Similarity fusion retrieval method, device and equipment for sounding profile and medium

Abstract

The application discloses a similarity fusion retrieval method, device and equipment for sounding profile, and a medium, and comprises the following steps: acquiring historical sounding profile data, and constructing an offline feature library based on the historical sounding profile data; collecting query sounding data and standardizing the query sounding data to obtain standardized query sounding data; performing multi-component feature extraction on the standardized query sounding data to obtain second multi-component features; comparing each component feature of the second multi-component features with the same type of component features of the first multi-component features of each historical sounding profile, and calculating the component similarity corresponding to each component feature; and fusing the component similarities to obtain the comprehensive similarity corresponding to each historical sounding profile. By introducing the profile structure component, the technical problem of profile mismatch caused by relying only on part of the numerical components in the existing sounding profile similarity retrieval is solved, and the accuracy of the sounding profile similarity retrieval result is improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological observation data processing technology, and in particular to a similarity fusion retrieval method, device, equipment and medium for sounding profiles. Background Technology

[0002] Radiosonde profiles characterize the vertical distribution features of meteorological elements such as temperature, humidity, and wind field. Similarity retrieval of radiosonde profiles can match historical observation samples with similar stratification features. On one hand, it can be used for extreme weather event analysis, retrieving historical cases with profile morphologies similar to those of severe convection, heavy rain, and tornadoes, assisting in weather identification and forecasting. On the other hand, it can be used for verifying operational meteorological data, facilitating the identification of key stratification features such as temperature inversions, moisture layers, and wind shear. Furthermore, batch similarity profile retrieval can analyze the atmospheric background conditions of extreme events, enabling climate statistics and causal analysis.

[0003] Meteorological sounding data are atmospheric vertical profile data obtained through sounding balloon observations. The data includes typical meteorological variables such as temperature (TEMP), dew point (DWPT), wet-bulb temperature (TWC), virtual temperature (TV), wind speed (WSPD), wind direction (WDIR), air pressure (P), and altitude (HGHT). Each variable changes vertically with air pressure or altitude, directly reflecting atmospheric stratification, humidity distribution characteristics, and upper-level wind field structure, among other atmospheric physical properties.

[0004] Radiosonde data possesses significant multidimensional and complex characteristics, specifically: the data exhibits a multivariate, multi-level structure; the distribution of vertical observation layers is uneven, with a large span between effective layers; the observation process is prone to missing data and outliers, requiring data preprocessing through quality control and interpolation; different meteorological variables show significant differences in dimensions and statistical distribution, resulting in strong data heterogeneity; the overall structural characteristics of the profile are prominent, with vertical structures such as inversion layers, wet layer thickness, dry intrusion, and wind shear having a crucial impact on weather evolution. Therefore, conventional data processing and matching methods in general fields cannot directly adapt to the processing needs of radiosonde profile data, making it difficult to effectively balance the numerical characteristics and vertical structural features of radiosonde data. Large-scale profile retrieval generally adopts a segmented approach of offline database construction and online querying. If the preprocessing rules for offline feature construction and online querying are not consistent, the features obtained from historical and real-time profiles are not comparable, introducing persistent matching errors; and reprocessing all historical samples for each retrieval incurs huge computational costs, making it difficult to meet operational timeliness requirements.

[0005] Current traditional similarity retrieval schemes for radiosonde profiles are mostly based on a single component, such as calculating distances by combining profile interpolation with a single distance or a small number of distance combinations. Specifically, the radiosonde profile is interpolated to a unified set of standard layers (e.g., 1000 hPa to 100 hPa, with a step size of 25 hPa), and then Euclidean distance, root mean square error (RMSE), dynamic time warping (DTW), or combinations thereof are calculated for each meteorological variable. Its advantages are simplicity and ease of batch processing; however, its disadvantages are also obvious. This component only reflects the similarity of meteorological values ​​and is difficult to characterize structural differences in the profile (e.g., inversion layers, wet layers, frontal structures, etc.); the algorithm is sensitive to missing observations and outliers, and errors introduced by interpolation can easily amplify distance calculation biases; furthermore, the significant differences in the dimensions of different variables can easily lead to uneven distance contributions, requiring additional normalization processing.

[0006] Another common approach is scalar index space retrieval. This involves calculating multiple meteorological stratification indicators, such as convective available potential energy (CAPE), convective insufficiency energy (CIN), raised condensation height (LCL), free convection height (LFC), K-index, precipitable water (PW), wind shear, and storm relative helicity (SRH), and then performing distance or similarity matching within a scalar feature space. While this method offers clear physical meaning and highly interpretable results, it also has significant drawbacks. Relying on a limited set of indicators for matching fails to cover the full vertical structure information contained in the sounding profile; information correlations and data redundancy exist among various stratification indicators, making it difficult to accurately characterize complex atmospheric stratification morphologies; and local fine-grained profile structures such as the location of the inversion layer and the thickness of the moist layer cannot be effectively characterized by a limited number of scalar indicators.

[0007] Another approach involves structural characterization, frequency domain analysis, or wavelet retrieval. Structural features such as frequency band distribution, multi-scale energy, and key feature points of the profile are extracted using techniques like Fourier transform and wavelet transform. Then, distribution distances such as Wasserstein distance, L1 distance, and L2 distance are used to measure sample similarity. This method can capture multi-scale morphological differences in the profile, but it has the following limitations: the structural feature extraction and distance calculation process is complex, resulting in high engineering implementation costs; when structural features cannot be effectively extracted from the profile, there is a lack of feasible degradation solutions, leading to poor system stability.

[0008] Matching based solely on single-dimensional meteorological values ​​is insufficient to simultaneously consider both the magnitude of element values ​​and the overall morphological structure of the profile, resulting in unreliable matching results. Furthermore, some existing technologies that fuse multiple features lack a complete fusion system adapted to the characteristics of radiosonde data, exhibiting significant technical deficiencies. The profile, scalar, and structural components of radiosonde profiles exhibit strong heterogeneity, corresponding to continuous time-series data, low-dimensional statistical parameters, and frequency domain morphological features, respectively. Their data dimensions, distribution patterns, and physical representations differ greatly, making data alignment and unified preprocessing highly challenging. Simultaneously, the similarity metrics for different components are independent and lack uniformity in dimensions, preventing direct coupling and matching. Existing fusion schemes often employ fixed-weight linear superposition, lacking adaptive adjustment capabilities and failing to adapt to complex conditions such as missing profile data, structural differences, and varying weather scenarios.

[0009] Therefore, a method that balances the fusion of heterogeneous profile information from multiple types with consistency constraints between offline and online processing is proposed, which has significant engineering value for improving the reliability of profile similarity retrieval under extreme weather conditions. Summary of the Invention

[0010] The purpose of this invention is to provide a similarity fusion retrieval method, device, equipment, and medium for sounding profiles. By introducing profile structural components and achieving collaborative fusion of numerical and structural components to output a comprehensive similarity, this invention solves the technical problem of profile mismatch caused by existing sounding profile similarity retrieval relying only on partial numerical components, thereby improving the accuracy of sounding profile similarity retrieval results.

[0011] To achieve the above objectives, the technical solution adopted by this invention is: a similarity fusion retrieval method for sounding profiles, comprising the following steps: S1: Acquire historical radiosonde profile data and construct an offline feature library based on the historical radiosonde profile data. The offline feature library includes at least one first multi-component feature corresponding to a historical radiosonde profile data. S2: Collect and standardize the radiosonde data to obtain standardized radiosonde data; wherein, the radiosonde data is collected by a radiosonde on a radiosonde balloon during its ascent. S3: Perform multi-component feature extraction on the standardized query radiosonde data to obtain the second multi-component feature; wherein, both the first multi-component feature and the second multi-component feature include profile component feature, scalar component feature and structural component feature; S4: Compare each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, and calculate the component similarity corresponding to each component feature. S5: Combine the similarity of each component to obtain the comprehensive similarity of each historical sounding profile.

[0012] This invention performs unified multi-component feature extraction on both queried and historical sounding profiles. It introduces a profile structure component in addition to the numerical component, enabling the fusion and matching of multiple heterogeneous components with significant differences in dimensions and distribution. Compared to traditional methods that only perform single-component analysis or only fuse numerical components, this invention simultaneously considers both profile numerical information and vertical stratigraphic morphology information. This effectively avoids the risk of mismatches between profiles with similar numerical values ​​but significantly different stratigraphic structures, significantly improving the reliability of sounding profile similarity retrieval results.

[0013] Furthermore, building an offline feature library includes standardizing historical sounding profile data; The historical sounding profile data and the queried sounding data are processed using the same standardization steps, which include deduplication, unified interpolation to a standard pressure layer, unit normalization, and data quality control.

[0014] This invention ensures that offline database construction and online queries follow a consistent preprocessing procedure, allowing querying radiosonde data to be directly matched with historical radiosonde profile data. This avoids the incomparability between historical and query features caused by inconsistent interpolation schemes and normalization strategies, which leads to systematic biases in the similarity results calculated when matching online and offline data, resulting in numerous incorrect matches. If preprocessing is re-performed on all historical profiles for each query to eliminate bias, it would incur significant computational overhead, failing to meet the timeliness requirements of real-time online meteorological retrieval.

[0015] Furthermore, in S3, both the first and second multi-component features also include a humidity enhancement component feature. The addition of the humidity enhancement component focuses on key information about water vapor distribution. The four types of component information complement each other to comprehensively mine the meteorological information included in the sounding profile, avoiding mismatch of the profile caused by the lack of information in a single dimension, and improving the comprehensiveness of similar profile matching.

[0016] Furthermore, S4 specifically includes each historical sounding profile: Calculate the weighted RMSE distance on the common valid layer set of historical sounding profile data and queried sounding data. Based on the weighted RMSE distance, the profile component similarity is calculated using an exponential kernel function. : ; Calculate the weighted L1 distance between historical sounding profile data and queried sounding data. Based on the weighted L1 distance, scalar component similarity is calculated using an exponential kernel function. : ; Calculate the difference in wavelet features of the profiles between historical sounding data and queried sounding data. Based on the differences in profile wavelet features, the structural component similarity is calculated using the exponential kernel function. : ; in, Represents the exponential kernel mapping function; v This represents the dimensional index of meteorological elements within the profile. Indicates the first v Weighting coefficients corresponding to each meteorological profile element; Represents the first value calculated on the common effective layer set. v Normalized root mean square error of each profile element; k This represents a scalar meteorological index. Indicates the first k The weighting coefficients corresponding to the itemized scalar indicators; , These represent the normalized versions of historical sounding profile data and queried sounding data, respectively. k The numerical values ​​of the scalar component characteristics.

[0017] The original distances are solved using differentiated distance metrics for profile components, scalar components, and structural components respectively. An exponential kernel function is uniformly adopted to map weighted RMSE distances, weighted L1 distances, and wavelet feature differences—which have completely different dimensions, value ranges, and physical meanings—to a consistent similarity value range. This solves the problem that most existing fusion schemes only apply simple weighting to distances of similar features, failing to be compatible with the profile structural difference metrics obtained from wavelet transforms. Furthermore, the exponential kernel forms a non-linear similarity decay law, which better aligns with the meteorological operational understanding that the greater the profile difference, the faster the similarity decreases, thus improving the physical rationality of the matching results.

[0018] Furthermore, in S4, when comparing each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, if the amount of data in the offline feature library exceeds a preset threshold, the following steps are performed: S4.1: Select the pre-selected component features from the component features, calculate the component similarity between the second multi-component feature and the first multi-component feature corresponding to each historical sounding profile; S4.2: Based on the component similarity corresponding to the selected component features, sort all historical radiosonde profiles in descending order and output historical radiosonde profile samples with matching scores within the first preset ranking range as candidate samples; S4.3: For the candidate sample, calculate the component similarity between the second multi-component feature and the first multi-component feature corresponding to the candidate historical sounding profile, and the component similarity between all component features.

[0019] This invention employs a two-stage matching method of coarse screening and fine sorting. While ensuring that the final matching accuracy is not significantly affected, it avoids performing complete feature comparisons, including structural components, on each of the massive historical profiles. This greatly reduces the overall computational overhead, shortens the online retrieval response time, and better meets the real-time query needs of meteorological services.

[0020] Furthermore, in S5, fusing the similarity of each component specifically includes the following steps: S5.1: Weighted fusion of the profile component similarity and scalar component similarity corresponding to each historical sounding profile to obtain the numerical similarity corresponding to each historical sounding profile; S5.2: Determine whether the structural component similarity corresponding to each historical sounding profile meets the fusion condition; if the fusion condition is met, proceed to S5.3; otherwise, use the numerical similarity result as the comprehensive similarity result corresponding to each historical sounding profile that does not meet the fusion condition. S5.3: Weighted fusion of the numerical similarity and structural component similarity of each historical sounding profile that meets the fusion conditions, to obtain the comprehensive similarity result of each historical sounding profile that meets the fusion conditions.

[0021] Because structural components differ from numerical components and cannot be completed using simple interpolation alone, this invention employs a hierarchical fusion logic. It first fuses various numerical components, then judges and fuses the structural components, reverting to numerical component fusion when structural components are unavailable. This solves the problem of structural feature extraction failure caused by wavelet decomposition's inability to effectively extract stable multi-scale structural features when there are smooth profiles, few vertical layers, and a large number of missing observations. This improves the model's adaptability to complex observation conditions.

[0022] Furthermore, after S5, all historical radiosonde profiles are sorted in descending order according to comprehensive similarity, and the matching degree of historical radiosonde profile samples within the second preset ranking range and the similarity of each component are output. This enables the display of matching information item by item, clearly showing the causes of profile similarity, enhancing the interpretability of search results, and better meeting the needs of actual meteorological business analysis.

[0023] Based on the same concept, the present invention also provides a similarity fusion retrieval device for sounding profiles, which is used to implement the above method, the device comprising: An offline feature library construction module is used to acquire historical sounding profile data and construct an offline feature library based on the historical sounding profile data. The offline feature library includes at least one first multi-component feature corresponding to a historical sounding profile data. The radiosonde data acquisition module is used to collect radiosonde data and standardize the radiosonde data to obtain standardized radiosonde data; wherein, the radiosonde data is collected by a radiosonde on a radiosonde balloon during its ascent. The multi-component feature extraction module is used to extract multi-component features from the standardized query radiosonde data to obtain a second multi-component feature; wherein, both the first and second multi-component features include profile component features, scalar component features, and structural component features. The component similarity calculation module is used to compare each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, and calculate the component similarity corresponding to each component feature. The comprehensive similarity calculation module is used to fuse the similarity of each component to obtain the comprehensive similarity corresponding to each historical sounding profile.

[0024] Based on the same concept, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0025] Based on the same concept, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0026] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing profile structure components and humidity enhancement components, and combining them with profile and scalar numerical components for fusion matching, this invention solves the shortcomings of traditional methods that rely solely on numerical information for matching and ignore atmospheric vertical structure information, effectively avoiding mismatches of profiles with similar values ​​but large structural differences. Simultaneously, for the introduced structural component features, corresponding distance algorithms are used for different types of components, and a kernel function is introduced to unify the similarity scale, solving the problem that traditional methods only perform simple weighted fusion, resulting in inconsistent dimensions between structural and numerical features, making it difficult to effectively integrate structural information into similarity calculations. Regarding the problem of structural components not being easily completed, this invention does not directly perform weighted fusion of all feature components, but instead designs a method of first fusing numerical features and then structural features, avoiding the inability to output effective matching results due to missing structural features, thus improving the applicability of this method. Attached Figure Description

[0027] To more clearly illustrate the technical method of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of the similarity fusion retrieval method for sounding profiles in an embodiment of the present invention; Figure 2 This is a schematic diagram of the Skew-T Log-P temperature and humidity profile of the sounding profile; Figure 3 This is a diagram showing the component characteristics of the DPD profile of different thermal anchor points and dew point temperature differences. Figure 4 It is a comparison chart of the detrended temperature perturbation profile of the query profile and the candidate profile; Figure 5 It is a comparison chart of the normalized power spectral density distribution of the query profile and the candidate profile; Figure 6 It is a visualization of the scaling coefficients of the continuous wavelet transform of the query profile. Detailed Implementation

[0029] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. For ease of description, the terms "upper," "lower," "left," and "right" used below only indicate that they correspond to the upper, lower, left, and right directions in the accompanying drawings and do not limit the structure.

[0030] The technical methods of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] The technical methods of this application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0032] Example 1

[0033] Figure 1 The flowchart of the similarity fusion retrieval method for sounding profiles provided in the embodiments of the present invention is shown, as follows: Figure 1 As shown, the method includes the following steps: S1: Obtain historical radiosonde profile data and construct an offline feature library based on the historical radiosonde profile data. The offline feature library includes at least one first multi-component feature corresponding to a historical radiosonde profile data. Figure 2 The original radiosonde profile is plotted using the meteorological standard Skew-T logarithmic barometric coordinates, including the vertical distribution of elements such as temperature and dew point. The original radiosonde profile is the basic data source for this method, used for offline feature library construction and online query of radiosonde input, and serves as the data foundation for subsequent preprocessing and feature extraction operations.

[0034] Specifically, the method in this embodiment of the invention supports access to multiple data sources such as local files, databases, and API interfaces, and automatically identifies data formats; it uniformly interpolates the original historical sounding data to a standard pressure layer set (e.g., 1000~100hPa, with preferred step sizes of 10, 25, and 50hPa), simultaneously standardizing physical quantity units, uniformly using degrees Celsius to represent temperature and meters per second to represent wind speed, while automatically removing or interpolating missing layers and abnormal observations; furthermore, it uses the median absolute deviation (MAD) robust normalization method to normalize each meteorological variable, avoiding distance amplification caused by extreme values, and effectively improving the algorithm's adaptability and robustness to abnormal samples and differences in the dimensions of different physical quantities. Normalization is performed according to the following formula: ; in, x These are the original meteorological element observations; z The values ​​represent the normalized meteorological variables; median is the median of the meteorological element sample set; and the constant 1.4826 is the correction coefficient. During the normalization statistical phase, the median and MAD are calculated independently for different meteorological variables and characteristic indicators, and the resulting statistics are persistently stored in metadata.

[0035] In one embodiment, the vertical coordinates used for the profile can be replaced by height coordinates instead of pressure coordinates; the set of profile variables involved in the calculation can be increased or decreased according to actual needs.

[0036] Extracting multi-component features from historical sounding profile data specifically includes extracting profile features, scalar features, and structural features. Specifically, profile features include temperature (TEMP), dew point (DWPT), wet-bulb temperature (TWC), virtual temperature (TV), surface parcel temperature (PARCEL_T), mixed-layer parcel temperature (PARCEL_T_MIX), wind speed (WSPD), and wind direction (WDIR), which are all interpolated to a standard layer. Scalar features include convective available potential energy (CAPE), convection suppression energy (CIN), raised condensation height (LCL), free convection height (LFC), K-index, precipitable water (PW), environmental wind shear, storm relative helicity (SRH), descending convective available potential energy (DCAPE), and wet-bulb zero-degree height (WBZ). After obtaining profile meteorological elements from a unified standard layer, scalar features are calculated in batches based on profile variables and atmospheric thermodynamic physics formulas. Extracting structural component features involves detrending the profile variables, then using Fourier transform or wavelet transform to decompose the signal and extract characteristic information representing the vertical structure of the atmospheric stratification, such as spectral distribution, scale energy, and key feature points. All features are stored in a columnar format (e.g., Parquet) and metadata (number of samples, layer set, variable list, whether it contains structural features, etc.) is generated.

[0037] In one embodiment, the structural features used to construct the structural components can be replaced by at least one of the following: frequency band energy features, key layer detection features, and profile morphology coding features.

[0038] In one embodiment, extracting multi-component features from historical sounding profile data further includes extracting humidity-enhanced component features, such as T-Td profiles, to enhance the ability to distinguish water vapor environments. Unlike existing technologies that use a single distance or single exponential space for similarity retrieval, this embodiment employs a multi-component orthogonal decomposition strategy, explicitly decomposing sounding profile features into four complementary feature channels: profile, scalar, structure, and humidity. Each channel supports independent configuration and can perform similarity calculations independently.

[0039] S2: Collect and standardize the radiosonde data to obtain standardized radiosonde data; wherein, the radiosonde data is collected by a radiosonde on a radiosonde balloon during its ascent. It includes the pressure layer and core variables (TEMP, DWPT, WSPD, WDIR, etc.), and allows for missing measurements.

[0040] The standardized radiosonde data query adopts the exact same standardization steps as the standardized historical radiosonde profile data. Unlike existing technologies, this invention employs an online-offline isomorphic constraint mechanism, forcing the query samples and historical samples to undergo completely identical preprocessing and feature extraction processes. Parameters such as the standard layer set, normalized statistics, and structural feature activation status are solidified into metadata. This allows the statistics obtained in the offline stage to be directly reused in online calculations. In one embodiment, isomorphic preprocessing uses linear interpolation, spline interpolation, or piecewise interpolation resampling methods to achieve vertical layer alignment of the radiosonde profile.

[0041] Inconsistent variable dimensions, hierarchies, and missing data handling methods can lead to difficulties in extending new features or adapting to diverse scenario requirements. By leveraging isomorphic preprocessing (unified standard layer grid, unit standardization, and missing data handling rules) combined with robust median / MAD normalization, the dimensions, comparison hierarchies, and missing data handling logic of various meteorological variables are unified. A multi-component orthogonal decomposition architecture is adopted, allowing for independent addition and deletion of components such as physical indicators and structural features without requiring reconstruction of the main computational process, facilitating adaptation to different forecasting operational scenarios.

[0042] S3: Multi-component feature extraction is performed on the standardized radiosonde data to obtain the second multi-component feature; wherein, both the first and second multi-component features include profile component features, scalar component features, and structural component features. The query radiosonde and historical samples use the same preprocessing and feature extraction process to ensure isomorphism. Batch queries and single interactive queries are supported. When extracting the spectral features of the structural components, the perturbation profile is first detrended under logarithmic pressure coordinates before the transformation is performed, and the power spectrum (PSD) is normalized to a unit mass distribution; if this normalization operation is not performed, the total spectral energy of different samples will differ, making it impossible to directly use Wasserstein distance for comparison calculations.

[0043] S4: Compare each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature for each historical sounding profile, and calculate the component similarity corresponding to each component feature. For each historical sounding profile: For the profile component, the weighted RMSE distance is calculated on the common valid layer set of historical sounding profile data and queried sounding data. Based on the weighted RMSE distance, the profile component similarity is calculated using an exponential kernel function. : ; in, This refers to the exponential kernel mapping function, and the same applies below; v This represents the dimensional index of meteorological elements within the profile. Indicates the firstv Weighting coefficients corresponding to each meteorological profile element; Represents the first value calculated on the common effective layer set. v Normalized root mean square error of each profile element.

[0044] This invention allows setting a common effective layer threshold, for example, requiring that the proportion of overlapping effective pressure layers between the queried profile and the historical profile is not less than 30%. It also supports custom configuration of the missing data penalty coefficient; if the proportion of common effective layers is lower than the threshold, the weight of that sample is reduced or it is removed. If a variable is missing, the profile distance is calculated only on the remaining variables, and the weights are re-normalized.

[0045] Figure 3 The retrieval capability of profile components is demonstrated. Taking dew point temperature difference (DPD) as an example, four typical profile centers are obtained by clustering based on historical radiosonde data, representing four different profile morphologies. n represents the number of historical data corresponding to different profile types. If new query radiosonde data is input, the distance difference between the query radiosonde data and the full historical DPD profile data in the feature library will be calculated. It should be noted that the above clustering division is only used for visualization and is not a necessary step in the retrieval process.

[0046] For the scalar component, calculate the weighted L1 distance between historical sounding profile data and queried sounding data. Based on the weighted L1 distance, scalar component similarity is calculated using an exponential kernel function. : ; in, k This represents a scalar meteorological index. Indicates the first k The weighting coefficients corresponding to the itemized scalar indicators; , These represent the normalized versions of historical sounding profile data and queried sounding data, respectively. k The numerical values ​​of the scalar component characteristics.

[0047] For the structural component, calculate the difference in profile wavelet features between historical sounding profile data and queried sounding data. Based on the differences in profile wavelet features, the structural component similarity is calculated using the exponential kernel function. : ; In one embodiment, the feature differences of the profile wavelet are obtained by selecting the corresponding distance functions to calculate the feature differences for different types of structural features obtained from wavelet decomposition and then jointly obtaining the differences. Specifically, Wasserstein distance metric is used for spectral distribution features, and L2 distance metric is used for wavelet energy features.

[0048] Figure 4 It involves comparing the detrended temperature perturbation profiles of the query profile and candidate profiles. By performing detrended processing on the sounding temperature profile under logarithmic barometric coordinates, a temperature perturbation sequence is obtained. This perturbation sequence is the basis for extracting profile structural features and solving for structural component similarity. Simultaneously, the similarity of the fine perturbation morphologies of the two profiles can be visually compared graphically. Based on... Figure 4 The detrended temperature perturbation profile shown is analyzed using the Welch method to obtain the power spectral density, which is then normalized to a unit mass distribution to obtain... Figure 5 The normalized power spectral density distribution curve. Figure 6 Therefore Figure 4 The detrended temperature perturbation profile shown is used as the input signal, and a two-dimensional wavelet coefficient matrix is ​​obtained by performing a continuous wavelet transform. Based on Figure 5 and Figure 6 The distances between these two types of features are calculated separately, and together they constitute the structural component similarity.

[0049] In one embodiment, humidity enhancement component similarity Using the same calculation method as the profile components and reusing the same distance metric algorithm, similarity discrimination of water vapor vertical profile features is achieved.

[0050] In one embodiment, the distance metric used for profile component similarity and structural component similarity can be at least one of DTW distance, layer-weighted distance, kernel metric, and learning metric.

[0051] The distance units and numerical ranges output by different channels differ significantly, making direct weighting impossible. This invention first uses a kernel function to convert distance to similarity, unifying the value range, and then configures weight fusion to give the fused weights comparable physical meaning, achieving effective combination of multiple complementary feature information. Furthermore, missing structural features can lead to system instability, and structural features cannot be completed through simple interpolation. This invention improves the compatibility of its method by smoothing structural weight values, automatically reverting to pure numerical channels when missing features are found, and padding with zeros through outer joins.

[0052] The method of this invention employs the same standardized processing approach for historical data and queries, along with offline database construction and online querying. It simultaneously considers the similarity of profile numerical morphology and perturbation structure, effectively improving the accuracy and stability of radiosonde similarity retrieval results. Offline feature pre-calculation, while computationally intensive, is performed only once, ensuring no duplication in new queries. This reduces the computational overhead of the online retrieval stage and adapts to interactive forecasting application scenarios.

[0053] In one embodiment, when comparing each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, if the amount of data in the offline feature library exceeds a preset threshold, the following steps are performed: S4.1: Select the pre-selected component features from the component features, calculate the component similarity between the second multi-component feature and the first multi-component feature corresponding to each historical sounding profile; S4.2: Based on the component similarity corresponding to the selected component features, sort all historical radiosonde profiles in descending order and output historical radiosonde profile samples with matching scores within the first preset ranking range Top-K as candidate samples; for example, the value of Top-K is one of 50 to 20000, and in one embodiment it is 200.

[0054] S4.3: For the candidate sample, calculate the component similarity between the second multi-component feature and the first multi-component feature corresponding to the candidate historical sounding profile, and the component similarity between all component features.

[0055] In large-database scenarios, low-cost features (e.g., a small number of scalar metrics or low-dimensional signature vectors) are first selected, and a rough similarity is calculated. Then, a Top-K candidate is selected, and the full component similarity (profile / scalar / structure / optional humidity) is calculated for the Top-K candidates for fine ranking. After fusion, the Top-N is output, which can significantly reduce the amount of online computation. The selection of Top-K can be adjusted according to the database size and response time requirements. For example, when Top-N is 20, Top-K can be 200 to 500.

[0056] S5: Merge the similarity of each component to obtain the comprehensive similarity for each historical sounding profile. The fusion of the similarity of each component specifically includes the following steps: S5.1: The numerical similarity of each historical sounding profile is obtained by weighted fusion of the profile component similarity and the scalar component similarity according to the following formula: ; S5.2: Determine whether the structural component similarity corresponding to each historical sounding profile meets the fusion condition; if the fusion condition is met, proceed to S5.3; otherwise, use the numerical similarity result as the comprehensive similarity result corresponding to each historical sounding profile that does not meet the fusion condition. S5.3: Based on the following formula, the numerical similarity and structural component similarity of each historical sounding profile that meets the fusion conditions are weighted and fused to obtain the comprehensive similarity of each historical sounding profile that meets the fusion conditions. result: .

[0057] in, For profile weights, Scalar weights; satisfying ; For structural component weights, satisfying The spectral and wavelet features support custom parameter configuration, allowing for flexible setting of the number of wavelet key scales (e.g., values ​​from 3 to 10), the number of key points extracted from the perturbation curve (e.g., values ​​from 3 to 8), and the sensitivity parameters of the kernel function used for similarity mapping.

[0058] By employing a step-by-step weighted fusion mechanism, it can automatically revert to numerical channel fusion (profile and scalar) when structural component features are missing or some samples are unavailable, and output a prompt. The similarity of each component is calculated based on configurable weights, supporting adjustments to the structural component weights. The multi-component weight vector can be flexibly customized.

[0059] In one embodiment, the similarity fusion method may employ at least one of gated fusion, nonlinear fusion, and learning-based fusion.

[0060] In one embodiment, an outer join zero-padding strategy is adopted: when data for any feature channel is missing and the corresponding similarity cannot be solved, the similarity of that channel is assigned a value of 0, and a weighted fusion operation is performed based on the remaining valid channels to ensure that the comprehensive similarity result can still be output normally even in scenarios where some channels are missing.

[0061] Existing technologies do not consider the fusion of structural features and data features. In this embodiment of the invention, while adding fused structural features, a rollback mechanism is set up. When the feature library has not pre-calculated or some structural components are missing, it automatically rolls back to the fusion of profile and scalar numerical channels and outputs a rollback prompt. With the outer connection zero-padding fusion strategy, it is ensured that the results can still be output normally when any channel is missing.

[0062] In one embodiment, in S5.2, it is further determined whether the calculated similarity of the humidity enhancement component corresponding to each historical sounding profile meets the fusion condition; if the fusion condition is not met, proceed to S5.3; otherwise, the numerical similarity and structural component similarity corresponding to each historical sounding profile that meets the condition are weighted and fused according to the following formula to obtain the comprehensive similarity corresponding to each historical sounding profile that meets the fusion condition. result: ; in, For structural component weights, To enhance the weighting of the humidity component, satisfying .

[0063] In a preferred embodiment, the activation of the structural component fusion process can be controlled by preset system configuration parameters. When the configuration parameters indicate that structural component fusion is enabled, the retrieval process automatically verifies whether structural features are persistently stored in the offline feature library. If the feature library lacks structural components, it is determined that the structural fusion conditions are not met, and the numerical similarity is directly used as the comprehensive similarity, thus automatically downgrading the calculation strategy. If the configuration parameters do not enable structural component fusion, the calculation steps related to structural components are skipped, and the retrieval results are output solely based on numerical similarity. The candidate set is merged through outer joins to ensure that no candidate is lost when any channel is missing.

[0064] Following S5, all historical sounding profiles are sorted in descending order according to their overall similarity. The matching scores of historical sounding profile samples within the second preset ranking range (Top-N) and the similarity scores of each component are output. The value of Top-N ranges from 1 to 200, and in one embodiment, it is 20. The final output is the Top-N search results, which include metadata such as sample identifiers (e.g., station / time / file_path), observation time, station information, and file path, as well as the overall similarity score and the similarity scores at each component level (e.g., profile component similarity). Scalar component similarity Structural component similarity (etc.); at the same time, it outputs visualization results such as component score bar charts, Skew-T profile comparison charts, and structural feature comparison charts.

[0065] In existing technologies, search results only output a comprehensive distance or similarity, lacking component-level scores and explanatory paths, making it difficult for business personnel to verify the search reasons. However, the method in this invention outputs component-level scores (profile, scalar, structure, humidity) along with the comprehensive similarity, and can include explanatory paths such as profile comparison diagrams and component score bar charts, enabling business personnel to clearly verify the search reasons channel by channel. It can also provide high-quality similar samples for AI models, improving training and inference performance.

[0066] The following details the online and offline phases of the process: The offline phase completes the feature library construction, supporting one-time construction or batch incremental updates. Specifically, all historical sounding samples are read, and isomorphic preprocessing (sample deduplication, interpolation to the standard pressure layer, unit standardization, and quality control) is performed. Multiple features are extracted, including profile components, scalar components, structural components, and optional humidity enhancement components. The median and MAD statistic corresponding to each meteorological variable and feature index are calculated and persistently saved. Finally, the data is written to the Parquet feature library and the metadata.json file. This step involves significant computation, but it is only performed during library construction or batch updates; new query tasks do not require repeating these calculations.

[0067] The online phase requires only a single query, meeting the need for interactive and rapid response. Specifically, it receives a query sounding profile Q; it extracts features from the query sounding profile Q using the same process as the offline phase, processing only a single sample, resulting in low computational overhead; it reads the columnar feature library built offline and uses vectorized operations to batch calculate the distances between the query profile and each component of all historical samples in the library; it calls the median and MAD values ​​pre-stored offline for normalization; after exponential kernel mapping and multi-component similarity weighted fusion, it sorts the results and outputs the Top-N search results and the similarity scores of each component.

[0068] Existing technologies require full-database interpolation, index calculation, and distance calculation, resulting in significant I / O and CPU / GPU overhead, making it difficult to meet the needs of interactive Top-N retrieval. The method of this invention employs offline feature library pre-computation combined with online vectorized comparison, transferring full-database interpolation, index calculation, and normalization statistics to be completed offline in one go; online, features are extracted only once for the query sample, eliminating the need for repeated profile interpolation and feature calculation; normalized statistics are solved offline and stored in metadata, eliminating the need for re-statistical analysis across the entire database online; and matrix and vectorized operations replace sequential comparisons, thereby accelerating online response speed.

[0069] The method of this invention can be deployed on computer hardware. The feature library uses Parquet for efficient columnar storage, enabling fast batch reading and vectorized comparison. All modules provide standardized API interfaces for easy access to web services, desktop applications, and batch scripts. Component weights and structural fusion weights can be flexibly adjusted via configuration files, command lines, and front-end interfaces. It supports parameters such as Top-N; it has strong scalability, and when adding new feature components, there is no need to reconstruct the main calculation process; it also supports multi-threaded and multi-process batch construction of feature libraries, and online retrieval accelerates the calculation speed and improves the overall running efficiency by leveraging vectorization and matrix operations.

[0070] The corresponding feature library is internally divided into multiple data partitions, and the field design can directly correspond to columnar storage or vectorized matrices, which facilitates online batch processing and multi-threaded computation. The specific field structure is as follows: catalog: Sample metadata (site, time, file path, quality control tags, etc.); profiles: Flattened columns of profile variables on the standard layer set (e.g., TEMP_1000hPa...TEMP_100hPa). scalars: scalar index series (such as MUCAPE, SFC_CAPE, PW, SRH_0_3km, etc.); structural: structural feature column (PSD, wavelet energy SAV, keypoints flattened field); metadata.json: Stores configuration information such as the standard layer set, variable list, whether structural features are enabled, and normalized statistics.

[0071] In API or service-oriented deployment scenarios, the output can be in JSON format. Example fields include: id, similarity, sim_profile, sim_scalar, sim_struct, sim_moist, station, timestamp, file_path, etc.

[0072] For typical weather events such as temperature inversion, dry intrusion, and wind shear, the method of this invention can effectively retrieve radiosonde samples with similar profile structure features; it significantly reduces the computational load of online retrieval based on the offline pre-computation architecture, and its response speed is adapted to the interactive Top-N retrieval business requirements; at the same time, it outputs the similarity scores of each component, improving the interpretability of the retrieval results and the practical value of meteorological business.

[0073] Example 2 Based on the same concept, embodiments of the present invention also provide a similarity fusion retrieval device for sounding profiles, the device comprising: An offline feature library construction module is used to acquire historical sounding profile data and construct an offline feature library based on the historical sounding profile data. The offline feature library includes at least one first multi-component feature corresponding to a historical sounding profile data. The radiosonde data acquisition module is used to collect radiosonde data and standardize the radiosonde data to obtain standardized radiosonde data; wherein, the radiosonde data is collected by a radiosonde on a radiosonde balloon during its ascent. The multi-component feature extraction module is used to extract multi-component features from the standardized query radiosonde data to obtain a second multi-component feature; wherein, both the first and second multi-component features include profile component features, scalar component features, and structural component features. The component similarity calculation module is used to compare each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, and calculate the component similarity corresponding to each component feature. The comprehensive similarity calculation module is used to fuse the similarity of each component to obtain the comprehensive similarity corresponding to each historical sounding profile.

[0074] Example 3 Based on the same concept, embodiments of the present invention also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in Embodiment 1 above.

[0075] Example 4 Based on the same concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in Embodiment 1 above.

[0076] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0077] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0078] The above embodiments should be understood as being used only to illustrate the present invention more clearly, and not to limit the scope of the present invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art fall within the scope defined by the appended claims.

Claims

1. A similarity fusion retrieval method for sounding profiles, characterized in that, Includes the following steps: S1: Acquire historical radiosonde profile data and construct an offline feature library based on the historical radiosonde profile data. The offline feature library includes at least one first multi-component feature corresponding to a historical radiosonde profile data. S2: Collect and standardize the radiosonde data to obtain standardized radiosonde data; wherein, the radiosonde data is collected by a radiosonde on a radiosonde balloon during its ascent. S3: Perform multi-component feature extraction on the standardized query radiosonde data to obtain the second multi-component feature; wherein, both the first multi-component feature and the second multi-component feature include profile component feature, scalar component feature and structural component feature; S4: Compare each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, and calculate the component similarity corresponding to each component feature. S5: Combine the similarity of each component to obtain the comprehensive similarity of each historical sounding profile.

2. The similarity fusion retrieval method for sounding profiles according to claim 1, characterized in that, Building an offline feature library includes standardizing historical sounding profile data; The historical sounding profile data and the queried sounding data are processed using the same standardization steps, which include deduplication, unified interpolation to a standard pressure layer, unit normalization, and data quality control.

3. The similarity fusion retrieval method for sounding profiles according to claim 1, characterized in that, In S3, both the first multi-component feature and the second multi-component feature further include a humidity enhancement component feature.

4. The similarity fusion retrieval method for sounding profiles according to claim 1, characterized in that, S4 specifically includes each historical sounding profile: Calculate the weighted RMSE distance on the common valid layer set of historical sounding profile data and queried sounding data. Based on the weighted RMSE distance, the profile component similarity is calculated using an exponential kernel function. : ； Calculate the weighted L1 distance between historical sounding profile data and queried sounding data. Based on the weighted L1 distance, scalar component similarity is calculated using an exponential kernel function. : ； Calculate the difference in wavelet features of the profiles between historical sounding data and queried sounding data. Based on the differences in profile wavelet features, the structural component similarity is calculated using the exponential kernel function. : ； in, Represents the exponential kernel mapping function; v This represents the dimensional index of meteorological elements within the profile. Indicates the first v Weighting coefficients corresponding to each meteorological profile element; Represents the first value calculated on the common effective layer set. v Normalized root mean square error of each profile element; k This represents a scalar meteorological index. Indicates the first k The weighting coefficients corresponding to the itemized scalar indicators; , These represent the normalized versions of historical sounding profile data and queried sounding data, respectively. k The numerical values ​​of the scalar component characteristics.

5. The similarity fusion retrieval method for sounding profiles according to claim 1, characterized in that, In S4, when comparing each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, if the amount of data in the offline feature library exceeds a preset threshold, the following steps are performed: S4.1: Select the pre-selected component features from the component features, calculate the component similarity between the second multi-component feature and the first multi-component feature corresponding to each historical sounding profile; S4.2: Based on the component similarity corresponding to the selected component features, sort all historical radiosonde profiles in descending order and output historical radiosonde profile samples with matching scores within the first preset ranking range as candidate samples; S4.3: For the candidate sample, calculate the component similarity between the second multi-component feature and the first multi-component feature corresponding to the candidate historical sounding profile, and the component similarity between all component features.

6. The similarity fusion retrieval method for sounding profiles according to claim 1, characterized in that, In S5, fusing the similarity of each component specifically includes the following steps: S5.1: Weighted fusion of the profile component similarity and scalar component similarity corresponding to each historical sounding profile to obtain the numerical similarity corresponding to each historical sounding profile; S5.2: Determine whether the structural component similarity corresponding to each historical sounding profile meets the fusion condition; if the fusion condition is met, proceed to S5.3; otherwise, use the numerical similarity result as the comprehensive similarity result corresponding to each historical sounding profile that does not meet the fusion condition. S5.3: Weighted fusion of the numerical similarity and structural component similarity of each historical sounding profile that meets the fusion conditions, to obtain the comprehensive similarity result of each historical sounding profile that meets the fusion conditions.

7. The similarity fusion retrieval method for sounding profiles according to claim 1, characterized in that, After S5, all historical sounding profiles are sorted in descending order according to their comprehensive similarity, and the matching scores of historical sounding profile samples within the second preset ranking range and the similarity scores of each component are output.

8. A similarity fusion retrieval device for sounding profiles, used to implement the method according to any one of claims 1 to 7, characterized in that, The device includes: An offline feature library construction module is used to acquire historical sounding profile data and construct an offline feature library based on the historical sounding profile data. The offline feature library includes at least one first multi-component feature corresponding to a historical sounding profile data. The radiosonde data acquisition module is used to collect radiosonde data and standardize the radiosonde data to obtain standardized radiosonde data; wherein, the radiosonde data is collected by a radiosonde on a radiosonde balloon during its ascent. The multi-component feature extraction module is used to extract multi-component features from the standardized query radiosonde data to obtain a second multi-component feature; wherein, both the first and second multi-component features include profile component features, scalar component features, and structural component features. The component similarity calculation module is used to compare each component feature of the second multi-component feature with the same type of component feature of the first multi-component feature of each historical sounding profile, and calculate the component similarity corresponding to each component feature. The comprehensive similarity calculation module is used to fuse the similarity of each component to obtain the comprehensive similarity corresponding to each historical sounding profile.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Images