Even though some companies suggest that they can offer predictive maintenance as a ready-to-use software solution, the reality is often different. Predictive maintenance with the help of AI is an automation task that requires a fundamental understanding of the process and the data generated from it, and thus usually requires an individual solution. In order to deal with heterogeneous circumstances of our customers, we have developed a 4-stage maturity model. This helps to better discuss the status quo and enables us to approach the project iteratively with you.
In the classic understanding of condition monitoring, AI does not yet play a role at this point. The focus here is on the presentation of sensor data in the form of dashboards.
However, it is possible that domain-specific logic for call-to-action functionality is integrated, such as an alarm when limit values are exceeded.
From this level onwards, machine learning methods are used to recognize changes in the data patterns. The goal is to identify changes that indicate potential damage, so repairs can be performed before a failure gets the chance to occur.
This layer is not about predicting the point of failure, but about detecting anomalies in order to perform maintenance steps near the potentially imminent failure.
incl. fault diagnostic
This level extends the anomaly detection to the point of fault diagnosis. After a degradation is recognized by machine learning algorithms, the diagnosis component is triggered.
This component is designed to enable fast and targeted maintenance by using Explainable AI to localize the root cause.
remaining useful life
In the highest maturity level, the pure detection of the ageing process is extended by the prediction of the expected process behaviour.
The problem to be solved is now no longer a classification problem (machine condition is ok / not ok), but a regression problem with the prediction of the point in time of the machine failure.
With this knowledge, the choice of the time of maintenance can then be made in the best possible way.
4 success factors for PdM projects
Predictive maintenance is inconceivable without comprehensive data. However, data can vary greatly in terms of quality and level of detail. This has a major influence on the development of AI models, as they can only achieve good results if the data have a large information volume.
In the case of production machines, the data is often available in different levels of detail:
• Raw data
• Event-based data
• High-level information
• High-level actions
For the development of AI models, it makes a big difference whether only error codes or the raw data from the sensors are available. A common problem is often the lack of important contextual information. Different batches of 2 kg and 2.5 kg workpiece weights have different effects on the sensor data and must be taken into account so that these differences are not wrongly classified as an anomaly.
A frictionless data processing and the use of trained AI models for maintenance optimization is only possible through an orchestrated interaction of several systems.
Here is a brief overview of the different components needed for an infrastructure:
- A connectware to connect sensors or machines from different manufacturers to a unified system.
- Scalable data storage and a Big Data processing engine to handle streaming data (if real-time data is required)
- Computing resources and environments for training AI models
- Services to deliver the predictions of the AI models to the end user.
A common problem in predictive maintenance is the large number of sensor manufacturers with different protocols for transmitting data. Together with our partner Cybus, who specializes in connectware, we can overcome this challenge.
A deep understanding of the basics of machine learning is something every Data Scientist should know. Besides the right choice of algorithms, however, the skill usually lies in correctly formulating the automation task as a machine learning problem. This is especially true in the area of predictive maintenance.
The fundamental problem of anomaly detection is the lack of a definition of how different a novelty must be before classifying it as abnormal. Marked data for training models is often not available to this context. Either it is not possible to observe normal and abnormal behaviour in all possible ways, or it is too expensive to obtain specific labels.
Due to the numerous projects we have carried out in the field of predictive maintenance, we can optimally support your project with a lot of experience.
User-centricity is important for all products with an AI component. Often, black box solutions are built that are not trusted in daily use. Human-in-the-loop is the central approach here. In a dashboard for the machine operator or in the control center, it must be clear why certain forecasts or anomaly detections were made.
By implementing AI technology, existing processes and ways of working are always changed. That’s why change management always plays an important role in the success of our customers project.
Our Predictive Maintenance Experts
Trainee Data Scientist
Since our foundation, our data & AI experts have successfully implemented numerous projects in the area of predictive maintenance and product quality control. Read some of our references on the topic of AI in the production industry here.
You want to know how specifically AI and Data Science can reduce maintenance costs and machine downtime in your production, but you still have many questions:
- For which machines do a predictive maintenance approach using AI make sense?
- What framework conditions must be in place to ensure that concrete added value is created in production from the anomaly detections?
- Is the data stored so far sufficient for training AI models?
- Which characteristics can be taken into account in addition to the sensor data?
- How do I integrate different production machines into the company-wide BigData infrastructure?
Then talk to our experts and get an initial assessment of your project plan!