3L: Machine Learning, Deep Learning and Active Learning
Apart from working with existing tools in the state of the art (tensorflow, keras, Caffe, sk-learn, etc.) and the different programming languages, at LORTEK we design specific algorithms for industrial problems. In the information explosion in the current I4.0, we address the problem of dealing with different origins and types of data, their traceability, and the detection of false data. We create ad-hoc algorithms, to generate high-precision and quality prediction models.
An aspect that concerns us is knowledge obsolescence. Thus, we accompany our methods with Self-Learning techniques that guarantee maximum performance both in the industrial installation, and in the following tolerance of context changes (changes in product, parameters, or components).
We apply these methods to: predictive maintenance, product optimization, energy optimization, and fault prediction.
Digital Image Processing is a specific case where Machine Learning methods can be applied, with the advantage that the images are a body of sorted data. In this context, we apply techniques that adapt to the problem of vision, such as analysis of reflectance, chromatic calibration, and automatic systems to correct changes in lighting.
Regarding computational methods, we continue working with traditional methods such as SVM, RF, ANN, Bayes, etc. But, concerning vision algorithms, we also work with the latest Deep Learning advances (CNN, RNN, GAN, etc.).
We support vision systems with Active Learning software architectures in many cases. This means that expert worker is still a key figure for the quality control. Today’s industry demands the manufacture of increasingly shorter series with an increasingly greater variety of products. Expert workers are the ones, who, with human interfaces/AI, manage this process of adapting the algorithms to new manufacturing contexts.
We essentially apply these methods to quality control systems. We observe the product with cameras, and then we conclude if the part is good or not.
Smart control systems use several artificial intelligence techniques to control complex processes. Processes, which models are very complex, non-linear, uncertain environments, or with control criteria and objectives that change over time. Smart control permits solving control problems that are impossible to address with traditional methods.
The four basic techniques that essentially comprise Smart Control are Expert Control, Fuzzy Control, Neuronal Networks, and Genetic Algorithms. At LORTEK, we are working on the implementation of these techniques in complex industrial processes with the main objective of improving quality.