Predictive equipment maintenance: where service life is lost
Scheduled maintenance is built on the calendar, not on the actual condition of the component.
A bearing is replaced at a fixed interval — and either a part with remaining service life is thrown away, or, worse, accelerated degradation is missed: incorrect lubrication, misalignment, rising vibration.
Equipment failure arrives off-schedule, and unplanned downtime halts the entire processing chain — from raw material loading to shipment. The cost of such downtime is calculated not in repair hours, but in ruined batches and contractual penalties.
The gap arises where sensors are already installed but there is no decision-making based on them. Vibration, temperature, drive current, hydraulic pressure are written to the archive, but they are analyzed after the fact — when the component has already failed.
The dispatcher sees a trend, does not understand the threshold, waits for confirmation from the mechanic, and the mechanic waits for a free window in the schedule. Between the signal and the action, a day is lost, during which component wear turns into destruction.
A mobile application closes exactly this gap: it turns the telemetry stream into a concrete action on a specific asset. Condition monitoring runs on the inspector's phone and in the chief mechanic's office — tied to a piece of equipment, not to an abstract schedule.
The engineer receives a notification while the deviation is still in the controlled-risk zone, sees the component's history, and assigns work before the emergency protection trips.
Next, we break down how this works: what data is collected, how thresholds and alerts are configured, what each role sees, and how the application fits into the existing maintenance loop.
What a mobile application for predictive maintenance delivers
A mobile application for predictive maintenance turns equipment telemetry into a decision that an engineer makes on site rather than in the office.
Sensors already output vibration, temperature, and current — the question is whether a deviation reaches a person within minutes or surfaces during an incident review.
The application closes this loop: it receives a stream from SCADA and historians, compares it against the asset's baseline, and shows not a "raw" chart but a specific asset with an out-of-tolerance condition.
The first measurable effect is the reaction time to a deviation.
Previously, the signal lived in the control room until a scheduled walkdown; now a push with an urgency level reaches the duty officer and the area supervisor: from the appearance of a trend to an inspection, hours pass rather than shifts.
The second is the share of sudden failures. A model on a sliding window catches bearing degradation or a rise in winding temperature before protection trips, and moves the event from the "shutdown" class to the "planned intervention" class.
The third effect is the overhaul interval. When an asset's condition is confirmed by data, scheduled replacement is shifted based on actual wear rather than the calendar, and the resource between services grows without loss of reliability.
The flip side is trust in the forecast: the application shows which features the warning was built on and how confidence changed, so the engineer makes a decision consciously rather than "on orders from above."
We separately address offline mode and access separation: a field inspector enters measurements without a network, they synchronize when a channel appears, and the supervisor and dispatcher roles see different views.
This is how predictive maintenance becomes a working shift practice rather than a reporting dashboard.
Operating Modes: Data Flow, Thresholds, and Manual Review
The application does not dictate a single data collection scenario — it supports three modes, and the choice between them is determined by the nature of the equipment and the quality of the channel at the site. Below is how each one works and where it starts to get in the way.
Continuous Stream
Telemetry comes from sensors over a constant channel, the application maintains the connection and processes readings as they arrive.
This is the only mode where predictive maintenance catches a developing defect — rising vibration, temperature drift — before it crosses the boundary.
Limitation: it requires a stable connection and constant power at the gateway, and if the channel breaks, gaps in the observation window are possible.
Threshold Rules
The application stores a set of conditions for each node and triggers when a value goes beyond a boundary or stays within it longer than a specified time. The monitoring mode does not require a wide channel — short event messages are enough, so it works even on a weak signal.
Limitation: the rule sees the fact of an exceedance, but not the rate of approach to it, so early signs of wear remain unnoticed.
Periodic Upload
Data accumulates on the device and is sent in a batch on a schedule — once per shift or when the network becomes available. It is suitable for remote sites and areas without a constant channel.
Limitation: between uploads the application is "blind," and a decision on a node is made with a delay of one interval.
| Mode | When it is suitable | Limitations |
|---|---|---|
| Continuous stream | Critical nodes, stable channel and power | Sensitive to connection breaks |
| Threshold rules | Weak signal, only the fact of an event is needed | Does not see the trend before an exceedance |
| Periodic upload | Remote sites, channel on a schedule | Delay of one interval |
Modes are combined at a single site: critical nodes go as a stream, the rest — by thresholds, remote areas — by upload. Switching between them does not require reinstalling the application — only the collection configuration changes.
How we implement the application: from fleet audit to handover
Implementation begins not with code, but with an inventory: which units actually produce a signal, which data is already being collected, and which will have to be captured from scratch.
At this stage we record the list of equipment, the available telemetry channels, and the points where sensors are missing — the entire subsequent data collection architecture depends on this.
Next, we build the data collection scheme and configure the models for a specific fleet, not for an abstract set. We validate everything on historical data before changing anything on a live line.
- Fleet audit. Equipment inventory: types of units, operating modes, critical failures, which data already exists in the customer's systems and which is missing.
- Data collection scheme. We determine where telemetry is captured from and at what frequency, how it is transmitted and stored, and how gaps and outliers are filtered out.
- Model configuration. Models are trained on the real operating modes of a specific fleet — warning thresholds are selected for each type of unit rather than taken as averaged values.
- Validation on historical data. We run the models on past failures: we look at how far in advance of a breakdown the system raised a signal and how many false alarms it produced.
- Pilot on the line. We launch on a limited section, compare predictions with the actual state of the units, and adjust thresholds and notification logic based on the results.
- Solution handover. We deploy across the entire fleet, train the operations service, and hand over the documentation and the regulations for working with the system's signals.
Each step ends with a verifiable result: a list of equipment, a data flow scheme, trained models, a report on historical failures, pilot results, a working system with regulations.
The transition to the next stage happens only after the previous one has been confirmed on real data.
What the customer receives upon completion of implementation
Implementing predictive maintenance is not just a working service, but also a set of artifacts that your team can use to operate the system without us.
We hand everything over to the customer's repository: configurations, schemas, policies, and procedures — not a "black box" dependent on the contractor.
- Service and model configurations. Deployment files, trigger threshold parameters for each type of equipment, retraining settings on new data — with a description of what each parameter affects.
- Integration schemas with sensors and process control systems. Telemetry collection topology, protocols and polling frequencies, buffering during communication loss, failure points and the procedure for working around them.
- Access policies and roles. Who sees raw data, who sees forecasts, who can change thresholds; segregation by site and department, action logging.
- Escalation procedures. What happens during an anomaly: who receives a notification, after how many minutes, at what criticality level the on-duty shift is engaged, how feedback on false alarms is recorded.
- Documentation and training materials. Architectural description, instructions for dispatchers and service engineers, analysis of typical incidents, recordings of training sessions.
- Model quality metrics and dashboards. Share of correct predictions, the time-to-failure that can be predicted, false alarm dynamics — so you can see whether the system works rather than taking our word for it.
- Support and handover plan. What we support ourselves, what transfers to your team, the procedure for retraining the model when new equipment is added.
We agree on each artifact during the design phase rather than delivering it at the end — this way your engineers have time to ask questions and accept the system into operation.
Describe your equipment fleet and current data collection loop — we will get back to you with an assessment and a handover plan.
Case study: how unplanned production downtime was reduced
What it was like before implementation
A continuous production line: three sections, drives and pump units, maintenance according to schedule. The schedule was rarely changed, so some components were serviced before their resource was used up, while others — only after the first signs of wear appeared.
Unplanned stoppages occurred 5-7 times a month, each one — from 40 minutes to several hours of downtime and an urgent reassignment of a crew.
Diagnostics relied on rounds: an inspector took readings once per shift, and a deviation was noticed only when the component was already overheating or making noise.
Our customer — a production site with a continuous cycle. The task was formulated not as "install sensors" but as "manage to react before failure."
What was changed
Vibration sensors and temperature sensors were installed on critical components, and the data went to an app on the tablets of inspectors and supervisors.
The model calculates deviation from the component's individual norm, not from a general threshold: each unit has its own baseline, so early signs of wear are visible against the background of normal variation.
The trigger threshold was tuned iteratively — in the first weeks the system gave too many reasons for checks. Confirmation by a second metric and a delay for trend confirmation were added: a signal is sent only if the deviation persists and grows.
The number of false triggers was reduced by about threefold, while not a single real failure was missed.
Metrics before and after
| Indicator | Before | After |
|---|---|---|
| Unplanned stoppages per month | 5-7 | 1-2 |
| Time from signal to decision | hours (via rounds) | minutes (push to supervisor) |
| False triggers | — | reduced threefold |
| Share of failures caught in advance | about one third | more than 80% |
A separate effect — predictability of repairs: the supervisor sees which components require attention this week and plans the crew in advance, rather than after a stoppage occurs.
Spare parts storage costs stopped growing — purchases are made for a specific forecast, not "just in case."
Let's discuss your line — we'll come back with an assessment and a pilot plan for one section.
How predictive maintenance differs from planned preventive maintenance
Planned preventive maintenance is built around a calendar schedule: the regulation is tied to operating hours or a due date, and a component is opened up, cleaned, or replaced regardless of how it is actually behaving. The logic is simple — build in a margin so that a failure does not catch you off guard.
The problem is that service life is consumed unevenly: some parts are replaced long before they are worn out, while others fail between two scheduled windows, because real conditions — load, temperature, raw material quality — are not accounted for in the schedule.
Condition-based maintenance flips the logic: the decision to intervene is made not by the calendar but by the actual condition of the component.
Vibration, temperature, current, and pressure sensors provide a continuous series, the model compares it against the baseline of a specific unit, and shows the deviation before it becomes a failure.
The regulation remains, but becomes event-driven: work is scheduled when the component approaches a threshold, not when a date arrives.
The difference shows up in two quantities. The first is the share of premature replacements: under a calendar approach it is high, under condition-based maintenance it drops, because service life is fully used.
The second is the share of failures between scheduled windows: a schedule does not catch them by definition, whereas a predictive model sees a developing defect days and weeks before failure and moves it into a scheduled window.
The planned preventive approach does not become outdated, however.
For components with predictable wear, inexpensive consumables, and equipment without sensors, the calendar remains the most reliable and cheapest option — where the cost of failure is lower than the cost of monitoring.
Condition-based maintenance is justified on critical assets, where downtime is expensive and failure develops gradually and leaves physical precursors.
In practice, a hybrid wins: the calendar regulation covers what it does well, while the predictive model takes on critical components and adjusts timelines based on actual condition.
In this scheme, the application is not a replacement for the regulation, but a tool that shows which work can be shifted and which cannot be postponed.
Common technical doubts before starting
Before starting, almost every team asks the same set of questions: whether the sensors already installed will work, what will happen if communication is lost on the shop floor, how much the forecast can be trusted, and where the data from the equipment goes. Below are brief answers to each, to remove uncertainty before design.
We already have sensors from different manufacturers — will we have to replace everything?
You do not need to replace the fleet. We collect data through a gateway that reads common industrial protocols and can pull readings from existing controllers and vibration sensors.
Where the protocol is closed, we install an intermediate module that takes the signal from the analog output. Hardware replacement is needed only if the sensor physically does not provide the required value — for example, it does not measure bearing temperature.
The internet on the shop floor is unstable. Will the application work?
Yes. The logic for collection and primary processing lives on the device: readings are written to local storage, and the buffer survives a restart and a communication outage.
When the channel returns, the accumulated data goes to the server in packets, without losses or duplicates. Critical deviation alerts are generated on site, without waiting for the network.
How accurate is the forecast and what happens in case of a false trigger?
The model is trained on the history of specific equipment, not on averaged data — accuracy grows as observations accumulate. The trigger threshold is configured for each unit, so the balance between a missed event and a false alarm is set by the operations service itself.
Each warning is accompanied by the contribution of features: you can see which value deviated and by how much.
How is production data protected?
Transmission goes over a secure channel, access is role-based, and actions in the system are logged. Data can be kept on your own infrastructure if policy prohibits external storage. Anonymized telemetry series are separated from employees' personal data.
How will the application integrate with our accounting systems?
Integration goes through a software interface: repair requests go to the existing maintenance system, equipment directories are synchronized, and statuses are returned back. We agree on the exchange format at the start so that accounting does not have to be restructured.
Describe your equipment fleet and current infrastructure — we will get back to you with an assessment and a pilot plan.
Let's discuss your task and come back with a plan
To come back with an estimate and an implementation plan rather than generalities, we need a minimum of input data about your fleet.
The more specific the initial picture, the more accurate the estimate of timelines and scope of work — and the fewer revisions at the start of the project.
What helps in the first conversation:
- A list of equipment and types of units for which a forecast is needed: pumps, compressors, drives, bearing units — indicating which data on them is already being collected.
- Available telemetry sources: vibration, temperature, pressure, current, and operating-hours sensors; how often readings are taken and where they are stored now.
- History of failures and repairs over the past period: what broke, when, and how it ended — this is the basis for training the predictive maintenance model.
- Network and infrastructure at the sites: connection stability, availability of gateways, restrictions on data transmission from the sites.
- What result is considered a success: reducing unplanned downtime, transitioning to condition-based maintenance, reducing emergency callouts.
- Who on your side is responsible for data and operations — an engineer, a process specialist, the IT service — and how access is coordinated.
- Existing systems: SCADA, accounting platforms, the repair service — so that the application integrates into current processes rather than living separately.
The first conversation takes about an hour: we break down the task, clarify which signals are actually available, and define the boundaries of the first version.
After that, we come back with an implementation plan — stages, scope of work, and what will be required from your team at each step. Describe the task and send what you have according to the list above — then we move forward concretely.