Johnson Case Study


Product quality monitoring using sound



The era of modern digital industrial technology, better known as Industry 4.0 or Industrial Internet of Things, is a transformation that enables gathering and complex data analysis across industrial processes, resulting in significant improvements in multiple domains and across different processes. Product quality control is one of the most important processes in every manufacturing industry. It has to be reliable and efficient at the same time to ensure constant product quality without being a bottleneck in production

Johnson Electric is a global international company, primarily active in the automotive industry, delivering various electric parts for modern vehicles. Amongst many products, Johnson delivers windshield wiper motors of various power and sizes.


Challenge


Quality control of the windshield wiper motors involves multiple steps, one of which includes acoustic motor inspection. Acoustic inspection is currently done by people trained to recognise noises that might indicate bad quality or potential malfunction in the tested motor. This approach is very subjective and error prone. It relies on persons ability to stay focused and undeterred by the events in the environment that might impact the sound perception.

Due to the described problems, Johnson Electric approached Netico to implement a pilot project and deploy Industry 4.0 technologies to automate the process and make it more reliable and efficient.


Solution


The project “Advanced DC Motor Noise Analysis System”, uses modern edge computing technologies combined with machine-learning (deep learning) algorithms and the real-time process control actions to implement automated sound-based quality control.

The main component of the system is Netico Edge Acoustic Sense 100 device, which records sound produced by the DC motors under quality tests. The Edge device pre-processes the acquired data and analyses it using the machine learning algorithms trained to recognise good and bad quality motors. The system is issuing appropriate alarms when bad quality is detected. All acquired sound data is stored in the Netico Hive environment for further statistic analysis and training of machine learning models. Initial machine learning training was done using supervised learning techniques, on a labeled data set created by people who initially did quality control.


Outcome


The described system is intensively tested by Johnson. The existing quality control process needs some modifications to create better data sets for training of the machine learning models. In addition, new, unsupervised, models are deployed and tested to improve recognition quality.

Machine leaning projects require multiple iterations and modification of data and models until full potential and desired quality is reached.