Aiming at the problems of high power consumption and complex scene for scene perception in universal transportation modes, a new transportation mode detection algorithm combining Residual Network (ResNet) and dilated convolution was proposed. Firstly, the 1D sensor data was converted into the 2D spectral image by using Fast Fourier Transform (FFT). Then, the Principal Component Analysis (PCA) algorithm was used to realize the downsampling of the spectral image. Finally, the ResNet was used to mine the local features of transportation modes, and the global features of transportation modes were mined with dilated convolution, so as to detect eight transportation modes. Experimental evaluation results show that, compared with 8 algorithms including decision tree, random forest and AlexNet, the transportation mode recognition algorithm combining ResNet and dilated convolution has the highest accuracy in eight traffic patterns including static, walking and running, and the proposed algorithm has good identification accuracy and robustness.

In the multi-step traffic flow prediction task, the spatial-temporal feature extraction effect is not good and the prediction accuracy of future traffic flow is low. In order to solve these problems, a fusion model combining Long-Short Term Memory (LSTM) network, convolutional residual network and attention mechanism was proposed. Firstly, an encoder-decoder-based architecture was used to mine the temporal domain features of different scales by adding LSTM network into the encoder-decoder. Secondly, a convolutional residual network based on the Squeeze-and-Excitation (SE) block of attention mechanism was constructed and embedded into the LSTM network structure to mine the spatial domain features of traffic flow data. Finally, the implicit state information obtained from the encoder was input into the decoder to realize the prediction of high-precision multi-step traffic flow. The real traffic data was used for the experimental testing and analysis. The results show that, compared with the original graph convolution-based model, the proposed model achieves the decrease of 1.622 and 0.08 on the Root Mean Square Error (RMSE) for Beijing and New York traffic flow public datasets, respectively. The proposed model can predict the traffic flow efficiently and accurately.

Multi-label feature selection has been widely used in many domains, such as image classification and disease diagnosis. However, there usually exist missing labels in the label space of data in practice, which destroys the structure and correlation between labels, so that the learning algorithms are difficult to exactly select important features. To address this problem, a Multi-label Feature Selection based on Label-specific feature with Missing Labels (MFSLML) algorithm was proposed. Firstly, the label-specific feature for each class label was obtained via sparse learning method. At the same time, the mapping relations between labels and label-specific features were constructed based on linear regression model, and were used to recover the missing labels. Finally, experiments were performed on 7 datasets with using 4 evaluation metrics. Experimental results show that compared to some state-of-the-art multi-label feature selection algorithms, such as multi-label feature selection algorithm based Max-Dependency and Min-Redundancy (MDMR) and the Multi-label Feature selection with Missing Labels via considering feature interaction (MFML), MFSLML can increase the average precision by 4.61-5.5 percentage points. It can be seen that MFSLML achieves better classification performance.

Aiming at the problem that the traditional Laplacian score for feature selection cannot be directly applied to multi-label tasks, a multi-label feature selection algorithm based on Laplacian score was proposed. Firstly, the sample similarity matrix was reconstructed by the correlation of the common and non-correlated correlations of the samples in the overall label space. Then, the correlation and redundancy between features were introduced into Laplacian score, and a forward greedy search strategy was designed to evaluate the co-operation ability between candidate features and selected features, which was used to evaluate the importance of candidate features. Finally, extensive experiments were conducted on six multi-label data sets with five different evaluation criteria. The experimental results show that compared with Multi-label Dimensionality reduction via Dependence Maximization (MDDM), Feature selection for Multi-Label Naive Bayes classification (MLNB) and feature selection for multi-label classification using multivariate mutual information (PMU), the proposed algorithm not only has the best classification performance, but also has a remarkable performance of up to 65%.