| As modern military technology continues to develop,drone swarms have become an increasingly prominent military tactic.Simulation platforms have also become an important tool for researchers to conduct simulated experiments and test new strategies.These platforms can simulate various combat scenarios,helping researchers test different attack strategies and planning methods for drone swarms.By collecting potential strategy data from the simulation platform and clustering it,it can provide strategic samples for future drone swarm attacks in real scenarios.Therefore,exploring user strategy data through simulated drone swarm combat tasks,and effectively extracting,processing,and clustering it has significant research value.This thesis is mainly devoted to analyzing the extraction and processing of strategy data and effectively classifying it into different strategy sets using unsupervised methods.Due to the characteristics of strategy data,two models were respectively constructed: a denoising and smoothing model based on time series data,and a deep clustering model based on multivariate time series data.The effectiveness of the models was demonstrated through extensive experiments.Finally,a simulation platform was built,which combined the two models.The research work is summarized as follows:Time series data itself often suffers from noise and asynchrony.Traditional denoising methods usually model the time data by filtering and other smoothing techniques,which largely ignore the characteristic representation of time series data.To address these issues,this thesis proposes a novel denoising autoencoder-based model(KDAE)for addressing the noise and asynchrony issues in time series data.Specifically,the KDAE model utilizes Gaussian noise to add noise to the time series data,and through learning the smoothed and original features of the original time series data,it achieves denoising reconstruction of the original data.In addition,this thesis constructs a loss function for the Kalman smoothing and denoising autoencoder and adds weighting coefficients to enhance the model’s generalization and learning capabilities.Numerous experiments conducted on ten publicly available datasets demonstrate that the KDAE model can effectively denoise and smooth time series data and improve clustering performance.Current deep clustering algorithms are insufficiently robust for multi-variate time series(MTS)data.To address this issue,this thesis proposes a new autoencoder structure,called MDTC,which employs one-dimensional convolution to encode the attributes and temporal features of MTS data.Additionally,to enhance feature extraction of MTS data,a temporal attention module,MCBAM,is added to pool and combine feature weights in the temporal dimension,thus improving temporal feature representation.The effectiveness of the MDTC structure and MCBAM module are demonstrated in experiments on nine publicly available UEA MTS datasets,and both outperform classical clustering algorithms and advanced MTS clustering algorithms.Based on the above research content,through the combination of KDAE model noise reduction and MDTC algorithm,the simulated strategy trajectory data is clustered into different categories,and sample data sets of different strategies are provided. |
