CN117312783B - Ship heave compensation prediction method, device and storage medium based on neural network - Google Patents

Abstract

The invention relates to a ship heave compensation prediction method and device based on a neural network and a storage medium, belonging to the field of ship heave compensation prediction. Comprising the following steps: acquiring historical heave data of a ship, reducing noise, and dividing a data set into a training set and a testing set according to a proportion at random; performing outlier processing on the historical heave data; performing time delay processing on the data processed by the abnormal value, converting the data into a vector form suitable for the neural network through a plurality of time delay modules and a Mux module, and taking the vector form as a series of inputs of the neural network; acquiring time characteristics through a neural network, then fusing the time characteristics with the characteristics of a longicorn group algorithm and a cosine annealing algorithm of the wall-up, constructing a characteristic learning model based on the neural network and training the model; and carrying out differential filtering processing on the output signal of the neural network to output a result, so as to reduce the shake problem of the output signal of the network, wherein the result is used as a prediction result of heave displacement. The invention solves the problems that the algorithm is easy to fall into a local extremum and the convergence speed is low.

Description

Ship heave compensation prediction method, device and storage medium based on neural network

Technical field

The invention relates to a neural network-based ship heave compensation prediction method, device and storage medium, and belongs to the technical field of ship heave compensation prediction.

Background technique

Mother ships are required for ocean engineering operations such as seabed drilling, lifeboat retracting and releasing, offshore hoisting, and underwater towing. However, due to environmental disturbances such as waves and ocean currents, the mother ship will produce complex rocking motions with six degrees of freedom, driving the load to move with the mother ship. The heave direction has the greatest impact on the safety of offshore operations. Therefore, it is necessary to add a heave compensation system to the operating system.

The time delay between sensor measurement and controller execution in the heave compensation system will lead to poor compensation results. The use of ship heave historical motion data for ship heave compensation prediction technology can effectively compensate for the impact of system delays.

There are currently some algorithms that can use machine learning and neural network technology to predict ship heave compensation. For example, Chinese patent CN116307273A proposes a real-time ship movement prediction method and system based on XGboost machine learning algorithm. Compared with the neural network model, the training time and prediction time of the machine learning algorithm are greatly shortened, but the prediction accuracy gap is large. At the same time, the XGboost algorithm cannot directly identify the interactive relationship between features. It only mechanically uses data splitting rules to split each feature. Since the output signal of the sensor in the heave compensation system is a series of univariate time-series-based historical data of ship heave displacement, it is more appropriate to use a time-series neural network to predict ship heave compensation motion.

Sequential neural networks include single sequential neural networks, convolutional sequential neural networks, and combined sequential neural networks. Single sequential neural networks in sequential neural networks include but are not limited to fully connected neural networks (DNN, such as BP), recurrent neural networks (RNN, such as NARX, Transformer, LSTM, GRU, and improved Bi-LSTM, Bi-GRU), Temporal convolutional neural network (TCN); convolutional temporal neural network adds a traditional single-dimensional convolutional layer CNN based on a single temporal neural network, such as the Conv-Bi-LSTM model proposed by Chinese patent CN116280094A. A combined neural network is a combination of any two or more single sequential neural networks, such as BP-LSTM. RNN is famous for being good at mining time series features, but it is unable to extract single-dimensional and one-way spatial features of time series, while CNN and TCN can extract single-dimensional and one-way spatial features of time series, but are not good at extracting sequential features of time series elements. Generally speaking, the combined time series neural network has more comprehensive capabilities than a single model, can adapt to a variety of complex sea conditions, and significantly improves prediction accuracy.

Current research on ship heave prediction mainly focuses on the free combination and experiment of sequential neural networks, but current research often ignores the overall performance improvement of sequential neural networks by optimization algorithms. Sequential neural networks use the idea of negative gradient descent to correct weights and thresholds, which can easily fall into local extreme values. The swarm intelligence optimization algorithm has better results than the traditional biologically inspired algorithm instead of the gradient descent algorithm, such as the longhorned beetle swarm algorithm and the whale optimization algorithm proposed by Chinese patent CN116861617A. The longhorned beetle swarm algorithm is easy to implement, has a small amount of calculation, and only needs A small number of beetle individuals are needed to complete the optimization of the algorithm. The learning speed of the optimization algorithm is mainly determined by two aspects: one is the learning rate parameter η , and the other is the size of the activation function derivative. If the learning rate parameter η is too large, it will cause oscillation or even divergence in the network training process, and if it is too small, it will affect the update speed of weights and thresholds. At present, the field of ship heave prediction focuses on fixed learning rate, piecewise constant decay, and exponential decay, resulting in slow convergence speed and inability to adapt the iterative process.

Therefore, there is an urgent need for a neural network-based ship heave compensation prediction method and device to solve the problems existing in the existing technology.

Contents of the invention

In view of the shortcomings of the existing technology, the present invention provides a neural network-based ship heave compensation prediction method, device and storage medium to solve the problem that the algorithm is easy to fall into local extreme values and the convergence speed is slow.

The core technology of the present invention is to perform heave information outlier detection, time delay processing, construction of feature learning models and differential filtering processing on historical heaving data. The feature learning model mainly uses natural gas in the process of optimizing the parameters of the neural network. The herd algorithm replaces the traditional gradient descent algorithm, thus avoiding the local extreme value phenomenon that is easily produced by the traditional gradient descent algorithm. The present invention introduces the warm-up cosine annealing algorithm to replace the fixed learning rate parameter eta , so that the error function can converge to the optimal value faster; the adaptive T distribution is introduced to enhance the global search and local search capabilities.

The present invention adopts the following technical solutions:

In a first aspect, the present invention provides a neural network-based ship heave compensation prediction method, which includes the following steps:

(1) Obtain the ship's one-dimensional time series historical heave data and perform noise reduction processing, and randomly split the data set into a training set and a test set in proportion;

(2) Detect outliers on the one-dimensional time series historical heave data of the platform displacement and process the outliers;

(3) Perform delay processing on the one-dimensional time series data after outlier processing, pass through multiple delay modules and then convert it into a vector form suitable for neural networks through the Mux module, and together serve as a series of inputs to the neural network; the delay module The number is related to the number of sampling time points and prediction time points;

(4) Acquire temporal features through the neural network, and then fuse them with the features of the beetle swarm algorithm and warm-up cosine annealing algorithm to build a feature learning model based on the neural network and train the model;

(5) Perform differential filtering on the neural network output signal to output the result to reduce the jitter problem of the network output signal. The result is used as the prediction result of the heave displacement.

Preferably, step (1) also includes: performing maximum and minimum normalization on the data to map the data to the interval [-1,1].

Preferably, in step (2), outliers are detected through the Raida criterion. Assuming that the data set X obeys a normal distribution, the outliers are judged according to the following formula:

;

;

;

Where , _{_} _ _ _ _ _ _ _ _ +3 δ ) interval, the data is treated as abnormal data, and the mean value is used to fill the outliers detected by Laida's rule.

Preferably, the neural network is a BP neural network, and step (4) includes the following steps:

Initialize the parameters of the beetle population algorithm, randomly set the beetle position, and use the beetle position as the initialization parameter of the BP neural network;

Make a random vector and normalize the direction of the beetle's whiskers;

Based on the normalization processing results, the coordinate relationship between the left and right whiskers and the center of mass of the beetle is created;

Determine the odor intensity of the beetle's left and right whiskers according to the fitness function. The fitness function is established based on the distance between the input sample feature vector and the neuron weight vector value, and calculates the individual position estimate of the beetle;

Based on the position coordinates of individual longhorned beetles in the longhorned beetle population, calculate the individual fitness of longhorned longhorned beetles and the average fitness of longhorned longhorned beetles and compare them;

Establish an iterative update method for the beetle position, use the attenuation strategy to update the optimal spatial position of the beetle, and use the warm-up cosine annealing algorithm to update the learning rate;

Form a new solution that conforms to the adaptive T distribution based on the position of the optimal solution, and update the optimal spatial position of the beetle;

When the fitness function value reaches the set accuracy or the maximum number of iterations is reached, the current position of the beetle is used as the parameter value of the BP neural network.

Preferably, the BP neural network includes an input layer, a hidden layer and an output layer. First, the dimension D of the search space of the longhorned swarm algorithm is determined:

;

In the formula, M is the number of neurons in the input layer, L is the number of neurons in the hidden layer, and O is the number of neurons in the output layer. The topological structure of this example is determined to be in the form of MLO, where M, L, and O represent the number of neurons and nodes in the input layer, hidden layer, and output layer of the BP neural network respectively.

Preferably, the initialized BP neural network is used to fit the heave displacement to obtain the fitted value of the ship's heave displacement; the mean square error MSE of the fitted value of the heave displacement of the test set and the expected value is used as the beetle swarm algorithm fitness function of Promote the spatial search of the algorithm. The smaller this value indicates that the prediction model has better accuracy. The formula is expressed as:

;

In the formula, y _i is the fitted value of heave displacement, is the expected value of heave displacement, and iterative optimization is performed in the entire space area, so that the place where the fitness function value is minimum is the space optimal solution.

Preferably, assuming that the inputs of M nodes in the input layer are x ₁ , x ₂ ,..., x _m , then the output of the L -th hidden layer node is g _L :

;

In the formula: is the hidden layer weight, a _L is the hidden layer bias, f ₁ is the hidden layer activation function, and the f ₁ activation function is set to be the Tanh function, that is, /> ;

The output layer output h is:

;

In the formula, is the output layer weight,/> is the output layer bias; f ₂ is the output layer activation function, and the f ₂ activation function is set to be a linear function, that is/> ;

According to the actual output h of the test set and the expected output H, the error of the BP neural network is determined as:

;

After the error back propagation of the BP neural network, the weight threshold is updated. According to the error E , the increment of the connection weight from the Lth neuron of the hidden layer to the output layer has the following formula:

;

In the formula: η is the learning rate parameter of the neural network to adjust the speed and degree of oscillation of the neural network in searching for optimal weights;

The threshold incremental update formula from the Lth neuron of the hidden layer to the output layer is as follows:

;

According to the error E , the increment of the connection weight from the m-th neuron in the input layer to the L -th neuron in the hidden layer has the following formula:

;

The threshold incremental update formula from the m-th neuron in the input layer to the L- th neuron in the hidden layer is as follows:

.

Preferably, the process of the beetle swarm algorithm to find the optimal beetle spatial position best X when the fitness function is minimized through continuous iteration is as follows:

S1: Define the random direction vector of the k-th longhorned beetle. And perform normalization processing:

;

In the formula: D is the dimension of the search space, rands is a random number, Random unit vectors must be explored for the k-th longhorned beetle, k∈1, 2, 3,...k, k is the longhorned beetle population size;

S2: Create the spatial positions of the left and right beetles:

;

;

In the formula: is the spatial position of the kth beetle’s right whisker,/> is the spatial position of the left whisker of the k-th beetle, Respectively represent the spatial positions of the left and right whiskers of the k-th beetle in the t-th iteration,/> is the distance between the two whiskers at t iterations;

S3: Calculate the fitness function values of the left and right whiskers and/> , and determine the size of the two, and update the spatial position of the beetle as:

;

In the formula: sign() is the sign function, c is the learning factor, that is, the movement step size and exploration step/> The proportional coefficient between them, and the learning factor is introduced to enhance the memory learning ability of longamis groups and historical individuals,/> is the position estimate of the k-th beetle centroid in the t+1 iteration, which is calculated from the fitness of the two whiskers obtained through exploration;

According to the coordinate position of individual longhorned beetles in the longhorned beetle group, the fitness function definition formula is used to calculate the individual longhorned beetle fitness f and the average fitness fag of the longhorned beetle group, compare the intensity of the two, and update the position of longhorned longhorned individuals. And calculate the fitness function value at the current coordinate position;

The centroid of each beetle is updated when and only when the optimal value of the population changes. At this time, the beetle exploration step remains unchanged; otherwise, the beetle will use the current optimal value of the population as the center of mass position, using the attenuation strategy. Update, when the centroids of all beetles in the population are estimated, their group optimal values and step sizes are updated according to the following strategy:

;

;

In the formula: The centroid position updated after considering the group extreme value for longhorn beetles. When this position cannot obtain a better optimal fitness of the population than the last iteration, the estimated value of its own individual remains unchanged. x _g is the optimal position state, f _g is the optimal fitness value;

S4: Update the step size and distance between the two whiskers:

;

;

;

In the formula, is the initial step size,/> To specify t search steps,/> is the maximum value of the learning rate,/> is the minimum value of the learning rate,/> is the total number of iterations,/> is the current number of iterations,/> Represents a random number;/> is the probability threshold;

Through the warm-up cosine annealing algorithm and update the learning rate, the formula is:

;

In the formula, is the initial learning rate,/> Represented as setting the number of iterations of Warmup;

New solution for calculating center of mass position via adaptive T-distribution:

;

;

In the formula: It is a new solution that conforms to the centroid position of the adaptive distribution, Max(t) is the maximum number of iterations, y(t) is the T distribution with the degree of freedom parameter being the number of iterations t, z and v are parameters, and their values are taken according to the specific situation;

Finally, it is judged whether the fitness function value of the beetle swarm algorithm reaches the set accuracy fbest or whether the number of iterations t exceeds the maximum number of iterations Max(t);

If the conditions are met, stop the iteration and use the best X of the beetle space position at this time as the optimal initial weight of the BP neural network. Otherwise, continue the iteration according to S1-S4;

S5: The search is over. The best centroid coordinates obtained by the search are the weights and thresholds of the neurons in the neural network, which are assigned to the BP neural network, that is

;

.

In a second aspect, the present invention provides a neural network-based ship heave compensation prediction device, including:

Acquisition module, used to obtain ship heave displacement data and perform noise reduction processing;

The outlier processing module is used to detect and process outliers in ship heave and displacement data;

The delay module is used to perform delay processing on the ship's heave displacement data and convert it with the Mux module into a vector form suitable for neural networks;

The feature learning model building module obtains temporal features through the BP neural network, and then performs feature fusion with the beetle swarm algorithm and warm-up cosine annealing algorithm to complete the construction of the feature learning model;

The output module performs differential filtering on the neural network output signal and outputs it, and the result is used as the prediction result of the heave displacement.

In a third aspect, the present invention provides a readable storage medium. A computer program is stored in the readable storage medium. The computer program includes a program code for controlling a process to execute the process. The process includes the above-mentioned neural network-based ship heave compensation prediction. method.

Where the present invention is not detailed, existing technologies can be used.

The beneficial effects of the present invention are:

1. Compared with the existing technology, this application can effectively improve the prediction accuracy of ship heave compensation. At the same time, the temporal neural network proposed in the feature learning stage uses a single hidden layer to avoid multiple hidden layers without losing prediction accuracy. The resulting gradient disappearance, gradient explosion, and overfitting phenomena. In addition, the algorithm model of the present invention has high robustness and generalization ability.

2. In the face of massive data sets, manual processing will inevitably cause excessive workload. Therefore, the present invention uses the Laida criterion to detect outliers. This method has the advantages of being simple, relatively accurate, and widely applicable.

3. The present invention uses delay processing to predict by comparing multiple delays, and selects a duration suitable for the algorithm model.

4. The present invention uses the sequential neural network to obtain time features, and then performs feature fusion with the beetle swarm algorithm and the warm-up cosine annealing algorithm. In the feature fusion stage, the present invention uses the beetle swarm algorithm to replace the traditional gradient descent algorithm as the parameter optimization method of the neural network, thereby avoiding the local extreme value phenomenon that is easily produced by the traditional gradient descent algorithm. The warm-up cosine annealing algorithm is introduced to replace the fixed learning rate parameter eta , so that the error function can converge to the optimal value faster. The adaptive T distribution is introduced to enhance global search and local search capabilities.

5. The present invention introduces differential filtering processing into the output module, and selects appropriate parameters such as filter factors and speed factors to avoid large phase delays and amplitude reductions.

Description of drawings

Figure 1 is an overall flow chart of the ship heave compensation prediction method of the present invention;

Figure 2 is a flow chart of a ship heave compensation prediction method according to an embodiment of the present invention;

Figure 3 is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention;

In the figure, 1-memory, 2-processor, 3-transmission device, 4-input and output device.

Detailed ways

In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments, but it is not limited thereto. Anything that is not described in detail in the present invention is in accordance with conventional techniques in the art.

The input of the BP neural network of the present invention is a sequence of univariate historical ship heave and displacement data based on time series. The present invention selects 5 sampling time points to correspond to 1 prediction time point, so 4 delay modules are required, and the output is one-dimensional. sequentially.

Example 1

A neural network-based ship heave compensation prediction method, as shown in Figure 1, includes the following steps:

(1) Obtain the ship's one-dimensional time series historical heave data and perform noise reduction processing, and randomly split the data set into a training set and a test set in proportion;

In this embodiment, since the ship heave signal directly measured by the sensor contains a large amount of noise, a band-pass filter is used to process the collected sensor signal before entering the controller. However, the bandpass filter causes phase lag, and this part can be processed with time delay. Data preprocessing can reduce the impact of noise data on experimental results and facilitate the measurement of experimental effects. Therefore, this application performs maximum and minimum normalization on the data used so that the data is mapped to the interval [-1,1], as shown in the following formula:

;

Among them, x is the input data, x _max is the maximum input data value, x _min is the minimum input data value, is the normalized output value.

(2) Detect outliers on the one-dimensional time series historical heave data of the platform displacement and process the outliers;

In this embodiment, some elements of the collected data have large or small values, which obviously deviate from the actual situation, which are so-called outliers. However, in the face of massive data sets, manual processing will inevitably cause excessive workload. Therefore, this application uses the Laida criterion to detect outliers. Assuming that the data set X obeys a normal distribution, the outliers are judged according to the following formula:

;

;

;

Where , _{_} _ _ _ _ _ _ _ _ +3 δ ) interval, the data is treated as abnormal data, and the mean value is used to fill the outliers detected by Laida's rule.

(3) Perform delay processing on the one-dimensional time series data after outlier processing, pass through multiple delay modules and then convert it into a vector form suitable for neural networks through the Mux module, and together serve as a series of inputs to the neural network; the delay module The number is related to the number of sampling time points and prediction time points;

(4) Obtain time features through the neural network, and then fuse them with the features of the beetle swarm algorithm and warm-up cosine annealing algorithm to build a feature learning model based on the neural network and train the model, as shown in Figure 2;

The neural network is a BP neural network, and step (4) includes the following steps:

Initialize the parameters of the beetle population algorithm, randomly set the beetle position, and use the beetle position as the initialization parameter of the BP neural network;

Make a random vector and normalize the direction of the beetle's whiskers;

Based on the normalization processing results, the coordinate relationship between the left and right whiskers and the center of mass of the beetle is created;

Determine the odor intensity of the left and right beetles of the beetle according to the fitness function. The fitness function is established based on the distance between the input sample feature vector and the neuron weight vector value, and calculates the individual position estimate of the beetle;

According to the position coordinates of individual longhorned beetles in the longhorned beetle population, calculate the individual longhorned beetle fitness and the average fitness of the longhorned beetle population and compare them;

Establish an iterative update method for the beetle position, use the attenuation strategy to update the optimal spatial position of the beetle, and use the warm-up cosine annealing algorithm to update the learning rate;

Form a new solution that conforms to the adaptive T distribution based on the position of the optimal solution, and update the optimal spatial position of the beetle;

When the fitness function value reaches the set accuracy or the maximum number of iterations is reached, the current position of the beetle is used as the parameter value of the BP neural network.

The BP neural network includes an input layer, a hidden layer and an output layer. First, the dimension D of the search space of the longhorned swarm algorithm is determined:

;

In the formula, M is the number of neurons in the input layer, L is the number of neurons in the hidden layer, and O is the number of neurons in the output layer. The topological structure of this example is determined to be in the form of MLO, where M, L, and O represent the number of neurons and nodes in the input layer, hidden layer, and output layer of the BP neural network respectively.

The initialized BP neural network is used to fit the heave displacement, and the fitted value of the ship's heave displacement is obtained; the mean square error MSE of the fitted value of the heave displacement of the test set and the expected value is used as the fitness of the beetle swarm algorithm function Promote the spatial search of the algorithm. The smaller this value indicates that the prediction model has better accuracy. The formula is expressed as:

;

In the formula, y _i is the fitted value of heave displacement, is the expected value of heave displacement, and iterative optimization is performed in the entire space area, so that the place where the fitness function value is minimum is the space optimal solution.

Assume that the inputs of M nodes in the input layer are x ₁ , x ₂ ,..., x _m , then the output of the L -th hidden layer node is g _L :

;

In the formula: is the hidden layer weight, a _L is the hidden layer bias, f ₁ is the hidden layer activation function, and the f ₁ activation function is set to be the Tanh function, that is, /> ;

The output layer output h is:

;

In the formula, is the output layer weight,/> is the output layer bias; f ₂ is the output layer activation function, and the f ₂ activation function is set to be a linear function, that is/> ;

According to the actual output h of the test set and the expected output H, the error of the BP neural network is determined as:

;

After the error back propagation of the BP neural network, the weight threshold is updated. According to the error E , the increment of the connection weight from the Lth neuron of the hidden layer to the output layer has the following formula:

;

In the formula: η is the learning rate parameter of the neural network to adjust the speed and degree of oscillation of the neural network in searching for optimal weights;

The threshold incremental update formula from the Lth neuron of the hidden layer to the output layer is as follows:

;

According to the error E , the increment of the connection weight from the m-th neuron in the input layer to the L -th neuron in the hidden layer has the following formula:

;

The threshold incremental update formula from the m-th neuron in the input layer to the L- th neuron in the hidden layer is as follows:

;

Furthermore, the long-term beetle swarm algorithm uses continuous iterations to find the optimal long-term beetle spatial position best

S1: Define the random direction vector of the k-th longhorned beetle. And perform normalization processing:

;

In the formula: D is the dimension of the search space, rands is a random number, Random unit vectors must be explored for the k-th longhorned beetle, k∈1, 2, 3,...k, k is the longhorned beetle population size;

S2: Create the spatial positions of the left and right beetles:

;

;

In the formula: is the spatial position of the kth beetle’s right whisker,/> is the spatial position of the left whisker of the k-th beetle, Respectively represent the spatial positions of the left and right whiskers of the k-th beetle in the t-th iteration,/> is the distance between the two whiskers at t iterations;

S3: Calculate the fitness function values of the left and right whiskers and/> , and determine the size of the two, and update the spatial position of the beetle as:

;

In the formula: sign() is the sign function, c is the learning factor, that is, the movement step size and exploration step/> The proportional coefficient between them, and the learning factor is introduced to enhance the memory learning ability of longamis groups and historical individuals,/> is the position estimate of the k-th beetle centroid in the t+1 iteration, which is calculated from the fitness of the two whiskers obtained through exploration;

According to the coordinate position of individual longhorned beetles in the longhorned beetle group, the fitness function definition formula is used to calculate the individual longhorned beetle fitness f and the average fitness fag of the longhorned beetle group, compare the intensity of the two, and update the position of longhorned longhorned individuals. And calculate the fitness function value at the current coordinate position;

Furthermore, the centroid of each beetle is updated when and only when the optimal value of the population changes. At this time, the beetle exploration step remains unchanged; otherwise, the centroid will use the current optimal value of the population as the centroid position. Use the attenuation strategy to update. After the centroids of all beetles in the population are estimated, their group optimal values and step sizes are updated according to the following strategy:

;

;

In the formula: The centroid position updated after considering the group extreme value for longhorn beetles. When this position cannot obtain a better optimal fitness of the population than the last iteration, the estimated value of its own individual remains unchanged. x _g is the optimal position state, f _g is the optimal fitness value;

S4: Update the step size and distance between the two whiskers:

;

;

;

In the formula, is the initial step size, is/> Specify t search steps,/> is the maximum value of the learning rate,/> is the minimum value of the learning rate,/> is the total number of iterations,/> is the current number of iterations,/> Represents a random number;/> is the probability threshold; this embodiment proposes to use/> Strategy to generate a probability threshold. When the population extreme value cannot be updated in the current exploration, the probability threshold/> to perform the step decay process. /> The probability that the group loses its optimal position when the update condition is not met and the current fitness function value of the beetle centroid is less than the optimal fitness.

In order to solve the problem of drastic changes in the initial stage of the cosine annealing algorithm, the model can be preheated with a very small learning rate. The learning rate will gradually increase to a larger initial learning rate in the annealing stage under a predetermined number of iterations, allowing the model to undergo a transitional training process. The stability of model training is ensured, and the learning rate is updated through the warm-up cosine annealing algorithm. The formula is:

;

In the formula, is the initial learning rate,/> Represented as setting the number of iterations of Warmup;

The adaptive T distribution introduces adaptive parameters on the basis of the T distribution, and forms a new solution that conforms to the adaptive T distribution near the position of the optimal solution, which can combine the advantages of the Gaussian distribution and the Cauchy distribution. Use larger values for parameters in the early stages of the iteration to take advantage of the Cauchy distribution and enrich population diversity to improve global search performance; use parameters in the middle and later stages of the iteration to reduce the impact of the T distribution on the optimal solution, show the advantages of the Gaussian distribution, and retain the elite solution to enhance local search capabilities. New solution for calculating center of mass position via adaptive T-distribution:

;

;

In the formula: It is a new solution that conforms to the centroid position of the adaptive distribution, Max(t) is the maximum number of iterations, y(t) is the T distribution with the degree of freedom parameter being the number of iterations t, z and v are parameters, and their values are taken according to the specific situation;

Finally, it is judged whether the fitness function value of the beetle swarm algorithm reaches the set accuracy fbest or whether the number of iterations t exceeds the maximum number of iterations Max(t);

If the conditions are met, stop the iteration and use the best X of the beetle space position at this time as the optimal initial weight of the BP neural network. Otherwise, continue the iteration according to S1-S4;

S5: The search is over. The best centroid coordinates obtained by the search are the weights and thresholds of the neurons in the neural network, which are assigned to the BP neural network, that is

;

;

The optimal centroid coordinates obtained in this embodiment are used as the initial weights and thresholds in the BP neural network for training. During the training process, the BP neural network continuously adjusts the weights and thresholds based on the training data set to find the best weights and thresholds.

(5) Perform differential filtering on the neural network output signal to output the result to reduce the jitter problem of the network output signal. The result is used as the prediction result of the heave displacement.

Example 2

A neural network-based ship heave compensation prediction device, including:

Acquisition module, used to obtain ship heave displacement data and perform noise reduction processing;

The outlier processing module is used to detect and process outliers in ship heave and displacement data;

The delay module is used to perform delay processing on the ship's heave displacement data and convert it with the Mux module into a vector form suitable for neural networks;

The feature learning model building module obtains temporal features through the BP neural network, and then performs feature fusion with the beetle swarm algorithm and warm-up cosine annealing algorithm to complete the construction of the feature learning model;

The output module performs differential filtering on the neural network output signal and outputs it, and the result is used as the prediction result of the heave displacement.

Example 3

An electronic device includes a memory 1 and a processor 2. A computer program is stored in the memory 1, and the processor 2 is configured to run the computer program to execute the above-mentioned neural network-based ship heave compensation prediction method.

Specifically, the above-mentioned processor 2 may be an integrated circuit chip with signal processing capabilities. During the implementation process, each step of Embodiment 1 can be completed by instructions in the form of hardware integrated logic circuits or software in the processor. The above-mentioned processor may be a general-purpose processor, a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. Each method, step and logical block diagram disclosed in Embodiment 1 of this application can be implemented or executed. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. Or one or more integrated circuits may be configured to implement embodiments of the present application.

Among them, the memory 1 may include large-capacity memory for data or instructions. By way of example and not limitation, the memory 1 may include a hard disk drive (HDD), a floppy disk drive, a solid state drive (SSD), flash memory, an optical disk, a magneto-optical disk, a magnetic tape or a universal serial bus (USB) drive, or two or more A combination of the above. Where appropriate, memory 1 may comprise removable or non-removable (or fixed) media. Where appropriate, the memory 1 may be internal or external to the data processing device. In a particular embodiment, memory 1 is a non-volatile memory. In a particular embodiment, memory 1 includes read only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), electrically rewritable ROM (EAROM) or flash memory ( FLASH) or a combination of two or more of these. Where appropriate, the RAM can be static random access memory (SRAM) or dynamic random access memory (DRAM), where DRAM can be fast page mode dynamic random access memory (FPMDRAM), extended data output dynamic random access memory (FPMDRAM), or extended data output dynamic random access memory (DRAM). access memory (EDOD RAM), synchronous dynamic random access memory (SDRAM), etc.

The memory 1 may be used to store or cache various data files required for processing and/or communication, as well as possible computer program instructions executed by the processor 2 .

The processor 2 reads and executes the computer program instructions stored in the memory 1 to implement any neural network-based ship heave compensation prediction method in the above embodiments.

Optionally, the above-mentioned electronic device may also include a transmission device 3 and an input-output device 4, wherein the transmission device 3 is connected to the processor 2, and the input-output device 4 is connected to the processor 2.

The transmission device 3 can be used to receive or send data via a network. Specific examples of the above-mentioned network may include a wired or wireless network provided by a communication provider of the electronic device. In one example, the transmission device includes a network adapter (NIC), which is connected to other network devices through a base station to communicate with the Internet. In one example, the transmission device 3 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

The input and output device 4 is used to input or output information. In this embodiment, the input information may be ship heave displacement data, etc., and the output information may be ship heave displacement prediction results, etc.

Example 4

A readable storage medium. A computer program is stored in the readable storage medium. The computer program includes a program code for controlling a process to execute a process. The process includes the above-mentioned neural network-based ship heave compensation prediction method.

The above is the preferred embodiment of the present invention. It should be pointed out that for those of ordinary skill in the art, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications can also be made. should be regarded as the protection scope of the present invention.

Claims (8)

1. A neural network-based prediction method for ship heave compensation, which is characterized by including the following steps: (1) Obtain the one-dimensional time series historical heave data of the ship and perform noise reduction processing, and randomly split the data set into a training set and a test set in proportion; (2) Detect outliers on the one-dimensional time series historical heave data of the platform displacement and process the outliers; (3) Perform delay processing on the one-dimensional time series data after outlier processing, pass through multiple delay modules and then convert it into a vector form suitable for neural networks through the Mux module, and together serve as a series of inputs to the neural network; (4) Acquire temporal features through the neural network, and then fuse them with the features of the beetle swarm algorithm and warm-up cosine annealing algorithm to build a feature learning model based on the neural network and train the model; The neural network is a BP neural network, and step (4) includes the following steps: Initialize the parameters of the beetle population algorithm, randomly set the beetle position, and use the beetle position as the initialization parameter of the BP neural network; Make a random vector and normalize the direction of the beetle's whiskers; Based on the normalization processing results, the coordinate relationship between the left and right whiskers and the center of mass of the beetle is created; Determine the odor intensity of the beetle's left and right whiskers according to the fitness function. The fitness function is established based on the distance between the input sample feature vector and the neuron weight vector value, and calculates the individual position estimate of the beetle; Based on the position coordinates of individual longhorned beetles in the longhorned beetle population, calculate the individual fitness of longhorned longhorned beetles and the average fitness of longhorned longhorned beetles and compare them; Establish an iterative update method for the beetle position, use the attenuation strategy to update the optimal spatial position of the beetle, and use the warm-up cosine annealing algorithm to update the learning rate; Form a new solution that conforms to the adaptive T distribution based on the position of the optimal solution, and update the optimal spatial position of the beetle; When the fitness function value reaches the set accuracy or the maximum number of iterations is reached, the current position of the beetle is used as the parameter value of the BP neural network; (5) Perform differential filtering on the neural network output signal to output the result, which is used as the prediction result of heave displacement. 2. The neural network-based ship heave compensation prediction method according to claim 1, characterized in that step (1) also includes: performing maximum and minimum normalization on the data to map the data to the interval [-1,1 ]between. 3. The neural network-based ship heave compensation prediction method according to claim 2, characterized in that in step (2), abnormal values are detected through the Raida criterion. Assuming that the data set X obeys the normal distribution, then according to The following formula determines outliers: ; ; ; Among them, x represents the data set data, x _i represents the i- th data, μ represents the mean, δ represents the standard deviation, and n is the number of samples. If the value of x exceeds the ( μ -3 δ , μ +3 δ ) interval, Treat the data as outliers and use the mean to fill in the outliers detected by Laida's rule. 4. The neural network-based ship heave compensation prediction method according to claim 3, characterized in that the BP neural network includes an input layer, a hidden layer and an output layer, and first determines the dimension D of the search space of the beetle swarm algorithm. : ; In the formula, M is the number of neurons in the input layer, L is the number of neurons in the hidden layer, and O is the number of neurons in the output layer; The initialized BP neural network is used to fit the heave displacement, and the fitted value of the ship's heave displacement is obtained; the mean square error MSE of the fitted value of the heave displacement of the test set and the expected value is used as the fitness of the beetle swarm algorithm function The spatial search of the advancement algorithm is expressed as: ; In the formula, y _i is the fitting value of heave displacement, is the expected value of heave displacement, and iterative optimization is performed in the entire space area, so that the place where the fitness function value is minimum is the space optimal solution. 5. The neural network-based ship heave compensation prediction method according to claim 4, characterized in that assuming that the inputs of M nodes in the input layer are x ₁ , x ₂ ,..., x _m , then the Lth implicit The layer node output is g _L : ; In the formula: is the hidden layer weight, a _L is the hidden layer bias, f ₁ is the hidden layer activation function, and the f ₁ activation function is set to be the Tanh function, that is, /> ; The output layer output h is: ; In the formula, is the output layer weight,/> is the output layer bias; f ₂ is the output layer activation function, and the f ₂ activation function is set to be a linear function, that is/> ; According to the actual output h of the test set and the expected output H, the error of the BP neural network is determined as: ; After the error back propagation of the BP neural network, the weight threshold is updated. According to the error E , the increment of the connection weight from the Lth neuron of the hidden layer to the output layer has the following formula: ; In the formula: is the learning rate parameter of the neural network; The threshold incremental update formula from the Lth neuron of the hidden layer to the output layer is as follows: ; According to the error E , the increment of the connection weight from the m-th neuron in the input layer to the L -th neuron in the hidden layer has the following formula: ; The threshold incremental update formula from the m-th neuron in the input layer to the L- th neuron in the hidden layer is as follows: . 6. The neural network-based ship heave compensation prediction method according to claim 5, characterized in that the process of the beetle swarm algorithm to find the optimal beetle spatial position best X when the fitness function is minimized through continuous iteration is as follows: S1: Define the random direction vector of the k-th longhorned beetle. And perform normalization processing: ; In the formula: D is the dimension of the search space, rands is a random number, Random unit vectors must be explored for the k-th longhorned beetle, k∈1, 2, 3,...k, k is the longhorned beetle population size; S2: Create the spatial positions of the left and right beetles: ; ; In the formula: is the spatial position of the kth beetle’s right whisker,/> is the spatial position of the left whisker of the k-th beetle,/> Respectively represent the spatial positions of the left and right whiskers of the k-th beetle in the t-th iteration,/> is the distance between the two whiskers at t iterations; S3: Calculate the fitness function values of the left and right whiskers and/> , and determine the size of the two, and update the spatial position of the beetle as: ; In the formula: sign() is the sign function, c is the learning factor, that is, the movement step size and exploration step/> the proportional coefficient between is the position estimate of the kth beetle centroid in the t+1 iteration; According to the coordinate position of individual longhorned beetles in the longhorned beetle group, the fitness function definition formula is used to calculate the individual longhorned beetle fitness f and the average fitness fag of the longhorned beetle group, compare the intensity of the two, and update the position of longhorned longhorned individuals. And calculate the fitness function value at the current coordinate position; The centroid of each beetle is updated when and only when the optimal value of the population changes. At this time, the beetle exploration step remains unchanged; otherwise, the beetle will use the current optimal value of the population as the center of mass position, using the attenuation strategy. Update, when the centroids of all beetles in the population are estimated, their group optimal values and step sizes are updated according to the following strategy: ; ; In the formula: The centroid position updated after considering the group extreme value for longhorn beetles. When this position cannot obtain a better optimal fitness of the population than the last iteration, the estimated value of its own individual remains unchanged. x _g is the optimal position state, f _g is the optimal fitness value; S4: Update the step length and distance between the two whiskers: ; ; ; In the formula, is the initial step size,/> To specify t search steps,/> is the maximum value of the learning rate,/> is the minimum value of the learning rate,/> is the total number of iterations,/> is the current number of iterations,/> Represents a random number;/> is the probability threshold; Through the warm-up cosine annealing algorithm and update the learning rate, the formula is: ; In the formula, is the initial learning rate,/> Represented as setting the number of iterations of Warmup; New solution for calculating center of mass position via adaptive T-distribution: ; ; In the formula: It is a new solution that conforms to the centroid position of the adaptive distribution, Max(t) is the maximum number of iterations, y(t) is the T distribution whose degree of freedom parameter is the number of iterations t, z, v are parameters; Finally, it is judged whether the fitness function value of the beetle swarm algorithm reaches the set accuracy fbest or whether the number of iterations t exceeds the maximum number of iterations Max(t); If the conditions are met, stop the iteration and use the best X of the beetle space position at this time as the optimal initial weight of the BP neural network. Otherwise, continue the iteration according to S1-S4; S5: The search is over. The best centroid coordinates obtained by the search are the weights and thresholds of the neurons in the neural network, which are assigned to the BP neural network, that is ; . 7. A neural network-based ship heave compensation prediction device, characterized by including: Acquisition module, used to obtain ship heave displacement data and perform noise reduction processing; The outlier processing module is used to detect and process outliers in ship heave and displacement data; The delay module is used to perform delay processing on the ship's heave displacement data and convert it with the Mux module into a vector form suitable for neural networks; The feature learning model building module obtains temporal features through the BP neural network, and then performs feature fusion with the beetle swarm algorithm and warm-up cosine annealing algorithm to complete the construction of the feature learning model; the neural network is a BP neural network, and the feature fusion process for: Initialize the parameters of the beetle population algorithm, randomly set the beetle position, and use the beetle position as the initialization parameter of the BP neural network; Make a random vector and normalize the direction of the beetle's whiskers; Based on the normalization processing results, the coordinate relationship between the left and right whiskers and the center of mass of the beetle is created; Determine the odor intensity of the beetle's left and right whiskers according to the fitness function. The fitness function is established based on the distance between the input sample feature vector and the neuron weight vector value, and calculates the individual position estimate of the beetle; Based on the position coordinates of individual longhorned beetles in the longhorned beetle population, calculate the individual fitness of longhorned longhorned beetles and the average fitness of longhorned longhorned beetles and compare them; Establish an iterative update method for the beetle position, use the attenuation strategy to update the optimal spatial position of the beetle, and use the warm-up cosine annealing algorithm to update the learning rate; Form a new solution that conforms to the adaptive T distribution based on the position of the optimal solution, and update the optimal spatial position of the beetle; When the fitness function value reaches the set accuracy or the maximum number of iterations is reached, the current position of the beetle is used as the parameter value of the BP neural network; The output module performs differential filtering on the neural network output signal and outputs it, and the result is used as the prediction result of the heave displacement. 8. A readable storage medium, characterized in that a computer program is stored in the readable storage medium, and the computer program includes a program code for controlling a process to execute the process, and the process includes a method according to any one of claims 1 to 6. Ship heave compensation prediction method based on neural network.