CN115640736A

CN115640736A - Key event detection-based method and device for mining key tactics in war game deduction process

Info

Publication number: CN115640736A
Application number: CN202211067595.3A
Authority: CN
Inventors: 徐志伟; 夏少杰; 包骐豪; 朱善胜; 瞿崇晓
Original assignee: CETC 52 Research Institute
Current assignee: CETC 52 Research Institute
Priority date: 2022-09-01
Filing date: 2022-09-01
Publication date: 2023-01-24

Abstract

The invention discloses a method and a device for excavating key tactics in a war game deduction process based on key event detection, wherein the method comprises the following steps: performing multi-resolution potential feature abstraction on the acquired war game deduction data to obtain a battlefield potential feature set of a unit of battle under unified dimensionality; extracting known key events in the war game deduction data by adopting a weak supervision self-training method based on the battlefield situation characteristic set, and obtaining a specific key event set; based on a battlefield situation characteristic set and a specific key event set, extracting novel key events in the war game deduction data by using a novel key event identification method based on a deep hidden layer characteristic, and marking the novel key events to obtain a novel key event set; and combining the key events in the specific key event set and the novel key event set according to the time sequence relationship to form a multi-element tactical set, and finishing the excavation of the key tactical set. The invention improves the timeliness and the accuracy of key tactical excavation.

Description

Key event detection-based method and device for mining key tactics in war game deduction process

Technical Field

The invention belongs to the technical field of large-scale battle-level military chess deduction and intelligent game, and particularly relates to a method and a device for excavating a key tactic in a military chess deduction process based on key event detection.

Background

The war deduction is carried out by utilizing war simulation, and the war deduction method is a scientific and effective method for designing the quality of fighting in the future and forging combat battles, and plays a great role in the fields of large-scale combined combat practice training in the military, daily command training of commanders, combat theory, war research and the like.

With the continuous acceleration of the rhythm of modern war and the continuous rise of complexity, the human brain decision is difficult to adapt to the rapid and alternating trend of battlefield situation, the rapid, automatic and autonomous decision is urgently needed in the future war, and the intelligent technology is urgently needed to extend the human brain so as to improve the capability of command information system, thereby adapting to the high-speed, complex and variable battlefield environment. In recent years, with the rapid development of artificial intelligence technology and the introduction of advanced technologies such as deep learning and reinforcement learning, the automation and intelligence of the deduction of battle are improved to a certain extent, and the artificial intelligence decision-making intelligent agent can make a partial operation in which Cheng Zhongchong is deduced, so that the artificial intelligence decision-making intelligent agent can provide reference for the command and control decision of a commander.

Because the simulation deduction scale of the weapons and chess is often large and the simulation time is long, the quantity of the involved combat entities and weapon platforms is huge and the types are various. Under the circumstances, it is very difficult for a commander to locate a key action from a long-time high-frequency decision of an artificial intelligence decision-making intelligent agent, namely, the consumed time is long, and the problem of inaccurate key tactics for location exists, so that an innovative tactics mining method is urgently needed, the key battlefield situation and the key action made by the commander in front of and behind the node are automatically and intelligently extracted from simulation engagement data after the deduction is finished, the understanding of game fighting decisions made by the commander and the fighter in an artificial intelligence algorithm is deepened, the command training efficiency of the commander by utilizing military chess deduction is improved, and the own command art is promoted by mining a new tactics from the game fighting data.

Disclosure of Invention

One of the purposes of the invention is to provide a key tactics mining method for a war game deduction process based on key event detection, which is used for meeting the abstract requirement of large-scale intelligent game strategies and solving the problems of poor interpretability and non-intuitive decision process of an intelligent decision scheme generated by utilizing reinforcement learning to play against in a typical battle scene.

In order to realize the purpose, the technical scheme adopted by the invention is as follows:

a method for excavating key tactics in a military chess deduction process based on key event detection comprises the following steps:

performing multi-resolution potential feature abstraction on the acquired war game deduction data to obtain a battlefield potential feature set of a unit of battle under unified dimensionality;

extracting known key events in the war playing data by adopting a weak supervision self-training method based on the battlefield situation characteristic set, and obtaining a specific key event set;

based on a battlefield situation characteristic set and a specific key event set, extracting novel key events in the war game deduction data by using a novel key event identification method based on a deep hidden layer characteristic, and marking the novel key events to obtain a novel key event set;

and combining the key events in the specific key event set and the novel key event set according to the time sequence relationship to form a multi-element tactical set, and finishing the excavation of the key tactical set.

Several alternatives are provided below, but not as an additional limitation to the above general solution, but merely as a further addition or preference, each alternative may be combined individually for the above general solution or between several alternatives without technical or logical contradictions.

Preferably, the performing multi-resolution potential feature abstraction on the acquired war game deduction data comprises:

processing original battlefield cooperation information in the war game deduction data through a dummy coding method, an interval scaling method and a discretization method to obtain cooperation characteristics;

increasing the dimension of the static features of the operational units in the cooperative features by using a dimension increasing method, and reducing the dimension of the dynamic features of the operational units in the cooperative features by using a dimension reducing method;

and processing the static characteristics and the dynamic characteristics based on a characteristic selection method to obtain a final battlefield potential characteristic set, wherein the battlefield potential characteristic set comprises a potential-instruction characteristic sequence which is arranged according to a decision moment sequence.

Preferably, the processing the original battlefield cooperation information in the war game deduction data through a dummy coding, interval scaling and discretization method to obtain the cooperation characteristics comprises the following steps:

carrying out characteristic dummy coding on the original battlefield cooperative information in the war game deduction data to obtain the cooperative characteristics of the combat unit when carrying out the combat mission;

processing the cooperative features by adopting a dimensionless method of interval scaling, and converting differentiated format data of different combat units into uniform format data;

and carrying out discretization processing on the collaborative characteristics after the interval scaling to obtain the final collaborative characteristics.

Preferably, the performing dimension enhancement on the static features of the combat units in the collaborative features by using a dimension enhancement method and performing dimension reduction on the dynamic features of the combat units in the collaborative features by using a dimension reduction method includes:

adopting a polynomial expansion and kernel function mode to increase the dimension of the static characteristics of the combat unit;

and reducing the dimension of the dynamic data of the combat unit by adopting principal component analysis and linear discriminant analysis.

Preferably, the method for extracting known key events in the war game deduction data by adopting a weak supervision self-training method based on the battlefield situation characteristic set and obtaining a specific key event set comprises the following steps:

clustering data in the battlefield situation characteristic set to obtain a clustering result;

according to the data space distribution represented by the clustering result, training a semi-supervised support vector machine classifier as a classification recognition model by combining a combat key event data set with a label, and labeling data in a battlefield situation characteristic set by using the classification recognition model;

and selecting data with labels conforming to the prior trigger marks from the marked battlefield potential state feature set according to the prior trigger marks to form a specific key event set.

Preferably, the method for extracting the novel key events in the military chess deduction data by using the novel key event identification method based on the deep hidden layer features based on the battlefield situation feature set and the specific key event set comprises the following steps:

training an LSTM network by taking a specific key event set, taking a feature extraction layer in the trained LSTM network, and combining the feature extraction layer with a multi-head attention network to obtain a hidden layer feature extraction model;

constructing a linear classification model according to the definition of a novel key event;

taking the battlefield situation characteristics in the battlefield situation characteristic set to input the hidden layer characteristic extraction model to obtain the hidden layer characteristics of the battlefield situation characteristics;

and inputting the extracted hidden layer features and the hidden layer features of the known key events into the linear classification model together, and determining a novel key event in the battlefield situation state feature set according to the output of the linear classification model.

Preferably, the constructing a linear classification model according to the definition of the novel key event includes:

if the similarity between the hidden layer feature to be judged and the hidden layer feature of the known key event meets a first threshold value, the action decision corresponding to the hidden layer feature to be judged is different from the action decision corresponding to the hidden layer feature of the known key event, and the reward return obtained by the hidden layer feature to be judged is higher than a second threshold value, the hidden layer feature to be judged is a novel key event;

or, if the degree of difference between the hidden layer feature to be determined and the hidden layer feature of the known key event satisfies the third threshold, the action decision corresponding to the hidden layer feature to be determined is the same as the action decision corresponding to the hidden layer feature of the known key event, and the reward return obtained by the hidden layer feature to be determined is higher than the fourth threshold, the hidden layer feature to be determined is a novel key event.

Preferably, the labeling of the novel key event to obtain a novel key event set includes:

limiting the elements of battlefield events in the process of war game deduction;

according to the battlefield element weight contained in a novel key event output by the multi-head attention network, determining a core element according to the battlefield element weight, and determining a discrete value for the core element to obtain a plurality of situation description keywords consisting of the core element and the discrete value;

and obtaining the labeling information of the novel key event by adopting a keyword template serial connection method based on a plurality of situation description keywords, and taking the labeled novel key event as a novel key event set.

The key tactics mining method based on the key event detection in the war game deduction process provided by the invention can automatically and intelligently extract key battlefield situations and key actions made by a commander in front of and behind the node from simulated engagement data after deduction is finished, thereby deepening the understanding of game confrontation decisions made by commanders on an artificial intelligent algorithm, improving the efficiency of commanding and training the commander by utilizing the war game deduction and mining new tactics to promote own command art.

The invention also aims to provide a key tactics mining device for the war game deduction process based on key event detection, which is used for meeting the abstract requirement of large-scale intelligent game strategies and solving the problems of poor interpretability and non-intuitive decision process of an intelligent decision scheme generated by carrying out game countermeasures by reinforcement learning in a typical battle scene.

a key event detection-based weapon and chess deduction process key tactic mining device comprises a processor and a memory, wherein the memory is used for storing a plurality of computer instructions, and the computer instructions are executed by the processor to realize the key event detection-based weapon and chess deduction process key tactic mining method.

Drawings

FIG. 1 is a flow chart of a key tactical mining method in a war game deduction process based on key event detection according to the present invention;

FIG. 2 is an exemplary diagram of linear discriminant analysis dimension reduction applied in the present invention;

FIG. 3 is a flowchart of key event extraction based on weak supervised self-training in the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

In order to better assist the commander to analyze and understand the large-scale wargame deduction data and improve the daily training efficiency and the wartime decision level of the commander, the embodiment provides a critical tactics mining method for the wargame deduction process based on key event detection, so as to provide a research and solution idea for understanding the difficulty of the behavior of the intelligent body under the large-scale wargame deduction and improve the mining accuracy of the critical tactics.

For convenience of understanding, the present embodiment is described by taking game deduction of military chess as an example, and as shown in fig. 1, the implementation steps of the method for mining the key tactics in the military chess deduction process based on key event detection are as follows:

step 1, performing multi-resolution potential feature abstraction on the acquired war game deduction data to obtain a battlefield potential feature set of a combat unit under unified dimensionality.

The large-scale military chess game deduction data has various problems of large information quantity, high redundancy, non-uniform data formats and the like, and the cooperative characteristics of tactics are processed by methods such as dummy coding, interval zooming, discretization and the like; applying a polynomial expansion and kernel function dimension increasing method and PCA and LDA dimension reducing methods to encode and process multi-dimensional static and dynamic characteristics such as longitude and latitude height, orientation, speed and the like of combat unit names, types and performance parameters; the multi-resolution situation characteristic engineering is carried out based on characteristic selection methods such as Filter, wrapper, embedded and the like, the influence caused by special mechanisms, simulation mechanisms and effect delay in a battle-level confrontation scene is solved, and effective abstraction of important situation information is realized.

1) And processing the cooperative characteristics of tactics.

Aiming at a large-scale battle deduction scene, firstly, the multi-weapon collaborative characteristics need to be abstracted:

(1) in order to convert qualitative multi-weapon-kind cooperative data into quantitative characteristics, the original battlefield cooperative information is subjected to characteristic dummy coding to obtain coding characteristics (cooperation characteristics are the same) of different war units such as different weapon kinds and troops when the different war units perform specific combat tasks.

(2) Because the formats of the cooperative features are different among different combat units, the cooperative features are processed by adopting a dimensionless method of interval scaling, so that the data in the differentiated formats of the different combat units are converted into uniform specifications.

(3) In order to improve the expression capability of tactical cooperative features and enhance the robustness of a model to abnormal data, discretization processing is carried out on the cooperative features after interval scaling, nonlinearity is introduced, and the problem of information redundancy is solved.

And (4) encoding qualitative data on the battlefield into quantitative data for processing through a dummy encoding process of battlefield situation information. Considering that battlefield environments are complex and various and abnormal data cannot be avoided, discretization processing is carried out on the data in an equal-frequency discretization mode so as to enhance robustness of a subsequent model to the abnormal data.

2) And (4) increasing and reducing dimensions of the features.

The data under the battlefield environment situation are complex and diverse, and comprise static characteristics such as names, types and performance parameters of combat units, and dynamic characteristics such as longitude and latitude, height, orientation and speed. Whether static or dynamic, most of them are nonlinear data. When the data is non-linear, using models such as linear regression directly can create a large degree of under-fitting problems. In addition, because the dimensions of static characteristics such as the names, types and performance parameters of the combat units are low, in order to comprehensively consider the static characteristics and the dynamic characteristics of battlefield situations, the static characteristics of the existing battlefield situations are converted by adopting polynomial expansion and kernel functions, and the static characteristics are mapped into a space with higher dimensions, so that the static characteristics can be fitted with wider data.

As the dimension of the dynamic characteristics such as longitude and latitude, height, orientation, speed and the like of the operation unit is higher, the dimension reduction method of Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) is adopted to reduce the dimension of the dynamic characteristics of the operation unit, so that the static and dynamic characteristics of the operation unit can be encoded and processed in a consistent manner.

In order to extract the key battlefield situation information, firstly, a covariance matrix of a data matrix formed by dynamic characteristic data of all the operation units is calculated by utilizing a principal component analysis method to obtain an eigenvalue and an eigenvector of the covariance matrix, a matrix formed by eigenvectors corresponding to k characteristics with the largest eigenvalue (namely, the largest variance) is selected, and the dynamic data matrix of the operation units is converted into a new characteristic space to realize the dimension reduction of the characteristics. In addition, the dynamic features of the combat units are subjected to dimension reduction coding by adopting a linear discriminant analysis mode on the features after the principal component analysis processing, data are projected on a low dimension, and after the projection, the projection points of each category are expected to be as close as possible, and the distance between the category centers of the data of different categories is expected to be as large as possible, as shown in fig. 2.

The static characteristics of the operation units are subjected to dimension increasing and coding through polynomial expansion and kernel functions, and the dynamic characteristics of the operation units are subjected to dimension reduction treatment in modes of PCA, LDA and the like, so that the battlefield situation characteristics of the operation units under unified dimensions can be obtained, and the abstraction of important battlefield situation information is facilitated.

3) The multi-resolution situation characteristic engineering is carried out based on characteristic selection methods such as Filter, wrapper, embedded and the like, the influence caused by special mechanisms, simulation mechanisms and effect delay in a battle-level confrontation scene is solved, and effective abstraction of important situation information is realized.

And 2, extracting known key events in the war game deduction data by adopting a weak supervision self-training method based on the battlefield situation characteristic set, and obtaining a specific key event set.

To realize the identification of specific key tactics and the excavation of novel (new-quality) key tactics, key events are firstly identified and extracted by using abstracted battlefield situation characteristics and intelligent agent operation.

On the basis of a situation-instruction characteristic sequence after digital abstraction, situation-instruction characteristics at a single decision moment and situation-instruction characteristic sequence data of a sequence form are comprehensively utilized, a weak supervision self-training algorithm is applied, and data generated in a large-scale long-time game deduction process is analyzed from the aspects of situation to action correlation, action instruction to battlefield change correlation and the like, and key events in a deduction bureau are positioned from the data.

For the event in the simulation deduction process of the military chess, the present embodiment imitates the event definition in the natural language processing, and consists of two parts: the trigger mark is an element which can represent the event occurrence in the event, in this example, a situation-instruction characteristic pair at a certain time is a main characteristic for determining the event category, and the component element is an auxiliary response process before and after the event trigger mark, and together with the trigger mark, the component element forms the whole frame of the event.

In consideration of the unrealistic situation that a large number of effective labels are effectively marked on events in the large-scale military chess deduction process, the weak supervision method is adopted for detecting the events, and a classification recognition model of the events is trained based on the existing smaller-scale battle key event data set constructed by technical guidance provided by professional military personnel and a large number of military chess deduction data sets by adopting a semi-supervision thought. As shown in fig. 3, the specific operation is as follows:

firstly, constrained clustering is carried out on unlabeled military chess deduction data, the correlation among time information, entity information and battlefield environment information is considered in the process of obtaining a clustering result, and the specific idea is that in a specific battlefield environment, if the entity states are similar in similar time, the same event is probably triggered, so that the constrained clustering is carried out.

After clustering in the mode, a simple semi-supervised support vector machine classifier is trained by combining a labeled battle key event data set, and the final classification hyperplane of the semi-supervised support vector machine classifier is consistent with the distribution of unlabelled data by using the data space distribution of the unlabelled reasoning data. And marking the unmarked data according to the classification result, and preliminarily selecting a trigger mark from the sequence data as a characteristic key event set according to a formed prior trigger mark (typical battlefield situation-typical behavior decision) set, namely the most representative battlefield situation in a time sequence data fragment and an instruction made by an intelligent agent. Therefore, complete event extracted labeling data can be obtained, and further combined training is carried out on the event extracted labeling data and labeled data, so that the model effect is improved.

And 3, based on the battlefield situation characteristic set and the specific key event set, extracting novel key events in the war game deduction data by using a novel key event identification method based on the deep hidden layer characteristics, and marking the novel key events to obtain a novel key event set.

And 31, identifying the novel key event based on the deep hidden layer feature extraction.

In the last step, the present embodiment uses a weak supervision self-training mode to preliminarily classify the key events in the large-scale simulation deduction data according to the prior categories, and the classification result inevitably includes the classification of the new key events into the existing categories or the omission of the new key events. Therefore, the embodiment provides a novel key event identification method based on a deep hidden layer feature, and the novel key event is further detected from three angles of situation features, behavior features and reward features by means of the high efficiency and accuracy of a deep circulation neural network on the extraction of the sequence data features. Missing key events are mined from the derived data and new key events are subdivided from previously located specific key events.

1) And (3) retraining the LSTM network which is subjected to pre-training on the public natural language processing data set by using the specific key event set obtained in the step (2), realizing optimization and correction, and taking out a feature extraction layer, namely a backbone network, from the LSTM network as a backbone of subsequent feature analysis.

2) Constructing a linear classification model according to the definition of a plurality of novel key events, wherein the core idea of the definition is as follows: under the condition that the similarity of the hidden layer characteristics of the known key events meets a specified threshold, making different action decisions, and making a reward report obtained subsequently higher than the specified threshold, wherein the reward report is a novel key event; in the case that the degree of difference of the hidden layer characteristics from the known key events meets the specified threshold, the same action decision is made as in the existing key event set, and the reward return obtained subsequently is higher than the specified threshold, so that the new key event is obtained. Based on the above assumptions, a linear classification model is formed.

3) And (3) according to the linear classification model constructed in the step 2) and the backhaul network obtained by training in the step 1), combining a multi-head attention network, setting a reasonable threshold value by using a fuzzy matching algorithm according to the similarity of hidden layer features, classifying the deduction data again, and extracting the key events which are not extracted originally and the novel key events which are partially mixed in the defined key events. The method comprises the following specific steps: inputting a piece of deduction data, namely battlefield potential state characteristics (potential state-instruction characteristics) in a battlefield potential state characteristic set, and extracting a characteristic extraction network formed by combining a backbone network and a multi-head attention network for the situation around a trigger mark of the deduction data to obtain hidden layer characteristics of the deduction data; sending the hidden layer characteristics and the hidden layer characteristics of the existing known key events into the linear classification model constructed in the step 2) together to judge whether the hidden layer characteristics and the hidden layer characteristics of the existing known key events are novel key events.

And 32, automatically labeling the novel key tactics based on keyword intelligent description and serial connection.

For the novel key events extracted in step 31, it is impractical to manually label each event segment, and this embodiment applies keyword intelligent description and concatenation technology to automatically label the new key events. Aiming at a situation-instruction characteristic sequence divided into novel key events, an intelligent description technology is applied, an attention mechanism is relied on, key words are automatically described for enemy combat entity information, combat intention information, own party strength information, own party instruction information and the like which are the most core in the situation, and label information of the novel key tactics is obtained based on a key word series connection method.

1) The method comprises the steps of utilizing knowledge prior to limit elements of battlefield events, such as types of target entities, the number of the target entities, positions of the target entities, cooperative relations among the target entities, whether the target entities are detected for the first time, current strategic deployment of one party and the like, and then limiting core behaviors of the events to be limited to a plurality of sub behaviors familiar to a commander.

2) Determining core elements to be described by combining with the battlefield element weight output by the multi-head attention network used in the sub-step 3) in the step 31, and determining discrete values for the core elements to obtain situation description keywords, wherein the elements are main bodies, the discrete values are descriptions, and the main bodies and the discrete values are combined to form the situation description keywords.

3) And obtaining the labeling information of the novel key event by applying a keyword template serial connection method.

And 4, combining the key events in the specific key event set and the novel key event set according to a time sequence relation to form a multi-element tactical set, and finishing the excavation of the key tactical set.

In the embodiment, the key battlefield situation and the key actions of the commander before and after the node are automatically and intelligently extracted from the simulated engagement data after deduction is finished, so that the timeliness and the accuracy of key battle mining are improved, the understanding of the commander to the game countermeasure decision made by the artificial intelligent algorithm by the commander is enhanced, the efficiency of carrying out command training by the commander through war push is improved, and the own command art is improved by exploring a new battle.

In another embodiment, the application further provides a critical-tactics mining device for a chess deduction process based on key event detection, which comprises a processor and a memory, wherein the memory stores a plurality of computer instructions, and the computer instructions are executed by the processor to realize the steps of the critical-tactics mining method for the chess deduction process based on key event detection.

For the specific limitation of the key tactics mining apparatus in the chess deduction process based on key event detection, reference may be made to the above-mentioned limitation on the key tactics mining method in the chess deduction process based on key event detection, and details are not repeated herein.

The memory and the processor are electrically connected, directly or indirectly, to enable transmission or interaction of data. For example, the components may be electrically connected to each other via one or more communication buses or signal lines. The processor runs the computer program stored in the memory, so that the key tactics mining method based on the key event detection in the weapon deduction process is realized.

The Memory may be, but is not limited to, a Random Access Memory (RAM), a Read Only Memory (ROM), a Programmable Read-Only Memory (PROM), an Erasable Read-Only Memory (EPROM), an electrically Erasable Read-Only Memory (EEPROM), and the like. The memory is used for storing programs, and the processor executes the programs after receiving the execution instructions.

The processor may be an integrated circuit chip having data processing capabilities. The Processor may be a general-purpose Processor including a Central Processing Unit (CPU), a Network Processor (NP), and the like. The various methods, steps and logic blocks disclosed in embodiments of the present invention may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above examples are merely illustrative of several embodiments of the present invention, and the description thereof is more specific and detailed, but not to be construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A key event detection-based key tactics mining method for a war game deduction process is characterized in that the key event detection-based key tactics mining method for the war game deduction process comprises the following steps:

extracting known key events in the war game deduction data by adopting a weak supervision self-training method based on the battlefield situation characteristic set, and obtaining a specific key event set;

2. The method of claim 1, wherein said extracting the collected military chess deduction data with multi-resolution potential features comprises:

3. The method for mining key tactics in the chess pursuit process based on key event detection as claimed in claim 2, wherein said processing original battlefield cooperative information in the chess pursuit data by dummy coding, interval scaling and discretization method to obtain cooperative features comprises:

4. The method for mining key tactics in the chess deduction process based on key event detection as claimed in claim 2, wherein said using dimension increasing method to increase dimension of static features of operational units in the collaborative features and using dimension reducing method to reduce dimension of dynamic features of operational units in the collaborative features comprises:

and reducing the dimension of the dynamic data of the combat units by adopting principal component analysis and linear discriminant analysis.

5. The method for mining key tactics in the chess deduction process based on key event detection as claimed in claim 1, wherein said extracting known key events in the chess deduction data by using weak supervision self-training method based on the battlefield situation characteristic set, and obtaining specific key event set, comprises:

6. The method for mining key tactics in the chess deduction process based on key event detection as claimed in claim 1, wherein said extracting new key events in the chess deduction data by using a new key event recognition method based on deep hidden layer features based on the battlefield situation feature set and specific key event set comprises:

7. The method of claim 6, wherein the construction of the linear classification model according to the definition of the novel key event comprises:

8. The method of claim 6, wherein said labeling new key events to obtain a new set of key events comprises:

according to the weight of battlefield elements contained in a novel key event output by the multi-head attention network, determining core elements according to the weight of the battlefield elements, and determining discrete values for the core elements to obtain a plurality of situation description keywords consisting of the core elements and the discrete values;

9. A critical tactics digging device of a weapon chess deduction process based on key event detection comprises a processor and a memory which is stored with a plurality of computer instructions, and is characterized in that the computer instructions are executed by the processor to realize the steps of the critical tactics digging method of the weapon chess deduction process based on key event detection as claimed in any one of the claims 1 to 8.