Offshore ocean observation data quality control method based on time-frequency domain characteristics

Abstract

The invention relates to the technical field of ocean observation data, in particular to an offshore ocean observation data quality control method based on time-frequency domain characteristics, which comprises the following steps: extracting the characteristics of first ocean data, including time sequence characteristics, frequency domain characteristics, inner time sequence characteristics and outer time sequence characteristics; fusing the internal time sequence features and the external time sequence features to obtain internal and external time sequence features; fusing the internal and external time sequence features and the frequency domain features to obtain time-frequency domain features; inputting a preset anomaly detection model to obtain abnormal first ocean data; and correcting the abnormal first ocean data to obtain the ocean data after quality control. According to the method, the time sequence features, the frequency domain features and the fusion of the internal and external time sequence features are introduced, the multi-dimensional features of offshore ocean observation data are comprehensively described, and the accuracy of anomaly detection is remarkably improved. A preset anomaly detection model is adopted to identify the time-frequency domain features, and an abnormal data correction mechanism is combined, so that the reliability and stability of observation data are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of marine observation data technology, and specifically to a method for quality control of offshore marine observation data based on time-frequency domain characteristics. Background Technology

[0002] With the rapid development of marine scientific research and marine engineering construction, offshore marine observation systems are widely used for real-time monitoring of various environmental and hydrological parameters, such as wind speed, air pressure, air temperature, humidity, wave height, and current velocity. This observational data is of great significance for marine environmental evolution analysis, disaster early warning, disaster prevention and mitigation, and navigation safety. However, in actual data acquisition, limitations in sensor accuracy, sampling conditions, complex marine environmental interference, and noise during communication transmission often result in missing values, outliers, and noise interference in the raw data. Failure to address these issues effectively and promptly will directly affect the reliability and accuracy of subsequent data analysis and applications.

[0003] Existing quality control methods mainly focus on anomaly detection based on threshold judgment and statistical regularity. For example, they identify and remove abnormal data by setting reasonable physical value ranges or using statistical features such as sliding window mean and variance. These methods are simple to implement, but they often fall short when dealing with nonlinear anomalies in complex environments, easily leading to misjudgments or omissions. They also fail to fully characterize the complex feature structure of marine observation data, thus limiting the accuracy and robustness of quality control. Summary of the Invention

[0004] (a) Purpose of the invention The purpose of this invention is to provide a quality control method for offshore marine observation data based on time-frequency domain features. By introducing the fusion of time-series features, frequency-domain features, and internal and external time-series features, the method comprehensively characterizes the multidimensional properties of offshore marine observation data, significantly improving the accuracy of anomaly detection. A pre-defined anomaly detection model is used to identify time-frequency domain features, and combined with an anomaly data correction mechanism, the reliability and stability of the observation data are effectively improved, providing high-quality data support for subsequent marine environmental analysis and applications.

[0005] (II) Technical Solution To address the above problems, this invention provides a method for quality control of offshore marine observation data based on time-frequency domain characteristics, comprising: The collected offshore marine observation data is preprocessed to obtain the first marine data; Features are extracted from the first ocean data, including time-series features, frequency-domain features, internal time-series features, and external time-series features, which are extracted based on the time-series features. The internal and external time series features are fused to obtain internal and external time series features; The internal and external time-series features and frequency domain features are fused to obtain time-frequency domain features; The time-frequency domain features are input into a preset anomaly detection model to obtain the first ocean data of the anomaly; The first ocean data with the aforementioned anomalies is corrected to obtain ocean data after quality control.

[0006] In another aspect of the present invention, preferably, the preprocessing of the collected offshore marine observation data to obtain the first marine data includes: Using a sliding window approach, multivariate ocean observation data are processed into a data sequence. Data from different variables form a multidimensional input with the same time period. When there are no valid values, they are represented as empty data to obtain the first data sequence. The empty data in the first data sequence is filled with the average value of ocean observation data from three consecutive time periods to obtain the second data sequence; The second data sequence is normalized using the Z-score normalization method to obtain the first ocean data.

[0007] In another aspect of the present invention, preferably, extracting the temporal features includes: The first ocean data is input into a preset time-series feature extraction module, and the output result is the time-series feature; The preset temporal feature extraction module includes a position encoding layer, a multi-head attention mechanism layer, a feedforward network layer, and a normalization layer; The location coding layer introduces temporal sequence information into the input data to identify the temporal positional relationship between different data points; The multi-head attention mechanism layer is used to calculate the correlation between data after the data has been processed by the position encoding layer by query vector, key vector and value vector; The forward feedback network layer performs non-linear mapping and dimension transformation based on the correlation between the data; The normalization layer is used to standardize the temporal features of the data and output after processing by the feedforward network layer.

[0008] In another aspect of the present invention, preferably, extracting frequency domain features includes: The time-series features are input into a preset frequency domain feature extraction module, and the output result is the frequency domain features; The preset frequency domain feature extraction module includes a fast Fourier transform layer, a first Transformer layer, and an inverse fast Fourier transform layer. Based on the Fast Fourier Transform layer, the time-series features are converted into frequency information; Based on the first Transformer layer, the periodic features of the frequency information are extracted; Based on the inverse fast Fourier transform layer, the frequency information with extracted periodic features is transformed to obtain frequency domain features, which include query features, key features and value features.

[0009] In another aspect of the present invention, preferably, extracting external temporal features includes: The time series features are input into a preset external time series feature extraction module, and the output result is the external time series features. The preset external temporal feature extraction module includes a second Transformer layer; Based on the second Transformer layer, long-term features of the temporal features are extracted, and the long-term features are the external temporal features.

[0010] In another aspect of the present invention, preferably, extracting intra-temporal features includes: The external temporal features are input into a preset internal temporal feature extraction module, and the output result is the internal temporal features; The preset internal temporal feature extraction module includes a third Transformer layer, and the window length of the third Transformer layer is smaller than the window length of the second Transformer layer. Based on the third Transformer layer, short-time features of the temporal features are extracted, and the short-time features are the internal temporal features.

[0011] In another aspect of the present invention, preferably, the internal and external time-series features are fused to obtain internal and external time-series features, including: The attention mechanism is used to calculate the weights of external and internal temporal features; Based on the weights of the external and internal time series features, the external and internal time series features are weighted and calculated to obtain the internal and external time series features, which include query features, key features, and value features.

[0012] In another aspect of the present invention, preferably, the internal and external time-series features and frequency domain features are fused to obtain time-frequency domain features, including: Based on the query features, key features, and value features of the internal and external time series features and the query features, key features, and value features of the frequency domain features, the interaction relationship between the internal and external time series features and the frequency domain features is obtained. The interaction relationships are concatenated based on a multi-head attention feature concatenation function to obtain the concatenated features; Based on the multi-head attention computation function, the concatenated features are transformed to obtain time-frequency domain features.

[0013] In another aspect of the present invention, preferably, the time-frequency domain features are input into a preset anomaly detection model to obtain first ocean data on the anomaly, including: The time-frequency domain features are input into a preset anomaly detection model to obtain prediction results, which include the category probabilities of each variable. The sigmoid function is used to calculate the anomaly score for each variable under the anomaly category; When the anomaly score exceeds the preset anomaly detection threshold range, the corresponding data is determined to be the first ocean data that is abnormal.

[0014] In another aspect of the present invention, preferably, the correction of the anomalous first ocean data to obtain quality-controlled ocean data includes: The first ocean data for the aforementioned anomaly is replaced with the average value of ocean observation data from three consecutive time periods.

[0015] (III) Beneficial Effects The above-described technical solution of the present invention has the following beneficial technical effects: This invention extracts temporal, frequency, and internal / external temporal features from observational data, performing multi-dimensional and multi-level feature fusion to more comprehensively reflect the inherent patterns of data in complex marine environments. Compared to traditional methods based on threshold judgment or single feature analysis, it can effectively identify data anomalies under nonlinear and non-stationary conditions, significantly reducing false positives and false negatives. By inputting the fused time-frequency domain features into a pre-defined anomaly detection model, intelligent identification of observational data can be achieved. Combined with subsequent anomaly correction steps, this not only ensures the integrity and continuity of the data but also significantly improves its accuracy and reliability, exhibiting strong robustness and adaptability. It can be widely applied to the real-time processing of offshore observation data, providing high-quality data support for marine environmental evolution research, disaster monitoring and early warning, and marine engineering safety. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of one embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0020] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0021] Example 1 A quality control method for offshore ocean observation data based on time-frequency domain characteristics. Figure 1 An overall flowchart of one embodiment of the present invention is shown, as follows: Figure 1 As shown, it includes: Preprocessing the collected offshore marine observation data to obtain first marine data; the preprocessing may include data cleaning, preliminary outlier removal, missing value imputation, data format standardization, and sampling frequency unification to ensure the integrity and consistency of the input data. Through preprocessing steps, the initial bias in the original observation data caused by sensor noise, communication errors, or extreme weather conditions can be effectively reduced, providing a reliable data foundation for subsequent feature extraction. In this embodiment, the preprocessing of the collected offshore marine observation data to obtain the first marine data includes: Using a sliding window approach, multivariate ocean observation data is processed into a data sequence. Data from different variables form a multidimensional input within the same time period. When no valid value exists, it is represented as empty data, thus obtaining the first data sequence. The multivariate ocean observation data includes multiple observations such as wind speed, wind direction, air temperature, air pressure, humidity, wave height, current velocity, and salinity. By setting the sliding window length, the original data is segmented, allowing the different variable data corresponding to each time window to be combined into a multidimensional input vector within the same time period. For missing or invalid observations at a certain time point, this embodiment represents them as empty data, thus obtaining the first data sequence. For example, the first data sequence is represented as X; , Where X i Let x represent the multivariate input data at time i, T represent the sliding window size, and x represent the input data at time i. iLet x represent the data for the i-th variable, and n represent the number of variables, such as temperature, humidity, wave height, etc. empty This indicates empty data.

[0022] For missing data in the first data sequence, the average of ocean observation data from three consecutive time periods is used to fill in the missing data, resulting in the second data sequence. For missing data in the first data sequence, a filling method based on the average of neighboring data is used. Specifically, taking the time point of the missing data as the center, valid observation data of the same variable within three consecutive time periods are selected, their arithmetic average is calculated, and this average is used to replace the missing data, thus obtaining the second data sequence. When missing values ​​still exist in neighboring periods, recursive filling can be further used, i.e., first partially replacing the missing data with existing valid data, then recalculating the average until filling is complete. This method effectively reduces the impact of missing data on the overall sequence continuity and stability. This can be expressed using the following formula: , Where, x empty This indicates empty data, and i represents the data number of the time period associated with the empty data.

[0023] For the second data sequence, the data dimensions are normalized using the Z-score normalization method to obtain the first oceanographic data, which can be calculated using the following formula: , , , in, This represents the data value of the j-th variable at time i. This represents all data of the j-th dimension variable throughout the entire acquisition period. This represents the data obtained after preprocessing the j-th variable at time i. and These represent calculating the mean and variance of the data, respectively. This represents the data obtained after preprocessing at time i. This represents the data obtained after preprocessing at time T.

[0024] Features are extracted from the first ocean data, including time-series features, frequency-domain features, internal time-series features, and external time-series features, which are extracted based on the time-series features. In this embodiment, the extraction of the temporal features includes: The first ocean data is input into a preset time-series feature extraction module, and the output result is the time-series feature; The preset temporal feature extraction module includes a position encoding layer, a multi-head attention mechanism layer, a feedforward network layer, and a normalization layer; The location encoding layer introduces temporal sequence information into the input data, identifying the temporal positional relationships between different data points. Since the original first oceanographic data is only represented as a multi-dimensional vector when input into the network, the location encoding layer introduces temporal positional information into the data. In this embodiment, fixed location encoding vectors can be generated using sine and cosine functions, or learnable location embedding vectors can be used to map the positional relationships of time points to the corresponding vector space. Through location encoding, the network can identify the chronological relationship of the same variable at different times, thereby establishing global dependencies in the temporal dimension.

[0025] The multi-head attention mechanism layer is used to calculate the correlation between data points after the location encoding layer by querying the key vector, value vector, and key vector. Through the multi-head mechanism, attention weights can be computed in parallel across different subspaces, thereby capturing multi-scale and multi-angle temporal dependencies. This layer can highlight the focus on key moments in the observation sequence, weaken redundant or irrelevant parts, and thus obtain more refined temporal dependency features.

[0026] The feedforward network layer performs nonlinear mapping and dimensionality transformation based on the correlation between the data. After obtaining the attention-weighted output, the data enters the feedforward network layer. This layer typically consists of two fully connected neural networks: the first layer is used for dimensionality increase and nonlinear mapping, and the second layer is used for dimensionality reduction or further feature compression. Its function is to perform nonlinear transformation and feature reorganization on the correlation results output by the attention mechanism layer, making the extracted temporal features more suitable for downstream tasks. This layer can improve the expressive power of the model and avoid the limitations of relying solely on linear transformations.

[0027] The normalization layer is used to standardize the data and output temporal features after processing by the feedforward network layer. The data processed by the feedforward network layer is finally input into the normalization layer. This layer uses layer-by-layer normalization to standardize the features at different time points, ensuring the stability of the data within the numerical range and avoiding model training instability caused by gradient vanishing or exploding. At the same time, the normalization layer can enhance the model's generalization ability, making the output temporal features more stable and robust.

[0028] The calculation formula is: , , , , in and These represent the output features of the first and last layers, respectively. The location coding layer indicates the temporal and positional relationships of the data. (Layer Normalization, LN) is the normalization layer; (Forward Feed Network, FFN) is the forward feedback network layer; Attn is the multi-head attention mechanism layer, and its calculation formula is: , in, The length of the feature encoding vector for the current input data; Q, K, and V represent the query, key, and value in the attention mechanism, respectively.

[0029] Furthermore, in this embodiment, extracting frequency domain features includes: The time-series features are input into a preset frequency domain feature extraction module, and the output result is the frequency domain features; The preset frequency domain feature extraction module includes a fast Fourier transform layer, a first Transformer layer, and an inverse fast Fourier transform layer. Based on the Fast Fourier Transform layer, the time-series features are converted into frequency information; , in, The rotation factor for the frequency transform is used to extract periodic feature factors from time-series data. `n` represents the index value in the time dimension, `k` represents the frequency index, and `N` represents the number of sampling points for the frequency in the Fast Fourier Transform (FFT). The FFT layer maps the input time-series features from the time domain to the frequency domain. By performing a Fourier transform on the time-series feature sequence, the original time series can be decomposed into different frequency components and their amplitude and phase information, thus obtaining the frequency representation of each time series. FFT can effectively capture periodic variations in ocean observation data, such as tidal cycles, diurnal variations, and the frequency characteristics of wind and waves. The frequency information obtained in this step is in complex form, including real and imaginary parts, reflecting the energy distribution and phase relationship of the signal at different frequencies.

[0030] Based on the first Transformer layer, the periodic features of the frequency information are extracted; , , Here, mean represents the calculated mean of the frequency feature signal; std represents the variance of the frequency signal. After the frequency information is input into the first Transformer layer, periodic features are extracted. The first Transformer layer uses a multi-head self-attention mechanism with query vector (Q), key vector (K), and value vector (V) to calculate the correlation and dependency between different frequency components, thereby capturing potential periodic patterns and multi-scale features in the frequency signal. This layer can identify components with significant energy within a specific frequency range, enhancing the model's sensitivity to periodic patterns. Through a feedforward network and normalization processing, the nonlinear mapping and numerical stabilization of frequency features are further realized, thus obtaining a rich periodic representation.

[0031] Based on the inverse fast Fourier transform layer, the frequency information with extracted periodic features is transformed to obtain frequency domain features, which include query features, key features and value features.

[0032] The calculation formula is: , The frequency information processed by the first Transformer layer is then mapped back from the frequency domain to the time domain using an Inverse Fast Fourier Transform (iFFT) layer. The iFFT converts the frequency components encoded with periodic features into a time-domain representation, aligning the extracted frequency patterns with the original time series and combining them with time-series features in subsequent feature fusion. The resulting frequency-domain features contain both the time sequence information of the original signal and enhance the periodicity and frequency dependence characteristics of the signal.

[0033] Furthermore, in this embodiment, unlike the large-scale single-window feature extraction mode in the frequency domain, this embodiment designs a nested dual-window (inner and outer windows) temporal local mutation feature extraction structure. Extraction of external temporal features includes: The time series features are input into a preset external time series feature extraction module, and the output result is the external time series features. The preset external temporal feature extraction module includes a second Transformer layer; Based on the second Transformer layer, long-term features of the temporal features are extracted, and these long-term features are the external temporal features. The formula is: , Extracting intra-temporal features, including: The external temporal features are input into a preset internal temporal feature extraction module, and the output result is the internal temporal features; The preset internal temporal feature extraction module includes a third Transformer layer, and the window length of the third Transformer layer is smaller than the window length of the second Transformer layer. Based on the third Transformer layer, short-time features of the temporal features are extracted, and the short-time features are the internal temporal features.

[0034] The inner window feature is then implemented using a smaller window length coefficient in the outer window feature. To capture more refined anomaly information (local mutation features), the inner window output features... Defined as:

[0035] in and These are hyperparameters, representing the number and size of the inner windows, respectively. This indicates the feature slicing operation performed on the external window.

[0036] The internal and external time series features are fused to obtain internal and external time series features, including: The attention mechanism is used to calculate the weights of external and internal temporal features; Based on the weights of the external and internal time series features, the external and internal time series features are weighted and calculated to obtain the internal and external time series features, which include query features, key features, and value features.

[0037] , , , , in and These represent the weights of the inner window on the outer window and the outer window on the inner window, respectively. This represents the internal and external time-series characteristics. An attention mechanism is used to weight the two types of features to fully utilize their complementarity. Internal time-series features mainly reflect the changing patterns of a single observation node within a short time window, while external time-series features mainly reflect long-term trends and correlations across nodes or long-term windows. By fusing the two types of features, a comprehensive feature that considers both local short-term fluctuations and reflects the overall long-term trend can be generated.

[0038] The internal and external time-series features and frequency domain features are fused to obtain time-frequency domain features. In this embodiment, the fusion of internal and external time-series features and frequency domain features aims to simultaneously utilize the multi-scale time-series information of the time series and the frequency features of the signal to generate comprehensive time-frequency domain features, providing rich input information for subsequent anomaly detection, including: Based on the query features, key features, and value features of the internal and external time-series features and the query features, key features, and value features of the frequency domain features, the interaction relationship between the internal and external time-series features and the frequency domain features is obtained. Using the internal and external time-series features and the frequency domain features obtained in the preceding steps as input, their respective query features (Q), key features (K), and value features (V) are obtained. The Q, K, and V vectors of the internal and external time-series features reflect multi-node, multi-scale time dependencies, while the Q, K, and V vectors of the frequency domain features reflect the periodicity and energy distribution of the signal in the frequency dimension. By interactively calculating the Q, K, and V vectors corresponding to the two types of features, the dependency relationship and complementary information between the time-series features and the frequency domain features can be captured, forming a multi-dimensional interaction relationship matrix.

[0039] Based on a multi-head attention feature concatenation function, the interaction relationships are concatenated to obtain concatenated features; the interaction relationship matrix is ​​then concatenated using the same function. Specifically, the interaction features calculated by multiple attention heads are concatenated along the feature dimension to retain multi-angle, multi-scale feature information captured by different heads. Each attention head can independently focus on different time periods or frequency ranges, thereby improving the comprehensiveness and expressive power of feature fusion. The concatenated features contain complementary information from internal and external temporal features and frequency domain features, forming a richer representation.

[0040] Based on a multi-head attention computation function, the concatenated features are transformed to obtain time-frequency domain features. Then, based on the same multi-head attention function, linear transformation and weighted summation operations are performed on the concatenated features to obtain the final time-frequency domain features. Through this step, the fused features retain both the local and long-term dependency information of the time-series features and enhance the periodicity and frequency energy distribution characteristics of the frequency-domain features, achieving deep interaction and integration between the time and frequency domains.

[0041] , , , in, For multi-head attention computation function, is the multi-head attention feature concatenation function, where N is the number of multi-heads; and These represent query, key, and value features under frequency domain features and internal / external time series features, respectively.

[0042] The time-frequency domain features are input into a preset anomaly detection model to obtain the first ocean data on the anomaly, including: The time-frequency domain features are input into a preset anomaly detection model to obtain prediction results, which include the category probabilities of each variable. The anomaly detection model can be a traditional machine learning model, such as Support Vector Machine (SVM), Random Forest, Isolation Forest, etc.; or a deep learning model, such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), Transformer, etc. The model output is the probability distribution of each variable under different categories, that is, the predicted probability of each observed variable corresponding to the normal category and the anomaly category.

[0043] The sigmoid function is used to calculate the anomaly score for each variable under the anomaly category; When the anomaly score exceeds the preset anomaly detection threshold range, the corresponding data is determined to be the first ocean data that is abnormal.

[0044] The specific calculation formula is as follows: , , , in, Indicates a fully connected layer. For feature flattening operation, This represents the outlier score of the j-th variable at time T+1. This indicates that the maximum score is calculated as the final output score for this variable; The threshold for outlier scores is denoted by n, which represents the number of multivariates. This is a label indicating whether the j-th variable is abnormal data.

[0045] The first oceanographic data containing the aforementioned anomalies is corrected to obtain quality-controlled oceanographic data, including: The first ocean data for the aforementioned anomaly is replaced with the average value of ocean observation data from three consecutive time periods.

[0046] Furthermore, in this embodiment, the sliding window update is also included; the data at time T+1 (and the corrected values ​​for abnormal data) is used as the end time point of the next sliding window, and the sliding window position is updated from [0,1,…,T] to [1,2,…,T+1]. The above steps are performed to predict the abnormal data points of multiple variables at time T+2 until the quality control process ends.

[0047] This invention extracts temporal, frequency, and internal / external temporal features from observational data, performing multi-dimensional and multi-level feature fusion to more comprehensively reflect the inherent patterns of data in complex marine environments. Compared to traditional methods based on threshold judgment or single feature analysis, it can effectively identify data anomalies under nonlinear and non-stationary conditions, significantly reducing false positives and false negatives. By inputting the fused time-frequency domain features into a pre-defined anomaly detection model, intelligent identification of observational data can be achieved. Combined with subsequent anomaly correction steps, this not only ensures the integrity and continuity of the data but also significantly improves its accuracy and reliability, exhibiting strong robustness and adaptability. It can be widely applied to the real-time processing of offshore observation data, providing high-quality data support for marine environmental evolution research, disaster monitoring and early warning, and marine engineering safety.

[0048] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

[0049] The present invention has been described above with reference to embodiments thereof. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. The scope of the invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

[0050] Although embodiments of the present invention have been described in detail, it should be understood that various changes, substitutions, and modifications can be made to the embodiments of the present invention without departing from the spirit and scope of the invention.

[0051] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for quality control of offshore marine observation data based on time-frequency domain characteristics, characterized in that, include: The collected offshore marine observation data is preprocessed to obtain the first marine data; Features are extracted from the first ocean data, including time-series features, frequency-domain features, internal time-series features, and external time-series features, which are extracted based on the time-series features. The internal and external time series features are fused to obtain internal and external time series features; The internal and external time-series features and frequency domain features are fused to obtain time-frequency domain features; The time-frequency domain features are input into a preset anomaly detection model to obtain the first ocean data of the anomaly; The first ocean data with the aforementioned anomalies is corrected to obtain ocean data after quality control.

2. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 1, characterized in that, The preprocessing of the collected offshore marine observation data to obtain the first marine data includes: Using a sliding window approach, multivariate ocean observation data are processed into a data sequence. Data from different variables form a multidimensional input with the same time period. When there are no valid values, they are represented as empty data to obtain the first data sequence. The empty data in the first data sequence is filled with the average value of ocean observation data from three consecutive time periods to obtain the second data sequence; The second data sequence is normalized using the Z-score normalization method to obtain the first ocean data.

3. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 1, characterized in that, Extracting the time-series features includes: The first ocean data is input into a preset time-series feature extraction module, and the output result is the time-series feature; The preset temporal feature extraction module includes a position encoding layer, a multi-head attention mechanism layer, a feedforward network layer, and a normalization layer; The location coding layer introduces temporal sequence information into the input data to identify the temporal positional relationship between different data points; The multi-head attention mechanism layer is used to calculate the correlation between data after the data has been processed by the position encoding layer by query vector, key vector and value vector; The forward feedback network layer performs non-linear mapping and dimension transformation based on the correlation between the data; The normalization layer is used to standardize the temporal features of the data and output after processing by the feedforward network layer.

4. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 3, characterized in that, Extracting frequency domain features includes: The time-series features are input into a preset frequency domain feature extraction module, and the output result is the frequency domain features; The preset frequency domain feature extraction module includes a fast Fourier transform layer, a first Transformer layer, and an inverse fast Fourier transform layer. Based on the Fast Fourier Transform layer, the time-series features are converted into frequency information; Based on the first Transformer layer, the periodic features of the frequency information are extracted; Based on the inverse fast Fourier transform layer, the frequency information with extracted periodic features is transformed to obtain frequency domain features, which include query features, key features and value features.

5. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 4, characterized in that, Extracting external temporal features, including: The time series features are input into a preset external time series feature extraction module, and the output result is the external time series features. The preset external temporal feature extraction module includes a second Transformer layer; Based on the second Transformer layer, long-term features of the temporal features are extracted, and the long-term features are the external temporal features.

6. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 5, characterized in that, Extracting intra-temporal features, including: The external temporal features are input into a preset internal temporal feature extraction module, and the output result is the internal temporal features; The preset internal temporal feature extraction module includes a third Transformer layer, and the window length of the third Transformer layer is smaller than the window length of the second Transformer layer. Based on the third Transformer layer, short-time features of the temporal features are extracted, and the short-time features are the internal temporal features.

7. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 6, characterized in that, The internal and external time series features are fused to obtain internal and external time series features, including: The attention mechanism is used to calculate the weights of external and internal temporal features; Based on the weights of the external and internal time series features, the external and internal time series features are weighted and calculated to obtain the internal and external time series features, which include query features, key features, and value features.

8. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 7, characterized in that, The internal and external time-series features and frequency domain features are fused to obtain time-frequency domain features, including: Based on the query features, key features, and value features of the internal and external time series features and the query features, key features, and value features of the frequency domain features, the interaction relationship between the internal and external time series features and the frequency domain features is obtained. The interaction relationships are concatenated based on a multi-head attention feature concatenation function to obtain the concatenated features; Based on the multi-head attention computation function, the concatenated features are transformed to obtain time-frequency domain features.

9. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 8, characterized in that, The time-frequency domain features are input into a preset anomaly detection model to obtain the first ocean data on the anomaly, including: The time-frequency domain features are input into a preset anomaly detection model to obtain prediction results, which include the category probabilities of each variable. The sigmoid function is used to calculate the anomaly score for each variable under the anomaly category; When the anomaly score exceeds the preset anomaly detection threshold range, the corresponding data is determined to be the first ocean data that is abnormal.

10. The method for quality control of offshore marine observation data based on time-frequency domain characteristics according to claim 9, characterized in that, The correction of the first oceanographic data for the anomaly to obtain quality-controlled oceanographic data includes: The first ocean data for the aforementioned anomaly is replaced with the average value of ocean observation data from three consecutive time periods.

Images