AI and our lives [Part 1] Solved! Energy cost halved by wisely melting snow

Artificial intelligence (AI), which is one of the most advanced technologies, has already entered our daily lives. I asked Professor Hidenori Kawamura of the Graduate School of Information Science and Technology, Hokkaido University to explain how useful it is and how far it can be expected in the future. First is a road heating device that solves the problem of snow accumulation on roads in snowy countries. It has been demonstrated that AI can become smarter by learning the snow condition on the road surface, control wasteful operation of the boiler and cut energy costs in half.

The summer in Hokkaido is short. As the night breeze cools past the tray, winter preparations spill over. It is a habit of people who live in northern countries that are locked in snow for months. Snow falls mercilessly without considering the convenience of people. If you don't clear the snow The trouble is the parking lot without roof. Work to dig up a car buried in snow is waiting. Muscle pain is an inevitable heavy labor. It is said that manual snow removal uses almost the same exercise intensity as volleyball and swimming, and uses 6 times more energy than at rest.

When it comes to large-scale sites, it's unmanageable. Snow on roads, public facilities, commercial facilities and hospitals is often removed by bulldozers. A snow melting device may be introduced. I am grateful that the snow will melt from the end that fell on the road and prevent it from accumulating without bothering people. However, fuel costs are high.

The "AI Road Heating Optimizer" is the result of joint research between Hokkaido University and Hokkaido Gas, which is a groundbreaking snow melting system that solves troublesome cost problems. AI is a device that optimizes snow melting by determining the snow condition on the road surface. It says that energy is not wasted and costs are reduced.

Judgment of snow on the road surface by image recognition

"The principle of snow melting is simple, and in the case of the hot water type, boiling water is heated with a boiler to warm the road surface and melt the snow. Therefore, the snow condition of the road surface is judged by image recognition, and if there is snow, the boiler is turned on. "We have developed a device that turns it on and turns it off when the snow melts. That's the AI ​​Road Heating Optimizer," explains Professor Kawamura.

The epoch-making thing is to judge whether there is snow on the road or not. The conventional snow melting device controls the boiler by checking "whether or not it is snowing". This difference is significant. "The conventional one moves the boiler when the snow sensor detects snow. At this time, the condition of the road surface is unknown. Whether or not there is snow, it will try to melt it with full force. "We often set a long timer to avoid things that don't exist, which means that the boiler runs endlessly after the snow stops, wasting energy," says Kawamura. Waste of gas, oil, and electricity is reflected in fuel costs.

It was Hokkaido Gas that wanted to eliminate waste of energy and fuel costs. Locally known as "Kitagasu," it is a comprehensive energy company that supplies natural gas and electricity. About four years ago, we have been developing an optimizer in collaboration with Professor Kawamura's laboratory.

When I was working on another collaborative research, the topic of "energy saving of snow melting" was talked about in the chat. Professor Kawamura said, "I also had a snow melting device at home, and I was always worried about excessive operation, so I suggested that I try it because it can be solved with image recognition technology."

Immediately, I made a prototype by attaching a camera to the ultra-small computer "Raspberry Pi". This is so-called AI (Artificial Intelligence). Then, I collected a large amount of images of road surfaces in various states and learned what kind of state "with/without snow" is. The method used at this time is "deep learning," which supports the remarkable evolution of AI. It is said that if AI is made to repeatedly learn, it will be possible to determine if there is snow or not. “In image recognition that determines whether something is present or not, it's already comparable to people. I thought that if people could see that snow was piled up, AI would know,” recalls Professor Kawamura.

"Deep learning" imitating the human brain

AI also has its strengths and weaknesses. The more and more people talked about taking or not taking jobs for other people, the more I became able to do things like people. Can recognize both natural language (a language used by humans as opposed to programming language) and voice. What he is particularly good at is image recognition.

In this field, “The movement to replace the work that people are doing now with AI will continue,” says Professor Kawamura. That doesn't mean that AI takes away people's work. It may be the division of labor or cooperation between AI and people. It can be seen from the fact that many of the issues facing companies brought into Professor Kawamura's laboratory are aimed at eliminating labor shortages and overwork. The challenge of snow melting is the reduction of energy and fuel costs, but also due to the limits of manual snow removal.

Education & Training

By the way, it is said that the academic field was established in the 1950s when the word AI was used. Research has been repeated until the present. "Deep learning" supports the remarkable development in recent years. So what is deep learning? Roughly speaking, it is a technology that AI learns by itself. Professor Kawamura taught me the mechanism in detail.

Deep learning has an underlying mathematical model. It is called a "neural network" and is a mathematical expression of the mechanism of the human brain. Here, please remember how the brain processes information. The information that a person sees becomes an electrical signal in the brain and is transmitted to a myriad of neurons (nerve cells). The synapse is responsible for this transmission.

What is important is that the information is weighted instead of flowing uniformly. When information is considered important, synapses increase in number, strengthen ties, and transmit efficiently. The change in synapse due to weighting is called "synaptic plasticity" and is considered to be a mechanism of memory and learning. In a neural network that reproduces this series of operations on a computer, those that correspond to nerve cells are called "nodes," and those that correspond to synapses are called "edges."

Now, let's look at the information processing of the neural network with a calculation example. First, enter the numbers 2, 3 and 5 for each of the three nodes. This number is weighted by the edge and transmitted to the next. Assuming that 2 is weighted 100 times, 3 is weighted 10 times, and 5 is weighted 1 times, the number 2/3/5 changes to the number 235. The next node receiving it concludes that 235 is the answer. The node that first inputs a number is called the “input layer”, and the node for which the calculation result appears is called the “output layer”.

Actually, there is a "hidden layer" (intermediate layer) between the input layer and the output layer, and the answer is calculated by multiplying the multiplication and the addition. Deep learning is a neural network in which hidden layers are stacked in multiple layers, that is, deepened. Although omitted in the calculation example, the weighted number 235 is calculated by the “activation function” and becomes the final answer.

Demonstration test in parking lot, accuracy is over 98%

How will AI learn using deep learning? When compared to humans, it is an image of solving problems by self-study. Repeat the steps of solving the problem, matching the answers, correcting the mistakes... to increase the correct answer rate. In the calculation example in the previous section, when the answer of 235 is incorrect, it is considered that the weighting was wrong. Then correct and recalculate. One of the learning methods of AI is to reduce the number of mistakes by doing many problem books by yourself. At this time, humans do not help. Give only the instruction "solve the 10,000 problems prepared here. Make the incorrect answers as few as possible" and let AI do the rest.

Let's take a look at AI image recognition methods. There is an image called handwritten 1. To recognize this as the number 1, do the following: First, the image is finely decomposed, and the white part is digitized as "0" and the black part as "1". Next, input the numerical value into the node and calculate it in a hidden layer in which many layers are connected. From the combination of "0" and "1", the characteristics of the image are judged, and which number is applicable is determined. When learning images of various numbers, AI repeats feedback even if it makes a mistake and finds a weighting with a high probability of being 1. Learning is complete when appropriate weighting with few incorrect answers can be done.

He said that he learned the same way with the image recognition of "with/without snow". "First, the number of snowflakes in a single image is digitized as "1" and the non-existence as "0". Then, when the snow percentage exceeded 50%, I trained until I was able to judge that there was snow," says Kawamura. The images used amount to 24,000. AI has completed learning and has conducted repeated field tests in parking lots, but there are few mistakes in judgment. The precision is indeed over 98%.

It used to be difficult for AI to distinguish snow from white and road from black. Asphalt that has become whitish over time and the road surface that glows white when illuminated by streetlights is mistaken for snow. The ability to improve it largely depends on deep learning, and it has been possible to improve the processing capacity of the computer that has enabled deep learning. Now that the accuracy of image recognition has improved, AI has caught up with human vision.

Energy saving and friendly to the environment

Let's check the snow conditions in Sapporo City, where Hokkaido University is located. The cumulative amount of snowfall in one winter is 597 cm (normal value). It is inconspicuous in the city, but it accumulates in its own way. The cost of snow countermeasures in FY2019 is 21,513 million yen. Among them, the budget related to snow melting exceeds 1.6 billion yen. When spring comes, the snow, which melts naturally, will be used for the cost of this to maintain the lives of citizens.

The AI ​​road heating optimizer has the potential to reduce fuel costs for snow melting equipment. According to Professor Kawamura, "Waste of energy occurs when the boiler is stopped and in the early spring. Conventional snowfall sensors cannot determine when to stop. In addition, the snow that falls from March to the beginning of April does not accumulate. It often melts, but we can't even make that decision, so we can't control the boiler well and waste energy. We have succeeded in reducing energy consumption by up to 50% or more in demonstration experiments." That. In large parking lots such as supermarkets, hospitals, and condominiums, costs can be reduced by about 50 to 1 million yen per winter. It is expected to be commercialized next winter.

For the time being, "The cost of one unit is high because it has a function that can be operated remotely when a problem occurs. If more data can be collected and the accuracy of image recognition will be improved by putting it into practical use, I think we can reduce the cost by removing the communication function. We want to be able to introduce it at home in the future," he said. I am happy that it is energy-saving and friendly to the environment. It seems that winter is nearing when snow-melting equipment can be used at a running cost of tens of thousands of yen.