Vibration data is a valuable source of information about the condition of a system and can be used to detect and predict damage. However, extracting useful information from vibration signals is a challenge. Recently, machine learning (ML) methods, especially neural networks (NNs), have been used to extract patterns and detect damage from vibration signals. However, designing an NN model is not a trivial task and, in most cases, we end up with an over-complex model that requires a long training time with a long latency. Neural Architecture Search (NAS) [1] provides an automated framework to design NN models with optimized performances. However, NAS algorithms usually are computationally expensive and can lead to a significant carbon footprint. This paper proposes a novel NAS algorithm with almost zero training cost. Any NAS algorithm uses a technique to validate models while searching for the best model. This Validation step takes the most time and energy because the algorithm usually needs to train the model fully to know its performance. The zero-cost NAS algorithm uses novel validation techniques (or proxies) instead of the standard validation step in NAS algorithms to reduce energy and time consumption. This reduces the computational cost of NAS to almost zero [2], without sacrificing accuracy. The proposed algorithm was evaluated on a vibration dataset to detect damage in wind turbine gearboxes. It showed its superiority over human-designed models in the sense of accuracy and complexity.