Adaptive Intelligent Systems for Extreme Environments

As embedded processors become powerful, a growing number of embedded systems equipped with artificial intelligence (AI) algorithms have been used in radiation environments to perform routine tasks to reduce radiation risk for human workers. On the one hand, because of the low price, commercial-off-the-shelf devices and components are becoming increasingly popular to make such tasks more affordable. Meanwhile, it also presents new challenges to improve radiation tolerance, the capability to conduct multiple AI tasks and deliver the power efficiency of the embedded systems in harsh environments. There are three aspects of research work that have been completed in this thesis: 1) a fast simulation method for analysis of single event effect (SEE) in integrated circuits, 2) a self-refresh scheme to detect and correct bit-flips in random access memory (RAM), and 3) a hardware AI system with dynamic hardware accelerators and AI models for increasing flexibility and efficiency. The variances of the physical parameters in practical implementation, such as the nature of the particle, linear energy transfer and circuit characteristics, may have a large impact on the final simulation accuracy, which will significantly increase the complexity and cost in the workflow of the transistor level simulation for large-scale circuits. It makes it difficult to conduct SEE simulations for large-scale circuits. Therefore, in the first research work, a new SEE simulation scheme is proposed, to offer a fast and cost-efficient method to evaluate and compare the performance of large-scale circuits which subject to the effects of radiation particles. The advantages of transistor and hardware description language (HDL) simulations are combined here to produce accurate SEE digital error models for rapid error analysis in large-scale circuits. Under the proposed scheme, time-consuming back-end steps are skipped. The SEE analysis for large-scale circuits can be completed in just few hours. In high-radiation environments, bit-flips in RAMs can not only occur but may also be accumulated. However, the typical error mitigation methods can not handle high error rates with low hardware costs. In the second work, an adaptive scheme combined with correcting codes and refreshing techniques is proposed, to correct errors and mitigate error accumulation in extreme radiation environments. This scheme is proposed to continuously refresh the data in RAMs so that errors can not be accumulated. Furthermore, because the proposed design can share the same ports with the user module without changing the timing sequence, it thus can be easily applied to the system where the hardware modules are designed with fixed reading and writing latency. It is a challenge to implement intelligent systems with constrained hardware resources. In the third work, an adaptive hardware resource management system for multiple AI tasks in harsh environments was designed. Inspired by the “refreshing” concept in the second work, we utilise a key feature of FPGAs, partial reconfiguration, to improve the reliability and efficiency of the AI system. More importantly, this feature provides the capability to manage the hardware resources for deep learning acceleration. In the proposed design, the on-chip hardware resources are dynamically managed to improve the flexibility, performance and power efficiency of deep learning inference systems. The deep learning units provided by Xilinx are used to perform multiple AI tasks simultaneously, and the experiments show significant improvements in power efficiency for a wide range of scenarios with different workloads. To further improve the performance of the system, the concept of reconfiguration was further extended. As a result, an adaptive DL software framework was designed. This framework can provide a significant level of adaptability support for various deep learning algorithms on an FPGA-based edge computing platform. To meet the specific accuracy and latency requirements derived from the running applications and operating environments, the platform may dynamically update hardware and software (e.g., processing pipelines) to achieve better cost, power, and processing efficiency compared to the static system.