Machine learning model-based acoustic signal-to-text translation

A system includes a processor and a memory storing software code and a machine learning (ML) model.

A system includes a processor and a memory storing software code and a machine learning (ML) model. The processor executes the software code to receive data including a sequence of acoustic signals or a corresponding spectrogram and an identifier of a source of the acoustic signals, translate, using the ML model, the acoustic signals or the spectrogram to one or more words each describing a respective acoustic signal pattern included in the sequence of acoustic signals or represented in the spectrogram, and obtain, using the identifier of the source, a predetermined word-based description of an acoustic signature of the source. The software code is further executed to determine, using the translated one or more words and the word-based description, whether an anomaly of the source is indicated, and output, when an anomaly is indicated, an alert.

BACKGROUND (1) Mechanical systems experience wear and tear on the components that make up those systems. As the components degrade, the performance of a system deteriorates and eventually the system can fail. A system failure has many undesirable consequences including lost productivity, lost revenue, safety concerns for human operators, costly repairs and potentially environmental impacts. Even substandard system performance has possible consequences including product defects and lower productivity. Although persistent monitoring of mechanical systems can keep systems operating at peak performance, such monitoring can involve modifications to the system itself (e.g., the addition of mechanical and electrical sensors) which can themselves introduce more opportunities for failures and mechanical problems. (2) One approach to addressing the problem described above is mechanical fault diagnosis based on audio signal analysis (MFDA). MFDA typically involves recording audio signals of a system and comparing the audio to recordings of the same system in a known good state. When the new audio differs enough from the known good audio, an alert is triggered. However, conventional approaches to employing MFDA merely perform a comparison that results in a pass/fail result, with no specificity regarding the nature of the difference in sound. Alternatively, a system operator is presented with audio waveforms or other scientific representations of audio signals that do not correlate readil