Visualizing Neural Network Decisions for Industrial Sound Analysis
Recent research has shown acoustic quality control using audio signal processing and neural networks to be a viable solution for detecting product faults in noisy factory environments. For industrial partners, it is important to be able to explain the network's decision making, however, there is limited research on this area in the field of industrial sound analysis (ISA). In this work, we visualize learned patterns of an existing network to gain insights about the decision making process. We show that unwanted biases can be discovered, and thus avoided, using this technique when validating acoustic quality control systems.