Abnormal sound detection for rotary parts in noisy environment by one-class SVM and non-negative matrix factorization Mr. Taichi KIATAMURA, Naoya TAKEISHI, Takehisa YAIRI and Koichi HORI (The University of Tokyo)
The purpose of this research is to develop a data-driven method of detecting abnormal sound coming from plant machinery‘s rotary part. For example, such abnormal sounds are caused by small scratches on the surface of bearings. While we can easily get normal sound data, it is difficult to obtain a sufficient amount of abnormal data because anomaly things rarely occur. Therefore, we assume that only normal data is available beforehand for anomaly detection in this research.
The main difficulty of this research is that there is a lot of surrounding noise which makes detecting small anomaly sound very hard.
In the proposed method, feature vectors are created by applying short time Fourier transform to the sound data. Then one-class SVM is trained on the normal data and is used to discriminate normal data and abnormal data. In this research, separation of surrounding noise is tried by using Non negative matrix factorization (NMF) before creating feature vectors to overcome the problem of noise superimposition and to improve the precision of discrimination. We claim that this noise separation is a unique characteristic of this research.
As a result, anomaly detection precision was greatly improved when applying the separation of noise by NMF, compared with not doing so.
Wavelet-like CNN Structure for Time-Series Data Classification Mr. Seungtae PARK, Haedong JEONG and Seungchul LEE (UNIST)
Vibration is one of the richest information in manufacturing field. Due to its cheap acquisition, vibration has become “big data” in manufacturing fields. Recently, deep learning models show the state-of-art performance on analyzing big data due to its sophisticated structure. Traditional models for a machinery monitoring system have been highly dependent on features selected by human experts. Besides, its representational power fails as data gets complicatedly distributed. On the other hand, deep learning models automatically select highly abstracted features while optimization process and its representational power overcomes traditional models. However, its applicability in PHM field has been investigated mainly based on image data. This paper introduces a CNN model for a ‘vibration data’ which is the richest data in manufacturing field. We integrate ‘residual fitting’ mechanism into the CNN structure. As a result, the architecture combines signal re-construction and classification procedures into a single model. Validation results with the rotor vibration data demonstrate our model outperforms any other off-the-shelf feature-based models and other deep learning models recently proposed in the PHM field.
Usage of Accelerated Test Data for Predicting Remaining Useful Life at Field Operating Conditions Dr. Dawn AN, Joo-Ho CHOI and Nam-Ho KIM (Korea Institute of Industrial Technology, Korea Aerospace University, University of Florida)
Prognostics predicts future damage/degradation and the remaining useful life of in-service systems based on damage data obtained during previous usage. General prognostics methods are physic-based approaches when physical models and loading conditions are available and data-driven approaches when only the damage data are available. The damage data are of great importance regardless of prognostics methods used, but it is very expensive to obtain data from in-service systems because of time and cost. Instead, companies frequently perform accelerated tests under much more severe operating conditions for the purpose of design. This paper presents a method of utilizing accelerated degradation data for the purpose of prognostics. As an example, crack growth data are synthetically generated under over-loaded conditions, which are utilized to predict damage growth and remaining useful life at field operating conditions. Four different scenarios are considered based on the availability of a physical model and field loading conditions. Using accelerated test data increases prediction accuracy in the early stages of physics-based prognostics and also compensates for the insufficient data problem in data-driven prognostics.
Imbalanced Classification for Fault Detection in Monitored Critical Infrastructures Prof. Yan-Fu LI (Tsinghua University)
Safety and reliability are among the most crucial factors for the critical infrastructures (CIs). For this reason, they are typically closely monitored and large amounts of data have been collected. Due to their importance, CIs are designed to be highly reliable such that fault cases are rare in the Big-data set. This renders the fault detection an imbalanced classification task. In this work, we developed accurate data mining classifier to tackle this problem. The imbalance ratio of the data can be more than 200.
A Study on Reliability Analysis and Life time Improvement Method by Field Failure of Vibro Hammer Mr. Hankwon LEE and Joo-Ho CHOI (Korea Aerospace University)
Vibro hammer is a construction equipment that uses vibration to drive the foundation piles and H-beams underground. On the work site, vibration is generated and tilting, swinging and jaw opening/closing is made by using the hydraulic pressure and flow rate supplied by the excavator. Since the equipment is usually used overseas or in remote location, the unanticipated failure may lead to excessive down time cost and customer complaint, which is the reason that high durability is demanded. In order to achieve longer life, on-site failure cases have been investigated that have occurred from January 2014 to December 2015. Statistical analysis is carried out to estimate the associated reliability parameters of each module in the equipment, from which the suggestions are made for improving the system reliability.