IMU gyroscope optimization method and system based on neural network

Abstract

The invention provides an IMU gyroscope optimization method and system based on a neural network, and relates to the technical field of IMU gyroscope optimization. The method comprises the following steps: firstly, acquiring a time sequence signal of the gyroscope, and processing the time sequence signal by utilizing linear filters with different bandwidths to obtain a multi-resolution signal branch; then, inputting each signal branch into an adaptive expansion convolutional network, generating a feature representation set, and calculating a cross-resolution attention weight; next, fusing the branch features by using the weights to obtain a fused feature set, inputting the fused feature set into a neural network to be optimized, and outputting the denoised angular velocity, bias estimation and attitude increment; and finally, constructing a multiple loss function based on the optimization parameters to carry out network training, and obtaining optimized angular velocity and attitude estimation by using actual measurement signals. According to the method, the problem that an IMU output signal is unstable due to the fact that attitude errors caused by time-varying bias and noise are difficult to effectively suppress in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of IMU gyroscope optimization technology, and in particular to an IMU gyroscope optimization method and system based on neural networks. Background Technology

[0002] Low-cost inertial measurement units (IMUs) are widely used in scenarios requiring high attitude stability, such as drones, wearable devices, service robots, and vehicle positioning, due to their advantages of high sampling rates and low cost. However, the gyroscope output of these IMUs is often affected by factors such as axis misalignment, time-varying biases caused by temperature and vibration, and random noise. Directly using integrated angular velocity to calculate attitude can lead to rapid drift. Therefore, traditional calibration and compensation schemes mainly target static or quasi-static errors, and have poor suppression capabilities for time-varying biases and transient non-Gaussian noise that occur during dynamic operation.

[0003] As applications expand from indoor services to highly mobile outdoor platforms, the system faces increasing demands for transient robustness in handling short-term sudden movements (such as high-speed turns and abrupt stops). Simultaneously, the resource constraints of edge computing are driving algorithms to evolve towards "efficient multi-scale perception" and "edge-deployability." Against this backdrop, deep learning methods are being widely used for gyroscope noise reduction and bias estimation. Dilated convolution and multi-scale feature fusion techniques are being widely adopted to achieve a larger temporal receptive field with lower computational cost, thereby effectively capturing error patterns at different time scales.

[0004] While existing dilated convolution-based networks have improved gyroscope signal processing capabilities to some extent, they still suffer from several significant drawbacks. First, these methods generally employ a fixed dilation rate, which decouples the temporal sampling interval from the motion state. This makes it difficult to dynamically focus key signals between low-speed, stable periods and high-speed, intense motion, increasing the risk of temporal aliasing and abrupt anomalies. Second, cross-scale fusion often uses simple splicing or fixed weighting, failing to adaptively adjust the information flow based on instantaneous motion intensity and energy distribution, resulting in insufficient generalization across different motion spectra. Furthermore, loss function design often relies on single-scale pose or trajectory supervision, neglecting the differences in the sensitivity of short, medium, and long time windows to pose errors. This prevents training from effectively directing gradients towards the more critical time windows for final pose drift. Therefore, the trade-off between accuracy and computational power, and between generalization and robustness, remains unresolved. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an IMU gyroscope optimization method and system based on neural networks. This invention solves the problem that existing technologies are unable to effectively suppress attitude errors caused by time-varying bias and noise, resulting in unstable IMU output signals.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A neural network-based IMU gyroscope optimization method includes:

[0008] Acquire the time-series signal from the gyroscope;

[0009] The time series signal is processed using multiple linear filters with different bandwidths to obtain multi-resolution signal branches;

[0010] Each multi-resolution signal branch is input into an adaptively dilated convolutional network to obtain a feature representation set;

[0011] Based on the feature representation set, calculate the corresponding cross-resolution attention weights;

[0012] The branch features of each multi-resolution signal branch are fused using the cross-resolution attention weights to obtain a fused feature set;

[0013] The fused feature set is input into the neural network to be optimized to obtain optimization parameters, wherein the optimization parameters include: denoised angular velocity, gyroscope bias estimate and attitude increment estimate;

[0014] Based on the optimization parameters, a multiple loss function is constructed to optimize and train the neural network to be optimized, thereby obtaining an optimized neural network.

[0015] The time-series signal of the gyroscope, which is currently being measured, is input into the optimized neural network to obtain the optimized angular velocity signal and attitude estimate.

[0016] Preferably, acquiring the time-series signal of the gyroscope includes:

[0017] Based on the preset measurement accuracy and application scenario, the sampling frequency, range, and filtering parameters of the gyroscope are configured to obtain the initial configuration;

[0018] Based on the initialization configuration, the gyroscope is powered on and the internal processing unit is started to obtain a usable gyroscope;

[0019] High-speed sampling technology is used to read the initial time-series signal of the available gyroscope in real time;

[0020] The initial time series signal is denoised and missing value interpolated to obtain the final time series signal.

[0021] Preferably, the step of processing the time series signal using multiple linear filters with different bandwidths to obtain multi-resolution signal branches includes:

[0022] The number and type of linear filters were determined, and the corresponding cutoff frequencies were set according to different frequency bandwidths, thus obtaining the filter configuration parameters;

[0023] Based on the filter configuration parameters, multiple linear filters are created, resulting in a usable linear filter bank, where each linear filter has different bandwidth characteristics.

[0024] Each linear filter was applied to the time series signal, and after filtering, the signal components of the corresponding frequency bands were extracted, resulting in filtered signals of multiple frequency bands.

[0025] Each filtered signal was segmented to obtain multiple signal branches with different resolutions;

[0026] The signal branches with different bandwidths are organized into structured data according to a preset format, resulting in the final multi-resolution signal branches.

[0027] Preferably, the type of the linear filter includes:

[0028] Low-pass filter, high-pass filter, and band-pass filter.

[0029] Preferably, the step of inputting each multi-resolution signal branch into an adaptively dilated convolutional network to obtain a feature representation set includes:

[0030] The architecture of an adaptive dilated convolutional network was designed, and the network structure design document was obtained. The architecture includes an input layer, multiple dilated convolutional layers, and an output layer.

[0031] Each multi-resolution signal branch is input into the adaptive dilated convolutional network one by one to obtain the network input stream;

[0032] Adaptive dilated convolutional networks process the network input stream through convolution operations and activation functions, dynamically adjusting to extract multi-scale features of the signal and obtaining intermediate representations of different signal features;

[0033] By integrating all the feature maps of the intermediate representations of different signal features, the final feature representation set is obtained.

[0034] Preferably, the calculation expression for the cross-resolution attention weight is:

[0035] ;

[0036] in, For the first Cross-resolution attention weights for each feature; For the first Feature representation of a multi-resolution signal branch; To adjust the temperature parameter that indicates the importance of the feature, This is the index for all signal branches.

[0037] Preferably, the calculation expression for the fused feature set is:

[0038] ;

[0039] in, The feature set after fusion; This represents the total number of signal branches.

[0040] Preferably, the expression for the multiple loss function is:

[0041] ;

[0042] in, This is the total loss function; For classification loss; For regression loss; For attention deficit; and Hyperparameters for adjusting the contribution of each loss term.

[0043] Preferably, the expression for the optimized neural network is:

[0044] ;

[0045] in, This is the output of the neural network; For activation functions; This is the weight matrix that connects the input features to the neural network layers; This is a bias term.

[0046] A neural network-based IMU gyroscope optimization system includes:

[0047] The raw data acquisition module is used to acquire the time-series signal from the gyroscope;

[0048] A multi-resolution signal processing module is used to process the time series signal using multiple linear filters with different bandwidths to obtain multi-resolution signal branches;

[0049] The feature extraction module is used to input each multi-resolution signal branch into the adaptive dilated convolutional network to obtain the feature representation set;

[0050] A cross-resolution attention calculation module is used to calculate the corresponding cross-resolution attention weights based on the feature representation set;

[0051] The feature fusion module is used to fuse the branch features of each multi-resolution signal branch using the cross-resolution attention weights to obtain a fused feature set;

[0052] An optimization parameter generation module is used to input the fused feature set into the neural network to be optimized to obtain optimization parameters, wherein the optimization parameters include the denoised angular velocity, gyroscope bias estimate, and attitude increment estimate.

[0053] The multiple loss function construction module is used to construct multiple loss functions based on the optimization parameters to optimize and train the neural network to obtain an optimized neural network.

[0054] The real-time signal processing module is used to input the time-series signal of the gyroscope currently being measured into the optimized neural network to obtain the optimized angular velocity signal and attitude estimate.

[0055] The present invention discloses the following technical effects:

[0056] This invention provides a neural network-based IMU gyroscope optimization method and system, comprising: acquiring a time-series signal from the gyroscope; processing the time-series signal using multiple linear filters with different bandwidths to obtain multi-resolution signal branches; inputting each multi-resolution signal branch into an adaptive dilated convolutional network to obtain a feature representation set; calculating corresponding cross-resolution attention weights based on the feature representation set; fusing the branch features of each multi-resolution signal branch using the cross-resolution attention weights to obtain a fused feature set; inputting the fused feature set into a neural network to be optimized to obtain optimization parameters, wherein the optimization parameters include: denoised angular velocity, gyroscope bias estimation, and attitude increment estimation; constructing a multiple loss function based on the optimization parameters to optimize and train the neural network to obtain an optimized neural network; and inputting the currently measured time-series signal from the gyroscope into the optimized neural network to obtain an optimized angular velocity signal and attitude estimation. This invention introduces multi-resolution signal processing technology, using linear filters with different bandwidths to more fully capture important features in time-series signals. The adaptive dilated convolutional network automatically selects the optimal feature representation, significantly improving the efficiency and accuracy of feature extraction. Furthermore, the introduction of a cross-resolution attention mechanism further optimizes the fusion process of each signal branch, ensuring the reliability of the final output's denoised angular velocity and attitude estimation. This method not only improves gyroscope performance but also provides effective support for real-time applications, better meeting the needs of modern high-precision positioning and attitude estimation, and solving key technical problems existing in the background technology. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 A flowchart of an IMU gyroscope optimization method based on a neural network is provided for an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1 As shown, this invention provides a neural network-based IMU gyroscope optimization method, comprising:

[0062] Step 100: Acquire the time-series signal from the gyroscope;

[0063] Step 200: Process the time series signal using multiple linear filters with different bandwidths to obtain multi-resolution signal branches;

[0064] Step 300: Input each multi-resolution signal branch into an adaptive dilated convolutional network to obtain a feature representation set;

[0065] Step 400: Calculate the corresponding cross-resolution attention weights based on the feature representation set;

[0066] Step 500: Use the cross-resolution attention weights to fuse the branch features of each multi-resolution signal branch to obtain a fused feature set;

[0067] Step 600: Input the fused feature set into the neural network to be optimized to obtain optimization parameters, wherein the optimization parameters include: denoised angular velocity, gyroscope bias estimate and attitude increment estimate;

[0068] Step 700: Based on the optimization parameters, construct a multiple loss function to optimize and train the neural network to be optimized, thereby obtaining an optimized neural network;

[0069] Step 800: Input the time series signal of the gyroscope currently being measured into the optimized neural network to obtain the optimized angular velocity signal and attitude estimate.

[0070] Furthermore, the specific implementation process of step 100 is as follows:

[0071] This embodiment first configures the gyroscope's sampling frequency, range, and filtering parameters according to preset measurement accuracy and application scenarios to obtain an initial configuration. This configuration step ensures that the gyroscope can operate with optimal performance under specific environmental conditions. Specifically, the selection of the sampling frequency affects the signal's timeliness and response speed, while the range determines the maximum angular velocity that the gyroscope can accurately measure. Furthermore, configuring the filtering parameters helps remove unnecessary noise interference in subsequent data processing, thereby improving signal quality and reliability.

[0072] After completing the initial configuration, this embodiment then powers on the gyroscope and activates its internal processing unit to obtain a usable gyroscope. During this process, the internal processing unit performs a self-test to ensure that all functions of the gyroscope are normal. This step is crucial because high-speed sampling technology can only be implemented after ensuring that the device is working properly. The high-speed sampling process reads the initial time-series signal of the usable gyroscope in real time to obtain dynamically changing real-time data, which provides the basis for subsequent analysis and processing.

[0073] Finally, this embodiment performs denoising and missing value imputation on the initial time series signal to obtain the final time series signal. In the denoising step, appropriate filtering techniques, such as weighted average filtering or median filtering, are used to eliminate random noise generated during the acquisition process while preserving the main features of the signal. In the missing value imputation stage, interpolation or other statistical methods are used to fill in missing data in the time series, ensuring the integrity and continuity of the data, thus laying a good foundation for subsequent feature extraction and analysis.

[0074] Furthermore, the specific implementation process of step 200 is as follows:

[0075] This embodiment first determines the number and type of linear filters, and sets corresponding cutoff frequencies according to different frequency bandwidths to obtain the filter configuration parameters. During this process, appropriate filter types, including low-pass, high-pass, and band-pass filters, are selected based on the characteristics of the time-series signal to be processed. The result is that an overall filter configuration scheme is developed for the signal characteristics of each frequency band to ensure that the final extracted signal components cover the diversity and complexity of the input data.

[0076] Next, based on the filter configuration parameters, this embodiment creates multiple linear filters with different bandwidth characteristics, forming a usable linear filter bank. The parameters of each linear filter are finely tuned to achieve optimal signal transmission characteristics. These filters are designed to maximize the retention of useful information while effectively suppressing noise when processing signals within a specific frequency range. Furthermore, the created filters can process time-series signals in parallel, thereby accelerating the entire filtering process.

[0077] After creating the filter bank, this embodiment applies each linear filter to the time-series signal and performs filtering to extract the signal components of the corresponding frequency band. By segmenting the filtered signals, multiple signal branches with different resolutions can be obtained. These signal branches reflect the characteristics of the time-series signal in different frequency ranges. Finally, the signal branches with different bandwidths are organized into structured data according to a preset format. The resulting multi-resolution signal branches can be used for subsequent feature extraction and deep learning applications, thereby further improving the efficiency and effectiveness of the overall gyroscope signal optimization.

[0078] Furthermore, the specific implementation process of step 300 is as follows:

[0079] This embodiment first designs the architecture of an adaptive dilated convolutional network, resulting in the network structure design document. This architecture includes an input layer, multiple dilated convolutional layers, and an output layer. The input layer is responsible for receiving data streams from various multi-resolution signal branches, while the dilated convolutional layers effectively capture multi-scale features of the signal by expanding the convolutional kernels while maintaining the feature receptive field. The network structure design document details the connection methods and parameter settings of each layer, laying a solid foundation for the subsequent model implementation.

[0080] After the network architecture design is completed, this embodiment inputs each multi-resolution signal branch into an adaptively dilated convolutional network to form the network input stream. The input of each signal branch first passes through the input layer and is then processed by the adaptively dilated convolutional layer. When processing these inputs, the network automatically adjusts the dilation rate of the convolutional kernel according to the different input features to ensure sensitivity to information in different frequency bands. This process not only extracts the features of the signal at different scales but also effectively reduces redundant information that may be generated during feature extraction, thereby improving the overall performance of the model.

[0081] Finally, this embodiment integrates all feature maps from the intermediate representations of different signal features to obtain the final feature representation set. In this stage, the integration process includes concatenating or weighted aggregation of the convolutional outputs of each signal feature to form a unified feature representation. These integrated feature representation sets contain rich information about the signal, providing more comprehensive feature support for subsequent learning tasks. Furthermore, this multi-level feature extraction method ensures the full capture of diverse signal features, laying the foundation for achieving an optimized neural network.

[0082] Specifically, in this embodiment, the architecture of the adaptive dilated convolutional network is designed by comprehensively considering the characteristics and complexity of multi-resolution signals. The input layer not only simply receives the data stream but also enhances the quality of the input signal through a data preprocessing module, ensuring that the features extracted from the multi-resolution signal branches have higher effectiveness in subsequent processing stages. In the design of the dilated convolutional layers, convolutional kernels with variable dilation factors are selected. This allows for dynamic adjustment of the receptive field size according to different scales of the input signal, optimizing the flexibility of feature extraction and ensuring a good balance between low-frequency and high-frequency components of the signal. This design approach enables the network to adapt to various application scenarios and meet the diverse needs of gyroscope signals.

[0083] Furthermore, the network architecture design document details the function and corresponding parameter settings of each layer to facilitate subsequent model training and optimization. The document not only lists the number and type of convolutional layers but also includes key parameters such as kernel size, stride, and padding method—all crucial factors for efficient training and improved model generalization ability. By clearly defining each stage of the network, the model construction and optimization process is systematized, thereby enhancing the network's performance in complex environments in practical applications. This innovative design ensures the accuracy and efficiency of subsequent feature extraction, guaranteeing that the final output feature representation faithfully reflects the dynamic characteristics of the gyroscope signal.

[0084] Furthermore, the specific implementation process of step 400 is as follows:

[0085] This embodiment calculates the corresponding cross-resolution attention weights based on the feature representation set to improve the applicability and importance of features from different signal branches. Specifically, the calculation method for the cross-resolution attention weights adopts a normalized exponential function form, which ensures that the role of features from different resolution signal branches in the model is optimized by adjusting the importance of the features. The feature representations here are extracted from each multi-resolution signal branch, representing the characteristics of the gyroscope signal at different processing scales.

[0086] In this computational method, the feature representations of each multi-resolution signal branch are exponentially weighted by adjusting parameters to obtain a corresponding attention weight. This adjustment parameter acts as a temperature parameter, regulating the importance and sensitivity of the feature representations. By appropriately adjusting the value of this parameter, the distribution of feature weights can be controlled. When the temperature parameter is low, the model will be more sensitive to smaller feature differences, resulting in higher attention weights; while when the temperature parameter is high, the model may tend to smooth the distribution of weights, reducing the impact of differences between features. The indices of all signal branches are used to traverse different feature representations to ensure that the features of each signal branch fairly reflect their importance during the decision-making process.

[0087] For example, setting the temperature parameter to 0.5 allows for clearer focus on differences in feature representations, with different signal branches reflecting more detailed feature proportions. This aims to enhance the model's flexibility in processing complex signals, ensuring that each multi-resolution signal branch performs optimally in the subsequent attention mechanism, thereby improving the overall performance of gyroscope signal processing. This cross-resolution attention mechanism will promote effective fusion of feature representations, contributing to improved accuracy and efficiency in subsequent data analysis.

[0088] Furthermore, the specific implementation process of step 500 is as follows:

[0089] This embodiment utilizes cross-resolution attention weights to fuse features from various multi-resolution signal branches to obtain a fused feature set. The core of this fusion process lies in multiplying the feature representation of each signal branch with its corresponding attention weight, thereby highlighting important features and suppressing secondary features. This effectively improves the expressive power of the comprehensive features. The fused feature set reflects the comprehensive feature representation of the input signal obtained after weighting across multiple resolutions, which is helpful for subsequent deep learning tasks.

[0090] In the specific implementation, the total number of signal branches to be processed is first determined, and an attention weight is assigned to each signal branch to reflect its importance. The feature representation of each signal branch is the result extracted in the previous steps, while the attention weight is derived from the calculation of the cross-resolution attention mechanism. In the fused feature set, the feature representation of each signal branch is weighted according to its attention weight, thereby achieving effective enhancement of relevant features. For example, if a feature representation is highly important, its corresponding attention weight is larger, and therefore, during the fusion process, this feature's contribution to the final fused feature set is more significant.

[0091] This weighted fusion method ensures that feature representations from different signal branches are merged into a new, richer feature set while preserving their individual characteristics. The resulting fused feature set not only contains key information from each multi-resolution signal branch but also effectively reduces redundant data, contributing to improved accuracy and capture capability of subsequent algorithms when processing gyroscope signals. This approach demonstrates the importance of feature fusion in deep learning and provides a more solid foundation for gyroscope signal optimization.

[0092] Furthermore, the specific implementation process of step 600 is as follows:

[0093] In this embodiment, the previously obtained fused feature set is input into the neural network to be optimized to calculate the optimization parameters. The key to this process is to design a reasonable neural network architecture that can effectively learn valuable information from the fused feature set. The input layer of the network receives the fused feature set, and then through multiple nonlinear transformations in the hidden layers, these features are gradually transformed into the final outputs, namely the denoised angular velocity, gyroscope bias estimate, and attitude increment estimate. These outputs will be used for precise control and state monitoring of the gyroscope.

[0094] In the design of the neural network, this embodiment adopts a multi-layer fully connected structure commonly used in deep learning. Each layer has an activation function to introduce non-linear characteristics, thereby improving the model's expressive power. Through the backpropagation algorithm, the neural network learns the relationship between different features and optimization parameters from the training data. During training, the network also uses an appropriate loss function to measure the difference between the output value and the actual observed value. Continuous iterative optimization training ensures that the network can accurately predict the denoised angular velocity, bias, and attitude increment of the gyroscope, improving measurement accuracy.

[0095] Ultimately, through the trained neural network, the optimized parameters output by the system will include the denoised angular velocity value, eliminating the interference of environmental noise on the signal and ensuring a more stable and reliable output. Simultaneously, accurate bias estimation can effectively improve the long-term stability of the gyroscope, avoiding system errors caused by device characteristics. Attitude increment estimation provides necessary dynamic feedback for real-time control, supporting subsequent navigation and positioning operations. This entire process achieves an effective transformation from fused features to specific output, providing crucial technical support for the application of gyroscope optimization methods.

[0096] Specifically, the denoising angular velocity is precisely extracted from the fused features through a neural network to eliminate background noise interference with the original signal, ensuring its stability in dynamic environments. Furthermore, the gyroscope bias estimation reflects the inherent biases of the device over long-term use and is optimized through an error compensation algorithm to improve measurement accuracy. The attitude increment estimation provides fine-grained dynamic feedback to the navigation system by updating angle changes in real time. The outputs of these optimized parameters are meticulously adjusted during network training to ensure that the system can reflect the true state of the gyroscope in real time and improve overall system performance when processing complex signals.

[0097] Furthermore, the specific implementation process of step 700 is as follows:

[0098] This embodiment constructs a multiple loss function based on previously obtained optimization parameters to optimize and train the neural network, thereby forming an optimized neural network. This multiple loss function comprehensively considers classification loss, regression loss, and attention loss to ensure the network's learning ability in multiple aspects. Classification loss evaluates the network's ability to recognize gyroscope states (such as motion patterns), regression loss measures the network's accuracy in estimating specific gyroscope parameters (such as denoised angular velocity and bias), and attention loss reflects the model's performance in effectively processing important features. These three losses work together to comprehensively optimize the network during training, achieving higher accuracy and reliability.

[0099] The weights of each loss term are determined by adjusting hyperparameters. These hyperparameters are set with prior knowledge before training begins and can be dynamically adjusted during training based on validation results. By setting appropriate hyperparameter values, the contribution of each loss term to the total loss can be controlled, thereby achieving a balance between classification accuracy, regression precision, and feature focus in the network. For example, if the weight of the classification loss is set too high, the neural network will tend to improve classification accuracy, but this may affect the optimization of the regression loss. Therefore, properly adjusting these hyperparameters helps to optimize model performance and avoid overfitting or underfitting.

[0100] By constructing multiple loss functions, this embodiment ensures that in the process of optimizing the neural network, not only is the accuracy of the output considered, but also the network's ability to effectively extract and utilize signal features is enhanced. This diversified loss function design strategy helps improve the adaptability and stability of the neural network in complex environments, thereby achieving accurate signal processing and state estimation, and providing solid technical support for the practical application of gyroscopes.

[0101] In this embodiment, the total loss function consists of three parts: category loss, regression loss, and attention loss. These parts work together to optimize the performance of the neural network. Specifically, the category loss is used to evaluate the network's performance in classification tasks, such as determining whether the current gyroscope data sample belongs to a specific motion pattern (e.g., stationary, horizontal rotation, or vertical rotation). This loss is optimized by calculating the difference between the predicted result and the true label, ensuring that the algorithm can accurately distinguish different motion states.

[0102] The regression loss is primarily used to optimize specific estimated parameters of the gyroscope, such as the denoised angular velocity and bias estimate. This loss calculates the difference between the model's output value and the actual measured value, prompting the network to learn how to predict these continuous values ​​more accurately. For example, if the network outputs a denoised angular velocity of 0.8 rad / s while the actual value is 0.75 rad / s, the regression loss will reflect the error between the two, thus guiding the network to improve its output.

[0103] Attention loss measures the appropriateness of the features a network prioritizes when processing information. This loss allows the network to better learn how to identify important information among multiple signal features, improving the representativeness and effectiveness of the overall features. For example, when processing gyroscope data, signals within a specific frequency range may be more important, and the network needs to emphasize these features through an attention mechanism. Hyperparameters λ1, λ2, and λ3 adjust the degree of influence of each component on the total loss. For example, setting λ1 to 0.5, λ2 to 0.3, and λ3 to 0.2 indicates that the network prioritizes classification accuracy, followed by regression performance, and finally the importance of the attention mechanism. This flexible loss function design allows the optimization process to consider multi-dimensional performance metrics, ensuring stable performance and accuracy of the network in complex environments.

[0104] Furthermore, the specific implementation process of step 800 is as follows:

[0105] In this embodiment, the currently measured gyroscope time-series signal is input into an optimized neural network to calculate an optimized angular velocity signal and attitude estimate. Specifically, the time-series signal includes continuous gyroscope readings, representing the dynamic changes of the device during the measurement process. By passing these signals to the optimized neural network, the model can process the input signal using features and patterns learned during previous training, thereby outputting accurate angular velocity values ​​and attitude change estimates.

[0106] In this computational process, the output of the neural network is generated by a linear transformation of an activation function through a weight matrix and bias terms. The activation function introduces non-linearity, enabling the network to learn more complex input-output relationships. For example, using activation functions like ReLU or Sigmoid can effectively handle the non-linear characteristics of signals, thereby improving the model's adaptability to different motion states. The weight matrix contains crucial parameters connecting the input features to the neural network layers. These parameters are continuously updated and adjusted during optimization using the backpropagation algorithm, ensuring that the network approximates the true output target.

[0107] Ultimately, the optimized neural network output provides an optimized angular velocity signal and attitude estimate for a given time-series signal. These outputs not only reflect the gyroscope's true state in a dynamic environment but also provide necessary data support for subsequent control and navigation systems. In this way, this embodiment ensures accurate extraction of the gyroscope's motion information, significantly improving the accuracy and real-time performance of signal processing—a key factor in enhancing gyroscope performance. The efficient execution of this step provides a reliable foundation for the real-time control and dynamic response capabilities of the entire system.

[0108] Specifically, in this embodiment, the output of the neural network is the prediction result obtained by processing the input features, specifically including the optimized angular velocity signal and attitude estimate. The activation function is an important component of a neural network, its function being to introduce nonlinear factors, enabling the network to learn and represent complex data features. For example, when using ReLU (Rectified Linear Unit) as the activation function, the output is zero when the input signal is less than zero, and directly outputs when the input signal is greater than zero, effectively improving the network's performance in handling nonlinear problems. Using, for example, Sigmoid as the activation function, its output ranges between 0 and 1, suitable for scenarios requiring probabilistic output, such as classification tasks.

[0109] The weight matrix W plays a crucial role in neural networks, connecting input features to the various layers within the network. The value of each element represents the degree of influence of the input feature on the neuron. For example, when the input features are angular velocity and attitude information, the weight matrix determines how these features are passed to the next layer and their importance in the final output. Through continuous training and optimization, the values ​​of the weight matrix elements are gradually adjusted to minimize the error between the network output and the true values, thereby ensuring that the neural network can accurately predict angular velocity and attitude.

[0110] The bias term 'b' is used to further adjust the neural network's output, increasing the model's flexibility. It is a constant value, typically added to the output of each neuron, allowing the model to be biased against the input of the activation function. For example, in practical applications, if the output signal of a sensor is consistently low when passed to the network, introducing a bias term can help the model correct this bias during training, thereby improving the final output. Therefore, the synergistic effect of the weight matrix and the bias term, combined with the nonlinear characteristics of the activation function, enables the neural network to accurately capture the complex relationships in the input data, ultimately providing optimized angular velocity signals and attitude estimates.

[0111] This embodiment also provides an IMU gyroscope optimization system based on a neural network, including:

[0112] The raw data acquisition module is used to acquire the time-series signal from the gyroscope;

[0113] A multi-resolution signal processing module is used to process the time series signal using multiple linear filters with different bandwidths to obtain multi-resolution signal branches;

[0114] The feature extraction module is used to input each multi-resolution signal branch into the adaptive dilated convolutional network to obtain the feature representation set;

[0115] A cross-resolution attention calculation module is used to calculate the corresponding cross-resolution attention weights based on the feature representation set;

[0116] The feature fusion module is used to fuse the branch features of each multi-resolution signal branch using the cross-resolution attention weights to obtain a fused feature set;

[0117] An optimization parameter generation module is used to input the fused feature set into the neural network to be optimized to obtain optimization parameters, wherein the optimization parameters include the denoised angular velocity, gyroscope bias estimate, and attitude increment estimate.

[0118] The multiple loss function construction module is used to construct multiple loss functions based on the optimization parameters to optimize and train the neural network to obtain an optimized neural network.

[0119] The real-time signal processing module is used to input the time-series signal of the gyroscope currently being measured into the optimized neural network to obtain the optimized angular velocity signal and attitude estimate.

[0120] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0121] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A neural network-based IMU gyroscope optimization method, characterized in that, include: Acquire the time-series signal from the gyroscope; The time series signal is processed using multiple linear filters with different bandwidths to obtain multi-resolution signal branches; Each multi-resolution signal branch is input into an adaptively dilated convolutional network to obtain a feature representation set; Based on the feature representation set, calculate the corresponding cross-resolution attention weights; The branch features of each multi-resolution signal branch are fused using the cross-resolution attention weights to obtain a fused feature set; The fused feature set is input into the neural network to be optimized to obtain optimization parameters, wherein the optimization parameters include: denoised angular velocity, gyroscope bias estimate and attitude increment estimate; Based on the optimization parameters, a multiple loss function is constructed to optimize and train the neural network to be optimized, thereby obtaining an optimized neural network. The time-series signal of the gyroscope, which is currently being measured, is input into the optimized neural network to obtain the optimized angular velocity signal and attitude estimate.

2. The IMU gyroscope optimization method based on neural networks according to claim 1, characterized in that, The acquisition of the time-series signal from the gyroscope includes: Based on the preset measurement accuracy and application scenario, the sampling frequency, range, and filtering parameters of the gyroscope are configured to obtain the initial configuration; Based on the initialization configuration, the gyroscope is powered on and the internal processing unit is started to obtain a usable gyroscope; High-speed sampling technology is used to read the initial time-series signal of the available gyroscope in real time; The initial time series signal is denoised and missing value interpolated to obtain the final time series signal.

3. The IMU gyroscope optimization method based on neural networks according to claim 1, characterized in that, The process of using multiple linear filters with different bandwidths to process the time series signal to obtain multi-resolution signal branches includes: The number and type of linear filters were determined, and the corresponding cutoff frequencies were set according to different frequency bandwidths, thus obtaining the filter configuration parameters; Based on the filter configuration parameters, multiple linear filters are created, resulting in a usable linear filter bank, where each linear filter has different bandwidth characteristics. Each linear filter was applied to the time series signal, and after filtering, the signal components of the corresponding frequency bands were extracted, resulting in filtered signals of multiple frequency bands. Each filtered signal was segmented to obtain multiple signal branches with different resolutions; The signal branches with different bandwidths are organized into structured data according to a preset format, resulting in the final multi-resolution signal branches.

4. The IMU gyroscope optimization method based on neural networks according to claim 3, characterized in that, The types of linear filters include: Low-pass filter, high-pass filter, and band-pass filter.

5. The IMU gyroscope optimization method based on neural networks according to claim 1, characterized in that, The process involves inputting each multi-resolution signal branch into an adaptively dilated convolutional network to obtain a feature representation set. include: The architecture of an adaptive dilated convolutional network was designed, and the network structure design document was obtained. The architecture includes an input layer, multiple dilated convolutional layers, and an output layer. Each multi-resolution signal branch is input into the adaptive dilated convolutional network one by one to obtain the network input stream; Adaptive dilated convolutional networks process the network input stream through convolution operations and activation functions, dynamically adjusting to extract multi-scale features of the signal and obtaining intermediate representations of different signal features; By integrating all the feature maps of the intermediate representations of different signal features, the final feature representation set is obtained.

6. The IMU gyroscope optimization method based on neural networks according to claim 1, characterized in that, The expression for calculating the cross-resolution attention weights is: ； in, For the first Cross-resolution attention weights for each feature; For the first Characteristic representation of a multi-resolution signal branch; To adjust the temperature parameter that indicates the importance of the feature, This is the index for all signal branches.

7. The IMU gyroscope optimization method based on neural networks according to claim 6, characterized in that, The calculation expression for the fused feature set is: ； in, The feature set after fusion; This represents the total number of signal branches.

8. The IMU gyroscope optimization method based on neural networks according to claim 1, characterized in that, The expression for the multiple loss function is: ； in, This is the total loss function; For classification loss; For regression loss; For attention deficit; and Hyperparameters for adjusting the contribution of each loss term.

9. The IMU gyroscope optimization method based on neural networks according to claim 8, characterized in that, The expression for the optimized neural network is: ； in, This is the output of the neural network; For activation functions; This is the weight matrix that connects the input features to the neural network layers; This is a bias term.

10. A neural network-based IMU gyroscope optimization system, characterized in that, include: The raw data acquisition module is used to acquire the time-series signal from the gyroscope; A multi-resolution signal processing module is used to process the time series signal using multiple linear filters with different bandwidths to obtain multi-resolution signal branches; The feature extraction module is used to input each multi-resolution signal branch into the adaptive dilated convolutional network to obtain the feature representation set; A cross-resolution attention calculation module is used to calculate the corresponding cross-resolution attention weights based on the feature representation set; The feature fusion module is used to fuse the branch features of each multi-resolution signal branch using the cross-resolution attention weights to obtain a fused feature set; An optimization parameter generation module is used to input the fused feature set into the neural network to be optimized to obtain optimization parameters, wherein the optimization parameters include the denoised angular velocity, gyroscope bias estimate, and attitude increment estimate. The multiple loss function construction module is used to construct multiple loss functions based on the optimization parameters to optimize and train the neural network to obtain an optimized neural network. The real-time signal processing module is used to input the time-series signal of the gyroscope currently being measured into the optimized neural network to obtain the optimized angular velocity signal and attitude estimate.

Images