BMT default image, depicting a supergraphic on a grey background

Data Analytics for Machinery Health

Machinery typically have all kind of moving parts and particularly gears and bearings. All the parts are exposed to wear and tear; thus, they could fail if not maintained properly or replaced on time.

23 January 2020

Energy and Resources

Data Analytics for Machinery Health

In this short study we will focus on identification of the abnormal bearing conditions that would result in bearing failure. The conditions under which a bearing works include external loads and vibration as well as the internal vibration created by bearing itself. In addition, internal friction as well as the temperature and humidity contribute to degradation of bearing physical integrity that in turn reflects as an increase of bearing vibration, which can be easily recorded by accelerometers. Based on characteristics of the measured accelerations, the health status of the bearing can be assessed (see Figure 1 and Figure 2). However, by visual observation, the first indication of the Bearing 1 failure could be observed in the morning on February 16, 2004 (Figure 2).


Figure 1: : Acceleration signals for healthy bearings



Figure 2: : Acceleration signals for failing bearings


Let’s see if Machine Learning (ML) process can be used to detect the health issues with the Bearing 1 earlier than February 16, 2004.

A TensorFlow unsupervised neural network model is used to reconstruct the measured incoming input acceleration data. The model was first trained on the set of acceleration signals from healthy bearings. We used four bearings that operate in the same kind environment, so using characteristics of all would improve the robustness and accuracy of the model through so called tribal learning. The healthy signals are split into training and validation data sets, which became independent. The training set is used to construct an unsupervised ML model. The validation set was used to validate the same model for bias and variability. The results of these operations are shown in Figure 3. It is obvious that the ML model seems to be accurate as the validation errors are very close to the training errors. Once the model is trained and validated it is used first to encode the acceleration data to a new space which represents the expected “healthy bearings” acceleration data.


Figure 3: Performance of training and validation processes


The differences between encoded and actual healthy data are shown as a histogram in Figure 4. From this histogram, the threshold between healthy and faulty data (concern them as outliers as well) can be easily determined. It is important to realize that this “threshold” value is not determined as a preset parameter based on the expected values, as it has been usually done in engineering practice, but rather based on a nonlinear ML model and the error between actual and modeled data.


Figure 4: Histogram of differences between measure healthy and encoded healthy data


Now, to identify if any of the bearings has tendency to fail, the new incoming data can be encoded using the trained ML model and to observe if errors between the real and encoded values exceed the established “healthy threshold value”. Results are shown in Figure 5.


Figure 5: Error trend between actual (measured) and encoded ML data


From Figure 5 it is obvious that the first visible indication that Bearing 1 tends to fail is associated with February 16, 2004. This indication is even much more visible from Figure 5 than from Figure 2.

A closer inspection of the results shows that a faulty condition for Bearing 1 could be traced 3 days earlier from the date when it would become obvious to the machinery operator. Figure 6 represents the sequence of Bearing 1 failure condition and potential indication for triggering the alarm. In this example, the alarm conditions are checked in 3-hour intervals. From Figure 6 is obvious that at the beginning there was a mix of healthy and faulty information and that alarms were sparsely triggered until February 16, 2004 when a fast progress of Bearing 1 started. The alarms were triggered regularly over the next 3-day period until Bearing 1 failed completely.


Figure 6: Alarm indicator for Bearing 1


In summary, the ML technique can be used to create a health checkup of the system with the goal that the change in health conditions can be detected as soon as possible. In this case, we demonstrated that the unhealthy condition of Bearing 1 had been detected 3 days before than it became obvious to the operator.

This example also demonstrates that ML method could be used in real time operation and equally successfully over other quite complicated engineering methods that require spectral and other more sophisticated techniques that cannot be used in real time and that require highly trained experts to interpret the results. Moreover, similar ML technique can be used for other systems than just for bearing health monitoring.

For more information please contact:

Igor Prislin, PhD OE/NA BMT CI North America




Share this