Industrial control network flow prediction method and system

Through the construction and attention mechanism processing of industrial control network traffic tensors and linear deviation tensors, the problem of low prediction accuracy in the existing technology is solved, and high-precision prediction of industrial control network traffic is achieved, which is suitable for complex industrial control network environments.

CN118965090APending Publication Date: 2024-11-15GUANGDONG OPEN UNIV (GUANGDONG POLYTECHNIC VOCATIONAL COLLEGE) +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410885965.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-03
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing industrial control network traffic prediction methods fail to effectively reflect the nonlinear structure in the data, resulting in low prediction accuracy and difficult to adapt to complex industrial control network environments.

Method used

The industrial control network traffic tensor structure, linear deviation tensor calculation and attention mechanism are adopted, and the tensor and output layer output tensor are represented through initialization processing, weighted and rank one tensor calculation, and attention mechanism is used to represent the tensor and output layer output tensor, and the training process is finally carried out to obtain the target Flow identification, attributes and hidden feature matrix, thereby predicting industrial control network traffic.

Benefits of technology

The accuracy of traffic prediction in industrial control networks is improved, and traffic data containing timing information and nonlinear structure can be predicted more accurately.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965090A_ABST
    Figure CN118965090A_ABST
Patent Text Reader

Abstract

The invention discloses an industrial control network traffic prediction method and system. The method comprises the following steps: acquiring historical industrial control network traffic data; constructing an industrial control network flow tensor; calculating an industrial control network flow linear deviation tensor; performing initialization processing according to the industrial control network flow tensor to obtain a first rank-one tensor and a model parameter; according to the first connection weight, the first rank-one tensor and the first nonlinear activation function, calculating a weight sum rank-one tensor; calculating an attention mechanism representation tensor through an attention mechanism; calculating an output tensor of an output layer; according to the industrial control network flow tensor, the output layer output tensor and the model parameters, performing training processing to obtain a target flow identification linear deviation vector, a target attribute linear deviation vector and a target hidden feature matrix; and calculating an industrial control network flow prediction value. According to the invention, industrial control network flow prediction is realized, and the prediction precision is improved. The method can be widely applied to the technical field of industrial control networks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of industrial control networks, and in particular to an industrial control network traffic prediction method and system. Background Art

[0002] With the rapid development of industrial informatization, the integration of modern information technology and industrial control systems has gradually deepened. However, security incidents such as internal information leakage and external network attacks frequently occur in industrial control systems. At present, traditional methods construct historical industrial control network traffic into industrial control network traffic tensors and predict unknown industrial control network traffic through simple linear calculations. However, industrial control network traffic data contains complex nonlinear structures, resulting in low prediction accuracy.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for predicting industrial control network traffic, which effectively improves the prediction accuracy.

[0005] On the one hand, an embodiment of the present invention provides an industrial control network traffic prediction method, comprising the following steps:

[0006] Obtain historical industrial control network traffic data;

[0007] Constructing an industrial control network traffic tensor according to the historical industrial control network traffic data;

[0008] Calculating the linear deviation tensor of the industrial control network traffic according to the historical industrial control network traffic data;

[0009] Performing initialization processing according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters;

[0010] Calculate a weighted sum rank tensor according to a first connection weight between the input layer and the hidden layer, the first rank tensor, and a first nonlinear activation function;

[0011] Calculating an attention mechanism representation tensor through an attention mechanism according to the weighted sum rank tensor;

[0012] Calculate the output layer output tensor according to the weighted sum rank tensor, the second connection weight between the hidden layer and the output layer, the attention mechanism representation tensor, the industrial control network flow linear deviation tensor and the second nonlinear activation function;

[0013] According to the industrial control network flow tensor, the output layer output tensor and the model parameters, training processing is performed to obtain a target flow identification linear deviation vector, a target attribute linear deviation vector and a target latent feature matrix;

[0014] The industrial control network traffic prediction value is calculated according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix.

[0015] In some embodiments, constructing the industrial control network traffic tensor according to the historical industrial control network traffic data includes:

[0016] Dividing the historical industrial control network traffic data into time periods according to a preset number of groups to obtain a plurality of four-tuple group fragments;

[0017] According to the plurality of quaternion slices, construct a corresponding slice matrix;

[0018] Sorting the plurality of slice matrices according to a preset sorting strategy to obtain a sparse tensor;

[0019] A known data set is selected from the sparse tensor to obtain the industrial control network flow tensor, wherein the known data set is used to characterize that a flow identifier has an interactive relationship with an attribute in a current time period.

[0020] In some embodiments, the calculating the industrial control network traffic linear deviation tensor according to the historical industrial control network traffic data includes:

[0021] Initializing an initial flow identification linear deviation vector according to the historical industrial control network traffic data;

[0022] Initializing an initial attribute linear deviation vector according to the historical industrial control network traffic data;

[0023] The industrial control network flow linear deviation tensor is calculated according to the initial flow identification linear deviation vector and the initial attribute linear deviation vector.

[0024] In some embodiments, the initialization process is performed according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters, including:

[0025] Performing tensor latent feature decomposition on the industrial control network traffic tensor to obtain multiple initial latent feature matrices;

[0026] Initializing the model parameters, wherein the model parameters include the latent feature dimension of the initial latent feature matrix, the maximum number of training iterations, an iteration control variable, a convergence termination threshold, a regularization factor, and a learning rate;

[0027] The first latent feature vectors with the same sequence number in the multiple initial latent feature matrices are combined to obtain the first rank tensor.

[0028] In some embodiments, the calculating the weighted sum rank tensor according to the first connection weight between the input layer and the hidden layer, the first rank tensor and the first nonlinear activation function comprises:

[0029] Calculate a first output tensor based on the first rank tensor and the first connection weight;

[0030] The weighted sum rank tensor is calculated according to the first output tensor by the first non-linear activation function.

[0031] In some embodiments, the step of calculating an attention mechanism representation tensor by an attention mechanism according to the weighted and rank tensor includes:

[0032] Performing a connection operation on the second latent eigenvector in the weighted and rank tensor to obtain a connection vector;

[0033] Calculating the inter-vector attention according to the connection vector;

[0034] Calculating a dense latent feature vector according to the attention between the vectors;

[0035] Calculate the multi-head attention mechanism latent feature vector according to the dense latent feature vector;

[0036] Calculate the attention mechanism embedding layer network input according to the multi-head attention mechanism hidden feature vector;

[0037] Calculate the attention mechanism embedding layer network output according to the attention mechanism embedding layer network input and the third nonlinear activation function;

[0038] According to the attention mechanism embedding layer network output, the attention mechanism representation tensor is calculated.

[0039] In some embodiments, the calculating the output layer output tensor according to the weighted sum rank tensor, the second connection weight between the hidden layer and the output layer, the attention mechanism representation tensor, the industrial control network traffic linear deviation tensor and the second nonlinear activation function includes:

[0040] Calculate a second output tensor based on the weighted sum rank tensor and the second connection weight;

[0041] The output layer output tensor is calculated according to the second output tensor, the attention mechanism representation tensor, the industrial control network traffic linear deviation tensor and the second nonlinear activation function.

[0042] In some embodiments, the training process is performed according to the industrial control network flow tensor, the output layer output tensor and the model parameters to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix, including:

[0043] Constructing a target loss function according to the industrial control network flow tensor and the output layer output tensor;

[0044] According to the training iteration learning strategy, the target loss function and the preset iteration requirements, the initial flow identification linear deviation vector, the initial attribute linear deviation vector and the initial latent feature matrix are iteratively optimized to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix; the preset iteration requirements include that the iteration round number control variable in the model parameters is greater than the maximum training iteration round number or the function value difference is less than the convergence termination threshold, and the function value difference is used to characterize the absolute value of the difference between the current round target loss function value and the previous round target loss function value.

[0045] In some embodiments, the calculating the industrial control network traffic prediction value according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix includes:

[0046] The industrial control network traffic prediction value is calculated according to the target flow identification linear deviation vector, the target attribute linear deviation vector, the target latent feature matrix, the fourth nonlinear activation function and the interaction value between the latent feature vectors.

[0047] On the other hand, an embodiment of the present invention provides an industrial control network traffic prediction system, including:

[0048] An industrial control network flow data receiving module, wherein the industrial control network flow data receiving module is used to obtain historical industrial control network flow data;

[0049] An industrial control network flow tensor construction module, the industrial control network flow tensor construction module is used to construct an industrial control network flow tensor according to the historical industrial control network flow data;

[0050] An industrial control network traffic linear deviation module, the industrial control network traffic linear deviation module is used to calculate the industrial control network traffic linear deviation tensor according to the historical industrial control network traffic data;

[0051] An industrial control network traffic prediction module, the industrial control network traffic prediction module comprising an input layer unit, a hidden layer unit, an attention mechanism embedding layer unit, an output layer unit, a training unit, and an industrial control network traffic prediction value calculation unit;

[0052] An industrial control network traffic data module, the industrial control network traffic data module is used to store the historical industrial control network traffic data and the industrial control network traffic prediction value;

[0053] The input layer unit is used to perform initialization processing according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters;

[0054] The hidden layer unit is used to calculate a weighted sum rank tensor according to a first connection weight between the input layer unit and the hidden layer unit, the first rank tensor and a first nonlinear activation function;

[0055] The attention mechanism embedding layer unit is used to calculate the attention mechanism representation tensor according to the weighted and rank tensor;

[0056] The output layer unit is used to calculate the output layer output tensor according to the weighted sum rank tensor, the second connection weight between the hidden layer unit and the output layer unit, the attention mechanism representation tensor, the industrial control network flow linear deviation tensor and the second nonlinear activation function;

[0057] The training unit is used to perform training processing according to the industrial control network flow tensor, the output layer output tensor and the model parameters to obtain a target flow identification linear deviation vector, a target attribute linear deviation vector and a target latent feature matrix;

[0058] The industrial control network traffic prediction value calculation unit is used to calculate the industrial control network traffic prediction value according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix.

[0059] The beneficial effects of the present invention are as follows:

[0060] The present invention first obtains historical industrial control network traffic data, constructs an industrial control network traffic tensor, calculates an industrial control network traffic linear deviation tensor, then performs initialization processing based on the industrial control network traffic tensor to obtain a first rank tensor and model parameters, and calculates a weighted sum rank tensor, then calculates an attention mechanism representation tensor through an attention mechanism, and calculates an output layer output tensor, and finally performs training processing based on the industrial control network traffic tensor, the output layer output tensor and the model parameters to obtain a target flow identification linear deviation vector, a target attribute linear deviation vector and a target latent feature matrix, and calculates an industrial control network traffic prediction value, so that the industrial control network traffic tensor and the industrial control network traffic linear deviation tensor can be used to predict the industrial control network traffic, thereby improving the prediction accuracy.

[0061] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0063] Figure 1 This is a flow chart of a method for predicting industrial control network traffic according to an embodiment of the present invention;

[0064] Figure 2 A schematic diagram of an overall process of predicting industrial control network traffic according to an embodiment of the present invention;

[0065] Figure 3 The present invention is a schematic diagram of the structure of an industrial control network traffic prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are only examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the attached claims.

[0067] It is understood that the terms "first", "second", etc. used in this application can be used to describe various concepts in this article, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another concept. For example, without departing from the scope of the embodiment of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein can be interpreted as "at the time of" or "when" or "in response to determination".

[0068] The terms "at least one", "multiple", "each", "any", etc. used in this application, at least one includes one, two or more, multiple includes two or more, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0069] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0070] Before describing the embodiments of the present application in detail, some nouns and terms involved in the embodiments of the present application are first described. The nouns and terms involved in the embodiments of the present application are subject to the following explanations.

[0071] Attention Mechanism: Originated from the study of human vision. In cognitive science, due to the bottleneck of information processing, humans selectively focus on part of all information while ignoring other visible information. The above mechanism is usually called the attention mechanism. Different parts of the human retina have different degrees of information processing capabilities, namely acuity, and only the fovea centralis has the strongest acuity. In order to make rational use of limited visual information processing resources, humans need to select a specific part of the visual area and then focus on it. For example, when people read, usually only a small number of words to be read will be paid attention to and processed. In summary, the attention mechanism has two main aspects: deciding which part of the input needs to be paid attention to; allocating limited information processing resources to important parts.

[0072] In the related technology, with the rapid development of industrial informatization, the integration of modern information technology and industrial control systems has gradually deepened, and the traditional protection mode of industrial control systems has become increasingly difficult to adapt to the complex industrial control network environment. In recent years, security incidents such as internal information leakage and external network attacks have occurred frequently in industrial control systems. Therefore, combining the characteristics of industrial control network traffic to accurately predict industrial control network traffic data is the key to industrial control network anomaly detection. At present, many researchers have proposed a tensor-based latent feature decomposition model for predicting unknown industrial control network traffic, which uses known historical industrial control network traffic to predict unknown industrial control network traffic. However, the industrial control network traffic prediction method based on tensor latent feature decomposition constructs historical industrial control network traffic into an industrial control network traffic tensor and predicts unknown industrial control network traffic through simple linear calculations. However, the industrial control network traffic data not only has a simple linear structure, but also contains a complex nonlinear structure. The prediction method based on tensor latent feature decomposition cannot effectively reflect the nonlinear structure of industrial control network traffic, thus causing the problem of reduced prediction accuracy.

[0073] An industrial control network traffic prediction method and system provided by an embodiment of the present invention are specifically used for industrial control network traffic data, and can perform accurate industrial control network traffic prediction in accordance with statistical laws to solve the problem of industrial control network traffic prediction that contains timing information and nonlinear structures, thereby improving prediction accuracy.

[0074] An industrial control network traffic prediction method provided in an embodiment of the present application relates to the field of industrial control network technology. An industrial control network traffic prediction method provided in an embodiment of the present application can be applied to a terminal, can also be applied to a server, and can also be software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and a car terminal, etc., but is not limited to this; the server side can be configured as an independent physical server, or it can be configured as a server cluster or distributed system composed of multiple physical servers, and can also be configured as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application that implements an industrial control network traffic prediction method, etc., but is not limited to the above forms.

[0075] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0076] The following is a detailed explanation of the embodiments of the present application in conjunction with the accompanying drawings:

[0077] Figure 1 is an optional flowchart of an industrial control network traffic prediction method provided in an embodiment of the present application. Figure 1 The method may include but is not limited to steps S101 to S109.

[0078] Step S101, obtaining historical industrial control network traffic data;

[0079] Step S102: construct an industrial control network traffic tensor based on historical industrial control network traffic data;

[0080] Step S103, calculating the linear deviation tensor of the industrial control network traffic according to the historical industrial control network traffic data;

[0081] Step S104: Perform initialization processing according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters;

[0082] Step S105, calculating a weighted sum rank tensor according to a first connection weight between the input layer and the hidden layer, a first rank tensor and a first nonlinear activation function;

[0083] Step S106, calculating an attention mechanism representation tensor through an attention mechanism according to the weighted sum rank tensor;

[0084] Step S107, calculating the output layer output tensor according to the weighted sum rank tensor, the second connection weight between the hidden layer and the output layer, the attention mechanism representation tensor, the industrial control network flow linear deviation tensor and the second nonlinear activation function;

[0085] Step S108, performing training processing according to the industrial control network flow tensor, the output layer output tensor and the model parameters to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix;

[0086] Step S109, calculating the industrial control network traffic prediction value according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix.

[0087] Steps S101 to S109 shown in the embodiment of the present application implement industrial control network traffic prediction and improve prediction accuracy.

[0088] In step S101 of some embodiments, historical industrial control network traffic data may be obtained through a traffic database or other methods, but are not limited thereto.

[0089] In some embodiments, in step S102, constructing an industrial control network traffic tensor based on historical industrial control network traffic data may include but is not limited to the following steps:

[0090] According to the preset number of groups, the historical industrial control network traffic data is divided into time periods to obtain multiple four-tuple group fragments;

[0091] According to the slices of multiple quaternion groups, a corresponding slice matrix is ​​constructed;

[0092] According to the preset sorting strategy, multiple slice matrices are sorted to obtain sparse tensors;

[0093] A known data set is selected from the sparse tensor to obtain the industrial control network traffic tensor. The known data set is used to characterize the interactive relationship between the flow identifier and the attribute in the current time period.

[0094] In some embodiments, the historical industrial control network traffic data can be divided into time periods according to a preset number of groups to obtain multiple quadruplet slices, and then a corresponding slice matrix can be constructed based on the multiple quadruplet slices. Then, the multiple slice matrices can be sorted according to a preset sorting strategy to obtain a sparse tensor, and finally a known data set is selected from the sparse tensor to obtain an industrial control network traffic tensor. Among them, the known data set is used to characterize the interactive relationship between the flow identifier and the attribute in the current time period. Exemplarily, the preset number of groups can be set to K, and the historical industrial control network traffic data can be divided into time periods. All quadruplets P = (u, v, t, w) are divided into K time period data according to time period t to obtain multiple quadruplet slices, where K represents a positive integer, and its size can be determined according to specific needs without limiting the specific value range. Then, the quadruplet P with t = 1 is selected from the multiple quadruplet slices. (1) =(u,v,t,w), using P (1) The data (u, v, w) is used to construct the slice matrix Q (1) , Q (1) The size of the matrix is ​​I rows and J columns, and the elements in the matrix Q (1) ij represents the interaction between the i-th flow identifier and the j-th attribute in the first time period, where 1≤i≤I, 1≤j≤J, I is the number of flow identifiers, and J is the number of attributes. Then select the four-tuple P from the four-tuple P=(u,v,t,w) in turn. (2) , P (3) ,…,P (K) Construct the slice matrix Q separately (2) , Q (3) , …, Q (K) , and then according to the preset sorting strategy, use K slice matrices to sort from small to large from front to back in three-dimensional space according to the K time periods divided, and obtain the sparse tensor Z0∈R I×J×K , R represents a real number set. Since each flow identifier in the network only interacts with some attributes in a time period, the constructed industrial control network flow tensor Z0 is a sparse tensor. The known data set Γ and the unknown data set L in Z0 are given. Since this embodiment is only trained on the known data set, the known data set Γ can be selected from the sparse tensor to represent the set composed of all data of the industrial control network flow tensor, and the industrial control network flow tensor Z is obtained.

[0095] In some embodiments, in step S103, the linear deviation tensor of the industrial control network traffic is calculated based on the historical industrial control network traffic data, which may include but is not limited to the following steps:

[0096] Initialize the initial flow identification linear deviation vector based on historical industrial control network traffic data;

[0097] Initialize the initial attribute linear deviation vector based on historical industrial control network traffic data;

[0098] The linear deviation tensor of industrial control network traffic is calculated according to the initial flow identification linear deviation vector and the initial attribute linear deviation vector.

[0099] In some embodiments, the initial flow identification linear deviation vector a=[a1, a2, ..., a i ], where a i represents the linear deviation of the i-th flow identifier, and the modulus of a is M, M = I. Then, according to the historical industrial control network traffic data, the initial attribute linear deviation vector b = [b1, b2, ..., b j ], where b j represents the linear deviation of the jth attribute, and the modulus of b is N, N = J. Finally, the linear deviation tensor of the industrial control network traffic is calculated based on the initial flow identification linear deviation vector and the initial attribute linear deviation vector. Among them, the calculation formula of the linear deviation tensor of the industrial control network traffic is: Where N is the linear deviation tensor of industrial control network traffic, I is the number of flow identifiers, J is the number of attributes, and a i is the linear deviation of the ith stream identifier, b j is the linear deviation of the j-th attribute.

[0100] In some embodiments, in step S104, initialization processing is performed according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters, which may include but is not limited to the following steps:

[0101] Perform tensor latent feature decomposition on the industrial control network traffic tensor to obtain multiple initial latent feature matrices;

[0102] Initialize model parameters, including the latent feature dimension of the initial latent feature matrix, the maximum number of training iterations, the iteration control variable, the convergence termination threshold, the regularization factor, and the learning rate;

[0103] The first latent eigenvectors with the same sequence number in multiple initial latent feature matrices are combined to obtain a first rank tensor.

[0104] In some embodiments, the industrial control network flow tensor Z can be used as an input neuron to perform tensor latent feature decomposition on the industrial control network flow tensor to obtain multiple initial latent feature matrices. For example, the initial latent feature matrix can include a flow identification latent feature matrix U = [u1, u2, ..., u f ], attribute latent feature matrix V = [v1, v2, ..., v f ], time latent feature matrix W = [w1,w2,…,w f ]. Where uf is the fth latent eigenvector of the flow identifier latent feature matrix, v f is the fth latent eigenvector of the attribute latent feature matrix, w f is the fth latent feature vector of the time latent feature matrix. Then initialize the model parameters, wherein the model parameters may include the latent feature dimension F of the initial latent feature matrix, the maximum number of training iterations D, the number of iterations control variable d during the training process, the convergence termination threshold c, the regularization factor g, and the learning rate η. Exemplarily, the latent feature dimension F determines the latent feature space dimension of the initial latent feature matrix, is initialized to an integer, and can be 20. The size of the initial latent feature matrix is ​​determined by each dimension value of the corresponding industrial control network flow tensor Z and the latent feature dimension F, the size of the flow identification latent feature matrix U can be I rows and F columns, the size of the attribute latent feature matrix V can be J rows and F columns, and the size of the time latent feature matrix W can be K rows and F columns, and the three latent feature matrices can be initialized by randomly taking values ​​in the interval [0, 0.05]; the size of the initial flow identification linear deviation vector a and the initial attribute linear deviation vector b is determined by the dimensions I and J of the corresponding flow identification latent feature matrix U and the attribute latent feature matrix V, and the two linear deviations can be initialized by randomly taking values ​​in the interval [0, 0.05]. The maximum number of training iterations D is a variable that controls the upper limit of the iteration process and can be set to 1000. The variable d that controls the number of iterations in the training process can be initialized to 0. The convergence termination threshold c is a parameter that determines whether the iteration process has converged. It is initialized with a very small positive number and can be set to 10. -5 The regularization factor g is a constant that controls the regularization effect of the corresponding linear deviation vector a, v and the initial latent feature matrix U, V, W in the iteration process. It can be taken as 0.001. The learning rate η is used to control the update degree of the model parameters, which directly affects the training speed and convergence performance of the model. It can be taken as 0.002. Finally, the first latent feature vectors with the same sequence number in multiple initial latent feature matrices are combined to obtain a first rank tensor

[0105]

[0106] In some embodiments, in step S105, calculating a weighted sum rank tensor according to the first connection weight between the input layer and the hidden layer, the first rank tensor, and the first nonlinear activation function may include but is not limited to the following steps:

[0107] Compute a first output tensor based on a first rank tensor and a first connection weight;

[0108] Based on the first output tensor, a weighted sum rank tensor is calculated by a first non-linear activation function.

[0109] In some embodiments, a first rank tensor X may be firstf and the first connection weight M 1f , calculate the first output tensor. The calculation formula of the first output tensor is: In the formula, is the first output tensor, M 1f is the first connection weight, x f is a first-rank tensor, u f is the fth latent eigenvector of the flow identifier latent feature matrix, v f is the fth latent eigenvector of the attribute latent feature matrix, w f is the fth latent eigenvector of the temporal latent feature matrix, Represents the outer product between vectors. Then, according to the first output tensor, a weighted sum rank tensor is calculated by the first nonlinear activation function S1(·). The calculation formula of the weighted sum rank tensor is: In the formula, G f is a weighted and rank tensor, S1(·) is the first nonlinear activation function, is the first output tensor.

[0110] In some embodiments, in step S106, the attention mechanism representation tensor is calculated by the attention mechanism according to the weighted sum rank tensor, which may include but is not limited to the following steps:

[0111] Performing a connection operation on the second latent eigenvector in the weighted and rank tensor to obtain a connection vector;

[0112] According to the connection vector, calculate the attention between vectors;

[0113] Calculate the dense latent feature vector according to the attention between vectors;

[0114] According to the dense latent feature vector, calculate the latent feature vector of the multi-head attention mechanism;

[0115] Calculate the network input of the attention mechanism embedding layer based on the hidden feature vector of the multi-head attention mechanism;

[0116] Calculate the output of the attention mechanism embedding layer network according to the attention mechanism embedding layer network input and the third nonlinear activation function;

[0117] According to the attention mechanism embedding layer network output, the attention mechanism representation tensor is calculated.

[0118] In some embodiments, the second latent feature vector in the weighted and rank tensor may be first concatenated to obtain a concatenated vector of the attention mechanism embedding layer, where the expression of the concatenated vector is: l = {u i ,v j ,w k}={h1,h2,h3}. Then, according to the connection vector, the inter-vector attention is calculated, where the calculation formula of the inter-vector attention is: In the formula, is the attention between hidden feature vectors in the o-th single-head attention mechanism, <,> represents the inner product operation, is the feature element in the o-th single-head attention mechanism and The similarity between is the connection vector h l The first characteristic element in is the connection vector h l The second characteristic element in is the parameter matrix queried in the self-attention mechanism, is the parameter matrix of the key in the self-attention mechanism. Then, according to the attention between vectors, the latent feature vector h is calculated l The dense latent feature vector under the o-th single-head attention mechanism, where the calculation formula of the dense latent feature vector is: In the formula, is the dense latent feature vector, K k is the latent feature vector h l A collection of is the parameter matrix of the self-attention mechanism, h j is the jth latent eigenvector, is the hidden feature vector h in the o-th single-head attention mechanism l and h j Then, according to the dense latent feature vector, the latent feature vector of the multi-head attention mechanism is calculated, where the calculation formula of the latent feature vector of the multi-head attention mechanism is: In the formula, is the hidden feature vector of the multi-head attention mechanism, is the oth dense latent feature vector, In order to capture the high-order interaction between latent feature vectors, the attention mechanism embedding layer network input m can be calculated based on the multi-head attention mechanism latent feature vector. (0) , that is, the connection vector h l The information H obtained by calculating the encoding through the multi-head attention mechanism prep , where the calculation formula of the attention mechanism embedding layer network input is: In the formula, m (0) H is the network input of the attention mechanism embedding layer. prep is the connection vector h l The information encoded by the multi-head attention mechanism is calculated. is the hidden feature vector of the first multi-head attention mechanism, is the hidden feature vector of the second multi-head attention mechanism, is the third multi-head attention mechanism hidden feature vector, [,] represents the connection operation. Then, according to the attention mechanism embedding layer network input and the third nonlinear activation function, the output of the vth layer is used as the input of the v+1th layer to calculate the attention mechanism embedding layer network output, where the calculation formula of the attention mechanism embedding layer network output is: (v+1) =S3(W (v) m (v) +c (v) ), where m (v+1) is the network output of the v+1th layer attention mechanism embedding layer, S3(·) is the third nonlinear activation function, W (v) is the parameter matrix of the vth layer, m (v) is the output of the vth layer, c (v) is the linear deviation of the attention mechanism of the vth layer. Finally, according to the output of the attention mechanism embedding layer network, the attention mechanism representation tensor is calculated, that is, the output of the multi-head attention mechanism of all layers is calculated by the multi-head attention mechanism, where the calculation formula of the attention mechanism representation tensor is: H = S3 (W (L) m (L) +c (L) ), where H is the attention mechanism representation tensor, S3(·) is the third nonlinear activation function, and W (L) is the parameter matrix of the final layer, m (L) is the input of the final layer, c (L) is the linear bias of the attention mechanism in the final layer.

[0119] In some embodiments, in step S107, the output layer output tensor is calculated according to the weighted sum rank tensor, the second connection weight between the hidden layer and the output layer, the attention mechanism representation tensor, the industrial control network flow linear deviation tensor, and the second nonlinear activation function, which may include but is not limited to the following steps:

[0120] Calculate a second output tensor based on the weighted sum rank tensor and the second connection weight;

[0121] The output layer output tensor is calculated based on the second output tensor, the attention mechanism representation tensor, the industrial control network traffic linear deviation tensor, and the second nonlinear activation function.

[0122] In some embodiments, a weighted sum rank tensor G may be first calculated based on f and the second connection weight M 2f , calculate the second output tensor, where the calculation formula of the second output tensor is: Where Y is the second output tensor, M 2fis the second connection weight, G f is a weighted sum rank tensor, and F is the hidden feature dimension. Then, according to the second output tensor, the attention mechanism representation tensor constructed by the interaction between the hidden feature vectors, the industrial control network traffic linear deviation tensor, and the second nonlinear activation function S2, the output layer output tensor is calculated, where the calculation formula of the output layer output tensor is: In the formula, is the output tensor of the output layer, S2(·) is the second nonlinear activation function for calculating the closed tensor, Y is the second output tensor, B is the linear deviation tensor of the industrial control network traffic, and H is the attention mechanism representation tensor.

[0123] In some embodiments, in step S108, training processing is performed based on the industrial control network flow tensor, the output layer output tensor and the model parameters to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix, which may include but is not limited to the following steps:

[0124] Construct the target loss function according to the industrial control network traffic tensor and the output layer output tensor;

[0125] According to the training iteration learning strategy, the target loss function and the preset iteration requirements, the initial flow identification linear deviation vector, the initial attribute linear deviation vector and the initial latent feature matrix are iteratively optimized to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix; the preset iteration requirements include that the iteration round number control variable in the model parameters is greater than the maximum training iteration round number or the function value difference is less than the convergence termination threshold, and the function value difference is used to characterize the absolute value of the difference between the current round target loss function value and the previous round target loss function value.

[0126] In some embodiments, a target loss function may be constructed based on the industrial control network traffic tensor and the output layer output tensor, wherein the expression of the target loss function is:

[0127]

[0128] Where ε is the target loss function, A is the linear deviation vector value of the flow identifier, B is the linear deviation vector value of the attribute, U is the latent feature matrix of the flow identifier, V is the latent feature matrix of the attribute, W is the temporal latent feature matrix, and H is the tensor representing the attention mechanism. is the predicted value of industrial control network traffic, z ijk is the industrial control network traffic value generated by the interaction between the known entity, i.e., the ith flow identifier, and the kth time and jth attribute in the industrial control network traffic tensor Z, F is the number of columns of the three latent feature matrices decomposed from the industrial control network traffic tensor Z, u if is the value of the i-th row and f-th column in the flow identifier latent feature matrix U, v jfis the value of the jth row and fth column in the attribute latent feature matrix V, w kf is the value of the kth row and fth column in the temporal latent feature matrix W, a i is the linear deviation of the ith stream identifier, b j is the linear deviation of the jth attribute, η is the learning rate, h ijk To obtain the interaction value between hidden feature vectors through the attention mechanism embedding layer, Γ is the known data set in the sparse matrix, M 1f is the first connection weight between the input layer and the hidden layer, M 2f is the second connection weight between the hidden layer and the output layer.

[0129] Then, according to the training iteration learning strategy, the target loss function and the preset iteration requirements, the initial flow identification linear deviation vector, the initial attribute linear deviation vector and the initial latent feature matrix are iteratively optimized to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix. Among them, the preset iteration requirements include that the iteration round number control variable d in the model parameters is greater than the maximum training iteration round number D or the function value difference is less than the convergence termination threshold c, and the function value difference is used to characterize the absolute value of the difference between the current round target loss function value and the previous round target loss function value. In the iterative optimization process, the Euclidean distance can be used to measure the above-mentioned optimization objectives, and Tikhonov regularization can be used to constrain the optimization process to prevent overfitting problems in the optimization process, and then the stochastic gradient descent update rule is used to iteratively optimize the target loss function ε, thereby ensuring the iteration efficiency during the update process. In order to minimize the value of the target loss function ε, the learning rule of the training iteration is as follows:

[0130]

[0131] M 1f ←M 1f -h ijk d ijk u if v jf w kf -h ijk lm 1f ,

[0132]

[0133] u if ←u if -h ijk d ijk M 1f v jf w kf -h ijk lu if ,

[0134] v jf ←v jf -h ijk d ijk M 1f u if w kf -h ijk lv jf ,

[0135] w kf ←w kf -h ijk d ijk M 1f u if v jf -h ijk 1 kf ,

[0136]

[0137] In the formula, ← means that it is approximately equal to, l is the regularization parameter of the latent feature matrix and the two linear deviation vectors, is the weighted and rank tensor G output by the hidden layer f The corresponding element, d ijk Auxiliary parameters set to simplify subsequent representation.

[0138] The above training process can be repeated on the known data set Γ until the target loss function ε converges on the known data set Γ. The convergence judgment condition is that the training iteration round control variable d reaches the maximum training iteration round number D, or the absolute value of the difference between the value of the target loss function ε calculated after the current round of iteration and the value of the target loss function ε of the previous round is less than the convergence termination threshold c, and the training is stopped to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix. Among them, the target latent feature matrix can include the target flow identification latent feature matrix, the target attribute latent feature matrix, and the target time latent feature matrix.

[0139] In some embodiments, in step S109, the industrial control network traffic prediction value is calculated according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix, which may include but is not limited to the following steps:

[0140] The industrial control network traffic prediction value is calculated according to the target flow identification linear deviation vector, the target attribute linear deviation vector, the target latent feature matrix, the fourth nonlinear activation function and the interaction value between the latent feature vectors.

[0141] In some embodiments, the target stream identification linear deviation vector, the target attribute linear deviation vector, the target latent feature matrix, the fourth nonlinear activation function S4(·), and the interaction value h between the latent feature vectors can be used. ijk, calculate the predicted value of industrial control network traffic, where the calculation formula of the predicted value of industrial control network traffic is: In the formula, is the predicted value of industrial control network traffic, a i is the linear deviation of the ith stream identifier, b j is the linear deviation of the jth attribute, F is the number of columns of the three latent feature matrices decomposed by the industrial control network traffic tensor Z, and u if is the value of the i-th row and f-th column in the flow identifier latent feature matrix U, v jf is the value of the jth row and fth column in the attribute latent feature matrix V, w kf is the value of the kth row and fth column in the temporal latent feature matrix W, h ijk In order to obtain the interaction value between latent feature vectors through the attention mechanism embedding layer, S4(·) is the fourth nonlinear activation function.

[0142] In some embodiments, the overall process of predicting industrial control network traffic is as follows: Figure 2 As shown, the instruction of the perceptron to predict the industrial control network traffic can be received first, the received historical industrial control network traffic data can be stored in the form of a quadruple, and then the industrial control network traffic tensor can be constructed, and the initial industrial control network traffic linear deviation data can be created and stored, and then the input and output representations of the input layer neurons can be calculated, the relevant parameters can be initialized and trained, the input and output representations of the hidden layer neurons can be calculated, the training representation of the attention mechanism embedding layer can be calculated, and the input and output representations of the output layer neurons can be calculated. Finally, the industrial control network traffic prediction target loss function can be constructed, and the stochastic gradient descent algorithm can be used to iteratively optimize the loss function. If the iteration condition is met, the predicted value of the industrial control network traffic can be calculated and stored; otherwise, the loss function can be iteratively optimized.

[0143] The beneficial effects of implementing the embodiments of the present invention include: the embodiments of the present invention first obtain historical industrial control network traffic data, construct an industrial control network traffic tensor, calculate the industrial control network traffic linear deviation tensor, then perform initialization processing based on the industrial control network traffic tensor to obtain a first rank tensor and model parameters, and calculate a weighted sum rank tensor, then calculate the attention mechanism representation tensor through the attention mechanism, and calculate the output layer output tensor, and finally perform training processing based on the industrial control network traffic tensor, the output layer output tensor and the model parameters to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix, and calculate the industrial control network traffic prediction value, so that the industrial control network traffic tensor and the industrial control network traffic linear deviation tensor can be used to predict the industrial control network traffic, thereby improving the prediction accuracy. At the same time, this embodiment specifically acts on industrial control network traffic data, and can perform accurate industrial control network traffic prediction that conforms to statistical laws to solve the problem of industrial control network traffic prediction that contains time series information and nonlinear structures.

[0144] like Figure 3 As shown, an industrial control network traffic prediction system provided by an embodiment of the present application includes: an industrial control network traffic data receiving module 201, an industrial control network traffic data module 202, an industrial control network traffic tensor construction module 203, an industrial control network traffic linear deviation module 204, and an industrial control network traffic prediction module 205. The industrial control network traffic data receiving module 201 can be used to obtain historical industrial control network traffic data, and then the industrial control network traffic tensor can be constructed by the industrial control network traffic tensor construction module 203 according to the historical industrial control network traffic data, and then the industrial control network traffic linear deviation tensor can be calculated by the industrial control network traffic linear deviation module 204 according to the historical industrial control network traffic data, and finally the industrial control network traffic prediction value can be calculated by the industrial control network traffic prediction module 205 according to the industrial control network traffic tensor and the industrial control network traffic linear deviation tensor. At the same time, the industrial control network traffic data module 202 can be used to store historical industrial control network traffic data and industrial control network traffic prediction values. The industrial control network traffic prediction module includes an input layer unit 501, a hidden layer unit 502, an attention mechanism embedding layer unit 503, an output layer unit 504, a training unit 505, and an industrial control network traffic prediction value calculation unit 506. The industrial control network traffic data module includes an industrial control network traffic data storage unit 301 and an industrial control network traffic prediction value storage unit 302.

[0145] In some embodiments, the industrial control network traffic tensor can be initialized through the input layer unit 501 to obtain a first rank tensor and model parameters; a weighted sum rank tensor is calculated through the hidden layer unit 502 according to the first connection weight between the input layer unit and the hidden layer unit, the first rank tensor and the first nonlinear activation function; the attention mechanism representation tensor is calculated through the attention mechanism embedding layer unit 503 according to the weighted sum rank tensor; the output layer output tensor is calculated through the output layer unit 504 according to the weighted sum rank tensor, the second connection weight between the hidden layer unit and the output layer unit, the attention mechanism representation tensor, the industrial control network traffic linear deviation tensor and the second nonlinear activation function; the industrial control network traffic tensor, the output layer output tensor and the model parameters are trained through the training unit 505 to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix; the industrial control network traffic prediction value is calculated through the industrial control network traffic prediction value calculation unit 506 according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix. In some embodiments, the historical industrial control network traffic data can be stored by the industrial control network traffic data storage unit 301, and can be stored in the form of a quadruple. The quadruple representation is T = (u, v, t, w), where u represents the flow identifier in the network traffic data, that is, different types of data flows in the network traffic data, v represents the attributes contained in the network traffic data, that is, the different attributes contained in the data flow in the network traffic data, t represents the time period when the flow identifier u contains the attribute v, and w represents the interactive relationship data between the flow identifier u in the time period t and the attribute v. The industrial control network traffic prediction value can be stored by the industrial control network traffic prediction value storage unit 302, and can also be stored in the form of a quadruple.

[0146] The beneficial effects of implementing the embodiments of the present invention include: the embodiments of the present invention first obtain historical industrial control network traffic data through an industrial control network traffic data receiving module, then construct an industrial control network traffic tensor through an industrial control network traffic tensor construction module based on the historical industrial control network traffic data, and calculate the industrial control network traffic linear deviation tensor through an industrial control network traffic linear deviation module, and finally calculate the industrial control network traffic prediction value through an industrial control network traffic prediction module, so that the industrial control network traffic tensor and the industrial control network traffic linear deviation tensor can be used to predict the industrial control network traffic, thereby improving the prediction accuracy.

[0147] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A method for predicting industrial control network traffic, characterized in that: The following steps are involved: Obtain historical industrial control network traffic data; Constructing an industrial control network traffic tensor according to the historical industrial control network traffic data; Calculating the linear deviation tensor of the industrial control network traffic according to the historical industrial control network traffic data; Performing initialization processing according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters; Calculate a weighted sum rank tensor according to a first connection weight between the input layer and the hidden layer, the first rank tensor, and a first nonlinear activation function; Calculating an attention mechanism representation tensor through an attention mechanism according to the weighted sum rank tensor; Calculate the output layer output tensor according to the weighted sum rank tensor, the second connection weight between the hidden layer and the output layer, the attention mechanism representation tensor, the industrial control network flow linear deviation tensor and the second nonlinear activation function; According to the industrial control network flow tensor, the output layer output tensor and the model parameters, training processing is performed to obtain a target flow identification linear deviation vector, a target attribute linear deviation vector and a target latent feature matrix; The industrial control network traffic prediction value is calculated according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix.

2. The method according to claim 1, characterized in that: The step of constructing an industrial control network traffic tensor according to the historical industrial control network traffic data includes: Dividing the historical industrial control network traffic data into time periods according to a preset number of groups to obtain a plurality of four-tuple group fragments; According to the plurality of quaternion slices, construct a corresponding slice matrix; Sorting the plurality of slice matrices according to a preset sorting strategy to obtain a sparse tensor; A known data set is selected from the sparse tensor to obtain the industrial control network flow tensor, wherein the known data set is used to characterize that a flow identifier has an interactive relationship with an attribute in a current time period.

3. The method according to claim 1, characterized in that The step of calculating the linear deviation tensor of the industrial control network traffic according to the historical industrial control network traffic data includes: Initializing an initial flow identification linear deviation vector according to the historical industrial control network traffic data; Initializing an initial attribute linear deviation vector according to the historical industrial control network traffic data; The industrial control network flow linear deviation tensor is calculated according to the initial flow identification linear deviation vector and the initial attribute linear deviation vector.

4. The method according to claim 3, characterized in that: The initialization process is performed according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters, including: Performing tensor latent feature decomposition on the industrial control network traffic tensor to obtain multiple initial latent feature matrices; Initializing the model parameters, wherein the model parameters include the latent feature dimension of the initial latent feature matrix, the maximum number of training iterations, an iteration control variable, a convergence termination threshold, a regularization factor, and a learning rate; The first latent feature vectors with the same sequence number in the multiple initial latent feature matrices are combined to obtain the first rank tensor.

5. The method according to claim 1, characterized in that The step of calculating a weighted sum rank tensor according to a first connection weight between the input layer and the hidden layer, the first rank tensor and a first nonlinear activation function comprises: Calculate a first output tensor based on the first rank tensor and the first connection weight; The weighted sum rank tensor is calculated according to the first output tensor by the first non-linear activation function.

6. The method according to claim 1, characterized in that The step of calculating the attention mechanism representation tensor by the attention mechanism according to the weighted sum rank tensor includes: Performing a connection operation on the second latent eigenvector in the weighted and rank tensor to obtain a connection vector; Calculating the inter-vector attention according to the connection vector; Calculating a dense latent feature vector according to the attention between the vectors; Calculate the multi-head attention mechanism latent feature vector according to the dense latent feature vector; Calculate the attention mechanism embedding layer network input according to the multi-head attention mechanism hidden feature vector; Calculate the attention mechanism embedding layer network output according to the attention mechanism embedding layer network input and the third nonlinear activation function; According to the attention mechanism embedding layer network output, the attention mechanism representation tensor is calculated.

7. The method according to claim 1, characterized in that The step of calculating the output layer output tensor according to the weighted sum rank tensor, the second connection weight between the hidden layer and the output layer, the attention mechanism representation tensor, the industrial control network flow linear deviation tensor and the second nonlinear activation function includes: Calculate a second output tensor based on the weighted sum rank tensor and the second connection weight; The output layer output tensor is calculated according to the second output tensor, the attention mechanism representation tensor, the industrial control network traffic linear deviation tensor and the second nonlinear activation function.

8. The method according to claim 4, characterized in that The training process is performed according to the industrial control network flow tensor, the output layer output tensor and the model parameters to obtain a target flow identification linear deviation vector, a target attribute linear deviation vector and a target latent feature matrix, including: Constructing a target loss function according to the industrial control network flow tensor and the output layer output tensor; According to the training iteration learning strategy, the target loss function and the preset iteration requirements, the initial flow identification linear deviation vector, the initial attribute linear deviation vector and the initial latent feature matrix are iteratively optimized to obtain the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix; the preset iteration requirements include that the iteration round number control variable in the model parameters is greater than the maximum training iteration round number or the function value difference is less than the convergence termination threshold, and the function value difference is used to characterize the absolute value of the difference between the current round target loss function value and the previous round target loss function value.

9. The method according to claim 1, characterized in that: The step of calculating the industrial control network traffic prediction value according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix includes: The industrial control network traffic prediction value is calculated according to the target flow identification linear deviation vector, the target attribute linear deviation vector, the target latent feature matrix, the fourth nonlinear activation function and the interaction value between the latent feature vectors.

10. An industrial control network traffic prediction system, characterized in that: include: An industrial control network flow data receiving module, wherein the industrial control network flow data receiving module is used to obtain historical industrial control network flow data; An industrial control network flow tensor construction module, the industrial control network flow tensor construction module is used to construct an industrial control network flow tensor according to the historical industrial control network flow data; An industrial control network traffic linear deviation module, the industrial control network traffic linear deviation module is used to calculate the industrial control network traffic linear deviation tensor according to the historical industrial control network traffic data; An industrial control network traffic prediction module, the industrial control network traffic prediction module comprising an input layer unit, a hidden layer unit, an attention mechanism embedding layer unit, an output layer unit, a training unit, and an industrial control network traffic prediction value calculation unit; An industrial control network traffic data module, the industrial control network traffic data module is used to store the historical industrial control network traffic data and the industrial control network traffic prediction value; The input layer unit is used to perform initialization processing according to the industrial control network traffic tensor to obtain a first rank tensor and model parameters; The hidden layer unit is used to calculate a weighted sum rank tensor according to a first connection weight between the input layer unit and the hidden layer unit, the first rank tensor and a first nonlinear activation function; The attention mechanism embedding layer unit is used to calculate the attention mechanism representation tensor according to the weighted and rank tensor; The output layer unit is used to calculate the output layer output tensor according to the weighted sum rank tensor, the second connection weight between the hidden layer unit and the output layer unit, the attention mechanism representation tensor, the industrial control network flow linear deviation tensor and the second nonlinear activation function; The training unit is used to perform training processing according to the industrial control network flow tensor, the output layer output tensor and the model parameters to obtain a target flow identification linear deviation vector, a target attribute linear deviation vector and a target latent feature matrix; The industrial control network traffic prediction value calculation unit is used to calculate the industrial control network traffic prediction value according to the target flow identification linear deviation vector, the target attribute linear deviation vector and the target latent feature matrix.