Guide to Predictive Maintenance

Unplanned asset maintenance causes operational havoc. There’s a halt in production, with downstream and upstream impacts. Quotas aren’t met, deadlines are missed, and productivity is lost. All of this translates to substantial material losses.

Unplanned downtime costs the world’s biggest manufacturing companies 11% of their annual revenues or $1.4 trillion in total. The better news? Asset downtime can be largely avoidable with a predictive maintenance strategy.

What is Predictive Maintenance?

Predictive maintenance (PdM) measures and monitors the health characteristics of different assets to estimate deterioration trends and forecast failures. It builds on condition monitoring practices of constant quantitative data collection and in-depth analysis to detect and predict issues before any visible signs of damage manifest.

In other words, PdM systems provide critical information to the right people at the right time. With predictive maintenance insights, asset managers can:

• Decrease asset failure risks
• Reduce maintenance costs
• Extend asset service life
• Improve staff safety
• Achieve greater sustainability
• Boost organizational productivity

Most (95%) of predictive maintenance adopters reported a positive ROI, with 27% reporting full amortization in less than a year.

PepsiCo added over 4,000 hours a year of manufacturing capacity to its Frito-Lay plants after adopting a PdM system thanks to a reduction in equipment breakdowns, workflow, interruptions, and incremental maintenance costs.

Suncor, a global integrated energy company, implemented predictive maintenance solutions for 20,500 critical assets across 14 sites including oil sands, downstream, exploration and production, and pipelines—and gained over CAD 37 million in value within a year.

Predictive Maintenance vs Preventive Maintenance: Key Differences

Asset managers rely on a quintet of maintenance strategies:

1. Reactive maintenance: Unplanned, immediate repairs after an asset failure or breakdown.

2. Corrective maintenance: Planned or scheduled repairs to an asset showing early signs of damage or degradation.

3. Preventive maintenance: Performed on a predefined schedule, based on historical data.

4. Predictive maintenance: Scheduled based on real-time insights on the likelihood of service degradation or failures.

5. Prescriptive maintenance: Planned around data-driven recommendations for failure prevention in the future.

In simpler terms: reactive and corrective maintenance are applied post-factum aka when the damage is already evident. Preventive and predictive maintenance strategies are future-oriented, but they differ in terms of precision.

Preventive maintenance relies on the expected asset conditions, based on a historical baseline, past maintenance records, and compliance requirements. Based on these, the team’s schedule planned corrective maintenance, without factoring in the actual need.

Predictive maintenance, in contrast, is informed by the actual (rather than anticipated) asset condition, determined with the help of advanced analytics models. Predictive maintenance tools compare historical records against various real-time data points (e.g., ultrasonic acoustics, thermal measures, lubrication levels) to measure the current asset health. Then generate time series forecasts to suggest the optimal maintenance schedules.

Here’s a side-by-side comparison of the key differences between preventive and predictive maintenance:

Core Predictive Maintenance Technologies

Predictive maintenance is a collection of processes and technologies to enable real-time data collection and high-fidelity analytics.

A typical PdM system usually has two key components: a data aggregation module and an advanced analytics engine. The former collects data from all conceded equipment, sensors, and other operating systems (e.g., an asset management platform, SCADA, CMMS, etc). After appropriate transformations—quality checks, standardization, etc—the aggregated insights are given to an algorithm, which performs binary classifications (e.g., normal or faulty status), detects anomalies (e.g., unexplained rise in operating temperatures), and issues time-series predictions (e.g., likely service degradation in 4 weeks).

To better understand the setup, let’s have a look at each technology component individually.

Sensing Technologies

Most companies rely on Supervisory Control and Data Acquisition (SCADA) systems for remote asset and equipment monitoring and performance tuning. Modern SCADA systems aggregate data from sensors, actuators, PLCs, and communication networks to provide performance baselines.

However, data collection is often limited to process variables (e.g., temperature or pressure decrease). SCADA systems also only provide reactive alerts—after the issue has already occurred. That’s why many leaders combine SCADA insights with extra data from sensors and IoT devices.

Thanks to advances in manufacturing, sensors have become smaller in size and cheaper to produce. The average sales price per sensor reduced in half, from $0.66 to $0.29 over the last two decades.

Moreover, new sensing equipment is more durable. For instance, moderate-quality accelerometers have an IM rate of just 2% and high-quality ones 0.3%, with a lifetime cost of $5,600 and $1,350. respectively, according to Wilcoxon estimates.

To enable predictive monitoring, asset managers usually invest in a combination of sensors to capture parameters like temperature, vibration, pressure, or humidity. These can be digital or analog sensors for vibration analysis, acoustic monitoring, oil analysis, motor circuit monitoring, magnetic field measurements, and more. There’s a sensor solution on the market for almost every monitoring need at different stages of the asset lifecycle.

Many predictive monitoring scenarios require Internet of Things (IoT) devices. IoT devices combine sensing, microprocessors, communication modules, and sometimes edge computing capabilities, to aggregate and transmit data to a centralized repository for analysis.

An industrial IoT network can include:

• Smart gateway(s) to aggregate data from multiple sensors
• Edge devices performing anomaly detection locally for monitored assets
• IoT-enabled machinery that shares condition data in real time and supports remote performance tuning

Regular sensors measure physical changes in monitored assets, equipment, and operating environment. IoT devices collect, transmit, and sometimes process sensor data from multiple systems to deliver predictive insights. The main purpose of an IoT system is to enable real-time data exchanges between the monitored equipment, predictive analytics systems, and other software used for asset management.

Inspection Solutions

Sensors provide ample condition data, but they don’t usually cover all monitoring needs. For example, detect thickness loss, corrosion under insulation, or surface-level cracking.

Many regulatory scenarios also explicitly require periodical non-destructive testing (NDT) to provide a more in-depth assessment. NDT data can (and should) be also used for optimizing predictive maintenance. Lastly, the sheer size of an industrial asset (e.g., a 400 ft fire stack or offshore rig structures) makes sensor deployments either too challenging or economically non-viable.

Hence, many companies invest in additional inspection systems. For example, industrial inspection drones can collect RGB and thermal imagery, take ultrasonic thickness measurements, or perform multispectral analysis. Specialized drone payloads can help asset managers obtain extra condition data for large-scale infrastructure in less time and with more precision.

Boasting 360-degree freedom of movement, Voliro T drone can perform close-to-structure inspections with high precision in 2X less time than using conventional methods (e.g., handheld NDT equipment and rope access or lifting platforms).

Featuring six interchangeable payloads, Voliro T can be used to perform wall thickness measurements (using ultrasonic A-scans or EMAT technology), coating thickness measurements, corrosion under insulation detection, and lighting protection system testing.

With our solution, you can gather reliable condition data from elevated piping, large storage tanks, transmission towers, flare stacks, and chimneys to inform your predictive maintenance strategy.

To monitor smaller production equipment, leaders are also investing in computer vision systems. For example, gain a real-time line of sight into conveyor belt operations to detect suboptimal equipment performance. For example, detect low liquid fill rates for containers, which may indicate either poor system configuration or damage caused by degraded water seals. Or monitor remote assets. A U.S.-based heat cable provider installed an autonomous computer vision system to detect ice formation on its assets. When ice levels go above and beyond the set threshold, the system automatically adjusts cable heating levels.

Machine Learning and Deep Learning

Machine learning (ML) and deep learning (DL) transform the aggregated conditioning data into operational insights: equipment statuses, event-based alerts, time-dependent predictions, in-depth defect descriptions, and more.

The algorithms, powering predictive maintenance solutions, include:

• Regression models (e.g., linear or support vector regression) for predicting continuous, variable outcomes like remaining useful life, forecast degradation trends, or estimating the impacts of environmental stressors on equipment performance.

• Classification models (e.g., decision trees or gradient boosting) for categorizing patterns in data. They’re used for binary classification for predicting failure, root cause analysis classification, and various performance anomaly classifications.

• Time-series models (e.g., ARIMA or LSTM ) use sequential data to build predictions about the assets’ future condition and performance trends. For example, forecast equipment failure, detect gradual performance degradation, or identify recurring failure patterns

• Convolutional neural networks (CNNs) can detect patterns in multidimensional data such as a combination of spatial data, asset images, and sensor readings to deliver more accurate classifications and predictions.

One PdM can combine several ML and DL techniques to deliver more accurate insights. Deep learning (e.g., a CNN) can be used to extract features from data and then a time-series ML model can churn the predictions.

Diagnostic and Decision-Support Tools

Predictive maintenance systems are usually integrated with other software to streamline scheduling and maintenance workflows. For example, an integration with a computerized maintenance management system (CMMS) allows you to inform scheduling with predictive insights. By implementing recommendations promptly, downtime is prevented and resources are used efficiently.

An integration with an asset management platform, in turn, helps you further optimize the asset life cycles. Asset performance can be balanced against maintenance costs by using predictive asset management tools that provide risk exposure scores for different asset classes. You can also optimize your supply chain process for replacement, spare, or new parts reordering, as well as resource scheduling for carrying out corrective action.

Predictive analytics outputs can be also steamed to root cause analysis (RCA) tools to aid technicians in investigating the underlying issues contributing to performance degradation or asset failure.

Lastly, predictive models serve as a bedrock for more advanced decision-support systems that can recommend (and sometimes automatically apply) specific maintenance or performance-tuning actions in real-time aka prescriptive maintenance solutions.

Types of Predictive Maintenance Use Cases

ML and DL enable a new pane of analysis, using a wider combination of parameters and data types. Unlike earlier data analytics systems, ML-based solutions can cross-correlate data from different sources (e.g., vibration sensors, thermal imagery, and oil monitoring systems) to provide richer operational insights.

Conditionally, all predictive maintenance use cases can be grouped as:

• Indirect failure predictions
• Annomal detection
• Remaining useful life estimates

The difference between these three types boils down to data inputs/outputs, applied methods, and delivered outcomes.

Indirect Failure Prediction

Failure mechanisms are complex because multiple factors are at play — chemical exposures, usage frequency, proper exploitation, etc. Some failures also occur rather rapidly like stress corrosion cracking.

Predictive maintenance tools help estimate the likelihood of failures based on different performance variables, operating conditions, running history, and past maintenance actions. A machine learning algorithm usually assigns risk scores, based on the combination of the above factors to suggest when failures may occur.

Boeing deployed a predictive machine learning model to detect early signs of corrosion in the hydraulic spoiler power control units that can cause aircraft-on-ground events. The model was trained on over 200 billion data points from flight sensor data and maintenance records. It can detect with high precision when one of the two redundant copper wires has corroded to the point of open circuit failure. This is a major advantage for Boeing since on-board aircraft systems never generate alerts until both wires fail.

In the Oil and Gas industry, indirect failure predictions can help ensure better oil rig maintenance. For example, a predictive maintenance system can process conditioning data from drilling equipment—vibration levels, temperature changes, corrosive fluids presence, past maintenance—to deliver alerts about likely malfunction. With such foresight, rig operators can optimize spare parts delivery to the structure to avoid operational disruptions.

In the Power and Energy sector, failure predictions can help extend the asset life of wind turbines and transformers. By combining control system logs, with lubrication system data, vibration analysis, and drone-captured blade and tower data, operators can detect early signs of wear and tear to minimize downtime and maximize energy outputs.

Anomaly Detection

Anomaly detection is another great capability of predictive maintenance solutions. In this case, the algorithm is trained to detect any deviations from the ‘normal’ asset profile, indicative of possible malfunctions, defects, or degradation. Anomaly detection provides teams with ‘early warnings’ about the risks of damage or downtime, allowing faster intervention.

Assets like industrial boilers and pressure vessels are more cost-effective to service than to replace. However, tiny cracks can be hard to spot, especially when they start proliferating internally. What’s possible to detect, however, is the signs of their presence like a reduction in pressure or flow rates.

An oil company implemented a condition-based monitoring solution to track the performance of its heat exchangers. The system monitors the equipment temperatures, hydrocarbon flow rates, pressure, and other parameters in real-time to detect early signs of leakage. During the test phase, the PdM system detected recurring anomalies, which led to immediate inspection and replacement of several aging heat exchangers. This saved the company thousands in direct and indirect costs like regulatory investigations and reputational damages in case of a rupture.

Remaining Useful Life (RUL) Estimates

RUL indicates how much time is left before the asset needs to be repaired or replaced. Depending on the asset type, RUL can be defined with parameters like distance traveled, repetition cycles performed, or the time since the start of usage.

If the monitored asset degrades in a predictable way, machine learning algorithms can be used to build accurate estimates from sensor data. The algorithm can continuously benchmark asset performance against its ‘normal’ baseline to alert about early signs of deterioration. Depending on the data availability, RUL models can be trained on either lifetime asset data, run-to-failure data, or prescribed threshold degradation values, when the former two datasets aren’t available. Degradation models estimate RUL by predicting the threshold crossing of the condition indicator.

You can build appropriate datasets for RUL models by determining the leading failure indicators for different types of tracked assets. For example, oil analysis, acoustic emission, and ultrasonic testing data can yield the best data for estimating RUL in centrifugal pumps at chemical plants, according to Preston Johnson, senior delivery manager at Novity.

In the mining industry, RUL models are often used to predict the remaining useful life of haul truck engines in open-pit mines. Since the vehicles are constantly exposed to harsh conditions, they require more frequent servicing. A RUL model can be trained on data points like:

• Engine temperature fluctuations
• Fuel consumption rates
• Engine vibration levels
• Oil quality and contaminants presence

Using sensor data collected under known conditions, an RUL model is trained to predict how these indicators evolve as the engine approaches failure. For instance, as engine components degrade, the model might recognize an increase in vibration intensity combined with reduced fuel efficiency as a sign of imminent failure.

Conclusion

Predictive maintenance streamlines the analysis of asset health, providing operators with real-time foresight into performance deterioration, structural reliability, and failure risks. Thanks to early warnings, your operators can optimize planned downtime windows for maintenance, reduce unplanned downtime due to failures, and improve staff productivity.

However, the success of predictive maintenance programs heavily hinges on data availability. Especially, if we’re talking about monitoring large-scale industrial assets like bridges, dams, storage tanks, or pipelines. Sensor and IoT deployments may not always be cost-effective in such cases, while manual data collection is time-consuming due to inspection complexities.

We aim to solve this challenge with our industrial inspection drone. Deployable in minutes, Voliro T enables NDT data collection at heights using a selection of interchangeable payloads—ultrasonic transducers, EMAT, DFT gauge, a lightning protection system testing toolkit, and pulsed eddy current gauge. Collect reliable asset conditioning data with smaller crews in 2X less time to gain the missing insights for your predictive maintenance program.