【转】机器学习在工业中的应用场景

原文地址：http://www2.le.ac.uk/departments/informatics/research/kdml

Potential Collaborative Projects with Industry

Within KDML we have a broad range of research interests and capabilities. Below are some examples of current projects. If you have any queries or ideas, please do not hesitate to contact us (see below for contact details).

Effort-estimation from cross-company data: We have developed Machine Learning algorithms that can enable organisations to accurately predict effort by using cross-company data, reducing the dependence upon internally recorded data.

Textile flaw detection: We have had a successful series of collaborations with an industrial partner in the textile industry. As a part of this, we inferred classifiers to more accurately detect textile flaws, flagging up fewer false-positives, and leading to a higher degree of automation.

Analysing live data streams to predict rail traffic build-up: The DfT funded PREPAReD project is fusing live rail data with computational models to enable the prediction of rail delays.

Multi-factor decision support for software safety case assessments: We have developed a tool-supported approach to aggregate multi-faceted safety assessments for critical software components, and to produce coherent overviews.

 

原文链接：https://link.springer.com/article/10.1007/s11740-017-0718-7

Improving the laser cutting process design by machine learning techniques

In the field of manufacturing engineering, process designers conduct numerical simulation experiments to observe the impact of varying input parameters on certain outputs of the production process. The disadvantage of these simulations is that they are very time consuming and their results do not help to fully understand the underlying process. For instance, a common problem in planning processes is the choice of an appropriate machine parameter set that results in desirable process outputs. One way to overcome this problem is to use data mining techniques that extract previously unknown but valuable knowledge from simulation results. This paper presents a hybrid machine learning approach for applying clustering and classification techniques in a laser cutting planning process.

In a first step, a clustering algorithm is used to divide large parts of the simulation data into groups of similar performance values and select those groups that are of major interest (e.g. high cut quality results).

Next, classification trees are used to identify regions in the multidimensional parameter space that are related to the found groups. The evaluation shows that the models accurately identify multidimensional relationships between the input parameters and the output values of the process.

In addition to that, a combination of appropriate visualization techniques for clustering with interpretable classification trees allows designers to gain valuable insights into the laser cutting process with the aim of optimizing it through visual exploration.