Underwater glider CTD sampling abnormal data detection method and system based on AE-LSTM and storage medium

Abstract

The invention discloses an AE-LSTM-based underwater glider CTD sampling abnormal data detection method and system and a storage medium, and belongs to the technical field of ocean exploration, the detection method comprises the steps that an AE-LSTM hybrid model is constructed, and the model comprises a bidirectional long short-term memory network Bi-LSTM encoder and a unidirectional LSTM decoder; preprocessing the original CTD time sequence data, wherein the preprocessing comprises edge auxiliary point expansion and upper-layer ocean data density interpolation; inputting the preprocessed data into the AE-LSTM hybrid model for encoding and decoding, and generating reconstructed data; calculating a reconstruction error of the original data and the reconstruction data; and dynamically setting a threshold value by a POT method based on an extreme value theory EVT, and judging abnormal data according to the reconstruction error. The method has the advantages of being high in detection precision, high in adaptability, low in misjudgment rate, suitable for large-scale automatic processing and the like, the requirement for obtaining high-quality observation data in the complex dynamic marine environment can be effectively met, and reliable data guarantee is provided for marine scientific research, environmental monitoring and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ocean exploration, and particularly relates to an underwater glider CTD sampling abnormal data detection method and system based on AE-LSTM and a storage medium. BACKGROUND

[0002] An underwater glider is a new type of autonomous underwater vehicle driven by buoyancy and without propeller propulsion. The design integrates buoy system, subsurface buoy system and underwater robot technology. Since the engineering of this type of equipment in the early 21st century, it has been widely used in global ocean scientific research and military tasks, and has gradually become one of the key technical equipment in the field of ocean exploration. The basic working principle of the underwater glider is to realize non-powered sliding through the conversion between buoyancy and gravitational potential energy. Specifically, the glider adjusts the oil bladder inside the body to change the displacement volume, thereby controlling the net buoyancy and realizing the heave motion in the vertical direction. At the same time, by moving the internal mass block (usually a battery pack), the pitch angle and roll angle are adjusted to obtain a sliding path along a zigzag trajectory. During the gliding process, the glider is in a non-active propulsion state most of the time, and only consumes energy for a short time when adjusting the buoyancy and controlling the attitude. This low-power design makes the glider have the advantages of long endurance, long range, low noise, etc., and can realize continuous observation tasks for several months.

[0003] When the glider approaches the sea surface, the tail communication antenna can float out of the water to establish a connection with the shore-based platform, transmit the observation data and running state information collected during the gliding process, and receive subsequent task instructions. Due to its excellent operating performance and multi-task adaptability, the underwater glider has become one of the important platforms for current ocean environment monitoring and data acquisition. During the ocean observation task of the glider, the Conductivity-Temperature-Depth sensor (CTD) can measure the conductivity, temperature and pressure of seawater, among which the salinity can be indirectly calculated from the conductivity, temperature and pressure, and the depth can be calculated according to the salinity, temperature and pressure. The above parameters constitute important indicators of the basic state of the ocean environment and have key significance for the study of ocean physical processes. However, the CTD sensor is easily affected by platform motion, environmental disturbance and its own structural characteristics during actual operation, resulting in inevitable outliers and noise in the observation data. These errors not only reduce the accuracy of the data, but also may interfere with the subsequent analysis of the ocean environment, numerical prediction and scientific research conclusions. Therefore, how to accurately and reliably control the quality of the CTD observation data has become an important link to ensure the effectiveness and scientificity of the ocean observation data.

[0004] The quality control method of current underwater glider CTD observation data is mainly borrowed from the Argo global ocean observation program. The program has built a real-time data acquisition network covering the global ocean by deploying self-sinking and floating profile floats. The floats dive and float according to the preset process, and continuously collect temperature, salinity and pressure profile data in the process, forming complete water structure information. The data quality control system of Argo project is divided into real-time mode quality control and delayed mode quality control. Real-time mode mainly faces the data released quickly, and usually completes the basic abnormal detection and screening within 12 to 72 hours after data acquisition; the delayed mode is based on historical data, sensor performance analysis and reference materials to carry out more strict and accurate quality evaluation and correction of observation data to meet the needs of scientific accuracy. The quality control method of underwater glider CTD data is similar, which also includes real-time quality control and delayed quality control. Real-time quality control emphasizes processing speed and faces the data returned from the glider in the task execution process, usually for task frame information at certain depth intervals. Due to the limited resolution and conservative threshold setting, real-time quality control is prone to miss some potential data anomalies. In order to make up for the shortcomings of real-time processing and improve the scientific application value of data, delayed quality control needs to be carried out after the task is completed. Delayed processing focuses on the in-depth analysis of high-resolution full-profile data, including systematic evaluation of sensor state, comparison of data with historical observation results, and correction of possible sensor drift. Its common data verification process includes time consistency check, geographic location verification, reverse pressure test, extreme value test, profile anomaly detection, burr removal, gradient change detection, standard deviation analysis, vertical velocity test, density inversion detection and comparison with climate average state, etc.

[0005] With the continuous improvement of the quality requirements of ocean science research on observation data, the traditional CTD data quality control method has gradually shown its limitations in applicability and performance. The widely used quality control method is mainly based on experience rules and statistical analysis methods, such as fixed threshold judgment method, gradient change detection method and manual visual inspection. Although these methods have the advantages of simple implementation and high computational efficiency, they have significant limitations when faced with large-scale, multi-source and heterogeneous observation data. Specifically, the traditional method usually cannot accurately identify abnormal data under nonlinear characteristics, and cannot effectively adapt to complex and dynamic changes in the marine environment. Especially in the case of various types of data anomalies and strong noise interference, the miss rate is high and manual secondary inspection is needed. In the actual inspection process, the abnormal judgment is highly dependent on expert experience, which is highly subjective, resulting in insufficient stability of the quality control results; at the same time, the manual visual inspection process is tedious, time-consuming and labor-intensive, and the overall efficiency is low.

[0006] In summary, the existing CTD data quality control technology is difficult to meet the development needs of high precision, automation and intelligentization, and it is urgent to introduce a new deep learning method with complex feature learning ability to break through the bottleneck of traditional technology in accuracy, adaptability and processing efficiency. SUMMARY

[0007] In view of the problems that the existing CTD data quality control method has high missing detection rate, strong subjectivity, low efficiency and is difficult to adapt to complex dynamic marine environment when facing large-scale, multi-source and heterogeneous observation data, the application provides an AE-LSTM-based underwater glider CTD sampling abnormal data detection method, system and storage medium.

[0008] The application is implemented in the following manner: an AE-LSTM-based underwater glider CTD sampling abnormal data detection method, characterized in that it comprises the following steps: constructing an AE-LSTM hybrid model, wherein the model comprises a bidirectional long short-term memory network Bi-LSTM encoder and a unidirectional LSTM decoder; preprocessing the original CTD time series data, including edge auxiliary point expansion and upper ocean data density interpolation; inputting the preprocessed data into the AE-LSTM hybrid model for encoding and decoding to generate reconstructed data; calculating the reconstruction error of the original data and the reconstructed data; dynamically setting a threshold based on the POT method of extreme value theory EVT, and determining abnormal data according to the reconstruction error.

[0009] In the above technical solution, preferably, the hidden state dimension of the Bi-LSTM encoder is 128, the input dimension of the unidirectional LSTM decoder is 64 and the hidden state dimension is 128, the encoder and the decoder both use a tanh activation function, the optimizer is Adam, and the learning rate is 0.001.

[0010] In the above technical solution, preferably, the edge auxiliary point expansion comprises predicting auxiliary points using an LSTM model, the LSTM model has a hidden layer dimension of 128, and 5 auxiliary points are predicted based on 100 historical data each time; the auxiliary points are spliced to the front and back of the original data, and only the effective output of the central region is retained after reconstruction.

[0011] In the above technical solution, preferably, the upper ocean data density interpolation comprises: defining the upper ocean depth as 0-120 meters, calculating the data density according to the absolute value of the adjacent salinity difference; distributing the number of interpolation points in proportion to the density, and the total number of interpolation points is 2000; using a spline interpolation method to complete the data in the sparse area to form a uniformly distributed data sequence.

[0012] In the technical scheme, preferably, the POT dynamic threshold setting method comprises: independently modeling each CTD profile, screening the reconstruction error exceeding the initial threshold; fitting extreme error data based on a generalized Pareto distribution (GPD); determining the dynamic threshold by calculating the exceedance probability, and realizing adaptive determination of the abnormal points.

[0013] In the technical scheme, preferably, the method takes salinity data as the main input parameter for anomaly detection, and if the salinity data is determined to be abnormal, the corresponding conductivity, temperature and pressure data are simultaneously marked as abnormal.

[0014] In the technical scheme, preferably, the input data is normalized to the interval [-1, 1], and different profile data is input into the model in the form of an overall sequence.

[0015] In the technical scheme, preferably, the method further comprises combining with a traditional quality control process, including thermal hysteresis correction, extreme value test and gradient change detection, to further verify the anomaly detection result.

[0016] The application provides an AE-LSTM-based underwater glider CTD sampling anomaly data detection method, which fully utilizes the advantages of deep learning in time series modeling and feature extraction, and effectively improves the detection accuracy, stability and intelligent level of CTD data quality control. The method adopts a deep hybrid model structure composed of an autoencoder, a bidirectional long short-term memory (Bi-LSTM) encoder and a unidirectional LSTM decoder to perform deep feature learning and reconstruction modeling on CTD time series data. The Bi-LSTM encoder can simultaneously capture the forward and backward dependencies of the time series, enhancing the model's understanding and modeling ability of nonlinear, multi-scale and dynamically changing ocean environmental characteristics. The LSTM decoder is responsible for restoring data and optimizing reconstruction error, so that the error of the model in reconstructing normal data is significantly smaller than that of abnormal data, thereby realizing high-precision identification of abnormal data. Compared with traditional threshold-based or experience rule-based detection methods, the deep model has self-learning ability and does not need to set complex rules manually, can accurately identify various types of anomalies under complex noise interference, and greatly improves the accuracy and automation level of anomaly detection.

[0017] In view of the defects of large reconstruction error and easy misjudgment of time series reconstruction model in the edge area of data (i.e. both ends of the sequence), the application further proposes an edge effect correction strategy based on auxiliary points. The specific method is to introduce a certain number of auxiliary points at both ends of the original CTD time series, and to carry out continuation processing combined with the existing data characteristics, so as to enhance the understanding ability of the model to the characteristics of the boundary area, thereby weakening the misjudgment phenomenon caused by the error of the edge area. This strategy significantly improves the detection stability and comprehensiveness of the model to the whole range of the sequence, especially in the case that the starting and ending sections of the actual glide profile are easily affected by disturbances, it has stronger adaptability and robustness.

[0018] In addition, considering that the reconstruction model is prone to large reconstruction error due to sparse data and drastic changes when processing upper layer seawater data (such as shallow low pressure area) on the profile, the application introduces an interpolation optimization method based on data variation density. This method dynamically determines the number and distribution density of interpolation points by analyzing the variation density of data at different water depth horizons, and carries out encryption and spline interpolation processing on the original data, so as to optimize the training data structure and improve the learning and fitting ability of the model to the upper profile data. The interpolation optimization method not only improves the balance of data distribution and reduces the reconstruction error of the model to the high gradient area, but also significantly improves the overall data reconstruction quality, further enhancing the accuracy and applicability of anomaly detection.

[0019] In summary, the application realizes anomaly detection of CTD observation data through the AE-LSTM hybrid deep model, and combines edge effect correction and density-based interpolation optimization strategy to comprehensively improve the intelligent level of underwater glider CTD sampling data quality control. This method has the advantages of high detection precision, strong adaptability, low misjudgment rate and suitability for large-scale automatic processing, and can effectively meet the demand for obtaining high-quality observation data in complex dynamic marine environment, providing reliable data guarantee for marine scientific research, environmental monitoring and numerical prediction.

[0020] A computer readable storage medium is proposed, which stores a computer program, and the program is executed by a processor to realize the steps of the above method.

[0021] An underwater glider CTD data anomaly detection system is proposed, comprising:

[0022] The preprocessing module is used to perform the edge auxiliary point extension and density interpolation of claims 2 to 4;

[0023] The AE-LSTM hybrid model module is used to perform the data reconstruction of claim 1;

[0024] The dynamic threshold determination module is used to perform the POT dynamic threshold calculation and anomaly marking of claim 5. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 AE-LSTM model structure diagram, showing the connection relationship between Bi-LSTM encoder and unidirectional LSTM decoder;

[0026] Figure 2 Edge error comparison between original data and reconstructed data;

[0027] Figure 3 Auxiliary point expansion and center area reservation schematic diagram;

[0028] Figure 4 Density interpolation flowchart based on salinity change;

[0029] Figure 5 Overall flowchart of anomaly detection and POT dynamic threshold setting;

[0030] Figure 6 Experimental results of salinity data sensitivity to anomalies;

[0031] Figure 7 Comparison chart of reconstructed data after edge correction;

[0032] Figure 8 Comparison of upper ocean data distribution before and after interpolation;

[0033] Figure 9 Traditional thermal hysteresis correction effect diagram;

[0034] Figure 10 Comparison of anomaly detection results between the present application and traditional methods. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0036] In order to solve the problems of high false negative rate, strong subjectivity, low efficiency and difficulty in adapting to complex dynamic marine environment of the existing CTD data quality control method when facing large-scale, multi-source and heterogeneous observation data, the present application provides a kind of underwater glider CTD sampling anomaly data detection method, system and storage medium based on AE-LSTM. In order to further illustrate the structure of the present application, the detailed description is as follows in combination with the drawings:

[0037] A kind of underwater glider CTD sampling anomaly data detection method based on AE-LSTM, comprising the following steps:

[0038] The AE-LSTM hybrid model is constructed, which includes a bidirectional long short-term memory network (Bi-LSTM) encoder and a unidirectional LSTM decoder.

[0039] A hybrid model of fusion of an autoencoder (AE), a bidirectional long short-term memory network (Bi-LSTM) and a unidirectional LSTM is constructed. The model uses the Bi-LSTM as an encoder to encode the input time series data and compress it into a low-dimensional latent vector, while extracting and retaining key dynamic features. The Bi-LSTM is composed of two forward and reverse LSTM layers, which can simultaneously capture the historical dependence and future dependence in the time series, thereby fully utilizing the context information of the sequence. The decoder part uses a unidirectional LSTM network to reconstruct the original sequence according to the encoded latent vector. The reconstruction error is optimized by back propagation, so that the latent space can accurately represent the distribution characteristics of the original data.

[0040] During the model training process, the input is a plurality of historical profile data sequences, and the model reconstructs the input data through compression and decoding process. The reconstruction error is used to measure the degree of abnormality of the data points, and the larger the error, the more likely it is an abnormal point.

[0041] The original CTD time series data is preprocessed, including edge auxiliary point expansion and upper ocean data density interpolation; the preprocessed data is input into the AE-LSTM hybrid model for encoding and decoding to generate reconstructed data; and the reconstruction error of the original data and the reconstructed data is calculated.

[0042] After reconstructing the data using the above model, the reconstruction error of the data on the left and right sides is much larger than that of the middle part of the data, as shown in Figure 2 This is because there is no previous data at the beginning of the data and no subsequent data at the end of the data, while the data in the middle part has both the data on the left and the data on the right. Therefore, it is necessary to improve this phenomenon. To solve this edge effect problem, the application proposes an auxiliary point prediction mechanism. During training and prediction, the LSTM model generates auxiliary points based on historical data, and then expands several auxiliary points before and after the original data as input, and only retains the output of the central region as the effective reconstruction result, as shown in Figure 3 This way effectively enhances the context information of the data in the edge area by improving the coverage of the input, thereby significantly improving the reconstruction accuracy of the edge data points and reducing the misjudgment rate.

[0043] The upper ocean (such as 0-120 meters) has less data and sparse distribution, while the lower ocean has large amount of data and dense distribution, so if the model is directly input for reconstruction, the model almost only reconstructs the upper ocean data according to the density of the lower ocean data, and the reconstruction error of the upper ocean data is much larger than that of the lower ocean data. Therefore, the reconstruction error of the normal data in the upper ocean is too large, which leads to misjudgment as an anomaly. Therefore, the application introduces an interpolation method based on density, which can make the data density of the upper ocean almost consistent with that of the lower ocean, so as to improve the reconstruction effect of the upper ocean data. The key of the interpolation method based on density lies in the calculation of the density and the distribution of the interpolation points, and the principle is as shown in Figure 4 The data density can be measured by the salinity change degree between adjacent data points: the larger the salinity difference, the more intense the data change, the sparser the distribution between points, and the smaller the density; on the contrary, the smaller the salinity difference, the more gentle the data change, the denser the distribution between points, and the larger the density. Therefore, the absolute value of the adjacent salinity difference can be normalized as the data density index of each position. Next, the distribution of the interpolation points is determined according to the density of each position. First, the total number of interpolation points is set, and then the number of interpolation points to be allocated is calculated according to the ratio of the density of each point to the overall density. Finally, interpolation is performed in each region according to the distribution using spline interpolation and other methods, so as to realize the density adaptive completion of the data.

[0044] The POT method based on the EVT dynamically sets the threshold value, and determines the abnormal data according to the reconstruction error. The POT dynamically sets the threshold value method includes: independently modeling each CTD profile, screening the reconstruction error exceeding the initial threshold value; fitting the extreme error data based on the GPD; determining the dynamic threshold value by calculating the exceeding probability, and realizing the adaptive determination of the abnormal points.

[0045] The application selects the POT (Peaks Over Threshold) method as the dynamic threshold value calculation means of the reconstruction error. The POT is one of the core methods in the EVT (Extreme Value Theory), and the GPD (Generalized Pareto Distribution) is used to model the extreme error exceeding the high threshold value, so as to depict the tail behavior and identify the abnormal value.

[0046] The method mainly includes three steps: first, a higher initial threshold value is selected, and the reconstruction error values exceeding the threshold value are screened out; second, the GPD is fitted to these extreme values; and finally, the abnormal points are determined by calculating the exceeding probability. Unlike the traditional static threshold method, the POT can adaptively adjust the threshold value according to the tail characteristics of the data, and has good dynamic adaptability.

[0047] In view of the significant time, space and platform differences of the CTD profile data, the application independently applies the POT method to model the reconstruction error of each profile, and dynamically selects the most suitable threshold, so as to more accurately identify abnormal data. This local modeling-based method effectively overcomes the lack of adaptability of the fixed threshold method when facing non-stationary data, and significantly improves the accuracy and flexibility of anomaly detection.

[0048] The technical solution of the application is shown in the anomaly detection process as Figure 5 The AE-LSTM hybrid model is constructed, first, the original data is increased with edge auxiliary points, and then the upper ocean data is interpolated according to the density. After data preprocessing, the AE-LSTM hybrid model is input for reconstruction, the reconstructed data is removed from the auxiliary points and interpolation points, then the reconstruction error is calculated, and finally the reconstruction error is judged according to the threshold to detect the abnormal points.

[0049] When an anomaly occurs, salinity is more sensitive to abnormal changes than conductivity, pressure and temperature, as shown in Figure 6 To improve the accuracy of anomaly detection, the method selects salinity as the main basis for anomaly discrimination, only inputs the salinity data into the anomaly detection model for reconstruction, and uses the reconstruction error as the criterion to identify abnormal points. If a salinity data point is determined to be abnormal, it can be inferred that the corresponding conductivity, pressure and temperature data also have abnormalities.

[0050] Data normalization is an important preprocessing step in machine learning, which can improve the training efficiency and effect of the model. Since the salinity data is generally distributed between 33 PSU and 35 PSU, the salinity interval [33, 35] is scaled to the interval [-1, 1]. Put the profile data of different lengths into a list, and when training the model, input the different profiles into the AE-LSTM hybrid model as a whole each time.

[0051] In the proposed AE-LSTM architecture, the dimension of the hidden state of the bidirectional LSTM layer in the encoder is 128, while the input dimension of the unidirectional LSTM layer in the decoder is set to 64, and the dimension of the hidden state is also 128. The number of layers of the bidirectional LSTM layer in the encoder and the unidirectional LSTM layer in the decoder is set to 1, and the activation function of all linear layers is set to tanh, and the learning rate is set to 0.001. Adam is used as the optimizer of the AE-LSTM hybrid model.

[0052] In the correction of the edge effect problem, LSTM model is used to predict auxiliary points, and the LSTM hidden dimension is set to 128, and 100 historical data are used to predict 5 auxiliary points at a time. The left and right sides of the data are spliced with the corresponding 5 auxiliary points. As can be seen from Figure 7 It can be seen that the data edge reconstruction accuracy is good, and the original data and the reconstructed data are almost consistent.

[0053] In order to improve the processing effect of upper ocean data, data less than 120m is determined as upper ocean data, and the total interpolation quantity is set to 2000, and spline interpolation method is used for interpolation. Figure 8 It can be seen that the right subgraph uses the density-based interpolation method, and the reconstructed data is almost consistent with the original data.

[0054] The method also includes combining with the traditional quality control process, including thermal hysteresis correction, extreme value test and gradient change detection, to further verify the abnormal detection result. Specifically, taking a profile as an example, the traditional quality control process is preliminarily used and the thermal hysteresis correction is performed, and the correction effect is as shown in Figure 9 Then, the method of the present application is used to detect the abnormal profile, and the traditional quality control process is combined, and finally the thermal hysteresis error is corrected. The effect is as shown in Figure 10 It can be seen that the method proposed in the present application removes a large number of abnormal points missed by the traditional quality control process at 34.58 PSU. Since the method proposed in the present application can effectively detect CTD data anomalies, there is no missed phenomenon, so it is not necessary to manually draw Figure Two Abnormal values are identified, so the efficiency of data quality control is greatly improved.

[0055] Embodiment two

[0056] A computer readable storage medium stores a computer program, the program is executed by a processor to realize the steps of the method of embodiment one.

[0057] Embodiment three

[0058] An underwater glider CTD data anomaly detection system, comprising a preprocessing module, an AE-LSTM hybrid model module and a dynamic threshold determination module: the preprocessing module is used for performing edge auxiliary point expansion and density interpolation; the AE-LSTM hybrid model module is used for performing data reconstruction; the dynamic threshold determination module is used for performing POT dynamic threshold calculation and abnormal marking.

[0059] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting abnormal data of underwater glider CTD sampling based on AE-LSTM, characterized in that: The following steps are involved: An AE-LSTM hybrid model is constructed, comprising a bidirectional long short-term memory network (Bi-LSTM) encoder and a unidirectional LSTM decoder. The original CTD time series data is preprocessed, including edge-assisted point expansion and upper ocean data density interpolation. The preprocessed data is input into the AE-LSTM hybrid model for encoding and decoding to generate reconstructed data. The reconstruction error between the original data and the reconstructed data is calculated. A threshold is dynamically set based on the POT method of extreme value theory (EVT), and abnormal data is determined based on the reconstruction error.

2. The underwater glider CTD sampling abnormal data detection method based on AE-LSTM according to claim 1 is characterized in that: The hidden state dimension of the Bi-LSTM encoder is 128, the input dimension of the unidirectional LSTM decoder is 64 and the hidden state dimension is 128, both the encoder and the decoder use the tanh activation function, the optimizer is Adam, and the learning rate is 0.

001.

3. The underwater glider CTD sampling abnormal data detection method based on AE-LSTM according to claim 1 is characterized in that: The edge auxiliary point expansion includes using an LSTM model to predict auxiliary points, where the hidden layer dimension of the LSTM model is 128, and 5 auxiliary points are predicted each time based on 100 historical data; the auxiliary points are spliced ​​to the front and back ends of the original data, and only the valid output of the central area is retained after reconstruction.

4. The underwater glider CTD sampling abnormal data detection method based on AE-LSTM according to claim 1 is characterized in that: The upper ocean data density interpolation includes: defining the upper ocean depth as 0 to 120 meters, calculating the data density based on the absolute value of the adjacent salinity difference; allocating the number of interpolation points according to the density ratio, with a total number of interpolation points of 2000; and using the spline interpolation method to complete the data in sparse areas to form a uniformly distributed data sequence.

5. The underwater glider CTD sampling abnormal data detection method based on AE-LSTM according to claim 1 is characterized in that: The POT dynamic threshold setting method includes: independently modeling each CTD profile and screening reconstruction errors that exceed the initial threshold; fitting extreme error data based on the generalized Pareto distribution (GPD); determining the dynamic threshold by calculating the exceedance probability to achieve adaptive judgment of outliers.

6. The underwater glider CTD sampling abnormal data detection method based on AE-LSTM according to claim 1 is characterized in that: The method uses salinity data as the main input parameter for anomaly detection. If the salinity data is judged to be abnormal, the corresponding conductivity, temperature and pressure data are simultaneously marked as abnormal.

7. The method for detecting abnormal data of underwater glider CTD sampling based on AE-LSTM according to claim 1, characterized in that: The input data are normalized to the interval [-1,1], and the data of different profiles are input into the model as an overall series.

8. The method for detecting abnormal data of underwater glider CTD sampling based on AE-LSTM according to claim 1, characterized in that: The method also includes integration with traditional quality control processes, including thermal hysteresis correction, extreme value testing, and gradient change detection, to further validate the anomaly detection results.

9. A computer-readable storage medium, characterized in that A computer program is stored, and when the program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. An underwater glider CTD data anomaly detection system, characterized in that: include: A pre-processing module for performing edge-assisted point expansion and density interpolation as described in claims 2 to 4; An AE-LSTM hybrid model module, configured to perform the data reconstruction described in claim 1; A dynamic threshold determination module is used to perform the POT dynamic threshold calculation and abnormality marking as described in claim 5.

Images