Fault Diagnosis Of Axle Acoustic Emission Data Based On Deep Learning

With the rapid development of China’s transportation field,especially in railway transportation,train axle,as one of the most basic and important components,has been in a high load and high strength working state for a long time,which is prone to wear cracks and other faults,and the fault location is not easy to occur.If it can not be solved in time,it will not only bring great benefits to the economy and personal safety of passengers It will also bring inestimable damage and impact to the safety credibility of relevant industries.Therefore,it is particularly important to detect the axle timely and accurately.Acoustic emission(AE)testing technology is the most common non-destructive testing method.It can test the equipment without external force on the testing part,which can ensure the safety of the equipment to a great extent.At the same time,as a one-dimensional time series data,AE signal can contain a lot of information.Considering the above advantages,AE data plays an important role in the field of axle fault diagnosis It has played a great role.According to the characteristics of acoustic emission data,a fault diagnosis method based on term and phase invariant neural network(TPINN)is proposed.The original data is decomposed by improved local mean decomposition(LMD)method,and then the model is multi-channel input,which improves the model’s ability of understanding and analyzing the data;the phase invariant feature is extracted by using the convolution layer of large and small cores in TPINN,which overcomes the influence of signal phase difference on the diagnosis result;the low risk loss function is defined in the output part of the model,which changes the model In the era of big data,the concept of data flow is introduced,and the Kafka framework is used to realize the data flow processing,which can meet the needs of real-time and fast processing without losing data.Finally,the real-time and accurate fault diagnosis of the axle is realized.Through the analysis of the experimental results and process visualization,it can be seen that the method proposed in this paper still has high response speed and recognition accuracy in the face of a large number of and diverse data,and has great advantages over the traditional statistical methods.