Predictive mould maintenance with acoustic AI

Challenge

Unplanned breakdowns of injection moulds disrupt production and drive maintenance cost. LEGO wanted a predictive classification model that eliminates unplanned stops and tells operators when — and only when — moulds need lubrication, using acoustic sensor audio data.

Solution

A proof-of-concept feasibility check showed that valuable insights could be extracted from the sensor packages and generalised across moulds.

• PoC model. Recurrent neural network trained on raw time-spectral audio features from test moulds.
• Scaling across moulds. Advanced feature extraction — dynamic range control, sample synchronisation, mel-spec features, Gaussian mixture models — to generalise one feature set across all moulds.
• Path to production. RNN tuning against the extracted features and evaluation of industrial-scale deployment options (Raspberry Pi / ARM + Azure ML).

Results

• Cost reduction among maintenance workers
• Lower spend on spare parts
• Optimised production output from reduced downtime
• Stabilised production & supply chain