Ship motion attitude prediction method and device, medium and electronic equipment

Abstract

The invention provides a ship motion attitude prediction method and device, a medium and electronic equipment, and relates to the technical field of artificial intelligence, and the method comprises the steps: collecting a plurality of motion feature parameters of a key position of a ship once every other time step; determining motion data of the ship based on the plurality of motion characteristic parameters; at any detection moment, acquiring latest collected motion data of continuous N time steps, and determining the motion data as a motion data sequence; performing normalization processing on the motion data sequence; processing each piece of motion data in the motion data sequence after normalization processing by a circulation unit according to a sliding window mode; outputting M pieces of target attitude prediction data in sequence to obtain a prediction data sequence; and outputting a prediction data sequence as an attitude prediction result at the detection moment. The motion attitude of the ship can be accurately and stably predicted.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, specifically to a method, device, medium, and electronic equipment for predicting ship motion attitude. Background Technology

[0002] During navigation, ships are affected by marine environmental factors, resulting in complex, multi-dimensional movements. Predicting a ship's motion in advance is a key technology for ensuring the safety and efficiency of maritime operations.

[0003] Currently, the main technology for predicting ship motion attitude is recurrent neural networks (RNNs). The RNN network structure includes a multilayer perceptron (MLP) module, which approximates the target function by combining a linear weight matrix with a fixed nonlinear activation function (such as the sigmoid or tanh function) applied to the neurons.

[0004] However, the fixed nonlinear expression method of using "linear weight matrix" and "fixed nonlinear activation function" on neuron nodes limits the model's ability to fit complex time series data and make accurate predictions. When faced with sudden and strongly nonlinear changes common in ship motion, it is difficult to fully and accurately capture its dynamic characteristics, resulting in insufficient accuracy and stability in the prediction of ship motion attitude. Summary of the Invention

[0005] The main objective of this disclosure is to provide a method, apparatus, medium, and electronic device for predicting ship motion attitude, in order to solve the problems of insufficient accuracy and stability in the prediction of ship motion attitude in related technologies.

[0006] To achieve the above objectives, the first aspect of this disclosure provides a method for predicting ship motion attitude, applied to an electronic device. The electronic device deploys a pre-trained Kolmogorov-Arnold recurrent network model, wherein the Kolmogorov-Arnold recurrent network model includes recurrent units, each including a predictor, a linear transformation layer, and a Kolmogorov-Arnold network (KAN) module; the method includes: Multiple motion feature parameters of key positions of the ship are collected at each time step; The ship's motion data is determined based on multiple motion characteristic parameters; At any given detection time, acquire the latest consecutive N time steps of motion data that have been collected and determine them as a motion data sequence; Normalize the motion data sequence; The motion data in the motion data sequence is processed sequentially through a loop unit using a sliding window approach, where each sliding window processes T motion data points, as follows: For any target motion data corresponding to any target time step in each of the T motion data, the target motion data is input into the KAN module to obtain key feature data; Obtain the hidden state feature data that has been output after the motion data of the previous time step of the target time step has been processed by the loop unit; The hidden state feature data from the previous time step is transformed by a linear transformation layer to obtain dynamic feature data. By fusing key feature data and dynamic feature data, the hidden state feature data of the target time step is obtained; The hidden state feature data is input into the predictor to obtain pose prediction data; The last output pose prediction data from the T motion data is used as the target pose prediction data. Output M target attitude prediction data points sequentially to obtain a prediction data sequence, where M = N / T; The output prediction data sequence is the attitude prediction result at the detection time.

[0007] A second aspect of this disclosure provides a ship motion attitude prediction device applied to an electronic device, wherein a trained Kolmogorov-Arnold recurrent network model is deployed in the electronic device, wherein the Kolmogorov-Arnold recurrent network model includes recurrent units, and the recurrent units include predictors, linear transformation layers, and Kolmogorov-Arnold network (KAN) modules; the device includes: The acquisition module is used to acquire multiple motion feature parameters of key positions of the ship at each time step; The preprocessing module is used to determine the ship's motion data based on multiple motion feature parameters; at any detection time, it acquires the latest consecutive N time steps of collected motion data and determines them as a motion data sequence; and it performs normalization processing on the motion data sequence. The loop processing module processes each motion data point in the motion data sequence sequentially through a sliding window, with each sliding window processing T motion data points, as follows: For the target motion data corresponding to any target time step in each of the T motion data points, the target motion data is input into the KAN module to obtain key feature data; the hidden state feature data output after the motion data of the previous time step of the target time step has been processed by the loop unit is obtained; the hidden state feature data of the previous time step is transformed through a linear transformation layer to obtain dynamic feature data; the key feature data and dynamic feature data are fused to obtain the hidden state feature data of the target time step; the hidden state feature data is input into the predictor to obtain the pose prediction data; and the pose prediction data output at the end of the T motion data points is used as the target pose prediction data. The prediction result output module is used to sequentially output M target attitude prediction data to obtain a prediction data sequence, where M=N / T; the output prediction data sequence is the attitude prediction result at the detection time.

[0008] A third aspect of this disclosure provides a computer-readable storage medium storing computer instructions for causing a computer to perform the method provided in the first aspect.

[0009] A fourth aspect of this disclosure provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the method provided in the first aspect.

[0010] This disclosure employs a pre-trained Kolmogorov-Arnold recurrent network (RNN) model, which includes recurrent units (RNNs) containing a predictor, a linearly varying layer, and a KAN module. The KAN module replaces the original Multilayer Perceptron (MLP) module in the RNN. The output of the KAN module serves as one of the inputs to the predictor of the entire KAN model, essentially acting as an "edge" rather than a "node." By leveraging the learnable nature of the KAN module—distinct from the "fixed nonlinear activation function" used in related techniques—the recurrent network model incorporating the KAN module can accurately and stably predict the ship's motion attitude. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the specific embodiments or related technologies of this disclosure, the accompanying drawings used in the description of the specific embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic diagram of the structure of the Kolmogorov-Arnold recurrent network model is provided for embodiments of this disclosure.

[0013] Figure 2 A flowchart illustrating the ship motion attitude prediction method provided in this embodiment of the disclosure. Figure 1 .

[0014] Figure 3 This is a schematic diagram illustrating the computational principle of the Kolmogorov-Arnold cyclic network model provided in an embodiment of this disclosure.

[0015] Figure 4 A flowchart illustrating the ship motion attitude prediction method provided in this embodiment of the disclosure. Figure 2 .

[0016] Figure 5 A schematic diagram of the network structure of the Kolmogorov-Arnold network module provided in an embodiment of this disclosure.

[0017] Figure 6 A block diagram of a ship motion attitude prediction device provided in an embodiment of this disclosure.

[0018] Figure 7 A block diagram of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] During navigation, ships are affected by complex marine environmental factors such as wind, waves, and currents, resulting in complex movements with multiple degrees of freedom, including heave, pitch, and roll. Predicting the ship's motion attitude is a crucial technology for ensuring safe maritime operations.

[0023] Current research indicates that traditional methods for predicting ship motion and attitude primarily employ autoregressive models, Kalman filters, and other mathematical-physical models or time-series models of ship motion. However, these models typically require repeated adjustments based on specific ship types and sea conditions, and their generalization ability is insufficient.

[0024] With the development of artificial intelligence technology, recurrent neural networks (RNNs) have gradually become a research hotspot for ship motion attitude prediction due to their advantages in processing time series data. However, these RNN models mainly follow the basic principles of multilayer perceptrons (MLPs), which rely on a combination of linear weight matrices and fixed nonlinear activation functions (such as sigmoid or tanh) applied to the neuron nodes to approximate complex objective functions. This fixed nonlinear expression limits the model's ability to fit and accurately predict complex time series data, making it difficult to fully and accurately capture the dynamic characteristics of ship motion when faced with sudden and strongly nonlinear changes common in ship motion.

[0025] To address the aforementioned problems, this disclosure provides a method for predicting ship motion attitude. This method is applied to an electronic device that deploys a trained Kolmogorov-Arnold recurrent network model, referencing... Figure 1 , Figure 1This disclosure provides a schematic diagram of the structure of a Kolmogorov-Arnold recurrent network model, which includes a recurrent unit 100. The recurrent unit 100 includes a predictor 101, a linear transformation layer 102, and a Kolmogorov-Arnold Network (KAN) module 103.

[0026] The initial Kolmogorov-Arnold recurrent network model is obtained by replacing the Multilayer Perceptron (MLP) module in the recurrent network (RNN) with a Kolmogorov-Arnold Network (KAN) module. During training, the Kolmogorov-Arnold network modules are updated, ultimately resulting in the fully trained Kolmogorov-Arnold recurrent network model.

[0027] refer to Figure 2 , Figure 2 A flowchart illustrating the ship motion attitude prediction method provided in this embodiment of the disclosure. Figure 1 The method includes the following steps S21 to S26, as follows: Step S21: Collect multiple motion feature parameters of the ship's key positions at every time step.

[0028] In this embodiment, the interval between time steps can be set as needed. Optionally, the interval between time steps is 100 milliseconds. That is, the data acquisition frequency of the motion feature parameters is 10 Hz.

[0029] In this embodiment, the key location can be the ship's center of gravity, or other key locations outside the center of gravity (e.g., the ship's deck). An attitude sensor is installed at this key location to measure data at that location. Optionally, the attitude sensor can be an inertial measurement unit (IMU) or other attitude sensors (such as linear displacement sensors, angular displacement sensors, etc.). Optionally, the attitude sensor can also be located on the ship's deck.

[0030] In this embodiment, measurement data is collected at time step intervals to form a time series of measurement data. The measurement data may include various parameters reflecting the ship's heave and sway, as well as various parameters reflecting the ship's tilt.

[0031] Among these, key parameters are selected from the measurement data as multiple motion characteristic parameters of the ship. Optionally, the multiple motion characteristic parameters can be the three degrees of freedom that are most critical to the safety of ship transmission, including: heave (heave displacement) reflecting the vertical movement of the deck, and pitch (longitudinal roll amplitude) and roll (lateral roll amplitude) reflecting the longitudinal and lateral tilt of the deck.

[0032] Step S22: Determine the ship's motion data based on multiple motion characteristic parameters.

[0033] In one embodiment of this disclosure, multiple motion characteristic parameters include: the ship's heave displacement, the ship's longitudinal roll amplitude, and the ship's lateral roll amplitude. Accordingly, based on multiple motion characteristic parameters, the ship's motion data is determined, including: The instantaneous vertical displacement of a ship is calculated based on its heave displacement, longitudinal roll amplitude, and lateral roll amplitude. These parameters, along with the instantaneous vertical displacement, are then used to determine the ship's motion data.

[0034] The instantaneous vertical displacement of the ship is calculated based on its heave, longitudinal roll, and lateral roll. The corresponding formula is as follows:

[0035] In the formula, This refers to the instantaneous vertical displacement of the ship; This refers to heave displacement; The vertical sway amplitude is ; the horizontal sway amplitude is . The longitudinal distance between the key location and the center of the ship; This refers to the lateral distance between the critical location and the center of the ship.

[0036] Step S23: At any detection time, acquire the latest consecutive N time steps of motion data that have been collected and determine them as a motion data sequence, where N is a positive integer.

[0037] In this embodiment, the latest consecutive N time steps of motion data collected refer to the N time steps of motion data collected before this detection time, and can be denoted as: .

[0038] Step S24: Normalize the motion data sequence.

[0039] In this embodiment, the normalization process is to convert data of different dimensions in the motion data sequence to a uniform scale, typically [0,1] or [-1,1].

[0040] Step S25: Process each motion data point in the normalized motion data sequence sequentially through a loop unit using a sliding window approach. Each sliding window processes T motion data points. The specific process is as follows: Sa: For any target motion data corresponding to any target time step in each of the T motion data, input the target motion data into the KAN module to obtain key feature data.

[0041] In this example, the Kolmogorov-Arnold network module (KAN module) is a component of the already trained Kolmogorov-Arnold recurrent network model. During the training of the Kolmogorov-Arnold recurrent network model, the network parameters of the KAN module have also been optimized.

[0042] In this embodiment, the KAN module is used to extract key technical features that can represent the ship's motion from the target motion data.

[0043] In this embodiment, key feature data at any target time step The corresponding target motion data is denoted as Accordingly, the key feature data are denoted as... .

[0044] Sb: Obtain the hidden state feature data that has been output after the motion data of the previous time step of the target time step has been processed by the loop unit.

[0045] In this embodiment, the Kolmogorov-Arnold recurrent network model is obtained by replacing the multilayer perceptron (MLP) module in the recurrent network (RNN) with the Kolmogorov-Arnold Network (KAN) module. The Kolmogorov-Arnold recurrent network model is an improved recurrent network (RNN).

[0046] The motion data of the previous time step of the target time step is denoted as... After processing by the improved recurrent neural network (RNN) recurrent unit of this embodiment, the hidden state feature data is output, denoted as... .

[0047] Sc: The hidden state feature data of the previous time step is transformed by a linear transformation layer to obtain the dynamic feature data of the previous time step.

[0048] In this embodiment, the linear change layer is used to capture the dynamic dependency between the target time step and the previous time step. The calculation formula for the linear change layer in this embodiment is as follows:

[0049] In the formula, For dynamic feature data; A linearly varying layer; This refers to the hidden state feature data from the previous time step.

[0050] Optionally, the linear transformation layer includes a learnable weight matrix and a bias term. The weight matrix controls the transmission strength of historical information, and the bias term adjusts the baseline value of the output (dynamic feature data). Accordingly, the above formula can be expanded as follows:

[0051] In the formula, For dynamic feature data; The weight matrix is ​​a learnable matrix; The hidden state feature data of the previous time step; This is a bias term.

[0052] Sd: The hidden state feature data of the target time step is obtained by fusing key feature data and dynamic feature data.

[0053] In this example, key feature data and dynamic feature data are fused to obtain the hidden state feature data corresponding to the target time step. The corresponding calculation formula is as follows:

[0054] In the formula, The hidden state feature data corresponding to the target time step; A linearly varying layer; The hidden state feature data of the previous time step; Record the key feature data.

[0055] Se: Input the hidden state feature data corresponding to the target time step into the predictor to obtain the pose prediction data corresponding to the target time step.

[0056] In this embodiment, the predictor is an output layer module. Optionally, the predictor can be a long short-term memory network and a gated recurrent unit.

[0057] In this embodiment, the mathematical formula for the predictor is expressed as follows:

[0058] In the formula, For the target time step t Corresponding attitude prediction data; For predictors; For the target time step t The corresponding hidden state feature data.

[0059] Sf: Use the last corresponding output attitude prediction data from the T motion data as the target attitude prediction data.

[0060] In this example, for each of the T motion data points, steps Sa to Sf are executed in order of priority from time step to time step. The last corresponding pose prediction data among the T motion data points is then used as the output of a sliding window.

[0061] Step S26: Output M target attitude prediction data in sequence to obtain the prediction data sequence, where M=N / T; the output prediction data sequence is the attitude prediction result at the detection time.

[0062] In this embodiment, after processing each motion data sequentially through the loop unit in step S25 using a sliding window method, M target attitude prediction data are obtained, which serve as the prediction data sequence. Here, M is the number of sliding windows, which is also the number of attitude prediction data.

[0063] In this embodiment, the predicted data sequence can be denoted as: .

[0064] It should be noted that this disclosure also provides a specific algorithmic example for predicting the motion attitude of a ship using a fully trained Kolmogorov-Arnold recurrent network model, detailed below: I. Data Definition: Definition For the trained Kolmogorov-Arnold recurrent network model; Latest collected motion data; N For collection N Motion data at each time step; T is the length of the sliding window, i.e., the data processed in each sliding window. T One motion data point; M The window length for the output predicted data sequence; To N The motion data sequence after normalizing the motion data at each time step; The output is the predicted data sequence; This refers to the attitude prediction data for a single time step.

[0065] II. Based on the above definition, the specific algorithm for predicting the ship's motion attitude is as follows: 1. Input: Input a pre-trained Kolmogorov-Arnold recurrent network model. The latest collected sports data (The length of the motion data is) N That is, including N (motion data at each time step), and the length of the sliding window, T.

[0066] 2. For the initial Normalization is performed: .

[0067] 3. Initialize an empty list to store the output predicted data sequence (window length is...). M ): ← [].

[0068] 4. Loop M Next, gradually generate M Individual target pose prediction data: for = 1 to M do ← .forward( ) .append( ) [1:] +

[0069] end for.

[0070] 5. Return ( M Target attitude prediction data.

[0071] As described in the above-disclosed embodiments, this disclosure employs a trained Kolmogorov-Arnold recurrent network (RNN) model, which includes recurrent units. Each recurrent unit includes a predictor, a linear transformation layer, and a KAN module. The KAN model is obtained by replacing the original multilayer perceptron (MLP) module in the recurrent network (RNN) with a KAN module. The output of the KAN module serves as one of the inputs to the predictor of the entire KAN model, essentially making the KAN module an "edge" rather than a "node." By leveraging the learnable nature of the KAN module, unlike the "fixed nonlinear activation function" used in related technologies, the recurrent network model incorporating the KAN module can accurately and stably predict the ship's motion attitude.

[0072] refer to Figure 3 , Figure 3This is a schematic diagram illustrating the computational principle of the Kolmogorov-Arnold recurrent network model provided in an embodiment of this disclosure. In one embodiment of this disclosure, the Kolmogorov-Arnold network (KAN) module includes a time basis function processing module 1031 and a time spline function processing module 1032. The specific processing procedure of the Kolmogorov-Arnold network module is detailed below: In step Sa, the target motion data is input into the KAN module to obtain key feature data, specifically including: The target motion data is input into the time basis function processing module to output the first key feature data; the target motion data is input into the time spline function processing module to output the second key feature data.

[0073] In one embodiment of this disclosure, the time basis function processing module employs... SiLU As a time basis function, its formula is expressed as follows:

[0074] In the formula, The target motion data corresponding to time t at the target time step. These are learnable weight parameters; This is the first key feature data.

[0075] in, .

[0076] In the formula, SiLU As a time basis function; It is a natural constant and an infinite non-repeating irrational number.

[0077] In one embodiment of this disclosure, the time spline function processing module employs a set of B Spline basis functions are expressed in the following formula:

[0078] In the formula, This is the second key feature data. For learnable time weights; These are the spline coefficients; As a group B spline basis functions; This represents the number of grid intervals for the time spline function. Let be the order of the time spline function; For summation index.

[0079] Accordingly, in step Sd, key feature data and dynamic feature data are fused to obtain the hidden state feature data for the target time step, specifically including: Based on the first key feature data, the second key feature data, and the dynamic feature data, the hidden state feature data corresponding to the target time step is determined.

[0080] In this embodiment, the hidden state feature data corresponding to the target time step is determined based on the first key feature data, the second key feature data, and the dynamic feature data. The corresponding calculation formula is as follows:

[0081] In the formula, The hidden state feature data corresponding to the target time step; For dynamic feature data; This is the first key feature data; This is the second key feature data; The weight matrix is ​​a learnable matrix; The hidden state feature data of the previous time step; This is the bias term, used to adjust the baseline value of the output.

[0082] In this embodiment, the hidden state feature data corresponding to the target time step is input into the predictor to obtain the pose prediction data corresponding to the target time step. The corresponding calculation formula is as follows:

[0083] In the formula, For attitude prediction data; The weight matrix is ​​used to map the hidden state to the output space; this weight matrix can be learned. The hidden state feature data corresponding to the target time step; This is the bias vector.

[0084] It should be noted that, using the above calculation formulas, the recurrent unit of the Kolmogorov-Arnold recurrent network model receives the target motion data and the hidden state feature data of the previous time step at time t, and calculates the new hidden state feature data. This calculation is completed by two parallel and complementary processing modules: the time basis function processing module and the time spline function processing module.

[0085] The time basis function in the time basis function processing module provides a basic and stable nonlinear transformation path, which acts similarly to a residual connection, helping to stabilize the learning process and improve gradient flow.

[0086] Among them, a set of time spline function processing modules BSpline basis functions (spline functions) are used to replace the single-layer fully connected network in the original recurrent network (RNN) on the input layer. They can realize the model learning of a "edge" of the entire Kolmogorov-Arnold recurrent network model, rather than the model learning of the "nodes" in the Kolmogorov-Arnold recurrent network model.

[0087] in, B The spline coefficients of the spline basis function are a set of learnable coefficients used to learn the shape of the spline function; essentially, they learn the nonlinear relationship between the input variables and the system state. Furthermore, the time-spline function processing module includes learnable time weights. This is used to adjust the overall influence intensity of the entire spline function at the current time step.

[0088] As described above, by extracting the first key feature data through the time basis function processing module and the second key feature data through the time spline function processing module, and based on the first key feature data, the second key feature data, and the dynamic feature data passed from the previous time step, the hidden state feature data corresponding to the target time step is obtained. This not only extracts multi-level and rich key features from the target motion data, but also captures the dynamic dependencies passed from the previous time step, so that the corresponding hidden state can accurately reflect the state of the current target time step, thereby making the final predicted attitude prediction data more accurate.

[0089] Meanwhile, the stability and controllability of the overall Kolmogorov-Arnold recurrent network model are enhanced by the learnability of parameters such as spline coefficients, learnable weight parameters, and learnable time weights.

[0090] refer to Figure 4 , Figure 4 A flowchart illustrating the ship motion attitude prediction method provided in this embodiment of the disclosure. Figure 2 Based on the above embodiments, this embodiment will focus on describing the training process of the Kolmogorov-Arnold recurrent network model, as detailed below: Step S31: Within a preset time period, collect measurement data from the attitude sensors at key positions of the ship at time step intervals.

[0091] In this embodiment, this step refers to the relevant description in step 21 of the above embodiment, and will not be repeated here.

[0092] In this embodiment, the preset duration can be set according to requirements, such as 90 days or 120 days.

[0093] Step S32: For each time step, extract multiple motion feature parameters from the measurement data.

[0094] In one embodiment of this disclosure, the motion of the aircraft on the ship's deck has a critical impact on the aircraft. Therefore, optionally, the motion characteristic parameters of this disclosure can be the heave displacement of the ship's deck, the longitudinal roll amplitude of the ship, and the lateral roll amplitude of the ship.

[0095] Step S33: Calculate the instantaneous vertical displacement of the corresponding ship based on multiple motion characteristic parameters.

[0096] In this embodiment, the instantaneous vertical displacement of the ship is calculated based on multiple motion characteristic parameters, and the corresponding calculation formula is as follows:

[0097] In the formula, This refers to the instantaneous vertical displacement of the ship; This refers to heave displacement; The vertical sway amplitude is ; the horizontal sway amplitude is . The longitudinal distance between the key location and the center of the ship; This refers to the lateral distance between the critical location and the center of the ship.

[0098] Step S34: Construct a ship motion attitude training dataset based on multiple motion feature parameters and their corresponding instantaneous vertical displacements.

[0099] In this embodiment, the motion data sequence composed of multiple motion feature parameters and instantaneous vertical displacement corresponding to each time step within a preset time period is used as the ship motion attitude training dataset.

[0100] Step S35: Based on the ship motion attitude training dataset, train the initial Kolmogorov-Arnold recurrent network model until the training error meets the preset error requirements, and obtain the trained Kolmogorov-Arnold recurrent network model.

[0101] In this embodiment, the loss function used when training the initial Kolmogorov-Arnold recurrent network model is as follows:

[0102] In the formula, This is the loss value; This is a predicted value; The actual value; s is the sample size. .

[0103] In this example, the Kolmogorov-Arnold network module (KAN module) is a component of the already trained Kolmogorov-Arnold recurrent network model. During the training of the Kolmogorov-Arnold recurrent network model, the network parameters of the KAN module have also been optimized. In one embodiment of this disclosure, the training (parameter update and optimization) process of the Kolmogorov-Arnold network module (KAN module) is described in detail below: The Kolmogorov-Arnold network module disclosed herein is theoretically derived from the Kolmogorov-Arnold representation theorem, which states as follows: The Kolmogorov-Arnold representation theorem states that any multi-dimensional continuous function on a bounded field can be represented by a composition of a finite number of single-variable functions and addition operations, i.e., for a function defined on a unit hypercube... Multivariate continuous functions on The theorem gives the following mathematical expression:

[0104] In the formula, It is a multivariate continuous function; n The number of input variables; For the first q An outer single-variable function; The index for summing the outer layer; For the first q Group, No. p An inner univariate function, p The index for the inner summation; For the first p One input variable.

[0105] It is understandable that in this expression, the function The summation is decomposed into two nested layers of functions: the inner layer consists of a set of univariate functions, each performing an independent nonlinear transformation on each input variable; the outer layer consists of another set of univariate functions, which further transform the summation of the inner layer functions. The Kolmogorov-Arnold representation theorem proves that arbitrarily complex multivariate continuous functions can be reconstructed (fitted) through a finite number of such combinations of inner and outer functions. This theorem provides the mathematical basis for the network structure of the Kolmogorov-Arnold network module disclosed in this paper.

[0106] refer to Figure 5 , Figure 5 This is a schematic diagram of the network structure of the Kolmogorov-Arnold network module provided in an embodiment of this disclosure. Figure 5 The image shows an example of a hypothetical KAN network structure, where... Figure 5The two leftmost black nodes represent the original two-dimensional input data. Each subsequent black node represents a summation operation on its inputs (defined as "edges"). Each curve represents the processing of a single-variable function of the data for one edge. This structure can be extended to arbitrary depths along the computational direction. Therefore, a structure with... L The general network structure of the KAN module with layers can be concisely represented as a composite function of these layers:

[0107] In the formula, This is represented as the output of the entire Kolmogorov-Arnold network module; This is represented as the input vector of the Kolmogorov-Arnold network module; Indicates the first The learnable function corresponding to each KAN layer takes the value of 0 to 1 .

[0108] During the calculation process of the KAN module, L The input vector of the Kolmogorov-Arnold network module. Set as ,and belong 3D real space (i.e., vector) ).

[0109] for L The first Kolmogorov-Arnold network module The first layer Each neuron has an output value. By the Output of all neurons in the layer After the corresponding learnable function (i.e., a univariate function). The transformed and summed result is expressed as follows:

[0110] In the formula, Represented as the first The first layer The output value of each neuron; Represented as the first The first layer One neuron, Represented as the first The number of neurons in the layer; For the first The first layer The first neuron to the second The first layer A univariate function of neurons; where ,when When it is not 0, it is represented as the first... The first layer The output value of each neuron, when for At that time, it directly represents the first input vector of the Kolmogorov-Arnold network module. The value of the dimension.

[0111] Each univariate function Represented as a linear combination of time spline functions, where the time spline functions include a set of... B spline basis functions That is, a single variable function Represented as a group B spline basis functions A linear combination of is represented as follows:

[0112] In the formula, For the first The first layer The first neuron to the second The first layer A univariate function of one neuron; For summation index; This represents the number of grid intervals for the time spline function. Let be the order of the time spline function; These are the learnable spline coefficients; for B Spline basis functions.

[0113] Based on the above two formulas, the calculation formula for the forward propagation of training Kolmogorov-Arnold network modules can be obtained:

[0114] It should be noted that the formula above represents the spline coefficients corresponding to one neuron; For the first The first layer The first neuron to the second The first layer Spline coefficients of each neuron.

[0115] In this embodiment, during the initial training of the Kolmogorov-Arnold recurrent network model, the Kolmogorov-Arnold network modules are also updated, ultimately resulting in a fully trained Kolmogorov-Arnold recurrent network model. The update process of the Kolmogorov-Arnold network modules involves optimizing the set of all spline coefficients. To minimize the difference between the predicted output and the true target value of the entire Kolmogorov-Arnold recurrent network model, the following is a detailed explanation: Calculate the loss function For each spline coefficient The partial derivative (i.e., gradient) .

[0116] According to the chain rule, the gradient is decomposed and backpropagated as follows:

[0117] The partial derivatives of each part can be obtained as follows: ; .

[0118] Therefore, the key to backpropagation calculation lies in the backpropagation error term. If gradient Once calculated, the spline coefficients can be iteratively updated using the following update formula:

[0119] in, It is the learning rate. This is achieved by repeatedly performing forward propagation and loss calculation [i.e., inputting the loss function and calculating the loss function]. The steps of [intermediate propagation], backpropagation, and iterative update are used to obtain results that meet the error requirements. Therefore, it can be seen that the optimal spline function shape can be learned step by step through the above steps, thereby accurately fitting the target function.

[0120] It should be noted that this disclosure also provides specific algorithmic examples for training the Kolmogorov-Arnold recurrent network model, detailed below: I. Data Definition: Defined as a training dataset for ship motion attitude; T The length of the sliding window, i.e., the amount of data processed in each sliding window. T One motion data point; M The window length for the output predicted data sequence; For the trained Kolmogorov-Arnold recurrent network model; The input sequence after sliding window processing (from) ); The target sequence after sliding window processing (from) ); This is the initial Kolmogorov-Arnold recurrent network model; For all learnable parameters in the Kolmogorov-Arnold recurrent network model; These are the input and target data for a batch during the training process, respectively. This is the model's predicted output for the input batch; This represents the mean squared error loss value.

[0121] II. Based on the above definition, the specific algorithm for training the Kolmogorov-Arnold recurrent network model is as follows: 1. Input: Ship motion attitude training dataset Length of the sliding window T and the window length of the output predicted data sequence. M .

[0122] 2. Construct supervised learning samples using a sliding window: X, Y ←SlidingWindows( , T, M) .

[0123] 3. Normalize the input: X ← Normalize(X) .

[0124] 4. Initialize the initial Kolmogorov-Arnold recurrent network model and all learnable parameters .

[0125] 5. Perform the following steps for each training cycle: for Each training cycle do for ( ) in (X, Y) Each batch ( Forward propagation to obtain predicted values (Calculate the loss based on the mean square error) Clear the gradient calculated in the previous round loss.backward() (Perform backpropagation and calculate the gradient) Update parameters using the optimizer based on gradients. Θ (Using the Adam optimizer) end for end for 6. Training ends: ← .

[0126] 7. Return (A trained Kolmogorov-Arnold recurrent network model).

[0127] As can be seen from the description of the above-disclosed embodiments, the instantaneous vertical displacement of the corresponding ship is calculated through multiple motion feature parameters. The instantaneous vertical displacement is one of the core variables of ship motion. By combining multiple motion feature parameters and the corresponding instantaneous vertical displacement, a ship motion attitude training dataset is constructed. The resulting ship motion attitude training dataset can be used to train an accurate Kolmogorov-Arnold recurrent network model.

[0128] Meanwhile, by using the parameter update and optimization method of the Kolmogorov-Arnold network module provided in this embodiment, a good spline function shape can be obtained, which in turn enables accurate fitting of the target function.

[0129] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0130] This disclosure also provides a ship motion attitude prediction device for implementing the above method embodiments, applied to an electronic device. The electronic device deploys a trained Kolmogorov-Arnold recurrent network model, wherein the Kolmogorov-Arnold recurrent network model includes recurrent units, each including a predictor, a linear transformation layer, and a Kolmogorov-Arnold network module; as shown... Figure 6 As shown, the ship motion attitude prediction device 60 includes: a data acquisition module 61, a preprocessing module 62, a loop processing module 63, and a prediction result output module 64.

[0131] Among them, the acquisition module 61 is used to acquire multiple motion feature parameters of the key positions of the ship at each time step.

[0132] The preprocessing module 62 is used to determine the ship's motion data based on multiple motion feature parameters; at any detection time, it acquires the latest consecutive N time steps of collected motion data and determines them as a motion data sequence; and performs normalization processing on the motion data sequence.

[0133] The loop processing module 63 is used to process each motion data in the motion data sequence sequentially through the loop unit in a sliding window manner. Each sliding window processes T motion data, as follows: For the target motion data corresponding to any target time step in each of the T motion data, the target motion data is input into the KAN module to obtain key feature data; the hidden state feature data output after the motion data of the previous time step of the target time step has been processed by the loop unit is obtained; the hidden state feature data of the previous time step is transformed through a linear transformation layer to obtain dynamic feature data; the key feature data and dynamic feature data are fused to obtain the hidden state feature data of the target time step; the hidden state feature data is input into the predictor to obtain attitude prediction data; and the attitude prediction data output last in the T motion data is used as the target attitude prediction data.

[0134] The prediction result output module 64 is used to sequentially output M target attitude prediction data to obtain a prediction data sequence, where M=N / T; the output prediction data sequence is the attitude prediction result at the detection time.

[0135] The specific methods of execution of each module or unit in the above device embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0136] This disclosure also provides an electronic device, such as... Figure 7 As shown, the electronic device includes one or more processors 71 and a memory 72. Figure 7 Take a processor 71 as an example.

[0137] The electronic device may also include an input device 73 and an output device 74.

[0138] The processor 71, memory 72, input device 73, and output device 74 can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0139] Processor 71 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips. The general-purpose processor can be a microprocessor or any conventional processor.

[0140] The memory 72, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 71 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions, and modules stored in the memory 72, thereby implementing the methods in the above-described method embodiments.

[0141] The memory 72 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the use of the processing device operated by the server. Furthermore, the memory 72 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 72 may optionally include memory remotely located relative to the processor 71, and these remote memories can be connected to a network connection device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0142] Input device 73 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the server's processing device. Output device 74 may include display devices such as a display screen.

[0143] One or more modules are stored in memory 72 and, when executed by one or more processors 71, perform the methods as described in the above embodiments.

[0144] Those skilled in the art will understand that all or part of the processes in the above method embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes as described in the above method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory (FM), hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0145] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for predicting the motion attitude of a ship, characterized in that, The method is applied to electronic devices, in which a pre-trained Kolmogorov-Arnold recurrent network model is deployed. The Kolmogorov-Arnold recurrent network model includes recurrent units, each comprising a predictor, a linear transformation layer, and a Kolmogorov-Arnold network (KAN) module. The method includes: Multiple motion feature parameters of key positions of the ship are collected at each time step; The ship's motion data is determined based on multiple motion characteristic parameters; At any given detection time, acquire the latest consecutive N time steps of motion data that have been collected and determine them as a motion data sequence; Normalize the motion data sequence; The motion data in the motion data sequence is processed sequentially through a loop unit using a sliding window approach, where each sliding window processes T motion data points, as follows: For any target motion data corresponding to any target time step in each of the T motion data, the target motion data is input into the KAN module to obtain key feature data; Obtain the hidden state feature data that has been output after the motion data of the previous time step of the target time step has been processed by the loop unit; The hidden state feature data from the previous time step is transformed by a linear transformation layer to obtain dynamic feature data. By fusing key feature data and dynamic feature data, the hidden state feature data of the target time step is obtained; The hidden state feature data is input into the predictor to obtain pose prediction data; The last output pose prediction data from the T motion data is used as the target pose prediction data. Output M target attitude prediction data points sequentially to obtain a prediction data sequence, where M = N / T; The output prediction data sequence is the attitude prediction result at the detection time.

2. The method according to claim 1, characterized in that, The Kolmogorov-Arnold network (KAN) module includes a time basis function processing module and a time spline function processing module. Accordingly, the step of inputting the target motion data into the KAN module to obtain key feature data includes: The target motion data is input into the time basis function processing module to output the first key feature data; The target motion data is input into the time spline function processing module to output the second key feature data; Accordingly, the fusion of key feature data and dynamic feature data to obtain hidden state feature data for the target time step includes: Based on the first key feature data, the second key feature data, and the dynamic feature data, the hidden state feature data corresponding to the target time step is determined.

3. The method according to claim 2, characterized in that, The time basis function processing module mentioned above adopts... SiLU As a time basis function, its formula is expressed as follows: In the formula, The target motion data corresponding to time t of the target step. These are learnable weight parameters; This is the first key feature data; in, .

4. The method according to claim 3, characterized in that, The time spline function processing module mentioned above uses a set of B Spline basis functions are expressed in the following formula: In the formula, This is the second key feature data. For learnable time weights; These are the spline coefficients; As a group B spline basis functions; This represents the number of grid intervals for the time spline function. Let be the order of the time spline function. For summation index.

5. The method according to claim 1, characterized in that, The training process of the Kolmogorov-Arnold recurrent network model includes: Within a preset time period, measurement data from various attitude sensors at key positions of the ship are collected according to time steps. For each time step, multiple motion feature parameters are extracted from the measurement data; Based on the aforementioned multiple motion characteristic parameters, the instantaneous vertical displacement of the corresponding ship is calculated; Based on the multiple motion feature parameters and their corresponding instantaneous vertical displacements, a ship motion attitude training dataset is constructed. Based on the ship motion attitude training dataset, the initial Kolmogorov-Arnold recurrent network model is trained until the training error meets the preset error requirements, thus obtaining the trained Kolmogorov-Arnold recurrent network model.

6. The method according to claim 5, characterized in that, The aforementioned multiple motion characteristic parameters include the ship's heave displacement, the ship's longitudinal roll amplitude, and the ship's lateral roll amplitude. Accordingly, the formula for calculating the instantaneous vertical displacement of the ship based on the multiple motion characteristic parameters is as follows: In the formula, This refers to the instantaneous vertical displacement of the ship; This refers to heave displacement; The vertical sway amplitude is ; the horizontal sway amplitude is . The longitudinal distance between the key location and the center of the ship; The distance between the key location and the center of the ship is lateral.

7. The method according to claim 5, characterized in that, The loss function used when training the initial Kolmogorov-Arnold recurrent network model is as follows: In the formula, This is the loss value; This is a predicted value; The actual value; s is the sample size. .

8. A ship motion attitude prediction device, characterized in that, An application in electronic devices, wherein a pre-trained Kolmogorov-Arnold recurrent network model is deployed in the electronic device, wherein the Kolmogorov-Arnold recurrent network model includes recurrent units, and the recurrent units include predictors, linear transformation layers, and Kolmogorov-Arnold network KAN modules; the device includes: The acquisition module is used to acquire multiple motion feature parameters of key positions of the ship at each time step; The preprocessing module is used to determine the ship's motion data based on multiple motion feature parameters; at any detection time, it acquires the latest consecutive N time steps of collected motion data and determines them as a motion data sequence; and it performs normalization processing on the motion data sequence. The loop processing module processes each motion data point in the motion data sequence sequentially through a sliding window, with each sliding window processing T motion data points, as follows: For the target motion data corresponding to any target time step in each of the T motion data points, the target motion data is input into the KAN module to obtain key feature data; the hidden state feature data output after the motion data of the previous time step of the target time step has been processed by the loop unit is obtained; the hidden state feature data of the previous time step is transformed through a linear transformation layer to obtain dynamic feature data; the key feature data and dynamic feature data are fused to obtain the hidden state feature data of the target time step; the hidden state feature data is input into the predictor to obtain the pose prediction data; and the pose prediction data output at the end of the T motion data points is used as the target pose prediction data. The prediction result output module is used to sequentially output M target attitude prediction data to obtain a prediction data sequence, where M=N / T; the output prediction data sequence is the attitude prediction result at the detection time.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the ship motion attitude prediction method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the ship motion attitude prediction method according to any one of claims 1 to 7.

Images