Detecting and Understanding Anomalies

What Are Anomalies in Sound Data?

Anomalies are deviations from expected patterns in sound data. These deviations can be caused by sudden faults, gradual wear, or environmental changes. Detecting anomalies is crucial in applications such as predictive maintenance, quality control, and safety monitoring.

Types of Anomalies in Sound Data

Generically, we consider two type of deviations:

• Short-term anomalies – Sudden outliers, such as an unexpected loud noise, may indicate a fault requiring immediate action.

• Long-term changes – Gradual shifts, such as wear on mechanical components, altering sound properties over time.

Resonyx anomaly score, anomalies in loudness being analysed based on audio and line voltage.

Methods for Anomaly Detection

We utilize a combination of statistical models and advanced neural network-based approaches to detect anomalies. These models can identify deviations from normal patterns and predict failures before they occur. Some common techniques include:

• Time-series analysis – Monitoring feature values over time to detect sudden spikes or gradual shifts.

• Spectrogram-based anomaly detection – Identifying new patterns in sound or spectral changes that indicate a fault.

By analysing deviations from normal patterns, we can predict failures and optimise maintenance processes, ultimately improving reliability and efficiency.

Anomalies and operational data

Sometimes, sound changes simply because the operation of the machine changes. If Resonyx has access to data describing these operations, this can be incorporated into our models, enhancing the insights that it can provide and helping distinguish between expected and unexpected sound changes.