$140,000 a year lost across five losses on one premix skid
The biggest was dosing: errors on the boric acid premix skid took 65 minutes to troubleshoot, so off-ratio material reached the FP tank and the whole batch was reblended or scrapped, at $80,000 a year. Four mechanical and electrical losses in and around the skid added another $60,000.
| Loss | Frequency | Value lost |
|---|---|---|
| Dosing, concentration and deviation alarms | ~10 a month | $80,000 a year |
| Pig stuck at changeover | ~2 a month | $10,000 a year |
| Valve failure | ~15 a month | $20,000 a year |
| Communication and network errors | ~5 a month | $5,000 a year |
| Cable damage | ~1 a month | $25,000 a year |
| Total | $140,000 a year |
- ~$75,000 a year in potential savings on this skid with Nova
- 10 min instead of 65 to root cause, catching off-ratio product before it reaches the FP tank
- $250,000 a year target across the full making process
Batch-made base, continuously dosed product
Liquid products are made in two stages. First the white base, the backbone of the product, is built batch by batch in the main mix tank (MMT). It then moves into continuous processing through the inline mixer and dosing system (ILM and LPD), where premixes, perfumes, enzymes and additional surfactants are dosed into the moving stream, each on its own flow setpoint. Finished product goes to the finished-product (FP) tanks and on to the packing lines. Off-spec product is reblended back into the ILM alongside the other dosed streams, or scrapped.
Each premix is prepared on its own skid. On the boric acid skid, powder from a hopper is dissolved in DI water in a 2.5 m³ premix tank, transferred to a buffer tank, then pumped to the ILM through a mass flowmeter and an on/off dosing valve. Because dosing is continuous, a dosing error isn't caught at the end of a batch. Every minute the ratio is off goes straight into the tank.
The alarm said what. Nobody could see why.
Five losses kept recurring in and around the boric acid premix skid.
The biggest was dosing. Around ten times a month the loop raised a dosing, concentration or deviation alarm, even with the setpoint and tolerance configured correctly. Finding the cause took 65 minutes, by which point off-ratio product had already reached the FP tank. The whole batch then had to be reblended or scrapped: $80,000 of product in a year.
The other four were mechanical and electrical. Pigs stuck at changeover because they were launched below the pressure the line needed, and had to be pulled out of the pipeline before the process could restart. Valves failed mechanically while the PLC still showed them open. Communication and network errors, including two IO cards sharing one IP address, dropped valve signals intermittently. Damaged cables cut feedback altogether.
Each time, the cause sat in a different system: a flowmeter trend in the historian, a feedback bit in the PLC, a comms fault in the event journal, a concentration result in the quality database. Pulling those together mid-shift took the most experienced person on the floor, if they were on shift at all.
Where the losses start on the skid
For each of the five losses, we defined the combination of signals that tells it apart from the others. These signatures are what Nova matches against when an alarm fires. Select a point on the drawing to see the signature, the root cause and the troubleshooting steps.
One skid, seven sources, one case timeline
Every source was connected read-only, through the plant's existing servers. Nothing on the PLC was changed. Each source has a dedicated tool the agent calls when it needs that piece of evidence, so every conclusion Nova draws can be traced back to the system it came from.
| Source | What it tells Nova on this skid | Read by |
|---|---|---|
| PLC | Valve command and feedback, pump run status, IO module and communication diagnostics | get_live_state |
| SCADA / HMI | Dosing setpoints and tolerances, live flow and pressure values, active recipe | get_live_state get_recipe |
| Historian | Second-by-second trends for flow, pressure, pump current and tank levels leading up to the alarm | get_timeseries |
| Alarms and events | Dosing, concentration and deviation alarms, IO comms faults, pigging sequence events | get_alarms get_network_events |
| Quality | Concentration, pH and viscosity results per batch and campaign, linked to the time window of an excursion | get_quality |
| UPS and power | Power and UPS status, to rule power events in or out when IO cards drop out | get_live_state |
| Operator experience | SOPs and manuals, plus troubleshooting know-how recorded from senior operators and every resolved case | search_docs recall_similar_cases |
From loss map to live assistant
The work moved in five steps, starting with the operators who live with the losses and ending with Nova on their phones.
Mapped the losses
Interviewed making operators and engineers, then ranked losses by frequency and cost. Premix dosing came out on top. In the same interviews, Nova also captured the changes operators had made to the process in the last few weeks.
Connected the data
Linked PLC, SCADA, historian, alarm, quality and UPS data read-only into one timeline per skid.
Defined signatures
Turned each failure mode into a signal pattern Nova can recognise, and checked it against past excursions.
Captured know-how
Recorded how senior operators troubleshoot each mode and turned it into playbook entries, linked to SOPs.
Went live
Nova now opens a case on every alarm, proposes the root cause with evidence, guides the fix and saves what worked. Every process change and fix, from the smallest to the largest, is recorded in the app, so the brain keeps growing and everything is documented in one place.
What Nova could change on the boric acid premix skid
Root cause in 10 minutes instead of 65 means off-ratio product is caught before it reaches the FP tank. That alone could cut off-quality losses by $60,000 a year, based on the on-site demo.
Off-quality product reblended or scrapped
Time from alarm to confirmed root cause
New operator training time
Based on guided troubleshooting from day one
Process knowledge
Scattered across structured and unstructured systems, with tribal know-how held by operators
One brain for the premix skid: 100+ SOPs and documents, thousands of centerlines and millions of live data points
Less product off-spec
Catching the signature early and routing output to reblend before the FP tank cuts the volume affected by each excursion.
Fewer repeat causes
Patterns across cases, like low buffer level before air blocks or a duplicate IP behind two "random" valve faults, become permanent fixes instead of recurring alarms.
Every shift as good as the best one
The way the most experienced operator troubleshoots is available to every operator, on every shift, on their phone.
What plant teams ask us
Does Nova change anything on the PLC or SCADA?
No. Every connection is read-only. Nova recommends and verifies; the operator acts. Changes like a new interlock go to the process engineer as a request.
How does Nova connect to plant systems?
Through the plant's existing OPC-UA server, historian and databases, rarely to individual PLCs. Fast signals are buffered and downsampled at the edge, on plant servers. OT security sign-off is usually the real gate, not the technology.
Where do the root-cause probabilities come from?
From how closely the live signals match each failure signature, weighted by how often that signature turned out to be that cause in past resolved cases on the same line.
What if the root cause is something new?
Nova still gathers the evidence, cites the relevant manual and SOP sections, and narrows the options. Once the operator resolves it, the fix is captured, so the next time it's in the playbook.
Which AI models does it use?
Nova is model-agnostic. It runs through a model gateway, so the plant can use Claude, Gemini or another approved model.
Have a loss that keeps coming back?
We start with a loss-mapping session with your operators, rank your losses by cost, and show where Nova pays back first.