Industrial predictive analytics
Predictive analytics built around an operating decision
Use historical and current evidence to estimate a future condition or prioritize an exception for review. NeoBram starts with the action a person can take, the cost of an error and a baseline the team can test.
Choose a decision before a model
Define who uses the result, the required lead time and the consequence of missed or false signals. A forecast without a response owner may add noise rather than value. Document the current rule, planning method or expert judgment as a baseline. Model complexity should be justified by better evidence for that decision, not by a more impressive demonstration.
Reconstruct the evidence available at prediction time
Potential inputs include historian signals, operating states, maintenance events, orders, alarms and project records. Match timestamps, units and asset or product identifiers. Distinguish measured values, missing data and corrected records. Evaluation must not use later information that would have been unavailable when a real prediction was needed.
- Represent relevant operating regimes, seasonal conditions and process changes, including difficult or infrequent cases.
- Review outcome labels with domain experts and identify periods where the truth is unknown.
- Document the permitted source boundary and how delayed, duplicate or out-of-range inputs are handled.
Make anomaly and risk priorities reviewable
Anomaly detection identifies unusual patterns; it does not establish a fault or its cause. A risk score should help a responsible person decide what to investigate, with supporting evidence and uncertainty visible. For alarm-related use, retain the existing alarm philosophy and escalation process. Prioritization must not silently suppress protective alarms or authorize operating changes.
Connect the estimate to a practical response
The interface should show the forecast horizon, baseline, relevant context and allowed next action. Reviewers need to challenge, defer or override a result and record their reasoning. Maintenance and planning workflows require different outputs. A predicted event is useful only if the evidence, lead time and response capacity support a meaningful decision.
Evaluate time, error costs and human effort
Use held-out time periods or assets and compare with a simple statistical or rule-based baseline. Report missed events, false alerts, forecast bias, calibration and useful lead time as appropriate. Break results down by operating regime and decision horizon. Test missing or shifted signals, unfamiliar conditions and the review workload; one overall accuracy score is not enough.
Agree the model and operating handover
Possible outputs include an evaluated prototype, baseline comparison, source-linked exception view and limitation record. Handover defines feature and model versions, thresholds, monitoring, support, rollback and the owner of each response. Changes to sensors, products or processes may invalidate earlier evidence. Retraining and release decisions require reviewed tests, rather than assuming that more recent data always improves the model.
An illustrative alarm-priority review
Several alarms occur near a changing operating state. The workflow assembles the timeline, related measurements and missing-data flags for an engineer. The engineer checks whether this is an expected transition or a condition needing investigation, following the established response procedure. This illustrative example does not prove fault diagnosis, earlier detection or reduced downtime.
Bring one decision and its failure conditions
Start with a recurring decision, representative records and the expert who owns the outcome. Agree the baseline, useful response window and conditions under which the model should decline to make a recommendation. If evidence is insufficient, define a data-readiness step or use a simpler method. Automated control needs a separate engineering and safety assurance decision.
More clarity
Questions and answers
Is anomaly detection the same as predictive maintenance?
No. An anomaly is an unusual pattern. Predictive maintenance also requires asset context, useful intervention time, an approved response workflow and evidence that the alerts support maintenance decisions.
Which accuracy metric should we use?
Choose metrics from the decision and error costs. Forecast error, bias, event coverage, calibration, lead time and reviewer workload may each matter; a universal percentage is insufficient.
How much history is required?
Enough to represent the relevant outcomes, regimes and changes, with trustworthy context. There is no universal duration; assess event coverage and source quality before committing to a model.
Can a model identify the root cause?
A model may surface associations or candidate explanations, but correlation does not establish cause. Domain experts determine what additional investigation is needed.
Can it automatically change a process setting?
That is a separate control and operational-assurance question. The initial boundary is commonly an estimate or recommendation reviewed by an authorized person, with established procedures retained.
Start with one business problem