Predictive maintenance is the ability of the system to predict a machine failure. Predictive maintenance phase comprises of 2 parts – one is the ability to predict when the machine/asset failure would happen and secondly to perform maintenance activity before the malfunction happens. Predictive maintenance is one of the most widely discussed topics in the IoT ecosystem.
The first two phases of the manufacturing IoT involved monitoring and condition-based maintenance. These phases can provide us with enough historical data, learnings, the correlation between the data, type of failures and corrective action taken and enabled to predict possible failures and what actions needs to be performed on the concerned asset.
In many places, you would read that predictive maintenance is same or a part of condition-based maintenance. We chose to call it out separately as the scope and implementations are quite different. Both deals with ensuring the maintenance are carried out before failure. The condition based maintenance primarily use monitoring, rules, and anomaly detection techniques; while predictive maintenance takes a step further to analyze volumes of historical or trend data, correlations, and machine specifications to predict an outcome. Predicting an outcome is very complex and an ongoing task, which requires being handled separately.
A simple use case is using the information of the assets and its lifecycle and actual ‘wear and tear’ data of the parts provided through the connected devices; one can possibly predict the remaining life cycle of an asset and when should the maintenance be required. Imagine a dashboard, which lists the assets and its metadata, like manufacturing date, installed date, type, etc. along with its actual usage and maintenance activity carried out during condition based maintenance phase. It also depicts external factors and predictions on remaining life cycle of the asset and a maintenance date. These factors can be used to plan a minimum maintenance downtime, schedule spare parts delivery and ensure maintenance is executed with least impact.
Secondly, every manufacturer typically has historical maintenance records of the systems and usage data in some form, which needs to be converted into required format and can be a valuable input to predict the maintenance activity.
Going back to the elevator use case, take the example of the elevator lift cables. Can a system predict when the elevator lift cables need to be changed? Manufacturing innovations are happening in elevator cables, like using super light carbon fiber ropes that increase the lifespan of the cables, but still changing the lift cables is a costly maintenance activity and at the same time its failure can have a considerable downtime. Ensuring availability of new lift cables, specialized technicians availability, compliance check and all these factors can impact the business operations considerably.
In order to carry out any predictive maintenance for elevator lift cables, the manufacturer needs to look at what data points would be required to predict the failure. As part of its Connect product design, the manufacturer had probably installed a sensor to track the running time or distance served by the cable, a sensor to detect if the elevator is descending faster than its designated speed and to monitor the start and stop instances of the elevator. Sensor input together with the cable’s specified life expectancy can be used to predict when the lift cables need to be replaced. In an actual scenario, many more such data sets need to be provided to predict outcomes.
Predictive maintenance involves building out machine learning models based on volumes of data. Developing machine learning models require considerable time and effort. It’s virtually impossible to expect a system to devise a predictive model which is always 100% accurate (not even human operate with that level of accuracy:)), but should be considerable enough to suggest a cause of possible failure with reasonable accuracy.
Open source scalable machine learning models like Spark MLlib or commercial offerings like SPSS from IBM or Azure ML for Microsoft can aid in building predictive models. The real challenge is building feature sets (attributes) and using algorithms like Support Vector Machines, Logistic Regression, and Decision Trees or an ensemble model using multiple machine learning algorithms to predict an outcome.
The model once developed can be integrated into your IoT platform (as part of the Analytics Platform layer –refer Chapter 1) to predict outcomes in real-time. We would talk about this in detail in our next chapter as part of the services offered by various IoT platforms.
In future, we should see specialized pre-shipped predictive maintenance services targeted for various industries/verticals like connected car, elevator maintenance, wind turbines, etc. These services would provide a generalized machine learning model developed using various factors we talked about earlier. System Integrators would play a key role in building the new machine learning model or use existing machine learning models and integrate with the IoT platform. For instance, take an example of a connected car, using the OBD device (actual diagnostic data at runtime) + GPS location, along with asset metadata (like type and make of car, manufacturing date of various parts and its specifications), a generalized machine learning model can be developed which can help predict maintenance activities and failures for any car type. This assumes that you should be able to look up the metadata for the car and its specifications, for instance, the AUDI car type, model, maintenance service requirements would be different as compared to BMW or an AUDI of a different model. The generalized data model (a connected car would have different input/output parameters as compared to a connected elevator) used by the machine learning model would also be a key component in helping to build predictive models effectively.
Many manufacturers are taking a step in this direction but building predictive models with a good amount of accuracy is not an easy task and this space would see a lot of competition, partnership and innovations from manufacturers to software platform provider to system integrators.