Predictive Maintenance

A manufacturing plant is a complex set of machines, processes and people. A day of production loss in the plant could cost the company millions of dollars. So maintenance becomes an important and critical practice. However, it is generally difficult to reconcile production and maintenance goals. The following charts shows MTTR and MTBF data as collected from a pharma plant in 10-weeks. The scales are all in minutes.

The typical failures could be categorized as:

As one can make out, other than #4, all other failures (total of 23 worksheets) could have been predicted. Besides we noticed many lapses in the maintenance planning and work allocation methods too. Technologies like MicroElectro-mechanical systems (MEMS), Product embedded information devices (PEID), supervisory control and data acquisition systems (SCADA), GPS, RFID and sensors are being incorporated in regular operations to provide for rich and near real-time data about a components operation and its environment. But on-field collection of data is not enough for predictive maintenance.

We implemented a generic maintenance strategy for that plant alongwith their maintenance service provider. Two different approaches were implemented – preventive and corrective; which could be represented in the following fashion:

Data was staged from various logs like parameters, message, energy etc. An extensive labeling and fusion techniques were applied to generate data feeders to the model. Association analysis was implemented to generate a-priori knowledge to identify principal components using knowledge available with the plant maintenance team. An ensemble of decision-tree, linear regression and deep-learning (MLP) was used to predict the outcome; which was fed to the job-sheet generation system. The ensemble was able to achieve 88% accuracy, 73% precision, 93% recall and 68% specificity.

The most interesting outcome of this implementation was that the number of job-sheets were brought down by 15% in the 4 weeks of observation period and at the same time mean time between failures went up from 90,300 minutes to 2,03,000 minutes.