Research On Failure Warning Of Combine Harvester Based On Deep Learning

The traditional fault warning method of combine harvester based on "artificial feature extraction + shallow classifier recognition" is highly dependent on professional knowledge and domain expert experience,and it is gradually difficult to meet the requirements of modern automated early warning.As a new force in the field of modern artificial intelligence,deep learning can automatically learn representative features from data,and to a large extent get rid of the dependence on diagnostic experts.Aiming at the limitations of traditional combine harvester fault early warning methods,this paper focuses on the research of combine harvester fault early warning technology based on deep learning based on artificial intelligence technology.The main contents of this paper are as follows:First of all,based on the analysis of research status at home and abroad and on the basis of the working principle of the combine harvester,this paper determines the various working parts as monitoring objects,and build the combine harvester on-board data monitoring system,working data of important components are collected through sensors,and real-time working data are stored locally and sent to remote servers at the same time,remote data are used as historical data for further learning and mining.Secondly,on the basis of combine harvester working characteristics,we have conducted a large number of working data collection experiments and simulated fault experiments,according to the characteristics of the original data,we have formulated data preprocessing rules and developed the data preprocessing programs,including data inspection,data cleaning,data integration,etc.By using the Fliter data feature selection method,the noise feature value is screened out,which greatly improves the purity and quality of data.Finally,this paper proposes a variant of Long Short-Term Memory(LSTM)Networks-Simplified Minimal Gated Unit(SMGU),the working data of harvester verify that SMGU has the same performance as LSTM,but only consumes 1/6 of the time and computing cost compared with LSTM,which greatly reduces resource consumption and fault warning response time.At the same time,aiming at the characteristics of combine harvester which has a large amount of working data,we propose a method that combines the Self-Organizing Feature Mapping(SOM)with the SMGU network.The SOM-SMGU network not only solves the problem of huge data volume through unsupervised learning,but also can better complete the classification problem based on time series tasks.Moreover,it can gradually update the fault warning system as the data volume changes.This paper makes a comprehensive performance measurement of the SOM-SMGU in the early warning task of the combine harvester through the three evaluation standards of loss function,accuracy,and f1-score,compared with supervised and unsupervised learning algorithms,the SMGU network has the best performance underthe condition of fault warning of harvester,and its comprehensive fault warning accuracy can reach 98.5% in the first two seconds before the fault occurs.In addition,the visualization results of warning classification data and the trend of warning error recognition are also presented to provide reference experience for the follow-up research.