Operational Causal & Modal Analysis

Dr. PG Madhavan

Managing Partner – Digital Twins; Enterprise Minds 

pg@eminds.ai ; https://www.garudax.id/in/pgmad/

#Causality #Modalanalysis #coupledsignal

Abstract: Operational Causal Analysis (OCA) and Operational Modal Analysis (OMA) use the same normal operating data and shared Blind Source Separation (BSS) math to deliver two views of an asset: a causal map showing which variables drive others and where harmful feedback loops form, and a modal picture showing how the structure flexes and its key vibration modes. Together, they provide early warning of emerging problems and its root causes as well as a safe way to explore “what-if” interventions in software, without disturbing plant operations.

(1) For the first time, OCA and OMA are shown to be two sides of a coin with BSS as the underpinning method; (2) “Coupled” or “cross-talk” signal in Causal Analysis is identified and uniquely exploited for deeper analysis; (3) OCA method is applied for the first time to Industry / Engineering vertical use cases.

Operational Causal Analysis (OCA) and Operational Modal Analysis (OMA) are two complementary ways to understand what is really happening inside complex industrial assets while they are running, without having to disturb operations or inject special test signals.

From Lab Tests to Real-World Operation

In traditional Modal Analysis, engineers study how a structure behaves when it vibrates: how an airplane wing flexes, how a bridge sways, or how a refinery vessel responds to flow-induced vibration. This can be done by exciting the structure in a controlled way—using a shaker table, a wind tunnel, or impact hammers—which is often called Experimental Modal Analysis.

However, real plants rarely allow that kind of controlled testing on critical equipment. In many cases, only measurements taken during normal operation are available. When engineers extract modal information (natural frequencies, mode shapes, and damping behavior) from these “business as usual” measurements, it is called Operational Modal Analysis (OMA). OMA lets engineers see how an asset is really behaving in its true environment—under actual fluid flow, pressure, and temperature conditions.

In the world of causality, a similar distinction exists. Much of academic causal analysis assumes controlled experiments: you “intervene” on a system, change one variable, and observe the effect. That is rarely feasible in a refinery, gasifier, or compressor train that must stay online. When causality is inferred only from operating data, without interventions, this article calls it Operational Causal Analysis (OCA); it is also called Causal “Discovery”.

What OCA Does in Plain Language

In most plants today, engineers see what can be called surface signals: time series from sensors—pressures, flows, accelerations, temperatures—collected by a control system or condition monitoring system. These measurements are mixtures of many underlying influences: process disturbances, mechanical defects, control actions, fluid instabilities, structural resonances, and more.

OCA aims to answer two practical questions:

• Which variables are truly driving others, and which are just reacting.
• What underlying mechanisms or “hidden” sources are producing the patterns seen on the sensors.

To do this, OCA uses a family of mathematical techniques known as Blind Source Separation (BSS). A classic illustration is the “cocktail party problem”: multiple people talk at once, multiple microphones record a jumble of voices, and BSS methods separate out the individual speakers—without knowing in advance what they are saying. In an industrial setting, the “voices” are the hidden sources (true causes), and the microphones are the plant sensors.

Mathematically, the plant measurements can be seen as mixtures of these hidden source signals, combined through a mixing matrix. BSS techniques estimate the un-mixing matrix, which effectively “unmixes” the surface signals into their underlying sources. When framed in a causal modeling formalism, this un-mixing matrix can be converted into a causal matrix, which tells you how each variable influences others—forming a directed causal graph that encodes the data-generating mechanisms in the system.

In plant terms: OCA takes the multivariate sensor time series and returns a causal map that shows directional relationships and loops among measurement points, revealing likely root causes and feedback structures that are not obvious from raw trends alone.

How OMA Fits In—and Why It Matters

OMA starts from the same multichannel measurements but looks at them through a structural dynamics lens. Instead of asking “who causes whom,” it asks:

• What are the dominant modes of vibration or dynamic behavior.
• How do these modes’ shapes (how different locations move relative to one another) evolve over time.

In standard OMA notation, the measurements are expressed as a modal matrix multiplying modal time histories. The important insight in this work is that the BSS mixing matrix, when viewed from a structural perspective, plays the same role as the modal matrix in OMA. With appropriate scaling, the columns of the mixing matrix become mode shape vectors—the familiar objects vibration engineers use to visualize how structures flex and twist.

This leads to a powerful unification:

• The un-mixing matrix from BSS feeds OCA and gives the causal matrix (structural interrelationships in a causal sense).
• The mixing matrix from BSS feeds OMA and gives the modal matrix (structural mode shapes and dynamic behavior).

In other words, the same data and the same class of algorithms can deliver both a causal view and a modal/dynamic view of the asset.

For example, in a syngas gasifier, high-velocity flow, turbulence, acoustic pulsations from pumps and compressors, and mechanical vibrations all interact. When a mechanical issue develops, engineers often see:

• Mode frequencies shifting downward.
• New modes emerging.
• Existing modes splitting or disappearing.

OMA captures these changes in the mode shapes and frequencies, making evolving structural problems visible.

Using OCA and OMA Together in Industry

The joint use of OCA and OMA is especially attractive for industrial assets monitored by multiple sensors, such as bearings, rotating equipment, gasifiers, columns, and piping systems. A typical workflow looks like this:

• Collect multichannel time series from sensors during normal operation - no special tests, no downtime.
• Apply a BSS-based algorithm to the data.
• From the result, derive: A causal matrix for OCA, which can be visualized as a directed graph showing how variables influence each other. A modal matrix for OMA, which yields mode shapes and, under standard assumptions, natural frequencies and damping.

Because there is a direct mathematical mapping between the causal and modal matrices, engineers can move between a cause-effect explanation and a vibration/dynamics explanation using the same underlying model. That means:

• Plant subject matter experts can use the causal graph to reason about process interactions, feedback loops, and potential interventions.
• Vibration specialists can use the mode shapes to understand how the structure is flexing and where resonance or imbalance is building up.

Both perspectives arise from the same operational data and the same analytical pipeline.

A Bearing Failure Example

A NASA bearing vibration experiment illustrates how this combined OCA–OMA approach works in practice. In this test, bearings were run continuously while vibration data were collected over many days, from initial healthy operation to eventual failure of one bearing.

Analyzing this data with the OCA–OMA framework shows:

• Causal graphs that start simple on early days of testing and become much more complex by the day before failure.
• The appearance of energy feedback loops in the causal graphs, which are known early warning signs of trouble in mechanical systems.

These loops highlight problematic couplings among components—places where energy is feeding back into the system instead of dissipating, potentially driving damaging vibrations. An engineer can then explore hypothetical design or operational changes and test their effect “in software” by modifying the causal model and seeing how the graph and behavior would change, before committing to physical interventions.

On the modal side, OMA results show four mode shapes, and changes in the first and third mode shapes become apparent as the failure approaches. These mode shape changes represent how the structure’s flexing pattern shifts as imbalance and structural correlation increase.

The analysis goes one step further by looking at coupled or cross-talk signals, defined simply as the difference between surface signals and the separated source signals. In mechanical systems, unusually strong or new couplings between components are red flags. The power spectra of these cross-talk signals on the day before failure show:

• Much higher peaks—more than ten times higher than earlier days.
• More peaks overall.
• Peaks shifted to lower frequencies.

These are classic signatures of unwanted couplings and impending failure. In effect, OCA–OMA turns raw vibration data into a storyline: how the bearing system’s internal interactions and structural behavior evolve from healthy operation to failure, with clear visual indicators along the way.

Why This Matters for Industry

For industrial companies, the practical benefits of combining OCA and OMA are significant:

• Both are operational methods: they work from measurements taken while the asset is running, with no need to inject test signals or interrupt production.
• They provide complementary views: OCA reveals structural causal interrelationships - who drives whom in the data-generating process. OMA reveals dynamic behavior - how the system vibrates and how mode shapes change as conditions evolve.
• A single data collection and algorithmic workflow yields both structural (causal) and dynamic (modal) insight.

Because these methods build on advanced but well-established algorithms developed over the past two decades in signal processing, causal discovery, and modal analysis, they offer a robust foundation for industrial digital twins and diagnostic tools. In high-stakes settings where failures can cost millions and interventions are risky, OCA–OMA enables engineers to identify root causes, test “what-if” scenarios in software, and uncover optimization opportunities—all from the data plants are already collecting.

#Causality #Modalanalysis #coupledsignal