Here’s a common scenario: a team uses data analysis to prevent equipment downtime at a factory by monitoring the vibration metrics of a bearing on an electric motor — frequency, amplitude, and others. Vibration analysts detect high-frequency spikes in the 2-6 kHz range, which usually indicate lubricant deficiency or contamination, and recommend that the plant engineer lubricate the bearing. After lubrication, these frequency spikes no longer appear, and the motor operates smoothly. The issue is resolved before it leads to costly repairs. For instance, a worn bearing in an onshore wind turbine can cause gearbox failure, and replacing it can cost $250,000 to $300,000.
Predictive Maintenance (PdM), which can reduce equipment downtime by up to 45%, is a comprehensive approach to equipment servicing. It combines mathematical algorithms, IIoT, AI, and cloud technologies with the expertise of vibration data analysts. In this article, I will explain how a PdM system works using the example of bearing vibration analysis, and also share how IT teams in manufacturing can prepare for implementing equipment monitoring and data analysis.
What Bearing Signals Reveal
Bearings are found in almost all industrial equipment: motors, fans, pumps, conveyors, and gearboxes. They are one of the most critical yet vulnerable components of industrial machinery, accounting for about 40% of equipment failures overall.
However, bearing defects can be detected at very early stages. This is done through vibration diagnostics — capturing vibration signals and then analyzing their frequency, amplitude, spectrum, and other parameters.
What does this accomplish? Normally, equipment vibrates evenly and predictably. But when a defect occurs, unusual frequencies appear in the vibration spectrum. For example, uneven or disproportionate vibration in three directions (along the X, Y, and Z axes) often indicates loosening of the bearing mounting or its housing, causing unstable rotation and overloading the equipment. To prevent failure, if dealt with in a timely fashion, it is enough to tighten the bearing mounting or replace the bearing.
How Predictive Maintenance Works
Vibration diagnostics have been around for over 60 years. Fifteen to twenty years ago, industrial vibrometers were used to measure vibration levels by manually placing the device on each machine. However, this method only provided a snapshot of equipment condition at the time of measurement.
Today, PdM systems perform remote, real-time equipment diagnostics thanks to IIoT, cloud computing, machine learning, and AI. Still, traditional engineering methods like vibration diagnostics and spectral analysis remain essential. Both algorithms and human experts use them to verify AI conclusions. Here are the key technologies behind modern PdM systems:
- IIoT. IIoT sensors measure and transmit bearing vibration signals to a gateway node at a frequency determined by the system settings. In addition to vibration data along two or three axes, the data packet includes the bearing temperature and device identifier. The gateway node compresses and encodes this information into a structured format optimized for fast cloud transmission via LTE, Bluetooth, Wi-Fi, or other protocols. Next, the data is decompressed — the compressed byte stream is unpacked into a full set of measurements, which are then deserialized.
- Mathematical Algorithms. The PdM system analyzes the time-domain vibration signal and performs a Fast Fourier Transform (FFT) — a mathematical operation that converts the signal from the time domain to the frequency domain. This transformation reveals signs of bearing defects, such as characteristic peaks, spikes, and noise.
- Cloud Technologies and Big Data. PdM systems are typically deployed in cloud environments that enable scalable data storage and processing. Big Data technologies are integrated into this cloud infrastructure, supporting the storage of large datasets and real-time streaming signal processing. For example, modern sensors, depending on their configuration, can capture vibration data at sampling rates of up to several tens of thousands of measurements per second, with each measurement representing a point in a time series that includes a timestamp and vibration direction (along the X, Y, and Z axes). As a result, the system processes massive volumes of structured data daily, which are stored in a relational database and classified by metric type, sensor installation point, and timestamp.
- ML and AI. The algorithms access a database containing extensive historical and current data, including computed values such as spectra (FFT), root mean square (RMS) vibration levels along axes, peak values, and trends. They compare new measurements against patterns of typical defects, detect anomalies, and generate preliminary conclusions about possible faults. However, PdM systems work not only with sensor data but also with values calculated using physical formulas that describe the behavior of rotating machinery. This enables AI to learn not only from real failures but also from simulated scenarios.
How to Prepare for Implementing Predictive Maintenance
The process of implementing a PdM system varies depending on whether you choose to build the infrastructure in-house or work with a ready-made solution provider. Here are the main stages for establishing a digital infrastructure for PdM:
- Selecting IIoT Sensors and Gateways. Choose sensors that support the required sampling frequencies and communication protocols (BLE, Wi-Fi, LTE). Verify that gateways are compatible with the company’s network or the cloud.
- Building Network Infrastructure. The system must have a reliable communication channel to the cloud for continuous data transmission. Configure routing, buffering, and packet delivery control. Transmission delays can affect the quality of analytics.
- Choosing a Platform for Data Processing and Storage. Decide where data processing and storage will occur: in a public cloud like AWS or Azure, a private cloud, or on a local server.
- Preparing Historical Data. Collect historical equipment performance metrics: vibration signals, RMS values, spectra (FFT), defect indicators. This data will be used for machine learning and equipment condition analysis.
- Integrating ML/AI Modules. Ensure connectivity between the database and the machine learning core. Algorithms must have access to both historical and current data.
- Developing Interfaces for Engineers. Set up dashboards displaying trends, peaks, spectra, and defect alerts so analysts can monitor history, statuses, and notifications.
PdM systems can increase manufacturing productivity by 5-20%, but the effectiveness of analytics depends on the quality level of the infrastructure built by the IT team. This infrastructure determines how quickly vibration analysts receive data about equipment defects and how promptly the maintenance team can respond.