JUL-AUG 2019

Issue link:

Contents of this Issue


Page 34 of 53

INTECH JULY/AUGUST 2019 35 AUTOMATION IT power plants, which is traditionally a point of con- cern. Connections routed through the Internet, in particular, are regulated by national standards and laws. However, as a customer pointed out, "solutions that address this concern are techni- cally possible but still face a psychological hurdle." Different monitoring approaches Classical monitoring Typical hydropower plant monitoring systems are fully integrated in the control system of the plant. Each monitoring sensor focuses on a singu- lar quantity at a specific location. The sensors are physically mounted to the equipment. Therefore, a retrofit requires some physical modifications of the equipment. During the design phase of the monitoring sys- tem, the engineer predefines static alarm limits, for example, in accordance with a norm like DIN ISO 7919-5, where maximum relative vibration dis- placement limits are defined. These limits are de- rived from statistical analysis of many hydropower plants. Although they gather the collective hydro experience, the limits are not necessarily suitable for a particular plant or even power unit. Adjust- ments to these static limits are usually done during the commissioning of the hydropower plant. Advanced integral monitoring Often, before equipment failure, early indicators like noise, heat, or odor are noticeable, even be - fore a classical monitoring system observes an indication in a specific signal, because the static alarm limit is not yet reached. Equipment failure starts very early, progresses slowly, and is only no - ticeable to either personnel inspecting the plant or where slowly changing trends are monitored explicitly. This also requires that the actual failing component is equipped with a sensor. Often, this is not the case, because too many sensors would be required. The new approach is based on microphones, w hich collect integral information from a larger space or a larger set of equipment, like inside the turbine pit. The system can detect anomalies or devia - tions from typical ma- chine behavior and slowly moving trends and can diagnose specific signal patterns and relate them to past events. Figure 1 shows a sketch of the sys- tem where sensor data is combined with process information (e.g., current active power). Such an approach becomes feasible today, because the storage concepts and algorithms required to im- plement them are now available. Some of the intelligence of this monitoring system is based on domain-specific preknowl - edge, like the knowledge of typical machine frequencies and their meaning. But, to achieve diagnostic or even predictive capabilities be - yond such preknowledge, there are three major steps that have to be taken during the so-called learning phase. l First, collect a large number of training sam- ples to get a fingerprint of the "normal" situa- tion. This is the basis for detecting any abnor- mal situation or anomaly. l Second, classify samples that show anoma- lies by the type. Later, if a reasonable number of such anomalies have been found, then the classification of similar types can be done automatically. l Third, do further classification, where the technical background of a certain sound event is included. This puts the sample in a specific technical context. This technical classification includes the analysis of actual events happen- ing locally as well as past experiences of either the involved experts or those already stored in the system. Only now can the system be trained to di - agnose specific events in addition to detect - ing anomalies. Predic- tion is still one step further. This requires that the classification explains whether the specific sound is an indication of an actual event to happen in the future. FAST FORWARD l A large amount of equipment can be simultaneously supervised with only a few sensors. l Cloud-based machine learning algorithms can detect upcoming failures. l An acoustic monitoring system prevents any interference with the safety and control system of the plant. Figure 1. Integral monitoring system connected to a central, cloud-based plat- form for assessment and analysis of the collected information. Continuous improvement Acoustic signal View Technical hydro expert Energy market expert VOITH OnPerformance Lab Your O&M team Data scientist team Data scientist Assess Acoustic signal Data recorder Operation data Shared knowledge PLC OnCumulus Anomaly detection Machine learning Warnings & alarms Advanced diagnosis VOITH Data analysis O&M optimization VOITH VOITH VOITH VOITH

Articles in this issue

Archives of this issue

view archives of InTech - JUL-AUG 2019