Fingerprints helping preventative maintenance

Machine-tool spindle head for crankshaft manufacturing

Using unsupervised machine learning to develop fingerprints to track assets health


This use case si deployed into Spanish machine-tool manufacturer premises and also into its main automotive O.E.M. customers around the world. Specifically, this use case is oriented to machines equipped with high power spindle heads.


Spindle heads are a critical part of a machine-tool, in charge in a large percentage of the material removing operation. If the spindle head fails, the machine is stopped. Because the machine is part of a manufacturing system, a breakdown stops the complete production line. Depending on the production facilities, the unexpected downtime could cost around $50k per hour additional to the +$20k spindle spare. A broken spindle head could take up to 40 hours (five 8-hour shifts) to be replaced, if the damage is contained only at the spindle part. As these elements are rated from 14 to 24 kW, and unexpected failure could also cause collateral damages, increasing the costs of replacement.

Therefore, a continuous analysis on the spindle head health and other critical parts is important. Even during installation on a brand-new machine, as the useful life could be compromised during spindle manufacturing or stockage.

Figure 1. Machine-tool spindle head for crankshaft manufacturing


In order to have a general view of the spindle head health, the Aingura IIoT approach is to develop a data-based fingerprint where all the related variables are analyzed in a multivariate manner during real operation. Therefore, this fingerprint is developed for each spindle head during first stages of real production, such as capacity certification among others. Once the fingerprint is developed, it could be used to compare the behavior pattern itself along its useful life or between other spindles performing the same operation.

In order to get the required data to build the fingerprints, a proprietary embedded edge device called Oberon, is connected to the machine. Oberon system is able to gather process variables its Sinumerik 840D, PLC and CNC system, and other external sensors that can bring a clear idea of spindle head, such as accelerometers and high sampling rate energy measurements.

The data required for these type of analysis has the following details:

  • Sampling rate:
    • NCU: 4 milliseconds
    • Accelerometer: 32 microseconds
    • Energy: 250 microseconds
  • Number of variables: 29
  • Main variables: time stamp, machine states, power, angular velocity, torque, temperature, bearings frequencies and energy consumption at the three phases.
  • Sampling time: 10 machining cycles
  • Dataset size: 270 MB


By Smason79 [CC BY-SA 3.0 (], from Wikimedia Commons

Figure 2. Gaussian Mixture Model | CC BY-SA 3.0

The machine learning solution to analyze data and produce the spindle head fingerprint is unsupervised technique based on Gaussian Mixture Models (GMM) clustering. GMM clustering performance is ideal for manufacturing data, as shown in our work titled: “Machine Learning-based CPS for Clustering High throughput Machining Cycle Conditions”, as it is able to partition different machine states during production. Other important aspect is that GMM is considered a technique of “soft” clustering, as it assign each data sample to a component using a probability. Therefore, one data sample could assigned to all the clustering components with different probabilities. Other clustering techniques force data sample assignment to each cluster, losing a critical amount of information. This feature is really powerful, as it let us inspect those data samples that are near the decision border..

Basically, GMM is a probability-based clustering technique that fit the sampled data density distribution to a finite mixture model, in this case, a Gaussian mixture model that has the form:

f(x; \theta) = \sum_{k=1}^{N} \pi_k f_k(x;\theta_k)

where k is the number of components to cluster the data and the parameters \pi and \theta are estimated using the Expectation-Maximization algorithm. Once the parameters for the GMM are found, the data samples are assigned, with a probability of belonging, to each component.

Right now, this solution is designed to work with historical/static data. However, Aingura IIoT is working to enable data stream processing using GMM, giving the option to monitor the fingerprint of an asset in real-time, as shown in our work titled: “Clustering of Data Streams with Dynamic Gaussian Mixture Models. An IoT Application in Industrial Processes”.

Figure 3. GMM of spindle head during crankshaft manufacturing


Figure 3 shows a 2-D view of the fingerprint pattern of a spindle head used for crankshaft manufacturing, built using a GMM clustering technique. In this case, the fingerprint is clustered into three components, where Cluster 0 contains data from the material removal operation, Cluster 1 contains data from the unload rotation and Cluster 2 contains data when the spindle is stopped. It is important to highlight that the real power of this fingerprint is that has 29 dimensions, that is, it has different views where the spindle behavior could be analyzed and/or compared.

Figure 4 shows a 2-D view of a different spindle head doing the same operation of the spindle in Figure 3. This figure has been done with the same data structure shown previously, where high speed sampling is able to catch different spindle head behavior, including some non-stationary states.

Comparing the shapes for each spindle, a similar behaviour of the spindle head mechanics could be concluded, where small differences could be caused by anisotropy related to spindle fabrication.

As these fingerprints were obtained during machine-tool capacity testing and certification (real production) a long term comparison could be done along the spindle useful life in order to detect changes in each of the different states.

Figure 4. A second GMM fingerprint of a different spindle head during crankshaft manufacturing

With these references created a generalization could be done using a consensus approach, where fingerprints coming from different spindles that work in the same operation could be joined together to create a reference pattern to compare during the initial stages of spindle operation.


The main benefit with is type of fingerprint is that it can help preventative maintenance to detect when a degradation is starting to develop in critical parts such as spindle heads. By doing this, a reduction of unexpected failures represents savings of $2M on downtime costs and $200k on spare parts, depending on the production facility.

© 2018 Aingura IIoT All Rights Reserved.