03. Uncertainty-Calibrated Graph Structure Learning for Robust Industrial Fault Diagnosis Under Missing, Noisy and Shifted Sensor Data

Executive Summary Working Title: Variance-Attenuated Latent Graph Structure Learning for Sensor Fault Diagnosis

Industrial cyber-physical systems—such as chemical reactors, smart power grids, and water treatment plants—are monitored by dozens of heterogeneous sensors whose physical interactions are rarely documented as an explicit adjacency matrix. Graph Structure Learning (GSL) methods (such as MTGNN and GDN) infer a latent directed sensor graph $\mathbf{A}$ end-to-end and apply Graph Attention (Graph Attention Networks) to diagnose faults. However, a major vulnerability surfaces in field deployments: message passing is a double-edged sword. When $10\text{–}20\%$ of sensors suffer from packet loss, calibration drift, or high-frequency noise, standard GNN layers propagate that corruption to every connected neighbor within $L$ hops, triggering false alarms across healthy nodes and ruining fault root-cause localization. In this topic, I develop Uncertainty-Calibrated Graph Structure Learning (UC-GSL), which dynamically re-weights latent edges using node-wise uncertainty estimates and produces calibrated conformal fault sets under missing and shifted sensor regimes.

Tags: #Graph Structure Learning #GATv2 #Fault Diagnosis #Sensor Networks #Missing Data #Conformal Prediction

Research Motivation & Feasibility

Strong low-compute alternative that directly leverages my graph-learning and smart-grid/feature-correlation background (ICCIT 2024, ICEEICT 2024). Multivariate sensor datasets train in under 30 minutes on a single mid-range GPU, allowing rapid mathematical iteration.

Core Identified Research Gap

Graph Structure Learning (GSL) models for multivariate sensor anomaly and fault diagnosis (e.g., GDN, Deng & Hooi, AAAI 2021; MTGNN, Wu et al., KDD 2020) learn a deterministic top-k adjacency matrix from clean training data. When a subset of sensors drops out, freezes, or drifts under a new operating regime, standard GNN message passing blindly aggregates and amplifies those corrupted sensor readings across L-hop neighborhoods—causing healthy subsystems to be misclassified as faulty.

Basic Information

Field	Details
Core Focus	Top Low-Compute & Graph Learning Direction
Final Research Title	Uncertainty-Calibrated Graph Structure Learning for Robust Industrial Fault Diagnosis Under Missing, Noisy and Shifted Sensor Data
Working Title	Variance-Attenuated Latent Graph Structure Learning for Robust Fault Localization
Research Area	Geometric Deep Learning, Time-Series Telemetry & Reliable AI
Subfield	Graph Structure Learning (GSL) & Spatio-Temporal Fault Diagnosis
Target Academic Level	MSc Thesis / PhD Track (Ultra-Low Compute Footprint)
Current Status	Validated Architecture & Datasets

Idea Summary

2.1 One-Sentence Core Thesis

Learn a probabilistic latent sensor graph where message-passing edge weights $\tilde{\alpha}_{ij}^{(t)}$ are dynamically damped by neighbor imputation uncertainty and operating-regime shift scores, preventing corrupted sensors from contaminating root-cause fault diagnosis.

2.2 Research Motivation & Strategic Fit

Strong Alignment With My Prior Publications: My published papers on Smart Grid Stability Prediction (ICCIT 2024), Correlation-Based Feature Extraction (ICCIT 2024), and Multi-Modal Protein Classification (ICEEICT 2024) give me a head start in correlation structure and relational modeling.
Unbeatable Compute Efficiency: Unlike Vision-Language Models or 14B LLMs, tabular/time-series sensor graphs have $N \in [50, 130]$ nodes. A full training run on Tennessee Eastman Process (TEP) or SWaT takes $12\text{–}20$ minutes on a single GPU, making extensive 10-seed ablation studies effortless.

Problem Exploration & Identified Research Gap

3.1 The "Contagious Corruption" Problem in Sensor GNNs

In Graph Deviation Network (GDN, Deng & Hooi, 2021), each sensor $i$ has a static learned embedding $\mathbf{v}_i \in \mathbb{R}^d$, and directed edges are selected via top-$k$ cosine similarity $e_{ij} = \mathbf{v}_i^\top \mathbf{v}_j / (\|\mathbf{v}_i\|\|\mathbf{v}_j\|)$. During inference:

1
If sensor $j$ experiences a hardware glitch (e.g., stuck-at-zero, spike noise, or $30\%$ missingness imputed via zero/mean fill), its temporal feature vector $\mathbf{x}_j^{(t)}$ becomes highly corrupted.
2
Because the graph topology $\mathcal{N}_i$ and attention weights rely on static node embeddings $\mathbf{v}_i$ (or uncalibrated GAT scores), node $i$ aggregates $\mathbf{x}_j^{(t)}$ with high weight, causing node $i$'s prediction residual $|x_i^{(t)} - \hat{x}_i^{(t)}|$ to spike.
3
Consequently, the fault localizer blames node $i$ (a healthy downstream sensor) instead of isolating node $j$, and overall fault classification ECE degrades sharply under load-regime shifts.

3.2 Explicit Research Gap

Existing GNN fault-diagnosis methods either assume complete, clean sensor streams or apply a separate pre-imputation module whose imputation uncertainty is discarded before graph message passing. *There is a clear gap for an end-to-end Graph Structure Learning framework that explicitly propagates observation and imputation variance through dynamic GATv2 edges—automatically severing or damping edges from unreliable sensors during inference while calibrating node-level fault probabilities via conformal prediction.*

Proposed Research Direction

4.1 Core Hypothesis

Hypothesis: Coupling a distributional temporal encoder $(\mu_j^{(t)}, \sigma_j^{2(t)})$ per sensor with a variance-penalized GATv2 structure learner will cut false-positive root-cause localization errors by $>40\%$ under $25\%$ random/block sensor dropout and operating-mode shifts on TEP and SWaT.

4.2 Mathematical Formulation

Let $\mathbf{X}^{(t-W:t)} \in \mathbb{R}^{N \times W}$ be the sliding window of $N$ sensors over $W$ time steps, accompanied by a binary observation mask $\mathbf{M}^{(t-W:t)} \in \{0, 1\}^{N \times W}$.

1
Uncertainty-Aware Temporal Node Encoder: Each sensor $i$ maps its masked window to a Gaussian latent state $\mathcal{N}(\mathbf{h}_i^{(t)}, \text{diag}(\boldsymbol{\sigma}_i^{2(t)}))$, where missing or spike-corrupted windows naturally induce high variance $\|\boldsymbol{\sigma}_i^{2(t)}\|_1$.
2
Variance-Damped Dynamic Graph Attention (UC-GATv2): Combining static structural affinity $\mathbf{v}_i^\top \mathbf{v}_j$ with dynamic GATv2 features (Graph Attention Networks) and a variance penalty $\gamma > 0$:

$$ e_{ij}^{(t)} = \mathbf{a}^\top \text{LeakyReLU}\Big(\mathbf{W}_l \mathbf{h}_i^{(t)} + \mathbf{W}_r \mathbf{h}_j^{(t)}\Big) + \beta \frac{\mathbf{v}_i^\top \mathbf{v}_j}{\|\mathbf{v}_i\|\|\mathbf{v}_j\|} - \gamma \big\|\boldsymbol{\sigma}_j^{2(t)}\big\|_1 $$

$$ \tilde{\alpha}_{ij}^{(t)} = \frac{\exp(e_{ij}^{(t)})}{\sum_{u \in \mathcal{N}_i^{\text{top-}k}} \exp(e_{iu}^{(t)})} $$

When sensor $j$ drops out or produces erratic out-of-range spikes, $\|\boldsymbol{\sigma}_j^{2(t)}\|_1$ increases, exponentially driving $\tilde{\alpha}_{ij}^{(t)} \to 0$ so neighbor $i$ relies on self-history and remaining clean sensors!

Interactive Diagram SpecificationMermaid Flow / Schedule

flowchart LR
    S["Multivariate Sensor Stream X_t<br/>+ Missingness Mask M_t"] --> ENC["Temporal Gaussian Encoder<br/>Outputs (h_i^(t), σ²_i^(t))"]
    ENC --> GSL["Variance-Damped GSL & GATv2<br/>Penalizes Edges from High-σ² Nodes"]
    GSL --> HEAD["Fault Classifier & Node Forecaster"]
    HEAD --> CONF["Conformal Fault Set C(X_t)<br/>Root-Cause Localization"]

Preliminary Technical Plan

Datasets:

1
Tennessee Eastman Process (TEP): 52 sensors/actuators, 21 fault classes, evaluated under 10%, 20%, and 30% random/block missingness and operating mode shifts (Modes 1–6).
2
SWaT (Secure Water Treatment): 51 sensors/actuators over 11 days of normal and attack/fault operations.

Baselines: MLP, 1D-CNN/TCN, GCN (Kipf & Welling, 2017), GATv2 (Brody et al., 2022), MTGNN (Wu et al., 2020), and GDN (Deng & Hooi, 2021).
Metrics: Macro-F1 & Top-$K$ Root-Cause Localization Hit-Rate ($\text{HR}@3$), Expected Calibration Error ($\text{ECE}$), and Conformal Fault Set Size at $95\%$ coverage.

Expected Contributions

1
Variance-Damped GSL Layer (UncertainGSLConv): A drop-in PyTorch Geometric layer that prevents fault smearing across corrupted sensor neighborhoods.
2
Comprehensive Robustness Benchmark: Evaluation of industrial GNN fault diagnosis under simultaneous missingness, noise injection, and operating-regime shift.
3
Conformal Root-Cause Sets: Replacing point fault guesses with calibrated candidate fault sets $\mathcal{C}(X_t)$ for human plant operators.

Feasibility, Scope, and Honest Risks

Why Low Risk: Tabular/sensor graphs require minimal VRAM ($< 4\text{ GB}$), no massive pretrained checkpoints, and have crystal-clear ground-truth fault labels in TEP.
Main Challenge & Mitigation: Preventing the variance head $\boldsymbol{\sigma}_j^{2(t)}$ from collapsing to zero or exploding during training. Solution: Train with Beta-NLL loss (Seitzer et al., ICLR 2022) and apply synthetic sensor dropout masking ($p_{\text{mask}} \in [0.05, 0.25]$) during training so the variance head explicitly learns to flag missing/corrupted channels.

Initial Literature Notes

[1]
A. Deng and B. Hooi, "Graph Neural Network-Based Anomaly Detection in Multivariate Time Series (GDN)," AAAI, 2021.
[2]
Z. Wu et al., "Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks (MTGNN)," ACM KDD, 2020.
[3]
S. Brody, U. Alon, and E. Yahav, "How Attentive are Graph Attention Networks? (GATv2)," ICLR, 2022.
[4]
M. Seitzer et al., "On the Pitfalls of Heteroscedastic Uncertainty Estimation with Probabilistic Neural Networks (Beta-NLL)," ICLR, 2022.

Idea Validation Checklist

Clear physical failure mode identified (fault smearing from corrupted neighbor sensors)Verified
Mathematical damping mechanism integrated into GATv2 edge logitsVerified
Public benchmarks (TEP, SWaT) identified with multi-mode shift splitsVerified
Run baseline GDN vs. UC-GATv2 under 20% synthetic sensor blackout on TEPPending

Current Assessment

Criterion	Score	My Assessment Notes
Problem clarity	9 / 10	Fault smearing across graph edges is easy to visualize and quantify.
Novelty potential	8 / 10	Clean theoretical bridge between heteroscedastic GNNs and GSL fault diagnosis.
Technical feasibility	9 / 10	Straightforward implementation in PyTorch / PyG.
Resource feasibility	10 / 10	Lowest compute requirement of all 15 topics.
Dataset availability	9 / 10	TEP and NASA C-MAPSS are immediately downloadable.
Publication potential	8 / 10	Great target for IEEE TIM, Reliability Engineering & System Safety, or AAAI/KDD.
Personal interest	9 / 10	Natural extension of my smart-grid and feature-correlation research.

Next Actions

1
Download the 21-fault Tennessee Eastman Process (TEP) dataset and build the sliding-window PyTorch Dataset with synthetic random/block sensor dropout masks.
2
Compare standard GATv2 against variance-damped GATv2 on root-cause localization $\text{HitRate}@3$.

Open Questions and Working Notes

Can we separate sensor fault (a broken thermocouple where node $j$'s value is wrong but the reactor is healthy) from process fault (where the physical reactor state actually shifted across connected nodes) by comparing self-residual vs. neighborhood-consensus residual?