Predictive Maintenance of Mining Machines—Problem of Non-Gaussian Noise

Abstract

Predictive maintenance has been recognized by companies as a promising approach for reducing machine operation costs. It could be done in online and offline mode. In any case, the key issue is how to process acquired data, how to interpret diagnostic features, and how robust the final decision is. Unfortunately, in case of mining machines, classical diagnostic techniques used commonly in various applications often fail in the sense of false alarms, lack of detection, etc. The problem has been first recognized as an influence of time-varying load/speed conditions. Recent research done for mining machines showed that one might obtain misleading results due to contamination by non-Gaussian and impulsive noise as well. Source of such non-Gaussian noise could be related to process (mechanical excavation, crushing, milling, etc.) or natural operation of the machine (compression, pumping), can be transmitted to the sensor from other machine operated nearby, and might have purely random, accidental nature, etc. In any case, one needs appropriate tools to process such noisy data. According to our experience, a good example to illustrate the problem and one of the most challenging machines is the copper ore crusher, where regime of work produces impulsive noise related to falling pieces of ore into crusher, as well as crushing process. Classical methods based on the detection of impulsivity will not provide satisfactory results. One needs to take into account cyclic nature of fault-related signal. For this class of diagnostic problems, we have developed a set of tools that is based on alpha-stable (heavy-tailed in general) distribution. Framework for localized damage detection in gears or bearings is as follows: we use time–frequency decomposition of the vibration signal to obtain set of sub-signals with simplest structure instead of one complex signal. Extracted time series from each of the frequency bands can be analyzed separately in the context of impulsive and cyclic nature. We have adopted advanced modeling techniques defined for signals with heavy-tailed distributions. We are able to identify so-called informative frequency band (to asses which sub-signals are informative) and identify cycle/period of disturbance by generalized measures of dependence for heavy-tailed signals (codifference, covariation, Fractional Lower Order Covariance—FLOC). Note that due to impulsive character of the signal, one cannot apply methods requiring finite variance (autocorrelation should not be used). Thus, authors proposed to apply appropriate counterparts for heavy-tailed data. In this paper, we will summarize the proposed approach and present the latest examples of applications to vibration data from mining machines.