⚡ electrical engineering

Topological Data Analysis for High-Dimensional Dynamic Process Monitoring

This paper proposes a novel process monitoring framework that integrates topological data analysis with neural ordinary differential equations to effectively detect diverse events in high-dimensional industrial time-series data by modeling the dynamic evolution of topological structures, outperforming both reconstruction-based and trajectory-based alternatives.

Original authors: Angan Mukherjee, Tyler A. Soderstrom, Michael J. Kurtz, Victor M. Zavala

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Angan Mukherjee, Tyler A. Soderstrom, Michael J. Kurtz, Victor M. Zavala

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a symphony orchestra, but instead of hearing the music, you are staring at a spreadsheet with thousands of numbers changing every second. This is what monitoring a massive chemical plant feels like. Hundreds of sensors (temperature, pressure, flow rates) are constantly shouting data, and the challenge is figuring out if the orchestra is playing a beautiful song (normal operation) or if a violinist is about to break their string (a fault or accident).

This paper presents a new way to listen to that "data orchestra" by combining two powerful tools: Topological Data Analysis (TDA) and Neural Ordinary Differential Equations (NODEs).

Here is how their method works, broken down into simple concepts:

1. The Problem: The "Noise" of Normal Life

Chemical plants are messy. They have high noise, complex connections, and things happen on different time scales.

Old Methods (The "Photo" Approach): Traditional tools like PCA or Autoencoders are like taking a single photo of the orchestra. They look at the data and ask, "Does this picture look like a normal picture?" If the picture is blurry or weird, they sound an alarm.
- The Flaw: This works great for big disasters (like the whole stage collapsing), but it often misses subtle problems, like a violinist playing slightly out of tune. The "noise" of the plant hides these small errors.
Other Methods (The "Video" Approach): Some newer tools try to watch the video of the orchestra to see how the music moves over time. However, even these struggle to catch the small, subtle glitches before they become big problems.

2. The New Solution: Turning Data into a Shape

The authors propose a different way to look at the data. Instead of treating the numbers as a list or a photo, they turn the data into a shape (a "manifold").

The Analogy: Imagine you take the last 24 hours of sensor data and lay it out on a giant 2D grid. One side of the grid is the different sensors, and the other side is time. The height of the surface at any point represents the value of that sensor.
- If the plant is running smoothly, this shape looks like a calm, rolling hill.
- If something goes wrong, the shape might develop a hole, a spike, or a weird twist.

3. The "Shape Shifter" (Topological Descriptors)

Once they have this 3D shape, they don't just look at it; they measure its topology.

The Analogy: Think of a coffee mug and a donut. To a topologist, they are the same thing because they both have exactly one hole.
The authors use a mathematical tool called Persistent Homology to count the "holes" and "connected pieces" in their data shape as they slowly change the "water level" (a threshold) rising over the shape.
This produces a simple line graph called an Euler Characteristic (EC) curve. This curve is a compact summary of the shape's structure. It tells you, "At this moment, the data shape has 3 holes and 5 islands."

4. The "Future Predictor" (NODE)

Here is the clever part. The authors know that the shape of the data changes over time.

They feed these EC curves (the shape summaries) into a Neural Ordinary Differential Equation (NODE).
The Analogy: Imagine you are watching a movie of a balloon inflating. A standard camera just shows you the balloon getting bigger. The NODE is like a super-smart observer who learns the rules of how the balloon inflates. It learns the "flow" of the shape.
The model learns what the "normal flow" of the shape looks like. It predicts: "If the plant is healthy, the shape should change this way in the next second."

5. Catching the Glitches

The magic happens when the prediction goes wrong.

If the actual shape of the data suddenly changes direction or speed in a way the model didn't expect, the "error" (the difference between the prediction and reality) spikes.
The Result: This spike is the alarm.
- Big Events: The model catches major shutdowns and startups easily (like a stage collapse).
- Small Events: Crucially, this method also catches minor events (like a slight temperature fluctuation or a rate cut) that the old "photo" methods missed. Even if the shape looks normal at a single glance, the direction it is moving is wrong, and the NODE catches that immediately.

6. The Proof

The team tested this on real data from an industrial olefins plant (which makes plastics).

They compared their method against the old "photo" methods (PCA, Autoencoders) and the "video" methods (Koopman Autoencoders).
The Outcome: The old methods mostly only screamed when the plant was about to shut down completely. They missed the small, subtle warnings.
The new TDA-NODE method successfully identified both the big disasters and the tiny, subtle instabilities.
Even more impressively, they trained the model on one set of data and tested it on a completely different time period with different patterns. The model still worked, proving it learned the rules of the plant's behavior, not just memorized the specific data.

Summary

In short, instead of just looking at a snapshot of the data to see if it looks weird, this new method:

Turns the data into a shape.
Counts the holes in that shape to get a simple summary.
Uses an AI to learn how that shape normally moves and evolves.
Sounds an alarm the moment the shape starts moving in an unexpected direction, catching problems early before they become catastrophes.

Technical Summary: Topological Data Analysis for High-Dimensional Dynamic Process Monitoring

Problem Statement
Real-time monitoring of chemical processes requires extracting actionable information from high-dimensional, multivariate time-series data. These datasets, often comprising hundreds of variables, are characterized by strong coupling, nonlinearity, multi-timescale dynamics, and high measurement noise. Detecting both major events (e.g., shutdowns, startups) and subtle, minor instabilities (e.g., rate cuts, temperature fluctuations) is challenging. Existing methods generally fall into two categories:

Reconstruction-based approaches (e.g., Principal Component Analysis [PCA], Autoencoders [AE]): These compress data into a latent space and detect faults via reconstruction errors. While effective for major, persistent disturbances, they often struggle to detect subtle, dynamically evolving instabilities embedded in noisy data.
Trajectory-based approaches (e.g., Koopman Autoencoders [KAE]): These exploit the temporal evolution of latent features. While they offer a dynamic perspective, benchmarks in this work suggest they may still lack sensitivity to minor, transient events compared to major regime shifts.

A fundamental limitation across these methods is the representation of data. Traditional approaches treat time-series as vectors or matrices, potentially overlooking the geometric and topological structures inherent in the data.

Methodology
The authors propose a hybrid framework combining Topological Data Analysis (TDA) and Neural Ordinary Differential Equations (NODE) to address these limitations. The methodology proceeds in three stages:

Manifold Representation:
Multivariate time-series data within a moving window are transformed into a 2D manifold. The process variables are normalized and arranged such that one axis represents variable indices and the other represents time steps. This creates a scalar field (heatmap) where pixel intensity reflects normalized process values, capturing spatiotemporal patterns.
Topological Descriptor Extraction:
Using sublevel-set filtration on the manifold representation, the authors compute the Euler Characteristic (EC) curve for each window. The EC curve, defined as $\chi_k(\ell) = \beta_0(\ell) - \beta_1(\ell)$ (where $\beta_0$ and $\beta_1$ are Betti numbers representing connected components and holes, respectively), serves as a compact topological descriptor summarizing the shape and structure of the data at various filtration thresholds. This sequence of EC curves, $\{\chi_k\}$ , forms a time-resolved topological summary of the process.
Learning Temporal Dynamics with NODE:
Instead of analyzing static EC curves, the framework models their temporal evolution. A Neural Ordinary Differential Equation (NODE) is trained to learn the continuous-time dynamics of the topological state. The NODE takes the EC curve of a window as input and predicts its instantaneous time derivative ( $\dot{\chi}$ ).
- Event Detection: Event signatures are inferred from the norm of the predicted derivative, $\|\dot{\chi}_k\|$ . Sharp increases in this norm indicate abrupt changes in the topological structure of the process, signaling the onset of an event.

Key Contributions

Novel Framework: The integration of TDA (specifically persistent homology via EC curves) with NODEs to model the dynamic evolution of topological structures in process data.
Trajectory-Based Sensitivity: Unlike reconstruction-based methods that rely on static mismatch, this approach explicitly models the directionality of topological changes, allowing for the detection of subtle instabilities that do not significantly alter the static state but do alter the dynamic trajectory.
Robustness to Noise and Permutation: The EC descriptors exhibit robustness to noise and variable reordering, as they encode global connectivity statistics rather than local adjacency.
Benchmarking: A comprehensive comparison against PCA, AE, and KAE using real industrial data.

Results
The framework was evaluated on data from an industrial olefins plant involving 56 variables over a 63-day period. Six representative events were analyzed, including major events (shutdown, startup) and minor events (coldbox temperature excursions, rate cuts, ramp rates).

Reconstruction-Based (PCA/AE): Successfully detected major events (shutdown/startup) but failed to produce distinguishable signatures for minor instabilities.
Trajectory-Based (KAE): Improved upon static methods by capturing dynamic changes but remained largely insensitive to minor events, showing clear excursions only for major transitions.
TDA-NODE:
- Training: The framework detected both major and minor events. While static EC curves for minor events overlapped with the normal operating region in PCA projections, the temporal derivatives ( $\|\dot{\chi}_k\|$ ) showed distinct peaks corresponding to the onset of minor instabilities.
- Testing: The model was tested on an independent dataset (62 days) with different transient patterns. Despite the lack of retraining, the pretrained NODE successfully identified instability regions through peaks in the predicted topological dynamics, demonstrating generalizability to unseen operating conditions.

Significance and Claims
The paper claims that the proposed TDA-NODE framework provides a scalable, interpretable, and robust approach for process monitoring. Its primary significance lies in its ability to detect subtle, minor instabilities that are often missed by conventional reconstruction-based or standard trajectory-based methods. By shifting the focus from static reconstruction errors to the directionality of topological evolution, the method captures dynamic signatures of process instability directly from high-dimensional data. The authors emphasize that this approach offers a unified trajectory-based monitoring framework capable of handling both major regime shifts and minor, transient deviations without the need for high-fidelity mechanistic models.