Interventional Causal Structure Discovery for Domain-Invariant Operating-Condition Classification in Multimodal Robotic Manufacturing
Year: 2026
Type: preprint
Abstract: Reliable condition-based maintenance (CBM) in smart manufacturing is severely limited by associative overfitting, where traditional models learn volatile statistical correlations that collapse under out-of-distribution (OOD) operational shifts. When routine changes in manufacturing system’s speed, trajectory geometry, or payload alter a system’s mechanical impedance, standard classifiers mistake environmental baseline shifts for genuine changes in machine state, generating unreliable operating-condition labels. This study introduces the Interventional Causal Framework for Condition-Based Maintenance (IC-CBM), which transitions predictive manufacturing analytics from statistical correlation to structural causation. Using the constraint-based Peter-Clark (PC) algorithm supported by bootstrap resampling, IC-CBM recovers the physical causal graph governing multimodal vibration and acoustic sensor streams from a controlled factorial interventional design. The recovered graph isolates the kinematic velocity backbone and cross-modal acoustic radiation mechanism as high-stability causal pathways. Structural Residual Normalization (SRN) then derives a targeted architectural correction directly from the recovered graph, subtracting mass-induced acoustic inflation along the identified causal path. The operating condition evaluated here is defined by a speed threshold, an operational proxy for mechanical stress rather than a measured fault or degradation state, so this study validates IC-CBM as a domain-invariant operating-condition classifier, a foundational step toward causally grounded CBM. Under a 10-trial bootstrapped held-out domain evaluation, the residual-optimized causal framework achieved 88.89% out-of-distribution accuracy, statistically indistinguishable from its in-distribution performance and outperforming the full-feature correlational baseline by 20.00 percentage points. By compressing the feature space by 84.61%, the framework achieves a high-stress-condition classification precision of 1.00 with a recall of 0.67 under severe payload covariate shift, meaning every flagged high-stress cycle is genuine, though roughly a third of true high-stress cycles are conservatively classified as standard. These results demonstrate that grounding predictive manufacturing analytics in interventionally discovered causal structure yields classification representations that generalize reliably to unseen operational contexts without sensor recalibration or data relabeling.
Source: Research Square
Authors: Md Omar Al Javed(Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Writing – original draft), Matthew Taylor(Data curation, Investigation, Resources, Writing – review & editing), Vivekanand Naikwadi(Project administration, Supervision, Writing – review & editing), Ayantha Senanayaka(Conceptualization, Methodology, Project administration, Supervision, Writing – review & editing)
Institution: Tennessee Technological University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203840966
