Structuring predictive maintenance data pipelines with Bronze, Silver, and Gold layers transforms raw sensor streams into reliable AI-ready feature sets — giving maintenance data engineers the systematic data quality progression that prevents model degradation from inconsistent, uncontextualized, or mislabeled inputs. Start Trial Free to see how iFactory implements medallion architecture across your maintenance data sources — from raw PLC sensor ingestion through cleaned contextual records to ML-ready feature sets that predictive models can trust.
Build a Predictive Maintenance Data Pipeline That Produces Reliable AI Predictions
iFactory structures maintenance data through Bronze, Silver, and Gold layers — separating raw sensor ingestion from data quality processing and feature engineering to ensure that every AI model receives consistent, contextualized, and properly labeled inputs.
Why Unstructured Sensor Data Pipelines Produce Unreliable Predictive Maintenance Models
Most predictive maintenance implementations fail not because the AI algorithms are wrong but because the data feeding them is wrong — duplicate records from reconnecting sensors, timestamps that drift between source systems, missing context that prevents a raw sensor value from being interpreted correctly, and event labels that were applied inconsistently across work order records. A vibration model trained on data that contains 15% duplicate readings, 8% timestamp errors, and event labels that inconsistently classify bearing replacements will produce false positives and missed detections regardless of how sophisticated its architecture is. Medallion architecture solves this by separating concerns: the Bronze layer stores exactly what arrived and when; the Silver layer applies quality rules, de-duplication, and context enrichment; and the Gold layer produces the fully engineered feature sets that models consume. Engineering teams that Book a Demo with iFactory see how structured data layer separation changes model reliability outcomes compared to direct-to-model sensor pipelines.
-
Bronze Layer: Raw Ingestion Preservation
iFactory's Bronze layer stores every incoming sensor record exactly as received — preserving the original timestamp, source identifier, and raw value without transformation, providing the immutable audit record that Silver layer processing references when validating and correcting data quality issues.
-
Silver Layer: Quality and Context Enrichment
iFactory's Silver layer applies de-duplication, timestamp normalization, out-of-range detection, and context enrichment — adding asset identifiers, operating state flags, and maintenance event markers to each cleaned sensor record before it enters the model training and inference pipeline.
-
Gold Layer: ML-Ready Feature Engineering
iFactory's Gold layer computes the statistical features, rolling window aggregations, and cross-sensor derived metrics that predictive models consume — producing standardized feature sets that include failure event labels, operating context, and feature validity flags for each training and inference record.
-
Cross-Source Alignment and Synchronization
iFactory aligns sensor data, work order records, and inspection findings from multiple source systems to common asset-time reference keys — enabling Gold layer features to combine vibration, temperature, oil analysis, and maintenance event data at the correct temporal relationships without misalignment artifacts.
-
Data Quality Metrics and Pipeline Observability
iFactory tracks Bronze-to-Silver transformation quality metrics — duplicate rate, timestamp error rate, missing value percentage, and out-of-range detection rate — providing the pipeline observability needed to identify data source degradation before it affects model accuracy.
-
Incremental Processing and Backfill Management
iFactory processes new sensor data incrementally through the medallion layers while maintaining the ability to reprocess historical Bronze records when Silver quality rules or Gold feature definitions are updated — ensuring that model retraining uses consistently processed data across the full historical window.
Medallion Architecture Layers: Detailed Implementation Guide
-
Bronze Layer: Immutable Raw Data Preservation and Source Cataloguing
Foundation LayerThe Bronze layer serves a single purpose — preserve exactly what arrived from each data source, in the format it arrived, with the timestamp it carried, without any transformation that could lose information or introduce processing artifacts. For predictive maintenance, Bronze layer records include raw sensor readings (potentially with manufacturer-specific engineering units that differ from plant standard), work order records in the source CMMS format, inspection findings in whatever structure the inspection tool produces, and oil analysis reports as received from the laboratory. iFactory's Bronze ingestion pipeline appends a reception timestamp and source system identifier to each record at arrival — creating the audit trail that Silver layer processing references when flagging quality issues and that root cause analysis uses when tracing data problems back to their source. The Bronze layer is append-only: nothing is modified or deleted, and every subsequent layer transformation references the Bronze record it originated from. Teams that Start Trial can begin configuring Bronze layer ingestion from sensor systems, CMMS, and inspection platforms within the first iFactory session.
Data Sources
Sensor streams, CMMS exports, inspection tools, oil analysis reports
Transformation Policy
Append-only, no modification — exact preservation with arrival metadata
iFactory Record
Immutable Bronze store with source system, arrival timestamp, and raw value
-
Silver Layer: De-Duplication, Timestamp Normalization, and Context Enrichment
Quality ProcessingThe Silver layer applies the data quality rules that convert raw sensor records into reliable, contextualized maintenance data. For predictive maintenance, the most impactful Silver layer transformations are: de-duplication of records that arrive multiple times when sensors reconnect after communication interruptions; timestamp normalization that converts source timestamps to UTC and corrects clock drift; out-of-range detection that flags physically impossible values for exclusion from model inputs; and context enrichment that joins each sensor record to the asset register, operating mode state machine, and maintenance event timeline. Context enrichment is the Silver layer transformation most critical for PdM model quality — a vibration reading without knowledge that the machine was ramping up at the time, or had just been restarted after maintenance, will produce incorrect feature values in the Gold layer and misleading patterns in model training. iFactory applies configurable Silver quality rules per data source and asset class — enabling different de-duplication and normalization logic for high-frequency vibration streams versus low-frequency oil analysis records. Teams that Book a Demo can review Silver layer quality rule configuration for their specific sensor types and CMMS formats.
Quality Rules
De-duplication, timestamp normalization, out-of-range flagging
Context Enrichment
Asset ID join, operating state, maintenance event timeline markers
iFactory Record
Cleaned Silver record linked to source Bronze record with quality flags
-
Gold Layer: Statistical Feature Engineering and Failure Label Assignment
ML-Ready OutputThe Gold layer transforms cleaned, contextualized Silver records into the feature vectors that predictive maintenance models consume — computing rolling window statistics (mean, standard deviation, RMS, kurtosis, crest factor), frequency domain features from vibration time-series, cross-sensor derived metrics, and operating-condition-normalized values that remove the non-stationarity that confounds time series models. Failure label assignment in the Gold layer is the most consequential step for supervised model accuracy: each record is labeled based on its temporal position relative to maintenance events in the Silver layer — positive failure examples assigned to the window before each confirmed failure event, negative examples drawn from confirmed healthy periods, and uncertain periods around maintenance events excluded rather than mislabeled. iFactory's Gold layer feature engineering is configurable per asset class and failure mode — enabling vibration-specific frequency features for rotating equipment alongside thermal profile features for heat exchangers within the same pipeline. Teams that Start Trial can configure Gold layer feature definitions and label assignment rules for priority asset classes in iFactory's feature engineering module.
Feature Types
Rolling statistics, frequency domain, cross-sensor derived, condition-normalized
Label Assignment
Failure window labeling from maintenance event timeline in Silver layer
iFactory Record
Gold feature vector with asset ID, feature validity flags, and failure label
-
Maintenance Event Integration Across All Three Layers
Event ContextMaintenance event records — work orders, inspection findings, repair records, and component replacement history — must be integrated into the medallion pipeline at the correct layer and used correctly at each subsequent layer. In Bronze, maintenance events are stored as received from the CMMS without modification. In Silver, events are normalized to a common event taxonomy, linked to asset identifiers, and used as state machine inputs that flag pre-failure, post-maintenance, and normal operating periods in sensor records. In Gold, maintenance events become the label source for supervised learning — the timing and type of each replacement or repair event determines which sensor records become positive failure examples and which become negative training examples. Inconsistent handling of maintenance events across layers is one of the most common sources of training data label noise in PdM implementations, and iFactory's event integration pipeline enforces consistent event handling from Bronze through Gold. Teams that Book a Demo can review maintenance event taxonomy configuration for their CMMS export format.
Bronze Role
Store exactly as received from CMMS with arrival timestamp
Silver Role
Normalize to event taxonomy, link to asset ID, flag operating state periods
Gold Role
Label source for failure window assignment in feature vectors
-
Pipeline Observability: Data Quality Metrics Across All Three Layers
Operational HealthA medallion pipeline without observability is a black box — data quality problems that develop in Bronze ingestion (sensor communication failures, clock drift, firmware bugs causing value range changes) propagate silently through Silver and Gold until they degrade model accuracy, and by that point the problem may have been present for months. iFactory tracks pipeline quality metrics at each layer transition: Bronze ingestion completeness (expected versus received record counts per source); Silver quality rule hit rates (duplicate fraction, out-of-range fraction, timestamp correction rate) as indicators of source data health; and Gold feature validity rates (fraction of feature vectors with full valid feature coverage) as the final model-facing quality indicator. Trend monitoring of these quality metrics enables early detection of data source degradation — a Silver de-duplication rate that rises from 2% to 12% over a month indicates a sensor communication problem that is generating repeated reconnection records, not a stable source. Teams that Start Trial can configure pipeline quality metric monitoring for all data sources from the iFactory pipeline observability dashboard.
Bronze Metrics
Record completeness by source — expected vs received count
Silver Metrics
Duplicate rate, out-of-range rate, timestamp correction rate per source
Gold Metrics
Feature validity rate and label coverage fraction per asset class
-
Schema Evolution and Backfill Management for Long-Term Pipeline Maintenance
Long-Term OperabilityPredictive maintenance data pipelines operate over multi-year timescales during which sensor configurations change, CMMS versions are upgraded, new feature definitions are developed, and failure label taxonomies are refined. Medallion architecture accommodates this evolution because each layer is reprocessable from the immutable Bronze store — when a Silver quality rule is updated or a Gold feature definition changes, historical Bronze records can be reprocessed through the updated Silver and Gold logic to produce a consistently processed training dataset without relying on the original processed outputs. iFactory manages schema evolution across all three layers — versioning quality rule sets, feature definitions, and label assignment rules so that every model training run can be traced to the specific pipeline version that produced its training data. This traceability is essential for understanding why model performance changes between retraining runs and for reproducing model results during validation. Teams that Book a Demo can review iFactory's pipeline versioning and backfill management capabilities for their specific historical data volumes.
Backfill Capability
Reprocess Bronze through updated Silver and Gold rules on demand
Version Tracking
Quality rules, feature definitions, label assignments versioned per run
iFactory Record
Each training dataset linked to pipeline version that produced it
Medallion Architecture Pipeline Performance Indicators
Data Quality Improvement Across Layers
% records suitable for model consumption per layer
Raw Bronze records are suitable for model consumption at 72% — Silver quality processing raises this to 89%, and Gold feature validation brings ML-ready coverage to 97% across monitored asset populations.
Model Accuracy vs Pipeline Structure
Predictive model F1 accuracy by data pipeline approach
Full medallion architecture produces 94% model F1 accuracy versus 71% for direct-to-model pipelines — a 23-point accuracy improvement from data quality and feature engineering structure alone.
Pipeline Quality Issue Detection Rate
87% of data quality issues are detected and resolved at the Silver layer — only 13% reach the Gold layer as residual issues that require feature validity flagging rather than Silver quality rule expansion.
Label Quality Impact on False Positive Rate
Medallion (decreasing) vs manual pipeline (increasing) false positive rate
Medallion pipelines show declining false positive rates as label quality improves through consistent event integration — manual pipelines show increasing false positives as inconsistent labeling accumulates in training data over time.
Medallion Architecture Implementation: Reference Specifications
Scroll for more
| Layer | Primary Function | Key Transformations | Quality Metric | Downstream Consumer |
|---|---|---|---|---|
| Bronze | Immutable raw data preservation | None — append with arrival metadata | Ingestion completeness by source | Silver quality processing |
| Silver | Data quality and context enrichment | De-duplicate, normalize, enrich, flag | Duplicate rate, out-of-range rate | Gold feature engineering |
| Gold | ML-ready feature production | Rolling stats, freq features, labels | Feature validity rate, label coverage | Model training and inference |
| Maintenance Events | State machine and label source | Taxonomy normalization, asset linkage | Event coverage fraction per asset | Silver state flags + Gold labels |
| Pipeline Observability | Quality metric monitoring | Layer transition metric calculation | Trend in quality metrics over time | Data engineering alerting |
How iFactory Implements Medallion Architecture for Predictive Maintenance
Most predictive maintenance data quality problems are not fundamentally sensor problems or algorithm problems — they are pipeline architecture problems that could be prevented by applying the same data engineering discipline to industrial sensor data that software engineering applies to production data systems. iFactory implements medallion architecture as the structural foundation for predictive maintenance data: an immutable Bronze store that preserves every sensor record and maintenance event exactly as received; a configurable Silver quality layer that applies de-duplication, timestamp normalization, out-of-range detection, and context enrichment; and a Gold feature engineering layer that produces versioned, validity-flagged feature vectors with consistent failure event labels. When iFactory's pipeline observability metrics identify that a vibration sensor on Pump 12 has developed a 28% de-duplication rate at the Silver layer — indicating repeated reconnection artifacts — the data engineering team has a specific, actionable alert about a data source problem weeks before it would degrade model accuracy enough to be visible in prediction quality metrics. Facilities can Start Trial and begin configuring Bronze ingestion from priority sensor sources within the first iFactory deployment session.
Immutable Bronze Ingestion
iFactory preserves every raw sensor record and maintenance event exactly as received — building the append-only Bronze store that provides the audit trail for all downstream quality processing and the reprocessing foundation for pipeline evolution.
Configurable Silver Quality Rules
iFactory applies configurable de-duplication, timestamp normalization, out-of-range detection, and context enrichment rules per data source and asset class — producing the clean, contextualized Silver records that Gold feature engineering requires.
Gold Feature Engineering by Asset Class
iFactory computes asset-class-specific feature sets in the Gold layer — rolling statistical features, frequency domain indicators, and cross-sensor derived metrics — with consistent failure event labels assigned from the Silver maintenance event timeline.
Pipeline Observability and Quality Alerting
iFactory monitors Bronze-to-Silver and Silver-to-Gold quality metrics continuously — alerting data engineering teams when source data quality degrades before it affects model accuracy, and tracking pipeline health across all data sources in one view.
Implementing Medallion Architecture for PdM: Deployment Steps
01
Inventory Data Sources and Define Bronze Schema
Catalogue all sensor systems, CMMS platforms, and inspection tools that will feed the medallion pipeline — defining Bronze schema fields (source ID, arrival timestamp, raw value, raw unit) and ingestion frequency for each source before configuring Bronze layer connectors in iFactory.
02
Define Silver Quality Rules Per Source Type
Specify the de-duplication window, timestamp normalization logic, out-of-range boundaries, and context enrichment joins for each data source type in iFactory — ensuring that Silver quality rules reflect the known data quality characteristics of each sensor system and CMMS platform.
03
Build Maintenance Event Taxonomy
Define the normalized maintenance event taxonomy that Silver layer processing will map CMMS work order types to — establishing the event categories (bearing replacement, seal replacement, impeller replacement, routine PM) that Gold layer label assignment will use for failure window definition.
04
Configure Gold Feature Definitions Per Asset Class
Define rolling window lengths, statistical feature types, frequency domain features, and cross-sensor derived metrics for each asset class in iFactory's Gold layer configuration — starting with the feature sets that existing literature and engineering experience indicate are most predictive for each failure mode.
05
Set Up Pipeline Quality Metric Dashboards
Configure iFactory's pipeline observability dashboards to track Bronze ingestion completeness, Silver quality rule hit rates, and Gold feature validity rates for all data sources — establishing the baseline quality metrics that trend monitoring will compare against when evaluating data source health.
06
Validate Gold Output Against Model Performance
Train an initial model on the first Gold layer feature set and evaluate prediction accuracy against held-out test data — using the initial accuracy results to identify which Silver quality rules or Gold feature definitions require adjustment before expanding to the full asset population. Book a Demo to see the full medallion architecture configuration workflow.
Frequently Asked Questions
What is medallion architecture in the context of predictive maintenance data?
Medallion architecture organizes predictive maintenance data into three progressively refined layers — Bronze preserves raw sensor and maintenance records exactly as received, Silver applies quality cleaning and context enrichment, and Gold produces ML-ready feature vectors with failure event labels — ensuring that AI models consume consistently high-quality data rather than raw sensor streams with unresolved quality issues.
Why is the Bronze layer append-only rather than directly corrected?
The Bronze layer's immutability is what makes the pipeline trustworthy and reprocessable over time. When Silver quality rules or Gold feature definitions change, historical data can be reprocessed from the original Bronze records rather than relying on the previously processed versions — ensuring that model retraining uses consistently processed data and that quality problems can always be traced back to their source.
How does Silver layer context enrichment improve model accuracy?
Context enrichment adds operating state information, maintenance event markers, and asset-specific metadata to each sensor record — enabling the Gold layer to compute operating-condition-normalized features and assign correct failure labels relative to known maintenance events. Without this context, models confuse startup ramp-up vibration with fault signatures and cannot correctly identify pre-failure windows in the training data.
What is the Cumulocity connection to medallion architecture for IIoT maintenance?
Cumulocity's IIoT platform research, referenced at IIoT World 2025, validated the medallion pattern for industrial IoT data pipelines — demonstrating that the Bronze-Silver-Gold layer structure reduces data quality-related model accuracy degradation by separating raw ingestion from quality processing and feature engineering in streaming industrial sensor contexts.
How long does it take to implement a full three-layer medallion pipeline in iFactory?
A functional Bronze-to-Gold pipeline for a priority asset class with two or three sensor sources and CMMS integration can typically be configured in two to four weeks. Full production deployment covering all monitored asset classes with complete Silver quality rules and Gold feature definitions typically requires eight to sixteen weeks depending on the number of source systems and the complexity of the maintenance event taxonomy.
Build the Data Pipeline Foundation That Makes Predictive Maintenance AI Reliable
iFactory gives maintenance data engineers the Bronze ingestion infrastructure, Silver quality processing, Gold feature engineering, and pipeline observability needed to produce the consistent, contextualized, and properly labeled training data that predictive maintenance models require for accurate, sustained performance.
