How Xiaomi Leveraged AI to Transform Air‑Conditioner Installation and Energy Efficiency

Facing the long‑standing problem that air‑conditioner performance depends heavily on correct installation, Xiaomi built an AI‑powered pipeline that turns the previously invisible installation stage into a measurable, closed‑loop process. The first target was indoor‑unit height recognition. Poor photo quality, inconsistent labeling, and regional scene variance were identified as three core obstacles. The team first introduced a lightweight image‑quality model to enforce photo standards, then unified annotation by assigning senior engineers to label data, and finally addressed systematic distribution bias by expanding the training set with diverse angles and lighting. After two weeks of iterative refinement, model accuracy rose from 60 % to 95 %, and after three deployment cycles the field accuracy stabilized at 93‑94 % nationwide.

For energy‑saving, Xiaomi rejected the traditional fixed‑parameter control and designed a cloud‑edge collaborative architecture. Over 80 % of devices stay online, allowing real‑time policy updates. The cloud performs deep reinforcement learning on compressor, expansion‑valve, and evaporator data, while the edge executes lightweight commands. Early lab tests (the first three‑day run) showed a 25 % reduction in power consumption versus conventional control. Subsequent extensive testing resolved temperature‑fluctuation issues, leading to a stable AI‑driven control that, by the end of 2025, delivered a cumulative 29 % energy‑saving rate, saving about 2.46 × 10⁸ kWh, cutting electricity costs by ¥1.56 billion and reducing CO₂ emissions by 14.268 kt.

The AI framework also enabled intelligent diagnosis. A filter‑clog detection model quantifies blockage severity and pushes alerts with cleaning instructions, while a refrigerant‑leak detection model fuses multi‑source data to identify low‑charge conditions without manual pressure tests. Both diagnostics reuse the data‑quality and annotation lessons from the installation model, shortening development cycles for new fault‑type models.

Crucial to these achievements was a two‑way collaboration between the appliance business units (hardware, installation, service) and the AI lab. Business teams provided domain knowledge—hardware structure, installation standards, control logic—while the algorithm team supplied modeling expertise. Weekly joint sessions, on‑site immersion, and a 24‑hour lab‑resident task force bridged the language gap, allowing rapid iteration and shared understanding of both business constraints and AI capabilities.

Overall, the six‑month effort produced a reusable AI engineering methodology: data‑first, annotation‑standardized, distribution‑aware modeling, and cloud‑edge co‑design. The approach cut model development time from six months to one month for new categories, scaled from a single air‑conditioner line to the entire home‑appliance portfolio, and set a roadmap for future AI‑enabled services such as proactive after‑sales and cross‑product intelligence.