CN113743509B - Online combat intent recognition method and device based on incomplete information - Google Patents
Online combat intent recognition method and device based on incomplete information Download PDFInfo
- Publication number
- CN113743509B CN113743509B CN202111041309.1A CN202111041309A CN113743509B CN 113743509 B CN113743509 B CN 113743509B CN 202111041309 A CN202111041309 A CN 202111041309A CN 113743509 B CN113743509 B CN 113743509B
- Authority
- CN
- China
- Prior art keywords
- information
- model
- time
- data
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 48
- 238000012545 processing Methods 0.000 claims abstract description 39
- 238000013136 deep learning model Methods 0.000 claims abstract description 35
- 238000007906 compression Methods 0.000 claims abstract description 34
- 230000000295 complement effect Effects 0.000 claims abstract description 30
- 230000006835 compression Effects 0.000 claims abstract description 24
- 238000001514 detection method Methods 0.000 claims abstract description 18
- 238000012512 characterization method Methods 0.000 claims description 32
- 230000006870 function Effects 0.000 claims description 16
- 108091026890 Coding region Proteins 0.000 claims description 12
- 230000004927 fusion Effects 0.000 claims description 4
- 239000003550 marker Substances 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 14
- 238000005065 mining Methods 0.000 abstract description 7
- 230000008569 process Effects 0.000 description 12
- 238000012360 testing method Methods 0.000 description 10
- 230000008901 benefit Effects 0.000 description 8
- 239000000523 sample Substances 0.000 description 8
- 230000008859 change Effects 0.000 description 5
- 230000009471 action Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 230000019771 cognition Effects 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 241000282414 Homo sapiens Species 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000001502 supplementing effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 230000036648 cognitive speed Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000001939 inductive effect Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000002035 prolonged effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009469 supplementation Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 239000013598 vector Substances 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention provides an on-line combat intention recognition method and device based on incomplete information, which are characterized in that information data are acquired through various detection and sensing equipment to obtain original time-varying situation information formed by continuous tracking signals of each target unit in a time period delta T; performing coding complement compression processing on the original time-varying situation information to obtain effective input data; inputting the training data into a deep learning model to obtain a trained deep learning model; and inputting the current information data into a trained deep learning model to obtain a target intention recognition result. The learner is used for mining global structures, learning the representation of potential shared information, mining more global structures from limited battlefield information, discarding low-level information and more local noise, taking the time characteristics of target information into consideration, designing a variable-length time sequence processing model for learning intention classification, and realizing the online intention recognition effect under incomplete information.
Description
Technical Field
The invention belongs to the technical field of target intention recognition, and particularly relates to a method and a device for recognizing incomplete information on-line combat intention.
Background
The situation understanding is a process of explaining the current situation according to situation feature vectors generated by situation awareness and combining military knowledge of field experts to identify enemy intention and combat plan. The battlefield target combat intention recognition is always the focus of attention of commanders at all levels, is a hotspot problem in the situation assessment field, and is an important basis for the commander to decide the next combat action.
With the continuous development of information technology, a large number of reconnaissance detection and sensing devices are applied to the battlefield, so that the capability of collecting information reconnaissance and battlefield data is greatly improved. However, all these need to be used by decision makers after data analysis, and the information interface is challenged by information overload, so that the cognitive speed and processing capacity of human beings have difficulty in keeping pace with the pace of battlefield data growth and change, and even more so, the tactical intent of the enemy target is rapidly and accurately identified from the battlefield situation of instant change. The target intention recognition belongs to the pattern recognition problem under dynamic and countermeasure conditions, key information such as battlefield environment, target attribute, target state and the like is comprehensively considered on the basis of military knowledge and combat experience, accurate recognition of the target intention is realized through a series of highly abstract complex thinking activities such as key feature extraction, contrast analysis, association, reasoning and the like, and the process is difficult to describe and generalize explicitly through a mathematical formula, so that an efficient intelligent recognition model is required to be designed to assist a commander in decision making, so that decision making time is shortened, and decision quality is improved. Existing target intent recognition studies have focused mainly on template matching, expert systems, bayesian networks, and neural networks. In general, the template matching method accords with the human cognition rule and is easy to realize, but the establishment of a template library depends on the acquisition of prior knowledge of domain experts, the objectivity and the credibility are difficult to ensure, and the updating of the template library is difficult; although expert system has strong knowledge expression and knowledge reasoning ability, it has great difficulty in realization and weak fault tolerance and learning ability because of the need of abstracting complete knowledge base and reasoning rule; the Bayesian network has strong causal probability reasoning capability, receives extensive attention, and can solve the problem of uncertainty reasoning of the intention, but the prior probability of each node event of the Bayesian network has certain difficulty in determining the conditional probability. The neural network is successfully applied to various fields, and the self-adaption and self-learning capabilities of the neural network are used for predicting the intention, so that the problem of target intention recognition when the prior knowledge of the expert in the field is insufficient can be better solved. However, the traditional shallow neural network has the defects of difficult network training, high feature extraction difficulty, low calculation precision and the like.
Most importantly, the existing research results rarely discuss the influence of uncertainty and incompleteness of battlefield situation information on an intelligent model, but war is a typical imperfect information game, on an opposing battlefield, the 'three-incompleteness' characteristic of the battlefield situation information can be caused by the own concealment, mutual deception, uncertainty of the war and the like of each party, and the problem of how to realize efficient online intention recognition tasks is urgent to be solved in the face of massive incomplete, untimely and inaccurate, even erroneous or with deceptive information.
In order to solve the problem, a deep learning model W-cpcstm is proposed herein, which learns the representation of potential shared information by means of Comparative Predictive Coding (CPC), can mine more global structures from limited battlefield information and discard low-level information and more local noise, comprehensively considers the time characteristics of target information, designs a variable-length time sequence processing model LSTM to learn intention classification, and then effectively combines the two based on weights representing training attention, and discusses the on-line intention recognition effect of the proposed model under incomplete information by means of information of three different detection degrees and a perfect situation under ideal conditions. In addition, the effect of different lengths of intelligence information on the model is discussed.
Disclosure of Invention
The invention aims to solve the technical problem of realizing efficient on-line intention recognition when facing massive incomplete, untimely and inaccurate, even wrong or deceptive information, and provides a method and a device for recognizing incomplete information on-line combat intention.
In order to solve the technical problems, the invention adopts the following technical scheme:
an incomplete information online combat intention recognition method comprises the following steps:
step 1: the various detecting and sensing devices are paired with the ith at the current moment tIndividual target unitsThe continuous tracking signal within the time period delta T is integrated and encoded to obtain the original time-varying situation information TU ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N }, whereintu t The original time-varying situation information at the moment T is represented, N refers to the number of targets detected in a time period delta T, and T is the time length of delta T;
step 2: for the original time-varying situation information TU ΔT Performing coding complement compression processing to obtain effective deep learning model input data TU ΔT,P P represents a complement compression process;
step 3: inputting data TU into effective deep learning model ΔT,P Inputting the information to a deep learning model for information characterization learning and intention classification;
step 4: and obtaining a target intention recognition result.
Further, the deep learning model includes:
a learner for characterizing underlying shared information between information situational data acquired by various detection and sensing devices;
the classifier is used for accurately identifying the fight intention of the detected target at the current moment according to the bottom shared information obtained by the learner at the current time period;
and a controller for rationally distributing training attention between the learner and the classifier.
Further, the main components of the learner CPC include:
a variable-length time sequence processing model LSTM for extracting situation information TU of N detection target units obtained in the current time period delta T ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N Coding, complementing and compressing to obtainWherein-> The model input sequence under the incomplete information after the completion and compression of the ith target unit is input into a variable-length time sequence processing model LSTM, fusion coding is carried out on the model input sequence, and a potential representation coding sequence EU= { Eu of the situation information of the target unit is obtained t-T ，…，Eu t-1 ，Eu t }，Eu t Potential representation codes output at the time t;
autoregressive model of GRU: efficient time-step information characterization for potential characterization of the target unit situation information coding sequence EUSummarizing, IL r Original information length before complementation;
fully connected prediction layer: bottom shared information SI between information situation data for representing current moment t based on summarized features obtained by autoregressive model t ；
The three parts are combined and optimized through InfoNCE loss, and a loss function L of the learner CPC model is defined CPC ：
Wherein, (Tu) p,t ,SI t ) Can be regarded as a positive sample pair, tu p,t Representing situation information after completion of compression at time t,can be regarded as a negative sample pair, f (Tu p,t ,SI t ) Is the density ratio, f (Tu p,t ,SI t )＝exp(Eu t SI t )。
Further, the model structure of the classifier is as follows: the variable length time series data processing model LSTM connected to a linear output layer, the loss function uses a basic cross entropy loss function:
wherein,target intention labels of N target units at the current time t under' God vision>To identify the intention of all targets detected, the final inference result is obtained based on the potential representation coding sequence EU learned by the learner.
Further, the model loss function of the classifier:
L w ＝αL CPC +βL LSTM
alpha and beta respectively represent the weight parameters of the learner characterization learning and the weight parameters of the classifier classification learning.
Further, the complementary compression processing means: performing 0-supplementing processing on target unit data with the length smaller than delta T, and simultaneously recording the original data length IL of the target unit r Then based on IL r And carrying out marker compression on the complement data, wherein the deep learning model only calculates the non-complement data.
The invention also provides an online combat intention recognition device based on incomplete information, which comprises the following modules:
the information data acquisition module: for pairing various detecting and sensing devices to an ith target unit at a current time tThe continuous tracking signal within the time period delta T is integrated and encoded to obtain the originalTime-varying situation information TU ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N }, wherein-> tu t The original time-varying situation information at the moment T is represented, N refers to the number of targets detected in a time period delta T, and T is the time length of delta T;
and the data processing complement module: for the original time-varying situation information TU ΔT Performing complement compression processing to obtain effective deep learning model input data TU ΔT,P P represents a complement compression process;
and a learning classification module: for inputting data TU into a deep learning model to be effective ΔT,P Inputting the information to a deep learning model for information characterization learning and intention classification;
the identification result output module: and the intent classification result obtained by the learning classification module is decoded and output.
By adopting the technical scheme, the invention has the following beneficial effects:
according to the on-line combat intent recognition method and device based on incomplete information, incomplete battlefield information is complemented, so that a time sequence data processing model LSTM can be used in batch processing, the applicability of the model is enhanced, computing resources and time are saved, and in order to avoid result errors, the complemented data is subjected to marking compression, so that the model only calculates effective information, namely non-complemented marked data. According to the invention, through mining global structures by using learner CPC, the representation of potential shared information is learned, more global structures can be mined from limited battlefield information, low-level information and more local noise can be discarded, the time characteristics of target information are comprehensively considered, a variable-length time sequence processing model LSTM is designed for learning intention classification, and then the learning intention classification and the learning intention classification are effectively combined based on the weight for representing training attention, so that the online intention recognition effect under incomplete information can be realized.
Drawings
FIG. 1 is a flow chart of online intent recognition based on incomplete information in accordance with the present invention;
FIG. 2 is a deep learning model framework diagram;
FIG. 3 is a comparison of intent recognition accuracy (a) and loss value (b) between LSTM and the deep learning model W-CPCLSTM of the present invention based on incomplete information of target position ambiguity
FIG. 4 is a graph comparing the intent recognition accuracy (a) and the penalty value (b) of LSTM and W-CPCLSTM based on incomplete information with a definite target standpoint;
FIG. 5 is a comparison of the accuracy of each training generation intent recognition (a) and the loss value (b) for W-CPCLSTM of different training attentions in the case of incomplete information with unclear target positions;
FIG. 6 is a graph comparing the accuracy of intent recognition (a) and the loss value (b) between LSTM and W-CPCLSTM (AB 3) based on incomplete information with ambiguous target location;
FIG. 7 is a graph comparing the intent recognition accuracy (a) and the penalty value (b) of LSTM and W-CPCLSTM based on incomplete information with a clear target location;
fig. 8 compares the intention recognition accuracy (a) and the loss value (b) between LSTM and W-cpcstm (AB 5) based on incomplete information with ambiguous target types.
FIG. 9 compares the accuracy of intent recognition (a) and the penalty value (b) for LSTM and W-CPCLSTM based on incomplete information that is explicit in the target class.
Detailed Description
The following description of the embodiments of the present invention will be made apparent and fully in view of the accompanying drawings, in which some, but not all embodiments of the invention are shown. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Fig. 1 to 9 show a specific embodiment of an on-line combat intent recognition method based on incomplete information according to the present invention, as shown in fig. 1, comprising the steps of:
step 1: the various detecting and sensing devices are directed to the ith target unit at the current moment tThe continuous tracking signal within the time period delta T is integrated and encoded to obtain the original time-varying situation information TU ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N }, whereintu t The original time-varying situation information at time T is represented, N refers to the number of targets detected during a time period Δt, and T is the time length of Δt.
In this embodiment, the original time-varying situation information TU of each target unit in the time period Δt= { T-1, …, T-1, T } is obtained by the detection device ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N The start time of the tracked target is different and the duration of observation is different within the time period deltat, and the acquired data lengths are inconsistent.
Step 2: for the original time-varying situation information TU ΔT Performing coding complement compression processing to obtain effective deep learning model input data TU ΔT,P P represents the completion compression process.
In this embodiment, the complementary compression process refers to: performing 0-supplementing processing on target unit data with the length smaller than delta T, and simultaneously recording the original data length IL of the target unit r Then based on IL r And carrying out marker compression on the complement data, wherein the deep learning model only calculates the non-complement data. The label compression refers to labeling the original data length of the target unit, so that the model only processes the effective data, and compared with the length after 0 supplementation, the processing of the non-complement data is equivalent to compression processing. Because the acquired data are inconsistent in length, the acquired data can be processed in batch to complement and integrate the data into the same length in order to enhance the applicability of the model, but the complemented data needs to be marked in order to save the computing resource and time and avoid the result error caused by the complemented dataAnd (5) recording compression, so that the deep learning model only calculates non-complement data. The completion and compression process is performed in this embodiment to obtain the final model input [ TU ] ΔT,p ∈R N×T×D ,IL r ∈R N ]Wherein D refers to the number of features, and the data is complemented to obtain a data input variable-length time sequence processing model LSTM with consistent length, but in actual calculation, only non-complemented data is calculated.
Step 3: inputting data TU into effective deep learning model ΔT,P Inputting the information to a deep learning model for information characterization learning and intention classification;
step 4: and obtaining and outputting a target intention recognition result.
As the battlefield situation is a dynamic continuous evolution process, the battlefield situation is the result of dynamic games of both sides of the enemy, and situation information coming from various reconnaissance detection and sensing means is mostly fuzzy, the uncertainty is extremely strong, and great difficulty is brought to the situation information characterization. Therefore, the invention firstly needs to integrate all information to perform characterization learning research on situation information, and then performs classification learning on the combat intention of the target at the current moment based on the mined global structure and history tracking information.
The deep learning model W-CPCLSTM comprises: a learner for characterizing underlying shared information between information situational data acquired by various detection and sensing devices; the classifier is used for accurately identifying the fight intention of the detected target at the current moment according to the bottom shared information obtained by the learner at the current time period; and a controller for rationally distributing training attention between the learner and the classifier.
The main components of the model of learner CPC in this embodiment include:
a variable-length time sequence processing model LSTM for extracting situation information TU of N detection target units obtained in the current time period delta T ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N Coding, complementing and compressing to obtainWherein-> The method comprises the steps of completing a model input sequence of an ith target unit after compression, inputting the model input sequence into a variable-length time sequence processing model LSTM, and performing fusion coding on the model input sequence to obtain a potential representation coding sequence EU= { Eu of target unit situation information t-T ，…，Eu t-1 ，Eu t }，Eu t And encoding the potential characterization output at the time t.
GRU autoregressive model: efficient time-step information characterization for potential characterization of the target unit situation information coding sequence EUSummarizing, IL r Is the original information length; in this embodiment, based on characteristics of battlefield situation information, the acquired information is fully compressed, and because the fully compressed data is marked in the fully compressing process, the potential information characterization of the target unit situation information, which contains valid time steps, in the potential characterization coding sequence EU is characterized as ∈ ->The GRU autoregressive model is used for inducing all potential characterization generation context associated information containing global structures, and in this embodiment, in order to save computing resources and avoid computing errors, the GRU autoregressive model only characterizes the potential information of the effective time step as +.> And (5) processing and summarizing.
Fully connected prediction layer: for being based on autoregressive modelsThe obtained summary features represent bottom shared information SI between information situation data at the current moment t t The method comprises the steps of carrying out a first treatment on the surface of the In this embodiment, the associated task indicated by the prediction layer composed of the fully connected layers is not a potential representation of predicting the future time, but represents the shared information SI at the current time based on the situation information obtained at the current stage t 。
The three parts are combined and optimized through InfoNCE loss, and a loss function L of the learner CPC model is defined CPC ：
Wherein, (Tu) p,t ,SI t ) Can be regarded as a positive sample pair, tu p,t Representing situation information after completion of compression at time t,can be regarded as a negative sample pair, f (Tu p,t ,SI t ) Is density ratio E ]]Represents a function of solving for expectations, f (Tu p,t ,SI t )＝exp(Eu t SI t ). To optimize this loss function, we want the numerator to be as large as possible and the denominator to be as small as possible. That is, it is desirable that the mutual information between the positive sample pairs is larger and the mutual information between the negative sample pairs is smaller. Optimizing the loss, in effect maximizing Tu p,t And SI (information and information) t Mutual information between the two; f (Tu) p,t ,SI t ) Is the density ratio by using the true code value Eu t And shared information SI t For similarity, the following approximations are possible: f (Tu) p,t ,SI t )＝exp(Eu t SI t )。
In this embodiment, the learner CPC includes three parts, a variable length time series data processing model LSTM, an autoregressive model that generalizes all potential token-generating context correlation information that contains global structures, and a predictive layer that can indicate the correlation tasks of potential space.
In this embodiment, the model structure of the classifier in the deep learning model is: the variable length time series data processing model LSTM connected to a linear output layer, the loss function uses a basic cross entropy loss function:
wherein,target intention labels of N target units at the current time t under' God vision>To identify the intention of all targets detected, the final inference result is obtained based on the potential representation coding sequence EU learned by the learner.
In this embodiment, in order to further improve the recognition efficiency of the algorithm and reasonably distribute training attention, we consider different weight parameters [ α, β ] for the two parts of characterization learning and classification learning, so as to obtain a final model loss function: model loss function of controller:
L w ＝αL CPC +βL LSTM
alpha and beta respectively represent the weight parameters of the learner characterization learning and the weight parameters of the classifier classification learning.
According to the invention, through mining global structures by using learner CPC, the representation of potential shared information is learned, more global structures can be mined from limited battlefield information, low-level information and more local noise can be discarded, the time characteristics of target unit information are comprehensively considered, a variable-length time sequence processing model LSTM is designed for learning intention classification, and then the learning intention classification and the learning intention classification are effectively combined based on weight representing training attention, so that the online intention recognition effect under incomplete information can be realized.
FIG. 2 shows a deep learning model W-CPCLSTM model frame diagram given incomplete informationFirstly, a learner CPC consisting of a nonlinear encoder of LSTM, an autoregressive model of GRU and a fully-connected prediction layer is used for mining a global structure, then the LSTM model with variable length is used as a classifier for training intention recognition, and finally weight parameters [ alpha, beta ] are based]The two parts are effectively combined, and the smooth training of the model is realized through reasonable distribution of training attention.
The invention also provides an online combat intention recognition device based on incomplete information, which comprises the following modules:
the information data acquisition module: for pairing various detecting and sensing devices to an ith target unit at a current time tThe continuous tracking signal within the time period delta T is integrated and encoded to obtain the original time-varying situation information TU ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N }, wherein-> tu t The original time-varying situation information at the moment T is represented, N refers to the number of targets detected in a time period delta T, and T is the time length of delta T;
and the data processing complement module: for the original time-varying situation information TU ΔT Performing complement compression processing to obtain effective deep learning model input data TU ΔT,P P represents a complement compression process;
and a learning classification module: for inputting data TU into a deep learning model to be effective ΔT,P Inputting the information to a deep learning model for information characterization learning and intention classification;
the identification result output module: and the intent classification result obtained by the learning classification module is decoded and output.
The effectiveness of the present invention is verified by experiments as follows.
The simulation data set is from a deduction platform, the simulation data is divided into two conditions of incomplete information and perfect information, the former is missing and wrong information collected by reconnaissance equipment such as a radar station, an unmanned aerial vehicle and the like, namely historical information of enemy weapons and enemy units detected by the me, and the simulation is real enemy situation information identified in a countering environment. The latter is deduction data of action tracks, environments, time sequence events and the like of all combat units recorded by the system, namely all combat real information of the units and weapons on the side of the so-called "God".
The data set contains 15 intents from 45 classes of 12290 targets in total, the monitoring time of each target is 5-6000 seconds, and the intention data set under incomplete information records 27 characteristics of the targets at each moment, wherein 3 dimensions, 4 dimensions of position information, 2 dimensions of target types, 5 dimensions of action parameters, 7 dimensions of equipment correlation and 6 dimensions of detected events are related to target position information. While the perfect dataset recorded 65-dimensional features of these targets including 45 equipment states and 7 task related information, etc. It is emphasized that for each detected target, there is only one combat intent at the same moment, but its combat intent may change over time as time progresses and the situation evolves.
In the process of fight deduction, situation data is per deltat up Once updated from moment to moment, to simulate its online access state, we time-period partitioned the dataset based on the formula defined below:
wherein i is up Refers to the ith update of situation information.
For each time period Δt detected target data, 70% was classified into a training set and 30% was classified into a test set. After 5500 iterative training, the test is switched to. In order to accelerate training speed and increase stability, 50 samples form one batch, each 100 batches are trained, each 10 batches are trained, 10 batches are randomly selected from a training set (during training) or a testing set (during testing) for verification, and the average value is used as the final recognition precision. In addition, the model considers the five weight parameters [ α, β ] defined by table 1 to learn the impact of different training attention patterns on the intent recognition effect.
Table 1 different weight parameters [ alpha, beta ] for characterizing learning and classifier learning, respectively
| Weight parameter | α | β |
| AB1 | 0.1 | 0.9 |
| AB2 | 0.9 | 0.1 |
| AB3 | 0.3 | 0.7 |
| AB4 | 0.7 | 0.3 |
| AB5 | 0.5 | 0.5 |
Carrying out intention recognition on targets aiming at information with three maximum detection degrees in the deduction process: 1) Detecting unknown entity-target type is unknown, and the threat degree cannot be judged; 2) The target position is blurred and cannot be locked; 3) The target is not clear, and the enemy can not be judged.
Example one, target position is unknown
In the scenario of multi-party combat, a situation that a detected new target cannot be determined often occurs, and the detected target is clearly hostile, which brings a certain challenge to the task of intention recognition. Thus, the present embodiment explores this problem to verify the applicability of the model of the present invention to this problem. First, in order to assign appropriate training attention, intent recognition was performed based on model weight parameters defined in table 1, and the obtained result found AB3 to be the best training attention assignment among these five configurations. Secondly, based on incomplete information with an ambiguous target position, the online intention recognition effect of the model W-CPCLSTM provided by the invention is verified through comparison with a traditional LSTM model, as shown in fig. 3 (a) and 3 (b), and compared with the traditional LSTM, the model provided by the invention can greatly improve the accuracy of intention recognition with fewer iteration times. It is observed from fig. 3 (a) that feature characterization of CPC can greatly improve the efficiency of on-line intent recognition of enemy targets in the face of incomplete intelligence without any information about the target standpoint. This observation can be confirmed when we compare the recognition accuracy and training trend of LSTM and W-cpclsm, respectively. In particular, it is apparent from fig. 4 (a) that the model W-cpclsm of the present invention can reach more than 90% compared to the accuracy of less than 80% of the conventional LSTM, and fig. 4 (b) tells us that the proposed model is more stable and reliable during training. This is a valid proof that the application of the learning structure by feature characterization is advantageous in the task of intended recognition. Finally, by supplementing the detected position information to verify the robustness of the two models to the disturbance, the result shown in fig. 4 (a) is obtained, and it can be observed that the model of the present invention stably improves the recognition accuracy at a lower cost when the target position is clear. By adding the detected information about the target position, the following two conclusions can be drawn: 1) From fig. 4 (a), it can be seen that W-cpclsm still maintains significant advantages over opponent LSTM in recognition accuracy and training speed. 2) Compared with fig. 3, it can be observed that after the target position information is determined, both LSTM and W-cpcstm have a certain improvement in recognition accuracy, and in addition, recognition stability of both is improved, however, if the target position information is detected or not, compared with LSTM, the recognition effect on the W-cpcstm model is not so obviously affected, which indicates that the model of the invention is less sensitive when the information disturbance facing the target position is performed, that is, feature characterization learning helps the model to mine more effective features or global structures when the information facing the limited combat situation is performed.
Embodiment two: unknown target position
On the battlefield of countermeasure, the characteristics of concealment, deception and limitation of detection equipment, time ductility and the like of each party cause difficulty in tracking the exact position of the target in real time, and many tactical intentions are closely related to the coordinate positions of the targets. In the event of this loss of important information, how does training attention rationally distribute so that the model maintains a steady output? Whether the model of the present invention can continue to maintain the advantages in the combat intent recognition task in the face of situation information where the target position is not well-detected? What is the intention model in turn given feedback if the location information from the probe device is obtained?
The present invention employs AB3, the best training attention allocation among these five configurations. As can be seen from fig. 5 (a) and 5 (b), the model W-cpclsm proposed by the present invention has a significant advantage in classification accuracy as compared with the conventional LSTM. Fig. 6 (a) and 6 (b) show that the proposed model stably improves accuracy of intention recognition at lower cost than the conventional LSTM when the target position is clear. In addition, whether the coordinate features of the targets are absent in fig. 5 or the related information of the positions of the targets is detected in fig. 6, the model of the invention is superior to the LSTM in training stability, identification accuracy and convergence speed, and the effective combination of characterization learning and time sequence classification learning is proved to be capable of aiming at improving the on-line intention recognition efficiency. Second, compared to fig. 5, it can be observed from fig. 6 that after supplementing the relevant information of the target location, both LSTM and W-cpcstm have improved training speed and recognition accuracy to some extent, even if the latter is less obvious. Finally, by combining fig. 3 and fig. 5, it can be seen that, whether the standing information is absent or the position information is lost, compared with the LSTM, the model of the present invention can perform the intention recognition more robustly, which illustrates that the characteristic learning structure can effectively overcome the fluctuation caused by the change of the information in the countermeasure process.
Example III, type of unknown object
In the fight process, a suspected track is often associated with the detected enemy target initially, the specific target type of the suspected track needs to be continuously tracked for a period of time to be confirmed, and how to efficiently identify the detected unknown target in an online manner is important content to be explored in modern informatization war. In this example, the proposed algorithm is validated against conventional LSTM models for its intended recognition effect in the face of situational information of unknown target type. And finally, exploring the contribution degree of the target type information to the intention recognition of the two models.
In the case of the incomplete situation information of the type of the unknown object, the present example adopts AB5 with the best recognition efficiency as its best training parameter configuration to compare with the LSTM network with the best effect as well, and the experimental results of fig. 7 (a) and 7 (b) show that, obviously, the W-cpcstm has a significant advantage in the training stability and speed, compared with the recognition accuracy of less than 80% of the opponent, with the advantage of about 10%. Likewise, even with the addition of the target type of probe information, the model of the present invention maintains its consistent advantages over LSTM in terms of the recognition accuracy given in FIG. 8 (a) and the loss value given in FIG. 8 (b). Moreover, by comparison with fig. 8, it can be found that the model of the present invention is not as dramatic as LSTM performance in terms of both recognition accuracy and training stability when no target information is detected (fig. 7). These again demonstrate that combining the characterization learning ability of CPC with the temporal information mining ability of LSTM based on training attention allocation can be a powerful tool for incomplete information online intent prediction.
Fourth embodiment, perfect situation
It can be seen from the above experiments that the model of the present invention has obvious advantages in on-line intention recognition task compared with the conventional LSTM in the face of incomplete situation of various detection degrees, but is in the face of perfect situation without any error, loss and more target information related to task actions when giving "god's vision", that is, real situation information in ideal state? The invention selects the optimal parameter in the AB3 configuration by comprehensively considering the recognition precision and the training stability. As is apparent from the comparison of LSTM and W-cpcstm shown in fig. 9 (a) and 9 (b), although the input information is sufficiently perfect, our model can still learn and mine more effective information through the characterization of situation characteristics, so that the model can have certain advantages in the situation cognition field, whether the accuracy of the intended identification, the convergence speed of training or the stability of output. Of course, compared with the recognition result of incomplete information, it can be seen that the recognition efficiency of the effective input information can be greatly improved no matter the information is LSTM or W-CPCLSTM.
Fifth embodiment, different target tracking times
In theory, as the tracking time of the target is prolonged, the more information is obtained, the clearer the sketch image' of the target is, the intention is also clear, and in order to verify the influence degree of the factor on the model of the invention under the complex countermeasure environment, in this part, the embodiment selects information and situation data with different lengths to identify the online intention of the combat target. In addition, in order to more clearly evaluate the practicability and effectiveness of the proposed W-CPCLSTM in the combat intention recognition task, the accuracy, standard deviation, effective speed and the like of the proposed W-CPCLSTM and a traditional time sequence processing model LSTM are evaluated on a training set and a test set, and the obtained results are shown in Table 2.
TABLE 2LSTM and W-CPCLSTM intent recognition results on training and test sets in terms of accuracy, standard deviation and speed based on perfect situation and incomplete intelligence information
1 Tr is a training set;
2 ts is a test set;
3 acc, intention recognition accuracy, the highest average accuracy obtained during the whole exercise process;
4 std is standard deviation, standard deviation of highest average precision;
5 spe, training speed, characterized by the number of iterations required to achieve 80% (90% in perfect situation) accuracy throughout the training process;
6 l: information data length;
7 * The required precision is not achieved;
it can be seen from table 2 that given the target information, whether there is a target type or lacks its type-related information, unexpectedly, for the LSTM model, as the time sequence length goes from 10 to 20 to 30, the recognition accuracy is not improved significantly on both the training set and the test set, and even a tendency of slipping down occurs. The same conclusion is also verified on the W-CPCLSTM model of the invention, and the change is less obvious, which shows that with the increase of the duration, the obtained false information and false information are increased along with more effective information, and the recognition result is stopped. In addition, it can be seen that the recognition accuracy of the model of the invention on the training set and the test set can exceed 7% -11% of LSTM, and the training speed for achieving effective accuracy can be 5-32 times faster than that of the other side, no matter the position information, the position information or the target type. From these conclusions, it can be seen that the model of the invention can be more robust and can rapidly realize higher-precision online intention recognition in the face of incomplete information, and is expected to become a powerful tool in the situation cognition field.
The invention provides a W-CPCLSTM model, which is composed of the following three parts: aiming at CPC models of characteristics such as incompleteness and deception of information data, a variable length LSTM model considering sequence characteristics of situation characteristics, and training attention weights considering recognition stability. The CPC model is used for mining a global structure from limited information data through feature characterization learning, the variable-length LSTM model is used for carrying out intention classification training through a sequence mechanism of feature characterization analysis, and training attention weights are used for carrying out stable intention recognition by effectively combining the first two through reasonable distribution of training attention.
In order to verify the intention recognition effect of the proposed model facing incomplete information in the countermeasure process, performance analysis and application evaluation are carried out on the model by comparing the model with LSTM based on three common conditions of unknown target position, position and type detection; meanwhile, in order to comprehensively test the practicability and effectiveness of the algorithm, a perfect situation under the view angle of a emperor is given, and on-line intention recognition experimental analysis is performed based on the ideal information; finally, the influence effect of incomplete information on the model is discussed by intercepting detection information with different lengths. All experiments show that the perfect model can improve the recognition accuracy by 7 to 11 percent, the recognition speed can be improved by 6 to 32 times, and the stability is also superior, which proves that the model is expected to become effective assistance in the task of intention recognition.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.
Claims (4)
1. An online combat intention recognition method based on incomplete information is characterized in that,
the method comprises the following steps:
step 1: the various detecting and sensing devices are directed to the ith target unit at the current moment tThe continuous tracking signal within the time period delta T is integrated and encoded to obtain the original time-varying situation information TU ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N }, whereintu t The original time-varying situation information at the moment T is represented, N refers to the number of targets detected in a time period delta T, and T is the time length of delta T;
step 2: for the original time-varying situation information TU ΔT Performing coding complement compression processing to obtain effective deep learning model input data TU ΔT,P P represents a complement compression process;
step 3: inputting data TU into effective deep learning model ΔT,P Inputting the information to a deep learning model for information characterization learning and intention classification;
step 4: decoding and outputting to obtain a target intention recognition result;
the deep learning model includes:
a learner for characterizing underlying shared information between information situational data acquired by various detection and sensing devices;
the classifier is used for accurately identifying the fight intention of the detected target at the current moment according to the bottom shared information obtained by the learner at the current time period;
and a controller for rationally distributing training attention between the learner and the classifier;
the main components of the learner include:
a variable-length time sequence processing model LSTM for extracting original time-varying situation information TU obtained by N detection target units in the current time period delta T ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N Coding, complementing and compressing to obtainWherein-> The model input sequence under the incomplete information after the completion and compression of the ith target unit is input into a variable-length time sequence processing model LSTM, fusion coding is carried out on the model input sequence, and a potential representation coding sequence EU= { Eu of the situation information of the target unit is obtained t-T ，…，Eu t-1 ，Eu t }，Eu t Potential representation codes output at the time t;
autoregressive model of GRU: efficient time-step information characterization for potential characterization of the target unit situation information coding sequence EUSummarizing, IL r Original information length before complementation;
fully connected prediction layer: bottom shared information SI between information situation data for representing current moment t based on summarized features obtained by autoregressive model t ；
The three parts are combined and optimized through InfoNCE loss, and a loss function L of the learner CPC model is defined CPC ：
Wherein, (Tu) p,t ,SI t ) Tu is the positive sample pair p,t Representing situation information after completion of compression at time t,is a negative sample pair, f (Tu p,t ,SI t ) Is the density ratio, f (Tu p,t ,SI t )＝exp(Eu t SI t )；
The complement compression process refers to: performing 0-supplementing processing on target unit data with the length smaller than delta T, and simultaneously recording the original data length IL of the target unit r Then based on IL r And carrying out marker compression on the complement data, wherein the deep learning model only calculates the non-complement data.
2. The on-line combat intent recognition method of claim 1, wherein the model structure of said classifier is: the variable length time series data processing model LSTM connected to a linear output layer, the loss function uses a basic cross entropy loss function:
wherein,is the target intention label of the current time t of N target units under ideal conditions,to identify the intention of all targets detected, the final inference result is obtained based on the potential representation coding sequence EU learned by the learner.
3. The on-line combat intent recognition method of claim 2, wherein a model loss function of said controller:
L w ＝αL CPC +βL LSTM
alpha and beta respectively represent the weight parameters of the learner characterization learning and the weight parameters of the classifier classification learning.
4. An online combat intention recognition device based on incomplete information is characterized by comprising the following modules:
the information data acquisition module: for pairing various detecting and sensing devices to an ith target unit at a current time tThe continuous tracking signal within the time period delta T is integrated and encoded to obtain the original time-varying situation information TU ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N }, wherein-> tu t The original time-varying situation information at the moment T is represented, N refers to the number of targets detected in a time period delta T, and T is the time length of delta T;
and the data processing complement module: for the original time-varying situation information TU ΔT Performing complement compression processing to obtain effective deep learning model input data TU ΔT,P P represents a complement compression process; the complement compression process refers to: performing 0-supplementing processing on target unit data with the length smaller than delta T, and simultaneously recording the original data length IL of the target unit r Then based on IL r The complement data is subjected to marking compression, and the deep learning model only calculates the non-complement data;
the deep learning model includes:
a learner for characterizing underlying shared information between information situational data acquired by various detection and sensing devices;
the classifier is used for accurately identifying the fight intention of the detected target at the current moment according to the bottom shared information obtained by the learner at the current time period;
and a controller for rationally distributing training attention between the learner and the classifier;
the main components of the learner include:
a variable-length time sequence processing model LSTM for extracting original time-varying situation information TU obtained by N detection target units in the current time period delta T ΔT ＝{Tu 1 ,Tu 2 ,…,Tu N Coding, complementing and compressing to obtainWherein-> The model input sequence under the incomplete information after the completion and compression of the ith target unit is input into a variable-length time sequence processing model LSTM, fusion coding is carried out on the model input sequence, and a potential representation coding sequence EU= { Eu of the situation information of the target unit is obtained t-T ，…，Eu t-1 ，Eu t }，Eu t Potential representation codes output at the time t;
autoregressive model of GRU: efficient time-step information characterization for potential characterization of the target unit situation information coding sequence EUSummarizing, IL f Original information length before complementation;
fully connected prediction layer: bottom shared information SI between information situation data for representing current moment t based on summarized features obtained by autoregressive model t ；
The three parts are combined and optimized through InfoNCE loss, and a loss function L of the learner CPC model is defined CPC ：
Wherein, (Tu) p,t ,SI t ) Tu is the positive sample pair p,t Representing situation information after completion of compression at time t,is a negative sample pair, f (Tu p,t ,SI t ) Is the density ratio, f (Tu p,t ,SI t )＝exp(Eu t SI t )；
And a learning classification module: for inputting data TU into a deep learning model to be effective ΔT,P Inputting the information to a deep learning model for information characterization learning and intention classification;
the identification result output module: and the intent classification result obtained by the learning classification module is decoded and output.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111041309.1A CN113743509B (en) | 2021-09-07 | 2021-09-07 | Online combat intent recognition method and device based on incomplete information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111041309.1A CN113743509B (en) | 2021-09-07 | 2021-09-07 | Online combat intent recognition method and device based on incomplete information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113743509A CN113743509A (en) | 2021-12-03 |
| CN113743509B true CN113743509B (en) | 2024-02-06 |
Family
ID=78736269
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111041309.1A Active CN113743509B (en) | 2021-09-07 | 2021-09-07 | Online combat intent recognition method and device based on incomplete information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113743509B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114139550B (en) * | 2022-02-08 | 2022-05-06 | 中国电子科技集团公司第五十四研究所 | Situation intelligent cognition method based on activity semantic text message |
| CN114818853B (en) * | 2022-03-10 | 2024-04-12 | 中国人民解放军空军工程大学 | Intention recognition method based on bidirectional gating circulating unit and conditional random field |
| CN115481702B (en) | 2022-10-28 | 2023-02-17 | 中国人民解放军国防科技大学 | Predictive comparison characterization method for multi-element time series data processing |
| CN116227952B (en) * | 2023-05-09 | 2023-07-25 | 中国人民解放军海军潜艇学院 | Method and device for selecting attack target defense strategy under key information deficiency |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329348A (en) * | 2020-11-06 | 2021-02-05 | 东北大学 | Intelligent decision-making method for military countermeasure game under incomplete information condition |
| CN112749761A (en) * | 2021-01-22 | 2021-05-04 | 上海机电工程研究所 | Enemy combat intention identification method and system based on attention mechanism and recurrent neural network |
-
2021
- 2021-09-07 CN CN202111041309.1A patent/CN113743509B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329348A (en) * | 2020-11-06 | 2021-02-05 | 东北大学 | Intelligent decision-making method for military countermeasure game under incomplete information condition |
| CN112749761A (en) * | 2021-01-22 | 2021-05-04 | 上海机电工程研究所 | Enemy combat intention identification method and system based on attention mechanism and recurrent neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113743509A (en) | 2021-12-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113743509B (en) | Online combat intent recognition method and device based on incomplete information | |
| Umer et al. | Machine learning for intrusion detection in industrial control systems: Applications, challenges, and recommendations | |
| Cai et al. | Trajectory‐based anomalous behaviour detection for intelligent traffic surveillance | |
| CN111310915A (en) | Data anomaly detection and defense method for reinforcement learning | |
| Chen et al. | Online intention recognition with incomplete information based on a weighted contrastive predictive coding model in wargame | |
| CN110768971B (en) | Confrontation sample rapid early warning method and system suitable for artificial intelligence system | |
| CN111178504B (en) | Information processing method and system of robust compression model based on deep neural network | |
| CN113204745A (en) | Deep learning back door defense method based on model pruning and reverse engineering | |
| CN111753300A (en) | Reinforcement learning-oriented abnormal data detection and defense method and device | |
| Liu et al. | Zero-bias deep-learning-enabled quickest abnormal event detection in IoT | |
| Shahid et al. | Label flipping data poisoning attack against wearable human activity recognition system | |
| CN116432514A (en) | Interception intention recognition strategy simulation system and method for unmanned aerial vehicle attack and defense game | |
| Zhang et al. | Many-objective optimization based intrusion detection for in-vehicle network security | |
| Yin et al. | A bagging strategy-based kernel extreme learning machine for complex network intrusion detection | |
| Rabie et al. | A security model for smart grid SCADA systems using stochastic neural network | |
| Şeker | Use of Artificial Intelligence Techniques/Applications in Cyber Defense | |
| Cao et al. | Discovering intrinsic spatial-temporal logic rules to explain human actions | |
| CN111882037A (en) | Deep learning model optimization method based on network addition/modification | |
| CN117036405A (en) | Anti-occlusion target tracking method integrating multi-granularity dynamic appearance | |
| CN111291962A (en) | Method for preventing and attacking AI crime and AI data infringement | |
| Narengbam et al. | Harris hawk optimization trained artificial neural network for anomaly based intrusion detection system | |
| Mulo et al. | Towards an adversarial machine learning framework in cyber-physical systems | |
| Zhu et al. | FocusedCleaner: Sanitizing Poisoned Graphs for Robust GNN-based Node Classification | |
| CN111343032A (en) | Industrial control network abnormal session detection method, device, electronic equipment and storage medium | |
| Glinton et al. | A markov random field model of context for high-level information fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |