Toward Foundation Models for Online Complex Event Detection in CPS-IoT: A Case Study
Liying Han, Gaofeng Dong, Xiaomin Ouyang, Lance Kaplan, Federico Cerutti, Mani Srivastava · 2025
Read Toward Foundation Models for Online Complex Event Detection in CPS-IoT: A Case Study by Liying Han, Gaofeng Dong, Xiaomin Ouyang, Lance Kaplan, Federico Cerutti, Mani Srivastava (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-12282.txt.
Introduction
Liying Han and co-authors test foundation models for spotting complex events in cyber-physical sensor systems. Complex events are chains of small atomic events linked across space and time. Sensors produce long, noisy streams, and most models handle only short spans.
The case study compares three routes to detection. One uses large language models directly on the streams. One trains neural networks to learn event rules from data. One joins neural models with symbolic engines that hold human knowledge. A state-space model called Mamba scored best overall. It also generalized best to longer unseen sensor traces. The results guide design of monitoring systems that reason over time.
Catalog record
| CREATOR | Liying Han, Gaofeng Dong, Xiaomin Ouyang, Lance Kaplan, Federico Cerutti, Mani Srivastava |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-12282
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Toward Foundation Models for Online Complex Event Detection in CPS-IoT: A Case Study Liying Han, Gaofeng Dong, Xiaomin Ouyang, Lance Kaplan, Federico Cerutti, Mani Srivastava Abstract . Complex events ( CE s) play a crucial role in CPS-IoT applications, enabling high-level decision-making in domains such as smart monitoring and autonomous systems. However, most existing models focus on short-span perception tasks, lacking the long-term reasoning required for CE detection. CE s consist of sequences of short-time atomic events ( AE s) governed by spatiotemporal dependencies. Detecting them is difficult due to long, noisy sensor data and the challenge of filtering out irrelevant AE s while capturing meaningful patterns. This work explores CE detection as a case study for CPS-IoT foundation models capable of long-term reasoning. We evaluate three approaches: (1) leveraging large language models (LLMs), (2) employing various neural architectures that learn CE rules from data, and (3) adopting a neurosymbolic approach that integrates neural models with symbolic engines embedding human knowledge. Our results show that the state-space model, Mamba, which belongs to the second category, outperforms all methods in accuracy and generalization to longer, unseen sensor traces. These findings suggest that state-space models could be a strong backbone for CPS-IoT foundation models for long-span reasoning tasks. Conference: The 2nd International Workshop on Foundation Models for Cyber-Physical Systems & Internet of Things ; May 6–9, 2025; Irvine, CA, USA The 2nd International Workshop on Foundation Models for Cyber-Physical Systems & Internet of Things (FMSys ’25), May 6–9, 2025, Irvine, CA, USA DOI: 10.1145/3722565.3727198 ISBN: /2025/05 CCS: Computing methodologies Machine learning CCS: Computer systems organization Embedded and cyber-physical systems CCS: Artificial intelligence Knowledge representation and reasoning CCS: Artificial intelligence Hybrid symbolic-numeric methods † † cc-license: by Figure 1. (a) An intelligent smartwatch assistant monitors workplace hygiene and alerts users of potential violations. (b) In a smart facility, a surveillance system detects suspicious activities, such as unusual parcel hand-offs, using distributed cameras. (c) An illustration of the online CED task showing sensor processing and CE labeling. 1. Introduction Current CPS-IoT applications excel in short-span perception tasks like human activity recognition and object detection, which require only a few seconds of sensor data. However, real-world scenarios often demand high-level contextual understanding over long periods for complex decision-making, a capability largely overlooked. For instance, an elderly person wearing a smartwatch for continuous health monitoring generates IMU data over hours. An 8-hour recording at 100 Hz produces 2.88 million data points, making it crucial to efficiently compress and retain key information to identify long-term, meaningful patterns related to health conditions. To effectively capture these long-term patterns, it is useful to represent them as Complex Events ( CE s), which consist of sequences of short-time Atomic Events ( AE s) governed by spatiotemporal dependencies. They are crucial for decision-making in CPS-IoT applications such as smart monitoring and autonomous systems that require long-term contextual understanding, as illustrated in Fig. 1 (a) and (b). Detecting CE s is challenging for three reasons. First, the model must identify relevant AE s while ignoring irrelevant ones, increasing sequence-matching complexity. For instance, a sanitary protocol can be represented as “Use restroom → X → Wash hands → X → Eat,” where “X” includes irrelevant AEs like “walking” or “sitting,” complicating detection. Second, CE s involve long-term dependencies with varying temporal gaps, making it difficult to model temporal and logical rules accurately. Third, many CPS-IoT applications require online detection for real-time responses to safety-critical events, necessitating efficient and accurate inference. Two main approaches exist for CE detection. Data-driven methods use neural networks to learn patterns from data but struggle with long-term dependencies due to fixed context windows or inefficient memory mechanisms. For example, Recurrent Neural Networks (RNNs) and Transformers capture context through sequential processing and attention mechanisms but face memory fading and computational inefficiency over long durations. Recent advances in state-space models (SSMs), such as Mamba ( Gu et al., 2022 ; Dao and Gu, 2024 ; Gu and Dao, 2024 ) , show promise by modeling event progression and capturing temporal dependencies. However, their effectiveness in noisy, high-dimensional sensor data for CE detection remains unexplored. Neurosymbolic frameworks, as the second approach, combine neural networks with symbolic reasoning to leverage predefined rules for CE patterns, reducing the need for extensive labeled data. However, symbolic engines are sensitive to noise, and probabilistic reasoning engines ( Manhaeve et al., 2018 ; Li et al., 2023 ; Yang et al., 2023 ; Skryagin et al., 2021 ) , although more robust through probabilistic programming, require complex temporal rule implementation and demand specialized expertise. Large Language Models (LLMs) are emerging as competitive alternatives due to their large-scale pre-training and enhanced reasoning abilities, enabling them to model long-term dependencies. Recent works like ( Xu et al., 2024 ; Ouyang and Srivastava, 2024 ) apply LLMs in sensor-based reasoning for CPS-IoT. However, their tasks do not address complex spatiotemporal dependencies or long temporal rule-based reasoning required for CE detection. This work explores online CE detection as a case study for building CPS-IoT foundation models capable of long-term reasoning, enabling improved situational awareness for downstream tasks. We investigate three approaches: (1) leveraging LLMs as CE detectors, (2) employing neural architectures that learn CE rules directly from data, and (3) integrating neural models with symbolic engines embedding human knowledge. Our evaluation shows that the state-space model, Mamba, from the second category, outperforms all methods in both accuracy and generalization to longer, unseen sensor traces. These findings highlight the potential of state-space models as robust backbones for CPS-IoT foundation models designed for long-term complex event detection. 2. Related Works Complex Event Detection (CED) has been extensively studied in traditional stream processing for structured databases. Early works ( Cugola and Margara, 2012 ; Schultz-Møller et al., 2009 ; Debar and Wespi, 2001 ) employed event engines such as Finite State Machines (FSMs) to identify complex events based on predefined patterns. More recently, research has shifted towards CED over unstructured, high-dimensional data. Approaches such as ( Xing et al., 2020 ; Roig Vilamala et al., 2023 ) integrate neural detectors with symbolic rule-based systems, enabling backpropagation through logical rules. However, these methods suffer from computational scalability issues. Moreover, most existing approaches focus on complex activities within short time spans ( Khan et al., 2023 ; Fabian Caba Heilbron and Niebles, 2015 ; Jiang et al., 2014 ) , limiting their applicability to longer sensor traces with extended temporal dependencies. State-Space Models (SSMs) capture long-span temporal dependencies through state transitions, maintaining state representations that implicitly encode temporal progression. Recent advances, such as Mamba ( Gu et al., 2022 ; Dao and Gu, 2024 ; Gu and Dao, 2024 ) , show superior performance in long-term contextual understanding by using selection mechanisms to process sequences more efficiently and adaptively model different temporal dependencies. However, SSMs remain underexplored in noisy, high-dimensional sensor environments, particularly for long-span temporal events in CPS-IoT applications. 3. Online Complex Event Detection 3.1. Complex Event Definitions Definition 3.1. Atomic events ( AE s) are short-duration, low-level events that serve as building blocks for complex events. They occur instantaneously or within a small time window and are directly detectable by models like image classification, object detection, or activity recognition. Definition 3.2. Complex events (