How AI Sensors Monitor Cold Storage Durian Quality

Step 1: Attach Ethylene and CO2 Sensor Arrays

Durian metabolizes even at the recommended storage window of 5–7°C. The fruit’s postharvest respiration shifts the internal gas balance, and the first reliable indicator of premature ripening is a spike in ethylene. In a working setup at a Shah Alam cold storage facility, the array typically uses Figaro TGS2602-B metal-oxide gas sensors mounted on each pallet’s ventilation duct. These sensors detect ethylene from 0.05 ppm upward. Next to them, a Sensirion SCD30 CO2 sensor tracks respiration rate.

The mounting position matters. Sensors placed on the return-air side of the evaporator unit read bulk cycles, not per-pallet conditions. For per-pallet reading, the sensor board must sit at the point where air exits the pallet’s bottom vent core. Malaysian operators who skip this detail end up measuring the cold room, not the fruit. Wiring runs through CPVC conduit because the acidic skin compounds corrode standard PVC. Calibration is done against a 10 ppm ethylene gas cylinder before each Musang King season batch.

Step 2: Scan Pallets with Thermal and Near-Infrared Cameras

Gas sensors tell you that the fruit is ripening. Thermal and NIR imaging tell you where on the pallet and how far along the process has progressed.

A FLIR A400 LWIR camera, mounted on the gantry above the conveyor belt between the night loading dock and the cold room entrance, captures a thermal map of every pallet surface. Durian pulp temperature stays around 6.5°C when stable; any area above 8.2°C suggests poor airflow bypass or a gap in the pallet wrap. The camera runs at 60 fps with a sensitivity of ±0.5°C. Because the warehouse floor in Puchong has thermal drift from afternoon sun hitting the dock door, the camera’s auto-focus is disabled. Instead, the image acquisition script performs a fixed-focus capture triggered by the pallet’s optical proximity sensor.

Then comes the Specim FX17 hyperspectral imager in the 900–1700 nm range. It scans the husk’s reflectance. The spectral signature at 1150 nm and 1450 nm correlates well with interior sugar density and moisture of the pulp—this has been validated on MD2 and Musang King samples. But there is a catch: the FX17 sees the surface, not the inside. The model compensates using a known transfer function derived from destructive tests of 300 fruit across three Malay Peninsula farms.

Step 3: Run Edge Inference on Grading Models

All of this sensor output lands on an edge device—typically a Coral Edge TPU for classification or a NVIDIA Jetson Orin Nano when the deployment also runs the NIR model. In most KL setups, the Jetson is preferred because it can execute both the thermal-image mask and the spectral classifier without sending data to a cloud backend.

The inference pipeline runs in five stages:

1. Normalize thermal frames using a per-pallet emissivity correction for green vs. brown husk surfaces.

2. Extract average ethylene and CO2 slope over the last 30 minutes.

3. Feed the spectral reflectance vector into a LightGBM classifier trained on the destructive-test dataset.

4. Combine the two outputs using a rule: if ethylene slope exceeds 0.3 ppm/hour and thermal variance above 1.1°C, the pallet is flagged as “entering climacteric phase.”

5. Emit a grade score from 0.85 to 0.20, where 0.85 means “fit for overnight delivery,” and anything below 0.45 triggers a re-grade.

The latency stays under 50 ms. No cloud, no 4G dependence. The warehouse operates a diesel backup generator anyway because Tenaga Nasional feeder lines in that industrial lane are unreliable during the monsoon.

Step 4: Adjust Cold Room Fans and Compressors Automatically

The output of the edge model connects to the cold room’s programmable logic controller via Modbus TCP. When the system detects a warm spot near one evaporator’s airflow shadow, it opens the corresponding variable air volume damper by 15% and raises that fan’s speed from 800 rpm to 1,100 rpm. This resolves the thermal imbalance without forcing the main compressor to cycle.

In the durian-specific case, ethylene is not just a signal—it is a catalyst. One ripening pallet can trigger its neighbors. So the Step 3 classifier also calculates a “spillover risk index.” When that index crosses a threshold, the system commands a 2-minute burst of the ammonia refrigeration compressor to drop the room from 6°C to 4.5°C. The low setting is a one-off maneuver; holding it there would damage the pulp’s cell walls and cause candy-like transparency in the flesh.

Malaysian installations need a robust control loop. The typical cold room control board in Shah Alam runs a direct expansion system with hot-gas defrost every 6 hours. The edge unit must suppress the AI trigger during those defrost windows, otherwise the temperature rise from the defrost cycle gets misread as a ripening event.

Step 5: Tag Batches with Predicted Pulp Shelf-Life

The final phase is output labeling. Every pallet gets a Grade A, B, or C sticker printed by a barcode printer at the chiller’s exit. The grade is not just visual—it embeds the model’s predicted remaining shelf life at 6°C.

– Grade A (sensor score 0.70+): ripening suppressed, predicted shelf life 14–16 days. Good for West Malaysia distribution.

– Grade B (0.45–0.69): mild climacteric activity; shelf life 8–12 days. Marked for immediate transfer to the Jalan Ipoh wholesale market tier.

– Grade C (below 0.45): high ethylene and internal fermentation signature; shelf life under 6 days. Redirected to a pulp-processing partner in Puchong that makes frozen durian puree.

The shelf-life tag is printed in both English and Malay, with the digital twin record stored in a structured query language table on the warehouse local area network. This gives the quality audit trail that the Malaysian Department of Standards’ MS 1501 fresh-fruit classification expects. Once the pallet reaches a retail buyer in SS15, the buyer can scan the QR code to see the original spectral curve—not a hand-written “checked on” date.