By David F. Nettleton (Ph.D.) and Elodie Bugnicourt (Ph.D.)
Industry 4.0 refers to an industrial revolution based on cyber-physical systems, in order to evolve towards “smart factories” which have a modular structure, and in which cyber-physical systems monitor physical processes by creating a virtual copy of the physical world thus facilitating decentralized decision making. Also, the Internet of Things will enable cyber-physical systems to communicate and cooperate with each other and with humans in real time. The drive towards Industry 4.0 has been defined as highly strategic by the European Union [1,2]. In the context of data analytics and modeling, Industry 4.0 promises solutions to real industrial problems and improvement of existing processes, such as in-line customized mass production, which has previously seemed impracticable. This is facilitated by the availability of machine learning and statistical function APIs which can be integrated into industrial PLCs, in closed loop or outside, as an automated initial calibration step.
Fig. 1. Depiction of the four phases of the Industrial Revolution, culminating in Industry 4.0
In the following we briefly look at two opportunities and challenges presented by Industry 4.0:
1. Process integration
2. Online retraining of data models
The following examples are adapted from real scenarios encountered in the OPENMIND applied research project, funded by the European Union’s Horizon 2020 research and innovation programme (grant agreement Nº 680820) . IRIS is participating with 9 other industrial partners, including Fraunhofer IPT and Blueacre Technology. IRIS is responsible for the data modeling part of the project.
1. Process integration.
Although this may sound obvious, a sequence of in-line processes can only go as fast as its slowest process. So when we integrate processes we find ways of speeding up the slowest process and slowing down the fastest process, until a feasible mutual speed is reached. If this is not possible, the processes will have to be performed in separate production lines.
For example, consider the production of a piece of polymer material (a guidewire), 1.5m long and 0.8mm in diameter. Two processes intervene: a pullwinding process which constructs the core (extrusion) and an outer layer (winding); then a laser process which ablates the material in certain sections to reduce the diameter. The laser ablation process is relatively slow (Feed Speed 0.05-0.1mm/s, that is 3-6mm per minute) whereas the pullwinding process is relatively fast (250 mm per minute). So, firstly, we have to evaluate if it is possible to put these in-line together or alternatively put them in separate (parallel) lines.
Fig. 2. Micro-pullwinding system at Fraunhofer Institute for Production Technology IPT
Fig. 3. Tube cutter machine – Blue Acre Technology
2. Machine learning/AI models.
Building data models is often considered an “off-line” activity, in which different historical datasets are analyzed and modeled until an acceptable and repeatable precision is obtained for the trained model on the test datasets. However, if we want intelligent embedded industrial systems to be autonomous or semi-autonomous, they have to be able to relearn and adapt “on the fly”, in the face of “concept drift”, going “out of tune”, and/or new unclassifiable cases.
Two approaches can be considered:
(i) Full retrain: when new cases appear which the model does not predict well, incorporate them in the train/test datasets and retrain the whole model. This makes sense when the data volumes are relatively small and the frequency of retraining is relatively low.
(ii) Incremental learning [3,4]: in which new cases are “incorporated” in the model without having to retrain with the whole dataset. This makes sense for big data processing such as online streaming where the computational cost of a full-retrain implies excessive down-time of a high-performance system.
For example, consider a calibration model which is initially trained offline for a pullwinding process. After a certain amount of time the machine behaviour (for example, the friction on a die) starts to go out of tolerance for a given set of parameters. Thus, the quality control module detects this and triggers a retrain process for the calibration model, in order to find the parameter set which satisfies the tolerance limits of the production processes. In this situation, the data volumes and the retrain time are relatively small so it makes sense to do a full retrain.
The OPENMIND project will develop a new process chain that makes possible an on-demand customisation of minimally invasive guidewire devices as a part of a continuous process. Processes such as pullwinding, laser ablation and structuring, coating, marking and visual inspection (quality control) will be integrated, together with machine learning algorithms to find the best control parameter set for a given customer production specification. Tackling the two aspects we have seen in this article (process integration and online machine learning data models) will place the OPENMIND project and IRIS at the forefront of Industry 4.0 applied research.
 Arif Budiman, Mohamad Ivan Fanany, Chan Basaruddin, Adaptive Online Sequential ELM for Concept Drift Tackling arXiv:1610.01922v1, [cs.AI] 6 Oct 2016, Available at: https://arxiv.org/pdf/1610.01922.pdf
 Deepak Agrawal (24 Inc.), Building and deploying real time big data prediction models. Strata+Hadoop World, December 1-3, 2015, Singapore. Available at: http://conferences.oreilly.com/strata/big-data-conference-sg-2015/public/schedule/detail/45045