What you’ll learn
- 边缘计算的概念及其如何使IIOT受益。
- Leveraging machine learning in edge computing.
Lots of the talk surrounding edge computing is in the context of industrial IoT, so first let’s define that term. Essentially, edge computing means that the data collected by a sensor or a group of sensors is processed at the source without the data being transferred to the central computing or cloud resource. In some cases, the processed data (a much smaller dataset than the raw data) would be passed onto the cloud for historical record-keeping or further processing such as trend analysis.
Let’s clarify with an example. Think about an industrial camera used in a manufacturing environment for inspecting parts. Such a camera would combine a CMOS sensor with a processor and software to inspect the product and make a quality determination at the edge without having to transfer the data to the cloud. In a bottling plant, for instance, a smart edge device (in this case a camera) could read labels, verify if the bottle is full, and maybe even check the seal—all while the bottles are moving on the assembly line.
在PCB制造的屁股embly line, this type of camera can flag gross routing errors and other potential defects. Oftentimes, these cameras would be hooked up to a local industrial PC (to store historical data) but would not be connected to the internet to minimize security risks（见图）.
This is classic edge computing, where the data can and must be processed locally (for latency) and the processing capability can be incorporated in a small form-factor device that’s deployed on the actual assembly line (i.e,. at the edge of the industrial network).
According to CB Insights, “Edge computing enables data to be processed closer to where it’s created (i.e., motors, pumps, generators, or other sensors), reducing the need to transfer data back and forth between the cloud.”4
Benefits of Edge Computing
延迟定义为数据从传感器传输到处理器所需的时间。在许多工业用例中，重要的总延迟（包括处理和采取关键措施的数据返回）很重要。考虑在焊接机周围的电子屏障（也称为浅窗帘）的情况。如果有人违反障碍物，则需要关闭焊接。另外，当屏障X毫米内发生某些东西时，必须发出警报以指示危险。
在这种用例中，没有理由将灯罩通过网络生成的数据发送。本地处理器可以处理漏洞数据和接近性数据并触发适当的响应。某些数据可以存储在内存中，并在预定的时间发送到云进行趋势分析。
Having the processor local within the edge device not only minimizes latency, but also guarantees deterministic latency. Data that travels over Ethernet to a processor that will schedule the data processing at an opportune time doesn’t work in a factory or process industry setting, where latency must be deterministic at all times.
将自动驾驶汽车视为最终的边缘设备。必须在一定量的毫秒内处理车辆过程的视觉和雷达数据必须在短的，确定性的时间窗口内。这种数据不适合云处理。
Michael Clegg, a vice president at Supermicro,2explained this in a very interesting way: “By processing incoming data at the edge, less information needs to be sent to the cloud and back. This also significantly reduces processing latency. A good analogy would be a popular pizza restaurant that opens smaller branches in more neighborhoods, since a pie baked at the main location would get cold on its way to a distant customer.”
最小化数据传输
To understand how much data we are talking about, this is what an IDC whitepaper states: “By 2025, 175 zettabytes (or 175 trillion gigabytes) of data will be generated around the globe. Edge devices will create more than 90 zettabytes of that data.”3Of course, this is for all data, not just in the industrial setting. Still, it underscores the magnitude of the potential data-transfer problem.
For an example of both latency and minimizing the data transfer, consider a vibration sensor on a motor. The frequency of vibration may be a good indicator of impending failure and can even flag an immediate problem that may necessitate a shutdown or alarm. The vibration data collected may be huge, and not all of it is worthwhile to store or analyze.
振动传感器可以具有局部处理器，例如在振动数据上进行快速傅立叶变换（FFT），并立即标记频率，这可能需要关闭，在设定的时间窗口内进行维护或某些纠正措施。通过在本地进行处理，仅需要将相关数据传输到云才进行离线分析。
While the summarized data can be sent to the cloud for later analysis, the wrong frequency must be flagged immediately by the vibration sensor to avoid any serious damage. Thus, the latency limitation.
A case study done by IHS Markit on a Duke Energy industrial IoT system notes: “To collect vibration information, it can be necessary to capture anywhere from 10,000 to 100,000 samples per second for several seconds to obtain a good measure of the machine condition.”1This data must be processed in real-time using edge-processing systems, which can then narrow down the data to that pertaining to the health status of the machine.”
Security
对于基于边缘的计算，这不是一个干净的好处。虽然开发空调，身体安全的边缘设备的安全风险较低，但与Internet连接或折衷的边缘设备可以大大增加攻击面。
众所周知，几乎可以将连接到网络连接的任何内容都可以入侵。虽然失去机密数据肯定是令人困惑的，但想象一下是否打开或关闭阀门或操作警报的边缘设备被黑客入侵以可能不安全的方式操作。
在工厂环境中，边缘设备在物理安全的环境中。如果从边缘设备到连接的工业PC的唯一通信通道是用于单向数据转储（例如，当天的振动数据频率分析流），则系统已得到充分保护。
However, if the edge devices can be controlled or reconfigured from the industrial PC or a microPLC, then the security depends on how well that system is protected against attacks. Many edge devices will have some kind of remote access and management capability. Or, an intelligent sensor in, say, an oil pipeline pressure-monitoring application is remote enough to open the door to physical tampering. In either of those cases, an intelligent edge-computing device will need to have standard security protocols built-in either via software or hardware.
制造的边缘计算趋势
Leveraging Machine Learning
Soon we will see machine-learning algorithms brought to bear on the software controlling edge devices. In the beginning, a large amount of visual data (going back to the industrial camera example in the introduction) for both good and bad parts can be fed into a cloud-based machine-learning system to train the neural network. The training data would be used to develop a good/bad classification algorithm. Typically, while the training algorithm requires massive computing resources, the inferred algorithms are smaller and can be run on a low-power microprocessor.
The only hitch in developing a deep-neural-network training algorithm for industrial applications is the availability of a large enough training dataset. For example, more than 350 million photos are uploaded every day on Facebook. That’s a huge training dataset. In any industrial setting, the training dataset is a small fraction of what’s available in the consumer world.
It remains to be seen how effective machine learning will be in creating inference algorithms for industrial systems. However, we will definitely see edge-computing devices leveraging the new machine-learning technologies in the near future.
Separating Useful Data from Just Data
As we start collecting terabytes, petabytes, and even more levels of data, it’s becoming clear that not all data is useful to be transmitted, stored, and analyzed. In the industrial setting, there’s a cost to keeping the data and running algorithms to analyze and correlate this humungous dataset to a specific interesting event (e.g., failure or energy efficiency).
We know from experience that some datasets correlate to what we’re interested in: predicting failure via vibration frequency, predicting optimum energy usage via current and phase, predicting optimum performance of a motor via variations in speed, to give a few examples. These select datasets can be stored and even transmitted to the cloud for further analysis and to develop optimum inference algorithms.
一些数据集可以被视为低价值for either prediction or optimization and may be discarded at the edge once the necessary action is taken.
Rise of Predictive Maintenance
一个有趣的预测维护案例研究是一家名为PING Monitor的澳大利亚初创公司，该初创企业已开发出一种声 - 涡轮监测系统。系统位于风力涡轮机的底部，并测量涡轮叶片旋转时产生的声音。利用这些声学，该公司具有基于AI的系统，可以预测风力涡轮机的任何故障或异常行为。这样的边缘设备节省了运营成本和检查成本。通过提前安排维护，它还可以优化能源生产。5
While predictive maintenance was always touted as the killer app for the industrial IoT, so far it has failed to live up to its promise. As edge devices grow in sophistication are able to incorporate more processing and sensing components, they’ll likely play an increasing role in finally giving us the elusive goal of predictive maintenance.
Conclusion
Edge computing is indeed the latest trend that everyone is talking about, but this is more than just a passing fad. This type of distributed computing—at the edge and in the cloud when it makes sense—is here to stay. Some data, especially in the industrial setting, is best processed and acted upon locally. Equally, some datasets are best analyzed in conjunction with a vast amount of historical data.
Since it’s abundantly clear that not all smart devices need access to the cloud to make sense and act upon the data, the era of edge computing is here to stay and exist in parallel with cloud computing.
Suhel Dhanani is Director of Business Development, Industrial & Healthcare Business Unit, atMaxim Integrated.
参考：
1.https://technology.ihs.com/599357/industrial-iot-case-study-duke-energy-january-2018
4.https://www.cbinsights.com/research/what-is-edge-computing/
