Solutions for predictive maintenance of machines
Manufacturing companies in Germany are investing in Federal Statistical Office invest over 50 billion euros a year in machinery and equipment. Efforts to prevent breakdowns are correspondingly high. Predictive maintenance is therefore becoming a central component of Industry 4.0.
Production chains are becoming increasingly complex. This means that they are affected by defects much more frequently and to a greater extent. An outdated maintenance strategy can reduce a site's production capacity by up to 20 percent. More and more companies are therefore turning to predictive maintenance. In contrast to preventive maintenance, predictive maintenance relies on the prediction of maintenance requirements for machines and systems in order to minimize downtimes and reduce maintenance costs. Sensor data is used for this purpose, with which many machines are already equipped at the factory.
Preventive maintenance means carrying out maintenance measures according to predefined criteria, for example changing a light bulb every 7,000 operating hours. Predictive maintenance, on the other hand, means that the maintenance effort is automatically determined on the basis of key figures and the relationships between these key figures and optimized for low downtimes and high machine effectiveness. Predictive maintenance can increase the availability of production systems by 10 to 20 percent compared to preventive maintenance, while reducing maintenance costs by 5 to 10 percent.
Example Rolls Royce - Extended time between maintenance by up to 50
The engine manufacturer optimizes the maintenance intervals for each engine individually and can increase the time between two maintenance intervals for an engine by up to 50 %. For other engines, the need for maintenance can also be identified earlier. This leads to an immense reduction in the stock of spare parts and an increase in the efficiency of engines with overdue maintenance requirements. The motors are monitored while they are in use. If maintenance is required, maintenance options are automatically checked and appropriate measures are planned. Expensive and unplanned breakdowns are thus avoided.
How does predictive maintenance work?
In short, predictive maintenance in Industry 4.0 means that the maintenance requirements of systems and machines are determined automatically and individually. In doing so, an algorithm constantly approaches the optimum time for maintenance in order to minimize downtimes and save costs. The optimum interval between two maintenance tasks can vary greatly between machines of the same model, as each machine is analyzed individually.
For predictive maintenance, an algorithm learns certain patterns from data that indicate a need for maintenance. Sensor data is used in almost every solution, but more and more systems are using additional image recognition or external information such as the type of material used. For example, fatigue in the material can be detected by a change in sound. Sensors then detect sound frequencies that are inaudible to humans. This allows signs of an impending defect to be detected up to two weeks earlier.
Collection of sensor data
The data is analyzed using sensor data. A third of all machines in use in Germany are already equipped with sensors. New models in particular are increasingly being fitted with integrated sensors. Data such as temperature, operating time, speeds, pressure and vibration are used particularly frequently for predictive maintenance.
Generally speaking, large amounts of data make the determination of maintenance costs much more reliable. Many SMEs that have had little contact with big data and artificial intelligence so far complain about hurdles in implementation, do not have the necessary skills within the company and are facing high costs. However, these challenges of Industry 4.0 can be met competently with the right long-term strategy.
Example ALCOA - 20 % reduction in machine downtimes
The American aluminum manufacturer is implementing a proof of concept with 50 machines in order to increase the level of service and save maintenance costs. The 50 machines are equipped with sensors that measure the current in the machine, among other things. For example, an unexpected increase in amperage is detected, indicating a faulty motor. The source of the fault, a damaged belt, is quickly found and rectified, so that twelve hours of unplanned downtime can be avoided. By implementing the PoC, potential downtimes of the selected machines are reduced by around 20 %. The investment pays for itself within the first six months.
Challenges and solutions for initial projects with predictive maintenance
Every company operates under individual circumstances, framework conditions, requirements and objectives. However, there are also a number of challenges that are regularly encountered in predictive maintenance projects. These challenges can be categorized and then systematically solved.
Data acquisition
Data forms the basis for every learning algorithm. But first of all, it has to be collected. Data acquisition and data management work when data is collected, transmitted and stored. More and more systems have integrated sensors that collect data automatically.
Solution for data acquisition
Data capture is a composition of several sub-processes. Each of these sub-processes should be subject to best practice standards. These standards offer two advantages: Reusability and robustness. Functioning sub-processes can be used for other cases with slight adaptations. This reduces the effort required to set up and further develop data collection systems, while they remain flexible. At the same time, a high quality standard is ensured. The individual components have been tried and tested in practice many times and optimized in such a way that the rapid setup of the system can be combined with high performance. The selection of specific tools is of the utmost relevance. Experience with suitable technologies can speed up the data acquisition process immensely and reduce long-term costs to a minimum.
Data preparation
The entire data capture and all sub-steps are optimized to enable a smooth and fast process. Data is packaged and compressed for this purpose. This allows data to be used very flexibly while keeping storage and transfer costs low. However, for an algorithm to learn from data, it must be converted into a different format. In addition, data can always be incorrect or incomplete. High data quality is a prerequisite for reliable artificial intelligence. The so-called "Rule of Ten" is one of the most important rules when working with data. It states that work steps with high-quality data are 90% cheaper than with incorrect data because they are completed much faster and produce far fewer errors.
Solution for data preparation
Choosing the right tools is also crucial for ensuring high data quality. The development of effective processes for controlling data quality is crucial here. An effective control process provides insights into which quality deficiencies exist and how these can be rectified.
The use of engineering best practices ensures that data quality can be maintained at a high level. Tools such as dbt make it possible for young, small data teams to implement best practices in data preparation. Properly configured, dbt ensures that defects are detected before they actually occur.
Data combination
Artificial intelligence is based on recognizing rules and patterns in data. In order for this to happen, these patterns must be contained in the data. One of the most important success factors for predictive maintenance projects is the enrichment of data. Certain regularities can only be uncovered by adding further information.
Solution for data combination
At first glance, combining data from different source systems may seem like an additional expense. In fact, this step can actually reduce the costs of predictive maintenance and provide additional flexibility.
In order to identify meaningful possibilities for data enrichment, technical experience in the use of systems and machines is of the utmost importance. Through the exchange between AI developers and production employees, possible correlations are exchanged directly and can also be integrated into the predictive maintenance system in the shortest possible time. All information used is assessed according to its added value by comparing the improvements in performance and effort. This comparison takes place automatically and is permanently available for further optimization and controlling.
Automotive manufacturing example 30 % increase in machine uptime
A car manufacturer increases the uptime of its machines by 30 % within 24 months of the start of the project. 30 %. Temporarily installed sensors are used, which only need a few minutes to scan each machine. The investment in sensor equipment therefore remains low. The sensor data used includes vibration, infrared temperature measurement, ultrasound and electrical voltage.
Success factors for projects with predictive maintenance
The success of predictive maintenance projects can certainly be predicted if important basic requirements are met. Companies therefore urgently need to do their homework in the area of classic data analysis. After all, there is one rule of thumb that must also be observed in Industry 4.0: First data analytics, then data science.
Transparency of the maintenance effort
Even if a sufficient database is available, this does not mean that it can be seen. To ensure that the data is not only collected on the server or in the cloud, but is also visible, it must be visualized and made accessible to all those responsible. This can be implemented via dashboards, automated emails and integration into existing communication tools such as Slack or Trello. The work of a production manager or technical asset manager is reliably supplemented with this data, which is often delivered in real time. Professional data analysis via a modern data stack is therefore a guarantee of success for predictive maintenance.
Prediction of the maintenance effort
When and how much effort is required to maintain machines can be predicted with the help of artificial intelligence and the use of corresponding machine learning algorithms. Algorithms make machines capable of learning. It is important that the measured data is monitored smoothly in order to create system diagnoses, automatically detect deviations and suspected problems and provide a forecast of the usable remaining service life. Which AI algorithm is used depends on the machine being monitored. A distinction is made between classification and anomaly detection. While classification is a sensible option for monitoring machines with a high failure rate, anomaly detection is used for devices that are hardly susceptible to faults. Monitoring them is much more complex, as there are rarely any signs of failure.
Raw data from sensors
For predictive maintenance, sensors are used that are already integrated in the system or can be installed with little effort. A distinction is made between necessary and additional data. Necessary data is a prerequisite for predictive maintenance. Additional data improves the quality of the predictions, but is not a must. The most important data to be collected is information about past failures, temperature, operating time, speeds, pressure, vibration and indirect parameters such as the material used or suppliers. Which of these are necessary is decided on a case-by-case basis.
Example Global Mining Company Reduction of transportation failures by 50 % through anomaly detection
The mining products company needs around 10,000 wagons to transport materials. Each breakdown of a wagon costs at least 6,000 US dollars. Originally, around 2,000 unplanned breakdowns were recorded each year. The aim of the project is to reduce downtime and repair costs using only data that has already been collected. Integrated sensors are already being used so that operational processes are not impaired by the collection of data such as temperature, sound and image recognition. With predictive maintenance and anomaly detection, failures are reduced to less than 1,000 per year using an anomaly detection algorithm.
Process of a predictive maintenance project
A predictive maintenance project requires careful preparation, but can then be quickly put on track. Basically, it can be divided into three phases: concept, MVP and analysis:
.webp)
