Utilization of acoustic signals with generative Gaussian and autoencoder modeling for condition-based maintenance of injection moulds

ABSTRACT

Predictive and condition-based maintenance is given more and more attention to further optimize the utilization of manufacturing and production equipment. Utilizing acoustic signals for equipment monitoring and condition-based maintenance has been proven effective in many applications. Many manufacturing and production setups consist of multiple alike machines (e.g. within injection moulding) where it would be beneficial to use the same monitoring setup and configuration on all machines. Based on an industrial application within injection moulding (using five different injection moulds), a methodology is proposed utilizing acoustic signals from injection moulds combined with generative Gaussian or autoencoder modeling. To improve the generalization ability of generative modeling to moulds not seen at training time, a simple yet effective model adaptation is proposed, which only requires a few faultless moulding cycles at runtime/test time. The best results are obtained using the Gaussian model, where area under the curve values close to one are achieved when employing a model adapted to the specific mould at test time to detect abnormal situations like mechanical-defective moulds (loose latch lock) and the need for lubrication. The proposed framework is light in terms of computation and makes the setup implementation practically feasible in a real industrial context with multiple similar machines.

KEYWORDS:

• Acoustic signal processing
• predictive maintenance
• injection moulding
• Industry 4.0
• machine learning
• autoencoder

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1. Particularly, in this work, $\mathit{\tau }=14$ was heuristically selected taking into consideration the present data scarcity scenario (see Table 1).

2. After autoencoder adaptation to a mould not seen at training time, performance on the training dataset is close to random guessing.