Skip to main content



The final networks comprise around 700 weights, including bias inputs for the nodes. All of these networks are executed once in every system cycle. Tested in isolation on the DSP device, using a minimal run-time support host program but requiring the full complement of DPRAM input and output operations, the DSP device could execute the 700 weight neural code in around 11.5 milliseconds.

The update rate of the entire WINNER system, with all components on-line, reached a maximum of 3.14 Hz.

Closed loop welding evaluation was performed with a control strategy of (1) avoid defects, (2) control bead dimensions, (3) control transfer mode, stability and spatter. The controller was very good at reaching and maintaining a required bead geometry, even after deliberately forcing the power supply into defect causing outputs. The system performs well with root gaps of up to 2 mm, after which weld quality rapidly deteriorates.

In post-weld assessments, bead geometry control proves to be the system's best feature. To an extent, this is self-fulfilling, as bead geometry is the primary control variable. The rule based controller actively drives the process into operating areas where the network outputs are stable and accurate. Different control strategies could easily shift the emphasis to transfer mode accuracy or spatter control, for instance.

As an overview, it can be said that, in spite of the need to review parts of the process control, namely standoff, travel speed and adaptation to varying gaps, the system performs very reasonably for butt welding, being capable of assuring a convenient level of weld quality.

As to fillet welding, the system needs some further work to ease up tasks like robot positioning and torch angle, but, apart from some deviations from the requested leg length in one of the tests, the quality that the prototype was able to assure can be said to be very good.
The proposed research is directed at developing a prototype system that will collect signals from welding equipment, process this data using neural networks and use the derived information to control the welding process in real time. Rapidly acquired data will be passed directly to a unique, innovative neural network realised in dedicated hardware. The network will interpret and correlate the data and pass on control information to an adaptable control system. Thus electrode position and welding parameters will be controlled in real time.

The major research tasks will be:

1. the development of a data acquisition system to collect information in real time
2. to design dedicated electronic hardware neural network architectures to process the data collected
3. the construction of a suitable universal control system
4. the investigation, testing and evaluation of the prototype system.

The realisation of such a system will bring about improvements in operator working conditions, enhance product quality, reduce remedial repair resulting in diminished environmental impact.

Funding Scheme

CSC - Cost-sharing contracts


Brunel University
Runnymede Campus
TW20 0JZ Egham
United Kingdom

Participants (3)

Cranfield University
United Kingdom

MK43 0AL Cranfield
Av. Rovisco Pais
1096 Lisboa Codex
Rheinisch-Westfälische Technische Hochschule Aachen (RWTH)
Reutershagweg 4
52074 Aachen