Predictive Maintenance in Pharmaceuticals | Starschema
Gradient starschema cut 02

Predictive Maintenance in Pharmaceuticals

Going from preventive to predictive maintenance


Producing high-value, high-volume pharmaceuticals on a 24/7 operational schedule means that breakdowns are time-consuming, interrupt production and are often extremely costly — damaged products must be discarded, thus directly affecting profitability.

Most modern pharmaceutical packaging solutions for solid unit-dose preparations employ single assembly line machines that perform sorting, cold or thermoforming, blister sealing and boxing in the same sterile enclosed manufacturing line. While this is a highly cost- and space efficient infrastructure and preferable from a safety perspective – the individual blister packs never leave the packaging line before being put into their final packaging – a breakdown in any one part of the workflow renders the entire production line inoperable.

Stringent quality standards

Consumers expect flawless packaging every time — packaging flaws such as failing blister seals can compromise the safety of the product and may even lead to costly recalls damaging the manufacturer’s reputation. With high-value pharmaceuticals, the goal is zero defects – even a single box of defective product can be an expensive wastage. To meet these demands, manufacturers follow aggressive preventive maintenance schedules with pre-scheduled shutdowns as often as several times a day. During these shutdowns, manual and optical examination of the machine's parts and maintenance operations, such as cleaning of adhesive dispensers and sealing machinery of adhesive or sealant residue, are performed.

Simple models do not suffice

Despite this time-consuming and laborious schedule, costly failures occur that result in improperly sealed medication, risking the need to recall affected batches and inflicting millions of dollars in losses. Fortunately, modern packaging machines are equipped with a wide array of devices that can monitor machine status, and complex predictive maintenance models can leverage this data to minimize outages while also preventing unwarranted shutdowns.

In addition to intelligent drives, which provide drive mechanical characteristics such as tension, load variability and overall load, a range of other sensors are available, such as temperature sensors on the thermosealing units and pressure weight sensors to detect under-filled blister packs. In addition, optical cameras and infrared thermal cameras (FLIR) installed at strategic locations can identify anomalies in heat seal output temperature, shape or labeling.

Manufacturers often set thresholds for sensor data to signal if a fault may occur. However, facilities often produce a range of different products, many with different parameters, including different forming areas, forming depth, forming materials (including PVC, polypropylene and laminated aluminum foils), and different lid materials, such as paper/aluminum, soft aluminum and hard aluminum, along with different blister stack size and number of blister packs per carton. With so many varying parameters, simple threshold models are not sufficient.


Gathering the right data

Data scientists and manufacturing specialists can examine production lines to determine what data can be used to predict future failures and how to best collect it. Where needed, recommendations can be given for additional sensors, such as high-resolution optical inspection cameras, infrared thermal cameras and additional vibration and pressure sensors.

Predicting equipment breakdowns and shutdowns

Machine-reported failure events can be collected and data scientists can build a Long Short-Term Memory (LSTM) model, a neural network uniquely suited for time series data analysis. This model can be combined with a deep convolutional subnetwork that processes input from a thermal camera's scheduled snapshots at each phase of manufacturing. Together, this combination creates a very accurate predictor of the likelihood that the production line will encounter a failure event within the upcoming 96 hours of operation, allowing plenty of time to call specialized maintenance and initiate a scheduled shutdown.

Predicting Equipment Errors Leading to Defects

A second LSTM model can be optimized for detecting the probability of non-breakdown errors, i.e. production of defective blister packs, with sensor data and optical inspection data, along with machine settings. The results can be correlated with the results of test runs and random sample QA spot checks. Like the breakdown-detection LSTM, this model, too, is agnostic as to packaging parameters, as these are part of the input variables during training.

Predicting Product Defects

One of the most significant barriers for the wider adoption of deep convolutional neural networks, trained on a set of 'normal' and a set of 'abnormal' packages, was the need to obtain large volumes of training data, i.e. images of defective and non-defective packaging for each different packaging type and product. This is both costly and time-consuming to gather, as defective packaging is encountered relatively infrequently, and defects are often visually. Instead, we employ autoencoders, which learn a compressed (dimensionality-reduced) representation of what normal packaging looks like, and comparethis against the observed images.

By using this method, data scientists can calculate the likelihood that the image will fall within the range of normal packages – the greater the reconstruction error, the more likely that the image shows a defective blister pack. This allows anomaly detection without a massive training set – as few as 50-100 images of normal blister packs are sufficient to train the model, which is approximately two orders of magnitude fewer than would be required with the traditional approach.

The Combined Solution

Together, these three metrics combine to give an estimate of how long the machine will be able to operate at adequate performance without a breakdown or producing faulty blister packs, while also being able to alert the operator immediately if an anomalous blister pack was detected. Designed to continuously accumulate data, the model continuously improves its performance over time.

The solution can be delivered with 'batteries included': training the anomaly detection model for a new blister pack configuration is as easy as providing a set of 50-100 images to a bundled shell script that generates the suitable autoencoder in a matter of minutes. The system maintains a memory of previous models, therefore changing to a different product with different parameters is a single-click operation. The solution is routinely configured to push performance, throughput and QA metrics into a SQL database, but it can also be configured to expose data as a REST endpoint or provide these metrics as a data stream.

Contact an expert:


Principal Data Scientist

[email protected]

Demand Forecasting with Latent Matrix Factorization

For many businesses, demand planning is essential. Profitability, cash flow, and customer satisfaction and retention all hinge on getting this right. This white paper will introduce latent matrix factorization to model demand curves and discuss how it can be used to achieve these outcomes.

Introducing the Stack of the Future for Modern Data Leaders

Fast unobstructed access to data and time to insight matters more now than ever. In these quickly changing times, businesses must innovate and implement a ‘Stack of the Future’ to be able to make accurate, data-driven decisions in minutes, not hours or days. The potential value of data is well known but in the new environment, the ability to easily share and collaborate on data is a competitive differentiator that will be leveraged by forward-thinking companies

COVID–19 Data Set Modeling and Analytics

During times of crisis, companies must look at the available data — both internal and external— and try to understand how that data can be used to determine how the business is currently being impacted, how it is likely to be affected in the future, what are most likely scenarios that will play out, what can be done to counter those scenarios and take advantage of hidden opportunities in this rapidly changing environment. The Starschema COVID-19 dataset ingests reliable data from multiple sources and makes it analytics-ready so it can be easily accessed and used.

A DataOps Journey

Keeping your data platforms running with operational efficiency is both paramount and can be a costly and complicated endeavor. Join us and learn how to apply strategies, techniques, and tools to build a reliable and effective DataOps practice in your organization.


This website uses cookies

To provide you with the best possible experience on our website, we may use cookies, as described here. By clicking accept, closing this banner, or continuing to browse our websites, you consent to the use of such cookies.

I agree