Maximo Health, Predict, and Reliability Strategies: Building a Predictive Maintenance Practice

Most organizations already capture enormous amounts of asset data in Maximo Manage. Work orders record what broke and what was fixed. Inspections capture condition observations. Meters record runtime, cycles, and sensor values. Purchases and inventory track what was spent and what was consumed. Yet very few organizations use that data to predict what will fail next. The gap between data collection and predictive action is where Asset Performance Management (APM) lives, and it is exactly the gap that Maximo Health, Predict, and Reliability Strategies are designed to close.

IBM Maximo Application Suite provides a unified APM capability that combines Health, Predict, Monitor, Visual Inspection, and Reliability Strategies in a single platform. Health transforms operational data into scores that show current condition and risk. Predict uses artificial intelligence and machine learning to forecast failure probability and time to failure. Monitor ingests time-series sensor data. Visual Inspection applies computer vision to defect detection. Reliability Strategies brings a library of RCM studies to maintenance planning. Together, these components enable a shift from reactive and calendar-based maintenance to condition-based and predictive strategies.

The 9.x release line has made APM more accessible and more powerful. MAS 9.0 laid the foundation for Health scoring and Predict integration. MAS 9.1 introduced Maximo Reliability Strategies as an add-on. MAS 9.2 added industry accelerators, configurable dashboards, dissolved gas analysis for transformers, Weibull distribution analysis, and tighter connections between Reliability Strategies, Health, Job Plans, and meters. For organizations serious about reliability, the upgrade to 9.x is not just a technical refresh; it is the delivery vehicle for predictive maintenance.

This article explores how to build a predictive maintenance practice using Maximo Health, Predict, and Reliability Strategies. We will look at how health scoring works, how predictive models are built and deployed, how RCM-based strategies are created, what is new in MAS 9.2, and why data quality is the single most important success factor. The content is written for reliability engineers, maintenance strategists, and the data scientists who support them.

From Reactive Maintenance to Predictive Asset Management

Reactive maintenance waits for something to break, then fixes it. Calendar-based maintenance replaces or services assets on a fixed schedule regardless of condition. Both approaches have roles, but both are inefficient. Reactive maintenance leads to unplanned downtime, safety incidents, and secondary damage. Calendar-based maintenance wastes resources on assets that are still healthy and misses assets that degrade faster than expected.

Predictive maintenance aims to intervene at the right time: before failure, but not so early that maintenance is wasted. It requires three things: data that describes asset condition, models that turn that data into predictions, and processes that act on those predictions. Maximo APM provides the platform for all three. Manage provides the work execution backbone. Monitor provides the time-series data. Health provides the scoring framework. Predict provides the machine learning. Reliability Strategies provides the maintenance strategy library.

The transition from reactive to predictive is not instantaneous. It requires cultural change as much as technology. Maintenance crews must trust that a model's recommendation is worth acting on. Planners must integrate predictive alerts into the work schedule. Capital planners must use health and cost data to justify replacement decisions. Data scientists must maintain models and explain their outputs. Executives must fund the data collection and cleanup work that makes the models possible.

A common mistake is to treat predictive maintenance as a one-time project. Models degrade over time as operating conditions change, assets are replaced, and new failure modes emerge. A predictive practice requires continuous monitoring of model performance, retraining, and feedback from the field. The platform can automate much of this, but domain expertise remains essential. A model may predict bearing failure, but only a reliability engineer can decide whether to replace the bearing, lubricate it, or run to failure based on criticality and operational context.

Maximo Health: Scoring Models and Contributors

Maximo Health is the foundation of the APM approach. It takes data already in Maximo Manage and other sources, then calculates scores for health, criticality, and risk. These scores help teams prioritize maintenance, optimize reliability, and make better operational decisions. Health is not a black box. It uses a structured framework of asset categories, scoring groups, contributors, and thresholds that organizations can configure to match their specific assets and operating contexts.

Asset categories group assets that share similar scoring logic. For example, a utility might create categories for "High-Voltage Transformers," "Distribution Poles," "Circuit Breakers," and "Critical Pumps." Each category has its own scoring model because the factors that determine transformer health are different from the factors that determine pump health. Categories are the starting point for meaningful scoring.

Contributors are the factors that influence a score. Common contributors include overdue preventive maintenance work orders, work order backlog, equipment age, inspection results, failure history, meter readings, and environmental conditions. Maximo Health includes out-of-the-box contributors, and organizations can extend them for specific needs. Each contributor is assigned a weight that reflects its impact on the overall score. A contributor with a weight of 30 percent has more influence than one with a weight of 10 percent.

Thresholds define what the scores mean. A health score of 90 might indicate excellent condition, 70 to 89 indicates good, 50 to 69 indicates fair, and below 50 indicates poor. These thresholds should be set based on operational experience and risk appetite, not arbitrary defaults. An asset in poor health is a candidate for urgent inspection, accelerated monitoring, or immediate work.

The scoring framework can be represented conceptually as follows:

{
  "assetCategory": "HIGH_VOLTAGE_TRANSFORMER",
  "scoringGroups": [
    {
      "name": "Health",
      "weight": 1.0,
      "contributors": [
        { "name": "overdue_pm", "weight": 0.25, "direction": "lower_is_better" },
        { "name": "inspection_defects", "weight": 0.20, "direction": "lower_is_better" },
        { "name": "dissolved_gas_trend", "weight": 0.30, "direction": "lower_is_better" },
        { "name": "age_adjusted_condition", "weight": 0.25, "direction": "lower_is_better" }
      ],
      "thresholds": [
        { "min": 90, "label": "Excellent" },
        { "min": 70, "label": "Good" },
        { "min": 50, "label": "Fair" },
        { "min": 0, "label": "Poor" }
      ]
    }
  ]
}

This JSON is a simplified view of how a scoring group might be configured. In practice, the contributor calculations are defined in Health and may pull from Manage records, Monitor time series, or custom data sources. The key point is that scoring is configurable and transparent. Teams can explain why an asset received a particular score, which is essential for building trust with maintenance crews and regulators.

Health scoring also supports criticality and risk. Criticality reflects the consequence of failure, independent of condition. A spare pump in a warehouse might be in excellent health but low criticality. A single transformer feeding a hospital might be in good health but very high criticality. Risk combines health and criticality to prioritize actions. An asset in poor health and high criticality is the highest priority. An asset in good health and low criticality is the lowest.

Maximo Predict: Failure Probability, Time to Failure, and Anomalies

Maximo Predict extends Health with machine learning. It uses time-series data from Maximo Monitor and failure data from Maximo Manage to build models that answer predictive questions. Common questions include: What is the probability that this asset will fail in the next 30, 60, or 90 days? How many days remain until a specific failure mode occurs? Are there operating-context-specific anomalies that suggest an impending problem? What maintenance schedule is needed to avoid failure?

The modeling workflow in Predict is collaborative. Data scientists work with asset experts to define asset groups, select failure modes, and choose modeling approaches. Predict provides default notebooks that data scientists can extend or replace with custom notebooks. Models are trained and deployed in Watson Machine Learning, and the results flow back into Maximo asset records. Each asset can display a Predictions section with failure probabilities, days to failure, and anomaly flags.

Asset groups are the starting point. A group is a set of assets that share enough characteristics to be modeled together. All pumps of a specific model in a refinery might form one group. All transformers of a specific vintage on a distribution feeder might form another. The group definition matters because a model trained on mixed asset types will be less accurate. Group IDs are used by notebooks to fetch the right training data and generate predictions for the right assets.

The default notebooks cover common predictive tasks. They include data preparation, feature engineering, model training, evaluation, and deployment steps. Data scientists can customize them to incorporate organization-specific features or algorithms. The models must be deployed to Watson Machine Learning to be consumed by Maximo Predict. This integration is one reason why MAS 9.1 and 9.2 support specific Cloud Pak for Data versions.

A simplified view of the Predict workflow might look like this:

# Conceptual Predict notebook fragment
import pandas as pd
from maximo_predict import AssetGroup, PredictModel

# Load asset group and sensor data
group = AssetGroup(group_id="REFINERY_PUMPS_4500RPM")
sensor_df = group.load_monitor_data(days=180)
failure_df = group.load_manage_failures()

# Engineer features
features = group.build_features(
    sensor_df,
    rolling_windows=["7d", "30d"],
    aggregations=["mean", "std", "max"]
)

# Train and evaluate model
model = PredictModel(model_type="regression", target="days_to_failure")
model.train(features, failure_df)
metrics = model.evaluate()
print(f"RMSE: {metrics['rmse']}, MAE: {metrics['mae']}")

# Deploy to Watson Machine Learning
model.deploy(space_id="wml-space-12345")

# Predictions now available in Maximo asset records

This Python snippet is illustrative, but it captures the essence of the workflow: load data, engineer features, train, evaluate, deploy, and consume. In practice, the notebooks provided by IBM handle much of the plumbing, but data scientists still need to understand the asset domain to build useful features.

Anomaly detection is another Predict capability. It identifies operating-context-specific anomalies that may signal a failure before traditional thresholds are breached. For example, a motor might show normal vibration levels in absolute terms, but abnormal vibration given its current load and speed. Context-aware anomalies are more useful than simple threshold alerts because they reduce false positives and catch subtle degradation patterns.

Work queues in Predict help reliability engineers track assets that have a high probability of failure or that will fail before the next scheduled preventive maintenance work order. These queues bridge the gap between prediction and action. An asset on the queue can be inspected, monitored more closely, or have a work order generated. Without this bridge, predictions become interesting but unused reports.

Reliability Strategies and RCM-Based Maintenance Planning

Maximo Reliability Strategies, introduced as an add-on in MAS 9.1, brings Reliability Centered Maintenance (RCM) capability into the suite. IBM acquired a library of RCM studies covering hundreds of asset types and tens of thousands of failure modes, along with suggested preventive maintenance activities and step-by-step tasks. This library gives organizations a starting point for building asset-specific maintenance strategies without conducting a full RCM study from scratch.

The core idea is simple but powerful. Instead of guessing what maintenance to perform, organizations can select an asset type, review the documented failure modes for that type, and choose mitigation strategies that match their operating context. The library includes recommended tasks, frequencies, and logic for each failure mode. Organizations can also create custom strategies on the Custom Strategies tab for assets that are not covered by the standard library or that have unique operating conditions.

MAS 9.2 extended Reliability Strategies with AI capabilities to build out failure and remediation descriptions. It also tightened the connection between Reliability Strategies, Health, Job Plans, and meters. This means a strategy can generate job plans, which generate preventive maintenance work orders, which feed back into health scores and failure data. The loop is closed: strategy drives execution, execution generates data, data informs strategy.

The value of Reliability Strategies is especially clear for organizations with large, diverse asset populations. Conducting a custom RCM study for every asset type is expensive and time-consuming. The library provides a baseline that can be adopted, adapted, or overridden as needed. For common asset types such as pumps, motors, valves, and transformers, the library covers the majority of failure modes that maintenance teams encounter.

However, the library is not a replacement for engineering judgment. Operating context matters. A pump running in a clean environment with redundant backup has different maintenance needs than the same pump running in a corrosive environment as the sole source of cooling. Reliability Strategies supports customization precisely because context cannot be fully standardized. Use the library as a starting point, then refine it with local knowledge.

MAS 9.2 Advancements: Accelerators, Weibull, and Duval Triangle

MAS 9.2 added several capabilities that matter specifically for predictive and reliability workflows. The lightweight Monitor engine scales more efficiently, which is important for organizations ingrowing large volumes of time-series data from sensors and historians. Configurable dashboards give reliability engineers more control over how health and predict data is visualized. These improvements reduce the friction of scaling APM from a pilot to a production deployment.

Industry accelerators are a major addition. MAS 9.2 includes scoring models for electrical distribution and transmission, covering health, criticality, effective age, end-of-life probability, and risk. These accelerators are not generic templates; they are built for the asset classes and failure modes common in electrical utilities. For transmission and distribution teams, this can slash the time needed to stand up a health scoring practice. The models can be adopted as-is or tuned to match local asset populations.

Dissolved gas analysis for transformers is another utility-focused enhancement. Transformer oil contains dissolved gases that indicate different types of internal faults. The Duval triangle visualization helps reliability engineers interpret gas ratios and classify faults such as partial discharge, overheating, or arcing. This is a specialized but critical capability for utilities managing large transformer fleets. Prior to MAS 9.2, this analysis might have lived in a standalone spreadsheet or a third-party tool. Now it is part of the suite.

Weibull distribution analysis supports probability-of-failure calculations over time. The Weibull distribution is widely used in reliability engineering because it can model different failure patterns, including early-life failures, random failures, and wear-out failures. MAS 9.2's Weibull analysis gives reliability engineers a rigorous way to estimate when assets are likely to fail, which feeds into replacement planning, inspection scheduling, and capital budgeting.

Asset Investment Planning (AIP) also improved in MAS 9.2. It now integrates with replacement plans in Health and Job Plans, replaces effective age modeling with more robust health score-based models, and supports enhanced asset selection using standard queries. Optimization models can run in the background for improved performance. For organizations trying to defend capital plans with data, these features provide a direct link between asset condition, risk, and investment decisions.

Data Quality and Operating Context as the Foundation

No APM capability works without good data. A health score with bad contributors is misleading. A predictive model trained on incomplete failure history is unreliable. A reliability strategy based on inaccurate asset hierarchies will recommend the wrong maintenance. Data quality is the foundation, and it is the most commonly underestimated part of an APM program.

Start with asset master data. Every asset should have a clear class, specification, location, and parent-child relationship. Rotating assets should be tracked correctly. Meters should be associated with the right assets and units. Work orders should have accurate problem codes, failure codes, and completion data. Inspections should be complete and consistently coded. If this data is not clean, do not expect APM to deliver miracles.

Operating context is equally important. The same asset can have different health and risk profiles depending on how it is used. A generator used for standby power has different failure patterns than one used for base load. A valve that cycles frequently wears differently than one that remains open. Health and Predict support context through asset categories, groups, and feature engineering, but the definitions must come from domain experts.

Data governance should include feedback loops. When a model predicts a failure and inspection shows the asset is healthy, capture that feedback. When a reliability strategy recommends a task and the field crew finds it unnecessary, capture that feedback. When a health score suddenly drops because of a data entry error, fix the data and document the cause. These feedback loops are how the APM practice improves over time.

Organizations should also invest in data lineage and monitoring. Understand where each data element comes from, how it is transformed, and where it is consumed. Monitor data freshness, completeness, and anomalies. A sensor that stops reporting should not silently degrade model accuracy. A missing work order should not make an asset look healthier than it is. Visibility into data health is as important as visibility into asset health.

Practical Implications

For reliability engineers, MAS 9.x provides a credible platform for moving from calendar-based to predictive maintenance. Health scoring gives a structured way to prioritize assets. Predict models forecast failure probability and time to failure. Reliability Strategies provides an RCM library that accelerates strategy development. MAS 9.2 adds industry accelerators and specialized analyses such as Duval triangle and Weibull. The tools are now mature enough for production use, but they require clean data and domain expertise.

For maintenance planners, the integration between Reliability Strategies, Health, and Manage means that predictive insights can flow directly into work plans. Job plans can be generated from strategies, and work orders can be triggered by health scores or predict alerts. Planners should work with reliability engineers to define thresholds that generate actionable work without creating alert fatigue.

For data scientists and IT teams, Predict and Monitor require infrastructure planning. Watson Machine Learning integration, Cloud Pak for Data compatibility, sensor data ingestion, and model lifecycle management all need support. Do not treat predictive modeling as a side project. Allocate resources for data engineering, model monitoring, retraining, and explainability. A model that cannot be explained will not be trusted.

Bottom Line

Maximo Health, Predict, and Reliability Strategies turn Maximo Manage data into predictive maintenance decisions. Health provides configurable scoring for condition, criticality, and risk. Predict uses machine learning to forecast failures and detect anomalies. Reliability Strategies brings RCM-based maintenance planning into the suite. MAS 9.2 strengthens the entire stack with industry accelerators, scalable monitoring, dissolved gas analysis, Weibull modeling, and better integration with asset investment planning. The technology is ready, but success depends on data quality, operating context, and a culture that acts on predictions. Start with clean asset data, build a small number of well-defined scoring models, create feedback loops, and scale incrementally. Predictive maintenance is not a feature you turn on; it is a practice you build.