Directly training spiking neural networks for cyber-physical systems: from supervised to reinforcement learning

Spiking Neural Networks (SNNs) are bio-inspired Artificial Neural Networks (ANNs) utilizing discrete spiking signals, akin to neuron communication in the brain, making them ideal for real-time and energy-efficient Cyber-Physical Systems (CPSs). This thesis explores their potential in Structural Health Monitoring (SHM), leveraging low-cost MEMS accelerometers for early damage detection in motorway bridges. The study focuses on Long Short-Term SNNs (LSNNs), although their complex learning processes pose challenges. Comparing LSNNs with other ANN models and training algorithms for SHM, findings indicate LSNNs' effectiveness in damage identification, comparable to ANNs trained using traditional methods. Additionally, an optimized embedded LSNN implementation demonstrates a 54% reduction in execution time, but with longer pre-processing due to spike-based encoding. Furthermore, SNNs are applied in UAV obstacle avoidance, trained directly using a Reinforcement Learning (RL) algorithm with event-based input from a Dynamic Vision Sensor (DVS). Performance evaluation against Convolutional Neural Networks (CNNs) highlights SNNs' superior energy efficiency, showing a 6x decrease in energy consumption. The study also investigates embedded SNN implementations' latency and throughput in real-world deployments, emphasizing their potential for energy-efficient monitoring systems. This research contributes to advancing SHM and UAV obstacle avoidance through SNNs' efficient information processing and decision-making capabilities within CPS domains.

Predictive Ability of NGAL as a Marker of Renal Damage: evaluation of Multiple Clinical Settings

Introduction. Neutrophil Gelatinase-Associated Lipocalin (NGAL) belongs to the family of lipocalins and it is produced by several cell types, including renal tubular epithelium. In the kidney its production increases during acute damage and this is reflected by the increase in serum and urine levels. In animal studies and clinical trials, NGAL was found to be a sensitive and specific indicator of acute kidney injury (AKI). Purpose. The aim of this work was to investigate, in a prospective manner, whether urine NGAL can be used as a marker in preeclampsia, kidney transplantation, VLBI and diabetic nephropathy. Materials and methods. The study involved 44 consecutive patients who received renal transplantation; 18 women affected by preeclampsia (PE); a total of 55 infants weighing ≤1500 g and 80 patients with Type 1 diabetes. Results. A positive correlation was found between urinary NGAL and 24 hours proteinuria within the PE group. The detection of higher uNGAL values in case of severe PE, even in absence of statistical significance, confirms that these women suffer from an initial renal damage. In our population of VLBW infants, we found a positive correlation of uNGAL values at birth with differences in sCreat and eGFR values from birth to day 21, but no correlation was found between uNGAL values at birth and sCreat and eGFR at day 7. systolic an diastolic blood pressure decreased with increasing levels of uNGAL. The patients with uNGAL <25 ng/ml had significantly higher levels of systolic blood pressure compared with the patients with uNGAL >50 ng/ml ( p<0.005). Our results indicate the ability of NGAL to predict the delay in functional recovery of the graft. Conclusions. In acute renal pathology, urinary NGAL confirms to be a valuable predictive marker of the progress and status of acute injury.

Human congenital cytomegalovirus infection: characteristics and pathogenesis of fetal brain damage

Human cytomegalovirus (HCMV) causes congenital neurological lifelong disabilities. The study analyzed 10 HCMV-infected human fetuses at 21 weeks of gestation to evaluate the characteristics and pathogenesis of brain injury related to congenital human CMV (cCMV) infection. Specifically, tissues from cortical and white matter areas, subventricular zone, thalamus, hypothalamus, hippocampus, basal ganglia and cerebellum were analysed by: i) immunohistochemistry (IHC) to detect HCMV-infected cell distribution, ii) hematoxylin-eosin staining to evaluate histological damage and iii) real-time PCR to quantify tissue viral load (HCMV-DNA). Viral tropism was assessed by double IHC to detect HCMV-antigens and neural/neuronal markers: nestin (expressed in early differentiation stage), doublecortin (DCX, identifying neuronal precursor cells) and neuronal nuclei (NeuN, identifying mature neurons). HCMV-positive cells and viral DNA were found in the brain of 8/10 (80%) fetuses. For these cases, brain damage was classified in mild (n=4, 50%), moderate (n=3, 37.5%) and severe (n=1, 12.5%) based on presence of i) diffuse astrocytosis, microglial activation and vascular changes; ii) occasional (in mild) or multiple (in moderate/severe) microglial nodules and iii) necrosis (in severe). The highest median HCMV-DNA level was found in the hippocampus (212 copies/5ng of humanDNA [hDNA], range: 10-7,505) as well as the highest mean HCMV-infected cell value (2.9 cells, range: 0-23), followed by that detected in subventricular zone (1.8 cells, range: 0-19). This suggests a preferential HCMV tropism for immature neuronal cells, residing in these regions, confirmed by the detection of DCX and nestin in 94% and 63.3% of HCMV-positive cells, respectively. NeuN was not found among HCMV-positive cells and was nearly absent in the brain with severe damage, suggesting HCMV does not infect mature neurons and immature HCMV-infected neuronal cells do not differentiate into neurons. HCMV preferential tropism in immature neural/neuronal cells delays/inhibits their differentiation interfering with brain development processes that lead to structural and functional brain defects.

Damage Identification of Structures through Vibration-Based Structural Monitoring Systems

The thesis has been carried out within the “SHAPE Project - Predicting Strength Changes in Bridges from Frequency Data Safety, Hazard, and Poly-harmonic Evaluation” (ERA-NET Plus Infravation Call 2014) which dealt with the structural assessment of existing bridges and laboratory structural reproductions through the use of vibration-based monitoring systems, for detecting changes in their natural frequencies and correlating them with the occurrence of damage. The main purpose of this PhD dissertation has been the detection of the variation of the main natural frequencies as a consequence of a previous-established damage configuration provided on a structure. Firstly, the effect of local damage on the modal feature has been discussed mainly concerning a steel frame and a composite steel-concrete bridge. Concerning the variation of the fundamental frequency of the small bridge, the increasing severity of two local damages has been investigated. Moreover, the comparison with a 3D FE model is even presented establishing a link between the dynamic properties and the damage features. Then, moving towards a diffused damage pattern, four concrete beams and a small concrete deck were loaded achieving the yielding of the steel reinforcement. The stiffness deterioration in terms of frequency shifts has been reconsidered by collecting a large set of dynamic experiments on simply supported R.C. beams discussed in the literature. The comparison of the load-frequency curves suggested a significant agreement among all the experiments. Thus, in the framework of damage mechanics, the “breathing cracks” phenomenon has been discussed leading to an analytical formula able to explain the frequency decay observed experimentally. Lastly, some dynamic investigations of two existing bridges and the corresponding FE Models are presented in Chapter 4. Moreover, concerning the bridge in Bologna, two prototypes of a network of accelerometers were installed and the data of a few months of monitoring have been discussed.