📘 Resource Overview & Documentation Details

You are viewing the technical directory listing for Predictive Maintenance for Boilers: What the Data Shows. This engineering resource is curated under our Predictive Maintenance division to support automation specialists, control systems engineers, and manufacturing technicians looking for verified reference materials.

🔍 Contextual Abstract & Technical Scope

Share Summary:LinkedInEmail

Predictive maintenance depends on more than installing sensors or selecting a machine-learning algorithm. Reliable forecasts require a connected data pipeline, careful preparation of equipment logs and evaluation methods that account for rare failures. A study of boilers and heat pumps illustrates how these pieces fit together—and why predicting severe faults remains difficult even when operational data is available.

The work used an Internet of Things (IoT) platform to collect and process boiler data, then tested models for identifying whether a fault might occur within the following seven days. For industrial automation teams, its central lesson is practical: data architecture and time-aware analysis matter, but a high overall accuracy score alone does not prove that a system can reliably detect the failures that matter most.

Predictive maintenance starts with an end-to-end data pipeline

A predictive maintenance system needs to move information from equipment to a point where it can inform maintenance decisions. The architecture described in the study includes data sources, acquisition, storage, processing, data mining and visualization. These functions connect device telemetry with analysis and, ultimately, information that operators or maintenance teams can use.

The researchers implemented this approach through the Smart Cloud of Things (SCoT), a platform built from containerized services. Its components support device connectivity and management, while a data gateway routes information to storage and real-time services. Rather than forcing every data type into one database, the implementation used Cassandra, PostgreSQL and InfluxDB for NoSQL, relational and time-series storage, respectively. A RabbitMQ broker provided access to streaming data.

This separation is relevant to automation projects that combine device telemetry, configuration information and time-stamped measurements. A gateway can transform and direct incoming data to the appropriate backend, while separately deployed services can be scaled to handle workloads. The platform provides the foundation for analytics; it does not, by itself, guarantee that fault predictions will be accurate.

Preparing boiler data for fault prediction

The dataset covered 16 months of information from 1,000 boilers across ten models. For the experiments, the researchers used a subset representing around three consecutive complete months. Each boiler was characterized by 91 features, including measurements such as temperature, operating counts and heating-cycle information. The logs were event-driven: a feature was recorded when its value changed, so timestamps could contain gaps for individual variables.

To make the data suitable for supervised learning, the team applied several preparation steps:

  • Fill gaps: Missing feature values were carried forward from the most recently recorded value.
  • Reduce log volume: Records were aggregated into consecutive, non-overlapping ten-minute intervals.
  • Assign a target: Each observation was labelled according to the worst fault level seen in the next seven days: no fault, light fault or severe fault.
  • Scale and select features: Features were standardized, and a random forest-based selection process removed about 75% of features considered not relevant to the classification task.

The class distribution made the task especially challenging. About 90.8% of the logs represented normal operation, 8.8% were associated with light faults and 0.4% with severe faults. In this kind of imbalanced dataset, a model can appear successful by predicting the common class while missing rare but operationally important events.

Why time-series models matter

Equipment behaviour unfolds over time, so an isolated reading may miss a pattern that becomes clearer across successive observations. To account for this temporal relationship, the study tested Long Short-Term Memory (LSTM) networks, a type of recurrent neural network designed to process sequences. The models used a one-day window of ten-minute observations to predict the fault class in the following week.

Several LSTM configurations varied how much consecutive days overlapped in the training samples. The researchers also compared them with models that treated each ten-minute observation independently, including neural networks and a decision tree. This comparison helped assess whether sequence information added value beyond individual snapshots.

Evaluation included precision and recall for each fault class, alongside accuracy, macro F1-score and Matthews correlation coefficient (MCC). These measures provide different views of performance. In particular, class-specific recall indicates how many actual faults a model finds, while precision indicates how often its fault alerts are correct. Macro F1 gives equal weight to each class, making it more informative than accuracy when class frequencies are highly uneven.

Results: useful signals, limited severe-fault detection

No model performed best across every metric. One LSTM configuration achieved the highest reported accuracy, 0.89, but it assigned the no-fault class to most test samples. That result demonstrates why accuracy should not be read in isolation: a score driven by the dominant class may say little about the ability to identify rare failures.

Another LSTM configuration produced a more balanced class assignment and achieved the best reported MCC, 0.21. It correctly identified a small proportion of severe-fault samples, with a severe-fault precision of 0.05. Across most tested models, severe-fault precision and recall were close to zero or otherwise weak. The study therefore found that forecasting severe faults up to a week ahead was still a difficult problem, despite some improvement from models that used temporal information.

For maintenance planning, this distinction has direct consequences. Missing a severe fault can leave equipment unavailable, while a false alert can trigger an unnecessary service response. Model selection and thresholds should therefore reflect the operational cost of each outcome, rather than relying on a single headline metric.

Practical implications for automation teams

The findings point to several considerations when building an equipment-monitoring program:

  • Design for the full data lifecycle. Device connectivity, storage, processing and presentation need to work together before predictions can support maintenance workflows.
  • Preserve temporal context. Trends and sequences may carry information that individual sensor readings do not, so time-aware models are worth evaluating.
  • Measure rare classes directly. Review precision and recall for critical fault categories, as well as macro F1 or MCC, instead of judging performance by accuracy alone.
  • Check the forecast horizon and data window. The authors identified the seven-day horizon, ten-minute aggregation and limited three-month subset as factors that could have constrained results.

Predictive maintenance for boilers is not simply a modelling exercise: it is a coordinated combination of IoT data acquisition, data quality work and operationally appropriate evaluation. In this study, sequence-based models showed promise, but severe-fault predictions remained limited. Automation and maintenance teams should treat such models as decision-support tools to be validated against their own data and response requirements, with particular attention to rare failures and the cost of missed or unnecessary interventions.

Access Official Technical PDF