Article | November 11, 2026
How AI vision and edge intelligence transform cycle process variability into actionable factory insights
By Russell Duggan-Rees, Chief Technologist, Global Smart Manufacturing, DXC Technology
Modern manufacturing organizations have spent decades optimizing automation, robotics and process control. Yet one of the most persistent operational challenges remains surprisingly difficult to answer: Why do cycle times change when nothing appears to have changed?
Production managers frequently experience situations where throughput declines, bottlenecks emerge, quality fluctuates or capacity forecasts become unreliable. Individual robots may still function normally, PLC alarms may not trigger and maintenance systems may report no issues. Yet production performance deteriorates.
The reality is that many factories operate with only partial visibility into what is physically happening on the shop floor. Minor pauses, motion drift, synchronization delays between robots, mechanical wear, environmental changes and process variability can all influence cycle performance long before traditional systems detect a problem.
Cycle Performance Analytics (CPA) was developed to address this challenge. By combining AI vision, edge computing and advanced machine learning, CPA creates a new layer of operational intelligence capable of revealing inefficient and unstable manufacturing behavior that is often invisible to conventional monitoring systems. The platform originated from highly automated automotive production environments where even small timing deviations can have significant business impact.
Manufacturing excellence depends on consistency. A production line is designed around expected cycle times, synchronized robotic movement, predictable transfer operations and tightly coordinated process flows. When those assumptions hold, throughput, quality and resource utilization remain stable.
However, reality is rarely so predictable. Cycle times and movements frequently vary between stations, between shifts and even between individual production cycles. These variations may be caused by: robot motion drift; minor equipment degradation; process instability; waiting conditions between stations; product variant complexity; area-clear or safety interlocks; material flow interruptions; or changes in environmental conditions.
The challenge is that many of these factors leave little or no trace in traditional operational systems. Manufacturers often know that a cycle became slower, but they do not know:
As production estates become increasingly automated and geographically distributed, this visibility gap becomes even more significant. Variability introduced at one robot cell can propagate across an entire factory and potentially across multiple plants.
The result is a familiar operational reality: Teams spend considerable time investigating symptoms but gain only limited evidence regarding root cause.
Solving this problem is considerably more difficult than it first appears.
The traditional approach would be to collect data directly from robots, PLCs, SCADA systems or industrial control platforms. Unfortunately, that is often not practical. Manufacturing environments frequently contain:
Even when data is available, the information captured often describes controller states rather than actual physical behavior.
A robot controller may indicate a process step completed successfully, yet the physical movement may have slowed, paused, oscillated or deviated significantly from normal operation. This creates a fundamental visibility gap.
To gain deeper insight, a different form of data must be collected: external observation data.
Video provides a uniquely valuable source of information because it captures what is physically happening regardless of vendor, controller type or machine age. A camera can observe robot motion, workpiece movement, human interaction, material flow, equipment behavior and environmental conditions. However, video introduces its own challenges. Unlike traditional industrial analytics projects, there is usually very little labeled training data and no predefined anomaly catalog. There is also significant variation between "good" cycles, multiple product variants, changing lighting and environmental conditions, and continuous process evolution.
At project initiation, there is rarely an obvious machine learning model suited to the problem. This then becomes both an engineering and research issue: How can a system learn what normal operation looks like when normal itself contains substantial variability?
To address this challenge, CPA begins by creating a reliable observation layer independent of existing factory systems.
Video streams are collected from strategically positioned industrial cameras and processed through DXC's SPARK architecture on AWS and Edge AI platform running close to the cameras. This approach allows analytics to be performed close to the production process while maintaining centralized management, governance and deployment capabilities.
The solution delivers several important advantages:
Once data collection is established, machine learning models can begin learning the operational behavior of the production process. Importantly, CPA does not assume that a single AI technique will solve every manufacturing challenge.
Manufacturing processes contain different forms of variability, requiring different analytical approaches. As AI research evolves, new techniques continuously emerge, improve and become practical for industrial deployment.
For this reason, CPA evolved through multiple complementary methodologies rather than a single algorithm.
The first approach establishes a reference understanding of what a healthy production cycle looks like.
Rather than relying on programmable logic controller (PLC) states or manually defined rules, CPA learns directly from observed video of successful operations. A known-good cycle becomes the baseline against which future cycles are compared.
By transforming video frames into machine-readable motion signatures, CPA creates a temporal representation of robot behavior. Live production cycles can then be aligned against this reference model. This enables:
Automatic cycle detection
Variant identification
Cycle phase tracking
Accurate cycle time measurement
Pause detection
Motion drift identification
The result is visibility into how production actually behaves rather than how controllers report that it behaves.
Manufacturing processes rarely execute at identical speeds. A robot may perform exactly the same task but complete it slightly faster or slower depending on upstream conditions, material characteristics or process interactions.
Traditional comparisons often struggle with these timing differences. CPA addresses this using Dynamic Time Warping, an alignment technique that compares process behavior even when execution speeds vary. DTW enables CPA to:
Compare current cycles against historical references
Detect timing drift
Identify slowdowns and pauses
Reveal synchronization problems
Highlight affected cycle phases
Rather than simply reporting that a cycle took longer, DTW helps identify precisely where and when the timing changed. This significantly improves root-cause investigations and supports earlier intervention before productivity losses become systemic.
The most advanced CPA research explores a fundamentally different question: Can the system predict what should happen next and detect when reality differs from expectation?
Drawing inspiration from emerging world-model research, this approach trains AI systems on healthy operational behavior and asks them to anticipate future visual states.
Instead of detecting predefined faults, the model learns how a process normally unfolds. Each incoming video frame is compared against the system's prediction, generating evidence of unexpected behavior. Spatial analysis identifies where deviations occur, while temporal analysis tracks how those deviations develop. This approach represents an important evolution:
From rule-based monitoring to learned behavior
From threshold detection to expectation modeling
From identifying known faults to discovering previously unknown anomalies
In practical terms, it moves industrial monitoring closer to the concept of a factory-floor coach: a system that understands healthy movement and recognizes when something appears unusual, even before formal failure conditions exist.
CPA has already demonstrated the ability to expose operational behaviors that are difficult to observe using conventional manufacturing systems alone.
Current findings indicate that AI vision can successfully:
The solution creates a bridge between traditional operational data and physical factory reality, allowing engineers to see not only that performance changed, but also understand why it changed.
The long-term opportunity extends far beyond individual robot cells. By deploying CPA across an entire manufacturing estate, organizations can establish a common operational intelligence layer that continuously observes, learns and evaluates production behavior. Such a platform could provide:
Ultimately, CPA represents a shift from reactive manufacturing analytics toward proactive operational intelligence. Rather than waiting for failures, alarms, or quality issues to reveal themselves, manufacturers gain the ability to continuously observe production reality and detect subtle signs of change before they become significant business problems.
In a world where competitive advantage increasingly depends on operational precision, CPA provides manufacturers with the capability to transform video into insight, insight into understanding and understanding into measurable performance improvement.
Russell Duggan-Rees is an accomplished Chief Technologist specializing in digital transformations within the manufacturing sector, boasting over 25 years of experience in the technology industry. At DXC, Russell spearheads the Smart Manufacturing initiative, where he crafts solutions and services like Smart Factory, leveraging his deep understanding of manufacturing, Industry 4.0 and legacy systems to furnish DXC clients with a tangible competitive edge. Connect with Russell at LinkedIn.
