CausalTrace: A Neurosymbolic Causal Analysis Agent for Smart Manufacturing¶

Conference: AAAI 2026 arXiv: 2510.12033 Code: GitHub Area: LLM Agents / Smart Manufacturing Keywords: Neurosymbolic Systems, Causal Analysis, Root Cause Analysis, Knowledge Graph, Industrial CoPilot

TL;DR¶

This paper proposes CausalTrace — a neurosymbolic causal analysis agent integrated into an industrial CoPilot (SmartPilot) that combines data-driven causal discovery with industrial ontologies and knowledge graphs, enabling real-time root cause analysis, counterfactual reasoning, and interpretable decision support.

Background & Motivation¶

Background: Manufacturing is moving toward hyper-autonomous operations, with AI-driven perception, control, and decision support becoming increasingly prevalent. Machine learning models excel at demand forecasting and anomaly detection, yet lack interpretability in high-stakes industrial environments.

Limitations of Prior Work: Existing AI systems typically operate as isolated black boxes, lacking seamless integration of prediction, explanation, and causal reasoning. Shop-floor operators and domain experts require not only accurate predictions but also actionable and comprehensible insights into system behavior.

Key Challenge: Industrial scenarios demand a unified capability encompassing "what happened + why it happened + what would change if we intervened," yet existing approaches are either purely symbolic (poor scalability), purely neural (lack of transparency), or agent-based systems (lacking semantic grounding and human-machine interaction support).

Goal: To construct a practical decision support system that unifies causal reasoning, neurosymbolic methods, and agentic AI into a single deployable framework.

Key Insight: Extending causal analysis capabilities on the existing SmartPilot multi-agent industrial CoPilot platform, following the C3AN (Customized, Compact, Composite AI + Neurosymbolic Integration) design paradigm.

Core Idea: CausalTrace transforms causal analysis from an academic method into a deployable industrial decision tool through a Bootstrap stability-enhanced causal discovery engine, knowledge graph/ontology-driven semantic augmentation, and an LLM-powered natural language interaction interface.

Method¶

Overall Architecture¶

CausalTrace serves as the fourth agent in the SmartPilot multi-agent architecture, operating alongside PredictX (anomaly prediction), ForeSight (throughput forecasting), and InfoGuide (question answering). Its causal analysis pipeline comprises: Data Loader → Feature Selector → Causal Discovery Engine → Root Cause Analysis → Neurosymbolic Integration → Interactive User Interface → Memory Module.

Key Designs¶

Module 1: Causal Discovery Engine (with Bootstrap Stability Analysis)

Function: Constructs directed acyclic graphs (DAGs) from multivariate sensor data and assesses the reliability of each causal edge.
Mechanism: Supports ICA-LiNGAM and DiffAN algorithms for causal structure discovery from data. The key innovation is the integration of Bootstrap stability analysis — repeatedly resampling data and re-running causal discovery to statistically measure the frequency and strength variance of each edge across different samples, yielding a stability score \(s = 1/(1+\sigma)\). Edges with \(s \geq 0.9\) are considered highly stable; those with \(s < 0.6\) are excluded. Retained edges are used to compute the total causal effect matrix \(\mathbf{T} = (\mathbf{I} - \mathbf{B})^{-1}\).
Design Motivation: Single-run causal discovery is sensitive to data perturbations; Bootstrap augmentation substantially improves the robustness and credibility of discovered results. The total causal effect matrix simultaneously captures direct and indirect (multi-hop) causal influences.

Module 2: Neurosymbolic Integration (Three-Layer Knowledge Injection)

Function: Injects structured domain knowledge into reasoning and explanation generation via multiple complementary representations.
Mechanism: (a) Knowledge Graph Layer: An RDF-encoded smart manufacturing knowledge graph providing semantic context (relationships among sensors, machines, parts, and anomalies) to InfoGuide responses via rdflib; (b) Process Ontology Layer: A dynamic process ontology implemented in Neo4j, with real-time Cypher queries retrieving explanations, tolerance ranges, and sensor-function mappings; (c) Causal Graph Prompt Injection: Serializing the total causal effect matrix and injecting it into LLM prompts, grounding generated explanations and reasoning in the causal graph structure.
Design Motivation: Purely data-driven causal analysis lacks domain semantics and may produce meaningless or misleading results. The three-layer knowledge injection ensures that reasoning outputs are meaningful and interpretable within industrial contexts.

Module 3: Memory Module (Persistent Context Reasoning)

Function: Stores and retrieves information across sessions to support context-aware continuous reasoning.
Mechanism: Three memory types — (a) Episodic Memory: timestamped logs of causal discovery and RCA runs for longitudinal tracking; (b) Semantic Memory: structured annotations of sensors and entities for context-enriched explanations; (c) Procedural Memory: user preferences (e.g., algorithm selection, display settings) for personalized interaction. Memories are stored in JSON format and injected into InfoGuide responses.
Design Motivation: Industrial analysis is typically ongoing; operators need analytical context to remain coherent across multiple sessions.

Loss & Training¶

CausalTrace does not involve end-to-end training. Its core relies on algorithmic causal discovery (ICA-LiNGAM / DiffAN) augmented by knowledge graph integration. Evaluation follows 10 selected principles from the C3AN framework's 14-criterion rubric, assessed via LLM-as-Judge (GPT-4o-mini + LLaMA3-70B) and 6 human evaluators (3 manufacturing domain experts + 3 computer scientists).

Key Experimental Results¶

Main Results¶

Dataset: Rocket assembly dataset from the Future Factories Lab at the University of South Carolina — 166K records, 30 hours, 285 complete assembly–disassembly cycles.

Method	ROUGE-1	Jaccard	MAP@3	PR@2	MRR
RCA Baseline (Correlation)	—	0.33	44%	51%	0.50
CausalTrace (w/o KG/Ontology)	0.56	—	—	—	—
CausalTrace (Full)	0.91	0.92	94%	97%	0.92

C3AN Principle Evaluation Overall Score: 4.59 / 5

Ablation Study¶

Removing the knowledge graph and ontology degrades ROUGE-1 from 0.91 to 0.56, demonstrating the significant contribution of knowledge integration to explanation quality.
The correlation baseline achieves only 0.33 Jaccard and 44% MAP@3, underscoring the importance of causal directionality (correlation ≠ causation).

Key Findings¶

Causal graphs (20 edges from LiNGAM, 15 from DiffAN) align closely with domain expert judgments.
The counterfactual validation module supports interactive verification of the credibility of causal links.
Ontology augmentation not only improves quantitative metrics but also renders explanations more credible in the eyes of domain evaluators.

Highlights & Insights¶

CausalTrace successfully translates causal reasoning from an academic concept into a deployable industrial decision support tool, demonstrating strong practical value.
Bootstrap stability analysis is a concise and effective approach to enhancing the robustness of causal discovery.
The three-layer neurosymbolic integration design (KG + Ontology + Causal Graph Prompt Injection) offers a replicable architectural pattern.
The three-type memory classification (episodic/semantic/procedural) is broadly applicable to agent systems requiring continuous interaction.

Limitations & Future Work¶

Validation is limited to an academic rocket assembly testbed; deployment on real industrial production lines has not been evaluated.
LiNGAM assumes a linear non-Gaussian model, which may be insufficient for complex nonlinear industrial systems.
The system depends on LLaMA3-70B for natural language understanding; performance under model substitution remains unknown.
A portion of the evaluation ground truth relies on manual expert annotation, limiting scalability.

SmartPilot (Shyalika et al.): The foundational multi-agent platform upon which this work builds.
C3AN Framework (Sheth et al.): Design paradigm for Customized, Compact, and Composite AI.
LiNGAM (Shimizu et al.): A classical ICA-based causal discovery method.
Insight: Industrial AI cannot be solved by a single model; it requires a systems engineering approach combining multi-agent collaboration, knowledge augmentation, and human-machine interaction.

Rating¶

⭐⭐⭐⭐ (4/5)

The system engineering is comprehensive, covering causal discovery, knowledge augmentation, and user interface design. The C3AN evaluation framework itself offers reference value. One point is deducted because the experimental setting is relatively simple (an academic testbed), and the strong assumptions underlying the causal discovery methods raise questions about generalizability to real industrial environments.