Machine learning applications in the manufacturing industry have been around for many years. In the early days (10 years ago), they were reserved for large corporations that had extensive resources. The applications were very expensive to purchase and deploy. Once deployed, the applications required extensive engineering staff to maintain the asset models and evaluate the results of the machine learning application. Today, there are a number of machine learning applications for industrial applications that can be deployed in a short period of time and, in some cases, on a trial basis. In this paper, I will give an overview of how to utilize machine learning applications to monitor assets, discuss some of the different machine learning applications that are currently offered and detail future applications.
Machine learning applications are utilized to identify machine failure points at the earliest occurrence. If you are familiar with a PF curve, you know that the earlier you identify a potential failure, the better. In the machine learning software applications, you begin by building a model of the asset. The model consists of all process parameters and manufacturing equipment parameters that are associated with that particular asset. These parameters are typically stored in a data historian that captures data from the plant DCS, associated PLCs, electronic logs, etc. If we use a pump as an example, suction pressure, discharge pressure, control valve position, bearing temperature and bearing vibration would be a few good examples of parameters to include in a model. Most models have between 10-30 parameters, but we have some models that have close to 100 parameters.
Once the model is created, historical operational data is imported into the model. This is typically known as the training data set and includes one year of data. One year of data allows the model to account for seasonal changes in operation. A person knowledgeable in the operation of the asset would then identify what data to “include” in the training data set (good operation) and what data to “exclude” from the training data set (bad operation). The machine learning application then uses the trained data set to develop an operational matrix for the asset. The matrix basically identifies how the machine should operate at any given time based on the training data that was used to create the matrix.
Now here’s where the magic occurs, the machine matrix is deployed within the software application to constantly monitor the operation of the machine and predict where the machine parameters should be running based on the matrix that it developed. If a parameter deviates away from the model prediction by a significant percentage, then the system creates an alert condition for that particular parameter. An engineering evaluation is then performed on the asset to evaluate the change in condition. Upon evaluation, three general outcomes are likely to occur: 1) the alert is valid, alert the plant of the condition and work with local resources to resolve the issue; 2) the alert needs additional research/operational data, continue to monitor the parameter and all associated parameters for alert conditions; and 3) the alert is a false positive, retrain the model with additional operational data and redeploy the model.
In any case, the machine learning application requires dedicated resources to maintain the models and resolve manufacturing equipment issues with plant equipment owners. A full-time resource will typically spend 40 percent of the time maintaining models, 40 percent of the time working with plant resources to resolve issues and 20 percent of the time valuing the assists generated from the program.
The results of a machine learning application are very powerful. These software applications identify changes in manufacturing machine conditions or process parameters that are not noticeable to the human eye. Figure 1 shows an increased bearing vibration on a primary air fan due to a loss of oil flow to the bearing. An alert condition came in on the outboard bearing of the fan. The machine learning application had predicted that the bearing vibration should have been around 3.5 mils for the current operating conditions. The bearing vibration had slowly deviated from the predicted value, and the alert condition occurred at 4.7 mils. The plant was notified of the alert condition, and visual inspection of the fan identified that the oil line had developed a leak at the connection to the bearing cap. The suction of the fan was from each end near the bearing pedestals. The fan was actually sucking the oil into fan housing, so there was no indication of the leak on the ground. The oil on the fan blades collect dirt and debris, causing the fan to become out of balance and subsequently increase the vibration. The plant resources were able to take corrective action to stop the leak before the bearing was damaged.
Figure 2 is of a slow decay in hydrogen purity on a large turbine generator set. The green line is the model predicted value. The blue line is the actual value, and the red dots are where the parameter went into alarm. During this month-long trend, the local crews had not noticed the slow decay of the hydrogen purity. The plant was notified in advance of a local alarm or plant shutdown, giving them time to manage the situation without operating in crisis mode.
Figure 3 is associated with the electro-hydraulic control (EHC) system that controls valve position, turbine speed and trip safety valves. In this case, the differential pressure across the EHC pump “A” strainer began to increase. The inspection of the strainer would typically be on the operators’ round sheet for local inspection once or twice per shift. The plant was contacted, and they were able to swap from EHC pump “A” to EHC pump “B.” This prevented a turbine trip and any associated damage that might have occurred.
Figure 4 is associated with the lubrication system on a large pulverizer, also known as a bowl mill. The lubrication system supplies oil to the integrated gearbox and all associated bearings. The asset model predicted the temperature to be 90 degrees F, but it actually trended up to 110 degrees F. Local plant resources were contacted, and they found that the cooling water control valve to the lube oil heat exchanger had failed. The control valve was replaced, and the system returned to normal.
Figure 4. Pulverizer oil temperature
The remaining set of examples are from a second machine learning software application. Although the user interface may be quite different, the principles of operation and the output of the software applications are very similar. The true value of the machine learning application is the communication and follow-up activities that occur once the application has identified a change from historical operating condition.
In the following graphs, the blue line is the actual value, the red line is the predicted value, the light green shaded area represents an alert condition and the magenta vertical line is where the parameter reached an alarm condition due to the deviation. The gray area is where the machine is offline. Notice that no prediction or alarms occur when the machine is offline.
In Figure 5 below, we are monitoring a steam turbine supplied by a heat-recovery steam generator (HRSG) in a combined cycle power plant. The process parameter is the high-pressure spray flow to the steam turbine. The red and the blue lines trend well together until 1,000 pounds per hour of steam flow. The actual and predicted begin to deviate at 1,000 pounds per hour, and a green alert condition occurs within the application.
The two values continue to deviate until an alarm condition, shown by the magenta vertical line. Plant resources were contacted to discuss the steam generator operation and the model prediction.
Figure 6 illustrates the change in vibration on a combustion turbine during start up. The graph shows five separate start-up scenarios, represented by the vertical stripped areas. In all cases, the vibration represented by the blue line is much higher when the unit initially returns to service. In the third start-up, the vibration is actually elevated for long enough to create an alarm condition for the parameter. The plant was notified, and it was found that the vibration sensor had become loose due to normal operation.
Figure 7 illustrates multiple start-up scenarios, with the fourth start-up resulting in an alarm condition. The software application had predicted that the intermediate steam drum pressure in the HRSG to be 278 psi, but the actual reading was 240 psi. The condition was reported to the local plant resources, and the unit was brought offline to make repairs to the control valve. Upon returning to service the following day, the model prediction and actual value returned to normal.
At some point in the near future, these types of software applications will be standard in what we know today as a DCS. The DCS system would build predictions of each parameter based on the correlation to other parameters within the process. As the predictions deviate from the actual, alerts would be sent to the operator with guidance as to what process parameters are driving the deviation. As the machine learns what corrective action to take, the operator requires less and less involvement until you reach a point where the machine can operate itself. This may seem out of reach, but we have examples of this today. Tesla’s autopilot is probably the best example of the capability of machine learning today. In December 2016, Tesla released a video of total autonomous driving from your home to your work with automatic parking and retrieval. It is an impressive video to watch and shows us what the future will hold.
Machine learning software applications can deliver powerful improvements in asset availability, process improvements and production increases when applied to a manufacturing process. Deploying and maintaining the software applications require specialized skills, but the barrier to entry is much lower in recent years. As these software applications become more mainstream, the cost of the application will decrease even further. At some point, applications built for the general public will make their way into the industrial environment.