A Pulsed Eddy Current Data-Driven Interpretable Neural Network for Metallic Foreign Objects’ Characteristics Identification

During the high-yield continuous crushing operations in coal mines, the detection of feeding metallic foreign objects (MFOs) faces challenges such as complex noise interference and varying liftoff distances. This article proposes a pulsed eddy current data-driven interpretable neural network (PECDINet) based on optimized variational mode decomposition (OVMD) and improved wavelet threshold denoising (IWTD) for multifeature classification and identification of MFOs. First, for the industrial dataset collected at the crushing station, the variational mode decomposition (VMD) is optimized using the snow geese algorithm (SGA) and envelope entropy (EE) function that incorporates environmental noise. Improved adaptive threshold function is then applied for wavelet threshold denoising (WTD) of the intrinsic mode functions (IMFs), effectively filtering out complex noise while preserving key characteristics of PEC signals corresponding to the time-domain solutions of PEC reflection fields in truncated regions of MFOs. Then, PECDINet with subtraction, square difference, and average layers is developed to extract start and end times, and time to peak of experimental signals, enabling characteristics identification of conductivity, thickness, and vertical width of MFOs. Finally, experiments demonstrate the proposed method’s advantages in accuracy and efficiency, addressing the challenges of multifeature classification and identification of MFOs in coal mine field environments.