Research On Multi-source Information Fusion Method For Test Data

Faced with the increasingly complex marine safety environment,China is vigorously developing target recognition equipment to address new challenges.Compared with single source recognition systems,the advantages of integrating multi-source information are more obvious:enhancing the system’s survivability;Expand the coverage of time and space;Reduce the ambiguity of information;Improve detection performance,etc.Currently,multi-source fusion of underwater targets is facing new challenges.With the continuous development and improvement of vibration and noise reduction technology,the noise level of underwater vehicles reaches or approaches the ocean background noise,making recognition difficult and accuracy difficult to guarantee.Moreover,ocean conditions are complex and variable,with significant differences in water quality and seabed quality across different regions,which also increases the difficulty of target recognition.With the arrival of the era of ocean big data,target fusion recognition has also ushered in new opportunities.With the rapid development of artificial intelligence and deep learning,applying these methods to multi-source fusion of ocean data has become a trend.The fusion of acoustic data from different sources was studied and promoted earlier,and has been widely applied in practical combat.Marine acoustic data has the characteristics of uncertainty,incompleteness,fuzziness,and variability.Traditional fusion methods have reached a bottleneck in the fusion of same and different sources of acoustic data.At the same time,modern naval warfare requires higher requirements for target recognition:faster or even real-time response;More accurate target positioning and stable output in different environments.In order to improve the performance of target recognition,there is an urgent need for new methods such as artificial intelligence to be applied to acoustic data fusion.With the increasing number of digital devices,there are more and more heterogeneous data from multiple sources in the ocean,such as text,sound,images,and videos.These multimodal data have strong complementarity.In recent years,significant achievements have been made in multimodal fusion research,but multimodal fusion also faces some difficulties.The main problem is that there is currently no unified mathematical tool and method to describe and analyze the characteristics of multimodal data,Multimodal fusion still has a long way to go.Ocean equipment is expensive and the number of sea trials is limited.It is unrealistic to collect a large amount of sea trial data,especially under different conditions.However,training depth models requires a large amount of data under different conditions,and how to expand the test data is also a problem that needs to be solved.Firstly,the on-site test data is limited,and simulation data and related test data are good supplements to the on-site test data.The prerequisite for using these data is to verify their credibility.In this paper,a method of reliability analysis of static test data based on statistical distribution is proposed.This method is based on the P-value test,and gives a calculation method of the optimal P-value reliability under Normal distribution.Through experimental analysis,this method can effectively verify the reliability of static data.A metric learning based Univariant Time Series Classification(ML-UTSC)method is proposed for analyzing the similarity of dynamic experimental data based on metric learning.ML-UTSC attempts to measure and learn a mahalanobis matrix to calculate the local distance of multivariate time series data,and combines it with dynamic time warping to calculate the global distance.Relevant experiments have shown that the algorithm proposed in this paper can effectively verify the credibility of dynamic experimental data.Secondly,aiming at the low fusion performance of traditional methods for homologous acoustic test data,an improved deep Canonical correlation feature level fusion method(IDCCA)is proposed.IDCCA uses de-noising autoencoder to remove noise and reduce dimensions,and uses one-dimensional Convolutional neural network to improve the depth Canonical correlation model.Relevant experiments have shown that the proposed method achieves good target recognition rate.A deep coupling recurrent autoencoder(DCRA)method for heterogeneous acoustic data fusion based on convolutional autoencoder structure is proposed to address the issue of low fusion performance in heterogeneous acoustic experimental data.DCRA uses a coupling layer to link multiple single mode autoencoder together,and constructs a joint objective Loss function optimization model.In order to ensure the gradient stability in the long sequence learning process,multi-level gated recurrent units are used to design the autoencoder model.Relevant experiments show that the deep coupled recurrent autoencoder method is better than the single mode method,and also better than the latest multi-mode fusion method.Once again,in response to the high difficulty of multimodal experimental data fusion,this article constructs a two-layer fusion structural model(Cross Attention Fusion by Exploring Inductive Bias,CAF_IB)based on Transformer,CAF_The first layer module of IB is constructed in parallel by Convolutional neural network and multi-head self-attention.Image and sound use independent modules respectively.The second layer module uses multi-head self-attention to achieve cross modal query,and image and sound achieve feature level fusion.This hierarchical parallel structure makes up for the shortcomings of the Transformer model,explores Inductive bias to improve the locality and scalability invariance,and the image and sound input alone reduces the dimension,which controls the Time complexity of the model.Relevant ablation experiments show that the algorithm in this paper is effective and feasible.Finally,in response to the common problems of insufficient accuracy,poor real-time performance,and even conflicting results in traditional methods for multi-source data decision level fusion,this paper proposes a D-S evidence theory decision level fusion method based on Shapelets.The first step of this method is to learn the base classifier Shapelets on single modal data,and the classification accuracy of the base classifier is determined as its weight.The second step is to propose a weighted probability assignment algorithm that increases the weight of the base classifier with high classification accuracy and reduces the weight of the base classifier with low classification accuracy.Two combination strategies are added to the algorithm,which eliminate evidence conflicts and improve efficiency.