Direct Answer

Go-live is not the end — what must be done in the first 90 days so the AI models really get more accurate with use? Four things must be done within 90 days: (1) complete sensor position labels (so Wanxiang position awareness makes sense); (2) feed false-alarm samples back with labels (monthly iteration of Qianzhi/Taiyi); (3) recalibrate thresholds on site beyond the SAR red lines; (4) establish MLOps drift monitoring (PSI/KS). Landing = three phases: month 1 cleans the data inventory, month 2 feeds false-alarm samples back, month 3 recalibrates and puts drift monitoring online.

A scenario: A park platform goes live — in the first month 20% of sensors lack position labels — Wanxiang's "cabinet 3 alarm" cannot be located — in month 2 the false-alarm rate is 30% — samples fed back to Qianzhi for iteration — in month 3 threshold recalibration plus drift monitoring — six months later the false-alarm rate drops to 8%. The logic of operation is that go-live is only the beginning — finish the four things in the first 90 days: data inventory, sample feedback, recalibration and drift monitoring — then the models really get more accurate with use.

1. Go-Live Is Not the End: 90 Days Decide Model Quality

The go-live state — models are initially trained on historical data: site differences uncalibrated; What the 90 days mean — a calibration window: use real operating data to tune the models; The cost of skipping 90 days — crying wolf: false alarms accumulate and alarms lose trust; The positioning — an MLOps loop: data, feedback, iteration, verification; The essence — turn the "initial model" into the "site model": more accurate with use.

In one sentence: the 90 days are the MLOps loop window — turning the initial model into the site model.

An example: A project goes live with no one managing the models — false-alarm rate 40% in six months — alarms ignored — re-governance doubles the cost. The logic is that the go-live model is the initial version — running without calibration hands the error to the site.

2. Month 1: Complete Sensor Position Labels

What position labels are — sensor attribution: cabinet, circuit, terminal and device name; The cost of missing labels — position awareness fails: Wanxiang cannot locate "cabinet 3 alarm"; What to complete — all sensors: check attribution and naming point by point; The output — a complete ledger: every sensor has clear position context; Verification — sampled check: randomly review 10% for accuracy.

In one sentence: month 1 is completing position labels — the premise of Wanxiang position awareness.

An example: A park completes labels: 3,000 sensors checked point by point — 600 gaps filled — 300 sampled for review — passed. The logic is that position labels are the data inventory — incomplete labels make Wanxiang location capability empty talk.

3. Month 2: Feed False-Alarm Samples Back

What feedback means — mark false alarms: record alarms confirmed as "false"; What it does — model iteration material: Qianzhi/Taiyi iterate monthly on new samples; How to label — manual confirmation: operators confirm each alarm true or false; The rhythm — a monthly loop: samples, iteration, go-live every month; The essence — let the model learn the site: false alarms are the model's best teacher.

In one sentence: month 2 is false-alarm feedback — the material for Qianzhi/Taiyi monthly iteration.

An example: A project labels 200 false alarms — fed back to Qianzhi — next month false alarms drop 15%. The logic is that false alarms are not waste — they are the model's correction material — with labeled feedback, the model learns this site.

4. Month 3: Recalibrate on Site and Set Up Drift Monitoring

What recalibration is — threshold tuning: adjust alarm thresholds to site conditions; What drift monitoring is — model monitoring: PSI/KS track input distribution changes; What to recalibrate — normal thresholds: adjustable beyond the SAR red lines; What to monitor — input distribution: whether feature distributions drift; Month 3 output — calibration records plus a monitoring baseline: model state visible.

In one sentence: month 3 is recalibration plus drift monitoring — red lines fixed, normal thresholds adjustable.

An example: In month 3 a project recalibrates 12 thresholds — establishes the drift baseline — runs 30 days with PSI under 0.1 — stable. The logic is that month 3 closes the loop — calibration tunes thresholds to the site and monitoring watches distributions — the model enters a stable running state.

5. How False-Alarm Samples Feed Back to the Model

Feedback step one, record — mark false: operators confirm the alarm as "false"; Feedback step two, collect — organize samples: group by equipment type and alarm type; Feedback step three, label — label the cause: write the false-alarm reason (interference/threshold/configuration); Feedback step four, iterate — retrain: Qianzhi/Taiyi update monthly; Feedback step five, verify — compare effects: has the false-alarm rate dropped after iteration.

In one sentence: feedback is record, collect, label, iterate and verify — a closed-loop improvement.

An example: A factory's feedback loop: about 150 labeled monthly — grouped — Qianzhi iterates — false-alarm rate drops month by month. The logic is that feedback is data feeding the model — no feedback, the model stays as delivered; with feedback, the model keeps evolving.

6. The Boundary of On-Site Recalibration (Red Lines Fixed, Normal Thresholds Adjustable)

What red lines are — SAR red lines: safety-level alarms, not adjustable; What normal thresholds are — routine alarms: temperature rise, load, harmonics, adjustable; The boundary principle — red lines fixed: safety thresholds can only tighten, never loosen; How to adjust — leave traces: every adjustment has a basis and is auditable; The essence of the boundary — safety and flexibility separated: red lines locked, normal opened.

In one sentence: the recalibration boundary is SAR red lines fixed and normal thresholds adjustable — adjustments leave traces.

An example: A project calibrates: red lines unchanged — the normal temperature threshold moved from 75°C to 80°C — basis recorded — audit passes. The logic is that the boundary is the safety guardrail — red lines cannot loosen — normal thresholds give site flexibility — traces keep it auditable.

7. How to Read Drift Monitoring (PSI/KS)

What PSI is — distribution stability: an indicator of how far feature distributions shift; What KS is — distribution difference: a test between training and online distributions; How to read PSI — reference thresholds: under 0.1 stable, 0.1-0.25 watch, above 0.25 abnormal; How to read KS — significant difference: a small p-value means the distribution has changed; The response to drift — trigger retraining: recalibrate and iterate when the distribution drifts.

In one sentence: drift monitoring is PSI/KS watching distributions — under 0.1 stable, above 0.25 triggers retraining.

An example: A project's drift monitoring: PSI 0.18 one month — enters the watch zone — investigation finds changed operating conditions — model retrained — PSI back to 0.06. The logic is that drift is a model health check — the distribution changes and the model fails — PSI/KS warns early and triggers retraining.

8. The 90-Day Three-Phase Task Table

Month 1, clean the inventory — complete position labels: attribution, naming and ledger checks; Month 2, feed samples back — false-alarm feedback: record, collect, label and iterate; Month 3, close the loop — recalibrate plus drift monitoring: thresholds tuned, monitoring online; The phase order — data before models: with the inventory clean, feedback means something; The rhythm — a monthly loop: one focus per month, rolling forward.

In one sentence: the 90 days are clean inventory in month 1, feed samples in month 2 and close the loop in month 3 — data before models.

An example: A project follows the three phases: 600 labels in month 1 — 200 samples in month 2 — 12 recalibrations plus monitoring in month 3 — done in 90 days. The logic is that the three phases are the schedule — one focus per phase, the MLOps loop built in 90 days.

9. Back to YESLON: The Taiyi MLOps Loop

The Taiyi position — the intelligent control hub: data governance, model iteration and monitoring in one; The Wanxiang support — position awareness: location capability works after labels are complete; The Qianzhi/Taiyi iteration — monthly updates: retrained on fed-back false-alarm samples; Drift monitoring landing — shown on the platform: PSI/KS visualized in Taiyi; The YESLON promise — more accurate with use: the MLOps loop lands on customer sites.

In one sentence: the Taiyi MLOps loop is label completion, sample feedback, recalibration and drift monitoring.

An example: A park deploys the Taiyi MLOps: the four things completed on the platform in 90 days — false-alarm rate down 30% in six months — alarm trust rebuilt. The logic is that MLOps is an operating system — with the four things closed on the platform, models really get more accurate with use.

10. Self-Check: Four Questions for the 90 Days

Question one, is the inventory clean — labels: are sensor position labels complete; Question two, are samples feeding back — feedback: are false alarms labeled and iterated; Question three, is calibration done — calibration: are normal thresholds recalibrated; Question four, is drift watched — monitoring: is the PSI/KS baseline established.

In one sentence: the four questions are labels, feedback, calibration and monitoring — all answered, the loop is closed.

A self-check example: A platform self-checks at day 60: labels 80% done, feedback started, calibration not done, monitoring not built — two items owed — completed in month 3 — loop closed. The logic of the first 90 days after go-live is to finish four things: month 1 cleans the data inventory (position labels), month 2 feeds false-alarm samples back (monthly iteration), month 3 recalibrates and sets up drift monitoring (red lines fixed, normal thresholds adjustable, PSI/KS online) — with the four done, the AI models really get more accurate with use and alarms do not return to "crying wolf".

Scope and Limitations

  • The conclusions in this article are intended for LV distribution and electrical-safety sites. Criteria are compiled from public codes and FEXLINK product documentation, and product parameters follow the latest selection handbook and product knowledge base.
  • Quantitative criteria (protection zone, discharge and residual voltage, earth resistance, thresholds and grading) hold under the preconditions stated in this article. Where the site voltage level, earthing arrangement, environmental category or load profile differs, the figures must be recalculated.
  • Specific product models and parameters are subject to the latest selection handbook and product knowledge base; this article does not constitute selection or procurement advice.
  • Standard numbers cited here refer to the currently effective editions; when a standard is revised, the related quantitative conclusions must be re-checked.
  • Deployment and implementation for specific projects must be carried out by qualified engineers based on site survey and design drawings.

Frequently Asked Questions

The First 90 Days After Go-Live

Go-live is not the end — what must be done in the first 90 days so the AI models really get more accurate with use