Research And Implementation Of Automotive Engine Fault Prediction System Based On Deep Learning

Automobile fault prediction is a very valuable research field.In order to ensure the stability and safety of the engine in the process of driving,it needs to spend a lot of costs to maintain and maintain it.Therefore,how to achieve the balance between economy and safety has become a difficult problem for the automobile enterprises.In the era of big data,the research of engine fault prediction system based on deep learning provides the direction to solve this problem.Fault prediction based on deep learning is a method to predict equipment failure using deep neural network.At present,most relevant studies mainly aim at predicting Remaining Useful Life(RUL),which refers to the time during which an object or system can operate normally in its current state.Deep learning learns the nonlinear relationship between historical time series data and real RUL through multi-layer neural network to predict the RUL of new time series data.Using this technology for preventive maintenance makes it easier to take timely maintenance measures.At present,most artificial intelligence fault prediction methods use the original sensor data directly in the prediction model(such as LSTM,CNN or hybrid model,etc.).Although it has good prediction effect on some public data sets,the automobile engine sensor data under complex working conditions has strong noise and non-constancy.In other words,the statistical characteristics of sensor data will be seriously affected by external factors such as time and environment,which will greatly reduce the prediction accuracy of the prediction model.Previous solutions mainly focus on reducing the noise and non-constancy of data through data processing methods such as time-frequency domain filtering,but these methods have limited effect.Therefore,aiming at engineering applications,this paper focuses on improving the prediction accuracy of deep learning model to automobile engine RUL.From the perspective of network structure design and super-parameter combination optimization,a series of experiments are conducted on the model,and a deep learning prediction model with dual module structure is established,which is suitable for automobile engine sensor data with strong noise and non-steadiness.Based on this model,an automobile engine failure prediction system is developed.The main work and innovative achievements of this paper are as follows:(1)The strong noise and non-constant characteristics of the sensor data of automobile engines pose a challenge to the high prediction accuracy of traditional prediction models.This thesis proposes a deep learning model,LSTM-VAE-1DCNN,with a dual-module structure consisting of an anomaly detection submodule(EDSM)and a remaining useful life(RUL)prediction submodule(RULPSM).The EDSM transforms the raw sequence into an anomaly sequence with smaller noise and more constant statistical characteristics,which serves as input to the prediction module to output the final RUL value.The experiment uses the sensor data of two engines from a certain automaker’s test bench that ran until failure as the experimental data,and finds the optimal hyperparameter combination for the dual-module structure through experimentation.By comparing the LSTM-VAE-1DCNN dual-module prediction model with a single prediction model 1DCNN,it is concluded that the former has lower prediction error.(2)Based on the deep learning algorithm model in(1),a deep learning-based car engine fault prediction system was developed.The front-end of the system uses the VUE development framework and Element UI framework,while the back-end uses Spring,Spring MVC,and My Batis development technologies.The database used is My SQL.The system uses a web page as the interaction entrance and has functions such as RUL prediction and display,sensor value monitoring,fault and maintenance record management,and user login.Finally,the system was functionally and non-functionally tested using black-box testing methods.