In this paper, we address the design of data-based Artificial Neural Networks (ANN) controllers. More specifically, we consider a scalable design based on a Transfer Learning approach where an ANN controller trained at a given source scenario is transferred to other target domains. In order to properly assess the transfer suitability of the controller, the adoption of a Transfer Suitability Metric (TSM) is required. And here resides the main goal of this paper: to develop a TSM able to measure the amount of information captured by a neural network to estimate a desired output from input data. To do so, we resort to Mutual Information (MI) studies addressing the learning process in a neural network. As shown in the paper, we propose a MI-based metric able to assess the transfer suitability while reducing metric computation complexity.

In the last years, the industrial digitalisation and the Industry 4.0 paradigm is no longer a fairy-tale but a reality. It is becoming more common to find industrial environments relying and adopting data-based approaches to perform some sorts of processes. Some of them are related to the industrial control, where the incursion of Artificial Neural Networks (ANNs) is promoting the usage of data-based solutions to substitute conventional control structures. Besides, one of the greatest issues related to the ANN time-consuming training process has been alleviated by means of Transfer Learning (TL) methods. However, in the industrial control domain TL cannot be freely adopted since the final performance of the transferred control structure cannot be known before substituting the conventional structure. This is an issue that needs to be tackled, especially in critical industrial scenarios where an incorrect control can produce huge disasters. For that reason we present here the Transfer Suitability Metric (TSM). Based on the environments similarities, its main aim is to compute the transference suitability of ANN-based controllers in order to transfer the ANN to the target domain without resorting to new control design and optimization. It provides the plant operators with an insight of the controller behaviour before it is finally substituting the conventional control structure. Results have shown that the metric is highly correlated with the final control behaviour in the sense that the higher the metric, the better the final ANN-based controller performance.

The incursion of the Industry 4.0 paradigm and the Artificial Neural Networks (ANNs) is changing the way as the industrial systems are conceived and controlled. Now, it is more common to talk about data-driven methods either supporting conventional industrial control strategies, or acting as the control itself. Thus, one can find that in the last years it is more common to find control systems which are purely based on data leaving aside the highly complex mathematical models. However, data-driven models and ANNs have to be correctly trained in order to offer a good performance and therefore, be contemplated as the core part of a control strategy. This can become a time-demanding and tedious process. For that reason, Transfer Learning (TL) techniques can be adopted to ease the conception, design and training processes of the data-based and ANNs methods, since the efforts have to be mainly focused on training a unique net which will be then transferred into the other scenarios. In that sense, we present here a TL approach to design and implement the whole control of a Wastewater Treatment Plant (WWTP). First, the control of the quickest dynamics under control is performed by means of a Long Short-Term Memory cell (LSTM) based Proportional Integral (PI) controller (LSTM-based PI). Once the LSTM is trained and tested, its knowledge will be transferred into the remaining WWTP control loops. In that way, an ease and reduction in the time involved in the design and training of the control as well as in its complexity is achieved. Results have shown a twofold achievement: (i) the LSTM-based PI achieves an improvement of the control performance with respect to a conventional PI controller around a 93.56% and a 99.07% in terms of the Integrated Absolute (IAE) and Integrated Squared (ISE) errors between the desired measurement and the obtained one, respectively, and (ii) the LSTM-based PI controller achieves an average improvement in the IAE and ISE around a 9.55% and 15.25%, respectively, when it is transferred into a different WWTP control loop.

Wastewater Treatment Plants (WWTPs) are industries where highly complex and non-linear processes are performed to reduce the pollutant concentrations of residual waters. However, some nitrogen and phosphorus derived pollutants are generated in these processes. As a consequence, certain control strategies have been developed to maintain these pollutants under certain limits. Benchmark Simulation Model No.1 (BSM1), a framework emulating the behaviour of a general purpose WWTP, considers a default controller strategy based on Proportional Integral (PI) controllers. Nevertheless, these controllers are based on linearised models of the WWTP behaviour. For that reason, this work proposes a new control approach based on Internal Model Controllers (IMC) adopting Artificial Neural Networks (ANNs), which are able to model the real plant behaviour without performing linearisation. Results show that the proposed IMC is improving the default controller performance around a 16% and a 53% in terms of the Integral Absolute Error (IAE) and the Integral Square Error (ISE), respectively.