Theory And Application Research On Wheel Rail Force Load Identification Based On Data Modeling

This dissertation focuses on wheel-rail load identification theory based on data modeling and its application. Including time-frequency feature extraction technology of track-vehicle system inspection data, sparse principal component analysis feature fusion technology of multi-nodes feature data, wheel-rail load identification method based on feature data modeling using artificial neural network and its application in the evaluation of the track quality.To analyse and extract the time-frequency features of track-vehicle system inspection data, this study starts from the classical time-frequency analysis theory, extends to the using of parametric time-frequency analysis method to obtain instantaneous frequency and signal s sparse decomposition. A change parameters domain and short time adaptive gaussian chirplet signal decomposition algorithm is proposed. The new algorithm is more accurate than the traditional time-frequency analysis methods. It can be used to provide effective feature data for the establishment of wheel-rail load identification data model.The input of the wheel-rail load identification data model is the time-frequency feature extraction inspection data of axle box, bogie, car body acceleration and track irregularity. The output of the model is the time-frequency features extraction wheel-rail force data. Although feature data has a consistent correspondence between time and frequency domain, there are correlation and multicollinearity interference. Using these data modeling directly will not be able to play its expected role. It will bring the curse of dimensionality, which spends a lot of computing time. To solve these problems, a multi-node and sparse principal component analysis feature data fusion method is proposed, which can eliminate the correlation and multicollinearity interference in multi-node feature data. Each main component has interpretability. It can be used to train wheel-rail force load identification data model. It also helps to identify the wheel-rail force more accurately.After fusing multi-node feature data, large quantity of feature data to train the data model is required. It raises new questions to the existing artificial neural networks and machine learning algorithms. Therefore, a wheel-rail load identification data modeling method based on multi-node L1/2-SpasePCA-ELM neural network is proposed. The L1/2 regular conditions are applied to increase sparsity constraint to determine the neuron nodes number of hidden layer. It provides a basis to select neuron nodes number of hidden layer in ELM learning algorithm. This algorithm has good stability, strong generalization ability and fast training speed. This algorithm is compared with other artificial neural network data modeling algorithms using simulation data and verified using field test data.Finally, a track quality status assessment method is proposed based on the track irregularity T2 statistic associate with wheel-rail force. The integrated variable consisted of seven track irregularity in the main subspace projection vector unit-normalization is accepted to obtain T2 statistics. It can be used to evaluate the track quality status in individual sampling point and 200 m section. Using track irregularity and the wheel-rail force data in high speed railway and existing lines to verify the method, the results show that T2 statistics can reflect the track quality more comprehensive.