Authors:
Eduardo Jr Piedad、Zherish Galvin Mayordo、Eduardo Prieto-Araujo、Oriol Gomis-Bellmunt
Paper:
https://arxiv.org/abs/2408.09649
Introduction
Machine fault detection is a critical aspect of industrial operations, as it helps in minimizing downtime and preventing operational interruptions. Traditional methods of fault detection often rely on vibration-based sensor data, which can be expensive and invasive. Recent advancements in artificial intelligence, particularly in machine learning (ML) and deep learning (DL), have shown promise in diagnosing motor conditions using motor phase current signals. This study focuses on converting time-series motor current signals into time-frequency 2D plots using Short-time Fourier Transform (STFT) methods and employing Convolutional Neural Networks (CNNs) to classify motor faults.
Related Work
Previous studies have explored various methods for motor fault detection. Traditional ML methods using 1D frequency-transformed datasets have shown good accuracy, with some achieving up to 93.20%. Recent approaches have transformed time-series motor current signals into 2D occurrence and recurrence plots, achieving accuracies of 80.25%, 74.80%, and 82.80% using CNNs. However, these methods still lag behind the best ML methods. The STFT method, which provides a time-frequency analysis of signals, has been effective in capturing transient and non-stationary characteristics typical of motor faults. This study aims to explore different STFT variants to improve fault detection accuracy.
Research Methodology
Electric Motor Current Dataset
The dataset consists of single-phase electric current signal data collected from five 2-HP induction motors. One motor is healthy, while the others exhibit synthetically imposed faults: bearing axis misalignment, stator inter-turn short circuit, broken rotor strip, and outer bearing defect. Data is collected under five loading conditions (0%, 25%, 50%, 75%, and 100%), resulting in a total of 3750 datasets, each with five seconds of data sampled at 10kHz. These datasets are then transformed into 2D plots using various STFT methods.
Short-time Fourier Transform and Its Variants
The STFT is a widely-used method for time-frequency analysis of signals. It involves sliding a window across the signal and computing the Fourier transform of the windowed signal at each position. Variants of the STFT, such as realigned STFT and synchrosqueezed STFT, offer refined time-frequency representations. This study uses five STFT methods: non-overlapping STFT, overlapping STFT, non-overlapping realigned STFT, overlapping realigned STFT, and synchrosqueezed STFT. These methods are applied to the motor current signal dataset to generate 2D time-frequency plots.
Experimental Design
Convolutional Neural Network (CNN)
The CNN model architecture used in this study is based on a previous architecture that has shown promising results in image-based analysis. The model takes 64×64 RGB images as input, extracted from each set of STFT methods. The training and testing steps, including 10-fold cross-validation, are reproduced from previous studies to allow for a fair comparison of results.
Training and Validation
Five CNN models were developed, trained, and validated using the generated 2D plots from the five STFT methods. The training and validation classification accuracy performances, as well as the loss function graphs, were analyzed to evaluate the models’ performance.
Results and Analysis
Performance Evaluation
The CNN models showed varying levels of performance across the different STFT methods. The overlapping STFT method achieved the highest average accuracy of 97.65%, followed by non-overlapping realigned STFT (96.32%), non-overlapping STFT (96.08%), and overlapping realigned STFT (96.03%). The synchrosqueezed STFT method struggled to converge, achieving an average accuracy of 88.27%.
Comparison with Previous Methods
The results were compared with previous studies using the same dataset and CNN architecture. Four of the five STFT methods outperformed the best ML method, which had an accuracy of 93.20%. All five STFT methods outperformed previous 2D-plot-based methods, which had accuracies of 80.25%, 74.80%, and 82.80%.
Confusion Matrix Analysis
The confusion matrices of the best-performing STFT method (overlapping STFT) and the previous best ML model (LightGBM) were analyzed. The overlapping STFT method showed perfect classification accuracy for healthy motors, bearing axis misalignment, and stator inter-turn short circuit faults, with minimal confusion between broken rotor strips and outer bearing defects.
Overall Conclusion
This study demonstrates the effectiveness of using Short-time Fourier Transform (STFT) methods to transform time-series motor current signals into time-frequency 2D plots for motor fault diagnosis. The CNN models trained on these plots showed promising performances, with four out of five STFT methods achieving over 96% average classification accuracy. These results highlight the potential of 2D-based methods in advanced signal processing and deep learning for motor fault diagnosis. Future work will explore other time-frequency analysis methods, such as the Wigner-Ville transform, to further improve fault detection accuracy.