An Edge AI Industrial Environment Supporting Reliability Assessment and Enhancement

The expansion of decentralized intelligence has intensified the need for Edge Artificial Intelligence (Edge AI) to provide real-time decision-making, local data privacy, and reduced communication overhead directly on embedded devices. However, the adoption of these solutions relies on development platforms that assist designers in balancing hardware selection and software implementation, specifically by balancing inference performance and memory footprint. In safety-critical domains like robotics and healthcare, designers must integrate reliability into the early design phases of Edge AI systems. This involves estimating the effectiveness of different solutions, quantifying reliability targets, and determining specific methods to improve system resilience. As Edge AI moves into safety-critical sectors, the need for reliable solutions must be balanced against accelerated development cycles. However, the absence of integrated reliability tools in current Edge AI frameworks often forces designers to skip these vital analyses or implement them manually, requiring significant designer effort. Consequently, there is an urgent requirement for platforms that support early-stage reliability assessment and automated hardening. This paper presents a reliability extension to the industrial ST Edge AI Core suite, augmenting the existing toolchain with reliability assessment capabilities. The extension can i) estimate the impact of possible faults affecting the inference phase of a Neural Network (NN) ii) harden the NN model by introducing software redundancy on its most critical layer. Experimental results on a set of benchmark NNs show that the extended suite can effectively be used to harden (with a very limited impact on the designer effort) an existing model, trading off the achieved accuracy, cost, and reliability.