Underwater thruster control method and related equipment based on multi-parameter fusion

Abstract

The present application provides a method and related equipment for underwater propeller control based on multi-parameter fusion, which improves the autonomous decision-making ability and adaptability of underwater propellers in different environments, realizes more intelligent and efficient control, and thus helps to improve the overall stability of underwater propellers. The method of the present application includes: collecting water flow velocity and regional terrain information of the target area where the underwater propeller is located; establishing a three-dimensional flow field real-time perception model based on the water flow velocity and the regional terrain information, and obtaining multimodal data according to the three-dimensional flow field real-time perception model, the multimodal data including water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates and terrain height of the target area; dispersing and filtering the multimodal data and then fusing them to obtain target multimodal data; constructing a multimodal neural network model based on the target multimodal data and through a temporal attention mechanism.

Description

Technical Field

[0001] The present application relates to the technical field of underwater thrusters, and in particular to an underwater thruster control method based on multi-parameter fusion and related equipment. Background Art

[0002] An underwater thruster is a device that provides propulsion and steering for underwater vehicles. It is a core component for underwater operations and navigation. It is widely used in ocean exploration, underwater operations, submarine navigation and other fields, and is of great significance in marine resource development, environmental monitoring and other aspects.

[0003] During the actual operation of the underwater thruster, the underwater environment is complex and the water flow speed, direction, and density are variable. Therefore, when the underwater thruster is working, it is necessary to continuously collect underwater environmental information in order to continuously adjust the control strategy according to the collected environmental information.

[0004] However, existing traditional underwater thrusters generally use single data or simple empirical strategies to control the operation of the underwater thrusters. Due to the complex underwater environment, single data cannot fully reflect the coupling factors such as water flow, pressure, terrain, etc. in the area where the underwater thruster is located. Therefore, the real environment of the area where the underwater thruster is located cannot be fully obtained, resulting in the inability to control the underwater thruster in real time according to environmental data. As a result, the underwater thruster may collide with obstacles due to inaccurate power output or recognition, and the overall stability is low. Summary of the Invention

[0005] The present application provides an underwater thruster control method and related equipment based on multi-parameter fusion to solve the above technical problems.

[0006] In a first aspect, the present application provides a method for controlling an underwater thruster based on multi-parameter fusion, the method comprising:

[0007] Collect water flow velocity and regional terrain information in the target area where the underwater propeller is located;

[0008] Establishing a three-dimensional flow field real-time perception model based on the water flow velocity and the regional terrain information, and acquiring multimodal data according to the three-dimensional flow field real-time perception model, wherein the multimodal data includes water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates, and terrain height of the target area;

[0009] Distributing and filtering the multimodal data and then fusing them to obtain target multimodal data;

[0010] Based on the target multimodal data, a multimodal neural network model is constructed through a temporal attention mechanism;

[0011] Inputting real-time multimodal data into the multimodal neural network model and generating control instructions using a two-layer reinforcement learning strategy;

[0012] generating power parameters of the underwater thruster according to the control instruction;

[0013] When it is determined that the power parameter is within a preset parameter range, the underwater propeller is controlled to move according to the power parameter.

[0014] Optionally, the multimodal data is dispersedly filtered and then fused to obtain target multimodal data, including:

[0015] Preprocessing the multimodal data to obtain preprocessed data;

[0016] extracting water flow data and terrain data from the pre-processed data;

[0017] According to the data characteristics of water flow data and terrain data, Kalman filter algorithm and Gaussian filter algorithm are used to eliminate outliers respectively;

[0018] Perform data registration on the filtered water flow data and terrain data;

[0019] The weighted average algorithm is used to fuse the data after data registration;

[0020] The weighted average algorithm is:

[0021]

[0022] Where S is the weighted average, n is the number of data points, and x i is the value of the i-th data point, w i is the weight of the i-th data point;

[0023] The compared fused data is verified, and when it is determined that the data verification passes, the target multimodal data is output after the data is smoothed.

[0024] Optionally, a multimodal neural network model is constructed based on the target multimodal data and through a temporal attention mechanism, including:

[0025] Extracting numerical, spatial, and temporal features from the target multimodal data and converting them into a unified dimensional vector;

[0026] Arranging the unified dimension vectors in time series to obtain a feature matrix;

[0027] According to the data characteristics of the feature matrix, a multimodal neural network model is constructed using a temporal attention mechanism.

[0028] Optionally, real-time multimodal data is input into the multimodal neural network model, and a two-layer reinforcement learning strategy is used to generate control instructions, including:

[0029] Acquire real-time multimodal data;

[0030] Preprocessing the real-time multimodal data to form an input feature vector adapted to the multimodal neural network model;

[0031] Inputting the input feature vector into the multimodal neural network model to obtain a target feature vector;

[0032] Building a two-layer reinforcement learning strategy network, the two-layer reinforcement learning strategy network includes a high-level decision network and a low-level execution network, the high-level decision network is used to receive the target feature vector output by the multimodal neural network model and plan the control strategy, and the low-level execution network is used to refine the control strategy into control parameters;

[0033] Inputting the target feature vector into the two-layer reinforcement learning policy network and updating the policy network parameters using the Actor-Critic algorithm;

[0034] A control instruction is generated based on the policy network parameters and in combination with the current environment state.

[0035] Optionally, generating the power parameters of the underwater propulsion device according to the control instruction includes:

[0036] Calculating the output power of the underwater thruster according to the control instruction;

[0037] comparing the output power of the underwater propeller with the maximum output power of the underwater propeller and obtaining a comparison result;

[0038] The power parameters of the underwater propeller are generated according to the comparison result.

[0039] Optionally, when it is determined that the power parameter is within a preset parameter range, controlling the underwater propeller to move according to the power parameter includes:

[0040] Analyze the power parameters and verify whether the power parameters are complete;

[0041] If so, the parsed dynamic parameters are compared with the preset parameter range;

[0042] When it is determined that the power parameter is within a preset parameter range, converting the power parameter into a control signal executable by the underwater propulsion device;

[0043] The underwater propeller is controlled to move according to the control signal.

[0044] Optionally, after comparing the parsed power parameters with the preset parameter range, the method further includes:

[0045] When it is determined that the power parameter is not within the preset parameter range, an alarm instruction is triggered;

[0046] generating an execution instruction according to the alarm instruction;

[0047] Retrieving the maximum value parameter within the preset parameter range according to the execution instruction;

[0048] The underwater propeller is controlled to move according to the maximum value parameter.

[0049] In a second aspect, the present application provides an underwater thruster control device based on multi-parameter fusion, the device comprising:

[0050] A collection unit, used to collect water flow velocity and regional terrain information in the target area where the underwater propeller is located;

[0051] a first acquisition unit, configured to establish a three-dimensional flow field real-time perception model based on the water flow velocity and the regional terrain information, and acquire multimodal data according to the three-dimensional flow field real-time perception model, wherein the multimodal data includes water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates, and terrain height of the target area;

[0052] A second acquisition unit is used to perform dispersion filtering and subsequent fusion on the multimodal data to obtain target multimodal data;

[0053] A construction unit, configured to construct a multimodal neural network model based on the target multimodal data and through a temporal attention mechanism;

[0054] a first generating unit, configured to input the real-time multimodal data of the underwater thruster into the multimodal neural network model and generate control instructions using a two-layer reinforcement learning strategy;

[0055] A second generating unit, configured to generate power parameters of the underwater propeller according to the control instruction;

[0056] A control unit is used to control the underwater propeller to move according to the power parameter when it is determined that the power parameter is within a preset parameter range.

[0057] In a third aspect, the present application provides an underwater thruster control device based on multi-parameter fusion, the device comprising:

[0058] processor, memory, input and output units, and buses;

[0059] The processor is connected to the memory, the input and output unit, and the bus;

[0060] The storage stores a program, and the processor calls the program to execute the first aspect and any optional method in the first aspect.

[0061] In a fourth aspect, the present application provides a computer-readable storage medium, on which a program is stored. When the program is executed on a computer, the program executes the first aspect and any optional method in the first aspect.

[0062] It can be seen from the above technical solutions that this application has the following advantages:

[0063] The following are the technical effects that can be achieved by this technical solution:

[0064] 1. This application collects water flow velocity and regional terrain information in the target area and establishes a three-dimensional flow field real-time perception model. The three-dimensional flow field real-time perception model can perceive the underwater environment more comprehensively and accurately, provide a more reliable data basis for subsequent control, and help underwater thrusters better adapt to the complex and changeable underwater environment.

[0065] 2. The multimodal data obtained from the three-dimensional flow field real-time perception model, including water velocity, water pressure, water intensity, three-dimensional terrain coordinates and terrain height, covers multiple key parameters of the underwater environment, enriches the data dimension, and enables the control method of this application to comprehensively consider more factors and improve the accuracy and adaptability of control.

[0066] 3. This application performs decentralized filtering and post-fusion on multimodal data, which can effectively remove noise and interference, improve the quality of the target multimodal data, and ensure that the subsequently constructed models and generated control instructions are more accurate and reliable, thereby better controlling the movement of the underwater thruster.

[0067] 4. Based on the target multimodal data, a multimodal neural network model is constructed through the temporal attention mechanism, which can better capture the temporal characteristics and important information in the data, improve the adaptability and prediction accuracy of the multimodal neural network model to the underwater thruster control task, and thus optimize the control effect.

[0068] 5. This application adopts a two-layer reinforcement learning strategy to generate control instructions, so that the control method can continuously learn and optimize the control strategy based on real-time multimodal data, improve the autonomous decision-making ability and adaptability of the underwater propeller in different environments, and achieve greater intelligence. It also generates the power parameters of the underwater propeller according to the control instructions, and controls the propeller operation only when it is determined that the power parameters are within the preset parameter range. It can effectively avoid propeller failure or damage caused by parameter abnormalities, and ensure the safety and stability of the underwater propeller operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in this application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0070] Figure 1 This is a flow chart of an embodiment of the underwater thruster control method based on multi-parameter fusion of the present application;

[0071] Figure 2 This is a schematic diagram of another embodiment of the underwater thruster control method based on multi-parameter fusion of the present application;

[0072] Figure 3 This is a schematic diagram of another embodiment of the underwater thruster control method based on multi-parameter fusion of the present application;

[0073] Figure 4 This is a schematic diagram of another embodiment of the underwater thruster control method based on multi-parameter fusion of the present application;

[0074] Figure 5 This is a schematic diagram of another embodiment of the underwater thruster control method based on multi-parameter fusion of the present application;

[0075] Figure 6 This is a schematic diagram of another embodiment of the underwater thruster control method based on multi-parameter fusion of the present application;

[0076] Figure 7 This is a schematic diagram of another embodiment of the underwater thruster control method based on multi-parameter fusion of the present application;

[0077] Figure 8 This is a schematic diagram of an embodiment of an underwater thruster control device based on multi-parameter fusion in the present application;

[0078] Figure 9 This is a schematic diagram of another embodiment of the underwater thruster control device based on multi-parameter fusion of the present application. DETAILED DESCRIPTION

[0079] The present application provides an underwater thruster control method and related equipment based on multi-parameter fusion, which improves the autonomous decision-making ability and adaptability of the underwater thruster in different environments, realizes more intelligent and efficient control, and thus helps to improve the propulsion efficiency of the underwater thruster.

[0080] See also Figure 1 In a first aspect, the present application provides an embodiment of an underwater thruster control method based on multi-parameter fusion, the embodiment comprising:

[0081] 101. Collect water flow velocity and regional terrain information in the target area where the underwater propeller is located;

[0082] 102. Establish a three-dimensional flow field real-time perception model based on the water flow velocity and the regional terrain information, and acquire multimodal data according to the three-dimensional flow field real-time perception model, wherein the multimodal data includes water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates, and terrain height of the target area;

[0083] 103. Distribute and filter the multimodal data and then fuse them to obtain target multimodal data;

[0084] 104. Based on the target multimodal data, a multimodal neural network model is constructed through a temporal attention mechanism;

[0085] 105. Inputting the real-time multimodal data of the underwater thruster into the multimodal neural network model, and generating control instructions using a two-layer reinforcement learning strategy;

[0086] 106. Generate power parameters of the underwater propulsion device according to the control instruction;

[0087] 107. When it is determined that the power parameter is within a preset parameter range, control the underwater propeller to move according to the power parameter.

[0088] In an embodiment of the present application, the water flow velocity and regional terrain information of the target area where the underwater propeller is located are first collected, and then a three-dimensional flow field real-time perception model is established based on the water flow velocity and regional terrain information, and multimodal data is obtained according to the three-dimensional flow field real-time perception model. The multimodal data includes the water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates and terrain height of the target area. The multimodal data is then dispersed filtered and fused to obtain target multimodal data. Based on the target multimodal data, a multimodal neural network model is constructed through a temporal attention mechanism. Furthermore, the real-time multimodal data is input into the multimodal neural network model, and a two-layer reinforcement learning strategy is used to generate control instructions. Then, the power parameters of the underwater propeller are generated according to the control instructions. When it is determined that the power parameters are within the preset parameter range, the underwater propeller is controlled to move according to the power parameters.

[0089] In step 101, the water flow velocity and regional terrain information of the target area where the underwater propeller is located are collected. Specifically, the control accuracy of the underwater propeller is highly dependent on the accuracy of the environmental data. Therefore, it is necessary to collect the environmental data of the area where the underwater propeller is located, and the main focus is on collecting water flow velocity and regional terrain information.

[0090] Among them, water flow velocity information can be collected through ADCP (Acoustic Doppler Current Profiler). ADCP uses the Doppler effect to emit 4-6 beams of sound waves into the water body. By analyzing the frequency offset of the echo reflected by suspended particles in the water body, it obtains the three-dimensional water flow velocity at different depths and angles, providing a real-time water flow velocity reference for the underwater propeller. It can also be collected through electromagnetic flow meters and data acquisition frequency methods, which are not specifically limited here.

[0091] The collection of regional terrain information requires the use of a multi-beam echo sounder. The multi-beam echo sounder emits multiple beams of sound waves and measures water depth data at multiple angles, thereby constructing a high-precision three-dimensional terrain model. The terrain undulation information it generates can be used for propeller path planning to avoid collisions with obstacles. For example, in areas with reefs on the seabed, multi-beam data can clearly show the location and height of the reefs, allowing the propeller to plan a detour. By emitting sound waves to the side and receiving backscattered signals from the seabed and objects to generate images, which complement the multi-beam data and improve the terrain details, Gaussian filtering is then used to remove terrain data noise, and the Kriging interpolation method is used to encrypt the data to improve the accuracy of the terrain model. The processed terrain data will serve as an important input for the real-time perception model of the three-dimensional flow field.

[0092] Then, using GPS or a high-precision clock device, the network time protocol provides a unified time reference for the data collection equipment, ensuring that the water flow velocity and terrain information are consistent in time. This ensures that the established 3D flow field real-time model accurately reflects the environmental conditions at the same moment. After the regional water flow velocity and terrain information are collected, step 102 is executed.

[0093] In step 102, a three-dimensional flow field real-time perception model is established based on water flow velocity and regional terrain information, and multimodal data is obtained according to the three-dimensional flow field real-time perception model. Specifically, establishing a three-dimensional flow field real-time perception model is a process of integrating discrete water flow velocity and terrain information into a continuous, dynamic three-dimensional spatial model.

[0094] First, the collected data needs to be preprocessed, including noise removal and interpolation to fill missing data. For water velocity data, there may be some outliers due to sensor failure or environmental interference. These can be removed by setting thresholds or using statistical methods. Missing portions of terrain data can be filled using kriging interpolation or inverse distance weighted interpolation to ensure data integrity.

[0095] The processed data is then fused into three-dimensional space using mathematical algorithms. Computational fluid dynamics (CFD) algorithms are commonly used, simulating the movement and distribution of water flow by solving the Navier-Stokes equations of fluid mechanics. It should be noted that when establishing the model, factors such as the boundary conditions and initial conditions of the target area must be considered. For example, the boundary of the target area is set as a fixed or movable wall, and the initial water flow velocity and pressure distribution are set according to the actual water flow conditions. Then, through continuous iterative calculations, a three-dimensional flow field real-time perception model is obtained that can reflect the water flow state and terrain characteristics of the target area in real time.

[0096] Based on the established three-dimensional flow field real-time perception model, multimodal data can be further obtained. In addition to the water flow velocity initially collected, parameters such as water flow pressure and water flow intensity can also be calculated through the model. The water flow pressure can be calculated using fluid mechanics formulas and determined based on the water flow velocity, density and position. The water flow intensity can be measured based on the size and direction of the water flow velocity.

[0097] The real-time perception model of the three-dimensional flow field already includes the regional terrain information, including three-dimensional terrain coordinates and terrain height. Multimodal data can be used to describe the environmental characteristics of the target area. For example, water flow velocity and pressure can reflect the dynamic characteristics of the water flow, while terrain coordinates and height can describe the geometric shape of the bottom of the water. These data provide a more comprehensive foundation for subsequent data processing and model construction.

[0098] In step 103, the multimodal data is decentralized filtered and then fused to obtain the target multimodal data. Specifically, after acquiring multimodal data through the 3D flow field real-time perception model, the acquisition process may be affected by various noise and interference. Directly using this raw data may lead to inaccurate and unstable subsequent models. Therefore, decentralized filtering of the multimodal data is necessary.

[0099] Kalman filtering is often used to process water flow velocity data. Kalman filtering establishes the system's state equation and observation equation, and uses a recursive method to estimate and update the system state. In water flow velocity filtering, the water flow velocity is regarded as the system's state variable, and the sensor measurement value is used as the observation value. Through continuous iterative calculation, the noise in the water flow velocity data can be effectively removed, and the smoothness and accuracy of the data can be improved.

[0100] For terrain height data, a median filter algorithm can be used. This is a nonlinear filtering method that sorts all data values within a data window and takes the median value as the filtered output. This method effectively removes outliers and impulse noise from terrain height data, preserving the data's key features. For data such as flow pressure and flow intensity, appropriate filtering algorithms can also be selected based on their characteristics, such as exponential smoothing and low-pass filtering, though these are not specifically defined here.

[0101] After filtering, the data from different modalities is fused. The purpose of data fusion is to integrate information from different modalities to obtain a more comprehensive and accurate description of the target area environment. During the data fusion process, weighted averaging or Bayesian estimation can be used, but these methods are not specifically limited here. The resulting fused target multimodal data will serve as the basis for the subsequent construction of a multimodal neural network model. This target multimodal data combines the advantages of different modal data and can more accurately reflect the environmental state of the target area.

[0102] In step 104, a multimodal neural network model is constructed based on the target multimodal data and through a temporal attention mechanism. Specifically, before inputting the target multimodal data into the neural network model, the target multimodal data needs to be normalized to unify the numerical ranges of different modal data in the target multimodal data into the same interval, for example, normalizing the data to between [0, 1] or [-1, 1]. After processing, it is possible to avoid the influence of different modal data on model training due to excessive differences in numerical ranges.

[0103] Then, feature extraction is performed on the target multimodal data. For continuous data such as water velocity, pressure, and intensity, the Fourier transform method is used to extract their frequency domain features or time-frequency features. For spatial data such as terrain coordinates and height, the principal component analysis method is used to extract their main features. Through feature extraction, the dimension of the data can be reduced and the training efficiency and generalization ability of the model can be improved.

[0104] A multimodal neural network model is constructed, which includes a convolutional neural network (CNN) subnetwork and a recurrent neural network (RNN) subnetwork. A convolutional neural network (CNN) subnetwork is constructed for flow-related data such as velocity, pressure, and intensity. CNNs have powerful feature extraction capabilities and can automatically learn local features and spatial structures in data. Through multi-layer convolution and pooling operations, they can transform raw data into high-level feature representations.

[0105] For spatial data such as terrain coordinates and altitude, a recurrent neural network (RNN) subnetwork is constructed. RNN can process data with temporal relationships and capture long-term dependencies in the data. When processing terrain data, the terrain coordinates can be regarded as time series data, and the changing patterns of terrain features can be learned through RNN.

[0106] In underwater propulsion control scenarios, environmental information at different times has varying degrees of influence on current control decisions. Therefore, the introduction of a temporal attention mechanism automatically focuses on important temporal segments in the data, allowing the model to prioritize information useful for the current decision. The temporal attention mechanism comprises an attention calculation layer and a weight distribution layer. In the attention calculation layer, an attention score is calculated based on the query vector at the current moment and key vectors at different moments. The query vector can be viewed as the model's focus at the current moment, while the key vector contains feature information about the data at different moments. The attention score is calculated by calculating the similarity between the query vector and the key vector.

[0107] In the weighting layer, data at different times are weighted and summed according to the attention scores to obtain a weighted feature representation. This allows the model to automatically give higher weight to data at important times when processing data, thereby improving the model's decision-making ability.

[0108] The processed target multimodal data is then divided into training set, validation set and test set. The training set is used to train the multimodal neural network model, and the parameters of the model are continuously adjusted through the back propagation algorithm to make the output of the model as close to the true value as possible.

[0109] During training, the model's prediction error is measured using functions such as cross-entropy loss and mean squared error loss. Simultaneously, the model is validated using a validation set, and the model's hyperparameters are adjusted based on the performance metrics on the validation set to prevent overfitting or underfitting. After multiple iterations of training and optimization, a multimodal neural network model is obtained that can learn the mapping relationship between target multimodal data and control decisions.

[0110] In step 105, the real-time multimodal data of the underwater thruster is input into the multimodal neural network model, and a two-layer reinforcement learning strategy is used to generate control instructions. Specifically, during actual operation, the underwater thruster will continuously collect real-time multimodal data, which include information such as water flow velocity, pressure, intensity, and terrain coordinates and altitude at the current moment. The collected real-time multimodal data also needs to be preprocessed, including normalization, feature extraction and other operations.

[0111] The preprocessed real-time multimodal data is input into a trained multimodal neural network model. The model will output a preliminary control decision suggestion based on the learned mapping relationship. This decision suggestion contains the probability distribution of different control actions. A two-layer reinforcement learning strategy is then used to generate control instructions. Specifically, the two-layer reinforcement learning strategy includes an upper-layer strategy and a lower-layer strategy, which work together to generate the optimal control instructions.

[0112] The upper-level policy is responsible for developing control objectives and strategies. For example, the upper-level policy might determine an overall control direction and priority based on the underwater thruster's mission requirements, such as reaching a specific location and avoiding obstacles, and the current environmental conditions. If the mission requires reaching the target location as quickly as possible, and the upper-level policy determines that the current water flow direction favors the thruster's forward motion, the upper-level policy will tend to select a control direction that accelerates the thruster's forward motion.

[0113] The lower-level strategy generates specific control instructions based on the guidance of the upper-level strategy, combined with the current environmental status and the dynamic characteristics of the thruster. The lower-level strategy adopts a more sophisticated control method, such as calculating the specific thruster speed, steering angle and other parameters based on the physical model of the thruster and real-time data.

[0114] In conjunction with the two-layer reinforcement learning strategy generation process, the terminal also continuously tries different control commands through interaction with the environment and adjusts its strategy based on the rewards or penalties received after executing the commands. This is achieved through a reward function, which can be defined based on the control objectives and task requirements. For example, if the thruster successfully reaches the target position, a large positive reward can be given; if a collision occurs or the thruster deviates from the target direction, a negative reward is given. After multiple iterations of learning and strategy adjustment, the terminal can learn the optimal control strategy, thereby generating more reasonable and efficient control commands.

[0115] In step 106, the power parameters of the underwater propeller are generated according to the control instructions. Specifically, the control instructions are direct signals that guide the movement of the underwater propeller, but these instructions are usually expressed in an abstract form, such as direction instructions such as forward, backward, turn left, turn right, and speed instructions. Therefore, it is necessary to first parse the control instructions and convert them into specific control parameters.

[0116] For example, a direction command can be converted into the propeller's steering angle, while a speed command can be converted into the propeller's target speed. The parsing process ensures that the converted control parameters accurately meet the control command's requirements.

[0117] Based on the analyzed control parameters, combined with the physical characteristics and dynamic model of the underwater thruster, the corresponding dynamic parameters can be calculated. The thruster's dynamic parameters mainly include speed and thrust. There is a certain relationship between the thrust and speed of the thruster, which can be described by the thruster's performance curve or mathematical model. For example, the thrust of the thruster can be expressed as a quadratic function or higher-order function of the speed. Based on the thrust required by the control command, the required speed of the thruster can be calculated by solving this functional relationship.

[0118] At the same time, in actual applications, it is necessary to consider the impact of factors such as water flow resistance on the thrust of the propeller. Water flow resistance is related to factors such as water flow velocity, shape and size of the propeller. When calculating the power parameters, water flow resistance must be taken into consideration to ensure that the propeller can generate sufficient thrust to overcome resistance in actual operation and achieve the expected motion effect.

[0119] In step 107, if the power parameters are determined to be within a preset parameter range, the underwater propeller is controlled to move according to the power parameters. Specifically, after the power parameters are generated, a safety check is performed on the power parameters. The preset parameter range is determined based on factors such as the design specifications, performance limitations, and safety requirements of the underwater propeller. For example, the speed of the underwater propeller cannot exceed its maximum allowable speed, otherwise it may cause the propeller to overheat, damage, or lose control. The thrust cannot exceed the range that its structural strength can withstand, otherwise it may damage the mechanical structure of the propeller.

[0120] Therefore, safety checks can be performed by comparing the generated power parameters with a preset parameter range. If the power parameters exceed the preset parameter range, these parameters are unsafe or infeasible, and the control instructions or the power parameter generation process need to be readjusted. For example, if the calculated propeller speed exceeds the maximum allowable speed, the speed requirement in the control instruction can be reduced, or the propeller control strategy can be adjusted to generate power parameters that meet safety requirements.

[0121] If the generated power parameters are determined to be within the preset parameter range, they are considered safe and feasible. At this point, the power parameters are transmitted to the underwater propulsion system's actuator, such as the motor controller. Based on the received power parameters, the motor controller controls the propulsion system's motor, adjusting its speed and direction. During the propulsion system's motion, its operating status and environmental changes must be monitored in real time to ensure its safe operation.

[0122] Therefore, this application adopts a two-layer reinforcement learning strategy to generate control instructions, so that the control method can continuously learn and optimize the control strategy based on real-time multimodal data, improve the autonomous decision-making ability and adaptability of the underwater propeller in different environments, achieve more intelligent and efficient control, and improve the propulsion efficiency of the underwater propeller. In addition, the power parameters of the underwater propeller are generated according to the control instructions, and the propeller operation is controlled only when it is determined that the power parameters are within the preset parameter range. This can effectively avoid propeller failure or damage caused by parameter abnormalities and ensure the safety and stability of the underwater propeller operation.

[0123] Please refer to Figure 2 According to some embodiments of the present invention, the steps of distributing and filtering the multimodal data and then fusing them to obtain target multimodal data in step 103 may specifically include, but are not limited to, the following:

[0124] 201. Preprocess the multimodal data to obtain preprocessed data;

[0125] 202. Extracting water flow data and terrain data from the pre-processed data;

[0126] 203. According to the data characteristics of water flow data and terrain data, Kalman filter algorithm and Gaussian filter algorithm are used to eliminate outliers respectively;

[0127] 204. Performing data registration on the filtered water flow data and terrain data;

[0128] 205. Using weighted average algorithm to fuse the data after data registration;

[0129] The weighted average algorithm is:

[0130]

[0131] Where S is the weighted average, n is the number of data points, and x i is the value of the i-th data point, w i is the weight of the i-th data point;

[0132] 206. Verify the compared and fused data. When it is determined that the data verification passes, smooth the data and output the target multimodal data.

[0133] In an embodiment of the present application, after obtaining the multimodal data for controlling the underwater thruster, in order to be able to perform analysis and processing more accurately and efficiently in the future, and thus achieve precise control, these multimodal data must first be preprocessed. The preprocessing stage will clean the multimodal data to remove those data points that are obviously wrong or do not conform to the actual situation. For example, some sensors may output abnormally high or abnormally low values due to external interference or their own failures. These abnormal values will affect the accuracy of subsequent data analysis, so they will be eliminated during preprocessing. At the same time, the data format will also be converted to ensure that data from different sources and in different formats can be unified into a standard format to facilitate subsequent unified processing and analysis. After the preprocessing operation, the preprocessed data is obtained, and the preprocessed data lays the foundation for subsequent data extraction and analysis.

[0134] Then, water flow data and terrain data are extracted from the preprocessed data. The water flow data contains information such as the water flow speed and direction in the underwater environment, while the terrain data describes the terrain features around the underwater propeller, such as the undulations of the bottom of the water and the location of obstacles. The obtained water flow data and terrain data help plan the movement path of the underwater propeller and avoid collisions with obstacles.

[0135] Because water flow data and terrain data have different characteristics, different filtering algorithms are required to eliminate outliers. Water flow data is characterized by a certain degree of dynamics and randomness, and the speed and direction of the flow may change over time and space. Therefore, the Kalman filter algorithm is used for filtering. The Kalman filter algorithm effectively removes noise and outliers from water flow data. It uses the state estimate at the previous moment and the observed value at the current moment to continuously update the state estimate through recursion, so that the estimated value can more accurately reflect the actual water flow state.

[0136] Topographic data, on the other hand, is characterized by a certain degree of spatial correlation and smoothness. Data such as terrain height typically does not experience significant abrupt changes between adjacent areas, so a Gaussian filter is used to process topographic data. Gaussian filtering is a linear smoothing filter that removes noise by performing a weighted average on the data. The Gaussian filter uses a Gaussian function as a weighting function, giving data points closer to the center a greater weight and data points farther from the center a smaller weight. This filtering process effectively removes noise while preserving the overall characteristics and edge information of the topographic data. By using the Kalman filter and Gaussian filter algorithms to process water flow data and topographic data, respectively, the quality and reliability of the data can be greatly improved.

[0137] After filtering, flow data and terrain data may still have some discrepancies, such as inconsistencies in the data's coordinate system and time base, which can affect subsequent data fusion and analysis. Therefore, it is necessary to perform data registration on the filtered flow data and terrain data. The goal of data registration is to unify data from different sources and formats into a common coordinate system and time base, enabling comparison and analysis within the same spatial and temporal framework.

[0138] During data registration, the corresponding relationship between water flow data and terrain data is identified. For example, through methods such as feature point matching and spatial transformation, the location information in the water flow data is accurately matched with the location information in the terrain data. At the same time, time information is synchronized to ensure that the different data are synchronized in time. Through data registration, the water flow data and terrain data can be better integrated, providing an accurate foundation for subsequent data fusion.

[0139] Finally, the weighted average algorithm is used to fuse the registered water flow data and terrain data. The weighted average algorithm is:

[0140]

[0141] Where S is the weighted average, n is the number of data points, and x i is the value of the i-th data point, w i is the weight of the i-th data point; the weighted average algorithm assigns different weights according to the importance and reliability of different data, and then performs weighted summation on each data to obtain the fused data.

[0142] It should be noted that when determining the weights, the degree of influence of water flow data and terrain data on the underwater thruster's control decisions must be taken into account. For example, if the water flow velocity is considered to have a greater impact on the thruster's motion, a higher weight can be assigned to the water flow data; whereas, if certain detailed features in the terrain data have less influence on the thruster's path planning, a lower weight can be assigned. By rationally assigning weights, the weighted average algorithm can combine the advantages of water flow data and terrain data to obtain a more comprehensive and accurate description of environmental information. The fused data will provide a more reliable basis for the underwater thruster's control decisions, enabling the thruster to better adapt to complex underwater environments and achieve safe and efficient operation.

[0143] Please refer to Figure 3 According to some embodiments of the present invention, in step 104, constructing a multimodal neural network model based on the target multimodal data and using a temporal attention mechanism may specifically include, but is not limited to, the following:

[0144] 301. Extract numerical, spatial, and temporal features from the target multimodal data and convert them into a unified dimensional vector;

[0145] 302. Arrange the unified dimension vectors in time series to obtain a feature matrix;

[0146] 303. Based on the data characteristics of the feature matrix, a multimodal neural network model is constructed using a temporal attention mechanism.

[0147] In an embodiment of the present application, various features in the target multimodal data are first extracted and converted. The target multimodal data covers a variety of information about the environment in which the underwater propeller is located, including numerical data such as water velocity, pressure, and intensity, spatial data such as terrain coordinates and altitude, and time series data such as water flow dynamic information that changes over time. For numerical features, such as water velocity, the speed values collected at different times and locations vary in size, and these values directly reflect the changes in water flow under different conditions. By performing statistical analysis on these values, such as calculating the mean, variance, maximum value, minimum value, etc., the overall characteristics and fluctuation characteristics of the water velocity can be extracted.

[0148] The spatial features are mainly reflected in the terrain coordinates and height. These data describe the geometric shape of the environment around the underwater propeller. The terrain data is processed using a spatial interpolation algorithm to obtain more detailed terrain features, such as the slope and undulation of the terrain.

[0149] In terms of time series characteristics, since data such as water flow velocity and pressure will change continuously over time, by calculating the autocorrelation function and power spectrum density of the data, we can extract characteristics such as periodicity and trend in the time series data.

[0150] After extracting various features, they must be converted into vectors of uniform dimensions. This is because different types of features have different data structures and dimensions, and direct processing can make it difficult for the model to effectively learn. For numerical features, they can be directly combined into a vector, but to align with other feature dimensions, normalization may be required to scale the values to an appropriate range.

[0151] For spatial features, we can map the terrain coordinates and height information into a specific spatial coordinate system and then convert them into vector form according to grid division. For example, the terrain area can be divided into several small grids, each grid corresponding to a feature value, thus forming a high-dimensional vector.

[0152] For time series features, the time series encoding method is adopted. For example, the time series data is sliced according to the time step, and the features of each time step form a sub-vector. These sub-vectors are then spliced into a long vector. It should be noted that during the conversion process, it is necessary to ensure that the position and weight of various features in the vector are reasonable so that the subsequent model can make full use of these feature information.

[0153] After the conversion of the feature vectors is completed, these vectors of uniform dimension are arranged in time series to obtain a feature matrix. For example: suppose that multiple sets of target multimodal data are collected over a period of time. After feature extraction and conversion, each set of data obtains a vector of uniform dimension. These vectors are arranged in chronological order, and each row represents the feature vector of a time point, thus forming a two-dimensional feature matrix. The number of rows in the feature matrix represents the number of time steps, and the number of columns represents the dimension of the feature vector. Through this arrangement, the feature matrix not only retains the feature information of each time point, but also reflects the relationship between the changes of features over time. In the feature matrix, the difference between the feature vectors of adjacent rows can reflect the temporal trend of parameters such as water flow velocity and pressure, which helps to understand the dynamic behavior of water flow and predict future changes.

[0154] Finally, based on the data characteristics of the feature matrix, a multimodal neural network model is constructed using the self-attention mechanism. The feature matrix has high data dimensionality and strong temporal correlation, and the self-attention mechanism is well suited to handle these characteristics. The self-attention mechanism enables the model to automatically focus on related features in other time steps while processing the features of each time step.

[0155] When building a neural network model, the feature matrix is first fed into an embedding layer, which converts the discrete feature values into a continuous vector representation that the model can better process. The embedded vector is then fed into a self-attention layer. The self-attention layer calculates the similarity between the query vector, key vector, and value vector to determine the attention weight of each time step feature on the features of other time steps.

[0156] The value vectors are then weighted and summed based on these attention weights to obtain a weighted feature representation. This allows the model to dynamically focus on important information at different time steps when processing the features of each time step, thereby better capturing the temporal dependencies between features.

[0157] To further improve model performance, a backpropagation algorithm and optimizer are used during model training to continuously adjust model parameters, ensuring that the model output is as close as possible to the actual control decision or target value. Furthermore, the model is validated using a validation set, and hyperparameters are adjusted based on performance metrics from the validation set to prevent overfitting or underfitting. After multiple iterations of training and optimization, a multimodal neural network model is developed that fully utilizes the feature matrix information and accurately predicts underwater thruster control commands.

[0158] Please refer to Figure 4 According to some embodiments of the present invention, in step 105, inputting real-time multimodal data into the multimodal neural network model and using a two-layer reinforcement learning strategy to generate control instructions may specifically include, but are not limited to, the following:

[0159] 401. Obtain real-time multimodal data;

[0160] 402. Preprocess the real-time multimodal data to form an input feature vector adapted to the multimodal neural network model;

[0161] 403. Input the input feature vector into the multimodal neural network model to obtain a target feature vector;

[0162] 404. Build a two-layer reinforcement learning strategy network, the two-layer reinforcement learning strategy network including a high-level decision network and a low-level execution network, the high-level decision network is used to receive the target feature vector output by the multimodal neural network model and plan a control strategy, and the low-level execution network is used to refine the control strategy into control parameters;

[0163] 405. Input the target feature vector into the two-layer reinforcement learning policy network, and use the Actor-Critic algorithm to update the policy network parameters;

[0164] 406. Generate a control instruction based on the strategic network parameters and in combination with the current environment state.

[0165] In the embodiments of the present application, after acquiring real-time multimodal data, it is necessary to preprocess it to form input feature vectors that are adapted to the multimodal neural network model. Since the data collected by different sensors differ in format and dimension, directly using this raw data will make model training difficult or ineffective. Therefore, it is necessary to normalize the data and unify the numerical ranges of the different modal data to an appropriate interval, such as normalizing the data to between [0, 1] or [-1, 1], to avoid affecting the model performance due to excessive differences in the numerical range.

[0166] Next, feature extraction is performed. For continuous data such as water velocity and pressure, methods such as Fourier transform and wavelet transform are used to extract their frequency domain features or time-frequency features. These features can better reflect the inherent laws of the data. For spatial data such as terrain coordinates and altitude, methods such as principal component analysis (PCA) and independent component analysis (ICA) are used to extract their key features, reducing the data dimension while retaining key information. Through preprocessing operations such as normalization and feature extraction, the raw real-time multimodal data is converted into input feature vectors that the multimodal neural network model can understand and process.

[0167] The input feature vector formed after preprocessing is input into the multimodal neural network model to obtain the target feature vector. Specifically, since the multimodal neural network model has undergone a lot of early training, it has learned the mapping relationship between the input feature vector and the target feature vector. When the input feature vector enters the model, the various network layers within the model will perform a series of complex calculations and transformations on it. For example, the convolutional neural network (CNN) subnetwork will perform convolution and pooling operations on features related to water flow, automatically extracting local features and spatial structures in the data, and the recurrent neural network (RNN) or its variant subnetwork will process features with temporal or spatial dependencies such as terrain to capture long-term dependencies in the data. Through the collaborative work of these network layers, the input feature vector is gradually converted into more representative and discriminative target feature vectors, which contain information about the environmental state and the propeller's motion requirements.

[0168] Building a two-layer reinforcement learning strategy network is an important part of the entire control process. The two-layer reinforcement learning strategy network includes a high-level decision network and a low-level execution network. The main function of the high-level decision network is to receive the target feature vectors output by the multimodal neural network model and plan the control strategy based on these feature vectors. It can comprehensively consider the mission objectives of the underwater thruster, such as reaching a designated location, avoiding obstacles, etc. and the current environmental state, and determine the overall control direction and priority of the thruster's movement direction, speed range, etc. from a macro perspective. The low-level execution network is responsible for refining the control strategy planned by the high-level decision network into specific control parameters. Combined with the physical model of the thruster and real-time data, it calculates the thruster's precise parameters such as speed and steering angle to ensure that the thruster can move accurately according to the strategy formulated by the high-level decision network.

[0169] After inputting the resulting target feature vector into the two-layer reinforcement learning policy network, the actor-critic algorithm is used to update the policy network parameters. The actor-critic algorithm is a reinforcement learning algorithm that combines the policy gradient method with the value function estimation method. In the two-layer reinforcement learning policy network, the actor network is responsible for generating the control policy or control parameters based on the current state, while the critic network is used to evaluate the quality of the actions selected by the actor network, that is, to estimate the value function of the state-action pair.

[0170] When the target feature vector is input into the network, the Actor network generates a control policy or parameters based on the current state. After the thruster executes the action, the environment returns a reward signal. The Critic network uses this reward signal and subsequent state transition information to update its estimate of the value of the state-action pair. Simultaneously, the Actor network adjusts its own policy parameters based on the value estimate provided by the Critic network, ensuring that it chooses the more valuable action when encountering similar states in the future.

[0171] By continuously interacting with the environment, the Actor-Critic algorithm can gradually optimize the parameters of the two-layer reinforcement learning policy network, improving the accuracy and effectiveness of the control policies and parameters generated by the network.

[0172] Finally, control instructions are generated based on the updated policy network parameters and the current environmental state. Specifically, after the parameters of the two-layer reinforcement learning policy network are continuously optimized by the Actor-Critic algorithm, the terminal can generate more reasonable and efficient control instructions based on the input target feature vector and the current environmental state to ensure the safe operation of the thruster.

[0173] Please refer to Figure 5 According to some embodiments of the present invention, generating the power parameters of the underwater propulsion device according to the control instruction in step 106 may specifically include, but is not limited to, the following:

[0174] 501. Calculate the output power of the underwater propulsion device according to the control instruction;

[0175] 502. Compare the output power of the underwater propeller with the maximum output power of the underwater propeller, and obtain a comparison result;

[0176] 503. Generate power parameters of the underwater propulsion device according to the comparison result.

[0177] In an embodiment of the present application, after a control instruction is generated, the output power of the underwater propeller is calculated based on the control instruction. It should be noted that the control instruction is a guiding signal in the entire control process, and the control instruction contains specific requirements for the motion state of the underwater propeller, such as the desired motion speed, direction, and acceleration.

[0178] In order to convert these abstract control requirements into actual operational output power values, the following factors must be comprehensively considered: on the one hand, the mechanical structure characteristics of the underwater thruster itself must be combined with the environmental factors on the other hand. Based on these factors, an accurate dynamic model should be established, and the motion parameters in the control instructions should be combined with the mechanical characteristics and environmental parameters of the thruster. Fluid mechanics calculations and dynamic equations should be used to calculate the output power required by the underwater thruster under the current control instructions.

[0179] After calculating the output power, compare the underwater propeller's output power with its maximum output power to obtain a comparison result. The maximum output power of an underwater propeller is a critical parameter determined by multiple factors, including the propeller's design specifications, material strength, and motor performance. It represents the maximum driving force the propeller can provide within its safe operating range.

[0180] Comparing the calculated current output power with the maximum output power provides a clear understanding of the propeller's power requirements under the current control requirements. The comparison results fall into three categories: First, the current output power is less than the maximum output power, indicating that the propeller can operate safely and stably under the current control instructions, without concern for power overload. Second, the current output power is equal to the maximum output power. At this point, the propeller is operating at full load. While this meets the current control requirements, attention should be paid to the propeller's operating status to avoid insufficient power or equipment damage due to environmental changes or mechanical failures. Third, the current output power is greater than the maximum output power, indicating that the propeller will exceed its capacity under the current control instructions. This situation is very likely to cause serious consequences such as equipment overheating, mechanical component damage, and even loss of control. This comparison process provides a clear basis for the subsequent generation of power parameters.

[0181] After the comparison, the results are obtained, and finally, the underwater propulsion system's power parameters are generated based on the comparison results. If the comparison results show that the current output power is less than the maximum output power, the power parameters can be appropriately adjusted based on actual needs and system optimization goals, without exceeding the maximum output power. For example, to improve the propulsion system's operating efficiency or response speed, the power output can be appropriately increased within a safe range to generate corresponding power parameters such as speed and thrust.

[0182] If the comparison result shows that the current output power is equal to the maximum output power, while ensuring that the thruster can complete the current task, factors such as the heat dissipation and mechanical fatigue of the equipment must be fully considered. The power parameters can be fine-tuned, such as reducing the speed or thrust slightly, to extend the service life of the thruster.

[0183] When the comparison result shows that the current output power is greater than the maximum output power, the control instructions must be re-evaluated and adjusted. One way is to reduce the motion requirements in the control instructions, such as reducing the expected motion speed or acceleration, so as to recalculate the output power so that it does not exceed the maximum output power, and then generate new power parameters. Another way is to optimize the propeller's operating strategy, such as changing the propeller's operating mode and adopting intermittent propulsion. While meeting certain mission requirements, it ensures that the power output is within a safe range. By generating reasonable power parameters based on different comparison results, it can ensure that the underwater propeller can operate safely and efficiently in various complex environments and achieve precise motion control.

[0184] Please refer to Figure 6 According to some embodiments of the present invention, in step 108, when it is determined that the power parameter is within a preset parameter range, controlling the underwater propeller to move according to the power parameter may specifically include, but is not limited to, the following:

[0185] 601. Analyze the power parameters and verify whether the power parameters are complete;

[0186] 602. If yes, then compare the analyzed dynamic parameters with the preset parameter range;

[0187] 603. When it is determined that the power parameter is within a preset parameter range, convert the power parameter into a control signal executable by the underwater propulsion device;

[0188] 604. Control the underwater propeller to move according to the control signal.

[0189] In the embodiment of the present application, after the generation of the underwater propeller power parameters is completed, in order to ensure that these parameters can accurately and effectively control the propeller movement, the power parameters need to be processed and verified.

[0190] First, the power parameters are parsed and their integrity verified. Power parameters are typically stored in specific data formats or encoded forms. Parsing involves breaking down these complex data structures into specific, understandable values or instructions. For example, power parameters may include information such as propeller speed, thrust, and steering angle. Parsing requires extracting this information from the overall data set according to protocols or rules. During parsing, the integrity of the power parameters must be verified. This integrity verification involves multiple aspects. First, data integrity must be checked—that is, all necessary parameters are included in the power parameter set. For example, if propeller motion control requires both speed and steering angle, verification must confirm that both parameters exist and are valid. Second, data corruption or errors must be checked. Data verification methods such as checksums and cyclic redundancy checks (CRCs) can be used to detect data errors. If the data is found to be incomplete or erroneous, the power parameters must be re-acquired or regenerated to ensure the accuracy of subsequent control processes.

[0191] If the power parameters are verified to be complete, the next step is to compare the parsed power parameters with the preset parameter range. The preset parameter range is determined based on factors such as the underwater thruster's design specifications, performance limitations, and safe operation requirements. It sets a reasonable range for each thruster's power parameters. For example, the thruster's speed may have a minimum and maximum value, and the thrust may also have upper and lower limits.

[0192] By comparing the parsed power parameters with these preset ranges one by one, it is possible to determine whether the current power parameters are within a safe and reasonable range. During the comparison process, the specific values of each parameter and the relationship between them are considered. For example, there may be a certain correlation between speed and thrust. When the speed is too high, the thrust may also increase. However, it is necessary to ensure that they are all within the preset range. If a power parameter is found to be outside the preset range during the comparison process, it means that there may be risks in operating the thruster according to the current parameters, such as equipment overheating, mechanical damage, or failure to achieve the expected movement effect. At this time, the power parameters need to be adjusted or regenerated until all parameters are within the preset range.

[0193] When the power parameters are determined to be within a preset range, they are converted into control signals that can be executed by the underwater propeller. Finally, the underwater propeller is controlled to move according to the control signals. The control signals are sent to the propeller's actuators, such as the motor controller and servo. The motor controller adjusts the motor's speed and direction based on the received control signals, thereby driving the propeller or other power components of the propeller to generate the corresponding thrust.

[0194] It should be noted that during the movement of the thruster, its operating status needs to be monitored in real time, including parameters such as speed, thrust, current, and temperature. These data are collected through sensors and fed back to the control system for comparison with the preset target values. If a deviation is found between the actual operating status and the target value, the control system can adjust the control signal in a timely manner and dynamically correct the power parameters of the thruster to ensure that the thruster can move according to the expected trajectory and speed. For example, when a sudden change in water flow causes the actual speed of the thruster to be lower than the target speed, the control system can increase the control signal, increase the speed of the thruster, and thus increase the thrust, so that the thruster returns to the expected state of motion. Through this closed-loop control method, accurate and stable control of the underwater thruster can be achieved, ensuring that it can complete various tasks safely and efficiently.

[0195] Please refer to Figure 7 According to some embodiments of the present invention, after comparing the analyzed power parameters with the preset parameter range in step 108, the following may be specifically included, but not limited to:

[0196] 701. When it is determined that the power parameter is not within the preset parameter range, triggering an alarm instruction;

[0197] 702. Generate an execution instruction according to the alarm instruction;

[0198] 703. Retrieve the maximum value parameter within the preset parameter range according to the execution instruction;

[0199] 704. Control the underwater propeller to move according to the maximum value parameter.

[0200] In an embodiment of the present application, when the generated power parameter is verified and it is determined that it is not within the preset parameter range, the system will quickly trigger an alarm instruction. After the alarm instruction is triggered, the system will immediately generate an execution instruction based on the instruction.

[0201] After generating the execution instruction, the system will call the maximum value parameter within the preset parameter range according to the instruction. The maximum value parameter within the preset parameter range has been selected and verified. It is the maximum driving force or the highest performance indicator that can be provided under the premise of ensuring the safe operation of the thruster. For example, among the speed parameters of the thruster, the maximum value parameter is determined under extreme conditions such as the heat dissipation capacity of the motor and the load-bearing capacity of the mechanical transmission components. After obtaining the maximum value parameter, the system will control the underwater thruster to move according to these parameters. This ensures that the thruster always operates in a safe and efficient state and completes the established underwater mission.

[0202] See also Figure 8 In a second aspect, the present application provides an underwater thruster control device based on multi-parameter fusion, comprising:

[0203] The acquisition unit 801 is used to collect water flow velocity and regional terrain information in the target area where the underwater propeller is located;

[0204] A first acquisition unit 802 is configured to establish a three-dimensional flow field real-time perception model based on the water flow velocity and the regional terrain information, and acquire multimodal data according to the three-dimensional flow field real-time perception model, wherein the multimodal data includes water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates, and terrain height of the target area;

[0205] A second acquisition unit 803 is configured to perform dispersed filtering and subsequent fusion on the multimodal data to obtain target multimodal data;

[0206] A construction unit 804 is configured to construct a multimodal neural network model based on the target multimodal data and through a temporal attention mechanism;

[0207] A first generating unit 805 is configured to input the real-time multimodal data of the underwater thruster into the multimodal neural network model and generate control instructions using a two-layer reinforcement learning strategy;

[0208] A second generating unit 806 is configured to generate power parameters of the underwater propulsion device according to the control instruction;

[0209] The control unit 807 is configured to control the underwater propulsion device to move according to the power parameter when it is determined that the power parameter is within a preset parameter range.

[0210] See also Figure 9 , the present application also provides an underwater thruster control device based on multi-parameter fusion, comprising:

[0211] Processor 901, memory 902, input / output unit 903, bus 904;

[0212] The processor 901 is connected to the memory 902 , the input and output unit 903 , and the bus 904 ;

[0213] The storage 902 stores a program, and the processor 901 calls the program to execute any of the above methods.

[0214] The present application also relates to a computer-readable storage medium, on which a program is stored. When the program is run on a computer, the computer is caused to execute any of the above methods.

[0215] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0216] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0217] Units described as separate components may or may not be physically separate, and 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 these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0218] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0219] If the integrated unit is implemented in the form of 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 the present application, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk, and other media that can store program code.

Claims

1. A method for controlling underwater thrusters based on multi-parameter fusion, characterized in that: The method comprises: Collect water flow velocity and regional terrain information in the target area where the underwater propeller is located; Establishing a three-dimensional flow field real-time perception model based on the water flow velocity and the regional terrain information, and acquiring multimodal data according to the three-dimensional flow field real-time perception model, wherein the multimodal data includes water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates, and terrain height of the target area; Distributing and filtering the multimodal data and then fusing them to obtain target multimodal data; Based on the target multimodal data, a multimodal neural network model is constructed through a temporal attention mechanism; Acquire real-time multimodal data; Preprocessing the real-time multimodal data to form an input feature vector adapted to the multimodal neural network model; Inputting the input feature vector into the multimodal neural network model to obtain a target feature vector; Building a two-layer reinforcement learning strategy network, the two-layer reinforcement learning strategy network includes a high-level decision network and a low-level execution network, the high-level decision network is used to receive the target feature vector output by the multimodal neural network model and plan the control strategy, and the low-level execution network is used to refine the control strategy into control parameters; Inputting the target feature vector into the two-layer reinforcement learning policy network and updating the policy network parameters using the Actor-Critic algorithm; Generate control instructions based on the policy network parameters and in combination with the current environment state; generating power parameters of the underwater thruster according to the control instruction; When it is determined that the power parameter is within a preset parameter range, the underwater propeller is controlled to move according to the power parameter.

2. The underwater thruster control method based on multi-parameter fusion according to claim 1 is characterized in that: The multimodal data is dispersedly filtered and then fused to obtain target multimodal data, including: Preprocessing the multimodal data to obtain preprocessed data; extracting water flow data and terrain data from the pre-processed data; According to the data characteristics of water flow data and terrain data, Kalman filter algorithm and Gaussian filter algorithm are used to eliminate outliers respectively; Perform data registration on the filtered water flow data and terrain data; The weighted average algorithm is used to fuse the data after data registration; The weighted average algorithm is: Where S is the weighted average, n is the number of data points, and x i is the value of the i-th data point, w i is the weight of the i-th data point; The compared fused data is verified, and when it is determined that the data verification passes, the target multimodal data is output after the data is smoothed.

3. The underwater thruster control method based on multi-parameter fusion according to claim 1, characterized in that: Based on the target multimodal data, a multimodal neural network model is constructed through a temporal attention mechanism, including: Extracting numerical, spatial, and temporal features from the target multimodal data and converting them into a unified dimensional vector; Arranging the unified dimension vectors in time series to obtain a feature matrix; According to the data characteristics of the feature matrix, a multimodal neural network model is constructed using a temporal attention mechanism.

4. The underwater thruster control method based on multi-parameter fusion according to claim 1, characterized in that: Generating the power parameters of the underwater propeller according to the control instruction includes: Calculating the output power of the underwater thruster according to the control instruction; comparing the output power of the underwater propeller with the maximum output power of the underwater propeller and obtaining a comparison result; The power parameters of the underwater propeller are generated according to the comparison result.

5. The underwater thruster control method based on multi-parameter fusion according to claim 1, characterized in that: When it is determined that the power parameter is within a preset parameter range, controlling the underwater propeller to move according to the power parameter includes: Analyze the power parameters and verify whether the power parameters are complete; If so, the parsed dynamic parameters are compared with the preset parameter range; When it is determined that the power parameter is within a preset parameter range, converting the power parameter into a control signal executable by the underwater propulsion device; The underwater propeller is controlled to move according to the control signal.

6. The underwater thruster control method based on multi-parameter fusion according to claim 5 is characterized in that: After comparing the analyzed power parameters with the preset parameter ranges, the method further includes: When it is determined that the power parameter is not within the preset parameter range, an alarm instruction is triggered; generating an execution instruction according to the alarm instruction; Retrieving the maximum value parameter within the preset parameter range according to the execution instruction; The underwater propeller is controlled to move according to the maximum value parameter.

7. An underwater thruster control device based on multi-parameter fusion, characterized in that: The device comprises: A collection unit, used to collect water flow velocity and regional terrain information in the target area where the underwater propeller is located; a first acquisition unit, configured to establish a three-dimensional flow field real-time perception model based on the water flow velocity and the regional terrain information, and acquire multimodal data according to the three-dimensional flow field real-time perception model, wherein the multimodal data includes water flow velocity, water flow pressure, water flow intensity, three-dimensional terrain coordinates, and terrain height of the target area; A second acquisition unit is used to perform dispersion filtering and subsequent fusion on the multimodal data to obtain target multimodal data; A construction unit, configured to construct a multimodal neural network model based on the target multimodal data and through a temporal attention mechanism; A first generation unit is configured to obtain real-time multimodal data, pre-process the real-time multimodal data to form an input feature vector adapted to the multimodal neural network model, input the input feature vector into the multimodal neural network model to obtain a target feature vector, and build a two-layer reinforcement learning policy network, wherein the two-layer reinforcement learning policy network includes a high-level decision network and a low-level execution network, the high-level decision network is configured to receive the target feature vector output by the multimodal neural network model and plan a control policy, the low-level execution network is configured to refine the control policy into control parameters, input the target feature vector into the two-layer reinforcement learning policy network, and use an Actor-Critic algorithm to update policy network parameters, and generate control instructions based on the policy network parameters and in combination with the current environment state; A second generating unit, configured to generate power parameters of the underwater propeller according to the control instruction; A control unit is used to control the underwater propeller to move according to the power parameter when it is determined that the power parameter is within a preset parameter range.

8. An underwater thruster control device based on multi-parameter fusion, characterized in that: The device comprises: processor, memory, input and output units, and buses; The processor is connected to the memory, the input and output unit, and the bus; The storage stores a program, and the processor calls the program to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, and when the program is executed on a computer, the method according to any one of claims 1 to 6 is executed.

Images