Time-Series Analysis and Computer Vision

Turning Audio Features into Time-Series Data

Many of the audio features computed by Resonyx are extracted using a sliding fixed-size window, generating time-series data where each point represents a specific feature value at a given timestamp. This enables the tracking of feature evolution over time, which is crucial for detecting trends and patterns.

Analysing Trends in Time-Series Data

Most of our time-series analysis models focus on tracking sound behaviour over time to identify patterns. These can include:

• Cyclical patterns – Repeating behaviours in the sound signature, such as machinery operating in cycles.

• Drift and gradual changes – Slowly evolving variations in sound properties due to wear and tear.

• Sudden deviations – Unusual spikes or drops in sound features that may indicate faults.

By analysing these trends, we can provide early warnings about potential failures, optimising maintenance efforts and reducing downtime.

Computer Vision Techniques for Sound Analysis

Another important sound representation used by Resonyx are spectrograms (see previous post). Since these resemble greyscale images, we have adapted computer vision techniques originally designed for image processing to analyse them. The main approaches include:

• Classification – Assigning audio samples to predefined categories, akin to image labelling.

• Event Detection – Identifying when specific sound events occur within a clip, similar to object detection in images.

• Anomaly Detection – Identifying audio samples that sound different from the ones recorded during the training period. (More details in posts to come)

We adapt state of the art vision deep learning models to efficiently analyse spectrograms, for example modern one-stage object detectors for real-time sound event detection.