How can predictive maintenance become a success?

02/21/2019 Know-How

Preparing today for tomorrow. If you postpone maintenance, you could well experience sudden machine and belt failure or faulty production. Regardless of whether the delay was due to cost savings, lack of skilled workers or loss of motivation: Thanks to predictive maintenance, this can no longer occur. However, a few points must be observed in order for this to succeed.

Each predictive maintenance project begins with the detection of machine characteristics during normal machine operation, such as acoustic or natural frequencies of structure-borne and air-borne noise, which are then digitized and transmitted to a computer or the cloud. In the case of local data processing, we speak of edge computing. This is where local AI solutions such as Intel Movidius can be applied. In terms of cloud computing, an unlimited offering for data analysis is available from third party service providers. The edge solution is able to respond in the millisecond range. An Internet connection is, nevertheless, indispensable for firmware updates and remote monitoring. In principle, however, swarm intelligence is not used for learning and improvement processes and is limited to local computing power and your own experience history. Cloud computing, on the other hand, enables comparisons to be made with all the systems operated in the field and can draw conclusions from changes in individual systems to other systems. In addition to this swarm intelligence, there are no limits in terms of computing power or memory capacity and you can flexibly change the logarithm used - for example, from static data analysis to machine learning or deep learning.

As different as the two implementations may seem, an Internet connection and a local initial analysis of the sensor data are always necessary in a real-life scenario. However, both the scaling of significant components and the ongoing maintenance costs vary accordingly, which makes a precise cost-benefit analysis even more complex.

Positioning of the sensor is a decisive criterion

Whether analyzed locally or in the cloud, it is important to clarify where damage can occur and where it can best be detected. But can the sensor also be installed here? Is the site accessible and is sufficient space available? Is there excessive or loud ambient noise? Is it constant or does it occur only at irregular intervals?

Once the ideal installation location has been clarified, the sensor type is often already determined: If everything speaks in favor of attachment to the device or the machine, it is about detecting structure-borne noise. Thus, a shock and vibration sensor or an acceleration sensor is the implement of choice. When placed outside the device or the machine, air-borne noise is detected. MEMS microphone sensors with a specific frequency range are available for this purpose, for example from STMicroelectronics and Infineon. As they always have an opening to absorb the sound waves and to reduce the sound pressure, they are not suitable, without special measures, in humid or dusty environments. In this case, shock and vibration sensors or acceleration sensors can be used.

To answer the question of which frequency range a predictive maintenance system should cover, the following rule of thumb can be applied: The higher the detected frequency, the earlier damage can be registered. In the ultrasonic range above 16kHz, initial signs can be detected months before the damage would actually occur. When detected in the audible range up to 16kHz, there may be only a few weeks left before damage happens. This may be sufficient time for some machines or equipment, but too late for others - this must be checked on a case-by-case basis: How serious can the potential damage be? How long does it take until spare parts are obtained and a service technician is available? Is it possible to stop the machine at any time to carry out repair or maintenance work? For instance, detection in the ultrasonic range is highly recommended for a leak test on gas pipes. If a leak generates audible noise, it is already too late for predictive maintenance.

If detection in the audible range is sufficient, the type of machine or machine part determines which frequency range the sensor should cover. The faster the relevant parts rotate, the higher the frequencies to be detected. For example, damage to air guidance systems is usually caused by unbalance, incorrect adjustment or loose connections. This takes place in a range of about 2kHz. With very slow moving parts, an acceleration sensor instead of a microphone sensor can provide better results.

Microphone, acceleration, shock, and vibration sensors can be combined to increase the number of hits during error detection. Even more information is possible when using other sensor types, e.g. for temperature, humidity or pressure. This type of combination offers the greatest benefits when the sensors are networked with each other. However, this not only pushes up the costs for the sensors and the connection but also results in more data and higher evaluation effort. The combination of several sensors is, thus, only worthwhile if there is a corresponding potential damage, for instance due to belt failure or faulty production, which may even go unnoticed for a longer period of time. This can also be useful for systems in remote areas, e.g. offshore wind farms, as unnecessary engineer call outs result in high costs here. Comprehensive detection of safety-critical systems, such as the braking system in a car, is particularly recommended.

New wireless technologies for data transfer

Depending on the application, individual sensors must first transfer their measurement data to a local data collector. Microcontrollers with integrated radio interfaces and integrated AD converters, so-called wireless SoCs, are ideal for this purpose. Quite often radio stacks are already supplied free of charge and tailored to the microcontroller, so that only the application, i.e. the digitization of the analog values and transfer to the data collector, still needs to be implemented with a few program lines. The data collector can now evaluate the data locally and use its gateway function only for software updates or occasional reporting. In this case, LTE would be a more than sufficiently fast Internet connection, which will also have a secure infrastructure for many years to come. For a time-critical analysis of data in the cloud, where feedback is required within a few milliseconds, 5G will be able to hold its own. The connection of the sensors to the data collector cannot always be achieved with cables. Radio technology is usually cheaper, more flexible, and more durable. With an nRF52840 from Nordic Semiconductor, you can easily choose between Bluetooth mesh, ZigBee or Gazell, a free open source stack for star topologies. NFC enables an uncomplicated connection of the sensors to the respective data collector. For the first time, sensors can be calibrated with a laptop via the integrated USB port. Users who know from the start that they will only use Bluetooth 5 or Bluetooth mesh can also switch to cheaper variants, such as the nRF52810. The latest Bluetooth 5 version enables a range of up to over one kilometer in long-range mode. This makes the technology interesting even where SubGHz technology was previously indispensable.

The new LTE categories are suitable for sensors that do not use data collectors, or data collectors that only need to transfer small amounts of data to the Internet due to strong data compression through edge computing. They make it possible to establish a direct Internet connection from the sensor to the cloud and to transfer the measured values to the cloud without a separate gateway.

The new LTE categories

The latest LTE categories NB1 and M1 - also known as NB-IoT and LTE M1 or LTE-M - are ideal for applications such as predictive maintenance, where small amounts of data need to be transferred in isolated cases.

Both LTE-M and NB-IoT are supported by Nordic Semiconductor's nRF91 family. The highly integrated SiP (system in package) comes with an ARM Cortex M33 microcontroller for custom programming of the application, sensors, and actuators. Its computing power enables the application of more complex algorithms for data analysis. This means: The wireless module generates information on site from the measurement data provided by the sensors, so that only a much smaller amount of data needs to be sent. This optimizes the overall energy balance and keeps online data consumption at a low level. In addition to the sensors, LEDs can also be connected via 32 GPIOs, for example as an on-site warning if a sensor detects a value that is too high. It is also possible to connect buttons or switch relays.
For example, the sensor point can switch off entire systems if required or the user can acknowledge machine states.

The nRF91 SiP is still available with integrated assisted GPS. Through the use of NB-IoT or LTE-M, this enables fast position determination during a cold start for monitoring vehicles or other mobile devices.

Protection against data theft

Since the measured values of the sensors can provide a wealth of information on the use of the machines, systems, and equipment concerned, they should be protected against unauthorized access. In this case, the nRF91 also already contains a solution: The host processor with TrustZone uses a trusted execution environment in the CPU and in the system, thereby contributing to the security of application data, firmware, and connected peripherals. ARM CryptoCell ensures secure memory access while TLS and SSL ensure end-to-end encryption of data transmission. The nRF91 is also perfectly suited for interaction with an nRF52, as implemented on the nRF91 development kit. Thus, both a short-range radio network for sensor connection and a cellular network for Internet connection are available with this multicore two-chip solution. If you choose the nRF52840 from the nRF52 family, it also features ARM TrustZone and CryptoCell technologies.

Factor of success - data analysis

Once the data have been transferred from the sensor, the trickiest task is data analysis. What does it mean when the frequency of a roller bearing has changed? Is it at risk of failing, was the production process simply altered or was the machine shut down for the weekend? Or is an interference factor responsible for the change? Which deviations still belong to normal fluctuations? And finally: How high is the probability of damage occurring, i.e. when does intervention become necessary?

This results in specific profiles, which are stored in the software by corresponding parameters and threshold values. Readjustments may be necessary after the first practical test. The predictive maintenance system also needs to be adapted in the event of production alteration, changes to machinery or similar. If you take all these points into account, you are on the right track to: Never again experiencing unexpected machine damage, downtime or belt failure due to undiscovered aging systems. The expenditure for maintenance work can be better planned in advance and only the spare parts actually required can be kept in stock automatically. This benefits not only users but also machine manufacturers. If they integrate a predictive maintenance system into their products, they offer customers real added value through greater machine availability. In addition, they can use evaluated field experience for further product development.

What is predictive maintenance?

In contrast to preventive maintenance, predictive maintenance is not based on fixed maintenance cycles but on demand-oriented maintenance that utilizes measured data collected continuously on site and the respective data evaluation. It registers vibrations or altered noises from machines, systems, and equipment that indicate problems during operation long before actual damage occurs.

Find components at <link _blank external-link-new-window "open internal link"></link>

Subscribe to our newsletter to stay updated: <link _blank external-link-new-window "Opens external link in new window"></link>