+ Machine Learning · E-Sentinel · Trained against a validated reference
Trained against the truth. Deployed without it. The inexpensive nodes infer what the expensive instrument measured.
E-Sentinel is the Moseley on-device ML model. It learns to recognise ethylene state and trajectory by training against a validated reference ethylene monitor during field trials — a relatively expensive, large, commercially deployed instrument that rides along with the E-Aegis on training voyages. Once trained, E-Sentinel runs on every E-Aegis without that reference instrument. A distilled TinyML student of E-Sentinel then runs on every E-MCP release node and every E-Nose Label pallet sensor — nodes that carry no ethylene sensor of their own, and do not need one.
+ Architectural pattern · load-bearing across Moseley
A source-of-truth instrument teaches once during training trials. Inexpensive distributed nodes infer state from whatever signals they already have — and act on that inference. Every Moseley product rides on this pattern. It is what makes ambient ethylene intelligence economic at fleet scale, and it is what allows low-cost sensors to deliver precision sensing that lets the ML model predict potential problems — across produce, pharma, semiconductors and beyond.
01 · How E-Sentinel is trained
A validated reference monitor rides along. E-Sentinel learns from it.
Ambient IoT trains E-Sentinel in the field, not in a simulator. During training deployments, an E-Aegis is co-located with a validated reference ethylene monitor — a relatively expensive, large, commercially deployed instrument already trusted in cold-storage operations. The two devices breathe the same air, inside the same refrigerated environment, over the same produce. The reference monitor produces validated ethylene readings; the E-Aegis logs its own multi-sensor response from the on-board E-Array. E-Sentinel learns the mapping between the two.
01 Co-locate
Reference monitor + E-Aegis in the same headspace
A validated reference ethylene monitor and an E-Aegis are installed in the same refrigerated environment over the same produce — controlled-atmosphere room, ripening room, refrigerated container or distribution centre — for the full duration of the trial. Both devices see the same molecules.
field trialsame airsame produce
02 Label
Reference readings become the training signal
The reference monitor’s validated ethylene readings act as the labelled signal for every moment of the trial. Rising ethylene, climacteric onset, spoilage-risk territory — each labelled by an instrument the industry already trusts. No synthetic data. No spec-sheet substitutes.
validatedground truthreal cold-chain
03 Learn
E-Sentinel learns the E-Array → state mapping
E-Sentinel is trained to map the E-Aegis’s multi-sensor E-Array response to the reference monitor’s labels. Over many trials, it learns to recognise from the E-Array signal alone the patterns associated with rising ethylene, climacteric activity, and spoilage risk — without needing the reference instrument in deployment.
supervisedmulti-sensoron-device
02 · The training pipeline
Four stages. One device emerges trained.
During training trials, two parallel signals are recorded in the same airspace. E-Sentinel learns the mapping between them, then emits a calibrated risk score that downstream nodes can act on.
▶ E-Sentinel training & surrogate sensing explainer · 2:29
03 · What E-Sentinel learns to output
State & trajectory recognition — reported as a calibrated risk score.
E-Sentinel does not claim to deliver laboratory-grade absolute parts-per-billion readings from the E-Array. That is not what it is trained to do, and not what downstream remediation actually needs. E-Sentinel is trained to recognise ethylene state and trajectory — rising, climacteric, spoilage-risk — from the E-Array signal it sees, and to emit a calibrated risk score that downstream nodes act on. The risk score is the contract between the brain and the rest of the architecture.
A · State recognition
Rising · climacteric · spoilage-risk.
E-Sentinel learns to identify which regime the produce is currently in: baseline pre-climacteric, ethylene rise, active climacteric, or spoilage-risk territory. Each regime triggers a different downstream response.
B · Trajectory inference
Where the curve is going, not just where it is.
Because E-Sentinel is trained on rate-of-change patterns labelled by a validated reference, it learns to anticipate climacteric events — reading the curve while it is still climbing, not just after it peaks. Lead time, measured in hours, becomes the dimension that matters.
C · Calibrated risk score
The output every downstream node consumes.
State and trajectory collapse into a single calibrated risk score. The E-MCP release node consumes it to decide when to act. The E-Nose Label reports against it. The risk score is the actionable language of the Moseley fleet.
What we are not claiming. We are not claiming the E-Array is a laboratory-grade ethylene analyser. We are claiming something more useful and more defensible — that E-Sentinel, once trained against a validated reference, recognises state and trajectory well enough to drive remediation decisions. That is a practical, believable development path.
04 · Surrogate sensing
The student does not need to share sensors with the teacher.
This is the bigger idea. In a teacher-student ML system, the student is not required to carry the same sensors as the teacher. It only has to carry enough correlated information for the model to learn the relationship. The student then performs latent inference — predicting the teacher’s state from a different, less expensive set of signals. The Moseley E-MCP release node is the worked example.
Worked example · E-MCP release node
The E-MCP carries no ethylene sensor at all. It does not need one. The TinyML student running on the E-MCP — distilled from E-Sentinel on the E-Aegis — has been trained to recognise climacteric state from the environmental fingerprint a release node already sees.
The teacher saw ethylene. The student infers it from everything else.
Signals the E-MCP already has
Temperature, humidity, time, neighbours, adsorbent state.
Every E-MCP has access to its own temperature, its own humidity, its own time-in-deployment, the BLE neighbour signals it broadcasts and receives, and the state of its own adsorbent loading. None of these is an ethylene sensor. Together, they are more than enough for a model to infer climacteric state.
What the model learns
Environmental fingerprints, not absolute readings.
During training, paired E-Aegis and E-MCP data are recorded in the same headspace. The TinyML student learns: when the environment behaves like this, the teacher historically saw rising ethylene or climacteric onset. Adsorption kinetics, moisture interactions, packaging microclimate and diffusion timing all become predictive features.
Why this is stronger
Leaner BOM. Bigger idea. Defensible IP.
No VOC channel to specify, calibrate or drift-compensate. No ethylene sensor to over-claim. The architectural argument — climacteric state inferred from distributed environmental signatures, not direct ethylene quantification — is a much bigger position than “inexpensive ethylene sensor” and far more defensible in claims.
The architectural claim. E-Sentinel infers climacteric state from distributed environmental signatures rather than relying solely on direct ethylene quantification. That is the load-bearing pattern across the Moseley fleet — and the reason a release node can act on climacteric intelligence without ever measuring ethylene itself.
05 · Once trained, the reference is gone
The reference monitor is training-time scaffold. It does not ship with the product.
The validated reference monitor is large, expensive, and impractical to put on every reefer journey. That is the entire point of training. Once E-Sentinel has been trained against enough labelled events across enough produce categories and cold-chain conditions, the reference instrument is removed. The E-Aegis goes to work on its own, running the trained E-Sentinel model entirely on-device.
01
Training trials are run in controlled-atmosphere rooms, ripening rooms, refrigerated containers and distribution centres — with the reference monitor and E-Aegis co-located in the same headspace.
02
Paired samples — reference readings and E-Array responses — are pushed back to the Moseley training pipeline along with operator context (temperature, humidity, CO2, produce category, arrival quality).
03
E-Sentinel is updated. The new model is validated against held-out trial data, then pushed over-the-air to every E-Aegis in deployment.
04
Deployment runs solo. No reference monitor in the reefer. No connectivity required. The E-Aegis emits its risk score on-device, in real time, against the journey it is actually on.
06 · E-Sentinel teacher → TinyML student
The trained model also distills down to the inexpensive nodes.
An E-Aegis is engineered to host E-Sentinel and a multi-sensor E-Array. A pallet-level sticker or a hockey-puck release node cannot. So a second compression step happens: the trained E-Sentinel becomes the teacher for a small TinyML student, distilled down to run on the modest hardware inside an E-MCP release node and an E-Nose Label pallet sensor. The student inherits E-Sentinel’s judgement, without inheriting its hardware bill.
What distillation means
A small model learns from a big one.
Knowledge distillation is a machine-learning technique in which a small, inexpensive student model is trained to imitate the decisions of a larger, more capable teacher. The student ends up reproducing nearly the same decisions on commodity hardware.
The teacher
The trained E-Sentinel.
E-Sentinel has already been calibrated against a validated reference and runs on the multi-sensor E-Aegis. It is the trained truth from which the students are distilled.
The students
TinyML on E-MCP & E-Nose Label.
The distilled student model is flashed onto every E-MCP release node and every E-Nose Label. They infer with the teacher’s judgement, but at sticker-pack hardware economics.
07 · Why this is the right development path
Modest about the sensor. Ambitious about what the model extracts from it.
A new ethylene sensor that claims laboratory-grade ppm accuracy out of the gate is asking customers for a leap of faith the industry will not give. A new ethylene sensor that claims to recognise state and trajectory after being trained against an instrument the industry already trusts — that is a development story cold-chain operators can actually engage with.
Honest about the sensor
We do not over-claim the E-Array.
The E-Array is a multi-sensor metal-cystine and supporting-element platform — not a benchtop analyser. We deliberately do not claim absolute ppm accuracy from it on day one.
Ambitious about the model
Trajectory beats absolute readings.
For remediation decisioning, knowing the curve is climbing matters more than knowing the absolute value. That is what E-Sentinel is trained to recognise — and what a validated reference instrument can teach it to recognise.
Capex spent once
Trial-time, not product-time.
The expensive reference instrument is amortised across training trials only. From there, the intelligence cascades — to every E-Aegis in deployment, every E-MCP release node, every E-Nose Label sticker.
08 · IP
E-Sentinel sits inside the Ambient IoT patent portfolio.
E-Sentinel and the reference-monitor training methodology are protected within the Moseley Climacteric™ portfolio — Australian provisional patent applications filed by Ambient IoT Pty Ltd, sole inventor Tony Raftis, ABN 42 669 457 783.
E-SEN
E-SentinelOn-device ML for ethylene state & trajectory recognition
E-Sentinel training methodology, reference-monitor co-location protocol, paired-sample pipeline and on-device inference budget available under NDA. Get in touch and we’ll set up a conversation with the model team.
