Our own product · live demo

Boljam — predictive maintenance for refining and oil production

Boljam is a predictive maintenance system built by 105 Industrial AI. It brings vibration diagnostics, telemetry and maintenance history into one loop: five machine-learning models classify defects, estimate remaining useful life and detect anomalies before RMS velocity crosses the ISO threshold. On a demo fleet of 450 machines, in 231 episodes on 183 units the detector fired before RMS velocity reached zone C, with a median lead time of 31 days. It opens in a browser and runs without a server round-trip.

days of median lead time over the RMS threshold231 episodes on 183 machines, synthetic data
31
pieces of equipment under monitoring1,302 measurement points, 147 MW
450
ROC-AUC of the anomaly detectorfalse-alarm rate 0.0012 at threshold 0.9
0.971
months of history — scrub with the cursorsignal recomputed in the browser
18

A bearing 85% destroyed — instruments still read “as new”

A pump at 1476 rpm with an 85% outer-race defect: RMS velocity of 1.86 mm/s — zone A under ISO 20816-3. Formally the machine is healthy. Meanwhile the envelope at the defect frequency has grown 95-fold and kurtosis has gone from 2.96 to 10.87. That is exactly what Boljam surfaces.

Demonstration on synthetic data: the figures come from a physics model, not from measurements on real equipment. Asset names are taken from public sources; the system is not the reporting of any company and uses no corporate branding.

Two assets in one loop

Refinery

refining, 5.5 mln t/year, three production units

  • 16 process units — from a 1969 crude distillation train to a 2018 deep-conversion complex
  • 267 machines, 709 measurement points, 89.0 MW
  • Pumps, compressors, flue-gas fans, air coolers, furnaces

Mature oil field

production, four field units, waterflood system, oil treatment

  • 183 machines, 593 measurement points, 58.4 MW
  • 3,698 wells: 92.5% rod pumps, 7.5% ESP; 121 under vibration monitoring
  • Rod-pump diagnostics from dynamometer cards — 8 condition classes

How it works

Three engineering decisions that make the demo fast, honest and server-free.

01

The signal is synthesised, not stored

No raw waveforms are stored anywhere. A physics model compiled from Rust to WebAssembly generates them deterministically in the browser from a key of tag + point + time. The same code drives the offline dataset generator, so stored features always agree with the signal you see when you drag the cursor to any date.

02

Consistency verified end to end

The median discrepancy in RMS between the stored dataset and the browser recomputation is 0.0000%, with a 95th percentile of 0.17%. Two generator runs produce bit-identical files. The DSP core is covered by 62 tests, the domain logic by 184 unit tests plus end-to-end scenarios.

03

Inference in the browser, data stays put

Training is offline and one-off; inference runs client-side through ONNX Runtime Web, with all models totalling 1.87 MB. The same architecture transfers to an enterprise perimeter: computation happens where the data sits and nothing leaves the site.

Five models and their measured quality

Train/test splits are by machine identity rather than by window, so there is no target leakage: the ISO zone is explicitly excluded from the inputs.

ModelTaskMeasured quality
FaultNetDefect classification, 12 classesAccuracy 0.92, F1 0.92, ROC-AUC 0.994
RULNetRemaining useful life P10/P50/P90MAE 44.7 days, P10–P90 coverage 0.93 against a nominal 0.80
HealthIndexAnomaly detectorROC-AUC 0.971, false alarms 0.0012
DynaCardRod-pump condition from dynamometer cards, 8 classesAccuracy 0.988 on the demo cards
ReliabilityReliability, optimal maintenance intervalWeibull with censoring, checked against Kaplan–Meier

What the system cannot do — stated plainly

FaultNet reliably detects bearing defects, looseness and gear mesh; it separates imbalance from misalignment using a second input built on the first running harmonics, but recall for imbalance is only 0.65 so far. At 746 of 1,302 points — on two- and four-pole motors — an electromagnetic fault cannot be told apart from misalignment. All models are trained and tested on synthetic data from the same physics model; validation on public recordings has not been completed, and the interface says so directly. Loaders for public datasets (CWRU, NASA IMS, MFPT, Paderborn, XJTU-SY, Petrobras 3W, MetroPT-3, C-MAPSS) are included — validation on real data is performed during the customer pilot.

Standards base

ISO 20816-3 — zones A/B/C/D by RMS velocity, with the band following shaft speed: from 2 Hz below 600 rpm, otherwise from 10 Hz. ISO 20816-8 and ISO 10816-6 cover reciprocating machines, where limits are 28–45 mm/s rather than 4.5–11. API 670 covers turbomachinery on sleeve bearings, where the governing criterion is relative shaft displacement rather than casing RMS. Also: ISO 13373, ISO 17359, ISO 13379, ISO 13381, API 610.

What it gives an enterprise

For the chief mechanic and reliability team

A ranked risk list across the whole fleet instead of a calendar round: you see which machines need intervention in the next 7 and 30 days and why — with the defect frequency and its development over time.

For planning and economics

The economic effect is computed on screen rather than in a slide deck: all 25 assumptions are displayed with a rationale for each, so the calculation can be challenged item by item and rebuilt with your own downtime and repair costs.

For IT and information security

No database, no outbound calls: computation happens in the browser. In an industrial setting the same architecture is deployed on-premise — telemetry and video never leave the plant perimeter.

The same approach — delivered as a service on your equipment

Predictive maintenance service

FAQ

Frequently asked questions

Short, straight answers to the questions we hear most often.

What is Boljam?

Boljam is a predictive maintenance system by 105 Industrial AI (“105kz” LLP, Kazakhstan). It combines vibration diagnostics, telemetry and maintenance history: five ML models classify defects from the envelope spectrum, estimate remaining useful life as P10/P50/P90 intervals and detect anomalies before RMS velocity crosses the ISO 20816-3 threshold. The demo runs in a browser on synthetic data.

Why is predictive diagnostics more accurate than an RMS velocity threshold?

An RMS threshold responds to total vibration energy and therefore reacts late. Measured in Boljam: a bearing with an 85% outer-race defect reads 1.86 mm/s RMS — formally zone A, “as new”. Meanwhile the envelope at the defect frequency has grown 95-fold and kurtosis has risen from 2.96 to 10.87. Across a demo fleet of 450 machines in 231 episodes on 183 machines the anomaly detector fired before RMS reached zone C, with a median lead time of 31 days.

What data does the demo run on?

All data is synthetic: the figures come from a Rust → WebAssembly physics model rather than measurements on real equipment. Asset names are taken from public sources to keep the demo recognisable; the system is not the reporting of any company and uses no corporate branding. In a customer pilot the models are retrained on the plant's real telemetry, vibration data and maintenance histories.

Can such a system be deployed inside an enterprise perimeter?

Yes, that is the default scenario. Training is offline and one-off; inference runs client-side through ONNX Runtime Web with all models totalling 1.87 MB. The same architecture is deployed on-premise on customer servers: computation happens where the data sits and nothing is transmitted outside.

Which standards does the vibration assessment follow?

ISO 20816-3 for A/B/C/D zones by RMS velocity, with the band following shaft speed. ISO 20816-8 and ISO 10816-6 for reciprocating machines, where limits are 28–45 mm/s rather than 4.5–11. API 670 for turbomachinery on sleeve bearings, where the governing criterion is relative shaft displacement rather than casing RMS. Additionally ISO 13373, ISO 17359, ISO 13379, ISO 13381 and API 610.

Which defects can the system not distinguish?

FaultNet separates imbalance from misalignment using a second input built on the first running harmonics, but imbalance recall is still 0.65; at 746 of 1,302 points an electromagnetic fault cannot be told apart from misalignment. All metrics come from synthetic data of the same physics model. The limitations are stated directly in the system interface.

How long does a pilot on real equipment take?

8–12 weeks, from data export to a working model with a measured effect. A pilot usually covers a group of similar critical machines — pumps, compressors, flue-gas fans — and compares the forecast against actual failures and work orders over a past period.

Next step

See Boljam applied to your own fleet

Tell us about your task — we'll come back with a data audit plan and an effect estimate within two business days.

  1. 1A 30-minute call: the task, your current systems, who makes the decision
  2. 2A review of data and processes, a staged plan with time and cost ranges
  3. 3A 2–4 week prototype or an 8–12 week pilot with success criteria agreed upfront

NDA from the first contact

Reply within one business day