World Record Results for Quality 

Predictive AI for Monitoring High-speed Autoinjector Assembly

An autoinjector is a complex mechanism that demands 100% quality to ensure each device functions correctly every time. At ATS, we were hearing from customers that they needed this quality but at faster rates. Given the ever-increasing demand for autoinjector assembly across all sectors, the engineers in the Innovation Center saw an opportunity to review our proven processes and technology, and challenge established monitoring rates.

Using our SymphoniTM high-speed digital assembly platform, and a process for data acquisition and analysis throughout the assembly process, ATS was able to reduce total assembly stations by 60%, lower reject rates, and beat world record quality rates.

As a result, autoinjector manufacturers can have increased confidence that their fully assembled devices are correct and safe, and their cost of production is reduced with increased throughput and less waste.

THE PROBLEM

When increased volume is required in a delicate process such as autoinjector assembly, issues can include:

  • Traditional autoinjector assembly processes that drive to a fixed stop position regardless of part variation,
  • Breakage from increased force,
  • Too many assembly stations for the desired throughput,
  • Downstream quality checks that aren’t as effective as in-line monitoring, and
  • No solid understanding of the assembly of a highly complex medical device.

The process is impacted by many factors: incoming part dimensional tolerance issues, insertion station alignment issues, component alignment issues, and the dynamic insertion forces on the product all impact the process. Furthermore, all of these factors have the potential to change over time. The monitoring system is required to identify and isolate the change, ideally before rejects and waste are created.

THE SOLUTION

In the Innovation Centre, the ATS team believed they could achieve high speeds, and have very precise control at all points of the assembly process. The desired process was to drive to a fixed force feedback position that is blind to variation from process to process. Using existing Symphoni technology as the foundation, they quickly designed tests to ensure monitoring on multiple variables was accurate and representative of forces, positions, and temperatures on actual products.

In order not to increase rejects with increased assembly speed, the process starts with a data driven, high-speed, repeatable and robust assembly process. The data need is realized with process data acquisition and monitoring for complete confidence that each device is being properly assembled at each station, while the high speed, repeatable and robust assembly platform need is met with the Symphoni platform which provides the capability to reliably assemble units at extremely high speeds, with a full set of process data monitoring tools.

Multiple tests were conducted to understand the characteristics of good and bad parts, and insertions, which allowed the engineers to set the limits between what was identified as good and bad. Engineers could monitor the force / displacement curve during parts insertion initially using Kistler equipment for classification of pass/fail decisioning.

Taking the next step, the team employed Yanomaly’s AI-based methods for detection, testing two syringe types, allowing them to:

  • monitor early detection of drift or trending towards spec limits / failure modes (ie. near bad)
  • monitor classification of parts in pass / fail (including failure mode)
  • develop advanced preprocessing and analysis methods to classify waveforms with robustness against variation due to high volume / speed, and inertia effects due to moving load cell

This early and continuous monitoring is useful for developing advanced preprocessing and analysis methods to classify waveforms with robustness against variation due to high volume / speed.

Yanomaly Algorithm 1: Drift detector / extrapolator:
The drift detector detects slow upward/downward drift in a signal while the drift extrapolator forecasts whether a drift, if continued, will cross a threshold within the future time window.

It is insensitive to short-term fluctuations making it useful for detecting:

  • Clogging of filters / pipes
  • Dirt buildup on sensors
  • Wear of a stamping / machining part

Yanomaly Algorithm 2: Multivariate anomaly detection

Abnormal Signal Relation Detector
The AI-based algorithm learns the relation between signals from good data and detects abnormalities; in this case between displacement and force

  • Predicts each monitored signal based on the other signals and the operating context
  • Compares the prediction with the actual value and computes an anomaly score
  • Learns how predictable certain features are and adjusts its sensitivity based on that

Sensitivity can depend on so-called ‘context features’; that is, a discrete variable that indicates the operational condition (for example, an integer indicating the section of the force versus displacement curve).

Compared to quality inspections of the completed device, it eliminates additional stations, and detects rejects that otherwise cannot be inspected unless destructive tests are performed.

Signal Validation
’Parallel univariate’ detector computes (configurable) the following properties of the raw signal:

  • A filtered version of the signal value to eliminate brief peaks
  • The signal fluctuation
  • The absolute and relative gradient of the feature
  • The signal travel

It learns from training data thresholds on each of these properties and detects when these thresholds are violated. The learned thresholds can be manually overridden.

THE RESULTS

Critical Learnings for the Yanomaly AI-enabled Process

The dataset has the capability to monitor incoming part dimensional tolerance issues, insertion station alignment issues, component alignment issues, insertion forces on the product.

This dataset is the enablement for classification of both good and rejected product.

Classification of the reject product enables immediate root cause analysis and faster problem resolution.

Classification of the good product enables predictive maintenance so that issues can be addressed prior to the rejection of product.

These results address the needs of manufacturers of drug delivery devices, ensuring they can increase production while increasing quality (more good products) and decreasing cost (less waste).

Pushing the limits of high-speed autoinjector assembly monitoring

Responding to calls from autoinjector manufacturers for increasing throughput without sacrificing quality, the ATS Innovation Team responded with a data-driven solution that delivered unexpected results

More Case Studies

BROCHURE

Your Pharma Projects, Our Proven Automation Solutions

Are you a pharma manufacturer? Learn about how we can help you with your project

BROCHURE

Your Medtech Projects, Our Proven Automation Solutions

Are you a medical device or diagnostic manufacturer?Learn about how we can help you with your project

Fact Sheet

Standard Autoinjector Manufacturing System

Pre-engineered for lower design costs and faster delivery; suitable for any common three-piece autoinjector