Sensors driven by machine learning sniff-out gas leaks fast | Hacker News Reader

Sensors driven by machine learning sniff-out gas leaks fast

sciencedaily.com

38 pointsbennyspaceman15 comments

Someone has already implemented an evolved neural network to detect gas leaks. Individual nodes are communicative, autonomous, airborne and self-supporting across multiple environments.

>> In the 1960s, ornithologist Kenneth Stager of the Los Angeles County Museum of Natural History in California confirmed that turkey vultures were responding to ethanethiol. This chemical is a pungent byproduct of decaying organic matter, and energy companies use it to give odourless gas an easily identifiable smell.

http://www.bbc.com/earth/story/20150713-the-truth-about-vult...

This uses a simple shallow network. They don't seem to disclose the network architecture but claim it has 100s of weights (parameters), a hidden layer of 25 neurons and 3 neuron ouput layer (x location, y location, leak rate) and lastly we are told it takes minutes to train on a 2018 MacBook (so probably trained with the CPU).

It is trained on simulated data so it essentially creates a small inverse model of the gas leak simulation, and it appears to work on the real world. This shallow model could easily be deployed on the edge.

What does ML have to do in the problem space at all?

I admit, I know nothing of the industry. That being said:

You want to detect leaks? If the expected pressures in your pipes are low, there’s a leak. Other sensors exist to detect the chemicals. Put them in places where those chemicals shouldn’t be.

This reads like a press release designed for grant/funding purposes.

I too feel / wonder the same

Is this ‘just’ using a set of sensor data as the input to a classifier? Or am I missing something more complex (and if so, could someone explain it?). If it’s ‘just’ the former, then why hasn’t this been done before?

(‘just’ in quotes bc I know the word always reduces a whole world of complexity)

Simply put miniaturization of neural networks coupled with more powerful yet more energy efficient microcontrollers has allowed projects like this to be feasible since '19, maybe '18. In retrospect it seems an obvious next step from edge ML but at one point it looked like ML was only possible in the domain of beefy GPUs or power hungry CPUs on the cloud.

A cornerstone of embedded ML aka TinyML is that not sending data using comms decreases energy consumption by many magnitudes when compared to chip ops, SRAM use, sensor I/O etc.

What we have is the culmination of many techniques applied in concert. A huge one is float32 -> int8 quantization to compress the network 4x with usually only around ~1-2% drop in accuracy.

I for one am incredibly excited about the prospects of this nascent field.

To be clear this tool (ALFaLDS) isn't that. The model runs on a full-fledged system, or as full-fledged as a laptop might seem compared to a microcontroller. It integrates with many type of gas and wind sensors which are the dispersed and/or low powered piece.

Yes, I am aware from my brief read of the paper. I took the logical leap into touching upon embedded ML. I should point out that the current achievable complexity of embedded neural networks doesn't at the moment remove edge ML from being the method of choice where on-device inferencing isn't applicable. Edge ML is already alleviating much backhaul traffic otherwise generated by cloud inferencing. There are many scenarios where the latency introduced by transmission over network to cloud servers is prohibitive. Also where power contraints are not applicable nor remote operation absolutely vital (in the forest, as an animal tag, on oil platforms, etc) an edge hub to the sensor spokes is the way to approach a problem.

Understood. I guess I didn't make that leap because I find that embedded ML isn't what I imagined in this type of scenario. I would assume you'd invest more heavily in power usage to make the mesh sensor network fast/resilient.

A "base station" with connectivity for alerting and utility power at the edge of the sensor network is where I imagine any ML to run, and that could power your typical CUDA-capable GPU if necessary.

I guess that's still edge ML if the point is to not backhaul everything to a datacenter somewhere, but I wouldn't call it embedded.

That brought mind low powered devices with many sensors or sensors that capture more local data (i.e. microphones, cameras) vs. just an anemometer, GPS, and chemical ppm.

Slightly related question: How do you find and keep track of interesting papers like this one? Does anyone know of an RSS feed that I can subscribe to?

I tag it as #ml in my https://histre.com/

You can also connect your Hacker News account to Histre so your upvotes are saved as bookmarks, so you don't have to explicitly do it: https://histre.com/features/share-hackernews-upvotes/

Here are my upvotes: https://histre.com/pub/notebooks/o34gelgt/hacker-news/

I'm working on recommendations, and when that works well, you will be automatically shown links related to what you liked.

Describe "interesting" algorithmically so that it applies generally

Sidely related: using machine learning to detect different types of booze with a gas sensor: https://twitter.com/kartben/status/1258791793073815552

Kairos Aerospace, founded by some Palantir and Lockheed alums, also operates in this sector, combining Lockheed's hardware expertise with Palantir's chops for satellite surveillance. Check them out!

Ping me when they make this into a who farted device. :P

Reply on news.ycombinator.com