Physical maintenance can lead to costly interruptions in the manufacturing process. By means of predictive analytics, repair and maintenance can be scheduled based on real-time predicted failure probabilities.
When customers ask us for advice on predictive maintenance, we tell them that Predictive Maintenance is simply regular maintenance strategy that is improved by Data Science.
Benefits
There is a strong chance that this is more obvious for you than us, but you can expect that improvements to your maintenance strategy lead to reduced downtimes and more robust production, lower maintenance costs, better understanding of reliability (regarding for example suppliers) and minimized warranty costs.
What is your current strategy?
Understanding the current strategy helps to identify what is needed for an improvement. The different levels of maintenance strategies can be described as follows:
- Reactive / Heuristic → No data, run to failure
- Preventive / Planned → Scheduled maintenance (by time or usage, only)
- Condition-based / Statistics-based → Monitor condition (sensors)
- Proactive → Root cause analysis to minimize failures / eliminate defects
- Predictive → Predict failures before they occur
- Self-Optimizing
When you have identified where you are with your strategy, you probably also have identified the corresponding data and tools available for it.
What is the next step?
You can begin to plan what is needed to move to the next level. This can require new data sources or more data, new tools and the introduction of new methodologies.
That can lead to simple methods (from Data Science) being the most sensible next step. Simple methods have the advantage of being easier to use and they are (in general) more commonly available. We suggest to aim for the low hanging fruits first.
What is Data Science?
Data Science is the collection of methods (statistical, machine learning, AI) that help improving the maintenance strategy. The selected method might stay as simple as being “only” improved data integration and preparation, visualisation, business rules, rudimentary statistics and/or continuous monitoring but can as well include advanced techniques from advanced statistical methods, data mining, machine learning and AI, Predictive modelling and model management if needed.
What methods fit your strategy?
In any scenario you should deploy the tools to integrate your data, to clean it and to have a proper way of visualizing it.
For the different levels of strategy some methods fit more nicely than others. Refer to this non-exhaustive list of popular methods:
- Reactive / Heuristic
- Cost-based analysis (e.g., comparison of strategies)
- Structured logging of events, reactions, and reasons
2. Preventive / Planned
- Automation via business rules
- Evaluation and optimization of current plans → Baseline for further improvements
- Reliability / Survival Analysis, Expected Lifetime, Weibull, Cox Regression
3. Condition-based / Statistics-based
- Real-time Monitoring, Alerting and Notifications
- SPC (Statistical Process Control)
- Advanced Visualization
4. Proactive
- Multivariate Analysis, PCA, Clustering
- Statistical methods: Correlation, Time Series Analysis, Regression Modelling, Changepoint Analysis, MASS, Root Cause Analysis
- Feature selection, reduction, and engineering
- Advanced Modelling (DM, ML, AI) to uncover hidden interactions
5. Predictive
- Predictive Modelling, ML, AI
- Model Deployment, Update and Execution: on Device (IoT), as Service (SOAP), Batch-oriented
- Model Management
6. Self-Optimizing
- Monitoring of the (predictive maintenance) system
- Automatic/Continuous model optimization
The Project
As you can see, predictive maintenance can be quite easy. It all depends on you and your environment. A simple project to use predictive maintenance starts with the assessment of your current strategy, identifying all its components and then decide what the next level of strategy should be and what the gap is that needs to be closed.
