Best practice approach developed by successful PdM implementations
Decide which business processes you want to improve and the desired outcomes you ultimately want to achieve. What you predict must be something you can take action on—otherwise, that prediction has no value. By starting with figuring out the outcome first this can help determine the predictive question you need to answer, and helps you measure the success of your effort.
It is important to understand that Predictive Maintenance cannot be used to effectively solve all maintenance problems. Some key criteria to consider when selecting the problems that you want to tackle include:
Predictive – There needs to be a target or outcome to the problem that can be predicted. There also needs to be a set of actions that can prevent failures once they are detected or occur. Some questions to try to answer in relation to candidate assets for PdM are:
- What are the impacts and associated costs of an unexpected failure of this asset?
- What is the failure history of this asset? What were the associated root causes of these failures?
- Are there tasks that can be undertaken proactively and cost-effectively to prevent or reduce the impact of such a failure?
- What are the costs of repair and/or replacement for this asset?
Operational History – A record of the operational history of the equipment involved that contains both good and bad outcomes is vital. Additionally, any steps taken to mitigate bad outcomes should also be available alongside details of any repairs and/or replacements.
Business Knowledge – A clear understanding of the problem and the associated internal processes and practises are essential to support the analysis and interpretation of the data sets.
Once you have assembled your list of potential candidate problems use the information you have available to you to prioritise the candidates and identify the most suitable to start with.
- Assess each one in terms of how easy or difficult it would be to tackle the problem e.g. do you know the data you might need; is that data available and accessible; etc.
- Assess each one in terms of the current impact on the business caused e.g. downtime costs; customer satisfaction impacts; workforce impacts; product quality impacts; etc.
- Quantify the impact of fixing each problem e.g. monies saved; time freed from manual activities (consider Business Outcomes); etc.
You can then plot the challenges on a four-box model with axes of Ease of Implementation and Benefits and find the optimum pilot candidate.
Define and agree the measurements and associated targets that will be used to validate the pilot success and how they will be measured and communicated.
As with any initiative getting the right skills and capabilities for all stages of the programme is vital.
One of the benefits of starting small (with a pilot) is that you can bring the problem, and the benefits to be realised from solving it, alive for stakeholders within your organization.
Identify the key stakeholder groups across your organisation who would be needed to get Preventive Maintenance implemented and/or would benefit from its implementation. These are likely to include – Operations, Engineering, Maintenance, IT, Executive Management. Engage with one or more member of each group to understand how supportive (or not) they are of what you are trying to do.
Implementing Predictive Maintenance successfully is definitely a team sport. A mix of skills are required to successfully implement PdM. These need to include component /machine engineering; maintenance engineering; data engineering; data science; and communication / stakeholder management.
What is key, particularly for the pilot, is to have people who can do the necessary analytical and technical work to analyse and model the data. Those involved should for example, be analytical, have an improvement mindset, be detail oriented; be excellent problem solvers; have broad familiarity with the equipment and systems. These skills could come from within your organisation or could come from external specialists but having people from within your organisation actively involved is more likely to result in a successful outcome and will grow knowledge and capability within the organisation aiding scaling of the initiative.
A Predictive Maintenance pilot should follow the basic workflow for PdM as shown below:
(I) Inventory and Acquire Data
Identify all potential sources (internal and external) and types (e.g. structured, unstructured) of relevant data. Gather and store the data sets in preparation for analysis. The business outcome being targeted will help identify what data is essential and what is optional.
If possible, get run-time operating data (from your Computerised Maintenance Management Systems [CMMS] if you have one) for a reasonable time period (e.g. 1-2 years). There should always be a date and timestamp on the data sets to aid alignment of the data sets.
If you don’t have enough normal usage operational data you can generate / synthesise data from a physical model of the machine. Make sure to vary parameter values, system dynamics, etc. to generate success and failure conditions.
(II) Pre-process Data
Combine the various data sets to get them ready for analysis. No data set is likely to be perfect so work needs to be done to try and get a realistic picture of normal behaviour for the machine. This involves identifying and removing outliers and noise from the data sets. In some cases you can replace anomalies with approximate values or work with a smaller data set.
(III) Develop and Iterate Model(s)
Analyse the data to identify meaningful patterns. Also identify and extract condition indicators from the sensor data. These are also known as features. These features are the input parameters to the models being trained. (A description of feature engineering is available here)
Next step is to train the model(s). You need to classify the data as being healthy or faulty; set thresholds for the conditions/failure scenarios you are trying to predict; adapt the features and simulating the model(s).
This is usually an iterative process where you may change / add to the features used in the model depending upon the results achieved.
There is much more detail on the methods and techniques for developing predictive and detection models in the Resources section.
(IV) Deploy and Integrate Models
You want to observe how your model works in a (close-to) live operational environment. Feed it with live, streaming data and monitor how it responds. This will refine the model and get it ready for full implementation.
The model you have built can then be converted to code that can be deployed either onto specific hardware / devices or onto other platforms / systems within the IT environment.
For a Pilot you may not undertake this step.
Once you have successfully completed your pilot the next step is to communicate the outcomes to stakeholders and build support and momentum for scaling up the pilot. The pilot should demonstrate the potential for PdM to both solve the candidate problem and more widely within the organisation.
The outputs of the pilot will deepen and validate the Business Case for PdM. Work with the key stakeholders to gain funding and approval to move forward. The next steps is to productionize the pilot but there may also be a desire to tackle other problem areas. Review your prioritised problem list to see if the ease of implementation or impact has changed based upon the outcomes delivered thus far. Adjust the priority as necessary and start working through them.
Productionising your pilot solution is likely to require a number of significant cultural, process, resource, and technology changes across your organisation. Though, as highlighted in the Business Outcomes section, the benefits to be gained from this type and level of change are also significant.