Information technologies for energy management in the smart factory

Power supply is an urgent problem in the perspective of deploying pervasive Wireless Sensor Networks (WSNs). If, on one side, it is desirable to save grid power, on the other hand it must be considered that, in many applications (e.g. seismic sensors installed on volcanos; weather monitoring stations for smart agriculture), the grid is unavailable, and therefore the use of alternative power supplies – usually batteries – is mandatory. In such cases, the replacement of exhausted batteries is expensive and time-consuming, and may be inconvenient if the installation site is hardly reachable. Thus, many energy harvesting (EH) methods, exploiting different energy sources, have been studied in order to extend as long as possible the battery lifetime, aiming at achieving a complete energy self-sufficiency. Among the various power sources, solar energy has driven, and is still driving, the attention of a big portion of the world marketplace of renewables. Consider that about half of the renewable energy produced in Italy in 2021 is attributable to the solar sector. The most common material for outdoor light energy harvesting is crystalline silicon (c-Si), in its polycrystalline (poly-Si) and monocrystalline (mono-Si) forms. Despite the fact that silicon technology is well-known and industrially standardized, a brisk research activity is still going in order to increase at maximum the efficiency, approaching the theoretical upper limit for silicon (about 30 % under 1 sun at 25 °C). For instance, passivating carrier-selective contacts have been studied recently to optimize the carrier transport characteristics of c-Si solar cells, leading to significant improvement in the efficiency. In the first part of the thesis, we present a theoretical method and an experimental method for evaluating the performance of a solar module and consequently optimizing the operation of a user device powered by the module. In particular, we propose these techniques applied to environmental pollution monitoring sensors, designed in our laboratories at the Department of Information Engineering and Mathematics of the University of Siena, equipped with a small mono-Si solar cell. This work was presented at conference and was extended as journal publication. Another key aspect of energy management is to ensure a continuous and high-quality power supply to the users. The term “continuous” refers to the necessity of certain appliances of relying on an uninterrupted power source. If, in a domestic context, the lack of electric energy can cause at worst discomfort, in other situations the consequences can be catastrophic: the lives of the patients in a hospital can be jeopardized, or petabytes of data can be lost in a data center. The term “high-quality” is linked to the waveform delivering the electric energy. In Italy, the standard is 50 Hz frequency, 230 Vrms between each phase and the neutral. For instance, harmonic distortion, i.e. the superposition of superior harmonics to the fundamental 50 Hz harmonic, usually due to the presence of non-linear loads connected to the grid, can provoke additional Joule dissipation in transformers, cables and appliances, with a subsequent temperature increment, and an abnormal current flowing in the neutral conductor. Over time, if these conditions persist, failures can occur. In this framework, fault diagnosis assumes a central role. When dealing with power electronic circuits, and analog circuits in general, it is important to distinguish between two classes of faults: hard faults and soft (or parametric) faults. Hard faults are typically linked to the breakdown of a component, usually manifesting as a short-circuit or an open-circuit, leading to a potentially destructive failure of the system. On the other hand, a fault is called soft when it is related to one or more circuit parameters exceeding the tolerance limits around the nominal value, due to aging, manufacturing defects or parasitic effects. As a result, the system functions outside the specifications, worsening the performance, and possibly leading to system ruptures in the long term, if the faulty condition is not individuated and fixed. With that being said, it is straightforward to understand that soft faults are subtler to be discovered than hard faults. In fact, in most cases, the quantities to be monitored are not directly measurable in a simple way, e.g. without disassembling the system, and interrupting the service for a long period. As an example, consider the pre-charge electrolytic capacitors in a UPS (Uninterruptible Power Supply) system . Electrolytic capacitors are particularly subject to aging, causing a deterioration of the capacitance value. Since to periodically measure the state of the capacitors is impractical, the idea is to deduce the information that is needed from other circuit quantities, which can be constantly monitored, can be easily measured with negligible impact on the system operation, or can be evaluated contextually to routine maintenance interventions. Such alternative measurements must be somehow affected by the variations of the parameter to be monitored. Back to the example of electrolytic capacitors in UPS systems, if we hypothesize to discharge the capacitive bank and recharge it, the steepness of the voltage curve across the capacitors should be influenced by the capacitance values. A common approach to this kind of problems, which has become more and more popular in the last years, relies on neural networks. These techniques exploit machine learning to abstract from the circuit topology, and to reduce the problem to a set of circuit “features” passed to a neural network, producing a certain output depending on how the neural network was trained. Indeed, these methods are sometimes called “data driven methods”, signifying that they are totally founded on measurements collected from the system rather than on mathematical models, simplifying considerably the design and the implementation of the fault diagnosis apparatus. In this work of thesis, we studied the application of a neural network based soft fault detection strategy to the already mentioned pre-charge stage of a UPS system. The activity can be divided into three phases. Initially, we designed a voltage down-scaled version of the circuit (the reason for this will be clear later) and simulated it via software to the aim of individuating a convenient set of circuit features to be fed to the neural network classifier, and we validated the classification performance on simulated data. We selected a Radial Basis Function (RBF) architecture as our neural network classifier. Our choice in favor of this kind of machine learning algorithm, which is relatively old and well-known in the literature, was dictated mainly by its simplicity: the three-layer structure of a RBF network makes it lighter than the majority of the neural network architectures currently employed, both computationally and for memory allocation requirements. As such, RBF networks represent a good compromise between performance and resource occupation, enabling the possibility to implement the entire fault detection algorithm on embedded devices (e.g. low-cost microcontrollers). This would be a decisive move towards automatic diagnostic systems. In addition, in this phase, the network was trained and the coefficients (or weights) of the network were determined. This set of coefficient values was kept also for the following stages of the activity, i.e. the RBF network was never retrained. It is important to underline this last aspect: the training data set, which notoriously has to be large, was obtained through simulations. If the training examples had to be obtained from measurements, the procedure would have been tediously long; moreover, it would have been unfeasible to retrain the network in case of mistakes, or in case of any modifications one wanted to introduce (add a new fault class, change the set of features, adjust the network parameters). The uncertainty arising from the use of simulated data can be mitigated by running a sufficient number of simulations, finely varying the circuit parameters to cover the feature space. The second step consisted in the realization of the voltage down-scaled version of the circuit to perform actual measurements on a physical system, to confirm or disconfirm the results of the simulations, and consequently decide if it was necessary to take a step back and reevaluate the choice of the features and retrain the network. The purpose of this method is to create a circuit which would present the same functionalities of the original one, but can be managed even in laboratories which do not have the required authorizations to work directly with the mains voltage, making it possible, for instance, to collaborate remotely with companies active in the field of power electronics, without the need for testing the real products. In this way, more comprehensive studies can be conducted separately, without interfering with the normal production chain of the company. This means that, with zero impact on the company activities, a complete prototype of the fault classification system can be realized, ready to be installed on a real machine for final field tests. Finally, we implemented the neural network on a commercial microcontroller board (in particular, a NUCLEO-H7A3ZI produced by STMicroelectronics). The classification performance of the network running on the microcontroller was estimated and compared with the results obtained with the network working on a classic PC, using actual measurements executed on the down-scaled circuit.