Predictive maintenance in a manufacturing plant is collecting of meaningful data to assess the condition of machinery and then using that data to keep the plant running in the most efficient way.
Condition monitoring is the monitoring of several parameters such as equipment vibration and temperature to identify potential issues such as misalignments or bearing failures. Condition monitoring tools can, for instance, map equipment degradation when a vibration analysis shows a change in the harmonic frequency of rotating equipment components.
Predictive maintenance is based on condition monitoring, anomaly detection, and classification algorithms, and integrates predictive models which can estimate the remaining machine runtime. This approach uses a wide range of tools, such as statistical analyses and machine learning to predict the state of the equipment.
STMicroelectronics (ST) has deployed advanced ICs and an ecosystem of evaluation tools, software, documentation, and online dashboards for remote monitoring, which can be used for condition monitoring and predictive maintenance.
Overview
Strategically placed sensors are the basic elements that provide raw data. Sensor nodes contain the raw sensors plus electronics that do initial processing and connect to gateways that coordinate the data, further process it, and send the processed information to a central controller and/or the cloud.
Sensor Nodes
Sensor nodes are designed to be mounted directly on a machine’s critical components, such as bearings, shafts, or motor housings. They collect data and perform initial preprocessing before transmitting the data to the gateway and cloud.
A typical sensor node includes:
Built-in sensors for such things as vibration, tilt, temperature, pressure, motion.
Signal conditioning and communication interfaces, enabling real-time decision-making and data exchange.
A microcontroller unit (MCU) to process the sensors’ data.
A communication module to transmit data either with wireless connectivity (e.g., Bluetooth® Low Energy, Wi-Fi®, LoRa, IoT, etc.) or wired (e.g., IO-Link, Modbus, CAN, EtherCAT, RS-485/RS-232, etc.).
A power source such as a battery or energy-harvesting system.
Memory for temporary data storage.
Gateways
A gateway acts as an intermediary device that connects sensor nodes to a central system or cloud platform. It aggregates, processes, and transmits data from multiple sensor nodes to ensure seamless communication between devices and networks.
A typical gateway includes:
An MPU to handle data aggregation and processing.
Communication interfaces such as Ethernet, Wi-Fi®, or cellular modules for cloud connectivity.
Local storage for temporary data storage.
A power supply to ensure continuous operation.
Security features to protect data during transmission.
Case Study
This case study explores the implementation of a wireless sensor node for condition monitoring and predictive maintenance of a high-power industrial motor. The focus is on high-accuracy, frequency-based vibration analysis.
In industrial environments, accelerometers are usually the primary choice for machine monitoring. If you use devices specifically created for the application, they can provide reliable vibration data in the whole vibration spectrum of interest, to detect the most important mechanical issues, like bearing failure, shaft misalignment, unbalance, lubrication failure, and mechanical wear and tear. It is important that the accelerometers are designed to be well-suited to the harsh conditions of the factory floor.
Accelerometers provide data in the form of acceleration (g) over time, typically sampled at high frequencies (20 kHz or higher). Some have the ability to process data in the edge, directly in the sensor.
Intelligent sensors from ST have embedded machine learning core (MLC) technology so preprocessing can be done inside the sensor node. An embedded intelligent sensor processing unit (ISPU) enables filtering, noise reduction, and fast Fourier transforms (FFTs) to convert time-domain data into frequency-domain data. This brings benefits in terms of power consumption reduction, low latency, and optimal partitioning of the computation.
Complex features, like envelope, integration, and rms value that are typically computed starting from time-domain signals are used together with the frequency-domain representations (FFTs) to identify different possible classes of failures associated with machines.
The cloud layer handles advanced, computationally intensive AI models that analyze larger aggregated datasets for deep insights, long-term trend analysis, and accurate predictive maintenance, such as estimating remaining useful life. This combined approach ensures fast local responsiveness while leveraging powerful centralized analytics for scalable and effective maintenance strategies. The features and the frequency-domain data computed in the edge are further analyzed in the cloud layer using AI to identify patterns, such as harmonic frequencies or sidebands, indicative of specific faults.
AI Model Adoption
In condition monitoring and predictive maintenance, AI model adoption typically follows a hierarchical approach between the edge and cloud layers. At the edge level — which in the case of ST’s MEMS sensors could be within the sensor itself with machine learning techniques, lightweight to moderately complex models process data close to the source, enabling real-time anomaly detection, contextual analysis, and preliminary diagnostics with low latency and reduced data transmission. This preprocessing reduces the amount of data that has to be sent to the cloud.
Sensor Fusion
Sensor fusion is the process where data from the different sensor nodes is combined to provide various useful outcomes. It is achieved through a multi-layered process involving data collection, preprocessing, integration, and analysis across different system levels.
Fusion begins at the edge, with filtering out of noise, calibrating signals, and performing preliminary fusion techniques such as averaging, weighted combination, and feature extraction to generate a coherent unified data representation. The fused or partially processed data is then transmitted via communication protocols such as Wi-Fi, LoRa, or cellular to the gateway and cloud infrastructure for further processing.
Data from multiple sensor nodes or gateways, and possibly other sources, are aggregated in the cloud. Sophisticated fusion algorithms — such as Kalman filters, Bayesian inference, machine learning models, or deep learning are applied to combine heterogeneous data streams into comprehensive, high-level insights.
The cloud processes the fused data to generate actionable information, predictions, or alerts. This information can be sent back to devices or operators for real-time control, optimization, or maintenance decisions.
This hierarchical approach ensures efficient use of bandwidth, reduces latency by performing edge-level fusion, and leverages powerful cloud computing for complex data integration and analysis.
Connecting to the User
Predictions might be channeled to the user via alerts and notifications indicating potential faults, their severity, and recommended actions or with detailed visualizations of vibration trends, fault classifications, and remaining useful life estimates. This depends on the system level and software design of the condition monitoring/ predictive maintenance provider, and it is not linked to the specific vibration sensor selected to collect the data. However, the quality of results strictly depends on the sensor and its ability to provide accurate, reliable, wide-bandwidth vibration data.
Using the Information
The information can be used for proactive maintenance planning, in order to strategically schedule maintenance activities, thereby minimizing unplanned downtime and maximizing equipment availability.
Predictive analytics can also be used to accurately pinpoint the fundamental sources of recurrent faults, enabling targeted corrective actions.
Prediction outcomes can be used to fine-tune machine parameters, reducing mechanical stress and wear, which contributes to extending the service life of critical components and optimizing overall system performance.
In Conclusion
Moving from condition-based monitoring based on sensor data to anomaly detection computed inside the sensors represents a major benefit for industrial applications, offering significant advantages for system-level power efficiency, downtime reduction, and reduced maintenance costs.
This article was written by Carlo Larghi, MEMS Sensors Product Marketing Manager, STMicroelectronics. For more information, contact him at
