Energy efficiency raises significant concerns as it is one of the most promising ways to mitigate climate change. Disaggregation and identification of individual electrical appliances activities is one of the essentials for energy preservation especially for smart buildings. This paper proposes a lightweight electrical appliance activity detection approach for smart building, which leverages a single smart metering device to establish a learning and detection processing for multiple appliances. In this system, data interpolation and transition detection algorithm are proposed to effectively reduce the cost of model training and optimize the detection accuracy. The concept of appliance fingerprint is proposed and a variety of fingerprints, including appliance-based and context-based, are defined to depict fine-grained appliance characteristics. Based on these fingerprints, the paper proposes a multi-source fingerprint-weighting KNN (FWKNN) classification algorithm and presents a boosting framework for continuous online learning and detection. A prototype system is implemented and demonstrated in IBM Bluemix PaaS cloud platform. Experimental result and analysis prove that FWKNN outperforms other benchmark methods in detection accuracy.
14 December 2021 in IEEE Transactions on Parallel and Distributed Systems
Event processing is one of the cornerstones to manage massive data streams in Cyber-Physical Systems (CPS). Due to CPS applications' increasing complexity, detecting highly complicated events (aka. expensive events) leads to significant performance degradation, particularly harmful to mission-critical systems. To tackle this challenge, we define a new task - dynamic event processing offloading to CPS-edges. This paper proves the problem NP-hard and proposes a solution DePo. DePo splits the expensive events into sub-models and offloads them to CPS edges. We design a long and short-term event memory mechanism in DePo that enables the edges and server to process expensive events collaboratively within their capabilities. Besides, we propose a concept called Edge Utility to measure the optimality of offloading schemes. A heuristic algorithm is presented in this study to guide how to dispatch events to edges, thereby helping DePo generate a sub-optimal solution in polynomial computational complexity. Our extensive experiments show that the performance gap between DePo and the optimal benchmark is less than 5%. DePo effectively reduces more than 40% redundant states and provides over 100% higher throughput than state-of-the-art approaches. Experimental results verified the high stability and scalability of DePo, especially when dealing with a large number of expensive events.
Reusing separate, pre-filled Key-Value (KV) Caches for multiple contexts has become a common practice in handling multi-context scenarios with Large Language Models. However, this leads to a lack of cross-attention mechanisms between contexts. To address this, we propose CatLLM, the first method that concatenates multiple contexts across requests offline to compensate for this deficiency. Specifically, during offline processing, CatLLM identifies contexts that severely lack cross-attention by incorporating the weighted inner products of Q and K vectors from tokens in an un-concatenated context into an equivalently transformed weighted formulation for concatenated Q and K inner products. This yields a weighting corresponding to the output vector difference, which can then be used to identify contexts with severe cross-attention deficiencies and concatenate them into a single context for KV Cache computation. Experimental results show that, compared to the baseline of separate caching (i.e., no concatenation), fully concatenating all contexts improves the F1 score by 6%. Meanwhile, the proposed method reduces the number of contexts requiring caching from 10 to 7 while achieving a 3% F1 score, thereby maximizing performance improvement while minimizing the degree of context compression.