An inertial measurement unit implicit noise estimation method and system

Abstract

The application provides an inertial measurement unit implicit noise estimation method and system, and relates to the technical field of inertial navigation, and the method comprises the following steps: obtaining an acceleration sequence and an angular velocity sequence; performing attribute-in-axis direction feature extraction respectively to obtain acceleration attribute features and angular velocity attribute features; performing cross-attribute feature interaction to obtain fusion features; performing noise estimation based on the fusion features to obtain estimated noise; performing noise removal on inertial measurement data based on the estimated noise to obtain denoised inertial measurement data; and performing integral processing on the denoised inertial measurement data to obtain velocity estimation and pose estimation. The application learns a noise estimation function in an implicit manner through an end-to-end neural network, does not need explicit noise modeling and does not depend on external sensors, can effectively suppress IMU integral drift, and significantly improves the long-term positioning accuracy and robustness of inertial navigation.

Description

Technical Field

[0001] This invention relates to the field of inertial navigation technology, and in particular to an implicit noise estimation method and system for inertial measurement units. Background Technology

[0002] An Inertial Measurement Unit (IMU) consists of an accelerometer and a gyroscope, capable of acquiring linear acceleration and angular velocity information of an object, respectively. It offers advantages such as low cost, small size, and immunity to ambient light and occlusion, and is widely used for localization and attitude estimation in quadruped robots, drones, autonomous driving, and wearable devices. Under ideal, noise-free conditions, the velocity and position of an object can be obtained by integrating the acceleration measurements. However, in reality, IMU measurements are inevitably affected by noise and bias. The measurement model for accelerometers and gyroscopes can be expressed as... , ,in and These represent the corresponding measurement noise. Performing a second integration on the noisy acceleration will continuously amplify the error, ultimately leading to unbounded drift in the positioning results.

[0003] In existing technologies, two approaches are typically used to address the aforementioned drift problem: one is to establish an explicit mathematical model of the noise (such as a random walk model or a Gauss-Markov model) and perform state estimation using filters such as Kalman filtering; the other is to rely on external sensors (such as GPS or visual cameras) to provide auxiliary correction information. However, IMU noise exhibits time-varying and nonlinear characteristics, with significant differences in noise properties under different motion states, making it extremely difficult to accurately establish a mathematical analytical model of the noise. Furthermore, relying on external sensors increases system cost and complexity, and external auxiliary information is unavailable in environments where GPS is denied or vision is degraded. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an implicit noise estimation method and system for inertial measurement units.

[0005] The first aspect of the present invention provides an implicit noise estimation method for an inertial measurement unit, comprising: Obtain the acceleration and angular velocity sequences; Axial features within the attribute are extracted from the acceleration sequence and the angular velocity sequence respectively to obtain acceleration attribute features and angular velocity attribute features; Cross-attribute feature interaction is performed on the acceleration attribute feature and the angular velocity attribute feature to obtain the fused feature; Based on the fusion features, noise estimation is performed to obtain the estimated noise; The inertial measurement data is denoised based on the estimated noise to obtain denoised inertial measurement data. The denoised inertial measurement data is integrated to obtain velocity and pose estimates.

[0006] Furthermore, noise estimation based on the fused features includes: using a graph attention network to capture local temporal dependencies of the fused features to obtain local temporal features; using a state-space model to capture long-range temporal dependencies of the fused features to obtain long-range temporal features; and predicting noise components based on the local temporal features and the long-range temporal features to obtain the estimated noise.

[0007] Furthermore, noise estimation is performed based on the fusion features to obtain estimated noise, including: predicting the accelerometer noise component and the gyroscope noise component based on the fusion features to obtain the predicted noise component; The inertial measurement data is denoised based on the estimated noise to obtain denoised inertial measurement data, including: subtracting the accelerometer noise component from the original acceleration measurement value and subtracting the gyroscope noise component from the original angular velocity measurement value to obtain the denoised inertial measurement data.

[0008] Further, noise component prediction based on the local time-series features and the long-range time-series features includes: concatenating the local time-series features and the long-range time-series features to calculate a gating signal; and performing an adaptive weighted summation on the long-range time-series features and the local time-series features based on the gating signal to obtain a time-series representation, wherein the adaptive weighted summation is expressed as: , For the long-range time series features, For the local temporal features, The gated signal is used; noise component prediction is performed based on the time-series characterization to obtain the estimated noise.

[0009] Furthermore, the method is implemented using an end-to-end neural network, which learns an implicit noise estimation function. The input to the implicit noise estimation function is the noisy acceleration measurement value. and noisy angular velocity measurement values The output is the estimated accelerometer noise component. and estimated gyroscope noise components ,satisfy ,in This refers to the set of trainable parameters obtained by the end-to-end neural network through training.

[0010] Furthermore, axial feature extraction within attributes is performed on the acceleration sequence and the angular velocity sequence respectively, including: processing the acceleration sequence and the angular velocity sequence separately using stacked residual depth separable time series blocks, each residual depth separable time series block being represented as follows: ,in For depthwise temporal convolutions that act independently on each axis, For pointwise convolution of cross-axis mixed channel information at each time step, For batch normalization, The acceleration attribute features and the angular velocity attribute features are obtained by stacking the functions point by point as nonlinear activation functions.

[0011] A second aspect of the present invention provides an implicit noise estimation system for an inertial measurement unit, comprising: The data acquisition module is used to acquire acceleration and angular velocity sequences. The attribute-in-attribute feature extraction module is used to extract axial features within the attribute from the acceleration sequence and the angular velocity sequence respectively, to obtain acceleration attribute features and angular velocity attribute features; A cross-attribute feature interaction module is used to perform cross-attribute feature interaction between the acceleration attribute feature and the angular velocity attribute feature to obtain fused features; The noise estimation module is used to perform noise estimation based on the fused features to obtain estimated noise; A denoising module is used to denoise the inertial measurement data based on the estimated noise to obtain denoised inertial measurement data. The integration module is used to integrate the denoised inertial measurement data to obtain velocity and pose estimates.

[0012] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor, when executing the program, implements the steps in the implicit noise estimation method for inertial measurement units as described in the first aspect of the present invention.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the implicit noise estimation method for inertial measurement units as described in the first aspect of the present invention.

[0014] A fifth aspect of the present invention provides a computer program product comprising software code, wherein the program in the software code performs steps in the implicit noise estimation method for inertial measurement units as described in the first aspect of the present invention.

[0015] Compared with existing technologies, the implicit noise estimation method and system for inertial measurement units provided by this invention, through a hierarchical processing mechanism of independent modeling within attributes and controllable interaction between attributes, enables the network to capture the coupling relationship between rotational motion and gravity-induced attributes while preserving the physical specificity of acceleration and angular velocity. Thus, without the need to explicitly establish a noise mathematical model or rely on external sensors, it achieves accurate estimation and effective suppression of IMU time-varying noise, significantly improving the long-term positioning accuracy and robustness of inertial navigation. Attached Figure Description

[0016] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0017] Figure 1 This is a flowchart of the implicit noise estimation method for inertial measurement units provided in Embodiment 1 of the present invention.

[0018] Figure 2 This is an architecture diagram of the implicit noise estimation system for inertial measurement units provided in Embodiment 1 of the present invention.

[0019] Figure 3 This is a schematic diagram of the motion-sensing inertial measurement unit odometer network framework provided in Embodiment 2 of the present invention. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0023] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0024] Example 1 This embodiment provides an implicit noise estimation method for inertial measurement units (IMUs). This method extracts intra-attribute axial features from acceleration and angular velocity sequences, then performs cross-attribute feature interaction to obtain fused features. Noise estimation is performed based on the fused features to obtain estimated noise. The estimated noise is then used to denoise the inertial measurement data, and the denoised data is integrated to obtain velocity and pose estimates. An inertial measurement unit (IMU) consists of an accelerometer and a gyroscope, used to measure linear acceleration and angular velocity, respectively. Under ideal noise-free conditions, the velocity of an object... With position By measuring the actual acceleration Earn points by playing:

[0025] However, in practical applications, IMU measurements are inevitably affected by noise and bias. The measurement model for accelerometers and gyroscopes can be expressed as:

[0026] in and These represent true linear acceleration and angular velocity, respectively. and This corresponds to the measurement noise. Performing a second integration on the noisy acceleration continuously amplifies the error, eventually leading to unbounded drift. To address this problem, this invention learns an estimation function. This implicitly captures complex time-varying noise without explicitly modeling it. Specifically, the goal is to learn the following function:

[0027] in and These are the estimated accelerometer and gyroscope noise components, respectively. θ is the set of all trainable weight parameters and bias parameters determined through training in the end-to-end neural network, including the learnable fusion parameters of the input denoising layer, the convolutional and normalization parameters in the acceleration and angular velocity branches, the parameters of the early fusion module, the parameters of the cross attention module, the parameters of the channel and temporal attention modules, the parameters of the graph attention network, the parameters of the state space model, the parameters of the gated fusion module, the parameters of the multi-scale temporal feature extraction module, the parameters of the noise feature correction module, and the parameters of the output layer.

[0028] like Figure 1 This embodiment provides an implicit noise estimation method for inertial measurement units, including: S101. Obtain the acceleration sequence and angular velocity sequence.

[0029] Specifically, set These are inertial measurement values ​​in the equipment coordinate system. To ensure coordinate consistency between inertial input and monitoring signals, all IMU measurements are transformed from the device coordinate system to a unified global coordinate system through corresponding attitude estimation. Given attitude quaternions... The rotational transformation is as follows:

[0030] For continuous inertial measurement sequences, a fixed-length sliding window is used to segment the raw IMU data stream to construct training samples:

[0031] Based on this, the IMU input is decomposed into linear acceleration flow and angular velocity flow at the attribute level to obtain acceleration sequence and angular velocity sequence.

[0032] S102. Extract the axial features within the attributes of the acceleration sequence and the angular velocity sequence respectively to obtain acceleration attribute features and angular velocity attribute features.

[0033] Specifically, in model-based inertial state estimation, gyroscopes and accelerometers play fundamentally different roles: gyroscopes are primarily used for attitude propagation, while accelerometers measure the specific force generated by both translational motion and gravity. To explicitly represent the physical heterogeneity of the inertial signal, the IMU input is decoupled at the attribute level, dividing the inertial sequence into acceleration flow and angular velocity flow:

[0034] The two signals are modeled through parallel processing branches. Within each branch, the three-axis signals are hierarchically integrated to capture their attribute-specific motion patterns and cooperative dynamics. Specifically, depthwise separable convolutions are used to achieve axial temporal modeling and instantaneous cross-axis feature fusion. Given the attribute stream... Attribute-specific representation Obtained through a series of temporal convolution and channel mixing operations:

[0035] in The number of stacked stages, each stage For a residual depth separable time block:

[0036] in For depthwise temporal convolutions that act independently on each axis, For pointwise convolutions that mix information across axes at each time step, For batch normalization, It is a pointwise nonlinear activation function.

[0037] Deep Temporal Convolution Perform convolution operations independently along each axis of time to capture the temporal evolution pattern of the signal on each axis; pointwise convolution. Linear mixing of the three-axis channels is performed at each time step to achieve instantaneous cross-axis feature fusion. By residual connections and stacking multiple temporal blocks, the network can extract hierarchical feature representations from local to global layers. Depthwise separable convolutions decouple axial temporal modeling from cross-axis channel mixing, reducing computational complexity while achieving efficient modeling of the collaborative relationships of three-axis signals within each attribute, providing high-quality, attribute-specific feature representations for subsequent cross-attribute interactions.

[0038] S103. Perform cross-attribute feature interaction on the acceleration attribute feature and the angular velocity attribute feature to obtain the fused feature.

[0039] Specifically, although acceleration and angular velocity are treated independently in the early stages, the coupling between rotational motion and gravity-induced properties necessitates controlled information interaction between them. To capture this interdependence without disrupting the property-specific representation, a cross-attribute interaction mechanism is introduced:

[0040] in and This is a cross-attention-based fusion operator that allows one attribute stream to selectively query complementary information from another attribute stream. Given query features... and key features First, calculate the cross-attribute temporal attention matrix:

[0041] The query representation is then updated using residuals:

[0042] in It is a learnable scalar used to regulate the strength of cross-attribute interactions.

[0043] Using acceleration features as queries and angular velocity features as keys, a cross-attribute attention matrix is ​​calculated. Information related to acceleration features in the angular velocity features is weighted, aggregated, and then superimposed onto the acceleration features as residuals; conversely, information not related to acceleration features is not weighted and aggregated. Learnable scalar. Control the intensity of interaction, when When it approaches 0, it degenerates into independent processing. When the size is large, sufficient interaction is allowed. This residual design achieves controlled information exchange between heterogeneous motion attributes while preserving the original attribute-specific representations. Through a cross-attention mechanism, one attribute stream can adaptively query complementary information related to the current task from another attribute stream, while simultaneously learning scalars. It provides flexible interaction intensity adjustment capabilities, avoiding the loss of attribute-specific information caused by excessive interaction. The result is an enhanced acceleration feature. and enhanced angular velocity characteristics , as a feature of fusion.

[0044] S104. Based on the fusion features, noise estimation is performed to obtain the estimated noise.

[0045] Specifically, the fused features are input into the temporal modeling part, and a band-limited graph attention network (GAT) and a state-space model (SSM) are used in parallel, combined with a gating fusion mechanism, to simultaneously capture local and long-range temporal dependencies in the sequence data. The graph attention network learns pairwise dependencies between time steps by focusing on other time steps within a fixed radius. The attention score is calculated as follows:

[0046] To focus on relevant temporal interactions, a band-limited mask is introduced to restrict attention to time steps within a radius r. The attention weights, after softmax normalization, are used to aggregate value features.

[0047]

[0048] in, For the first The time step for the first Normalized attention weights for each neighborhood time step; , for the first Centered on a time step, with a domain radius of [missing information] The set of local time steps The length of the input sequence; The local temporal features obtained by weighted aggregation of neighborhood value features in a graph attention network; , for the first Input features at time steps Value mapping matrix The characteristics of the values ​​obtained after projection.

[0049] State-space models effectively capture long-range temporal dependencies by modeling the evolution of hidden states over time. System dynamics are described jointly by state transition equations and observation equations.

[0050]

[0051] in, For the first Hidden states at each time step For the first The fusion features of the input state space model at each time step For the first Output features at each time step , , These are the state transition matrix, input mapping matrix, and output mapping matrix, respectively. For the first The process perturbation term at each time step For the first The observed perturbation term at each time step.

[0052] GAT captures local temporal dependencies within a fixed radius by explicitly modeling pairwise attention relationships between time steps, excelling at capturing transient shocks and rapid changes during motion. SSM, through recursive evolution of hidden states, compresses historical information into the hidden states, excelling at capturing long-term motion trends and slow drift. Both processes the same fused feature in parallel, extracting temporal information from different time scales. This parallel architecture of GAT and SSM simultaneously captures dynamic features at two time scales, providing richer temporal representations for subsequent noise estimation.

[0053] Furthermore, the gating fusion mechanism combines the outputs of GAT and SSM in parallel, with the former capturing local timing dependencies and the latter encoding long-range memory. The gating signal is then calculated after concatenating the two outputs. The final output is an adaptive weighted combination of the two:

[0054] in For the long-range memory features output by SSM, These are the short-range local features output by GAT.

[0055] Gating signal The values ​​are calculated from the concatenated features using a fully connected layer and a sigmoid activation function, with a range of (0,1). When the motion state changes drastically... When the motion approaches zero, the model focuses more on local features to respond quickly to transient changes; when the motion tends to be stable, Approaching 1, the model prioritizes long-range memory to maintain global consistency. This adaptive mechanism allows the model to dynamically adjust the emphasis on short-term and long-term features based on contextual information. The entire module is constructed as residual blocks and incorporates batch normalization and Dropout for regularization. Through gating, the model can dynamically adjust feature weights based on the contextual information of the input data, maintaining responsiveness during rapid maneuvers and smoothness of estimation during steady motion, thus achieving efficient modeling of multi-scale temporal dynamics.

[0056] S105. Based on the estimated noise, the inertial measurement data is denoised to obtain denoised inertial measurement data.

[0057] Specifically, based on the temporal representation output by the temporal modeling part, motion-related noise components are explicitly predicted. Specifically, accelerometer noise components and gyroscope noise components are predicted based on fused features to obtain the predicted noise components; then, the accelerometer noise component is subtracted from the original acceleration measurement value, and the gyroscope noise component is subtracted from the original angular velocity measurement value to obtain the denoised inertial measurement data. The entire method is implemented through an end-to-end neural network, which learns an implicit noise estimation function. The input is a noisy acceleration measurement value. and noisy angular velocity measurement values The output is the estimated accelerometer noise component. and estimated gyroscope noise components ,satisfy:

[0058] Unlike traditional explicit noise modeling methods such as Kalman filtering, this invention does not presuppose that noise follows a specific distribution (such as a Gaussian distribution). Instead, it learns the implicit representation of noise directly from the data through a neural network. The network takes noisy IMU measurements as input, performs hierarchical attribute fusion and temporal modeling, explicitly predicts the noise component at the output, and then removes the noise from the original signal through a subtraction operation. This "predict noise → subtract noise" paradigm, compared to directly predicting a clean signal end-to-end, preserves the high-frequency motion information in the original signal and avoids over-smoothing. Through implicit noise estimation, without the need for explicit noise modeling or external sensors, it effectively suppresses IMU integral drift and achieves high-precision velocity and pose estimation.

[0059] S106. Integrate the denoised inertial measurement data to obtain velocity and pose estimates.

[0060] Specifically, the denoised IMU sequence is ultimately fed into the integration section. Velocity is estimated by performing a first integration on the denoised acceleration data, and position is estimated by performing a second integration. Attitude is estimated by integrating the denoised angular velocity data. The integration process can be implemented using numerical integration methods such as trapezoidal integration, Simpson integration, or median integration.

[0061] In this embodiment, intra-attribute modeling and cross-attribute interaction provide high-quality, attribute-aware feature representations for temporal modeling, which in turn captures multi-scale temporal dynamics. In existing technologies, attribute-level decoupling and temporal modeling are typically considered two independent research directions. Those skilled in the art usually focus either on improving the attribute fusion method for IMU signals or on improving the temporal modeling architecture. Combining the two organically is not a simple superposition, because the features output by hierarchical attribute fusion have specific attribute semantic structures, and directly inputting them into a general temporal modeling module may dilute attribute-specific information during the temporal modeling process. This invention, by concatenating attribute fusion and temporal modeling and employing a parallel architecture of GAT and SSM in temporal modeling, enables temporal modeling to capture multi-scale temporal dynamics while preserving attribute-specific information. The synergy between the two produces better results than using them alone: ​​attribute fusion provides richer input features for temporal modeling, and temporal modeling provides more global contextual information for attribute fusion, jointly improving the accuracy of noise estimation.

[0062] To further verify the technical effectiveness of this invention, comparative experiments were conducted on the publicly available datasets RoNIN, RIDI, OzIOD, and IMUNet-Proposed. Table 1. Performance comparison of ATE and RTE on four datasets.

[0063] As shown in Table 1, the end-to-end neural network proposed in this invention achieves state-of-the-art performance on all five public datasets and both evaluation metrics, with its absolute trajectory error (ATE) and relative trajectory error (RTE) significantly lower than all existing comparative methods. Specifically, on the RoNIN-Seen dataset, the ATE of this invention is reduced by approximately 9.31% to 36.14% compared to existing solutions; on the RoNIN-Unseen dataset, which has the highest generalization requirement, this invention still maintains the best results, indicating that the model has excellent cross-scene generalization ability. On datasets such as IMUNet-Proposed, RIDI, and OxIOD, which cover different motion modes and sensor noise characteristics, this invention also achieves the lowest trajectory drift error, fully verifying its adaptability and robustness to complex and variable motion states.

[0064] Example 2 like Figure 2 This embodiment provides an implicit noise estimation system for inertial measurement units, including: The 201 data acquisition module is used to acquire acceleration and angular velocity sequences. 202 In-attribute feature extraction module, used to extract axial features within attributes from the acceleration sequence and the angular velocity sequence respectively, to obtain acceleration attribute features and angular velocity attribute features; 203 Cross-attribute feature interaction module, used to perform cross-attribute feature interaction on the acceleration attribute feature and the angular velocity attribute feature to obtain fused features; 204 Noise estimation module, used to perform noise estimation based on the fused features to obtain estimated noise; 205 Denoising module, used to denoise the inertial measurement data based on the estimated noise, to obtain denoised inertial measurement data; The 206 integration module is used to perform integration processing on the denoised inertial measurement data to obtain velocity estimation and pose estimation.

[0065] In one specific embodiment, such as Figure 3 As shown, the system architecture includes a complete processing flow of input sequence processing, hierarchical attribute fusion, temporal modeling, adaptive gating, noise prediction, denoising and integral estimation. The modules are connected through data flow and signal flow to form an end-to-end motion-sensing IMU odometry network framework.

[0066] In the input sequence processing section, the input sequence contains two types of physical quantities: acceleration and angular velocity, each containing triaxial components. The acceleration sequence contains... , , Three components, the angular velocity sequence contains , , Three components. The input sequence employs an attribute decoupling strategy to separate the two heterogeneous sensing signals, acceleration and angular velocity, at the input level, representing them as sequence dots and small feature blocks, thus providing a structured input data foundation for subsequent hierarchical attribute fusion.

[0067] In the layered attribute fusion module section, such as Figure 3As shown in the dashed box on the lower left, the hierarchical attribute fusion module comprises two levels: intra-attribute fusion and inter-attribute fusion. In the intra-attribute fusion stage, the components of acceleration along the three axes and angular velocity along the three axes undergo independent feature extraction processing. A depthwise separable convolutional structure is used to perform temporal modeling and channel mixing of the three-axis signals within each attribute. In the inter-attribute fusion stage, an attention mechanism with query, key, and value structures enables cross-attribute interaction. After contextual information processing and context modulation, acceleration and angular velocity attribute features exchange information through the inter-attribute fusion module, outputting enhanced fused features.

[0068] In the time series modeling module, such as Figure 3 As shown in the dashed box on the lower right, the temporal modeling module receives acceleration and angular velocity features, arranged in three consecutive time steps: t-1, t, and t+1. It employs a graph attention network and a state-space model for parallel processing. The graph attention network (upper part) learns pairwise dependencies between time steps by focusing on other time steps within a fixed radius, extracting dynamic features of these relationships. It excels at capturing transient impacts and rapid changes during motion. The state-space model (lower part) compresses historical information into hidden states through recursive evolution of hidden states, extracting temporal pattern features. It excels at capturing long-range motion trends and slow drifts. Both outputs converge into the gating feature section (removing quotes) marked by the dashed ellipse on the right, providing multi-scale temporal representations for subsequent adaptive gating.

[0069] In the adaptive gating section, the adaptive gating module receives context acceleration and context angular velocity information from the temporal modeling module. Through a gating fusion mechanism, it adaptively weights and combines the local temporal features output by the graph attention network and the long-range temporal features output by the state-space model. The calculation formula is as follows: Among them, the gating signal The concatenated features are calculated using a fully connected layer and a sigmoid activation function. When the motion state changes drastically... As the value approaches zero, the model focuses more on local features to respond quickly to transient changes. When the motion tends to be stationary, As the value approaches 1, the model places greater emphasis on long-range memory to maintain global consistency.

[0070] In the noise prediction and denoising section, the timing representation of the adaptive gated output enters the vertical noise prediction module, explicitly predicting the noise components in the IMU measurements, including accelerometer and gyroscope noise components. The noise prediction result is subtracted from the long-path signal from the input side using a circular operation node with a minus sign to remove noise, resulting in the denoised IMU data as shown on the right. The denoised IMU data is represented in the form of a sequence of dots, providing high-quality input data for subsequent integration processing.

[0071] In the integral estimation and trajectory output section, the denoised IMU data sequentially passes through two vertical modules with integral symbol patterns. The first integration integrates the denoised acceleration to obtain a velocity estimate, and the second integration integrates the velocity to obtain position and attitude estimates. Finally, the data enters the estimation result module to output the estimated trajectory.

[0072] like Figure 3 The data flow of the entire network framework is as follows: the input sequence (acceleration and angular velocity) enters the hierarchical attribute fusion module after attribute decoupling, then undergoes intra-attribute fusion and inter-attribute fusion, followed by the temporal modeling module (containing a graph attention network and a state-space model). After adaptive gating, it enters the noise prediction stage. The noise prediction result is subtracted from the original data to obtain the denoised IMU data. Finally, after two integrations, the estimated trajectory is output and compared with the actual trajectory. This end-to-end deep learning architecture, through hierarchical attribute fusion, multi-scale temporal modeling, adaptive gating, and implicit noise estimation, achieves accurate estimation and effective suppression of IMU time-varying noise without explicit noise modeling or external sensors, significantly improving the long-term positioning accuracy and robustness of inertial navigation.

[0073] Example 3 Embodiment 3 of the present invention provides an electronic device.

[0074] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. The processor includes, but is not limited to, at least one of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a tensor processor (TPU), or an artificial intelligence acceleration chip. When executing the program, the processor implements the steps in the implicit noise estimation method for an inertial measurement unit as described in Embodiment 1 of the present invention.

[0075] The detailed steps are the same as the implicit noise estimation method for inertial measurement units provided in Example 1, and will not be repeated here.

[0076] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0077] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the implicit noise estimation method for inertial measurement units as described in Embodiment 1 of the present invention.

[0078] The detailed steps are the same as the implicit noise estimation method for inertial measurement units provided in Example 1, and will not be repeated here.

[0079] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0080] A computer program product includes software code, wherein the program in the software code performs the steps of the implicit noise estimation method for inertial measurement units as described in Embodiment 1 of the present invention.

[0081] The detailed steps are the same as the implicit noise estimation method for inertial measurement units provided in Example 1, and will not be repeated here.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages. For example, in one implementation, the methods and systems can be developed based on deep learning frameworks (such as TensorFlow, PyTorch, etc.) and using the Python language. Those skilled in the art will understand that other suitable programming languages ​​or tools can also be used for implementation without departing from the core ideas of the present invention.

[0083] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for implicit noise estimation in an inertial measurement unit, characterized in that, include: Obtain the acceleration and angular velocity sequences; Axial features within the attribute are extracted from the acceleration sequence and the angular velocity sequence respectively to obtain acceleration attribute features and angular velocity attribute features; Cross-attribute feature interaction is performed on the acceleration attribute feature and the angular velocity attribute feature to obtain the fused feature; Based on the fusion features, noise estimation is performed to obtain the estimated noise; The inertial measurement data is denoised based on the estimated noise to obtain denoised inertial measurement data. The denoised inertial measurement data is integrated to obtain velocity and pose estimates.

2. The method according to claim 1, characterized in that, Noise estimation based on the fused features includes: using a graph attention network to capture local temporal dependencies of the fused features to obtain local temporal features; using a state-space model to capture long-range temporal dependencies of the fused features to obtain long-range temporal features; and predicting noise components based on the local temporal features and the long-range temporal features to obtain the estimated noise.

3. The method according to claim 1, characterized in that, Based on the fusion features, noise estimation is performed to obtain estimated noise, including: predicting accelerometer noise components and gyroscope noise components based on the fusion features to obtain predicted noise components; The inertial measurement data is denoised based on the estimated noise to obtain denoised inertial measurement data, including: subtracting the accelerometer noise component from the original acceleration measurement value and subtracting the gyroscope noise component from the original angular velocity measurement value to obtain the denoised inertial measurement data.

4. The method according to claim 2, characterized in that, Noise component prediction based on the local time-series features and the long-range time-series features includes: concatenating the local time-series features and the long-range time-series features to calculate a gated signal; and performing an adaptive weighted summation on the long-range time-series features and the local time-series features based on the gated signal to obtain a time-series representation, wherein the adaptive weighted summation is expressed as: , For the long-range time series features, For the local temporal features, The gated signal is used; noise component prediction is performed based on the time-series characterization to obtain the estimated noise.

5. The method according to claim 1, characterized in that, The method is implemented using an end-to-end neural network, which learns an implicit noise estimation function. The input to the implicit noise estimation function is the noisy acceleration measurement value. and noisy angular velocity measurement values The output is the estimated accelerometer noise component. and estimated gyroscope noise components ,satisfy ,in This refers to the set of all trainable weight parameters and bias parameters determined through training in the end-to-end neural network.

6. The method according to claim 1, characterized in that, In-attribute axial feature extraction is performed on the acceleration sequence and the angular velocity sequence, respectively, including: processing the acceleration sequence and the angular velocity sequence separately using stacked residual depth separable time series blocks, each residual depth separable time series block being represented as follows. ,in For depthwise temporal convolutions that act independently on each axis, For pointwise convolution of cross-axis mixed channel information at each time step, For batch normalization, The acceleration attribute features and the angular velocity attribute features are obtained by stacking the functions point by point as nonlinear activation functions.

7. An implicit noise estimation system for an inertial measurement unit, characterized in that, include: The data acquisition module is used to acquire acceleration and angular velocity sequences. The attribute-in-attribute feature extraction module is used to extract axial features within the attribute from the acceleration sequence and the angular velocity sequence respectively, to obtain acceleration attribute features and angular velocity attribute features; A cross-attribute feature interaction module is used to perform cross-attribute feature interaction between the acceleration attribute feature and the angular velocity attribute feature to obtain fused features; The noise estimation module is used to perform noise estimation based on the fused features to obtain estimated noise; A denoising module is used to denoise the inertial measurement data based on the estimated noise to obtain denoised inertial measurement data. The integration module is used to integrate the denoised inertial measurement data to obtain velocity and pose estimates.

8. An electronic device comprising an inertial measurement unit sensor, a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the implicit noise estimation method for inertial measurement units according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the implicit noise estimation method for inertial measurement units according to any one of claims 1 to 6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps of the implicit noise estimation method for inertial measurement units according to any one of claims 1 to 6.

Images