Deep learning scheme based on CNO inference engine based on IoT equipment

Through deep learning technology, Internet of Things (IoT) devices can parse unstructured multimedia data and intelligently respond to user and environmental events, but with demanding performance and power requirements. The authors explored two ways to successfully integrate deep learning and low-power IoT devices.

In recent years, more and more IoT products have appeared on the market, collecting environmental data around them and using traditional machine learning techniques to understand the data. An example is Google's Nest thermostat, which records temperature data in a structured manner and uses algorithms to grasp the user's temperature preferences and schedules. However, it does not work for unstructured multimedia data, such as audio signals and visual images.

Emerging IoT devices use more sophisticated deep learning techniques to explore their environment through neural networks. For example, Amazon Echo can understand human voice commands, convert audio signals into word strings through speech recognition, and then use those words to search for relevant information. Recently, Microsoft's Windows IoT team released a security system based on facial recognition, which uses deep learning technology to automatically unlock the door when the user's face is recognized.

Deep learning applications on IoT devices often have demanding real-time requirements. For example, security cameras based on object recognition typically require detection delays of less than 500 milliseconds to capture and process target events in order to respond to strangers present in the home. Consumer-grade IoT devices typically use cloud services to provide some kind of intelligence, but the high-quality Internet connections they rely on are only available in some areas and often require higher costs, which raises the need for devices to meet real-time requirements. The challenge. In contrast, deep learning directly on IoT devices may be a better option to avoid the quality of the connection.

However, implementing deep learning directly on embedded devices is difficult. In fact, low power consumption is a major feature of mobile IoT devices, and this usually means limited computing power and small memory capacity. On the software side, to reduce memory footprint, applications typically run directly on bare metal or on a lightweight operating system that contains a very small number of third-party libraries. In contrast, deep learning means high performance computing with high power consumption. In addition, existing deep learning libraries often require the invocation of many third-party libraries that are difficult to migrate to IoT devices.

The most widely used neural network in deep learning tasks is Convolutional Neural Networks (CNNs), which convert unstructured image data into structured object tag data. In general, the workflow of CNNs is as follows: First, the convolution layer scans the input image to generate feature vectors; in the second step, the activation layer determines which feature vectors should be activated during image inference; the third step, using pooling The layer reduces the size of the feature vector; finally, all the output and output layers of the pooled layer are connected using a fully connected layer.

In this article, we will discuss how to implement deep learning on IoT devices using CNN inference engines.

Migrate services to the cloud

For low-power IoT devices, the question is whether there is a reliable solution for deploying deep learning in the cloud while meeting power and performance requirements. To answer this question, we implemented CNN-based object reasoning on an Nvidia Jetson TX1 device and compared its performance and power consumption with the migration of these services to the cloud.

To determine if we can reduce power consumption and meet the real-time requirements for object recognition tasks after migrating services to the cloud, we send the image to the cloud and wait for the cloud to return the results. Studies have shown that for object recognition tasks, the local execution consumes 7 W, while the power consumption is reduced to 2 W after migration to the cloud. This shows that migrating services to the cloud is indeed an effective way to reduce power consumption.

However, migrating to the cloud can result in a delay of at least 2 seconds, and possibly as much as 5 seconds, which does not meet our real-time requirements of 500ms. In addition, the violent delay of the delay makes the service very unreliable (for comparison, we run these experiments separately in the US and China for observation). Through these experiments, we conclude that in the current network environment, migrating real-time deep learning tasks to the cloud is an unworkable solution.

Porting deep learning platform to embedded devices

One option is to port existing deep learning platforms to IoT devices compared to the impractical migration to the cloud. To this end, we chose to transplant TesnsorFlow, a deep learning platform developed and open by Google, to build the bare-core ARM system-on-chip of the Internet of Things device Zuluko, PercepTIn, with object reasoning capabilities. Zuluko consists of four ARM v7 cores running at 1 GHz and 512 MB of RAM, with a peak power consumption of approximately 3W. According to our research, TensorFlow provides the best performance on ARM-Linux-based system-on-chip, which is why we chose it.

Data Center Power Distribution Monitoring System

Data Center Power Distribution Monitoring System is a multi-circuit measuring system by modular design. It measures 60 circuits for the incomer and feeder cabinet. Connected with the high accuracy hall sensor, the BCM201 Dc Data Center Power Monitoring System provides 0.2S Class measuring parameters such as fundamental parameters and harmonic data. A color 7" HIM with Ethernet interface makes the user easy to over view and control the system in real time.

Data Center Power Distribution Monitoring System,Data Center Dc Power Distribution System,Data Center Power Measuring System,Dc Data Center Power Monitoring System

Jiangsu Sfere Electric Co., Ltd , https://www.elecnova-global.com