is used to obtain the residual correction according to the input residual, including: Build a two-layer LSTM network; The historical data set is input into the LSTM network model. The nonlinear error pattern related to the terrain is learned by LSTM. The model parameters are optimized using the mean square error loss function to obtain the trained LSTM network model: The residual is input into the trained LSTM network model. The input layer receives the residual sequence, and the output layer generates a nonlinear error compensation as the residual correction amount.

4. The fusion optimization method based on Kalman filtering and LSTM cascade according to claim 1, characterized in that: The state estimation value output by the Kalman filter and the residual correction value obtained by the LSTM network are fed back to the Kalman filter update stage to perform collaborative optimization of error compensation and state estimation, including: The state estimate and the residual correction are used as inputs in the Kalman filter update phase to correct the state and obtain the corrected IMU state. Adopting adaptive noise covariance adjustment strategy to reset the covariance matrix; The corrected state and the reset covariance matrix are used as the initial values ​​of the next filtering moment to form a closed-loop optimization and obtain the corrected IMU data.

5. The fusion optimization method based on Kalman filtering and LSTM cascade according to claim 1 is characterized in that: The state estimate and residual correction are used as inputs in the Kalman filter update phase to correct the state and obtain the corrected IMU state, including: in, is the corrected IMU state, is the state estimate, is the residual correction amount.

6. The fusion optimization method based on Kalman filtering and LSTM cascade according to claim 1, characterized in that: The covariance matrix is ​​reset using an adaptive noise covariance adjustment strategy, including: in, is the reset covariance matrix, is the reset factor.

7. A combined navigation method based on Kalman filtering and LSTM cascade, characterized in that: The corrected IMU data obtained by any one of claims 1 to 6 is used with the positioning result of PPP to perform IMU-GNSS loose combined navigation to obtain a combined navigation result.

8. A fusion optimization device based on Kalman filtering and LSTM cascade, characterized in that: include: The Kalman filter dynamic state estimation module is used to take IMU data as input, use Kalman filtering to perform dynamic state estimation and separate modelable system errors, and output state estimation values ​​and residual sequences, where the residual sequence is a sequence consisting of the difference between the observed value and the state estimation value; The LSTM time series error modeling module is used to obtain the residual correction value based on the input residual using the pre-built LSTM network model; The fusion optimization module is used to feed back the state estimation value output by the Kalman filter and the residual correction obtained by the LSTM network to the Kalman filter update stage, perform collaborative optimization of error compensation and state estimation, and obtain the corrected IMU data.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the fusion optimization method based on Kalman filtering and LSTM cascade is implemented as described in any one of claims 1 to 6.

10. An integrated navigation system, characterized in that: It includes a fusion optimization device based on Kalman filtering and LSTM cascade, a PPP solution module and a fusion module as described in claim 8, wherein the PPP solution module is used to receive observation quantities and ephemeris information broadcast by GNNS satellites, perform PPP solution, and output position information. The fusion module is used to correct navigation parameters through Kalman filtering, and feed back the IMU data output by the fusion optimization device based on Kalman filtering and LSTM cascade as an error into the inertial navigation solution to obtain a positioning result.

Images