University of Twente Student Theses


An audio based feature detector for shavers using Artificial Intelligence

Nijhuis, W.H. (2022) An audio based feature detector for shavers using Artificial Intelligence.

[img] PDF
Abstract:Current techniques to explore unit recognition in personal care embedded devices have failed, have an accuracy which is too low or are too expensive in production cost (From personal communication with employees at Philips Drachten, 2021). Previous examples are unit recognition by measuring the motor current, using a NFC communication or an electrical pin connection with a specific slot. The next alternative which will be explored is to mount a microphone inside the body of the shaver and use the audio, combined with a machine learning algorithm, to detect which unit is connected. The aim of the paper is to discover whether units can be recognized using sound and machine learning on personal care embedded devices, e.g. unit recognition. There are different ways to prepare raw audio for a machine learning network, each with its own type of machine learning network. In this paper there will be investigated which method is the most appropriate to prepare the audio for the machine learning network. By this method is also investigated which type of machine learning network is appropriate to detect the specific units. The validation of how well an unit is detected or labelled is expressed in terms of accuracy, precision and recall. The machine learning method which is chosen is to train and predict a feed-forward artificial neural network with audio. The output layer consist out of number of neurons equal to the number of units. The input layer depends on the output of preparing the audio. The audio samples are normalised in multiple ways and the output is the input of the neural network. The best result is achieved in combination with a Finite Impulse Response filter before the audio is filtered. The overall accuracy which is achieved is 91.1% for this combination. The precision values vary between 79.2% and 100%. The recall values vary between 81.3% and 97.5%. The combination consist out of the filter with a supervised feed-forward Artificial Neural Network. The results are measured for the situation when device with an unit has no load. The first results for the model while using the product looks promising but must be worked out further on.
Item Type:Essay (Master)
Philips, Drachten, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:53 electrotechnology, 54 computer science
Programme:Embedded Systems MSc (60331)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page