The embedded systems developed for modern automotive systems, including autonomous driving controllers and advanced driver assistance systems (ADAS), have stringent demands for high-performance computing with strict real-time constraints, ensuring passenger safety and global reliability. Nonetheless, these systems are required to operate under harsh conditions in a dynamic environment, where temporary failures due to electromagnetic interference, temperature fluctuations, and hardware wear significantly constrain their performance. These weaknesses may cause system crashes or slow down the system, and such a scenario could prove disastrous. Classical fault-tolerance designs, such as static Triple Modular Redundancy (TMR), and simpler microcontrollers tend to be weak in offering high reliability, low latency, and energy efficiency simultaneously. This paper will address FPGA accelerators, but to overcome the challenges posed by the system, the Fault-Tolerant FPGA Accelerator (FTFA) is proposed in this paper to suit critical, real-time embedded automotive systems. The FTFA structure combines an evolvable TMR scheme and adaptive voting mechanisms, which intelligently and adaptively avoid failed modules without compromising their performance. Through this, a self-test (BIST) implementation system constantly checks the health of the hardware, and an online partial reconfiguration controller provides a form of on-the-fly recovery by isolating and reprogramming malfunctioning software-defined areas of an FPGA without disrupting the system's operation. Moreover, lightweight machine learning models are also introduced for real-time fault prediction, enabling proactive alleviation when faults affect system functionality. It is also designed to be deterministically scheduled and utilise low-latency interconnects, ensuring that responding to hard real-time deadlines can be guaranteed. Additionally, power-conscious fault tolerance enables the achievement of minimum energy consumption. The testing of sample automotive workloads has shown the FTFA to have hundreds to thousands of faults, meaning that the volume of faults is significantly decreased based on the lag, resulting in better fault detection and recovery relative to current solutions, without sacrificing real-time and energy efficiency. These findings shed light on the capacity of the FTFA as an effective, scalable, and robust accelerator platform to enhance the safety and integrity of next-generation automotive embedded systems that operate in high-stakes, real-time conditions.