How To Predict Machine Failure Using Data Science

• Read original article here

It is well known, how annoying a machine breakdown can be. Production takes a direct hit because of equipment failures. A great deal of money is lost by the time production restarts. It also impacts OEMs and dealers in terms of lost reputation and business opportunity. Fortunately, these issue can be tackled to a major extent by using data science.

Cost of breakdown is not just the opportunity loss (of potential profit from production), but it also includes fixed cost of the machine. Further, delay in production can attract penalties and lost orders. At times, when other machines also depend on the failed machinery, the cost escalates through the roof. The cost of single breakdown can easily exceed thousands of dollars. The worst part is, this loss can hardly ever be recovered.

The figure above shows some of the components of cost of downtime.

Using predictive models, one can now estimate failure probability. This gives us two abilities. First, the ability to plan maintenance in a manner to minimize loss. Second, to optimize inventory better. Instead of keeping a lot of spare parts in inventory, it becomes possible to keep only the ones that will be required in near future.

Breakdowns may not directly impact an OEM, but it harms the reputation and may also end up in lost business. If a critical item is not available at the nearest point, customers may not hesitate to procure the item from local market. Further, manpower may not be available to repair the machine immediately.

These problems can be avoided if you already has an idea about possible breakdowns. OEMs/Dealers can then, either plan a maintenance and replace the parts or immediately support in case of breakdown. Furthermore, it can help OEMs launch new revenue models of maintenance contracts. This can also ensure that customers do not buy spare parts from local markets.

Perhaps the biggest advantage these models give an OEM is the ability to improve their products. The models indicate which are the factors that impact a component failure, giving them a direction on how to improve component life.

The process starts with identifying the problems to solve. Usually the problem should be broad and broken down into specifics. For e.g. objective could be to reduce cost of operations. One of the several ways to achieve it is by reducing downtime and optimizing spare parts inventory.

Once the problems are identified, the data needs to be collected for analysis. Often the data will not be available. In such case, infrastructure to collect data needs to be built. While building the infrastructure and processes, attempt should be made to enhance the utility of such infrastructure/processes. This can be done by assessing other possible utilities of the data collected. If the marginal cost of adding more data to solve a significant problem is low, it should be pursued. 

Once the data starts getting collected, it needs to be cleaned and visualized. This is not just for the data science team but also for other business stakeholders. If possible and feasible, the dashboard(s) may be built for different stakeholders, depending on the need. The predictive aspects of analyses is a natural progression from exploratory analyses. The dashboards and visualizations are to be followed by the creation of predictive model(s), which are to be tested, reviewed and deployed.

Code heavy!!! Skip if you are not interested in the codes. However, you may like to understand how well the below model performed in results

To illustrate a possible model, real life (normalized) data has been used. The data contains log of certain parameters, taken hourly. In addition, log of failures and maintenance will also be used. The parameters recorded hourly are voltage, vibration, rotation and pressure. Data from 100 machines of 4 types are recorded. Failure of 4 components are recorded. For the sake of simplification, we will analyze failure of ‘comp2’.

Several exploratory steps have been skipped to keep it short and focus on the predictive model.

# Following data are available # 1. Telemetry – Logs hourly parameters (Voltage, Pressure, Rotation and Vibration) for each machine # 2. Failures – Log of component failures. Contains the time slot (that matches the telemetry log) and machineID # 3. Maint – Log of maintainance. Contains time slot, machineID and the component replaced # 4. Assets – Information about the machines – machineID, Model and age # Function to calculate number of periods since last maintenance of a component timeslm %  #creating column with binary (failed or not)   left_join(maint2, by = c(“machineID” = “machineID”, “datetime” = “datetime”)) %>% #Joining maintenance data   mutate(maint = ifelse(is.na(comp), 0, 1)) %>% #New column with binary (maintained or not)   inner_join(assets, by = “machineID”) #Joining machine details df$datetime