Deep unsupervised learning approach , device and storage medium for airspace complexity evaluation

A method for evaluating airspace complexity based on deep unsupervised learning is used to process airspace complexity factors extracted from airspace operational situation data, comprising:

• step 1:
• establish an airspace complexity evaluation model based on the deep learning network and obtain the initial centroids of the clusters;
• a stacked autoencoder model is adopted for the deep learning network;
• the stacked autoencoder comprises an encoder network and a decoder network, which are mirror symmetry;
• the original airspace complexity factors sample is input into the encoder network of the stacked autoencoder and the low-dimensional embedded representations of the airspace complexity factors are obtained;
• the encoder network performs non-linear dimension reduction on the original data;
• the decoder network reconstructs the data according to the low-dimensional embedded representation;
• a nonlinear mapping function is obtained by training the stacked autoencoder model;
• the objective function of training is ∥y−x∥2, where x represents the input data of encoder network and it is a collection of history airspace complexity factors and z represents the corresponding low-dimensional embedding points set;
• y represents the reconstructed data of the decoder network;
• θ represents the network parameters of the stacked autoencoder;
• three complexity levels of the airspace operational situation are corresponding to three clusters;
• K-means clustering is used to obtain the initial centroids of the training sample set represented by low-dimensional embedded data;
• step 2:
• construct a loss function used to train the deep unsupervised model for airspace complexity evaluation;
• assume that the low-dimensional embedded points have a real assignment distribution for clustering tasks;
• use the t-distribution to construct a soft assignment, which measures the similarity between low-dimensional embedded points and cluster centroids;
• the training goal of the airspace complexity evaluation model is to make the soft assignment distribution closer to the real assignment distribution;
• the relative entropy KL-divergence in information theory is used to measure the difference between the soft assignment distribution and the real assignment distribution and construct the loss function;
• step 3:
• optimize the loss function in step 2 by the gradient descent algorithm using the training sample set x;
• the gradient descent method is used to minimize the loss function of step 2 to train the airspace complexity evaluation model;
• the three cluster centroids of airspace complexity levels are updated during the training process;
• after step 3, a trained airspace complexity evaluation model and three cluster centroids of airspace complexity are obtained;
• step 4:
• use the trained airspace complexity evaluation model to evaluate the airspace complexity in real time;
• the air traffic controller extracts the airspace complexity factors from the current airspace operational situation data;
• and inputs them into the trained airspace complexity evaluation model to obtain a low-dimensional embedded point;
• calculate the soft assignment distribution between the low-dimensional embedded point and the centroids of the three airspace complexity clusters;
• the airspace complexity corresponding to the centroid with the largest soft assignment distribution is the complexity level of the current input airspace complexity factors.