Este archivo ha sido creado el 20-01-2025 por Juan Manuel Montes Sánchez GENERAL INFORMATION ------------------ 1. Dataset title: Neuromorphic Audio for Predictive Maintenance in Peristaltic Pumps 2. Authorship: Name: Juan Manuel Montes Sánchez Institution: Universidad de Sevilla Email: juanmanuelmontes@us.es ORCID: 0000-0002-0983-2386 Name: Juan Pedro Domínguez Morales Institution: Universidad de Sevilla Email: jpdominguez@us.es ORCID: 0000-0002-5474-107X Name: Saturnino Vicente Díaz Institution: Universidad de Sevilla Email: satur@us.es ORCID: 0000-0001-9466-485X Name: Ángel Francisco Jiménez Fernández Institution: Universidad de Sevilla Email: angel@us.es ORCID: 0000-0003-3061-5922 DESCRIPTION ---------- 1. Dataset language: English 2. Abstract: This dataset contains processed audio samples coming from a hydraulic block from a biomedical equipment. The block mounts 3 Thomas SR10/30 DC standard perisltaltic pumps, which were filled with distilled water. Only one pump was running at the same time during these recordings. There are two different predictive maintenance scenarios. In the first one, the cassettes of the pumps were changed before each recording. We used cassettes with 2 different levels of degradation: NEW (unused) and OLD (lifetime already expired). We defined 3 different classes: Class 1 is STOP (no pump running), class 2 is NEW (one pump running with a new cassette), and class 3 is OLD (one pump running with an old cassette). In the second scenario, air bubbles were introduced into the tube. This second scenario also has 3 classes: Class 1 is STOP (no pump running), class 2 is NORMAL (no air bubbles), and class 3 is BUBBLE (air bubbles present). A single microphone was used for all recordings. The .wav audio files were processed using a 64 channel Neuromorphic Auditory Sensor (NAS) into .aedat files, which are the present in this dataset. This neuromorphic audio data were also converted into cochleogram images using the software pyNAVIS, and they are also present in this format (.png files). Este dataset contiene muestras procesadas de audio de un bloque hidráulico de un equipo biomédico. El bloque monta 3 bombas peristálticas estándar de corriente continua modelo Thomas SR10/30, que fueron llenadas con agua destilada. Solo una bomba estuvo funcionando a la vez durante la recogida de datos. Hay dos escenarios de mantenimiento predictivo diferentes. En el primero, los cassettes de las bombas fueron cambiados antes de cada recogida de datos. Usamos cassettes con dos niveles distintos de degradación: NEW (sin previo uso) y OLD (vida útil agotada). Definimos 3 clases diferentes: la clase 1 es STOP (ninguna bomba funcionando), la clase 2 es NEW (una bomba funcionando con un cassette nuevo) y la clase 3 es OLD (una bomba funcionando con un cassete viejo). En el segundo escenario, se introdujeron burbujas de aire en el tubo. Este segundo escenario también tiene 3 clases: la clase 1 es STOP (ninguna bomba funcionando), la clase 2 es NORMAL (sin burbujas de aire) y la clase 3 es BUBBLE (con burbujas de aire). Un solo micrófono fue usado para todas las grabaciones. Los archivos de audio .wav fueron procesados usando un Sensor Neuromórfico de Audición (NAS) en archivos .aedat, que están presentes en este set de datos. Estos datos de audio neuromórfico también fueron convertidos en imágenes tipo cocleograma usando el software pyNAVIS, y también están presentes en este formato (archivos .png). 3. Keywords: Neuromorphic, Neuromórfico, Microphone, Micrófono, Audio, IIoT, peristaltic pump, bomba peristáltica, predictive maintenance, mantenimiento predictivo 4. Date of data collection: 28-09-2024 5. Publication Date: 10-02-2025 6. Grant information: Grant Agency: Agencia Estatal de Investigación - Ministerio de Ciencia, Innovación y Universidades (Gobierno de España) Grant Number: Proyecto PREDICAR (ref. PID2023-149777OB-I00) 7. Geographical location/s of data collection: Universidad de Sevilla, Sevilla, SPAIN ACCESS INFORMATION ------------------------ 1. Creative Commons License of the dataset: CC BY-NC 4.0 2. Dataset DOI: 10.12795/11441/162880 3. Related publication: Title: Neuromorphic Audio Dataset for Predictive Maintenance in Peristaltic Pumps DOI: PENDIENTE 4. Link to related datasets: DOI/URL: 10.12795/11441/162880 VERSIONING AND PROVENANCE --------------- 1. Last modification date: 28-09-2024 for data, 21-01-2025 for this README 2. Were data derived from another source?: No 3. Additional related data not included in this dataset: DOI/URL: 10.12795/11441/162880 METHODOLOGICAL INFORMATION ----------------------- 1. Description of the methods used to collect and generate the data: This dataset contains processed audio samples coming from a hydraulic block from a biomedical equipment. The block mounts 3 Thomas SR10/30 DC standard perisltaltic pumps, which were filled with distilled water. Only one pump was running at the same time during these recordings. A single microphone was used for all recordings. There are two different predictive maintenance scenarios: AGING SCENARIO: The cassettes of the pumps were changed before each recording. We used cassettes with 2 different levels of degradation: NEW (unused) and OLD (lifetime already expired). We defined 3 different classes: Class 1 is STOP (no pump running), class 2 is NEW (one pump running with a new cassette), and class 3 is OLD (one pump running with an old cassette). BUBBLE SCENARIO: Air bubbles were introduced into the tube. Class 1 is STOP (no pump running), class 2 is NORMAL (no air bubbles), and class 3 is BUBBLE (air bubbles present). 2. Data processing methods: The raw audio data has already been processed into sepparate different .wav files using python code. These .wav files were also processed with a 64 channel Neuromorphic Auditory Sensor (NAS) into .aedat files (AER encoding). The .aedat files were finally converted into .png cochleogram images using pyNAVIS software. The cochleogram images are 0.5s each, with 50% overlapping, and have been normalized. 3. Software or instruments needed to interpret the data: Any audio player for .wav format, any image viewer for .png files, pyNAVIS or NAVIS (https://github.com/jpdominguez/pyNAVIS) for .aedat files. 4. Information about instruments, calibration and standards: Information for Neuromorphic Auditory Sensor at https://github.com/RTC-research-group/OpenNAS. For this dataset, OpenNAS configuration was: mono, cascade, 64 channel, target FPGA device ZTEX USB-FPGA Module 2.13. Rest of values left by default. 5. Environmental or experimental conditions: Closed room. No external equipment running besides lighting, power supply and PC. No external audible noises like human voices. 6. Quality-assurance procedures performed on the data: Empty or non valid values removed, if any. FILE OVERVIEW ---------------------- 1. Explain the file naming conversion: There is one folder for each of the two scenarios (aging and bubble). Inside those folders there are 3 subfolders, one for each data type (AEDAT, PNG and WAV). For WAV and AEDAT, each file is a unique sample tagged with one class (1, 2 or 3). At the end of each filename this class is also included. For example, 0008_03.wav is the sample number 8, which corresponds to class 3 tagged data. PNG cochleogram images represent 500ms audio time each and are named after their source AEDAT file followed by their starting time mark in microseconds. 2. File list: This dataset includes a large ammount of individual PNG image files (1640 for aging scenario and 1252 for bubble scenario), but also their corresponding source WAV and AEDAT files, in separate folders. Better see (1) for naming convention and (5) for folder structure. 3. Relationship between files: all samples with the same sample number come from the .wav file of the same name. 4. File format: .wav format for audio data, .aedat format for AER encoded neuromorphic files (result of processing each .wav file with OpenNAS's Neuromorphic Auditory Sensor), and .png images for cochleogram data (result of processing each .aedat file with pyNAVIS software). 5. If the dataset includes multiple files, specify the directory structure and relationships between the files: There is one folder for each of the two scenarios. Each scenario folder contains 3 subfolders, each one for one type of data (.aedat, .png and .wav). The structure is as follows: -DATA | |-dataset.aging | | | |- aedats.aging | | | | | |- 1_STOP_aedats | | | | | | | |- 8 .aedat files (STOP class) | | | | | |- 2_NEW_aedats | | | | | | | |- 4 .aedat files (NEW class) | | | | | |- 3_OLD_aedats | | | | | | | |- 6 .aedat files (OLD class) | | | |- plots.aging | | | | | |- 1_STOP_plots | | | | | | | |- 8 subfolders with .png files inside (STOP class) | | | | | |- 2_NEW_plots | | | | | | | |- 4 subfolders with .png files inside (NEW class) | | | | | |- 3_OLD_plots | | | | | | | |- 6 subfolders with .png files inside (OLD class) | | | |- wavs.aging | | | | | |- 1_STOP_wavs | | | | | | | |- 8 .wav files (STOP class) | | | | | |- 2_NEW_wavs | | | | | | | |- 4 .wav files (NEW class) | | | | | |- 3_OLD_wavs | | | | | | | |- 6 .wav files (OLD class) | | |-dataset.bubble | | | |- aedats.bubble | | | | | |- 1_STOP_aedats | | | | | | | |- 5 .aedat files (STOP class) | | | | | |- 2_NORMAL_aedats | | | | | | | |- 4 .aedat files (NORMAL class) | | | | | |- 3_BUBBLE_aedats | | | | | | | |- 4 .aedat files (BUBBLE class) | | | |- plots.bubble | | | | | |- 1_STOP_plots | | | | | | | |- 5 subfolders with .png files inside (STOP class) | | | | | |- 2_NORMAL_plots | | | | | | | |- 4 subfolders with .png files inside (NORMAL class) | | | | | |- 3_BUBBLE_plots | | | | | | | |- 4 subfolders with .png files inside (BUBBLE class) | | | |- wavs.bubble | | | | | |- 1_STOP_wavs | | | | | | | |- 5 .wav files (STOP class) | | | | | |- 2_NORMAL_wavs | | | | | | | |- 4 .wav files (NORMAL class) | | | | | |- 3_BUBBLE_wavs | | | | | | | |- 4 .wav files (BUBBLE class) MORE INFORMATION -------------- For the aging scenario, the following sample numbers are recommended for the holdout dataset when performing holdout validation: 0000, 0001, 0002, 0008, 0009 and 0010. This data has been taken using unique cassetes that are not present in the rest of the samples. For the bubble scenario, any set of three files with a different identifier and class is suitable as holdout set.