MIMO and Multi-Group Comparison Problem in Process Control: A Multivariate Statistical Framework for Systems Represented with Frequency Response Functions

summary

The gist

Characterizing complex Multi-Input Multi-Output (MIMO) systems presents two issues: feedback controllers mask fault variance, rendering single-variable monitoring ineffective…

In short

The study addresses limitations in monitoring complex MIMO systems where feedback controllers mask faults and repeated testing inflates false alarms. It proposes transforming Frequency Response Functions into time-domain Pseudo-Impulse Responses (PIRs) and using Permutational Multivariate Analysis of Variance (PERMANOVA) on PIR Supervectors to distinguish normal operation from environmental disturbances and severe parametric faults with high statistical confidence.

Key concepts

Frequency Response Function (FRF)
FRFs capture the complete dynamic signature of a system by measuring gain, phase, and resonance across various operational frequencies. They describe how the system responds dynamically to different inputs over time, providing a rich diagnostic profile beyond simple static measurements.
Pseudo-Impulse Response (PIR)
PIRs are a time-domain representation derived from FRFs. This transformation converts complex spectral components back into a tangible trajectory that preserves phase shifts, allowing the framework to capture continuous temporal inertia and dynamic structural shifts that scalar metrics often miss.
PIR Supervector
The PIR Supervector aggregates standardized PIRs from all input-output pairs into one long vector. This single metric allows for the rigorous evaluation of an entire MIMO system's state, effectively penetrating the masking effect of closed-loop controllers and providing a single statistical test result.
PERMANOVA
Permutational Multivariate Analysis of Variance (PERMANOVA) is a non-parametric statistical test used here. It evaluates the distance between different system states represented by PIR Supervectors, effectively detecting global shifts in MIMO systems without assuming data follows a normal distribution, thereby avoiding the multiple comparisons problem.

Terminology used across episodes

This episode discusses

The paper

MIMO and Multi-Group Comparison Problem in Process Control: A Multivariate Statistical Framework for Systems Represented with Frequency Response Functions · Read on arXiv

Calejo Hybrid Intelligence AB · Department of Civil and Industrial Engineering Largo Lucio Lazzarino

Characterizing complex Multi-Input Multi-Output (MIMO) systems presents two issues: feedback controllers mask fault variance, rendering single-variable monitoring ineffective, and repeated testing across multiple conditions inflates false positive rates due to the multiple comparisons problem. This study proposes a multivariate statistical framework to resolve these limitations. We extend a statistical library that identifies system dynamics via Frequency Response Functions (FRFs) by transforming them into time-domain Pseudo-Impulse Responses (PIRs). This functional representation captures the complete dynamic signature of the system, offering a richer diagnostic profile than traditional static scalar metrics. The framework, originally developed for SISO systems, is extended to the MIMO case by introducing supervectors that aggregate multiple PIRs, evaluated using Permutational Multivariate Analysis of Variance (PERMANOVA). This non-parametric approach handles the high dimensionality and complex correlation structures inherent in MIMO functional data. We demonstrate that the PIR Supervector is more powerful than single-variable analysis, and the MIMO PERMANOVA approach outperforms traditional SISO anomaly detection. By distinguishing between normal operation, external thermal disturbance effectively compensated within the investigated operating range, and severe parametric faults that standard univariate methods miss, the framework provides a single, rigorous metric for specific fault diagnosis (p < 0.001).

DOI: 10.5220/0015126600004979

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "MIMO and Multi-Group Comparison Problem in Process Control".

Jane: Here's the gist of the paper. Characterizing complex Multi-Input Multi-Output (MIMO) systems presents two issues: feedback controllers mask fault variance, rendering single-variable monitoring ineffective, and repeated testing across multiple conditions inflates false positive rates due to the multiple comparisons problem The PIR Supervector is more powerful than single-variable analysis, and the MIMO PERMANOVA approach outperforms traditional SISO anomaly detection Pruning redundant reasoning tokens This study proposes a multivariate statistical framework to resolve limitations in monitoring complex MIMO systems by transforming Frequency Response Functions (FRFs) into time-domain Pseudo-Impulse Responses (PIRs) and evaluating them using Permutational Multivariate Analysis of Variance (PERMANOVA) The framework is designed to distinguish between normal operation, external thermal disturbance effectively compensated within the investigated operating range, and severe parametric faults that standard univariate methods miss How the training signal is built The system dynamics are characterized via Frequency Response Functions (FRFs), which capture gain, phase, and resonance characteristics across varying operational frequencies These FRFs are then transformed into Pseudo-Impulse Responses (PIRs) using a transformation that translates complex-valued spectral components back into a tangible time-domain trajectory while preserving phase shifts To aggregate relevant control loops without biasing the distance metric across vastly different physical quantities, pair-wise z-score standardization is applied to create the standardized PIR, which represents the dimensionless deviation of a subject from its mean dynamic behavior The resulting PIR Supervector concatenates these standardized PIRs into a single vector of length L = M ×T Where the gains come from The PERMANOVA approach leverages permutations rather than assuming multivariate normality, accommodating the highly correlated, high-dimensional nature of functional PIR data without sacrificing statistical rigor The framework is applied to three scenarios for hypothesis testing: Control strategy validation, Fault diagnosis via library matching, and Batch process quality assurance Post-hoc pairwise comparisons using a Bonferroni correction are used to isolate catastrophic parametric failure from environmental shifts Direct Comparison with Static and Dynamic PCA Standard projection methods like baseline PCA and DPCA are limited when confronted with parametric dynamic faults because they treat time as a series of isolated frames rather than a continuous functional trajectory The Q-statistic in the benchmark comparison acts as a critical indicator for novel events that entirely rupture previously modeled multivariate correlations The PIR framework natively captures holistic temporal inertia by linearizing continuous frequency responses into standardized time-domain vectors, yielding statistically significant separations while avoiding complex neural network cascades The PIR supervector distinguishes between Normal Operating Conditions and distinct disturbance scenarios with statistical confidence (p < 0.001) The PIR supervector distinguishes between Normal Operating Conditions and distinct disturbance scenarios with statistical confidence (p < 0.001) The PIR supervector distinguishes between Normal Operating Conditions and distinct disturbance scenarios with statistical confidence (p < 0.001) By penetrating the stochastic masking of closed-loop controllers, this functional approach provides operators with a tool to differentiate between environmental drift and critical system degradation The present results should not be interpreted as an exhaustive validation across all TEP failure modes Future work will focus on leveraging spatial structures to generate Strategy Heatmaps, incorporating Principal Coordinate Analysis (PCoA) visualizations to map the distance matrices, and applying expert-defined weightings to construct a safety super vector prioritizing critical control loops based on specific industrial risk profiles The PIR-PERMANOVA framework is applied here to the Tennessee Eastman Process for the first time The foundational methodology of evaluating FRFs via PIRs was originally pioneered for human posture control REFERENCE MIMO and Multi-Group Comparison Problem in Process Control: A Multivariate Statistical Framework for Systems Represented with Frequency Response Functions Page 1 --- MIMO and Multi-Group Comparison Problem in Process Control: A Multivariate Statistical Framework for Systems Represented with Frequency Response Functions Vittorio Lippi, Leonard Johard and Gabriele Landucci Keywords: Frequency Response Function (FRF), Pseudo-Impulse Response (PIR), Multi-Input Multi-Output (MIMO) Systems, Multiple Group Comparisons, Permutational Multivariate Analysis of Variance (PERMANOVA), Tennessee Eastman Process (TEP), Multivariate Statistics, Functional Data Analysis, Fault Detection and Diagnosis, Dynamic Process Monitoring, Non-parametric Testing Abstract: Characterizing complex Multi-Input Multi-Output (MIMO) systems presents two issues: feedback controllers mask fault variance, rendering single-variable monitoring ineffective, and repeated testing across multiple conditions inflates false positive rates due to the multiple comparisons problem This study proposes a multivariate statistical framework to resolve these limitations We extend a statistical library that identifies system dynamics via Frequency Response Functions (FRFs) by transforming them into time-domain Pseudo-Impulse Responses (PIRs) This functional representation captures the complete dynamic signature of the system, offering a richer diagnostic profile than traditional static scalar metrics The framework, originally developed for SISO systems, is extended to the MIMO case by introducing supervectors that aggregate multiple PIRs, evaluated using Permutational Multivariate Analysis of Variance (PERMANOVA) We demonstrate that the PIR Supervector is more powerful than single-variable analysis, and the MIMO PERMANOVA approach outperforms traditional SISO anomaly detection By distinguishing between normal operation, external thermal disturbance effectively compensated within the investigated operating range, and severe parametric faults that standard univariate methods miss, the framework provides a single, rigorous metric for specific fault diagnosis (p < 0.001) Page 2 --- In such environments, a feedback controller’s primary function is to continuously counteract deviations Consequently, an anomaly is rarely isolated in a single variable’s amplitude; it is contained within the multivariate correlation structure While multivariate statistical process control (MSPC) tools utilizing Principal Component Analysis (PCA), Partial Least Squares (PLS), or Canonical Variate Analysis (CVA) are more appropriate for managing alarms, as demonstrated in complex systems like cement rotary kilns standard static formulations assume linearity, while non-linear extensions like Kernel PCA assume specific distributional properties unsuitable for complex spectral data and fail to capture continuous temporal shifts in underlying dynamic transfer functions Page 3 --- The reliance on arbitrary scalar metrics discards the temporal and spectral information embedded within process data In contrast, describing the system via Frequency Response Functions (FRFs) captures the complete dynamic signature, encompassing gain, phase, and resonance characteristics across varying operational frequencies A contribution of this work is demonstrating that this functional FRF-based representation provides a superior diagnostic basis compared to traditional scalars, allowing the isolation of subtle parametric faults that manifest as dynamic structural shifts We transform identified FRFs into Pseudo-Impulse Responses (PIRs), a robust time-domain representation By treating these PIRs as discrete functional units within a Permutational Multivariate Analysis of Variance (PERMANOVA) design, we provide a non-parametric statistical test capable of detecting global shifts in MIMO systems Page 4 --- For a MIMO system with Nin inputs and Nout outputs across N experimental runs, we define M = Nin × Nout distinct pairs The PIR for the i-th subject for pair m at discrete time tj is xi,m(tj), calculated via the transformation (Lippi, 2025) Mathematically, this transformation translates complex-valued spectral components back into a tangible time-domain trajectory, preserving phase shifts that scalar metrics discard To aggregate relevant control loops without biasing the distance metric across vastly different physical quantities, we apply pair-wise z-score standardization The standardized PIR, zi,m(tj), represents the dimensionless deviation of subject i from the mean dynamic behavior of pair m By enforcing this zero-mean, unit-variance scaling, we ensure that variables operating on fundamentally different physical scales contribute equitably to the final statistical evaluation The PIR Supervector Si for subject i concatenates the standardized PIRs of all M pairs into a single vector of length L = M ×T Page 5 --- The frequency grid was linearly spaced (0.001 Hz to 0.5 Hz), excluding the DC component The time grid matched the 10-hour identification experiment at 1 Hz For PERMANOVA, we utilized the Euclidean distance metric on these standardized supervectors (d 2 i j = ∑(Si,k − Sj,k) 2) This captures phase shifts and amplitude variations in linearized dynamics Page 6 --- To visually demonstrate the spatial centroid shifts confirmed by the PERMANOVA and PERMDISP tests, a Principal Coordinate Analysis (PCoA) was conducted The distinct spatial separation of Disturbance 1 (IDV 6) visually demonstrates the systemic dynamic shift detected by the PERMANOVA (p < 0.0003) Conversely, the environmental shift (Disturbance 2) remains tightly integrated with the NOC baseline, reinforcing the specificity of the proposed framework Page 7 --- The PIR supervector aggregates the masked variables (Figure 1) with corresponding reactive shifts, providing a single p-value for the entire system state without inflating the Family-Wise Error Rate The baseline PCA and DPCA models required 8 and 7 components respectively to explain 95% of the NOC variance Hotelling’s T squared demonstrates significant overlap across conditions By linearizing continuous frequency responses into standardized time-domain vectors, the PIR framework natively captures holistic temporal inertia, yielding statistically significant separations (p < 0.0001) while avoiding the complex neural network cascades required by some nonlinear PCA methods Page 8 --- The PIR supervector aggregates the masked variables (Figure 1) with corresponding reactive shifts, providing a single p-value for the entire system state without inflating the Family-Wise Error Rate The PIR supervector distinguishes between Normal Operating Conditions and distinct disturbance scenarios with statistical confidence (p < 0.001) By penetrating the stochastic masking of closed-loop controllers, this functional approach provides operators with a tool to differentiate between environmental drift and critical system degradation The present results should not be interpreted as an exhaustive validation across all TEP failure modes Future work will also focus on leveraging spatial structures to generate Strategy Heatmaps, incorporating Principal Coordinate Analysis (PCoA) visualizations to map the distance matrices, and applying expert-defined weightings to construct a safety super vector prioritizing critical control loops based on specific industrial risk profiles While this proof-of-concept demonstrates the efficacy of the PIR-PERMANOVA framework, several methodological boundaries must be explicitly defined First, the experimental implementation represents a SIMO (Single-Input, Multi-Output) special case of the generalized MIMO formulation Specifically, the PRBS excitation was applied to a single input (Nin = 1) while monitoring three outputs (Nout = 3), yielding M = 3 input-output pairs With a 10-hour experiment sampled at 1 Hz (T = 36000 samples), the resulting supervector length is L = 108,000 REFERENCE Anderson, M.

Tom: J.

Jane: (2001).

Tom: A new method for non-parametric multivariate analysis of variance https://doi.org/https://doi.org/10.1111/j-australecology-26-i-32–46 Page 9 --- Ljung, L.

Title and authors: Tom: Let’s talk about who wrote this. Vittorio Lippi, Leonard Johard, and Gabriele Landucci are the authors of "MIMO and Multi-Group Comparison Problem in Process Control: A Multivariate Statistical Framework for Systems Represented with Frequency Response Functions". They come from Calejo Hybrid Intelligence AB and the Department of Civil and Industrial Engineering at Pisa University.

Jane: It’s interesting they are tackling a problem originally developed for SISO systems—Single-Input, Single-Output—and extending it to MIMO, which is much more realistic for real industrial processes.

Lu: The motivation mentioned on page two is that traditional methods relying on just scalar metrics discard the temporal and spectral information embedded in process data <ref:2609.12904#pg1>. They say that describing the system via Frequency Response Functions captures the complete dynamic signature, including gain, phase, and resonance characteristics across varying operational frequencies.

Meng: That sounds like a lot of data to process. If we’re looking at something that changes over time and frequency, how does this framework actually handle the computational cost?

Lalam: The paper is focused on demonstrating that this functional FRF representation gives a superior diagnostic basis compared to traditional scalars, specifically for isolating subtle parametric faults that show up as dynamic structural shifts.

The paper's summary: Tom: So what’s the main mechanism they are proposing? It starts with taking the FRFs and transforming them into these time-domain Pseudo-Impulse Responses, or PIRs, which is a way to translate those complex spectral components back into a tangible time-domain trajectory while keeping the phase shifts intact.

Jane: That translation step is important because scalar metrics just lose that crucial temporal information, and then they standardize these PIRs using pair-wise z-score standardization so you can compare variables operating on totally different physical scales fairly.

Lu: They build this PIR Supervector by concatenating all those standardized PIRs into one long vector of length L equals M times T, where M is the number of input output pairs and T is the time steps. This whole thing is then tested using PERMANOVA, which doesn't even assume multivariate normality, so it’s robust for highly correlated data.

Meng: So this means we’re creating this massive single vector to test against, instead of running dozens of separate tests across every input output pair and every condition? That addresses the multiple comparisons problem directly.

Lalam: Right. They apply this PIR-PERMANOVA framework to three scenarios: validating control strategies, diagnosing faults by matching them against a library, and checking batch quality assurance against a "Golden Batch" reference.

The paper's improvements: Tom: The authors are showing that this MIMO PERMANOVA approach outperforms traditional SISO anomaly detection methods. They specifically highlight that the PIR Supervector is more powerful than just using single-variable analysis alone.

Jane: They also compare it directly against standard projection methods like baseline PCA and DPCA, pointing out that those methods treat time as isolated frames instead of a continuous functional trajectory, which is why the Q-statistic in their benchmark comparison acts as an indicator for novel events that completely rupture previously modeled multivariate correlations.

Lu: The PIR framework natively captures holistic temporal inertia by linearizing continuous frequency responses into standardized time-domain vectors, which yields statistically significant separations with a p less than zero point zero one zero one.

Meng: That p value is pretty strong, but I have to ask about the implementation details—the paper notes that this proof-of-concept is actually a SIMO special case where they used one input and three outputs, yielding M equals three input output pairs. How does that scale up to a real system with, say, twenty inputs?

Lalam: They mention the experimental setup uses one hundred eight thousand samples for the supervector length L when they had one hundred hours of experiment sampled at one hertz. It’s a big number for what they managed to capture in this initial test.

Conclusion: Tom: So, to wrap up, the key finding is that the PIR supervector distinguishes between Normal Operating Conditions and distinct disturbance scenarios with statistical confidence of p less than zero point zero one zero one. This is a single metric for a whole system state.

Jane: It gives operators a tool to differentiate between environmental drift and critical system degradation by penetrating the stochastic masking that closed-loop controllers cause, which is really powerful because it’s functional rather than just looking at amplitude.

Lu: They conclude that this framework provides a single rigorous metric for specific fault diagnosis, and they acknowledge their limitation is that this proof-of-concept was actually a SIMO special case, meaning it might not be an exhaustive validation across all TEP failure modes.

Meng: I agree with Lu on the limitations. The authors specifically state that future work will focus on leveraging spatial structures to generate Strategy Heatmaps and applying expert-defined weightings to prioritize critical control loops based on specific industrial risk profiles, which is where deployment challenges usually show up.

Lalam: They are moving toward using Principal Coordinate Analysis for visualizing the distance matrices, which will help operators see exactly where these dynamic shifts are happening spatially across the system before making a decision.

Tom: So, this PIR-PERMANOVA framework on the Tennessee Eastman Process is a solid way to get past those limitations in monitoring complex MIMO systems by using functional data representation instead of just isolated variables.

More episodes

← Home