Acoustic Signal Dataset of Homogeneous Balls under Steady Excitation in Sealed Containers

By analyzing the characteristics of collision acoustic signals, a machine learning model is established to achieve high-precision object counting. The following is a detailed description of the dataset: Data Collection and Structure Experimental setup: Containers and objects: Use containers made of different materials (wood, aluminum, acrylic) and small balls (glass, wood, quartz) to simulate various physical interaction conditions. Excitation device: By using a circular motion vibration table with a constant speed to excite the movement of the ball, data is collected after the system reaches steady state. Signal recording: A standard microphone (with a sampling rate of 44.1 kHz) records collision sounds for several seconds.

During training, it is divided into 3-second samples, and each sample is stored as a 132300 dimensional time-domain signal. Data composition: Categories and Samples: There are a total of 36 main categories (with ball sizes ranging from 1 to 36), each containing ≥ 200 samples, resulting in a total of 7263 audio files. Control group expansion: includes multiple sets of experimental data (such as aluminum box+glass ball, acrylic box+quartz ball, etc.), with a total sample size of 22133. Data Features and Preprocessing Time domain and frequency domain analysis: The original time-domain signal is transformed into frequency-domain features through Fast Fourier Transform (FFT), enhancing inter class linear separability. Principal Component Analysis (PCA) is used for dimensionality reduction (to 2D/3D) to reveal the high-dimensional spherical distribution characteristics of data. Key physical characteristics: The total amount of sound energy within a single category (Eeergy=∑ xi ^ 2) remains constant and shows an increasing and then decreasing trend with the number of balls. The frequency domain features exhibit strong linear separability, supporting classification models based on logistic regression. Application and Value Target task: Supervised learning (classification), predict the number of small balls in a closed container, and achieve recognition accuracy of over 95% within an allowable error range (such as ± 1). Advantages: Cross material and size robustness: suitable for different containers, object materials, and stacking densities. Low hardware dependency: only standard microphone is required, no expensive optical/electromagnetic equipment is needed. Potential applications: Industrial quality inspection, logistics monitoring, intelligent sensing and other non-contact object counting scenarios. Data format and access File structure: Classified and stored according to experimental conditions (container+object material) and the number of balls, including raw audio (WAV), FFT spectrum, and PCA dimensionality reduction results. Metadata: Annotate vibration frequency, environmental noise, ball size, and material parameters. License Agreement: Following CC BY-NC 4.0, non-commercial research and secondary analysis are allowed.