Underwater multi-body dynamic forecasting method

Abstract

An underwater multi-body power forecasting method comprises the following steps: constructing a coupling force forecasting basic model, and carrying out single-working-condition and multi-working-condition model training; training the coupling force prediction basic model by using the experimental data mixed by various working conditions to obtain a multi-working-condition prediction model with high generalization; constructing an integrated coupling force prediction model; selecting a final coupling force prediction model; comparing errors between the actually measured poses of the sensors and the predicted poses of the models in real time, and dynamically selecting the model with the minimum error as a final coupling force prediction model; determining a final coupling force predicted value; the coupling force prediction model is constructed based on the convolutional neural network, the coupling force at the joint of the underwater equipment and the mechanical arm is trained and learned, the complex nonlinear relation between the coupling force and underwater equipment movement and the complex nonlinear relation between the coupling force and mechanical arm movement can be more effectively captured, an accurate dynamic model and an accurate physical model do not need to be relied on, and the method is easy to implement. Therefore, the complexity of the model and the dependence on priori knowledge are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater equipment prediction methods, in particular to an underwater multi-body dynamic prediction method. Background Art

[0002] Underwater equipment has important application value in the fields of marine resource exploration and seabed environmental monitoring. Among them, underwater equipment equipped with robotic arms is key equipment, which can complete complex seabed operation tasks, such as grabbing samples and installing equipment. However, in actual operations, the movement of the robotic arm will produce significant coupling force interference on the overall movement of the underwater equipment. This dynamic coupling effect is even more complicated in scenarios involving the load of mooring facilities (such as laying out and maintaining mooring systems or working in conjunction with fixed offshore platforms). The coupling force will not only change the posture and trajectory of the underwater equipment, but also transmit additional loads through the rigid connection of the mooring facilities, further reducing the operation accuracy and causing systematic errors during task execution. To solve this problem, it is crucial to accurately predict and evaluate the coupling force.

[0003] Although torque sensors can currently measure coupling forces, their high cost and limited operating depth make practical application challenging. Therefore, exploring more economical and practical methods for predicting the coupling forces between underwater equipment and robotic arms under the load of mooring facilities, and constructing experimental devices and methods capable of simulating the dynamic constraints of mooring systems, are of great theoretical and engineering significance for improving operational accuracy under complex load scenarios.

[0004] Publication number CN115618574A discloses an Adams / Matlab co-simulation method for underwater equipment and robotic arm systems. This method primarily utilizes Adams software for multi-body dynamics modeling and Matlab for control system design, thereby achieving co-simulation of the entire system. During the simulation, the effect of the robotic arm's motion on the robot's body posture and motion state can be observed, allowing the magnitude and direction of the disturbance force exerted by the robotic arm on the body to be predicted. This method avoids the high cost of sensors, but in engineering applications, establishing the connection between the underwater equipment and robotic arm entities and the computer simulation environment remains to be studied.

[0005] Publication number CN118404584A discloses an adaptive disturbance dynamics control method suitable for operational underwater equipment. This method establishes a dynamic model that includes the motion of the manipulator and uses an adaptive control algorithm to estimate changes in the system's center of gravity and center of buoyancy in real time, thereby predicting the static coupling force exerted by the manipulator on the carrier. Simultaneously, a nonlinear disturbance observer is designed to estimate system model uncertainties and external disturbances such as ocean currents, and to compensate for them in real time. This method is highly applicable but requires a precise dynamic model and extensive computation, and is difficult to adapt to changes in model parameters and environmental disturbances.

[0006] Publication number CN113093771A discloses a neural network-based modeling method and system for underwater equipment-manipulator systems. This method utilizes a single hidden layer feedforward neural network to approximate unknown dynamic functions, combined with position error feedback and a weight matrix update law, to achieve overall dynamic modeling of the underwater equipment-manipulator system. This method reduces reliance on model accuracy. Furthermore, while a significant amount of computation is required during model training, the computational effort is minimal during use. This method considers coupling forces to be internal to the underwater equipment-manipulator system and implicitly embedded in the overall system model, without specifically targeting coupling forces for modeling, training, or output.

[0007] Publication number CN118862617A discloses a neural network-based method for predicting the coupling force of underwater working devices. This method utilizes a neural network regression model combined with a particle swarm algorithm, using operating parameters as input and the actual coupling force values ​​measured by tension and pressure sensors as output. This method trains a coupling force prediction model to predict coupling force under future operating conditions. While this method models the coupling force separately, the coupling force prediction model constructed is based on the tension of the buoy on the anchor chain, the tension of the anchor on the anchor chain, and the tension of the anchor chain on the robotic arm, and does not establish a coupling force prediction model for the connection between the underwater equipment and the robotic arm. Summary of the Invention

[0008] In response to the shortcomings of the above-mentioned existing production technologies, the applicant provides an underwater multi-body dynamic prediction method, thereby constructing a coupling force prediction model based on a convolutional neural network, training and learning the coupling force at the connection between underwater equipment and robotic arms, and more effectively capturing the complex nonlinear relationship between the coupling force and the movement of underwater equipment and robotic arms, without relying on precise dynamic models and precise physical models, thereby reducing the complexity of the model and dependence on prior knowledge.

[0009] The technical solutions adopted in the present invention are as follows:

[0010] A method for underwater multi-body dynamic prediction includes the following steps: S1: constructing a coupling force prediction basic model; constructing the coupling force prediction basic model based on a convolutional neural network, wherein the network architecture is a multi-layer cascade structure consisting of an input layer, multiple convolution modules, a regularization module, and a regression output layer; S2: training single-condition and multi-condition models; using experimental data of a mixture of multiple conditions to train the coupling force prediction basic model to obtain a multi-condition prediction model with strong generalization; S3: constructing an integrated coupling force prediction model; integrating the single-condition prediction model set obtained by training in S2 with the multi-condition prediction model to obtain an integrated coupling force prediction model, wherein the integrated model can compare the error between the actual measured posture of the sensor and the predicted posture of each model in real time; S4: selecting a final coupling force prediction model; by comparing the error between the actual measured posture of the sensor and the predicted posture of each model in real time, dynamically selecting the model with the smallest error as the final coupling force prediction model; S5: determining the final coupling force prediction value; and using the coupling force value output by the final coupling force prediction model as the final coupling force prediction value.

[0011] Its further technical solution is:

[0012] In S1, the input layer receives multi-dimensional data including thruster control quantities, robot posture state quantities and robot arm joint angles. In S1, the convolution module uses zero-filled convolution kernels and maximum pooling layers to alternately stack to achieve feature extraction and dimensionality reduction. In S1, the end of the network is mapped to the output space through a fully connected layer to simultaneously predict the coupling force and robot posture state quantities. In S2, for each independent working condition, a coupling force prediction basic model is trained using a single working condition data to obtain a highly targeted single working condition prediction model. In S5, the actual posture measured by the underwater equipment sensor is compared with the underwater equipment posture prediction value output by each regression layer to obtain the prediction error. The multi-working condition coupling force prediction model or the single working condition coupling force prediction model corresponding to the minimum prediction error is the final coupling force prediction model.

[0013] The beneficial effects of the present invention are as follows:

[0014] The present invention constructs a coupling force prediction model based on a convolutional neural network, and trains and learns the coupling force at the connection between underwater equipment and the robotic arm. It can more effectively capture the complex nonlinear relationship between the coupling force and the movement of the underwater equipment and the robotic arm, and does not need to rely on precise dynamic models and precise physical models, thereby reducing the complexity of the model and dependence on prior knowledge.

[0015] The present invention uses a known method to simultaneously predict the output values ​​of three independent sensors, which makes accurate prediction difficult. The present invention only predicts the output value of one independent sensor, which makes accurate prediction less difficult.

[0016] In the actual underwater equipment-mechanical arm system and anchoring facility load of the present invention, three sensors, namely the sensor between the anchor chain and the buoy, the sensor between the anchor chain and the anchor, and the sensor between the anchor chain and the mechanical arm, do not exist. The coupling force value output by the known method prediction model is difficult to evaluate its prediction error due to the lack of actual measurement values ​​of relevant sensors for comparison. The prediction model established by the present invention not only outputs the coupling force value, but also outputs the posture of the underwater equipment, and compares the underwater equipment posture output by the prediction model with the actual posture measured by the underwater equipment sensor to obtain a prediction error, and indirectly evaluates the coupling force prediction error based on this prediction error.

[0017] The present invention expands a multi-working condition coupling force prediction model of a known method into a combination of a multi-working condition coupling force prediction model and multiple single-working condition coupling force prediction models, and indirectly determines the final prediction model with the smallest coupling force prediction error by evaluating the underwater equipment posture prediction error. The coupling force value output by the final prediction model is used as the coupling force prediction value at the current moment, further improving the prediction accuracy.

[0018] The present invention is applicable to a method for predicting the coupling force between underwater equipment and a mechanical arm under the load of an anchoring facility. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Flowchart of the present invention.

[0020] Figure 2 This is an overall schematic diagram of the basic model for coupling force prediction of the present invention.

[0021] Figure 3 This is an overall schematic diagram of the integrated coupling force prediction model of the present invention.

[0022] Figure 4 This is a flow chart of the experimental method for predicting the coupling force between underwater equipment and robotic arms under the load of the anchoring facility of the present invention.

[0023] Figure 5 This is a schematic diagram of the overall training of the coupling force prediction model in the present invention.

[0024] Figure 6a This is a comparison chart of the output of the training set coupling force prediction model and the actual results of the sensor under the working condition that the longitudinal thruster thrust of the present invention is 0.5V (feature 1).

[0025] Figure 6b This is a comparison chart of the output of the training set coupling force prediction model and the actual results of the sensor under the working condition that the longitudinal thruster thrust of the present invention is 0.5V (feature 2).

[0026] Figure 6cThis is a comparison chart of the output of the training set coupling force prediction model and the actual results of the sensor under the working condition that the longitudinal thruster thrust of the present invention is 0.5V (feature 3).

[0027] Figure 6d This is a comparison chart of the output of the training set coupling force prediction model and the actual results of the sensor under the working condition that the longitudinal thruster thrust of the present invention is 0.5V (feature 4).

[0028] Figure 6e This is a comparison chart of the output of the training set coupling force prediction model and the actual results of the sensor under the working condition that the longitudinal thruster thrust of the present invention is 0.5V (feature 5).

[0029] Figure 6f This is a comparison chart of the output of the training set coupling force prediction model and the actual results of the sensor under the working condition that the longitudinal thruster thrust of the present invention is 0.5V (feature 6).

[0030] Figure 7a This is a comparison chart of the output of the coupling force prediction model of the test set and the actual result of the sensor under the working condition that the thrust of the longitudinal thruster of the present invention is 0.5V (feature 1).

[0031] Figure 7b This is a comparison chart of the output of the coupling force prediction model of the test set and the actual result of the sensor under the working condition that the thrust of the longitudinal thruster of the present invention is 0.5V (feature 2).

[0032] Figure 7c This is a comparison chart of the output of the coupling force prediction model of the test set and the actual result of the sensor under the working condition that the thrust of the longitudinal thruster of the present invention is 0.5V (feature 3).

[0033] Figure 7d This is a comparison chart of the output of the coupling force prediction model of the test set and the actual result of the sensor under the working condition that the thrust of the longitudinal thruster of the present invention is 0.5V (feature 4).

[0034] Figure 7e This is a comparison chart of the output of the coupling force prediction model of the test set and the actual result of the sensor under the working condition that the thrust of the longitudinal thruster of the present invention is 0.5V (feature 5).

[0035] Figure 7f This is a comparison chart of the output of the coupling force prediction model of the test set and the actual result of the sensor under the working condition that the thrust of the longitudinal thruster of the present invention is 0.5V (feature 6).

[0036] Figure 8a This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.5V (x-direction coupling force).

[0037] Figure 8bThis is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.5V (z-direction coupling force).

[0038] Figure 8c This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.5V (coupling force around the y direction).

[0039] Figure 8d This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.5V (longitudinal displacement).

[0040] Figure 8e This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.5V (vertical displacement).

[0041] Figure 8f This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.5V (pitch angle).

[0042] Figure 9a This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.6V (x-direction coupling force).

[0043] Figure 9b This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.6V (z-direction coupling force).

[0044] Figure 9c This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.6V (coupling force around the y direction).

[0045] Figure 9d This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.6V (longitudinal displacement).

[0046] Figure 9e This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.6V (vertical displacement).

[0047] Figure 9f This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal propeller thrust is 0.6V (pitch angle).

[0048] Figure 10a This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.7V (x-direction coupling force).

[0049] Figure 10bThis is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.7V (z-direction coupling force).

[0050] Figure 10c This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.7V (coupling force around the y direction).

[0051] Figure 10d This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.7V (longitudinal displacement).

[0052] Figure 10e This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.7V (vertical displacement).

[0053] Figure 10f This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.7V (pitch angle).

[0054] Figure 11a This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.8V (x-direction coupling force).

[0055] Figure 11b This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.8V (coupling force in the z direction).

[0056] Figure 11c This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.8V (coupling force around the y direction).

[0057] Figure 11d This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.8V (longitudinal displacement).

[0058] Figure 11e This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.8V (vertical displacement).

[0059] Figure 11f This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal propeller thrust is 0.8V (pitch angle).

[0060] Figure 12a This is a comparison diagram of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.9V (x-direction coupling force).

[0061] Figure 12bThis is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.9V (coupling force in the z direction).

[0062] Figure 12c This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.9V (coupling force around the y direction).

[0063] Figure 12d This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.9V (longitudinal displacement).

[0064] Figure 12e This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.9V (vertical displacement).

[0065] Figure 12f This is a comparison chart of the prediction effects of various coupling force prediction models of the present invention under the condition that the longitudinal thruster thrust is 0.9V (pitch angle). DETAILED DESCRIPTION

[0066] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0067] The underwater multi-body dynamic prediction method of this embodiment includes the following operating steps: S1: constructing a basic model for coupling force prediction; constructing a basic model for coupling force prediction based on a convolutional neural network, the network architecture is a multi-layer cascade structure consisting of an input layer, multiple convolution modules, a regularization module and a regression output layer; the input layer receives multi-dimensional data including thruster control quantities, robot posture state quantities and robot arm joint angles; the convolution module uses zero-filled convolution kernels and maximum pooling layers to alternately stack to achieve feature extraction and dimensionality reduction; the network end is mapped to the output space through a fully connected layer to synchronously predict the coupling force and robot posture state quantities.

[0068] S2: Conduct single-condition and multi-condition model training; use experimental data of multiple working conditions to train the coupling force prediction basic model to obtain a multi-condition prediction model with strong generalization; for each independent working condition, use single working condition data to train the coupling force prediction basic model to obtain a highly targeted single-condition prediction model.

[0069] S3: Construct an integrated coupling force prediction model. The single-condition prediction model set trained in S2 is integrated with the multi-condition prediction model to obtain an integrated coupling force prediction model. This integrated model can compare the error between the sensor's measured pose and the pose predicted by each model in real time.

[0070] S4: Select the final coupling force prediction model; by comparing the errors between the actual sensor posture and the posture predicted by each model in real time, dynamically select the model with the smallest error as the final coupling force prediction model;

[0071] S5: Determine the final coupling force prediction value; use the coupling force value output by the final coupling force prediction model as the final coupling force prediction value, compare the actual posture measured by the underwater equipment sensor with the underwater equipment posture prediction value output by each regression layer, and obtain the prediction error. The multi-condition coupling force prediction model or the single-condition coupling force prediction model corresponding to the minimum prediction error is the final coupling force prediction model.

[0072] When applied to the prediction of the coupling force between underwater equipment and the manipulator under the load of mooring facilities, the steps are as follows:

[0073] The first step is to build a basic model for coupling force prediction based on convolutional neural network.

[0074] like Figure 2 As shown in the figure, the network architecture of the model consists of an N-layer cascade structure: after the input layer receives M×1 single-channel data, it first passes through the first convolution layer containing a K×1 convolution kernel. This layer uses zero-padding technology to keep the number of features unchanged and generates feature maps of P channels. It is then connected to a batch normalization layer and a ReLU activation function to accelerate training convergence; then a Q×1 maximum pooling layer with a step size of S is configured to achieve feature dimensionality reduction. The second convolution layer then uses a K×1 convolution kernel extended to 2×P channels for deep feature extraction, and is also equipped with batch normalization and nonlinear activation modules; to prevent the model from overfitting, a Dropout regularization layer with a drop probability of D% is added in the middle of the network. At the end, the high-level features are mapped to the R-dimensional output space through a fully connected layer, and finally the numerical prediction task is completed with a regression layer. Among them, N, M, K, P, S, Q, D, and R are integers. M is the number of input variables, including the underwater equipment thruster control quantity, posture state quantity, and robotic arm joint angle. R is the number of output variables, including the coupling force at the connection between the underwater equipment and the robotic arm, and the underwater equipment posture state quantity.

[0075] Specifically, such as Figure 2 As shown in the figure, the network architecture of the model consists of a multi-layer cascade structure: after the input layer receives 18×1 single-channel data, it first passes through the first convolution layer containing a 3×1 convolution kernel. This layer uses zero-padding technology to keep the number of features unchanged and generates a feature map of 16 channels. It is then connected to a batch normalization layer and a ReLU activation function to accelerate training convergence; then a 2×1 maximum pooling layer with a step size of 1 is configured to achieve feature dimensionality reduction. The second convolution layer then uses a 3×1 convolution kernel expanded to 2×16 channels for deep feature extraction, and is also equipped with batch normalization and nonlinear activation modules; to prevent the model from overfitting, a Dropout regularization layer with a discarding probability of 10% is added in the middle of the network. At the end, the high-level features are mapped to a 6-dimensional output space through a fully connected layer, and finally the numerical prediction task is completed with a regression layer.

[0076] The second step is to train a multi-condition coupling force prediction model. Randomly mix the experimental data under multiple working conditions, and train the coupling force prediction basic model based on the mixed data. When the training accuracy converges, a multi-condition coupling force prediction model is obtained. The third step is to train a single-condition coupling force prediction model. Select the experimental data under one working condition, and train the coupling force prediction basic model based on the experimental data under the working condition. When the training accuracy converges, a single-condition coupling force prediction model is obtained. The fourth step is to train multiple single-condition coupling force prediction models. Repeat the third step, and train the coupling force prediction basic model based on the experimental data under each working condition. When the training accuracy converges, multiple single-condition coupling force prediction models are obtained. The fifth step is to construct an integrated coupling force prediction model. Integrate the multi-condition coupling force prediction model obtained in the second step with the n single-condition coupling force prediction models obtained in the fourth step to obtain an integrated coupling force prediction model, such as Figure 3 As shown. The integrated coupling force prediction model uses the same input layer. Each regression layer outputs the coupling force prediction value and the underwater equipment posture prediction value respectively. The actual posture measured by the underwater equipment sensor is introduced and compared with the underwater equipment posture prediction value output by each regression layer. The sixth step is to select the final coupling force prediction model. The actual posture measured by the underwater equipment sensor is compared with the underwater equipment posture prediction value output by each regression layer to obtain the prediction error. The multi-condition coupling force prediction model or the single-condition coupling force prediction model corresponding to the minimum prediction error is the final coupling force prediction model. The seventh step is to determine the final coupling force prediction value. The coupling force value output by the final coupling force prediction model is used as the final coupling force prediction value.

[0077] like Figure 4 As shown, an experimental method for analyzing the prediction method of the coupling force between underwater equipment and a manipulator under the load of an anchoring facility includes the following steps:

[0078] The first step is to build the anchoring facility load test device and fill the pool with water to complete the configuration of the experimental environment and experimental device. The experimental device of this embodiment mainly includes an underwater equipment-manipulator system, a pool, and an anchoring facility load simulation device. The anchoring facility load simulation device includes a float, an anchor chain, and a fixing; wherein one end of the anchor chain is connected to the float, the other end is connected to the fixing, and the middle part of the anchor chain is connected to the end of the manipulator. The underwater equipment-manipulator system includes underwater equipment, an underwater manipulator, and a six-axis force sensor. Among them, the six-axis force sensor is installed between the underwater equipment and the manipulator base. The top of the six-axis force sensor is connected to the underwater equipment and the bottom is connected to the manipulator base, so that the three-axis coupling force and coupling torque between the underwater equipment and the manipulator can be measured in real time. The underwater equipment includes an outer shell, an external frame, an electronic cabin, a vertical thruster, a lateral thruster, a longitudinal thruster, an attitude sensor, a Doppler velocity sensor, a depth gauge, etc. The thrusters are installed at the head and tail of the external frame, the electronic compartment is installed in the internal cavity of the external frame, the Doppler velocity sensor is fixed to the bottom of the external frame, the attitude sensor is installed inside the electronic compartment, the depth gauge is fixed to the tail of the external frame, and the outer shell is covered on the external frame as a whole by screws. The second step is to conduct a water tank test of the underwater equipment-manipulator system's mooring facility load. The underwater equipment-manipulator system is used to grab the anchor chain of the mooring facility load simulation device. Under the conditions of longitudinal thruster control voltage of 0.5v, 0.6v, 0.7v, 0.8v, and 0.9v, and vertical thruster control voltage of 0.6V, the anchoring facility load simulation device is carried out in the vertical plane movement. During the movement, the bow angle of the underwater equipment is kept unchanged by two lateral thrusters. The vertical displacement of the underwater equipment is measured using a depth gauge and the vertical velocity is derived from the derivative. The longitudinal velocity is measured using a Doppler velocity sensor and integrated to obtain the longitudinal displacement. The pitch angle of the underwater equipment is measured using an attitude sensor and the pitch angular velocity is derived. The voltage returned by the six servos of the manipulator is converted into the angles of the corresponding manipulator joints. The six-axis force sensor between the manipulator base and the underwater equipment is used to determine the coupling forces on the X and Z axes and the coupling torque about the Y axis. The experimental data obtained under each working condition is randomly divided into training and test sets for future use. The third step is to build a convolutional neural network model.Its network architecture consists of a multi-layer cascade structure: after the input layer receives 18×1 single-channel data, it first passes through the first convolution layer containing a 3×1 convolution kernel. This layer uses zero-padding technology to keep the number of features unchanged and generates a 16-channel feature map. It is then connected to a batch normalization layer and a ReLU activation function to accelerate training convergence; then a 2×1 maximum pooling layer with a step size of 1 is configured to achieve feature dimensionality reduction. The second convolution layer then uses a 3×1 convolution kernel expanded to 32 channels for deep feature extraction, and is also equipped with batch normalization and nonlinear activation modules; to prevent the model from overfitting, a Dropout regularization layer with a discarding probability of 10% is added in the middle of the network. At the end, the high-level features are mapped to a 6-dimensional output space through a fully connected layer, and finally the numerical prediction task is completed with a regression layer.

[0079] The fourth step is to train a convolutional neural network regression model to obtain a coupling force prediction model. Using the control voltages of the four vertical thrusters, the control voltages of the two longitudinal thrusters, the longitudinal displacement and longitudinal velocity of the underwater equipment, the vertical displacement and vertical velocity of the underwater equipment, the pitch angle and pitch angular velocity of the underwater equipment, and the angles of the six joints of the manipulator at the previous moment in the training set under different working conditions as outputs, and using the X- and Z-axis coupling forces, coupling torque around the Y-axis, longitudinal displacement, vertical displacement, and pitch angle between the underwater equipment and the manipulator at the current moment in the training set under the corresponding working conditions as outputs, a convolutional neural network regression model is trained to obtain one multi-working condition coupling force prediction model and five single-working condition coupling force prediction models.

[0080] In this embodiment, the experimental results of the longitudinal thruster control voltage of 0.5V are used as an example for explanation. Based on the performance of the convolutional neural network coupling force prediction model training set constructed under the longitudinal thruster control voltage of 0.5V, a multi-subgraph structure is used to compare and analyze the model prediction effect: each subgraph presents the dynamic matching relationship between the nine sets of output parameters and the sensor measured data through the timing curve of the red dotted line (predicted value) and the blue solid line (measured data). The root mean square error (RMSE) between the corresponding parameter and the actual data is marked in the title of each subgraph to quantify the prediction deviation. According to the performance evaluation results of the test set, the model still maintains the synchronous fluctuation characteristics of the red and blue curves in samples that did not participate in training. The fifth step is to obtain the parameters of the underwater equipment-manipulator system in the working condition in actual application, including the underwater equipment vertical thruster control voltage, longitudinal thruster control voltage, underwater equipment longitudinal displacement and longitudinal velocity, underwater equipment vertical displacement and vertical velocity, underwater equipment pitch angle and pitch angular velocity, and manipulator joint angle at the previous moment. These parameters are input into one multi-working condition coupling force prediction model and five single-working condition coupling force prediction models. Each prediction model outputs six parameters, including the current moment x-direction coupling force, z-direction coupling force, coupling torque around the y-axis, underwater equipment longitudinal displacement, underwater equipment vertical displacement, and underwater equipment pitch angle. By comparing the longitudinal displacement, vertical displacement, and pitch angle output by each coupling force prediction model with the actual longitudinal displacement, vertical displacement, and pitch angle measured by the underwater equipment sensor at the current moment, the underwater equipment posture prediction error is obtained. The coupling force prediction model with the smallest underwater equipment posture prediction error is selected as the final coupling force prediction model under the current working condition, thereby predicting the coupling force and coupling torque with higher accuracy.

[0081] In this embodiment, another set of original experimental data obtained under the condition of longitudinal thruster control voltage 0.5V is input into a multi-condition coupling force prediction model and five single-condition coupling force prediction models trained by the anchoring facility load tank test, and six sets of output parameters are obtained. The six sets of output parameters are compared with the corresponding parameters in the original experimental data, and the six sets of root mean square errors are calculated, thereby obtaining a comparison of the prediction errors of the six networks under the same output parameters, as shown in FIG. Figure 8a-8f As shown in the figure, the network version numbers 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0 represent the single-condition coupling force prediction model under the longitudinal thruster control voltage of 0.5V, 0.6V, 0.7V, 0.8V, and 0.9V, respectively, and a multi-condition coupling force prediction model. Figure 9a-9fAmong the longitudinal displacement, vertical displacement, and pitch angle of the underwater equipment output by each version of the network, the root mean square error of these three parameters output by the network with network version 0.5 is the smallest. According to the patented method of the present invention, that is, the network with network version 0.5 is selected as the final coupling force prediction model under this working condition, under the network prediction of network version 0.5, the root mean square errors of the x-direction coupling force, the z-direction coupling force, and the coupling torque around the y-axis are also smaller than the root mean square errors predicted by other network versions.

[0082] Another set of original experimental data obtained under the condition of longitudinal thruster control voltage of 0.6V was input into a multi-condition coupling force prediction model and five single-condition coupling force prediction models trained by the anchoring facility load tank test. Six sets of output parameters were obtained. The six sets of output parameters were compared with the corresponding parameters in the original experimental data, and the six sets of root mean square errors were calculated, thereby obtaining a comparison of the prediction errors of the six networks under the same output parameters, as shown in the figure. Figure 9a-9f As shown in the figure, it can be concluded that among the longitudinal displacement, vertical displacement, and pitch angle of the underwater equipment output by each network, the root mean square error of the longitudinal displacement and vertical displacement parameters output by the network with network version 0.6 is the smallest, and the root mean square error of the pitch angle parameter output by the network with network version 1.0 is the smallest. In summary, the network with network version 0.6 is selected as the final coupling force prediction model under this working condition. Under the prediction of the network with network version 0.6, the root mean square errors of the x-direction coupling force, the z-direction coupling force and the coupling torque around the y-axis are also smaller than the mean square errors predicted by other network versions.

[0083] As all Figure 10a-10f 、 Figure 11a-Figure 11f 、 Figure 12a-12f As shown in the figure, relevant conclusions can be drawn. This proves the effectiveness of the proposed method and improves the prediction accuracy to a certain extent.

[0084] The above description is an explanation of the present invention, not a limitation of the present invention. The scope of the present invention is defined in the claims. Any modifications may be made within the scope of protection of the present invention.

Claims

1. A method for underwater multi-body dynamic prediction, characterized by: The steps are as follows: S1: Construct a basic model for coupling force prediction; A coupling force prediction basic model is constructed based on a convolutional neural network. The network architecture consists of a multi-layer cascade structure consisting of an input layer, multiple convolution modules, a regularization module, and a regression output layer. S2: Conduct single-condition and multi-condition model training; The coupling force prediction basic model is trained using experimental data from a variety of working conditions to obtain a multi-working condition prediction model with strong generalization ability. S3: Construct an integrated coupling force prediction model; The single-condition prediction model set obtained through S2 training is integrated with the multi-condition prediction model to obtain an integrated coupling force prediction model. This integrated model can compare the error between the actual sensor pose and the pose predicted by each model in real time. S4: Select the final coupling force prediction model; By comparing the errors between the actual pose measured by the sensor and the pose predicted by each model in real time, the model with the smallest error is dynamically selected as the final coupling force prediction model; S5: Determine the final coupling force prediction value; The coupling force value output by the final coupling force prediction model is used as the final coupling force prediction value.

2. The underwater multi-body dynamic prediction method according to claim 1, characterized in that: In S1, the input layer receives multidimensional data including thruster control quantity, robot posture state quantity and manipulator joint angle.

3. The underwater multi-body dynamic prediction method according to claim 1, characterized in that: In S1, the convolution module uses zero-filled convolution kernels and maximum pooling layers stacked alternately to achieve feature extraction and dimensionality reduction.

4. The underwater multi-body dynamic prediction method according to claim 1, characterized in that: In S1, the network end is mapped to the output space through a fully connected layer, and the coupling force and robot posture state are predicted simultaneously.

5. The underwater multi-body dynamic prediction method according to claim 1, characterized in that: In S2, for each independent working condition, a coupling force prediction basic model is trained using single working condition data to obtain a highly targeted single working condition prediction model.

6. The underwater multi-body dynamic prediction method according to claim 1, characterized in that: In S5, the actual posture measured by the underwater equipment sensor is compared with the underwater equipment posture prediction value output by each regression layer to obtain the prediction error. The multi-condition coupling force prediction model or the single-condition coupling force prediction model corresponding to the minimum prediction error is the final coupling force prediction model.

Images