Quantization of the weights and activations of a neural network is a way to drastically reduce necessary memory accesses and to replace arithmetic operations with bit-wise operations. This is especially beneficial for the implementation on field-programmable gate array (FPGA) technology that is particularly suitable for embedded systems due to its low power consumption. In this paper, we propose an in-situ defect detection system utilizing a quantized neural network implemented on an FPGA for an automated surface inspection of wind turbine rotor blades using unpiloted aerial vehicles (UAVs). Contrary to the usual approach of offline defect detection, our approach prevents major downtimes and hence expenses. To our best knowledge, our work is among the first to transfer neural networks with weight and activation quantization into a tangible application. We achieve promising results with our network trained on our dataset consisting of 8024 good and defected rotor blade patches. Compared to a conventional network using floating-point arithmetic, we show that the classification accuracy we achieve is only slightly reduced by approximately 0.6%. With this work, we present a basic system for in-situ defect detection with versatile usability.
Observer design and optimization for model-based condition monitoring of the wind turbine rotor blades using genetic algorithm
