In March 2021 our AI detected an impending critical fault in the piston compressor, which later culminated in machine downtime. Early detection of this fault represents a concrete result and proof of how machine learning algorithms can assist in predictive maintenance.
AI deployed on critical manufacturing equipment
In 2021 recognized European automotive manufacturer chose acoustic anomaly detection from Neuron soundware for its automotive manufacturing facility to maintain continuity of operations for the tempering furnace.
Two piston helium compressors are equipped with this technology and pressurizing helium tanks. Pressurized helium is used for cooling (hardening) of the workpieces on gearbox parts in the hardening chamber.
If this equipment fails, the manufacturer runs the risk of delays in production. Extended unplanned outages can even lead to losses of millions of Czech crowns. This makes it essential to prevent such outages before they occur.
Fig. 1 J.P. Sauer&Sohn model WP 318L cylinder piston compressor with sensors installed on the electric motor and pump pistons.
Detecting anomalies and upcoming faults in practice
Fig.: timeline of events from installing IoT equipment through to the fault
- In January 2021, the sensors and IoT equipment were deployed on the compressor.
- February 1, 2021: the initial data set began recording on the compressor.
- February 5, 2021: after collecting the initial data set, the AI detection model was deployed on the compressor. This means that the AI spent five days “practicing” on the compressor’s nominal data (behavior) so it would recognize the sounds of normal operation and learn to distinguish sounds that are not normal and may indicate a fault.
- March 12, 2021: Scratches were discovered on piston and cylinder 2 during the client’s regular maintenance on March 12; these were repaired and the compressor tested. When analyzing the signal afterwards, our Neuron soundware diagnostician determined that it was probably pre-existing damage that made it into the initial data, which is why the algorithm did not detect it.
- March 14, 2021: the compressor was brought back into operation between 8 and 10 p.m.
- March 16, 2021: the Client performed control measurements on the compressor for vibration and temperature. The results were within normal parameters.
- March 16, 2021: at 12:11 p.m., shortly after the control measurements, the anomaly score on the Neuron soundware platform increased, indicating a degree of machine damage above the threshold at which the model sends out an alert. Alerts went out three times:
- 12:11 – 12:21 p.m. (within a span of 10 minutes) on March 16
- 3:30 – 3:41 p.m. (within a span of 11 minutes) on March 16
- 4:00 – 5:31 a.m. (within a span of 91 minutes) on March 17
Unfortunately, due to other tasks the diagnostic shift did not have time to go check the compressor again.
- March 17, 2021, about 4:00 a.m.: cylinder 2 showed signs of loss of pressure. Because of the check that had just been performed and because a similar fault occurred in the compressor last year, maintenance staff decided to replace the compressor instead of repairing it.
- March 17, 2021, 5:00 p.m.: after about 12 hours out of service, the new compressor is put into operation.
Fig. Sample anomaly detection in nShield
What is the value of installed monitoring services?
NSW diagnostics repeatedly predicted the nominal status on cylinder 2, compressor, even though the conventional means of evaluating machine condition via vibrodiagnostics and temperature measurement, performed several hours earlier, did not detect the impending fault. The first alert came 16 hours before the fault occurred and the compressor was removed from operation.
Since the assessment of the alerts did not lead to a suspension of production, this case shows us the value of the monitoring solution in practice. This also provides Neuron soundware with a valuable data sample to help improve the algorithm further.
The case also offers a good illustration of the consequences of machine faults that are not detected in time.
This fault caused 12 hours of idle time for this recognized European automotive manufacturer and the gearbox hardening process, including four hours emptying out the hardening chamber and 80% of the production waste intended for disposal due to the interrupted hardening process.
Key outcomes of this case study
Since compressor faults occur relatively frequently and the equipment plays an essential role in operations, the technology installed has a great deal of added value. Early detection of an upcoming failure can help detect a machine fault and prevent breakdowns in production that can result in losses of millions of Czech crowns.
The best way to prevent faults on industrial machines is this: Install IoT equipment and sensors on a critical machine, collect nominal data during machine operations to “train” the AI model (takes a matter of days), and then immediately deploy a continuous monitoring service on the chosen machine. It is important to respond quickly to alerts and decide on the appropriate response to minimize losses.
The service’s value to the customer grows over time. The more anomalies it detects, the better the AI gets at diagnosis.
At first, when the monitoring solution is deployed on the machine, the customer receives value in the form of remote equipment monitoring. They know whether the machine is up and running or experiencing a failure, and they have access to data for their own analyses and inspections. They then receive information on anomalies in the equipment. This information is first expanded with the qualified advice of a human diagnostician. Eventually, with an advanced trained model, AI is able to assess these anomalies on its own.
AI-assisted monitoring is an investment in the future and only grows over time. Production directors who make this investment have a clear advantage over those with preventive maintenance only, who have to resort to expensive human staff for all maintenance tasks.
Since fewer experts are available on the market and pressure on innovation and digitalization is growing, this is the time to invest in AI-assisted monitoring. If you would like more information, please do not hesitate to contact us at www.neuronsw.com.
Appendix: Key equipment parameters – piston compressor and sensor installation details
The monitored machine is a J.P. Sauer&Sohn piston compressor, model WP 318L, with label values: capacity 235 m3/h, power 58 kW, working pressure 40 bar, rotations 1480 rpm. The compressor works in start/stop mode, spending about five minutes bringing helium tanks to pressure with about 10 minutes of downtime afterwards. The compressor starts up gradually: level 1 – cylinders 1 and 2 create pressure of 2.6 bar, level 2 – cylinder 3 creates pressure of 12 bar, and level 3 – creates pressure of 36 bar.
The piston compressor is made up of a cooling mechanism connected to a block with four cylinders attached in a semicircle 45 degrees apart. The block is also connected by flange to the electric motor.
Fig. 1 J.P. Sauer&Sohn cylinder piston compressor model WP 318L.
Fig. 2 Disassembled, damaged piston from cylinder 2 with jammed piston rings.
Measurement points – sensor placement
Fig. Cylinder 1 – Sensor 4
Fig. Cylinder 2 – Sensor 3
Fig. Cylinder 3 – Sensor 2
Fig. Cylinder 4 – Sensor 1
Fig. Motor – Sensor 5
Fig. Motor flange – Sensor 6