Predictive maintenance analytics
NDDev builds predictive maintenance analytics: models that read sensor and repair history, show early signs of failure and suggest when to plan maintenance. Our RnD practice starts with simple rules and statistics and moves to machine learning when the data supports it. Our team has 100+ projects behind it, and NDDev has worked since 2019.
Who it is for
Manufacturers and service teams that have an equipment history and a repair log, and for whom unplanned downtime is expensive.
- Forecasts from your failure history
- Methods from rules to machine learning
- Accuracy in figures after measurement
- Prompts for the maintenance team
What a forecast rests on
A forecast starts with data: equipment telemetry, the repair and failure log, operating modes. First we check whether failures show up in the data, and how long before a breakdown they appear. Answering that question is the research phase, and you get the answer before the main contract.
- a review of failure history and telemetry
- a check of whether failures are visible in the data
- baseline rules and anomaly detection
- a forecasting model with measured accuracy
- prompts and alarms for the maintenance team
- model monitoring and retraining on new data
Methods go from simple to complex: thresholds and rules, statistics and anomaly detection, then models in scikit-learn, XGBoost, LightGBM or CatBoost. A model that has to run next to the equipment is converted to ONNX. Accuracy depends on the data, so we guarantee the process, the iterations and the documentation, and give accuracy figures after measurement.
From repair log to forecast
Data audit
We collect telemetry and the repair log, align them in time and see how many failures the history holds.
Baseline
We build simple rules and measure how many failures they catch. Everything we do next is compared with that number.
Models
We try machine learning models and keep the one that beats the baseline on held-out data.
Rollout
We connect the forecast to the data stream and send prompts to where the maintenance team will see them.
Monitoring
We track forecast quality and retrain the model when the equipment or the operating mode changes.
Questions about failure forecasting
How much data does a forecast need?
It depends on how often failures happen and how detailed the telemetry is. We answer this in the research phase, on your data, and tell you plainly if there is too little of it so far.
What if the history has almost no failures?
Then we start with anomaly detection: the model learns normal equipment behaviour and flags deviations from it. Along the way, this builds the labelled history a future forecast needs.
Do you guarantee forecast accuracy?
We guarantee an agreed process, iterations, measurements and documentation. We state accuracy after testing on held-out data the model never saw in training.
Will it work if we have few sensors?
Sometimes the data you already have is enough: the repair log, running hours, operating modes from your business system. If it falls short, we advise which measurements to add and write the software to collect them.
How much does predictive analytics cost?
The cost comes from the volume of data, the number of equipment types and whether the forecast has to run in real time. The starting price is shown in the price block; we fix the total after the data audit.
from $8,900
- Free warranty
- 80 days
Prices are starting prices for information only and are not a public offer. The final price is fixed in the specification.
Software for equipment and industry
Related
Show us your repair log and equipment data
Describe the equipment and the breakdowns that cost you most. We will assess whether they are visible in your data.
We reply within 24 hours