Tech

AI learns quickly to accurately ‘feel’ home appliances


Few Shot Learning AI “cảm nhận” chính xác đồ gia dụng

Researchers from the University of Johannesburg have developed a Few Shot Learning (FSL) algorithm to recognize the power signal of a home appliance. Credits: Therese van Wyk and Liston Matindife. Data from the study at Computational intelligence and neuroscience. (DOI: 10.1155/2022/2142935) creativecommons.org/licenses/by/4.0/

NIALM (Non Intrusive Equipment Load Monitoring) can “sense” devices using electrical energy. NIALM is used in homes and small buildings. For this, NIALM may require hundreds of labeled source signal images from each type of device for training. But there is a much faster and more cost effective approach than “traditional” machine learning.

Researchers from the University of Johannesburg have implemented the repeated learning (FSL) approach to NIALM. Classic FSL requires only 10 classified and labeled images to identify the device with very high accuracy.

They adjusted the process so that the AI ​​(artificial neural network) could choose the best training images for itself. This makes the training process even faster. When they also adjusted for some of the hyperparameters, just seven test images were enough for FSL to identify the devices with 97.83% accuracy. A training image (learning once) gives an accuracy of 88.17 to 91.343% depending on the number of layers during testing.

Their research is published in Computational intelligence and neuroscience.

FSL: When Machines Learn to Learn

Teaching AI to recognize a device’s source signal often requires a lot of data. Typically, an AI will need hundreds of images labeled by humans, to identify each type of device, despite different power and operating status. An example of two operating states for an appliance is the washing and spinning cycle of a washing machine.

All training data since AI has to be created and labeled by humans, this gets slow and expensive very quickly.

But there is another way for AI to learn, which actually requires very little labeled data. Just 10 labeled training images are enough to classify images with high accuracy.

For example, suppose an AI is trained this way with ten images each of an elephant, a tiger, and a bear. When the AI ​​is tested with an unlabelled image of a large male lion, the AI ​​recognizes that the lion resembles a tiger, but is not the same. Then the AI ​​will decide to create a new feature class for the lion.

Also, when that AI is confronted with an unlabelled image of a lion cub, it will be able to put that cub in the same class as the male lion.

This type of AI machine learning (ML) algorithm is called learning few times (FSL). It is a form of Meta-Learning, or “learning to learn.”

FSL has supported massive language models for globally dominating technology companies. Computer vision systems that check passports with travelers’ faces at some airports also use FSL.

Parts of a Cat

Professor Yanxia Sun from the Department of Electrical and Electronic Engineering Sciences, at UJ, said that learning less is really training an AI neural network with some data, even incomplete data about a particular object. object class. Sun is the lead author of the study.

“While we train a neural network With the training image, the AI ​​will learn the characteristics of each animal or object on its own.”

In the tiger versus lion example, FSL’s AI learns about a cat’s whiskers, eyes, fur, and tail from the tiger image. It had never seen a lion before. But when the AI ​​is tested with an image of a lion, it recognizes lions as a class of objects similar to, but not identical to, tigers.

NIALM: One power consumption meter for multiple devices

NIALM is used in small commercial buildings or indoors, to measure the amount of power consumed by each device or part of it.

NIALM uses power separation to separate the combined power consumption signals of multiple devices that are switched on simultaneously on the same power phase. NIALM uses only one measuring device. This is much easier and faster than having to connect a power meter to each device in turn.

Home smart meters in some countries store data about the power consumption of each device and send that data to the power companies. In other countries, smart meters also provide energy consumption data to homeowners.

Power consumption signal into digital image

In this study, the researchers trained their FSL AI on NIALM images of electrical load signals from a variety of home appliances.

They get the power consumption signal by plugging the power analyzer (Tektronix PA1000) and each in turn into a multi-plug extension power outlet. Then they turned on the energy analyzer. The device is then switched on and off, while the power analyzer records power consumption over time. For laptops and desktops, the entire boot sequence was recorded.

Few Shot Learning AI 'senses' correctly household items

Researchers from the University of Johannesburg tested the Few Shot Learning (FSL) algorithm to recognize the power signal of a home appliance. As the number of test images increased, the average accuracy increased from a minimum of 91.343% to a maximum of 97.83%. Credits: Liston Matindife and Therese van Wyk. Data from the study in Computational intelligence and neuroscience. (DOI: 10.1155/2022/2142935) creativecommons.org/licenses/by/4.0/

The power analyzer converted the device’s analog power consumption signal into digital data. That data is then converted into Gramian angle summation fields (GASFs), which look like brightly colored cells.

The 400 X 400 pixel color GASF images are then converted to grayscale and downsized to 28 X 28 pixels. This reduces the complexity of the algorithms and uses less computational resources.

Accelerate training

Classic FSL is a two-stage process—training and testing. In this study, the researchers added a tweak to speed up the selection of data that is very suitable for training, which will also speed up the training process.

Dr Liston Matindife, first author of the study, said: “We increased the accuracy of the FSL algorithm by performing an initial assessment of the ease or relevance of their data. me for metric learning. We call this the analogy test.”

During her studies, Matindife was a Ph.D. candidate at the University of Johannesburg. He currently teaches at the National University of Science and Technology in Zimbabwe.

“If the GASF image does not pass the similarity test, it means that the data needs to be preprocessed more, especially in time series or waveform format, before converting to an image. GASF. Images that pass the similarity test allow our model to learn faster.” Add Matindife.

Training and testing FSL AI

To train the FSL AI, the researchers fed it GASF images of 10 of the 14 types of devices in the study. They used 10 images for each type of device: Laptops on, laptops running MSWord, desktop computers, refrigerators, two-disc stoves, and a variety of low-energy lights.

They then tested FSL’s AI to see how well it learned to recognize or classify devices; and learn how to create new layers for devices it has never seen before.

For the test, they provided FSL AI images of 4 new layers, 10 images each: Video display laptop, microwave oven, kettle, and compact fluorescent lamp (CFL).

High accuracy with few images

In the case of laptop test images, the FSL algorithm was 97.83% accurate in classifying (recognizing) test images coming from a Laptopsignal consumes power, but is in a new operating state. This new working state is to show the video, not start or run MSWord as shown in the training images.

FSL AI achieves this accuracy with just seven test images to boot and another seven to run MS Word.

From studying several times to learning once

The researchers also varied the number of training images, then measured the classification accuracy of the algorithm. The test showed that as the number of test images per layer increased, the average accuracy increased from the minimum of 91.343% to the maximum of 97.83%. This shows that FSL can be applied in NIALM identification.

“The evolution of the NIALM algorithm is data scaling. For devices with different activation times, we will generate a different number of dataset images for each device. The imbalance Equal number of training images for different devices will often affect the training algorithm,” Matindife said.

“Our algorithm reduces the need for costly device-specific data collection. The prototype network’s use of the FSL algorithm allows for importing device datasets with varying numbers of sample images without encountering problems. difficult,” he added.

This study shows that an average accuracy of 90% can be achieved with only one training image per class, when using computer vision based prototypal Siamese and FSL algorithms applied to GASF graph. When a training image gives accurate enough results, it is called one-time learning. One-time learning, Matindife says, can solve a major challenge in NIALM—the large number of training images required.

See failure before it happens

Sun says this FSL NIALM AI can be used to identify high-value devices that are not working properly, such as computers or refrigerators where the mains or the compressor motor is starting slowly leading to failure.

An FSL AI can be trained on a few images of a well-functioning home refrigerator source signal. The first refrigerator is of brand A and capacity B.

Then in a house or a small commercial building, if that AI sees a power signal from a household refrigerator with a compressor motor problem from brand C and capacity D, it can signal that the second device has a power signal problem, she said.

More information:
L. Matindife et al., Learn several times to recognize non-intrusive device signals based on images, Computational intelligence and neuroscience (2022). DOI: 10.1155/2022/2142935

Provided by the University of Johannesburg

quote: Short-term learning AI accurately ‘senses’ home appliances (2022, Nov 21) retrieved Nov 21, 2022 from https://techxplore.com/news/2022-11-few- shot-ai-accurately-home-appliances.html

This document is the subject for the collection of authors. Other than any fair dealing for private learning or research purposes, no part may be reproduced without written permission. The content provided is for informational purposes only.

news7f

News7F: Update the world's latest breaking news online of the day, breaking news, politics, society today, international mainstream news .Updated news 24/7: Entertainment, Sports...at the World everyday world. Hot news, images, video clips that are updated quickly and reliably

Related Articles

Back to top button