Skin cancer has been recognised as a significant global health concern, with millions of cases diagnosed annually. Early detection, particularly for melanoma, has improved survival rates. In recent years, artificial intelligence (AI) techniques have gained attention in skin cancer classification and have shown potential for automated skin cancer diagnosis. However, the research mainly uses dermoscopic or macroscopic images for AI model training. This research addresses these limitations by developing a generalised deep learning model that integrates dermoscopic and macroscopic image datasets, utilising transfer learning with VGG16 to enhance accuracy and accessibility. The model is designed to generalise features from the macroscopic and dermoscopic perspective, making skin cancer diagnosis more scalable. The model was trained and tested on two diverse datasets, including HAM10000 [15] and PAD-UFES-20 [14], ensuring adaptability to clinical settings and patient populations. We achieved an average Receiver Operating Characteristic Area Under the Curve (ROC AUC) score of 0.97, which indicates a good performance of the classifier. Furthermore, we also test the model in the real-time data application by testing the model’s ability to classify the lesion from a web-camera stream for real-world scenarios.

Adaptation is crucial in time-triggered systems for maintaining system performance and reliability under varying conditions, such as dynamic workloads or resource failures. Time-triggered systems rely on metascheduling techniques for adaptation; however, utilizing existing metascheduling schemes for time-triggered systems faces storage challenges for adaptation using the resulting schedules. This work presents a Genetic Algorithm (GA)-based metascheduler to tackle the state explosion problem in time-triggered systems. Our proposed method is designed to optimize meeting the application deadline and memory utilization by employing a Multi-Objective Genetic Algorithm (MOGA). Meeting the deadlines is crucial for safety-critical applications. Additionally, minimizing the schedule changes after a context event reduces the memory overhead. We leverage the similarity between successive schedules to store schedules incrementally, retaining only the differences between parent and child schedules. This approach significantly reduces redundant data storage and helps mitigate the state explosion problem. Comparative experiments demonstrate that our MOGA-based metascheduler achieves up to 90% memory savings, outperforming a makespan-optimized metascheduler and a metascheduler optimizing the lateness in a scenario with twenty-two context events. These results highlight the potential of our approach in enhancing the scalability and efficiency of metascheduling techniques in memory-coznstrained environments.

Embedded systems powered by artificial intelligence (AI) are widely employed in diverse domains. However, the lack of inherent predictability in existing AI accelerators poses significant challenges, especially in safety-critical applications where deterministic safety specifications are essential. Moreover, temporal unpredictability caused by memory access contention further hinders the suitability of current platforms for safety-critical tasks. To address these issues, we propose a time-triggered memory access approach for the versatile tensor accelerator (VTA) that provides temporal predictability guarantees for load data. Our research focuses on investigating the temporal predictability of Neural Network load input, weight, and operations. We introduce a time-triggered memory access mechanism that pre-fetches data from DDR memory ahead of the VTA load module request. Through extensive experimentation and analytical evaluations in the VTA runtime environment, we demonstrate the effectiveness of our time-triggered memory access concept. The results reveal an 8% performance improvement and reduced execution time while upholding strict safety specifications and predictability. These findings establish the feasibility of employing a time-triggered approach for runtime neural network prediction.

This research presents a fault-tolerant distributed Run-Time Infrastructure (RTI) using the High-Level Architecture (HLA) standard, emphasizing its importance for safety-critical use cases such as automotive systems. In this context, HLA serves as an enabler for hardware-in-the-loop testing, particularly vehicle-in-the-loop testing. Ensuring system reliability, managing real-time responses, facilitating seamless integration and interoperability of various components, effectively handling potential component failures, and maintaining the scalability of complex systems is paramount. Furthermore, the ability to comply with stringent safety regulations also underlines the necessity for safety mechanisms. These safety measures are crucial in protecting users and ensuring dependable system performance, particularly in safety-critical automotive applications where user safety is paramount. The study develops a fault-tolerant distributed RTI, validates it with an automotive use case, and provides a comparative assessment of centralized and distributed HLA implementations. The findings offer insight into large-scale distributed simulations' efficiency, dependability, and safety in various safety-critical applications.

Meta-scheduling algorithms are used for adaptation in time-triggered systems as they adapt to different scenarios such as failures or different environmental conditions. Most meta-scheduling algorithms demand a considerable amount of storage space from the host cyber-physical system due to the state-space explosion problem in covering a reasonable number of scenarios. This work deploys the Graph Neural Network (GNN) to learn the multi-schedules from the meta-scheduling algorithm required for adaptation. The GNN is used to learn the scheduling mechanism of a Genetic Algorithm (GA) so that at runtime, adaptation is achieved using the GNN model. We further investigate the impact of modifiable tasks during a meta-scheduling operation on the overall makespan. Finally, a comparison of the makespans is made between a List Scheduler (LS), GA and the proposed GNN-based technique to evaluate the impact of our approach. Our proposed GNN-based approach outperforms the LS scheduler as the number of modifiable tasks increases. The results show that the proposed GNN-based meta-scheduling can be suitable for real-time scenario adaptation in cyber-physical systems.

Time triggered devices are increasingly deployed in safety-critical distributed applications. Failures that manifest during the startup process of the synchronisation service pose a challenge to diagnose. The difficulty stems from the fact that most implemented diagnostic services for time-triggered systems employ the prior knowledge of the schedule to provide diagnosis. However, at the beginning of the startup process, when the global time base is not yet established, these diagnostic services are not viable. This work proposes the use of a fault injection framework to generate data that resembles the behaviour of failed components during startup. The data generated can then be used for developing fault diagnostic mechanisms. Due to the large data set that can be provided by fault injection frameworks, deep learning is proposed as a strategy for failure identification. The data generated from a fault injection framework is used to train the neural network to distinguish between correct behaviour, corruption and omission failures during startup.

In safety critical communication systems, faulty nodes can monopolize a channel by transmitting untimely messages at random intervals and thus resulting in the failure of the system. This failure is known as a babbling idiot failure. This could be costly and catastrophic for safety critical systems, therefore these failures are avoided in time triggered communication systems by implementing fault tolerant functions such as local or central guardians. Research works evaluating the guardian functionality for real time networks such as Flexray and TTP have been carried out. This work evaluates the guardian functionality of the TTEthernet protocol. TTEthernet enforces a TDMA scheme for time triggered traffic and traffic policing for rate constrained traffic thereby protecting the network against babbling idiot failures. However these guardian functionality was not extensively evaluated. Dependability evaluation by fault injection on the entire TTEthernet communication system as a whole has not been extensively studied. In this paper we exploit a novel fault injection framework to generate babbling idiot failures for the purpose of verifying TTEthernet implementations with respect to fault containment. The framework adopts a novel cut-through approach abstracting the fault injector from both the end systems and switches, thereby facilitating portability and ease of use. This work introduces a fault injection framework to effectively verify babbling idiot fault tolerance of TTEthernet hardware implementations. In addition, it provides a means to evaluate the effect of various network faults on applications running on top the protocol. Test results carried out indicate the effect of babbling idiot messages on the latency and jitter of traffic over the TTEthernet network.