A sensor drift correction method and system based on transfer learning
By using a Gaussian process regression model based on transfer learning and a joint decision matrix of slope intercept, the problem of zero-point drift in MEMS sensors is solved, achieving high-precision and adaptive correction, which is applicable to fields such as inertial navigation, attitude monitoring, and structural health monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN BEIDOU COMM TECH CO
- Filing Date
- 2026-03-06
- Publication Date
- 2026-06-19
Smart Images
Figure CN121783233B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of sensors, and more particularly to a sensor drift correction method and system based on transfer learning. Background Technology
[0002] MEMS (Micro-Electro-Mechanical Systems) sensors have been widely used in inertial navigation, attitude monitoring, and structural health monitoring due to their advantages such as small size, low cost, and low power consumption. However, MEMS sensors commonly suffer from zero-point drift during long-term operation, which is the phenomenon where the sensor's output value gradually deviates from the true zero point over time when the sensor is stationary. This drift is mainly caused by the coupling of multiple factors, including temperature changes, mechanical stress relaxation, packaging aging, and environmental factors (humidity, air pressure). Especially after continuous operation across quarters or throughout the year, the cumulative effect of drift is significantly aggravated, seriously affecting measurement accuracy and system reliability.
[0003] Existing correction methods are mainly based on linear regression models of environmental parameters, assuming that drift and environmental factors have a fixed linear relationship. However, long-term operation shows that although the variation of the total modulus of the sensor's three-axis composite vector is related to the average temperature, it cannot be completely regressed by a simple linear model. Parameters such as the temperature coefficient themselves change significantly over time, leading to a sharp drop in compensation accuracy after the quarter ends, and even the phenomenon of increasing deviation with compensation. At the same time, when the monitored object actually tilts or moves, the change in total modulus is physically real. Existing methods cannot effectively distinguish between modulus offset caused by zero-point drift and real motion events. Blind compensation will lead to the erroneous suppression of the real signal, resulting in distorted monitoring data and misjudgment of decisions, which is particularly dangerous in high-reliability scenarios such as structural health monitoring and geological disaster early warning. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a sensor drift correction method and system based on transfer learning, which solves the above problems.
[0005] To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps:
[0006] S1. Collect sensor data, including triaxial acceleration data, temperature, humidity, and air pressure environmental parameters;
[0007] S2. Calculate the magnitude of the triaxial composite vector. ,in Let be the magnitude of the triaxial acceleration vector. Sensor triaxial raw output values, Zero-point drift for each axis;
[0008] S3. Using the Gaussian process regression model as a base, perform source domain pre-training, feature extraction, and regression;
[0009] S4. Employ a transfer learning strategy in the target domain, using the maximum mean difference loss function. To achieve distribution alignment and adaptive migration, This represents the mean squared error loss in Gaussian process regression. For the trade-off hyperparameter of MMD loss, This represents the maximum mean difference between the feature distributions of the source and target domains.
[0010] S5. Distinguish between zero-point drift and real motion using the slope intercept joint judgment matrix. If it is determined to be real motion, trigger environmental baseline reset; otherwise, perform drift compensation and output correction data. ,in These are the sensor's raw triaxial readings. It is the drift compensation amount predicted by the transfer learning model. This is the ambient temperature value at the current sampling time. This is the ambient relative humidity value at the current sampling time. t is the ambient atmospheric pressure value at the current sampling time, and t is the current sampling time.
[0011] Preferably, the feature extraction includes time segmentation and constructing feature vectors. ,in This is the ambient temperature value at the current sampling time. This is the ambient relative humidity value at the current sampling time. This is the ambient atmospheric pressure value at the current sampling time. It is the rate of temperature change. It represents the time of day in which the sampling point is located.
[0012] Preferably, the kernel function of the Gaussian process regression is a combined kernel, including a periodic kernel, a linear kernel, and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends, and short-term random noise, respectively.
[0013] Preferably, the transfer learning strategy includes cold start and hot update, using Adaptive BatchNormalization to update the mean and variance of the model with the current small sample data to achieve rapid adaptation.
[0014] Preferably, the joint determination matrix of slope and intercept includes regression slope stability. and intercept jump intensity ,in This represents the slope of the linear regression of modulus and temperature for the current time period. This is the difference between the current slope and the initial slope k_0 of the source domain. The standard deviation of the historical slope of the source domain. This represents the regression intercept of modulus length versus temperature for the current time period. This is the difference between the current intercept and the initial intercept b_0 of the source domain. The average value of the historical intercept of the source domain, if and ,but ,in yes The threshold, yes The threshold.
[0015] Preferably, the threshold range is adjusted according to the sensor noise level and application scenario. The typical threshold range is 0.10 to 0.20. The typical threshold range is 2.5–4.0; when When the value is less than 0.15 and D > 3.0, it is considered a real motion.
[0016] A sensor drift correction system based on transfer learning includes:
[0017] The data acquisition module is used to collect raw data from the sensors and environmental parameters;
[0018] The feature extraction module is used to construct feature vectors and perform time-segment alignment.
[0019] The regression model module constructs a zero-point mapping based on Gaussian process regression.
[0020] The transfer learning module uses the maximum mean difference loss function to align the distributions of the source and target domains.
[0021] The judgment module distinguishes between drift and real motion through a joint judgment matrix of slope intercept and outputs correction data.
[0022] Preferably, the feature extraction module utilizes the stable regression coefficients of data from the same time period each day to perform spatiotemporal data cleaning, including time period slicing and feature vector construction. The feature vector contains temperature, humidity, air pressure, temperature change rate, and time period information.
[0023] Preferably, the kernel function of the regression model module is designed as a combined kernel, including a periodic kernel, a linear kernel, and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends, and short-term random noise, respectively; in the cold start and hot update strategy, the transfer learning module uses AdaBN to update the mean and variance of the model with the current small sample data to achieve rapid adaptation.
[0024] Preferably, it also includes an online monitoring module, which calculates the regression parameters for the same time period every 24 hours and calculates the KL divergence between the current distribution and the source domain distribution. If the divergence exceeds the threshold, migration compensation is initiated and corrected data is output.
[0025] This invention provides a sensor drift correction method and system based on transfer learning. Compared with existing technologies, it has the following advantages:
[0026] 1. In this invention, a zero-point mapping model is constructed by using Gaussian process regression. The combined kernel function of periodic kernel, linear kernel and RBF kernel can simultaneously capture the periodic fluctuations of environmental factors, the long-term trend of sensor aging and short-term random noise. It has a stronger nonlinear fitting ability than the traditional linear temperature compensation method. The maximum mean difference (MMD) loss function is used to achieve adaptive distribution alignment between the source domain and the target domain, so that the model can quickly adapt to the target domain. After 4500 hours of operation across quarters, the modulus deviation decreased from 0.08g to less than 0.015g, an improvement of 81%. The temperature-related error coefficient decreased from 0.45mg / ℃ to 0.09mg / ℃, an improvement of about 80%. The Allan deviation stability increased from 15μg to 4μg. At the same time, the slope intercept joint judgment matrix is introduced to achieve accurate differentiation between zero-point drift and real motion and tilt events, with an accuracy of over 96%. It effectively avoids measurement distortion caused by miscompensation and significantly improves the reliability and data credibility of the system.
[0027] 2. In this invention, regression parameters and KL divergence are calculated every 24 hours, and AdaBN technology is used to achieve rapid updates of the model mean and variance. Cold start and hot update can be completed without manual intervention, and the model adaptation time is shortened to within 24 hours, which is significantly more efficient than traditional recalibration methods. This method is not only applicable to single-sensor scenarios, but can also be extended to multi-sensor array collaborative calibration. The robustness of the system is further improved through consistency verification between nodes. In addition, the comprehensive utilization of multi-dimensional environmental parameters such as temperature, humidity, and air pressure in this invention enables it to maintain high-precision calibration effects in complex environments (such as outdoor meteorological monitoring, marine platforms, and infrastructure health monitoring). It is easy to integrate into existing sensor embedded systems and provides an efficient, economical, and scalable solution for applications that require long-term stable measurement, such as inertial navigation, attitude monitoring, and structural health monitoring. It has important engineering practical value and industrialization prospects. Attached Figure Description
[0028] Figure 1 This is a flowchart illustrating the sensor drift correction method and system based on transfer learning proposed in this invention. Detailed Implementation
[0029] 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.
[0030] Please see Figure 1 The present invention provides the following technical solutions, specifically including the following embodiments:
[0031] Example 1:
[0032] A sensor drift correction method based on transfer learning includes the following steps:
[0033] S1. Collect sensor data, including triaxial acceleration data, temperature, humidity, and air pressure environmental parameters;
[0034] S2. Calculate the magnitude of the triaxial composite vector. ,in Let be the magnitude of the triaxial acceleration vector. Sensor triaxial raw output values, Zero-point drift for each axis;
[0035] S3. Using a Gaussian process regression model as a foundation, perform source domain pre-training, feature extraction, and regression. The feature extraction includes time-slicing and constructing feature vectors. ,in This is the ambient temperature value at the current sampling time. This is the ambient relative humidity value at the current sampling time. This is the ambient atmospheric pressure value at the current sampling time. It is the rate of temperature change. It is the time representation of the sampling time in a day. The kernel function of the Gaussian process regression is a combined kernel, including a periodic kernel, a linear kernel, and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends, and short-term random noise, respectively.
[0036] S4. Employ a transfer learning strategy in the target domain, using the maximum mean difference loss function. To achieve distribution alignment and adaptive migration, This represents the mean squared error loss in Gaussian process regression. For the trade-off hyperparameter of MMD loss, To determine the maximum mean difference between the feature distributions of the source domain and the target domain, the transfer learning strategy includes cold start and hot update. Adaptive Batch Normalization is used to update the mean and variance of the model using the current small sample data to achieve rapid adaptation.
[0037] S5. Distinguish between zero-point drift and real motion using the slope intercept joint judgment matrix. If it is determined to be real motion, trigger environmental baseline reset; otherwise, perform drift compensation and output correction data. ,in These are the sensor's raw triaxial readings. It is the drift compensation amount predicted by the transfer learning model, and the joint decision matrix of slope intercept includes the regression slope stability. and intercept jump intensity ,in This represents the slope of the linear regression of modulus and temperature for the current time period. This is the difference between the current slope and the initial slope k_0 of the source domain. The standard deviation of the historical slope of the source domain. This represents the regression intercept of modulus length versus temperature for the current time period. This is the difference between the current intercept and the initial intercept b_0 of the source domain. The average value of the historical intercept of the source domain, if and ,but ,in yes The threshold, yes The threshold value, the range of which is adjusted according to the sensor noise level and the application scenario, The typical threshold range is 0.10 to 0.20. The typical threshold range is 2.5–4.0; when When the value is less than 0.15 and D > 3.0, it is considered a real motion.
[0038] A sensor drift correction system based on transfer learning includes:
[0039] The data acquisition module is used to collect raw data from the sensors and environmental parameters;
[0040] The feature extraction module is used to construct feature vectors and perform time period alignment. The feature extraction module utilizes the stable regression coefficient of data in the same time period of each day to perform spatiotemporal data cleaning, including time period slicing and feature vector construction. The feature vector contains temperature, humidity, air pressure, temperature change rate and time period information.
[0041] The regression model module constructs a zero-point mapping based on Gaussian process regression. The kernel function of the regression model module is designed as a combined kernel, including a periodic kernel, a linear kernel, and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends, and short-term random noise, respectively. In the cold start and hot update strategy, the transfer learning module uses AdaBN to update the mean and variance of the model with the current small sample data to achieve rapid adaptation.
[0042] The transfer learning module uses the maximum mean difference loss function to align the distributions of the source and target domains.
[0043] The judgment module distinguishes between drift and real motion through a joint judgment matrix of slope intercept and outputs correction data.
[0044] It also includes an online monitoring module, which calculates the regression parameters for the same time period every 24 hours and calculates the KL divergence between the current distribution and the source domain distribution. If the divergence exceeds the threshold, migration compensation is initiated and corrected data is output.
[0045] Example 2:
[0046] Based on Example 1, to verify the effectiveness of the present invention, long-term testing was conducted on a MEMS triaxial accelerometer (typical noise density of approximately 150 μg / √Hz). The experiment was divided into a source domain (first month of deployment, temperature range 20–28°C) and a target domain (6 months after quarterly operation, temperature range -10–40°C, cumulative operation of approximately 4500 hours). This example uses the same MEMS triaxial accelerometer hardware platform (noise density 150 μg / √Hz) and core algorithm architecture (GPR combined kernel + MMD transfer learning + slope / intercept decision matrix) as Example 1, but verifies the sensor under accelerated aging scenarios to simulate longer-term drift effects.
[0047] S1. Collect sensor data, including triaxial acceleration data, temperature, humidity, and air pressure environmental parameters;
[0048] S2. Calculate the magnitude of the triaxial composite vector. ,in Let be the magnitude of the triaxial acceleration vector. Sensor triaxial raw output values, Zero-point drift for each axis;
[0049] S3. Using a Gaussian process regression model as a foundation, perform source domain pre-training, feature extraction, and regression. The feature extraction includes time-slicing and constructing feature vectors. ,in This is the ambient temperature value at the current sampling time. This is the ambient relative humidity value at the current sampling time. This is the ambient atmospheric pressure value at the current sampling time. It is the rate of temperature change. It is the time representation of the sampling time in a day. The kernel function of the Gaussian process regression is a combined kernel, including a periodic kernel, a linear kernel, and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends, and short-term random noise, respectively.
[0050] S4. Employ a transfer learning strategy in the target domain, using the maximum mean difference loss function. To achieve distribution alignment and adaptive migration, This represents the mean squared error loss in Gaussian process regression. For the trade-off hyperparameter of MMD loss, To determine the maximum mean difference between the feature distributions of the source domain and the target domain, the transfer learning strategy includes cold start and hot update. Adaptive Batch Normalization is used to update the mean and variance of the model using the current small sample data to achieve rapid adaptation.
[0051] S5. Distinguish between zero-point drift and real motion using the slope intercept joint judgment matrix. If it is determined to be real motion, trigger environmental baseline reset; otherwise, perform drift compensation and output correction data. ,in These are the sensor's raw triaxial readings. It is the drift compensation amount predicted by the transfer learning model, and the joint decision matrix of slope intercept includes the regression slope stability. and intercept jump intensity ,in This represents the slope of the linear regression of modulus and temperature for the current time period. This is the difference between the current slope and the initial slope k_0 of the source domain. The standard deviation of the historical slope of the source domain. This represents the regression intercept of modulus length versus temperature for the current time period. This is the difference between the current intercept and the initial intercept b_0 of the source domain. The average value of the historical intercept of the source domain, if and ,but ,in yes The threshold, yes The threshold value, the range of which is adjusted according to the sensor noise level and the application scenario, The typical threshold range is 0.10 to 0.20. The typical threshold range is 2.5–4.0; when When the value is less than 0.15 and D > 3.0, it is considered a real motion.
[0052] Key parameter adjustment: To address the rapid temperature changes that accelerate aging, the daily anchor point period has been adjusted from "2:00-4:00 AM" to the temperature stabilization plateau period (when | (Data acquisition is triggered when the temperature is less than 0.1℃ / min and lasts for more than 30 minutes) to ensure that the regression parameter calculation is not affected by temperature transients. Due to the more severe distribution shift caused by accelerated aging, λ is adjusted from 0.5 to 0.8 to enhance the distribution alignment strength. In response to the increased noise after accelerated aging, the S threshold is relaxed from 0.15 to 0.20 and the D threshold is tightened from 3.0 to 3.5 to improve the conservatism of the real motion judgment.
[0053] Example 3: This example extends the single-sensor solution of Example 1 to a 3-node accelerometer array, verifying the effectiveness of transfer learning in multi-sensor collaborative drift correction. It is applicable to structural health monitoring scenarios requiring spatial consistency monitoring. Three identical MEMS triaxial accelerometers (noise density 150 μg / √Hz) are rigidly mounted on the same base (10cm spacing, temperature gradient <0.5℃). During the first month of deployment, the temperature range is 20-28℃, and the three nodes collect data synchronously to establish a joint source domain model. After six months of operation across quarters, the temperature range is -10 to 40℃, with a cumulative operating time of approximately 4500 hours.
[0054] S1. Collect sensor data, including triaxial acceleration data, temperature, humidity, and air pressure environmental parameters;
[0055] S2. Calculate the magnitude of the triaxial composite vector. ,in Let be the magnitude of the triaxial acceleration vector. Sensor triaxial raw output values, Zero-point drift for each axis;
[0056] S3. Using a Gaussian process regression model as a foundation, perform source domain pre-training, feature extraction, and regression. The feature extraction includes time-slicing and constructing feature vectors. ,in This is the ambient temperature value at the current sampling time. This is the ambient relative humidity value at the current sampling time. This is the ambient atmospheric pressure value at the current sampling time. It is the rate of temperature change. It is the time representation of the sampling time in a day. The kernel function of the Gaussian process regression is a combined kernel, including a periodic kernel, a linear kernel, and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends, and short-term random noise, respectively.
[0057] S4. Employ a transfer learning strategy in the target domain, using the maximum mean difference loss function. To achieve distribution alignment and adaptive migration, This represents the mean squared error loss in Gaussian process regression. For the trade-off hyperparameter of MMD loss, To determine the maximum mean difference between the feature distributions of the source domain and the target domain, the transfer learning strategy includes cold start and hot update. Adaptive Batch Normalization is used to update the mean and variance of the model using the current small sample data to achieve rapid adaptation.
[0058] S5. Distinguish between zero-point drift and real motion using the slope intercept joint judgment matrix. If it is determined to be real motion, trigger environmental baseline reset; otherwise, perform drift compensation and output correction data. ,in These are the sensor's raw triaxial readings. It is the drift compensation amount predicted by the transfer learning model, and the joint decision matrix of slope intercept includes the regression slope stability. and intercept jump intensity ,in This represents the slope of the linear regression of modulus and temperature for the current time period. This is the difference between the current slope and the initial slope k_0 of the source domain. The standard deviation of the historical slope of the source domain. This represents the regression intercept of modulus length versus temperature for the current time period. This is the difference between the current intercept and the initial intercept b_0 of the source domain. The average value of the historical intercept of the source domain, if and ,but ,in yes The threshold, yes The threshold value, the range of which is adjusted according to the sensor noise level and the application scenario, The typical threshold range is 0.10 to 0.20. The typical threshold range is 2.5–4.0; when When the value is less than 0.15 and D > 3.0, it is considered a real motion. After a single node is judged, a consistency check between nodes is added: if a node is judged as a real motion but the other two nodes are judged as drifting, then the node triggers "isolated event review", and is judged again after a delay of 2 hours, to avoid misjudgment caused by single node abnormality.
[0059] The following are the test results of the above embodiments:
[0060]
[0061] Conclusion: By employing Gaussian process regression combined with the maximum mean difference (MMD) loss function, adaptive distribution alignment from the source domain to the target domain is achieved, enabling the sensor to maintain high-precision measurements even after cross-seasonal or long-term operation. Experiments show that after 4500 hours of cross-seasonal operation, the modulus deviation is improved by approximately 81%, and the temperature-related error coefficient is improved by approximately 80%, significantly outperforming the traditional linear temperature compensation method (accuracy improvement of 65%–75%). The introduction of a slope / intercept joint decision matrix, using a binary decision based on the regression slope stability S and intercept jump intensity D, effectively distinguishes zero-point drift from actual motion / tilt events, achieving a distinction accuracy of over 96% and a misclassification rate controlled within 4%. This avoids the miscompensation problem of traditional methods and significantly improves the system's reliability and robustness.
[0062] In summary, this invention achieves high-precision, adaptive, and intelligent correction of sensor zero-point drift by combining transfer learning strategies with physical constraints, and has significant engineering application value and promising prospects for widespread application.
[0063] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A sensor drift correction method based on transfer learning, characterized in that: Includes the following steps: S1. Collect sensor data, including triaxial acceleration data, temperature, humidity, and air pressure environmental parameters; S2. Calculate the magnitude of the triaxial composite vector. ,in Let be the magnitude of the triaxial acceleration vector. Sensor triaxial raw output values, Zero-point drift for each axis; S3. Using the Gaussian process regression model as a base, source domain pre-training, feature extraction and regression are performed. The kernel function of the Gaussian process regression is a combined kernel, including a periodic kernel, a linear kernel and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends and short-term random noise, respectively. S4. Employ a transfer learning strategy in the target domain, using the maximum mean difference loss function. To achieve distribution alignment and adaptive migration, This represents the mean squared error loss in Gaussian process regression. For the trade-off hyperparameter of MMD loss, This represents the maximum mean difference between the feature distributions of the source and target domains. S5. Distinguish between zero-point drift and real motion using the slope intercept joint judgment matrix. If it is determined to be real motion, trigger environmental baseline reset; otherwise, perform drift compensation and output correction data. ,in These are the sensor's raw triaxial readings. It is the drift compensation amount predicted by the transfer learning model. This is the ambient temperature value at the current sampling time. This is the ambient relative humidity value at the current sampling time. The current sampling time is the ambient atmospheric pressure value, where t is the current sampling time. The slope intercept joint determination matrix includes regression slope stability. and intercept jump intensity ,in The slope of the linear regression of modulus and temperature for the current time period. This is the difference between the current slope and the initial slope k_0 of the source domain. The standard deviation of the historical slope of the source domain. This represents the regression intercept of modulus length versus temperature for the current time period. This is the difference between the current intercept and the initial intercept b_0 of the source domain. The average value of the historical intercept of the source domain, if and ,but ,in yes The threshold, yes The threshold.
2. The sensor drift correction method based on transfer learning according to claim 1, characterized in that: The feature extraction includes time segmentation and constructing feature vectors. ,in This is the ambient temperature value at the current sampling time. This is the ambient relative humidity value at the current sampling time. This is the ambient atmospheric pressure value at the current sampling time. It is the rate of temperature change. It represents the time of day in which the sampling point is located.
3. The sensor drift correction method based on transfer learning according to claim 1, characterized in that: The transfer learning strategy includes cold start and hot update, using Adaptive Batch Normalization to update the mean and variance of the model with the current small sample data to achieve rapid adaptation.
4. The sensor drift correction method based on transfer learning according to claim 1, characterized in that: The threshold range is adjusted according to the sensor noise level and application scenario. The typical threshold range is 0.10 to 0.
20. The typical threshold range is 2.5–4.0; when When the value is less than 0.15 and D > 3.0, it is considered a real motion.
5. A sensor drift correction system based on transfer learning, based on any one of the sensor drift correction methods based on transfer learning according to claims 1-4, characterized in that: include: The data acquisition module is used to collect raw data from the sensors and environmental parameters; The feature extraction module is used to construct feature vectors and perform time-segment alignment. The regression model module constructs a zero-point mapping based on Gaussian process regression. The transfer learning module uses the maximum mean difference loss function to align the distributions of the source and target domains. The determination module distinguishes between drift and real motion using a slope intercept joint determination matrix and outputs correction data.
6. The sensor drift correction system based on transfer learning of claim 5, wherein: The feature extraction module utilizes the stable regression coefficients of data from the same time period each day to perform spatiotemporal data cleaning, including time period slicing and feature vector construction. The feature vector contains temperature, humidity, air pressure, temperature change rate, and time period information.
7. The sensor drift correction system based on transfer learning of claim 5, wherein: The kernel function of the regression model module is designed as a combined kernel, including a periodic kernel, a linear kernel, and an RBF kernel, which correspond to periodic environmental fluctuations, long-term aging trends, and short-term random noise, respectively. In the cold start and hot update strategy, the transfer learning module uses AdaBN to update the mean and variance of the model with the current small sample data, achieving rapid adaptation.
8. The sensor drift correction system based on transfer learning of claim 5, wherein: It also includes an online monitoring module, which calculates the regression parameters for the same time period every 24 hours and calculates the KL divergence between the current distribution and the source domain distribution. If the divergence exceeds the threshold, migration compensation is initiated and corrected data is output.