Acoustic Spectral Feature Dataset for Non-Destructive Defect Detection in Powder Metallurgy Automotive Oil Pump Stators

The dataset contains 20 normalized acoustic spectral features extracted from powder-metallurgy (PM) automotive oil pump stators for the purpose of non-destructive defect detection using machine learning. Acoustic signals were recorded from 40 intact and 62 defective components (26 cracked, 16 with tooth breakage, and 20 completely fractured) after controlled impact excitation. Each signal was filtered, transformed into the frequency domain (Fast Fourier Transform), and analyzed within four frequency bands (1–5 kHz, 5–10 kHz, 10–15 kHz, and 15–20 kHz). From every band, five descriptors were computed—Peak Frequency (P1–P4), Peak Amplitude (PA1–PA4), Spectral Centroid (C1–C4), Mean Amplitude (MA1–MA4), and Skewness (S1–S4)—to capture shifts in natural frequency and spectral energy caused by cracks or mass loss. The resulting 20-feature matrix, with labels for intact and three defect categories, provides a structured basis for training and evaluating machine-learning classifiers (SVM, KNN, MLP, RBF) to distinguish defective stators from healthy ones and to identify specific defect types with high accuracy.