Representative End User Clients
Representative Automation Clients
Representative Software Clients
Predictive maintenance (PdM) has become a key element of modern Asset Performance Management (APM) because it shifts maintenance from a reactive or calendar-based activity to risk-informed, data-driven decision-making. By using condition data, analytics, and machine learning to identify early signs of degradation, organizations can intervene before failures disrupt production, reduce unnecessary maintenance on healthy assets, extend equipment life, and improve the planning of labor and spare parts.
Regular model updates are essential to the long-term success of PdM because assets, operating conditions, maintenance histories, and failure modes do not remain static. In a standard industrial plant, it is becoming common for AI/ML models for PdM to continuously learn. As these models ingest more data, the algorithms adjust their baselines to become more accurate.
In life sciences, an algorithm that changes itself could be a major compliance issue. Here, data sets are often limited or highly context-specific, assets may operate under changing batch conditions and cleaning cycles, and even small changes to monitoring logic or predictive models can trigger formal change control and revalidation requirements. As a result, life sciences companies must balance innovation in AI/ML for PdM with strict controls over data integrity, traceability, and validated system performance.
In life sciences, PdM systems used in GxP-relevant contexts may be subject to computerized system validation, change control, and data integrity requirements. FDA 21 CFR Part 11 may apply where electronic records and signatures are involved, while GAMP 5 provides widely used risk-based guidance for compliant computerized systems. Any update to a validated PdM system should go through change assessment and, depending on its risk and impact, may require regression testing, partial requalification, or formal revalidation.
Four-Step Approach to Updating PdM Models
One practical approach for upgrading a validated PdM model may include change assessment, controlled retraining or updating in a non-production environment, verification/validation testing, and controlled release under QA oversight.
1. Triggering the Change Control
Before any code is changed in production, a formal change control process is followed. Typically, the change control dictates exactly what the new model aims to fix, its scope, and a formal impact assessment detailing how the change will affect patient safety, product quality, and data integrity. Model updates are usually triggered by one of two scenarios:
Model Drift: The physical asset has aged, undergone a major mechanical overhaul, such as a bearing replacement, or process parameters have shifted. The old model’s accuracy is dropping because the operational baseline has changed.
Vendor Optimization: The software supplier has released a more advanced algorithm that promises better accuracy.
2. Parallel Sandbox Retraining
Once authorized to proceed, data engineers may pull historical process data and the latest sensor data into an isolated development environment or sandbox.
Retraining on the New Baseline: The new model is trained on the updated data set.
Retrospective Testing: The new model is tested against historical failure events to prove its statistically superior accuracy over the current production model.
3. The Shadow Deployment and Revalidation Phase
To prove the updated model functions reliably in the actual manufacturing environment, it may be moved to a shadow deployment state. As established during initial validation, the model is fed live production data streams, but its outputs are completely walled off from the other plant systems.
During this trial, validation engineers execute targeted revalidation protocols that typically include:
IQ/OQ (Installation Qualification/Operational Qualification): Verifying that the new model version is securely locked, correctly containerized, and that its unique cryptographic hash matches the development sign-off.
PQ (Performance Qualification): Analyzing the logs to compile data-driven evidence proving that the updated model does not trigger false alerts that could disrupt a batch and successfully maintains its required sensitivity.
4. Predefined Change Procedures
The final phase outlines the acceptable performance boundaries and testing metrics the updated model must achieve during the shadow phase.
QA Sign-Off: If the shadow data successfully meets the criteria, the validation team compiles the data into a Validation Summary Report. Quality Assurance can then fast-track the approval because the update methodology was already prevalidated.
Promotion Toggle: Upon formal QA release, a secure configuration switch is flipped. The old model version is decommissioned or archived, the updated model is adopted, and its predictive alerts begin flowing directly into the plant’s active maintenance workflows.
This rigorous process ensures that as algorithms update, the intelligence used to predict asset health remains traceable, auditable, and compliant. However, the burden of revalidation has historically led many life sciences end users to keep outdated software in place rather than risk the time and effort required for change control and revalidation.
As a result, life sciences companies often adopt PdM and APM changes far more slowly than other industries because of compliance change control issues. To close this gap between fast-moving AI innovation and static regulatory expectations, software suppliers are developing new strategies to reduce validation changes and accelerate deployment. We will take a closer look at some of these strategies in upcoming blog posts.