“Get new technology first” is the strapline of Hannover Messe Industrie 2017, which is taking place on April 24–28. With 6500 international exhibitors – including The unbelievable Machine Company – and more than 400 industry applications, the world-leading industry trade fair is the Mecca of modern, digital-based engineering art. It is therefore the perfect place for us to showcase innovative artificial intelligence opportunities for Industry 4.0:
Deep Learning showcase “Component recognition”
at Hannover Messe Industrie 2017
at the Hewlett Packard Enterprise (HPE) stand
in the Industry 4.0 Forum in Hall 8
Deep Learning case: identifying components using image recognition (live @ #HMI17)
When assemblies are complex, the part number of each component is often hidden or difficult to see. If devices are old these part numbers may be illegible or even missing all together. In addition, individual components look very similar and it is often barely possible to distinguish between them.
So how can components be identified easily? To create an application for visually recognizing components, we have implemented and further developed processes from the field of image recognition – this involves using pre-trained deep neural networks and adjusting them for application.
The effect: Assistance for production employees who require alerts for the maintenance and repair of components. They are able to directly call up machine data by simply photographing the component with a mobile device, rather than typing in the component number or device number.
Image recognition can also be applied for detecting defects, such as damage to the surface structure of components, or for analyzing correlations between features of product images and sales data.
Deep Learning case: predictive maintenance using sound recognition
In the field of predictive maintenance we have already implemented several use cases, in which vibration data or time series of vibrations were analyzed for the early recognition of defects in industry facilities.
How can potential defects be recognized by sound? Defective machines show anomalies. In the simplest case this can be a different signal intensity, while less obvious cases may include characteristic patterns being missing or new, previously unobserved (sound) patterns being observed.
In the event of a defect it is possible to observe isolated high amplitudes, known as peaks, and their multiples, known as harmonics, which are not present during normal operation. Vibrating or rotating components in the device may therefore indicate a defect.
Before analyzing the data there is a pre-processing stage in order to correct the influence of the operating mode on the vibrations. Machine learning models are then trained to record the normal state.
The models were evaluated on a small number of proven defect cases in order to be able to characterize the majority of the spectra as accurately as possible, and to differentiate between defective and functioning spectra.
The effect: The application case results in a feasibility analysis and implementation into operation, whereby the suitable model processes the machine data in real time and supplies information to a reliable early warning system.