Research On Multivariate Time-series Classification Problems With Incomplete Information Under Strong Confrontation Conditions

Decision-making tasks with strong confrontation attributes,such as joint warfare,anti-terrorist assault,network attack and defense,owing to their outstanding characteristics of dynamic and variable opponent strategies,scarce and intermittent data information,mixed intelligence resources of truth and fiction,etc.,make it difficult for decision-makers to fully grasp the real global situation.Especially with the development of intelligent technology,strong confrontation tends to be unmanned and clustered,and the trend of situation evolution is becoming more rapid and variable,how to analyse the real scenario under the fog of the situation according to fragmented and uncertain data has become a key issue to improve the quality of decision-making.Situational information is a kind of serial data that evolves over time,and situation analysis is often modelled as a classification mining problem for multivariate time-series data.However,most of the traditional multivariate time-series classification algorithms are based on the premise of “complete information”:on the one hand,large-scale sample data with high quality labels are required to support model training,the more labeled samples,the more accurate the model classification;on the other hand,complete features from the ideal environment are required,and the sample features need to be complete and accurate.However,the high intensity of the confrontation and the opaque nature of the confrontation not only make the annotation of labels extremely difficult,but also pose a great challenge to the integrity,authenticity and certainty of the entity unit features.To address these problems,the paper develops a research on multivariate time-series classification problem with incomplete information under strong confrontation conditions,and the main work and innovations include:(1)Built a framework for multivariate time-series classification with incomplete information under strong confrontation conditionsThe existing multivariate time-series classification methods suffer from the conflict between general skills and specialized domain data in the application scenarios of strong adversarial environments,and the low sample efficiency makes it difficult to apply standard machine learning methods to many realistic adversarial scenarios.In this paper,we propose active learning combined with reinforcement learning and representation learning combined with time-series classification for the task of time-series classification of poor labeling and incomplete information under strong confrontation conditions,respectively.At the same time,a framework for intelligent identification of key elements of the situation with the combination of offline training analysis and online deduction analysis has been constructed for a general strong adversarial application structure.(2)Proposed a human-machine agent-based multivariate time-series classification method for poor labelingFor the multivariate time-series classification task of poor labeling of incomplete information under strong confrontation conditions,to minimize the cost of label annotation,this paper combines active learning and reinforcement learning to propose a humanmachine agent MTCARq＿H-M that can intelligently infer when to introduce manual label annotation and how to perform the classification task autonomously,which achieves selective annotation of sample labels by defining active learning time-series classification task settings under poor labeling,modeling its query strategy based on reinforcement learning sequential decision making,and proposing a DQQN policy update method that introduces a double-Q cost function to improve the sampling efficiency of reinforcement learning.Meanwhile,to analyse the respective roles of guided learning of manual labels and data learning of machines in the time-series classification task,the human intervener MTCARq＿H,based on manual annotation,is proposed by changing the action settings of the model.Finally,based on a public dataset and replay data from a typical strong adversarial wargame platform,the proposed models MTCARq＿H-M and MTCARq＿H are evaluated for generic performance and analysed for strong adversarial applications by the ablation experiments with MTCAR＿H-M and MTCAR＿H and the comparison experiments with the supervised model Supervised,the traditional active learning model QBC and the uncertainty sampling model in terms of the four aspects of time-series memory capability,classification performance,uncertainty analysis and reward value factors,respectively.The experiments demonstrate that,both for generic object classification and type recognition of adversarial targets,the agent can achieve high classification accuracy with a low label demand rate,substantially reducing the manual annotation cost.(3)Proposed a multivariate time-series classification method based on weighted contrast coding with incomplete featuresFacing the massive incomplete,inaccurate,even wrong or deceptive feature information,this paper combines representation learning and time-series classification to propose a weighted contrast coding-based time series classifier W-CPCLSTM.By constructing a contrast coding-based representator to deeply mine the global features and shallow shared information of the adversarial situation and discarded the low-level noise in intelligence information;for the characteristic of varying time sequence length of the situation information,a variable-length LSTM classifier model is constructed to develop time-series classification learning;a weighted allocator is formed with different weight parameters to reasonably allocate training attention to the representator and classifier.Finally,based on the large-scale simulation replay dataset from the wargame platform,performance evaluation and application analysis of W-CPCLSTM are conducted by comparing it with the classical time series models LSTM and FCN,as well as the latest work Oct Conv and Oct FCN networks in three aspects,namely ideal environment,noisy environment and information absence,respectively.The experiments show that the proposed model has different degrees of advantages in terms of classification accuracy and output stability in most cases,regardless of different types of error information or different degrees of missing data.(4)Conducted offline training + online deduction case analysis on the intelligent identification of critical elements of the situationBased on the platform of the National Wargame Competition,which simulates the process of battlefield confrontation,this paper develops a case study on the intelligent identification of critical elements of situation for cluster operations.Firstly,a public dataset is developed to study the situational awareness for strong confrontation environment.Secondly,by constructing a thermal grid to characterize the spatial characteristics of the target and an empirical rule for the temporal characteristics of historical information,and introducing empirical parameters with expert experience and domain knowledge,a spatio-temporal rule body based on the thermal grid is proposed for the target formation judgment task.Then,the proposed spatio-temporal rule body,human-machine agent and time series classifier are used to carry out offline training analysis for target formation judgement,unknown target classification of formation and formation combat intent recognition respectively,and the experiments prove that the proposed models can achieve better results than the traditional time-series classification models for incomplete information under strong confrontation conditions.Finally,the three models were embedded into the wargame platform to conduct an online deduction analysis on the identification of critical elements of the situation.The simulation scenario from the red side’s perspective confirmed that the blue side’s AI with the situational awareness models is able to discriminate the enemy’s situational elements in a timely and effective manner,which provides a good cognitive basis for the commander’s command decision,proving that the proposed models have good practical value in multivariate classification tasks with incomplete information under strong confrontation conditions.