Protecting critical infrastructures like municipal water supplies requires reliable equipment, transport routes, and storage facilities – and any defects or errors in these systems can quickly have serious consequences. The digitalisation of devices, equipment, and measuring systems can prevent – or predict – precisely such anomalies. This enables maintenance work to be proactively scheduled and saves costs.
As part of their Service Meister research project, KROHNE Messtechnik GmbH and inovex developed a solution for the automated, real-time detection of anomalies in measuring instruments. The resulting solution is, for example, capable of replacing threshold-value-based, non-automated water supply technology. Capturing and analysing sensor data using machine learning algorithms enables automated, accurate anomaly detection – in real time and during live operations.
Machine learning opens up new opportunities for small-to medium-sized businesses
The system developed was designed specifically to be applied to similar problems in industry, and the inovex-KROHNE Messtechnik GmbH collaboration illustrates how its implementation might look in practice. The defined and partly developed concepts, interfaces, and software/machine learning components (blueprints) provide companies – especially SMEs – with easy access to machine-learning methods. This lays the foundation for:
- reducing service costs and increasing customer satisfaction through targeted anomaly detection and the definition of cost-optimised maintenance intervals,
- safeguarding industrial production processes – which in turn facilitates
- the development of innovative digital products and services.
Typical use cases for the machine learning solution
The flow meters manufactured by KROHNE Messtechnik GmbH are used in both the industrial sector and in critical urban infrastructure. They are, for example, used to monitor drinking water utilisation or the load on supply lines within a supply network, or to provide valuable information for the optimized control of industrial processes.
These highly precise measuring systems are often exposed to the weather, while the monitored processes themselves are sometimes subject to creeping, unwanted changes, such as mineral build-up or dirt clogs in connecting lines. Undetected, these issues can, in extreme cases, lead to sections of the drinking water supply being affected or result in industrial production processes being unavailable for anywhere from hours to days. The consequences of such failures range from high costs due to production downtime to jeopardization of the municipal water supply.
Until now, manual searching has been the method of choice for detecting potential errors. Troubleshooting using this method is, however, extremely time-consuming. KROHNE Messtechnik GmbH and inovex used the blueprints created in their joint research project to develop a prototype anomaly detection solution for flow meters.
Intelligent measurement technology, reimagined
In order to monitor measuring sections automatically and in real time, inovex trained machine learning models using previously collected sensor data. These models can use the sensor data to identify the various flow meter statuses in real time and to differentiate them from one another. This enables short circuits, air bubbles in the water, and supply line blockages to be quickly detected.
As use cases and error types vary, the models can be retrained or be given supplemental training to cover a wide range of potential anomalies.
A clear dashboard displays the sensor data captured and the statuses detected. This enables service technicians to quickly identify anomalies and carry out targeted maintenance work. This, in turn, reduces unnecessary costs and ensures maximum operational uptime.
The solution developed consists of three components: a connector to record and forward the sensor data from the measuring instruments, the anomaly detection function to detect anomalies based on the sensor data, and the service distribution, a system which delivers the predictions to target systems and end users.
The anomaly detection solution developed was implemented using free and open software libraries. This decision allowed the machine-learning lifecycle to be automated to the maximum possible extent and also enables it to be applied to other use cases.
Predictive characteristic values are automatically extracted from the (high-dimensional) sensor data using tsfresh, which accelerates the complex feature engineering. Based on the features identified, inovex uses H2O’s AutoML functionality to carry out the machine learning training. This allows the automatic selection of machine learning models and hyperparameter optimisation. Trained models, including their associated evaluation results, are stored versioned in a model repository based on MLflow. Sensor data and the associated status predictions are stored in an InfluxDB – a database optimised for time series. Standardized interfaces are used to connect the measurement systems and transfer the predictions. The services developed have been containerised and can be run on a variety of platforms.
The prototype anomaly detection solution for KROHNE Messtechnik GmbH’s flow meters was implemented on this architecture, and the anomalies detected, together with the sensor data, were displayed clearly and in real time in a Grafana dashboard.
As part of their joint research for the Service Meister project, KROHNE Messtechnik GmbH and inovex developed a generic anomaly detection system for measuring instruments and a prototype solution specifically for detecting anomalies in flow meters. The solution for capturing and classifying malfunctions and displaying that information in real time is currently being evaluated and iteratively advanced. Blueprints and information on the infrastructure developed are available on request and can be adapted for other use cases.
The Projekt „Service-Meister“ is supported by the Bundeswirtschaftsministerium.