"Ubiquitous IoT has been painstaking deployment and operating costs for a long time. "
With the increasing level of digitalization and transformation needs of various industries, data and high-speed network deployment have become "heavy" operational assets. Massive terminal data is transmitted to the cloud, and more and more resources and power are invaded. Lightweight and low-cost deployment have become the desire of many enterprises. And this "suffering" may be truly resolved with the development of technologies such as micro machine learning (hereinafter referred to as TinyML ) and LPWAN.
Based on the obvious advantages of terminal computing in terms of energy consumption and efficiency, especially in the context of the Internet of Everything IoT era, TinyML has been described as "hot" in the past two years. Recently, the new generation of LPWAN technology ZETA launched the "ZETA+TinyML" end intelligent series of products, which are accelerating the application of industrial predictive maintenance, digital logistics, smart cities and other fields. With the incremental OTA remote upgrade technology customized for the LPWAN network, the deployment and operating costs of IoT software are further reduced. ZETA+TinyML's end-to-end intelligent application may solve the problem of IoT deployment and software upgrade from the root.
01
What is ZETA+TinyML?
As we all know, LPWAN is a long-distance, low-power, narrow-bandwidth wireless communication technology that solves the problems of high terminal power consumption, massive terminal connections, insufficient wide-area coverage and high cost in the Internet of Things industry, and is suitable for large-scale deployment. But even so, the application of the Internet of Things is still a problem. Take the industrial scene as an example. There are many industrial equipments and the environment is extremely complex. The penetration of technology and the convenience of deployment have become rigid demands. Therefore, as a new generation of LPWAN technology ZETA launched the "LPWAN2.0 Ubiquitous Internet of Things", which aims to accelerate the application of the Internet of Things in cost-sensitive industries such as industry and logistics with higher performance, stronger penetration, and lower cost.
At the same time, ZETA also aims to limit the application of the Internet of Things technology to the pain point of "high-bandwidth data analysis is not timely". According to IDC's latest research report, 68% of data is wasted in industrial scenarios. For example, vibration monitoring and image recognition applications. These scenarios generally use high-speed communications such as 5G and wifi to return the original waveforms and images. While network costs have soared, they also rely heavily on background algorithm development and secondary analysis. TinyML has built-in micro-machine learning technology in the end-test. While meeting the increasingly urgent need for real-time analysis, it combines ZETA's low-cost and strong penetration advantages to create "end-intelligence" applications, which may break through the limitations of traditional IoT technology architecture and expand more LPWAN application scenarios.
TinyML is giving hundreds of millions of edge devices the ability to process data in real time, thereby getting rid of the dependence on gateways and computing centers. Data-centric nodes will become "end-intelligence" units, integrating data acquisition, signal processing, model reasoning, and result transmission, using low-cost, low-power, and more penetrating LPWAN communication technology as a whole The carrier makes the intelligent ubiquitous Internet of Things ubiquitous. The integration of TinyML and ZETA LPWAN technology will greatly broaden the original application scenarios of LPWAN and bring huge imagination to the road to AIoT upgrade in the Internet of Things market.
02
ZETA incremental upgrade helps TinyML ubiquitous deployment
Most IoT devices may be just an MCU, which does not have enough storage resources like mobile phones. The "deep compression" technology has made the TinyML model "slim down", but it is far from enough.
The TinyML model is not once and for all after training, and it is necessary to update it on demand, which is mainly reflected in the following aspects:
The fragmented application scenarios of the Internet of Things lead to serious differentiation. Even for tasks of the same type, it is necessary to retrain the model or adjust the model parameters.
Data accumulation is needed to improve model accuracy. The TinyML model is generally trained by a neural network algorithm driven by pure data. The accuracy of the model is positively related to the amount of data. As the data continues to accumulate, the model also has a need for regular iteration.
The need for relearning brought about by the change of state. For example, if the vibration level of factory equipment changes after overhaul, it is necessary to re-collect a period of data for learning and generate a new model benchmark.
Upgrade needs due to changes in task output. For typical tasks such as classification model, adding the output of a category requires re-collecting the "gene data" of that category.
Incremental OTA (Over-the-Air, over-the-air download technology) is also called differential remote upgrade technology, which can solve the problem of TinyML remote upgrade. Compared with the traditional whole package upgrade and compressed package upgrade method, incremental upgrade can reduce the model upgrade cost to the greatest extent, including the memory cost in the space dimension and the time cost in the time dimension.
The key to incremental upgrade is the file difference algorithm. The file difference algorithm takes the source data and the target data as input, extracts the common part, and packs the remaining part in the target file into patch package difference data as the output. If the difference between the two files is small, data redundancy can be greatly reduced, saving data storage space and transmission bandwidth. bsdiff is currently the most commonly used differential algorithm, with high compression efficiency and good algorithm stability.
The bsdiff differential uses the bzip2 compression algorithm by default. Although its compression efficiency is high, it also needs to solve the minimum memory of 400KB on the end side, so it is not suitable for resource-constrained embedded chips. In order to adapt to more IoT terminals, ZETA self-developed high-efficiency compression algorithms, taking into account compression efficiency and platform compatibility, and using the feature of bsdiff differential package 0 to simplify compression-only 0 compression. Taking the TinyML model with 100kB and 10% difference, the final differential packet size will be below 11kB, and only 10kB of additional memory space is required, so that it can be used on all IoT chips.
Incremental upgrade technology combined with ZETA's low power consumption and wide coverage communication advantages will further connect data links such as transmission, processing, and upgrades in series to solve the industry's pain points such as fragmented application scenarios and high online upgrade costs of the Internet of Things, and create end-to-end Data value-added services.
03
ZETA+TinyML end intelligent application accelerates landing
ZETA "end computing" smart sensor, through the "artificial intelligence compression" of vibration, sound, image and other signals on the sensor side (that is, relying on mechanism + deep learning/machine learning feature extraction and event detection), and fusing TinyML technology to further realize " Ubiquitous deployment of "machine intelligence". At present, the ZETA-end intelligent series of sensors are gradually being implemented in scenarios such as predictive maintenance of industrial equipment, smart buildings, and urban management.
ZETA端智能振温传感器
目前的设备故障预警和诊断极度依赖平台大数据算法和后台专家分析的方式，系统整体实施成本居高不下：不仅仅是传感器成本，还包括高昂的网络部署、算法二次开发和人力成本，很大程度上造成设备预测性维护难以大规模落地，渐成行业“理想”。
ZETA端智能振温传感器将机理先验知识和TinyML技术融为一体，在端侧提取敏感度高的特征参数和运行推理模型（而非原始波形），并通过低功耗、强穿透的ZETA网络实现大范围覆盖和高效传输。在部署阶段有个短暂的“学习期”，由平台自动训练生成异常检测和故障识别的TinyML模型，并通过ZETA server的增量升级技术进行批量远程部署。报警和诊断完全在端侧执行，保证了数据的时效性，并节省了网络带宽占用和电量消耗。在典型工业环境下，单网关接入容量300+，电池寿命3年以上。
ZETA端智能振温传感器是LPWAN通信技术和TinyML算法深度结合的一次创新，由于振动信号的通用性，该产品构架也可拓展为其他横向场景，诸如结构安全评估、可穿戴设备、运动和健康监测等，继续丰富ZETA产品生态圈和泛在物联的价值圈。
ZETA端智能幕墙检测器
幕墙玻璃检测器是ZETA端智能的另一个TinyML应用示例。图像识别玻璃幕墙是否破碎是最直接的方式，但是实际应用中存在成本高昂、布线不便和隐私泄露等痛点。ZETA端智能幕墙检测器集成低成本摄像模组，将图像识别的算法“压缩”成TinyML模型植入模组，仅1/3信用卡大小，占用内存仅几十kB，可由电池供电长期运行。图像识别的结果由ZETag协议回传，实现功耗更低的通信传输，一个网关甚至可以覆盖附近的2~3栋大楼。
端智能的图像识别TinyML应用还有许多场景可扩展，比如数字抄表、工厂异常情况检测、城市安防监控等等，摄像头将不再是依赖于有线供电和填满硬盘的一个仪器，而是具备“自我唤醒”意识和“自主辨识“能力的守望者。
As a new generation of LPWAN technology, through deep integration with TinyML technology, ZETA "end intelligence" will connect in series a complete data chain of transmission, application and upgrade to achieve a lower cost, lower power consumption and smarter network . It is applied to all aspects of social fields such as smart cities, industrial Internet, and smart agriculture, and its vision of creating a transparent, efficient, and ubiquitous LPWAN2.0 ubiquitous IoT ecosystem is just around the corner.