Continuous Monitoring of Assets Integrity

First steps towards extreme IIoT enabled by machine learning technology applications on data streams to monitor system integrity


This use case has been deployed in a German automotive O.E.M. as a monitoring support device to ramp-up a new powertrain manufacturing line. Specifically, this use case is deployed into crankshaft manufacturing machine tool.


When a new machine is integrated into the production line, it has to meet the required cycle time in order to avoid bottleneck effects. During production line ramp-up, this tuning has to be done. Additionally, during the machine useful life there are many changes on the machine configuration done by machine manufacturer and end-users to continuously improve the machine behavior, that is, adapt to new production requirements or restart after maintenance operations. Along this time, there are many parameters that could be moved in order to meet those requirements affecting the machine behavior increasing its performance, but sometimes reducing drastically its performance jeopardizing the machine integrity. This critical challenge is explained through a real use case, where a machine-tool used for crankshaft manufacturing has to be integrated into a production line with 60 seconds of cycle-time.

Figure 1. Raw data coming from the machine during production.


In order to continuously monitor the machine-tool behavior, the solution approach has to take into account that different combination of machine parameters is ables to produce different results, reflected on the variable readings. That is why, a multivariate data mining approach is followed, where all those effects are analyzed together. In this case, there are relatively high number of variables to have to be monitored at high speed sampling rate. Therefore, extreme IIoT technologies have to be applied in order to have robust and high quality actionable insights about the machine integrity.

In order to measure the effects of each configuration into the machine behavior, a proprietary embedded edge device called Oberon, is connected to the machine. Oberon is able to gather process variables its Sinumerik 840D PLC and CNC system during the production ramp-up. There are other external variables related with peripherals that also have to be gathered. Figure 1 shows raw data coming from the machine, where clear machining cycles could be seen.

In this case, with variables coming from different domains, special commitment has to be given during the acquiring and synchronization stage. Basically, it is very important to have the variables values at the right time, with a well-known acquisition error that let the system compensate it. For this reason, specific computing capabilities from Oberon based on FPGA technologies are used. Thanks them, a complete procedure of sensor fusion is able to be done, give a sub-nanosecond timestamp resolution to each data example.

During this analysis, data has been acquired with the following details:

  • Sampling rate: 100 milliseconds
  • Number of variables: 98
  • Main variables: machine states, cycle time, tool number, spindle and servos data such as angular speed, power and temperature.
  • Sampling time: 3 months
  • Data stream size: 50 MB per hour
By Tdunningvectorization: Own work [GFDL ( or CC BY 3.0 (], from Wikimedia Commons

Figure 2. HMM | Reproduced under: GFDL

The machine learning technology applied in this case are the Hidden Markov Models or HMM, specially tailored for this application.  The main reason for this selection is that HMM are one of the most important sequence processing machine learning models, and in this case, the objective is to analyze the machining sequence to get deeper knowledge on its process depending on the parameter configuration.

HMM, similar to Figure 2, are based are based on Markov chains that are usually specified by the following components:

  1. a set of N states \left \{ x_1, x_2, \ldots, x_N \right \};
  2. a transition probability matrix A with elements a_{ij} each representing the probability of moving from state x_i to state x_j, that is, a_{ij} = p\left ( x_j|x_i \right ), with \sum_{j=1}^N a_{ij}=1;
  3. and special start, x_0, and final, x_F, states that are not associated with observations.

The useful part of these transitional algorithm is that it can detect unexpected changes in the hidden states as the probability matrix to change from one state to the next one does not hold. That means, the likelihood starts to deviate, levaring does unexpected behavior in time.

Figure 3. Power consumption each cycle. Red cycle was taking 2 seconds more than green.


In this case, the analysis done by the HMM algorithm detected that every odd cycle of the machine took 2 seconds more than the even cycle, when it was designed to do the same operation in the same cycle time. A graphical representation of this detected issue is shown in Figure 3. Those two seconds are completely imperceptible to other types of traditional analysis as the difference is extremely lower than the normal operation time.

Therefore, a close investigation on the machine itself found that side A of the clamping system was taking 2 seconds longer because the drilling process in this side was going unnecessarily down into the hole. The main reason of this behavior was related with a configuration parameter, represented in only one line of CNC code, that was not correctly replicated from side B to A.


The CNC program issue found that was penalizing the machining availability by 2 seconds in cycle time has a direct impact on machine productivity of 3%, which in this case means an increase of production capacity in 10,950 crankshafts per year (around $21,9M per year). It is important to highlight that the issue was found during the production line ramp-up phase, however, this system is continuously monitoring the machine to detect possible future system issues that could compromise machine integrity.

© 2018 Aingura IIoT All Rights Reserved.