Implementing convolutional neural networks in has gained significant relevance in recent years. The study of the reliability of this type of networks has become a hot topic due to its great impact on a growing range of applications. Observability and controllability play a primary role when considering the systems reliability, and even more so when thinking about complex CNNs implemented in hardware in applications in which failures are not allowed, known as mission-critical applications. This work proposes an experimental setup for implementing convolutional neural networks in hardware, in which observability and controllability characteristics were included from the design phase. The proposed setup was implemented on a ZYBO Z7 APSoC development board for a LeNet-5 convolutional network architecture and a 32x32 input image dataset. The results demonstrate its usefulness since it provides the CNN designer with more detailed information about the operation of each of its internal layers.