Research Of Fault Diagnosis Of Turnout Switch Machine Based On Apoptosis Mechanism Spiking Neural Network

With the continuous development of science and technology and industrial automation,the speed of trains has rapidly increased.As the core equipment to control the running direction of the train,the fast and accurate diagnosis of faults of the turnout switch machine is an important foundation to ensure the safety of the train operation.At present,the railway operation and maintenance mainly use the microcomputer monitoring system to monitor the fault of the turnout switch equipment.The system alarms often correspond to a variety of different fault modes.The actual troubleshooting is still through the skylight detection of the workers on the spot.In order to ensure the safety of the equipment,sensor equipment cannot installed inside the turnout switch machine on site,which requires continuous communication between signal and track workers.When the number of equipment to be detected increases,the time for detection will increase.In addition,the occurrence of faults has a certain order in time,and the current commonly used fault classifiers cannot well reflect the time characteristics of faults of the turnout switch machine.The acquisition of sound signals has the characteristics of non-contact measurement,which can provide very convenient conditions for the detection of on-site equipment.This thesis takes the ZDJ9 turnout switch machine as the research object,and uses the sound information during its operation to detect faults.A hybrid method is used to extract sound features.A spiking neural network model combined with the idea of brain-inspired intelligence is established as a fault classifier,which can reflect time characteristics.The main research work is as follows:(1)Aiming at the feature extraction of sound,a feature extraction method combining empirical mode decomposition,wavelet packet decomposition and added energy entropy is proposed,which can well solve the problems existing in traditional time-frequency decomposition methods.The average fault diagnosis accuracy can reach91% using the support vector machine.(2)Aiming at the shortcomings of the existing spiking coding methods,a coding method based on time delay frequency is proposed,which combines time-based time delay coding with frequency-based Poisson coding.It can embeds the time characteristics of sound information in spikes.Besides,the effect of random noise caused by delay coding is effectively solved,and the continuous stimulation of the pulse to the neuron is guaranteed.(3)A spiking neural network model that can reflect the time characteristics of faults of the turnout switch machine is established.Competition rules and dynamic ignition mechanisms are added to increase the competitive relationship between neurons and ensure that they will not always be in a dominant position.The network can learn autonomously through local unsupervised learning rules.Compared with the traditional neural network,under the same number of network layers,the fault diagnosis effect of the established spiking neural network model is better than that of the traditional neural network.Their average failure diagnostic accuracies are 79% and 64.3%,respectively.(4)Aiming at the weight connection problem of the established model,a apoptosis mechanism based spikes stimulation is proposed.It calculates the importance of neurons by counting the number of pulse sequences.By judging whether they meet the significant characteristics of the fault,the influence of the weight connection is reduced.Finally,the average fault diagnosis accuracy reaches 87.6%.