Development of self-learning condition monitoring using visual information and statistical data by application of analog acoustic and magnetic sensors

Speaker:
Barteldes, Soeren; QASS GmbH; Germany

Authors:
Barteldes, S.; QASS GmbH; Germany
Schiffner, J.; QASS GmbH; Germany
Seuthe, U.; QASS GmbH; Germany

ID: ECNDT-0437-2018
Download: PDF
Session: Structural Health Monitoring 4
Room: H1
Date: 2018-06-13
Time: 11:10 - 11:30

The spread of production process needs innovative and intuitive operable software for sensor data. Based on new sensor concepts QASS provides solutions to detect quantitative dimensions like mechanical hardness or qualitative information like cracks in components, wear of tool or striations, using both, magnetic and piezoelectric sensors. Sensor data is transformed via Short-Time-Fourier-Transformation into visual data by creating a spectral-data-diagram, which gives information about the amplitude and high-frequency-distribution over time. This visual representation makes the production process better accessible to the applicant to get a visual input for cognitive treatment.
QASS designed a pattern recognition system for spectral transformed data based on an own visual work- and data-flow design software. Any data stream, recorded by any type of sensor is compared with a reference pattern, based on its visual form from the spectral-data-diagram or its energy content. In calculating the similarity to the reference pattern the comparer function tolerates deviations in length and amplitude up to a specified level in order to account for fluctuations in process speed and loudness. The comparer function allows variations in time and amplitude spread and calculates a similarity to the reference pattern. Low similarity to the reference pattern correlates mostly with damaged parts.
For an automated self-learning analysis of unknown data/processes an incremental scheme is used. This described method has to be self-executed. First the spectral-energy of the production data is calculated to get a general idea of the energy distribution per process. Then the results over a whole production batch are statistically analyzed and used to create reference patterns that are searched on further production batches. Significant areas, that mark a damage or deviation from the average are used to provide a new reference pattern for actual production. The reference patterns can be refined in later steps based on further search results.