Application Cases

Experiment Name: Research on Data-Driven Characterization Method of Magnetostrictive Barkhausen Noise for Surface Stress in Structural Steel

Research Direction: Material Testing

Experimental Objective:
Magnetostrictive Barkhausen Noise (MBN) technology can be used to quantitatively evaluate the surface stress of ferromagnetic materials. Current MBN-based stress evaluation techniques face challenges such as difficulty in selecting characteristic parameters, complexity in quantitative prediction models, and low fitting accuracy to calibration datasets. This paper proposes a data-driven nonlinear mapping algorithm to fit the relationship between MBN noise and stress. It explores the use of wavelet packet transform coefficients as time-frequency features to replace statistical parameters, reducing the computational load of sample data. The wavelet packet transform coefficients of MBN noise in the time-frequency domain are used as feature vectors, and a data dimensionality reduction algorithm based on singular value decomposition is applied to reduce the dimensionality of the feature vectors. The reduced-dimensional feature vectors are then input into a BP neural network for model training to establish a prediction model. The results show that the data dimensionality reduction algorithm based on singular value decomposition reduces the complexity of the model. Using the reduced-dimensional wavelet packet transform coefficients as feature vectors to train the BP neural network enables high-precision prediction of surface stress in ferromagnetic materials. The characterization method established in this paper effectively addresses the issue of stress distribution imaging in ferromagnetic components and holds broad application prospects in damage prevention, such as stress corrosion prevention and fatigue strength enhancement.

Testing Equipment:
The system consists of hardware and software components. The hardware includes an excitation module, an MBN sensor, a signal conditioning module, and an AD acquisition module. The software comprises an upper computer designed based on LabVIEW, featuring data acquisition and analysis capabilities. The excitation module consists of an arbitrary waveform function signal generator and an ATA-4014 high-voltage power amplifier.

Figure 1: Schematic Diagram of the MBN Detection System

Experimental Procedure:
The MBN sensor consists of three parts: a U-shaped magnetic yoke, an excitation coil, and a detection coil. The U-shaped magnetic yoke is made of manganese-zinc ferrite with high magnetic permeability and low electrical conductivity. The magnetic pole faces are square, with a distance of 19 mm between the centers of the two poles. Both the excitation and detection coils are made of enameled wire, wound around the U-shaped yoke and a magnetic rod, respectively. The detection coil is positioned directly below the center of the U-shaped yoke. When a sinusoidal signal passes through the excitation coil, an approximately uniform alternating magnetic field is generated between the poles of the U-shaped yoke, forming a magnetic circuit between the yoke and the test specimen, producing millivolt-level weak MBN signals. The detection coil transmits the captured MBN signals via cable to the signal conditioning module for filtering and amplification, making them suitable for acquisition by the AD module and subsequent display and storage on the upper computer. The figure below compares the MBN signal waveforms before and after filtering.

Figure 2: Comparison of MBN Signal Waveforms Before and After Filtering

Experimental Results:

1. By applying wavelet packet transform to convert the MBN signal from the time domain to the time-frequency domain, time-frequency redundancy features were extracted, which qualitatively reflect different stress levels.

2. Feature vectors of wavelet packet transform coefficients from different frequency bands were used to construct a wavelet packet transform coefficient matrix. Singular value decomposition reduced the dimensionality of the MBN feature vectors, thereby lowering the complexity of the BP neural network model.

3. A gradient stress calibration experiment was conducted on a cantilever beam. MBN signals from gradient stress on the upper surface of the cantilever beam were collected, and a training dataset for the BP neural network was established based on the stress levels corresponding to different measurement points.

4. Using MBN signals from gradient stress on the lower surface of a structural steel specimen under three-point bending, a test dataset for the BP neural network was established to evaluate the prediction accuracy of the data-driven model.

5. Using wavelet packet coefficients as time-frequency redundancy features and reducing their dimensionality via singular value decomposition, a BP neural network stress prediction model was established. This model exhibits low complexity and fast response, with prediction errors for most samples below 10%, achieving quantitative evaluation of surface stress in ferromagnetic components based on data-driven methods. In practical engineering applications, combining technologies such as robotic arm automated scanning, sensor adaptive profiling optimization, and signal processing can enable high-precision imaging of stress distribution on ferromagnetic component surfaces.

Aigtek ATA-4014C High-Voltage Power Amplifier:

Figure: Specifications of the ATA-4014C High-Voltage Power Amplifier

The experimental materials in this article were compiled and published by Xi’an Aigtek Electronics.