A Hidden Markov Framework for Physically Interpretable Arc Stability Dynamics in Welding Systems

1. Problem Statement

Electric Arc Welding (EAW) is a complex industrial process governed by nonlinear interactions between electromagnetic fields, thermal dynamics, and plasma behavior. The stability of the electric arc is critical for weld quality and structural integrity. However, the arc signal is inherently non-stationary and stochastic, exhibiting rapid temporal fluctuations due to plasma turbulence and external disturbances.

Key Limitations of Existing Approaches:

• Static Classification: Traditional methods rely on frame-based analysis, treating observations as independent snapshots. This ignores temporal dependencies and the sequential nature of arc evolution.
• Feature Overlap: Spectral features (e.g., energy, entropy) often exhibit partial overlap between different operational regimes (transient, stable, extinction), making static classifiers unreliable.
• Lack of Interpretability: While data-driven models exist, they often fail to explicitly link extracted features to underlying physical arc dynamics, limiting their practical applicability in real-time monitoring.

2. Methodology

The proposed framework models arc dynamics as a probabilistic state-space problem using a Hidden Markov Model (HMM) to capture temporal evolution and regime transitions.

A. Signal Acquisition and Preprocessing

• Data Source: Welding current signals are acquired from a Metal Active Gas (MAG) welding platform using a Hall-effect sensor.
• Modeling: The signal is modeled as $i(t) = s(t) + n(t)$, where $s(t)$ is the underlying arc dynamics and $n(t)$ is additive noise.
• Segmentation: The continuous signal is segmented into overlapping windows to preserve temporal continuity.

B. Time-Frequency Feature Extraction

The Short-Time Fourier Transform (STFT) is applied to convert the signal into the time-frequency domain. Three physically meaningful spectral descriptors are extracted to form the observation vector $v_t$:

1. Spectral Energy ($E(t)$): Reflects total energy content and power transfer characteristics.
2. Spectral Entropy ($H_s(t)$): Quantifies the disorder or complexity of the spectral distribution.
3. Spectral Centroid ($C(t)$): Represents the dominant frequency location.

C. Hidden Markov Model (HMM) Formulation

The arc process is modeled as a discrete-time stochastic system with three latent states:

1. Transient: Rapid ignition dynamics, broadband spectral content.
2. Stable: Energy concentrated around fundamental frequencies; sustained conduction.
3. Extinction: Irregular fluctuations, dispersed energy, and arc collapse.

• State Transition: Modeled via a transition matrix $A$, assuming a first-order Markov process where the next state depends only on the current state.
• Emission Model: Observations are modeled as state-dependent Gaussian distributions ($N(\mu_k, \Sigma_k)$).
• Inference: The Viterbi algorithm is used to decode the most probable sequence of hidden states ($\hat{Z}$) given the observation sequence ($V$).
• Parameter Estimation: The Baum-Welch algorithm is employed to maximize the likelihood of the model parameters.

3. Key Contributions

1. Probabilistic State-Space Formulation: Shifts arc analysis from static classification to a temporal probabilistic inference problem, modeling the welding process as a sequence of hidden regimes.
2. Physically-Informed Observation Model: Explicitly links STFT-based spectral features (energy, entropy, centroid) to underlying arc physics, ensuring the model is interpretable.
3. Temporal Inference Framework: Proposes an HMM-based approach to estimate state transitions and persistence, capturing the gradual evolution of arc behavior rather than instantaneous labeling.
4. Demonstration of Limitations of Static Methods: Provides empirical evidence (via PCA and Silhouette analysis) that feature distributions overlap significantly, proving that temporal modeling is necessary for robust state inference.

4. Experimental Results and Analysis

The framework was evaluated using experimental MAG welding data.

• Feature Space Analysis:

  • PCA Projection: Revealed three distinct but partially overlapping clusters corresponding to the three regimes.
  • Silhouette Analysis: Confirmed that while structure exists, significant overlap (especially between Transient and Stable states) renders static clustering insufficient.
  • Statistical Trends:
    • Transient: Low mean energy, high variability.
    • Stable: Moderate energy, low entropy variance (quasi-stationary).
    • Extinction: Highest energy concentration, lowest entropy (highly constrained state before collapse).

• Transition Dynamics:

  • State Persistence: The diagonal elements of the transition matrix were high (Transient: 0.95, Stable: 0.90, Extinction: 0.87), indicating that once a regime is entered, the system tends to persist in that state.
  • Instability Risk: The probability of transitioning from Stable to Extinction was estimated at ~0.10, providing a quantitative metric for failure likelihood.

• Temporal Validation Metrics:

  • Average State Persistence: 0.8802 (High continuity).
  • Mean State Duration: ~19.7 frames (Indicates meaningful temporal segments rather than noise).
  • Transition Ratio: 0.0435 (Low frequency of state switching, consistent with physical inertia).

5. Significance and Conclusion

This study demonstrates that arc stability is a temporally evolving process rather than a set of independent observations. By integrating time-frequency analysis with Hidden Markov Models, the proposed framework offers several advantages:

• Physical Consistency: The model aligns with the physics of plasma dynamics, where established arc conditions persist due to inertia, and transitions occur gradually.
• Robustness: It overcomes the limitations of static classifiers by leveraging temporal coherence to resolve ambiguities caused by overlapping feature distributions.
• Predictive Capability: The transition matrix allows for multi-step prediction of arc behavior, enabling early detection of instability (e.g., predicting the likelihood of arc extinction).

The work establishes a new paradigm for welding monitoring, moving from static feature-based classification toward interpretable temporal modeling, which is essential for real-time control and quality assurance in non-stationary industrial environments. Future work will focus on expanding datasets across different operating conditions and refining labeling strategies.