Predicting Motor Failures with Vibration Sensor Data (Predictive Maintenance)

Challenge

• Previously, vibration sensors were installed on power plant motors, and inspections were conducted based on predefined rules.
• However, since the specific influence of each vibration component on motor failure was unclear, plants were required to halt operations whenever vibration exceeded a certain threshold, leading to reduced operational efficiency.
• Although Statistical Process Control (SPC) methods were applied, they showed significant limitations in accuracy when the data did not follow a normal distribution.

Approach

• Data Collection with Data CAMP
  Motor sensor data was consolidated using the Data CAMP platform.
• Anomaly Detection with LISA
  LISA’s Anomaly Detector trained the model using only normal data, enabling the detection of motor anomalies by analyzing shifts in specific frequency bands.
• Feature Importance Analysis
  The feature importance of various data was analyzed and visualized, allowing users to intuitively identify components requiring immediate attention.

Result

• Achieved a 97% reduction in false alarms.
  

Full Story

The Need for Predictive Maintenance to Prevent Motor Failures

Rotating machinery like motors is extensively used in mechanical systems. Power plant turbines, which generate electricity through rotational motion, are a prime example.

When such rotating machinery fails, the power plant must shut down for repairs. If the failure is sudden and unexpected, obtaining replacement parts can take significant time, causing serious disruptions to electricity production. (It takes about two months to purchase and replace the broken part, resulting in a loss of hundreds of thousands of dollars.)

What if it was possible to determine how much the current condition of the equipment deviates from its normal state? This would enable the identification and maintenance of potentially problematic parts without halting operations, allowing for proactive maintenance. This concept is known as Predictive Maintenance.

Addressing Limitations in the Existing Inspection System

P Corporation, a supplier of bearings for power plant equipment, adopted AHHA Labs’ deep learning model to preemptively detect motor anomalies.

Traditionally, vibration sensors were attached to power plant motors, and inspections were conducted based on predefined rules. When vibration frequency exceeded a set threshold, an alarm was triggered, prompting an immediate shutdown for inspection.

However, this method had inherent drawbacks: it only provided information on vibration frequency. There was no way to identify the exact cause of the increased vibration or assess its potential severity on motor failure. In some cases, vibration exceeded the threshold without indicating a serious issue. As a result, frequent inspections unnecessarily reduced operational efficiency.

While Statistical Process Control (SPC) was used to analyze motor data, its accuracy diminished significantly when the data did not follow a normal distribution.

Anomaly Detection Using LISA

To detect anomalies without dismantling operational equipment, AHHA Labs proposed attaching external sensors to motors. Considering the structural and mechanical characteristics of power plant equipment, multiple six-axis sensors (3-axis accelerometers and 3-axis gyroscopes) were installed at key points, such as bearings, generator load sides, and gearboxes.

The collected sensor data was integrated using AHHA Labs’ big data platform, Data CAMP.

Example dashboard of the Data CAMP solution that automatically collects and visualises a variety of facility data. Image Credit: AHHA Labs

LISA’s Anomaly Detector, part of AHHA Labs’ industrial MLOps platform, trains models using only normal data, enabling predictive maintenance. Sensor data collected via Data CAMP was used to train the AI model, which analyzed shifts in specific frequency bands to detect motor anomalies in combination with other sets of data such as power output, voltage, and wind speed.

Automating Priority Identification with Feature Importance Analysis

To assist users in identifying high-priority inspection targets, AHHA Labs analyzed feature importance across the data.

Feature Importance is a machine learning technique that assigns scores to input features based on their usefulness in predicting the target variable.

The results were visualized on the Data CAMP dashboard, enabling end-users to easily monitor motor conditions and utilize the analysis results for decision-making.

Example graph visualizing Feature Importance. A way to assess how important each feature is when a machine learning model makes a prediction. Image Credit: Wikimedia commons Image Credit: Wikimedia commons

Reduced False Alarms by 97%

As a result, false alarms were reduced by 97%. For instance, out of 100 previously unnecessary inspection cases, 97 were classified as normal by LISA’s anomaly detection and feature importance analysis. This reduction in unnecessary inspections improved operational efficiency.