Research On State Recognition Method Of S700K Switch Machine Based On Sound Signal

The S700 K switch machine is an important signal infrastructure commonly used for high-speed railway turnouts in China.And its function is to be responsible for pulling the railway turnout and changing the travelling direction of the train.The operating status of the switch machine is related to driving safety.The timely and accurately identified abnormal condition of the switch machine can help to prevent accidents in train operation.At present,the main method for recognizing the working status of switches is manual judgement of their microcomputer monitoring data,but this method has the problems of low identification accuracy and long processing time,which makes it difficult to meet the requirements of efficiency and accuracy for high-speed railway equipment status recognition.Therefore,it is of great value and relevance to study an accurate and efficient method for identifying the operating status of switch machines.Based on the advantages of sound signals in the field of non-contact fault diagnosis,this thesis investigates the front-end processing and feature extraction algorithms of the S700 K switch machine under different states,establishes and improves the sound signal state recognition model,and realizes the state recognition of the switch machine by action sound signals.The main research work of the thesis is as follows.(1)Based on the basic structure and working principle of the switch machine,combined with literature and field research,nine common states of the switch machine are summarized,including eight fault states and one normal state.and the characteristics of sound signals corresponding to each state and the causes of the fault states are analyzed.The front-end processing of the sound signal is then introduced,and a feature extraction algorithm that combines the overall features of the signal in the time-frequency domain,the local features of the Meier cepstral coefficients and the dynamic features obtained by their differencing is proposed to provide a more comprehensive and detailed characterization of the mechanical sound signal of the switch machine,in view of the shortcomings of the common sound feature extraction methods.(2)Based on the basic structure of a convolutional neural network,three different depths of switch machine condition recognition models are first constructed for the extracted sound feature parameters.Afterwards,K-fold cross-validation and grid search methods are combined to search for the optimal depth and parameters of the recognition models.Finally,the asymmetric spatial splitting idea of convolutional kernel is used to improve the network structure with the optimal depth.The experimental results show that the improved SFACNN model can improve the recognition accuracy by 3.5%.(3)To further improve the recognition efficiency and accuracy,the SFACNN is improved by using the extreme learning machine technique.The experimental results show that the recognition accuracy of the improved ELM-SFACNN hybrid model can reach 98.6%,and the model robustness has been improved while reducing the model training time.In the comprehensive comparison experiments,the improved feature parameters combined with the ELM-SFACNN hybrid model gave the best results.In summary,this thesis improves and optimizes the feature extraction and model classification algorithm for the action sound signal of switch machines,which achieving the fast and accurate identification of different operating states of switch machines,and providing a guarantee for equipment maintenance and safe train operation.