An Adaptive Solution Method and System for Inertial Navigation Based on Motion Feature Recognition

Abstract

This invention discloses an adaptive inertial navigation solution method and system based on motion feature recognition, comprising: extracting multidimensional features reflecting the motion state of the carrier based on angular velocity and specific force data collected by the inertial measurement unit, and constructing a feature matrix; performing dimensionality reduction processing on the feature matrix to obtain low-dimensional feature vectors; performing cluster analysis based on a clustering algorithm that initializes cluster centers using prior samples to identify the motion state of the carrier; and adaptively selecting an appropriate combined navigation filtering strategy for navigation solution. This invention effectively suppresses the accumulation of inertial navigation errors, improving the autonomous navigation accuracy and long-term stability of the system in scenarios of satellite signal rejection or visual failure; it can finely distinguish different motion modes such as hovering, uniform motion, and transitional states, avoiding the shortcomings of traditional threshold methods that are susceptible to noise interference and have poor adaptability; by introducing continuous frame mode voting and physical anti-jump mechanisms, it further eliminates single-frame misjudgments, ensuring the continuity and robustness of state recognition.

Description

Technical Field

[0001] This invention relates to the field of inertial navigation technology, specifically to an adaptive solution method and system for inertial navigation based on motion feature recognition. Background Technology

[0002] Inertial navigation systems (INS) are a fully autonomous navigation method that uses angular velocity and specific force data measured by gyroscopes and accelerometers to calculate the attitude, velocity, and position of a vehicle in real time through integration. They are unaffected by external electromagnetic interference and do not rely on external signals, giving them irreplaceable advantages in scenarios where GNSS signals are limited or denied, such as UAV flight and underground tunnel exploration. However, pure inertial navigation systems inherently suffer from error accumulation: sensor errors such as gyroscope drift and accelerometer bias accumulate over integration time, causing rapid divergence in velocity and position estimates. To address this issue, traditional methods typically employ integrated navigation techniques, such as loosely or tightly coupled GNSS / INS systems, utilizing external auxiliary information (satellites, vision, barometers, etc.) to periodically correct inertial navigation errors. However, the following technical problems exist in practical applications:

[0003] When the vehicle enters tunnels, dense forests, urban canyons, or is subjected to electromagnetic interference, satellite signals are lost, visual features are missing, and external auxiliary information becomes unavailable. Traditional integrated navigation degenerates into pure inertial calculation, and errors accumulate rapidly, making it difficult to meet the requirements of long-term high-precision navigation. Existing methods typically use fixed Kalman filter models, failing to fully consider the impact of different motion states of the vehicle (such as hovering, uniform motion, and maneuvering turns) on error propagation characteristics. Some schemes attempt to simply classify motion states based on velocity or angular velocity amplitude, but the thresholds are difficult to adjust adaptively and are susceptible to noise interference, leading to misjudgment of states and affecting the accuracy of filter correction. The parameters of traditional methods (such as noise variance and initial deviation) largely rely on offline calibration and cannot be dynamically adjusted according to the actual motion environment of the vehicle and changes in sensor characteristics, resulting in decreased accuracy after long-term operation. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive solution method and system for inertial navigation based on motion feature recognition, so as to solve the problems mentioned in the background art, such as failure correction under denied conditions, difficulty of adapting a single filtering strategy to multiple motion states, reliance on human experience or simple thresholds for state judgment, and lack of online self-optimization capability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] According to one aspect of the present invention, an adaptive solution method for inertial navigation based on motion feature recognition is provided, the method comprising:

[0007] Based on the angular velocity and specific force data collected by the inertial measurement unit, multidimensional features reflecting the motion state of the carrier are extracted, and a feature matrix is ​​constructed.

[0008] The feature matrix is ​​reduced in dimensionality to obtain a low-dimensional feature vector;

[0009] A clustering algorithm based on prior samples initializes cluster centers, performs cluster analysis on the low-dimensional feature vector, and identifies the motion state of the carrier.

[0010] Based on the identified motion state, the appropriate combined navigation filtering strategy is adaptively selected for navigation calculation, with different filtering correction methods corresponding to different motion states.

[0011] Based on the aforementioned scheme, the extraction of multidimensional features reflecting the motion state of the carrier includes:

[0012] The raw angular velocity and specific force data are preprocessed. The preprocessing includes removing outliers from the static initialization data using the 3σ criterion, and calculating the gyroscope zero bias and accelerometer zero bias using the static data after removing outliers, and performing zero bias compensation on the raw data.

[0013] The preprocessed data is divided into data frames by using a sliding time window. Temporal features, statistical features, and correlation features are extracted within each window to form the feature matrix.

[0014] Based on the aforementioned scheme, the time-domain features include the mean of the three-axis angular velocity and the mean of the three-axis specific force within the window; the statistical features include the variance of the three-axis angular velocity, the variance of the three-axis specific force, and the energy of the gyroscope data and the energy of the accelerometer data within the window; the correlation features include the Pearson correlation coefficient between the corresponding axis data of the gyroscope and the accelerometer within the window.

[0015] Based on the aforementioned scheme, the dimensionality reduction processing of the feature matrix includes:

[0016] The feature matrix is ​​standardized, the covariance matrix of the standardized feature matrix is ​​calculated, and the covariance matrix is ​​decomposed to obtain eigenvalues ​​and eigenvectors.

[0017] The number of principal components to be retained is determined based on the cumulative variance contribution rate, and the eigenvectors corresponding to the first k eigenvalues ​​are selected to form the projection matrix.

[0018] The standardized feature matrix is ​​projected onto the k-dimensional space spanned by the projection matrix to obtain the dimensionality-reduced feature matrix.

[0019] Based on the aforementioned scheme, the clustering algorithm for initializing cluster centers based on prior samples includes:

[0020] During the effective period of integrated navigation, sample data corresponding to known motion states are collected, and the mean of the dimensionality-reduced feature vector of each motion state sample is calculated as the initial cluster center of the corresponding motion state.

[0021] Starting from the initial cluster center, iterative clustering is performed on the dimensionality-reduced feature matrix until the cluster center converges or the maximum number of iterations is reached, and the initial cluster label of each frame of data is output.

[0022] Based on the aforementioned scheme, the iterative clustering includes:

[0023] Calculate the Euclidean distance from each feature vector to each of the initial cluster centers, and assign them to the corresponding clusters according to the principle of minimum distance; update the cluster centers to the mean of all feature vectors in the current cluster;

[0024] Repeat the clustering steps until the change in cluster centers is less than a preset threshold.

[0025] Based on the aforementioned scheme, after outputting the initial clustering labels, a post-processing step is also included:

[0026] A continuous frame majority voting mechanism is adopted, and the category that appears most frequently in the initial clustering labels of N consecutive frames centered on the current frame is taken as the final state label of the current frame.

[0027] An anti-jump mechanism is set up based on the physical characteristics of motion. When there is a direct jump between a hovering state and a uniform motion state in adjacent frames, the label of the next frame is forcibly corrected to the hovering to motion state.

[0028] Based on the aforementioned scheme, the adaptive selection of the integrated navigation filtering strategy includes:

[0029] When the external auxiliary information is valid, the position provided by the external auxiliary information is used as the measurement for Kalman filtering correction.

[0030] When the external auxiliary information is invalid, different measurement methods are selected based on the identified motion state:

[0031] If the state is identified as hovering, a zero-speed correction strategy is adopted, using the velocity calculated by the inertial navigation system as the measurement; if the state is identified as uniform motion, a velocity correction strategy is adopted, using the difference between the current velocity of the inertial navigation system and the average velocity at the initial moment of uniform motion as the measurement; if the state is identified as hovering to motion, no filtering correction is performed, and only pure inertial calculation is performed.

[0032] Based on the aforementioned scheme, the method further includes using the combined navigation output when the external auxiliary information is effective to determine the real motion state of the current frame in real time, and combining the real motion state with the corresponding dimensionality-reduced feature vector to form a new sample, and dynamically updating the initial cluster center.

[0033] According to another aspect of the present invention, an adaptive inertial navigation solution system based on motion feature recognition is provided, the system comprising an inertial measurement unit, a feature extraction module, a dimensionality reduction module, a state recognition module, and an adaptive navigation solution module;

[0034] The inertial measurement unit is used to collect the angular velocity and specific force data of the carrier;

[0035] The feature extraction module is used to extract multi-dimensional features reflecting the motion state of the carrier based on the angular velocity and specific force data, and construct a feature matrix.

[0036] The dimensionality reduction module is used to perform dimensionality reduction processing on the feature matrix to obtain a low-dimensional feature vector;

[0037] The state recognition module is used to perform cluster analysis on the low-dimensional feature vector based on the clustering algorithm that initializes cluster centers based on prior samples, and to identify the motion state of the carrier.

[0038] The adaptive navigation solution module is used to adaptively select the corresponding combined navigation filtering strategy for navigation solution based on the identified motion state, with different filtering correction methods corresponding to different motion states.

[0039] As can be seen from the above technical solution, the present invention has at least the following advantages and positive effects compared with the prior art: Through multi-dimensional feature extraction and cluster analysis, the motion state of the carrier can be accurately identified without external auxiliary information, and the filtering correction strategy (zero speed correction during hovering, speed correction during uniform motion, and pure inertial calculation during transition) can be adaptively selected according to the state, effectively suppressing the accumulation of inertial navigation errors and significantly improving the autonomous navigation accuracy and long-term stability of the system in scenarios of satellite signal rejection or visual failure. By constructing a multi-dimensional feature matrix in the time domain, statistical domain, and correlation domain, and combining the K-means clustering algorithm with principal component analysis dimensionality reduction and prior sample initialization, the present invention can finely distinguish different motion modes such as hovering, uniform motion, and transition states, avoiding the defects of traditional threshold methods that are easily affected by noise interference and have poor adaptability; after introducing continuous frame mode voting and physical anti-jump mechanism, single-frame misjudgment is further eliminated, ensuring the continuity and robustness of state recognition. When external auxiliary information (satellite, vision, etc.) is available, this invention utilizes the high-precision output of integrated navigation to determine the true motion state in real time, and combines the state label with the corresponding feature vector to form a new sample, dynamically updating the initial cluster centers. This allows the state recognition model to adaptively optimize itself according to environmental changes and sensor characteristic drift, maintaining long-term operational accuracy. Differentiated filtering strategies are adopted for different motion states: zero-velocity correction is used to suppress divergence in static scenarios such as hovering; velocity correction is used to control error drift in uniform motion scenarios; and short-term accuracy of inertial navigation is trusted and no correction is made in transitional dynamic scenarios. This ensures the system's dynamic response capability while minimizing steady-state error.

[0040] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0042] Figure 1 This is a schematic diagram of the adaptive solution method for inertial navigation based on motion feature recognition according to the present invention;

[0043] Figure 2 This is a schematic diagram of the data preprocessing and feature extraction process of the present invention;

[0044] Figure 3 This is a schematic diagram of the feature dimensionality reduction process of the present invention;

[0045] Figure 4This is a schematic diagram of the adaptive navigation settlement process of the present invention. Detailed Implementation

[0046] To more clearly illustrate the purpose, technical solutions, and advantages of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein. On the contrary, these embodiments are provided so that the present invention will be more comprehensive and complete, and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0047] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.

[0048] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0049] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0050] The present invention will now be described in detail with reference to specific embodiments.

[0051] Example 1

[0052] like Figure 1 As shown in the figure, this embodiment provides an adaptive solution method for inertial navigation based on motion feature recognition. The specific steps of this method are as follows:

[0053] S1: Based on the angular velocity and specific force data collected by the inertial measurement unit, multidimensional features reflecting the motion state of the carrier are extracted and a feature matrix is ​​constructed.

[0054] like Figure 2As shown, the collected raw angular velocity and specific force data are preprocessed. The preprocessing includes using the 3σ criterion to remove outliers in the static initialization data, and using the static data after removing outliers to calculate the gyroscope zero bias and accelerometer zero bias to compensate for the zero bias of the raw data. The preprocessed data is divided into data frames by a sliding time window, and time-domain features, statistical features and correlation features are extracted in each window to form the feature matrix.

[0055] Specifically, during actual aircraft operation, due to instantaneous impacts, electromagnetic interference, or transient sensor malfunctions, the raw data output by gyroscopes and accelerometers may contain outliers with amplitudes far exceeding the normal range, i.e., outliers. To ensure the accuracy of subsequent zero-bias estimation, the static data used for calibration must first be cleaned. In this embodiment, the 3σ criterion (Laida criterion) is preferably used to remove outliers from a segment of static initialization data of length L collected after the aircraft is powered on; this criterion is based on the statistical characteristic that random noise from sensors in a static state approximately follows a normal distribution. Specifically, the raw gyroscope data in the static segment is... accelerometer data is (i=x, y, z represent the three axes of the carrier coordinate system; t=1,2,...,L are the sampling time indices); calculate the mean of the data for each axis. , and standard deviation , :

[0056] ;

[0057] ;

[0058] Iterate through each data point in the static sequence. If the value of a certain point satisfies... or If a point is identified as an outlier, it is removed and replaced by interpolation using valid data points before and after it, resulting in a clean static data sequence. , It effectively distinguishes between normal noise fluctuations caused by environmental vibrations and sudden abnormal interference, ensuring the reliability of subsequent data processing.

[0059] Both gyroscopes and accelerometers have inherent zero-bias, meaning their output values ​​are non-zero when there is no input. This zero-bias leads to constant or slowly changing errors in angular velocity and specific force measurements, directly affecting the accuracy of inertial calculations. To eliminate this systematic bias, this embodiment uses the static data sequence after outlier removal for zero-bias estimation. The static data sequence is calculated separately. , The arithmetic mean across the three axes is used as the gyroscope zero bias for the current cycle. and accelerometer zero bias :

[0060] ;

[0061] ;

[0062] All raw data collected during the subsequent dynamic flight of the aircraft ( , In this embodiment, the dynamic data segment will not undergo additional outlier removal, or robust methods such as median filtering can be used to further suppress impulse noise, depending on the actual situation. Subtracting the corresponding zero bias, the zero bias compensated inertial data is obtained:

[0063] ;

[0064] ;

[0065] In the formula, and This is the effective gyro angular velocity and acceleration ratio data that can be used for subsequent feature extraction after preprocessing (static segment outlier removal and full-time zero bias compensation).

[0066] Furthermore, the preprocessed inertial data is analyzed from three dimensions: time domain, statistical domain, and correlation domain. and Feature extraction is performed to comprehensively characterize the motion characteristics of the carrier within a short time window. A continuous data stream is divided into a series of data frames using a sliding time window, with each window having a length of N (N=100 sampling points in this embodiment). The windows can slide at fixed steps, and some data overlap is allowed between adjacent windows to ensure the continuity of state recognition. Within each window, temporal features, statistical features, and correlation features are calculated respectively.

[0067] Specifically, time-domain characteristics reflect the overall level and average trend of motion; these include the mean angular velocity and the mean specific force. Mean angular velocity The average value of the three-axis angular velocities within the calculation window is used to characterize the average intensity of the carrier's rotational motion. The calculation formula is:

[0068] ;

[0069] in, ω is the angular velocity of the i-th axis after preprocessing.

[0070] Specific force mean The average value of the triaxial specific force within the calculation window represents the average level of linear acceleration (including the gravitational component). The calculation formula is:

[0071] ;

[0072] in, This represents the preprocessed specific force along the i-th axis.

[0073] Statistical characteristics quantify the intensity of data fluctuations within a window, reflecting the stability and dynamic changes of motion, including the variance of three-axis angular velocity, the variance of three-axis specific force, and inertial data energy within the window. Angular velocity variance The variance measures the fluctuation of angular velocity around its mean. A small variance indicates smooth rotation, while a large variance indicates the presence of maneuvering or vibration. The formula is:

[0074] ;

[0075] Comparison variance The specific force measures the fluctuation around its mean value and is related to the changes in the linear acceleration and vibration of the carrier. The calculation formula is:

[0076] ;

[0077] Inertial data energy, including gyroscope data energy and accelerometer data energy, is calculated by summing the squares of the angular velocities and specific forces of all axes within a calculation window. This serves as a comprehensive measure of the overall dynamic intensity of the motion, and the calculation formula is as follows:

[0078] ;

[0079] For gyroscope data energy, This is the energy of the accelerometer data; this characteristic is sensitive to the intensity of motion, approaching zero when stationary, and increasing significantly during violent maneuvers.

[0080] The correlation characteristic represents the cooperative relationship between the outputs of the gyroscope and the accelerometer, which exhibits specific patterns under different motion modes. The Pearson correlation coefficient between the data sequences of the gyroscope and the accelerometer along the x, y, and z axes is calculated. :

[0081] ;

[0082] in, , indicating the degree of linear correlation between the angular velocity and the specific force along the i-th axis; when When the value approaches 1, it indicates a strong linear correlation between rotational motion and linear acceleration along that axis, typically corresponding to specific maneuvers (such as the correlation between centripetal acceleration and angular velocity during turning); when... When the value approaches -1, it indicates a strong negative linear correlation between the two; when When the value is close to 0, it indicates that the motion relationship between the two along that axis is weak, which is commonly seen in uniform linear motion or hovering.

[0083] Through the extraction of the above three types of features (time domain, statistical domain, and correlation domain), each sliding window obtains a 17-dimensional feature vector, including: the mean of the three-axis angular velocities. , , Mean of triaxial specific force , , Three-axis angular velocity variance , , Three-axis specific force variance , , ; Gyroscope data energy Accelerometer data energy triaxial correlation coefficient , , .

[0084] Arranging the feature vectors of M consecutive windows in chronological order forms the feature matrix for subsequent analysis. each of the lines , which corresponds to the 17-dimensional feature vector of the m-th frame of data.

[0085] S2: Perform dimensionality reduction processing on the feature matrix to obtain low-dimensional feature vectors.

[0086] The multidimensional feature matrix constructed in step S1 is subjected to dimensionality reduction. Principal Component Analysis (PCA) is used to remove the linear correlation between the original features and extract the core components that best represent the variability of the data. This reduces the feature dimensionality while retaining the main information, simplifying the complexity of subsequent clustering calculations. Figure 3 As shown, the feature dimensionality reduction process includes standardizing the feature matrix, calculating the covariance matrix of the standardized feature matrix, and performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors; determining the number of principal components to retain based on the cumulative variance contribution rate, selecting the eigenvectors corresponding to the first k eigenvalues ​​to form a projection matrix; and projecting the standardized feature matrix onto the k-dimensional space spanned by the projection matrix to obtain the dimensionality-reduced feature matrix, where k is less than the original feature dimension.

[0087] Specifically, before performing principal component analysis, the feature matrix is ​​first standardized so that the mean of each dimension is 0 and the standard deviation is 1, in order to eliminate the influence of dimensions. Let the multidimensional feature matrix constructed in step S1 be... , where M is the number of samples in the sliding window (i.e., the number of data frames), and D is the original feature dimension (D=17 in this embodiment). Matrix each line This corresponds to the D-dimensional feature vector of the m-th frame of data. For each feature d (d=1, 2, ..., D), calculate its mean. and standard deviation :

[0088] ;

[0089] ;

[0090] in, Let d be the sample mean of the d-th feature. Let be the sample standard deviation of the d-th feature. Let be the original feature value of the d-th dimension of the m-th sample. Then the standardized feature matrix is... The elements are given by the following formula:

[0091] ;

[0092] These are the standardized feature values. After standardization, the mean of each feature dimension is 0, and the standard deviation is 1.

[0093] Furthermore, the correlation structure between features is analyzed to determine the main direction of data variation.

[0094] Calculate the covariance matrix of the standardized feature matrix The elements of the covariance matrix represent the covariance between different features. The formula for calculating the covariance matrix is:

[0095] ;

[0096] in Indicates the first Dimensional features and the first The covariance of the features. Since the features have been standardized, the covariance matrix is ​​actually equivalent to the correlation coefficient matrix, with all diagonal elements being 1, and the off-diagonal elements reflecting the degree of linear correlation between different features.

[0097] For covariance matrix Perform eigenvalue decomposition (i.e., solve for eigenvalues ​​and eigenvectors) to obtain eigenvalues. and its corresponding eigenvectors ; Eigenvalues ​​satisfy Each eigenvalue measures the magnitude of the variance in the direction of the corresponding eigenvector, which is the amount of information carried by the principal component; the eigenvector defines the projection direction from the original feature space to the principal component space.

[0098] The number of principal components to retain is selected based on the cumulative variance contribution rate, aiming to preserve as much original information as possible while reducing dimensionality. The variance contribution rate of the i-th eigenvalue. Defined as the proportion of this eigenvalue to the sum of all eigenvalues:

[0099] ;

[0100] For the i-th eigenvalue, The variance contribution rate of the i-th principal component. The cumulative variance contribution rate of the first k principal components. for:

[0101] ;

[0102] This embodiment sets a threshold for the cumulative variance contribution rate. That is, select the one that satisfies The minimum value of k is used as the number of principal components to be retained. This criterion allows for maximum compression of the feature dimension while ensuring that information loss does not exceed 10%. In practical applications, k is usually much smaller than the original dimension D (k=3 in this example).

[0103] The projection matrix is ​​constructed by selecting the eigenvectors corresponding to the first k eigenvalues. Standardize the feature matrix Projected onto by In Zhang Cheng's k-dimensional principal component space, the dimensionality-reduced feature matrix is ​​obtained. :

[0104] ;

[0105] in, each line This is the k-dimensional principal component feature vector of the m-th frame data. This vector integrates the core information of the original multidimensional features, and the principal components are uncorrelated with each other, which can more effectively represent the intrinsic differences between different motion states and provide high-quality input for subsequent cluster-based state recognition.

[0106] S3: A clustering algorithm based on prior samples initializes cluster centers, performs cluster analysis on the low-dimensional feature vector, and identifies the motion state of the carrier.

[0107] The reduced-dimensional feature vector obtained in step S2 is used as input, and the flight state of the quadcopter UAV is automatically identified using the K-means clustering algorithm. To improve the accuracy and stability of clustering, a supervised initialization method based on prior samples, an iterative clustering process, and a post-processing mechanism are designed. Finally, the motion state label corresponding to each data frame is output, providing a decision-making basis for subsequent adaptive navigation calculations. The clustering algorithm based on prior samples to initialize cluster centers includes collecting sample data corresponding to known motion states within the effective time period of integrated navigation, calculating the mean of the reduced-dimensional feature vector of each state sample as the initial cluster center for the corresponding state; starting from the initial cluster center, iterative clustering of the reduced-dimensional feature matrix is ​​performed until the cluster center converges or the maximum number of iterations is reached, and the initial cluster label for each frame of data is output.

[0108] Specifically, based on typical UAV flight modes, the number of clusters is set to K=3, corresponding to hovering, hover-to-motion, and uniform motion states, respectively. Among them, the basic characteristics of hovering are small angular velocity variance, stable specific force mean, and ground velocity approaching zero, which can be corrected by Kalman filtering using near-zero velocity. The basic characteristics of hover-to-motion state are significant changes in angular velocity variance and gradual changes in specific force mean. When the carrier transitions from hovering to motion, there are changes in horizontal tilt angle and velocity. The short-time calculation accuracy of the inertial navigation system can be trusted, and no correction is made to ensure dynamic performance. The basic characteristics of uniform motion state are that the angular velocity variance is stable at a moderate level and the specific force mean conforms to the law of uniform motion, which can be corrected by Kalman filtering using historical stable velocity.

[0109] Furthermore, a supervised initialization method based on prior samples is adopted, utilizing known state data within the effective time period of integrated navigation (e.g., satellite / vision) to accurately anchor the feature intervals of three types of states. Specifically, during the normal operation period of the inertial navigation and satellite / vision integrated navigation, data under typical motion scenarios are collected as a sample set, including hovering state samples, hovering-to-motion state samples, and uniform motion state samples. Among them, hovering state samples are taken from the period when the velocity output by the satellite / vision is approximately zero, with a dynamic cumulative sample size of no less than 20,000 frames (corresponding to more than 100 seconds of data), and their dimensionality-reduced feature vectors are extracted; hovering-to-motion state samples are taken from the period when the velocity output by the satellite / vision changes in combination with the pitch or roll angle of the inertial navigation, with a dynamic cumulative sample size of no less than 20,000 frames; uniform motion state samples are taken from the period when the velocity output by the satellite / vision is stable, with a dynamic cumulative sample size of no less than 20,000 frames. The mean of the dimensionality-reduced feature vectors of the three types of samples is calculated respectively to obtain three initial cluster centers, denoted as hovering state center C1, hovering-to-motion state center C2, and uniform motion state center C3. These initial centers directly correspond to the average level of each state feature, ensuring that the starting point of clustering is located in a reasonable feature range.

[0110] The dimensionality-reduced feature matrix output in step S2 Each feature vector in ( Using these as samples to be clustered, iterative clustering is performed, including: calculating the Euclidean distance from each feature vector to each initial cluster center, and assigning them to the corresponding clusters according to the principle of minimum distance; updating the cluster centers to the mean of all feature vectors within the current cluster; repeating the iterative steps until the change in cluster centers is less than a preset threshold. Specifically:

[0111] 1) Calculate each feature vector separately. To the three initial cluster centers , , Euclidean distance:

[0112] ;

[0113] Where k is the dimension of the principal components after dimensionality reduction (k=3 in this embodiment). The score of the j-th principal component of the m-th frame of data. This represents the j-th dimension coordinate of the i-th cluster center. Let be the Euclidean distance from the feature vector of the m-th frame to the i-th cluster center.

[0114] 2) Based on the principle of minimum distance, the feature vectors are... Assign to the cluster category with the smallest distance. If If it is the smallest, it is marked as hovering; if If the minimum value is reached, it is marked as a hovering rotation state; if If the value is the smallest, it is marked as a state of uniform motion.

[0115] 3) For the three assigned clusters, calculate the mean of all feature vectors within each cluster, and use this mean as the new cluster center to update the cluster. , , Coordinates:

[0116] ;

[0117] in Let i be the set of all samples contained in the i-th cluster. This represents the number of samples in the cluster.

[0118] 4) Calculate the change in the coordinates of the cluster centers before and after the update. If the coordinate changes of all centers are less than a preset threshold (in this embodiment, we take...), then... If the condition is not met and the number of iterations has not reached the maximum number of iterations (set to 50 in this embodiment), then the iteration terminates and the current cluster label is output; if the condition is not met and the number of iterations has not reached the maximum number of iterations (set to 50 in this embodiment), then repeat the above steps 1)-3); if the termination condition is not met after reaching the maximum number of iterations, then the label corresponding to the current cluster center is used as the temporary result and enters the subsequent verification stage.

[0119] Furthermore, after outputting the initial clustering labels, post-processing steps are also included: adopting a continuous frame majority voting mechanism, taking the category that appears most frequently in the initial clustering labels of N consecutive frames centered on the current frame as the final state label of the current frame; setting an anti-jump mechanism based on motion physics characteristics, when an adjacent frame directly jumps from a hovering state to a uniform motion state, the label of the next frame is forcibly corrected to the transition state.

[0120] Specifically, a continuous frame majority voting mechanism and a physical law-based state recognition anti-jump mechanism are introduced to avoid state jitter caused by single-frame clustering errors. For the continuous frame majority voting mechanism, the initial clustering labels of N consecutive frames (N=11 in this embodiment) are selected to form a window. :

[0121] ;

[0122] in, The initial clustering label for the data in frame t (with values ​​of...) , , one); To ensure the window is symmetrical about the current frame, a floor sign is used for rounding down. The number of times each category appears within the window is counted.

[0123] ;

[0124] in, A label window consisting of N consecutive frames centered at frame t; For window Medium category Number of times it appears; The function is an indicator function that takes the value 1 when the condition in parentheses is true, and 0 otherwise. The category that appears most frequently is taken as the final state label of the data in frame t. If there is a tie for the most frequent category, the original label of the current frame or the closest category can be selected first. This strategy can effectively smooth out random fluctuations in a single frame and enhance the stability of state judgment.

[0125] For the state recognition anti-jump mechanism, based on the motion physics characteristics of the quadcopter UAV, there is no direct physical process between the hovering state and the uniform motion state (the two states have significantly different characteristics and require acceleration or deceleration transition). If the (t-1)th frame is in the hovering state and the tth frame is in the uniform motion state, or the (t-1)th frame is in the uniform motion state and the tth frame is in the hovering state, then the state recognition of the tth frame is determined to be abnormal, and the label of the tth frame is forcibly corrected to the hovering to motion state to ensure the physical rationality of the state sequence.

[0126] After the above post-processing, each data frame obtains a stable and reliable state label, which will serve as the core input for adaptive navigation calculation in step S4.

[0127] S4: Based on the identified motion state, adaptively select the appropriate combined navigation filtering strategy to perform navigation calculation.

[0128] Based on the current motion state of the UAV identified in step S3, a differentiated combined navigation filtering correction strategy is dynamically selected to minimize the accumulation of errors in the inertial navigation system and improve the accuracy and stability of pure inertial navigation in different scenarios with and without external auxiliary information. The specific implementation includes basic inertial calculation, Kalman filter model, motion feature recognition sample update, and adaptive selection of measurement equations.

[0129] like Figure 4 As shown, the adaptive selection of the integrated navigation filtering strategy includes: when external auxiliary information is valid, using the position provided by the auxiliary information as a measurement for Kalman filtering correction; when external auxiliary information is invalid, selecting different measurement methods according to the identified motion state: if identified as a hovering state, a zero-velocity correction strategy is adopted, using the velocity calculated by the inertial navigation system as the measurement; if identified as a uniform motion state, a velocity correction strategy is adopted, using the difference between the current velocity of the inertial navigation system and the average velocity at the initial moment of uniform motion as the measurement; if identified as a transition state (hovering to motion state), no filtering correction is performed, only pure inertial calculation is performed. Furthermore, it also includes using the integrated navigation output when external auxiliary information is valid to determine the true motion state of the current frame in real time, and combining this state with the corresponding dimensionality-reduced feature vector to form a new sample, dynamically updating the initial cluster centers.

[0130] Specifically, the inertial navigation system uses the angular velocity output from the gyroscope and the specific force output from the accelerometer as inputs, and updates the carrier's attitude, velocity, and position in real time through a strapdown inertial navigation algorithm. The rotation relationship between the carrier's coordinate system (b-frame) and the navigation coordinate system (n-frame, selected in this embodiment as the Northeast-Northeast Geographic coordinate system) is described using the quaternion method. The attitude quaternion update equation is:

[0131] ;

[0132] in , These are the attitude quaternions from the vehicle coordinate system to the navigation coordinate system at times t-1 and t, respectively. Represents quaternion multiplication; Let be the quaternion representing the attitude change from time t-1 to time t, describing the change in the carrier coordinate system over the time interval t-1. The rotation within the gyroscope is calculated from the angular increment output by the gyroscope:

[0133] ;

[0134] in, The angular increment vector (three-dimensional) from time t-1 to time t is obtained by integrating the angular velocity output by the gyroscope. The value is the pre-processed angular velocity of the three-axis gyroscope. The superscript b indicates that the angular velocity is defined in the carrier coordinate system, and the subscript ib indicates the rotation of the carrier relative to inertial space. The magnitude of the angular increment vector, i.e., the total rotation angle; fraction The unit vector representing the direction of the rotation axis.

[0135] The velocity update considers the force integral term and harmful acceleration terms (including Coriolis acceleration and gravitational acceleration):

[0136] ;

[0137] in , These are the velocity vectors (three-dimensional) of the vehicle in the navigation coordinate system at times t-1 and t, respectively, where n represents the navigation coordinate system; The projection of the velocity increment caused by specific force acceleration within the update cycle onto the navigation system; This is the projection of the velocity increment caused by harmful accelerations (including Coriolis acceleration and gravitational acceleration) within the update period onto the navigation frame. The velocity increment caused by specific force is obtained by performing a coordinate transformation on the specific force output by the accelerometer and integrating it. The harmful acceleration term needs to be calculated based on the vehicle velocity, the Earth's rotational angular velocity, and local gravity.

[0138] The location update uses the trapezoidal integral method to calculate longitude separately. ,latitude and height H:

[0139] ;

[0140] ;

[0141] ;

[0142] In the formula , , These are the speeds in the east, north, and sky directions, respectively. , These are the radii of curvature of the Earth's meridian and circumference, respectively. For the update cycle.

[0143] , , These represent the longitude, latitude, and altitude at time t, respectively. , , These represent the longitude, latitude, and altitude at time t-1, respectively. , , These represent the eastward velocity, northward velocity, and celestial velocity in the navigation system at time t, respectively. , , These represent the eastward, northward, and celestial velocities at time t-1, respectively. The radius of curvature of the Earth's meridian at the location of the carrier at time t-1 is related to latitude; The radius of curvature of the Earth's zonal circle at the location of the carrier at time t-1 is related to latitude; Update period: The time interval from t-1 to t.

[0144] This embodiment employs a 15-dimensional Kalman filter to achieve multi-source information fusion and error correction. The system state variables are defined as follows:

[0145] ;

[0146] in The platform misalignment angle (3D) describes the degree of misalignment between the navigation coordinate system and the real geographic coordinate system. The velocity error (three-dimensional) is the deviation between the velocity calculated by the inertial navigation system and the actual velocity in the east, north, and sky directions. The positional error (three-dimensional) corresponds to longitude error, latitude error, and altitude error, respectively. This refers to gyroscope drift (three-dimensional), which is the constant or slowly varying deviation between the gyroscope measurement and the actual angular velocity; This represents the accelerometer zero bias (three-dimensional), i.e., the constant or slowly varying deviation between the accelerometer measurement and the true specific force. The general forms of the system state equation and measurement equation are:

[0147] ;

[0148] F is the state transition matrix, which is established using the strapdown inertial navigation error model; The system noise consists of white noise from the gyroscope and accelerometer measurements, with a covariance matrix of Q; Z is the observation vector; and H is the measurement matrix, which maps the state vector to the observation space. The noise is measured, and its covariance matrix is ​​R.

[0149] The Kalman filter performs time and measurement updates according to a standard procedure. Time update (discrete form):

[0150] ;

[0151] ;

[0152] in, This is the estimated value of the state vector at time k-1; (The state transition matrix is ​​obtained by discretizing the continuous-time state transition matrix F). Let K be the mean square error matrix of the state estimation at time k-1; The system noise driving matrix is ​​used to discretize continuous noise. The system noise variance matrix (discrete form); This is a one-step prediction of the state at time k; This is the mean square error matrix for one-step prediction.

[0153] Measurement Update:

[0154] ;

[0155] ;

[0156] ;

[0157] In the formula Here is the filter gain matrix. Let k be the measurement matrix at time k. Let be the measurement noise variance matrix at time k. Let k be the actual observation vector at time k. This is the optimal estimate of the state at time k. Let I be the mean square error matrix of the state estimation at time k, and let I be the identity matrix.

[0158] When external auxiliary information such as satellite navigation, vision, or maps is available, the integrated navigation system can output high-precision velocity, position, and attitude information. This embodiment utilizes this information to dynamically optimize the initial cluster centers in step S3, making the clustering results more closely match the actual motion characteristics of the current aircraft.

[0159] Specifically, using the ground speed output by the integrated navigation system (in , These are the eastward and northward velocities and the horizontal attitude angle (pitch angle), respectively. Roll angle The stability of ) is used to determine the true label of the current flight status in real time: if If so, then the current frame is determined to be in a hovering state; if And the standard deviation of the horizontal attitude angle within 0.1 seconds is less than If the state is constant, it is classified as uniform motion; all other states are classified as hovering and turning motion. The real-time determined state label and the corresponding reduced-dimensional feature vector are combined to form a new sample, which is used to dynamically adjust the three initial centers of K-means clustering. , , This enables online adaptive optimization of cluster centers, further improving the accuracy of state recognition.

[0160] Based on the availability of external auxiliary information and the state recognition result output in step S3, different measurement equations are adaptively selected for Kalman filtering correction. Specifically, when satellite / visual or other auxiliary information is available, the position provided by the auxiliary system is directly used as the measurement to correct the inertial navigation position error. The measurement vector and measurement matrix are as follows:

[0161] ;

[0162] in , , These represent the longitude, latitude, and altitude calculated by the inertial navigation system at time t, respectively. , , These are the longitude, latitude, and altitude provided by the external auxiliary system at time t, respectively. It is a three-dimensional identity matrix; in the H matrix express The zero matrix corresponds to the first 6 dimensions of the state ( and It does not participate in measurement updates.

[0163] In scenarios where satellite signals are denied or visual failure occurs, based on the motion state identified in step S3, a virtual quantity measurement is constructed using the inertial navigation system's own velocity information to suppress error divergence. For hovering states ( The zero-velocity correction strategy is adopted, directly using the velocity calculated by the inertial navigation system as the measurement, forcing the velocity error to converge to zero; the measurement vector and measurement matrix are:

[0164] ;

[0165] in , , The inertial navigation system calculates the eastward, northward, and upward velocities at time t; the H matrix contains... correspond , , States in a certain state do not participate in measurement updates.

[0166] For uniform motion ( Once the drone enters uniform motion, its speed theoretically remains constant. In this embodiment, after determining that the state is uniform motion, speed data is continuously collected for 5 seconds, and the average speed within this period is taken as the reference speed. , , Subsequently, the difference between the current inertial navigation velocity and the reference velocity is used as a measurement to correct the velocity error.

[0167] ;

[0168] For hovering and rotating motion state ( This state is a dynamic transition process, where the inertial navigation system has high short-term accuracy and no stable velocity reference. Therefore, in this state, no Kalman filter correction is performed, and only pure inertial calculation is performed to ensure the dynamic response performance of the system.

[0169] By employing the aforementioned adaptive measurement selection strategy, the system can intelligently select the optimal correction method based on the UAV's own motion state in the absence of external information, significantly suppressing the accumulation of speed and position errors in pure inertial navigation and enhancing the system's autonomous navigation capability in denied environments.

[0170] Example 2

[0171] This embodiment exemplifies an adaptive inertial navigation solution system based on motion feature recognition, including an inertial measurement unit, a feature extraction module, a dimensionality reduction module, a state recognition module, and an adaptive navigation solution module.

[0172] An inertial measurement unit (IMU) is used to acquire the angular velocity and specific force data of the carrier. The IMU consists of a three-axis gyroscope and a three-axis accelerometer, and is used to acquire raw angular velocity data of the carrier in real time during its motion. And the original data (i=x,y,z). This unit outputs the raw measurement values ​​at a fixed frequency, which serve as the input basis for all subsequent processing stages.

[0173] The feature extraction module is used to extract multidimensional features reflecting the motion state of the carrier based on angular velocity and specific force data, and construct a feature matrix. The feature extraction module receives the raw data output from the inertial measurement unit and first performs preprocessing: outlier removal is performed on the static initialization data after the aircraft is powered on using the 3σ criterion, and the gyroscope zero bias is calculated using the static data after outlier removal. and accelerometer zero bias Then, zero-bias compensation is performed on all the original data to obtain a clean angular velocity. And comparison The data sequence is then divided into continuous data frames using a sliding time window. Multidimensional features are extracted within each window, including time-domain features (mean values ​​of three-axis angular velocities and three-axis specific forces), statistical features (variance of three-axis angular velocities and three-axis specific forces, gyroscope data energy, and accelerometer data energy), and correlation features (Pearson correlation coefficients between the corresponding axes of the gyroscope and accelerometer). Finally, a feature matrix is ​​constructed. Where M is the number of windows and D is the feature dimension (D=17 in this embodiment).

[0174] The dimensionality reduction module is used to reduce the dimensionality of the feature matrix to obtain lower-dimensional feature vectors. The module also standardizes the feature matrix to eliminate the influence of different feature units, resulting in a standardized matrix. Then, the covariance matrix Σ of the normalized matrix is ​​calculated, and its eigenvalues ​​are obtained through eigenvalue decomposition. and the corresponding feature vectors The number of principal components to be retained, k (k=3 in this embodiment), is determined based on the cumulative variance contribution rate threshold (90% in this example). The first k eigenvectors are selected to form the projection matrix U. The standardized matrix is ​​then... Projecting onto the low-dimensional space spanned by U, we obtain the dimensionality-reduced feature matrix. each of the lines Let be the k-dimensional principal component feature vector of the m-th frame of data.

[0175] The state recognition module is used to perform cluster analysis on low-dimensional feature vectors based on a clustering algorithm that initializes cluster centers using prior samples, thereby identifying the motion state of the vehicle. The state recognition module takes the dimensionality-reduced feature matrix Y as input and uses a K-means clustering algorithm based on prior samples to identify the vehicle's motion state. First, within the effective time period of integrated navigation, sample data of known states (hovering, hovering-to-motion, uniform motion) are collected, and the mean of their dimensionality-reduced feature vectors is calculated for each state, serving as three initial cluster centers. , , Then, the clustering process (distance calculation, category assignment, center update) is iterated until convergence or the maximum number of iterations is reached, outputting the initial clustering label for each frame of data. To avoid misjudgment in a single frame, the module further introduces a post-processing mechanism: it uses continuous frame mode voting to smooth label fluctuations, and sets anti-jump logic based on motion physics characteristics. When a direct jump occurs between hovering and constant speed states, it is forcibly corrected to a transition state, and finally outputs stable and reliable state labels.

[0176] The adaptive navigation solution module adaptively selects the appropriate combined navigation filtering strategy for navigation solution based on the identified motion state. Different motion states correspond to different filtering correction methods. The adaptive navigation solution module integrates basic inertial calculation and Kalman filtering, dynamically selecting the filtering correction strategy based on the availability of motion labels output by the state recognition module and external auxiliary information. Basic inertial calculation uses preprocessed angular velocity and specific force as input, updating attitude, velocity, and position using quaternions. When external auxiliary information such as satellite / vision is effective, the module directly uses the auxiliary position as a measurement for Kalman filtering to correct inertial navigation errors. It then uses the high-precision output at this point to determine the true motion state in real time, forming new samples by combining the state labels with the corresponding dimensionality-reduced feature vectors, and feeding this data back to the state recognition module to dynamically optimize cluster centers. When external assistance fails, the module selects a correction method based on the current state label: if in hovering, a zero-velocity correction strategy is used, with the inertial navigation velocity as the measurement; if in uniform motion, a velocity correction strategy is used, with the difference between the current inertial navigation velocity and the average velocity at the initial moment of uniform motion as the measurement; if in a transitional state, no filtering correction is performed, and only pure inertial calculation is executed. Through the above adaptive mechanism, the system can effectively suppress error accumulation and improve navigation accuracy and stability even in denied environments.

[0177] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims. It should be understood that the invention is not limited to the precise structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. An adaptive solution method for inertial navigation based on motion feature recognition, characterized in that, The method includes: Based on the angular velocity and specific force data collected by the inertial measurement unit, multidimensional features reflecting the motion state of the carrier are extracted, and a feature matrix is ​​constructed. The feature matrix is ​​reduced in dimensionality to obtain a low-dimensional feature vector; The clustering algorithm based on prior sample initialization of cluster centers performs cluster analysis on the low-dimensional feature vector to identify the motion state of the carrier and output the initial cluster label for each frame of data. The prior sample initialization of cluster centers is obtained by collecting sample data corresponding to known motion states during the effective time period of integrated navigation and calculating the mean of the reduced feature vector of each motion state sample. The initial clustering labels are post-processed, including: using a continuous frame majority voting mechanism, the category that appears most frequently in the initial clustering labels of N consecutive frames centered on the current frame is taken as the final state label of the current frame; and setting an anti-jump mechanism based on motion physics characteristics, when an adjacent frame jumps directly from a hovering state to a uniform motion state, the label of the next frame is forcibly corrected to the hovering to motion state. Based on the final state label, the corresponding combined navigation filtering strategy is adaptively selected for navigation calculation, and different filtering correction methods correspond to different motion states.

2. The adaptive solution method for inertial navigation based on motion feature recognition according to claim 1, characterized in that, The extraction of multidimensional features reflecting the motion state of the carrier includes: The raw angular velocity and specific force data are preprocessed. The preprocessing includes removing outliers from the static initialization data using the 3σ criterion, and calculating the gyroscope zero bias and accelerometer zero bias using the static data after removing outliers, and performing zero bias compensation on the raw data. The preprocessed data is divided into data frames by using a sliding time window. Temporal features, statistical features, and correlation features are extracted within each window to form the feature matrix.

3. The adaptive inertial navigation solution method based on motion feature recognition according to claim 2, characterized in that, The time-domain features include the mean angular velocity and the mean specific force of the three axes within the window; the statistical features include the variance of the angular velocity of the three axes, the variance of the specific force of the three axes, and the energy of the gyroscope data and the energy of the accelerometer data within the window; the correlation features include the Pearson correlation coefficient between the corresponding axis data of the gyroscope and the accelerometer within the window.

4. The adaptive inertial navigation solution method based on motion feature recognition according to claim 1, characterized in that, The dimensionality reduction process for the feature matrix includes: The feature matrix is ​​standardized, the covariance matrix of the standardized feature matrix is ​​calculated, and the covariance matrix is ​​decomposed to obtain eigenvalues ​​and eigenvectors. The number of principal components to be retained is determined based on the cumulative variance contribution rate, and the eigenvectors corresponding to the first k eigenvalues ​​are selected to form the projection matrix. The standardized feature matrix is ​​projected onto the k-dimensional space spanned by the projection matrix to obtain the dimensionality-reduced feature matrix.

5. The adaptive solution method for inertial navigation based on motion feature recognition according to claim 1, characterized in that, Starting from the initial cluster centers, iterative clustering is performed on the dimensionality-reduced feature matrix until the cluster centers converge or the maximum number of iterations is reached, and the initial cluster labels of each frame of data are output.

6. The adaptive solution method for inertial navigation based on motion feature recognition according to claim 5, characterized in that, The iterative clustering includes: Calculate the Euclidean distance from each feature vector to each of the initial cluster centers, and assign them to the corresponding clusters according to the principle of minimum distance; update the cluster centers to the mean of all feature vectors in the current cluster; Repeat the clustering steps until the change in cluster centers is less than a preset threshold.

7. The adaptive solution method for inertial navigation based on motion feature recognition according to claim 1, characterized in that, The adaptive selection of the appropriate integrated navigation filtering strategy includes: When the external auxiliary information is valid, the position provided by the external auxiliary information is used as the measurement for Kalman filtering correction. When the external auxiliary information is invalid, different measurement methods are selected based on the identified motion state: If the state is identified as hovering, a zero-speed correction strategy is adopted, using the velocity calculated by the inertial navigation system as the measurement; if the state is identified as uniform motion, a velocity correction strategy is adopted, using the difference between the current velocity of the inertial navigation system and the average velocity at the initial moment of uniform motion as the measurement; if the state is identified as hovering to motion, no filtering correction is performed, and only pure inertial calculation is performed.

8. The adaptive solution method for inertial navigation based on motion feature recognition according to claim 7, characterized in that, The method further includes using the combined navigation output when the external auxiliary information is effective to determine the real motion state of the current frame in real time, and combining the real motion state with the corresponding dimensionality-reduced feature vector to form a new sample, and dynamically updating the initial cluster center.

9. An adaptive inertial navigation solution system based on motion feature recognition, used to implement the method described in any one of claims 1-8, characterized in that, It includes an inertial measurement unit, a feature extraction module, a dimensionality reduction module, a state recognition module, and an adaptive navigation solution module; The inertial measurement unit is used to collect the angular velocity and specific force data of the carrier; The feature extraction module is used to extract multi-dimensional features reflecting the motion state of the carrier based on the angular velocity and specific force data, and construct a feature matrix. The dimensionality reduction module is used to perform dimensionality reduction processing on the feature matrix to obtain a low-dimensional feature vector; The state recognition module is used to perform cluster analysis on the low-dimensional feature vector based on the clustering algorithm that initializes cluster centers based on prior samples, and to identify the motion state of the carrier. The adaptive navigation solution module is used to adaptively select the corresponding combined navigation filtering strategy for navigation solution based on the identified motion state, with different filtering correction methods corresponding to different motion states.

Images