Research On Remaing Useful Life Prediction Method Of Medium Carbon Steel Based On Recurrent Neural Network And Deep Belief Network

In the modern industry,the condition of the machine is related to production safety and production progress,especially for some important equipment.If the important equipment fails suddenly,it will cause huge economic losses to the whole enterprise and even the whole region.Therefore,Prognostics and Health Management(PHM)has developed rapidly in recent years.As a part of PHM,remaining useful life(RUL)prediction has gradually moved from theory to practice.Accurate life prediction of critical equipments will greatly enhance the safety of the entire industry and have great economic value.In the era of Internet of Things,engineers can easily collect large amounts of data and then use deep learning method to make decision analysis.It become a new direction in the study of metal fatigue remaining useful life prediction.This paper focuses on life prediction.The main contents of this paper include the following parts.At the system design level,the hardware and software of the system are determined.The hardware part includes:computer,acoustic emission sensor,signal amplifier,acquisition card,infrared sensor,etc.The software includes Labview,FLIR Research IR,data processing program,feature extraction,prediction program and database programs,etc written by python.Through analysis,the three-tier system architecture is determined,the function lines of the system are described and the shared database is established according to the actual needs.At the data processing level,this paper uses acoustic emission sensors and infrared sensors as data acquisition devices,and uses wavelet denoising method to denoise the signals collected by multi-sensors.And then further de-sampled in order to reduce the difficulty of model calculation.The temperature matrix signal collected by infrared sensors is segmented,and the temperature data unrelated to the research object is eliminated.Better and more reliable data are used as the next prediction model,which is the basis of remaining useful life predictionAt the model algorithm level,this paper mainly uses two mature deep learning algorithms for life prediction,namely,Recurrent Neural Network(RNN)and Deep Belief Network(DBN).According to the actual problems such as limited samples and over-fitting,this paper puts forward the solution,and carries out a step-by-step analysis of model structural parameters.Two models with good prediction effect are obtained.Finally,combined with multiple physical fields,the final prediction model is established by using ensemble learning.It has obvious improvement effect by experiment.At the test verification level,eight groups of experiments were completed with the widely used concave medium carbon steel specimens,cooperating with the established life prediction system.The fatigue test bench to simulate the fatigue process of metals under cyclically varying loads.In the course of the test,the related characteristics of the specimens during the fatigue process is consistent with the conclusions of the previous analysis work.At the end,the actual experimental data verify that the fatigue life prediction value is consistent with the true value,the prediction error is within acceptable range,and the system has high ease of use and strong generalization ability.Based on the analysis of the development status and shortcomings of the remaining useful life prediction methods,this paper puts forward a prediction method based on multi-sensor and discusses how to improve the prediction accuracy and generalization performance.The topic is relatively novel,which has theoretical value and practical significance.