A method, system, electronic device and storage medium for predicting seabed topographic deformation

Abstract

This application provides a method, system, electronic device, and storage medium for predicting seabed topography deformation, belonging to the field of topography monitoring technology. The method includes: acquiring seabed topography data collected at the current time and multiple historical times to form a current multi-scale time series; decomposing the current multi-scale time series to obtain a current trend series, a current seasonal series, and a current residual series; performing gradient descent optimization on the current residual series to obtain a current optimal residual series; generating a current optimal multi-scale time series based on the current trend series, the current seasonal series, and the current optimal residual series; performing multi-scale partitioning and encoding on the current optimal multi-scale time series to obtain multiple sets of sub-sequences; and inputting the multiple sets of sub-sequences into a seabed deformation prediction model based on an adaptive multi-head attention mechanism for analysis to obtain seabed topography prediction data for future times. This application can improve the accuracy of seabed topography deformation prediction.

Description

Technical Field

[0001] This application relates to the field of terrain monitoring technology, and in particular to a method, system, electronic device and storage medium for predicting seabed terrain deformation. Background Technology

[0002] As a ubiquitous geological phenomenon, seafloor topography changes make accurate monitoring and prediction of these changes a crucial research topic in geographic information science and engineering. With advancements in sensor technology, acquiring seafloor topography data has become increasingly convenient. Effectively extracting and fusing this data is key to improving the accuracy of seafloor deformation prediction. Current technologies primarily employ the Fast Fourier Transform (FFT) algorithm for feature extraction, followed by convolution and folding of the extracted features. However, this approach is typically only suitable for stationary sequences and struggles to capture dynamic changes and complex patterns in seafloor topography, leading to inaccurate subsequent seafloor deformation predictions. Summary of the Invention

[0003] The main objective of this application is to propose a method, system, electronic device, and storage medium for predicting seabed topographic deformation, which can more comprehensively capture the characteristics of seabed topographic changes and thus improve the accuracy of seabed topographic deformation prediction.

[0004] To achieve the above objectives, one aspect of this application proposes a method for predicting seabed topographic deformation, the method comprising:

[0005] The seabed topography data collected at the current moment and at multiple historical moments are acquired to form a current multi-scale time series; wherein, the seabed topography data includes the height variation values ​​of multiple different locations relative to the sea level within the seabed area to be measured;

[0006] The current multi-scale time series is decomposed to obtain the current trend series, current seasonal series, and current residual series related to seafloor topography changes;

[0007] Gradient descent optimization is performed on the current residual sequence to obtain the current optimal residual sequence;

[0008] Generate the current optimal multi-scale time series based on the current trend sequence, the current seasonality sequence, and the current optimal residual sequence;

[0009] The current optimal multi-scale time series is divided and encoded into multiple scales to obtain multiple sets of subsequences;

[0010] The multiple sets of subsequences are input into the seabed deformation prediction model based on the adaptive multi-head attention mechanism for analysis to obtain seabed topography prediction data at future times.

[0011] Further, the step of performing gradient descent optimization on the current residual sequence to obtain the current optimal residual sequence includes:

[0012] The current residual sequence is used as the current residual sequence to be optimized;

[0013] Calculate the gradient value of the loss function with respect to the current residual sequence to be optimized. The loss function is used to measure the difference between the current multi-scale time series and the optimized current multi-scale time series. The optimized current multi-scale time series is obtained by optimizing the current residual sequence contained in the current multi-scale time series.

[0014] The gradient value is used to optimize the current residual sequence to be optimized, resulting in an optimized current residual sequence.

[0015] Determine whether the preset iteration termination condition is met; if yes, then use the optimized current residual sequence as the current optimal residual sequence; if no, then use the optimized current residual sequence as the current residual sequence to be optimized, and then return to the step of calculating the gradient value of the loss function on the current residual sequence to be optimized.

[0016] Further, generating the current optimal multi-scale time series based on the current trend series, the current seasonal series, and the current optimal residual series includes:

[0017] Obtain the historical trend sequence at a first historical moment and the historical seasonal sequence at a second historical moment, wherein the first historical moment is the moment before the current moment, and the time interval between the second historical moment and the current moment is a seasonal cycle;

[0018] Based on the difference between the current trend sequence and the historical trend sequence, the current trend sequence is optimized to obtain the current optimal trend sequence;

[0019] Based on the difference between the current seasonal sequence and the historical seasonal sequence, the current seasonal sequence is optimized to obtain the current optimal seasonal sequence;

[0020] The current optimal trend sequence, the current optimal seasonal sequence, and the current optimal residual sequence are fused to obtain the current optimal multi-scale time series.

[0021] Furthermore, the step of performing multi-scale partitioning and encoding on the current optimal multi-scale time series to obtain multiple sets of subsequences includes:

[0022] Based on a preset scale range, determine multiple different scale factors;

[0023] For each of the scale factors, the current optimal multi-scale time series is divided according to the scale factor to obtain all initial subsequences;

[0024] Based on the current optimal multi-scale time series, the boundaries of all initial subsequences are processed by moving average to obtain all optimal subsequences;

[0025] After position encoding is performed on all the optimal subsequences, a set of subsequences is formed.

[0026] Furthermore, the seafloor deformation prediction model based on the adaptive multi-head attention mechanism includes an encoder network, an adaptive multi-head attention network, and a feedforward neural network; the multiple sets of sub-sequences contain several optimal sub-sequences, and the several optimal sub-sequences carry several location encoding information; the step of inputting the multiple sets of sub-sequences into the seafloor deformation prediction model based on the adaptive multi-head attention mechanism for analysis to obtain seafloor topography prediction data at future times includes:

[0027] The encoder network is used to extract features from the several optimal subsequences to obtain several initial potential feature data related to seabed topographic changes;

[0028] Based on the aforementioned initial latent feature data and the aforementioned location encoding information, several final latent feature data are generated;

[0029] The adaptive multi-head attention network is used to extract features from the several final potential feature data to obtain several attention feature data related to seabed topographic changes;

[0030] The feedforward neural network is used to fuse and analyze the several attention feature data to obtain the seabed topography prediction data at the future time.

[0031] Furthermore, the step of generating several final latent feature data based on the several initial latent feature data and the several location encoding information includes:

[0032] The initial latent feature data and the location encoding information are concatenated to obtain a number of intermediate latent feature data.

[0033] The intermediate latent feature data are linearly transformed to obtain the final latent feature data.

[0034] Furthermore, the adaptive multi-head attention network includes multiple attention heads; the process of extracting features from the several final potential feature data using the adaptive multi-head attention network to obtain several attention feature data related to seabed topographic changes includes:

[0035] For each of the final latent feature data, the multiple attention heads are used to process the final latent feature data respectively to obtain multiple attention outputs;

[0036] The multiple attention outputs are weighted and summed to obtain the final attention output.

[0037] The final attention output is residually connected and normalized with the final latent feature data to obtain attention feature data.

[0038] To achieve the above objectives, another aspect of this application proposes a seabed topography deformation prediction system, the system comprising:

[0039] The acquisition module is used to acquire seabed topographic data collected at the current time and at multiple historical times to form a current multi-scale time series; wherein, the seabed topographic data includes the height variation values ​​of multiple different locations relative to the sea level within the seabed area to be measured;

[0040] The decomposition module is used to decompose the current multi-scale time series to obtain the current trend series, current seasonal series, and current residual series related to seafloor topography changes.

[0041] The optimization module is used to perform gradient descent optimization on the current residual sequence to obtain the current optimal residual sequence;

[0042] The generation module is used to generate the current optimal multi-scale time series based on the current trend series, the current seasonal series, and the current optimal residual series;

[0043] The partitioning module is used to perform multi-scale partitioning and encoding on the current optimal multi-scale time series to obtain multiple sets of subsequences;

[0044] The prediction module is used to input the multiple sets of subsequences into the seabed deformation prediction model based on the adaptive multi-head attention mechanism for analysis, so as to obtain seabed topography prediction data at future times.

[0045] To achieve the above objectives, another aspect of this application proposes an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0046] To achieve the above objectives, another aspect of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0047] This application includes at least the following beneficial effects: by decomposing the current trend sequence, current seasonal sequence, and current residual sequence from the current multi-scale time series formed by seabed topography data collected at different times, and then performing gradient descent optimization on the current residual sequence which covers the subtle features of seabed topography changes, the current trend sequence and current seasonal sequence are re-fused together to generate the current optimal multi-scale time series. Finally, the current optimal multi-scale time series is divided and encoded into multiple scales and then predicted and analyzed by the seabed deformation prediction model. Furthermore, an adaptive multi-head attention mechanism is introduced in the prediction and analysis process to more comprehensively capture the features of seabed topography changes, which can improve the prediction accuracy of seabed topography deformation. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating a method for predicting seabed topographic deformation provided in an embodiment of this application;

[0049] Figure 2 This is a schematic diagram of the structure of a seabed topography deformation prediction system provided in an embodiment of this application;

[0050] Figure 3 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of systems and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0052] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0053] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0055] Changes in seafloor topography not only affect the health of ecosystems, such as leading to accelerated climate change, sea-level rise, and more frequent extreme weather events, but also relate to the effective development and utilization of marine resources. As a ubiquitous geological phenomenon, accurate monitoring and prediction of seafloor topographic changes has become a crucial research topic in the field of geographic information science and engineering.

[0056] With advancements in sensor technology, acquiring seafloor topography data has become increasingly convenient, greatly facilitating the prediction of seafloor topography deformation. However, this data often exhibits complex trends and seasonal variations. Effectively extracting and fusing this data is crucial for improving the accuracy of seafloor topography deformation predictions. Therefore, multi-scale analysis is of great significance in seafloor topography research, enabling the analysis of topographic changes at multiple levels, such as sediment movement, the formation of seafloor fissures, and the interaction between climate change and tectonic movements. By integrating data from different scales, the complex mechanisms behind these changes can be revealed, thus providing a scientific basis for predicting seafloor topography changes.

[0057] In existing technologies, such as the MSGNet (Multi-Scale Graph Network) model and the TimesNet model, the fast Fourier transform algorithm is mainly used to extract features from these seabed topographic data, and then the extracted feature data is convolved and folded. However, this processing method is usually only applicable to stationary sequences and is difficult to capture the dynamic changes and complex patterns of seabed topography, resulting in inaccurate prediction of subsequent seabed topographic deformation.

[0058] In view of this, embodiments of this application provide a method, system, electronic device, and storage medium for predicting seabed topography deformation. This scheme decomposes the current trend sequence, current seasonal sequence, and current residual sequence from the current multi-scale time series formed by seabed topography data collected at different times. Then, after performing gradient descent optimization on the current residual sequence which covers subtle features of seabed topography changes, it is re-fused with the current trend sequence and current seasonal sequence to generate the current optimal multi-scale time series. Finally, after multi-scale division and encoding of the current optimal multi-scale time series, the prediction analysis is completed by the seabed deformation prediction model. Furthermore, an adaptive multi-head attention mechanism is introduced in the prediction analysis process to more comprehensively capture the features of seabed topography changes, which can improve the prediction accuracy of seabed topography deformation.

[0059] This application provides a method for predicting seabed topography deformation, relating to the field of topography monitoring technology. It can be applied to terminals, servers, or software running on either a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle-mounted terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the above-mentioned seabed topography deformation prediction method, but is not limited to these forms.

[0060] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0061] Figure 1 This is a schematic diagram of an optional process for predicting seabed topography deformation provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S106:

[0062] Step S101: Acquire seabed topographic data collected at the current time and at multiple historical times to form the current multi-scale time series;

[0063] Step S102: Decompose the current multi-scale time series to obtain the current trend series, current seasonal series and current residual series related to seafloor topography changes;

[0064] Step S103: Perform gradient descent optimization on the current residual sequence to obtain the current optimal residual sequence;

[0065] Step S104: Generate the current optimal multi-scale time series based on the current trend series, the current seasonal series, and the current optimal residual series;

[0066] Step S105: Perform multi-scale partitioning and encoding on the current optimal multi-scale time series to obtain multiple sets of subsequences;

[0067] Step S106: Input multiple sets of subsequences into the seabed deformation prediction model based on the adaptive multi-head attention mechanism for analysis to obtain seabed topography prediction data at future times.

[0068] Steps S101 to S106 shown in the embodiments of this application can capture the characteristics of seabed topographic changes more comprehensively, thereby improving the accuracy of seabed topographic deformation prediction.

[0069] In step S101 of some embodiments, multiple historical moments are all before the current moment, and the multiple historical moments and the current moment are consecutive moments. The seabed topography data collected at each moment includes the height change values ​​of multiple different locations within the seabed area to be measured relative to the sea level at that moment. This can be collected by deploying displacement sensors at multiple different locations within the seabed area to be measured. Furthermore, the current multi-scale time series can be understood as being formed by arranging all seabed topography data in the order of the collection time from first to last.

[0070] In step S102 of some embodiments, the STL (Seasonal and Trend decomposition using Loess) algorithm is used to decompose the current multi-scale time series into a current seasonal series, a current trend series, and a current residual series related to seafloor topography changes. This can be represented by the following expression:

[0071] X t =S t +T t +R t ;

[0072] In the formula, X t For the current multi-scale time series, S t This is the current seasonal sequence, used to characterize the seasonal fluctuations in seafloor subsidence, T. t This is the current trend sequence, used to characterize the long-term trend of seafloor subsidence, R. t This is the current residual sequence, used to characterize the random fluctuations in seabed subsidence.

[0073] Specifically, the STL algorithm, as a general and robust time series decomposition method, involves the following steps: After selecting a suitable seasonal period, the current multi-scale time series is first analyzed using a locally weighted regression smoothing method to fit a trend line. This allows for the decomposition of the current trend sequence from the current multi-scale time series to remove long-term variations. In this process, the focus is on transient effects related to seafloor topography changes, such as sediment movement or ocean current influences. Next, the initially decomposed current multi-scale time series is smoothed using a locally weighted regression smoothing method to identify short-term seasonal fluctuations. This allows for further decomposition of the current seasonal sequence from the initially decomposed current multi-scale time series. In this process, the focus is on regular changes in the shape and structure of the seafloor area under test caused by water level changes in certain situations. Finally, the re-decomposed current multi-scale time series is output as the current residual sequence, which represents the random fluctuations in the seafloor topography data and helps identify anomalies or subtle short-term changes that have not been captured.

[0074] It should be noted that the number of all trend component values ​​in the current trend series, the number of all seasonal component values ​​in the current seasonal series, and the number of all residual component values ​​in the current residual series are all the same as the number of all seafloor topography data in the current multi-scale time series. There is a time correspondence between the data contained in these four series.

[0075] In step S102 above, the current multi-scale time series is a non-stationary series. By introducing the STL algorithm to perform feature decomposition on the current multi-scale time series, spatiotemporal data related to seabed topographic deformation (such as tides, ocean currents, etc.) can be better processed.

[0076] As an optional implementation, the current multi-scale time series can be preprocessed before decomposition to ensure the integrity and accuracy of the collected data.

[0077] In this application, the Z-score algorithm is preferably used to preprocess the current multi-scale time series, including cleaning and imputation. The corresponding implementation process includes: for all height change values ​​collected by the same displacement sensor at different times in the current multi-scale time series, the average value and standard deviation of all height change values ​​are calculated. Then, each height change value is subtracted from the average value, and the result of the subtraction is divided by the standard deviation. It is determined whether the absolute value of the division result is greater than a preset value as an empirical rule. The preset value is preferably set to 3. If it is, the height change value is retained (that is, the height change value is judged as a normal value). If not, the height change value is replaced with the average value (that is, the height change value is judged as an outlier). This mean imputation strategy can avoid data bias.

[0078] In some embodiments, step S103 may include, but is not limited to, steps S201 to S204:

[0079] Step S201: Use the decomposed current residual sequence as the current residual sequence to be optimized.

[0080] Step S202: Calculate the gradient value generated by the loss function for the current residual sequence to be optimized; wherein, the loss function is mainly used to measure the difference between the current multi-scale time series and the optimized current multi-scale time series, which is obtained by optimizing the current residual sequence contained in the current multi-scale time series.

[0081] In this step, the mean square error (MSE) function is preferably used as the loss function, and the corresponding mathematical expression is as follows:

[0082]

[0083] Where: X i =S i +T i +R i ,

[0084] In the formula, L(R) t For the current residual sequence R t The loss function set for the optimization problem, where N is the current multi-scale time series X. t The amount of all seabed topographic data included. For the optimized current multi-scale time series, R t Let X' be the optimized current residual sequence. i For the current multi-scale time series X t The i-th seabed topography data included For the i-th optimized seabed topography data contained in the current optimized multi-scale time series, S i For the current seasonal sequence S t The i-th seasonal component value contained in T i For the current trend sequence T t The i-th trend component value contained in R i For the current residual sequence R t The i-th residual component value contained in R i R' is the optimized current residual sequence. t The i-th optimized residual component value contained in ′, and X i , S i T i R i and R i There is a time correspondence between these data points, meaning that these data points correspond to the same time.

[0085] In this step, the gradient value generated by the loss function for the current residual sequence to be optimized can be determined using the following expression:

[0086]

[0087] In the formula, The loss function L(R) t The current residual sequence R to be optimized t The resulting gradient value.

[0088] Step S203: Use the gradient value to optimize the current residual sequence to obtain the optimized current residual sequence, which can be determined by the following expression:

[0089]

[0090] In the formula, α is the learning rate used to control the update step size.

[0091] Step S204: Determine whether the preset iteration termination condition is met; if it is met, then use the optimized current residual sequence R′. t Output the current optimal residual sequence; if not, output the optimized current residual sequence R′. t The current residual sequence R to be optimized t Then return to execute the above step S202.

[0092] The iteration termination condition can be set as the change in the loss function being less than a preset change threshold. In this case, the current loss value needs to be calculated using the optimized current residual sequence obtained in the current optimization process, and the historical loss value needs to be calculated using the optimized current residual sequence obtained in the previous optimization process. The absolute value of the difference between the current loss value and the historical loss value is then calculated and used as the change in the loss function. The iteration termination condition can also be set as the current loss value converging to be less than the preset loss threshold, or it can be set as the current number of optimization executions reaching the preset maximum number of iterations. This application does not make any specific limitations on this.

[0093] In steps S201 to S204 of the embodiments of this application, considering that the current residual sequence decomposed by the STL algorithm is often ignored, that is, the subtle features of seabed topographic changes it covers are easily regarded as noise, this application uses the gradient descent method to optimize the current residual sequence. This can avoid information loss when the current optimal multi-scale time series is generated by subsequent fusion, and can also enable the subsequent model to better capture the potential patterns and structures in the residual, thereby improving the interpretability of the prediction.

[0094] In some embodiments, step S104 may include, but is not limited to, steps S301 to S304:

[0095] Step S301: Obtain the historical trend sequence at the first historical moment, where the first historical moment is the previous moment of the current moment. The historical trend sequence is obtained by decomposing the multi-scale time series collected at the first historical moment using the STL algorithm. And obtain the historical seasonal sequence at the second historical moment, where the time interval between the second historical moment and the current moment is a seasonal cycle. The historical seasonal sequence is also obtained by decomposing the multi-scale time series collected at the second historical moment using the STL algorithm.

[0096] Step S302: Based on the difference between the current trend sequence and the historical trend sequence, optimize the current trend sequence to obtain the current optimal trend sequence, which can be determined using the following expression:

[0097] ΔT t =T t -T t-1 , T′ t =T t +ΔT t ;

[0098] In the formula, T′ t For the current optimal trend sequence, T t-1 For the historical trend sequence at the first historical moment, Δt t Let T be the current trend sequence. t and the historical trend sequence T t-1 The first-order difference results between them.

[0099] Step S303: Based on the difference between the current seasonal sequence and the historical seasonal sequence, optimize the current seasonal sequence to obtain the current optimal seasonal sequence, which can be determined using the following expression:

[0100] ΔS t =S t -S t-s S′ t =S t +ΔS t ;

[0101] In the formula, S′ t S is the current optimal seasonal sequence. t-s Let s be the historical seasonal sequence at the second historical moment, where s is the seasonal period and ΔS is the value of ΔS. t For the current seasonal sequence S t and the historical seasonal sequence S t-s The first-order difference results between them.

[0102] Step S304: Merge the current best trend series, the current best seasonal series, and the current best residual series to obtain the current best multi-scale time series, which can be determined using the following expression:

[0103] X′ t =S′ t +T′ t +R′ t ;

[0104] In the formula, X′ t This is the current best multi-scale time series.

[0105] Steps S301 to S304 as shown in the embodiments of this application optimize the current trend sequence and current seasonal sequence decomposed by the STL algorithm by using a first-order difference processing method. This can highlight the dynamic characteristics of seabed topographic changes, avoid static influences, and facilitate subsequent models to better capture time series features.

[0106] In some embodiments, step S105 may include, but is not limited to, steps S401 to S402:

[0107] Step S401: Determine multiple different scale factors according to the preset scale range.

[0108] In this step, the scale range is preferably set to [1, 10], and the number of all seabed topographic data contained in the current optimal multi-scale time series should be much greater than 10. The number of all seabed topographic data contained in the current optimal multi-scale time series is the same as the number of all seabed topographic data contained in the current multi-scale time series. Multiple different scale factors are selected from this scale range by random selection. The value of each scale factor is an integer. This scale factor mainly represents the amount of data to be included in each initial subsequence that is subsequently divided.

[0109] Step S402: For each scale factor, firstly, the current best multi-scale time series is divided according to the scale factor to obtain all initial subsequences; secondly, based on the current best multi-scale time series, the boundaries of all initial subsequences are processed by moving average to obtain all best subsequences; finally, all best subsequences are positionally encoded to form a set of subsequences. According to this implementation method, by using multiple scale factors to divide and encode the current best multi-scale time series, multiple sets of corresponding subsequences can be obtained.

[0110] In this step, the number of all initial subsequences derived from the current optimal multi-scale time series can be determined using the following expression:

[0111]

[0112] In the formula, M is the total number of initial subsequences, N is the total number of seafloor topography data contained in the current optimal multi-scale time series, and s is the scale factor. This is the floor symbol.

[0113] For example, starting with the first seafloor topography data contained in the current optimal multi-scale time series, s seafloor topography data are extracted from the current optimal multi-scale time series to form the first initial subsequence; starting with the (s+1)th seafloor topography data contained in the current optimal multi-scale time series, s seafloor topography data are extracted from the current optimal multi-scale time series to form the second initial subsequence; starting with the (2s+1)th seafloor topography data contained in the current optimal multi-scale time series, s seafloor topography data are extracted from the current optimal multi-scale time series to form the third initial subsequence; and so on, until the Mth initial subsequence is formed.

[0114] In this step, for each initial subsequence, the boundary of the initial subsequence is processed by a moving average based on the current best multi-scale time series to obtain the optimal subsequence, which can be determined using the following expression:

[0115]

[0116] In the formula, x′ i Let x be the boundary of the initial subsequence. i The result of the moving average processing, x i This can be understood as the i-th seafloor topography data contained in the current optimal multi-scale time series, x i-m Let be the i-th seabed topographic data in the current optimal multi-scale time series, and k be the preset sliding window size.

[0117] For example, assuming a scale factor of 3, the current best multiscale time series X′ t All seabed topographic data included are denoted as {x0, x1, ..., x j ,...,x N-1}, j = 0, 1, ..., N-1, i = j + 1, at this time, the scaling factor is used to analyze the current optimal multi-scale time series X′. t Divide into, and obtain There are x0, x1, x2, x3, x4, x5, x6, x7, x8, ... ... N-3 ,x N-2 ,x N-1};

[0118] Assuming the sliding window size is 2, considering the current optimal multi-scale time series X′ t The query cannot find a seabed topography data point preceding the boundary x0 of the first initial subsequence {x0,x1,x2}. Therefore, it is assumed that no moving average processing is needed for the first initial subsequence {x0,x1,x2}, meaning it is considered the first optimal subsequence. For the second initial subsequence {x3,x4,x5}, its boundary x3 is updated to x′3 = (x2 + x3) / 2, and its boundary x5 is updated to x′5 = (x4 + x5) / 2, resulting in the second optimal subsequence {x′3,x4,x′5}. For the third initial subsequence {x6,x7,x8}, its boundary x6 is updated to x′6 = (x5 + x6) / 2, and its boundary x8 is updated to x′8 = (x7 + x8) / 2, resulting in the third optimal subsequence {x′6,x7,x′8}. This process continues until a more complete and accurate analysis is performed. The optimal subsequence.

[0119] In this step, for each optimal subsequence, it is preferable to use a sine function and a cosine function to generate the positional encoding information corresponding to that optimal subsequence. The expression for the sine function is as follows:

[0120]

[0121] The expression for the cosine function is as follows:

[0122]

[0123] In the formula, PE(pos,2r) and PE(pos,2r+1) are the location-coded data calculated using different functions, pos is the location index, r is the dimension index, and d is the position index. model d represents the dimension of the location encoding information, and this application preferably sets d to... model =2.

[0124] For example, for the first optimal subsequence {x0,x1,x2} mentioned above, the dimension index r ranges from [0,1,d]. model -1], the value range of the position index pos is [0,1,2], which corresponds to the data index of the first optimal subsequence {x0,x1,x2}. First, it is determined that only when xr=0 can the values ​​of 2r and 2r+1 fall within the value range of the dimension index r. Then, the relevant positional encoding data is calculated using the above sine function and the above cosine function, specifically as follows:

[0125] When pos = 0:

[0126]

[0127] When pos = 1:

[0128]

[0129] When pos = 2:

[0130]

[0131] Therefore, the positional encoding information corresponding to the first optimal subsequence {x0,x1,x2} can be determined to include [0,1], [0.8415,0.5403] and [0.9093,-0.4161].

[0132] Steps S401 to S402 as illustrated in the embodiments of this application divide the current optimal multi-scale time series into multiple scales, making it easier for subsequent models to capture features at different levels and time scales, thereby enhancing the model's overall understanding and prediction ability of seafloor topography changes; by performing boundary moving average processing on the initial subsequences obtained from the division, it is possible to avoid feature abrupt changes at the boundaries that could negatively impact subsequent model predictions; by performing position encoding on the optimized subsequences, it is easier for subsequent models to capture the sequential features of seafloor topography changes.

[0133] In some embodiments, step S106 may include, but is not limited to, steps S501 to S505:

[0134] Step S501: Input multiple sets of subsequences into the seabed deformation prediction model based on the adaptive multi-head attention mechanism. The multiple sets of subsequences include several optimal subsequences that carry several position encoding information. The seabed deformation prediction model based on the adaptive multi-head attention mechanism mainly includes an encoder network, an adaptive multi-head attention network and a feedforward neural network, and multiple attention heads are set in the adaptive multi-head attention network.

[0135] Step S502: The encoder network is used to extract features from several optimal subsequences to obtain several initial potential feature data related to seabed topographic changes.

[0136] In this step, for each optimal subsequence, the optimal subsequence is input into the encoder network for feature extraction to obtain initial latent feature data, which can be determined using the following expression:

[0137] P = Sigmoid(W1X′+b1);

[0138] In the formula, P is the initial latent feature data, X′ is the optimal subsequence, W1 is the weight matrix of the encoder network, b1 is the bias of the encoder network, and Sigmoid is the activation function, or ReLU function can be used.

[0139] Step S503: Generate several final latent feature data based on several initial latent feature data and several location encoding information.

[0140] In this step, several initial latent feature data and several location encoding information are concatenated to obtain several intermediate latent feature data; then, the several intermediate latent feature data are linearly transformed to obtain several final latent feature data.

[0141] Specifically, for each initial latent feature data, the initial latent feature data and its corresponding positional encoding information are concatenated to obtain intermediate latent feature data, which can be determined using the following expression:

[0142] P′=[P;PE];

[0143] Then, a linear transformation is performed on the intermediate latent feature data to obtain the final latent feature data, which can be determined using the following expression:

[0144] P″=W3P′+b3;

[0145] In the formula, P′ is the intermediate latent feature data, PE is the position encoding information corresponding to the initial latent feature data P, [;] is the splicing symbol, P″ is the final latent feature data, W3 is the preset first weight matrix, and b3 is the preset first bias.

[0146] By concatenating and linearly transforming the initial latent feature data with its corresponding location encoding information, the subsequent model can better learn and understand the diversity and complexity of seabed topographic changes, which helps to improve the model's predictive ability and also ensures data consistency in subsequent calculations.

[0147] Step S504: The adaptive multi-head attention network is used to extract features from several final potential feature data to obtain several attention feature data related to seabed topographic changes.

[0148] In this step, for each final latent feature data, the final latent feature data is input into the adaptive multi-head attention network for feature extraction to obtain attention feature data. The corresponding implementation methods include the following:

[0149] First, the final latent feature data is processed by multiple attention heads to obtain multiple attention outputs, which can be determined using the following expression:

[0150]

[0151]

[0152] Secondly, the weighted sum of multiple attention outputs yields the final attention output, which can be determined using the following expression:

[0153]

[0154] Finally, the final attention output is residually concatenated with the final latent feature data and normalized to obtain the attention feature data, which can be determined using the following expression:

[0155] U′ end =LayerNorm(U end +P″);

[0156] In the formula, Q h This is the query matrix corresponding to the h-th attention head. Let K be the query weight matrix corresponding to the h-th attention head. h Let h be the key matrix corresponding to the h-th attention head. V is the key weight matrix corresponding to the h-th attention head. h This is the value matrix corresponding to the h-th attention head. Let d be the value weight matrix corresponding to the h-th attention head. k The key matrix K h The dimension, and query matrix Q h The dimension and value matrix V h The dimension is also d k K j The key matrix K h The j-th dimension of data contained therein can also be understood as the key matrix K. h It contains data in the j-th column, where j = 1, 2, ..., d k T is the transpose symbol, A hj For query matrix Q h Key matrix K h The j-th dimension data K contained therein j Attention score, α hj U is the weight value. h U is the attention output obtained by processing the final latent feature data P″ by the h-th attention head. end For the final attention output, H is the number of attention heads, and β is the value of the attention head. h U′ represents the weight coefficient corresponding to the h-th attention head, which characterizes the feature contribution of the h-th attention head, where h = 1, 2, ..., H. endFor attention feature data, LayerNorm is the normalization function.

[0157] By introducing an adaptive multi-head attention mechanism, different attention heads learn different feature representations, and weight coefficients are used to adjust the contribution of each attention head. This allows the model to focus on more important attention heads to learn more important feature representations, thereby more accurately capturing the dynamic features of seabed topographic changes and enhancing the model's ability to predict seabed topographic deformation.

[0158] Step S505: The feedforward neural network is used to fuse and analyze several attention feature data to obtain the seabed topography prediction data at future times.

[0159] In this step, the feedforward neural network comprises an input layer, four hidden layers, and an output layer. Each hidden layer is used to map the input to a new space through a linear transformation. Specific implementation methods include:

[0160] This input layer concatenates several attention feature data to obtain a feature data to be tested.

[0161] The feature data to be tested is passed sequentially to four hidden layers for processing to obtain the final feature data, which can be determined using the following expression:

[0162] Hidden1 = ReLU(W4·Input) FFN +b4),

[0163] Hidden2=ReLU(W5·Hidden1+b5),

[0164] Hidden3=ReLU(W6·Hidden2+b6),

[0165] Hidden4=ReLU(W7·Hidden3+b7);

[0166] The output layer processes the final feature data to obtain the predicted seabed topography data for future times, which can be determined using the following expression:

[0167] Y = W8·Hidden4 + b 8 / ;

[0168] In the formula, Input FFNFor the feature data to be tested, W4 is the weight matrix of the first hidden layer, b4 is the bias of the first hidden layer, Hidden1 is the output of the first hidden layer, W5 is the weight matrix of the second hidden layer, b5 is the bias of the second hidden layer, Hidden2 is the output of the second hidden layer, W6 is the weight matrix of the third hidden layer, b6 is the bias of the third hidden layer, Hidden3 is the output of the third hidden layer, W7 is the weight matrix of the fourth hidden layer, b7 is the bias of the fourth hidden layer, Hidden4 is the output of the fourth hidden layer, which is the final feature data. W8 is the weight matrix of the output layer, b8 is the bias of the output layer, and Y is the predicted seabed topography data.

[0169] Steps S501 to S505, as shown in the embodiments of this application, can effectively improve the prediction accuracy of seabed topographic changes by sequentially inputting multiple sets of subsequences into the encoder network, the adaptive multi-head attention network, and the feedforward neural network for analysis.

[0170] In some embodiments, the training process for the seabed deformation prediction model is mainly divided into two parts. The first part is to train the encoder network separately. The second part is to use the trained encoder network, the adaptive multi-head attention network, and the feedforward neural network to build an initial seabed deformation prediction model and then train it as a whole. In this whole training process, only the parameters of the adaptive multi-head attention network and the feedforward neural network are updated.

[0171] Specifically, the training process for this encoder network includes the following:

[0172] The initial autoencoder model is built using this encoder network and an additional decoder network. The data processing within this autoencoder model can be represented by the following expression:

[0173] Y m =Sigmoid(W1X) in +b1),

[0174] Y out =Sigmoid(W2Y) m +b2);

[0175] In the formula, X in Y is the input to the encoder network, i.e., the input to the autoencoder model. m Y is the output of the encoder network, W2 is the weight matrix of the decoder network, b2 is the bias of the decoder network, and Y is the weight matrix of the decoder network. out This is the output of the decoder network, i.e., the output of the autoencoder model.

[0176] The first training dataset is obtained as follows: several different historical time periods are determined; for each historical time period, historical seabed topography data collected for the same seabed area under test during that historical time period are obtained and corresponding historical multi-scale time series are formed; the historical multi-scale time series is decomposed to obtain historical trend series, historical seasonal series, and historical residual series related to seabed topography changes; gradient descent optimization is performed on the historical residual series to obtain the historical optimal residual series; based on the historical trend series, the historical seasonal series, and the historical optimal residual series, a historical optimal multi-scale time series is generated; the historical optimal multi-scale time series is divided and encoded into multiple scales to obtain multiple sets of historical subsequences; the multiple sets of historical subsequences corresponding to each historical time period are used as training samples, thereby forming the first training dataset.

[0177] The reconstruction loss function is determined, preferably the mean squared error loss function, which is mainly used to characterize the input X of the autoencoder model. in With output Y out The differences between them.

[0178] When the autoencoder model is trained forward using the first training dataset, the reconstruction loss function is used to optimize the first network parameters contained in the autoencoder model. The first network parameters include two weight matrices (W1, W2) and two biases (b1, b2). The first network parameters are updated through the backpropagation algorithm until the reconstruction loss function reaches a convergent state, thereby obtaining the trained autoencoder model, and then the trained encoder network is obtained from it.

[0179] Specifically, the overall training process for this seabed deformation prediction model includes the following:

[0180] A second training dataset is obtained, which is updated based on the first training dataset. Specifically, for each historical time period, historical seabed topography data collected for the same seabed area to be measured at the next moment of that historical time period is obtained. The next moment of that historical time period is taken as the future moment, and the historical seabed topography data at the next moment of that historical time period is taken as the true seabed topography data at the future moment. Furthermore, the true seabed topography data at the future moment is used as the true label carried by multiple historical subsequences corresponding to that historical time period. In other words, by assigning a corresponding true label to each training sample contained in the first training dataset through the above implementation method, the second training dataset can be obtained.

[0181] Determine the training loss function, which can be either the mean squared error function or the mean absolute error function. This function is mainly used to characterize the difference between the output of the seabed deformation prediction model and the true label.

[0182] When forward training the seabed deformation prediction model using this second training dataset, the training loss function is used to optimize the second network parameters included in the seabed deformation prediction model. These second network parameters include five weight matrices (W4, W5, W6, W7, W8), five biases (b4, b5, b6, b7, b8), and weight coefficients (β1, β2, ..., β) corresponding to multiple attention heads. H The parameters of the second network are updated using the backpropagation algorithm or the Adam optimization algorithm until the training loss function reaches a convergent state, thereby obtaining a trained seabed deformation prediction model.

[0183] For example, the training loss function can be represented by the following expression:

[0184]

[0185] In the formula, L final Let S be the training loss function, and T be the number of all training samples in the second training dataset. s Y represents the true label of the s-th training sample in the second training dataset. s The prediction result is obtained by analyzing the s-th training sample of the seabed deformation prediction model.

[0186] The parameters of the second network are updated using the backpropagation algorithm, specifically as follows:

[0187]

[0188]

[0189]

[0190] In the formula, For the loss gradient, W p b′ is the updated weight matrix. p For the updated bias, p = 4, 5, 6, 7, 8, β′ h U represents the updated weight coefficients corresponding to the h-th attention head. h Let η be the output of the h-th attention head, and η be the learning rate.

[0191] The seafloor topography deformation prediction method proposed in this application decomposes the current trend sequence, current seasonal sequence, and current residual sequence from the current multi-scale time series formed by seafloor topography data collected at different times. After performing gradient descent optimization on the current residual sequence which covers subtle features of seafloor topography changes, it is then re-fused with the current trend sequence and current seasonal sequence to generate the current optimal multi-scale time series. Finally, the current optimal multi-scale time series is divided and encoded into multiple scales and then predicted and analyzed by the seafloor deformation prediction model. Furthermore, an adaptive multi-head attention mechanism is introduced during the prediction and analysis process to more comprehensively capture the features of seafloor topography changes, thereby improving the prediction accuracy of seafloor topography deformation.

[0192] Please see Figure 2 This application also provides a schematic diagram of a seabed topography deformation prediction system, which can implement the above-mentioned seabed topography deformation prediction method. The system includes:

[0193] The acquisition module 601 is used to acquire seabed topographic data collected at the current time and at multiple historical times to form the current multi-scale time series;

[0194] Among them, multiple historical moments are all before the current moment, and multiple historical moments and the current moment are consecutive moments. The seabed topography data collected at each moment includes the height change values ​​of multiple different locations in the seabed area to be measured relative to the sea level at that moment. This data can be collected by deploying displacement sensors at multiple different locations in the seabed area to be measured.

[0195] The decomposition module 602 is used to decompose the current multi-scale time series to obtain the current seasonal series, current trend series and current residual series related to seafloor topography changes;

[0196] Optimization module 603 is used to perform gradient descent optimization on the current residual sequence to obtain the current optimal residual sequence;

[0197] The generation module 604 is used to generate the current optimal multi-scale time series based on the current seasonal series, the current trend series, and the current optimal residual series.

[0198] The partitioning module 605 is used to perform multi-scale partitioning and encoding on the current best multi-scale time series to obtain multiple sets of subsequences;

[0199] The prediction module 606 is used to input multiple sets of subsequences into the seabed deformation prediction model based on an adaptive multi-head attention mechanism for analysis, so as to obtain seabed topography prediction data at future times.

[0200] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0201] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for predicting seabed topographic deformation. This electronic device can include any smart terminal such as a tablet computer or a vehicle-mounted computer.

[0202] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0203] Please see Figure 3 , Figure 3 This illustrates the hardware structure of an electronic device according to another embodiment, the electronic device comprising:

[0204] The processor 701 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0205] The memory 702 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 702 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this application are implemented through software or firmware, the relevant program code is stored in the memory 702 and is called and executed by the processor 701.

[0206] The input / output interface 703 is used to implement information input and output;

[0207] The communication interface 704 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0208] Bus 705 transmits information between various components of the device (e.g., processor 701, memory 702, input / output interface 703, and communication interface 704);

[0209] The processor 701, memory 702, input / output interface 703 and communication interface 704 are connected to each other within the device via bus 705.

[0210] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting seabed topographic deformation.

[0211] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0212] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0213] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0214] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0215] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0216] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0217] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0218] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0219] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0220] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0221] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0222] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0223] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for predicting seabed topographic deformation, characterized in that, The method includes: The seabed topography data collected at the current moment and at multiple historical moments are acquired to form a current multi-scale time series; wherein, the seabed topography data includes the height variation values ​​of multiple different locations relative to the sea level within the seabed area to be measured; The current multi-scale time series is decomposed to obtain the current trend series, current seasonal series, and current residual series related to seafloor topography changes; Gradient descent optimization is performed on the current residual sequence to obtain the current optimal residual sequence; Generate the current optimal multi-scale time series based on the current trend sequence, the current seasonality sequence, and the current optimal residual sequence; The current optimal multi-scale time series is divided and encoded into multiple scales to obtain multiple sets of subsequences; The multiple sets of subsequences are input into the seabed deformation prediction model based on the adaptive multi-head attention mechanism for analysis to obtain seabed topography prediction data at future times. The step of performing gradient descent optimization on the current residual sequence to obtain the current optimal residual sequence includes: The current residual sequence is used as the current residual sequence to be optimized; Calculate the gradient value of the loss function with respect to the current residual sequence to be optimized. The loss function is used to measure the difference between the current multi-scale time series and the optimized current multi-scale time series. The optimized current multi-scale time series is obtained by optimizing the current residual sequence contained in the current multi-scale time series. The gradient value is used to optimize the current residual sequence to be optimized, resulting in an optimized current residual sequence. Determine whether the preset iteration termination condition is met; if yes, then use the optimized current residual sequence as the current optimal residual sequence; if no, then use the optimized current residual sequence as the current residual sequence to be optimized, and then return to the step of calculating the gradient value of the loss function on the current residual sequence to be optimized.

2. The method for predicting seabed topographic deformation according to claim 1, characterized in that, The step of generating the current optimal multi-scale time series based on the current trend series, the current seasonal series, and the current optimal residual series includes: Obtain the historical trend sequence at a first historical moment and the historical seasonal sequence at a second historical moment, wherein the first historical moment is the moment before the current moment, and the time interval between the second historical moment and the current moment is a seasonal cycle; Based on the difference between the current trend sequence and the historical trend sequence, the current trend sequence is optimized to obtain the current optimal trend sequence; Based on the difference between the current seasonal sequence and the historical seasonal sequence, the current seasonal sequence is optimized to obtain the current optimal seasonal sequence; The current optimal trend sequence, the current optimal seasonal sequence, and the current optimal residual sequence are fused to obtain the current optimal multi-scale time series.

3. The method for predicting seabed topographic deformation according to claim 1, characterized in that, The process of multi-scale partitioning and encoding the current optimal multi-scale time series to obtain multiple sets of subsequences includes: Based on a preset scale range, determine multiple different scale factors; For each of the scale factors, the current optimal multi-scale time series is divided according to the scale factor to obtain all initial subsequences; Based on the current optimal multi-scale time series, the boundaries of all initial subsequences are processed by moving average to obtain all optimal subsequences; After position encoding is performed on all the optimal subsequences, a set of subsequences is formed.

4. The method for predicting seabed topographic deformation according to claim 3, characterized in that, The seafloor deformation prediction model based on the adaptive multi-head attention mechanism includes an encoder network, an adaptive multi-head attention network, and a feedforward neural network; the multiple sets of sub-sequences contain several optimal sub-sequences, and the several optimal sub-sequences carry several location encoding information; the process of inputting the multiple sets of sub-sequences into the seafloor deformation prediction model based on the adaptive multi-head attention mechanism for analysis to obtain seafloor topography prediction data at future times includes: The encoder network is used to extract features from the several optimal subsequences to obtain several initial potential feature data related to seabed topographic changes; Based on the aforementioned initial latent feature data and the aforementioned location encoding information, several final latent feature data are generated; The adaptive multi-head attention network is used to extract features from the several final potential feature data to obtain several attention feature data related to seabed topographic changes; The feedforward neural network is used to fuse and analyze the several attention feature data to obtain the seabed topography prediction data at the future time.

5. The method for predicting seabed topographic deformation according to claim 4, characterized in that, The step of generating several final latent feature data based on the several initial latent feature data and the several location encoding information includes: The initial latent feature data and the location encoding information are concatenated to obtain a number of intermediate latent feature data. The intermediate latent feature data are linearly transformed to obtain the final latent feature data.

6. The method for predicting seabed topographic deformation according to claim 4, characterized in that, The adaptive multi-head attention network is configured with multiple attention heads; the adaptive multi-head attention network is used to extract features from the several final potential feature data to obtain several attention feature data related to seabed topographic changes, including: For each of the final latent feature data, the multiple attention heads are used to process the final latent feature data respectively to obtain multiple attention outputs; The multiple attention outputs are weighted and summed to obtain the final attention output. The final attention output is residually connected and normalized with the final latent feature data to obtain attention feature data.

7. A seabed topography deformation prediction system, characterized in that, The system includes: The acquisition module is used to acquire seabed topographic data collected at the current time and at multiple historical times to form a current multi-scale time series; wherein, the seabed topographic data includes the height variation values ​​of multiple different locations relative to the sea level within the seabed area to be measured; The decomposition module is used to decompose the current multi-scale time series to obtain the current trend series, current seasonal series, and current residual series related to seafloor topography changes. The optimization module is used to perform gradient descent optimization on the current residual sequence to obtain the current optimal residual sequence; The generation module is used to generate the current optimal multi-scale time series based on the current trend series, the current seasonal series, and the current optimal residual series; The partitioning module is used to perform multi-scale partitioning and encoding on the current optimal multi-scale time series to obtain multiple sets of subsequences; The prediction module is used to input the multiple sets of subsequences into the seabed deformation prediction model based on the adaptive multi-head attention mechanism for analysis, so as to obtain seabed topography prediction data at future times. The step of performing gradient descent optimization on the current residual sequence to obtain the current optimal residual sequence includes: The current residual sequence is used as the current residual sequence to be optimized; Calculate the gradient value of the loss function with respect to the current residual sequence to be optimized. The loss function is used to measure the difference between the current multi-scale time series and the optimized current multi-scale time series. The optimized current multi-scale time series is obtained by optimizing the current residual sequence contained in the current multi-scale time series. The gradient value is used to optimize the current residual sequence to be optimized, resulting in an optimized current residual sequence. Determine whether the preset iteration termination condition is met; if yes, then use the optimized current residual sequence as the current optimal residual sequence; if no, then use the optimized current residual sequence as the current residual sequence to be optimized, and then return to the step of calculating the gradient value of the loss function on the current residual sequence to be optimized.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images