-
RCA necessitates the integration of causal associations (via causal models) among the data and
knowledge from diverse sources, ofering enhanced explainability, scalability, and the ability to
perform intervention and counterfactual analysis.
Causal Neuro-Symbolic AI (NSAI) is a hybrid framework that blends the strengths of causal and
NSAI representations and techniques [
4
]. Our Causal NSAI-enhanced RCA approach supports
RCA by expressing the causal association among the data into symbolic representation. A KG
that encodes causal Bayesian network (CBN) based representation explicitly expresses the causal
association. The framework utilizes NSAI methods such as KG link prediction, providing better
scalability for inferring causal relations within large volumes of data. Causal NSAI provides the
following benefits: 1) Explainability- it enhances explanations for failures by combining causality
with symbolic (i.e., ontology and KG) reasoning and neural networks; 2) Robustness- it leverages
causal associations to improve the robustness of AI models to changes in data distribution,
and out-of-distribution data; 3) Generalization- it enhances generalizability by incorporating
prior domain knowledge with causal associations; 4) Intervention- it leads to models which
can predict the impact of interventions in the production line and make informed decisions to
achieve desired outcomes. This integration of causal associations with NSAI enables dynamic
adaptation to new manufacturing environments. The framework is applied to an
industrygrade dataset for a rocket assembly line1 and the causalAssembly2 based simulated dataset
[
5
]. The rocket assembly is a multi-modal data setup with four robots, rocket parts, conveyor
belts, a material handling station, stoppers, image and video recording, and sensors such as
temperature, potentiometer, load cells, robot angles, programmable logic controls, etc. A failure
in the assembly is defined as the absence of a rocket part in the final product. The Causal NSAI
framework is utilized to 1) provide explanations for the cause of the failure, 2) suggest measures
that can be taken to obtain a final product despite the failure, and 3) identify interventions and
counterfactuals using the causal associations to prevent future failures. Acknowledgments:
NSF Awards #2335967 and #2119654.
1Our public data is at https://www.kaggle.com/datasets/ramyharik/f-2023-12-12-analog-dataset
2https://github.com/boschresearch/causalAssembly