arXiv:2609.21110v1 [cs.DC] 17 Sep 2026
SensorWF: A FAIR Generalizable Workflow Framework for Scientific Time-Series Analysis Logan Luna
Joseph Rigo, Kellan Shew, Raul Alejandro Vargas-Acosta
School of Computer Science, College of Computing Georgia Institute of Technology Atlanta, Georgia, USA [email protected]
Department of Electrical Engineering and Computer Science Embry-Riddle Aeronautical University Daytona Beach, Florida, USA {rigoj1, shewk1}@my.erau.edu, [email protected]
Abstract—Scientific sensor data is foundational across a wide range of disciplines, including spacecraft engineering, clinical medicine, and atmospheric science. In each context, pipelines are constructed to ingest raw archives, assess data quality, perform feature engineering, perform semantic annotation, and record provenance. However, these pipelines are often implemented as monolithic, domain-specific scripts with implicit assumptions and limited reusability across fields. This work introduces SensorWF, a FAIR-annotated workflow framework designed for generalizable scientific time-series analysis. The framework features a fivemodule reusable core (M1–M5) with typed input/output contracts that constitute the analytical backbone. A domain adapter pattern isolates all domain-specific logic within M1, enabling modules M2–M5 to operate identically across disciplines. Domain assumptions are encoded in a machine-readable component registry, facilitating integration and reuse of any sensor domain without modifying the analytical core. Domain-specific analyses are incorporated as use-case extensions without altering the core modules. The framework generates runtime PROV-O/ProvONE provenance traces, including SHA-256 checksums for all file-path entities, and emits SSN/SOSA-aligned OWL ontologies as primary outputs. To assess generalizability, SensorWF is instantiated in three distinct scientific domains: spacecraft telemetry, ambulatory ECG, and atmospheric climate. Synthetic fault injection and multidetector anomaly detection are demonstrated as representative use-case extensions. Results indicate that a single codebase, parameterized solely through M1 adapters of approximately 170–500 lines each depending on the domain’s complexity, supports comprehensive analytical pipelines across domains with varying sampling rates, channel counts, and fault taxonomies. All code, component registry, ontology artifacts, and datasets are made available as an open scientific object. Our codebase is publicly available at https://purl.archive.org/sensor-wf. Index Terms—computational workflows, FAIR principles, domain adaptation, sensor analytics, anomaly detection, PROV-O, ProvONE, knowledge graph, reproducibility
I. I NTRODUCTION Disciplines as different as spacecraft operations, clinical monitoring, and atmospheric science commonly run their sensor data through similar stages. Raw measurements are ingested This paper has been accepted for publication in the Proceedings of the ReWorDS26 Workshop, held in conjunction with the 22nd IEEE International Conference on eScience (eScience 2026). © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
from an archive, screened and corrected for artifacts, and transformed into analysis-ready feature representations before any domain question is asked of them. What differs between fields is not the shape of this pipeline but the assumptions embedded at each stage: expected sampling rates, quality thresholds, and the semantics of individual channels. Those assumptions determine how a result should be interpreted, yet they are seldom documented explicitly and almost never expressed in machine-readable form. Despite this shared structure, pipelines are typically rebuilt from scratch in each domain as monolithic, singlepurpose scripts. The cost is threefold: analytical code cannot be reused across fields, cross-domain comparisons rest on implementations that were never verified to behave alike, and post-publication auditing becomes impractical because the assumptions behind a result are recoverable only by reading the source. The eScience and open-science communities have responded by arguing that pipelines should be treated as first-class scientific objects, with provenance that is traceable, shareable, and reusable [1], [2]. Acting on that position requires an architecture that separates the analytical stages every domain shares from the assumptions each domain supplies. This paper presents SensorWF, a framework for multi-domain scientific sensor time-series analysis built on exactly that separation. Its reusable core comprises five modules, ingestion (M1), quality assessment (M2), feature engineering (M3), semantic annotation (M4), and provenance export (M5), of which only M1 is domain-specific; the remaining four are configured by the adapter rather than rewritten for it. Analyses particular to a discipline attach as use-case extensions instead of modifications to the core, with synthetic fault injection (E1) and automated anomaly detection (E2) serving as demonstrations throughout. Figure 1 illustrates the complete architecture. A. Key Contributions The primary contributions of this work are as follows: 1) A generalizable five-module workflow core (M1– M5) featuring typed input/output contracts and explicit scientific assumptions. The domain adapter pattern confines domain-specific logic to M1, while M2–M5
Satellite
ECG
Climate
SCOTTI/SATLL
MIT-BIH
Jena Reusable Core (M2–M5)
Domain Adapters
M1
M2
M3
M4
M5
Domain Adapter
Quality
Features
Semantic
Provenance
E1
E2
Injection
Detection
provenance.ttl kg.ttl / .owl
Use-Case Extensions (opt.)
Fig. 1. SensorWF architecture. The domain adapter (teal, M1) normalizes raw sensor archives from three domains into a standardized DataFrame. The reusable core (blue, M2–M5) performs quality assessment, feature engineering, semantic annotation, and provenance export identically across all domains. Optional use-case extensions (orange, E1–E2) inject synthetic faults and evaluate ML detectors; E2 results optionally enhance M4’s knowledge graph (dashed arrow). Switching domains requires only a new M1 adapter (≈150 lines of Python); M2–M5 require no modification.
operate uniformly across disciplines through adapter- workflow-specific constructs such as pipeline programs and driven configuration. execution states [7]. 2) A machine-readable component registry Sensor-oriented ontologies and vocabularies, such as SSN (components.json) that encodes domain and SOSA, provide semantic foundations for representing assumptions as typed ports, configuration parameters, heterogeneous measurements, including voltage channels, ECG domain tags, and ProvONE URIs for each module. signals, and temperature observations, with consistent and This design enables integration and reuse of any sensor domain-specific meaning [3], [4]. Knowledge graphs operadomain without modifying the analytical core, thereby tionalize these models by linking engineered features to source supporting automated composition, discovery, and channels, domain concepts, and analytical outputs, thereby extension. improving both interoperability and reproducibility [6], [7]. 3) Runtime provenance and semantic annotation. SenSensorWF contributes to this ecosystem in two primary sorWF generates PROV-O/ProvONE Turtle provenance ways: first, by encoding domain assumptions in a machinefor each execution, including SHA-256 checksums for readable component registry, which enables sensor-domain all file-path entities. Analytical evidence is mapped to integration without modifying the analytical core; and second, domain OWL ontologies (aligned with SSN/SOSA [3], by treating semantic annotation as an integral workflow [4]), resulting in SPARQL-queryable knowledge graphs. function rather than as a post-processing step. Features are 4) Cross-domain evaluation through an anomaly detec- mapped to ontology classes, and relationships are exported tion use case. SensorWF is evaluated on three established as machine-readable graph artifacts. This design enhances datasets using E1 fault injection and E2 multi-detector FAIR compliance, particularly in terms of interoperability and evaluation as concrete analytical extensions. reusability, by linking outputs from diverse sensor domains, such as satellite telemetry and climate observation, to explicit II. BACKGROUND AND R ELATED W ORK scientific concepts [1], [5]. This approach builds upon previous work on assumption-aware composition [8]–[10] by shifting the A. FAIR Workflows, Provenance, and Semantic Foundations focus from composing non-machine-learning scientific models The FAIR principles (Findable, Accessible, Interoperable, to constructing reusable machine learning-enabled workflows. Reusable) mandate that computational workflows provide typed module specifications, explicitly document scientific assump- B. Domain Adapter Patterns in Scientific Computing tions, and include machine-readable metadata to facilitate The separation of domain-specific components from reusable discovery and reuse without requiring source code inspection algorithmic cores is a foundational principle in scientific [1], [2]. WorkflowHub operationalizes these principles through computing. The Galaxy platform [11] implements this principle RO-Crate packaging and adoption of open standards such as using standalone Tool XML wrappers. Each domain tool CWL, Bioschemas, and the GA4GH TRS API to support specifies its inputs, outputs, and execution commands, while workflow dissemination [5]. the scheduler, data management layer, and workflow engine In scientific computing, provenance, defined as the compre- remain unchanged. This design confines the integration effort hensive record of inputs, transformations, software versions, for new bioinformatics tools to a single wrapper definition. and timing, is critical for verification and downstream reliability However, a key limitation of Galaxy’s model is the lack of assessment. The W3C PROV Ontology (PROV-O) establishes runtime enforcement of typed contracts between tools, and the the standard for machine-readable provenance as queryable recording of provenance only at the workflow level rather than graphs [6], and is further extended by ProvONE to support as first-class semantic artifacts.
Time-series toolkits such as TSFEL [12] employ a com3) Climate Observation: The Jena Climate Dataset [20] is plementary, configuration-driven approach. Feature extraction a widely used multivariate meteorological benchmark, with functions are fully decoupled from the signal source, and data recorded at 10-minute intervals from 2009 to 2016. a JSON configuration specifies which feature families are Climate anomaly detection in this context presents challenges computed and at what window size. This design separates the distinct from those associated with higher-frequency signals. specification of computation from its implementation. However, Specifically, low sampling rates result in contextual anomalies, TSFEL lacks abstractions for data ingestion, quality assessment defined as deviations from expected seasonal patterns rather pipelines, and semantic or provenance outputs, functioning than absolute value thresholds, being more prevalent than point as a feature library rather than a comprehensive workflow anomalies. Autoregressive methods have been employed to architecture. identify instrumental drift and calibration failures; however, SensorWF integrates these two architectural patterns while these methods require stationarity assumptions that often addressing their respective limitations. The DomainAdapter fail during seasonal transitions [21]. In contrast, unsuperabstract interface formalizes domain isolation, as seen in vised reconstruction-based and isolation-based approaches that Galaxy, by establishing explicit architectural invariants with operate on rolling windows are better suited to the nonmachine-readable typed input and output contracts. M3’s stationary, multi-variable characteristics of meteorological data. configuration-driven feature families extend TSFEL’s approach Wu and Keogh [22] show that single-axis difficulty scaling can into a comprehensive pipeline that incorporates spectral produce misleading detector rankings, which motivates E1’s features, semantic knowledge graph annotation, and PROV- simultaneous and independent scaling of amplitude, duration, O provenance export. Portability boundaries, such as M2’s and channel spread across all three domains. assumption of temporally regular sampling, are documented as SPARQL-queryable sensorwf:hasAssumption predi- D. KISPE SATLL The KISPE Satellite Learning Laboratory (SATLL) provides cates instead of remaining implicit within the source code. telemetry data from multiple onboard experiments designed to C. Anomaly Detection Benchmarks and Methods demonstrate spacecraft subsystem functionality and generate In this work, we employ time series anomaly detection as a sensor datasets for analysis. These experiments focus primarily use case for SensorWF. on the Attitude Determination and Control System (ADCS), Anomaly detection refers to the identification of observations including accelerometer, gyroscope, and reaction wheel tests, that deviate significantly from expected behavior, including as well as thermal monitoring experiments representative of signal spikes, sensor drift, dropout, or correlated multi-channel environmental testing commonly performed on spacecraft deviations that may indicate equipment faults or physiological systems. Together, these experiments generate telemetry data events. Two principal approaches are utilized: supervised across multiple subsystems and sensor types, providing a useful methods, which are trained on labeled examples of known foundation for analysis, visualization, and semantic modeling fault types, and unsupervised methods, which model normal of CubeSat telemetry data. behavior and flag deviations. In most cases, labeled fault data The laboratory supports four experiment families. Three exerare scarce for scientific instruments, making unsupervised cise the ADCS: an accelerometer test that sweeps the CubeSat detection the predominant strategy. Techniques such as statis- through successive orientations to record acceleration on the tical thresholding (Z-score, rolling robust variants), ensemble x, y, and z axes; a gyroscope test that rotates the suspended isolation (IsolationForest [13]), density-based local outlier satellite through a fixed sequence (90◦ counterclockwise, 180◦ detection (LOF [14]), and reconstruction-error neural networks clockwise, then back to origin), holding each position for (Autoencoders) each address distinct aspects of anomalous several seconds; and a reaction wheel test that actuates the behavior. wheel to rotate the satellite while the gyroscope records the 1) Spacecraft Telemetry: In spacecraft telemetry, Hundman resulting motion. The fourth is a longer-duration thermal test, et al. [15] introduced LSTM-based detection for NASA in which the satellite is warmed in direct sunlight, cooled SMAP and MSL missions. The OPS-SAT benchmark [16] indoors, and then heated by the onboard heater, producing provides segment-level ground truth, whereas the ESA Anomaly a high-volume record across multiple independent thermal Database (ESA-ADB) [17] supplies event-level ground truth nodes. Together the four families span both stable regimes with affiliation-aware metrics that more accurately represent (accelerometer, thermal) and oscillatory ones (reaction wheel, operational detection performance compared to point-wise gyroscope) at approximately 1 Hz, which makes them a useful AUC-ROC. stress test for a domain-agnostic analytical core. 2) Biomedical Data: In the biomedical field, the MITIII. M ETHODOLOGY BIH Arrhythmia Database [18] is the standard benchmark for arrhythmia detection. Supervised deep learning approaches Within this section the framework is described in three such as CNN-based heartbeat classification [19] demonstrate segments. The design principles below state the architectural strong cross-dataset transferability on this corpus; SensorWF invariants that make a single core reusable across domains. targets unsupervised fault detection without reliance on rhythm The component registry and module descriptions then specify labels. what each module consumes and produces, and the domain
application subsections show how those specifications are instantiated for satellite, ECG, and climate data. A. Design Principles The SensorWF framework is based upon five core design principles. Core/extension separation. The framework differentiates a reusable core (M1 to M5) from optional use-case extensions. The core uniformly manages ingestion, quality assessment, feature extraction, semantic annotation, and provenance across all domains. Extensions, such as fault injection (E1) and anomaly detection (E2), interface with core typed outputs without modifying any core module. This separation is explicitly specified in components.json using an extension flag. Adapter isolation. All domain-specific logic is encapsulated within M1 via the DomainAdapter abstract interface. Each adapter implements four methods: load(), get_quality_config(), get_feature_config(), and get_ontology_path(), thereby providing comprehensive configuration for all core modules. Integrating a new domain requires implementing only these four abstract methods in a new M1 adapter class (approximately 150–200 lines for the DomainAdapter interface); M2 through M5 require no modification. Use-case extensions may additionally invoke get_fault_types() as required. Explicit I/O contracts. Each module defines typed input and output ports that adhere to the ProvONE provone:Program pattern [7]. These contracts are documented in components.json to facilitate programmatic discovery and composition. Scientific assumption transparency. Domain-specific assumptions, including sampling rates, quality thresholds, and window sizes, are recorded as sensorwf:hasAssumption predicates in both the static workflow specification and the runtime provenance trace. This method ensures that assumptions are directly SPARQL-queryable. Reproducibility by design. All sources of randomness originate from a single seeded generator chain (seed = 42). Repeated execution of any domain entry point produces bit-forbit identical clean data, quality reports, feature matrices, and knowledge graph artifacts. Result caching, indexed by sentinel files (run_summary.json, ml_results.csv), prevents redundant recomputation unless otherwise specified. B. Component Registry and Module Descriptions M1: Data Ingestion. Three concrete adapters are provided: SatelliteAdapter (SCOTTI v2 hex-encoded telemetry at ∼1 Hz); ECGAdapter (MIT-BIH CSV at 360 Hz, decimated 7× to ≈51 Hz by uniform subsampling, 5-minute sessions yielding 15,000 samples); and ClimateAdapter (Jena CSV at 10-minute resolution, 14 channels, half-year slices). All three return a standardized DataFrame with mandatory columns timestamp, elapsed_s, and one or more signal channels. M2: Quality Assessment. Domain-agnostic quality checks are configured via the adapter’s get_quality_config() dictionary: NaN rates, stuck-sensor detection (≤ k unique
values in a window), timing regularity (> n× expected interval), per-channel Z-score flags, and optional linear-trend detection. Outputs a quality report JSON and a cleaned DataFrame. M3: Feature Engineering. Nine feature families are generated per channel: (i) raw channel value, (ii) first-order difference, (iii) rolling mean, (iv) rolling standard deviation, (v) rolling skewness, (vi) rolling excess kurtosis, (vii) zerocrossing rate relative to the channel mean, (viii) spectral entropy (Shannon entropy of the rolling power spectrum computed via vectorized batched FFT), and (ix) dominant frequency in Hz (peak non-DC frequency). Rolling skewness and kurtosis are calculated using a Cython-compiled O(n) sliding-window accumulator (scripts/_core_cy.pyx, -O3 -ffast-math). Spectral features are computed using a three-dimensional stride-tricks FFT across all channels, and lastly, a sample-interval timing feature (dt_sample) is appended. The resulting feature matrix is N × D, where D depends on channel count and window size. M4: Semantic Annotation. Feature importances are mapped to a domain-specific OWL ontology using lexical prefix matching on feature names. The resulting knowledge graph includes OWL class-instance nodes, feature nodes, and evidence edges with the if:featureImportance and if:evidenceForClass predicates, enabling SPARQLqueryable subsystem attribution. When executed after E2, anomaly tag nodes are incorporated; in core-only mode, edges represent feature-to-class evidence derived directly from the M3 feature matrix. Outputs include RDF/Turtle files and CSV node and edge lists. M5: Provenance Export. The ProvenanceRecorder accumulates one prov:Activity per module call, capturing wall-clock timestamps, input row counts, output file paths, and parameter values. SHA-256 checksums are automatically computed for all file-path entities and stored as telwf:sha256 triples, enabling bit-level reproducibility verification. The serialized PROV-O/ProvONE Turtle document is produced as a primary workflow output, regardless of whether use-case extensions are executed. E1: Fault Injection (use-case only). The procedure reads get_fault_types() and applies each morphology across three difficulty tiers, scaling amplitude, duration, and channel spread simultaneously. This multi-axis approach mitigates confounds associated with single-axis separability [22]. With two variants per (type, tier) pair, E1 generates Nfaults × 3 × 2 labeled sessions per domain. E2: Anomaly Detection (use-case only). Five machine learning detectors are trained on the first 60% of the clean session, which is temporally prior to any injected segment: (i) Z-Score with MAD-robust fusion (Iglewicz-Hoaglin threshold [23]); (ii) RobustRollingZScore with CUSUM persistence [24]; (iii) IsolationForest with a two-member random-rotation ensemble [13]; (iv) a multi-scale MLP Autoencoder with Gaussian denoising (σ = 0.06), dual sequence windows (lengths 8 and 16), and a five-layer symmetric encoder-decoder (widths ⌊n/3⌋, ⌊n/8⌋, ⌊n/16⌋, ⌊n/8⌋, ⌊n/3⌋ for flattened window dimension n, tanh activations) [25]; and (v) Local Outlier
Factor (LOF) with novelty detection and PCA pre-processing (95% variance retained) [14]. Ensemble and density-based detectors (IF, LOF, PCA) use scikit-learn [26]. For densitybased detectors (IF and LOF), a Peaks-over-Threshold (POT) extreme-value theory calibration [27] fits a Generalized Pareto Distribution to the tails of the training-set scores, whereas statistical detectors use the 99th percentile threshold. Per-fault and aggregate metrics, including AUC-ROC, AUC-PR, F1, FPR, recall, and the binary event_detected metric, are reported by tier. C. Domain Applications
ventricular contractions (PVCs) and bigeminy (105, 119, 200, 205, 215), and atrial fibrillation with complex ventricular rhythms (201, 208, 213, 221). The ECGAdapter applies uniform subsampling by a factor of 7 (360 Hz →≈ 51 Hz) with no anti-aliasing filter. Six ECG-specific fault morphologies are introduced: baseline wander, electrode dropout, EMG burst, powerline noise (10 Hz, representing the alias of 60 Hz mains interference after decimation: 60 mod 50 = 10 Hz), amplitude scaling, and lead inversion. For each record, 6 × 3 × 2 = 36 labeled sessions are generated. 3) Atmospheric Climate (Jena Climate Dataset): The Jena Climate Dataset [20] records 14 meteorological variables, including temperature, pressure, humidity, and wind speed or direction, at 10-minute intervals from 2009 to 2016 (420,551 rows). The ClimateAdapter extracts half-year segments (approximately 26,000 rows each) and removes −9999 sentinel values. M3 employs a 24-sample rolling window (4 hours). Seven half-year sessions from 2009 to 2015 are processed, collectively covering cold and warm seasons, distinct instrument drift periods, and seven years of inter-annual variability. Six climate-specific fault morphologies are introduced: sensor drift, stuck sensor, spike burst, sensor dropout, scale error, and sign flip. For each session, 6 × 3 × 2 = 36 labeled sessions are produced.
Section II provided an overview of each domain at the levels of physical experimentation and dataset description. The following subsections examine the specifics of the M1 adapter, present fault taxonomies, and detail module-level parameter selections that enable the operationalization of each domain within SensorWF. 1) Satellite Telemetry (KISPE SATLL): The KISPE Satellite Learning Laboratory (SATLL) provides authentic telemetry (real-time measurement data transmitted from remote sensors) from four experiment families: AccelerometerTest, GyroTest, ReactionWheelTest, and ThermalTest [28]. Each session generates CDH (Command and Data Handling) data, encompassing 94 columns including measurements of thermal, power, and IV. D OMAIN A PPLICATION R ESULTS timing, as well as ADCS (Attitude Determination and Control System) data, covering 32 columns that include sensors for In this section we describe our empirical evaluation of inertia measurement units (IMUs), reaction wheels, and sun SensorWF in addition to our results. Section IV-A fixes the sensors, at approximately 1 Hz (one measurement per second). train/test discipline shared by all three domains. The three E1 introduces 18 fault types: 16 single-channel faults (in domain subsections then report detection performance and which only one measurement channel is affected) and 2 explain each ranking in terms of session length, channel count, compound faults involving multiple channels. CDH faults and fault morphology. The final two subsections cover outputs encompass issues such as power-rail drift (gradual voltage that are produced regardless of whether the detection extension changes), thermal ramp (continuous temperature change), runs at all, those being the M4 knowledge graph and the and packet dropout (missing data segments), while ADCS runtime cost of the core itself. faults include gyro clipping (sensor saturation), magnetometer inversion (incorrect field polarity), and wheel runaway or A. Training and Evaluation Protocol stiction (uncontrolled or stuck motion). For each experiment E2 applies a strict temporal train/test split to prevent data family, there are 108 labeled sessions (18 × 3 × 2 = 108). leakage. The initial 60% of each clean baseline session is The satellite OWL ontology is generated by the designated as training data. E1 evaluates only the remaining pipeline at runtime. The script satellite_ontology.py 40% suffix; injected windows always commence at or after examines the CDH and ADCS channel lists identi- sample ⌊0.6 · N ⌋, which ensures no temporal overlap between fied during M1 and produces an OWL/XML document the training prefix and any evaluated anomalous segment. containing subsystem classes (sat:ADCS, sat:OBDH, Rolling-window features, with a maximum of 100 samples sat:EPS, sat:ThermalControlSubsystem), sensor- at 50 Hz, are negligible compared to the ≥ 6,000-sample clean type subclasses, and individual channels. Alignment with SSN gap, thereby preventing roll-back leakage. Evaluation metrics and SOSA is achieved through the use of sosa:Platform include sample-wise AUC-ROC, AUC-PR, F1, FPR, recall, and and ssn:Sensor superclasses. A single-session M4 execu- event_detected (a binary indicator of whether the injected tion yields 1,104 knowledge graph nodes and 846 edges. anomaly window was intercepted by at least one true positive). 2) Biomedical ECG (MIT-BIH Arrhythmia Database): The AUC-PR serves as the primary imbalance-aware metric, given MIT-BIH Arrhythmia Database [18] comprises 48 thirty-minute, the anomaly prevalence of ≈ 20–40% per injected session. two-lead ambulatory ECG recordings at 360 Hz from PhysioNet [29], each annotated beat-by-beat by cardiologists. Twenty B. Satellite Telemetry (KISPE SATLL) records are selected to represent all major rhythm classes: In the satellite domain, the Autoencoder outperforms all five normal sinus rhythm (100, 101, 103, 112, 113, 115), bundle detectors, achieving an AUC-ROC of 0.772 and an AUC-PR of branch block (106, 108, 109, 111, 118), frequent premature 0.581, which aligns with its temporal reconstruction capabilities
standard deviations for IsolationForest (±0.047) and LOF (±0.042) indicate stable performance across the 20-record rhythm-class mix. RobustRollingZScore demonstrates poor AUC-ROC / AUC-PR performance (AUC-ROC 0.487, AUC-PR 0.394), consistent Domain Sess. Faults AE IF LOF RRZS ZS with the breakdown of its fixed-window assumption when Satellite 1 18 0.772/0.581 0.564/0.298 0.549/0.280 0.597/0.359 0.610/0.394 ECG 20 6 0.915/0.837 0.880/0.762 0.918/0.826 0.487/0.394 0.785/0.637 confronted with varied ECG artifact morphologies. Per-tier Climate 7 6 0.835/0.725 0.720/0.577 0.737/0.609 0.536/0.399 0.681/0.552 results indicate strong performance across all difficulty levels AE=Autoencoder, IF=IsolationForest, LOF=Local Outlier Factor, for both the Autoencoder and LOF, with AUC-ROC ranging RRZS=RobustRollingZScore, ZS=ZScore. Bold: best per domain. from 0.876 (Hard) to 0.943 (Easy) and 0.886 (Hard) to 0.947 ECG std: AE ±0.061, LOF ±0.048; Climate: AE ±0.034, LOF ±0.044. (Easy), respectively. Full per-tier numeric results are provided in Appendix C. TABLE I C ROSS -D OMAIN D ETECTION P ERFORMANCE
D. Atmospheric Climate (Jena Climate Dataset)
Fig. 2. Cross-domain detection performance for all five E2 detectors across satellite (SATLL), biomedical ECG (MIT-BIH), and atmospheric climate (Jena). Values are macro-averaged over all sessions per domain. AE=Autoencoder, IF=IsolationForest, RRZS=RobustRollingZScore, ZS=ZScore.
In the climate domain, the Autoencoder achieves the highest performance with an AUC-ROC of 0.835 and an AUC-PR of 0.725, exhibiting notably low variance across the seven half-year sessions (±0.023). This consistency confirms reliable detection of sustained meteorological sensor faults over seven years of seasonal and inter-annual variability. LOF ranks second (AUC-ROC 0.737 ± 0.041, AUC-PR 0.609 ± 0.055), followed by IsolationForest (0.720/0.577). ZScore (0.681/0.552) outperforms RobustRollingZScore (0.536/0.399), with the latter’s degradation attributed to its sensitivity to variable anomaly duration across tiers. Hard-tier AUC-PR values substantially exceed those of the Easy-tier for all detectors (e.g., Autoencoder: 0.856 vs. 0.549), driven by the larger fractional window occupied by Hard-tier injections, which increases the positive class rate. IsolationForest demonstrates the largest gap between AUC-ROC and AUC-PR (0.720 vs. 0.577 overall), indicating weaker calibration at the precision-recall operating point for climate fault distributions.
for multi-channel CDH and ADCS fault morphologies. The detector ranking for the satellite domain differs substantially from those observed in the ECG and Climate domains: ZScore ranks second (AUC-ROC 0.610, AUC-PR 0.394), followed by RobustRollingZScore (0.597/0.359), while IsolationForest (0.564/0.298) and LOF (0.549/0.280) rank lowest. This reversal in the performance of density-based detectors is attributed to the short session length (333 rows), which restricts the training E. Semantic Knowledge Graph prefix to approximately 200 samples, insufficient for LOF and The satellite core M4 run generates a knowledge IsolationForest to construct reliable density estimates across 122 graph comprising 1,104 nodes and 846 edges. Following channels. The Autoencoder’s reconstruction loss demonstrates E2, the machine learning-enhanced knowledge graph reduced sensitivity to training-set size in low-row scenarios, incorporates IsolationForest feature importances, resulting establishing it as the most robust option for this domain. Easyin 23 ML-weighted nodes linked to subsystem classes. tier faults yield the highest Autoencoder AUC-ROC (0.801) The most frequently queried evidence class via SPARQL is due to larger amplitude injections, whereas Hard-tier faults sat:AttitudeDeterminationAndControlSubsystem, result in the highest AUC-PR (0.646) as the longer fractional which aligns with the ADCS-heavy channel structure. window increases the positive class rate. Figure 3 presents the resulting subgraph. The C. Biomedical ECG (MIT-BIH Arrhythmia Database) if:featureImportance predicate enables queries In the ECG domain, LOF achieves the highest AUC-ROC of the form SELECT ?feat ?cls WHERE {?feat (0.918 ± 0.042) and AUC-PR (0.826 ± 0.088) across all 20 if:evidenceForClass ?cls} without the need to records, narrowly surpassing the Autoencoder (AUC-ROC rerun the detector, thereby supporting post-hoc subsystem 0.915 ± 0.055, AUC-PR 0.837 ± 0.097) and IsolationForest attribution; channel-level attribution methods [30], [31] are (0.880/0.762). The two leading detectors are effectively tied on identified as a complementary extension. In the ECG domain, AUC-ROC, while the Autoencoder marginally leads on AUC- per-record M4 runs yield compact knowledge graphs averaging PR (0.837 vs. 0.826), indicating an advantage on harder-tier 20 nodes and 9 edges each (401 nodes and 189 edges in total faults where longer injected windows elevate the precision- across 20 records), reflecting the two-channel signal structure recall operating point. Six morphologically distinct fault types (19 features per record). In the Climate domain, per-session generate compact, well-separated clusters in the 19-feature M4 runs produce substantially larger graphs (987 total nodes rolling-window space, which benefits both neighborhood- and 882 total edges across 7 sessions) due to the 127 features density estimation and reconstruction-based approaches. Low derived from 14 meteorological channels.
the climate and ECG adapters need only 167 and 185 lines respectively, while the satellite adapter reaches 502 because it decodes hex-encoded SCOTTI frames across 122 channels and declares 18 fault morphologies. Adapter effort therefore tracks the complexity of the raw archive format rather than the complexity of the framework. The component registry enables programmatic composition: consumers query components.json for modules with a specified domain tag and receive fully specified input and output contracts, facilitating automated workflow instantiation and validation. Although only M1 is domain-specific at the code level, it is important to note that the behavior of M2 through M5 depends on adapter-supplied configuration dictionaries that encode domain knowledge, which must be provided explicitly by the adapter author. Portability limitations, such as M2’s assumption of temporally regular sampling, are documented as sensorwf:hasAssumption predicates in the registry. For instance, irregular-rate or eventdriven streams require time-indexed rolling windows in M3 and resampling extensions in M2. B. FAIR Assessment Fig. 3. M4 knowledge graph subgraph for the satellite domain (AccelerometerTest family). Nodes represent OWL class instances (sat:ADCS, sat:OBDH, etc.), anomaly tags, and IsolationForest feature nodes. Edges carry if:featureImportance and if:evidenceForClass predicates, enabling SPARQL queries that link detector evidence to satellite engineering subsystems without re-executing the detector.
F. Runtime Characterization Table II in Appendix B presents per-module wall-clock times measured on a standard laptop (Apple M-series, single thread). E2 accounts for approximately 80–90% of the total pipeline time. A detector-fit cache, keyed on detector configuration rather than fault tag, ensures that each family incurs only one training pass and re-evaluates all Nfaults ×3×2 variants without redundant model refits. Addressing a previous defect that retrained all five detectors once per fault tag (resulting in 6× redundant Autoencoder fits per record) reduced ECG E2 time from over 1,000 seconds per record to the values reported. Core modules M1–M5 collectively consume less than 10% of runtime, indicating that framework overhead is negligible compared to domain analysis costs. M3 consistently remains under 1 second across all domains due to the Cython-compiled skew/kurtosis accumulator and the batched stride-tricks FFT. V. D ISCUSSION Having established that the core transfers across three domains, this section examines what that transfer actually costs, how far the FAIR objectives were met in practice, and where the evaluation protocol falls short of operational satellite benchmarking practice. A. Generalizability and Adapter Effort For the three validated domains, implementing each adapter required 167 to 502 lines of Python code, whereas M2 through M5 required no code modifications. The spread is informative:
During the creation of SensorWF, aligning with the four FAIR principles was a primary objective. Reproducibility is ensured at the code level: executing any domain entry point with –seed 42 produces bit-identical CSVs, injection summaries, metrics, and knowledge graph artifacts. The only output that varies between runs is provenance.ttl, in which timestamps are intentionally different. Workflow registration on WorkflowHub with RO-Crate metadata and immutable Zenodo archives is planned for the open scientific object release. C. Relation to Satellite Anomaly Benchmarks The ESA-ADB benchmark [17] recommends event-level ground truth, affiliation-aware scoring, and dynamic or extreme value theory (EVT) thresholding with postprocessing for operational satellite telemetry. OPS-SAT [16] emphasizes segment-level precision at low operating false positive rates. SensorWF partially addresses these recommendations: densitybased detectors (IF, LOF) utilize peaks-over-threshold (POT) EVT calibration (Section IV-A), and an event-level event detected metric is generated for each variant. Full affiliation-aware scoring and dynamic postprocessing remain priority extensions. E2’s threshold strategy and detector suite are registry-declared parameters, which supports incremental alignment with ESAADB protocols without requiring modifications to any core module. VI. L IMITATIONS AND F UTURE W ORK Synthetic-only evaluation. All ground-truth labels are generated by E1, and no real annotated anomalies are included. This limitation may overstate the generalizability of the results compared to operational benchmark performance (e.g., ESAADB, OPS-SAT). Evaluation using real annotated datasets is therefore a priority for future research. Sample-wise metrics. AUC-ROC and AUC-PR are computed at the sample level, while the binary event_detected
flag provides an event-level complement for each variant. Event- on synthetic evaluation, use of point-wise metrics, and the wise, affiliation-aware, and range-based metrics [17] more absence of transformer baselines, are explicitly documented accurately reflect detection quality in operational contexts and as sensorwf:hasAssumption predicates. KISPE SATLL are therefore priorities for future extension. data and all code will be made available on GitHub as an open Deep learning baselines. The five E2 detectors do not scientific object. incorporate transformer-based approaches such as iTransformer [32] or TranAD [33], which require separate deep learning VIII. ACKNOWLEDGEMENTS frameworks. The registry-based E2 architecture is compatible with these methods, and integration of at least one transformer Special thanks to Dr. Jerry Sellers for explaining the baseline is planned as the next step. Kipse CubeSat dashboard. Special thanks to the College Assumption-aware and decision-aware ML workflows. of Engineering at Embry-Riddle Aeronautical University for Future work will extend SensorWF by representing additional providing access to the CubeSat. This project is supported by workflow assumptions and analytical choices as machine- the Center for Aerospace Resilient Systems at Embry-Riddle readable metadata, including preprocessing choices, feature- Aeronautical University under an internal grant. construction parameters, detector configurations, thresholding strategies, and evaluation settings. This would support inspecR EFERENCES tion, comparison, and reuse of ML workflows across domains by linking selected workflow choices to downstream analytical [1] M. D. Wilkinson, M. Dumontier, I. J. Aalbersberg, G. Appleton, M. Axton, A. Baak, N. Blomberg, J.-W. Boiten, L. B. da Silva Santos, P. E. products. VII. C ONCLUSION SensorWF demonstrates that a FAIR-annotated, generalizable workflow framework for scientific sensor time-series analysis is both feasible and practically valuable across multiple disciplines. The separation of domain-specific ingestion (M1 adapter) from a reusable analytical core (M2 to M5), along with the encoding of domain assumptions in a machinereadable component registry, allows integration and reuse of any sensor domain without modification to the analytical core. This design enables researchers to adopt, adapt, and extend a unified pipeline without the need to reimplement domainspecific components. Use-case extensions, including E1 fault injection and E2 anomaly detection, interface with core typed outputs without requiring changes to any core module, thereby demonstrating that the architecture accommodates a wide range of downstream analyses. Cross-domain validation involving spacecraft telemetry (KISPE SATLL, one session, 18 fault types), ambulatory ECG (20 MIT-BIH records covering all major rhythm classes), and atmospheric climate data (seven Jena half-year sessions) demonstrates that a consistent detection architecture can be maintained from a single codebase, with domain-specific configuration restricted to M1. For satellite data, the Autoencoder achieves the highest performance (AUC-ROC 0.772, AUCPR 0.581), whereas density-based detectors perform least effectively due to the short session, which limits the training prefix to approximately 200 samples. In ECG analysis, LOF and the Autoencoder yield comparable results (AUC-ROC 0.918 ± 0.042 and 0.915 ± 0.055, AUC-PR 0.826 and 0.837, respectively), with the Autoencoder narrowly surpassing in AUC-PR. For climate data, the Autoencoder again leads (AUCROC 0.835 ± 0.023, AUC-PR 0.725) and demonstrates low variance across seven years of sessions. The evaluation protocol is enhanced by EVT-calibrated thresholding for density-based detectors, rolling-window sensitivity analysis, and an eventlevel detection metric. Key limitations, such as reliance
Bourne, J. Bouwman, A. J. Brookes, T. Clark, M. Crosas, I. Dillo, O. Dumon, S. Edmunds, C. T. Evelo, R. Finkers, A. Gonzalez-Beltran, A. J. G. Gray, P. Groth, C. Goble, J. S. Grethe, J. Heringa, P. A. C. ’t Hoen, R. Hooft, T. Kuhn, R. Kok, J. Kok, S. J. Lusher, M. E. Martone, A. Mons, A. L. Packer, B. Persson, P. Rocca-Serra, M. Roos, R. van Schaik, S.-A. Sansone, E. Schultes, T. Sengstag, T. Slater, G. Strawn, M. A. Swertz, M. Thompson, J. van der Lei, E. van Mulligen, J. Velterop, A. Waagmeester, P. Wittenburg, K. Wolstencroft, J. Zhao, and B. Mons, “The FAIR guiding principles for scientific data management and stewardship,” Scientific Data, vol. 3, p. 160018, 2016. [2] D. Garijo, S. Kinnings, S. Picart-Armada, R. Zamora-Resendiz, Y. Liu, A. Salama, S. Dey, C. Goble, A. R. Pico, T. Nationalparks, and N. H. Shah, “Applying the FAIR principles to computational workflows,” Patterns, vol. 3, no. 9, p. 100603, 2022. [3] A. Haller, K. Janowicz, S. J. D. Cox, M. Lefrançois, K. Taylor, D. Le Phuoc, J. Lieberman, R. Garcı́a-Castro, R. Atkinson, and C. Stadler, “The modular SSN ontology: A joint W3C and OGC standard specifying the semantics of sensors, observations, sampling, and actuation,” Semantic Web, vol. 10, no. 1, pp. 9–32, 2019. [4] M. Compton, P. Barnaghi, L. Bermudez, R. Garcı́a-Castro, O. Corcho, S. Cox, J. Graybeal, M. Hauswirth, C. Henson, A. Herzog, V. Huang, K. Janowicz, W. D. Kelsey, D. Le Phuoc, L. Lefort, M. Leggieri, H. Neuhaus, A. Nikolov, K. Page, A. Passant, A. Sheth, and K. Taylor, “The SSN ontology of the W3C semantic sensor network incubator group,” Journal of Web Semantics, vol. 17, pp. 25–32, 2012. [5] O. J. R. Gustafsson, C. Goble, S. Soiland-Reyes, S. R. Wilkinson et al., “WorkflowHub: A registry for computational workflows,” Scientific Data, vol. 12, p. 837, 2025. [6] W3C Provenance Working Group, “PROV-O: The PROV ontology,” W3C Recommendation, Apr. 2013. [Online]. Available: https: //www.w3.org/TR/prov-o/ [7] P. Missier, K. Belhajjame, and J. Cheney, “ProvONE: A PROV extension for scientific workflows,” DataONE, Tech. Rep., 2013. [Online]. Available: http://purl.dataone.org/provone-v1-dev [8] R. A. Vargas-Acosta, L. Garnica Chavira, N. Villanueva-Rosales, and D. D. Pennington, “Automating Multivariable Workflow Composition for Model-to-Model Integration,” in Proceedings of the 2022 IEEE 18th International Conference on e-Science (e-Science), Salt Lake City, Utah, 2022. [9] R. A. Vargas Acosta, “Automated Composition of Multivariable Scientific Workflows Considering Scientific Assumptions,” Doctoral Dissertation, The University of Texas at El Paso, El Paso, Texas, 2024. [Online]. Available: https://scholarworks.utep.edu/open etd/4162/ [10] R. A. Vargas-Acosta, L. A. Garnica Chavira, N. Villanueva-Rosales, and D. D. Pennington, “Towards Improving Automatic Composition of Scientific Workflows: Variable Reconciliation and Tie-Breaking Strategy,” in 2025 IEEE 19th International Conference on Semantic Computing (ICSC). California, United States of America: Institute of Electrical and Electronics Engineers, Feb. 2025.
[11] E. Afgan, D. Baker, B. Batut, M. van den Beek, D. Bouvier, M. Čech, J. Chilton, D. Clements, N. Coraor, B. A. Grüning, A. Guerler, J. HillmanJackson, S. Hiltemann, V. Jalili, H. Rasche, N. Soranzo, J. Goecks, J. Taylor, A. Nekrutenko, and D. Blankenberg, “The Galaxy platform for accessible, reproducible and collaborative biomedical analyses: 2018 update,” Nucleic Acids Research, vol. 46, no. W1, pp. W537–W544, 2018. [12] M. Barandas, D. Folgado, L. Fernandes, S. Santos, M. Abreu, P. Bota, H. Liu, T. Schultz, and H. Gamboa, “TSFEL: Time series feature extraction library,” SoftwareX, vol. 11, p. 100456, 2020. [13] F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in Proceedings of the 8th IEEE International Conference on Data Mining (ICDM), 2008, pp. 413–422. [14] M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “LOF: Identifying density-based local outliers,” in Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data, 2000, pp. 93–104. [15] K. Hundman, V. Constantinou, C. Laporte, I. Colwell, and T. Soderstrom, “Detecting spacecraft anomalies using LSTMs and nonparametric dynamic thresholding,” in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’18), 2018, pp. 387–395. [16] B. Ruszczak, M. Mrochen, and J. Nalepa, “The OPS-SAT benchmark for detecting anomalies in satellite telemetry,” Scientific Data, vol. 12, no. 1, p. 710, 2025. [17] K. Kotowski, B. Ruszczak, and J. Nalepa, “ESA-ADB: A benchmark dataset for unsupervised anomaly detection of satellite telemetry,” Scientific Data, vol. 11, p. 853, 2024. [18] G. B. Moody and R. G. Mark, “The impact of the MIT-BIH arrhythmia database,” IEEE Engineering in Medicine and Biology Magazine, vol. 20, no. 3, pp. 45–50, 2001. [19] M. Kachuee, S. Fazeli, and M. Sarrafzadeh, “ECG heartbeat classification: A deep transferable representation,” in Proceedings of the IEEE International Conference on Healthcare Informatics (ICHI), 2018, pp. 443–444. [20] Max Planck Institute for Biogeochemistry, “The Jena climate dataset,” Online dataset, 2016, weather station data 2009–2016, 14 meteorological variables at 10-minute intervals. Available at https://www.bgc-jena.mpg. de/wetter/. [21] F. Martı́nez-Álvarez, A. Morales-Esteban, and J. Reyes, “Machine learning and big data for anomaly detection in environmental monitoring: A review,” Neural Computing and Applications, vol. 32, pp. 11 893– 11 907, 2020. [22] R. Wu and E. J. Keogh, “Current time series anomaly detection benchmarks are flawed and creating the illusion of progress,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 3, pp. 2421–2429, 2022. [23] B. Iglewicz and D. C. Hoaglin, How to Detect and Handle Outliers. ASQ Quality Press, 1993. [24] E. S. Page, “Continuous inspection schemes,” Biometrika, vol. 41, no. 1/2, pp. 100–115, 1954. [25] M. Sakurada and T. Yairi, “Anomaly detection using autoencoders with nonlinear dimensionality reduction,” in Proceedings of the MLSDA 2014 Workshop on Machine Learning for Sensory Data Analysis, 2014, pp. 4–11. [26] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and É. Duchesnay, “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [27] A. Siffer, P.-A. Fouque, A. Termier, and C. Largouet, “Anomaly detection in streams with extreme value theory,” in Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2017, pp. 1067–1075. [28] KISPE Space Systems Ltd., “Satellite learning laboratory (SATLL),” Product description, 2024. [Online]. Available: https://www.kispe.co.uk [29] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. e215–e220, 2000. [30] F. Schroeder, T. Borchert, and A. Schulz, “Conditional attribution for anomaly explanation in multivariate time series,” arXiv preprint, 2026, arXiv:2604.17616.
[31] Y. Chen, Z. Wang, and H. Li, “Multi-view autoencoder decomposition for root cause analysis in multivariate time series,” arXiv preprint, 2026, arXiv:2601.09287. [32] Y. Liu, T. Hu, H. Zhang, A. Wu, J. Wang, and M. Long, “iTransformer: Inverted transformers are effective for time series forecasting,” arXiv preprint, 2024, arXiv:2310.06625. [33] S. Tuli, G. Casale, and N. R. Jennings, “TranAD: Deep transformer networks for anomaly detection in multivariate time series data,” Proceedings of the VLDB Endowment, vol. 15, no. 6, pp. 1201–1214, 2022.
A PPENDIX A M1 D OMAIN A DAPTER I MPLEMENTATION Listing 1 gives an abridged ClimateAdapter, the shortest of the three validated M1 adapters at 167 lines. The four methods shown constitute the entire domain-specific surface of the framework; M2–M5 consume their outputs without modification.
Listing 1. Abridged ClimateAdapter
class ClimateAdapter(DomainAdapter): channels = _CHANNELS # 14 met. channels native_hz = 1.0 / 600.0 # 10-min sampling def load(self, path, **kw) -> pd.DataFrame: df = pd.read_csv(path, parse_dates=["Date Time"]) df = df.rename(columns=_COL_MAP) df[wv] = df[wv].replace(-9999.0, np.nan) df["elapsed_s"] = (df.timestamp - t0).dt.total_seconds() return df[["timestamp","elapsed_s"]+_CHANNELS] def get_quality_config(self) -> dict: return {"stuck_unique_max": 3, "zscore_threshold": 3.5, "expected_dt_s": 600.0, "gap_multiplier": 3.0} def get_feature_config(self) -> dict: return {"channels": _CHANNELS, "window": 24}
def get_ontology_path(self) -> str: return "results/ontologies/climate.owl"
A PPENDIX B P ER -M ODULE RUNTIME B REAKDOWN Table II disaggregates wall-clock time by module for a full satellite run (4 experiment families, 18 fault types, 3 difficulty tiers, 2 variants each) on a single laptop core. It is the basis for the <10% core-overhead result reported in Section IV-F. TABLE II P ER -M ODULE RUNTIME (S ATELLITE , 4 FAMILIES , 2 VARIANTS , 3 T IERS ) Module
Time (s)
%
Notes
M1 Ingestion M2 Quality M3 Features E1 Injection E2 Detection M4 Semantic M5 Provenance
<1 <1 Parse SCOTTI archive <1 <1 Clean + flag channels <1 <1 9 families; Cython O(n) + batched FFT ∼5 ∼1 Generate 432 labelled CSVs ∼400 ∼90 Train 5 ML detectors; config-keyed cache ∼30 ∼7 Build KG from ML evidence + OWL <1 <1 Serialize PROV-O trace + SHA-256
Total
∼440
100
Single laptop, no GPU
A PPENDIX C ECG D ETECTION P ERFORMANCE BY D IFFICULTY T IER Table III disaggregates the ECG AUC-ROC and AUC-PR results (averaged across all 20 MIT-BIH records) by the three
E1 difficulty tiers. AUC-PR increases monotonically from Easy A PPENDIX E to Hard for all detectors because Hard-tier injections occupy a M3 ROLLING -W INDOW A BLATION – F ULL N UMERIC larger fractional window, reducing effective class imbalance. R ESULTS AUC-ROC is highest on Easy for AE and LOF, reflecting their Table IV gives the complete per-detector AUC-ROC and superior detection of the largest-amplitude faults; IF follows AUC-PR values. Results are averaged over all fault types and the same trend. RobustRollingZScore is near-random (≈ 0.49) three difficulty tiers for 5 MIT-BIH ECG records. across all tiers, confirming its insensitivity to the varied ECG artifact morphologies. TABLE IV M3 W INDOW A BLATION – M EAN AUC-ROC / AUC-PR (ECG, 5 R ECORDS )
TABLE III ECG AUC-ROC / AUC-PR BY D IFFICULTY T IER (20 R ECORDS , 6 FAULT T YPES ) Tier
Det.
Easy
AE IF LOF RRZS ZS
0.943 0.906 0.947 0.482 0.809
0.762 0.638 0.729 0.153 0.441
Medium AE IF LOF RRZS ZS
0.927 0.889 0.921 0.483 0.789
Hard
0.876 0.844 0.886 0.495 0.757
AE IF LOF RRZS ZS
AUC-ROC AUC-PR Recall
AUC-ROC (AUC-PR)
FPR
F1
0.806 0.552 0.841 0.055 0.242
0.072 0.037 0.113 0.028 0.048
0.674 0.517 0.631 0.075 0.218
0.861 0.793 0.851 0.399 0.671
0.768 0.496 0.798 0.059 0.219
0.087 0.044 0.135 0.028 0.048
0.753 0.514 0.763 0.094 0.239
0.888 0.854 0.899 0.630 0.799
0.660 0.388 0.701 0.062 0.160
0.078 0.043 0.097 0.029 0.043
0.690 0.430 0.742 0.104 0.202
AE=Autoencoder, IF=IsolationForest, RRZS=RobustRollingZScore, ZS=ZScore. Bold: best AUC-ROC and AUC-PR per tier. Averaged across 20 records, 6 fault types.
w
AE
IF
LOF
RRZS
ZS
15 30 50 100
0.775 (0.651) 0.797 (0.671) 0.793 (0.671) 0.807 (0.694)
0.707 (0.554) 0.758 (0.594) 0.803 (0.647) 0.816 (0.671)
0.854 (0.768) 0.869 (0.777) 0.868 (0.773) 0.876 (0.780)
0.449 (0.370) 0.452 (0.371) 0.462 (0.382) 0.464 (0.388)
0.688 (0.523) 0.756 (0.575) 0.794 (0.612) 0.806 (0.625)
A PPENDIX F M5 PROV-O RUNTIME P ROVENANCE T RACE Fig. 4 shows the PROV-O/ProvONE runtime trace produced by M5 for the satellite anomaly detection use case, parsed directly from anomaly_provenance.ttl. Three prov:Activity nodes, E1 (fault injection), E2 (anomaly detection), and M4 (semantic annotation), are arranged in execution order with their input and output prov:Entity nodes. SHA-256 checksums are stored as telwf:sha256 triples for all file-path entities. The trace is fully queryable via any SPARQL 1.1 endpoint.
A PPENDIX D M4 K NOWLEDGE G RAPH SPARQL Q UERY E XAMPLES The following queries illustrate the post-hoc analytical value of the M4 knowledge graph without re-executing any ML detector. Q1: Which satellite subsystems carry the strongest anomaly evidence? PREFIX if: <http://sensorwf.org/interpretability#> PREFIX sat: <http://sensorwf.org/satellite#> SELECT ?subsystem (SUM(?imp) AS ?total_evidence) WHERE { ?feat if:belongsToSubsystem ?subsystem ; if:featureImportance ?imp . } GROUP BY ?subsystem ORDER BY DESC(?total_evidence)
Fig. 4. PROV-O runtime provenance trace for the satellite anomaly detection use case, parsed from anomaly_provenance.ttl. Blue boxes: prov:Activity nodes (E1 fault injection, E2 anomaly detection, M4 semantic annotation). Green boxes: input prov:Entity nodes (raw telemetry files, parameters). Orange boxes: output prov:Entity nodes (injected Q2: Top-5 features by IsolationForest importance for gyro variants, ML evaluation JSON, knowledge graph CSVs). Arrows show prov:used (green) and prov:wasGeneratedBy (orange) relations. SHAfaults: 256 checksums are stored for all file-path entities. PREFIX if: <http://sensorwf.org/interpretability#> SELECT ?feat ?cls ?imp WHERE { if:tag:A2_gyro_clipping if:importantFeature ?feat . ?feat if:featureImportance ?imp ; if:mapsToSensorClass ?cls . } ORDER BY DESC(?imp) LIMIT 5
These queries execute against the RDF/Turtle graph (if_ontology_graph.ttl) produced by M4 using any SPARQL 1.1 endpoint (e.g., Apache Jena Fuseki, RDFLib). No Python or ML libraries are required for post-hoc subsystem attribution.