Deep Dive

Classic Approaches to Observability

Historically, manufacturing lines were visually inspected manually by human workers, or by computer vision models reliant on rule-based image processing. These systems used fixed thresholds and simple pattern recognition to detect defects such as mislabeling, cracked vials, particulate contamination, fill level deviations, or broken tablets. While effective in stable environments, such an approach would struggle with variability, as well as subtle, non-standard defects.

Some of their limitations include:

• Sensitivity to noise: Small changes in lighting, positioning, or background could trigger false positives or miss real issues.
• Hardcoded logic: Required extensive manual reprogramming for each product variation or packaging type.
• Limited detection: Could only catch known, pre-programmed defect types, often missing edge cases.
• Scalability issues: As product complexity and regulatory requirements increased, these systems failed to keep pace.

New AI-Enabled Observability

With the rise of both deep learning and computer vision, modern AI-driven observability systems use convolutional neural networks (CNNs) and transformers to learn from large datasets of images. This enables more adaptive, accurate, and scalable inspection. The use of these new vision systems have enabled key advancements in manufacturing.

Imagery may be collected using:

• Conventional cameras, which can be cheaply purchased and positioned to observe manufacturing lines, without requiring any modification to existing machinery (which may otherwise inadvertently trigger machine revalidation requirements under "Good Practice" GxP regulations).
• Line-integrated industrial X-ray Computed Tomography (CT) scanners, which provide 3D scans of manufactured objects, and work by taking hundreds of individual X-ray images from different angles, to capture both the internal and external structures of objects.

Other sensor data may also be integrated, such as those that support real-time in-process monitoring and control systems. This allows for temperature, humidity, and other key information to be cross-referenced and further used to support quality assessments.

Real-time Defect Detection

While off-the-shelf video cameras are capable of capturing hundreds of frames per second, high-speed cameras can be used to capture thousands or even millions of frames per second. Even the latest generation of fast industrial CT scanners, on the other hand, are typically only capable of performing ~thousands of scans per shift. This is, however, a dramatic speed-up compared to state-of-the-art just a few short years ago, unlocking such devices for use in a wider array of manufacturing use-cases, allowing them to play a real role in production batch sampling, high-value electronic component verification, and more.

New AI models trained on high-resolution images or industrial CT scans can detect micro-defects that aren't detectable using traditional vision systems, at millisecond speeds, keeping up with high-throughput manufacturing lines.

As such, anomaly detection is today typically achievable in real-time, with few exceptions.

Self-Learning & Adaptability

Modern AI vision systems continually improve their performance through:

• Active learning loops: Operators validate uncertain predictions, and the system retrains on new labeled data.
• Transfer learning: Base models pre-trained on similar pharmaceutical products can be quickly adapted to new drugs or packaging formats.
• Few-shot learning: AI can generalize from very few labeled examples of rare defect types, crucial in regulated environments where such defects are uncommon by design.

Enabling Smart Manufacturing

Manufacturing observability helps provide the base data required to layer in an understanding of product quality within manufacturing digital twins, in combination with upstream process data and downstream Quality Assurance reports, enabling counterfactual simulations to be run, such as “What if we switch foil supplier?” or “What’s the cost impact of raising the defect threshold?”

As well as powering digital twins and allowing for the simulation of manufacturing lines, this data also enables automated anomaly alerting, log generation, and prescriptive remediation/action suggestions.

Advanced AI platforms like HASH can combine AI-driven vision systems with data from MES, LIMS, environmental monitoring, or operator input, to build a knowledge graph that provides a complete picture of quality across the production process.

Rather than simply classify images as defective or not, AI models are able to probabilistically assess why defects are likely to have occurred and recommend actions, incorporating:

• Root cause pattern recognition: By correlating defect signatures with sensor data (e.g. vibration, temperature, torque), AI can highlight probable causes like equipment drift or material variability.
• Anomaly anticipation: Computer vision fused with production telemetry enables early warnings for line maintenance or raw material substitution.
• Decision support: Predictive analytics can guide whether to hold or release a batch, retrain a model, or escalate for manual review.

GenAI and Copilot Use

Generative AI expands the power of vision systems:

• Natural-language interfaces: Operators can ask, “Show all defects from the last batch with blister seal issues” or “Explain why line 3 is failing more inspections.”
• Incident generation: GenAI copilots can auto-generate deviation reports pre-populated with defect images, logs, and likely root causes.
• Training document generation: Automatically extract visual examples from production runs to create operator SOPs or QA training guides.

Quality-Aware Agentic AI

Manufacturing line observability ensures the availability of reliable, real-time data regarding product quality that may not only be useful to humans, but autonomous agents as well. This allows agentic AI to perform actions with confidence, such as:

• Halting or adjusting production in the event of costly or catastrophic emergent anomalies
• Adjusting process parameters dynamically, as detailed in the Real-Time In-Process Monitoring & Control use case
• Dynamically switching to alternate product formats or failover equipment, enabling "graceful degradation" of manufacturing in the event of observed or inferred equipment failure

Implementation & Enhancement

Selecting A Solution

When researching manufacturing line observability solutions, consider:

• Data readiness: Do you have sufficient labeled image data (defective and good)? Can it be collected and stored securely? Will you need help bootstrapping this dataset?
• Model validation: Are you compliant with any applicable manufacturing regulations such as GxP, and can you demonstrate AI explainability and audit trails to regulators?
• Hardware alignment: Is your camera infrastructure adequate (lighting, resolution, angle, speed)?
• Sensor augmentation: Will you be linking captured vision data to other sensor readings, and if so how?
• Human-in-the-loop: Do you have protocols for manual override, escalation, and false positive correction?
• Agentic operation: Might you want to support the autonomous agentic operation of your manufacturing line, either now or in the future?
• Integration: Can inspection results be tied to batch records, and shared with QA, QA Ops, and production teams?

HASH offers an open-source, highly integrated AI platform capable of unifying structured and unstructured data sources into trusted knowledge graphs, facilitating intelligent, AI-driven smart manufacturing solutions tailored to highly-regulated environments (including GxP requirements). To find out more about our platform, visit hash.ai or contact us to learn more about how our technology and services can support your supply chain.

Roadmap To Value

A typical manufacturing line observability and vision system deployment looks like:

1. Baseline: Identify high-value inspection points; install or upgrade imaging hardware.
2. Train AI models: Use historical image data or conduct controlled defect simulation campaigns.
3. Launch pilot: Run in shadow mode with human oversight.
4. Automate alerts & reject logic: Tie into line control for auto-sorting.
5. Link to MES/QMS: Enable deviation documentation and root-cause feedback.
6. Unlock further predictive and prescriptive capabilities: Utilise the data alongside other sources to further optimise manufacturing processes.
7. Embed into digital twin: Fuse with broader quality and performance data for predictive optimization.