Semi-supervised anomaly detection for rotating machinery under variable working conditions based on non-intrusive measurements and deep probabilistic models

Abstract A new semi-supervised anomaly detection method is developed in this paper to detect the status anomalies for rotating machinery under variable working conditions using the non-intrusive measurements. In fact, detecting anomalies in rotating machinery under diverse application requirements can be a challenge due to the difficulty of establishing an accurate mathematical model to describe the rotational motion process. In addition, although the existing anomaly detection methods relying on vibration sensors have yielded certain results, the intrusive measurements utilized in these cases may possess problems like inconsistent communication protocols, restricted installation space, as well as causing damage to equipment. To address such issues, a novel semi-supervised anomaly detection method relying on non-intrusive measurements is devised, where a deep compression network is developed to generate a low dimensional feature representation and a linear estimation network is devised to generate the parameters of probabilistic models. Specifically, the latent feature representation of non-intrusive measurements under variable working conditions is captured by the deep convolutional autoencoder (DCAE) in the compression network, and its changes over time are modeled as a class of probability models that follow a Gaussian mixture hidden Markov process (GMHMM). Moreover, the compression and estimation networks are trained in an end-to-end manner, followed by a sample energy being designed to accomplish anomaly detection. The experimental results on an open access dataset named MCC5-THU and a real dataset collected from the self-developed spindle-rotating machinery show that the devised method has superior detection performance and a fairly comparable detection accuracy to intrusive measurements-based methods.