Frequently asked questions

Our strengths lie in the rapid and customized retrofitting of machines and systems with sensors, including in hazardous areas. Our business model focuses on customized solutions based on standardized components, ensuring they integrate seamlessly into your existing production processes. A key advantage of our solution is the use of high-quality IEPE vibration sensors, which enable measurements down into the ultrasonic range. It is precisely in the ultrasonic range that the earliest signs of incipient wear are often found. The analog sensors we use, including IEPE sensors, can measure at up to 200,000 samples per second—a capability particularly necessary for applications in the ultrasonic range to reliably capture the finest nuances in the measurement signal. This type of high-resolution measurement generates data streams on a scale that would overwhelm conventional transmission channels. Direct preprocessing at the sensor is therefore our key to achieving maximum system speed. This not only significantly reduces the load on networks but also saves you substantial cloud costs. In other words, only by using Edge AI can we fully leverage the potential of your high-quality data.

Our approach makes a clear distinction between data science and machine learning. Using data science, we first analyze and understand the data in order to develop intelligent algorithms for predictive maintenance based on that analysis. Machine learning then allows us to train our intelligent algorithms using existing data, rather than programming them explicitly. Training requires more effort to implement but can scale more effectively. Which method is used depends on the requirements of the specific use case.

The AI functions performed by the sensor node depend largely on the chosen system architecture. If the AI is executed directly on the sensor node - which is recommended in many cases - the resulting decision-making logic can be customized to the specific use case. This ranges from recommendations for action to automated interventions, such as shutting down a machine under defined conditions.

We translate and capture human experience into a digital system comprising sensors and AI, creating a digital assistant that can support people. Sensors serve as the system’s sensory input by detecting physical signals, while AI interprets this data to generate condition assessments and maintenance recommendations.

The choice of methods used depends on the specific application and the available data sources. In addition to sensor data, we also take into account process information from control systems as well as metadata (e.g., machine type, age of the equipment, existing operational data, position data, or maintenance history). For CNC machines, for example, the program currently being executed can also be included in the evaluation.

Our domain model is intentionally kept concise and is primarily designed to support the operation of the monitoring solution. It maps which hardware is installed on which machine and includes basic metadata such as machine type and system age. This information assists both system management and the AI in interpreting the data. However, detailed modeling of complex digital twins is not a focus of our work.

Currently, we rely on static models, which are, however, regularly refined and retrained through our AI Academy. In doing so, we use human-in-the-loop processes: users provide feedback on the AI’s performance as well as on any errors or damage to the machine. Based on this feedback, we optimize or retrain the models and make updated versions available on the devices via an update mechanism. Self-learning systems are technically feasible, but we currently still consider them a research topic.

Updates to AI models and knowledge bases are transmitted remotely to the devices via the provided MQTT interfaces.

Especially when we use data science methods in combination with custom-developed algorithms, decisions can generally be explained in a transparent manner. These explanations are based directly on the collected sensor data but are, by their very nature, highly technical. Simple relationships, such as a rise in temperature, are easy to understand and communicate effectively. More complex signals, for example from vibration analysis, on the other hand, are primarily interpretable by data scientists. In clear-cut cases, however, justifications for a decision that are understandable to users can also be provided. Updates to the underlying models and knowledge bases are delivered remotely via the available MQTT interfaces.

If no clearly defined thresholds exist, the system uses probability-based assessments. For example, it indicates the probability that an anomaly is present or that a specific component is affected. As a problem progresses, the clarity of the analysis generally increases. Updates to the necessary knowledge bases and models are provided remotely via the available MQTT interfaces.

Our AI primarily serves in a supportive role and provides maintenance recommendations. Technically, however, autonomous interventions in the plant control system are also possible. Such functions are generally implemented only after the algorithms used have been reliably validated and trained.

Through our AI Academy, we offer a program in which we conduct regular feedback sessions with your maintenance staff. During these sessions, we work together to analyze the AI’s performance and record relevant maintenance events. We use this information to identify the causes of failures based on the available sensor data and to refine the models in a targeted manner. Alternatively, maintenance events can also be automatically imported via a CMMS. In more highly automated environments, feedback can also be captured directly via software. Updates to the models and knowledge bases are provided remotely via MQTT.

All five analog sensors can be measured in parallel and in sync at a sampling rate of up to 200 kHz. The necessary updates can be installed remotely over a network via the existing MQTT interfaces.

We rely on an AI management system that provides complete transparency regarding model versions, training data, and the machines on which the algorithm is deployed. This allows us to manage models and roll out updates to the respective target systems. At the same time, we continuously refine our models in collaboration with maintenance personnel as part of the AI Academy. Since our AI systems primarily generate supportive recommendations for maintenance, they generally do not intervene in safety-critical processes. However, if AI results are used for the active control of machines, we recommend doing so only after the algorithms have been comprehensively validated and their outputs have been regularly and reliably tested.