G-codes-coupled wireless vibration trajectory-based side-channel monitoring for raster-angle intrusion detection in fused filament fabrication

Technical Summary: G-codes-coupled wireless vibration trajectory-based side-channel monitoring for raster-angle intrusion detection in fused filament fabrication

Problem Statement
Fused Filament Fabrication (FFF), also known as Fused Deposition Modeling (FDM), is a cyber-physical system (CPS) where digital instructions (G-code) drive physical manufacturing processes. A critical vulnerability in this domain is the potential for cyber-physical intrusions that alter internal structural parameters, such as raster orientation, without producing observable changes in the external geometry of the printed part. Such alterations can compromise the mechanical integrity and anisotropy of mission-critical components. While side-channel monitoring (e.g., vibration, acoustic, thermal) offers a method to verify process execution, existing approaches often rely on visual inspection or fail to distinguish between closely spaced raster angles (e.g., 5° increments). Furthermore, previous studies have not adequately separated the evaluation of "intended" digital commands from "observed" physical responses, risking data leakage where classifiers recognize the command file rather than the physical execution. The core challenge addressed is whether low-cost, non-intrusive wireless vibration sensing can reliably distinguish between closely spaced raster-angle states in previously unseen specimens and how coupling this with G-code trajectory information affects identification performance.

Methodology
The study proposes a side-channel monitoring framework that integrates digital command descriptors with physical vibration responses.

• Experimental Setup:

  • Hardware: A Prusa MK4S 3D printer was used to fabricate 70 independent dogbone specimens (10 per class) representing seven raster-angle classes (45° to 75° in 5° increments).
  • Sensing Architecture: Four distributed wireless nodes were employed, each consisting of an ADXL345 tri-axial accelerometer connected to an ESP32 microcontroller. One node was mounted near the nozzle, and three were mounted on the print bed. Data was transmitted via Wi-Fi (HTTP POST) to an external server without modifying the printer's internal controller.
  • Data Collection: 280 total observations were generated (4 sensors × 70 specimens).

• Feature Engineering:

  • G-code Trajectory Descriptors: 34 features were extracted from the G-code files, including planar displacement, circular statistics (mean direction, resultant length), axis dominance ratios, and toolpath complexity metrics.
  • Vibration Features: Raw signals (resampled to 8192 points) were processed to extract:
    • Time-domain features (RMS, standard deviation, crest factor, etc.) for X, Y, and Z axes.
    • Directional cross-axis features (correlations, energy ratios, orientation angles).
    • Wavelet Packet Transform (WPT) features: Energy and energy ratios from 16 terminal nodes at decomposition level 4.

• Analytical Pipelines:
  Two distinct pipelines were evaluated to isolate the contribution of digital vs. physical data:

  1. Proposed Pipeline (Trajectory-Informed): Fused G-code trajectory descriptors with vibration features. Feature selection was performed using Minimum Redundancy Maximum Relevance (mRMR). Classification was executed via a Bayesian-optimized Gaussian Error-Correcting Output Codes (ECOC) Support Vector Machine (SVM).
  2. Benchmark Pipeline (Vibration-Only): Used only vibration features. Dimensionality reduction was performed using Principal Component Analysis (PCA). Classification used the same Bayesian-optimized Gaussian ECOC-SVM.

• Validation Strategy:
  To prevent data leakage and ensure specimen independence, the study employed a strict specimen-level partitioning strategy. For five repeated runs, specimens were split into training (49), validation (14), and testing (7) sets. The training and validation sets were combined into a development set for model tuning (standardization, feature selection, hyperparameter optimization). Crucially, the 7 held-out test specimens were completely excluded from all model development steps.

Key Contributions

1. Non-Intrusive Retrofit Architecture: The study demonstrates a retrofit wireless sensing architecture that monitors physical execution without accessing or modifying the printer's internal motion-control system, preserving logical separation between command generation and execution verification.
2. Trajectory-Informed vs. Physical-Only Comparison: By benchmarking a fused G-code/vibration pipeline against a vibration-only pipeline, the study quantifies the specific contribution of commanded trajectory information versus independent physical side-channel evidence in detecting raster-angle alterations.
3. Rigorous Specimen-Level Validation: The research addresses common machine learning leakage issues in additive manufacturing by enforcing specimen-level data partitioning and ensuring that preprocessing and feature selection are fitted solely on development data, excluding held-out test specimens.

Results

• Proposed Method (Trajectory-Informed): Achieved 100% test accuracy across all five repeated specimen-level evaluations. The inclusion of G-code trajectory descriptors allowed the model to perfectly distinguish the seven closely spaced raster-angle states.
• Benchmark Method (Vibration-Only): Achieved a mean test accuracy of 90.00% with a standard deviation of 7.32% across the five runs. While the vibration-only approach demonstrated the ability to estimate raster angles from physical responses, it showed greater variability and lower performance, particularly in distinguishing intermediate angles (e.g., 60°–70°).
• Feature Analysis: t-SNE visualizations indicated that while vibration data alone contained class-discriminative information, the trajectory-informed features provided more compact and well-separated clusters. The vibration-only benchmark showed greater overlap among intermediate raster angles.

Significance and Claims
The paper claims that the proposed framework successfully demonstrates that wireless vibration side channels can distinguish closely spaced raster-angle states, supporting the feasibility of raster-angle intrusion monitoring. The primary significance lies in the complementary nature of the two information pathways:

• G-code features provide a deterministic representation of the intended manufacturing state.
• Wireless vibration features provide independent evidence of the physical execution response.

The authors emphasize that while the proposed method achieved perfect accuracy, this result reflects the strong separability of the commanded state when combined with physical data. The vibration-only benchmark's high (but imperfect) accuracy confirms that physical side channels carry sufficient information to infer process states, even without direct access to the G-code.

The study explicitly states that it does not demonstrate anomaly or attack detection performance, as no command-process mismatches or cyberattacks were introduced. Instead, the work establishes the necessary digital and physical information pathways for future verification systems. The authors conclude that such a system could support future command-execution verification by comparing a trusted reference (intended G-code) against an independently estimated physical state (vibration response) to detect deviations. The framework is presented as a practical, low-cost retrofit for enhancing the integrity monitoring of mechanically mission-critical FFF components.