A ship side wind speed correction prediction method and system based on self-supervised contrast learning

Abstract

The present application belongs to the technical field of ship side wind speed prediction, and provides a ship side wind speed correction prediction method and system based on self-supervised contrast learning, comprising: obtaining marine observation data; extracting data features of the obtained marine observation data; performing multi-granularity contrast learning on the extracted data features to obtain feature vectors of different granularity time series data of the marine observation data; and completing correction prediction of the ship side wind speed according to the obtained time series data feature vectors. The present application uses ultrasonic wind speed data to correct left and right ship side wind speeds, automatically extracts different granularity feature vectors from marine observation time series by constructing positive and negative sample pairs, without the need for manual parameter adjustment or reliance on domain expertise; has the ability to automatically learn the internal structure and patterns of data, improves the perception of errors in left and right ship side wind speed data, and enhances the correction capability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of ship side wind speed prediction, and particularly relates to a ship side wind speed correction and prediction method and system based on self-supervised contrastive learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Offshore wind speed data plays a crucial role in weather and sea state prediction, safe navigation, power generation, hydrological simulation, and other applications; it is usually obtained through regular marine observations on ships. However, wind speed is random and intermittent, and is easily affected by ship movement and marine environmental factors, resulting in a certain deviation between the wind speed values measured by ship-borne sensors and the actual wind field values in the area where the ship is located. Therefore, the need to correct these data to obtain more accurate actual wind speed information becomes increasingly important.

[0004] According to the inventors, the actual wind speed of a ship can be obtained through two apparently different approaches, traditional statistical methods and emerging artificial intelligence techniques. Traditional methods usually rely on establishing mathematical models or physical laws based on the relationship between sensor data and ship motion information; for example, Autoregressive Integrated Moving Average (ARIMA), Exponential Time Smoothing (ETS), High-Resolution Limited Area Model (HIRLAM), Fifth-Generation Mesoscale Model (MM5), and Weather Research and Forecasting (WRF) models, which are commonly used for wind speed prediction within a range of 48-72 hours; solve complex numerical systems to derive various meteorological parameters. However, due to inherent model assumptions, these traditional methods often require extensive domain knowledge and manual parameter tuning, especially wind speed correction based on physical models relies on supercomputer support, thus limiting their real-time performance and prediction accuracy. Compared with traditional methods, artificial intelligence techniques provide more powerful tools by utilizing machine learning algorithms, especially deep learning techniques, to integrate ship motion information, sensor readings, and advanced models to achieve a more comprehensive understanding and simulation of the complexity of the marine environment; can autonomously learn the patterns of sensor errors and continuously improve over time. Models such as Recurrent Neural Network (RNN) and Gate Recurrent Unit (GRU) learn long-term dependencies and nonlinear relationships in time series data, enabling more accurate predictions. Convolutional Neural Network (CNN) uses different size convolution kernels to extract features of different scales, making it suitable for capturing the granularity characteristics of wind speed sequences. Some studies also attempt to handle the volatility of wind speed through hybrid models, such as combining CNN with RNN.

[0005] Although the existing wind speed prediction and correction methods show superior performance, there are still inherent defects that need to be improved: the RNN and GRU models require a large amount of time series data, and are calculated in time steps, resulting in low computational efficiency and high computational complexity; the RNN and GRU models are usually good at short-term prediction, but are prone to drift in long-term prediction, resulting in poor prediction results; the hybrid model requires selecting appropriate sub-model combinations to adapt to the complex and changing environment, increasing the complexity of the training and adjustment process. SUMMARY

[0006] To solve the above problems, the present application provides a ship side wind speed correction prediction method and system based on self-supervised contrast learning, which uses ultrasonic wind speed data to correct the left and right ship side wind speed, extracts different granularity feature vectors from the marine observation time series automatically without manual parameter adjustment or relying on domain expertise by constructing positive and negative sample pairs; has the ability to automatically learn the internal structure and pattern of the data, improves the perception of left and right ship side wind speed data errors, and enhances the correction ability.

[0007] According to some embodiments, the first aspect of the present application provides a ship side wind speed correction prediction method based on self-supervised contrast learning, which adopts the following technical solution:

[0008] A ship side wind speed correction prediction method based on self-supervised contrast learning, comprising:

[0009] Obtaining marine observation data;

[0010] Extracting data features of the obtained marine observation data;

[0011] Performing multi-granularity contrast learning on the extracted data features to obtain feature vectors of different granularity time series data of the marine observation data;

[0012] According to the obtained time series data feature vectors, the correction prediction of the ship side wind speed is completed.

[0013] As a further technical limitation, according to the obtained time series data feature vectors and a linear regression model, the prediction of the ship side wind speed is performed.

[0014] Further, based on the extraction of the time series data feature vectors, a linear regression model is trained, a ridge regression algorithm is used to reduce the overfitting of the model, and a cross-validation technique is used to select the optimal regularization parameter; the ship side wind speed is predicted by the trained linear regression model.

[0015] As a further technical limitation, in the process of extracting data features of the acquired marine observation data, the sine-cosine decomposition is adopted to smooth the wind direction fluctuation, the periodic component of the acquired marine observation data is extracted, and the high-frequency noise is removed to obtain the data features of the marine observation data.

[0016] As a further technical limitation, the multi-granularity contrast learning is to learn an unsupervised non-parametric function through an encoder, to aggregate the data features of the marine observation data of different granularities by applying a continuous max-pooling operation along the time axis through training the learned unsupervised non-parametric function.

[0017] As a further technical limitation, in each time window, the maximum value of all numerical values in the window is taken as the feature representation of the time window; the noise features and unimportant features in the time series are filtered out through max-pooling, and each time series is converted into a fixed-length vector, i.e., the feature vector of the different-granularity time series data of the marine observation data is obtained.

[0018] As a further technical limitation, the marine observation data is acquired by using a ship-borne sensor measurement technology and an ultrasonic wind speed measurement technology.

[0019] According to some embodiments, a second aspect of the present application provides a ship side wind speed correction prediction system based on self-supervised contrast learning, which adopts the following technical solution:

[0020] A ship side wind speed correction prediction system based on self-supervised contrast learning, comprising:

[0021] An acquisition module configured to acquire marine observation data;

[0022] An extraction module configured to extract data features of the acquired marine observation data;

[0023] A learning module configured to perform multi-granularity contrast learning on the extracted data features to obtain feature vectors of different-granularity time series data of the marine observation data;

[0024] A prediction module configured to complete correction prediction of the ship side wind speed according to the obtained time series data feature vectors.

[0025] According to some embodiments, a third aspect of the present application provides a computer-readable storage medium, which adopts the following technical solution:

[0026] A computer-readable storage medium having a program stored thereon, the program being executed by a processor to implement the steps in the ship side wind speed correction prediction method based on self-supervised contrast learning according to the first aspect of the present application.

[0027] In some embodiments, the fourth aspect of the present application provides an electronic device, which adopts the following technical scheme:

[0028] An electronic device includes a memory, a processor, and a program stored on the memory and running on the processor, and the processor implements the steps in the shipboard wind speed correction prediction method based on self-supervised contrast learning according to the first aspect of the present application when executing the program.

[0029] Compared with the prior art, the present application has the following beneficial effects:

[0030] The present application converts time series data into vector representation to achieve prediction, and exhibits excellent performance in long and short term prediction tasks on the basis of considering data continuity, and has lower computational complexity; the present application smoothes wind direction fluctuations by using sine and cosine decomposition, extracts periodic components, and at the same time eliminates high-frequency noise, to obtain better ocean observation data information and features, effectively enhances the feature extraction capability of the model, and improves the overall robustness; by performing multi-granularity contrast learning on ocean observation data, the influence of different prediction granularities is fully considered, which can effectively capture rich feature information under different granularities, help the model more comprehensively understand and utilize patterns and trends in time series data, achieve accurate prediction of wind speed, and significantly improve the prediction effect. BRIEF DESCRIPTION OF DRAWINGS

[0031] The accompanying drawings, which form a part of this implementation, are used to provide further understanding of this implementation, and the schematic embodiments of this implementation and the description thereof are used to explain this implementation, and do not constitute an improper limitation on this implementation.

[0032] Figure 1 A flowchart of the shipboard wind speed correction prediction method based on self-supervised contrast learning in the first embodiment of the present application;

[0033] Figure 2 A framework structure diagram of the shipboard wind speed correction prediction model in the first embodiment of the present application;

[0034] Fig. 3(a) is a comparison diagram of original wind speed and true wind speed on the left and right shipboard in the first embodiment of the present application;

[0035] Fig. 3(b) is a comparison diagram of corrected wind speed and true wind speed in the first embodiment of the present application;

[0036] Figure 4 A structure block diagram of the shipboard wind speed correction prediction system based on self-supervised contrast learning in the second embodiment of the present application. DETAILED DESCRIPTION

[0037] The present application will be further described below in combination with the drawings and embodiments.

[0038] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, 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 pertains.

[0039] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0040] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0041] In this invention, terms such as "fixed connection," "connected," and "linked" should be interpreted broadly, indicating a fixed connection, an integral connection, or a detachable connection; a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can determine the specific meaning of these terms in this invention based on the specific circumstances, and they should not be construed as limitations on the invention.

[0042] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0043] Example 1

[0044] Embodiment 1 of this invention introduces a ship hull wind speed correction and prediction method based on self-supervised comparative learning.

[0045] like Figure 1 The method shown is a ship hull wind speed correction prediction method based on self-supervised contrastive learning, including:

[0046] Acquire ocean observation data;

[0047] Extract the data features from the acquired ocean observation data;

[0048] Multi-granularity comparative learning is performed on the extracted data features to obtain feature vectors of time series data of different granularities in marine observation data;

[0049] Based on the feature vectors of the obtained time series data, the corrected prediction of the ship's side wind speed is completed.

[0050] The embodiment adopts the framework structure of the ship side wind speed correction prediction model as shown in Figure 2 , which includes a main feature selection module, an encoder module, and a multi-granularity comparison module.

[0051] The embodiment focuses on predicting the ship side wind speed in marine observation data, as shown in Figure 2 , the input marine observation data is processed by the main feature selection module to extract features closely related to ultrasonic wind intensity and implement wind direction fluctuation smoothing; context views are obtained using random cropping technology, and abstract representations are extracted from these context views through the self-encoder module; a multi-granularity comparison module is introduced to perform temporal comparison learning and instance comparison learning at different time granularities, strengthening the model's representation ability and ensuring effective capture of rich multi-scale information.

[0052] The embodiment uses the Pearson Correlation Coefficient (PCC) method to process marine observation data to extract high-correlation elements. PCC is defined as the ratio of the covariance between two variables to the product of their respective standard deviations, i.e.: where γ represents the Pearson correlation coefficient, -1≤γ≤1. X i and Y i represent two specific samples from X and Y, respectively, and and represent the mean of samples X and Y.

[0053] In shipboard wind speed measurement, there is a problem of severe fluctuations in wind direction between 360° and 0°. To solve these instantaneous and sharp changes while preserving the periodic trends in wind direction data, the embodiment decomposes the wind direction into left and right ship side components, obtaining four orthogonal components, i.e. The decomposed wind direction information is then fused with the strongly correlated features to form the input data for subsequent processing.

[0054] As one or more implementations, the goal of contrastive learning is to learn an unsupervised non-parametric function t = f(s) through an encoder. By training the encoder function f, the goal of the model is to make the representation of the input sample s in the embedding space closer to the representation of the positive sample, while keeping a relatively distant distance from the representation of the negative sample.

[0055] 1) Constructing positive and negative sample pairs: In contrastive learning, constructing positive and negative sample pairs is a challenging task. To avoid changing the size of the marine time series during construction, positive and negative sample pairs are generated by randomly cropping the input marine time series. For a given time series xi ∈R T×C Where T is the number of timestamps and C is the number of features, we perform a random cropping operation along the time axis, producing two overlapping context views [a1, b1] and [a2, b2]. We ensure that 0 < a1 ≤ a2 ≤ b1 ≤ b2 ≤ T, and that the overlapping segment [a2, b1] has a consistent representation in both views.

[0056] 2) Input Projection Layer: The input projection layer is a fully connected layer that projects the observation x at time stamp t. i，t Mapping to a high-dimensional latent vector k i，t Its purpose is to transform the original observation x i，t Transform it into a more meaningful and learnable vector representation.

[0057] 3) Timestamp mask: for the output vector k of the input projection layer i ={k i，t Apply a timestamp mask. Use a binary encoded vector q∈{0,1} T The masking location is determined following a Bernoulli distribution, where p = 0.5. The timestamp masking module randomly masks some time steps of the vector representation, forcing the model to use the remaining information for inference and encouraging the learning of more robust feature representations. By masking the latent vector at randomly selected timestamps, an enhanced contextual view can be generated. This masking operation enables the model to learn position-independent feature representations, prevents over-focusing on specific timestamps, and enhances the model's ability to generalize to unseen timestamps.

[0058] 4) Dilated Convolutional Neural Network: A dilated convolutional neural network consists of five residual blocks, each containing two one-dimensional convolutional layers, where the dilation parameter is denoted as l (the dilation parameter of the l-th block is 2). l This module is used to extract the contextual representation of each timestamp. Using dilated convolutions allows for a larger receptive field, better capturing long-range dependencies and contextual information in time-series data.

[0059] As one or more implementation methods. Figure 2 In the multi-granularity comparison module, each parallelogram represents a timestamp representation vector for an instance. When processing time series data, this embodiment employs a multi-granularity approach, aggregating data of different granularities by applying continuous max-pooling operations along the time axis. Specifically, within each time window, the maximum value of all values ​​within the window is taken, and these maximum values ​​are used as the feature representation of that time window; max-pooling filters out noise and unimportant features in the time series, retaining only the most representative features; each time series is converted into a fixed-length vector representation for subsequent prediction tasks; this reduces the sequence length while obtaining time series data representations with different temporal resolutions.

[0060] This embodiment utilizes the encoder to apply "temporal contrast learning" and "instance contrast learning" at different time granularity levels to capture different feature information of marine observation time series data at different time granularities:

[0061] 1) Temporal contrast learning: To capture the temporal trend of marine observation data over time, the model adopts a temporal contrast loss. We regard the representations k i，t and k′ i，t of the same timestamp in the two context views of the ith time series as positive sample pairs, while the representations from different timestamps are regarded as negative sample pairs. The temporal contrast loss is: where Ω represents the set of overlapping timestamps between the two views, k i，t′ and k′ i，t′ represent the context representations at the overlapping timestamp t′. I is an indicator function, set to 1 if t≠t′; otherwise set to 0.

[0062] 2) Instance contrast learning: To learn feature representations that can distinguish different instances, we choose the representations of different instances at the same timestamp t in the same batch as negative sample pairs. The instance contrast loss is: where B represents the batch size, k j，t and k′ j，t represent the context representations at the jth timestamp t in the same batch. Set to 1 if i≠j; otherwise set to 0.

[0063] 3) Multi-granularity contrast total loss: The above two loss calculation methods complement each other, enabling the model to capture the context information of time data in a multi-granularity form and generate fine-grained representations. Define the multi-granularity contrast total loss as: where N represents the total number of instances, and T represents the time step of each sample.

[0064] In time series data analysis, it is often necessary to predict time series data to better understand and predict future trends and patterns. In this embodiment, the time series data is converted into a feature vector representation, and then a linear regression model is used for prediction. An autoencoder-based neural network model is employed to effectively extract key features from the time series data. In the feature extraction stage, the time series data is mapped into a vector representation using an improved model; a linear regression model is trained using these feature vectors. During model training, the ridge regression algorithm is used to reduce overfitting, and cross-validation is combined to select the optimal regularization parameter. The trained linear regression model is used to predict the feature data of the test set, and the final prediction results are obtained. The prediction results are shown in Figures 3(a) and 3(b). Figure 3(a) shows the original port and starboard wind speeds and the actual wind speeds, as well as the error comparison between the original port and starboard wind speeds and the actual wind speeds; Figure 3(b) shows the corrected wind speeds and the actual wind speeds, as well as the error comparison between the corrected wind speeds and the actual wind speeds.

[0065] This embodiment transforms time-series data into vector representations for prediction. While considering data continuity, it demonstrates superior performance in both short- and long-term prediction tasks with low computational complexity. It utilizes sine and cosine decomposition to smooth wind direction fluctuations, extract periodic components, and simultaneously remove high-frequency noise to obtain higher-quality ocean observation data information and features, effectively enhancing the model's feature extraction capabilities and improving overall robustness. Through multi-granularity comparative learning of ocean observation data, it fully considers the impact of different prediction granularities, effectively capturing rich feature information at different granularities. This helps the model to more comprehensively understand and utilize patterns and trends in time-series data, achieving accurate wind speed prediction and significantly improving prediction performance.

[0066] Example 2

[0067] Embodiment 2 of the present invention introduces a ship hull wind speed correction and prediction system based on self-supervised comparative learning.

[0068] like Figure 4 The system shown is a ship hull wind speed correction prediction system based on self-supervised contrastive learning, comprising:

[0069] The acquisition module is configured to acquire ocean observation data;

[0070] The extraction module is configured to extract data features from the acquired ocean observation data;

[0071] The learning module is configured to perform multi-granularity comparative learning on the extracted data features to obtain feature vectors of time series data of different granularities of ocean observation data.

[0072] The prediction module is configured to complete the corrected prediction of the ship side wind speed according to the obtained time series data feature vector.

[0073] The detailed steps are the same as the ship side wind speed corrected prediction method based on self-supervised contrast learning provided in Embodiment One, and will not be repeated here.

[0074] Embodiment Three

[0075] The embodiment three of the present application provides a computer readable storage medium.

[0076] A computer readable storage medium, which stores a program, the program is executed by a processor to realize the steps in the ship side wind speed corrected prediction method based on self-supervised contrast learning as described in Embodiment One of the present application.

[0077] The detailed steps are the same as the ship side wind speed corrected prediction method based on self-supervised contrast learning provided in Embodiment One, and will not be repeated here.

[0078] Embodiment Four

[0079] The embodiment four of the present application provides an electronic device.

[0080] An electronic device, comprising a memory, a processor and a program stored in the memory and running on the processor, wherein the processor executes the program to realize the steps in the ship side wind speed corrected prediction method based on self-supervised contrast learning as described in Embodiment One of the present application.

[0081] The detailed steps are the same as the ship side wind speed corrected prediction method based on self-supervised contrast learning provided in Embodiment One, and will not be repeated here.

[0082] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for ship hull wind speed correction and prediction based on self-supervised contrastive learning, characterized in that, include: Acquire ocean observation data; Extract the data features from the acquired ocean observation data; Multi-granularity comparative learning is performed on the extracted data features to obtain feature vectors of time series data of different granularities in marine observation data; Based on the obtained time series data feature vector, the correction prediction of ship hull wind speed is completed, including: predicting ship hull wind speed based on the obtained time series data feature vector and linear regression model.

2. The ship hull wind speed correction and prediction method based on self-supervised contrastive learning as described in claim 1, characterized in that, A linear regression model is trained based on the feature vectors extracted from time series data. The ridge regression algorithm is used to reduce overfitting of the model, and the optimal regularization parameter is selected by combining cross-validation technology. The trained linear regression model is then used to predict the wind speed at the ship's side.

3. The ship hull wind speed correction and prediction method based on self-supervised contrastive learning as described in claim 1, characterized in that, In the process of extracting data features from the acquired marine observation data, sine and cosine decomposition is used to smooth wind direction fluctuations, extract the periodic components of the acquired marine observation data, and remove high-frequency noise to obtain the data features of the marine observation data.

4. The ship hull wind speed correction and prediction method based on self-supervised contrastive learning as described in claim 1, characterized in that, The multi-granularity contrastive learning method learns unsupervised nonparametric functions through an encoder, and then aggregates data features from ocean observation data of different granularities by applying continuous max pooling operations along the time axis after training the learned unsupervised nonparametric functions.

5. The ship hull wind speed correction and prediction method based on self-supervised contrastive learning as described in claim 1, characterized in that, Within each time window, the maximum value of all values ​​in the window is taken and used as the feature representation of that time window. Noise and unimportant features in the time series are filtered out by max pooling, and each time series is converted into a vector of fixed length, thus obtaining the feature vectors of time series data of different granularities of ocean observation data.

6. The ship hull wind speed correction and prediction method based on self-supervised contrastive learning as described in claim 1, characterized in that, Marine observation data are acquired using shipborne sensor measurement technology and ultrasonic wind speed measurement technology.

7. A ship hull wind speed correction and prediction system based on self-supervised contrastive learning, characterized in that, include: The acquisition module is configured to acquire ocean observation data; The extraction module is configured to extract data features from the acquired ocean observation data; The learning module is configured to perform multi-granularity comparative learning on the extracted data features to obtain feature vectors of time series data of different granularities of ocean observation data. The prediction module is configured to perform corrected prediction of ship hull wind speed based on the obtained time series data feature vector, including: predicting ship hull wind speed based on the obtained time series data feature vector and linear regression model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the ship hull wind speed correction prediction method based on self-supervised contrastive learning as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the ship hull wind speed correction and prediction method based on self-supervised contrastive learning as described in any one of claims 1-6.

Images