A method, model and gateway for predicting satellite network traffic of ocean-going ships

Abstract

The present invention discloses a method for predicting network traffic of ocean-going ships based on satellite communications, comprising the steps of: obtaining network traffic data samples generated by ocean-going ships based on satellite communications, generating a first time series data set from the network traffic data samples according to the traffic time series; preprocessing the first time series data set to obtain a second time series data set; extracting short-term local features of the second time series data set using TCN; establishing a prediction model using LSTM: obtaining a predicted value of the ship network traffic through the prediction model. The present invention extracts high and low frequency information through short-term local dependencies in TCN learning time series and effectively captures long-term dependencies in time series through LSTM, providing a method for predicting network traffic of ocean-going ships with high accuracy and good robustness.

Description

Technical Field

[0001] The invention belongs to the technical field of ships, and in particular relates to a method, a model and a gateway for predicting satellite network traffic of ocean-going ships. Background Art

[0002] With the continuous development of ship technology, various ship equipment is undergoing large-scale improvement and innovation. At the same time, maritime satellite communication technology is developing rapidly, the scale of communication is expanding, the application of communication is growing sharply, and the communication behavior is becoming increasingly complex. Since ships are in the ocean and far away from land signal base stations, most ship network communications are based on maritime satellite communications. The characteristic of satellite communication systems is long communication distances. However, maritime satellite communications have the disadvantages of limited available resources, high operating costs, and being affected by the natural environment (such as solar eclipse, rainy and foggy weather, etc.). Satellite signals will also have signal delay problems when transmitted in space.

[0003] The stability and reliability of ship satellite communications and the navigation involved are directly related to the safety and stability of ship navigation. Faced with the complex satellite network communication environment, the analysis and distribution of ship satellite communication network traffic has become urgent. Summary of the invention

[0004] One of the embodiments of the present invention is a method for predicting satellite network traffic of ocean-going ships based on TCN and LSTM, which is used to predict the real-time traffic of network communication established between ocean-going ships and maritime satellites through onboard routing gateways. The prediction model predicts network traffic and helps optimize and solve the network resource allocation problem of ocean-going ships. The steps for establishing the prediction model are as follows:

[0005] The network communication traffic data of a large container ship is captured using a packet sniffing tool. All packets within a unit of time are taken as a sample, and all samples are combined to create an original time series data set.

[0006] After preprocessing the original time series data set, the features are extracted. After obtaining training samples from the time series data set, the neural network model based on the time convolution network TCN and the long short-term memory network LSTM is trained to obtain a prediction model to predict the network traffic of ocean-going ships.

[0007] The beneficial effect of the present invention is that a method for predicting network traffic of ocean-going ships is provided by extracting high and low frequency information through short-term local dependencies in TCN learning time series and effectively capturing long-term dependencies in time series through LSTM. The method of the present invention has high accuracy and good robustness. Experimental results show that the detection accuracy of the method can effectively improve the accuracy of ship network traffic prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, in which:

[0009] Figure 1 Flow chart of a method for predicting network traffic of ocean-going vessels according to one embodiment of the present invention.

[0010] Figure 2 A schematic diagram of a satellite network communication system for ocean-going vessels according to one embodiment of the present invention.

[0011] Figure 3 According to one embodiment of the present invention, time series data is converted into a supervised data flow graph.

[0012] Figure 4 A structural diagram of a long short-term neural network (LSTM) unit according to one embodiment of the present invention.

[0013] Figure 5 A diagram of the TCN and LSTM neural network model architecture according to one embodiment of the present invention. DETAILED DESCRIPTION

[0014] Since ship satellite communications have problems such as high latency, low bandwidth, packet loss, and connection interruption, accurately predicting ship network traffic in the future can help to know the future network traffic demand of ocean-going ships in advance, and adjust network strategies in a targeted manner and reasonably allocate network resources, so as to prevent network congestion and blockage to the greatest extent, improve the utilization of network resources, and better solve the problem of network resource allocation. By predicting ship network communications, problems such as unreasonable network resource allocation in the process of ocean-going ship satellite communications can be solved.

[0015] Network traffic prediction is a time series problem. Some mathematical methods and machine learning algorithms have been studied for this problem. For example, the AR model (Autoregressive Model) and the ARMA model (Autoregressive Moving Average Model) are used to predict bitstream network traffic. However, the above methods are all linear fits of network traffic sequences and cannot effectively capture the nonlinear characteristics of network traffic sequences.

[0016] In order to solve the problem that the above methods cannot capture the nonlinear characteristics in the network traffic sequence, the recurrent neural network (RNN) can capture the long-term dependencies in the sequence and has a natural advantage in the time series modeling task. Among the many variants of the recurrent neural network, the long short-term memory network (LSTM) is the most popular, overcoming the two key problems in the RNN: gradient explosion and gradient disappearance. At the same time, the temporal convolutional network (TCN) also shows good performance in some sequence modeling tasks, using causal hole convolution to obtain the long-term dependencies in the sequence, and using residual connections in the network to prevent network degradation. Therefore, the neural network model combining TCN and LSTM can be applied to the field of ocean-going ship network traffic prediction technology. This method is conducive to evaluating the carrying capacity of the ship network, and establishing an effective traffic prediction model has a positive effect on the prediction of ocean-going ship network communication traffic.

[0017] According to one or more embodiments, a method for predicting network traffic of a satellite network communication gateway of an ocean-going ship based on TCN and LSTM comprises the following steps:

[0018] Step 1: Use a packet sniffing tool to capture the traffic data of a satellite network communication gateway of a large ocean-going ship, and make it into original time series data based on all packets in a unit time as a sample.

[0019] Step 2: preprocess the original data to make it approximately conform to the normal distribution and more effectively predict network traffic. Then, the original sequence is smoothed using mean filtering to eliminate strong noise.

[0020] Step 3: Convert the preprocessed unlabeled dataset into a labeled supervised dataset.

[0021] Step 4: Apply TCN to the processed training dataset to extract short-term local features.

[0022] Step 5: Use the long short-term memory model LSTM to model the data.

[0023] Step 6: Use the model to predict new data and obtain the predicted value of ship network traffic.

[0024] In step 1, the packet sniffing tool is used to capture the network communication flow data of a large ship, and all packets per unit time are taken as a sample to make a data set. The details are as follows:

[0025] Packet sniffing tools are deployed on ocean-going ship communication equipment to periodically capture network traffic data on routing nodes in network communications within a unit time. All data packets within a unit time are taken as a sample, and all samples are combined together to use the model to make the data into an initial time series data set.

[0026] The original data set described in step 2 is transformed to make it approximately conform to the normal distribution and more effectively predict network traffic. Then, the original sequence is smoothed using a mean filter to eliminate strong noise. The details are as follows:

[0027] The order of magnitude of time series data is particularly large. After converting univariate temporal data into supervised data, the feature latitude of the data increases. The prediction task is a regression process, which is a nonlinear fitting for all features. If each latitude feature data is not mapped to the same range between the same regions and the same dimension is maintained, the operation will be extremely complicated, causing numerical problems, and slowing down the model fitting process, seriously affecting the convergence speed of the model. Therefore, this embodiment uses Min-Max Scaling to process time series data, as shown in the following formula:

[0028]

[0029] The mean filter is a high-frequency linear filter. It specifies a window of length m=5, replaces the middle value with the average value of all data in the window, and then moves the window with a step size of 1. Continue to repeat the above operation until the time series data ends, as shown in the following formula:

[0030]

[0031] Among them, m represents the window size, x i Represents the i-th data in the window, Represents the mean of all data in the window.

[0032] The preprocessed unlabeled dataset described in step 3 is converted into a labeled supervised dataset as follows:

[0033] In this embodiment, the characteristic value refers to the input sequence in the sliding window of step 2, which is used as the input data of the model, and the target value is the flow at the next moment, which is used as the output of the model. This embodiment adopts a window sliding method for time series data, sliding the window from left to right, sliding one time step each time, to obtain a new row of characteristic value and target value pairs, so that supervised data can be constructed. The data set is divided into a training set, a validation set, and a test set.

[0034] Apply the temporal convolutional network TCN to the processed data as described in step 4 to extract short-term local features, as follows:

[0035] TCN has certain advantages in processing time series data, which is defined as:

[0036] X T+1 =x 0 ,x 1 ,…,x T ,Y T+1 =y 0 ,y 1 ,…,y T ,

[0037] There exists a function f:X T+1 →Y T+1 ,

[0038] So that f(x 0 ,x 1 ,…,x T )=y 0 ,y 1 ,…,y T .

[0039] Note that for any y t Only with x 0 ,x 1 ,…,x T is related to x(t+1),x(t+2),…,x T Not relevant.

[0040] More formally, TCN defines a dilated convolution operation F on a sequence element s for a one-dimensional sequence input x and a filter f:{0,...,k-1}→R:

[0041]

[0042] Use TCN to smooth the sequence after mean filtering For feature extraction, TCN contains a residual module, and the output of the TCN layer can be calculated as follows:

[0043]

[0044] Among them, k represents the convolution kernel size, d represents the dilated convolution coefficient, ResidualBlock(*) represents the residual module of TCN, and L 1 Representation sequence Output of the residual module.

[0045] The data is modeled using the long short-term memory model LSTM as described in step 5, as follows:

[0046] LSTM is a neural network that has the ability to memorize long-term and short-term information, which can effectively avoid the gradient vanishing and gradient exploding problems in RNN. The calculation formula from input to output of LSTM unit is:

[0047]

[0048] i t =σ(W i [h t-1 ,x t ]+b i )

[0049]

[0050]

[0051]

[0052] h t =o t ⊙tanh(c t )

[0053] Among them, f,i,o, Respectively represent the forget gate, input gate, output gate and candidate vector, W c represents the weight matrix, σ(*) represents the sigmoid function, b c represents the deviation vector, tanh(*) represents the hyperbolic tangent function, and ⊙ represents the dot product.

[0054] Replace L in step 4 1 As the input of the LSTM layer, LSTM is used to extract the long-term dependencies in the sequence, thereby obtaining the output h of the last time step T Then the output h T As the input of the fully connected layer, the output z of this layer is obtained T Finally, z T As the input of the output layer, the traffic prediction value at the next moment is obtained

[0055] The model described in step 6 is used to predict new data to obtain the predicted value of ship network traffic, as follows:

[0056] The strategy of this embodiment is to make full use of the three technologies of the seamless integration solution, including: the mean filter smoothing and denoising capability, the feature extraction capability of TCN, and the long-term dependency capture capability of LSTM. Therefore, the obtained model can be named T-LSTM.

[0057] For the supervised data set obtained in step 3, it is processed into a format suitable for the TCN model, and then the parameters of the SLTM obtained by training are used to substitute the test set and parameters into the forward propagation of the loop layer in step 5 to obtain the prediction result of network traffic. For the evaluation index of the predicted value and the true value of network traffic, the root mean square logarithmic error (RMSLE) is used, as shown in the following formula:

[0058]

[0059] in, represents the predicted value of the i-th value, y i Represents the true value of the i-th value.

[0060] Compared with the existing technology, the significant advantage of this embodiment is that it integrates the time convolution network TCN and the long short-term neural network LSTM, thereby improving the dynamic allocation problem of network resources for ocean-going ships in satellite communication networks.

[0061] According to one or more embodiments, a method for predicting ocean-going ship satellite communication network traffic based on TCN and LSTM is implemented as follows: Figure 1 As shown, the following steps are included:

[0062] Step 101, capture the original data, using a packet sniffing tool to capture the original traffic data of the satellite network gateway communication of a large ocean-going ship.

[0063] Step 102, making a data set, taking the captured original traffic data and taking all packets in a unit time as a sample, making an original data set.

[0064] Step 103, data preprocessing, performs data preprocessing on the original data set to make it approximately conform to the normal distribution and more effectively predict network traffic. Then, the original sequence is smoothed using mean filtering to eliminate strong noise. The preprocessed unlabeled data set is converted into a labeled supervised training data set.

[0065] Step 104: Apply TCN to the training data set to extract short-term local features.

[0066] Step 105, using a long short-term memory model LSTM to model the data.

[0067] Step 106, using the model to predict new data, to obtain a predicted value of the ship network traffic.

[0068] In this embodiment, an implementation platform of a method for predicting ocean-going ship network traffic based on TCN and LSTM is a computer, the operating system may be Ubuntu 20.04, the deep learning framework may be Pytorch, the integrated development environment may be Pycharm, the development language may be Python, and the image acceleration unit may use NVIDIA GeForce RTX 3090.

[0069] The data capture in this embodiment is to deploy a packet sniffing tool on the ocean-going ship communication equipment, regularly capture the network traffic data on the routing node in the network communication in a unit time, take all the data packets in the unit time as a sample, and collect all the samples together, and use the model to make the data into an initial time series data set. The data preprocessing adopts the minimum and maximum normalization (Min-Max Scaling), as shown in the following formula:

[0070]

[0071] A mean filter with a window length of 5 is used to replace the middle value with the average value of all data in the window, and then the window is moved with a step size of 1. Continue to repeat the above operation until the time series data ends.

[0072] The preprocessed unlabeled dataset is converted into a labeled supervised dataset. In this embodiment, the feature value refers to Figure 2 The input sequence in the sliding window is used as the input data of the model, and the target value is the flow rate at the next moment, which is used as the output of the model. The window sliding method is used for time series data. The window is slid from left to right, one time step at a time, to obtain a new row of feature value and target value pairs, so that a supervised data set can be constructed.

[0073] After processing the obtained dataset, TCN is applied to extract short-term local features. TCN uses causal convolutions, where the output at time t is convolved only with the earlier elements in time t and the previous layer. If we look back at consecutive time steps, we can only look back at the number of layers in the network. To overcome this problem, dilated convolutions are used, taking input from every d steps from distance t, where k is the kernel size:

[0074] x t-(k-1)d ,…,x t-2d ,x t-d ,x t

[0075] Atrous convolutions let the network look back to (k-1)d time steps, making the receptive field grow exponentially with each layer. Each residual block has two layers of atrous causal convolutions, weight normalization, ReLU activations, and dropout. If the number of input channels is different from the number of output channels of the dilated causal convolution (the number of filters of the second dilated convolution), there is an optional 1x1 convolution. It is to ensure that the residual connection (element-wise addition of the convolution output and input) is valid.

[0076] Use TCN to smooth the sequence after mean filtering For feature extraction, TCN contains a residual module, and the output of the TCN layer can be calculated as follows:

[0077]

[0078] Among them, k represents the convolution kernel size, d represents the dilated convolution coefficient, ResidualBlock(*) represents the residual module of TCN, and L 1 Representation sequence Output of the residual module.

[0079] In this embodiment, the long short-term memory model LSTM is used to model the data. LSTM is a neural network that has the ability to memorize long-term and short-term information, which can effectively avoid the gradient disappearance and gradient explosion in RNN. The calculation formula from input to output of LSTM unit is:

[0080]

[0081] i t =σ(W i [h t-1 ,x t ]+b i )

[0082]

[0083]

[0084]

[0085] h t =o t ⊙tanh(c t )

[0086] Among them, f,i,o, Respectively represent the forget gate, input gate, output gate and candidate vector, W c represents the weight matrix, σ(*) represents the sigmoid function, b crepresents the deviation vector, tanh(*) represents the hyperbolic tangent function, and ⊙ represents the dot product.

[0087] L 1 As the input of the LSTM layer, LSTM is used to extract the long-term dependencies in the sequence, thereby obtaining the output h of the last time step T Then the output h T As the input of the fully connected layer, the output z of this layer is obtained T ,

[0088] z T =relu(vh T +bias)

[0089] Among them, v is the weight matrix, bias is the bias vector, and relu(*) represents the relu activation function.

[0090] Finally, z T As the input of the output layer, the traffic prediction value at the next moment is obtained

[0091]

[0092] Among them, u represents the weight parameter of the fully connected layer, q represents the bias parameter, and linear(*) represents the linear function.

[0093] At this point, the structure of the model has been introduced, and now we start predictive modeling.

[0094] In this example, the model is used to predict new data and obtain the predicted value of ship network traffic. The strategy is to make full use of the three technologies of the seamless integration solution: mean filter smoothing and denoising capability, TCN feature extraction capability and LSTM long-term dependency capture capability. Therefore, the obtained model can be named T-LSTM.

[0095] For the supervised data set obtained in step 3, it is processed into a format suitable for the TCN model, and then the parameters of the SLTM obtained by training are used to substitute the test set and parameters into the forward propagation of the loop layer in step 5 to obtain the prediction result of network traffic. For the evaluation index of the predicted value and the true value of network traffic, the root mean square logarithmic error (RMSLE) is used, as shown in the following formula:

[0096]

[0097] in represents the predicted value of the i-th value, y i Represents the true value of the i-th value.

[0098] The solution of this embodiment is to conduct comparative analysis on all models based on experimental data, so as to verify the effectiveness and robustness of the model proposed in this embodiment.

[0099] Compared with ordinary network communication, maritime satellite communication has the disadvantages of limited available resources, high operating costs, being affected by natural environments such as solar eclipse, rainy and foggy weather, and signal delay caused by satellite signal transmission in space. The ocean-going ship network traffic prediction method and model based on TCN and LSTM provided by the present invention are used to accurately and effectively predict ship network traffic, and can timely grasp the traffic fluctuation trend of the ship wireless network in operation, and provide guarantees for effectively solving problems such as congestion of ship wireless network. Since the ship is traveling in the ocean, other factors such as the longitude and latitude of the ship will have a certain impact on the network traffic. The TCN and LSTM neural network models are effectively used to capture the time and space characteristic attributes in the maritime satellite communication network traffic, and extract the high coupling correlation in the traffic characteristic data, which provides certain help for realizing the network traffic prediction with a longer time step in the complex and changeable marine communication environment.

[0100] It is worth noting that, although the foregoing content has described the spirit and principle of the invention with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined, and such division is only for the convenience of expression. The invention is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the appended claims.

Claims

1. A method for predicting traffic in a satellite communication network of an ocean-going ship, characterized in that: The following steps are involved: Step 1: deploy a packet sniffing tool on the ocean-going ship communication equipment, regularly capture network traffic data on the routing nodes in the network communication, take all data packets in a unit time as a sample, and collect all samples together to make the data samples into a first time series data set; Step 2, transforming the first time series data set to approximately conform to a normal distribution, and smoothing the transformed data set using a mean filter to obtain a second time series data set; Step 3, converting the second time series dataset into a labeled supervised dataset; Step 4, using TCN to extract short-term local features of the second time series data set; Step 5: Use the long short-term memory model LSTM to establish a network traffic prediction model; Step 6, obtaining a predicted value of the ship network flow through the prediction model; Wherein, the step 2 further comprises: Use minimum and maximum normalization to process the first time series data x and obtain: Specify a window of length m = 5, replace the middle value with the average of all data in the window, and then move the window in steps of 1 until the end of the time series data to obtain: Among them, m represents the window size, x i Represents the i-th data in the window, Represents the mean of all data in the window; In step 4, a temporal convolutional network (TCN) is used to extract short-term local features, and the definition is: duìX T+1 =x0,x1,...,x T ,Y T+1 =y0,y1,...,y T , There exists a function f:X T+1 →Y T+1 , So that f(x0, x1, ..., x T )=y0,y1,...,y T , Here, any y t Only with x0, x1, ..., x T is related to x(t+1), x(t+2), ..., x T No matter, Use TCN to smooth the sequence after mean filtering For feature extraction, TCN contains a residual module, and the output of the TCN layer can be calculated as follows: Among them, k represents the convolution kernel size, d represents the dilated convolution coefficient, ResidualBlock(*) represents the residual module of TCN, and L 1 Representation sequence Output of the residual module; The data is modeled using the long short-term memory model LSTM as described in step 5. The calculation formula from input to output of the LSTM unit is: i t =σ(W i [h t-1 ,x t ]+b i ) h t =o t ⊙tanh(c t ) Among them, f, i, o, Respectively represent the forget gate, input gate, output gate and candidate vector, W c represents the weight matrix, σ(*) represents the sigmoid function, b c represents the deviation vector, tanh(*) represents the hyperbolic tangent function, ⊙ represents the dot product, Replace L in step 4 1 As the input of the LSTM layer, LSTM is used to extract the long-term dependencies in the sequence, thereby obtaining the output h of the last time step T , Then, the output h T As the input of the fully connected layer, the output z of this layer is obtained T , z T =relu(vh T +bias) Among them, v is the weight matrix, bias is the bias vector, relu(*) represents the relu activation function, and finally, z T As the input of the output layer, the traffic prediction value at the next moment is obtained Among them, u represents the weight parameter of the fully connected layer, q represents the bias parameter, and linear(*) represents the linear function. The network traffic prediction model described in step 6 is used to predict new data to obtain the predicted value of the ship network traffic, specifically including: The supervised data set obtained in step 3 is processed into a format suitable for the TCN model. Then, the parameters of the SLTM obtained through training are used to substitute the test set and parameters into the forward propagation of the loop layer in step 5 to obtain the prediction results of the network traffic.

2. A satellite communication routing gateway for ocean-going vessels, characterized in that: The gateway comprises a memory and a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method according to claim 1.

Images