I spent three weeks in an automotive plant resolving a dimensional deviation on a gearbox housing. The engineering team insisted the problem was the tool. Then the material. Then the press machine. Every department had a theory, every expert had a favourite cause, and every one of them was wrong. They were searching for a single failure point where a system of interconnected factors was actually at work.
We mapped a network of relationships between mould temperature, granulate moisture, injection speed, holding pressure, cooling circuit settings, and operator capability. The whole picture changed. The defect was not located in one parameter. It emerged from the dynamics of the network itself, and the solution was hiding in those same dynamics.
This is Quality Network Analysis (QNA). It is a methodology grounded in graph theory that maps the relationships between inputs, process variables, human factors, and outputs. In modern manufacturing, where IATF 16949 and AS9100 environments push process complexity far beyond what linear tools can handle, QNA reveals the hidden dependencies that traditional problem-solving overlooks.
The limits of linear root-cause tools
I have implemented and transitioned ISO 9001 systems across automotive and aerospace plants for over twenty years. In that time I have relied on PFMEA, Ishikawa diagrams, 5 Whys, and 8D. They all have a critical place in quality engineering, but they share a structural blind spot: they assume causes are relatively isolated and linear. They present a neat hierarchy of inputs leading to a single output.
Manufacturing reality is different. Increased dryer temperature does not just reduce material moisture. It alters flow properties, which shifts injection parameters, which changes internal stress, which degrades dimensional stability after cooling. The chain of causality is longer and more branched than any fishbone diagram can capture. When you rely solely on linear tools, you risk optimising one node while destabilising three others.
Nonlinear interactions are the rule, not the exception. Anyone who has run a Design of Experiments (DOE) knows that interaction effects between factors are frequently more significant than main effects. Yet traditional problem-solving frameworks ignore these interactions until they are proven, which usually only happens after multiple failed attempts at a containment action.
Building the relationship matrix
Applying QNA begins with identifying every relevant node in your process. Nodes are the variables in your system. They include input variables like temperature and pressure, machine settings, human factors like shift rotation, measurement characteristics taken from your PPAP documentation, environmental factors, and system factors like maintenance schedules and foundry batch changes. During one pressing operation analysis, we identified forty-seven distinct nodes. In complex manufacturing, that is a conservative estimate.
You then map the relationships by creating an adjacency matrix. For each pair of nodes, you ask if node A affects node B, in which direction, and with what strength. You populate this matrix using expert assessment from process engineers and operators, correlation analysis of historical SPC data, results from previous DOEs, and your existing PFMEA database. Your PFMEA already contains implicit relationships; they simply need to be extracted and structured.
The QNA implementation sequence
- 01Node identificationList all process, material, environmental, and human variables.
- 02Relationship mappingBuild the adjacency matrix defining direction and strength.
- 03Network visualisationMap the structure to find clusters and central nodes.
- 04Data verificationTest expert assumptions against statistical correlations.
- 05Targeted interventionAct on high-centrality nodes to maximise systemic effect.
Network structure and node centrality
Once the data is mapped and visualised, previously invisible patterns emerge. You will find star nodes, which are factors connected to many others. These are your strategic control points. A shift in a star node propagates through the entire network, altering process conditions downstream. Ignoring a star node during process control guarantees ongoing instability.
You will also identify high-centrality nodes. These are factors through which the most influence paths travel. They might not have the largest direct impact on the final dimension, but they control the flow of influence across the network. Optimising a high-centrality node stabilises the surrounding variables, dampening process oscillation.
Clusters are tightly interconnected groups of factors that function as super-nodes. When you alter one factor inside a cluster, you alter the entire group. This structural insight is vital for solution design. If your intervention triggers an entire cluster of negative side effects, your solution will fail in production. You also find isolated nodes. These variables barely affect the system and can be safely deprioritised to save engineering resources.
Locating the systemic intervention point
QNA fundamentally changes how you approach containment. You stop looking for a single root cause and start looking for key intervention points where a single action yields the greatest systemic effect. In the gearbox housing case I mentioned earlier, traditional analysis pointed to the mould tool. Network analysis revealed that the actual intervention point was the mould cooling circuit.
You are not looking for a single root cause. You are looking for the intervention point with the greatest systemic effect.
The cooling circuit did not have the largest direct impact on the dimension, but it possessed high centrality in the network. When we optimised the cooling water flow, the mould temperature profile stabilised. That stability cascaded through the network, stabilising viscosity, cavity filling, and ultimately the housing dimensions. One targeted process change produced a systemic effect.
Traditional analysis vs. network analysis
Traditional root-cause approach
- Assumes a single, isolated defect trigger
- Relies on linear progression from cause to effect
- Optimises specific parameters in isolation
- Often results in recurring chronic defects
Quality Network Analysis
- Identifies systemic interaction points
- Maps feedback loops and variable dependencies
- Optimises high-centrality nodes for wider stability
- Targets the structural dynamics of the process
Validating expert intuition with data
Modern factories generate enormous volumes of operational data, but that data only becomes useful when structured by a framework. QNA provides that framework. You use Pearson and Spearman correlations to determine if two factors are related. Partial correlation reveals whether two factors are related independently of a third variable. For temporal relationships, Granger causality can reveal how shifts in one parameter predict deviations downstream.
In one project, we used historical data extracted directly from the MES to verify the expert-generated network. Statistical analysis supported roughly seventy percent of the relationships the engineers had identified. The remaining thirty percent were either unsupported by data, or the relationships operated in the opposite direction the experts assumed. This is the core value of QNA: it forces you to verify engineering intuition against operational reality.
You do not need a PhD in statistics to execute this. You need a disciplined process, access to your historical SPC data, and the willingness to treat manufacturing problems as complex systems. When the engineering team accepts that their mental model of the process might be structurally incomplete, the data will guide them to the true intervention points.
When to deploy QNA on the shop floor
Quality Network Analysis is not for every problem. If a parameter drifts out of specification, 5 Whys is sufficient. But you must deploy QNA when a chronic defect keeps returning after it has been officially closed in your 8D system. That recurrence is a signal that your team is treating a symptom, not the underlying systemic cause.
If multiple departments have conflicting theories about a defect, that is another trigger. It means each department sees only a fraction of the process. QNA forces these isolated observations into a single, unified map. When production, maintenance, and quality engineering all look at the same network map, the artificial boundaries between their responsibilities disappear.
A factory producing automotive connectors faced intermittent cracking on castings after storage. The material lab, process engineers, and maintenance all confirmed their specific domains were within specification. But QNA mapped a cluster of four nodes: granulate moisture, storage time, barrel jacket temperature, and injection speed. Individually, each was within tolerance. Operating simultaneously at their upper limits, they created microscopic porosity that led to cracking.
The solution was systemic. We tightened the tolerance windows on those four parameters simultaneously and implemented a dynamic control rule. If any one parameter approached its upper limit, the system automatically adjusted the others. This networked approach eliminated a defect that was costing the plant €180,000 annually in warranty claims.