An error identification method and system for an inertial measurement unit

Abstract

The application discloses an error identification method and system for an inertial measurement unit, and relates to the technical field of inertial measurement, which comprises the following steps: modal analysis is performed on the inertial measurement unit by using finite element simulation, a distributed sensor array is arranged in a resonance sensitive area, multi-source signals are synchronously collected, and a standardized signal set is output; short-time Fourier transform and multi-channel time-frequency matrix fusion denoising are performed on the standardized signal set, a special sparse dictionary is constructed and trained based on finite element modal parameters, candidate resonance modal characteristics are screened through sparse decomposition, a resonance modal cross-axis error mapping model is constructed, and a spatial frequency characteristic set is output; an adaptive wave trap filter set is initialized based on the spatial frequency characteristic set, a real signal is output, a cross-axis interference signal is obtained by calculating the difference between the signals before and after filtering, and error source positioning lists are generated through quantitative analysis and tracing in combination with the error mapping model, so that the precision and reliability of the inertial measurement unit are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of inertial measurement technology, specifically to a method and system for error identification of inertial measurement units. Background Technology

[0002] As a core component of navigation and motion control systems, the accuracy of the inertial measurement unit (IMU) directly determines the performance of the system. In the existing technology, the error identification of the IMU mainly relies on two methods: signal filtering and hardware compensation. Signal filtering methods usually use notch filters or Kalman filtering algorithms with fixed parameters to reduce errors by suppressing interference signals in specific frequency bands. Hardware compensation reduces the impact of resonance on sensor output by optimizing the mechanical structure design of the IMU or adding damping materials.

[0003] However, existing technologies have significant limitations. First, traditional filtering methods rely on fixed parameters and cannot dynamically adapt to changes in the resonant frequency of the IMU under different operating conditions, resulting in unstable error suppression effects. Second, while hardware compensation can partially reduce the impact of resonance, it lacks precise location of the error source, making it difficult to optimize the structural design and is also costly. Furthermore, existing technologies have failed to effectively address the identification and tracing of cross-axis coupling errors. The resonant modes of the IMU often cause signal interference between multiple axes, while existing methods are mostly limited to single-axis analysis and lack the ability to model the relationship between spatial frequency characteristics and error mapping. This leads to inaccurate location of error sources, making it difficult to provide a reliable basis for subsequent compensation algorithms and limiting further improvements in IMU performance. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for error identification in inertial measurement units (IMUs). By employing finite element simulation to locate the resonant sensitive area, multi-source signal collaborative acquisition and processing, adaptive filtering, and error source tracing techniques, this invention solves the problems of traditional methods being unable to dynamically adapt to changes in resonant frequency, inaccurate cross-axis coupling error identification, and difficulty in locating error sources. This significantly improves the accuracy and reliability of IMUs.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for error identification in inertial measurement units, comprising:

[0008] Finite element simulation was used to perform modal analysis on the inertial measurement unit, locate the resonance sensitive area and deploy a distributed sensor array, synchronously acquire multi-source signals, and output a standardized signal set after preprocessing the signals.

[0009] Short-time Fourier transform and multi-channel time-frequency matrix fusion denoising are performed on the standardized signal set. A dedicated sparse dictionary is constructed and trained based on finite element mode parameters. Candidate resonant mode features are screened through sparse decomposition. A BP neural network is trained in combination with a hardware-in-the-loop simulation platform to construct a resonant mode cross-axis error mapping model and output a spatial frequency feature set.

[0010] An adaptive notch filter bank is initialized based on the spatial frequency feature set. The signal is then subjected to collaborative filtering and motion consistency verification to output the real signal. The difference between the signals before and after filtering is calculated to obtain the cross-axis interference signal. Combined with the error mapping model, quantitative analysis and source tracing are performed to generate a list of error source locations.

[0011] Furthermore, by simulating and calculating the mode shapes and mode amplitudes at different frequencies, several resonant sensitive regions where the mode amplitudes are greater than the preset mode amplitude threshold are located; N piezoelectric thin film sensors are deployed in each sensitive region according to a 2×2 spatial gradient.

[0012] Furthermore, multiple source signals, including vibration signals, acceleration signals, and gyroscope signals, are acquired simultaneously. The acquired raw signals are preprocessed: the moving average method is used to adaptively eliminate the DC component of the vibration, acceleration, and gyroscope signals; a 5th-order Butterworth low-pass filter is used to filter the multiple source signals. After preprocessing, a standardized set of vibration signals, preprocessed acceleration signals, and preprocessed gyroscope signals are output.

[0013] Furthermore, the denoising process involving short-time Fourier transform (STFT) and multi-channel time-frequency matrix fusion is as follows: STFT is applied to each channel of the standardized vibration signal set, using a Hanning window and variable window length strategy. The window length for the resonant prediction frequency band is set to 128 points, and the window length for the non-resonant prediction frequency band is set to 256 points. The overlap rate of all frequency band signals is uniformly set to 60%, resulting in N high-resolution time-frequency matrices. The covariance matrix between any two channel time-frequency matrices is calculated, and the covariance coefficients between each channel are extracted from the covariance matrix. Effective signal channels with covariance coefficients greater than the coefficient threshold are selected, and a weighted average fusion algorithm is used to fuse the effective channel time-frequency matrices.

[0014] Furthermore, a dedicated sparse dictionary is constructed and trained. Candidate resonant mode features are screened through sparse decomposition. Specifically, an initial dictionary is constructed based on the resonant mode parameters from finite element simulation. The dictionary size is 2048×4096, and the dictionary atoms are two-dimensional feature vectors of frequency and spatial location. An improved K-SVD algorithm is used to train the dictionary, with the fused time-frequency matrix as the training sample. A resonant mode constraint term based on cosine similarity is introduced, and the process is iterated several times. Training stops when the joint objective function is less than the target threshold. An improved orthogonal matching pursuit algorithm is used for sparse decomposition. Spatial consistency constraints are introduced during the decomposition process. After the decomposition, a sparse coefficient matrix is ​​obtained. Dictionary atoms corresponding to sparse coefficients whose amplitudes are greater than the sparse coefficient threshold are selected to obtain a candidate resonant mode feature set. The feature set contains the frequency and spatial features of each candidate mode.

[0015] Furthermore, a resonant mode cross-axis error mapping model is constructed to output a spatial frequency feature set. This includes: building an IMU hardware-in-the-loop simulation platform to simulate resonant modes of different frequencies and modes, and obtaining cross-axis coupling error data for different resonant modes; using the spatial and frequency features of candidate resonant modes as input parameters and the cross-axis coupling error data as output parameters, training a BP neural network, optimizing the network weights using a stochastic gradient descent algorithm, and stopping training when the training error is less than 0.001, thus constructing the resonant mode cross-axis error mapping model; inputting the candidate resonant mode feature set into this mapping model, and the model outputs reference error values, including reference accelerometer cross-axis error values ​​and reference gyroscope cross-axis error values, to obtain a spatial frequency feature set containing resonant frequency, spatial position, and reference error values.

[0016] Furthermore, the adaptive notch filter bank is initialized as follows: m notch filters are initialized based on the spatial frequency feature set, the center frequency of each filter is aligned with the resonant frequency of the corresponding resonant mode, and the filter bandwidth is adaptively weighted by error; according to the inter-axis correlation relationship corresponding to the reference error of each mode in the spatial frequency feature set, the inter-axis cooperative suppression weight of the filter bank is set.

[0017] Furthermore, the signals undergo collaborative filtering and motion consistency verification. Specifically, FFT is performed on the gyroscope preprocessed signal, and the initial estimate of the actual cross-axis error is calculated using the frequency domain matching method in conjunction with the acceleration preprocessed signal. The error deviation between the initial estimate and the reference error value is calculated. If the error deviation is greater than the error deviation threshold, the deviation is used as feedback, and the filter center frequency, bandwidth, and inter-axis collaborative weights are adjusted using a PID algorithm. The preprocessed signal is input into the adjusted filter bank, and collaborative filtering is performed according to the inter-axis collaborative suppression weights. Based on the rigid body kinematics principle, the physical and logical matching between the filtered angular velocity signal and the filtered acceleration signal is verified. When the mean square error is less than the mean square error threshold, the filtered acceleration signal is output as the true acceleration signal, and the filtered angular velocity signal is output as the true angular velocity signal.

[0018] Furthermore, the difference between the preprocessed acceleration signal and the true acceleration signal is the actual acceleration cross-axis interference signal, and the difference between the preprocessed gyroscope signal and the true angular velocity signal is the actual gyroscope cross-axis interference signal. The peak value, effective value, main frequency, harmonic component ratio, and inter-axis coupling coefficient of the interference signal are extracted to form a quantization error report. The quantized interference signal is associated with the spatial coordinates of the resonant sensitive area in the feature set to determine the IMU mechanical structure resonant area and resonant frequency corresponding to each actual cross-axis error, and an error source location list is generated.

[0019] An error identification system for inertial measurement units, comprising:

[0020] The signal acquisition module uses finite element simulation to perform modal analysis on the inertial measurement unit, locates the resonant sensitive area and deploys a distributed sensor array, synchronously acquires multi-source signals, and outputs a standardized signal set after preprocessing the signals.

[0021] The resonant mode feature extraction module performs short-time Fourier transform and multi-channel time-frequency matrix fusion denoising on the standardized signal set. It constructs and trains a dedicated sparse dictionary based on finite element mode parameters, filters candidate resonant mode features through sparse decomposition, trains a BP neural network in conjunction with a hardware-in-the-loop simulation platform, constructs a resonant mode cross-axis error mapping model, and outputs a spatial frequency feature set.

[0022] The error filtering and source tracing module initializes an adaptive notch filter bank based on the spatial frequency feature set, performs collaborative filtering and motion consistency verification on the signal, outputs the real signal, calculates the difference between the signal before and after filtering to obtain the cross-axis interference signal, combines the error mapping model for quantitative analysis and source tracing, and generates a list of error source locations.

[0023] (III) Beneficial Effects

[0024] This invention provides a method and system for error identification in inertial measurement units, which has the following advantages:

[0025] (1) The resonance sensitive area of ​​the IMU was accurately located by finite element simulation, and a distributed sensor array was deployed. Combined with the adaptive sampling strategy and synchronous triggering system, high-precision acquisition and preprocessing of multi-source signals were realized, effectively eliminating DC components and high-frequency noise in the signal, and outputting a standardized signal set, which provided a high-quality data foundation for subsequent error identification and significantly improved the accuracy and reliability of resonance mode analysis.

[0026] (2) By using short-time Fourier transform and multi-channel time-frequency matrix fusion for denoising, a dedicated sparse dictionary is constructed by combining finite element modal parameters, and candidate resonant mode features are screened using sparse decomposition. This significantly improves the accuracy of signal denoising and modal feature extraction. The cross-axis error mapping model constructed by training a BP neural network through a hardware-in-the-loop simulation platform is used to achieve accurate correlation between resonant modes and errors. This provides a highly reliable spatial frequency feature set for subsequent error filtering and source tracing, effectively solving the problem of inaccurate cross-axis coupling error identification in traditional methods.

[0027] (3) By dynamically suppressing resonance interference through an adaptive notch filter bank and combining motion consistency verification to ensure the physical and logical matching of the filtered signal, the real signal and the cross-axis interference signal are effectively separated. Based on the error mapping model, the amplitude, frequency and inter-axis coupling characteristics of the interference signal are quantitatively analyzed and accurately associated with the resonance sensitive area of ​​the IMU mechanical structure. An error source location list is generated, which significantly improves the pertinence and dynamic adaptability of error suppression. It provides a high-precision error source tracing basis for subsequent compensation algorithms and solves the problem that cross-axis coupling error is difficult to accurately identify and locate in traditional methods. Attached Figure Description

[0028] Figure 1 This is a schematic diagram illustrating the steps of the error identification method for inertial measurement units according to the present invention;

[0029] Figure 2 This is a schematic diagram of the error identification method for inertial measurement units according to the present invention;

[0030] Figure 3 This is a schematic diagram of the error identification system for inertial measurement units according to the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-2This invention provides a method for error identification in inertial measurement units, comprising the following steps:

[0033] Step 1: Use finite element simulation to perform modal analysis on the inertial measurement unit, locate the resonant sensitive area and deploy a distributed sensor array, synchronously acquire multi-source signals, preprocess the signals and output a standardized signal set;

[0034] Step one includes the following:

[0035] Step 101: Use ANSYS finite element simulation tool to perform modal analysis on the mechanical structure of the inertial measurement unit (IMU). Input the material parameters of the IMU shell and internal support structure, including elastic modulus, Poisson's ratio, density, and specific geometric dimensions. Set the vibration frequency scanning range to 50 Hz to 10 kHz. Calculate the mode shapes and mode amplitudes at different frequencies through simulation, locate 3-5 resonant sensitive regions where the mode amplitude is greater than the preset mode amplitude threshold, and output the spatial coordinate range of each sensitive region. The mode amplitude threshold is 0.01 by default and can be adjusted according to specific conditions to determine the resonant sensitive regions. Within each resonant sensitive region, arrange piezoelectric thin film sensors in a 2×2 spatial gradient to form a distributed sensor array. The number of sensors N is determined according to the total area of ​​the sensitive region, with each 1 cm² sensor array containing piezoelectric thin film sensors. 2 Deploy one sensor; fix the IMU on a standard vibration table, select several characteristic frequency points covering the estimated resonant frequency range for sensitivity calibration, apply three sets of standard vibration excitations with different amplitudes at each frequency point, such as 0.1, 0.3, and 0.5, record the output voltage and corresponding vibration acceleration of each sensor, and use the least squares method to fit the nonlinear conversion relationship between the output voltage and vibration acceleration of each sensor to correct the sensitivity deviation caused by frequency changes, and ensure that the sensitivity deviation after calibration does not exceed 0.02.

[0036] Step 102: Based on the resonant frequency prediction range obtained from finite element simulation, from the lowest to the highest predicted resonant frequency, an adaptive sampling strategy is adopted to configure the acquisition parameters: the reference sampling frequency is set to twice the highest predicted resonant frequency to satisfy the Nyquist theorem and avoid frequency aliasing; at the same time, the vibration amplitude threshold is set to 1.2 times the rated operating vibration amplitude of the IMU; the vibration amplitude is monitored in real time through a sensor array; when the vibration amplitude exceeds the vibration amplitude threshold, the sampling frequency is automatically triggered to increase the sampling frequency to three times the highest predicted resonant frequency, achieving accurate capture of high-frequency strong vibration segments and efficient acquisition of conventional vibration segments; the sampling duration is determined according to the resonant period to ensure that the sampling duration is not less than 10 periods corresponding to the lowest resonant frequency;

[0037] Step 103: Build a dual-synchronization trigger system of FPGA and clock synchronization chip. The FPGA generates a reference clock signal, and the clock synchronization chip divides and calibrates the reference clock signal to generate a synchronization trigger signal. The synchronization trigger signal is sent to the signal acquisition terminals of the piezoelectric thin film sensor array, the IMU built-in 3-axis accelerometer and 3-axis gyroscope respectively to realize the synchronization of the acquisition start time of multiple devices.

[0038] Step 104: Preprocess the acquired multi-source raw signals: Adaptively eliminate the DC component of the vibration signal, acceleration signal, and gyroscope signal using the moving average method, with the sliding window length set to 50 sampling points to remove signal baseline drift; Filter the above signals using a 5th-order Butterworth low-pass filter, with the filter cutoff frequency set to 1.5 times the highest estimated resonant frequency to filter out high-frequency noise introduced during sampling; After preprocessing, output the standardized vibration signal set, acceleration preprocessed signal, and gyroscope preprocessed signal.

[0039] When using this method, refer to steps 101 to 104:

[0040] By accurately locating the resonant sensitive area of ​​the IMU through finite element simulation and deploying a distributed sensor array, combined with an adaptive sampling strategy and a synchronous triggering system, high-precision acquisition and preprocessing of multi-source signals were achieved. This effectively eliminated the DC component and high-frequency noise in the signal, outputting a standardized signal set, providing a high-quality data foundation for subsequent error identification, and significantly improving the accuracy and reliability of resonant mode analysis.

[0041] Step 2: Perform short-time Fourier transform and multi-channel time-frequency matrix fusion denoising on the standardized signal set, construct and train a dedicated sparse dictionary based on finite element modal parameters, screen candidate resonant mode features through sparse decomposition, train a BP neural network in combination with a hardware-in-the-loop simulation platform, construct a resonant mode cross-axis error mapping model, and output a spatial frequency feature set.

[0042] Step two includes the following:

[0043] Step 201: Perform Short-Time Fourier Transform (STFT) on the N-channel vibration signal set preprocessed in Step 1, using the Hanning window as the window function and implementing a variable window length strategy: divide the signal frequency band into a resonant prediction band and a non-resonant prediction band. The resonant prediction band is the resonant sensitive frequency band within the 50Hz~10kHz scanning range obtained from the finite element analysis in Step 1, and the non-resonant prediction band is the remaining frequency bands within this scanning range excluding the resonant sensitive frequency band. The window length within the resonant prediction band is set to 128 points, and the window length within the non-resonant prediction band is set to 256 points. The signal overlap rate for all frequency bands is uniformly set to 60%, meaning the number of overlapping data points between two adjacent analysis windows accounts for 60% of the number of data points within a single window length. The final result is... N high-resolution time-frequency matrices are obtained; multi-channel time-frequency matrix fusion denoising is carried out based on the spatial correlation of the sensor array: the covariance matrix between any two channel time-frequency matrices is calculated, and the covariance coefficient between each channel is extracted from the covariance matrix, specifically the Pearson correlation coefficient. The covariance coefficient of the same source signal channels with a covariance coefficient greater than the coefficient threshold is selected as the effective signal channels, and the interference signal channels with a covariance coefficient less than 0.3 are removed; a weighted average fusion algorithm is used to fuse the time-frequency matrices of the effective signal channels. The fusion weight is set according to the covariance coefficient. The larger the covariance coefficient, the higher the weight ratio. The sum of the weights of all effective channels is 1. The coefficient threshold is 0.8 by default and is adjusted as needed to ensure that effective signal channels are selected.

[0044] Step 202: Based on the resonant modal parameters obtained from the finite element analysis in Step 1, including frequency and mode shape spatial location, construct an initial dictionary D_init. The dictionary atoms are dual-dimensional feature vectors of frequency and spatial location. The frequency dimension is represented by the normalized values ​​of characteristic frequencies within the estimated resonant frequency band, and the spatial location dimension is represented by the normalized values ​​of the sensor deployment coordinates. The frequency dimension covers all characteristic frequency points within the estimated resonant frequency band, and the spatial location dimension corresponds to the deployment coordinates of the distributed sensor array. The initial dictionary size is set to 2048×4096, with rows representing the number of atoms and columns representing feature dimensions. The first 2048 columns correspond to the frequency dimension features, and the last 2048 columns correspond to the spatial location dimension features. An improved K-SVD algorithm is used to train the initial dictionary: to achieve fusion... The combined time-frequency matrix is ​​used as a training sample. A resonant mode constraint term is introduced into the objective function of the traditional K-SVD algorithm. This constraint term is constructed by calculating the matching degree between dictionary atoms and mode shape data obtained from finite element analysis. The matching degree is obtained by calculating the cosine similarity between the spatial dimension vector of the dictionary atom and the displacement vector of the corresponding sensitive region of the mode shape data, so that the trained dictionary atoms can accurately match the features of the IMU resonant modes. The number of iterations is set to 150 during training. The joint objective function is to minimize the sum of reconstruction error and mode matching error. Training stops when the value of the joint objective function is less than the target threshold. Finally, a dedicated sparse dictionary adapted for IMU resonant mode recognition is obtained. The target threshold is 0.008 by default and can be adjusted according to specific circumstances to ensure the adaptation result.

[0045] Step 203: Input the fused time-frequency matrix into a dedicated sparse dictionary and perform sparse decomposition using an improved orthogonal matching pursuit (OMP) algorithm. During the decomposition process, a spatial consistency constraint is introduced: the correlation threshold between different channel sparse coefficients corresponding to the same resonant mode is set to 0.8. The correlation is obtained by calculating the Pearson correlation coefficient of different channel sparse coefficient vectors. When the correlation between multiple channel sparse coefficients corresponding to any frequency point is greater than the correlation threshold, they are considered valid sparse coefficients; otherwise, they are considered noise interference and discarded. After decomposition, a sparse coefficient matrix is ​​obtained. A sparse coefficient threshold is set, which is adaptively adjusted according to the acquired vibration amplitude: when the acquired vibration amplitude is not lower than the vibration amplitude threshold, the sparse coefficient threshold is set to 0.7; when the acquired vibration amplitude is lower than the vibration amplitude threshold, the sparse coefficient threshold is set to 0.8. Dictionary atoms corresponding to sparse coefficients with amplitudes greater than the sparse coefficient threshold are selected to obtain a candidate resonant mode feature set. The feature set contains the frequency and spatial features of each candidate mode.

[0046] Step 204: Construct an IMU hardware-in-the-loop simulation platform. This platform consists of an IMU physical module, a vibration excitation module, a data acquisition module, and a host computer. The vibration excitation module can simulate resonant modes of different frequencies and modes, covering the 50Hz~10kHz vibration frequency scanning range determined in Step 1, and the simulated modes cover the modes corresponding to the 3~5 resonant sensitive regions obtained from finite element analysis. The platform is used to acquire cross-axis coupling error data for different resonant modes. The resonant modes cover all resonant frequencies and modes corresponding to the 3~5 resonant sensitive regions located in Step 1, including accelerometer cross-axis errors and gyroscope cross-axis errors. The spatial and frequency characteristics of the candidate resonant modes are used as input parameters, including the resonant frequency and the spatial coordinates of the corresponding sensitive regions, to obtain the cross-axis coupling error data. The combined error data is used as the output parameter to train a backpropagation (BP) neural network. This BP neural network adopts a three-layer network structure, specifically an input layer, a hidden layer, and an output layer. The number of nodes in the input layer is equal to the spatial and frequency feature dimensions, the number of nodes in the hidden layer is set to 20, and the number of nodes in the output layer is equal to the cross-axis error parameter dimension. The training process uses the stochastic gradient descent (SGD) algorithm to optimize the network weights. Training stops when the training error value is less than 0.001, thus constructing a resonant mode cross-axis error mapping model. Candidate resonant mode features are input into this mapping model, and the model outputs reference error values ​​strongly correlated with each candidate feature, including the reference accelerometer cross-axis error value and the reference gyroscope cross-axis error value. Finally, a spatial frequency feature set containing resonant frequency, spatial position, and reference error values ​​is obtained.

[0047] When using this method, refer to steps 201 to 204:

[0048] By fusing short-time Fourier transform with multi-channel time-frequency matrix denoising, and constructing a dedicated sparse dictionary based on finite element modal parameters, and using sparse decomposition to screen candidate resonant mode features, the accuracy of signal denoising and modal feature extraction is significantly improved. A cross-axis error mapping model constructed by training a BP neural network through a hardware-in-the-loop simulation platform is used to achieve accurate correlation between resonant modes and errors, providing a highly reliable spatial frequency feature set for subsequent error filtering and source tracing, and effectively solving the problem of inaccurate cross-axis coupling error identification in traditional methods.

[0049] Step 3: Initialize the adaptive notch filter bank based on the spatial frequency feature set, perform collaborative filtering and motion consistency verification on the signal, output the real signal, calculate the difference between the signal before and after filtering to obtain the cross-axis interference signal, combine the error mapping model for quantitative analysis and source tracing, and generate an error source location list.

[0050] Step three includes the following:

[0051] Step 301: Based on the spatial frequency feature set obtained in Step 2, initialize m notch filters to form an adaptive notch filter bank; the center frequency of each notch filter is precisely aligned with the resonant frequency of the corresponding resonant mode in the feature set. The filter bandwidth adopts an adaptive error weighting method, specifically: calculate the sum of reference errors corresponding to each resonant mode, that is, add the cross-axis error value of the reference acceleration of that mode to the cross-axis error value of the reference gyroscope, divide the sum of reference errors of a single mode by the sum of reference errors of all modes to obtain the weight ratio corresponding to each mode; calculate the bandwidth of each filter based on the reference bandwidth, wherein the reference bandwidth is set to 80Hz, which can be adjusted according to the actual operation of the IMU. Fine-tuning of the frequency range is sufficient to keep it within the operating frequency range. The bandwidth of a single filter is equal to 1 minus the weight ratio of that mode multiplied by the reference bandwidth. Simultaneously, based on the inter-axis correlation of the reference errors of each mode in the feature set, the inter-axis cooperative suppression weight of the filter bank is set. The specific setting rule is as follows: find the maximum value among all reference acceleration cross-axis error values ​​and the maximum value among all reference gyroscope cross-axis error values. Divide the reference acceleration cross-axis error value of a single axis by the maximum acceleration cross-axis error value and the reference gyroscope cross-axis error value by the maximum gyroscope cross-axis error value, and then add them together to obtain the inter-axis cooperative suppression weight. The value range of this cooperative suppression weight is normalized to 0~1.

[0052] Step 302: Perform a Fast Fourier Transform (FFT) on the gyroscope preprocessing signal to obtain the frequency domain signal. Extract the amplitude of the frequency components corresponding to each resonant frequency in the feature set from this frequency domain signal. Combined with the acceleration preprocessing signal, calculate the initial estimate of the actual cross-axis error using the frequency domain matching method. The specific calculation rule is as follows: the initial estimate of the actual gyroscope cross-axis error is equal to the amplitude of the corresponding resonant frequency in the current frequency domain signal, multiplied by the cross-axis error value of the reference gyroscope corresponding to that resonant frequency in the feature set, and then divided by the amplitude of the standard gyroscope signal at the same resonant frequency in the hardware-in-the-loop simulation platform. The standard gyroscope signal amplitude is the gyroscope signal under ideal conditions without cross-axis error, synchronously acquired when building a hardware-in-the-loop simulation platform to obtain cross-axis coupling error data. After FFT transformation, the amplitude at the corresponding resonant frequency is obtained and stored along with the cross-axis coupling error data, associated with the corresponding resonant frequency. The calculation rule for the initial estimate of the actual acceleration cross-axis error is consistent with that of the gyroscope, only replacing relevant parameters with those corresponding to acceleration, including the current frequency domain signal amplitude, the feature set reference acceleration cross-axis error value, and the simulation platform's standard acceleration signal amplitude. The calculated actual... The initial estimates of the acceleration cross-axis error and the actual gyroscope cross-axis error are compared with the corresponding reference error values ​​in the feature set to calculate the absolute value difference, resulting in two error deviations. The maximum of these two error deviations is taken as the final error deviation. An error deviation threshold is set according to the IMU accuracy requirements, with the specific value matching the error type. If the final error deviation is greater than the error deviation threshold, this deviation is used as feedback to dynamically adjust the parameters of the corresponding notch filter through a PID algorithm. The proportional coefficient of the PID algorithm is set to 0.8, the integral coefficient to 0.2, and the derivative coefficient to 0.1. The specific adjustment rules are as follows: the center frequency adjustment is equal to the product of the proportional coefficient and the error deviation, plus the product of the integral coefficient and the integral value of the error deviation, plus the product of the derivative coefficient and the rate of change of the error deviation; the bandwidth adjustment is equal to half of the product of the proportional coefficient and the error deviation; the inter-axis cooperative weight adjustment is equal to one-third of the product of the proportional coefficient and the error deviation, ensuring that the filter suppression frequency band and the actual resonant interference frequency band are completely matched. If the final error deviation is less than or equal to the error deviation threshold, the filter parameters remain unchanged.

[0053] Step 303: Synchronously input the acceleration preprocessed signal and gyroscope preprocessed signal into the adjusted adaptive notch filter bank, and perform multi-channel collaborative filtering in combination with inter-axis collaborative suppression weights: allocate the filtering intensity to the signal of each channel according to the corresponding inter-axis collaborative suppression weights. After filtering, perform motion consistency verification on the output signal: based on the rigid body kinematics principle, verify the physical and logical matching of the filtered angular velocity signal and the filtered acceleration signal. The specific verification rules are as follows: calculate the angular acceleration, i.e., the rate of change of the angular velocity signal, through the filtered angular velocity signal, derive the theoretical acceleration signal from the angular acceleration, and the acceleration is the product of the angular acceleration and the radius of rotation. Calculate the mean square error between the filtered acceleration signal and the theoretical acceleration signal; set the mean square error threshold to 0.001. If the calculated mean square error is less than the mean square error threshold, the motion consistency is deemed qualified, and the filtered acceleration signal is output as the real acceleration signal and the filtered angular velocity signal is output as the real angular velocity signal; if the mean square error is greater than or equal to the threshold, return to step 302 to readjust the filter parameters until the consistency verification is qualified.

[0054] Step 304: Calculate the difference between the corresponding signals before and after filtering to obtain the actual cross-axis interference signal. Specifically, the difference between the pre-processed acceleration signal and the real acceleration signal is the actual acceleration cross-axis interference signal, and the difference between the pre-processed gyroscope signal and the real angular velocity signal is the actual gyroscope cross-axis interference signal. Combining the output results of the constructed resonant mode cross-axis error mapping model, a quantitative analysis of the actual cross-axis interference signal is performed. The specific analysis content and rules are as follows: Error amplitude extraction: The peak value is the maximum amplitude of the interference signal within the acquisition period, and the effective value is the root mean square value of the interference signal within the acquisition period; Frequency characteristic extraction: The dominant frequency is the frequency component with the highest power spectral density of the interference signal, and the harmonic component ratio is the ratio of the total amplitude of each harmonic component to the amplitude of the dominant frequency; Inter-axis Coupling relationship extraction: The inter-axis coupling coefficient is the ratio of the amplitude of the interference signal on the cross axis to the amplitude of the interference signal on the main axis. Based on the above analysis, a quantization error report is generated, which clearly records the amplitude, frequency characteristics, inter-axis coupling coefficient, and corresponding resonant frequency of each actual cross-axis error. At the same time, based on the spatial coordinates of the sensitive areas in the feature set, the quantized actual cross-axis interference signals are associated with the corresponding resonant sensitive areas one by one, completing the resonant mode tracing of the cross-axis coupling error. After the tracing is completed, the resonant region and resonant frequency of the IMU mechanical structure corresponding to each actual cross-axis error are determined, down to the name of the structural component, such as the gyroscope mounting base, accelerometer bracket, etc., forming an error source location list, which provides accurate error source location information for the design of subsequent error compensation algorithms.

[0055] When using this method, refer to steps 301 to 304:

[0056] By dynamically suppressing resonant interference through an adaptive notch filter bank and combining motion consistency verification to ensure the physical and logical matching of the filtered signal, the real signal and cross-axis interference signal are effectively separated. Based on the error mapping model, the amplitude, frequency and inter-axis coupling characteristics of the interference signal are quantitatively analyzed and accurately correlated to the resonant sensitive area of ​​the IMU mechanical structure, generating an error source location list. This significantly improves the pertinence and dynamic adaptability of error suppression, provides a high-precision error source tracing basis for subsequent compensation algorithms, and solves the problem of difficult accurate identification and location of cross-axis coupling errors in traditional methods.

[0057] Please see Figure 3 This invention provides an error identification system for inertial measurement units, comprising: a signal acquisition module, a resonant mode feature extraction module, and an error filtering and source tracing module, wherein:

[0058] The signal acquisition module uses finite element simulation to perform modal analysis on the inertial measurement unit, locates the resonant sensitive area and deploys a distributed sensor array, synchronously acquires multi-source signals, and outputs a standardized signal set after preprocessing the signals.

[0059] The resonant mode feature extraction module performs short-time Fourier transform and multi-channel time-frequency matrix fusion denoising on the standardized signal set. It constructs and trains a dedicated sparse dictionary based on finite element mode parameters, filters candidate resonant mode features through sparse decomposition, trains a BP neural network in conjunction with a hardware-in-the-loop simulation platform, constructs a resonant mode cross-axis error mapping model, and outputs a spatial frequency feature set.

[0060] The error filtering and source tracing module initializes an adaptive notch filter bank based on the spatial frequency feature set, performs collaborative filtering and motion consistency verification on the signal, outputs the real signal, calculates the difference between the signal before and after filtering to obtain the cross-axis interference signal, combines the error mapping model for quantitative analysis and source tracing, and generates a list of error source locations.

[0061] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0062] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0063] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for error identification in inertial measurement units, characterized in that: include: Finite element simulation was used to perform modal analysis on the inertial measurement unit, locate the resonance sensitive area and deploy a distributed sensor array, synchronously acquire multi-source signals, and output a standardized signal set after preprocessing the signals. The standardized signal set is denoised by short-time Fourier transform and multi-channel time-frequency matrix fusion. A dedicated sparse dictionary is constructed and trained based on finite element mode parameters. Candidate resonant mode features are screened through sparse decomposition, including: constructing an initial dictionary based on resonant mode parameters from finite element simulation, with dictionary atoms being two-dimensional feature vectors of frequency and spatial location; training the dictionary using an improved K-SVD algorithm, using the fused time-frequency matrix as training samples, introducing resonant mode constraint terms based on cosine similarity, iterating several times, and stopping training when the joint objective function is less than the target threshold; performing sparse decomposition using an improved orthogonal matching pursuit algorithm, introducing spatial consistency constraints during the decomposition process, obtaining a sparse coefficient matrix after decomposition, and selecting dictionary atoms corresponding to sparse coefficients with amplitudes greater than the sparse coefficient threshold to obtain a candidate resonant mode feature set. The feature set contains each candidate mode... The frequency and spatial characteristics of the resonant modes are analyzed. A backpropagation (BP) neural network is trained using a hardware-in-the-loop (HIFU) simulation platform to construct a resonant mode cross-axis error mapping model, outputting a spatial frequency feature set. This process includes: building an IMU HIFU platform to simulate resonant modes of different frequencies and modes, obtaining cross-axis coupling error data for different resonant modes; training a BP neural network using the spatial and frequency characteristics of candidate resonant modes as input parameters and the cross-axis coupling error data as output parameters, optimizing the network weights using a stochastic gradient descent algorithm, stopping training when the training error is less than 0.001, thus constructing the resonant mode cross-axis error mapping model; inputting the candidate resonant mode feature set into this mapping model, the model outputs reference error values, including reference accelerometer cross-axis error values ​​and reference gyroscope cross-axis error values, resulting in a spatial frequency feature set containing resonant frequency, spatial position, and reference error values. An adaptive notch filter bank is initialized based on a spatial frequency feature set. Cooperative filtering and motion consistency verification of the signal are performed, including: performing an FFT on the gyroscope preprocessed signal; calculating the initial estimate of the actual cross-axis error using a frequency domain matching method in conjunction with the acceleration preprocessed signal; calculating the error deviation between the initial estimate and the reference error value; if the error deviation is greater than the error deviation threshold, using this deviation as feedback, adjusting the filter center frequency, bandwidth, and inter-axis cooperative weights using a PID algorithm; inputting the preprocessed signal into the adjusted filter bank, performing cooperative filtering according to the inter-axis cooperative suppression weights; verifying the physical and logical matching of the filtered angular velocity signal and the filtered acceleration signal based on rigid body kinematics principles; and outputting the filtered acceleration signal when the mean square error is less than the mean square error threshold. The system uses the real acceleration signal and the filtered angular velocity signal as the real angular velocity signal; it outputs the real signal and calculates the difference between the signals before and after filtering to obtain the cross-axis interference signal. It then uses an error mapping model for quantification analysis and source tracing, including: the difference between the pre-processed acceleration signal and the real acceleration signal is the actual acceleration cross-axis interference signal; the difference between the pre-processed gyroscope signal and the real angular velocity signal is the actual gyroscope cross-axis interference signal; it extracts the peak value, effective value, main frequency, harmonic component ratio, and inter-axis coupling coefficient of the interference signal to form a quantization error report; it associates the quantized interference signal with the spatial coordinates of the resonant sensitive area in the feature set to determine the IMU mechanical structure resonant area and resonant frequency corresponding to each actual cross-axis error, generating an error source location list.

2. The method for error identification of an inertial measurement unit according to claim 1, characterized in that: By simulating and calculating the mode shapes and mode amplitudes at different frequencies, several resonant sensitive regions where the mode amplitudes are greater than the preset mode amplitude threshold are located; N piezoelectric thin film sensors are deployed in each sensitive region according to a 2×2 spatial gradient.

3. The method for error identification of an inertial measurement unit according to claim 2, characterized in that: Simultaneously acquire multi-source signals, including vibration signals, acceleration signals, and gyroscope signals. Preprocess the acquired multi-source raw signals: use the moving average method to adaptively eliminate the DC component of the vibration signals, acceleration signals, and gyroscope signals; use a 5th-order Butterworth low-pass filter to filter the multi-source signals. After preprocessing, output a standardized set of vibration signals, preprocessed acceleration signals, and preprocessed gyroscope signals.

4. The method for error identification of an inertial measurement unit according to claim 1, characterized in that: The denoising method using short-time Fourier transform (STFT) and multi-channel time-frequency matrix fusion is as follows: STFT is applied to each channel of the standardized vibration signal set, using a Hanning window and variable window length strategy. The window length for the resonant prediction frequency band is set to 128 points, and the window length for the non-resonant prediction frequency band is set to 256 points. The overlap rate of all frequency band signals is uniformly set to 60%, resulting in N high-resolution time-frequency matrices. The covariance matrix between any two channel time-frequency matrices is calculated, and the covariance coefficients between each channel are extracted from the covariance matrix. Effective signal channels with covariance coefficients greater than the coefficient threshold are selected, and a weighted average fusion algorithm is used to fuse the effective channel time-frequency matrices.

5. The method for error identification of an inertial measurement unit according to claim 1, characterized in that: Initialize the adaptive notch filter bank, specifically: initialize m notch filters based on the spatial frequency feature set, align the center frequency of each filter with the resonant frequency of the corresponding resonant mode, and adopt error weighting for adaptive filter bandwidth; set the inter-axis cooperative suppression weight of the filter bank according to the inter-axis correlation relationship corresponding to the reference error of each mode in the spatial frequency feature set.

6. An error identification system for inertial measurement units, used to implement the method according to any one of claims 1 to 5, characterized in that: include: The signal acquisition module uses finite element simulation to perform modal analysis on the inertial measurement unit, locates the resonant sensitive area and deploys a distributed sensor array, synchronously acquires multi-source signals, and outputs a standardized signal set after preprocessing the signals. The resonant mode feature extraction module performs short-time Fourier transform and multi-channel time-frequency matrix fusion denoising on the standardized signal set. It constructs and trains a dedicated sparse dictionary based on finite element mode parameters, filters candidate resonant mode features through sparse decomposition, trains a BP neural network in conjunction with a hardware-in-the-loop simulation platform, constructs a resonant mode cross-axis error mapping model, and outputs a spatial frequency feature set. The error filtering and source tracing module initializes an adaptive notch filter bank based on the spatial frequency feature set, performs collaborative filtering and motion consistency verification on the signal, outputs the real signal, calculates the difference between the signal before and after filtering to obtain the cross-axis interference signal, combines the error mapping model for quantitative analysis and source tracing, and generates a list of error source locations.

Images