A method, system, medium, and apparatus for identifying an endo-wave in a taut mooring buoy

Abstract

The application belongs to the field of marine monitoring, and provides a tight anchor line buoy end internal wave identification method, system, medium and equipment, the technical scheme is based on the obtained underwater temperature chain historical data and the buoy attitude information at the corresponding moment to train the constructed internal wave identification model, wherein the construction process of the internal wave identification model includes: in the convolution layer, increase the sampling step, calculate the calculation result of each convolution kernel corresponding to each temperature chain historical data; combining the calculation result of each convolution kernel corresponding to the temperature chain historical data and the buoy attitude information at the corresponding moment, the attitude change of the buoy is corrected, the internal wave characteristics of the corrected buoy are classified to obtain the classification result; the trained internal wave identification model is pruned and quantized to obtain a light internal wave identification model; the light internal wave identification model is deployed to process the real-time collected temperature chain data, and the internal wave identification result is obtained. While ensuring the identification accuracy, the calculation power consumption is reduced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of ocean monitoring, and particularly relates to a method and system for identifying internal waves at the end of a taut mooring buoy. BACKGROUND

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

[0003] In the field of ocean monitoring, accurate identification of internal waves based on buoy ends is crucial, as it plays a key role in various applications such as ocean environment monitoring, ocean engineering, and ocean security. By deploying an internal wave identification system at the buoy end, not only can data transmission delays be reduced and reliance on remote servers be reduced, but data security and real-time performance can also be improved, thereby providing more accurate and reliable technical support for ocean research and applications.

[0004] Currently, the commonly used internal wave identification method at the edge end mostly adopts the threshold method, but it is limited by rule setting and is difficult to adapt to complex and changing ocean environments, resulting in insufficient identification accuracy. Although the introduction of artificial intelligence methods can improve identification accuracy, the large amount of neural network inference calculation and high power consumption become major challenges for buoy end applications. At the same time, fixed buoys represented by taut mooring buoys have their anchor lines in a high tension state, which, compared to slack mooring buoys, has less attitude change under flow field disturbance, and can improve measurement stability to a certain extent. However, even taut mooring buoys may experience slight tilting, vibration, or attitude drift when affected by external factors such as ocean currents, wind, and waves. These changes can cause the measurement depth and angle of the temperature chain (temperature profile chain) sensor to shift, thereby affecting the accuracy of internal wave identification. SUMMARY

[0005] To solve at least one of the technical problems in the background art, the present application provides a method and system for identifying internal waves at the end of a taut mooring buoy, which aims to reduce computational power consumption while ensuring identification accuracy.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0007] The first aspect of the present application provides a method for identifying internal waves at the end of a taut mooring buoy, comprising the following steps:

[0008] Obtain historical data of underwater temperature chains and buoy attitude information at the corresponding time;

[0009] Train the constructed internal wave identification model based on the obtained historical data of underwater temperature chains and buoy attitude information at the corresponding time; wherein the construction process of the internal wave identification model comprises:

[0010] In the convolutional layer, the sampling stride is increased, and the calculation result of each convolutional kernel is calculated for each historical data of the temperature chain.

[0011] By combining the calculation results of each convolution kernel corresponding to the historical data of the temperature chain and the buoy attitude information at the corresponding time, the attitude change of the buoy is corrected to obtain the corrected buoy end internal wave characteristics. Based on the corrected buoy end internal wave characteristics, classification is performed to obtain the classification results.

[0012] The trained internal wave recognition model is pruned and quantized to obtain a lightweight internal wave recognition model.

[0013] A lightweight internal wave identification model is deployed, and the real-time temperature chain data is processed based on the deployed internal wave identification model to obtain the internal wave identification results.

[0014] Furthermore, the calculation formula for each convolutional kernel corresponding to each historical data point of the temperature chain is as follows:

[0015] ,

[0016] in, Indicates the first The first original data corresponding to the Strip convolution output, Representing the The first original data corresponding to the Strip convolution kernel, This represents the number of bits in the convolution output. The sampling stride represents the convolution operation. The formula for calculating representative length. The weight index represents the convolution kernel.

[0017] Furthermore, by combining the calculation results of each convolution kernel corresponding to the historical data of the temperature chain and the buoy attitude information at the corresponding time, when correcting the buoy attitude change, a fixed time window is defined for each feature data point. The attitude data within the window will be corrected according to the weighting strategy, and the weighted average of the attitude data within the window will be calculated to correct the feature data point.

[0018] Furthermore, classification results are obtained based on the corrected internal wave characteristics of the buoy end, including:

[0019] The modified buoy end internal wave characteristics are processed by a multilayer perceptron to generate an attitude guidance vector.

[0020] The modified buoy end internal wave characteristics and attitude guidance vector are multiplied element by element, and then the attention weights are calculated using the Sigmoid activation function.

[0021] The modified buoy end internal wave feature and attention weight are element-wise weighted to obtain a weighted feature vector;

[0022] The weighted feature vector is input into a hidden fully connected layer and outputs a final classification probability through a Softmax function.

[0023] Further, the pruning and quantization of the trained internal wave recognition model to obtain a lightweight internal wave recognition model comprises:

[0024] The connections with small contribution weight to the network output are pruned, an L1 norm-based pruning method is used to analyze the sparsity of the weight, the L1 norm of each weight is calculated, and the weights below a set threshold are pruned;

[0025] The input underwater temperature chain data, internal wave recognition model activation value and internal wave recognition model weight value are quantized, and the input value of the dequantization activation function is dequantized; the internal wave recognition model is activated after dequantization to obtain a lightweight internal wave recognition model.

[0026] Further, the quantization formula of the internal wave recognition model weight value is:

[0027] ,

[0028] wherein, represents the quantization bit width, represents the scaling factor, represents the quantized integer weight value, represents the original weight value, represents the minimum value of the weight value, represents the maximum value of the weight value.

[0029] Further, the deployment of the lightweight internal wave recognition model comprises:

[0030] The producer task collects underwater temperature chain data by communicating with the buoy end controller, writes the data into a shared queue after integrity check and preprocessing;

[0031] The producer task runs according to a fixed sampling period, and the response order of data processing is determined according to the priority level of the consumer task;

[0032] When the queue data is full, an interrupt signal is triggered, and the consumer

[0033] The second aspect of the present application provides a tight anchor line buoy end internal wave recognition device, comprising:

[0034] A data acquisition module is configured to acquire underwater temperature chain historical data and buoy attitude information at a corresponding time.

[0035] An internal wave identification model training module is configured to train the constructed internal wave identification model based on the obtained underwater temperature chain historical data and the buoy attitude information at the corresponding time; wherein the construction process of the internal wave identification model comprises:

[0036] In the convolution layer, the sampling step is increased, and the calculation result of each convolution kernel corresponding to each temperature chain historical data is calculated;

[0037] The attitude change of the buoy is corrected to obtain the corrected buoy end internal wave feature by combining the calculation result of each convolution kernel corresponding to the temperature chain historical data and the buoy attitude information at the corresponding time, and the classification result is obtained based on the corrected buoy end internal wave feature;

[0038] A light module is configured to prune and quantize the trained internal wave identification model to obtain a light internal wave identification model;

[0039] A model deployment module is configured to deploy the light internal wave identification model, process the real-time collected temperature chain data based on the deployed internal wave identification model, and obtain the internal wave identification result.

[0040] The third aspect of the present application provides a computer readable storage medium.

[0041] A computer readable storage medium has a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned internal wave identification method for a buoy end of a tight mooring system.

[0042] The fourth aspect of the present application provides a computer device.

[0043] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-mentioned internal wave identification method for a buoy end of a tight mooring system.

[0044] Compared with the prior art, the present application has the following advantages:

[0045] 1、The internal wave identification method for a buoy end of a tight mooring system introduced in the present application reduces the influence of buoy attitude change on temperature chain measurement data by introducing an attitude correction module in the identification model, and realizes the compression and acceleration of neural networks through model quantization, pruning and hardware optimization, so that it can run efficiently on a low-power single-chip microcomputer, solving the problems of insufficient precision and high power consumption in traditional technology. The artificial intelligence method using neural networks for internal wave identification significantly improves the identification accuracy and reduces misidentification compared with the traditional threshold method, and performs better in complex marine environments.

[0046] 2、The feature classification module of the present application introduces an attention mechanism based on a posture auxiliary vector, and adaptively adjusts the importance distribution of the features, unlike the static correction of the buoy end internal wave features in the feature extraction module, the feature classification module dynamically guides the feature attention direction in the classification stage, further refines the compensation of the posture influence, and significantly improves the stability of the system in complex sea conditions.

[0047] 3、The present application reduces power consumption in the neural network inference process through model quantization, pruning and hardware optimization, and adapts to low-power embedded platforms.

[0048] The advantages of the additional aspects of the present application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0049] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description of the exemplary embodiments of the application given below, serve to explain the application, and do not constitute an improper limitation of the application.

[0050] Figure 1 is a tight mooring buoy end internal wave recognition method flow chart provided by the embodiment of the present application;

[0051] Figure 2 is a feature extraction module flow chart provided by the embodiment of the present application;

[0052] Figure 3 is a feature classification module flow chart provided by the embodiment of the present application;

[0053] Figure 4 is a model training, pruning and quantization flow chart provided by the embodiment of the present application;

[0054] Figure 5 is a producer-consumer model flow chart provided by the embodiment of the present application. DETAILED DESCRIPTION

[0055] The present application will be further described below in conjunction with the drawings and embodiments.

[0056] It should be pointed out that the following detailed description is all exemplary, and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0057] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0058] As mentioned in the background, the edge end commonly used internal wave recognition method at present adopts threshold method, but it is limited by rule setting and is difficult to adapt to complex and changeable marine environment, resulting in insufficient recognition accuracy. Although the introduction of artificial intelligence method can improve the recognition accuracy, due to the large amount of neural network inference calculation, the high power consumption problem becomes the main challenge of the buoy end application. At the same time, the fixed buoy represented by the tight anchor buoy has the anchor cable in the high tension state, compared with the slack anchor buoy, the attitude change under the flow field disturbance is smaller, which can improve the measurement stability to a certain extent. However, even the tight anchor buoy, when affected by external factors such as sea current, wind and wave, may still have slight tilt, vibration or attitude drift, which will cause the measurement depth and angle of the temperature chain sensor to deviate, thereby affecting the accuracy of internal wave recognition.

[0059] In order to make the system suitable for the buoy end with limited computing resources, the present application realizes the compression and acceleration of neural network through model quantization, pruning and hardware optimization, so that it can run efficiently on the single-chip microcomputer with low power consumption. In addition, the system integrates the real-time operating system (RTOS) to optimize task management, improve computing efficiency and further reduce power consumption. The whole system is an independent module, which directly interacts with the buoy end controller to realize high-precision, low-power internal wave recognition capability and improve the adaptability of the system in complex sea conditions. At the same time, a posture correction module is introduced in the recognition model to reduce the influence of buoy attitude change on temperature chain measurement data, thereby improving the stability of data and the accuracy of recognition.

[0060] Embodiment one

[0061] The present embodiment provides a tight anchor buoy end internal wave recognition method, which comprises the following steps:

[0062] Step 1: Obtain the historical data of underwater temperature chain and the buoy attitude information at the corresponding time;

[0063] The historical data of the lower temperature chain and the buoy attitude information at the corresponding time in the present embodiment are obtained at the buoy end; according to the change rule of vertical distribution of water temperature, finally the underwater temperature sequence of 14 layers of sensors and 30 sampling time points is selected as the input of feature extraction layer.

[0064] Since the temperature change within the ocean is usually a slow and continuous process, especially between adjacent time points, the temperature gradient change trend is highly consistent, so there is significant correlation in the time dimension of the temperature data. The Pearson Correlation Coefficient between adjacent time points is actually calculated:

[0065] ,

[0066] wherein, and represent the temperature vectors of each depth layer at the t th time point and the t +1th time point, is the covariance function, and are the temperature standard deviations at the t th time point and the t +1th time point;

[0067] The results show that in most time points, the temperature distribution between adjacent time points has strong correlation.

[0068] Step 2: Training the constructed internal wave recognition model based on the obtained underwater temperature chain historical data and the corresponding time point buoy attitude information;

[0069] The construction process of the internal wave recognition model specifically includes the following steps:

[0070] Step 201, extracting the internal wave features of the buoy end based on the constructed feature extraction model;

[0071] As shown in Figure 2 , the feature extraction model uses a convolutional neural network CNN, including an input layer, a convolution kernel, and a convolution output;

[0072] The specific feature extraction process includes the following steps:

[0073] Step 2011, inputting the ocean temperature profile data having strong correlation between adjacent time points into the input layer;

[0074] Step 2012, in the convolution layer, increasing the sampling step , calculating the calculation result of each convolution kernel corresponding to each data;

[0075] The input ocean temperature profile data has strong correlation between adjacent time points, which indicates that the temperature change trend of consecutive depth layers or time points is similar. Therefore, appropriately increasing the sampling step in the convolution operation has little effect on the algorithm performance.

[0076] Specifically, the first The original data and the corresponding first The convolution kernel convolution calculation formula is as follows:

[0077] (1),

[0078] wherein, The first The first The first The first The first The first The number of bits of the convolution output, The sampling step of the convolution operation, The length calculation formula, The weight index of the convolution kernel.

[0079] According to formula (1), the relationship between the input feature number The output feature number And the sampling step :

[0080] (2),

[0081] This shows that increasing the step The output feature number But because the correlation between adjacent time data is strong, it has little effect on feature extraction.

[0082] Step 2013, output the extracted The buoy end internal wave features.

[0083] Step 202: Based on the obtained IMU sensor data on the buoy and the extracted buoy end internal wave features, the attitude change of the buoy is corrected to obtain the corrected buoy end internal wave features, and based on the corrected buoy end internal wave features, the classification result is obtained.

[0084] In order to improve the internal wave recognition ability of the tight anchor buoy under the disturbance of sea current, wind wave and other disturbances, and handle the influence of the buoy attitude change on the underwater temperature chain sensor data, an improved attitude correction method is proposed to ensure the accuracy and reliability of the water temperature data. After the convolution feature extraction, the buoy end internal wave features do not consider the deviation of the temperature data caused by the buoy attitude change (such as inclination, yaw, etc.), so

[0085] The attitude correction module is added to correct the deviation of the temperature data caused by the buoy attitude change (such as inclination, yaw, etc.).

[0086] Specifically, the steps include the following:

[0087] Step 2021: Use a time window to perform weighted correction on the attitude data to obtain the corrected buoy end internal wave characteristics;

[0088] The attitude correction module's input includes IMU sensor data (pitch angle) from the buoy. Roll angle Heading angle The data includes features derived from the original underwater temperature chain data through convolutional feature extraction. Since the buoy attitude data is sampled at a high frequency, and the temperature feature data may decrease in volume during extraction, a time window is used to weight and correct the attitude data. Specifically, for each feature data point, a fixed time window is defined. The attitude data within the window is corrected according to a certain weighting strategy, and the weighted average of the attitude data within the window is calculated to correct the feature data point.

[0089] Specifically, the feature is corrected based on the buoy's attitude change using a weighted correction method. The formula for calculating the corrected feature is as follows:

[0090] (3),

[0091] in, The corrected internal wave characteristics of the buoy end. This is a correction factor, representing the degree to which changes in buoy attitude affect the characteristics. These are the pitch angle, roll angle, and heading angle at the corresponding moment of the output feature, respectively. The features are adjusted by weighting the attitude data, thereby correcting the influence of attitude changes on the feature values.

[0092] Step 2022: Obtain the classification result based on the corrected buoy end internal wave characteristics and the constructed feature classification module;

[0093] like Figure 3 As shown, the feature classification module consists of a two-layer fully connected neural network (FCNN) and a pose-guided attention module.

[0094] First, the feature classification input layer receives the buoy end internal wave features, which have been processed by the feature extraction module and undergone attitude correction. These features are vectors with a shape of 1×40. ;

[0095] Subsequently, the features are input to the pose-guided attention module, where the original pose angles are first processed by a small multilayer perceptron (MLP). Perform a dimensionality increase operation to generate a 1×40 attitude guidance vector. This process can be represented as:

[0096] ,

[0097] Then, the feature vector F is multiplied element-wise with the pose guidance vector g (Hadamard product), and the attention weight w is calculated by the Sigmoid activation function, as follows:

[0098] ,

[0099] where σ(⋅) denotes the Sigmoid function, and ⊙ denotes the element-wise multiplication operation.

[0100] Next, the input feature vector F is weighted element-wise with the attention weight w to obtain the weighted feature vector :

[0101] ,

[0102] The weighted feature vector is input into a hidden fully connected layer (FCNN) with a size of 1 × 32 to further extract deep features. Subsequently, the features are passed through an output fully connected layer with a size of 1 × 2, and the final classification probability is output by the Softmax function:

[0103] ,

[0104] where , , and b1, b2 are the weight matrices and bias terms of the hidden layer and the output layer, respectively.

[0105] The feature classification module adaptively adjusts the importance distribution of the features by introducing an attention mechanism based on the pose auxiliary vector. Specifically, the pose information is fused with the feature extraction results to calculate the attention weight, which is used to weight and correct the original features. This method enables the model to highlight more discriminative features under pose changes, improving the classification accuracy and robustness. Unlike the static correction of the buoy end internal wave features in the feature extraction module, the feature classification module dynamically guides the feature attention direction during the classification stage, further refining the compensation of the pose influence. The combination of the two achieves a transition from static compensation to dynamic adaptation, significantly improving the stability of the system in complex sea conditions.

[0106] The fully connected neural network itself does not have independent feature extraction capability. Its main function is to realize feature classification by performing nonlinear combination on the input features. In this model, a Softmax layer is added to the output layer to calculate the belonging probability of each class, achieving the purpose of classification.

[0107] The specific form of Softmax is:

[0108] (4),

[0109] in, Represents the output of the neural network. These represent two nodes in the output layer. Classification network output. for The matrix, These represent the probabilities of identifying it as not an internal wave and the probabilities of identifying it as an internal wave, respectively.

[0110] The final classification result is determined by setting a threshold. The determination will be made according to the following rules:

[0111] (5),

[0112] Based on underwater temperature chain data, a neural network model is used for internal wave identification. At the same time, the deviation caused by changes in buoy attitude (such as tilting, yaw, etc.) on the extracted feature values ​​is corrected, thereby improving the accuracy of the final internal wave detection.

[0113] Step 3: Prune and quantize the trained internal wave recognition model to obtain a lightweight internal wave recognition model;

[0114] like Figure 4 As shown, to meet the needs of the microcontroller platform, the trained model is pruned and quantized. By removing redundant parts and reducing the parameter bit width, the resource consumption and computational efficiency of the model are greatly optimized, enabling it to run efficiently on embedded platforms such as microcontrollers while maintaining a certain level of inference accuracy.

[0115] Step 301: Prune connections with small contribution weights to the network output;

[0116] During the pruning process, for weights that contribute little to the network output, an L1-norm-based pruning method is used to perform sparsity analysis on the weights. By reducing connections with smaller weight values, the neural network structure becomes sparser while effectively maintaining model accuracy.

[0117] Specifically, L1 norm pruning calculates the L1 norm of each weight and removes weights below a set threshold, thereby eliminating redundant connections, optimizing the model structure, and improving runtime efficiency. The L1 norm formula is as follows:

[0118] (6),

[0119] in, Represents the weight matrix The L1 norm of the sum of the absolute values ​​of all weights. Represents the weight matrix The Middle Each weight value.

[0120] Step 302, quantizing input values and weight values of the model;

[0121] In the quantization process, due to the limitation of the running resources and memory of the single-chip platform, directly using 32-bit floating-point numbers for operation and storage will significantly increase the time and energy consumption, and even may not be able to execute. This is contrary to the original intention of high efficiency and low power consumption of the buoy end processing. Therefore, the quantization process of the single-chip platform of the system is essentially the process of quantizing 32-bit floating-point numbers to 8-bit integers. By running data in the form of 8-bit integers, the storage space can be reduced by about 75%, and the inference speed is also significantly improved.

[0122] The quantization of the buoy end internal wave recognition model includes the quantization of the input underwater temperature chain data and the model activation value, and the quantization of the model weight value. This design effectively balances the accuracy and efficiency requirements of the model.

[0123] The quantization formula is as follows, representing 8-bit width, , respectively, represent real-time temperature chain data, activation value and model weight value:

[0124] (7),

[0125] wherein, represents the quantization bit width (8 in this embodiment, representing that the quantization target is 8-bit integer), represents the scaling factor, represents the quantized integer weight value, represents the original weight value, represents the minimum value of the weight value, represents the maximum value of the weight value.

[0126] Step 303, dequantizing the input value of the activation function;

[0127] In particular, before the model performs the activation operation, the input value of the activation function is an integer value after quantization processing, which lacks direct mathematical meaning. In order to ensure that the activation function can correctly perform nonlinear transformation, a short dequantization operation is needed to restore the quantized value to the floating-point domain before applying the activation function, and finally the floating-point output is obtained. This process is a necessary step for the single-chip platform to balance between accuracy and efficiency.

[0128] The input value dequantization formula of the activation function is as follows:

[0129] (8),

[0130] wherein, represents the dequantized floating-point value.

[0131] Through model quantization, pruning and hardware optimization, the power consumption in the neural network inference process is reduced to adapt to low-power embedded platforms. Combined with real-time operating system (RTOS) optimization task management, the system efficiency and real-time data processing capability are improved, and the overall power consumption is effectively reduced.

[0132] Step 4: Deploy the lightweight internal wave identification model, and process the real-time collected temperature chain data based on the deployed internal wave identification model to obtain the internal wave identification result.

[0133] As shown in Figure 5 The system is based on the internal wave identification of the buoy end of the single-chip microcomputer, combined with the real-time operating system (RTOS) and the producer-consumer model, and constructs an efficient real-time data processing flow.

[0134] Specifically, the following steps are included:

[0135] Step 401, the producer task collects underwater temperature chain data by communicating with the buoy end controller, and writes the data into the shared queue after integrity check and preprocessing;

[0136] Step 402, the producer task runs according to a fixed sampling period, and determines the response order of data processing according to the priority level of the consumer task;

[0137] In order to improve the real-time performance, the producer task runs according to a fixed sampling period, and the consumer task has a high priority and can respond to data processing immediately after the queue full interrupt is triggered.

[0138] Step 403, when the queue data is full, an interrupt signal is triggered, the consumer task reads the data from the queue and inputs the internal wave identification model for inference, and then feeds back the identification result to the buoy controller for storage and forwarding.

[0139] Through the producer-consumer model and the queue interrupt mechanism, the data transmission and processing flow are optimized, the stability and efficiency of data processing are improved, the identification result is quickly fed back, and high-precision and low-power internal wave identification is realized.

[0140] Embodiment Two

[0141] The embodiment provides a tight anchoring buoy end internal wave identification system, comprising:

[0142] A data acquisition module is configured to acquire underwater temperature chain historical data and buoy attitude information at a corresponding time;

[0143] An internal wave identification model training module is configured to train the constructed internal wave identification model based on the acquired underwater temperature chain historical data and buoy attitude information at a corresponding time; wherein the construction process of the internal wave identification model comprises:

[0144] In the convolution layer, the sampling step is increased, and the calculation result of each convolution kernel corresponding to each piece of temperature chain historical data is calculated;

[0145] The posture change of the buoy is corrected by combining the calculation result of each convolution kernel corresponding to the temperature chain historical data and the buoy posture information at the corresponding moment to obtain a corrected buoy internal wave feature, and classification is performed based on the corrected buoy internal wave feature to obtain a classification result;

[0146] A light module is used to prune and quantize the trained internal wave recognition model to obtain a light internal wave recognition model;

[0147] A model deployment module is used to deploy the light internal wave recognition model, and the temperature chain data collected in real time is processed based on the deployed internal wave recognition model to obtain an internal wave recognition result.

[0148] It should be noted that the specific implementation of the tight mooring buoy internal wave recognition system of the embodiment of the present application is similar to that of the tight mooring buoy internal wave recognition method of the embodiment of the present application, and specific reference can be made to the description in the method part. In order to reduce redundancy, this part will not be repeated.

[0149] Embodiment three

[0150] The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps in the tight mooring buoy internal wave recognition method.

[0151] Embodiment four

[0152] The embodiment provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor executes the program to realize the steps in the tight mooring buoy internal wave recognition method.

[0153] Embodiment five

[0154] The embodiment provides a program product, which is a computer program product, including a computer program, and the computer program is executed by a processor to realize the steps in the tight mooring buoy internal wave recognition method.

[0155] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. 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 identifying the internal wave at the end of a taut moored buoy, characterized in that, include: Acquire historical underwater temperature chain data and corresponding buoy attitude information at the corresponding time; The internal wave recognition model is trained based on the acquired historical underwater temperature chain data and the corresponding buoy attitude information; the construction process of the internal wave recognition model includes: In the convolutional layer, the sampling stride is increased, and the calculation result of each convolutional kernel is calculated for each historical data of the temperature chain. By combining the calculation results of each convolution kernel corresponding to the historical data of the temperature chain and the buoy attitude information at the corresponding time, the attitude change of the buoy is corrected to obtain the corrected buoy end internal wave characteristics. Based on the corrected buoy end internal wave characteristics, classification is performed to obtain the classification results. The trained internal wave recognition model is pruned and quantized to obtain a lightweight internal wave recognition model. A lightweight internal wave identification model is deployed, and the real-time temperature chain data is processed based on the deployed internal wave identification model to obtain the internal wave identification results.

2. The method for identifying the internal wave at the end of a taut moored buoy as described in claim 1, characterized in that, The calculation formula for each convolutional kernel corresponding to each historical data point of the temperature chain is: ， in, Indicates the first The first original data corresponding to the Strip convolution output, Representing the The first original data corresponding to the Strip convolution kernel, This represents the number of bits in the convolution output. The sampling stride represents the convolution operation. The formula for calculating representative length. The weight index represents the convolution kernel.

3. The method for identifying the internal wave at the end of a taut moored buoy as described in claim 1, characterized in that, By combining the calculation results of each convolution kernel corresponding to the historical data of the temperature chain and the buoy attitude information at the corresponding time, when correcting the buoy attitude change, a fixed time window is defined for each feature data point. The attitude data within the window will be corrected according to the weighting strategy. The weighted average of the attitude data within the window is calculated to correct the feature data point.

4. The method for identifying the internal wave at the end of a taut moored buoy as described in claim 1, characterized in that, The classification results are obtained based on the corrected internal wave characteristics of the buoy end, including: The modified buoy end internal wave characteristics are processed by a multilayer perceptron to generate an attitude guidance vector. The modified buoy end internal wave characteristics and attitude guidance vector are multiplied element by element, and then the attention weights are calculated using the Sigmoid activation function. The modified buoy end internal wave characteristics and attention weights are weighted element by element to obtain a weighted feature vector; The weighted feature vectors are then input into a hidden fully connected layer, and the final classification probability is output through the Softmax function.

5. The method for identifying the internal wave at the end of a taut moored buoy as described in claim 1, characterized in that, The step of pruning and quantizing the trained internal wave recognition model to obtain a lightweight internal wave recognition model includes: Connections with small contribution weights to the network output are pruned. The L1 norm-based pruning method is used to perform sparsity analysis on the weights, calculate the L1 norm of each weight, and remove weights that are below a set threshold. The input underwater temperature chain data, the activation value of the internal wave identification model, and the weight value of the internal wave identification model are quantized and dequantized to obtain the input value of the activation function. The dequantized internal wave identification model is then activated to obtain a lightweight internal wave identification model.

6. The method for identifying the internal wave at the end of a taut moored buoy as described in claim 2, characterized in that, The quantization formula for the weight values ​​of the internal wave identification model is: ， in, Indicates the quantization bit width. Indicates the scaling factor. This represents the quantized integer weight value. This represents the original weight values. This represents the minimum value of the weight. This represents the maximum value of the weight.

7. The method for identifying the internal wave at the end of a taut moored buoy as described in claim 1, characterized in that, The process of deploying a lightweight internal wave identification model and processing the real-time acquired temperature chain data based on the deployed internal wave identification model includes: The producer task collects underwater temperature chain data by communicating with the buoy end controller, and after integrity verification and preprocessing, writes the data into a shared queue. The producer task runs according to a fixed sampling period, and the response order of data processing is determined according to the priority level of the consumer task. When the queue is full, an interrupt signal is triggered. The consumer task then reads the data from the queue and inputs it into the internal wave recognition model for inference, thus obtaining the internal wave recognition result.

8. A tensioned mooring buoy end internal wave identification system, characterized in that, include: The data acquisition module is used to acquire historical underwater temperature chain data and buoy attitude information at the corresponding time. The internal wave recognition model training module is used to train the constructed internal wave recognition model based on the acquired underwater temperature chain historical data and the buoy attitude information at the corresponding time. The construction process of the internal wave recognition model includes: In the convolutional layer, the sampling stride is increased, and the calculation result of each convolutional kernel is calculated for each historical data of the temperature chain. By combining the calculation results of each convolution kernel corresponding to the historical data of the temperature chain and the buoy attitude information at the corresponding time, the attitude change of the buoy is corrected to obtain the corrected buoy end internal wave characteristics. Based on the corrected buoy end internal wave characteristics, classification is performed to obtain the classification results. Lightweight module, which is used to prune and quantize the trained internal wave recognition model to obtain a lightweight internal wave recognition model; The model deployment module is used to deploy a lightweight internal wave identification model, and then process the real-time acquired temperature chain data based on the deployed internal wave identification model to obtain the internal wave identification result.

9. 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 in the method for identifying the internal wave at the end of a taut moored buoy as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for identifying the internal wave at the end of a taut moored buoy as described in any one of claims 1-7.

Images