One of the key challenges pharma manufacturing facilities will (and do) face is unscheduled downtime. A 2011 study has found that 23% of drug shortages in pharma are cause by manufacturing difficulties, of which loss of production is a contributing factor.
How can pharma address the issue of unplanned downtime?
The key lies in moving from a reactive, fix on failure strategy to a predictive and preventative maintenance strategy. According to McKinsey, the use of a predictive maintenance strategy can reduce asset downtime by 30-50% and increase asset lifespan by 20-40%.
What is predictive maintenance?
Predictive maintenance relies on data from assets collected during normal operation periods and aims to predict when maintenance will be required, maximising the lifespan of different assets, mitigating against unplanned downtime and reducing operational costs.
How does predictive maintenance work?
In practical terms, predictive maintenance evaluates the condition of the equipment by performing either periodic (when equipment is offline) or continuous (when equipment is online) equipment monitoring, with a view to allow for maintenance to be planned when it’s most cost-effective and before the equipment losses performance (as opposed to time-based/scheduled maintenance, where an asset is maintained whether it needs it or not).
Predictive maintenance relies on the principles of statistical process control to determine (predict) the future trend of the equipment’s condition, and therefore when maintenance will be required.
From a data science perspective, the development of a predictive maintenance strategy follows the following steps:
i) Data inventory and acquisition
Potential sources of measurements (both from assets and from process performance) are identified. Special consideration here goes to the timespan covered by available data – it is imperative that collected measurements cover both “normal” and “failure” events. Data is then collected in a centralised repository for processing.
ii) Data pre-processing
Once all data has been collected, it needs to be prepared to be analysed. This often consists of identifying & dealing with outliers and imputing or removing missing observations. During the data pre-processing step, we also aim to identify the minimum common period across all data points by uniformising the intervals at which different measurements are collected.
iii) Model development and evaluation
During the model development steps, the main goal is to identify meaningful patterns in the set of available measurements that will allow us to understand and quantify their predictable connection with the asset life cycle and any “failure” events.
To do so, measurements are often labelled as being taken during a “normal” or during a “failure” event. This information is then used to train and test a number of statistical and machine learning models, which are then evaluated based on a number of pre-set and pre-agreed metrics, such as accuracy, false positive or false negative rate, predictive power, etc.
iv) Model deployment
The best performing model from the previous step is then deployed in a live operational environment for monitoring. Depending on performance, the model may need to be refined and we may find ourselves going back to the previous step and testing different contenders. Once a model has proven to work within the expected parameters in a live operational environment, the model can be fully deployed and used as part of day-to-day operational tasks, such as equipment monitoring and maintenance scheduling.
How can pharma successfully adopt predictive maintenance?
Although data collection & processing and model development & deployment are integral and indispensable steps in the development of a predictive maintenance strategy, the implementation and long term adoption of such a strategy relies much more on business buy-in and ownership than on technical expertise.
For a predictive maintenance strategy to be successfully adopted, there are a few considerations to have prior, during and post the development of such a strategy.
The first thing to keep in mind, prior to starting the development of a predictive maintenance strategy, is the business goal. Identifying the key maintenance challenges to be addressed, the degree of difficulty of tackling each one of those challenges and the impact that these challenges have in financial, personnel or time aspects ensures the resulting strategy has a defined scope and application.
Once challenges have been identified and defined, it is often a good approach to select a small number of challenges to focus on. Developing and implementing a predictive maintenance strategy may significantly change the way a business operates, so starting with a controlled environment will drastically increase chances of success.
Starting with a controlled environment also helps ensuring the right people are involved and on board with the strategy being developed. Identifying the key stakeholder groups within an organisation, involving them from the early stages and ensuring they are part of the development process ensures buy-in and instils a sense of ownership across several parts of the organisation, which is fundamental when it comes to the long-term adoption of a predictive maintenance strategy.
Where to start?
As with many data science applications, there is a big misconception that embarking in the development of a predictive maintenance strategy is only suitable for big companies with large resources to invest. However, this does not have to be the case – pilots can be run by a small team of a couple of data scientists and subject matter experts in a matter of a couple months. The only indispensable resource is data and as long as an organisation has plenty of it, the sky is the limit.
If you’d like to understand the potential your company has for predictive maintenance, get in touch to book a data audit. This is usually a short discovery exercise that maps out a roadmap of what will be possible and gives confidence to embark on the journey.