How Smart Warehouses Use AI Sensors for Stock MY

Depth Cameras, Load Cells, and Passive Counting

The old method in Shah Alam and Pulau Indah warehouses is a human with a Zebra RF scanner walking every rack. A 10,000-pallet site burns about 220 man-hours per month on cycle counting. The passive-counting retrofit replaces that with two sensor families. First, 3D depth cameras (SICK PLR, Orbbec, or Keyence) mounted on rack uprights pointing down each aisle, capturing pallet footprint changes every 30 seconds. Second, load cells rated at 2.5 tonnes embedded under rack feet or pallet positions, giving a live weight delta per bay.

The AI layer runs on edge boxes (NVIDIA Jetson Orin, Hailo-8) inside the warehouse, not in the cloud. This matters in Malaysia because CelcomDigi and Maxis 5G coverage inside metal-clad buildings in Bukit Raja is unreliable, and pushing 4K frame data over LTE is a non-starter. The edge model uses a YOLOv8-class detector to distinguish full pallet, partial pallet, empty bay, and loose carton sprawl. When the camera frame disagrees with the WMS bin state by more than one pallet footprint, it raises a discrepancy ticket. That ticket is the “Stock MY” variance — the gap between what SAP EWM thinks is on the floor and what the sensor net physically observes.

Humidity, 5G Dead Zones, MPOB Audit Trails

Sensor calibration in Malaysia is not a temperate-climate problem. Non-air-conditioned rack space in Port Klang runs at 33°C with 80% relative humidity for six months of the year. Cheap consumer LiDAR units fog up and drift. Working units in this region are IP65-rated industrial scanners, and installers in the Klang Valley route them through PoE switches with surge protection because lightning storms around the Strait of Melaka kill under-spec power supplies.

The audit trail is the harder part. If the warehouse stores palm olein fractions, the Malaysian Palm Oil Board (MPOB) license requires traceability from weighbridge ticket to tank and drum movement. AI sensor data has to synchronise with the weighbridge PLC, not just the WMS. Integration is typically via MQTT to a local broker, then an OPC-UA connector upstream to the ERP. The sensor log becomes a timestamped record of weight change per bay, which maps cleanly to MPOB’s movement documentation requirements. Warehouses that skip this sync get flagged during the annual licence renewal audit.

Damaged Lot Masking: The Real Stock MY Leak

The most common Stock MY failure mode is not phantom pallets. It is damaged lot masking — the sensor sees a pallet in the correct bay, so the count looks right, but the cartons underneath are short-dated or damaged and have been sitting there for 90 days past their sell-by window. FMCG warehouses in Pulau Indah and Klang handle goods with 30 to 60-day shelf lives. A camera that counts cartons solves nothing if the oldest stock is buried at the back.

The fix deployed in practice is a two-stage sensor pass. The ground camera reads the SSCC barcode on the pallet label. A secondary OCR model reads the expiry date stamp on exposed cartons at the camera field of view. When the date mask exceeds the WMS’s FEFO (First-Expiry-First-Out) threshold, the AI flags a “masked lot” and blocks that bay from being selected for outbound picking. This is not a theoretical capability; the logic runs on the same edge node, and the blocking instruction propagates to the warehouse control system via a simple REST webhook to the pick-to-light or voice-picking middleware.

The Wiring Cost That Kills the Business Case

The economics are brutally local. Retrofitting a 10,000-pallet rack block in Shah Alam with a sensor mesh — roughly 60 depth cameras, 900 load cells, four edge boxes, PoE cabling, and mounting — costs between RM250,000 and RM400,000 installed. The payback math depends on eliminating two manual cycle counter positions (RM1,700 minimum wage, plus 1.5x overtime during month-end stocktake) and cutting the annual stock-shutdown day. A single shutdown day at a busy KL Valley DC, where outbound trucks idle at the dock, costs more than the sensor capex in lost throughput and demurrage.

What kills the business case is building cabling into rack aisles that move. Warehouse operators who think they can shift racks quarterly to chase rental rates in PIK or Elmina should not install cabling. The workable deployments are in fixed-layout racks with anchor-bolted bay positions. For the mid-tier operators running HashMicro or Odoo WMS, the sensor integration is a CSV export per reconciliation cycle. For the bigger players on SAP EWM, the sensor net posts inventory differences via IDoc message type MBGM and auto-generates 701/702 postings for the warehouse clerk to approve.

KLIA Cargo and Penang Precision Deployments

The two native-species deployments in Malaysia right now tell two different stories.

At KLIA Aeropark, a freight forwarder uses volume-dimensional scanning (CubiScan-style) at the bonded warehouse inbound bay. The AI classifies each incoming unit by SKU family using 3D silhouette, then the load cell confirms gross weight, and the WMS gets a confirmed carton count within 11 seconds of the unit passing the scanner frame. This kills the old practice where bonded stock counts were done by staff physically opening bales and counting inner cartons — a process that delayed Customs Form 2 clearance.

In Penang’s Bayan Lepas industrial zone, a precision-parts electronics warehouse runs the opposite deployment: vibration and tilt sensors on high-value bare PCBA trays. The AI detects when a tray is lifted off its storage shelf by the slight vibration signature and tilt angle change. If the tray moves to a location not tied to an open pick order, the system raises a misappropriation alert. This is specific to the high-value, low-volume inventory profile of Malaysian electronics contract manufacturers, and it works because the stock is worth more per gram than the sensor network around it.