Unmanned ship path tracking method for multi-scene randomized dynamics simulation and DNN proxy model

Abstract

The invention discloses an unmanned ship path tracking method for multi-scene randomized dynamics simulation and a DNN proxy model. The method comprises the following steps: randomly generating multiple groups of sea condition parameters based on a real ocean statistical model; under each group of sea condition parameters, dynamic simulation of the unmanned ship is carried out by changing the combination of controller parameters and filter coefficients, and error indexes are recorded; the controller parameters, the filter coefficients and the sea condition parameters are used as inputs, the error indexes are used as outputs, and the inputs and the corresponding outputs are used as a group of training samples; constructing a deep neural network model taking the sea condition parameters and the controller parameters as inputs and the error indexes as outputs, and training the deep neural network model by using the training samples to obtain an agent model for predicting the error indexes; and automatically adjusting controller parameters or filter coefficients of the unmanned ship by adopting a particle swarm optimization algorithm according to the error index predicted value.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned watercraft path tracking, and in particular to an unmanned watercraft path tracking method for multi-scenario randomized dynamics simulation and a DNN proxy model. Background Art

[0002] Unmanned vehicles (UAVs) are a key component of intelligent autonomous systems and are widely used in ocean exploration, environmental monitoring, security patrols, and other fields. Path tracking control is a key issue in autonomous navigation for UAVs, requiring precise movement along a predetermined path while overcoming interference from complex hydrodynamic environments (such as wind, waves, and currents).

[0003] PID control is a classic control algorithm widely used in industrial control, with the advantages of simple structure and easy implementation. However, in the path tracking of unmanned aerial vehicles, PID control has the following shortcomings: Difficulty in parameter adjustment: Traditional PID control relies on manual or trial-and-error parameter tuning, which is difficult to adapt to different environments and mission requirements. Poor adaptability to nonlinear systems: The dynamic models of unmanned aerial vehicles typically have strong nonlinear and time-varying characteristics, which traditional PID cannot effectively cope with. Susceptibility to external interference: Environmental interference such as wind and waves can cause PID control accuracy to decrease, requiring additional compensation mechanisms.

[0004] Although the PSO algorithm can optimize PID parameters, in complex environments, the optimization directly based on the real unmanned boat model has high computational overhead and slow convergence speed.

[0005] Since traditional controllers have poor adaptability to complex sea conditions, PSO and traditional methods have large computational complexity and slow convergence speed, a comprehensive solution that integrates multi-scenario simulation, agent models and optimization algorithms is needed. Summary of the Invention

[0006] Purpose of the invention: To solve the problems of poor adaptability of traditional controllers in complex sea conditions, large computational complexity and slow convergence speed of both PSO and traditional methods, the present invention proposes an unmanned boat path tracking method based on multi-scenario randomized dynamics simulation and DNN agent model.

[0007] Technical solution: A path tracking method for unmanned aerial vehicles based on multi-scenario randomized dynamics simulation and DNN agent model, including the following steps:

[0008] Step 1: Set the value range and distribution of environmental parameters, and randomly generate multiple sets of sea state parameters based on the real ocean statistical model;

[0009] Step 2: Construct a dynamic model of the unmanned boat and perform dynamic simulation of the unmanned boat under each set of sea state parameters by changing the combination of controller parameters and filter coefficients, and record the error indicators;

[0010] Step 3: Take the controller parameters, filter coefficients and sea state parameters as input, the error index as output, and the input and corresponding output as a set of training samples;

[0011] Step 4: Construct a deep neural network model with sea state parameters and controller parameters as input and error indicators as output. Use the deep neural network model to learn the nonlinear mapping relationship between input and output. In this deep neural network model, a multi-layer perceptron structure is used for the input and output layers, and an attention mechanism is introduced in the embedding layer or fusion layer. The training samples are 10,000 groups and the number of iterations is 30. The deep neural network model is trained to obtain a proxy model with the prediction error (MSE) as the indicator.

[0012] Step 5: The current sea state parameters (wave direction, wave height, flow speed, load, wind direction and wind speed) and controller parameters (K p ,K i ,K d ) and the filter coefficient (α) are input into the proxy model input terminal, replacing the original simulation model calculation to directly obtain the error index prediction value (MSE);

[0013] Step 6: Based on the error index prediction value (MSE) obtained at the output end, the particle swarm optimization algorithm is used to automatically adjust the controller parameters or filter coefficients of the unmanned boat until the optimal parameters are found.

[0014] Furthermore, the environmental parameters include wave direction, wave height, flow velocity, load, wind direction and wind speed.

[0015] Furthermore, based on the MGG model, the unmanned boat dynamics model is constructed and expressed as:

[0016]

[0017] Where, are state variables, x, y, ψ are the north and east positions and heading angle of the UAV respectively; u, v, r are the longitudinal, transverse and heading angular velocities of the UAV respectively; d u d v d r They represent the longitudinal interference force, lateral interference force and bow interference moment generated by environmental interference respectively; τ u , τ r Represents longitudinal propulsion force and bow moment; m ii d ii are the components of the inertia and damping matrix of the unmanned boat on each axis.

[0018] Furthermore, suppose that the input x and output Respectively expressed as:

[0019] x=[θ w ,H W ,V current ,v wind ,θ wind ,L,K P ,K i ,K d ,α] T

[0020]

[0021] Where θ w is the wave direction, H w is the wave height, v current is the flow velocity, v wind is the wind speed, θ wind is the wind direction, L is the load of the unmanned boat, K p is the proportional gain, K i is the integral gain, K d is the differential gain, α is the filter coefficient, is the tracking error, For energy consumption;

[0022] The multi-layer perceptron structure is trained according to the following steps:

[0023] z [l] =W [l] a [l-1] +b [l]

[0024] a [l] =ReLU(z l )=max(0,z [l] )

[0025]

[0026] Where W [l] is the weight matrix of the lth layer, b [l] is the bias vector of the lth layer, z [l] is the linear input of the lth layer, a [l] is the activation output of layer l.

[0027] Among them, a [0] =x, representing the input of the deep neural network model;

[0028]

[0029] Where η is the learning rate, y is the true output, is the actual output, a [0]is the activation value of the 0th layer input, x is the input vector of the neural network, Θ is the set of all parameters, e track is the actual (trajectory) error or mileage deviation, E consunmed is the actual energy consumption, ζ is the loss function, and α1 and α2 are two positive weight coefficients used to balance the importance of the two errors.

[0030] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0031] (1) The present invention comprehensively considers environmental parameters such as wave direction, wave height, flow velocity, load, wind direction, and wind speed during data collection, thereby improving the accuracy of multi-scenario randomized path tracking from the source of the data.

[0032] (2) The present invention improves the intelligence and efficiency of PID parameter optimization and adopts the particle swarm optimization (PSO) algorithm to automatically adjust the PID parameters, thus avoiding the inefficiency and uncertainty of traditional PID parameter tuning methods (such as trial and error and empirical methods). PSO can globally search for the optimal parameter combination, improve the adaptive ability of PID control, and enable the unmanned vehicle to efficiently track the path in different waters and environmental conditions.

[0033] (3) The present invention reduces computational costs and improves optimization speed. The introduction of the DNN proxy model reduces dependence on the real physical model, making the optimization process more efficient. The proxy model can establish an approximate relationship between the path tracking error and the PID parameters through machine learning (such as Gaussian process regression, neural network, etc.), significantly reducing the number of real system calculations and accelerating the optimization process.

[0034] (4) Because the training data of the present invention covers a wide range of sea conditions, when encountering new combinations of wave direction, wave height, and current speed during actual navigation, the DNN proxy model can still maintain a high prediction accuracy and reduce the need for online retraining. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a training flow chart of the DNN agent model of the present invention. DETAILED DESCRIPTION

[0036] The technical solution of this embodiment will now be further described with reference to the accompanying drawings and embodiments.

[0037] This embodiment proposes a path tracking method for an unmanned vehicle based on multi-scenario randomized dynamics simulation and a DNN agent model, which mainly includes the following steps:

[0038] Step 1: Set the value ranges and distribution patterns of environmental parameters such as wave direction, wave height, current velocity, load, wind direction, and wind speed, and randomly generate multiple sets of sea state parameters based on a real ocean statistical model. Each set of sea state parameters can represent different conditions in the real ocean environment.

[0039] Step 2: Run the UAV dynamics simulation several times under each set of sea state parameters. By changing the combination of controller (such as PID) parameters and filter coefficients, record the final tracking error, energy consumption, or other performance indicators. The specific operations include:

[0040] Based on the improvement of the MGG model, the unmanned boat dynamics model is constructed as follows:

[0041]

[0042] Where x, y, ψ are the north and east positions and heading angle of the UAV respectively; u, v, r are the longitudinal, transverse and heading angular velocities of the UAV respectively; d u d v d r They represent the longitudinal interference force, lateral interference force and bow interference moment generated by environmental interference respectively; τ u , τ r Represents longitudinal propulsion force and bow moment; m ii d ii are the components of the inertia and damping matrix of the unmanned boat on each axis.

[0043] Step 3: Based on all simulation results, a large number of training samples are generated; the training samples are in the form of input (controller parameters, filter coefficients, sea state parameters) - output (error index).

[0044] Step 4: Divide the training samples into training set, validation set and test set, and use deep neural network to learn the nonlinear mapping relationship between input and output. For the multimodal input strategy, introduce feature embedding and channel processing mechanism, and introduce attention mechanism in the embedding layer or subsequent fusion layer to obtain a proxy model that can quickly predict error indicators. Use the training set to train the proxy model.

[0045] A surrogate model is a technique used to accelerate optimization calculations. It typically uses machine learning methods (such as Gaussian process regression, neural networks, and support vector machines) to construct a low-cost approximate model to replace the actual system for optimization calculations. In unmanned vehicle path tracking, surrogate models have the following uses: Reducing computational overhead: Using a surrogate model to replace some of the actual model calculations improves optimization speed. Smoothing the optimization space: Surrogate models can remove noise, making the optimization process more stable. Improving convergence: By learning the dynamic characteristics of the unmanned vehicle, the PSO algorithm can more quickly find the optimal PID parameters.

[0046] Specifically: A weighted combination of mean square error (MSE) and mean absolute error (MAE) is used as the loss function with a weight ratio of 3:7, which is more inclined to the overall MAE, reducing sensitivity to abnormal samples. It is suitable for complex sea conditions with some extreme noise points in the samples, can maintain a good fit for most common samples, and converges more stably. In most sea condition combinations and random samples of controller parameters, the weighted combination loss achieves a better balance between the final average error and the 90% quantile error of the model. The Adam optimization algorithm is used for network parameter optimization, training and validation losses are detected in real time, and early stopping is used to prevent overfitting.

[0047] The training steps of the proxy model are further explained:

[0048] Step 1: Clean the training samples, use the isolation forest algorithm to remove outliers or missing data, and perform normalization.

[0049] Step 2: Divide the samples into training, validation, and test sets. The validation set is used for model selection, hyperparameter tuning, and early stopping. The test set is used only after the network structure and training parameters are finalized to evaluate the actual generalization performance of the model and perform cross-validation.

[0050] Step 3: Build a deep neural network model based on a multi-layer perceptron (MLP), accepting both sea state parameters and controller parameters, taking into account the coupling between the environment and control, and covering a wide range of scenarios. Set a correlation coefficient matrix and perform pairwise correlation analysis (such as the Pearson correlation coefficient) on sea state parameters and controller parameters. If the correlation between certain features is too high (such as an absolute value higher than 0.9), they will be merged and removed to reduce interference with the network structure. Feature embedding is also performed, using PCA or embedding layer dimensionality reduction to remove redundant information. Set the model input and output:

[0051] x=[θ w ,H W ,V current ,v wind ,θ wind ,L,KP ,K i ,K d ,α] T

[0052]

[0053] Where θ w is the main direction of the wave (usually expressed as an angle), H w is the effective wave height, v current The magnitude of the ocean current speed can further include wind direction information, v wind The wind speed can also include wind direction information, θ wind The main direction of the wind, L is the load or load condition of the unmanned boat, K p is the proportional gain, K i is the integral gain, K d is the differential gain, α is the filter coefficient, is the tracking error, For energy consumption;

[0054] Step 4: Use a multi-layer perceptron (MLP) structure for the input and output layers and set the loss function for training:

[0055] z [l] =W [l] a [l-1] +b [l]

[0056] a [l] =ReLU(z l )=max(0,z [l] )

[0057]

[0058] Where: W [l] is the weight matrix of the lth layer, b [l] is the bias vector of the lth layer, z [l] is the linear input of the lth layer, a [l] is the activation output of layer l.

[0059] Where a [0] =x is the original input of the network (sea state, load, PID, filter parameters).

[0060]

[0061] Where η is the learning rate, y is the true output, is the actual output, a [0] is the activation value of the 0th layer input, x is the input vector of the neural network, Θ is the set of all parameters, e trackis the actual (trajectory) error or mileage deviation, E consunmed is the actual energy consumption, ζ is the loss function, and α1 and α2 are two positive weight coefficients used to balance the importance of the two errors.

[0062] Step 5: Particle swarm optimization (PSO) is used on the proxy model to quickly screen controller parameters, sea state parameters or filter coefficients, thereby improving the real-time performance and control accuracy of the unmanned boat in complex sea conditions.

[0063] The traditional full simulation combined with PSO takes several hours or even longer for each global optimization. However, the DNN combined with PSO in this invention can reduce the average time to minutes or tens of seconds, significantly shortening the iteration cycle and quickly screening parameters. Specifically:

[0064] v i (t+1)=w·v i (t)+c1·r1·(p i -x i (t))+c2r2·(gx i (t))

[0065] x i (t+1)=x i (t)+v i (t+1)

[0066] Among them, v i (t) is the velocity of particle i at time t, x i (t) is the position of particle i at time t, p i is the historical optimal position of particle i, g is the global optimal position, w is the inertia weight, which controls the inertia of the particle speed, c1 and c2 are acceleration constants, which control the influence of the individual and the group respectively, and r1 and r2 are random numbers between [0,1].

Claims

1. A path tracking method for unmanned aerial vehicles based on multi-scenario randomized dynamics simulation and DNN agent model, characterized by: The following steps are involved: Step 1: Set the value range and distribution of environmental parameters, and randomly generate multiple sets of sea state parameters based on the real ocean statistical model; Step 2: Construct a dynamic model of the unmanned boat and perform dynamic simulation of the unmanned boat under each set of sea state parameters by changing the combination of controller parameters and filter coefficients, and record the error indicators; Step 3: Take the controller parameters, filter coefficients and sea state parameters as input, the error index as output, and the input and corresponding output as a set of training samples; Step 4: Construct a deep neural network model with sea state parameters and controller parameters as input and error indicators as output. Use the deep neural network model to learn the nonlinear mapping relationship between input and output. In this deep neural network model, a multi-layer perceptron structure is used for the input and output layers. An attention mechanism is introduced in the embedding layer or fusion layer. The deep neural network model is trained to obtain a proxy model with the prediction error as the indicator. Step 5: Input the current sea state parameters, controller parameters and filter coefficients of the unmanned boat into the proxy model to obtain the error index prediction value; Step 6: Based on the obtained error index prediction value, the particle swarm optimization algorithm is used to automatically adjust the controller parameters or filter coefficients of the unmanned boat until the optimal parameters are found; Step 7: Control the unmanned boat according to the optimal parameters.

2. The unmanned vehicle path tracking method for multi-scenario randomized dynamics simulation and DNN agent model according to claim 1 is characterized by: The environmental parameters include wave direction, wave height, flow velocity, load, wind direction and wind speed.

3. The unmanned vehicle path tracking method for multi-scenario randomized dynamics simulation and DNN agent model according to claim 2 is characterized by: Based on the MGG model, the unmanned boat dynamics model is constructed and expressed as: Where, are state variables, x, y, ψ are the north and east positions and heading angle of the UAV respectively; u, v, r are the longitudinal, transverse and heading angular velocities of the UAV respectively; d u d v d r They represent the longitudinal interference force, lateral interference force and bow interference moment generated by environmental interference respectively; τ u , τ r Represents longitudinal propulsion force and bow moment; m ii d ii are the components of the inertia and damping matrix of the unmanned boat on each axis.

4. The unmanned vehicle path tracking method for multi-scenario randomized dynamics simulation and DNN agent model according to claim 3 is characterized by: Assume that the input x and output Respectively expressed as: x=[θ w ,H W ,V current ,v wind ,i wind ,L,K P ,K i ,K d ,a] T Where θ w is the wave direction, H w is the wave height, v current is the flow velocity, v wind is the wind speed, θ wind is the wind direction, L is the load of the unmanned boat, K p is the proportional gain, K i is the integral gain, K d is the differential gain, α is the filter coefficient, is the tracking error, For energy consumption; The multi-layer perceptron structure is trained according to the following steps: z [l] =W [l] a [l-1] +b [l] and [l] =ReLU(z l )=max(0,z [l] ) Where W [l] is the weight matrix of the lth layer, b [l] is the bias vector of the lth layer, z [l] is the linear input of the lth layer, a [l] is the activation output of layer l; Among them, a [0] =x, representing the input of the deep neural network model; Where η is the learning rate, y is the true output, is the actual output, a [0] is the activation value of the 0th layer input, x is the input vector of the neural network, Θ is the set of all parameters, e track is the actual error or mileage deviation, E consunmed is the actual energy consumption, ζ is the loss function, and α1 and α2 are two positive weight coefficients used to balance the importance of the two errors.

Images