Theory And System Research On Precise Quantitative Recognition Of Steel Tube's Small Defects Using Magnetic Flux Leakage Testing Method

When test steel tube's defects with MFL method, the information of defects may be interfered and influenced with many paths. Based on the analysis of these factors, programmable high-pass and anti-aliasing filters whose cut-off frequency can be automatically adjusted according to the relative velocity was well designed. When magnification is 70 times, the signal-to-noise-ratio (SNR) of signal is above 30dB which reaches or is closed to that of measuring apparatus. Unique single-channel double-amplitude (70 and 30 separately) circuit was designed so that different flaws with different size can be analyzed with different signal.Digital signal processing methods were applied to correct the phase and time-delay error so that tendency part in the signal was removed. Smoothing and filtering rejected the singularities and casual interference factors. By fitting the cross-section signal along the tube one by one, a series of"slice"signals, which reconstructed the space distribution of the defects'leakage flux, were obtained. Virtual signal of best position sensor, which would be used to analyze and calculate the fault quantitatively, was so extracted by re-sampling the reconstructed leakage flux signal of the defects.Biorthogonal spline wavelet method was applied to extract and compute 15 main eigenvalues including Up-p, Dv-v, Sa, E, and so on. Comparison analysis of eigenvalues between same kinds of defects with different size and between different kinds of defects was carried out. These eigenvalues provide the classification and quantification of defects with rich and effective information.BP neural network used to realize pattern recognition of defects was erected, one for each kind of fault. Single output modal for each network was adopted to improve the accuracy of classification. Gradient descent with momentum was employed in training the network to provide faster convergence. Classification result shows that classifier based on BP neural network has a relatively high accuracy. BP neural network used to realize quantitative recognition of defects was established, one for each kind of fault. Improvement has been made to networks'mean square error function to force the training process to be smoother and morestable. Result shows that the absolute deviation when quantifying trained samples is below 0.1mm, and is below 0.2mm when quantifying non-trained samples when considering mismachining tolerance. This result shows that quantification accuracy is high, and precise quantitative recognition for small defects can be realized.