Power management is a cornerstone for many Cyber-Physical Systems (CPSs), which relies on low-power circuits, dynamic power management algorithms and energy-aware software to match their requirements in terms of energy. As CPSs evolve towards data-centric designs to more promptly accommodate AI models and integration, traditional power management techniques must also be improved. In this paper, we build on SmartData to introduce a data-centric Power Manager (PM) framework that allows CPSs to model energy in terms of data. SmartData defines a high-level interface for sensing, actuation, and control in data-centric CPSs. It abstracts the myriad of features of modern embedded platforms related to processing, scheduling, synchronization, and communication. These Energetic SmartData encapsulate the components of a CPS, which interact in a publish–subscribe fashion, declaring interest on other SmartData and responding to other SmartData interests. We introduce an algorithm to extract a Directed Acyclic Graph (DAG) from these Interest relationships, with vertices representing the involved components and edges representing the associated cost in terms of energy. We also introduce a Power Manager that uses such DAGs to monitor the state of the system, eventually overriding low-priority Interests to reach the specified lifetime. We evaluated the proposed framework through a case study with Ocean-Bottom Nodes (OBNs) under realistic, dynamic energy conditions. Results show that without any power management, the system fails 12 days before its target operational lifetime. The proposed data-driven PM was then benchmarked against a fixed-schedule Static PM and a reactive Threshold PM. Our approach was the only strategy to guarantee a 365-day lifetime in all scenarios. With an ideal initial battery capacity of 260 Ah, it achieved a high utility of 23.1%. It also proved its adaptability in an energy-deficit scenario with an initial capacity of 257 Ah, where it reduced utility to 2.8% to survive, a condition in which the other strategies failed.

Ensuring fast, deterministic data exchange is vital for Advanced Driver Assistance System (ADAS), making Time-Sensitive Networking (TSN) a key enabler in real-time automotive environments. ADAS heavily depends on Machine Learning and Computer Vision to analyze information from various sensors, such as LiDAR, radar, and cameras. These technologies support essential functions like adaptive cruise control, collision prevention, and lane-keeping assistance. However, the need to handle and transmit vast amounts of high-resolution sensor data presents substantial computational and networking challenges. This study introduces an predictive model to refine transient subscription management in a publish-subscribe TSN setup for ADAS. The proposed system utilizes predictor to anticipate how long temporary data subscriptions should persist, particularly when triggered by critical scenarios—such as detecting an object at the edge of a sensor’s Field-of-View. By enhancing data exchange efficiency and reliability, this approach bolsters safety-critical CV operations while maintaining the stability of ongoing communication streams. Ultimately, this method improves the resilience and scalability of ADAS in highly dynamic, real-time conditions.

Power management is a cornerstone for many Cyber-Physical Systems (CPSs), which relies on low-power circuits, dynamic power management algorithms and energy-aware software to match their requirements in terms of energy. As CPSs evolve towards data-centric designs to more promptly accommodate AI models and integration, traditional power management techniques must also be improved. In this paper, we build on SmartData to introduce a data-centric Power Manager (PM) framework that allows CPSs to model energy in terms of data. SmartData defines a high-level interface for sensing, actuation, and control in data-centric CPSs. It abstracts the myriad of features of modern embedded platforms related to processing, scheduling, synchronization, and communication. These Energetic SmartData encapsulate the components of a CPS, which interact in a publish-subscribe fashion, declaring interest on other SmartData and responding to other SmartData interests. We introduce an algorithm to extract a Directed Acyclic Graph (DAG) from these Interest relationships, with vertices representing the involved components and edges representing the associated cost in terms of energy. We also introduce a Power Manager that uses such DAGs to monitor the state of the system, eventually overriding low-priority Interests to reach the specified lifetime. We evaluated the proposed framework through a case study with Ocean-Bottom Nodes (OBNs), demonstrating the system's ability to dynamically adapt to achieve its operational time. In an experiment with variability impacting the available energy budget, while the CPS without PM would exhaust its energy 12 days before the intended lifetime, the proposed PM successfully guaranteed the designed lifetime by only reconfiguring low-criticality Interest relationships, without canceling any CPS functionality.

Much effort has been put into Time-Sensitive Networking for Cyber- Physical Systems. Methods and tools have been proposed to support the design and operation of such networks. However, the incorporation of novel technologies, such as computer vision, cloud integration, and decision-making for autonomy, requires an operation regimen that cannot always be modeled at design-time. In this paper, we introduce algorithms to determine whether the load imposed on a Time-Sensitive Networking is within bounds, and to extract formal properties from message headers in order to dynamically verify the network's temporal requirements. We also introduce a formalism based on Signal Temporal Logic to express such properties, a set of Time-Sensitive Networking-specific property monitors, and a lightweight verification engine that can check them at runtime with little overhead. These mechanisms are used by a Safety Enforcement Unit to continuously monitor the network, triggering actions such as best-effort and low-priority traffic cancellations whenever a property fails verification. We evaluated the proposed mechanisms on a prototype autonomous vehicle that uses a Time-Sensitive Networking to interconnect GNSS, IMU, LiDAR, RADAR, and Camera to an ECU. Results demon-strate that these mechanisms can run in real-time, consuming 3% of the processing power of one of the cores in the ECU.

This work investigated two statistical properties, namely stationarity, and entropy, of a real-world publicly available battery dataset considering features like Current, Charge Capacity, Discharge Capacity, Temperature, and Voltage, with the objective of providing insights for the development of models that would be best suited for early End-of-life (EOL) prediction. From the characteristics of the data, we hypothesized that the lack of stationarity and higher entropy would deteriorate the performance of LSTM models while having less of an impact on CNNs. To fortify this hypothesis we developed 4 types of models and investigated their performances. The results for this case study indicate that CNN-based models are more robust to these properties of the data, while the LSTM-based ones are more sensible and therefore have worse performance. We discuss this sensibility by analyzing the correlation of these statistics with model performance. The paper presents a detailed process for preprocessing, model generation, and comparison. Our best LSTM-based model had 18.3% error while the best CNN-based model presented 3.5% error when considering unseen test data, using only the first 100 cycles of the batteries.

Modern embedded systems integrate software components onto a single computing platform to meet stringent non-functional requirements of cost, space, weight, and power consumption, amongst others. Moreover, the growing demand for computational power pushed for the adoption of multicore platforms. At the same time, those platforms are often connected to the external world to support a variety of applications. In this context, Machine Learning-based Intrusion Detection Systems (IDS) are of significant importance to guarantee the system’s security during its operation. One approach to be adopted by IDS is to model the behavior of the applications on an embedded system through Performance Monitoring Counters (PMC) and operate during runtime by detecting deviations to the modeled behavior. Notwithstanding, the execution of multiple tasks onto the same multicore platform often incurs shared-resource contention between tasks, which may impair the execution of software components and possibly affect the behavior observed through PMC. In this paper, we assess the impacts of lacking proper resource isolation mechanisms on multicore embedded systems over two Machine Learning-based Intrusion Detection Systems (IDS) solutions that rely on PMC. We use a relevant dataset in the scope of embedded systems control with both tasks monitored while executing without and with the interference of shared-resources contention. Results demonstrate that the lack of isolation can lead to the IDS mechanism losing the ability to recognize the behavior of target software components.

Cyber-Physical Systems (CPS) rely on sensing to control and optimize their operation. Nevertheless, sensing itself is prone to errors that can originate at several stages, from sampling to communication. In this context, several systems adopt multivariate predictors to assess the quality of the sensed data, to replace data from faulty sensors, or to derive variables that cannot be directly sensed. These predictors are often evaluated based on their accuracy and computing demands, however, such evaluations often do not consider the system's architecture from a broader perspective, ignoring the way components are interconnected and how they cascade as inputs of other Machine Learning (ML) models. In this work, we introduce a method to evaluate the performance of interdependent predictors based on the stability of the estimation error dynamics in faulty scenarios. The proposed method estimates the ability of a predictor to produce accurate predictions while accounting for the impacts of cascading predicted values as its inputs. The prediction correctness is estimated based solely on information acquired during the training of the multivariate predictors and mathematical properties of the ML activation functions. The proposed method is evaluated with a meaningful dataset in the scope of monitoring and control of a Cyber-Physical System, and the evaluation demonstrates the ability of the proposed method to account for the interdependence of data predictors.

In this paper, we address the problem of intrusion detection in multicore embedded systems through a self-nonself discrimination scheme based on Artificial Immune Systems. We collect runtime data to build a model in which the T-cells work as detectors for the system’s sane behavior. The T-cells are represented by N-dimensional data points composed of samples of the N variables monitored during model building. A pre-established binding threshold is used for the T-cells generation. The difference between data points is measured as the distance between them. While training, whenever a collected sample fails to bind to an existing T-cell, it becomes a new one. After training, the threshold is adjusted to the maximum distance observed in the model. Therefore, the model definition follows an iterative clustering algorithm where each T-cell is a cluster centroid with threshold as the radius. Nonself detection consists of comparing collected samples to the T-cells in the model through a cluster membership verification. Whenever the incoming sample is not a member of any of the clusters, the sample is classified as nonself. A time complexity analysis indicates the suitability of the proposed technique for runtime operation, and offline experiments show this approach achieved a 97.17% nonself detection rate.

In this work, we propose a secure bootstrap protocol for Cyber-Physical Systems (CPS) that compose IIoT Networks. The main contribution of our work is a solution to establish secure communication channels in CPSs through a protocol that enables authentication and confidentiality without the need for constant external verification or pre-stored keys. The proposed protocol relies on the unclonable property of Physical Unclonable Functions (PUF) to build authentication tokens to establish trust between the devices, the gateway, and the Cloud. Devices registration is triggered by an authenticated operator, which informs the PUF responses of the respective device to an External Security Agent (ESA) alongside the identification of the target gateway. ESA and gateway are mutually authenticated using a Certificate Authority and communicate via a secure channel built with HTTPS. The device registration relies on the properties of PUFs to avoid the establishment of security channels via key agreement protocols (e.g., ECDH) and the usage of pre-stored keys. In this way, the PUF challenge response can be used as a secret between the gateway and the device to build trust and establish a secure channel. The presented solution addresses attacks like message replication, Man-in-the-Middle (MITM), and nodes impersonation while supporting gateway integrity check solutions and being free of pre-stored key vulnerabilities.

In this paper, we introduce a Framework to Design and Implement Real-time Multicore Schedulers using Machine Learning techniques applied to the very own data such systems produce as they operate. The framework builds on sensors and event counters present in modern hardware platforms and on variables kept by the operating system to capture run-time data that are subsequently subjected to ML tools to produce scheduling heuristics targeting specific optimization goals. It provides non-intrusive mechanisms to collect such data while the system runs real task sets with real workloads, thus preserving the quality of the captured data. It abstracts the Performance Monitoring Unit, thermal sensing, energy monitoring, and Dynamic Voltage and Frequency Scaling available on such platforms through a lean, architecture-independent API. After describing the framework in details, we demonstrate its applicability with the implementation of an energy-efficient, load balancing, real-time, multicore heuristic for a PEDF scheduler. The measured overhead imposed by the framework on the tasks it schedule is at most 0,0003583% and the maximum added jitter is less than 40μs, corroborating the ability of the framework to support the development of effective domain-specific schedulers using machine learning techniques.