How Durian Processing Teams Use AI for Food Safety

Step 1: Scan Incoming Husks for BGYF and Micro-Cracks

The intake line at a Musang King sorting shed in Raub looks nothing like a sterile factory. Fruit arrives at 6 a.m. on open lorries, still carrying soil from the orchard floor, and the first food safety decision happens before any flesh is exposed. Teams have started bolting Keyence CV-X series AI vision sensors above the grading conveyor. The model is trained to spot bright greenish-yellow fluorescence (BGYF) on the husk under UV light — a known marker for aflatoxin-producing Aspergillus strains. The AI also flags micro-cracks that would allow chlorinated wash water to seep into the flesh and inflate free moisture during the 14-hour transit from Bentong to the processing plant in Nilai. Diverted fruit goes straight to pulping rather than the whole-fruit export line for the GACC market.

Step 2: Fuse Moisture and Metal Detection on the Paste Line

On the IQF flake and paste lines, food safety risk moves from mold to physical contaminants. Calcified seed shards, steel bristles from wash brushes, and fragments of palm kernel from the orchard sorting yard are the common failures. Instead of relying purely on a single metal detector, teams fuse the load cell data with an Ishida X-ray unit and a moisture sensor reading. The AI algorithm adjusts the rejection threshold in real time based on husk moisture content — a wet batch triggers a stricter logic to prevent soft, dense shards from slipping through. One facility in Segamat reported its false-reject rate dropping from 1.8% to below 0.4% after calibrating the fusion model against three months of historical reject data.

Step 3: Train Cold Chain Models on Freezer Defrost Downtime

The most underappreciated food safety risk is the blast freezer. Its defrost cycle runs automatically, often at night, and a single extended defrost can push the pulp internal temperature past 18°C. Re-freezing that pulp creates weeping flesh and free water on the surface — an ideal growth medium for Listeria monocytogenes. Processing teams now log time-temperature records from Tempmate wireless loggers alongside the blast freezer’s kW draw and door-open events from the chiller stores at Pulau Indah. The machine learning model predicts exactly when a batch crossed the growth threshold, so the quality manager can hold the affected trays for coliform testing instead of releasing them into a 40-ft reefer container bound for Shanghai.

Step 4: Auto-Draft MAQIS HACCP and GACC Documents

The documentation burden is where AI saves the most labour. Handwritten wash-bay chlorination logs are digitized using an OCR pipeline — Google Vision API or Abbyy — that reads the chlorine residual entries and flags any reading below the 50 ppm free-chlorine threshold for the wash water. Every change to a freezer setpoint is automatically version-stamped and appended to the HACCP plan. When a MAQIS inspector or a GACC auditor visits a plant in Simpang Renggam or Layang-Layang, the system can generate a date-sealed PDF log of the entire cold chain for a named lot, completely bypassing the shoebox of damp notebooks that most facilities used to produce.

Step 5: Track Flesh Packets to Harvest Lots and Courier Runs

Full traceability is a recall drill that the big exporters drill quarterly. Each flesh packet leaving the grading line in a chilled van gets a QR label carrying the orchard lot code assigned by the sourcing team at the Raub hillside. The AI matching engine in the plant’s ERP — a customized Yonyou SCM deployment — links those codes to the outbound packing list and the container number. If a buyer in Guangzhou files a complaint about an off-flavor, the quality team can pull up the exact harvest date, the pulp temperature profile, and the last-mile dispatch record from the GrabForBusiness or Lalamove trucking run that sat at the Sultan Abu Bakar CIQ checkpoint before crossing to Singapore. The complete isolation and hold decision takes under 30 minutes.