The Design And Implementation Of Predictive Maintenance System Based On Multiple Data Sources

Nowadays,with the upgrading of manufacturing industry and the continuous development and application of power machinery such as vehicles and aircraft,many enterprises are facing the pressure of continuous cost control.How to effectively improve the availability of machines,reduce the maintenance costs and the possible outage losses of machines caused by faults or non-faults has become a research topic in recent years.At the same time,effective analysis of the underlying causes also makes sustainable optimization of products for manufacturing enterprises from the design and manufacturing links.There is a consensus in the industry that product maintenance plays a vital role for enterprises.In the broad sense,maintenance is divided into Corrective Maintenance,preventive maintenance(Preventive Maintenance)and predictive maintenance(Predictive Maintenance,PdM).Corrective maintenance is a non-planned maintenance.In recent years with the rise of the Internet of Things,it is possible to obtain and analyze machine status information in real time.Preventive maintenance(Preventive Maintenance),also known as regular maintenance,is a maintenance based on time.At a specified time interval,a shutdown inspection,disintegration,and replacement of parts are carried out to prevent damage and production loss.The new predictive maintenance(Predictive Maintenance)combines the log of the machine sensor and the maintenance records of the product,and uses the popular machine learning and the widely-used survival analysis model in the pharmaceutical industry to predict the service life of the whole product and the important parts.Take door-to-door inspection initiatively or replace the related parts before the fatal error take place and then to reduce the downtime loss.The advantages of predictive maintenance were as bellows,overcoming the blindness of prevention and maintenance,adopting different methods according to the different state,reducing the long time problem caused by the post maintenance,improving the availability of equipment,reducing the cost of monitoring,reducing the workload of maintenance and improving the economic efficiency.Predictive maintenance is a forward-looking technology.The automation solution of data acquisition has its own standards per different industry,and it is difficult to have the same standard and production line using the same product and technology.The paper proposes an innovation on top of the existing experience and knowledge,designs and implements an effective end-to-end solution to integrate data sources,build prediction model,and visualize the prediction results.As for data integration,we use the traditional data warehouse theory for reference,clean and transform different data sources,build up prediction model by selecting characteristic variables,and then compare the prediction methods based on Cox Proportional Hazards and deep learning,and draw a conclusion that Cox Proportional Hazards has better prediction result under experimental conditions.On this basis,combined with the existing research results and related mechanisms,the future improvement and promotion solutions are discussed.Firstly,this paper introduces a predictive maintenance maturity model based on the overview of predictive maintenance technology and introduces the general process of predictive maintenance.Taking the fault prediction model as an example,this paper introduces some common statistical models such as prediction based on Cox model,logistic regression,support vector machine,deep learning and prediction model based on decision tree.Taking the analysis of key problems as example,two solutions are proposed to predict the failure probability and available time of large-scale mechanical equipment.(1)Establish a prediction model for Cox PH and integrate the algorithm model into Vertica data warehouse for invocation and analysis through R-DUF,using large data storage to improve performance.(2)A survival analysis framework based on deep learning is attempted,and the analysis and effect evaluation based on the analysis of test data sets are carried out.Meanwhile,the accuracy of the model is evaluated based on the actual project data.The results show that this prediction method is more effective than the existing prediction methods which based on single data source and traditional statistics model.On the other hand,the paper introduces scheduled task based on Tidal scheduling tool,which guarantees the consistency and consistency of the whole system from data loading to model building and then to prediction results visualization.Finally,based on the analysis of other similar prediction solutions in the industry and the conclusions of the forecasting results,some future improvement solutions,such as real-time streaming data access and dynamic updating of model parameters are proposed.