Gyroscope asymmetric error prediction method based on cross-scale informer neural network

Abstract

The present invention discloses a gyroscope asymmetric error prediction method based on a cross-scale informer neural network, which belongs to the field of gyroscope technology. The method first constructs a data set for preprocessing and builds a gyroscope asymmetric error prediction model. The model inputs the preprocessed data into a cross-scale attention module to output a first cross-channel interaction feature and a second cross-channel interaction feature. The first cross-channel interaction feature is input into the informer encoder, and then fused with the second cross-channel interaction feature. The cross-channel interaction feature is then point-multiplied with the preprocessed gyroscope resonator motion trajectory parameter and input into the decoder. Finally, the predicted values ​​of four asymmetric errors are obtained through a fully connected layer. The method enables the gyroscope asymmetric error prediction model to maintain high accuracy and stability when processing long sequences by designing a cross-scale informer neural network, and can accurately simulate the behavior of the gyroscope under different working conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of gyroscopes, and in particular relates to a gyroscope asymmetric error prediction method based on a cross-scale informer neural network. Background Art

[0002] As a high-precision angular velocity sensor, full-angle gyroscopes are widely used in aviation, aerospace, navigation, autonomous driving, and consumer electronics. Their core function is to provide accurate angular velocity information by detecting the rotational motion of an object in three-dimensional space. However, in practical applications, due to factors such as manufacturing processes, environmental conditions, and physical properties, full-angle gyroscopes inevitably introduce various errors. These errors directly affect the measurement accuracy and reliability of full-angle gyroscopes. Therefore, how to effectively identify and correct these errors has become a key issue in improving full-angle gyroscope performance.

[0003] Currently, traditional methods for identifying asymmetric errors in full-angle gyroscopes rely primarily on theoretical modeling and laboratory testing. This approach can provide a basic understanding of the error mechanism, but it is difficult to capture complex nonlinear and time-varying errors and requires extensive experimental data support. Traditional simulation methods are primarily used to verify and optimize the design parameters of full-angle gyroscopes, but they have the following limitations:

[0004] Traditional simulation methods are typically based on simplified mathematical models, ignoring the complex nonlinear and time-varying effects of actual systems, such as damping cross-coupling and stiffness cross-coupling. This results in a certain deviation between the gyroscope's angular velocity measurement and the actual situation. Furthermore, they often focus on static or quasi-static analysis, making it difficult to capture dynamic behavior and transient response, and thus fail to fully reflect the gyroscope's performance under actual operating conditions. Furthermore, existing simulation methods generate limited data, mostly under ideal conditions, making it difficult to meet the needs of large-scale machine learning training.

[0005] In recent years, with the development of deep learning technology, neural network-based error identification methods have gradually attracted attention. These methods train neural network models to learn gyroscope error patterns and achieve automatic identification. Common neural network architectures include multi-layer perceptrons (MLPs), convolutional neural networks (CNNs), and long short-term memory networks (LSTMs). While these methods have advantages in handling nonlinear errors, model complexity and inference speed become significant challenges due to computational resource and memory limitations.

[0006] Therefore, it is necessary to propose a gyroscope asymmetric error prediction method based on a cross-scale Informer neural network to solve the above technical problems existing in the prior art. Summary of the Invention

[0007] The purpose of the present invention is to provide a gyroscope asymmetric error prediction method based on a cross-scale informer neural network. By building a gyroscope asymmetric error prediction model, the behavior of the gyroscope under different working conditions can be accurately simulated to provide data support for subsequent error identification and correction.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The gyroscope asymmetric error prediction method based on the cross-scale informer neural network includes the following steps:

[0010] Step 1. Build a gyro resonator simulation model, set asymmetric error as input to solve the gyro resonator motion trajectory parameters, construct a data set containing asymmetric error and gyro resonator motion trajectory parameters, and preprocess the data set; the asymmetric error includes frequency difference, damping difference, stiffness coupling degree, and damping coupling degree; the gyro resonator motion trajectory parameters include the in-phase component, orthogonal component, amplitude, standing wave angle, and energy of the X mode and Y mode;

[0011] Step 2: Build a gyroscope asymmetric error prediction model including an encoder and a decoder;

[0012] Among them, the encoder includes a cross-scale attention module and an Informer encoder;

[0013] The preprocessed gyro resonator motion trajectory parameters are first grouped and reorganized, then passed through the cross-scale attention module to output the first cross-channel interaction feature and the second cross-channel interaction feature. The first cross-channel interaction feature is input into the informer encoder. The feature information processed by the informer encoder is then fused with the second cross-channel interaction feature. The fused feature is dot-multiplied with the preprocessed gyro resonator motion trajectory parameters and input into the decoder. Finally, it passes through the fully connected layer to obtain the predicted values ​​of the four asymmetric errors.

[0014] Step 3: Use the preprocessed data set in step 1 to train and evaluate the gyroscope asymmetric error prediction model in step 2, and then use the trained gyroscope asymmetric error prediction model to predict the asymmetric error of the gyroscope.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] As described above, the present invention describes a gyroscope asymmetric error prediction method based on a cross-scale informer neural network. This method uses simulation technology to highly fit the physical mechanism of the gyroscope, achieves accurate identification of the gyroscope asymmetric error (frequency difference, damping difference, stiffness coupling degree, damping coupling degree), and designs a cross-scale informer neural network, so that the gyroscope asymmetric error prediction model can maintain high accuracy and stability when processing long sequences; at the same time, through edge computing-oriented optimization, the gyroscope asymmetric error prediction model can realize real-time inference on resource-constrained devices, meeting the application scenarios of gyroscopes with high response speed requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments.

[0018] Figure 1 A physical model diagram of a gyroscope in an embodiment;

[0019] Figure 2 This is a process diagram of coherent demodulation under the condition of equal phase in the embodiment;

[0020] Figure 3 2. It is a structural diagram of the gyro resonator simulation model in the embodiment;

[0021] Figure 4 A schematic diagram of a sliding window when constructing input samples for a gyroscope asymmetric error prediction model in an embodiment;

[0022] Figure 5 Flowchart of a gyroscope asymmetric error prediction method based on a cross-scale informer neural network in an embodiment;

[0023] Figure 6 Flowchart of grouping and reorganizing the pre-processed gyro resonator motion trajectory parameters in the embodiment;

[0024] Figure 7 Flowchart of data processing in the cross-scale attention module and the informer encoder in the embodiment;

[0025] Figure 8 This is a process diagram of edge computing in an embodiment. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0027] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0028] Example 1

[0029] like Figures 1 to 7 As shown, this embodiment describes a gyroscope asymmetric error prediction method based on a cross-scale informer neural network. The method first constructs a dataset containing asymmetric errors and gyroscope resonator motion trajectory parameters and preprocesses the dataset; then builds a gyroscope asymmetric error prediction model to obtain predicted values ​​of four asymmetric errors; finally, the gyroscope asymmetric error prediction model is trained and evaluated using the preprocessed dataset, and then the trained gyroscope asymmetric error prediction model is used to predict the gyroscope asymmetric error. The specific steps are as follows:

[0030] Step 1. Build a gyro resonator simulation model, set asymmetric error as input to solve the gyro resonator motion trajectory parameters, construct a data set containing asymmetric error and gyro resonator motion trajectory parameters, and preprocess the data set; the asymmetric error includes frequency difference, damping difference, stiffness coupling degree, and damping coupling degree; the gyro resonator motion trajectory parameters include the in-phase component, orthogonal component, amplitude, standing wave angle, and energy of the X mode and Y mode.

[0031] Firstly, a mathematical model describing the dynamic behavior of the full-angle gyroscope is established based on physical principles, and four asymmetric errors are set as variables, including the frequency difference ( ), damping difference (Δζ), stiffness coupling degree ( ) and the degree of damping coupling ( ) to create different gyroscope operating environments. Coherent demodulation technology is used to extract the in-phase and quadrature components of the X and Y modes. The characterization parameter calculation module then calculates information such as the X and Y modal amplitudes, standing wave angles (orbital angles), and energy, generating rich training data for intelligent error identification.

[0032] The steps to build a gyro resonator simulation model are as follows:

[0033] Step 1.1. In an actual full-angle gyroscope, due to non-ideal factors in the manufacturing process and physical characteristics, there are frequency differences, damping differences, damping cross-coupling, and stiffness cross-coupling between the two gyroscope modes (X mode and Y mode). These problems will introduce additional error sources and affect the accuracy of the gyroscope. Specifically, an ideal axisymmetric gyroscope requires the same stiffness of the two modes, that is, the same resonant frequency. This state of the gyro is called mode matching, but in actual axisymmetric gyros there is often a frequency difference. This state of the gyro is known as modal splitting. Damping differential refers to the difference in damping coefficients between the two orthogonal modes (X and Y) in a full-angle gyro. The damping coefficient describes how quickly a system's vibrations decay, directly impacting its response speed and stability. Damping cross-coupling arises from non-proportional damping, while stiffness cross-coupling is caused by misalignment and manufacturing defects.

[0034] A mathematical model is established based on the dynamic equations of the gyroscope resonator, which is based on the following differential equations:

[0035] ;

[0036] ;

[0037] Where, is the stiffness coupling coefficient term between the X and Y modes, is the damping coupling coefficient term between the X and Y modes; where, and Represent the displacements in X mode and Y mode respectively, and Represent the damping coefficients of X mode and Y mode respectively, and Represent the stiffness coefficients of X mode and Y mode respectively, and Represent the driving forces of X mode and Y mode respectively, represents the equivalent mass, and Represent the speed in X mode and Y mode respectively, and represent the acceleration in X mode and Y mode respectively, and Represent the Coriolis force sensitivity items sensed in X mode and Y mode respectively, where is the angular gain factor, which is determined by the geometry of the gyro head and the gyro effect mode shape.

[0038] Step 1.2: In the simulation, the in-phase component and quadrature component of the X-mode and Y-mode are extracted by coherent demodulation, which are the in-phase component of the X-mode and the quadrature component of the Y-mode. , X modal quadrature components , Y mode in-phase component and the Y-mode orthogonal component Coherent demodulation not only preserves the temporal characteristics of the signal, but also improves the signal-to-noise ratio, making subsequent parameter calculations more accurate.

[0039] First, a reference signal with the same frequency as the gyro vibration signal is generated. Then, the reference signal is multiplied by the gyro vibration signal to obtain a baseband signal. Finally, a low-pass filter is used to remove the high-frequency component to obtain the X-mode in-phase component. , X modal quadrature components , Y-mode in-phase component and the Y-mode orthogonal component .

[0040] Under non-ideal conditions, the dynamic standing wave angle output solution of the full-angle mode gyroscope is:

[0041] ;

[0042] in, is the standing wave angle of the gyro resonator, represents the angular velocity of the gyroscope resonator, is the gyro angular rate gain, Represents the damping error coefficient between the X and Y modes, Represents the resonant frequency difference between the X and Y modes, represents the quadrature error of the resonator, represents the vibration energy of the gyroscope resonator, Represents the physical rotation angular rate of the Z axis.

[0043] The standing wave angle of the full-angle mode gyro resonator is calculated based on the detected gyro response signal. The output signals detected by the X and Y modes are coherently demodulated to demodulate the signals of each mode into in-phase signals (in-phase components) and orthogonal signals (orthogonal components), which are the in-phase components of the X mode and , X modal quadrature components , Y mode in-phase component and the Y-mode orthogonal component .

[0044] Step 1.3: According to the four demodulated signals in step 1.2 ( 、 、 and ) to calculate the parameters of the full-angle mode gyroscope, the parameters are as follows:

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] in, represents the oscillation energy of the gyroscope resonator, represents the orthogonal components of the gyro resonator, and They represent the X-mode component and the Y-mode component of the gyroscope resonator respectively.

[0050] According to the above analysis, a gyroscope resonator simulation model is built, such as Figure 3 As shown in FIG, the gyro resonator simulation model is divided into three parts: a gyro resonator dynamic modeling module, a coherent demodulation module, and a characterization parameter solution module.

[0051] After completing the simulation of the gyro resonator motion trajectory, extract the data to create a data set.

[0052] This example constructs a representative and easy-to-analyze gyroscope simulation data set by discretizing the parameter space, using a single / multiple error-dominated data segmentation strategy, and setting a 10-fold spacing. Multiple error dominance includes two error-dominated, three error-dominated, and four error-dominated. Based on the normal range of gyroscope parameters: frequency difference 0.001-1, damping difference 0-0.1, damping coupling gain 1×10 -9 ~1×10 -5 , stiffness coupling gain 1×10 -9 ~1×10 -5 By modifying the numerical values ​​of the four asymmetric errors, namely frequency difference, damping difference, stiffness coupling degree, and damping coupling degree, several different error combinations are generated, that is, a series of different test scenarios are created to fully cover possible error conditions.

[0053] The following describes in detail the generation process of different error combinations:

[0054] The four asymmetric errors of a gyroscope (frequency difference, damping difference, stiffness coupling, and damping coupling) may appear to have varying degrees of dominance in practical applications. "Dominance" means that under certain operating conditions, one asymmetric error significantly impacts gyroscope performance more than the others. To comprehensively cover all possible error scenarios, this embodiment permutes and combines the four asymmetric errors to generate 115 different error combinations. The specific combinations are as follows:

[0055] Single-error dominance: As shown in Table 1, error dominance is indicated by a √. Single-error dominance focuses on analyzing the impact of each asymmetric error individually. For each dominant asymmetric error, this embodiment maintains the other three asymmetric errors near the middle of their normal ranges. Then, multiple discrete values ​​are uniformly selected within the normal range of the dominant asymmetric error, spaced 10 times apart. This approach clearly demonstrates the independent impact of individual asymmetric error changes on system performance, thereby determining which asymmetric error has the greatest impact on gyroscope performance.

[0056] Table 1: Single error dominant dataset construction

[0057]

[0058] Multi-error Dominance: This examines the gyroscope's performance when multiple asymmetric errors vary simultaneously. As shown in Table 2, asymmetric errors can interact with each other, meaning that changes in one asymmetric error can affect the impact of another asymmetric error on system performance. To capture this interaction, this embodiment uses a more complex data segmentation strategy, selecting two asymmetric errors separately and selecting multiple discrete values ​​within their respective ranges at intervals of 10, thus forming a grid of asymmetric error combinations. Each combination in this grid is then simulated and the gyroscope's performance analyzed. The same applies to triple and quadruple errors.

[0059] Through this segmentation method, the impact of single error sources and mixed error sources on the system can be clearly distinguished.

[0060] Table 2: Construction of multi-error dominated gyro dataset

[0061]

[0062] Through the above combination, this embodiment generates 115 different test scenarios, each scenario corresponds to a specific error combination, and one combination includes 100,000 pieces of error data.

[0063] Then, for each set of error combinations, the corresponding data set is generated through the following steps:

[0064] (1) Setting error values: According to the dominance of asymmetric errors, set the specific values ​​of frequency difference, damping difference, stiffness coupling degree and damping coupling degree.

[0065] (2) Applying excitation signals: Apply different excitation signals to the gyro resonator, such as sinusoidal signals, step signals, and random signals, to simulate its dynamic response under different working conditions.

[0066] (3) Data collection: The sensor collects the gyro resonator motion trajectory information (gyro resonator motion trajectory parameters under different excitation signals), including the in-phase components of the X mode and the Y mode. 、 , orthogonal components 、 , amplitude, standing wave angle and energy and other information.

[0067] (4) Record data: Record the collected data as a time series, and mark the error combination and excitation signal type corresponding to each time series as the data sample of the data set.

[0068] The generated dataset primarily contains the following information: First, the input data for the gyroscope asymmetric error prediction model—the time series signal of the gyroscope resonator's motion trajectory, including the in-phase and quadrature components, amplitude, standing wave angle / orbital angle, and energy of the X and Y modes. Second, four key error parameters—frequency error, damping error, stiffness coupling, and damping coupling—are used as a comparison for the four asymmetric error prediction values ​​output by the gyroscope asymmetric error prediction model.

[0069] The following describes in detail the process of preprocessing the dataset:

[0070] (1) Data cleaning: The purpose of data cleaning is to remove noise and outliers from the original data set to ensure the accuracy and reliability of the data.

[0071] The gyro resonator motion trajectory parameters may be interfered with by high-frequency noise during the acquisition process. To remove the noise, this embodiment adopts the following method:

[0072] Low-pass filter: By setting a cutoff frequency, it filters out signal components with frequencies higher than that. For the gyroscope vibration signal, the cutoff frequency is set to 100 Hz to remove high-frequency noise.

[0073] Sliding average method: Perform sliding window averaging on the signal to smooth out the noise.

[0074] (2) Outlier detection

[0075] In order to detect and efficiently eliminate outliers in the gyroscope resonator motion trajectory parameters, this study adopts the Isolation Forest method, which can effectively identify outliers in the data.

[0076] The specific steps are as follows: First, isolation trees are constructed by randomly selecting gyroscope features and segmentation values ​​to construct multiple isolation trees. Second, anomaly scores are calculated by calculating the path length of a data point in the isolation tree. The shorter the path, the higher the anomaly score, indicating that the point is more likely to be an outlier. Finally, data points with an anomaly score above the threshold are removed based on a set threshold, ultimately obtaining a dataset free of outliers.

[0077] (3) Data normalization

[0078] The purpose of data normalization is to unify the numerical ranges of different features to accelerate the convergence of the gyroscope asymmetric error prediction model and improve training stability. Data normalization is used to scale the data to the range of [0, 1] or [-1, 1]. The collected gyroscope resonator motion trajectory parameters are normalized using the following formula:

[0079] ;

[0080] Where, represents the normalized value, Indicates the data points that need to be normalized. It represents the maximum value of all the collected gyro resonator motion trajectory parameter values. Indicates the minimum value of all the collected gyro resonator motion trajectory parameter values.

[0081] (4) Standardization

[0082] Transform the data into a distribution with mean 0 and standard deviation 1 and normalize it using the following formula:

[0083] ;

[0084] Where, represents the normalized value, Indicates the data points that need to be normalized. Indicates the standard deviation of all the collected gyro resonator motion trajectory parameter values, It represents the average value of all the collected gyro resonator motion trajectory parameter values.

[0085] After the above processing, the gyro resonator motion trajectory parameters in the original data set are converted into high-quality data suitable for the input of the gyroscope asymmetric error prediction model, including the in-phase components of the X mode and the Y mode. 、 , orthogonal components 、 , amplitude, standing wave angle / orbital angle and energy.

[0086] The data format in the dataset is: Each sample is a time series, containing feature data of multiple time steps. One sample contains the gyro resonator motion trajectory parameters of 100,000 time steps, and each time step contains 11 features (the in-phase components of the X mode and the Y mode). 、 , orthogonal components 、 , amplitude, standing wave angle / orbital angle, energy, frequency difference, damping difference, stiffness coupling degree and damping coupling degree).

[0087] To evaluate the performance of the gyroscope asymmetric error prediction model to be trained, the entire dataset was divided into three subsets, with a 70%:15%:15% distribution. The training set was used to train the gyroscope asymmetric error prediction model; the validation set was used to adjust hyperparameters to prevent overfitting; and the test set was used to ultimately evaluate the actual performance of the gyroscope asymmetric error prediction model.

[0088] Furthermore, considering the characteristics of time series data, this embodiment uses a sliding window approach to construct input samples for the gyroscope asymmetric error prediction model. Each window contains a fixed number of time steps (64 time points), and there is a certain overlap between windows (50% overlap). Assuming the time series length is T, the window size is 64, and the overlap is 50%, approximately (T64) / 32 samples are obtained. Each sample contains data from 64 time points, covering the baseband signals of the gyroscope's X and Y modes (i.e., the in-phase and quadrature components of the X and Y modes), amplitude, standing wave angle, and energy. The 50% overlap means that 32 time points overlap between two adjacent windows, ensuring that the gyroscope asymmetric error prediction model can capture continuous changes in the gyroscope time series. This approach not only preserves the dynamic characteristics of the gyroscope time series but also effectively increases the number of gyroscope training samples.

[0089] Step 2: Build a gyroscope asymmetric error prediction model. This model includes an encoder and a decoder. The encoder includes a cross-scale attention module and an informer encoder. The preprocessed gyroscope resonator motion trajectory parameters are first grouped and reorganized. Then, they pass through the cross-scale attention module to output the first cross-channel interaction feature and the second cross-channel interaction feature. The first cross-channel interaction feature is input into the informer encoder. The feature information processed by the informer encoder is then fused with the second cross-channel interaction feature. The fused feature is then dot-multiplied with the preprocessed gyroscope resonator motion trajectory parameters and input into the decoder. Finally, the four asymmetric error prediction values ​​are obtained through a fully connected layer.

[0090] The following describes in detail the processing of the pre-processed gyro resonator motion trajectory parameters in the gyroscope asymmetric error prediction model:

[0091] The input of the preprocessed gyro resonator motion trajectory parameters is (h, w, c), and then the input tensor is divided into g groups, each group has c / / g channels, and finally reorganized.

[0092] This is then fed into the cross-scale attention module in the encoder. To increase encoding capacity, the preprocessed features X from the gyro resonator trajectory parameters are grouped into 1×1 and 3×3 convolution branches. The 1×1 convolution branches undergo one-dimensional horizontal and one-dimensional vertical global pooling, followed by concatenation. The activation values ​​of the concatenated feature maps are compressed to the 0-1 range using a sigmoid function. This is then dot-producted with the preprocessed gyro resonator trajectory parameters X to preserve more of the original feature information. The group is then normalized, followed by softmax activation and average pooling, laying the foundation for the next steps of modeling and cross-fusion aggregation. The 3×3 convolution branches undergo softmax activation and average pooling.

[0093] The features of the 1×1 convolution branch after the Softmax activation function are then fused and multiplied with the features of the 3×3 convolution branch after average pooling to obtain the first cross-channel interaction feature. The features of the 1×1 convolution branch after average pooling and the features of the 3×3 convolution branch after the Softmax activation function are then fused and multiplied to obtain the second cross-channel interaction feature.

[0094] For further encoding, the first feature of cross-channel interaction is input into the Informer encoder, and multi-head sparse attention operation is performed respectively. Then, the second feature of cross-channel interaction is added to the output of the Informer encoder to obtain the fused feature. The fused feature is then point-multiplied with the pre-processed gyro resonator motion trajectory parameter X, which can retain the original feature information to a certain extent. The data after point multiplication is then input into the decoder.

[0095] After the data is input to the decoder, it undergoes joint decoding using masked multi-head probabilistic sparse self-attention and multi-head attention. The decoder gradually generates an output sequence. At each step, the masked multi-head probabilistic sparse self-attention layer processes the currently generated portion of the sequence, while the multi-head attention layer in the decoder uses the encoder's output information to guide the generation of the next output sequence. With each step, the decoder continuously generates a new output sequence and feeds it back to the masked multi-head probabilistic sparse self-attention layer for the next time step. By introducing the masking operation, information about future gyro time steps is shielded, thereby ensuring the temporal consistency of the gyro asymmetric error prediction task.

[0096] Finally, after passing through the fully connected layer, the predicted values ​​of four asymmetric errors (frequency difference, damping difference, stiffness coupling degree and damping coupling degree) are obtained.

[0097] This embodiment constructs a cross-scale gyroscope attention mechanism based on a cross-scale Informer neural network without dimensionality reduction, which can effectively retain the information of each gyroscope channel. At the same time, it combines the masked multi-head probabilistic sparse self-attention mechanism in the Informer architecture to achieve efficient processing of long gyroscope sequence data. Through the designed cross-space learning method, spatial features of different scales are integrated without losing gyroscope information, thereby improving the overall performance of the model.

[0098] Step 3: Use the preprocessed data set in step 1 to train and evaluate the gyroscope asymmetric error prediction model in step 2, and then use the trained gyroscope asymmetric error prediction model to predict the asymmetric error of the gyroscope.

[0099] After building the gyroscope asymmetric error prediction model, this embodiment trains the model using the training set in the dataset. During model training, this embodiment uses the mean squared error (MSE) as the loss function and optimizes it using the Adam optimizer. Finally, the predicted values ​​of the four asymmetric errors are regressed and compared with their actual values. The root mean square error (RMSE) is used to measure the accuracy of the model, reflecting the degree of deviation between the actual and predicted values ​​of the four asymmetric errors, and to evaluate model performance.

[0100] During the training process, this embodiment regularly evaluates the model performance on the validation set to monitor overfitting and adjust hyperparameters in a timely manner. Once the training is completed, the test set will be used to finally evaluate the performance of the model.

[0101] For each data in the test set, the trained gyroscope asymmetric error prediction model is used to predict the corresponding four asymmetric errors and compare them with the actual values. In order to quantify the accuracy of the model, this embodiment calculates the following evaluation indicators:

[0102] (1) Average accuracy: Calculate the absolute percentage error between the predicted value and the true value of the four asymmetric errors, and then calculate the average.

[0103] (2) Root mean square error (RMSE), which calculates the square root of the mean square error between the predicted value and the true value. The formula is:

[0104] ;

[0105] in, represents the total number of data points, Indicates the asymmetric error The true value of the data point, Indicates the The predicted value of a data point.

[0106] Through comprehensive analysis of these indicators, we can fully understand the performance of the gyroscope asymmetric error prediction model in the full-angle gyroscope error identification task and provide direction for subsequent optimization.

[0107] In summary, this embodiment proposes a gyroscope asymmetric error prediction method based on a cross-scale informer neural network. This method integrates an efficient cross-scale informer neural network to accurately model the complex dynamic characteristics of the full-angle gyroscope resonator motion trajectory parameters (including in-phase component, orthogonal component, amplitude, standing wave angle, and energy) in the time series, thereby obtaining a gyroscope asymmetric error prediction model.

[0108] First, the input data of the gyroscope asymmetric error prediction model consists of the gyroscope motion trajectory information (i.e., the gyroscope resonator motion trajectory parameters, including the in-phase component, quadrature component, amplitude, standing wave angle, and energy). These parameters not only reflect the current state of the gyroscope, but also imply potential asymmetric error information.

[0109] In order to effectively extract and utilize this gyro information, this embodiment designs an innovative cross-scale attention module, which can learn the effective gyro representation of each channel without channel dimensionality reduction, and establish short-range and long-range dependencies of gyro motion information through cross-space learning methods, thereby generating more precise attention.

[0110] Furthermore, this embodiment combines a cross-scale attention module with the Informer architecture to efficiently process long sequences of gyroscope data. Specifically, the cross-scale attention module reshapes some of the channel dimensions of the input feature map into batch dimensions through channel reshaping and grouping. The channel dimensions are then grouped into multiple sub-features, ensuring a uniform distribution of spatial semantic features within each feature group.

[0111] The system also incorporates a parallel sub-network design. By grouping 1×1 and 3×3 convolutional branches, it can input the first cross-channel interaction feature into the informer encoder for further processing. This feature information processed by the informer encoder is then fused with the original second cross-channel interaction feature and the preprocessed gyro resonator motion trajectory parameter X, providing rapid response capabilities. The output feature maps of the two parallel sub-networks are fused through a cross-space learning method to establish short-range and long-range dependencies.

[0112] This embodiment introduces a multi-head sparse attention mechanism in the encoder to efficiently handle long-range dependencies in gyroscope time series data. This mechanism optimizes the traditional self-attention mechanism by selecting only probabilistic sparse points with high information contribution for computation, significantly reducing computational overhead while preserving the expressiveness of key features. Adding the encoder output to the concatenated features of the activation function and average pooling layer outputs enhances the expressiveness of gyroscope time series features.

[0113] In the decoder, this embodiment employs a masked multi-head probabilistic sparse self-attention mechanism, which means that current predictions rely solely on past information. By introducing a masking operation, this mechanism blocks information from future gyroscope time steps, thereby ensuring temporal consistency in the gyroscope error prediction task. This design is particularly important for gyroscope scenarios, which require high real-time performance. It ensures the isolation of future information, allowing the model to make predictions based solely on past gyroscope information.

[0114] Furthermore, to further enhance the model's ability to capture long-range dependencies, this embodiment introduces a dependency pyramid structure into the encoder. This structure constructs a multi-layered feature representation framework by aggregating feature information at different time scales layer by layer. Specifically, the dependency pyramid extracts short-term and long-term dependencies from gyroscope time series data and organically combines them to more comprehensively reflect the dynamic behavior of the system.

[0115] This embodiment can retain the original feature information to a certain extent and prevent gradient disappearance by cascading the output of the encoder and the original input features.

[0116] After the data is input into the decoder, this embodiment jointly decodes the features output by the encoder using masked multi-head probabilistic sparse self-attention and multi-head attention, and then uses a fully connected layer to map the features to the final output dimensions, namely, four asymmetric errors (frequency difference, damping difference, stiffness coupling degree, and damping coupling degree).

[0117] Finally, after a series of transformations, the gyroscope asymmetric error prediction model outputs predicted values ​​of asymmetric errors such as frequency difference, damping difference, stiffness coupling degree, and damping coupling degree. These predicted values ​​can directly reflect the error state of the current system and provide a basis for subsequent error correction.

[0118] Since gyroscopes are usually used in embedded systems or mobile platforms, their computing resources are limited. Therefore, it is necessary to significantly reduce the computational complexity while ensuring the accuracy of the model. To this end, this embodiment adopts model pruning, quantization and other technologies, and combines them with edge computing devices (such as Jetson Nano) for efficient deployment to meet the real-time and low power consumption requirements of gyroscopes in practical applications. Figure 8 shown.

[0119] Edge computing is a distributed computing paradigm whose core concept is to shift data processing, storage, and computing capabilities from traditional centralized cloud computing (such as data centers) to locations closer to the data source (i.e., the "edge") to reduce latency, improve efficiency, and optimize the use of network resources. Because gyroscopes have relatively low computing power in practical applications, deep learning models are deployed in edge computing. This embodiment implements the following deployment and optimizations:

[0120] First, in terms of hardware selection, by leveraging the powerful hardware acceleration capabilities of the Jetson Nano edge computing device, the model can further improve inference speed and reduce power consumption while maintaining high accuracy.

[0121] Secondly, we compress the model through quantization, pruning, and knowledge distillation techniques to ensure that the model can adapt to the computing power and storage limitations of edge devices while maintaining high accuracy. The formula is as follows:

[0122] ;

[0123] Through the above optimization and deployment strategies, the gyroscope asymmetric error prediction method based on the cross-scale informer neural network can not only achieve efficient inference on the edge computing platform, but also meet the real-time and reliability requirements of full-angle gyroscope applications. It provides a powerful tool for improving the accuracy and reliability of full-angle gyroscopes, and also demonstrates the powerful potential of combining deep learning with edge computing.

[0124] This embodiment establishes a mathematical model based on the dynamic equations of the gyroscope resonator, fully considering asymmetric error factors such as damping cross-coupling, stiffness cross-coupling, and frequency difference. It can accurately simulate the behavior of the gyroscope under different working conditions, providing a solid foundation for subsequent error identification and correction. At the same time, a cross-scale informer neural network is designed to enable the gyroscope asymmetric error prediction model to maintain high accuracy and stability when processing long sequences. Moreover, through optimization for edge computing, the gyroscope asymmetric error prediction model can achieve real-time inference on resource-constrained devices, meeting the gyroscope application scenarios with high response speed requirements.

[0125] The embodiments of the present invention are only used to illustrate the technical solutions of the present invention rather than to limit the present invention. Those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A gyroscope asymmetric error prediction method based on a cross-scale informer neural network, characterized by: The method includes the following steps: Step 1. Build a gyro resonator simulation model, set asymmetric error as input to solve the gyro resonator motion trajectory parameters, construct a data set containing asymmetric error and gyro resonator motion trajectory parameters, and preprocess the data set; the asymmetric error includes frequency difference, damping difference, stiffness coupling degree, and damping coupling degree; the gyro resonator motion trajectory parameters include the in-phase component, orthogonal component, amplitude, standing wave angle, and energy of the X mode and Y mode; Step 2: Build a gyroscope asymmetric error prediction model including an encoder and a decoder; the encoder includes a cross-scale attention module and an informer encoder; The preprocessed gyro resonator motion trajectory parameters are first grouped and reorganized, then passed through the cross-scale attention module to output the first cross-channel interaction feature and the second cross-channel interaction feature. The first cross-channel interaction feature is input into the informer encoder. The feature information processed by the informer encoder is then fused with the second cross-channel interaction feature. The fused feature is dot-multiplied with the preprocessed gyro resonator motion trajectory parameters and input into the decoder. Finally, it passes through the fully connected layer to obtain the predicted values ​​of the four asymmetric errors. Among them, in the cross-scale attention module, the features are first grouped into the first convolution branch and the second convolution branch. The first convolution branch is subjected to global pooling, feature splicing and compression operations in sequence, and then point multiplication is performed with the preprocessed gyroscope resonator motion trajectory parameters and then group normalization operations are performed. Finally, activation function and average pooling are respectively processed; then the features of the first convolution branch after the activation function are fused and multiplied with the features of the second convolution branch after the average pooling to obtain the first cross-channel interaction feature; the features of the first convolution branch after the average pooling are fused and multiplied with the features of the second convolution branch after the activation function to obtain the second cross-channel interaction feature; Step 3: Use the preprocessed data set in step 1 to train and evaluate the gyroscope asymmetric error prediction model in step 2, and then use the trained gyroscope asymmetric error prediction model to predict the asymmetric error of the gyroscope.

2. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 1, characterized in that: In step 1, the steps to build the gyro resonator simulation model are as follows: Step 1.1: Build a mathematical model based on the dynamic equations of the gyro resonator. The model is based on the following differential equation: Where k yx =k xy is the stiffness coupling coefficient term between the X and Y modes, c yx =c xy is the damping coupling coefficient term between the X and Y modes; where x and y represent the displacements in the X and Y modes respectively, and c x and c y Represent the damping coefficients of X mode and Y mode respectively, k x and k y Represent the stiffness coefficients of X mode and Y mode respectively, F x and F y Represent the driving forces on X mode and Y mode respectively, m represents the equivalent mass, and Represent the speed in X mode and Y mode respectively, and represent the acceleration in X mode and Y mode respectively, and represent the Coriolis force sensitivity terms sensed in the X mode and the Y mode, respectively, where λ is the angular gain coefficient, which is determined by the geometry of the gyro head and the gyro effect vibration mode; Step 1.2: Extract the in-phase component and quadrature component of the X-mode and Y-mode through coherent demodulation, which are the in-phase component i of the X-mode and the quadrature component i of the Y-mode. x , X modal quadrature component q x , Y mode in-phase component i y and the Y-mode quadrature component q y ; The standing wave angle of the full-angle mode gyroscope resonator is calculated based on the detected gyroscope response signal. Under non-ideal conditions, the dynamic standing wave angle output solution of the full-angle mode gyroscope is: Where θ is the standing wave angle of the gyro resonator, represents the angular velocity of the gyro resonator, k is the gyro angular velocity gain, represents the damping error coefficient between the X and Y modes, Δω represents the resonant frequency difference between the X and Y modes, Q represents the orthogonal error of the resonator, E represents the vibration energy of the gyro resonator, Ω z Represents the physical rotation angular rate of the Z axis; Step 1.3: According to i in step 1.2 x ,q x 、i y and q y Calculate the parameters of the full-angle mode gyroscope.

3. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 2, characterized in that: In step 1.2, the process of extracting the in-phase and quadrature components of the X and Y modes is as follows: First, a reference signal with the same frequency as the gyro vibration signal is generated, and then the reference signal is multiplied by the gyro vibration signal to obtain a baseband signal. Finally, the high-frequency component is removed by a low-pass filter to obtain the X-mode in-phase component i x , X modal quadrature component q x , Y-mode in-phase component i y and the Y-mode quadrature component q y .

4. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 2, characterized in that: The formula for calculating the motion trajectory parameters of the gyro resonator in full-angle mode in step 1.3 is: Q=2(i x q y -i y q x ); Among them, E represents the oscillation energy of the gyro resonator, Q represents the orthogonal component of the gyro resonator, R and S represent the component of the gyro resonator in the X mode and the component in the Y mode, respectively.

5. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 1, characterized in that: The process of constructing the dataset is as follows: First, the four asymmetric errors are combined in a single-error-dominated or multi-error-dominated manner. Based on the dominance of the asymmetric errors, the specific values ​​of the four asymmetric errors are set to generate several different error combinations. Multi-error dominance includes two-error dominance, three-error dominance, and four-error dominance. Different excitation signals are applied to the gyro resonator under each error combination, and the gyro resonator motion trajectory parameters under different excitation signals are collected by sensors, including the in-phase components i of the X mode and Y mode. x 、i y , orthogonal component q x ,q y , amplitude, standing wave angle and energy; The collected data are recorded as a time series and the corresponding error combination and excitation signal type are marked as data samples of the data set.

6. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 1, characterized in that: In step 1, the process of preprocessing the data set is as follows: (1) Data cleaning to remove noise and outliers in the gyro resonator motion trajectory parameters; (2) Detect and eliminate abnormal values ​​in the gyro resonator motion trajectory parameters; (3) Data normalization, scaling the data to the range of [0, 1] or [-1, 1]; (4) Standardize the data and convert it into a distribution with a mean of 0 and a standard deviation of 1; The preprocessed dataset is then divided into training set, validation set, and test set in a ratio of 70%:15%:15%; The training set is used to train the gyroscope asymmetric error prediction model; the validation set is used to adjust hyperparameters to prevent overfitting; and the test set is used to ultimately evaluate the actual performance of the gyroscope asymmetric error prediction model. When constructing the input samples of the gyroscope asymmetric error prediction model, a sliding window method is used. Each window contains a fixed number of time steps, and there is a certain overlap between windows.

7. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 1, characterized in that: In step 2, the pre-processed gyro resonator motion trajectory parameters are processed in the cross-scale attention module as follows: First, the features of the gyro resonator motion trajectory parameters after preprocessing are grouped into the first convolution branch and the second convolution branch; the first convolution branch is a 1×1 convolution branch, and the second convolution branch is a 3×3 convolution branch; The 1×1 convolution branch is subjected to one-dimensional horizontal global pooling and one-dimensional vertical global pooling operations, followed by splicing. The activation value of the spliced ​​feature map is compressed to the range of 0-1 through the Sigmoid function, and then point multiplication is performed with the preprocessed gyro resonator motion trajectory parameters. The group is normalized, and then the Softmax activation function and average pooling operations are performed respectively. The 3×3 convolution branches are subjected to Softmax activation function and average pooling operations respectively.

8. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 1, characterized in that: In step 2, the process of obtaining the fused features from the first cross-channel interaction features is as follows: After the first cross-channel interaction feature is input into the Informer encoder, a multi-head sparse attention operation is performed on each feature. Then, the second cross-channel interaction feature is added to the output of the Informer encoder to obtain the fused feature. In addition, a dependency pyramid structure is introduced in the Informer encoder, which can extract short-term and long-term dependencies from gyroscope time series data and organically combine them to reflect the dynamic behavior of the system.

9. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 1, characterized in that: In step 2, the features after dot multiplication are processed in the decoder as follows: After the feature after the dot product is input into the decoder, the masked multi-head probabilistic sparse self-attention and multi-head attention are jointly decoded, and the decoder gradually generates the output sequence; at each step, the masked multi-head probabilistic sparse self-attention layer processes the current part of the sequence generated, and the multi-head attention layer in the decoder uses the output information of the encoder to guide the generation of the next output sequence. As each step proceeds, the decoder continuously generates a new output sequence and feeds it back to the masked multi-head probabilistic sparse self-attention layer of the next time step.

10. The gyroscope asymmetric error prediction method based on a cross-scale informer neural network according to claim 6, characterized in that: In step 3, the process of training and evaluating the gyroscope asymmetric error prediction model is as follows: During the training process of the gyroscope asymmetric error prediction model, the training set is used for training, the mean square error is used as the loss function, and the Adam optimizer is used for optimization. Finally, four asymmetric errors are predicted by regression. These are then compared with the true values ​​of the four asymmetric errors. The root mean square error is used to measure the accuracy of the gyroscope asymmetric error prediction model. During training, the performance of the gyroscope asymmetric error prediction model is regularly evaluated on the validation set to monitor overfitting and adjust hyperparameters as appropriate. After training is complete, the test set is used to finally evaluate the performance of the model. In order to quantify the accuracy of the gyroscope asymmetric error prediction model, the following evaluation indicators are calculated: (1) Calculate the absolute error percentage between the predicted value and the true value of the four asymmetric errors, and then calculate the average; (2) Calculate the square root of the mean square error between the predicted value and the true value. The formula is: Where n represents the total number of data points, y i represents the true value of the asymmetric error i-th data point, Represents the predicted value of the i-th data point.

Images