Acceleration sensor signal correction method based on zero offset correction
By simultaneously acquiring multiple data sources for high-order signal feature extraction and dynamic zero-bias prediction, the zero-bias drift problem of accelerometers in dynamic environments was solved, achieving accurate calibration and improving measurement accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QINGDAO ZITN MICROELECTRONICS CO LTD
- Filing Date
- 2026-03-24
- Publication Date
- 2026-05-19
AI Technical Summary
Accelerometers exhibit inaccurate modeling of time-varying nonlinear zero-bias drift in dynamic environments, and their correction lags and lacks adaptability under complex motion and environmental coupling, affecting measurement accuracy and stability.
By synchronously collecting gyroscope angular velocity, GNSS displacement information and environmental variable data, and performing bandpass filtering, high-order signal features are extracted. The dynamic zero-bias prediction model is used for weighted filtering, and the dynamic zero-bias prediction results and weighted zero-bias filtering results are fused to output the corrected zero bias.
It improves the real-time performance and accuracy of zero bias estimation, enhances the measurement precision and stability of the accelerometer, and enables precise online correction of zero bias drift.
Smart Images
Figure CN121899438B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of accelerometer signal correction, and in particular to an accelerometer signal correction method based on zero-bias drift correction. Background Technology
[0002] Accelerometers, as core components of inertial measurement units, are widely used in tasks such as motion capture, attitude estimation, vibration detection, and navigation and positioning. However, in practical applications, the output signal of accelerometers is affected by various error sources, especially zero-bias drift, which seriously affects measurement accuracy and reliability. Zero-bias drift refers to the non-zero, slowly changing deviation of the sensor output when there is no external acceleration input. It may be caused by internal temperature changes, circuit noise, material aging, and dynamic changes in the external environment such as motion state, mechanical vibration, and electromagnetic interference. Under long-term operation or drastic environmental fluctuations, zero-bias drift may lead to the accumulation of integral errors, further causing significant deviations in velocity and position estimation. In real dynamic environments, abrupt changes in motion state, temperature fluctuations, and multi-axis coupling effects cause zero bias to exhibit obvious time-varying and nonlinear characteristics. Traditional zero-bias correction relies on static calibration or fixed-parameter filtering, which is difficult to adapt to complex changes and does not fully solve the problem of abrupt changes and hysteresis response of zero bias under complex motion and environmental coupling. It cannot achieve accurate online correction of zero-bias drift, thus affecting the measurement accuracy and stability of accelerometers in dynamic scenarios.
[0003] At present, the relevant technologies have technical problems such as inaccurate modeling of time-varying nonlinear zero-bias drift of accelerometers, and lagging correction and insufficient adaptability under complex motion and environmental coupling. Summary of the Invention
[0004] This application provides an accelerometer signal correction method based on zero-bias drift correction, which solves the technical problems in the prior art such as inaccurate modeling of time-varying nonlinear zero-bias drift in accelerometers, and lag and insufficient adaptability in correction under complex motion and environmental coupling. It achieves the technical effect of improving the real-time performance and accuracy of zero-bias estimation and enhancing the measurement accuracy and stability of accelerometers.
[0005] This application provides an accelerometer signal correction method based on zero-bias drift correction. The method includes: reading the raw signal from the accelerometer and simultaneously acquiring gyroscope angular velocity, GNSS displacement information, and environmental variable data; performing bandpass filtering on the raw signal to establish a preprocessed acceleration signal; calculating the angular velocity amplitude and velocity change rate using the gyroscope angular velocity and GNSS displacement information, and configuring a motion state vector; extracting high-order signal features from the preprocessed acceleration signal to establish first input data, wherein the high-order signal feature extraction includes instantaneous signal entropy, power spectrum centroid drift, and sliding window statistics; extracting the motion state vector and the environmental variable data as second input data; inputting the first input data and the second input data into a dynamic zero-bias prediction model and outputting a dynamic zero-bias prediction result; configuring the length of the filtering sliding window according to the motion state vector and autocorrelation, performing weighted zero-bias filtering within the filtering sliding window, fusing the dynamic zero-bias prediction result and the weighted zero-bias filtering result, and outputting a corrected zero-bias.
[0006] In a possible implementation, the accelerometer signal correction method based on zero-bias drift correction further performs the following processing: after selecting a sliding window, a sliding window signal segment is extracted within the sliding window based on an instantaneous time node; the sliding window signal segment is subjected to a short-time Fourier transform to obtain the frequency band energy of each frequency band; the instantaneous signal entropy is calculated based on the ratio of the frequency band energy to the total energy as a probability; and the instantaneous signal entropy is output as a higher-order signal feature.
[0007] In a possible implementation, the accelerometer signal correction method based on zero-bias drift correction further performs the following processing: performing a fast Fourier transform on the sliding window signal segment to construct a power spectrum; and calculating the centroid of the power spectrum using the following formula:
[0008] ;
[0009] in, Representing an instantaneous time node, Characterizing frequency, For the signal at frequency The power spectral density value, Characterizing total energy, It represents the weighted frequency sum; outputs the centroid of the sequence spectrum, and outputs the centroid of the sequence spectrum as a higher-order signal feature; performs variance, skewness, and kurtosis calculations on the window signal within the sliding window signal segment, and outputs the sliding window statistics; and outputs the sliding window statistics as a higher-order signal feature.
[0010] In a possible implementation, the accelerometer signal correction method based on zero-bias drift correction further performs the following processing: The first input data and the second input data are interpolated to unify the time grid; the environmental variable data in the second input data is sent to the environmental modulation channel within the multi-source fusion layer to establish environmental bias parameters; the high-order signal features in the first input data are linearly modulated channel-by-channel using the environmental bias parameters to establish a first zero-bias prediction result; the motion state vector in the second input data and the first input data are sent to the motion consistency channel; zero-bias drift detection is performed through the cross-attention module within the motion consistency channel to establish a second zero-bias prediction result; the first input data is sent to the temporal consistency channel; temporal consistency residual analysis is performed under multiple time windows to establish a third zero-bias prediction result; the first zero-bias prediction result, the second zero-bias prediction result, and the third zero-bias prediction result are fused by the multi-source fusion layer to output a dynamic zero-bias prediction result.
[0011] In a possible implementation, the accelerometer signal correction method based on zero-bias drift correction further performs the following processing: configuring the autocorrelation function of the acceleration signal according to the preprocessed acceleration signal, performing window adaptation matching using the autocorrelation function of the acceleration signal and the motion state vector, and establishing a filtering sliding window; extracting signal segments within the filtering sliding window, setting the weighted filter weights using an exponential decay strategy and an autocorrelation strategy, and then performing weighted zero-bias filtering.
[0012] In a possible implementation, the acceleration sensor signal correction method based on zero-bias drift correction further performs the following processing: motion state stability identification is performed based on the preprocessed acceleration signal and motion state vector; fusion coefficients are configured using the motion state stability identification results; dynamic zero-bias prediction results and weighted zero-bias filtering results are fused using the fusion coefficients to output corrected zero-bias.
[0013] In a possible implementation, the accelerometer signal correction method based on zero-bias drift correction further performs the following processing: performing drift level trigger analysis of zero-bias drift based on the corrected zero bias, and establishing a drift level trigger record; performing continuous cumulative analysis on the drift level trigger record, and reporting a drift anomaly warning.
[0014] In a possible implementation, the accelerometer signal correction method based on zero-bias drift correction also performs the following processing: the environmental variable data includes temperature data, vibration amplitude, and power supply voltage data.
[0015] This application proposes an accelerometer signal correction method based on zero-bias drift correction. The method reads the original accelerometer signal and simultaneously acquires gyroscope angular velocity, GNSS displacement information, and environmental variable data. It then uses bandpass filtering to establish a preprocessed acceleration signal; calculates the angular velocity amplitude and velocity change rate; configures the motion state vector; extracts higher-order signal features; extracts the second input data; outputs the dynamic zero-bias prediction result using a dynamic zero-bias prediction model; configures the filtering sliding window length for weighted zero-bias filtering; and fuses the corrected zero-bias output. This method solves the technical problems of inaccurate modeling of time-varying nonlinear zero-bias drift in accelerometers, and the resulting correction lag and insufficient adaptability under complex motion and environmental coupling conditions. It achieves the technical effects of improving the real-time performance and accuracy of zero-bias estimation, and enhancing the measurement accuracy and stability of accelerometers. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.
[0017] Figure 1 This is a flowchart illustrating the acceleration sensor signal correction method based on zero-bias drift correction provided in an embodiment of this application.
[0018] Figure 2 This is a schematic diagram of the process for establishing the first input data in the accelerometer signal correction method based on zero-bias drift correction provided in the embodiments of this application. Detailed Implementation
[0019] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.
[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0021] In the following description, references to "some embodiments" describe a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same or different subsets of all possible embodiments and can be combined with each other without conflict. The terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, product, or server that includes a series of steps is not necessarily limited to those steps explicitly listed, but may include other steps not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.
[0022] This application provides an accelerometer signal correction method based on zero-bias drift correction, such as... Figure 1 As shown, the method includes:
[0023] Step S100: Read the raw signal from the accelerometer and simultaneously collect gyroscope angular velocity, GNSS displacement information and environmental variable data.
[0024] Step S100 further includes the environmental variable data including temperature data, vibration amplitude, and power supply voltage data.
[0025] Preferably, the discrete-time voltage value sequence is read from the analog-to-digital converter of the accelerometer chip as the raw signal of the accelerometer sensor, containing interference such as zero bias, noise, scale factor error, and cross-axis sensitivity, to provide the most complete acceleration information; and gyroscope angular velocity, GNSS displacement information, and environmental variable data are acquired simultaneously. Synchronous acquisition means aligning the timestamps of several types of data. Specifically, gyroscope angular velocity refers to the angular velocity value read from the gyroscope chip, reflecting the device's rotation around the X, Y, and Z axes, used to determine the motion state, such as calculating the angular velocity amplitude to determine whether the device is rotating and the degree of rotation, and correlating with accelerometer data to identify zero bias estimation errors; GNSS displacement information refers to the velocity and position changes obtained from the Global Navigation Satellite System receiver, to provide... A low-speed, high-precision reference standard can be used to determine a relatively reliable external acceleration estimate by calculating the rate of change of velocity. This estimate is then compared with the value read by the accelerometer, and the difference includes the accelerometer's zero-bias error. Environmental variable data refers to the physical quantities of the working environment and conditions of the accelerometer, including temperature data, vibration amplitude, and power supply voltage data. Among these, the resistance of semiconductor materials and the stress characteristics of silicon lattices are extremely sensitive to temperature, and the zero bias will drift significantly with temperature changes. Strong vibrations can cause the sensor to enter the nonlinear operating region or induce resonance in the internal structure, leading to an increase in output error. Monitoring the vibration amplitude can identify periods of low reliability. Fluctuations in power supply voltage data directly affect the stability of the sensor's internal reference voltage and amplification circuit, thus introducing additional zero-bias variations.
[0026] Step S200: After bandpass filtering the original signal, a preprocessed acceleration signal is established.
[0027] Preferably, a digital bandpass filter is used to filter the original signal sequence to retain signal components within a specific frequency band while suppressing noise and interference outside that band. Specifically, the high-pass filter removes extremely low-frequency zero-bias drift and DC components, separating the slowly changing zero bias from the dynamic acceleration signal, resulting in a clear zero-bias target and a clean pre-processed acceleration signal. The low-pass filter removes high-frequency noise and interference, preventing noise from interfering with feature extraction and improving the signal-to-noise ratio. The passband frequency range of the filter is preset according to the application scenario of the accelerometer. For example, when used for human motion capture, the effective signal is usually between 0.1Hz and 20Hz; when used for mechanical vibration monitoring, it may be between 10Hz and 1000Hz. The final output is a pre-processed acceleration signal, i.e., a bandpass-filtered, centered acceleration signal sequence, in which high-frequency noise and low-frequency drift are significantly suppressed, ensuring that the extracted features can truly reflect the motion state and improving the signal-to-noise ratio of the accelerometer signal.
[0028] Step S300: Calculate the angular velocity amplitude and velocity change rate using the gyroscope angular velocity and GNSS displacement information, and configure the motion state vector.
[0029] Preferably, the 2-norm of the gyroscope's angular velocity is calculated as the angular velocity amplitude, representing the degree of rotation of the device as a whole. An angular velocity amplitude close to 0 indicates that the device is stationary or almost without rotation, indicating a stable motion state. A larger angular velocity amplitude indicates that the device is rotating rapidly or jittering, indicating an unstable motion state. The velocity signal is determined based on GNSS displacement information, and the velocity change rate is calculated by numerical differentiation. Then, the angular velocity amplitude and velocity change rate are combined into a motion state vector to represent the current motion state. This vector is used to adaptively configure the length of the filtering sliding window and determine whether the current motion state is stationary, low-speed motion, high-speed motion, uniform speed, acceleration, deceleration, or violent rotation / jittering. The zero-drift characteristics and reliable correction strategies differ under different motion states.
[0030] Step S400: After extracting high-order signal features from the preprocessed acceleration signal, first input data is established. The high-order signal feature extraction includes instantaneous signal entropy, power spectrum centroid drift, and sliding window statistics.
[0031] Step S400 further includes step S410, after selecting the sliding window, extracting a sliding window signal segment within the sliding window based on the instantaneous time node; step S420, performing a short-time Fourier transform on the sliding window signal segment to obtain the frequency band energy of each frequency band; step S430, calculating the instantaneous signal entropy based on the ratio of the frequency band energy to the total energy as a probability; and step S440, outputting the instantaneous signal entropy as a higher-order signal feature.
[0032] Preferably, high-order signal feature extraction is performed on the preprocessed acceleration signal, including instantaneous signal entropy, power spectrum centroid drift, and sliding window statistics. Specifically, selecting a sliding window means determining a window of fixed length to slide along the time axis of the acceleration signal, dividing the non-stationary continuous-time signal into multiple short time intervals that can be considered locally stationary for time-frequency analysis. The window length determines the trade-off between time resolution and frequency resolution. Then, based on the instantaneous time node, a sliding window signal segment is extracted within the sliding window. That is, for the instantaneous time node where features need to be calculated, the acceleration signal sequence within the window length is extracted with that node as the center, and the discrete signal segment is output as the sliding window signal segment. Then, a short-time Fourier transform is performed on the sliding window signal segment, that is, a Fourier transform is performed within this local window to obtain the local spectrum of the signal during that time interval. The entire spectrum is then divided into several... For each frequency band or frequency zone, the amplitudes of all frequency components within that band are squared and summed to calculate the energy of each band. Then, the signal power spectrum is treated as a probability distribution, and the proportion of energy in each band to the total energy is defined as the probability of that band's occurrence. The more concentrated the energy in a band, the higher its probability. The Shannon entropy formula is then used to calculate the entropy of the probability distribution across the entire spectrum, determining the instantaneous signal entropy. This entropy is used to quantify the frequency component complexity of the acceleration signal over a short period near the current moment. A low entropy value indicates that the signal energy is highly concentrated in a few frequency bands; for example, when the device is stationary or moving at a constant speed, the signal is mainly low-frequency noise, with a concentrated spectrum and low entropy. A high entropy value indicates that the signal energy is evenly distributed across multiple frequency bands, with a flat and chaotic spectrum; for example, when the device is experiencing violent, irregular shaking or collisions, the signal contains a large number of broadband components, resulting in high entropy. Finally, the instantaneous signal entropy is output as a higher-order signal feature.
[0033] Furthermore, step S400 also includes step S450, performing a fast Fourier transform on the sliding window signal segment to construct the power spectrum; step S460, calculating the centroid of the power spectrum using the following formula:
[0034] ;
[0035] in, Representing an instantaneous time node, Characterizing frequency, For the signal at frequency The power spectral density value, Characterizing total energy, Characterize the weighted frequency sum; Step S470, output the centroid of the sequence spectrum and output the centroid of the sequence spectrum as a higher-order signal feature; Step S480, perform variance, skewness, and kurtosis calculations on the window signal within the sliding window signal segment and output the sliding window statistics; Step S490, output the sliding window statistics as a higher-order signal feature.
[0036] Preferably, a Fast Fourier Transform is performed on the sliding window signal segment to obtain its complex spectrum, and then the square of the spectral amplitude is calculated to construct the power spectrum, reflecting the distribution of signal energy at different frequency components; then, the power spectrum is constructed using the formula... Calculate the centroid of the power spectrum, where, Representing an instantaneous time node, Characterizing frequency, For the signal at frequency The power spectral density value, It is the sum of the power spectrum at all frequency points, representing the total energy of the signal segment within the sliding window. It involves weighting the power by frequency and then summing the results. The centroid or average frequency representing the energy of the entire signal in the frequency domain is the instantaneous power spectrum centroid. A lower value indicates that the energy of the signal is mainly concentrated in the low-frequency region, such as slow motion or steady vibration; a higher value indicates that the energy of the signal is concentrated in the high-frequency region, such as sudden impact, high-frequency jitter or sharp collision events; and then outputs the sequence spectrum centroid as a higher-order signal feature.
[0037] Preferably, variance, skewness, and kurtosis are calculated for the window signal within the sliding window signal segment, and sliding window statistics are output. Specifically, the mean of the squares of the deviations of each point in the signal segment from the mean is calculated as the variance, which measures the magnitude of the signal power or energy, representing the degree of dispersion or fluctuation intensity of the signal amplitude around its mean. A large variance indicates a large dynamic range of the signal and drastic amplitude changes, corresponding to strong motion or vibration; a small variance indicates a stable signal with weak changes, usually corresponding to a static or uniform motion state. The mean of the cube of the deviations of each point in the signal segment from the mean is calculated, and then standardized by dividing by the cube of the standard deviation to obtain the skewness, which measures the asymmetry of the signal amplitude distribution. A skewness of approximately 0 indicates that the distribution is symmetrical, similar to a normal distribution; a skewness of non-zero indicates that there are more positive or negative pulses with large amplitudes in the signal. The deviation of each point in the signal segment from the mean is calculated to the fourth power. The mean is then standardized by dividing the standard deviation by the fourth power. The difference is then compared to the kurtosis (3) of a normal distribution as a benchmark to obtain kurtosis, which measures the steepness and tail thickness of the signal amplitude distribution. A kurtosis of approximately 0 indicates that the signal segment's distribution is as steep as a normal distribution; a kurtosis greater than 0 indicates a steeper distribution with a thicker tail, suggesting the presence of large-amplitude impulses or pulses; a kurtosis less than 0 indicates a smoother distribution with a thinner tail, suggesting a more uniform signal. Finally, the sliding window statistics are output as higher-order signal features to more precisely perceive the current motion state and make more accurate zero-bias predictions. The first input data is constructed using instantaneous signal entropy, power spectrum centroid drift, and sliding window statistics to represent various characteristics exhibited by the acceleration signal.
[0038] Step S500: Extract the motion state vector and the environmental variable data as second input data, input the first input data and the second input data into the dynamic zero-bias prediction model, and output the dynamic zero-bias prediction result.
[0039] Preferably, the motion state vector and environmental variable data are fused to construct the second input data, which describes the external factors and states that may cause changes in the zero bias. The dynamic zero bias prediction model may be a neural network or a gradient boosting tree, which is used to learn the mapping relationship between the input data and the current zero bias value. The first input data and the second input data are input into the dynamic zero bias prediction model, and the predicted output is the zero bias estimate of the current accelerometer on the X, Y, and Z axes, which is used as the dynamic zero bias prediction result, thereby achieving more accurate and forward-looking prediction correction.
[0040] Furthermore, such as Figure 2 As shown, step S500 further includes step S510, which involves interpolating the first input data and the second input data to unify the time grid; step S520, which involves sending the environmental variable data in the second input data to the environmental modulation channel within the multi-source fusion layer to establish environmental bias parameters, and using the environmental bias parameters to perform channel-by-channel linear modulation on the high-order signal features in the first input data to establish a first zero-bias prediction result; step S530, which involves sending the motion state vector in the second input data and the first input data to the motion consistency channel, and performing zero-bias drift detection through the cross-attention module within the motion consistency channel to establish a second zero-bias prediction result; step S540, which involves sending the first input data to the temporal consistency channel, performing temporal consistency residual analysis under multiple time windows to establish a third zero-bias prediction result; and step S550, which involves fusing the first zero-bias prediction result, the second zero-bias prediction result, and the third zero-bias prediction result through the multi-source fusion layer to output a dynamic zero-bias prediction result.
[0041] Preferably, the unified time grid refers to resampling the first and second input data to a unified high-frequency time series through linear interpolation or spline interpolation. The dynamic zero-bias prediction model contains three independent sub-networks: an environmental modulation channel, a motion consistency channel, and a temporal consistency channel. Specifically, the environmental variable data in the second input data is sent to the environmental modulation channel in the multi-source fusion layer to learn how the environmental variables affect the zero bias, and then outputs the environmental bias parameter. The environmental bias parameter is then used to perform channel-by-channel linear modulation on the high-order signal features in the first input data, including scaling and offsetting each dimension of the high-order signal features, and outputting a zero-bias prediction value as the first zero-bias prediction result, reflecting the baseline influence of environmental factors on the zero bias.
[0042] Preferably, the motion state vector in the second input data and the first input data are sent to the motion consistency channel. Zero-bias drift detection is performed through the cross-attention module in the motion consistency channel. Specifically, the correlation between motion state and signal features is calculated based on the cross-attention mechanism and the attention weight is determined. Then, the zero-bias drift detection is performed using the motion consistency channel through the attention weight to identify signal features that do not match the current motion state. For example, when the motion state display device is rotating violently, but the acceleration signal features show that the energy is very low, a zero-bias prediction value is output as the second zero-bias prediction result, which mainly captures the zero-bias changes caused by abnormal motion or dynamic coupling.
[0043] Preferably, the first input data is sent to the time-consistency channel to perform time-consistency residual analysis under multiple time windows. The time-consistency channel predicts current features based on historical features. Specifically, it uses time-series models such as recurrent neural networks or Transformers to analyze changes in signal features within recent time windows. The actual feature values are subtracted from the predicted feature values by the time-series model to obtain residuals. For example, zero-bias drift may cause abrupt changes in the temporal evolution of signal features, thus outputting a zero-bias prediction value as the third zero-bias prediction result. This primarily captures unpredictable anomalies that occur during the temporal evolution of the signal itself. Finally, the first, second, and third zero-bias prediction results are weighted and fused using a multi-source fusion layer. This multi-source fusion layer learns the confidence weights of each channel's results under different scenarios, ultimately outputting a dynamic zero-bias prediction result. This is the optimal zero-bias estimate that integrates environmental baseline, motion anomalies, and temporal abrupt changes, ensuring the real-time performance and accuracy of the zero-bias estimate.
[0044] Step S600: After configuring the length of the filtering sliding window according to the motion state vector and autocorrelation, weighted zero-bias filtering is performed within the filtering sliding window. The dynamic zero-bias prediction result and the weighted zero-bias filtering result are fused to output the corrected zero-bias.
[0045] Step S600 further includes step S610, configuring the autocorrelation function of the acceleration signal according to the preprocessed acceleration signal, and using the autocorrelation function of the acceleration signal and the motion state vector to perform window adaptation matching and establish a filtering sliding window; step S620, extracting signal segments within the filtering sliding window, setting the weighted filter weights using the exponential decay strategy and the autocorrelation strategy, and then performing weighted zero-bias filtering.
[0046] Preferably, the autocorrelation function of the preprocessed acceleration signal within a certain time delay range is calculated to quantify the signal's memory length or periodicity. A slower decay of the autocorrelation function indicates a stronger correlation between signal samples and a slower change; a faster decay indicates the signal is closer to uncorrelated noise and changes faster. Then, window adaptation matching is performed using the acceleration signal's autocorrelation function and the motion state vector, dynamically determining the window length of the weighted filter. Specifically, if the autocorrelation function decays slowly, it indicates a high correlation between historical and current data, so a longer window length is matched to utilize more historical data for smoothing and improve estimation accuracy; if the motion state vector shows a state of rapid motion, it indicates rapid signal change, so a shorter window length is matched to ensure the filter can quickly track changes in zero bias; finally, a combined output filter sliding window is used. Next, signal segments are extracted within the filtering sliding window, including a preprocessed acceleration signal truncated forward from the current time to a length equal to the filtering sliding window. Then, weighted filter weights are set using exponential decay and autocorrelation strategies. The exponential decay strategy assigns weights based on the newness of the data signal, with newer data receiving higher weights. The autocorrelation strategy assigns weights based on the correlation between the data point and the signal at the current time, with historical data points having higher correlations to the current signal receiving higher weights. The weights assigned by the two strategies are then multiplied and normalized to obtain the weighted filter weights. Finally, weighted zero-bias filtering is performed, which involves calculating a weighted average of the signal segments within the window and outputting the zero-bias estimate for the current time.
[0047] Furthermore, step S600 also includes step S630, which involves identifying motion state stability based on the preprocessed acceleration signal and motion state vector, and configuring fusion coefficients using the motion state stability identification results; step S640, which involves fusing the dynamic zero-bias prediction results and the weighted zero-bias filtering results using the fusion coefficients, and outputting the corrected zero-bias.
[0048] Preferably, motion state stability is identified based on the preprocessed acceleration signal and motion state vector, and a stability metric is output to quantify the stability of the motion state at the current moment. Specifically, when the variance of the acceleration signal is very small and the values of each component of the motion state vector are all below a preset threshold, it is judged as high stability; when the signal feature value is large, indicating that the device is in a state of violent speed change or rotation, it is judged as low stability, thus obtaining the motion state stability identification result. Then, through a predefined mapping function, a fusion coefficient is calculated based on the motion state stability identification result, where the fusion coefficient is between 0 and 1. For example, high stability indicates that the weighted zero-bias filtering result is accurate and smooth, so the fusion coefficient is configured to approach 0, indicating greater trust in the filtering result; low stability indicates that the motion state is violent and changes rapidly, and historical data may be irrelevant to the current state, and the result of the dynamic zero-bias prediction model responds faster, so the fusion coefficient is configured to approach 1, indicating greater trust in the model prediction result; finally, the fusion coefficient is used to linearly weight and fuse the dynamic zero-bias prediction result and the weighted zero-bias filtering result to output the corrected zero bias, thereby realizing high-precision sensor calibration under various working conditions and enhancing the measurement accuracy and stability of the acceleration sensor.
[0049] Furthermore, step S600 also includes step S650, performing drift level trigger analysis of zero bias drift based on the corrected zero bias, and establishing a drift level trigger record; step S660, performing continuous cumulative analysis on the drift level trigger record, and reporting a drift anomaly warning.
[0050] Preferably, drift level trigger analysis is performed based on the zero bias correction. Specifically, multiple drift level thresholds are preset to classify different drift severity levels, such as level 0 normal, level 1 slight drift, level 2 significant drift, and level 3 severe drift. At each time point or in each calculation cycle, the absolute value of the current zero bias correction is compared with the preset drift level threshold to determine the drift level and establish a drift level trigger record, which includes at least a timestamp, the triggered drift level, and the specific zero bias value. Then, the drift level trigger record is continuously and cumulatively analyzed, including frequency analysis, pattern recognition, and duration analysis. That is, the number of times the drift level is triggered within the set time window is counted, whether the trigger record shows accelerated drift, and whether high-level drift continues for more than the predetermined time limit. When the cumulative analysis result meets multiple preset alarm conditions, a drift anomaly warning is issued. The preset alarm conditions may include frequency exceeding the limit, pattern alarm, or continuous timeout, indicating that the accelerometer has performance degradation or malfunction, thereby realizing fault warning from transient error compensation to long-term sensor health status monitoring.
[0051] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
Claims
1. An accelerometer signal correction method based on zero-bias drift correction, characterized in that, The method includes: Read the raw signal from the accelerometer and simultaneously acquire gyroscope angular velocity, GNSS displacement information and environmental variable data; After bandpass filtering the original signal, a preprocessed acceleration signal is established. The angular velocity amplitude and velocity change rate are calculated using the gyroscope angular velocity and GNSS displacement information, and the motion state vector is configured. After extracting high-order signal features from the preprocessed acceleration signal, the first input data is established. The high-order signal feature extraction includes instantaneous signal entropy, power spectrum centroid drift, and sliding window statistics. The motion state vector and the environmental variable data are extracted as second input data. The first input data and the second input data are input into the dynamic zero-bias prediction model, and the dynamic zero-bias prediction result is output. After configuring the length of the filtering sliding window based on the motion state vector and autocorrelation, weighted zero-bias filtering is performed within the filtering sliding window. The dynamic zero-bias prediction result and the weighted zero-bias filtering result are fused to output the corrected zero-bias. After extracting high-order signal features from the preprocessed acceleration signal, the first input data is established, including: After selecting the sliding window, the sliding window signal segment is extracted within the sliding window based on the instantaneous time node; The sliding window signal segment is subjected to a short-time Fourier transform to obtain the frequency band energy of each frequency band; The instantaneous signal entropy is calculated based on the ratio of the energy in the frequency band to the total energy as a probability. The instantaneous signal entropy is output as a higher-order signal feature; After extracting high-order signal features from the preprocessed acceleration signal, the process of establishing the first input data further includes: Perform a Fast Fourier Transform on the sliding window signal segment to construct the power spectrum; The centroid of the power spectrum is calculated using the following formula: ； in, Representing an instantaneous time node, Characterizing frequency, For the signal at frequency The power spectral density value, Characterizing total energy, Characterized by the weighted frequency sum; Output the centroid of the sequence spectrum and output it as a higher-order signal feature. Variance, skewness, and kurtosis are calculated for the window signal within the sliding window signal segment, and sliding window statistics are output. The sliding window statistics are output as higher-order signal features. The step of inputting the first input data and the second input data into the dynamic zero-bias prediction model and outputting the dynamic zero-bias prediction result includes: The first input data and the second input data are interpolated to form a unified time grid. The environmental variable data in the second input data is sent to the environmental modulation channel in the multi-source fusion layer to establish environmental bias parameters. The high-order signal features in the first input data are linearly modulated channel by channel using the environmental bias parameters to establish the first zero-bias prediction result. The motion state vector in the second input data and the first input data are sent to the motion consistency channel. Zero-bias drift detection is performed through the cross-attention module in the motion consistency channel to establish a second zero-bias prediction result. The first input data is sent to the time consistency channel, and time consistency residual analysis is performed under multiple time windows to establish the third zero-bias prediction result. The first zero-bias prediction result, the second zero-bias prediction result, and the third zero-bias prediction result are fused by a multi-source fusion layer to output a dynamic zero-bias prediction result.
2. The accelerometer signal correction method based on zero-bias drift correction as described in claim 1, characterized in that, The step of configuring the filter sliding window length based on the motion state vector and autocorrelation, and then performing weighted zero-bias filtering within the filter sliding window, includes: Based on the preprocessed acceleration signal, configure the acceleration signal autocorrelation function, and use the acceleration signal autocorrelation function and the motion state vector to perform window adaptation matching and establish a filtered sliding window; Signal segments are extracted within the filtering sliding window. After setting the weights of the weighted filter using the exponential decay strategy and the autocorrelation strategy, weighted zero-bias filtering is performed.
3. The accelerometer signal correction method based on zero-bias drift correction as described in claim 1, characterized in that, The fusion of dynamic zero-bias prediction results and weighted zero-bias filtering results outputs corrected zero-bias, including: Motion state stability is identified based on the preprocessed acceleration signal and motion state vector, and fusion coefficients are configured using the motion state stability identification results. The dynamic zero-bias prediction result and the weighted zero-bias filtering result are fused using the fusion coefficients to output the corrected zero-bias.
4. The accelerometer signal correction method based on zero-bias drift correction as described in claim 1, characterized in that, After the output is corrected to zero bias, it includes: Based on the corrected zero bias, perform drift level trigger analysis of zero bias drift and establish drift level trigger records; The drift level trigger records are continuously and cumulatively analyzed to report a drift anomaly warning.
5. The accelerometer signal correction method based on zero-bias drift correction as described in claim 1, characterized in that, The environmental variable data includes temperature data, vibration amplitude, and power supply voltage data.