Never a screw loose again

Back to blog

The use of the digital twin to predict the failure of device components

Over time, a crucial screw connection in medical devices will loosen during extended periods of operation. The affected devices then overheat to the point of emergency shutdown, and thermal defects occur in individual components. The manufacturer commissions its software service provider to develop an early detection system, but statistical analysis of the collected temperature sensor data does not reveal a clear pattern. Only the analysis of additional data and the deployment of a digital twin will lead to a breakthrough. A report from the field.

  • Frank Müller, Director Data Science & Analytics, infoteam Software AG
  • Patrick Kraus, Marketing Communications Manager, infoteam Software AG

Time and again, the screw components within the power connections in medical devices that have been in use for a long period of time become loose as a result of mechanical and thermal stress. This is critical for both the device manufacturer and for their customers as device operators, because the loosened screw connections lead to overheating of the device, followed by emergency shutdown and to thermal defects on individual components. This results in unplanned downtime and costly repairs by a service technician. The obvious solution for medical device manufacturers is a system that informs them as soon as a device is affected by loosened screw connections before the first damage occurs. This enables technicians to tighten the bolted joint in good time during comparatively short, cost-effective, and easier-to-plan service assignments at the customer’s site.

Such planned maintenance can be specifically scheduled outside the regular operating hours of hospitals and medical practices, avoiding the need for patient examinations to be canceled. At the same time, a targeted service assignment reduces the number of examination interruptions as a result of device malfunctions and thus also reduces the additional burden on patients due to follow-up examinations.

Deploying software, statistics, and machine learning to establish a reliable early warning system

The medical device manufacturer commissions its software service provider to develop a corresponding software solution. For this purpose, it provides temperature data collected by sensors at relevant positions in and on the devices over a period of several years. The data are supplemented by recorded defects, i.e. the period before the defect as well as the time of failure and repair. From this so-called historical data (i.e. data from the past), statistical as well as machine learning methods are intended to derive conclusions or recognizable patterns that are significant for devices with loosened screw connections. If a device exhibits a similar temperature pattern in the future, the software is able to warn the device manufacturer in time. That’s the theory. In practice, methods such as Random Forest and time series analysis reveal that each device behaves individually with respect to temperature development. In addition, the temperature sensors record varying degrees of temperature fluctuation depending on their position. However, clearly recognizable patterns for loose bolted joints cannot be identified from the data.

What is necessary to develop a digital twin?

To this end, the device manufacturer provides additional data in the form of operating data (time series on executed device functions), which the software service provider divides into operation and pause times. It then assigns the measured temperatures of all sensors to these times. The linking of operating data and temperature data now for the first time offers the possibility of deriving a simple model for temperature development during pause times. It serves as a basis for the development of an extended physical model of the temperature development as a function of the operating data (Fig. 1). This so-called digital twin makes it possible to compare the theoretical temperature curve with the actual one and to detect deviations.

Figure 1: The depicted theoretical temperature development of the physical model (orange) compared to the actual temperature development (blue) under normal system behavior already demonstrates a high degree of consistency. © infoteam Software AG

Due to the complexity of the devices, however, this physical model is subject to large tolerances, since not all dependencies hidden in the data can be identified and modeled mathematically. For this reason, and due to the enormously large amount of data (data collected from 10,000 devices over several years), the software service provider is pursuing the approach of a model based on artificial intelligence (AI) algorithms in addition to the physical model in order to predict usage-dependent temperature profiles. This model does not require expert knowledge but evolves into an expert. In addition, its prediction results can also be used to detect other temperature anomalies (which have been successfully implemented in a follow-up project).

Physical model vs. artificial intelligence

The self-learning AI model deploys methods from natural language processing, i.e. computational linguistics. The algorithms used are capable of processing natural language and understanding contextual relationships. In the present use case, a neural network learns to interpret the operating data translated into a “language-like” form. Since the operating data sometimes contain longer work sequences, long short-term memory (LSTM) networks are particularly well suited to refine the quality of results. The comparison with operating and temperature data of fault-free devices proves that the trained AI model leads to significantly more precise predictions than the manually created physical model (Fig. 2).

Figure 2: The depicted theoretical temperature development of the AI model (orange) compared to the actual temperature development (blue) under normal system behavior is even more precise than the physical model. © infoteam Software AG

Comparing the actual temperature profile of a device with the theoretical temperature profile of its digital twin now makes it possible to identify significant deviations from the standard. Subsequently, the comparison between the temperature prediction from the AI model and the actual data measured at the device proves: It is not exceeding a fixed temperature-related threshold that is characteristic for loosened bolted joints, but a slowly increasing offset of the temperature trajectories over several weeks (Fig. 3).

Figure 3: The depicted theoretical temperature development of the AI model (orange) demonstrates a significant deviation with slow screw loosening compared to the actual temperature development (blue). © infoteam Software AG

Automated AI model for each device as a practical solution

This insight helps to extend the time frame for action by several weeks compared to the solution originally stipulated by the medical device manufacturer, during which the slow loosening process is already visible and remediable. To address the proven individual behavior of each device in practice, the software service provider implements an automation process that derives an individual model for each device in use from the developed AI base model. The device manufacturer has since been able to:

  • monitor all devices with regard to temperature anomalies that occur,
  • visualize the analyses via dashboards,
  • create automatic service tickets in case of temperature anomalies,
  • repair the equipment in good time during scheduled service assignments, avoiding damage.

Θέλεις να γίνεις μέλος της ομάδας μας;

infoteam Hellas MIKE

Dr. Alexis Papadimitriou

Vasileos Irakliou 53 | 54623 Thessaloniki | Greece info@infoteam-software.gr