Research Council of Norway approves Bluestream's remaining-useful-life R&D project under SkatteFUNN

SkatteFUNN approval is not automatic. A case officer at Forskningsrådet assesses every application against three criteria and the project period, and the decision letter records a verdict on each:

The letter then gives two categories for each work package: the one the applicant chose and the one the case officer assigned. One of the categories open to the case officer is “not R&D, ordinary product development”. All six packages were submitted as experimental development, the case officer's column reads “not changed” for each, and all six were approved.

The scientific problem at the centre of the project is the cold start: how to give a calibrated estimate of remaining useful life (RUL), one whose stated 80% interval contains the actual failure about 80 times in 100, for one specific machine when there is little or no failure history for that machine. Data-driven platforms answer it with a long learning period per asset and a black-box output. Traditional reliability tools hold the standardised data, but use it offline and do not connect it to live sensors.

The hypothesis is that the two can be joined, as drawn at the top of this page. Standardised reliability data (OREDA failure data, organised by the ISO 14224 taxonomy) serves as the population prior: what is known about machines of this class before this one has been observed. Live data from the plant's own heterogeneous, non-proprietary sensors then updates that prior into a posterior for the unit: an RUL with an uncertainty band, per maintainable item, with the failure mode named.

None of this is established. It is not known whether the combined estimate stays statistically calibrated when failures are rare and the data right-censored, or whether degradation seen in shared signals can be assigned to the right maintainable item. The application states that the outcome is uncertain at the start, which is what makes it research and development.

Each work package takes one open question and states how the answer will be judged. Five name a pass criterion; WP5 delivers a measured trade-off curve instead. The application fixes the pass marks, not the number of cases behind them.

1. WP1

   Prior and evidence fusion

   Jan 2026 to Aug 2027 · 20 months

   A Bayesian model that starts from the OREDA prior and updates it with condition data into an RUL posterior with uncertainty bands. Calibration is tested on right-censored runs: observed failures should fall inside the predicted interval as often as the stated confidence level says.

   TargetBeats 2 baselinesthe prior alone and a naive threshold alarm, with calibrated intervals, on at least 2 equipment classes
2. WP2

   Attribution

   May 2026 to Aug 2028 · 28 months

   A motor and its pump share vibration, temperature and current signals. The method decomposes the shared signals and assigns degradation to the right maintainable item and ISO 14224 failure mode, with seeded faults as the answer key.

   Target≥ 80%of seeded, known-answer cases attributed to the correct item and failure mode, on at least 2 unit types
3. WP3

   Comparable health indicators

   May 2026 to Oct 2027 · 18 months

   Low-cost sensors differ in type, sampling rate, placement and protocol (OPC UA, Modbus, MQTT, I²C), so their raw signals are not comparable. The work is feature extraction that yields comparable health indicators regardless of setup.

   TargetWithin 10%agreement between two different sensor setups on the same unit
4. WP4

   Faithful explainability

   Apr 2027 to Dec 2028 · 21 months

   Explanations derived from the model's actual inference instead of rationalised afterwards, expressed in ISO 14224 vocabulary and as a position on the P–F curve. The application calls this unsolved in the explainability literature.

   Target≥ 85%of a reviewed set of cases accepted as technically correct by one or more reliability engineers
5. WP5

   Inference at the edge

   Oct 2027 to Dec 2028 · 15 months

   Accuracy, latency and energy measured on constrained edge hardware against an unconstrained reference, under intermittent connectivity.

   OutputA trade-off curveNo pass mark: the trade-off is not known in advance, so the deliverable is a quantified curve showing where fidelity holds
6. WP6

   Field validation

   Apr 2027 to Dec 2028 · 21 months

   Experiments with induced and accelerated degradation to create ground truth, then the whole method on live data from real industrial equipment, scored with statistics for right-censored data.

   Target≥ 2 siteswith a map of where the method fails as well as where it succeeds

The method is tested on three equipment classes chosen to differ in failure physics and in sensor type, so a result on one cannot be assumed for the others:

WP1 started in January 2026: a first prior-to-posterior RUL estimator is wired into the platform, and a separate calibration harness scores it against simulated run-to-failure data with known answers. WP2 and WP3 followed in May 2026. For WP3, a Raspberry Pi edge gateway now carries data from a first low-cost sensor to the platform.

WP4 and WP6 start in April 2027, WP5 in October 2027. Field validation is planned with U.S. company Optec, LLC under a non-binding Letter of Intent.

SkatteFUNN is a rights-based scheme: once a project is approved, the company is entitled to claim a tax credit of 19% of the eligible, documented R&D cost in the project, within a ceiling of NOK 25 million in project cost per company per income year.

The approved application budgets NOK 3,150,000 of R&D cost over the three years, all of it personnel and indirect cost. At 19% the estimated credit is NOK 598,500:

For an early-stage company the question is what happens when there is little or no tax to set the credit against. The Tax Act settles it: if the credit exceeds the company's assessed tax, the excess is paid out in the tax settlement (skatteloven § 16-40, sixth paragraph). Skatteetaten, the Norwegian Tax Administration, makes that payment. For a company with no tax payable, the whole credit therefore arrives as cash, in the year after the cost was incurred.

Bluestream is building the PdM platform without outside equity. A 19% credit on eligible cost lowers the net cost of three years of R&D by up to 19%, and by less wherever real cost exceeds the capped hourly rate. It can be combined, within the state-aid ceiling, with the other non-dilutive funding the company is pursuing in Norway and the EU.

What the approval is, and is not

It is a reimbursement, not an advance.

Bluestream carries the R&D cost first. The credit is claimed on each year's corporate tax return and settled the following year. Because the approved period starts on January 1st, 2026, eligible cost from that date can go into the 2026 claim, provided the hours were logged as they were worked.

Combined public support is capped.

SkatteFUNN can be combined with other public support for the same project, but state-aid rules set a maximum for the total. If the sum exceeds it, the SkatteFUNN credit is reduced in the tax settlement.

It assumes a company not in financial difficulty.

The approval presupposes that Bluestream AS was not an undertaking in difficulty, in the state-aid sense, on the decision date.

It is tied to this plan and this calendar.

The project cannot be extended past December 31st, 2028. If its goals or work packages change materially, it has to be closed with a final report and applied for again.

Only documented cost counts.

The whole budget is personnel and indirect cost, so the claim is built from timesheets: hours per person, per day, per work package. The hourly rate is set by regulation at 1.2 per mille (0.12%) of annual salary, with a maximum of NOK 700 per hour, and at most 1,850 hours per person per year count. An auditor has to attest the cost. Skatteetaten decides how the basis is calculated and can audit the claim.

It comes with reporting.

A status report by March 1st each year, a final report no later than March 1st, 2029, and a follow-up report no later than two years after the project ends. If a report is missed, Forskningsrådet notifies Skatteetaten and can withdraw the approval (SkatteFUNN regulation § 16-40-10).

The credit is the part with a number on it. The other part is what the project commits Bluestream to. The research questions sit under the claims the platform makes to customers: that an RUL interval can be calibrated rather than guessed, that a prediction can be traced to a named failure mode on a named maintainable item, and that none of it depends on one sensor vendor. The platform has a first version of each today: an RUL estimate with a P10 to P90 band (P10 is the earliest edge, the reverse of the reserves-reporting convention), a failure-mode label on the alert and the work order, and ingest over open protocols. Whether they hold on real equipment, and how well, is not established.

Forskningsrådet's approval is narrower than an endorsement. It means a case officer accepted three things: that the project aims at something new, that it has to produce new knowledge to get there, and that it is targeted and delimited. It says nothing about whether the method will work, and it is not a review of the product.

The tests written into each work package are how Bluestream, and its customers, will find out.