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

Abstract

The invention discloses a gyroscope asymmetric error prediction method based on a cross-scale Informer neural network, and belongs to the technical field of gyroscopes. The method comprises the following steps: firstly, constructing a data set for preprocessing, establishing a gyroscope asymmetric error prediction model, inputting preprocessed data into a cross-scale attention module by the model so as to output a first cross-channel interaction feature and a second cross-channel interaction feature, inputting the first cross-channel interaction feature into an Informer encoder, and inputting the second cross-channel interaction feature into the Informer encoder; and fusing with the second cross-channel interaction feature, performing point multiplication with the preprocessed motion track parameter of the gyro harmonic oscillator, inputting into a decoder, and finally obtaining predicted values of four asymmetric errors through a full connection layer. According to the method, the cross-scale Informer neural network is designed, so that the asymmetric error prediction model of the gyroscope keeps relatively high precision and stability when processing a long sequence, and behaviors of the gyroscope under different working conditions can be accurately simulated.

Description

Technical Field

[0001] The present invention belongs to the technical field of gyroscopes, and particularly relates to a method for predicting the asymmetric error of a gyroscope based on a cross-scale Informer neural network. Background Art

[0002] As a high-precision angular velocity sensor, the full-angle gyroscope is widely used in fields such as aviation, aerospace, navigation, autonomous driving, and consumer electronics. Its 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 characteristics, the full-angle gyroscope inevitably introduces various errors, which directly affect the measurement accuracy and reliability of the full-angle gyroscope. Therefore, how to effectively identify and correct these errors has become a key issue in improving the performance of the full-angle gyroscope.

[0003] Currently, traditional methods for identifying the asymmetric error of a full-angle gyroscope mainly rely on theoretical modeling and laboratory tests. This method can provide a basic understanding of the error mechanism, but it is difficult to capture complex non-linear and time-varying errors, and requires a large amount of experimental data support. Traditional simulation methods are mainly used to verify and optimize the design parameters of the full-angle gyroscope, but they have the following limitations: Traditional simulation methods usually rely on simplified mathematical models, ignoring complex non-linear and time-varying effects in the actual system, such as damping cross-coupling and stiffness cross-coupling. This leads to a certain deviation between the measured angular velocity of the gyro and the actual situation, and mostly focuses on static or quasi-static analysis, making it difficult to capture dynamic behavior and transient response, and unable to fully reflect the performance of the gyroscope under actual working conditions. Secondly, the data volume generated by existing simulation methods is limited, and most of the data is under ideal conditions, making it difficult to meet the needs of large-scale machine learning training.

[0004] In recent years, with the development of deep learning technology, error identification methods based on neural networks have gradually attracted attention. These methods learn the error patterns of the gyroscope by training neural network models to achieve automatic identification. Common neural network structures include multi-layer perceptron (MLP), convolutional neural network (CNN), and long short-term memory network (LSTM). Although these methods have advantages in dealing with non-linear errors, due to the limitations of computing resources and memory, the complexity and inference speed of the models have become important challenges.

[0005] Therefore, it is necessary to propose a method for predicting the asymmetric error of a gyroscope based on a cross-scale Informer neural network to solve the above technical problems existing in the prior art. Summary of the Invention

[0006] The object 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, providing data support for subsequent error identification and correction.

[0007] To achieve the above object, the present invention provides the following technical solutions: A gyroscope asymmetric error prediction method based on a cross-scale Informer neural network, comprising the following steps: Step 1: Build a gyroscope resonator simulation model, set the asymmetric error as the input, solve the gyroscope resonator motion trajectory parameters, construct a data set containing the asymmetric error and the gyroscope resonator motion trajectory parameters, and preprocess the data set; the asymmetric error includes frequency difference, damping difference, stiffness coupling degree and damping coupling degree, and the gyroscope resonator motion trajectory parameters include in-phase components, quadrature components, amplitudes, standing wave angles and energies of the X mode and the Y mode; Step 2: Build a gyroscope asymmetric error prediction model including an encoder and a decoder; Among them, the encoder includes a cross-scale attention module and an Informer encoder; The preprocessed gyroscope resonator motion trajectory parameters first undergo grouping and recombination operations, and then 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. Then, the feature information processed by the Informer encoder is fused with the second cross-channel interaction feature. The fused feature is dot-multiplied with the preprocessed gyroscope resonator motion trajectory parameters and input into the decoder. Finally, the predicted values of the four asymmetric errors are obtained through a fully connected layer; 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.

[0008] Compared with the prior art, the present invention has the following beneficial effects: As described above, the present invention relates to a method for predicting the asymmetric error of a gyroscope based on a cross-scale Informer neural network. This method highly fits the gyro physical mechanism through simulation technology, realizes the accurate identification of the asymmetric error of the gyroscope (frequency difference, damping difference, stiffness coupling degree, damping coupling degree), and designs a cross-scale Informer neural network, enabling the gyroscope asymmetric error prediction model to maintain high accuracy and stability when processing long sequences; at the same time, through optimization for edge computing, the gyroscope asymmetric error prediction model can perform real-time inference on resource-constrained devices, meeting the application scenarios of gyroscopes with high requirements for response speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments.

[0010] Figure 1 It is the physical model diagram of the gyroscope in the embodiment; Figure 2 It is the process diagram of coherent demodulation under the condition of equal phase in the embodiment; Figure 3 It is the structural diagram of the gyroscope resonator simulation model in the embodiment; Figure 4 It is the schematic diagram of the sliding window when constructing the input samples of the gyroscope asymmetric error prediction model in the embodiment; Figure 5 It is the flowchart of the method for predicting the asymmetric error of the gyroscope based on the cross-scale Informer neural network in the embodiment; Figure 6 It is the flowchart of the grouping and recombination operations on the preprocessed gyroscope resonator motion trajectory parameters in the embodiment; Figure 7 It is the flowchart of the data processing in the cross-scale attention module and the Informer encoder in the embodiment; Figure 8 It is the process diagram of edge computing in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0012] Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0013] Embodiment 1 AsFigures 1 to 7 As shown, this embodiment describes a method for predicting the asymmetric error of a gyroscope based on a cross-scale Informer neural network. The method first constructs a data set containing asymmetric errors and the motion trajectory parameters of the gyro resonator, and preprocesses the data set; then builds a gyroscope asymmetric error prediction model to obtain the predicted values of four asymmetric errors; finally, uses the preprocessed data set to train and evaluate the gyroscope asymmetric error prediction model, and then uses the trained gyroscope asymmetric error prediction model to predict the asymmetric error of the gyroscope. The specific steps are as follows: Step 1: Build a gyro resonator simulation model, set the asymmetric error as the input, solve the motion trajectory parameters of the gyro resonator, construct a data set containing asymmetric errors and the motion trajectory parameters of the gyro resonator, and preprocess the data set; the asymmetric errors include frequency difference, damping difference, stiffness coupling degree and damping coupling degree, and the motion trajectory parameters of the gyro resonator include the in-phase component, quadrature component, amplitude, standing wave angle and energy of the X mode and Y mode.

[0014] First, establish a mathematical model describing the dynamic behavior of the full-angle gyro through physical principles, and set four asymmetric errors as variables, including frequency difference ( ), damping difference (Δζ), stiffness coupling degree ( ), and damping coupling degree ( ), to construct different gyroscope working environments. Through coherent demodulation technology, extract the in-phase component and quadrature component of the X mode and Y mode, and then solve information such as the amplitude of the X mode, the amplitude of the Y mode, the standing wave angle (orbital angle), and energy through the characterization parameter calculation module, so as to generate rich training data for intelligent error identification.

[0015] The steps to build the gyro resonator simulation model are as follows: Step 1.1: In an actual full-angle gyro, due to non-ideal factors and physical characteristics during the manufacturing process, there are problems of frequency difference, damping difference, damping cross-coupling and stiffness cross-coupling between the two gyro modes (X mode and Y mode). These problems will introduce additional error sources and affect the accuracy of the gyroscope. Specifically, an ideal axisymmetric gyro requires the same stiffness of the two modes, that is, the same resonance frequency , and the gyro in this state is called mode matching, while in an actual axisymmetric gyro, there is often a frequency difference , and the gyro in this state is called mode splitting. The damping difference refers to the difference in damping coefficients between the two orthogonal modes (X mode and Y mode) in the full-angle gyro. The damping coefficient is a parameter describing the speed of vibration decay of the system, and it directly affects the response speed and stability of the system; the appearance of damping cross-coupling is caused by the non-proportional damping problem, while the stiffness cross-coupling is caused by misalignment and manufacturing defects.

[0016] A mathematical model is established based on the dynamic equation of the gyro resonator, and this model is based on the following differential equation: ; ; In the formula, is the stiffness coupling coefficient term between the X and Y modes, is the damping coupling coefficient term between the X and Y modes; among them, and respectively represent the displacements on the X mode and the Y mode, and respectively represent the damping coefficients of the X mode and the Y mode, and respectively represent the stiffness coefficients on the X mode and the Y mode, and respectively represent the driving forces on the X mode and the Y mode, represents the equivalent mass, and respectively represent the velocities on the X mode and the Y mode, and respectively represent the accelerations on the X mode and the Y mode, and respectively represent the Coriolis force sensitive terms sensed on the X mode and the Y mode, where is the angular gain coefficient, which is determined by the geometric shape of the gyro head and the vibration mode of the gyro effect.

[0017] Step 1.2: Extract the in-phase component and the quadrature component of the X mode and the Y mode through coherent demodulation in the simulation, which are the in-phase component of the X mode, the quadrature component of the X mode, the in-phase component of the Y mode, and the quadrature component of the Y mode. Coherent demodulation not only retains the time characteristics of the signal but also improves the signal-to-noise ratio, making the subsequent parameter calculation more accurate.

[0018] 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 the baseband signal, and finally, the high-frequency components are removed through a low-pass filter to obtain the in-phase component of the X mode, the quadrature component of the X mode, the in-phase component of the Y mode, and the quadrature component of the Y mode.

[0019] Under non-ideal conditions, the dynamic standing wave angle output solution of the full-angle mode gyroscope is: ; Among them, is the standing wave angle of the gyro resonator, represents the angular rate of the gyro 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 amount of the resonator, represents the vibration energy of the gyro resonator, represents the physical rotation angular rate of the Z axis.

[0020] The standing wave angle of the full-angle mode gyro resonator is calculated according to the detected gyro response signal. The output signals detected by the X and Y modes are demodulated by coherent demodulation to demodulate the signals of each mode into in-phase signals (in-phase components) and quadrature signals (quadrature components), which are the in-phase component of the X mode , the quadrature component of the X mode , the in-phase component of the Y mode and the quadrature component of the Y mode .

[0021] Step 1.3. According to the four demodulated signals ( , , and ) in Step 1.2, perform parameter calculation of the full-angle mode gyroscope. The parameters are as follows: ; ; ; ; Among them, represents the oscillation energy of the gyro resonator, represents the quadrature component of the gyro resonator, and respectively represent the components of the gyro resonator in the X mode and the Y mode.

[0022] Based on the above analysis, build a gyro resonator simulation model. As Figure 3 shown, the gyro resonator simulation model is divided into three parts: the gyro resonator dynamics modeling module, the coherent demodulation module, and the characterization parameter calculation module.

[0023] After completing the gyro resonator motion trajectory simulation, extract data to make a dataset.

[0024] In this embodiment, a representative and easy-to-analyze gyroscope simulation data set is constructed through parameter space discretization, single / multi-error-dominated data cutting strategies, and a 10-fold spacing setting. Among them, multi-error dominance includes two-error dominance, three-error dominance, and four-error dominance. According to the normal ranges of gyro 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 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 comprehensively cover possible error situations.

[0025] The generation process of different error combinations is described in detail below: The four asymmetric errors (frequency difference, damping difference, stiffness coupling degree, damping coupling degree) of the gyroscope may appear with different dominance in practical applications. The so-called "dominance" means that under a certain specific working condition, one type of asymmetric error has a significantly greater impact on the performance of the gyroscope than other asymmetric errors. To comprehensively cover possible error situations, this embodiment performs permutations and combinations on the four asymmetric errors to generate 115 different error combinations. The specific combination methods are as follows: Single-error dominance: As shown in Table 1, √ is used to indicate the use of error dominance. Single-error dominance focuses on separately analyzing the influence of each asymmetric error. For each dominant asymmetric error, this embodiment will keep the other three asymmetric errors near the middle value of their normal ranges, and then uniformly select multiple discrete values within the normal range of the dominant asymmetric error at a 10-fold spacing. In this way, the independent influence of the change of a single asymmetric error on the system performance can be clearly observed, so as to determine which asymmetric error has the greatest impact on the performance of the gyroscope.

[0026] Table 1: Construction of single-error-dominated data set

[0027] Multi-error dominance: Examine the performance of the gyroscope when multiple asymmetric errors change simultaneously. As shown in Table 2, there may be an interaction between asymmetric errors, that is, the change of one asymmetric error will affect the influence of another asymmetric error on the system performance. To capture this interaction, this embodiment uses a more complex data cutting strategy, selects two asymmetric errors respectively, and selects multiple discrete values at a 10-fold spacing within their respective ranges, thus forming a grid of asymmetric error combinations; then, simulate each combination in this grid and analyze the performance of the gyroscope. The same applies to three errors and four errors.

[0028] By this segmentation method, the impacts of single error sources and mixed error sources on the system can be clearly distinguished.

[0029] Table 2: Construction of Multi-Error-Dominated Gyroscope Dataset

[0030] Through the above combination method, 115 different test scenarios are generated in this embodiment. Each scenario corresponds to a specific error combination, and one combination includes 100,000 error data.

[0031] Then, for each group of error combinations, the corresponding dataset is generated through the following steps: (1) Set error values: According to the dominance of the asymmetric error, set the specific values of the frequency difference, damping difference, stiffness coupling degree, and damping coupling degree.

[0032] (2) Apply excitation signals: Apply different excitation signals, such as sine signals, step signals, and random signals, etc., to the gyroscope resonator to simulate its dynamic response under different working conditions.

[0033] (3) Collect data: Collect the motion trajectory information of the gyroscope resonator (the motion trajectory parameters of the gyroscope resonator under different excitation signals) through sensors, including the in-phase components of the X mode and Y mode , , quadrature components , , amplitude, standing wave angle, and energy and other information.

[0034] (4) Record data: Record the collected data as a time series, and label the error combination and excitation signal type corresponding to each time series as the data samples of the dataset.

[0035] The generated dataset mainly contains the following information: First, it is the input data of the gyroscope asymmetric error prediction model, that is, the time series signal of the motion trajectory of the gyroscope resonator, including the in-phase components, quadrature components, amplitude, standing wave angle / orbital angle, and energy of the X mode and Y mode; second, it is four key error parameters, including the frequency difference, damping difference, stiffness coupling degree, and damping coupling degree, which are used as the comparison for the four asymmetric error prediction values output by the gyroscope asymmetric error prediction model.

[0036] The following details the process of preprocessing the dataset: (1) Data cleaning: The purpose of data cleaning is to remove the noise and outliers in the original dataset to ensure the accuracy and reliability of the data.

[0037] The trajectory parameters of the gyro resonator may be disturbed by high-frequency noise during the acquisition process. To remove the noise, the following method is adopted in this embodiment: Low-pass filter: By setting the cut-off frequency, the components above this frequency in the signal are filtered out. For the vibration signal of the gyroscope, the cut-off frequency is set to 100 Hz to remove high-frequency noise.

[0038] Moving average method: The signal is processed by moving window averaging to smooth the noise.

[0039] (2) Outlier detection To detect and efficiently remove outliers in the trajectory parameters of the gyro resonator, the Isolation Forest method is adopted in this study, which can effectively identify the outliers in the data.

[0040] The specific steps are as follows: First, construct isolation trees by randomly selecting the features and splitting values of the gyroscope to construct multiple isolation trees; second, calculate the outlier scores by calculating the path lengths of the data points in the isolation trees to obtain the outlier scores. The shorter the path, the higher the outlier score, indicating that the point is more likely to be an outlier; finally, according to the set threshold, the data points with outlier scores higher than the threshold are removed, and finally a data set without outliers is obtained.

[0041] (3) Data normalization 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 the training stability. The data is scaled to the range of [0, 1] or [-1, 1] through data normalization. For the acquired trajectory parameters of the gyro resonator, the following formula is used for normalization: ; In the formula, represents the normalized value, represents the data point to be normalized, represents the maximum value among all the acquired trajectory parameter values of the gyro resonator, represents the minimum value among all the acquired trajectory parameter values of the gyro resonator.

[0042] (4) Standardization The data is converted into a distribution with a mean of 0 and a standard deviation of 1, and the following formula is used for standardization: ; In the formula, represents the standardized value, represents the data point to be standardized, represents the standard deviation of all the acquired trajectory parameter values of the gyro resonator, Represents the average value of all the gyro resonator motion trajectory parameter values collected.

[0043] After the above processing, the gyro resonator motion trajectory parameters in the original dataset are transformed into high-quality data suitable for input to the gyroscope asymmetric error prediction model, including the in-phase components of the X-mode and Y-mode , , the quadrature components , , amplitude, standing wave angle / orbital angle, and energy.

[0044] Data format in the dataset: Each sample is a time series containing feature data for multiple time steps. One sample contains the gyro resonator motion trajectory parameters for 100,000 time steps, and each time step contains 11 features (the in-phase components of the X-mode and Y-mode , , the quadrature components , , amplitude, standing wave angle / orbital angle, energy, frequency difference, damping difference, degree of stiffness coupling, and degree of damping coupling).

[0045] To evaluate the performance of the gyroscope asymmetric error prediction model to be trained, the entire dataset is divided into a training set, a validation set, and a test set, with the three subsets allocated in a ratio of 70%:15%:15% respectively. The training set is used to train the gyroscope asymmetric error prediction model; the validation set is used to adjust the hyperparameters to prevent overfitting; and the test set is used to finally evaluate the true performance of the gyroscope asymmetric error prediction model.

[0046] In addition, considering the characteristics of time series data, this embodiment uses the sliding window method to construct the 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 the windows (an overlap rate of 50%). Assuming the length of the time series is T, the window size is 64, and the overlap rate is 50%, then approximately (T - 64) / 32 samples can be obtained. Each sample contains data for 64 time points, covering the baseband signals of the X-mode and Y-mode of the gyroscope (i.e., the in-phase components and quadrature components of the X-mode and Y-mode), amplitude, standing wave angle, and energy information. The 50% overlap rate means that there are 32 time points overlapping between two adjacent windows, which can ensure that the gyroscope asymmetric error prediction model can capture the continuous changes in the gyro time series. This method can not only retain the dynamic characteristics of the gyro time series but also effectively increase the number of gyro training samples.

[0047] Step 2: Build a gyroscope asymmetric error prediction model. The gyroscope asymmetric error prediction model includes an encoder and a decoder. The encoder includes a cross-scale attention module and an Informer encoder. The preprocessed gyro resonator motion trajectory parameters first undergo grouping and recombination operations, and then 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. Then, the feature information processed by the Informer encoder is 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, the predicted values of four asymmetric errors are obtained through a fully connected layer.

[0048] The following details the processing process of the preprocessed gyro resonator motion trajectory parameters in the gyroscope asymmetric error prediction model: The input of the preprocessed gyro resonator motion trajectory parameters is (h, w, c). Then, the input tensor is divided into g groups, with each group having c / / g channels, and finally recombination is performed.

[0049] Then it is input into the cross-scale attention module in the encoder. To increase the encoding ability, the feature X in the preprocessed gyro resonator motion trajectory parameters is grouped into a 1×1 convolution branch and a 3×3 convolution branch. One-dimensional horizontal global pooling operation and one-dimensional vertical global pooling operation are respectively performed on the 1×1 convolution branch, and then a concatenation operation is carried out. Then, the activation values of the concatenated feature map are compressed to the range of 0-1 through the Sigmoid function, and then dot-multiplied with the preprocessed gyro resonator motion trajectory parameter X to retain more original feature information. Then, group normalization operation is performed, and then the Softmax activation function and average pooling operation are respectively carried out, laying a foundation for the next step of modeling and aggregating cross-fusion. The 3×3 convolution branch respectively undergoes the Softmax activation function and average pooling operation.

[0050] Then, the feature of the 1×1 convolution branch after passing through the Softmax activation function and the feature of the 3×3 convolution branch after passing through the average pooling operation are fused and multiplied to obtain the first cross-channel interaction feature. The feature of the 1×1 convolution branch after passing through the average pooling operation and the feature of the 3×3 convolution branch after passing through the Softmax activation function are fused and multiplied to obtain the second cross-channel interaction feature.

[0051] For further encoding, the cross-channel interaction first feature is input into the Informer encoder, and multi-head sparse attention operations are performed respectively. Then, the cross-channel interaction second feature is added to the output of the Informer encoder to obtain the fused feature. Next, the fused feature is dot-multiplied with the preprocessed gyro resonator motion trajectory parameter X, which can preserve the original feature information to a certain extent. Then, the data after dot-multiplication is input into the decoder.

[0052] After the data is input into the decoder, 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 currently generated partial sequence, 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 progresses, the decoder continuously generates new output sequences and feeds them back to the masked multi-head probabilistic sparse self-attention layer at the next time step. By introducing the masking operation, the information of future gyro time steps is masked, thus ensuring the temporal consistency of the gyro asymmetric error prediction task.

[0053] Finally, 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.

[0054] In this embodiment, by designing a cross-scale Informer neural network, a cross-scale gyro attention mechanism without dimensionality reduction is constructed, which can effectively retain the information of each gyro channel. At the same time, by combining the masked multi-head probabilistic sparse self-attention mechanism in the Informer architecture, the efficient processing of long-sequence gyro data is realized. Through the designed cross-space learning method, spatial features of different scales are fused without losing gyro information, thus improving the overall performance of the model.

[0055] 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.

[0056] After building the gyroscope asymmetric error prediction model, this embodiment uses the training set in the data set to train the gyroscope asymmetric error prediction model. During the model training process, this embodiment uses the mean square error (MSE) as the loss function and optimizes it through the Adam optimizer. Finally, the predicted values of four asymmetric errors are regressively predicted and compared with the true values of the four asymmetric errors. At the same time, the root mean square error (RMSE) is used to measure the accuracy of the model, reflecting the deviation degree between the actual values and the predicted values of the four asymmetric errors, and evaluating the model performance.

[0057] During the training process of this embodiment, the model performance is evaluated regularly on the validation set to monitor overfitting and adjust the hyperparameters in a timely manner. Once the training is completed, the test set will be used to finally evaluate the performance of the model.

[0058] 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. To quantify the accuracy of the model, this embodiment calculates the following evaluation metrics: (1)Average accuracy: Calculate the percentage of the absolute error between the predicted value and the true value of the four asymmetric errors, and then take the average.

[0059] (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: ; where, represents the total number of data points, represents the true value of the -th data point of the asymmetric error, represents the -th data point of the predicted value.

[0060] Through the comprehensive analysis of these metrics, the performance of the gyroscope asymmetric error prediction model in the full-angle gyro error identification task can be comprehensively understood, and directions for subsequent optimization can be provided.

[0061] In summary, this embodiment proposes a method for predicting gyroscope asymmetric errors 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 gyro resonator motion trajectory parameters (including in-phase component, quadrature component, amplitude, standing wave angle, and energy) in the time series, and obtains a gyroscope asymmetric error prediction model.

[0062] First, the input data of the gyroscope asymmetric error prediction model consists of gyro motion trajectory information (i.e., gyro resonator motion trajectory parameters, including in-phase component, quadrature component, amplitude, standing wave angle, and energy). These parameters not only reflect the current state of the gyro but also implicitly contain potential asymmetric error information.

[0063] To effectively extract and utilize this gyro information, this embodiment designs an innovative cross-scale attention module. This module can learn the effective gyro representations of each channel without channel dimensionality reduction and establish short-range and long-range dependencies of gyro motion information through a cross-space learning method, thereby generating more accurate attention.

[0064] Furthermore, in this embodiment, the cross-scale attention module is combined with the Informer architecture to achieve efficient processing of long sequences of gyroscope data. Specifically, in the cross-scale attention module, through channel reshaping and grouping, a partial channel dimension of the input feature map is reshaped into a batch dimension, while the channel dimension is grouped into multiple sub-features to ensure uniform distribution of spatial semantic features in each feature group.

[0065] Meanwhile, the design of parallel sub-networks is incorporated. That is, through grouped 1×1 convolution branches and 3×3 convolution branches, the first feature of cross-channel interaction can be input into the Informer encoder for further processing. Then, the feature information processed by the Informer encoder is fused with the original second feature of cross-channel interaction and the preprocessed gyroscope resonator motion trajectory parameter X to provide fast 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.

[0066] In this embodiment, a multi-head sparse attention mechanism is introduced in the encoder part to efficiently process long-range dependencies in gyroscope time series data. This mechanism optimizes the traditional self-attention mechanism and only selects probabilistically sparse points with high information contribution to participate in the calculation, thus significantly reducing the computational overhead while retaining the expressive ability of key features. Adding the output of the encoder to the feature concatenated with the output of the activation function and the average pooling layer can make the expressive ability of gyroscope time series features stronger.

[0067] In the decoder part, this embodiment adopts a masked multi-head probabilistic sparse self-attention mechanism, that is, the prediction at the current time can only depend on past information. By introducing a masking operation, this mechanism blocks the information of future gyroscope time steps, thus ensuring the temporal consistency of the gyroscope error prediction task. This design is particularly important for scenarios with high real-time requirements such as gyroscopes, ensuring the isolation of future information and enabling the model to make predictions only based on past gyroscope information.

[0068] In addition, to further enhance the model's ability to capture long-range dependencies, a dependency pyramid structure is introduced in the encoder in this embodiment. This structure constructs a multi-level feature representation framework by aggregating feature information at different time scales layer by layer. Specifically, the dependency pyramid can extract short-term and long-term dependencies from gyroscope time series data and organically combine them to more comprehensively reflect the dynamic behavior of the system.

[0069] By cascading the output of the encoder with the original input features, this embodiment can retain the original feature information to a certain extent and prevent gradient disappearance.

[0070] After the data input decoder, in this embodiment, the features output by the encoder are decoded by masked multi-head probabilistic sparse self-attention and multi-head attention jointly, and then a fully connected layer is used to map the features to the final output dimension, that is, four asymmetric errors (frequency difference, damping difference, stiffness coupling degree, and damping coupling degree).

[0071] Finally, after a series of transformations, the gyroscope asymmetric error prediction model outputs the 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.

[0072] Since gyroscopes are usually applied to embedded systems or mobile platforms with limited computing resources, it is necessary to significantly reduce the computational complexity while ensuring the model accuracy. For this purpose, this embodiment adopts technologies such as model pruning and quantization, and combines edge computing devices (such as Jetson Nano) for efficient deployment to meet the real-time and low-power requirements of gyroscopes in practical applications, as Figure 8 shown.

[0073] Edge Computing is a distributed computing paradigm. Its core concept is to transfer data processing, storage, and computing capabilities from traditional centralized clouds (such as data centers) to the source closer to where the data is generated (i.e., the "edge") to reduce latency, improve efficiency, and optimize the use of network resources. Since the computing power of gyroscopes is limited during actual application, when deploying a deep learning model in edge computing, this embodiment makes the following deployments and optimizations: Firstly, in terms of hardware selection, by leveraging the powerful hardware acceleration capabilities of the Jetson Nano edge computing device, the model can further improve the inference speed and reduce power consumption while maintaining high accuracy.

[0074] Secondly, for model compression, through quantization, pruning, and knowledge distillation techniques, it is ensured that the model can adapt to the computing power and storage limitations of edge devices while maintaining high accuracy. The formula is as follows: ; 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 requirements of full-angle gyroscope applications for real-time and reliability, providing a powerful tool for improving the accuracy and reliability of full-angle gyroscopes, and at the same time demonstrating the great potential of the combination of deep learning and edge computing.

[0075] In this embodiment, a mathematical model is established based on the dynamic equation of the gyro resonator, fully considering asymmetric error factors such as damping cross-coupling, stiffness cross-coupling, and frequency difference, which can accurately simulate the behavior of the gyroscope under different working conditions and provide 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 requirements for response speed.

[0076] The embodiments of the present invention are only used to illustrate the technical solutions of the present invention rather than to limit them. For those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can 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 in that It includes the following steps; Step 1: Build a simulation model of the gyro resonator, set the asymmetric error as the input to solve the motion trajectory parameters of the gyro resonator, construct a dataset containing the asymmetric error and the motion trajectory parameters of the gyro resonator, and preprocess the dataset; the asymmetric error includes frequency difference, damping difference, stiffness coupling degree, and damping coupling degree, and the motion trajectory parameters of the gyro resonator include the in-phase component, quadrature 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; Among them, the encoder includes a cross-scale attention module and an Informer encoder; The preprocessed motion trajectory parameters of the gyro resonator first go through grouping and recombination operations, and then 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. Then, the feature information processed by the Informer encoder is fused with the second cross-channel interaction feature. The fused feature is multiplied by the preprocessed motion trajectory parameters of the gyro resonator and input into the decoder. Finally, the predicted values of the four asymmetric errors are obtained through a fully connected layer; Step 3: Use the preprocessed dataset 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 the cross-scale Informer neural network according to claim 1, wherein, In Step 1, the steps to build the gyro resonator simulation model are as follows: Step 1.1: Establish a mathematical model based on the dynamic equation of the gyro resonator. This model is based on the following differential equation: ； ； In the formula, is the stiffness coupling coefficient term between the X and Y modes, is the damping coupling coefficient term between the X and Y modes; among them, and respectively represent the displacements on the X mode and the Y mode, and respectively represent the damping coefficients of the X mode and the Y mode, and respectively represent the stiffness coefficients on the X mode and the Y mode, and respectively represent the driving forces on the X mode and the Y mode, represents the equivalent mass, and respectively represent the velocities on the X mode and the Y mode, and respectively represent the accelerations on the X mode and the Y mode, and respectively represent the Coriolis force sensitive terms sensed on the X mode and the Y mode, where is the angular gain coefficient, which is determined by the geometry of the gyroscope head and the gyroscopic effect vibration mode; Step 1.2: Extract the in-phase component and the quadrature component of the X mode and the Y mode through coherent demodulation, which are the in-phase component of the X mode , the quadrature component of the X mode , the in-phase component of the Y mode and the quadrature component of the Y mode ; Calculate the standing wave angle of the full-angle mode gyro resonator according to the detected gyro response signal. Under non-ideal conditions, the dynamic standing wave angle output solution of the full-angle mode gyroscope is: ； Among them, is the standing wave angle of the gyro resonator, represents the angular rate of the gyro resonator, is the gyro angular rate gain, represents the damping error coefficient between the X and Y modes, represents the difference in resonance frequencies between the X and Y modes, represents the quadrature error amount of the resonator, represents the vibration energy of the gyro resonator, represents the physical rotation angular rate of the Z axis; Step 1.3: According to , , and in Step 1.2, calculate the parameters of the full-angle mode gyroscope. , , and perform parameter calculation of the full-angle mode gyroscope.

3. The gyroscope asymmetric error prediction method based on the cross-scale Informer neural network according to claim 2, characterized in that, In Step 1.2, the process of extracting the in-phase component and quadrature component of the X mode and Y mode is as follows: 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, the high-frequency components are removed through a low-pass filter to obtain the in-phase component of the X mode , the quadrature component of the X mode , the in-phase component of the Y mode and the quadrature component of the Y mode .

4. The gyroscope asymmetric error prediction method based on the cross-scale Informer neural network according to claim 2, wherein The formula for calculating the motion trajectory parameters of the full-angle mode gyro resonator in Step 1.3 is: ； ； ； ； Among them, represents the oscillation energy of the gyro resonator, represents the orthogonal component of the gyro resonator, and respectively represent the component of the gyro resonator in the X mode and the component in the Y mode.

5. The gyroscope asymmetric error prediction method based on the cross-scale Informer neural network according to claim 1, characterized in that The process of constructing the dataset is as follows: First, combine the four asymmetric errors in the way of single-error dominance and multi-error dominance, and set the specific values of the four asymmetric errors according to the dominance of the asymmetric error to generate several different error combinations; among them, multi-error dominance includes two-error dominance, three-error dominance, and four-error dominance; Apply different excitation signals to the gyro resonator under each error combination, and collect the gyro resonator motion trajectory parameters of the gyro resonator under different excitation signals through sensors, including the in-phase components of the X mode and the Y mode , , quadrature components , , amplitude, standing wave angle and energy; Record the collected data as a time series, and label the corresponding error combination and excitation signal type as the data samples of the dataset.

6. The gyroscope asymmetric error prediction method based on the cross-scale Informer neural network according to claim 1, wherein, In Step 1, the process of preprocessing the dataset is as follows: (1) Data cleaning, removing the noise and outliers in the motion trajectory parameters of the gyro resonator; (2) Detect and remove the outliers in the motion trajectory parameters of the gyro resonator; (3) Data normalization processing, scaling the data to the range of [0, 1] or [-1, 1]; (4) Standardize the data, converting the data into a distribution with a mean of 0 and a standard deviation of 1; Then, the preprocessed dataset is divided into a training set, a validation set, and a test set according to the 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. The test set is used to finally evaluate the true 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 adopted. 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 the cross-scale Informer neural network according to claim 1, characterized in that, In step 2, the processing process of the preprocessed gyroscope resonator motion trajectory parameters in the cross-scale attention module is as follows: First, the features in the preprocessed gyroscope resonator motion trajectory parameters are grouped into a 1×1 convolution branch and a 3×3 convolution branch. One-dimensional horizontal global pooling operation and one-dimensional vertical global pooling operation are respectively performed on the 1×1 convolution branch, and then a concatenation operation is carried out. Then, the activation values of the concatenated feature map are compressed to the range of 0-1 through the Sigmoid function, and then multiplied pointwise with the preprocessed gyroscope resonator motion trajectory parameters. Then, group normalization operation is performed, and then the Softmax activation function and average pooling operation are respectively carried out. The 3×3 convolution branch respectively undergoes the Softmax activation function and average pooling operation. The features after the 1×1 convolution branch passes through the Softmax activation function are fused and multiplied with the features after the 3×3 convolution branch passes through the average pooling operation to obtain the first cross-channel interaction feature. The features after the 1×1 convolution branch passes through the average pooling operation are fused and multiplied with the features after the 3×3 convolution branch passes through the Softmax activation function to obtain the second cross-channel interaction feature.

8. The gyroscope asymmetric error prediction method based on the cross-scale Informer neural network according to claim 1, wherein In step 2, the processing process of obtaining the fused features from the first cross-channel interaction feature is as follows: After the first cross-channel interaction feature is input into the Informer encoder, multi-head sparse attention operations are respectively performed. Then, the second cross-channel interaction feature is added to the output of the Informer encoder to obtain the fused features. In addition, a dependence pyramid structure is introduced in the Informer encoder. The dependence pyramid can extract short-term and long-term dependencies from the 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 the cross-scale Informer neural network according to claim 1, characterized in that In step 2, the processing process of the features after pointwise multiplication in the decoder is as follows: After the features after pointwise multiplication are input into the decoder, 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 currently generated partial sequence, 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 progresses, the decoder continuously generates a new output sequence and feeds it back to the masked multi-head probabilistic sparse self-attention layer at the next time step.

10. The gyroscope asymmetric error prediction method based on the cross-scale Informer neural network according to claim 6, wherein 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 squared error is used as the loss function, and optimization is carried out through the Adam optimizer. Finally, four asymmetric errors are predicted by regression, and then compared with the true values of the four asymmetric errors. At the same time, the root mean squared error is used to measure the accuracy of the gyroscope asymmetric error prediction model; During the training process, the performance of the gyroscope asymmetric error prediction model is evaluated on the validation set regularly to monitor the overfitting phenomenon and adjust the hyperparameters in a timely manner; After the training is completed, 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 metrics are calculated: (1) Calculate the percentage of absolute error between the predicted values and the true values of the four asymmetric errors, and then find the average value; (2) Calculate the square root of the mean squared error between the predicted values and the true values. The formula is: ； Among them, represents the total number of data points, represents the true value of the th data point of the asymmetric error, represents the th predicted value of the data point.

Images