Case C3 - Audio surveillance
The aim of this case study is to use airborne sound monitoring to detect various failures. This idea partly can be considered to mimic operators knowledge accumulation via auditive stimuli. In a hydro power plant, the standard instrumentation consists in dedicated sensors to monitor specific components such as accelerometers and proximity probes for shaft line vibration or pressure probes to monitor flow conditions. In contrast to these examples, sound monitoring offers the possibility to gather information from a large number of components with a single sensor (microphone) or a sensor array. This enables the recording of the sound patterns (signatures) emitted by the regarded components. The patterns may be associated with different conditions of the components. Then, the evolution of these patterns can be analyzed, thus providing characteristic features of various sound developments. This kind of monitoring is investigated as it offers various advantages in comparison to other methods: the simplicity of the installation and operation is a cost-effective solution, anomalies of several machines may be regarded accumulative and there is no need of cabling sensors directly on the machine. Additionally, an early on recognition of approaching issues shall be enabled.
In order to prove the concept, measurements were taken by Andritz at Statkraft's Svorka power plant during the first months of 2019. At the power plant, studio microphones with a sampling frequency of 44,1 kHz and ultrasonic microphones with a sampling frequency of 200 kHz were used. Recording covered short sound samples at regular intervals. The samples were collected by edge devices which can handle 1-4 sensors each and sent to a central server on the server's request. Processing and analysis were performed in Graz with dedicated HPC servers. The results of the analysis were then made accessible via the Andritz digitalization platform METRIS, which enables the management, control and investigation of data and processes.
The initial objective was to detect specific events according to their sound signatures:
- Stone impact
- Early detection of bearing failure
To distinguish the sound signatures of various events, advanced machine learning technics are necessary. First, audio signals are converted from time domain signals to spectrogram representations. Spectrograms are a widely used approach to compress audio data and to identify and compare respective features. In a second step, the number of dimensions describing the features of each sound sample is reduced using a neuronal network. Eventually, in this reduced space, the distance between the different samples is evaluated to quantify the anomaly score of each sample. Furthermore, a clustering algorithm is applied to group samples with similar features allowing a fast labelling with a limited amount of user interaction.
The figure shows that similar results are obtained using standard and studio microphones. Some additional events are detected using the ultrasonic microphones, which can be largely accounted to noises emerging in the ultrasonic range, which were not detectable by the studio microphones. Further investigation is needed to assess if those events are related with the presence of cavitation in the runner and if long-term trends can be derived from these findings. A limited number of clusters have been identified and are most probably related to various mode of operation of the units.
Not only specific targets like definite recognition of cavitation and other data assignments could be considered in the future, but also further hardware- and software-technical developments. Such improvements and research may address a better ruling to filter human-voice out of signals, implementation and exploitation of sensor arrays for finer signal separation or pre-calculations in edge devices.
Time evolution of anomaly score using studio and ultrasonic microphone.