Machinery Fault Detection via Multi-Domain Features and ML–DL Stacking
Reliable machinery fault classification is essential for reducing downtime and maintaining operational efficiency in industrial systems. This study evaluates single-model and stacking-based learning approaches for ten-class machinery fault classification using multi-sensor vibration data. Each raw signal record with a size of 250,000×8 was transformed into a compact 136-dimensional feature vector using multi-domain descriptors extracted from the time, frequency, short-time Fourier transform (STFT), and wavelet domains. The discriminative capability of the extracted features was statistically validated using one-way analysis of variance (ANOVA). Based on the extracted feature representation, four classifiers—Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), and Multi-Layer Perceptron (MLP)—were evaluated together with several stacking configurations, including MLP-based stacking, machine-learning-guided stacking, and a multi-base stacking model combining RF, SVM, and KNN with MLP as the meta-classifier. Model performance was evaluated using a leakage-aware stratified 10-fold cross-validation framework with accuracy, precision, recall, F1-score, and computational time. Among the single models, RF achieved the best standalone performance with 97.74% accuracy and 97.72% F1-score. Among the stacking models, RF+MLP achieved the highest mean performance with 98.05% accuracy and 97.99% F1-score, followed closely by the multi-base stacking model with 97.90% accuracy and 97.88% F1-score. Statistical comparison using fold-wise model scores showed that RF+MLP significantly outperformed weaker baseline models, although its improvement over RF and the multi-base stacking model was not statistically significant. Overall, the results indicate that multi-domain feature extraction provides an effective compact representation for machinery fault diagnosis, while RF-based models offer the most reliable balance between classification performance, stability, and computational efficiency.