The Artificial Intelligence of Things (AIoT) empowers IoT devices to leverage the advantages of AI near data-sources, reducing data movement, latency, and mitigating privacy issues. However, AI workloads are notoriously energy-intensive, posing significant challenges for energy-constrained IoT devices. Since such devices are often deployed in thousands of instances, even minor inefficiencies can significantly increase carbon emissions and energy consumptions. Model compression techniques have been employed to enable AI inference in resource-constrained environments. For example, Knowledge Distillation (KD) is an elaborate approach targeting low-footprint and high-accuracy models, although introducing further complexity during training due to inefficient grid searches of additional hyperparameters. The emerging wave of AIoT, however, calls for prioritizing energy-awareness both in inference and training. To address this shortcoming, this work proposes a three-stage design workflow for low-energy AIoT applications, driven primarily by an input energy budget characterizing the target IoT scenario. Given a specific CNN architecture and IoT platform, our workflow identifies the most effective student under the imposed energy constrained and derives an efficient configuration of the KD hyperparameters that maximizes student accuracy, while avoiding inefficient and expensive grid-search. Hence, this approach enable energy-efficient CNN inference while substantially reducing overall training costs. We validate our workflow with a systematic experimental campaign using ResNets and DenseNets on CIFAR-10, CIFAR-100,and Tiny Imagenet datasets, on an AMD Xilinx Zynq Ultrascale+ ZCU102 MPSoC. Our proposal maintains high accuracy while lowering energy consumption by up to 80%, highlighting the potential of our flow for real-world AIoT applications.

Nowadays, predicting public transport passenger flow (PF) is essential to optimize service planning and provide information to commuters. However, while current research focuses on enhancing accuracy using advanced models, like recurrent and graph neural networks, other key aspects, such as interpretability and computational efficiency, are often neglected. To fill this gap, we propose a framework to design on-board explainable PF prediction based on eXtreme Gradient Boosting (XGBoost). This framework enhances model interpretability and reduces computational costs, making the resulting PF predictive model suitable for inference on low-end devices. The proposed framework is validated on a real-world dataset from a major Italian city, involving 25 buses. Our results show that the framework achieves performance comparable to Convolutional Neural Networks (CNNs), with only a 0.2% percentage increase in Mean Absolute Percentage Error (MAPE). Additionally, it significantly reduces both inference time and energy consumption, with percentage decrease of 63%. Finally, we present examples to illustrate the interpretability of the predictions using the SHapley Additive exPlanation (SHAP) method. • Passenger flow (PF) prediction supports transport companies and passengers. • Currently, the main objective of research on PF is maximizing accuracy. • Our XGBoost-based framework provide accurate and explainable predictions. • The resulting model can be effectively executed on low end devices.

Abstract In the Intelligent Public Transportation Systems (IPTS) domain, predicting the number of commuters on-board, entering or leaving a metro train or a bus, i.e. the Passenger Flow (PF), is crucial for optimizing resource allocation and enhancing commuter satisfaction. In urban scenarios, the public transport system is often managed by distinct competing mobility providers. Traditional centralized machine learning models for PF prediction usually require data sharing among such competitors, leading to privacy and economic concerns. To overcome these issues, we propose exploiting Federated Learning (FL) in the PF predictions problem, as only model parameters must be shared among entities. Still, a straightforward application of FL can have some pitfalls. On one hand, it is widely recognized that FL can struggle with data heterogeneity, which is likely in the case of data acquired by distinct companies managing different public mobility services. Moreover, spatio-temporal features are not explicitly handled by classical FL. In this paper, we propose FedFlow: a personalized federated learning framework tailored for PF prediction. The proposed framework encompasses a personalized mechanism meant to refine local models based on client similarities, calculated by only leveraging publicly available domain-dependent information. The proposed framework has been experimentally validated on mobility data collected in a major Italian city, comparing FL predictions obtained by FedFlow against those obtained by LSTM models trained on local data, centralized data, FedAvg, and PerFedAvg. Results show that FedFlow outperforms all the considered adversary techniques. This work demonstrates that our proposal of personalized FL is effective in predicting PF while ensuring data privacy.

The Knowledge Discovery from Data (KDD) process is widely used across various domains to get valuable insights from data. Many platforms, like KNIME or RapidMiner, offer effective tools for KDD analysts, allowing them to perform data analytics tasks in a visual fashion, without writing code. In recent years, the increasing availability of mobility data has led to a surge in KDD-based initiatives from both industry and academia in the Intelligent Transportation Systems (ITS) domain. Still, KDD platforms lack comprehensive support for some typical mobility data manipulation tasks. As a result, mobility data analysis still requires a significant coding phase, with reduced productivity and hindered replicability of results. To address this gap, this paper presents a novel solution aimed at supporting ITS data analysts in defining KDD processes more efficiently. More in detail, we extended the KNIME platform by introducing a collection of new components explicitly tailored to facilitate some peculiar KDD tasks from mobility data. These components encompass critical functionalities such as map coverage analysis, trajectory partitioning and map-matching. To showcase the effectiveness of the proposed solution, we used it to replicate a study published in the ITS data analytics domain. Thanks to our proposal, such replication can be accomplished in a few minutes and with just a few clicks, without any manual coding, resulting in a pipeline that is easier to understand, distribute and re-execute, also for domain experts with no programming experience. Our solution is open-source and freely downloadable from the Knime Hub. In this way, we aim to foster data-driven research and practice in the ITS field, by providing researchers and practitioners with more effective analytics tools to handle mobility data.

Deep Learning is ubiquitous today and is increasingly moving from the cloud down to the edge of networked infrastructures, where it enables embedded applications to perform complex inference tasks close to the data sources, reducing long-distance data movement and alleviating the need for a powerful cloud infrastructure. Edge-class multi-processor system on chip (MPSoC) devices featuring an on-chip FPGA fabric offer key advantages for Deep Learning inference tasks, especially for complex applications where multiple models may be run concurrently in the same platform. In this work, we propose an approach and a practical framework for the systematic characterization of multithreaded Deep Learning inference on edge FPGA MPSoCs. We instantiate the framework into a real-world MPSoC platform, targeting Xilinx Vitis-AI as a representative example of a commercial Deep Learning acceleration toolkit for edge environments. We design a comprehensive experimental campaign and apply it to the platform for several convolutional neural networks, each trained on three different datasets. We show that our approach can be used for both hardware- and software-level analysis of a target system. Among other findings, the analysis revealed a suboptimal behavior of the underlying toolkit runtime, involving the utilization of the accelerator cores and the uneven software latency of the support library, influenced by the shapes of the input tensors.

With increasing urbanisation, and a growing population, transport within cities has never been more important. Buses are the most widespread form of transport worldwide, often being cheaper and more flexible than rail, but also less reliable. Long term bus journey time predictions are important for advanced journey planning and scheduling of bus services. For this reason, several machine/deep learning techniques have been defined to predict bus journey time. Still, due to the number of involved factors, such as complexity and noise in bus data, road network topology, etc., accurate predictions remain elusive. In this paper we aim at validating some Machine Learning methods recently shown to be effective in the literature, on new bus datasets from Dublin and Genoa. The analysis of the results shows some interesting insights into bus networks, highlighting that the accuracy of the predictions is strongly related to the standard deviation of the whole journey times. It emerges that some bus routes show consistency in the prediction error across methods, and for these routes it makes sense to use methods that are fast and computationally efficient, as there is no benefit to applying more complex algorithms. We use features of the route data distribution to develop an explanatory model for the consistency of the route across methods, with a coefficient of determination ( $$R^2$$ ) of 0.94. Finally, we identify a systematic anomaly in the data in Dublin that alters the performance of the methods.