Zero-speed detection method based on dynamic weighted generalized likelihood ratio detection

Abstract

The invention belongs to the technical field of inertial navigation, and particularly provides a zero-speed detection method based on dynamic weighted generalized likelihood ratio detection, which is used for solving the problem of insufficient precision of traditional zero-speed detection and improving the precision of systems such as pedestrian inertial navigation. The method comprises the following steps: firstly, establishing a zero-speed (H0) and non-zero-speed (H1) binary hypothesis test model based on a hypothesis that sensor data obeys Gaussian distribution within a short time; deriving a generalized likelihood ratio test (GLRT) statistic of each single sensor; secondly, dynamic adjustable weight factors are introduced to different sensor shafts, the GLRT statistics of all the single sensor shafts are subjected to weighted fusion according to weights, and weighted GLRT statistics are constructed to serve as a zero-speed judgment basis; through dynamic weighted fusion, the zero-speed detection accuracy is improved, the misjudgment rate is reduced, a more reliable opportunity is provided for navigation error correction, and the navigation error under the pure inertial navigation condition is effectively relieved.

Description

Technical Field

[0001] This invention belongs to the field of inertial navigation and integrated positioning technology, specifically relating to a zero-velocity detection method based on multi-source sensor reliability difference modeling. For navigation scenarios of typical biological motion carriers such as foot-strapped pedestrian navigation systems and bipedal robot autonomous localization, a zero-velocity detection method based on the Dynamically Weighted Generalized Likelihood Ratio Test (WGLRT) is proposed. This method effectively solves the zero-velocity detection bias problem caused by the heterogeneity of inertial sensor arrays by constructing a two-level weighted architecture of prior type weights and dynamic noise weights, significantly improving the zero-velocity interval detection accuracy in complex motion scenarios. Background Technology

[0002] Inertial Navigation Systems (INS) integrate autonomous sensors such as accelerometers, gyroscopes, and magnetometers to independently calculate the position and attitude information of a vehicle. They possess strong anti-interference capabilities and are a core positioning method in GNSS-denied environments (such as indoors and underground). However, the inherent error accumulation effects of INS (such as accelerometer bias and gyroscope angle random walk) cause position drift to increase cubically over time, severely limiting their long-term independent positioning capabilities.

[0003] In the field of foot-attached pedestrian inertial navigation, researchers have proposed Zero Velocity Update (ZUPT) to suppress the accumulation of systematic errors, specifically addressing the characteristics of the foot-on-ground stationary phase during the gait cycle. This technique detects the zero-velocity interval formed by foot contact in real time. Based on the theoretical constraint of a stationary vehicle, it resets the velocity error observations output by the inertial measurement unit (IMU) to zero, generating an error compensation value that is fed back to the navigation calculation module, forming a closed-loop correction mechanism. Studies show that the accuracy of zero-velocity detection directly determines the effectiveness of ZUPT, and its optimization can significantly improve long-term independent positioning accuracy.

[0004] Zero velocity detection is a core component of ZUPT, and existing methods mainly include: thresholding based on single-sensor signals (amplitude, energy, or variance of acceleration / angular velocity signals), multi-sensor combination judgment methods and generalized likelihood ratio detection (GLRT), and machine learning methods. However, single-sensor thresholding has inherent defects due to threshold rigidity—too high a threshold easily misjudges motion as zero velocity (false detection), while too low a threshold ignores the true stationary region (false detection). Although multi-sensor fusion methods theoretically improve detection performance, existing methods usually assume that data from different sensors have equal reliability and cannot adaptively adjust the weight of each sensor in zero velocity detection, resulting in significant fluctuations in detection performance under different walking scenarios (normal walking, running, climbing stairs). Machine learning methods face problems such as heavy reliance on high-quality labeled data, limited generalization ability, contradiction between model complexity and real-time requirements, and poor interpretability leading to debugging difficulties. Summary of the Invention

[0005] This invention addresses the shortcomings of traditional multi-sensor GLRT detection methods—namely, the failure to quantify the confidence differences between different sensors (such as accelerometers / gyroscopes) and across different axes of the same sensor in zero-velocity detection, and the lack of consideration for real-time changes in sensor confidence during motion. It proposes a zero-velocity detection method based on dynamically weighted generalized likelihood ratio (GLRT) detection. This method constructs a two-level weighting mechanism (prior type weight + dynamic noise weight) by dynamically assigning confidence weights to each sensor and its axis. This effectively suppresses false zero-velocity detection caused by differences in sensor performance, significantly improving the navigation and positioning accuracy of legged vehicles such as bipedal robots and pedestrian navigation systems in complex motions.

[0006] The technical solution adopted in this invention is as follows:

[0007] A zero-rate detection method based on dynamic weighted generalized likelihood ratio detection includes the following steps:

[0008] Step 1, Calculation of static parameters of multi-source sensors: In a controllable static calibration environment, synchronously acquire the triaxial (x, y, z) raw output data of gyroscope G, accelerometer A, and magnetometer M, and calculate the static mean value of each axis of each sensor. With variance s∈{G,A,M},i∈{x,y,z};

[0009] Step 2, Real-time generalized likelihood ratio calculation: During the operation of the navigation system, real-time data streams from each sensor are continuously collected. A generalized likelihood ratio detection statistic Λ is constructed for each axial data of each sensor. s,i ;

[0010] Step 3, Dynamic weighting coefficient generation: Based on the initial prior weights assigned to each axis for each sensor. And the weight correction amount calculated based on real-time data. The real-time weighting coefficient ω for each axis of each sensor was calculated. s,i ;

[0011] Step 4, Weighted Fusion and Decision: Based on the generalized likelihood ratio detection statistic Λ of the three sensors calculated in Step 2. s,i and the corresponding real-time weighting coefficient ω in step 3 s,i The weighted fusion is performed to generate the final decision value Λ; and Λ is compared with the detection threshold γ to determine whether the carrier is in a static or dynamic state.

[0012] Furthermore, in step 2, the generalized likelihood ratio detection statistic Λ for each axis of data from each sensor... s,i The calculation formula is:

[0013]

[0014] In the formula, N is the number of samples contained in the sliding detection window during the operation of the navigation system; The variance of the sample data is calculated using the following formula:

[0015]

[0016] In the formula, The sample mean is calculated using the following formula:

[0017]

[0018] Furthermore, the real-time weighting coefficient ω in step 3 s,i The calculation formula is:

[0019]

[0020] in, The prior weights are set based on the inherent characteristics of the sensor, satisfying The dynamic weight correction amount, calculated based on real-time collected sample data, is given by the following formula:

[0021]

[0022] Where, α s,i The reliability degradation factor is defined as:

[0023]

[0024] Here, exp(·) represents the exponential function.

[0025] Furthermore, the formula for calculating the final judgment value Λ in step 4 is as follows:

[0026]

[0027] The innovative advantages of the technical solution of this invention are reflected in:

[0028] This invention overcomes the limitations of traditional GLRT zero-velocity detection methods that assume equal confidence levels for different types of sensors and different axes of the same sensor in zero-velocity detection. By combining prior sensor characteristics with real-time motion states, it achieves dynamic allocation of weights for each sensor and axis. Secondly, it proposes a reliability attenuation model based on noise power ratio to automatically suppress weight coefficients under abnormal sensor conditions. Experiments show that this method reduces zero-velocity detection errors by over 30% compared to traditional GLRT in dynamic scenarios such as climbing stairs and running, significantly improving the navigation and positioning accuracy of legged motion vehicles. Attached Figure Description

[0029] Figure 1 This is a schematic diagram illustrating the implementation principle of an embodiment of the present invention.

[0030] Figure 2 This is a diagram illustrating the 3D trajectory effect of inertial navigation based on zero-velocity detection (ZUPT) in a stair-climbing scenario according to an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments.

[0032] like Figure 1 As shown, this invention provides a zero-rate detection method based on dynamic weighted generalized likelihood ratio detection, comprising the following core steps:

[0033] Step 1, Calculation of Static Parameters of Multi-Source Sensors: In a controllable static calibration environment, synchronously acquire the raw output data of the three axes (x, y, z) of the gyroscope (G), accelerometer (A), and magnetometer (M), and calculate the static mean value of each axis of each sensor. With variance To characterize the inherent noise level of the sensor.

[0034] Preferably, in step 1, to ensure the accuracy of variance calculation, the controllable static calibration environment should meet the conditions of no external interference and complete stationary carrier, the data acquisition duration should be no less than 10 seconds, and the sampling frequency should be no less than 100Hz. After acquisition, the variance of each axis for each type of sensor is calculated using the following formula:

[0035]

[0036] In the formula, M represents the total number of data collected under static conditions; Let represent the data collected on the m-th time by the i∈{x,y,z} axis of the sensor s∈{G,A,M}. The average value of the data collected along the i-th axis of the s-th sensor is represented by the following formula:

[0037]

[0038] Step 2, Real-time generalized likelihood ratio calculation: During the operation of the navigation system, real-time data streams from the sensors are continuously collected. For each axial data point from each sensor, construct the generalized likelihood ratio detection (GLRT) statistic Λ. s,i To assess how well the current data fits the quiescent state assumption;

[0039] Preferably, in step 2, the i-axis GLRT statistic Λ of the s-th sensor... s,i The calculation formula is:

[0040]

[0041] In the formula, N represents the number of samples contained in the sliding detection window during the operation of the navigation system. The variance of the sample data is calculated using the following formula:

[0042]

[0043] In the formula, in the formula, The sample mean is calculated using the following formula:

[0044]

[0045] Preferably, the statistic Λ in step 2 s,i It is derived based on the principle of generalized likelihood ratio detection, and the specific steps are as follows:

[0046] (1) Establish a binary hypothesis testing model for determining the motion state of the carrier; the binary hypothesis testing model is as follows:

[0047]

[0048] In the formula, H0 represents rest, H1 represents motion, and N represents a Gaussian distribution. and These represent the mean and variance of the Gaussian distribution under hypotheses H0 and H1, respectively. In this invention, Assuming it is a known quantity, which has already been obtained in a controllable static environment, and and All of these are unknown quantities.

[0049] Based on the hypothesis testing model described above, the probability density function of the i-th axis data of the s-sensor under hypotheses H0 and H1 can be expressed as:

[0050]

[0051] In the formula, The data represents the sliding window data of the i-th axis of the s-sensor; the symbol Π represents multiplication.

[0052] (2) Derive the single-axis GLRT statistic Λ for each sensor s,i The GLRT statistic for the i-th axis of the s-th sensor can be expressed as:

[0053]

[0054] According to the monotonically increasing property of logarithmic functions, Λ s,i It can be expressed in logarithmic form, and its statistical properties remain unchanged. In this case, Λ s,i Equivalent to:

[0055]

[0056] Regarding Λ s,i For unknown parameters, their maximum likelihood estimates are used instead. The maximum likelihood estimate is:

[0057]

[0058] Λ s,i The generalized likelihood ratio test statistic Λ is obtained by replacing the unknown variable with its maximum likelihood estimate and simplifying the result. s,i .

[0059] Step 3, Dynamic weighting coefficient generation: Based on the initial prior weights assigned to each axis of each sensor. And the weight correction amount calculated based on real-time data. The real-time weighting coefficient ω for each axis of each sensor was calculated. s,i ;

[0060] Preferably, the real-time weighting coefficient ω in step 3 s,i The calculation formula is:

[0061]

[0062] in, The prior weights are set based on the inherent characteristics of the sensor, satisfying like The dynamic weight correction amount, calculated based on real-time collected sample data, is given by the following formula:

[0063]

[0064] Where, α s,i The confidence decay factor is defined as:

[0065]

[0066] Where exp(·) represents the exponential function; when the instantaneous noise of a certain sensor axis deviates significantly from the static calibration value (e.g., the accelerometer is subjected to an impact or the magnetometer is subjected to a disturbance), α s,i This reduces the weighting coefficients of the corresponding axes, causing them to decrease automatically.

[0067] Step 4, Weighted Fusion and State Determination: Based on the generalized likelihood ratio detection statistic Λ of the three sensors calculated in Step 2. s,i and the dynamic weighting coefficient ω in step 3 s,i The weighted summation is performed to obtain the weighted generalized likelihood ratio statistic Λ. The weighted GLRT statistic Λ is compared with the preset detection threshold γ to determine whether the carrier is in a static or dynamic state.

[0068] Preferably, the carrier motion state discrimination model in step 4 can be expressed as:

[0069]

[0070] Preferably, the detection threshold γ in step 4 is determined by the Monte Carlo method, specifically: in a controllable static calibration environment, the false alarm probability P is calculated. fa =P(Λ>γ|H0), by choosing an appropriate value of γ, the actual P fa Myopia equals the set value (typical value is 10). -3 ).

[0071] Figure 2 This is a three-dimensional trajectory effect diagram of inertial navigation based on zero velocity detection (ZUPT) in the scenario of climbing stairs according to an embodiment of the present invention. It shows the motion trajectory when using this technology for walking navigation up stairs, including the starting point, the ending point and the actual motion path, intuitively demonstrating the effect of zero velocity detection in inertial navigation applications.

Claims

1. A zero-rate detection method based on dynamic weighted generalized likelihood ratio detection, characterized in that, Includes the following steps: Step 1, Calculation of static parameters of multi-source sensors: In a controllable static calibration environment, synchronously acquire the triaxial (x, y, z) raw output data of gyroscope G, accelerometer A, and magnetometer M, and calculate the static mean value of each axis of each sensor. With variance s∈{G,A,M},i∈{x,y,z}; Step 2, Real-time generalized likelihood ratio calculation: During the operation of the navigation system, real-time data streams from each sensor are continuously collected. A generalized likelihood ratio detection statistic Λ is constructed for each axial data of each sensor. s,i ; Step 3, Dynamic weighting coefficient generation: Based on the initial prior weights assigned to each axis for each sensor. And the weight correction amount calculated based on real-time data. The real-time weighting coefficient ω for each axis of each sensor was calculated. s,i ; Step 4, Weighted Fusion and Decision: Based on the generalized likelihood ratio detection statistic Λ of the three sensors calculated in Step 2. s,i and the corresponding real-time weighting coefficient ω in step 3 s,i The weighted fusion is performed to generate the final decision value Λ; and Λ is compared with the detection threshold γ to determine whether the carrier is in a static or dynamic state.

2. The zero-rate detection method based on dynamic weighted generalized likelihood ratio detection according to claim 1, characterized in that, In step 2, the generalized likelihood ratio detection statistic Λ for each axis of data from each sensor s,i The calculation formula is: In the formula, N is the number of samples contained in the sliding detection window during the operation of the navigation system; The variance of the sample data is calculated using the following formula: In the formula, The sample mean is calculated using the following formula:

3. The zero-rate detection method based on dynamic weighted generalized likelihood ratio detection according to claim 2, characterized in that, The real-time weighting coefficient ω in step 3 s,i The calculation formula is: in, The prior weights are set based on the inherent characteristics of the sensor, satisfying The dynamic weight correction amount, calculated based on real-time collected sample data, is given by the following formula: Where, α s,i The reliability degradation factor is defined as: Here, exp(·) represents the exponential function.

4. The zero-rate detection method based on dynamic weighted generalized likelihood ratio detection according to claim 1, characterized in that, The formula for calculating the final judgment value Λ in step 4 is as follows:

Images