設計可偵測時序性姿態變化之低耗能智慧感測系統

Abstract

隨著近年 IoT 裝置大量的應用於各類的地方，到 2024 年已經有約 200 億個 IoT 裝置，且這個數量每 6 到 7 年倍增的速度增長。 資料的收集與上傳會花費大量的電力資源。然而這類感測器通常使用電池作為供電並且僅有極少的運算能力，而仰賴智慧型手機或是個人電腦去收集偵測後的資料並從中提取資訊。這會需要耗費相當大的網路頻寬及能源用於傳遞資料串流。但是受限於有限的運算資源，要在感測器端去處理龐大的資料也非常的困難且會使得電池充電的頻率大幅上升，造成使用上的不便。 本論文中討論了不同計算平台與場景下會遇到的挑戰與解決方式。在使用多組感測器做定位系統時會面臨在感測邊緣時，感測器的定位不一致的誤差。使用類比的計算記憶體平台時，受限於輸出精度及非理想的誤差，模型的訓練會需要提出大量的修改。數位計算記憶體平台，提供了整數的矩陣乘法計算，與額外的非線性計算功能，可以用於完成模型層與層之間的計算需求，而需要微調模型來符合不同的計算模式。 本論文提出的日常健康監測系統中使用的極低解析度感應器去減少感測時的能耗，並且使用多種運算裝置分散其運算工作來達到低耗能的即時事件偵測系統。將資料於感測器端計算並僅上傳辨識結果可以大幅降低所需的網路頻寬，然而在缺乏跨感測器之間資料的情況下，辨識的結果容易有死角，而需要結合各別感測器的部分辨識依據來權衡各感測器結果的可信度與比重。低解析度感測器相較於高解析度感測器，裝置成本與耗電量都較低，在使用上若能搭配合適的人工智慧模型或是結合多個低解析度感測器，可以達到與直接使用高解析度感測器相近的結果。在部分情境下，需要從連續的資料中辨識動作變化的時序事件，此類事件由於需處理複數影像資料，所以需要較大的計算資源，而計算記憶體平台可以在感測器端提供低耗能的計算能力，但是需針對其不同的計算模式微調人工智慧模型。 在偵測步行速度時，在兩個感測器重疊的感測範圍可以減少步行速度的誤差至42.7%。在動作辨識的研究中，使用極低解析度感測器可以只使用高解析度感測器四成的成本與耗電量做到其97%相近的準確度。所提出使用數位計算記憶體平台的翻身偵測系統，相比於其他物聯網平台僅需6.3%的耗電量，且辨識結果有93.9%的近似度。

With the widespread use of Internet of Things (IoT) devices, their number reached 20 billion in 2024, doubling every six to seven years. Collecting and uploading data consumes significant energy and network bandwidth. However, these sensors are typically powered by batteries and employ smartphones or personal computers to process data. Continuously transmitting data quickly drains the battery. Data transmission requires a substantial network bandwidth when the number of sensors or their resolutions increase. However, processing high-resolution data on edge devices is challenging due to limited computational resources. Computing-in-memory (CIM) technology offers a highly energy-efficient computing platform for edge devices to process sensing data in real time. This work explores three computing platforms and optimizes energy consumption and sensing accuracy. When multiple sensors are employed in a location tracking system, inconsistencies in sensor-positioning errors arise at the sensing edge. Analog CIM offers a highly energy- and area-efficient computing platform but is constrained by its output precision and nonideal errors. Digital CIM offers an integer matrix-multiplication computing platform with nonlinear functions that perform calculations between model layers, requiring fine-tuning to accommodate various computational functions. This work proposes ultralow-resolution sensors with CIM for computation. The network bandwidth and energy consumption for data transfer can be significantly reduced by processing the data at the sensor. However, without fusing data from multiple nearby sensors, the results may have blind spots or inconsistencies between neighboring sensors. Therefore, the detection results and evidence could reduce errors. Low-resolution sensors have lower power consumption and cost than high-resolution sensors. Fusing low-resolution data and using suitable artificial intelligence (AI) models can realize performance similar to that of a high-resolution sensor but with better cost efficiency. When recognizing temporal events of posture change from continuous data, more computing resources are necessary to process the streaming data. The CIM platform provides highly energy-efficient computing resources; however, AI models must be fine-tuned based on the hardware design. Fusing the detection results on the overlapping area of two sensors reduced the gait velocity error to 42.7%. The accuracy of posture recognition using ultralow-resolution sensors achieved 97% of that of a high-resolution sensor at only 40% of the cost and energy consumption. The proposed turnover detection system with a digital CIM platform consumes only 6.3% of energy compared to other IoT platforms while achieving 93.9% accuracy.