Midrange Speakers Car Audio,Midbass Speaker,Shopping Mall Conference Speakers,High Performance Broadcast Loudspeaker NINGBO RFUN AUDIO TECHNOLOGY CO.,LTD , https://www.mosensound.com
With the rapid advancement of electronic technology, neural networks, and artificial intelligence, new methods for processing magnetic flux leakage signals have emerged globally. Traditional techniques are heavily influenced by human factors, leading to false positives and reduced detection accuracy. This has highlighted the urgent need for intelligent signal processing solutions. Multi-sensor data fusion is an emerging information processing technology that leverages the complementarity of multiple data sources and the computational power of modern computers. It enables multi-level, multi-faceted data analysis, generating insights that no single sensor could provide on its own. By combining redundant information from different sensors, it effectively reduces errors caused by individual sensor failures, making it a key focus in signal processing research [1].
This paper employs a wavelet threshold denoising algorithm to enhance defect signal features, followed by RBF neural network-based data fusion to improve detection accuracy and reduce measurement uncertainty. Two types of sensors—electromagnetic induction and Hall sensors—are used to take advantage of their complementary data. A 32-sensor array, consisting of 16 electromagnetic induction sensors and 16 Hall sensors alternately arranged, is applied to the steel pipe surface to capture high-quality data. These signals are then preprocessed and sent to the fusion center for integration.
The electromagnetic induction sensor operates by detecting changes in the magnetic flux passing through its coil as it scans the steel pipe. This induces an electromotive force that corresponds to the presence of defects. In contrast, the Hall sensor measures the magnetic flux density directly, producing a voltage proportional to the field strength. Unlike the electromagnetic sensor, the Hall sensor is not affected by the movement speed of the magnetic field, making it more stable in varying conditions.
Signal preprocessing is essential to remove noise and external disturbances before fusion. Techniques such as median smoothing and correlation-based culling help refine the data. The wavelet threshold denoising method is particularly effective in preserving defect features while reducing noise. After wavelet analysis, the local maximum modulus of the signal can be used as a feature for classification.
In the signal model, each sensor provides a measured value affected by internal and external noise. The true signal is estimated using statistical methods. The wavelet threshold is optimized using Stein’s Unbiased Risk Estimation (SURE), allowing adaptive noise reduction. This processed signal is then fed into the RBF neural network for feature-level fusion.
The RBF neural network structure consists of an input layer, a hidden layer with radial basis functions, and an output layer. The hidden layer maps input data to a high-dimensional space, while the output layer computes the final result. Training involves determining the center and width of the basis functions, followed by weight optimization using gradient descent.
Experimental results using MATLAB demonstrate that the RBF network outperforms conventional BP networks in terms of learning speed and detection accuracy. The average absolute error was reduced to 2.69% compared to 5.47% for BP networks. This confirms the effectiveness of the proposed approach in enhancing detection reliability and accuracy. The combination of wavelet denoising and RBF-based fusion offers a robust solution for magnetic flux leakage signal analysis, significantly improving the performance of non-destructive testing systems.