Novel Machinery Monitoring Strategy Based on Time–Frequency Domain Similarity Measurement With Limited Labeled Data

Abstract

Machinery condition monitoring methods that can automatically classify the system health condition using machine learning algorithms often require a large amount of labeled data, which are not always available in real industrial processes. This article proposes a different solution to automatic machinery condition monitoring based on the signal similarity measurement in the time–frequency domain. In the first stage, using short-time Fourier Transform (STFT), time–frequency representation (TFRs) of vibration signals issued for different machinery health conditions are used to create a baseline. In the next stage, the similarities in the time–frequency domain between very limited labeled and test signals are measured using a novel technique. In the third stage, the similarities are compared in order to classify the test signals. In the proposed method, the baseline of the labeled samples could be from only one operating condition, while the test samples are from all possible operating conditions. In addition, to improve the robustness against noise and make the proposed method sensitive to local differences in signals, a multiscale analysis (MSA) using the wavelet packet decomposition (WPD) technique is adopted. As a specific example, effectiveness and robustness of the proposed method in a noisy environment is validated through various bearing defective levels and for different speeds and workloads. Experimental results demonstrate that the proposed solution achieves a high classification accuracy even with very limited labeled samples.