Sky Chenwei Wan

K\mathcal{Y}=\{y_{k}\}_{k=1}^{K}=\{(t_{\texttt{on}}^{(k)},t_{\texttt{off}}^{(k)},e^{(k)})\}_{k=1}^{K} (1) Objective. Minimizing the discrepancy between the predicted set 𝒴\mathcal{Y} and the ground truth 𝒴^\hat{\mathcal{Y}}, which requires: (1) Detection Accuracy: For any ground-truth y^j∈𝒴^\hat{y}_{j}\in\hat{\mathcal{Y}}, a matched prediction yi∈𝒴y_{i}\in\mathcal{Y} (e(i)=e^(j)e^{(i)}=\hat{e}^{(j)}) should reach a certain alignment threshold (e.g. temporal IoU) with y^j\hat{y}_{j}. (2) Localization Precision: For each matched event instance pair (yi,y^j)(y_{i},\hat{y}_{j}) under the objective (1), the deviation between the predicted temporal boundaries [ton(i),toff(i)][t_{\texttt{on}}^{(i)},t_{\texttt{off}}^{(i)}] and the ground truth [t^on(j),t^off(j)][\hat{t}_{\texttt{on}}^{(j)},\hat{t}_{\texttt{off}}^{(j)}] should be minimized. 3 Related Work 3.1 Time Series Symbolic Representation To achieve the goal of accurate event detection with explainability, the internal structures of events need to be explicitly modeled. Based on the example introduced in Section 1, we set the following desiderata for designing the logic structure: D1. Hierarchical Representation: the ability to capture hierarchical structures of events: the way atomic patterns recursively form sub-patterns, and finally the global pattern. D2. Semantic Quantification: the ability to quantify the coherence between the signal morphology and the semantics. D3. Topological Elasticity: the ability to define events based on the internal temporal-logic structure, agnostic to actual temporal durations: the representation must support “time warping” (elastic stretching or compressing of time) while maintaining logical validity. We analyze existing symbolic time series representation frameworks based on our desiderata (summarized in Table 1). (1) SAX [Lin et al., 2003; Malinowski et al., 2013] and ABBA (Elsworth and Güttel, 2020) map time series into string sequences using unsupervised learning. Each character represents a specific shape or amplitude. However, the unsupervised learning is performed on the dataset level, making the symbolic representation dataset-dependent. Moreover, the repeated sub-strings carry semantics but do not quantify coherence (partially satisfying D2), and do not allow hierarchical structure (failing D1) or complex topological relationships (failing D3). (2) Logical-Shapelets (Mueen et al., 2011) represents entire time series with a series of frequently appearing sub-sequences with fixed length, i.e., shapelets (failing D3). Z-Time (Lee et al., 2024) improves the representation by making the length variable (satisfying D3). Both the shapelets and discrete representations of Z-Time carry no semantic abstraction (failing D2). Regarding structural modeling, Logical-Shapelets only considers boolean relations, and while Z-Time uses Allen’s algebra (Allen, 1983), only limited hierarchy structures can be represented by stacking temporal relation pairs. Therefore, both approaches partially satisfy D1. (3) Chronicle System (Dousson and Le Maigat, 2007) applies graph structure to represent temporal events, where each sub-event is considered as a node in the graph, and STL (Maler and Nickovic, 2004) represents events as recursive logical formulas over signals intervals, both satisfying D1. However, both systems model semantic coherence as binary values (true/false) (partially satisfying D2), and their reliance on actual interval durations does not meet D3. D1: Hierarchical, AI Lab, SLB chenwei.wan@telecom-paris.fr, thou2@slb.com Abstract Time Series Event Detection (TSED) has long been an important task with critical applications across many high-stakes domains. Unlike statistical anomalies, events are defined by semantics with complex internal structures, which are difficult to learn inductively from scarce labeled data in real-world settings. In light of this, we introduce Knowledge-Guided TSED, a new setting where a model is given a natural-language event description and must ground it to intervals in multivariate signals with little or no training data. To tackle this challenge, we introduce Event Logic Tree (ELT), a novel knowledge representation framework to bridge linguistic descriptions and physical time series data via modeling the intrinsic temporal-logic structures of events. Based on ELT, we present a neuro-symbolic VLM agent framework that iteratively instantiates primitives from signal visualizations and composes them under ELT constraints, producing both detected intervals and faithful explanations in the form of instantiated trees. To validate the effectiveness of our approach, we release a benchmark based on real-world time series data with expert knowledge and annotations. Experiments and human evaluation demonstrate the superiority of our method compared to supervised fine-tuning baselines and existing zero-shot time series reasoning frameworks based on LLMs/VLMs. We also show that ELT is critical in mitigating VLMs’ inherent hallucination in matching signal morphology with event semantics. 1 Introduction Time Series Event Detection (TSED) aims to locate segments in time series data where specific events happen. It has long been an important task across many high-stakes domains, including health [Supratak et al., 2017; Perslev et al., 2019], and energy production (Khaouja et al., 2025). Compared to other tasks in time series analysis, TSED is especially challenging. Unlike classification, which aims to assign single labels to time series sequences (Mohammadi Foumani et al., 2024), TSED requires fine-grained discrimination to be made on small timeframes or timestamps. Compared to anomaly detection, which identifies statistical drifts (Zamanzadeh Darban et al., 2024), events usually carry specific semantic meanings (e.g., “a sharp rise in A followed by a plateau in B”) with internal temporal-logic structures across multiple physical channels. The prevalent paradigm for TSED task is inductive pattern recognition. Models are trained to learn patterns from a large number of labeled examples. However, in certain domains, acquiring annotated data is time and resource consuming, making the effective training of such models difficult. Explainability is another shortcoming of inductive learning. Although gradients (Sundararajan et al., 2017) or intermediate activations (Selvaraju et al., 2017) of neural models can provide transparency, humans do not trust the results without logically organized clues. Such AI systems are difficult to be adopted in the domains where trust between AI and human experts is indispensable. In this study, we mitigate the issues by looking at TSED from a cognitive point of view: describing the patterns of events on time series with natural language is generally easier than collecting a large amount of data and annotations. This is in line with the cognitive process when human experts interpret time series data: read unstructured linguistic descriptions, comprehend the internal compositional structure, and deduce the occurrence of events from the time series measurements. Therefore, it naturally provides better explainability. We formalize the problem as Knowledge-Guided Time Series Event Detection (K-TSED). As no training data and annotations are needed, we switch the task setting from inductive pattern recognition to deductive knowledge grounding. However, how to accurately and logically ground the knowledge to time series is still a problem. The example below demonstrates an event description in the energy domain: During the Buildup event, the Volume should remain stable across the entire phase. Simultaneously, the Pressure should either recover slightly with the rate of change decreasing or bounce up quickly, and then reach a steady phase. When humans describe such events, atomic morphological patterns (e.g. “Volume remains stable”, “pressure recovers slightly”) are defined, along with temporal-logic relations (e.g. “simultaneously”, “and then”) hierarchically forming the global pattern of the event type “buildup”. Therefore, an explainable solution should not only map the global event to the time series interval correctly, but also map sub-patterns to their instances accurately, as evidence supporting the global decision that could easily be verified by humans. To achieve the joint goal of accuracy and explainability, we introduce Event Logic Tree (ELT), a novel knowledge representation framework that translates unstructured event descriptions to tree structures. The tree structures faithfully model the temporal-logic relationships between atomic patterns, clearly demonstrating the hierarchy that constitutes the global event. Based on ELT, we create SELA, a neuro-symbolic VLM agent system for zero-shot K-TSED, comprising two types of agents: Logic Analyst agents construct ELT schema from the unstructured event description, and Signal Inspector agents incrementally locate and refine time series intervals guided by the ELT schema. In SELA, ELT plays the critical role in the system as the structured intermediate representation of the target event. It not only guides the agents’ reasoning and behavior, but also provides well-organized clues that can easily be verified by human experts. To validate the effectiveness of our approach, we create the first K-TSED benchmark based on real-world time series data collected from the energy domain. The event descriptions and ground truth labels are created and verified by human experts. We demonstrate that our method significantly outperforms the previous inductive pattern recognition methods in low resource settings, and the existing deductive knowledge grounding solutions without ELT, approaching human-level performance. Through an ablation study, we also demonstrate the critical role of ELT in mitigating VLM’s inherent hallucination problem. Our contributions can be summarized as follows: (1) A new problem setting K-TSED. (2) Our solution: A knowledge representation framework with an implementation based on neural-symbolic VLM agents. (3) A new benchmark built from real-world time series data, with experiments showing the superiority of our method. 2 Problem Formulation The objective of the task K-TSED is to localize all event instances in a multivariate signal that semantically align with the linguistic event descriptions. Input. A multivariate time series 𝐗∈ℝT×C\mathbf{X}\in\mathbb{R}^{T\times C} (TT time steps, CC channels), and a textual description ℒ\mathcal{L} that characterizes signal features for each event type e∈ℰe\in\mathcal{E}, where ℰ\mathcal{E} denotes the set of all event types. Ground Truth. Let 𝒴^={y^m}m=