DAYPSCI: Event-based dataset for anomaly detection through fault injection in PLC-controlled industrial cyber-physical systems Authors: Juan Vicente Martín-Fraile, Nuño Basurto Hornillos, Jesús Enrique Sierra-García, Álvaro Herrero Published: 2026 ABSTRACT AbstractThis article presents DAYPSCI, an event-based dataset generated using an industrial cyber-physical system (CPS) testbed based on a PLC-controlled part marking station with Siemens S7-1200 and S7-1500 devices. The system integrates real industrial hardware with digital twin technologies, enabling controlled and repeatable experiments. Data acquisition follows an event-based logging approach, where only changes in system variables are recorded rather than using fixed sampling rates, and each event is associated with its corresponding inter-event time (Δt), enabling precise temporal characterization of system dynamics. The dataset also includes scan-level identifiers (scan_id) and event ordering (event_order), preserving the logical execution order within PLC scan cycles. It contains time-stamped records of digital sensors, solenoid valve control signals, and process states under both normal operation and controlled fault injection scenarios affecting sensors, actuators, or both. Ground truth is generated through a hybrid approach combining externally defined labels from the experimental configuration and labels derived at runtime from control system signals (e.g., GEMMA states), ensuring a clear separation between CPS execution and the labeling layer while enabling traceability between injected faults (cause) and observable system behavior (effect), and allowing differentiation between fault activation and observable anomaly manifestation, which may be temporally decoupled. The dataset is organized into independent experimental batches, each including processed data (CSV), network traffic captures (PCAPNG) from a Profinet-based industrial communication network, and detailed documentation, facilitating sequence-based analysis and reproducibility. The dataset supports the development and evaluation of machine learning methods for anomaly detection and fault classification in industrial CPS.Keywords: Event-based logging, Digital twin, Ground truth labeling, Discrete-event systems, PLC scan cycles, Industrial cybersecurity, Fault injection, Sensor and actuator faults Specifications Table SubjectComputer SciencesSpecific subject areaAnomaly detection and fault diagnosis in industrial cyber-physical systems using event-based data and PLC execution modelling.Type of dataTabular data (event-based time series); processed data (CSV); raw network traffic data (PCAPNG); labeled data; discrete-event sequences.Data collectionData were collected from an industrial cyber-physical testbed comprising Siemens S7-1200 and S7-1500 PLCs, digital sensors, and solenoid valves. Events were generated directly at the PLC level using an event-driven logging strategy. Each event was timestamped using the PLC system time (RD_SYS_T, UTC), assigned a monotonic sequence identifier, and stored in an internal ring buffer before being consumed and inserted into an SQL Server database. Derived temporal features, such as inter-event time (Δt), were computed using SQL views without modifying the original data. Scan-level execution was captured by assigning a scan identifier and ordering events within each PLC cycle, enabling reconstruction of the logical execution order of the control system.Data source locationUniversidad de Burgos, Av. Cantabria s/n, 09006 Burgos, Spain.Data accessibilityRepository name: Institutional Repository of the University of BurgosData identification number: https://doi.org/10.71486/2drf-1g53Direct URL to data: https://hdl.handle.net/10259/11686Instructions for accessing these data: The dataset is publicly available and can be accessed directly through the provided DOI or repository link without restrictions.Related research articleJ.V. Martín-Fraile, N. Basurto, J.E. Sierra-García, Álvaro Herrero. A testbed for evaluation of machine learning techniques in anomaly detection in industrial cyber-physical systems, DYNA, vol. 101, no. 2, 2026. Value of the Data •The dataset provides high-resolution, event-based records of an industrial cyber-physical system, capturing discrete transitions of sensors, solenoid valve actuators, and control states directly at the PLC level. Unlike time-sampled logs, the data preserve both temporal information (Δt) and the logical execution order of events through scan-level identifiers (scan_id) and event ordering, enabling detailed temporal and logical analysis of industrial processes. These variables are rarely available in conventional industrial datasets and provide a unique representation of temporal dynamics, PLC execution semantics, GEMMA-based operational states, and fault-injection behaviour. Furthermore, the integration of synchronized process data and industrial network traffic enables multi-level analysis of industrial CPS operation and anomalies. The dataset provides high-resolution, event-based records of an industrial cyber-physical system, capturing discrete transitions of sensors, solenoid valve actuators, and control states directly at the PLC level. Unlike time-sampled logs, the data preserve both temporal information (Δt) and the logical execution order of events through scan-level identifiers (scan_id) and event ordering, enabling detailed temporal and logical analysis of industrial processes. These variables are rarely available in conventional industrial datasets and provide a unique representation of temporal dynamics, PLC execution semantics, GEMMA-based operational states, and fault-injection behaviour. Furthermore, the integration of synchronized process data and industrial network traffic enables multi-level analysis of industrial CPS operation and anomalies. •The data include both normal operation and controlled fault injection scenarios affecting sensors and solenoid valves, with ground truth labels generated independently from the industrial control logic. This enables reproducible evaluation of anomaly detection and fault diagnosis methods under realistic industrial conditions, including class imbalance and timing constraints. The data include both normal operation and controlled fault injection scenarios affecting sensors and solenoid valves, with ground truth labels generated independently from the industrial control logic. This enables reproducible evaluation of anomaly detection and fault diagnosis methods under realistic industrial conditions, including class imbalance and timing constraints. •The dataset is valuable for researchers in industrial automation, cyber-physical systems, and data-driven monitoring, as it represents the behavior of a real industrial process implemented on commercial Siemens S7 PLCs. The inclusion of process states based on the GEMMA model enables analysis of system dynamics at both process and operational levels. The dataset is valuable for researchers in industrial automation, cyber-physical systems, and data-driven monitoring, as it represents the behavior of a real industrial process implemented on commercial Siemens S7 PLCs. The inclusion of process states based on the GEMMA model enables analysis of system dynamics at both process and operational levels. •The dataset supports machine learning and data-driven approaches, enabling sequence-based modeling, temporal feature extraction, and evaluation of event-driven algorithms. Its structure facilitates the development of models that exploit both timing and ordering of events, which are typically unavailable in conventional time-sampled datasets. The dataset supports machine learning and data-driven approaches, enabling sequence-based modeling, temporal feature extraction, and evaluation of event-driven algorithms. Its structure facilitates the development of models that exploit both timing and ordering of events, which are typically unavailable in conventional time-sampled datasets. •The dataset can be reused as a benchmark for anomaly detection, fault classification, and industrial cybersecurity applications, particularly those based on event logs or discrete-event representations. The availability of synchronized process data (CSV) and network traffic captures (PCAPNG format) enables joint analysis of control behavior, process evolution, and industrial communication patterns within a consistent experimental CPS environment. The dataset can be reused as a benchmark for anomaly detection, fault classification, and industrial cybersecurity applications, particularly those based on event logs or discrete-event representations. The availability of synchronized process data (CSV) and network traffic captures (PCAPNG format) enables joint analysis of control behavior, process evolution, and industrial communication patterns within a consistent experimental CPS environment. Furthermore, the dataset supports reproducible benchmarking and comparative evaluation of machine learning, anomaly detection, and cybersecurity methods, providing a valuable resource for researchers and educators working on Industry 4.0 and industrial CPS applications. Background The increasing adoption of cyber-physical systems (CPS) in industrial environments has created a growing demand for datasets that accurately represent the behavior of discrete-event processes under both normal and faulty conditions. In this context, recent studies have highlighted the importance of data representation in anomaly detection for industrial CPS, particularly distinguishing between time-sampled and event-based approaches [1]. In many industrial applications, data are commonly collected using fixed sampling rates, which may fail to capture the precise timing and ordering of events occurring within PLC-based control systems. Furthermore, existing research has pointed out limitations in currently available datasets, including restricted diversity, limited representation of industrial processes, and challenges in capturing realistic cyber–physical interactions. These limitations hinder the evaluation and generalization of anomaly detection methods across different industrial scenarios. To address these challenges, this dataset was created to provide realistic, event-based data acquired directly from industrial control systems. The dataset focuses on an event-based data acquisition strategy that captures both the temporal dynamics and logical execution order of industrial processes, enabling a more accurate representation of discrete-event behavior in PLC-controlled systems. The data were generated using a PLC-controlled part marking station integrating Siemens S7-1200 and S7-1500 devices, combined with digital twin technologies to enable controlled and repeatable experimentation. The data acquisition process follows an event-based logging paradigm, where only changes in system variables are recorded, and each event is associated with a duration (Δt). The dataset design is aligned with discrete-event system modeling approaches commonly used in industrial automation. In addition, system operation is described using process states based on the GEMMA reference model, enabling the representation of normal operation, fault conditions, and recovery phases. Controlled fault injection was applied to sensors and solenoid valves to generate labeled anomaly scenarios under consistent experimental conditions. This data article provides a detailed description of the dataset structure, acquisition methodology, and labeling strategy, complementing related research focused on anomaly detection and machine learning applications in industrial CPS. Data Description The DAYPSCI dataset described in this article is organized as a collection of independent experimental batches and is publicly available at [2], each corresponding to a specific operating condition of the industrial cyber-physical system. The dataset repository is structured into batches, where each batch represents a complete and self-contained experiment, as illustrated in Fig. 1. These experiments include baseline operation (normal behavior) and controlled fault injection scenarios affecting sensors, solenoid valves, or both. Directory structure of the DAYPSCI_v1 dataset. Each batch contains CSV, PCAPNG, and README files. Each batch contains three main files: •A CSV file containing the processed event-based dataset. This file includes system variables, temporal features, labels, and scan-level information. A CSV file containing the processed event-based dataset. This file includes system variables, temporal features, labels, and scan-level information. •A PCAPNG file containing raw network traffic captures collected from the industrial Profinet-based communication network during the experiment. A PCAPNG file containing raw network traffic captures collected from the industrial Profinet-based communication network during the experiment. •A Markdown (.md) file providing a detailed description of the experimental setup and the specific fault configuration associated with the batch. A Markdown (.md) file providing a detailed description of the experimental setup and the specific fault configuration associated with the batch. The naming convention follows a structured format that encodes the type of scenario and the batch identifier. For example: pn_baseline_batch01.csv, pn_sensor_T1_batch02.csv, pn_valve_T1_batch08.csv, pn_combined_T1_batch11.csv (combined sensor–solenoid valve faults). Within each CSV file, the dataset is organized as a sequence of events. Each row corresponds to a system event, defined as a change in one or more variables. Events are grouped into sequences associated with individual processed parts, identified by the field part_id. Table 1 summarizes the variables included in the dataset along with their types and descriptions. Table 1Dataset variables and description.Table 1 dummy alt textLabelTypeDescriptionevent_seqIntegerEvent sequence indexts_plcDatetimePLC timestampdelta_time_msIntegerTime between consecutive events (Δt) in millisecondsscan_idIntegerIdentifier of the PLC scan cycleevent_orderIntegerExecution order within the scan cycleevent_idIntegerUnique event identifierevent_typeBinaryEdge type: rising (1) or falling (0)part_idIntegerPart identifierGemmaIntegerGEMMA state (encoded)AnomalyIntegerAnomaly labelFaultIntegerFault presence indicatorFSxIntegerSensor fault indicator (per sensor channel)TFSxIntegerSensor fault type descriptorFCVxIntegerActuator fault indicator (per actuator channel)TFCVxIntegerActuator fault type descriptorsensor signalsBinaryBinary states of sensors (a0–c1, B1, B2)actuators signalsBinaryBinary states of actuators (YA_p, YA_m, YB_p, YB_m, YC_p, YC_m)batch_idIntegerIdentifier of the experimental batch (fault configuration scenario)experiment_idIntegerIdentifier of the experiment instance Dataset variables and description. The dataset follows an event-based representation, where events are recorded only when a change occurs in the system. Each event is associated with a duration (Δt), representing the time elapsed since the previous event. Scan-level execution is explicitly represented through the fields scan_id and event_order. These variables indicate the PLC scan cycle and the execution order of events within the same cycle, enabling the reconstruction of the logical behavior of the control system, including cases where multiple events occur within a single scan. The dataset includes temporal variables (event sequence, timestamps, Δt), process variables (system states based on the GEMMA model), discrete state indicators, solenoid valve control signals, and labeling fields. Labels describe both the presence of anomalies (anomaly) and the occurrence of faults (fault), as well as detailed fault descriptors indicating the affected component and fault type. Table 2 defines the anomaly and fault labels used in the dataset. Table 2Anomaly and fault label definitions.Table 2 dummy alt textLabelValueDescriptionAnomaly0No anomaly injected (normal operation)Anomaly1Sensor fault injectedAnomaly2Actuator fault injectedAnomaly3Combined fault injected (sensor + actuator)Anomaly4Operation indirectly affected by a faultFault0No faultFault1Fault present Anomaly and fault label definitions. The data are structured to enable sequence-level analysis, as each processed part is represented as an ordered set of events. The number of events per part may vary depending on system dynamics. In addition to process data, the PCAPNG files provide synchronized network traffic captures corresponding to the same experiments. This enables joint analysis of industrial communication patterns and control system behavior within the same experimental scenario. All batches are generated under consistent experimental conditions, including the same hardware configuration, communication infrastructure, and control logic, enabling direct comparison across different scenarios. Experimental Design, Materials and Methods The dataset was generated using an industrial cyber-physical system (CPS) testbed based on a PLC-controlled part marking station, previously described and validated in [3]. The testbed integrates Siemens S7-1200 and S7-1500 programmable logic controllers, digital sensors, and electromechanical actuators. A digital twin environment was incorporated to enable controlled, repeatable, and extensible experimentation. All data included in the DAYPSCI dataset were generated by the authors through controlled experimental campaigns conducted on the proposed CPS testbed. No external datasets or third-party sources were used for dataset generation. To improve clarity and reproducibility, Fig. 2 illustrates the architecture of the CPS-based testbed and the data acquisition pipeline used for dataset generation. CPS-based testbed architecture for event-based dataset generation. Control logic and process modeling The control program was implemented following standard PLC programming practices [4], based on cyclic scan execution. The system behavior was modeled using a discrete-event approach, representing the industrial process as a sequence of state transitions. The operational workflow was defined according to the GEMMA guideline [5], explicitly modelling modes such as automatic operation, fault occurrence, and recovery sequences. This approach ensures a structured and reproducible representation of system dynamics under both nominal and faulty conditions. Data acquisition strategy Data acquisition was performed at the PLC level using an event-based approach. Events were generated upon changes in relevant system variables, including sensor signals, actuator states, and internal control variables. Each event includes a timestamp obtained from the PLC system time (RD_SYS_T, UTC), a unique sequential identifier, and the corresponding variable change. Events were temporarily stored in an internal ring buffer within the PLC memory and periodically transmitted to an external storage system. Fig. 3 illustrates the event-based data acquisition and transformation process, improving the clarity and interpretability of the dataset. Raw binary signals from sensors and actuators, together with process-state information based on the GEMMA methodology, are monitored at the PLC level, where events are generated through rising and falling edge detection. Event-based data acquisition pipeline with SQL-based Δt computation. Events are temporarily stored in a circular buffer and subsequently transferred to a SQL Server database, where the time between consecutive events (Δt) is computed to obtain a compact and temporally meaningful representation of system behavior. The buffer is implemented within the PLC memory and consumed by a dedicated acquisition application, decoupling event acquisition from data storage operations and preventing event loss caused by communication or database access delays. The dataset generation workflow follows a structured process. First, experimental conditions and fault-injection scenarios are defined through the supervisory layer. The CPS testbed then executes the corresponding production cycles using Siemens PLCs and the digital twin environment. In addition to process data acquisition, the CPS architecture illustrated in Fig. 2 enables the collection of industrial network traffic through a managed switch configured for traffic mirroring. Using a Wireshark-based acquisition system connected to the PROFINET network, the network traffic is captured and stored in PCAPNG format. The resulting traffic traces are linked to the corresponding experimental batch, enabling joint analysis of process events, system behavior, and network traffic. Finally, processed datasets (CSV files), network traffic captures (PCAPNG files), and metadata describing the experimental configuration are organized into independent experimental batches and made publicly available through the institutional repository. Data storage and feature engineering All data were stored in an SQL Server database. The original event records were preserved without modification to ensure data integrity. Additional temporal features, such as inter-event time (Δt), were computed through SQL-based post-processing. To capture execution semantics at scan level, each event was annotated with: scan_id (identifier of the PLC scan cycle) and event_order (relative execution order within the scan). This allows the representation of multiple events occurring within the same scan cycle while preserving the logical execution order of the control program. Fault injection and experimental conditions Controlled fault injection was applied to generate anomalous operating conditions by introducing faults in digital sensors and solenoid valves. The dataset includes multiple fault types, comprising six sensor fault scenarios, three solenoid valve fault configurations, and one combined fault conditions affecting both components, providing representative coverage of common fault conditions encountered in PLC-controlled automated systems. To ensure reproducible experimental campaigns, a dedicated fault-planning tool was developed. The tool allows researchers to configure batch size, random seed, class distribution, candidate sensors and actuators, and fault-injection parameters before data acquisition. Based on this configuration, a fault-injection plan is automatically generated, specifying the affected parts, target components, fault types, and associated parameters for each experimental batch. Fig. 4 illustrates the configuration and planning stage of the fault-injection campaigns. The tool distinguishes between eligible components that may be affected by faults and the actual components selected for a given experimental batch, ensuring controlled and reproducible fault distributions. Once the experimental plan is generated, the corresponding fault scenarios are applied to the CPS testbed, as illustrated in the next Figure. Experimental planning and configuration of fault-injection campaigns. Fig. 5 presents the fault injection and labeling workflow used to generate the labeled dataset. Faults are externally configured and applied to the system, triggering transitions through GEMMA states. The labeling process assigns anomaly labels, fault indicators, and detailed fault descriptors, including the affected component (e.g., sensor or actuator) and fault type, enabling a fine-grained characterization of the system behavior. Workflow for dataset labeling based on fault injection and GEMMA state transitions. Faults were configured externally to the control logic, ensuring that the PLC program remained unchanged during all experiments. This separation ensures that anomalies are externally defined and not derived from the control logic, providing a clear distinction between process execution and labeling. Experiments were conducted under consistent conditions, including identical hardware configuration, communication settings, and control logic. Ground truth labeling Ground truth labeling is generated through a hybrid approach, combining externally defined labels from the experimental configuration and labels derived at runtime from control system signals, as illustrated in Fig. 5. The experimental configuration, defined by a higher-level supervisory layer, specifies the intended fault scenarios, while runtime signals from the PLC (e.g., GEMMA states and system responses) capture the observable effects of these faults. During data acquisition, configuration parameters are associated with each event, enabling the integration of planned fault information with runtime behavior. This allows each event to be labeled according to anomaly type, fault occurrence, affected component, and fault characteristics. This approach ensures a clear separation between the industrial process (CPS) and the labeling layer, while enabling traceability between injected faults (cause) and observable system behavior (effect). Furthermore, it enables differentiation between fault activation and observable anomaly manifestation, which may be temporally decoupled. Network data acquisition Network traffic was captured simultaneously during each experiment using Wireshark (version 4.4.5) from a mirrored network interface in a Profinet-based industrial communication environment and stored in PCAPNG format for subsequent analysis. The captured traffic includes cyclic communication and industrial network interactions between PLCs and field devices operating over Profinet. Dataset organization The dataset is structured into independent experimental batches, as illustrated in Fig. 1. Each batch corresponds to a specific operating condition, including nominal operation and fault injection scenarios, and contains the corresponding processed event-based dataset (CSV), synchronized industrial network traffic captures (PCAPNG), and a README.md file describing the associated experimental configuration. This organization facilitates reproducibility and supports benchmarking of anomaly detection, fault diagnosis, machine-learning, and cybersecurity methods. Limitations The dataset was generated using an industrial CPS testbed based on a PLC-controlled part marking station, and therefore reflects its specific characteristics, architecture, and operational behavior, potentially limiting generalization to other industrial systems. Experiments were conducted under controlled conditions with deterministic fault injection, ensuring reproducibility and precise ground truth labeling, but potentially not capturing the full variability of real production environments. Furthermore, the range of fault types and experimental scenarios is limited to a predefined set of sensor and actuator faults. Other types of anomalies, such as complex multi-component failures or network-based attacks, are not included in the current version of the dataset. Despite these limitations, the dataset provides a consistent and well-controlled representation of industrial process behavior under both nominal and faulty conditions, enabling systematic experimentation and benchmarking. CRediT Author Statement Juan Vicente Martín-Fraile: Conceptualization, Methodology, Software, Data curation, Formal analysis, Investigation, Visualization, Writing – original draft, Writing – review & editing, Resources, Validation, Supervision, Project administration. Nuño Basurto: Supervision, Writing – review & editing. Jesús Enrique Sierra-García: Supervision, Writing – review & editing. Álvaro Herrero: Supervision, Writing – review & editing. Ethics Statement The authors confirm that they have read and adhere to the ethical requirements for publication in Data in Brief. The current work does not involve human subjects, animal experiments, or any data collected from social media platforms. Acknowledgements The funding for this work was provided by AI4SECIoT project (``Artificial Intelligence for Securing IoT Devices'' - C032.23), funded by the National Cibersecurity Institute (INCIBE), derived from a collaboration agreement signed between the National Institute of Cybersecurity (INCIBE) and the University of Burgos. This initiative is carried out within the framework of the Recovery, Transformation and Resilience Plan funds, financed by the European Union (Next Generation), the project of the Government of Spain that outlines the roadmap for the modernization of the Spanish economy, the recovery of economic growth and job creation, for solid, inclusive and resilient economic reconstruction after the COVID19 crisis, and to respond to the challenges of the next decade. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.dib.2026.113146. Supplementary Materials Data Availability RIUBU – Universidad de Burgos RepositoryDAYPSCI v1: An Event-Based Dataset of Fault Injection Scenarios in PLC-Controlled Industrial Cyber-Physical Systems (Original data) RIUBU – Universidad de Burgos RepositoryDAYPSCI v1: An Event-Based Dataset of Fault Injection Scenarios in PLC-Controlled Industrial Cyber-Physical Systems (Original data) References Associated Data Supplementary Materials Data Availability Statement RIUBU – Universidad de Burgos RepositoryDAYPSCI v1: An Event-Based Dataset of Fault Injection Scenarios in PLC-Controlled Industrial Cyber-Physical Systems (Original data) RIUBU – Universidad de Burgos RepositoryDAYPSCI v1: An Event-Based Dataset of Fault Injection Scenarios in PLC-Controlled Industrial Cyber-Physical Systems (Original data)