A method for monitoring ocean internal waves based on deep learning

Abstract

The present invention discloses a method for monitoring ocean internal waves based on deep learning, which belongs to the technical field of ocean internal wave monitoring and is used for monitoring ocean internal waves, including collecting a velocity data set of an observed sea area as an original data set, performing quality control on the original data set, distinguishing normal sequences and abnormal sequences of the data after quality control, and dividing the data into a training set, a validation set, and a test set; constructing an ocean internal wave anomaly monitoring model, optimizing anomaly scores and reconstructing error parameters, using the trained ocean internal wave anomaly monitoring model to monitor the velocity observation data of the sea area, and identifying the ocean internal wave sequence; performing internal wave dynamics feature analysis on the ocean internal wave sequence extracted by the ocean internal wave anomaly monitoring model and outputting features. The present invention uses a deep learning model to accurately capture the behavior pattern of normal velocity, and realizes abnormal detection of internal wave flow by reconstructing errors, effectively distinguishing normal velocity from internal wave velocity, and reducing misjudgment and missed judgment caused by traditional methods.

Description

Technical Field

[0001] The invention discloses an ocean internal wave monitoring method based on deep learning, belonging to the technical field of ocean internal wave monitoring. Background Art

[0002] Ocean internal waves are a type of fluctuation that occurs in the ocean with stable density stratification, caused by the vertical vibration of seawater. Ocean internal waves and currents are fluctuation phenomena in the ocean, usually caused by temperature, salinity and flow velocity gradients. The occurrence of internal waves and currents has an impact on the marine ecological environment, offshore operating facilities, marine engineering structures and the navigation safety of ships. Therefore, identifying and monitoring internal waves and currents is an important issue in marine environmental protection and safe production. The identification of internal waves and currents in the existing technology mainly relies on ocean monitoring buoys and submerged buoy systems. The traditional identification method has problems such as low recognition rate, high misjudgment rate, and insufficient real-time performance, and it is difficult to effectively distinguish normal flow velocity from abnormal internal waves and currents. Summary of the invention

[0003] The purpose of the present invention is to provide a method for monitoring ocean internal waves based on deep learning to solve the problem of low accuracy of internal wave monitoring in the prior art.

[0004] A method for monitoring ocean internal waves based on deep learning, comprising:

[0005] S1. Collect the velocity data set of the observed sea area as the original data set, perform quality control on the original data set, distinguish the normal sequence and abnormal sequence of the quality-controlled data, and divide it into training set, validation set and test set;

[0006] S2. Improve the autoencoder AE model, build an ocean internal wave anomaly monitoring model, optimize the anomaly score and reconstruct the error parameter, use the trained ocean internal wave anomaly monitoring model to monitor the velocity observation data of the sea area, and identify the ocean internal wave sequence;

[0007] S3. Analyze the internal wave dynamic characteristics of the ocean internal wave sequence extracted by the ocean internal wave anomaly monitoring model and output the characteristics.

[0008] Quality control of the original data set includes uniqueness verification of the original data set, length alignment of the original data, and singular value detection of the original data. The singular value detection includes identifying and removing singular values, filling in missing values ​​or singular values, to ensure the accuracy and completeness of the data set.

[0009] The division of training set, validation set and test set includes:

[0010] The normal sequence and abnormal sequence are screened by the internal wave velocity impact single layer evaluation formula:

[0011] ;

[0012] In the formula, Indicates the first The velocity data of the layer, It represents the mean value of the difference in flow velocity changes in all directions. is the amplitude, is the wave speed, To influence the duration, , , is a constant coefficient, which is used to adjust the contribution of wave amplitude, wave velocity and impact duration to the impact assessment of each layer flow velocity. It is the single-layer evaluation formula for the effect of internal wave velocity;

[0013] Filter internal wave passing time The largest 7-dimensional time series with four layers in the east-west direction, two layers in the north-south direction, and one layer in the vertical direction is used as the model data set. Then 80% of the normal sequences are used as the training set, 10% of the normal sequences and 30% of the abnormal sequences are used as the validation set, and 10% of the normal sequences and 70% of the abnormal sequences are used as the test set.

[0014] The ocean internal wave anomaly monitoring model includes a data input layer, an encoder layer, a potential feature layer, an encoder layer and a reconstructed data layer;

[0015] The data input layer inputs the time series data into the encoder layer, and the encoder layer includes a multi-scale convolution layer and an LSTM layer. The multi-scale convolution layer includes multiple convolution layers with receptive fields from small to large. The multi-scale convolution layer outputs features of different receptive fields. The LSTM layer fuses features of different receptive fields and maps the features to the potential feature layer. The time step attention mechanism is introduced in the encoder layer to dynamically pay attention to important time points and generate weighted features for decoder decoding.

[0016] The encoder layer includes a multi-scale convolutional layer and an LSTM layer. The encoder layer extracts information from the latent feature layer, reproduces a sequence similar to the original input, and inputs the generated reconstruction result into the reconstructed data layer.

[0017] Introducing anomaly score calculation and reconstruction error calculation into the ocean internal wave anomaly monitoring model;

[0018] Anomaly scores are calculated for features in the latent feature layer, and a hypersphere is established to capture the distribution of normal data. During training, the radius of the hypersphere is automatically adjusted to adapt to changes in the distribution of normal data. The radius of the hypersphere is dynamically updated based on the distribution of normal data. When the network infers the input data, the distance between the latent feature and the center of the hypersphere is calculated:

[0019] ;

[0020] In the formula, is the potential feature and the center of the hypersphere The distance is a potential feature, according to Determine whether the data deviates from the normal distribution and whether it is abnormal;

[0021] Reconstruction error is calculated using mean square error:

[0022] ;

[0023] In the formula, is the reconstruction error calculation result, is the flow rate value, is the estimated flow rate, is the total number of samples, is the total feature dimension, It is The samples of the layer are at time step and feature dimensions The actual value of It is The samples of the layer are at time step and feature dimensions The reconstruction value on ;

[0024] The anomaly score and the reconstruction error are summed to generate the final anomaly score:

[0025] ;

[0026] In the formula, is the final anomaly score, is the weighting parameter;

[0027] Set the judgment threshold. If If the data point exceeds the judgment threshold, it is determined to be an abnormal data point;

[0028] Compare the data in the data input layer with the data in the reconstructed data layer, calculate the final anomaly score between them, and filter out abnormal data.

[0029] The analysis of internal wave dynamic characteristics includes calculation of internal wave flow duration, flow velocity processing, flow direction data processing and wave velocity calculation.

[0030] Calculating the internal current duration involves calculating the difference between the end time and the start time in minutes:

[0031] ;

[0032] ;

[0033] ;

[0034] In the formula, , , are the east-west, north-south, and vertical velocity components of the inner solitary wave after removing the background flow, , , The east-west, north-south, and vertical velocity components of the flow were measured on site. , , are the average values ​​of the east-west, north-south, and vertical velocity components within 30 minutes before the onset of the internal solitary wave, is the fluctuation velocity of the Doppler acoustic current profiler itself, It is calculated using the water depth data from the temperature, salinity and depth meter fixed on the buoy.

[0035] Flow velocity processing includes internal solitary wave flow direction for:

[0036] ;

[0037] ;

[0038] In the formula, when and hour, Take 0; when and when and hour, Take 1; when and hour, Take 2;

[0039] The propagation direction of the internal solitary wave is the same as the surface flow direction. The surface flow direction at the peak of the internal wave is taken as the propagation direction of the internal solitary wave. .

[0040] The flow direction data processing includes using the spatial projection method to project the latitudinal and longitudinal internal wave velocity on the entire section to various azimuths. The total horizontal kinetic energy density KED at each direction angle is calculated. The direction corresponding to the maximum value of KED is the propagation direction of the internal solitary wave. The calculation formula of KED is:

[0041] ;

[0042] ;

[0043] The horizontal and vertical flow velocities along the propagation direction of the internal solitary wave are:

[0044] ;

[0045] ;

[0046] In the formula, is the horizontal velocity along the propagation direction of the internal solitary wave, is the vertical velocity along the propagation direction of the internal solitary wave, is the propagation direction of the internal solitary wave angle.

[0047] The wave velocity calculation includes the KdV equation under the assumption of background flow and continuous stratification of seawater:

[0048] ;

[0049] ;

[0050] ;

[0051] In the KdV equation, the coordinate system is positive in the upward direction. is the vertical displacement of the jump layer, is the linear phase velocity, is the nonlinear term coefficient, is the dispersion term coefficient, and Depends on the background field density and horizontal flow velocity, The local water depth, is the background horizontal velocity, is the standardized vertical mode function, is the seawater depth;

[0052] Solving the homogeneous eigenvalue problem yields and :

[0053] ;

[0054] ;

[0055] The internal solitary wave steady-state solution of the KdV equation is in the form of:

[0056] ;

[0057] ;

[0058] ;

[0059] In the formula, is the maximum amplitude of the internal solitary wave, the concave type takes a negative value, and the convex type takes a positive value, is the nonlinear phase velocity, which is equivalent to the positive wave velocity, is the characteristic half-wave width.

[0060] Compared with the prior art, the present invention has the following beneficial effects: the deep learning model can accurately capture the behavior pattern of normal flow velocity, and realize the abnormal detection of internal wave flow by reconstructing the error; effectively distinguish the normal flow velocity from the internal wave flow velocity, reduce the misjudgment and missed judgment caused by the traditional method, and provide higher accuracy and stability for the monitoring of internal wave flow in complex marine environment. Through automated data preprocessing and internal wave flow anomaly detection, the state of internal wave flow in the ocean can be monitored in real time; the automated preprocessing of internal wave flow data and the detection of internal wave flow anomaly are realized, and the internal wave flow in the ocean is monitored and analyzed in real time through the automated processing flow, so as to provide rapid warning when abnormal fluctuations occur, and provide timely safety information guarantee for marine operations and navigation; analyze the key characteristics of internal wave flow, including dynamic characteristics such as flow velocity peak value, amplitude, and direction change. It can generate detailed reports on parameters such as duration and intensity of internal wave flow, providing important information for in-depth understanding of the dynamic characteristics of internal wave flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the process of the present invention;

[0062] Figure 2 It is a schematic diagram of the change of the flow velocity of each layer of the east-west flow during the passage of the internal wave in step S1 of the present invention;

[0063] Figure 3 It is a schematic diagram of the change of the flow velocity of each layer of the north-south flow during the passage of the internal wave in step S1 of the present invention;

[0064] Figure 4 It is a schematic diagram of the change of the flow velocity of each layer of the vertical flow during the passage of the internal wave in step S1 of the present invention;

[0065] Figure 5 A schematic diagram of constructing the sixth layer of the east-west flow of the multidimensional time series data set in step S1 of the present invention;

[0066] Figure 6 A schematic diagram of constructing the 12th layer of the east-west flow of the multidimensional time series data set in step S1 of the present invention;

[0067] Figure 7 A schematic diagram of constructing the 25th layer of the east-west flow of the multidimensional time series data set in step S1 of the present invention;

[0068] Figure 8 A schematic diagram of constructing the 31st layer of the east-west flow of the multidimensional time series data set in step S1 of the present invention;

[0069] Fig. 9 A schematic diagram of constructing the sixth layer of the north-south flow of the multidimensional time series data set in step S1 of the present invention;

[0070] Fig.10 A schematic diagram of constructing the 25th layer of the north-south flow of the multidimensional time series data set in step S1 of the present invention;

[0071] Fig.11 A schematic diagram of constructing the 18th layer of the vertical flow of the multidimensional time series data set in step S1 of the present invention;

[0072] Fig.12 It is a schematic diagram of the process flow of the ocean internal wave anomaly monitoring model in step S2 of the present invention;

[0073] Fig.13 This is a network structure diagram of the ocean internal wave anomaly monitoring model in step S2 of the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] A method for monitoring ocean internal waves based on deep learning, comprising:

[0076] S1. Collect the velocity data set of the observed sea area as the original data set, perform quality control on the original data set, distinguish the normal sequence and abnormal sequence of the quality-controlled data, and divide it into training set, validation set and test set;

[0077] S2. Improve the autoencoder AE model, build an ocean internal wave anomaly monitoring model, optimize the anomaly score and reconstruct the error parameter, use the trained ocean internal wave anomaly monitoring model to monitor the velocity observation data of the sea area, and identify the ocean internal wave sequence;

[0078] S3. Analyze the internal wave dynamic characteristics of the ocean internal wave sequence extracted by the ocean internal wave anomaly monitoring model and output the characteristics.

[0079] Quality control of the original data set includes uniqueness verification of the original data set, length alignment of the original data, and singular value detection of the original data. The singular value detection includes identifying and removing singular values, filling in missing values ​​or singular values, to ensure the accuracy and completeness of the data set.

[0080] The division of training set, validation set and test set includes:

[0081] The normal sequence and abnormal sequence are screened by the internal wave velocity impact single layer evaluation formula:

[0082] ;

[0083] In the formula, Indicates the first The velocity data of the layer, It represents the mean value of the difference in flow velocity changes in all directions. is the amplitude, is the wave speed, To influence the duration, , , is a constant coefficient, which is used to adjust the contribution of wave amplitude, wave velocity and impact duration to the impact assessment of each layer flow velocity. It is the single-layer evaluation formula for the effect of internal wave velocity;

[0084] Filter internal wave passing time The largest 7-dimensional time series with four layers in the east-west direction, two layers in the north-south direction, and one layer in the vertical direction is used as the model data set. Then 80% of the normal sequences are used as the training set, 10% of the normal sequences and 30% of the abnormal sequences are used as the validation set, and 10% of the normal sequences and 70% of the abnormal sequences are used as the test set.

[0085] The ocean internal wave anomaly monitoring model includes a data input layer, an encoder layer, a potential feature layer, an encoder layer and a reconstructed data layer;

[0086] The data input layer inputs the time series data into the encoder layer, and the encoder layer includes a multi-scale convolution layer and an LSTM layer. The multi-scale convolution layer includes multiple convolution layers with receptive fields from small to large. The multi-scale convolution layer outputs features of different receptive fields. The LSTM layer fuses features of different receptive fields and maps the features to the potential feature layer. The time step attention mechanism is introduced in the encoder layer to dynamically pay attention to important time points and generate weighted features for decoder decoding.

[0087] The encoder layer includes a multi-scale convolutional layer and an LSTM layer. The encoder layer extracts information from the latent feature layer, reproduces a sequence similar to the original input, and inputs the generated reconstruction result into the reconstructed data layer.

[0088] Introducing anomaly score calculation and reconstruction error calculation into the ocean internal wave anomaly monitoring model;

[0089] Anomaly scores are calculated for features in the latent feature layer, and a hypersphere is established to capture the distribution of normal data. During training, the radius of the hypersphere is automatically adjusted to adapt to changes in the distribution of normal data. The radius of the hypersphere is dynamically updated based on the distribution of normal data. When the network infers the input data, the distance between the latent feature and the center of the hypersphere is calculated:

[0090] ;

[0091] In the formula, is the potential feature and the center of the hypersphere The distance is a potential feature, according to Determine whether the data deviates from the normal distribution and whether it is abnormal;

[0092] Reconstruction error is calculated using mean square error:

[0093] ;

[0094] In the formula, is the reconstruction error calculation result, is the flow rate value, is the estimated flow rate, is the total number of samples, is the total feature dimension, It is The samples of the layer are at time step and feature dimensions The actual value of It is The samples of the layer are at time step and feature dimensions The reconstruction value on ;

[0095] The anomaly score and the reconstruction error are summed to generate the final anomaly score:

[0096] ;

[0097] In the formula, is the final anomaly score, is the weighting parameter;

[0098] Set the judgment threshold. If If the data point exceeds the judgment threshold, it is determined to be an abnormal data point;

[0099] Compare the data in the data input layer with the data in the reconstructed data layer, calculate the final anomaly score between them, and filter out abnormal data.

[0100] The analysis of internal wave dynamic characteristics includes calculation of internal wave flow duration, flow velocity processing, flow direction data processing and wave velocity calculation.

[0101] Calculating the internal current duration involves calculating the difference between the end time and the start time in minutes:

[0102] ;

[0103] ;

[0104] ;

[0105] In the formula, , , are the east-west, north-south, and vertical velocity components of the inner solitary wave after removing the background flow, , , The east-west, north-south, and vertical velocity components of the flow were measured on site. , , are the average values ​​of the east-west, north-south, and vertical velocity components within 30 minutes before the onset of the internal solitary wave, is the fluctuation velocity of the Doppler acoustic current profiler itself, It is calculated using the water depth data from the temperature, salinity and depth meter fixed on the buoy.

[0106] Flow velocity processing includes internal solitary wave flow direction for:

[0107] ;

[0108] ;

[0109] In the formula, when and hour, Take 0; when and when and hour, Take 1; when and hour, Take 2;

[0110] The propagation direction of the internal solitary wave is the same as the surface flow direction. The surface flow direction at the peak of the internal wave is taken as the propagation direction of the internal solitary wave. .

[0111] The flow direction data processing includes using the spatial projection method to project the latitudinal and longitudinal internal wave velocity on the entire section to various azimuths. The total horizontal kinetic energy density KED at each direction angle is calculated. The direction corresponding to the maximum value of KED is the propagation direction of the internal solitary wave. The calculation formula of KED is:

[0112] ;

[0113] ;

[0114] The horizontal and vertical flow velocities along the propagation direction of the internal solitary wave are:

[0115] ;

[0116] ;

[0117] In the formula, is the horizontal velocity along the propagation direction of the internal solitary wave, is the vertical velocity along the propagation direction of the internal solitary wave, is the propagation direction of the internal solitary wave angle.

[0118] The wave velocity calculation includes the KdV equation under the assumption of background flow and continuous stratification of seawater:

[0119] ;

[0120] ;

[0121] ;

[0122] In the KdV equation, the coordinate system is positive in the upward direction. is the vertical displacement of the jump layer, is the linear phase velocity, is the nonlinear term coefficient, is the dispersion term coefficient, and Depends on the background field density and horizontal flow velocity, The local water depth, is the background horizontal velocity, is the standardized vertical mode function, is the seawater depth;

[0123] Solving the homogeneous eigenvalue problem yields and :

[0124] ;

[0125] ;

[0126] The internal solitary wave steady-state solution of the KdV equation is in the form of:

[0127] ;

[0128] ;

[0129] ;

[0130] In the formula, is the maximum amplitude of the internal solitary wave, the concave type takes a negative value, and the convex type takes a positive value, is the nonlinear phase velocity, which is equivalent to the positive wave velocity, is the characteristic half-wave width.

[0131] The original data of the present invention is taken from a 75KHzADCP carried by a set of sea surface real-time transmission buoys deployed on the west side of a certain archipelago in the first half of 2024. The velocity data of the 75kHZADCP in the real-time transmission buoy of the internal wave current, the temperature, salinity and depth data of the CTD and the position data of the GPS are collected as the original data set, wherein the velocity data includes one layer of velocity every 8 meters in the east-west direction, one layer of velocity every 8 meters in the north-south direction, and one layer of velocity every 8 meters in the vertical direction. The quality control method of the original data set includes: performing a uniqueness test on the original data set to ensure that there is only one piece of data at each moment; aligning the length of the data to ensure that the length of all time series data is consistent. If there is any inconsistency in the length of the data, it is data missing, and the Kalman filter formula is used to fill the data of the entire moment; the data after quality control is distinguished between normal stable flow and abnormal internal wave flow time series, and the distinction method is as follows: the speed of the normal stable flow is relatively stable in a short time, the fluctuation amplitude is small, the change is relatively slow, and there will be no obvious sudden increase or decrease. During the passage of ocean internal waves, the flow velocity will increase rapidly, reaching a higher value than normal smooth flow. The instantaneous change in flow velocity is large, the flow velocity undergoes a sudden change, and an obvious sudden increase curve appears.

[0132] The division of time series signals of normal steady flow and abnormal internal wave flow includes: the speed of normal steady flow is relatively stable in a short period of time, with a small fluctuation amplitude and slow changes, and there will be no obvious sudden increase or decrease. In the time series diagram, the waveform of normal steady flow is relatively smooth, without large fluctuations, and appears as a low-frequency stable curve. During the passage of internal waves, the flow velocity usually increases rapidly, reaching a higher value than the normal steady flow, the instantaneous change of the flow velocity is large, and the flow velocity produces a sudden change in polarity, that is, the direction or magnitude of the flow velocity reverses sharply. In the time series diagram, the internal wave flow appears as obvious fluctuation peaks and valleys, forming periodic or non-periodic large fluctuations. In data preprocessing, the input data is standardized; the data is proportionally mapped to the range of [0,1]. This standardization method can maintain the relationship between the distribution form and characteristics of the data.

[0133] The collected abnormal internal wave flow data are analyzed as follows, and the following conclusions are drawn:

[0134] The surface water flows horizontally and is affected by wind and surface currents, resulting in significant changes in velocity. The surface flow direction is usually less affected by internal waves, but may change direction briefly when the wave crest propagates. Due to the vertical mixing of internal waves, the surface temperature drops slightly, while salinity and density changes are usually small.

[0135] The velocity of the middle layer of water is significantly affected by internal waves, especially in the crest and trough areas, where strong forward and backward flow and velocity reversal phenomena will occur. The velocity fluctuations may be large, sometimes even approaching or exceeding the velocity of the surface layer. When the internal wave current rolls, the middle layer of water will experience a significant reversal of flow direction, which is manifested as first moving forward and then flowing back in the trough. The cold water carried by the internal wave current is pushed into the middle layer, causing the temperature to drop, while in the trough area, warmer water may flow in, forming periodic temperature fluctuations. At the same time, salinity and density will also change accordingly, especially when crests and troughs appear alternately.

[0136] In deep water bodies, the energy propagation of internal waves is weakened, but it still affects the flow velocity. The flow velocity of deep water bodies will show periodic enhancement and weakening as the waves propagate, especially in the wave crest area, the flow velocity will increase. The flow direction of deep water bodies changes relatively little, but it will still be affected by internal waves and produce a certain horizontal swing, which manifests as a weak forward and backward flow. The temperature, salinity and density of deep water bodies change relatively little, but as internal waves propagate, the temperature will fluctuate to a certain extent, causing slight fluctuations in the temperature structure of deep water bodies.

[0137] When the internal waves reach the bottom layer, their energy has been significantly attenuated, but when they are close to the seafloor slope, they may still cause certain flow velocity fluctuations. The flow velocity changes are mainly manifested as periodic low-amplitude fluctuations. The flow direction of the bottom water is less affected by internal waves, but there may still be slight horizontal flow direction changes when the wave crest propagates. The temperature, salinity and density of the bottom water body have almost no significant changes, and the internal waves have a weak rolling and mixing effect on the bottom water body, but under specific terrain (such as steep seafloor slopes), internal waves may cause suspension and local mixing of bottom sediments.

[0138] The technical process of the present invention is as follows Figure 1As shown, it includes collecting the velocity data set of the observed sea area as the original data set, performing quality control on the original data set, distinguishing the normal sequence and the abnormal sequence of the quality-controlled data, and dividing the training set, the validation set and the test set; improving the autoencoder AE model, building an ocean internal wave anomaly monitoring model, optimizing the anomaly score and reconstructing the error parameter, using the trained ocean internal wave anomaly monitoring model to monitor the velocity observation data of the sea area, and identifying the ocean internal wave sequence; performing internal wave dynamics feature analysis on the ocean internal wave sequence extracted by the ocean internal wave anomaly monitoring model and outputting the features.

[0139] The velocity data above 100m are averaged to obtain the time series signals of normal steady flow and abnormal internal wave flow. Figure 2 , Figure 3 , Figure 4 As shown in the figure, the velocity of normal steady flow is relatively stable in a short period of time, with a small fluctuation amplitude and slow changes, and there will be no obvious sudden increase or decrease. In the time series diagram, the waveform of normal steady flow is relatively smooth, without large fluctuations, and it is manifested as a low-frequency stable curve. During the passage of internal waves, the flow velocity usually rises rapidly, reaching a higher value than the normal steady flow, and the instantaneous change of the flow velocity is large, and the flow velocity produces a sudden change in polarity, that is, the direction or magnitude of the flow velocity is sharply reversed. In the time series diagram, the internal wave flow is manifested as obvious fluctuation peaks and valleys, forming periodic or non-periodic large fluctuations.

[0140] The normal steady flow sequence and the abnormal internal wave flow sequence are distinguished. Figure 5 , Figure 6 , Figure 7 , Figure 8 As shown, the north-south direction is Fig. 9 , Fig.10 As shown, vertically Fig.11 As shown in the figure, when the internal wave flow passes, the influence on the screened layers is as follows: under the influence of the internal wave flow, the velocity time series of the 6th and 12th layers in the east-west direction and the 6th layer in the north-south direction experience an abnormal process, which is manifested as a sharp drop in the time series curve caused by a rapid increase in negative velocity when the internal wave front passes, and finally returns to the original steady flow state; the velocity of the 25th and 31st layers in the east-west direction and the 25th layer in the north-south direction is manifested as a sharp rise in the time series curve caused by a rapid increase in positive velocity when the internal wave front passes, and finally returns to the original steady flow state; the velocity of the 18th layer in the vertical direction is manifested in that the time series first drops rapidly, then rises sharply, and finally returns to the original steady flow state.

[0141] The process of the ocean internal wave anomaly monitoring model of the present invention is as follows: Fig.12As shown in the figure, data preprocessing is first performed, time series data is input, data normalization and sliding window are performed, and then the time series data is mapped to the latent space through LSTM and encoder, and the time dependency of the time series data is extracted through LSTM. The convolution layer TCN with different receptive fields is introduced in the multi-scale convolution layer to extract features of different time scales, and the time step attention mechanism is introduced to dynamically adjust the time step weights of the LSTM encoder and convolution features, and the important time steps are weighted by calculating the attention weights. The multi-scale convolution layer outputs the latent features, enters the LSTM decoder, reconstructs the latent features to the original input data, decodes the encoded latent features through LSTM, calculates the reconstruction error, obtains the difference between the original input and the reconstructed output, starts the single calibration mechanism, uses SVDD to calculate the abnormal score of the latent feature, and calculates whether the data is within the normal range. Finally, anomaly judgment is performed, and the reconstruction error and SVDD output are combined to determine whether the data is abnormal.

[0142] The structure of the ocean internal wave anomaly monitoring model of the present invention is as follows: Fig.13 As shown, the ocean internal wave anomaly monitoring model includes a data input layer, an encoder layer, a potential feature layer, an encoder layer and a reconstructed data layer, where X1, X2...Xt are the features;

[0143] The data input layer inputs the time series data into the encoder layer, and the encoder layer includes a multi-scale convolution layer and an LSTM layer. The multi-scale convolution layer includes multiple convolution layers with receptive fields from small to large. The multi-scale convolution layer outputs features of different receptive fields. The LSTM layer fuses features of different receptive fields and maps the features to the potential feature layer. The time step attention mechanism is introduced in the encoder layer to dynamically pay attention to important time points and generate weighted features for decoder decoding.

[0144] The encoder layer includes a multi-scale convolutional layer and an LSTM layer. The encoder layer extracts information from the latent feature layer, reproduces a sequence similar to the original input, and inputs the generated reconstruction result into the reconstructed data layer.

[0145] Anomaly score calculation and reconstruction error calculation are introduced into the ocean internal wave anomaly monitoring model; anomaly score calculation is performed for the features in the potential feature layer, and a hypersphere is established to capture the distribution of normal data. During the training process, the radius of the hypersphere is automatically adjusted to adapt to the distribution changes of normal data, and the radius of the hypersphere is dynamically updated based on the distribution of normal data. When the network infers the input data, the distance between the potential feature and the center of the hypersphere is calculated, and the reconstruction error is calculated using the mean square error. The anomaly score and the reconstruction error are summed to generate the final anomaly score, and a judgment threshold is set. If the judgment threshold is exceeded, it is determined to be an abnormal data point; the data in the data input layer is compared with the data in the reconstructed data layer, and the final anomaly score between them is calculated to filter out abnormal data.

[0146] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring ocean internal waves based on deep learning, characterized in that: include: S1. Collect the velocity data set of the observed sea area as the original data set, perform quality control on the original data set, distinguish the normal sequence and abnormal sequence of the quality-controlled data, and divide it into training set, validation set and test set; S2. Improve the autoencoder AE model, build an ocean internal wave anomaly monitoring model, optimize the anomaly score and reconstruct the error parameter, use the trained ocean internal wave anomaly monitoring model to monitor the velocity observation data of the sea area, and identify the ocean internal wave sequence; S3, analyzing the internal wave dynamic characteristics of the ocean internal wave sequence extracted by the ocean internal wave anomaly monitoring model and outputting the characteristics; The ocean internal wave anomaly monitoring model includes a data input layer, an encoder layer, a potential feature layer, an encoder layer and a reconstructed data layer; The data input layer inputs the time series data into the encoder layer, and the encoder layer includes a multi-scale convolution layer and an LSTM layer. The multi-scale convolution layer includes multiple convolution layers with receptive fields from small to large. The multi-scale convolution layer outputs features of different receptive fields. The LSTM layer fuses features of different receptive fields and maps the features to the potential feature layer. The time step attention mechanism is introduced in the encoder layer to dynamically pay attention to important time points and generate weighted features for decoder decoding. The encoder layer includes a multi-scale convolutional layer and an LSTM layer. The encoder layer extracts information from the latent feature layer, reproduces a sequence similar to the original input, and inputs the generated reconstruction result into the reconstruction data layer; Introducing anomaly score calculation and reconstruction error calculation into the ocean internal wave anomaly monitoring model; Anomaly scores are calculated for features in the latent feature layer, and a hypersphere is established to capture the distribution of normal data. During training, the radius of the hypersphere is automatically adjusted to adapt to changes in the distribution of normal data. The radius of the hypersphere is dynamically updated based on the distribution of normal data. When the network infers the input data, the distance between the latent feature and the center of the hypersphere is calculated: ； In the formula, is the potential feature and the center of the hypersphere The distance is a potential feature, according to Determine whether the data deviates from the normal distribution and whether it is abnormal; Reconstruction error is calculated using mean square error: ； In the formula, is the reconstruction error calculation result, is the flow rate value, is the estimated flow rate, is the total number of samples, is the total feature dimension, It is The samples of the layer are at time step and feature dimensions The actual value on It is The samples of the layer are at time step and feature dimensions The reconstruction value on ; The anomaly score and the reconstruction error are summed to generate the final anomaly score: ； In the formula, is the final anomaly score, is the weighting parameter; Set the judgment threshold. If If the data point exceeds the judgment threshold, it is determined to be an abnormal data point; Compare the data in the data input layer with the data in the reconstructed data layer, calculate the final anomaly score between them, and filter out abnormal data.

2. The method for monitoring ocean internal waves based on deep learning according to claim 1, characterized in that: Quality control of the original data set includes uniqueness verification of the original data set, length alignment of the original data, and singular value detection of the original data. The singular value detection includes identifying and removing singular values, filling in missing values ​​or singular values, to ensure the accuracy and completeness of the data set.

3. The method for monitoring ocean internal waves based on deep learning according to claim 1, characterized in that: The division of training set, validation set and test set includes: The normal sequence and abnormal sequence are screened by the internal wave velocity impact single layer evaluation formula: ； In the formula, Indicates the first The velocity data of the layer, It represents the mean value of the difference in flow velocity changes in all directions. is the amplitude, is the wave speed, To influence the duration, , , is a constant coefficient, which is used to adjust the contribution of wave amplitude, wave velocity and impact duration to the impact assessment of each layer flow velocity. It is the single-layer evaluation formula for the effect of internal wave velocity; Filter internal wave passing time The largest 7-dimensional time series with four layers in the east-west direction, two layers in the north-south direction, and one layer in the vertical direction is used as the model data set. Then 80% of the normal sequences are used as the training set, 10% of the normal sequences and 30% of the abnormal sequences are used as the validation set, and 10% of the normal sequences and 70% of the abnormal sequences are used as the test set.

4. The method for monitoring ocean internal waves based on deep learning according to claim 1, characterized in that: The analysis of internal wave dynamic characteristics includes calculation of internal wave flow duration, flow velocity processing, flow direction data processing and wave velocity calculation.

5. The method for monitoring ocean internal waves based on deep learning according to claim 4, characterized in that: Calculating the internal current duration involves calculating the difference between the end time and the start time in minutes: ； ； ； In the formula, , , are the east-west, north-south, and vertical velocity components of the inner solitary wave after removing the background flow, , , The east-west, north-south, and vertical velocity components of the flow were measured on site. , , are the average values ​​of the east-west, north-south, and vertical velocity components within 30 minutes before the onset of the internal solitary wave, is the fluctuation velocity of the Doppler acoustic current profiler itself, It is calculated using the water depth data from the temperature, salinity and depth meter fixed on the buoy.

6. The method for monitoring ocean internal waves based on deep learning according to claim 5, characterized in that: Flow velocity processing includes internal solitary wave flow direction for: ； ； In the formula, when and hour, Take 0; when and when and hour, Take 1; when and hour, Take 2; The propagation direction of the internal solitary wave is the same as the surface flow direction. The surface flow direction at the peak of the internal wave is taken as the propagation direction of the internal solitary wave. .

7. The method for monitoring ocean internal waves based on deep learning according to claim 6, characterized in that: The flow direction data processing includes using the spatial projection method to project the latitudinal and longitudinal internal wave velocity on the entire section to various azimuths. The total horizontal kinetic energy density KED at each direction angle is calculated. The direction corresponding to the maximum value of KED is the propagation direction of the internal solitary wave. The calculation formula of KED is: ； ； The horizontal and vertical flow velocities along the propagation direction of the internal solitary wave are: ； ； In the formula, is the horizontal velocity along the propagation direction of the internal solitary wave, is the vertical velocity along the propagation direction of the internal solitary wave, is the propagation direction of the internal solitary wave angle.

8. The method for monitoring ocean internal waves based on deep learning according to claim 7, characterized in that: The wave velocity calculation includes the KdV equation under the assumption of background flow and continuous stratification of seawater: ； ； ； In the KdV equation, the coordinate system is positive in the upward direction. is the vertical displacement of the jump layer, is the linear phase velocity, is the nonlinear term coefficient, is the dispersion term coefficient, and Depends on the background field density and horizontal flow velocity, The local water depth, is the background horizontal velocity, is the standardized vertical mode function, is the seawater depth; Solving the homogeneous eigenvalue problem yields and : ； ； The internal solitary wave steady-state solution of the KdV equation is in the form of: ； ； ； In the formula, is the maximum amplitude of the internal solitary wave, the concave type takes a negative value, and the convex type takes a positive value, is the nonlinear phase velocity, which is equivalent to the positive wave velocity, is the characteristic half-wave width.

Images