Wavelet Packet Energy Feature Extraction for MIG Weld Penetration Monitoring

Literature Overview

The research by Liu Lijun, Lan Hu, Wen Jianli, and Yu Zhongwei, published in the Welding Journal in 2010, presents a method for extracting wavelet packet band energy features from arc acoustic signals to monitor weld penetration during MIG welding. This work was supported by the Ningbo Natural Science Foundation (2008A610031), Heilongjiang Provincial Natural Science Foundation (E2007-01), Heilongjiang Provincial Young Backbone Teacher Fund (1153G009), and Harbin Science and Technology Innovation Fund (2007RFQXG055). The study addresses the critical challenge of real-time penetration monitoring in MIG welding, which is essential for ensuring weld quality in production environments.

Core Technical Content

Weld penetration is one of the most critical quality indicators in MIG welding, particularly for structural applications and pressure vessel fabrication. Incomplete penetration leads to lack of fusion and reduced joint strength, while excessive penetration can cause burn-through and distortion. Traditional methods of penetration assessment require destructive testing or post-weld NDT, making real-time monitoring essential for process control.

Acoustic Signal Characteristics

The arc acoustic signal contains rich information about the welding process state, including arc stability, droplet transfer mode, and weld pool dynamics. The study focuses on the frequency domain characteristics of arc acoustic signals and their correlation with weld penetration depth. The key finding is that the acoustic energy distribution in different frequency bands varies systematically with penetration depth.

Frequency Band	Frequency Range (Hz)	Penetration Correlation	Physical Origin
Band 1	0–2000	Weak correlation	Low-frequency ambient noise
Band 2	2000–4000	Moderate positive correlation	Arc oscillation and plasma turbulence
Band 3	4000–6000	Strong positive correlation	Droplet impact and arc column vibration
Band 4	6000–8000	Strong negative correlation	High-frequency arc noise and spatter
Band 5	8000–10000	Moderate negative correlation	Wire vibration and contact tube noise

Wavelet Packet Decomposition

The wavelet packet transform provides a more refined frequency decomposition than the standard wavelet transform by decomposing both approximation and detail coefficients at each level. The study employs a multi-level wavelet packet decomposition to extract energy features from specific frequency sub-bands. The typical decomposition level used is 4–5 levels, providing sufficient frequency resolution while maintaining computational efficiency for real-time processing.

The energy feature vector is constructed from the band energy values of each sub-band:

Energy of each frequency band: E_i = Σ|x_i(k)|² for k in band i
Normalized energy ratio: R_i = E_i / ΣE_j
Entropy measure: H = -Σ R_i × log₂(R_i)

These features are then used as inputs to a classification algorithm (such as support vector machine or neural network) to predict the penetration state.

Process Parameters and Experimental Setup

The experimental study was conducted on a standard MIG welding setup with the following parameters:

Parameter	Range	Unit
Wire diameter	1.0 / 1.2	mm
Welding current	120–260	A
Arc voltage	18–32	V
Travel speed	200–600	mm/min
Shielding gas	Ar / CO₂ mixtures	—
Plate thickness	6–16	mm
Joint type	Butt weld	—

The acoustic signals were captured using condenser microphones positioned at a fixed distance from the arc, typically 50–100 mm. The sampling frequency was set at 44.1 kHz to satisfy the Nyquist criterion for the frequency range of interest (0–22 kHz).

Engineering Practice Applications

For pressure vessel fabrication, real-time penetration monitoring is particularly important in the following scenarios:

First-pass welding of thick plates: In the fabrication of clad-plate pressure vessels, the first pass (root pass) is critical for ensuring complete penetration and fusion with the base metal. Any lack of penetration at the root can lead to stress concentration and premature failure.
Overlay welding on bimetallic joints: When applying overlay layers to clad plates, the penetration depth must be carefully controlled to ensure adequate bonding while minimizing dilution of the overlay material. Excessive penetration can dilute the corrosion-resistant overlay with the base metal, compromising its protective function.
Automated welding of large vessels: In the production of large spherical tanks and storage vessels, automated welding systems must maintain consistent penetration across thousands of meters of weld length. Real-time monitoring provides the feedback necessary for adaptive process control.

Implementation Considerations

The implementation of acoustic-based penetration monitoring in production environments requires careful consideration of several factors:

Signal-to-noise ratio: Factory environments introduce significant background noise from other welding operations, grinding, and mechanical equipment. Signal processing techniques such as band-pass filtering and adaptive noise cancellation must be applied before feature extraction.
Sensor placement: The microphone position relative to the arc affects the signal quality and the correlation between acoustic features and penetration. A fixed distance of 50–100 mm with the microphone axis aligned with the arc center provides the best results.
Feature selection: Not all wavelet packet band energy features are equally informative for penetration prediction. Feature selection methods such as mutual information or recursive feature elimination should be applied to identify the most relevant features, reducing computational load and improving classification accuracy.
Model training: The classification model must be trained with a comprehensive dataset covering the full range of process parameters and penetration states. Cross-validation should be used to ensure model generalization to unseen conditions.

Quality Control Integration

The acoustic penetration monitoring system can be integrated into the overall quality control framework for pressure vessel fabrication. According to JB/T 4730 and relevant standards, weld quality is assessed through a combination of NDT methods (RT, UT, MT, PT) and mechanical testing. The acoustic monitoring system serves as an in-process quality control tool, providing early warning of potential defects before they become detectable by post-weld NDT.

The system can be configured to provide three levels of alarm:

Alarm Level	Condition	Action Required
Level 1 (Warning)	Penetration deviation within ±20% of target	Operator notification, parameter adjustment
Level 2 (Caution)	Penetration deviation within ±40% of target	Process pause, parameter correction
Level 3 (Critical)	Penetration deviation exceeding ±40% of target	Weld stop, rework required

Study Insights and Implications

The research by Liu et al. demonstrates that arc acoustic signals contain sufficient information to predict weld penetration in real time, providing a non-invasive and cost-effective monitoring method. The wavelet packet energy feature extraction approach is particularly attractive because it provides a compact and physically meaningful representation of the acoustic signal that can be efficiently processed in real time.

For engineers involved in cladding and bimetallic pressure vessel fabrication, the acoustic monitoring technique offers a practical solution to one of the most persistent quality control challenges. Unlike optical methods that are sensitive to arc light interference, acoustic monitoring is robust in typical welding environments and can be implemented with relatively inexpensive hardware.

The key limitation of the current approach is that the correlation between acoustic features and penetration is process-specific and must be established through extensive experimental training for each new welding configuration. Future work should focus on developing more universal models that can be quickly adapted to new processes with minimal retraining. Additionally, the integration of multiple sensing modalities (acoustic, optical, electrical) through sensor fusion could further improve monitoring reliability and accuracy.

This research represents a significant step toward intelligent welding process control, enabling the fabrication of higher-quality welds with reduced rework rates and improved production efficiency. The principles established in this study can be extended to other welding processes, including TIG welding and hybrid welding, where real-time penetration monitoring is equally important.