Insight / Blog

Sound and vibration are essential indicators of equipment condition. Rule-based diagnostics typically monitor fixed frequency thresholds, making them vulnerable to background noise and changes in operating speed. They can raise false alarms during normal operation or miss subtle signs of wear.

Industrial AI can analyze multiple frequency characteristics, including FFT-based features, and learn normal operating patterns. Company A, an electronics manufacturer, wanted a foundation that would let domain engineers train and operate diagnostic models without writing new code for every iteration.

Background : False Alarms and Developer-Dependent Model Operations

Noise and vibration can reveal mechanical faults and abnormal conditions early, but data collection, preprocessing, model training, and operational management are often handled in separate tools.

When model development remains limited to specialist developers, engineers who understand the equipment cannot easily define problems or improve models. Company A needed an operating structure built around industrial signal data and field users.

Approach : Integrated Preprocessing, Training, and Deployment

SURROMIND connected preprocessing, feature extraction, model training, and deployment in one repeatable workflow for noise and vibration data.

• Ingested and integrated noise and vibration data
• Automated data cleaning, preprocessing, and feature extraction
• Extracted FFT, Log-Mel Energy, and other signal features
• Trained Autoencoder-based anomaly-detection models
• Validated the performance of candidate diagnostic models
• Automated model deployment and operational management

The objective was not a one-time model. It was an environment where domain engineers could load data, compare models, and review operating results without building a new software workflow each time.

Results : Validated Accuracy and 400-1600x Faster Training and Inference

Model	Performance
Anomaly detection	Approximately 0.99 accuracy under Autoencoder validation conditions
Classification	100% accuracy on validation data with CNN-based classification
Optimization	400-1,600x reduction in training and inference time after dimensionality reduction

Operational Impact : Sustainable MLOps Led by Domain Engineers

The main bottleneck in industrial AI is not a single algorithm.

New data continues to arrive, and equipment condition continues to change.

Models therefore need to be evaluated and improved during operation.

This project treated noise and vibration as a reusable operating asset, not a one-time analysis dataset.

Preprocessing and feature extraction were automated.

Field teams could compare multiple models and select the option that best matched operating conditions.

The resulting workflow supported anomaly detection, fault diagnosis, and predictive-maintenance use cases.

It also established a repeatable path from raw signal data to deployed diagnostic models.