Optimized data-driven measurement methods and devices for modeling diagnosis and prediction of batteries behavior

Batteries are playing an increasingly central role in today's society, driven by national and supranational entities as a means to facilitate the energy transition from fossil fuels to renewable sources, reducing emissions and pollution while aiming to enhance the quality of life in cities. An example of this push is the European Union's mandate that, by 2035, all vehicles sold in the European common market must be battery-powered. Simultaneously, the transition to an economy no longer reliant on fossil fuels is occurring, particularly in large-scale energy production, where renewable resources are being incentivized due to their intermittent energy supply nature, making energy storage devices crucial. Redox reactions-based batteries currently stand as the most efficient system for energy storage. They provide a higher energy density compared to any other available storage method. However, electrochemical cells are not without flaws, and improper management can lead to an increased risk of accidents with potentially tragic consequences. In this context, it is vital to develop methods that ensure proper battery operation, effective diagnostic tools, and predictive modeling. The topics addressed by predictive diagnostics align perfectly with the needs of electric energy storage systems. Notably, predictive diagnostics have gained significant traction in recent years, becoming a part of Industry 4.0, representing the latest industrial revolution focused on incorporating new production technologies to improve working conditions, increase the interconnection of production systems, and enhance plant resilience. Predictive diagnostics play a pivotal role in each of these aspects and are, therefore, fundamental within the framework of Industry 4.0. Predictive diagnostics, also known as predictive maintenance, is defined by two fundamental components: "data acquisition" and "condition monitoring." The "data acquisition" process encompasses all the tools and systems used to acquire knowledge about the system. Without adequate measurement of the phenomenon that accurately reflects reality, any subsequent action is unlikely to succeed and can even be detrimental. The thesis work conducted spans both areas characterizing predictive maintenance. In the realm of measurements on electrochemical cells, data acquisition involves sensor-based systems, measuring intrinsic parameters (such as voltage, current, temperature, and internal impedance) and derived parameters (such as State of Charge and State of Health). This thesis addresses the issue of data quality, seeking methods that not only provide an understanding of the measured data but also its associated quality. Specifically, a measurement system for battery impedance was created, capable of stimulating by injecting an alternating current and measuring both the current and voltage to derive impedance. The system's flexibility allowed for the study of various signal forms (stepped sine and multisine). An analysis of processing was also carried out, introducing sinusoidal fitting as a method to overcome the time constraints dictated by Fast Fourier Transform (FFT). The effectiveness of the created measurement system led to the development of a distributed embedded measurement system designed specifically for UPS (Uninterruptible Power Supply). "Condition monitoring" can be divided into two significant sectors: Diagnosis, which involves understanding the current state of the battery through data acquisition systems, and Prognosis, which entails forecasting the future state of the battery. In this thesis, diagnosis is addressed through the development of a methodology for optimizing measurement times in estimating the state of charge of lithium iron phosphate batteries. Knowledge of the state of charge is crucial not only for providing users with information about system autonomy but also for supplying the Battery Management System (BMS) with insights into the state of individual cells and enabling actions to protect battery modules. Currently, state of charge estimation relies on techniques such as Coulomb Counting, which are based on prior knowledge of the state of charge and battery capacity. Impedance measurement has excellent potential to surpass the limits imposed by current methodologies but has the significant drawback of requiring extended measurement times, especially when low frequencies are used (typically, these systems have stimulus frequencies ranging from 10 mHz to 10 kHz). A Support Vector Machine (SVM) is trained to estimate the state of charge based on impedance spectroscopy measurements within a genetic algorithm framework designed to reduce the number of stimulus frequencies for measurement time optimization. The results obtained enable a reduction in measurement time while preserving the accuracy of state of charge estimation. In the context of diagnosis, a preliminary study of non-destructive testing techniques for batteries has been conducted. In particular, a correlation between state of charge and the inductance of an inductor, printed on a PCB, attached to the side of a VRLA lead-acid battery, has been established. The proposed system is intriguing as it allows state of charge estimation without any galvanic connection to the battery. Once the current battery state is determined, the challenge is predicting the future state, i.e., the so-called prognosis. Prognosis cannot be separated from system modeling. Initially, the focus shifted away from batteries to model electric loads using state-based algorithms based on Markov Chain. Each device was characterized metrologically and then modeled in a series of states that vary through coefficients listed in a probability matrix. The result is a synthetic dataset capable of emulating the simultaneous use of multiple loads and allowing the training of algorithms for recognizing devices connected to an electrical network by measuring the electrical signature of the system. In the case of batteries, the modeling concentrated on the parametric identification of equivalent electrical models (ECMs) with the aim of predicting terminal voltage knowing the state of charge. The problem was approached in two ways. The first, already known in the literature, involves pulsed characterization tests in which batteries are discharged with current pulses, and the voltage response provides information about the ECM. The second method uses a genetic algorithm to provide a closed form of a function that relates voltage to state of charge. The pulsed approach has the major disadvantage of requiring specific characterization with precision instruments capable of discharging the battery while measuring voltage and current at a high sampling frequency. The approach based on Genetic Programming (GP) uses a simpler characterization to train the algorithm and also provides a closed form output that can be easily integrated into a BMS. The system can be constrained to provide an invertible solution, meaning it can provide the battery's state of charge based on terminal voltage as input. In conclusion, this thesis work has covered all the distinctive aspects of predictive maintenance by providing methodologies, defining and refining hardware tools, and creating datasets useful for the development of software tools.