Remaining Useful Life Estimation of Ball Bearings

Remaining Useful Life Estimation of Ball Bearings

Why Predictive Maintenance is needed?

Maintenance and reliability professionals in the manufacturing industry face a number of challenges, but the goal of any maintenance organization is always the same i.e. to maximize equipment availability. In order to achieve maximum availability, run-to-failure maintenance suggests that the equipment must be kept running until it fails. However, this strategy may result in catastrophic machine damage as components begin to vibrate excessively, overheat and break. This consequently stops the machine operation causing a highly expensive and time-consuming unplanned downtime. Conversely, frequent planned machine downtime and inspection increases replacement costs and causes disruption to operations. Therefore, there is a need for a well-designed prediction algorithm to preempt component failures and help plan machine downtime minimizing both the expensive repairs as well as preventing adverse outcomes from a component failure.

Machine Learning opens new horizons in Predictive Maintenance

Almost all mechanical systems involve rotating machinery such as motors, pumps, bearings and gearboxes in various industrial processes. They are, generally, designed to run for many years without unplanned downtime. However, these components sometimes malfunction prematurely due to improper installation, irregular operating conditions, increased mechanical or electrical load or some other reason. Therefore, condition-based monitoring of these types of equipment helps to boost productivity and profitability. Currently due to advent of Industry 4.0, machines and sensors are augmented with wireless connectivity and the condition governing parameters of machines such as vibrations, heat and energy consumption are monitored at all times. The collection of this sensor data opens up the opportunity to use machine learning based prognosis techniques to enable the machinery to intelligently make decisions on its own.

How Qult helps in industrial fault prognosis?

This type of data collected from actual mechanical systems is processed  by our experts to develop fully customized solutions for the problem under study. We designed an end-to-end solution optimized to enforce the need to schedule maintenance and replacement for anticipated failure events by monitoring vibration signals generated by the operating machine. 

Our solution uses advanced signal processing algorithms to process vibration signals and distill  the meaningful trends from raw sensor data. Several well-designed approaches are deployed to identify degeneration effects and spot the characteristic vibration signature before failure.

Rolling elements when repeatedly pass over a defect such as a spall or a crack, generate impulsive vibrations. Our solution can identify different harmonic components associated with the defect frequency and a peak at defect frequency or at harmonics indicates a possible bearing problem. Hence, spectral analysis methods such as wavelet transform play a crucial part in extracting the useful information from the vibration signals, that further leads to formulation of a relevant feature set. Another way to extract meaningful features from data is to count upon probability and statistical theory. Due to the uncertainty associated with vibration signals, these vibrations can be modelled as Gaussian Random Process. Suitable signal features such as spectral kurtosis, Reni entropy etc can then be extracted from the chosen model that, ideally, must exhibit a continuous trend which can be directly related to the deterioration of bearing condition. Finally, regression-based machine learning approaches assist in constructing a health indicator as a quantitative measure to quantify bearing degradation. The failure threshold for the health indicator represents the end of life point. Since neural networks are capable of approximating a function of their inputs really well, a non-linear mapping from health indicator to the leftover useful lifetime of the rolling element can further be acquired. Furthermore, Long Short-Term Memory (LSTM) neural networks can also be used for deep feature extraction and end-of-life estimation and they have shown to perform well due to the presence of memory units and their ability to exploit temporal dependency in time-series data.