Non-stationary platform sensor data error correction method and system based on pose perception and environmental background constraint

Abstract

The application discloses a non-stationary platform sensor data error correction method and system based on attitude sensing and environmental background constraint, and comprises the following steps: acquiring target physical quantity observation data, platform attitude statistical data, environmental state data and independent external reference data within a preset time window; constructing a multi-dimensional state vector which fuses attitude features, environmental features, observation features and environmental background constraint features, retrieving similar state samples from a historical sample library based on similarity measurement to form a similar sample set; calculating error correction parameters by using the difference between the observation data and the external reference data in the similar sample set; and correcting and outputting the current observation data. The method only relies on low-frequency statistical data to realize quantitative estimation and correction output of non-stationary platform sensor measurement deviation, improves the self-adaptive ability to platform state change and background field condition difference, and solves the problem that state-dependent time-varying measurement deviation cannot be objectively quantified and dynamically corrected when high-frequency original motion sequences are lacking.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints. Background Technology

[0002] In the fields of engineering observation and scientific measurement, a large number of sensors are deployed on non-stationary platforms to perform long-term continuous observation tasks. These non-stationary platforms refer to observation carriers that are in continuous motion on the water surface, underwater, in the air, or on the ground, with their attitude dynamically changing over time. Typical examples include oceanographic data buoys, shipborne observation platforms, drifting buoys, tethered balloons, and UAV-borne sensor platforms. Sensors mounted on such platforms experience continuous changes in spatial orientation and position as the platform moves, leading to time-varying deviations between the sensor's measurement coordinate system and the geographic reference coordinate system, resulting in systematic errors in the observation results. Taking vector physical quantities as an example, their observed values ​​may deviate from the true values ​​in both amplitude and direction, and the amount of deviation varies non-stationarily with the platform's instantaneous attitude state, environmental conditions, and the distribution characteristics of the background physical quantity field. How to quantify, estimate, and correct these time-varying, state-dependent measurement deviations from a data processing perspective is a key technical issue for improving the usability of observation data from non-stationary platforms.

[0003] Several methods have been developed in the existing technology for error processing of observation data from non-stationary platforms, but all of them have limitations to varying degrees.

[0004] The first type of method is motion compensation based on high-frequency inertial measurement data. This type of method synchronously acquires the platform's high-frequency attitude angle, angular velocity, and acceleration raw sequences, and uses coordinate rotation matrices and velocity vector compensation algorithms to reconstruct the geographic coordinate system of the sensor measurements. However, this type of method heavily relies on the integrity and continuity of the high-frequency raw motion sequences. In practical operational observation systems, due to limitations in data storage capacity, communication bandwidth, and power consumption, the observation data that is stored and transmitted over a long period is usually only low-frequency statistical products calculated according to a preset time window, such as the average value, extreme values, and corresponding statistical characteristics of 10-minute or 30-minute intervals. The high-frequency raw motion sequences are often not retained or transmitted back. Therefore, motion compensation-based methods are difficult to retrospectively apply to historical data and cannot be deployed and operated in regular operational flows that only have low-frequency statistical data.

[0005] The second category of methods is data quality control based on physical thresholds or empirical rules. These methods apply preset rules such as wind speed range checks, abrupt change checks, and continuity checks to the observed data, or construct dynamic quality thresholds based on the platform motion impact index and historical statistical quantiles, to classify the observed data into quality levels and identify and remove high-risk data affected by motion interference. However, the output of these methods is data quality labels or cleaned subsets of data. Their core function is to determine "which data is unusable," but they cannot quantitatively estimate systematic biases in the data, nor can they output corrected, usable data. For a large amount of intermediate data that is "suspicious but does not meet the removal criteria," these methods cannot provide further processing for bias quantification.

[0006] The third category of methods is post-hoc comparative verification based on external reference data. This type of method compares the observation data of a non-stationary platform with reference data from an independent source, such as reanalysis data fields, regional numerical model outputs, observation data from nearby fixed stations, or other quality-controlled reliable data sources. Existing comparative verification methods typically use the comparison results for overall statistical bias analysis or post-hoc data quality checks, lacking a bias modeling mechanism that couples external reference constraints with the platform's own attitude state and environmental state. Because a "state-bias" mapping relationship is not established, this type of method struggles to generate adaptive correction parameters for different attitude conditions, environmental backgrounds, and physical field conditions, and also cannot output correction data for a single time window.

[0007] The fourth category of methods consists of empirical correction methods based on single platform attitude indicators. These methods use the platform tilt angle or its statistics as independent variables and the attenuation ratio or deviation of the observed quantity as dependent variables, correcting through pre-defined fixed empirical relationships. However, these methods have a single dimension of state representation and do not fully incorporate the characteristic information of the environmental state and background physical field, resulting in insufficient adaptability of the same fixed empirical relationship under different sea states and background field conditions. Furthermore, for directional cyclic physical quantities, existing empirical correction methods typically do not specifically address the spurious deviation problem when the angle crosses the period boundary, and generally lack quantitative assessment of the reliability of the correction results, posing a risk of blind correction during operational automation.

[0008] In summary, existing technologies still lack a technical approach that can achieve data deviation quantification and correction output without relying on the constraints of high-frequency original motion sequences: there is a lack of effective multi-source data fusion mechanisms between external reference data and platform attitude and environmental states; the use of conventional arithmetic differences for cyclic angle measurements in similarity measurement leads to misjudgments in distance calculation when crossing cycle boundaries; there is a lack of effective suppression methods for abnormal historical samples in deviation estimation; and there is a lack of quantitative evaluation and hierarchical decision-making mechanisms for the reliability of correction results while outputting correction data. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for correcting sensor data errors on non-stationary platforms based on attitude perception and environmental background constraints. By constructing a multi-dimensional state vector that integrates attitude features, environmental features, observation features, and external reference constraint features, and by adaptively selecting similar state samples from a historical sample library based on similarity retrieval to dynamically calculate error correction parameters, this method achieves quantitative estimation and correction output of measurement deviations of non-stationary platform sensors under the condition of relying only on low-frequency statistical data. This improves the adaptive ability of the data processing method to platform state changes and background field condition differences, and solves the problem that existing data processing methods cannot objectively quantify and dynamically correct state-dependent time-varying measurement deviations when high-frequency original motion sequences are lacking.

[0010] To address the aforementioned technical problems, a first aspect of this invention provides a method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints, comprising the following steps: S1: Acquire the target physical quantity observation data collected by the non-stationary platform sensor within a preset time window, the attitude statistics of the sensor-mounted platform, the environmental state data of the environment where the sensor is located, and external reference data independent of the non-stationary platform sensor; the external reference data is a reliable reference measurement value of the same type as the target physical quantity. S2: Construct the state vector for the current time window, and calculate the similarity between the current state vector and each historical state vector from a pre-constructed historical sample library based on a preset similarity metric. Select historical samples that meet the preset similarity conditions to form a similar sample set. The state vector includes at least the attitude features extracted from the attitude statistics data, the environmental features extracted from the environmental state data, the observation features extracted from the target physical quantity observation data, and the environmental background constraint features extracted from the external reference data. Each historical sample in the historical sample library contains data of the same type as the current time window and a corresponding state vector. S3: Based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data, the error correction parameter for the current time window is calculated; S4: Use the error correction parameters to correct the target physical quantity observation data in the current time window to obtain the corrected target physical quantity data.

[0011] Further, the calculation of the similarity between the current state vector and each historical state vector in step S2 includes: S211: Identify the value type of each component in the state vector, and mark the components whose values ​​have cyclic boundary characteristics as cyclic angle components. S212: For each cyclic angle component, calculate the angle difference between the component in the current state vector and the corresponding component in the historical state vector, obtain the first intermediate value by taking the sine of the angle difference, and obtain the second intermediate value by taking the cosine of the angle difference. S213: Calculate the arctangent value of the first intermediate quantity and the second intermediate quantity, and take the absolute value as the component distance of the cyclic angle component; S214: Combine the component distances of the cyclic angle components and the component distances of the non-cyclic angle components after a preset standardization process to obtain the normalized distances, and obtain the comprehensive state distance between the current state vector and the historical state vector.

[0012] Further, in step S214, the component distance of the cyclic angle component and the normalized distance of the non-cyclic angle component are combined to obtain the comprehensive state distance between the current state vector and the historical state vector, including: S2141: Obtain the cyclic angular component distance and non-cyclic angular component distance for each historical sample, and calculate the first statistical dispersion of the cyclic angular component distance and the second statistical dispersion of the non-cyclic angular component distance respectively. S2142: Determine a first merging weight for the cyclic angular component distance and a second merging weight for the non-cyclic angular component distance based on the ratio of the first statistical dispersion to the second statistical dispersion; the first merging weight decreases as the ratio increases, and the second merging weight increases as the ratio increases; S2143: Using the first merging weight and the second merging weight, the component distance of the cyclic angle component and the normalized distance of the non-cyclic angle component are weighted and merged to obtain the comprehensive state distance.

[0013] Further, step S2, which involves selecting historical samples that meet preset similarity conditions to form a similar sample set, includes: S221: Set a similarity distance threshold, and select only historical samples whose overall state distance is less than the similarity distance threshold; S222: When the number of selected historical samples is less than the preset minimum number, expand the time range of the historical sample library or reduce the similarity distance threshold, and reselect historical samples; S223: When the number of historical samples after reselection is still less than the preset minimum number, the current time window is marked as unsuitable for correction, and the original target physical quantity observation data and the corresponding low confidence flag are output, and subsequent steps S3 and S4 are no longer executed.

[0014] Further, in step S3, the calculation of the error correction parameters for the current time window based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data includes: S311: Decompose the target physical quantity observation data of each historical sample in the similar sample set into a first observation component and a second observation component that are mutually orthogonal, and decompose the external reference data of the historical sample into a first reference component and a second reference component that are mutually orthogonal. S312: Calculate the first component difference between the first observed component and the first reference component, and the second component difference between the second observed component and the second reference component, respectively; S313: Calculate the weight corresponding to each historical sample, wherein the weight decreases as the comprehensive state distance between the historical sample and the current state vector increases; S314: The first component difference and the second component difference of each historical sample are weighted and combined according to the weights to obtain the first component correction parameter and the second component correction parameter of the current time window, which together constitute the error correction parameter.

[0015] Further, the calculation of the weights corresponding to each historical sample in step S313 includes: S3131: Calculate the initial weight of each historical sample based on the comprehensive state distance between each historical sample and the current state vector. The initial weight decreases as the comprehensive state distance increases. S3132: Calculate the third statistical dispersion of the first component difference of all historical samples in the similar sample set, and the fourth statistical dispersion of the second component difference; S3133: For each historical sample, the stability correction coefficient of the historical sample is calculated based on the degree of deviation between the first component difference and the third statistical dispersion, and the degree of deviation between the second component difference and the fourth statistical dispersion; the stability correction coefficient decreases as the degree of deviation increases. S3134: The initial weights are corrected using the stability correction coefficient to obtain the weights corresponding to the historical sample.

[0016] Further, in step S3, the calculation of the error correction parameters for the current time window based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data includes: S321: Select at least some components from the state vectors of each historical sample in the similar sample set as independent variables, and use the difference between the target physical quantity observation data of each historical sample and its external reference data as the dependent variable to construct a regression model; S322: Fit the regression model using the similar sample set to obtain the regression coefficients; S323: Substitute the corresponding components in the state vector of the current time window into the fitted regression model to obtain the error correction parameters for the current time window.

[0017] Further, step S4, which involves using the error correction parameter to correct the target physical quantity observation data within the current time window to obtain corrected target physical quantity data, includes: S41: Calculate the sample quantity factor based on the number of samples in the similar sample set. The value of the sample quantity factor increases monotonically with the increase of the number of samples and approaches the upper limit. S42: Calculate the state distance factor based on the comprehensive state distance between each historical sample in the similar sample set and the current state vector. The value of the state distance factor increases monotonically as the comprehensive state distance decreases. S43: Calculate the deviation stability factor based on the degree of dispersion between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data; S44: The sample quantity factor, state distance factor, and bias stability factor are fused to obtain the correction confidence level; S45: When the correction confidence level is higher than the first threshold, output the corrected target physical quantity data with a high confidence indicator; when the correction confidence level is lower than or equal to the first threshold, output the original target physical quantity observation data with a low confidence indicator.

[0018] Furthermore, the process described in S44, which fuses the sample quantity factor, state distance factor, and bias stability factor to obtain the correction confidence level, includes: S441: Decompose the deviation stability factor into a first observation deviation stability factor corresponding to the first observation component and a second observation deviation stability factor corresponding to the second observation component, and calculate the reference consistency factor; the reference consistency factor is determined based on at least one of the following: the source type of the external reference data, the spatiotemporal resolution, and the uncertainty index; S442: The sample quantity factor, state distance factor, first observation bias stability factor, second observation bias stability factor and reference consistency factor are fused to obtain the correction confidence.

[0019] Further, step S45, which involves outputting the corrected target physical quantity data with a high-confidence flag when the correction confidence level is higher than the first threshold, and outputting the original target physical quantity observation data with a low-confidence flag when the correction confidence level is lower than or equal to the first threshold, includes: S451: Determine whether the correction confidence level is higher than the first threshold; S452: When the correction confidence level is higher than the first threshold, output the corrected target physical quantity data and attach a high confidence mark; S453: When the correction confidence is lower than or equal to the first threshold, it is determined whether the correction confidence is higher than the second threshold, wherein the second threshold is lower than the first threshold; S454: When the correction confidence level is higher than the second threshold, output the corrected target physical quantity data and attach a reference correction identifier; S455: When the correction confidence is lower than or equal to the second threshold, output the original target physical quantity observation data with a low confidence flag.

[0020] Furthermore, the attitude features extracted from the attitude statistics in step S2 include a comprehensive tilt angle, which is calculated based on the square root of the sum of the squares of the pitch angle statistics and the roll angle statistics. The state vector includes the integrated tilt angle, azimuth angle, significant wave height, wave spectrum peak period, main wave direction, observed values ​​of target physical quantities, and external reference data values.

[0021] Furthermore, the pre-constructed historical sample library mentioned in step S2 has at least one of the following sources: historical observation data of the current sensor platform, historical observation data of other platforms of the same model as the current sensor platform under similar environmental conditions, and stable observation period data confirmed by quality control.

[0022] Furthermore, the external reference data mentioned in step S1 includes one or more of the following: reanalysis wind field data, regional numerical model wind field data, observation data from nearby fixed stations, observation data from offshore wind towers, and reference data after quality control.

[0023] Accordingly, a second aspect of the present invention provides a system for correcting errors in sensor data of a non-stationary platform based on attitude perception and environmental background constraints. This system corrects errors in sensor data of a non-stationary platform based on the aforementioned method for correcting errors in sensor data of a non-stationary platform based on attitude perception and environmental background constraints, including: The data acquisition module acquires target physical quantity observation data collected by the non-stationary platform sensor within a preset time window, attitude statistics of the sensor mounting platform, environmental state data of the environment where the sensor is located, and external reference data independent of the non-stationary platform sensor; the external reference data are reliable reference measurements of the same type as the target physical quantity. The vector calculation module constructs the state vector for the current time window and calculates the similarity between the current state vector and each historical state vector from a pre-built historical sample library based on a preset similarity metric. Historical samples that meet preset similarity conditions are selected to form a similar sample set. The state vector includes at least attitude features extracted from the attitude statistics data, environmental features extracted from the environmental state data, observation features extracted from the target physical quantity observation data, and environmental background constraint features extracted from the external reference data. Each historical sample in the historical sample library contains data of the same type as the current time window and a corresponding state vector. The error calculation module calculates the error correction parameters for the current time window based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data. The data correction module uses the error correction parameters to correct the target physical quantity observation data in the current time window, and obtains the corrected target physical quantity data.

[0024] Accordingly, a third aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-described non-stationary platform sensor data error correction method based on attitude perception and environmental background constraints.

[0025] Accordingly, a fourth aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints.

[0026] The above-described technical solutions of the embodiments of the present invention have the following beneficial technical effects: 1. By constructing a multi-dimensional state vector that integrates attitude features, environmental features, observation features, and external reference constraints, and using external reference data independent of the sensor to be calibrated as a component of the state vector, the selection of similar historical samples not only depends on the sensor's own observation data and platform state data, but is also constrained by an independent and reliable reference field. This achieves collaborative representation and joint retrieval of multi-source data from platform attitude, environmental background, and external reference field. Compared with existing technologies that only use sensor data for state identification or only use external references for post-event comparison, this method establishes a dynamic mapping relationship between "state and deviation" at the data processing level. This allows each time window to be calibrated to be matched with historical state samples that are highly similar to its attitude conditions, environmental conditions, and background physical quantity field distribution. This improves the adaptability of deviation estimation to complex working conditions and solves the problem of inaccurate deviation estimation in existing methods when platform state and background field conditions are coupled and change. 2. In the process of similarity measurement and deviation estimation, a cyclic difference measurement method based on the arctangent function is adopted for cyclic angle type physical quantities. The observed data and reference data are decomposed into mutually orthogonal vector components, and the component difference is calculated and weighted. This fundamentally avoids the technical defects of angular physical quantities such as wind direction and azimuth when crossing the 0° / 360° cycle boundary, which produce pseudo distance or pseudo deviation due to conventional arithmetic subtraction. In the error correction parameter generation stage, a stability correction mechanism based on the degree of deviation dispersion is introduced. Anomalous historical samples that deviate from the overall deviation distribution in the similar sample set are softly suppressed, rather than simply removed, so that the deviation estimation results remain numerically stable under the condition of uneven sample quality. Compared with the shortcomings of existing empirical correction methods that ignore the angular cycle characteristics and lack effective control over anomalous samples, a special processing mechanism for cyclic physical quantities and anomalous samples is built in the complete data processing chain from distance calculation and deviation estimation to parameter generation, which improves the numerical stability and physical consistency of the correction output. 3. A multi-factor fusion-based calibration reliability quantification evaluation system and a hierarchical output decision mechanism were established in the calibration output stage. By comprehensively considering factors such as the number of similar samples, the degree of state similarity, deviation stability, and the reliability of external reference data itself, a normalized calibration reliability index is generated. Based on the relationship between this index and preset multi-level thresholds, a hierarchical output strategy is executed. When the reliability is high, the calibration data is output with a high reliability label; when the reliability is medium, the calibration data is output with a reference label; and when the reliability is low, only the original observation data is output without performing forced calibration. Compared with the common problems of lacking calibration reliability assessment and having the risk of blind calibration in existing technologies, this system deeply binds calibration actions with reliability assessment, builds a safety decision barrier at the end of the data processing flow, and enables the automated calibration system to have self-assessment and risk avoidance capabilities, thereby improving the security, controllability, and traceability of business data calibration processing. Attached Figure Description

[0027] Figure 1 This is a flowchart of a non-stationary platform sensor data error correction method based on attitude perception and environmental background constraints provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of attitude-sea state similarity sample screening provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of wind vector deviation estimation and correction provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the correction confidence calculation logic provided in an embodiment of the present invention; Figure 5 This is a block diagram of a non-stationary platform sensor data error correction system based on attitude perception and environmental background constraints provided in an embodiment of the present invention.

[0028] Figure label: 1. Data acquisition module; 2. Vector calculation module; 3. Error calculation module; 4. Data correction module. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0030] This invention proposes a data error correction processing architecture for non-stationary observation platforms. This architecture integrates multi-source data acquisition, multi-dimensional state representation, similarity retrieval, deviation quantification, and dynamic correction into a closed-loop data processing chain at the system level. Using a long-term continuous observation scenario of an oceanographic buoy as the main example, the buoy continuously generates six degrees of freedom motion under the combined effects of wind, waves, and currents on the sea surface. The wind speed and direction sensors installed at the top of the buoy's mast constantly change their spatial orientation as the platform sways, resulting in dynamic deviations between the measurement coordinate system and the geographic coordinate system. The system architecture of this invention constructs a four-stage cascaded processing mechanism at the data processing level: "state awareness—similarity retrieval—deviation transfer—reliable correction." Its core improvements and innovations are reflected in two aspects. First, at the state representation level, this method breaks through the single-dimensional model of existing methods that rely solely on platform attitude or external references. It proposes to uniformly encode four types of heterogeneous data sources—platform attitude characteristics, marine environment characteristics, sensor observation characteristics, and independent external reference wind field characteristics—into a multi-dimensional state vector. This ensures that each observation time window obtains a comprehensive state label integrating platform motion state, environmental background conditions, and external reference constraints, providing a complete query foundation for subsequent similarity retrieval. Second, at the processing mechanism level, a deviation transfer model is established that "infers the current unknown deviation from the known deviations in historical similar states." By retrieving several historical windows from the historical sample database that are closest to the current state vector, and utilizing the known differences between sensor observations and independent reference values ​​in these historical windows, the error correction parameters for the current window are weighted and extrapolated. This achieves adaptive dynamic estimation of measurement deviations under different attitudes, sea states, and wind field backgrounds without needing to reconstruct the sensor physical measurement chain or acquire high-frequency original motion sequences. This architecture transforms the role of external reference data from a traditional post-hoc comparative benchmark to an active component participating in state representation and deviation constraints. This ensures that the selection of similar samples not only matches the platform's own state but also the distribution conditions of the wind field background, effectively solving the state blind spot problem encountered when relying solely on platform data or solely on external references. This embodiment scenario is hereinafter referred to as the marine buoy wind measurement scenario.

[0031] Please refer to Figure 1 The first aspect of this invention provides a method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints, comprising the following steps: S1: Acquire the target physical quantity observation data collected by the non-stationary platform sensor within a preset time window, the attitude statistics of the sensor-mounted platform, the environmental state data of the sensor's environment, and external reference data independent of the non-stationary platform sensor. The external reference data consists of reliable reference measurements of the same type as the target physical quantity.

[0032] In practical non-stationary observation scenarios, such as long-term continuous observation by ocean data buoys, the various sensors on board continuously generate observation records according to preset time windows. The preset time window referred to in this invention is typically a ten-minute or thirty-minute statistical period commonly used in operational observation systems. Within this period, the sensors continuously sample the target physical quantity and statistically obtain the average or extreme values ​​within the window. The target physical quantity observation data refers to the physical quantity values ​​directly measured and output by sensors installed on non-stationary platforms. In typical marine meteorological observation scenarios, this data can be sea surface wind speed and direction, or other vector physical quantities with amplitude and directional characteristics.

[0033] Simultaneously, the attitude measurement unit on the sensor platform also outputs attitude statistics according to time windows. These data characterize the platform's motion state within that window. The attitude statistics are statistical quantities within the window, such as pitch and roll angle statistics, rather than continuous high-frequency raw attitude sequences. This characteristic allows the invention to be compatible with most operational observation systems that only store statistical products. Environmental state data describing the sensor's environment describes the background environmental conditions during the observation period. In marine scenarios, this may include wave parameters such as significant wave height, wave peak period, and main wave direction. External reference data, independent of the non-stationary platform sensor, is a key constraint introduced in this invention. This data comes from observation sources or analysis systems completely independent of the sensor to be calibrated, such as global or regional reanalysis data fields, numerical weather prediction model outputs, and measurement data from nearby fixed observation stations. These are reliable reference measurements of the same type as the target physical quantity, constituting a reference benchmark for objectively quantifying sensor measurement bias.

[0034] In specific implementations of this invention, the duration of the preset time window can be set according to the actual configuration of the observation system and the data storage strategy. In typical operational scenarios of ocean buoy wind measurement, the preset time window is usually 10 minutes or 30 minutes, which is consistent with the recommendations of the International Meteorological Organization for the statistical cycle of routine meteorological observations. The attitude statistics within the time window are the statistics of attitude measurements within that window. For example, the pitch angle statistics can be the average or median of the pitch angle time series within that window, and the roll angle statistics are similarly calculated. This setting indicates that this invention does not require obtaining the high-frequency raw attitude sequence within the window, nor does it require knowing the continuous time history information of attitude angles, angular velocities, or angular accelerations. All deviation estimation and correction calculations can be completed using only the low-frequency statistical products commonly stored in operational observation systems. This invention can be directly applied to the reprocessing of existing historical data and can also be seamlessly integrated into existing operational data streams for real-time or near-real-time automatic correction.

[0035] The acquisition of these four types of data forms the input basis for all subsequent processing steps. It is worth noting that these data are all routine data products commonly found in operational observation systems, requiring no additional sensors or historical high-frequency data supplementation, and are highly operable and universally applicable in engineering implementation.

[0036] Before comparing external reference data with buoy observation data, it is necessary to ensure consistency in measurement height between the two. In ocean buoy wind measurement scenarios, the wind speed sensor on the buoy is usually installed at a fixed height at the top of the buoy mast or above the buoy deck, while the wind speed output from external reference wind field data, such as reanalysis products or numerical models, is usually defined at the standard height or model plane height. If the buoy sensor's wind measurement height is inconsistent with the height of the external reference wind field, directly calculating the deviation will introduce additional errors due to differences in wind profiles. Therefore, a preset wind profile relationship can be used to convert the reference wind speed from its original height to the buoy sensor's wind measurement height. Commonly used wind profile relationships include logarithmic wind profile models and exponential wind profile models. The logarithmic wind profile model is suitable for sea surface wind fields under near-neutral atmospheric stratification conditions, while the exponential wind profile model is simpler and has certain applicability under different stratification conditions. After height conversion, a reference wind speed value matching the buoy sensor's wind measurement height can be obtained, thereby eliminating the impact of height differences on the accuracy of deviation estimation. When the external reference wind field is consistent with or differs from the wind height measured by the buoy sensor within an acceptable range, this height conversion step can be omitted.

[0037] S2: Construct the state vector for the current time window, and calculate the similarity between the current state vector and each historical state vector from a pre-built historical sample library based on a preset similarity metric. Select historical samples that meet the preset similarity conditions to form a similar sample set. The state vector includes at least attitude features extracted from attitude statistics, environmental features extracted from environmental state data, observation features extracted from target physical quantity observation data, and environmental background constraint features extracted from external reference data. Each historical sample in the historical sample library contains data of the same type as the current time window and its corresponding state vector.

[0038] like Figure 2 As shown, Figure 2 The principle and mechanism of constructing a two-dimensional state space based on attitude and sea state environmental features and performing similar sample screening are demonstrated. Figure 2 In the coordinate system, the horizontal axis represents the "comprehensive tilt angle" and the vertical axis represents the "effective wave height". These two axes represent the key features of the platform's attitude state and the environmental background state, respectively, and constitute the basis vectors of the two-dimensional state space. Figure 2The numerous hollow circles represent historical observation samples stored in the historical sample library, distributed in different locations in space, depicting the combined state of platform attitude and sea state conditions at different historical moments. The solid pentagrams represent the sample state of the current time window to be calibrated, while the related dashed circles represent the similarity search range set based on distance metrics. Figure 2 The core similarity retrieval logic is intuitively explained: by calculating the comprehensive state distance between the current sample and each historical sample in the above state space, and using a preset similarity distance threshold, the K historical samples (or those with a distance less than the threshold) that are closest to the current sample in state are selected. These selected samples constitute the exclusive similar sample set for subsequent bias estimation.

[0039] This step is the core of the adaptive bias estimation in this invention. After acquiring multi-source data for the current time window, the first step is to construct a multi-dimensional feature vector that comprehensively characterizes the state of that window. This state vector is not a simple accumulation of various data, but rather a systematic characterization of the current observation window's state from four dimensions. Attitude features are key parameters extracted from attitude statistics that reflect the platform's motion state, such as the comprehensive tilt magnitude characterizing the platform's tilt intensity. Environmental features are parameters extracted from environmental state data that reflect the background conditions of the observation environment, such as the effective wave height characterizing wave intensity. Observation features are parameters extracted from the target physical quantity's observation data itself, reflecting the magnitude and direction of the current observation value. Environmental background constraint features are parameters extracted from external reference data, reflecting the magnitude constraints of the independent reference field on the target physical quantity within the current window. These four types of features are fused into a single state vector, giving each time window a unique high-dimensional state label.

[0040] Subsequently, using the state vector of the current time window as the query condition, a similarity search is performed in a pre-built historical sample database. The historical sample database stores a large number of complete data records for historical time windows, with each historical sample containing data of the same type as the current time window and its corresponding state vector. The similarity search is based on a preset similarity metric. By calculating the state distance between the current state vector and each historical state vector, the closeness between the two in the multi-dimensional state space is quantified. The smaller the state distance, the closer the platform attitude conditions, environmental background conditions, and wind field background conditions of the historical sample are to the current window, and the higher the reference value of the deviation relationship between its observed and reference values ​​for the current window. Based on the preset similarity criteria, several historical samples with sufficiently small state distances or high rankings are selected to form a dedicated similar sample set for the current time window. The physical significance of this mechanism is that the measurement deviation of sensors on non-stationary platforms is not random noise, but rather has an inherent regularity related to the platform's motion state and environmental conditions. The observation deviations of historically similar states have direct reference value for estimating the deviation of the current state.

[0041] S3: Based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data, the error correction parameters for the current time window are calculated.

[0042] For each historical sample in the similar sample set, there exists a pair of known data: the target physical quantity observation data of the historical sample itself, and the corresponding external reference data. The difference between the two reflects the degree of deviation of the sensor measurement value from the independent and reliable reference value under the specific attitude and environmental conditions within that historical time window. This difference, in a physical sense, is the measurement bias of that historical window.

[0043] Once the set of similar samples is determined, the error correction parameter for the current time window is derived from the known deviations of these historical samples. The basic logic is that the closer a historical sample is to the current window in the state space, the greater its reference weight should be for the deviation estimation of the current window; conversely, historical samples with greater state distance, even if included in the similar sample set, have relatively limited reference value. By comprehensively calculating the deviations of each historical sample in the similar sample set, an error correction parameter suitable for the current time window is finally obtained. This correction parameter quantitatively describes the systematic deviation of the sensor observations relative to the external reference value under the current attitude conditions and environmental background, providing a direct numerical basis for subsequent correction operations.

[0044] S4: Use error correction parameters to correct the target physical quantity observation data in the current time window to obtain the corrected target physical quantity data.

[0045] like Figure 3 As shown, Figure 3 This paper demonstrates the complete operational process for bias estimation and correction of wind vectors acquired by wind speed and direction sensors on non-stationary platforms. The figure first illustrates how to project the original observed wind vector, expressed in polar coordinates as wind speed (amplitude) and wind direction (angle), and an independent external reference wind vector onto mutually orthogonal geographic coordinate systems (e.g., east-west U component and north-south V component), thereby converting the cyclic angle physical quantity with periodic boundary characteristics into two independent scalar components. The figure further illustrates how to calculate the difference between the observed wind component and the reference wind component in Cartesian coordinates, thus obtaining the first component correction parameter (east-west bias) and the second component correction parameter (north-south bias). Finally, the figure shows how to use the above two orthogonal component correction parameters to perform component-level correction on the current observed wind vector, and then resynthesize the corrected orthogonal components into the corrected wind vector output. This process fundamentally avoids the 0° / 360° cross-bias problem that may occur when directly subtracting wind direction in polar coordinates.

[0046] This step involves applying error correction parameters to the raw observation data to obtain the final usable data product. After calculating the error correction parameters for the current time window in step S3, these parameters are used to correct the target physical quantity observation data for the current time window. The basic correction method is to adjust the values ​​of the raw observation data accordingly to offset or compensate for the systematic measurement bias introduced by the sensor due to the non-stationary motion of the platform. After correction, the bias components coupled into the raw observation data by attitude motion and environmental conditions are effectively removed or suppressed, and the resulting corrected target physical quantity data is numerically closer to the true value of the physical quantity. This correction output provides a more reliable observation data foundation for subsequent applications such as meteorological analysis, marine engineering design, and numerical model verification.

[0047] This invention constructs a complete data processing chain from multi-source data acquisition, state awareness and similarity retrieval, deviation quantification calculation to dynamic correction output. This scheme does not rely on high-frequency raw motion sequences; it only utilizes low-frequency statistical data and independent external reference data commonly found in operational observation systems. It can adaptively estimate measurement deviations and complete corrections for the attitude conditions, environmental background, and reference field constraints of each time window. This represents a technological leap from "passive rejection" to "active repair" of sensor data from non-stationary platforms, significantly improving the availability and reliability of long-term observation data from non-stationary platforms.

[0048] Further, step S2 calculates the similarity between the current state vector and each historical state vector, including: S211: Identify the value type of each component in the state vector, and mark the components whose values ​​have cyclic boundary characteristics as cyclic angle components.

[0049] Before calculating the similarity between the current state vector and each historical state vector, it is necessary to first identify and classify the numerical characteristics of each component within the state vector. The state vector is a multi-dimensional vector composed of attitude features, environmental features, observation features, and environmental background constraints. Some components represent physical quantities with cyclic boundary characteristics in their numerical definitions, while others take linear values ​​in the real number domain. The numerical properties of these two types of components are fundamentally different. If a unified distance calculation method is used, the cyclic components will suffer severe distortion at the boundaries. Taking the scenario of ocean buoy wind measurement as an example, wind direction, main wave direction, and the azimuth of the buoy platform are all typical cyclic angle components, with their value range usually standardized to 0° to 360°. Under this value system, the actual spatial angle between the wind direction of 359° and 1° is only 2°, but if the numerical arithmetic difference is directly calculated, the result is 358°, which completely fails to reflect the actual direction difference in physical space. If this spurious bias is incorporated into subsequent similarity measurements, it will directly lead to inaccurate state distance calculations, causing the selection results of similar samples to deviate from true state similarity. Therefore, before calculating similarity, it is necessary to distinguish the value types of each component in the state vector, marking components with cyclic boundary characteristics as cyclic angle components and classifying components without this characteristic as non-cyclic angle components. This allows for matching distance calculation methods appropriate to the numerical properties of each type of component. This identification step avoids calculation errors caused by confusion in numerical types from the outset, providing a fundamental guarantee for the accuracy of subsequent similarity measurements.

[0050] S212: For each cyclic angle component, calculate the angle difference between the current state vector and the corresponding component in the historical state vector, obtain the first intermediate value by taking the sine of the angle difference, and obtain the second intermediate value by taking the cosine of the angle difference.

[0051] For each feature quantity labeled as a cyclic angle component, the difference between the component value in the current state vector and the corresponding component value in the historical state vector needs to be calculated. Conventional arithmetic difference calculation methods face a fundamental difficulty here: when the two angles are located on opposite sides of the cyclic boundary, there is a significant deviation between the numerical difference and the true spatial angle. The technical principle of this step lies in using trigonometric function transformations to map the angle values ​​defined on the cyclic interval to the unit circle, thereby eliminating the ambiguity caused by the cyclic boundary. Specifically, for the angle difference between the current angle and the historical angle, its sine and cosine values ​​are calculated respectively. Geometrically, the sine value reflects the difference in the orthogonal directions of the two angles on the unit circle, while the cosine value reflects the degree of consistency in the same direction of the two angles on the unit circle. Combined, they form a two-dimensional vector that accurately describes the geometric relationship between the two angles, and this vector is unaffected by the boundaries of the angle's domain. In the scenario of ocean buoy wind measurement, the angular difference between the current wind direction of 350° and the historical wind direction of 10° is -20°, with a sine value of approximately -0.342 and a cosine value of approximately 0.940. These two values ​​precisely correspond to the geometric position of the 20° angle on the unit circle. However, if the direct arithmetic difference is used, the result is 340°, with a sine value of approximately -0.342 and a cosine value of approximately 0.940. But 340° is not the true spatial angle between the two directions; rather, it is the conjugate angle of the sum of the true minimum angle of 20° and 360°. If 340° is directly used as the angular distance, the angular distance will be mistakenly taken as the dominant angle of 340°, while the true minimum angle is only 20°. Through sine and cosine transformations, this invention transforms the cyclic angle difference from a one-dimensional scalar into a pair of orthogonal components with clear geometric meaning and continuous values, fundamentally overcoming the numerical jump problem caused by the cyclic boundary and providing unambiguous intermediate information for subsequent steps to accurately reconstruct the angle difference.

[0052] S213: Calculate the arctangent value of the first intermediate quantity and the second intermediate quantity, and take the absolute value as the component distance of the cyclic angle component.

[0053] After obtaining the two orthogonal components, sine and cosine, they need to be converted back into a single distance value to be used as the component distance of the cyclic angle component in the calculation of the overall state distance. The technical principle of this step lies in using the arctangent function to combine a pair of sine and cosine components into a single angle value, which is the angle difference corresponding to the shortest arc length of the two original angles on the unit circle. The arctangent function, with the sine value as the ordinate and the cosine value as the abscissa, uniquely determines an angle within the principal value range of -180° to 180°, and the magnitude of this angle is strictly proportional to the geometric distance between the two original angles on the unit circle. Since the sine and cosine values ​​generated in the previous step have eliminated interference from the cyclic boundary, the component distance obtained after taking the absolute value in this step naturally possesses cyclic continuity. In the scenario of ocean buoy wind measurement, the wind direction values ​​of 350° and 10°, after sine and cosine transformation, have an arctangent of -20°, which, when taken as an absolute value, is 20°, perfectly consistent with the actual spatial angle between the two. For the wind direction values ​​of 180° and 180°, the angle difference is 0°, the sine value is 0, the cosine value is 1, the arctangent is 0°, and the distance is zero, perfectly matching the actual situation. Compared to conventional arithmetic differences, which often produce pseudo-distances of hundreds of degrees at the cyclic boundary, the cyclic angle component distance determined in this step is numerically accurate, stable, and has a clear physical meaning. It provides a physically correct distance measurement basis for the state vector containing cyclic angle components such as wind direction, main wave direction, and azimuth, ensuring that the cyclic angle component no longer becomes a source of error in distance calculation during similarity retrieval.

[0054] S214: Combine the component distances of the cyclic angle components and the component distances of the non-cyclic angle components after pre-standardization to obtain the normalized distance, and obtain the comprehensive state distance between the current state vector and the historical state vector.

[0055] After obtaining the component distances of all cyclic angular components and all non-cyclic angular components in the state vector, these two types of distances need to be merged into a unified composite state distance. This merging faces two levels of technical challenges. The first level is the issue of component type differences: the distances of cyclic angular components have already been specially processed in the aforementioned steps, while the distances of non-cyclic angular components are obtained through conventional arithmetic differences; both need to be merged within the same framework. The second level is the issue of dimensional and scale differences: the physical meanings of the components in the state vector are different, and their values ​​differ significantly in terms of dimensions and numerical scales. For example, significant wave height is measured in meters with typical values ​​between 0 and 10, while wind speed is measured in meters per second with typical values ​​between 0 and 30. Directly merging the original component distances would result in components with larger numerical scales occupying an unreasonable dominant position in the composite distance. To solve these two problems, this step first standardizes the component distances of non-cyclic angular components, converting them into dimensionless normalized distances. Standardization can be performed based on the mean and standard deviation of the component in the historical sample database, or based on a preset reference range. The technical effect is to ensure that components with different physical dimensions contribute equally during merging. Subsequently, the component distances of the cyclic angle components and the normalized distances of the non-cyclic angle components are weighted and merged according to preset weights. These preset weights can be configured based on the correlation strength between each component and the sensor measurement deviations. The final comprehensive state distance is a scalar value that comprehensively reflects the overall similarity between the current state vector and the historical state vector across multiple dimensions, including platform attitude, sea state environment, physical quantity observations, and external reference constraints. The smaller the comprehensive state distance, the closer the historical sample is to the state of the current time window, and the higher the reference value of the deviation relationship between the observed values ​​and the reference values ​​recorded in the historical sample for the deviation estimation of the current window. This merging mechanism solves the technical challenge of overall distance evaluation of multi-dimensional state vectors with mixed quantity types under a unified metric framework, providing a quantitative ranking basis for the reliable screening of subsequent similar sample sets and the accurate calculation of deviation estimates.

[0056] This invention establishes a comprehensive distance calculation mechanism for multidimensional state vectors with mixed quantity types in the state similarity measurement stage. This mechanism categorizes the components of the state vector into cyclic angle components and non-cyclic angle components based on their quantity type. For cyclic angle components, a cyclic difference metric based on sine-cosine decomposition and arctangent synthesis is used. For non-cyclic angle components, a normalized distance after standardization is applied. Finally, the two are weighted and combined into a comprehensive state distance. This mechanism ensures that state vectors containing cyclic physical quantities such as wind direction, main wave direction, and azimuth are no longer affected by cyclic boundary spurious biases during similarity retrieval. It also eliminates the imbalance between different physical dimensions and numerical scales, ensuring that the similarity sample selection results are consistent with the true proximity of the current time window in the multidimensional state space. This provides an accurate and reliable sample basis for subsequent bias estimation based on similar samples.

[0057] In step S2, the component distances of the cyclic angle components and the normalized distances of the non-cyclic angle components are weighted and merged using preset weights to obtain the comprehensive state distance. However, the preset weights are usually fixed values ​​set once before the system runs, based on experience or prior knowledge, and cannot be dynamically adjusted according to the inherent distribution characteristics of the actual data. When the ability of the cyclic and non-cyclic angle components to distinguish different states changes significantly in the historical sample database—for example, the spatial variation range of wind direction is much greater than the variation range of wave height in some sea states, or the opposite is true in other sea states—the fixed preset weights will not be able to adapt to this change, leading to a systematic overestimation or underestimation of the contribution of one type of component in the comprehensive state distance. To solve this problem, this invention introduces an adaptive weight adjustment mechanism based on data statistical characteristics, enabling the contribution ratio of the two types of components during merging to be dynamically adjusted according to the dispersion of the actual data in similar sample sets.

[0058] Specifically, in step S214, the component distances of the cyclic angle components and the normalized distances of the non-cyclic angle components are combined to obtain the comprehensive state distance between the current state vector and the historical state vector, including: S2141: Obtain the cyclic angular component distance and non-cyclic angular component distance for each historical sample, and calculate the first statistical dispersion of the cyclic angular component distance and the second statistical dispersion of the non-cyclic angular component distance respectively.

[0059] After obtaining the cyclic angular component distance and non-cyclic angular component distance for each historical sample, the statistical dispersion of these two types of distances is calculated separately. The first statistical dispersion is calculated for the cyclic angular component distance of all historical samples included in the consideration, and the second statistical dispersion is calculated for the normalized non-cyclic angular component distance of the same historical samples. Statistical dispersion is a basic statistic that measures the degree of concentration or dispersion of a set of numerical distributions. Its specific implementation can be variance, standard deviation, mean absolute deviation, interquartile range, or other statistical indicators that can reflect the dispersion characteristics of data. This statistic does not depend on the specific value of a single historical sample, but rather characterizes the overall variation and discriminative power of this type of component distance within the currently considered sample set. In the scenario of ocean buoy wind measurement, if the distances of several historical samples currently considered are generally small and close to each other in the cyclic angular component of wind direction, then the value of the first statistical dispersion will be low, indicating that these historical samples have high homogeneity in the wind direction dimension; conversely, if the wind direction distances of these historical samples are widely distributed and vary significantly, then the value of the first statistical dispersion will be high. Similarly, the second statistical dispersion reflects the dispersion of the distance distribution of non-cyclic angular components such as significant wave height and wind speed within the same historical sample set. By calculating the statistical dispersion of these two types of distances separately, a data-driven quantitative basis is provided for the subsequent determination of adaptive weights.

[0060] S2142: Determine the first combined weight of the cyclic angular component distance and the second combined weight of the non-cyclic angular component distance based on the ratio of the first statistical dispersion to the second statistical dispersion. The first combined weight decreases as the ratio increases, while the second combined weight increases as the ratio increases.

[0061] After obtaining the first and second statistical dispersions, the ratio between them is calculated, and the first merging weight of the cyclic angle component distance and the second merging weight of the non-cyclic angle component distance are dynamically determined based on this ratio. The core logic of this adaptive weight determination mechanism is that the larger the statistical dispersion of a certain type of component distance, the more dispersed the distribution of that type of component distance in the current similar sample set, and the larger the range of differences between each historical sample and the current state in that component dimension, indicating that the component has a relatively strong ability to distinguish sample similarity; conversely, the smaller the statistical dispersion, the smaller the difference between that type of component distance and the samples, and the relatively weaker the distinguishing ability. In order to make the comprehensive state distance more accurately reflect the real differences between samples, the statistical dispersion ratio is used as an adjustment signal. When the first statistical dispersion is larger than the second statistical dispersion, it indicates that the cyclic angle component has a higher distinguishability in the current sample set. However, in order to avoid one component from overly dominating the distance calculation, this invention adopts a strategy contrary to traditional intuition: the first merging weight decreases as the ratio increases, and the second merging weight increases as the ratio increases. The technical principle behind this reverse adjustment mechanism is that components with higher statistical dispersion have a wider range of distance value fluctuations. Giving them greater weight during merging would decrease the numerical stability of the overall state distance and make it overly sensitive to individual extreme samples. By appropriately tilting the weights towards components with lower dispersion, the robustness of the overall state distance can be enhanced, preventing accidental fluctuations in a single component from excessively disturbing the overall similarity measure. In ocean buoy wind measurement scenarios, when sea state changes drastically, wave height differences are significant, and wind direction is relatively stable, the second statistical dispersion will be greater than the first. In this case, the merging weight of non-cyclic angle components will increase accordingly, allowing the overall state distance to more sensitively capture state changes caused by differences in sea state conditions. Conversely, when wind direction changes rapidly while sea state is stable, the weight of cyclic angle components will be appropriately increased, ensuring the overall state distance maintains sufficient responsiveness to wind direction changes. This adaptive adjustment mechanism ensures that the weight configuration of the overall state distance always dynamically matches the inherent distribution characteristics of the current data, overcoming the limitations of fixed preset weights in adapting to the changing marine environment.

[0062] S2143: Using the first and second merging weights, the component distances of the cyclic angle components and the normalized distances of the non-cyclic angle components are weighted and merged to obtain the comprehensive state distance.

[0063] After determining the first and second merging weights, the component distances of the cyclic angle components and the normalized distances of the non-cyclic angle components are weighted and merged using these two weights to obtain the final comprehensive state distance. Mathematically, the merging process is a weighted summation operation, multiplying the distance of each type of component by its corresponding weight and summing the results to obtain a single scalar value. Since the weights determined in the previous step have been adaptively adjusted based on the statistical characteristics of the data itself, the resulting comprehensive state distance, compared to a scheme using fixed preset weights, can more accurately reflect the true proximity of the current state vector to each historical state vector in the multidimensional state space. In the long-term operational use of ocean buoy wind measurement scenarios, the statistical distribution characteristics of sea state parameters and wind direction parameters differ significantly under the influence of different seasons and weather systems. This adaptive weighted merging enables the similarity measure to automatically adapt to these changes, maintaining a balanced perception of the differences among various state components.

[0064] This invention implements a data-driven adaptive weight adjustment mechanism in the merging stage of the comprehensive state distance calculation. This mechanism does not rely on pre-set fixed empirical coefficients, but instead starts from the statistical characteristics of the similar sample set itself. By measuring the dispersion of the cyclic angular component distance and the relative ratio of their respective distances, it dynamically generates merging weights that match the current data distribution. This mechanism ensures that the calculation of the comprehensive state distance is neither affected by cyclic boundary pseudo-biases nor limited by the insufficient adaptability of fixed weight configurations under different operating conditions, thus ensuring that the similar sample selection results always remain consistent with the proximity relationships in the actual multidimensional state space.

[0065] After calculating the comprehensive state distance, a quantified similarity index is obtained between the current state vector and each historical state vector in the historical sample library. However, it is not yet determined which historical samples should be selected for the subsequent bias estimation stage, nor what strategy should be adopted when the number of available similar samples is insufficient. The following content progressively constructs a complete sample selection and protection mechanism, from conditional screening and dynamic fault tolerance to safety blocking.

[0066] Further, in step S2, historical samples that meet preset similarity conditions are selected to form a similar sample set, including: S221: Set a similarity distance threshold to select only historical samples whose overall state distance is less than the similarity distance threshold.

[0067] After obtaining the combined state distance between the current state vector and each historical state vector, a similarity distance threshold is set, which defines the upper limit of distance for "sufficient similarity." The combined state distance of each historical sample is compared with this threshold, and only historical samples with a combined state distance less than the threshold are selected. These selected samples constitute the initial similarity sample set. The similarity distance threshold can be determined based on the overall distribution characteristics of the combined state distance in the historical sample library, or it can be configured according to a preset empirical value. In the scenario of ocean buoy wind measurement, if the combined state distance of a historical sample is less than the threshold, it indicates that the buoy tilt attitude, wave conditions, wind speed, wind direction, level, and external reference wind field characteristics of the historical sample are highly consistent with the current time window, and the deviation relationship between its observed values ​​and reference values ​​has reliable reference significance for the current window. Historical samples with a combined state distance greater than or equal to the threshold are considered to have too large a difference from the current state, and their deviation information may not be applicable to the correction of the current window, so they are not included in the subsequent deviation estimation calculation. This threshold screening mechanism ensures that only historical samples that are truly close in state can participate in the generation of correction parameters, avoiding interference from irrelevant samples to the deviation estimation from the source.

[0068] S222: When the number of selected historical samples is less than the preset minimum number, expand the time range of the historical sample database or lower the similarity distance threshold, and reselect historical samples.

[0069] After initial screening, the number of historical samples in the screening results is counted. If the number of selected historical samples reaches or exceeds the preset minimum, it indicates that the current window has enough similar historical samples to support reliable bias estimation, and the process can directly proceed to the subsequent error correction parameter calculation step. If the number of selected historical samples is less than the preset minimum, it means that the current window's state lacks enough similar reference samples in the existing historical sample library. This may occur when the buoy encounters extreme sea conditions, experiences historically rare attitude and wave combinations, or when the time span of the historical sample library is insufficient to cover all typical operating conditions. In this case, continuing to force bias estimation based on a small number of samples will introduce significant uncertainty. Therefore, this step initiates the first-level dynamic fault-tolerant adjustment strategy, which includes two operations that can be executed independently or in combination. The first operation is to expand the time range of the historical sample library, for example, by extending the search scope from recent data to historical data with a longer time span, in order to increase the sample base that may match. The second operation is to lower the similarity distance threshold, appropriately relaxing the definition criteria for "similarity," so that more historical samples have the opportunity to enter the candidate set. The choice of operation, or a combination thereof, depends on the specific structure of the historical sample database and the characteristics of the current state vector. In ocean buoy wind measurement scenarios, if the buoy's combined tilt angle and significant wave height reach historically rare extreme values ​​during a strong typhoon, making it difficult to find a sufficient number of similar samples in the historical sample database within a normal timeframe, the search scope can be automatically expanded to all historical archived data, or the allowable upper limit of the distance threshold can be appropriately increased to obtain a sufficient number of reference samples. After adjustment, historical samples are reselected according to the new search criteria.

[0070] S223: When the number of historical samples after reselection is still less than the preset minimum number, mark the current time window as unsuitable for correction, output the original target physical quantity observation data and the corresponding low confidence flag, and no longer execute subsequent steps S3 and S4.

[0071] After completing the dynamic fault-tolerant adjustment and reselecting historical samples, the number of selected historical samples is counted again. If the number of historical samples after reselection reaches the preset minimum, it indicates that the adjustment strategy is successful, and the subsequent bias estimation and correction process continues. If the number of historical samples after reselection is still less than the preset minimum, it means that even if the search scope is expanded or the similarity criteria are relaxed, the current window's state is still in a severely sparse region within the existing historical sample database. In this case, if correction continues based on too few samples, the obtained error correction parameters will have great uncertainty. The correction result may not only fail to improve data quality but may also introduce new errors. Therefore, this step activates a safety blocking mechanism, marking the current time window as unsuitable for correction, directly outputting the original target physical quantity observation data and the corresponding low-confidence flag, and explicitly terminating the execution of subsequent steps S3 and S4, no longer entering the error correction parameter calculation and correction processing stage. This blocking decision has clear protective significance: when the data foundation on which the correction depends is insufficient, it is better to abandon the correction and retain the original observation values ​​than to risk outputting an unreliable correction result. In the operational use of ocean buoy wind measurement, this mechanism can effectively avoid erroneous corrections caused by the scarcity of similar samples under rare and extreme sea conditions, ensuring that the output of the entire automatic correction system is within a controllable safety boundary under any conditions.

[0072] Through the progressive implementation of the above three steps, a three-tiered processing mechanism of "threshold screening—dynamic fault tolerance—safety blocking" is constructed in the similarity sample screening stage. This mechanism first ensures the relevance of the selected samples by using a similarity distance threshold. Then, it dynamically adjusts the search conditions to address the boundary conditions of sample sparsity. Finally, when adjustments are ineffective, it decisively terminates the correction process and outputs the original data with a low-confidence identifier. Compared to traditional methods that use fixed K-nearest neighbors or fixed thresholds and forcibly perform correction based on irrelevant samples when there are insufficient samples, this mechanism enables the correction system to have self-assessment and risk avoidance capabilities. It sets up a safety guarantee at the intermediate node of the data processing flow, avoiding correction runaway problems caused by unmet data conditions.

[0073] After obtaining several historical samples similar to the current time window state through similarity retrieval, it is necessary to use the differences between the known observations and reference values ​​in these historical samples to deduce the error correction parameters for the current window. When the target physical quantity is a vector physical quantity with amplitude and direction characteristics, such as sea surface wind speed and direction in the scenario of ocean buoy wind measurement, directly calculating the deviation in the angular dimension faces the problem of cyclic boundaries, namely the numerical jump phenomenon at the critical point between 0° and 360° mentioned above. If the deviation statistics of wind speed and wind direction are simply performed independently, the wind direction deviation will produce spurious deviations when crossing the 0° or 360° boundary, thus leading to the distortion of the correction results. To solve this problem, this invention introduces a vector component decomposition mechanism in the deviation estimation stage, projecting the two-dimensional wind vector onto two mutually orthogonal coordinate axes in the east-west and north-south directions, converting the two coupled physical quantities of wind speed and wind direction into two independent scalar components, thereby transforming the deviation estimation from polar coordinate space to Cartesian coordinate space, eliminating the ambiguity caused by angular cycles at the root.

[0074] Specifically, in step S3, based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data, the error correction parameters for the current time window are calculated, including: S311: Decompose the target physical quantity observation data of each historical sample in the similar sample set into a first observation component and a second observation component that are mutually orthogonal, and decompose the external reference data of the historical sample into a first reference component and a second reference component that are mutually orthogonal.

[0075] For each historical sample in the similar sample set, its target physical quantity observation data is decomposed into a first observation component and a second observation component that are mutually orthogonal. In the scenario of ocean buoy wind measurement, the first observation component typically corresponds to the horizontal wind component in the east-west direction, and the second observation component corresponds to the horizontal wind component in the north-south direction. These two components represent the projected lengths of the wind vector on two orthogonal coordinate axes in the geographic coordinate system, and their magnitudes are calculated by projecting the observed wind speed and observed wind direction of the historical sample using trigonometric functions. At the same time, the external reference data corresponding to the historical sample is also decomposed into a first reference component and a second reference component in the same way. In this way, the original observation value and the independent reference value of each historical sample are expressed as a pair of orthogonal components in the same coordinate system, so that the differences between the two can be accurately and unambiguously compared in the component dimension.

[0076] S312: Calculate the first component difference between the first observed component and the first reference component, and the second component difference between the second observed component and the second reference component, respectively.

[0077] After component decomposition, the difference between each historical sample and the two component dimensions is calculated separately. Specifically, the first observed component is subtracted from the first reference component to obtain the first component difference; the second observed component is subtracted from the second reference component to obtain the second component difference. These two component differences together describe the deviation vector of the sensor measurement value relative to the independent reference value in the two-dimensional space of the horizontal wind field in that historical sample. Since the calculation of the component differences is performed independently in two mutually orthogonal directions, and the difference in each direction is a simple algebraic operation in the real number field, there is no angular cyclic boundary problem. Therefore, the obtained first and second component differences are accurate and unambiguous in physical meaning. In the scenario of ocean buoy wind measurement, the first component difference reflects the difference between the observed wind speed and the reference wind speed in the east-west direction of the historical sample, and the second component difference reflects the corresponding difference in the north-south direction. The combination of the two completely describes the magnitude and direction of the wind field measurement deviation within that historical window.

[0078] S313: Calculate the weight corresponding to each historical sample. The weight decreases as the combined state distance between the historical sample and the current state vector increases.

[0079] A weight is calculated for each historical sample in the similar sample set. This weight adjusts the contribution of the deviation information of each historical sample to the comprehensive calculation of the current window correction parameters. The weight is determined based on the comprehensive state distance between the historical sample and the current state vector. The smaller the comprehensive state distance, the closer the buoy attitude, sea state conditions, and wind field background of the historical sample are to the current time window, and the stronger the applicability of its deviation information to the current window; therefore, it should be given a greater weight. Conversely, the larger the comprehensive state distance, the more different the state of the historical sample is from the current window, and the relatively lower the reference value of its deviation information; therefore, its weight should be reduced accordingly. The weight can be calculated using an exponential decay function or other distance-decreasing mapping methods, ensuring that historical samples with close state distances receive sufficient contribution to the calculation, while samples that are farther away but still within the similar sample set contribute less. This distance-based weighting strategy ensures that the generation of error correction parameters is always supported by the historical sample closest to the current state, balancing computational stability and local adaptability.

[0080] S314: The first component difference and the second component difference of each historical sample are weighted and combined to obtain the first component correction parameter and the second component correction parameter of the current time window, which together constitute the error correction parameter.

[0081] The first and second component differences of each historical sample are weighted and summed using the aforementioned weights. The weighted summation process involves multiplying the first component difference of each historical sample by its corresponding weight, summing the results, and then dividing by the sum of all weights to obtain the first component correction parameter for the current time window. Similarly, the second component differences of all historical samples are weighted and summed to obtain the second component correction parameter. These two component correction parameters together constitute the error correction parameters for the current time window, describing the systematic deviation of the sensor relative to the independent reference value in the east-west and north-south directions under the current attitude conditions and environmental background. In subsequent correction stages, the current observation value is decomposed into corresponding orthogonal components, and these two component correction parameters are subtracted from each component. This achieves deviation correction based on vector decomposition, fundamentally avoiding the period-crossing pseudo-bias problem that may occur when directly correcting wind direction angles in polar coordinates.

[0082] By organically linking the four stages of component decomposition, difference calculation, distance weighting, and component synthesis, the deviation estimation mechanism of this invention achieves accurate quantification of deviations in vector-type physical quantities. Compared to traditional methods that independently calculate wind speed and wind direction deviations or use a single empirical formula for correction, this mechanism transforms the observed and reference values ​​into Cartesian component space for processing. This couples the deviation estimates of wind speed and wind direction into a unified two-dimensional vector correction parameter, ensuring both the numerical stability of the deviation calculation and the clear physical consistency of the correction process. This significantly improves the accuracy and robustness of wind measurement data correction for non-stationary platforms.

[0083] In the basic implementation, the weight of each historical sample is directly determined by the comprehensive state distance; the closer the distance, the greater the weight, and the farther the distance, the smaller the weight. This single-distance-dependent weighting strategy works effectively in most cases. However, when the deviation of individual historical samples in a similar sample set deviates significantly from that of other samples, even if the state distance of the sample is relatively close, its deviation information may be unreliable due to accidental anomalies of the sensor within that historical window or instantaneous fluctuations of the reference data. In this case, if a large weight is still assigned based solely on state distance, the abnormal sample will have a disproportionate impact on the overall weighted synthesis result, causing the correction parameter to deviate from the true deviation value. To address this, the present invention further introduces a dual-weighted correction mechanism based on deviation stability, superimposing stability constraints on the distance weight, so that the weight reflects both state similarity and deviation reliability.

[0084] Further, step S313 calculates the weights corresponding to each historical sample, including: S3131: Calculate the initial weight of each historical sample based on the comprehensive state distance between each historical sample and the current state vector. The initial weight decreases as the comprehensive state distance increases.

[0085] For each historical sample in the similar sample set, its initial weight is calculated based on the comprehensive state distance between that historical sample and the current state vector. The initial weight and the comprehensive state distance have a monotonically decreasing relationship; that is, the smaller the comprehensive state distance, the larger the initial weight. This monotonically decreasing relationship can be achieved using an exponential decay function, or by using the reciprocal of the distance or other mapping functions that converge to zero as the distance increases. The initial weight only measures the reference value of each historical sample for estimating the current window bias from the perspective of state similarity, and does not yet consider the statistical consistency of the historical sample's own bias across the entire similar sample set.

[0086] S3132: Calculate the third statistical dispersion of the first component difference and the fourth statistical dispersion of the second component difference for all historical samples in the similar sample set.

[0087] After obtaining the first and second component differences of all historical samples in the similar sample set, the statistical dispersion of these two component differences is calculated respectively. The third statistical dispersion is calculated for the first component differences of all historical samples, and the fourth statistical dispersion is calculated for the second component differences of all historical samples. Statistical dispersion can characterize the degree of concentration or dispersion of the deviation in each component dimension of the similar sample set as a whole. Its specific implementation can be standard deviation, variance, median absolute deviation, or other statistics that can reflect the data dispersion characteristics. In the scenario of ocean buoy wind measurement, if the east-west wind component deviations of each historical sample in the similar sample set are generally concentrated within a small range, the third statistical dispersion is low, indicating that the deviations of these historical samples in this direction are relatively consistent; conversely, if the east-west deviations of each sample are scattered and vary greatly, the third statistical dispersion is high, suggesting that the deviations of some historical samples may be abnormal.

[0088] S3133: For each historical sample, calculate the stability correction coefficient based on the degree of deviation between the first component difference and the third statistical dispersion, and the degree of deviation between the second component difference and the fourth statistical dispersion. The stability correction coefficient decreases as the degree of deviation increases.

[0089] For each historical sample in the similar sample set, the deviation of its bias in the first and second component dimensions from the overall distribution is evaluated. Specifically, the first component difference of the historical sample is compared with the third statistical dispersion to obtain the relative bias of the sample in the first component dimension; the second component difference of the historical sample is compared with the fourth statistical dispersion to obtain the relative bias of the sample in the second component dimension. The stability correction coefficient of the historical sample is determined based on the bias in these two dimensions. The stability correction coefficient has a monotonically decreasing relationship with the bias, that is, the greater the deviation of the component difference of the historical sample from the overall dispersion, the less representative its bias information is in the entire similar sample set, and the smaller the stability correction coefficient of the sample; conversely, the smaller the bias, the larger the stability correction coefficient, approaching 1. In the scenario of ocean buoy wind measurement, if the east-west wind component deviation of a certain historical sample is significantly greater than the corresponding deviation of other similar samples, it may mean that the buoy encountered an instantaneous abnormal gust of wind within that historical window or that there was a local error in the reference wind field at that moment. In this case, the stability correction coefficient will effectively reduce the influence of the sample in the subsequent weighted generalization, thereby protecting the overall stability of the correction parameters.

[0090] S3134: Correct the initial weights with a stability correction coefficient to obtain the weights corresponding to the historical sample.

[0091] After obtaining the initial weights and stability correction coefficients, the two are multiplied or combined using other fusion methods to obtain the final weights corresponding to the historical samples. The corrected weights integrate information from two dimensions: state similarity and deviation stability. A historical sample can only obtain a high weight if it is sufficiently close to the current window in terms of state and consistent with the overall distribution of the similar sample set in terms of deviation performance. Conversely, even if the state distance is very close, if its deviation is significantly abnormal in the sample set, its weight will be significantly suppressed by the stability correction coefficient. In the long-term operation of ocean buoy wind measurement scenarios, due to factors such as sensor performance drift, instantaneous sea state changes, or accidental fluctuations in reference data, some historical samples with abnormal deviation performance will inevitably be mixed into the similar sample set. This dual-weighted correction mechanism can automatically identify and softly suppress the contribution of these abnormal samples without manual intervention, ensuring that the generation process of correction parameters is robust to local data contamination.

[0092] This invention constructs a weighting mechanism that combines distance and stability weights in the historical sample weight calculation process. This mechanism introduces an adaptive stability correction based on the statistical distribution of deviations, building upon traditional distance weighting. This ensures that the weights reflect not only the similarity of states but also the reliability of deviation information within the overall sample. Compared to existing techniques that rely solely on a single distance factor or directly remove outlier samples using hard thresholds, this soft suppression strategy retains valid information while suppressing abnormal interference. This achieves robust output of the correction parameter estimates under fluctuating data quality conditions, providing a more stable and reliable parameter foundation for subsequent correction stages.

[0093] In the aforementioned implementation, the error correction parameter for the current time window is obtained by distance-weighted averaging of the component differences of each historical sample in the similar sample set. This method directly interpolates the known deviations of historical samples to the current window according to weights, resulting in a simple calculation process and clear physical meaning. However, in some application scenarios, the relationship between measurement deviation and platform attitude, environmental conditions, and background physical fields is not a simple approximation of local constants, but rather involves a more complex correlation structure. For example, when the buoy's overall tilt angle is large, the wind speed measurement deviation may exhibit a nonlinear growth trend with the tilt angle, or the deviation may be simultaneously affected by the interaction of the tilt angle and the significant wave height. In such cases, relying solely on distance-weighted averaging may be insufficient to fully capture the regularity of deviation changes with state. To address this, the present invention provides another deviation estimation implementation based on a local regression model, which establishes a regression mapping relationship between state characteristics and deviation quantities on a similar sample set to predict the error correction parameter for the current window.

[0094] Furthermore, in step S3, based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data, the error correction parameters for the current time window are calculated, including: S321: Select at least some components from the state vectors of each historical sample in the similar sample set as independent variables, and use the difference between the target physical quantity observation data of each historical sample and its external reference data as the dependent variable to construct a regression model.

[0095] At least some components are selected as independent variables from the state vectors of historical samples in a similar sample set. These selected components constitute the feature inputs of the regression model. In the scenario of ocean buoy wind measurement, the state vector components as independent variables typically include characteristic parameters that are physically related to the measurement deviation, such as the composite tilt angle, significant wave height, observed wind speed, and observed wind direction. The strategy for selecting independent variables can be to select all or part of the state vector components, or it can be determined based on a pre-analysis of the correlation between each component and the deviation. The dependent variable of the regression model is the difference between the observed data of the target physical quantity of each historical sample and its external reference data, which is numerically equal to the known deviation of the sensor measurement value relative to the independent reference value within that historical window. The regression model can be a linear regression model or a nonlinear regression model selected based on the empirical relationship between the deviation and the state components.

[0096] In the specific construction of the local regression model, the selection of independent variables needs to consider both the explanatory power of the bias and the actual availability of component data. A typical configuration of independent variables is to select the composite tilt angle, significant wave height, observed wind speed, and the interaction term between the composite tilt angle and significant wave height as regression independent variables from the state vector. The composite tilt angle directly reflects the tilt amplitude of the buoy platform and has a direct physical correlation with the wind speed measurement bias. The larger the tilt angle, the more serious the deviation of the sensor from the geographic reference direction, and the more significant the attenuation effect of the wind speed measurement is usually. The significant wave height reflects the severity of the sea state. Under high sea state conditions, in addition to large tilt, the buoy also experiences violent heave and swaying motions, which generate additional dynamic interference to the wind sensor. Therefore, the significant wave height also has an independent explanatory power for the bias. The interaction term between the composite tilt angle and significant wave height is used to characterize the synergistic amplification effect of the two on the bias: when a large tilt angle and high sea state occur simultaneously, the working conditions of the sensor are extremely harsh, and the measurement bias may exceed the sum of the individual contributions of the two. Incorporating observed wind speed as an independent variable into the model can capture the regularity of deviation changes with wind speed level, such as the sensor's activation threshold effect under low wind speed conditions or the overshoot effect of the cup anemometer under high wind speed conditions. The dependent variables are the wind speed deviation and wind direction deviation for each historical sample, and independent regression models are established for each. The regression modeling of wind direction deviation needs to be combined with a cyclic angle processing mechanism to ensure the accuracy of the regression residual calculation.

[0097] In the implementation of linear regression, the deviation is expressed as a linear combination of the state components plus a constant term and an interaction term. The interaction term is used to characterize the synergistic effect of two or more state components on the deviation. For example, the product of the tilt angle and the significant wave height can reflect the coupling amplification effect on the wind speed measurement deviation when a large tilt angle is superimposed on high sea state.

[0098] S322: Fit the regression model using a similar sample set to obtain the regression coefficients.

[0099] After determining the structure of the regression model, it is fitted using all historical sample data from a similar sample set. The essence of the fitting process is to find the optimal estimates of each regression coefficient in the regression model, given the independent and dependent variables. For linear regression models, the least squares method can be used to solve for the regression coefficients, minimizing the sum of squared residuals between the model's predicted deviation and the actual deviation of historical samples. For nonlinear regression models, corresponding nonlinear optimization methods can be used for coefficient estimation. All data used in the fitting process comes from the selected similar sample set, meaning that the established regression model is a local approximation of the deviation pattern within the current window's neighborhood in the state space, rather than seeking a global optimum across the entire state space. The advantage of this local regression strategy is that it does not require a uniform functional relationship between deviation and state under all possible operating conditions, but allows this relationship to vary naturally with different state regions, thus enabling a more refined characterization of deviation features under different posture conditions and environmental backgrounds. In the scenario of ocean buoy wind measurement, the deviation pattern under moderate sea state may be significantly different from that under extreme sea state. Local regression can automatically adapt to this regional difference without the need for pre-segmentation of the working conditions.

[0100] S323: Substitute the corresponding components in the state vector of the current time window into the fitted regression model to obtain the error correction parameters for the current time window.

[0101] After fitting the regression model and obtaining the regression coefficients, the components corresponding to the state vector of the current time window are substituted into the fitted regression model to calculate the error correction parameters for the current time window. Since the input to the regression model is the actual state characteristics of the current window, and the output is the model's predicted values ​​of the deviations from these states, the obtained error correction parameters integrate the deviation patterns of multiple historical samples in a similar sample set, and match these patterns with the specific state of the current window through regression coefficients. In the implementation using vector component decomposition, two independent regression models can be established for the first component difference and the second component difference, respectively predicting the component correction parameters of the current window in two orthogonal directions. These two models together constitute the complete error correction parameters. In the implementation using direct deviation regression, wind speed deviation and wind direction deviation can also be modeled separately. In this case, the regression modeling of wind direction deviation needs to be combined with a cyclic angle processing mechanism to ensure the correctness of the regression residual calculation.

[0102] This invention provides a deviation estimation implementation method based on a local regression model in the error correction parameter generation stage. This method utilizes the local proximity of similar sample sets in the state space, automatically learning the quantitative mapping relationship between state characteristics and measurement deviation within the current operating area through regression analysis, and predicting the correction parameters for the current window based on this mapping relationship. Compared with the distance-weighted averaging method, the local regression method can more fully utilize the deviation interpretation information contained in each component of the state vector, and more meticulously characterize the changing trend of deviation in the multidimensional state space. It is particularly suitable for buoy observation scenarios with a sufficient number of historical samples and a stable correlation between deviation and state. The two deviation estimation methods can be selected independently or automatically switched based on the characteristics of similar sample sets, providing a technical foundation for the flexible deployment of the correction system under different data conditions.

[0103] After obtaining the error correction parameters for the current time window and completing the data correction through the aforementioned steps, a key unresolved issue remains: the reliability of the correction results cannot be determined. In non-stationary platform observation scenarios such as ocean buoy wind measurement, the correction conditions for different time windows vary significantly—some windows may have sufficient and highly similar reference samples, resulting in high reliability of the correction results; while other windows may experience a decline in correction quality due to sparse historical samples or discrete bias distribution. If correction data is output indiscriminately for all windows, operational users will be unable to determine which correction values ​​are reliable, which are only for reference, and which should not actually be adopted. To address this, this invention establishes a multi-factor fusion-based quantitative evaluation system for correction reliability and executes tiered output decisions driven by reliability indicators.

[0104] Further, in step S4, the target physical quantity observation data of the current time window is corrected using error correction parameters to obtain corrected target physical quantity data, including: S41: Calculate the sample quantity factor based on the number of samples in the similar sample set. The value of the sample quantity factor increases monotonically with the increase of the number of samples and approaches the upper limit.

[0105] like Figure 4 As shown, Figure 4This diagram illustrates the complete logical flow of multi-factor quantitative evaluation and hierarchical output decision-making for correction reliability. The left side of the diagram lists five core evaluation input factors affecting correction reliability: "number of similar samples," "state distance," "wind speed deviation dispersion," "wind direction deviation dispersion," and "reference wind field consistency." These five dimensions comprehensively characterize the correction conditions from aspects such as sample statistical support, state matching degree, deviation stability, and external reference benchmark reliability. After fusion processing in the middle box ("Correction Reliability Calculation"), a normalized correction reliability index is output. Based on the hierarchical comparison of this index with preset multiple thresholds (first threshold and second threshold), the right-hand output branch provides three differentiated processing strategies: for windows with high reliability, "high-reliability correction" data is output with a high-reliability label; for windows with intermediate reliability, "reference correction" data is output with a reference correction label; and for windows with insufficient reliability, a "low-reliability label" is output, directly retaining the original observation data without performing correction. This flowchart systematically demonstrates the security barrier and hierarchical control mechanism established by this invention in the output stage.

[0106] The number of historical samples included in the similar sample set is a fundamental factor affecting the reliability of correction. When the number of similar samples is sufficient, the estimation of error correction parameters has sufficient statistical support, the random fluctuations of individual samples have limited impact on the overall result, and the correction result tends to be stable. Conversely, when the number of similar samples is small, the sensitivity of the correction parameters to individual samples increases, the variance of the estimation result rises, and the reliability decreases accordingly. Based on this principle, the value of the sample size factor is designed to increase monotonically with the increase of the sample size, and gradually approach an upper limit value after the sample size reaches a certain scale, for example, approaching 1. In the range of small sample sizes, the growth of the sample size factor is relatively rapid, reflecting the rapid improvement in the reliability of correction from the stage of "insufficient samples" to "basically sufficient samples"; when the sample size continues to increase, the growth of the factor slows down and approaches saturation, which conforms to the marginal diminishing law of increasing sample size for improving accuracy in statistical estimation. In the scenario of ocean buoy wind measurement, if there are only a few historical samples in the similar sample set of the current window, the sample number factor will be at a low level, indicating that the statistical basis of this correction is weak; if the number of similar samples reaches dozens or more, the sample number factor will be close to the upper limit, indicating that the correction parameters have sufficient statistical support.

[0107] S42: Calculate the state distance factor based on the comprehensive state distance between each historical sample in the similar sample set and the current state vector. The value of the state distance factor increases monotonically as the comprehensive state distance decreases.

[0108] The comprehensive state distance between each historical sample in the similar sample set and the current state vector reflects the overall similarity between these historical samples and the current window in multiple dimensions such as platform attitude, sea state environment, observation characteristics, and background wind field. The smaller the comprehensive state distance, the more consistent the conditions of the historical samples are with the current window, the stronger the applicability of their deviation information to the extrapolation of the current window, and the higher the reliability of the correction result. Based on this principle, the value of the state distance factor is designed to monotonically increase as the comprehensive state distance decreases. In specific implementation, the average or median of the comprehensive state distance of all historical samples in the similar sample set can be taken as the representation of the overall state distance of the sample set, and the state distance factor decreases as the representation value increases. When the overall state distance of the similar sample set is small, it indicates that these samples are closely surrounding the current window in the state space, and the correction parameters estimated based on them have high local consistency; when the overall state distance is large, even if the number of samples is sufficient, the state differences between these samples and the current window cannot be ignored, and the correction reliability is correspondingly reduced.

[0109] S43: Calculate the deviation stability factor based on the degree of dispersion between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data.

[0110] The differences between the observed values ​​and reference values ​​of historical samples in a similar sample set constitute the direct basis for bias estimation. If the biases of these historical samples are close to each other and concentrated in distribution, it indicates that the sensor's measurement bias has good consistency and repeatability within this state area, and the estimated correction parameters are stable and reliable. Conversely, if the biases of each historical sample are discrete and vary greatly, it means that even under similar state conditions, the sensor's measurement bias still exhibits large random fluctuations, and the correction parameters estimated based on these samples have high uncertainty. The bias stability factor is designed to decrease as the degree of bias dispersion increases; that is, the higher the dispersion, the lower the stability factor. The degree of bias dispersion can be obtained by calculating the variance, standard deviation, or other dispersion measures of the biases of each historical sample. In the scenario of ocean buoy wind measurement, if the east-west wind component bias and the north-south wind component bias of a certain set of similar samples are concentrated within a narrow numerical range, the bias stability factor will be close to the upper limit, indicating that the bias pattern within this state area is clear and reliable; if the biases are dispersed over a large range, the bias stability factor decreases, and the correction reliability decreases accordingly.

[0111] S44: The sample size factor, state distance factor, and bias stability factor are fused to obtain the correction confidence.

[0112] After obtaining the sample size factor, state distance factor, and bias stability factor separately, these three factors are fused to obtain a comprehensive calibration credibility index. The fusion method can be the product of the factors, a weighted sum, or other fusion functions that comprehensively reflect multi-dimensional information. When using product fusion, a low value in any factor will significantly lower the overall credibility, which aligns with the practical needs of calibration credibility assessment—a deficiency in any of the three dimensions—sample size, state similarity, and bias stability—is sufficient to undermine the reliability of the calibration results. The calibration credibility value can be normalized to between 0 and 1, with a higher value indicating a higher degree of credibility in the calibration result. Through this fusion mechanism, calibration credibility comprehensively quantifies calibration quality from multiple independent and complementary dimensions, rather than relying solely on a single indicator for a biased judgment.

[0113] S45: When the correction confidence level is higher than the first threshold, output the corrected target physical quantity data with a high confidence indicator. When the correction confidence level is lower than or equal to the first threshold, output the original target physical quantity observation data with a low confidence indicator.

[0114] After obtaining the correction confidence level, it is compared with a preset first threshold to determine the output strategy. The first threshold is the decision boundary distinguishing between "reliable correction" and "unreliable correction". When the correction confidence level is higher than the first threshold, it indicates that the sample size, state similarity, and bias stability on which the correction is based are all at a good level, and the correction result is trustworthy. In this case, the corrected target physical quantity data is output with a high confidence label to inform downstream users that the data has been reliably corrected and can be directly used for subsequent meteorological analysis, model validation, or engineering design. When the correction confidence level is lower than or equal to the first threshold, it indicates that one or more conditions for the correction are obviously insufficient, and the correction result has a large uncertainty. In this case, forced correction is not performed, and the original target physical quantity observation data is directly output with a low confidence label to remind downstream users that the data has not been corrected or the correction confidence level is insufficient, and they need to assess the risks themselves when using it.

[0115] After correcting the current observation data using error correction parameters, the reasonableness of the correction results should be verified. In rare cases, due to poor quality of similar samples or an unusual window state, the corrected wind speed value may be physically unreasonable to be negative, and the corrected wind direction value may also exceed the standard definition range of 0° to 360°. For cases where the corrected wind speed is negative, it can be set to zero, as negative wind speed does not physically exist, and this situation itself indicates a potential problem with the correction. For cases where the corrected wind direction exceeds the standard range, it can be converted to the 0° to 360° definition range using angle modulo operations. When such abnormal correction results occur, even if the correction confidence does not trigger a low confidence flag, a corresponding anomaly marker should be included in the output to alert downstream users that the data quality of this window may have potential problems.

[0116] This invention constructs a credibility assessment and output control mechanism based on "factor quantification—multidimensional fusion—hierarchical decision-making" in the correction output stage. This mechanism makes the previously implicit judgment of correction reliability explicit and quantifiable, enabling the correction system to have self-evaluation capabilities. Compared with the common practice in existing technologies that lack correction credibility assessment and output correction values ​​indiscriminately for all windows, this mechanism sets up a credibility-based quality control checkpoint at the end of the automatic correction process, ensuring that correction data is output and marked with a high-credibility label only when correction conditions are sufficient, and conservatively retaining the original observation values ​​when conditions are insufficient. This hierarchical output strategy takes into account both correction coverage and correction security, providing traceable and controllable automated correction capabilities for operational data post-processing systems.

[0117] In the aforementioned implementation, the calibration reliability is obtained by fusing three factors: the sample quantity factor, the state distance factor, and the deviation stability factor. The deviation stability factor is a holistic measure of the dispersion of deviations in each historical sample within a similar sample set. However, when using vector component decomposition to estimate the deviation of a target physical quantity, the observed data and reference data are decomposed into a first observation component and a second observation component, which are mutually orthogonal. The deviations in the two component directions may exhibit different statistical characteristics. For example, in an ocean buoy wind measurement scenario, the dispersion of the east-west wind component deviation may not be consistent with the dispersion of the north-south wind component deviation. This difference may conceal the different response characteristics of the buoy platform to wind field measurements in different directions. Mixing the deviations in the two component directions together to calculate a single deviation stability factor would obscure the diagnostic information carried by this directional difference. Furthermore, the three factors in the aforementioned reliability assessment system are all based on the internal statistical characteristics of the similar sample set itself and do not yet incorporate consideration of the quality of the external reference data itself. External reference data, serving as the benchmark for bias estimation, directly determines the upper limit of the accuracy of the correction results based on its reliability. If the external reference data itself has significant errors or uncertainties, even with a sufficient number of similar samples, high state matching, and highly stable bias, the credibility of the correction results should be questioned. Therefore, this invention further decomposes the bias stability factor according to the component directions and introduces a reference consistency factor to construct a more complete five-factor credibility fusion system.

[0118] Furthermore, in S44, the sample size factor, state distance factor, and bias stability factor are fused to obtain the correction confidence level, including: S441: Decompose the bias stability factor into a first observation bias stability factor corresponding to the first observation component and a second observation bias stability factor corresponding to the second observation component, and calculate the reference consistency factor. The reference consistency factor is determined based on at least one of the following: the source type of the external reference data, the spatiotemporal resolution, and the uncertainty index.

[0119] The first operation involves splitting the original single bias stability factor into a first observation bias stability factor and a second observation bias stability factor. The first observation bias stability factor specifically calculates the dispersion of the component differences in the first observation component direction for each historical sample in the similar sample set, while the second observation bias stability factor specifically calculates the dispersion of the component differences in the second observation component direction. After splitting, the bias stability in each component direction is independently evaluated, which can more accurately reflect the consistency of the sensor's bias in different spatial directions. In the scenario of ocean buoy wind measurement, if there is a systematic difference between the east-west and north-south sway amplitudes of the buoy platform, it may cause the wind component measurement bias in the two directions to exhibit different stability characteristics. The split bias stability factors can capture the reliability signals in these two directions separately.

[0120] The second step is to calculate the reference consistency factor. The reference consistency factor is the only factor in the calibration credibility assessment system that incorporates the quality information of the external reference data itself. Its role is to include the credibility of the reference benchmark on which the bias estimation relies in the overall assessment. The determination of the reference consistency factor can include at least one of the following: the source type of the external reference data, its spatiotemporal resolution, and uncertainty indices. In terms of source type, external reference data from different sources have different levels of authority and accuracy. For example, the credibility of global reanalysis data, regional high-resolution numerical model data, measured data from nearby fixed stations, and measured data from offshore wind towers increases sequentially, allowing for different basic consistency scores to be assigned to different source types. In terms of spatiotemporal resolution, the higher the temporal and spatial resolution of the external reference data, the stronger its representativeness of the current observation window, and the higher the reference consistency factor; conversely, coarse-resolution reference data may not accurately depict the instantaneous changes in small-scale wind fields, resulting in a moderately lower reference consistency factor. In terms of uncertainty indices, some external reference data products contain uncertainty estimation information per grid point or per time interval, which can be directly mapped to the value of the reference consistency factor; the lower the uncertainty, the higher the factor. The final value of the reference consistency factor can be determined jointly by multiple dimensions through weighted summation or conditional judgment. By introducing the reference consistency factor, the credibility assessment system is corrected to expand from focusing only on the quality of internal samples to also taking into account the quality of external reference benchmarks, making the assessment dimensions more complete.

[0121] S442: The sample size factor, state distance factor, first observation bias stability factor, second observation bias stability factor, and reference consistency factor are fused together to obtain the correction confidence.

[0122] After decomposing the bias stability factor and calculating the reference consistency factor, the five factors—sample quantity factor, state distance factor, first observation bias stability factor, second observation bias stability factor, and reference consistency factor—are fused to obtain the final calibration confidence level. The fusion method can be the product of the factors, weighted summation, or other comprehensive functions. When using product fusion, any factor at a low level will significantly suppress the overall calibration confidence level. This characteristic aligns closely with the practical logic of calibration confidence assessment. For example, in an ocean buoy wind measurement scenario, if a calibration uses a sufficient number of similar samples, has close state distances, and stable biases in both component directions, but the external reference data comes from a global reanalysis product with coarse spatiotemporal resolution, the reference consistency factor will be at a moderately low level. This will appropriately lower the overall calibration confidence level, avoiding an overly high confidence assessment due to limited reference accuracy. Conversely, if all five factors are at a high level, it indicates that the correction has performed well in all four dimensions of sample sufficiency, state similarity, bias stability, and reference benchmark reliability. The correction credibility is close to the upper limit, and the correction results are highly reliable.

[0123] This invention expands the evaluation dimensions from three to five factors in the credibility fusion process. The directional splitting of the deviation stability factor allows for the independent expression of deviation statistical characteristics in different component directions, avoiding the mixing and loss of information regarding differences between directions. The introduction of the reference consistency factor incorporates the quality of external reference data into the evaluation of calibration credibility, filling the logical gap in previous credibility systems that only focused on internal sample statistics while neglecting the quality of external benchmarks. This five-factor fusion mechanism enables the calibration credibility index to comprehensively quantify calibration quality across four dimensions and five factors: sample sufficiency, state matching degree, component deviation stability, and reference benchmark reliability. This provides a more rigorous and complete basis for subsequent graded output decisions.

[0124] The comparison between the correction confidence level and the first threshold determines the output strategy. If the result is higher than the threshold, corrected data is output with a high confidence label; if it is lower than or equal to the threshold, the original data is output with a low confidence label. This two-level output strategy is simple and clear, effectively distinguishing between "confidential" and "unreliable" states. However, in actual operational data processing of ocean buoy wind measurement, the distribution of correction confidence level is often a continuous and gradual change rather than a clear dichotomy. A considerable portion of the time window has correction conditions in the intermediate zone between "completely reliable" and "completely unreliable," with a reasonable but not abundant number of similar samples, relatively close but not tightly coupled state distances, and moderate but not highly consistent deviation stability. For this portion of the window, completely abandoning correction means losing potential data improvement opportunities, while equating it with the high-confidence correction window for indiscriminate output may mislead downstream users into overestimating data accuracy. Therefore, in a further refined embodiment of S45, this invention introduces a second threshold, expanding the two-level output to a three-level hierarchical output, adding an intermediate output state of "reference correction," making the output decision granularity driven by correction confidence more refined.

[0125] Further, in step S45, when the correction confidence level is higher than the first threshold, the corrected target physical quantity data is output with a high confidence indicator. When the correction confidence level is lower than or equal to the first threshold, the original target physical quantity observation data is output with a low confidence indicator, including: S451: Determine whether the calibration confidence level is higher than the first threshold.

[0126] The correction confidence level calculated in S44 is compared with a preset first threshold. The first threshold is the decision boundary distinguishing between "high-confidence correction" and "non-high-confidence state," and its specific value can be determined based on statistical analysis of historical data or the specific requirements of operational applications for correction accuracy. In the scenario of ocean buoy wind measurement, the first threshold is usually set to a relatively high value, ensuring that the correction confidence level crosses this threshold only when the correction conditions perform well in multiple dimensions such as sample size, state similarity, component deviation stability, and external reference reliability. This judgment step is the starting point of the entire hierarchical decision-making process, determining whether the subsequent path leads to the high-confidence output path of S452 or the further judgment path of S453.

[0127] S452: When the correction confidence level is higher than the first threshold, output the corrected target physical quantity data and attach a high confidence indicator.

[0128] When the correction confidence level exceeds the first threshold, it indicates that the current time window is at a good level across all dimensions evaluated in S441 and S442, the error correction parameters have sufficient statistical support, consistent state matching and bias patterns, and the quality of the external reference data is trustworthy. At this point, the corrected target physical quantity data is output, accompanied by a high confidence indicator. This high confidence indicator explicitly informs downstream data users that the correction results for this time window have undergone rigorous multi-dimensional confidence evaluation and can be directly used as high-quality observational data for subsequent meteorological analysis, numerical model assimilation, wind energy resource assessment, or other applications requiring high data accuracy.

[0129] S453: When the correction confidence is lower than or equal to the first threshold, determine whether the correction confidence is higher than the second threshold, and the second threshold is lower than the first threshold.

[0130] When the calibration confidence level is lower than or equal to the first threshold, it does not mean that the calibration result is completely unusable, but rather that the calibration has not yet reached the highest confidence level. At this point, a second threshold is introduced for secondary judgment; the value of the second threshold is lower than the first threshold. The second threshold is the decision boundary that distinguishes between "reference calibration" and "low confidence no calibration." This judgment step allows intermediate states where calibration conditions are acceptable but have not reached a high standard to enter the reference calibration path of S454, rather than being directly assigned to the low confidence output path of S455. In the scenario of ocean buoy wind measurement, the second threshold can be set to a relatively lenient but still discriminative value, ensuring that calibration is only abandoned when the calibration conditions are truly poor.

[0131] S454: When the correction confidence level is higher than the second threshold, output the corrected target physical quantity data and attach a reference correction identifier.

[0132] When the correction confidence level is lower than or equal to the first threshold but higher than the second threshold, it indicates that the correction conditions in the current time window are at an intermediate level. The number of similar samples may be barely sufficient, the state distance may be slightly large but not to the point of losing its reference value, the stability of the deviation in each component direction may be moderate, and the quality of the external reference data may be average. In this case, the error correction parameters still have some reference value, and the correction direction for the original data is likely correct, but the accuracy of the correction magnitude may not be as good as that of the high-confidence correction window. At this time, the corrected target physical quantity data is still output, but with a reference correction label instead of a high-confidence label. The reference correction label conveys a clear signal to the data user: the data has been corrected and is statistically closer to the true value than the original observation. However, since the correction conditions are not optimal, it is recommended that users use it as reference data and carefully evaluate it in conjunction with other information in critical applications with strict requirements for data accuracy. The core value of this intermediate output strategy is that it opens a buffer between "corrected" and "uncorrected," allowing a considerable portion of the observation data in the ambiguity zone to receive a partial improvement, while preventing users from misjudging the correction accuracy through label differentiation.

[0133] S455: When the correction confidence is lower than or equal to the second threshold, output the original target physical quantity observation data with a low confidence flag.

[0134] When the correction confidence level is lower than or equal to the second threshold, it indicates that the correction conditions for the current time window are significantly insufficient across multiple evaluation dimensions. This could be due to excessively sparse similar samples, excessively large state distances, overly discrete distribution of deviations in each component direction, or excessively low reliability of external reference data. In such cases, the uncertainty of the error correction parameters exceeds acceptable limits. Forcibly applying correction not only fails to improve data quality but may also introduce directional errors or inappropriate magnitude corrections, making the data deviate from the true value more severely than the original observations. In this situation, no correction operation is performed; the original target physical quantity observation data is directly output with a low-confidence flag. The low-confidence flag clearly informs the data user that this time window was not corrected due to insufficient correction conditions, and the output data is the original sensor observation, which may contain systematic biases caused by non-stationary platform motion. The user must assess the applicability of this data independently. This safety blocking strategy ensures that the automated correction system does not produce misleading outputs even under the most unfavorable conditions.

[0135] This invention constructs a three-tiered output system in the hierarchical output decision-making process: "high-confidence correction—reference correction—low-confidence no correction." This system uses correction confidence as the core decision variable, mapping continuous confidence values ​​to three discrete output levels through a first and second threshold. Each level corresponds to a different correction strategy and labeling method. Compared to a two-tiered output strategy, the three-tiered output system more precisely defines the transition range of correction quality, ensuring that data in intermediate states receives correction improvement without misleading users due to improper labeling. In the long-term operational processing of ocean buoy wind measurement data, this mechanism strikes a balance between maximizing usable data output and strictly controlling data quality risks, providing tiered data service capabilities for the diverse needs of downstream applications.

[0136] Furthermore, the attitude features extracted in step S2 based on attitude statistics include the comprehensive tilt angle, which is calculated by taking the square root of the sum of the squares of the pitch angle statistics and the roll angle statistics.

[0137] In step S2, when extracting attitude features based on attitude statistics, this invention uses the comprehensive tilt angle as the core attitude feature characterizing the severity of platform tilt. The comprehensive tilt angle is calculated by taking the square root of the sum of the squares of the pitch angle and roll angle statistics. In the scenario of ocean buoy wind measurement, the buoy platform simultaneously generates swaying motions around multiple axes under the combined action of wind, waves, and currents. The pitch angle describes the buoy's forward and backward tilt amplitude in the longitudinal axis direction, and the roll angle describes the buoy's left and right tilt amplitude in the transverse axis direction. Each of these only reflects the degree of tilt in a single direction and cannot comprehensively characterize the overall tilt state of the buoy's deviation from the vertical line. By taking the square root of the sum of the squares of the pitch angle and the roll angle, the two-dimensional tilt information is compressed into a scalar index. This comprehensive tilt angle is physically equivalent to the angle between the buoy's vertical axis and the geographic vertical line. The larger the value, the more severe the overall tilt of the buoy, and the more serious the deviation of the wind measurement sensor from the geographic coordinate system. Compared to incorporating pitch and roll angles as two separate attitude features into the state vector, using a combined tilt angle can reduce the dimensionality of the state vector without losing key tilt information, reduce the redundancy of attitude-related components in similarity measurement, and make the relationship between attitude features and measurement deviations more direct and straightforward.

[0138] The state vector includes the composite tilt angle, azimuth angle, significant wave height, wave peak period, dominant wave direction, observed target physical quantities, and external reference data. The composite tilt angle and azimuth angle together constitute the attitude characteristics, characterizing the platform's spatial attitude from the perspectives of tilt amplitude and horizontal direction, respectively. Significant wave height, wave peak period, and dominant wave direction together constitute the environmental characteristics. Significant wave height represents the wave intensity level, wave peak period represents the wave frequency characteristics, and dominant wave direction represents the dominant direction of wave propagation; these three together describe the overall sea state within the current window. The observed target physical quantities serve as observation features, while the external reference data values ​​serve as environmental background constraint features. Together, they incorporate the measured values ​​of the current window and the independent reference benchmark into the state representation. The key to this invention is directly incorporating the external reference data values ​​into the state vector, rather than merely using them as an external benchmark for deviation calculation. This design ensures that the selection of similar samples depends not only on the buoy's own attitude and sea state, but also on the independent wind field background. This guarantees that the retrieved historical samples are consistent with the current window in terms of the overall wind field environment, avoiding the background wind field mismatch problem that may occur when searching solely based on the buoy's own state. Nine components provide a comprehensive and compact state encoding of the current time window from four dimensions, providing sufficient and dimensionally reasonable query input for subsequent similarity searches.

[0139] It should be understood that the complete nine-component configuration of the state vector described above is the preferred embodiment of the present invention under conditions of complete data availability. In actual engineering deployments, due to differences in sensor configuration and data transmission capabilities among different buoy systems, some components may be unavailable at certain times. For example, some buoys may not be equipped with a wave spectrum measurement module, thus failing to provide wave spectrum peak period data, or they may not be equipped with a directional wave measurement module, thus failing to provide main wave direction data. Under such conditions of limited data availability, the state vector can be constructed in a simplified form. A typical simplified form is a four-component state vector, which only includes the composite tilt angle, significant wave height, observed values ​​of target physical quantities, and external reference data values. This simplified form retains the core tilt information in attitude characteristics, wave intensity information in environmental characteristics, observation characteristics, and external reference constraint characteristics. Although it reduces the accuracy of state characterization to some extent, it can still provide effective similarity retrieval capabilities under most operating conditions. As the available data dimensions increase, the state vector can be gradually expanded to the complete form. State vectors of different simplification levels can be mixed and used in the same historical sample library, requiring only adaptation processing for missing components when performing similarity measurements.

[0140] Specifically, the historical sample library pre-constructed in step S2 includes at least one of the following sources: historical observation data from the current sensor-mounted platform, historical observation data from other platforms of the same model under similar environmental conditions, and data from stable observation periods confirmed by quality control. The pre-constructed historical sample library is the data foundation for similarity retrieval, and its quality and coverage directly affect the reliability and applicability of the correction effect. The current platform's own historical observation data is the most direct and reliable sample source because the sensor installation method, platform response characteristics, and measurement environment of the same platform at different times are highly consistent, and the deviation patterns accumulated in historical periods have the strongest transferability to the current period. Historical observation data from other platforms of the same model under similar environmental conditions serves as a supplementary source, providing additional reference samples when the current platform's own historical data is insufficient or encounters rare operating conditions. Platforms of the same model have the same structural design, sensor configuration, and similar dynamic response characteristics, and the measurement deviation patterns exhibited when operating in similar sea areas or under similar environmental conditions have high consistency; therefore, their historical data can serve as an effective reference for cross-platform sharing. Stable observation period data verified through quality control refers to samples from specific time periods that have been confirmed through manual review or automated quality assessment as indicating good sensor performance, smooth platform movement, and reliable measurement results. The data quality in these samples is verified, and the deviation relationship between their observed values ​​and reference values ​​is more reliable. As historical samples, they can provide a higher-quality reference benchmark for deviation estimation. These three sources can be used independently or combined to construct a richer and more robust historical sample library. In practical deployments, the historical sample library can be continuously expanded and updated as the observation system continues to operate, resulting in a wider range of matching historical states over time and a gradually enhanced adaptability of the calibration system.

[0141] Specifically, the external reference data in step S1 includes one or more of the following: reanalysis wind field data, regional numerical model wind field data, observation data from nearby fixed stations, observation data from offshore wind towers, and quality-controlled reference data. In step S1, the external reference data are reliable reference measurements independent of the sensors of the non-stationary platform to be calibrated, providing an objective quantitative benchmark for bias estimation.

[0142] Reanalysis wind field data is a long-term gridded wind field product generated by fusing historical observation data with numerical weather prediction models through data assimilation techniques. It offers advantages such as complete spatial coverage and a long time span, making it suitable for retrospective bias correction of historical buoy data. Regional numerical model wind field data is a wind field product output from a high-resolution regional meteorological model that performs a detailed simulation of a specific sea area. Its spatiotemporal resolution is typically higher than global reanalysis data, enabling it to better characterize the detailed changes in local wind fields. Nearby fixed-site observation data comes from fixed observation platforms deployed near buoys, such as coastal meteorological stations, island observation stations, or fixed offshore platforms. Because the sensors at fixed sites are installed in a stable orientation and are unaffected by platform movement, their wind field measurements have high accuracy and stability, serving as a high-quality reference benchmark for buoy wind measurement data correction. Offshore wind tower data comes from fixed wind towers specifically used for wind energy resource assessment. The layered wind speed and direction sensors installed on these towers have high measurement accuracy and are unaffected by platform movement, making them an authoritative data source for nearshore wind field observation. The quality-controlled reference data refers to a more reliable reference data product obtained by automatically or manually reviewing data from any of the above-mentioned sources through quality screening, anomaly removal, and consistency checks. The availability of multiple reference data sources allows the method of this invention to flexibly adapt to data availability conditions in different sea areas and time periods. For example, at deep-water buoy stations far from the coast, data from nearby fixed stations and wind towers may not be available; in this case, reanalysis wind field data or numerical model wind field data can be used as an external reference. At nearshore buoy stations, more accurate fixed station observation data or wind tower data can be preferentially selected as the reference benchmark. This multi-source selection design improves the applicability and portability of the entire correction method under different deployment conditions.

[0143] Furthermore, through the organic combination of the above three supplementary features, this invention has made targeted designs in the three basic aspects of the correction method: attitude characterization, sample library construction, and reference data acquisition. The introduction of the comprehensive tilt angle simplifies the attitude feature dimension and enhances the correlation between features and deviations. The inclusion of external reference data values ​​in the state vector ensures the consistency of wind field background in the screening of similar samples. The multi-source composition of the historical sample library expands the sample coverage and improves the retrieval robustness under sparse conditions. The multiple source types of external reference data give the method flexible configuration capabilities under different data availability conditions. The design of these basic aspects, together with the aforementioned core mechanisms such as cyclic angle distance measurement, adaptive weight adjustment, dynamic fault-tolerant screening, vector component deviation estimation, and confidence level output, jointly constitute the complete technical solution of this invention for sensor data error correction on non-stationary platforms.

[0144] It should be understood that although the above specific embodiments use the marine buoy wind measurement scenario as the main example to describe the method of the present invention in detail, the application scope of the present invention is not limited to this specific scenario. The core concept of the present invention lies in constructing a general correction framework for sensor data errors of non-stationary observation platforms. This framework constructs a multi-dimensional state vector by fusing platform attitude characteristics, environmental state characteristics, sensor observation characteristics, and independent external reference constraint characteristics. Based on state similarity retrieval, similar samples are selected from the historical sample library. The error correction parameters are dynamically calculated using the differences between the observed values ​​and independent reference values ​​in the similar samples. The correction results are then evaluated for multi-factor reliability and output in a hierarchical manner. The core problem addressed by this technical framework, namely the dynamic estimation and correction of measurement deviations caused by platform motion on non-stationary platforms, is widely present in many engineering observation and scientific research fields.

[0145] In the field of marine observation, besides wind measurement using marine data buoys, this invention is also applicable to shipborne meteorological observation platforms, drifting buoys, the surfacing observation sections of underwater moorings, environmental monitoring modules of wave energy power generation devices, and integrated meteorological monitoring systems for offshore oil platforms. The observed physical quantities are not limited to wind speed and direction, but can also include vector or directional marine environmental parameters such as ocean current profiles and wave direction spectra, which are also affected by the platform's non-stationary motion. In the field of atmospheric observation, this invention can be applied to scenarios such as meteorological sensors carried by tethered balloons, UAV-borne atmospheric composition analyzers, and radial wind inversion by mobile weather radars. The continuous swaying and attitude changes of these platforms in the air will also cause measurement deviations. In the field of ground-based mobile observation, the method of this invention can be extended to scenarios such as vehicle-mounted air quality monitoring stations, mobile noise monitoring platforms, and shipborne hydrological measurement systems. The impact of platform motion and attitude changes on sensor measurements also needs to be quantitatively estimated and corrected. In the broader field of industrial measurement, any sensor system deployed on a non-stationary carrier that needs to acquire reliable measurement data continuously over a long period of time, as long as it can acquire platform attitude statistics, environmental state data, and external reference data independent of the sensor, can be incorporated into the method framework provided by this invention for deviation correction processing.

[0146] Accordingly, please refer to Figure 5 The second aspect of this invention provides a system for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints. This system corrects errors in sensor data on a non-stationary platform based on the aforementioned method, including: Data acquisition module 1 acquires target physical quantity observation data collected by the non-stationary platform sensor within a preset time window, attitude statistics of the sensor-mounted platform, environmental state data of the sensor's environment, and external reference data independent of the non-stationary platform sensor. The external reference data consists of reliable reference measurements of the same type as the target physical quantity.

[0147] Vector calculation module 2 constructs the state vector for the current time window and calculates the similarity between the current state vector and each historical state vector from a pre-built historical sample library based on a preset similarity metric. Historical samples that meet the preset similarity conditions are selected to form a similar sample set. The state vector includes at least attitude features extracted from attitude statistics, environmental features extracted from environmental state data, observation features extracted from target physical quantity observation data, and environmental background constraint features extracted from external reference data. Each historical sample in the historical sample library contains data of the same type as the current time window and its corresponding state vector.

[0148] Error calculation module 3 calculates the error correction parameters for the current time window based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data.

[0149] Data correction module 4 uses error correction parameters to correct the target physical quantity observation data in the current time window, and obtains the corrected target physical quantity data.

[0150] Accordingly, a third aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-described non-stationary platform sensor data error correction method based on attitude perception and environmental background constraints.

[0151] Accordingly, a fourth aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints.

[0152] The embodiments of the present invention aim to protect a method and system for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints, which has the following effects: 1. By constructing a multi-dimensional state vector that integrates attitude features, environmental features, observation features, and external reference constraints, and using external reference data independent of the sensor to be calibrated as a component of the state vector, the selection of similar historical samples not only depends on the sensor's own observation data and platform state data, but is also constrained by an independent and reliable reference field. This achieves collaborative representation and joint retrieval of multi-source data from platform attitude, environmental background, and external reference field. Compared with existing technologies that only use sensor data for state identification or only use external references for post-event comparison, this method establishes a dynamic mapping relationship between "state and deviation" at the data processing level. This allows each time window to be calibrated to be matched with historical state samples that are highly similar to its attitude conditions, environmental conditions, and background physical quantity field distribution. This improves the adaptability of deviation estimation to complex working conditions and solves the problem of inaccurate deviation estimation in existing methods when platform state and background field conditions are coupled and change. 2. In the process of similarity measurement and deviation estimation, a cyclic difference measurement method based on the arctangent function is adopted for cyclic angle type physical quantities. The observed data and reference data are decomposed into mutually orthogonal vector components, and the component difference is calculated and weighted. This fundamentally avoids the technical defects of angular physical quantities such as wind direction and azimuth when crossing the 0° / 360° cycle boundary, which produce pseudo distance or pseudo deviation due to conventional arithmetic subtraction. In the error correction parameter generation stage, a stability correction mechanism based on the degree of deviation dispersion is introduced. Anomalous historical samples that deviate from the overall deviation distribution in the similar sample set are softly suppressed, rather than simply removed, so that the deviation estimation results remain numerically stable under the condition of uneven sample quality. Compared with the shortcomings of existing empirical correction methods that ignore the angular cycle characteristics and lack effective control over anomalous samples, a special processing mechanism for cyclic physical quantities and anomalous samples is built in the complete data processing chain from distance calculation and deviation estimation to parameter generation, which improves the numerical stability and physical consistency of the correction output. 3. A multi-factor fusion-based calibration reliability quantification evaluation system and a hierarchical output decision mechanism were established in the calibration output stage. By comprehensively considering factors such as the number of similar samples, the degree of state similarity, deviation stability, and the reliability of external reference data itself, a normalized calibration reliability index is generated. Based on the relationship between this index and preset multi-level thresholds, a hierarchical output strategy is executed. When the reliability is high, the calibration data is output with a high reliability label; when the reliability is medium, the calibration data is output with a reference label; and when the reliability is low, only the original observation data is output without performing forced calibration. Compared with the common problems of lacking calibration reliability assessment and having the risk of blind calibration in existing technologies, this system deeply binds calibration actions with reliability assessment, builds a safety decision barrier at the end of the data processing flow, and enables the automated calibration system to have self-assessment and risk avoidance capabilities, thereby improving the security, controllability, and traceability of business data calibration processing.

[0153] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0155] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0156] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints, characterized in that, Includes the following steps: S1: Acquire the target physical quantity observation data collected by the non-stationary platform sensor within a preset time window, the attitude statistics of the sensor-mounted platform, the environmental state data of the environment where the sensor is located, and external reference data independent of the non-stationary platform sensor; the external reference data is a reliable reference measurement value of the same type as the target physical quantity. S2: Construct the state vector of the current time window, and calculate the similarity between the current state vector and each historical state vector from the pre-constructed historical sample library based on the preset similarity measurement criteria. Select historical samples that meet the preset similarity conditions to form a similar sample set. The state vector includes at least the attitude features extracted from the attitude statistics, the environmental features extracted from the environmental state data, the observation features extracted from the target physical quantity observation data, and the environmental background constraint features extracted from the external reference data; each historical sample in the historical sample library contains data of the same type as the current time window and a corresponding state vector; S3: Based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data, the error correction parameter for the current time window is calculated; S4: Use the error correction parameters to correct the target physical quantity observation data in the current time window to obtain the corrected target physical quantity data.

2. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 1, characterized in that, Step S2, which involves calculating the similarity between the current state vector and each historical state vector, includes: S211: Identify the value type of each component in the state vector, and mark the components whose values ​​have cyclic boundary characteristics as cyclic angle components. S212: For each cyclic angle component, calculate the angle difference between the component in the current state vector and the corresponding component in the historical state vector, obtain the first intermediate value by taking the sine of the angle difference, and obtain the second intermediate value by taking the cosine of the angle difference. S213: Calculate the arctangent value of the first intermediate quantity and the second intermediate quantity, and take the absolute value as the component distance of the cyclic angle component; S214: Combine the component distances of the cyclic angle components and the component distances of the non-cyclic angle components after a preset standardization process to obtain the normalized distances, and obtain the comprehensive state distance between the current state vector and the historical state vector.

3. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 2, characterized in that, In step S214, the component distance of the cyclic angle component is combined with the normalized distance of the non-cyclic angle component to obtain the comprehensive state distance between the current state vector and the historical state vector, including: S2141: Obtain the cyclic angular component distance and non-cyclic angular component distance for each historical sample, and calculate the first statistical dispersion of the cyclic angular component distance and the second statistical dispersion of the non-cyclic angular component distance respectively. S2142: Determine a first merging weight for the cyclic angular component distance and a second merging weight for the non-cyclic angular component distance based on the ratio of the first statistical dispersion to the second statistical dispersion; the first merging weight decreases as the ratio increases, and the second merging weight increases as the ratio increases; S2143: Using the first merging weight and the second merging weight, the component distance of the cyclic angle component and the normalized distance of the non-cyclic angle component are weighted and merged to obtain the comprehensive state distance.

4. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 1, characterized in that, Step S2, which involves selecting historical samples that meet preset similarity conditions to form a similar sample set, includes: S221: Set a similarity distance threshold, and select only historical samples whose overall state distance is less than the similarity distance threshold; S222: When the number of selected historical samples is less than the preset minimum number, expand the time range of the historical sample library or reduce the similarity distance threshold, and reselect historical samples; S223: When the number of historical samples after reselection is still less than the preset minimum number, the current time window is marked as unsuitable for correction, and the original target physical quantity observation data and the corresponding low confidence flag are output, and subsequent steps S3 and S4 are no longer executed.

5. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 1, characterized in that, Step S3, which calculates the error correction parameters for the current time window based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data, includes: S311: Decompose the target physical quantity observation data of each historical sample in the similar sample set into a first observation component and a second observation component that are mutually orthogonal, and decompose the external reference data of the historical sample into a first reference component and a second reference component that are mutually orthogonal. S312: Calculate the first component difference between the first observed component and the first reference component, and the second component difference between the second observed component and the second reference component, respectively; S313: Calculate the weight corresponding to each historical sample, wherein the weight decreases as the comprehensive state distance between the historical sample and the current state vector increases; S314: The first component difference and the second component difference of each historical sample are weighted and combined according to the weights to obtain the first component correction parameter and the second component correction parameter of the current time window, which together constitute the error correction parameter.

6. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 5, characterized in that, The calculation of the weights corresponding to each historical sample in step S313 includes: S3131: Calculate the initial weight of each historical sample based on the comprehensive state distance between each historical sample and the current state vector. The initial weight decreases as the comprehensive state distance increases. S3132: Calculate the third statistical dispersion of the first component difference of all historical samples in the similar sample set, and the fourth statistical dispersion of the second component difference; S3133: For each historical sample, the stability correction coefficient of the historical sample is calculated based on the degree of deviation between the first component difference and the third statistical dispersion, and the degree of deviation between the second component difference and the fourth statistical dispersion; the stability correction coefficient decreases as the degree of deviation increases. S3134: The initial weights are corrected using the stability correction coefficient to obtain the weights corresponding to the historical sample.

7. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 1, characterized in that, Step S3, which calculates the error correction parameters for the current time window based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data, includes: S321: Select at least some components from the state vectors of each historical sample in the similar sample set as independent variables, and use the difference between the target physical quantity observation data of each historical sample and its external reference data as the dependent variable to construct a regression model; S322: Fit the regression model using the similar sample set to obtain the regression coefficients; S323: Substitute the corresponding components in the state vector of the current time window into the fitted regression model to obtain the error correction parameters for the current time window.

8. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 1, characterized in that, Step S4, which involves using the error correction parameter to correct the target physical quantity observation data within the current time window to obtain corrected target physical quantity data, includes: S41: Calculate the sample quantity factor based on the number of samples in the similar sample set. The value of the sample quantity factor increases monotonically with the increase of the number of samples and approaches the upper limit. S42: Calculate the state distance factor based on the comprehensive state distance between each historical sample in the similar sample set and the current state vector. The value of the state distance factor increases monotonically as the comprehensive state distance decreases. S43: Calculate the deviation stability factor based on the degree of dispersion between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data; S44: The sample quantity factor, state distance factor, and bias stability factor are fused to obtain the correction confidence level; S45: When the correction confidence level is higher than the first threshold, output the corrected target physical quantity data with a high confidence indicator; when the correction confidence level is lower than or equal to the first threshold, output the original target physical quantity observation data with a low confidence indicator.

9. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 8, characterized in that, Step S44, which involves fusing the sample quantity factor, state distance factor, and bias stability factor to obtain the correction confidence, includes: S441: Decompose the deviation stability factor into a first observation deviation stability factor corresponding to the first observation component and a second observation deviation stability factor corresponding to the second observation component, and calculate the reference consistency factor; the reference consistency factor is determined based on at least one of the following: the source type of the external reference data, the spatiotemporal resolution, and the uncertainty index; S442: The sample quantity factor, state distance factor, first observation bias stability factor, second observation bias stability factor and reference consistency factor are fused to obtain the correction confidence.

10. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to claim 8, characterized in that, In step S45, when the correction confidence level is higher than the first threshold, the corrected target physical quantity data is output and accompanied by a high confidence indicator. When the correction confidence level is lower than or equal to the first threshold, the original target physical quantity observation data is output along with a low confidence flag, including: S451: Determine whether the correction confidence level is higher than the first threshold; S452: When the correction confidence level is higher than the first threshold, output the corrected target physical quantity data and attach a high confidence mark; S453: When the correction confidence is lower than or equal to the first threshold, it is determined whether the correction confidence is higher than the second threshold, wherein the second threshold is lower than the first threshold; S454: When the correction confidence level is higher than the second threshold, output the corrected target physical quantity data and attach a reference correction identifier; S455: When the correction confidence is lower than or equal to the second threshold, output the original target physical quantity observation data with a low confidence flag.

11. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to any one of claims 1-10, characterized in that, The attitude features extracted from the attitude statistics in step S2 include the comprehensive tilt angle, which is calculated by taking the square root of the sum of the squares of the pitch angle statistics and the roll angle statistics. The state vector includes the integrated tilt angle, azimuth angle, significant wave height, wave spectrum peak period, main wave direction, observed values ​​of target physical quantities, and external reference data values.

12. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to any one of claims 1-10, characterized in that, The pre-built historical sample library mentioned in step S2 has at least one of the following sources: historical observation data of the current sensor platform, historical observation data of other platforms of the same model as the current sensor platform under similar environmental conditions, and stable observation period data confirmed by quality control.

13. The method for correcting sensor data errors on a non-stationary platform based on attitude perception and environmental background constraints according to any one of claims 1-10, characterized in that, The external reference data mentioned in step S1 includes one or more of the following: reanalysis wind field data, regional numerical model wind field data, observation data from nearby fixed stations, observation data from offshore wind towers, and quality-controlled reference data.

14. A sensor data error correction system for a non-stationary platform based on attitude perception and environmental background constraints, characterized in that, Error correction of non-stationary platform sensor data is performed based on the non-stationary platform sensor data error correction method based on attitude perception and environmental background constraints as described in any one of claims 1-13, including: The data acquisition module acquires target physical quantity observation data collected by the non-stationary platform sensor within a preset time window, attitude statistics of the sensor mounting platform, environmental state data of the environment where the sensor is located, and external reference data independent of the non-stationary platform sensor; the external reference data are reliable reference measurements of the same type as the target physical quantity. The vector calculation module constructs the state vector for the current time window and calculates the similarity between the current state vector and each historical state vector from a pre-built historical sample library based on a preset similarity metric. Historical samples that meet preset similarity conditions are selected to form a similar sample set. The state vector includes at least attitude features extracted from the attitude statistics data, environmental features extracted from the environmental state data, observation features extracted from the target physical quantity observation data, and environmental background constraint features extracted from the external reference data. Each historical sample in the historical sample library contains data of the same type as the current time window and a corresponding state vector. The error calculation module calculates the error correction parameters for the current time window based on the difference between the target physical quantity observation data corresponding to each historical sample in the similar sample set and its external reference data. The data correction module uses the error correction parameters to correct the target physical quantity observation data in the current time window, and obtains the corrected target physical quantity data.

Images