Real-time slurry characteristic analysis during ball milling using vibration data

ABSTRACT The characteristics of an internal slurry were analyzed during ball milling, which is commonly utilized in ceramic processing. We used a device with a capacity of 50 L because this is the size employed in industries and built a circulation system to collect the slurry during milling. The slurry was characterized in terms of the particle size and viscosity, and vibration data were collected from the side of the ball mill drum in real time for 24 h. By applying a fast Fourier transform to the vibration data, the energy was calculated and compared with the slurry characteristics. The vibration data in the (3–4) kHz range showed a strong negative correlation with the slurry viscosity at a shear rate of 24s−1. Our results confirm that the characteristics of the internal slurry can be monitored in real time using the vibration data collected during ball milling. Graphical abstract


Introduction
Concomitant with the advancements in data processing and machine learning technologies, big data are being increasingly used for process optimization and maintenance of equipment in the field of manufacturing [1][2][3]. The introduction of big data-based technologies requires high-quality data collected under various conditions. Consequently, such technologies are mainly applied in mass-producing industries, such as the semiconductor and automobile industries, which are led by large companies. Big data-based technologies are difficult to apply in industries led by small-and medium-sized enterprises, such as the ceramics industry. Therefore, a technology capable of a unified method of data collection is desirable for big data-based production in the field of ceramics.
The ball milling process is widely used in both industry and academia for various applications, including pulverizing or mixing of raw materials. Industries involving mass production require large-scale ball milling equipment. Thus, shortening the ball milling time while maintaining the quality of the produced powder is important for reducing production costs, increasing productivity, and ensuring quality control. To this end, it is necessary to optimize the ball milling conditions to maximize its effectiveness while minimizing the milling time by analyzing the change in the properties of the material during the milling process. Currently, field workers determine the process conditions empirically, and the process consumes more energy and time than necessary. Therefore, to optimize or manage the milling process, a method that can analyze the changes in the properties of the slurry in real time is required.
Because the characteristics of the internal material cannot currently be determined during the ball milling process, process optimization studies have been conducted via ex situ characterization of material properties such as particle size, particle shape, and slurry viscosity [4][5][6][7][8][9]. However, an ex situ analysis requires several experiments to optimize the milling process, and experimental errors are unavoidable because of the time difference between the acquisition of the slurry and the measurements. Conversely, simulations using the discrete element method have been performed to calculate the motion of the balls and raw material, which depends on the milling conditions [10][11][12][13][14][15][16][17][18][19], and estimate the changes in the properties of the internal material by collecting sound or vibration signals [20][21][22][23][24][25][26][27][28]. In the field of mineral production, process control through acoustic data monitoring has afforded ball milling conditions with the highest efficiency by using the acoustic signal intensity [20][21][22]. These studies are meaningful because they optimized and controlled the ball milling process using real-time data; however, these studies did not correlate the data with the actual material properties.
In this study, the changes in the particle size of the raw material and the viscosity of the slurry with the ball milling time were measured. The properties were evaluated by preparing spray-dried granules from slurries milled over different durations. In addition, vibration data generated during the ball milling process were collected in real time, and their correlation with viscosity was analyzed. This method allows for the realtime analysis of the characteristics of the internal slurry, which vary during the ball milling process.

Materials and methods
Ball milling was performed to prepare a slurry by mixing Al 2 O 3 (99.9%, D50: 0.43 μm, AES-11, Sumitomo Chem, Japan) and deionized water (DIW) with a dispersant (5468CF, San Nopco, Japan) using a high purity alumina ball (99.5%, 10 mm, SSA-995, Nikkato, Japan). In the milling process, alumina balls (60 kg), Al 2 O 3 (13.6 kg), DIW (9 kg), and a dispersant (81.6 g, 0.6 wt. % of powder) were mixed in a 50 L ball mill for 24 h at a rotational speed of 40 rpm. Figure 1 shows an image and schematic of the milling equipment used in this experiment. A circulation path was constructed in the milling machine using rotary joints at both ends of the rotating shaft, and the slurry was obtained during the milling process. After 1, 4, 12, and 24 h of milling, approximately 100 ml of the slurry was obtained, to which 3 wt.% of a binder (HS-BD 20A, San Nopco, Japan) was added before being stirred for 1 h. Granules were then obtained by drying the slurries using a mini spray dryer (B-290, Büchi, Switzerland). An analysis of the particle size and viscosity of the slurry according to the ball milling time was conducted using a particle size analyzer (LA-350, Horiba, Japan) and viscometer (DV2T, Brookfield, USA), respectively. The granule powder particle size was analyzed using a particle size analyzer, and scanning electron microscopy (SEM, JSM-6390, JEOL, Japan) was used to visualize the particle surface topography and cross-section. For cross-section analysis, the granule powder was impregnated in a polymeric resin and polished using a diamond suspension. During the ball milling process (Figure 1(b)), a vibration sensor (352C33, PCB Piezotronics, USA), DAQ (data acquisition, Ni-9234, National Instrument, USA) with a chassis (cDAQ-9191, National Instruments, USA), and battery pack were attached to the side of the ball mill to collect vibration data in real time. Vibration data were recorded using LabVIEW software (National Instrument, USA) for 10s at 30 min intervals, at a sampling rate of 51.2 kS/s, yielding a total of 48 data points collected over 24 h. Figure 2 shows the variations in the particle size of the alumina slurry with milling time. Particle size distribution D10, D50, and D90 corresponding to the percentages 10%, 50%, and 90% of particles under the reported particle size. The particle size variation results without ultrasonic treatment ( Figure 2(a)) clearly show the dispersion of soft agglomerations of the alumina powder during the milling process, indicating a rapid size reduction up to 8 h. Conversely, with sonication treatment, the particle size (Figure 2(b)) did not change significantly (0.48 to 0.44 μm) based on the D50 value. This is because the milling process in this experiment was not designed to grind particles but to disperse them in the DIW as a slurry for the spray dryer.

Results and discussions
The viscosity results based on the shear rate after 1, 4, 12, and 24 h of milling (Figure 3(a)) show a shear thinning behavior, and the viscosity increases proportionally with milling time when the shear rate is fixed at 24s −1 (Figure 3(b)). The change in the viscosity with time increased rapidly up to approximately 4 h, after which it increased at a slower rate until the end of the milling process regardless of the measuring shear rate (see supplementary Fig. S1). Considering the observed particle sizes and viscosities (Figures 2 and 3), the reason for the initial increase in the viscosity is likely the breaking of the soft agglomerations of the particles and their dispersion through the DIW. Conversely, after the initial rapid increase in the viscosity, the subsequent gradual increase can be attributed to the grinding of alumina particles and wear of the used balls. Figure 4 shows the particle size analysis of the alumina granules subjected to spray drying with the slurry after 1, 4, 12, and 24 h of ball milling, and hereafter referred to as G01, G04, G12, and G24, respectively. In the case of G01 and G04, particles similar in size to the raw powder were produced in addition to the granulated particles, indicating a bimodal particle  size distribution. This occurred because the particles in the slurry were not sufficiently dispersed [29][30][31], evidenced by the variations in the particle size and viscosity (Figures 2 and 3). Conversely, G12 showed a granule size of approximately 18 μm, larger than those of G01 and G04, and no fine particles of less than 1 μm were observed. Although G24 did not produce fine particles, the particle size was lower than that of G12. The increased viscosity as milling continued was considered to reduce the fluidity of the slurry, which reduced the feeding rate of the slurry and droplet size under similar spray dry conditions. Figure 5 shows the SEM images of the surface morphology and polished cross-section of the granules after impregnation in a polymeric resin. The G01 powder included several irregularly shaped particles, such as dimpled and hollow particles, instead of spherical granules ( Figure 5(a)). These atypical granules readily break down to produce fine particles which presented in particle size distribution of G01 powder (Figure 4(a)). Conversely, the number of irregularly shaped particles decreased with an increase in the milling time of the slurry ( Figure 5(b-d)) because the Al 2 O 3 raw powder was better dispersed through the DIW with longer milling times, causing homogeneous shrinkage during drying.
During the ball milling process described above, vibration signals were collected from the side of the ball mill drum. Before collecting the vibration data during the actual process, the data were collected from operations with only balls, balls + DIW, and balls + slurry (DIW + powder), as shown in Figure 6, to determine the cause of the observed signals. The intensity of the raw vibration data decreased in the order: balls + DIW (Figure 6(b)) > balls + slurry ( Figure 6(c)) > only balls (Figure 6(a)); this is because the propagation of vibrational energy to the drum wall is greater through a medium such as DIW than that through air, and the vibration is attenuated with the balls + slurry mixture owing to the powder. Moreover, strong signals were periodically observed in the vibration data of the three different cases, indicated by the red arrows in Figure 6(a-c). These periodic signals were observed at approximately 0.6 Hz in the fast Fourier transform (FFT) data ( Figure 6(d-f)), which corresponds to a rotational speed of 40 rpm, indicating that the periodic vibrations originated from the equipment due to friction, whenever the ball mill rotated. Excluding the signal caused by the rotation of the ball mill, the case where milling was performed with only the ball ( Figure 6(d)) shows high vibrational intensity in the (0-7) kHz range. Conversely, when fluids such as DIW or slurry were included in the milling process ( Figure 6(e, f)), the vibrational intensity in the (7-10) kHz range was high. It can be inferred that the signal due to the collision of the balls appears in the range of (0-7) kHz, while the signal due to the fluid appears in the  range of (7-10) kHz. In addition, when a fluid was present, the vibration signal in the (0-7) kHz range caused by the balls was attenuated. Additionally, the attenuation was greater in the case of the slurry than in the case of water. Figure 7 shows the vibration data collected at 1 and 24 h after ball milling. The raw vibration data (Figure 7(a)) show no significant difference over time, which is similar to the result shown in Figure 6(c), for which no milling was performed. After the FFT (Figure 7(b)), the shape was similar to that shown in Figure 6(f), which was obtained before milling, and a signal due to drum rotation was detected at 0.6 Hz. Conversely, there exists a difference with respect to the milling time in the data after FFT, which is a reduction in the signal intensity in the (2-7) kHz region.
To analyze the energy change with respect to the vibration frequency, the frequency band was divided into intervals of 1 kHz. The energy changes in each band are shown in Figure 8. The energy was calculated by summing the power values taken from the power spectrum obtained from the vibration data after the FFT. The calculated energies showed three different characteristics in the (0-2), (2-7), and (7-10) kHz frequency ranges with different energy scales (Note that the scale of y-axis for (0-1), (7)(8), (8)(9), and (9-10) kHz bands are different). In the (0-2) kHz region, the energy tended to increase; in the (2-7) kHz region, the energy gradually decreased after an initial rapid reduction in energy; and in the (7-10) kHz region, irregular energy changes with a high deviation were observed before the energy decreased after approximately 12 h of  milling. The change in energy, evidenced by the signal in the (2-7) kHz range, was inversely proportional to the change in viscosity, as can be observed from Figure 3(b).
The vibrational signals in the low-frequency region of (0-7) kHz are due to the collision of the balls; thus, the reduction in the intensity is assumed to be caused by the buffering action of the increased slurry viscosity on the collisions of the balls. Conversely, the signals in the (7-10) kHz region, which are assumed to be caused by the propagation of the ball collision signal through the slurry, are complex because the energy change is affected by the collision energy of the ball, and the propagation characteristics of the vibration signal are changed according to the viscosity of the slurry [32,33]. Figure 9(a) shows a scatter plot of the viscosity change (in Figure 3   S2b, S2c, and S2d). However, the correlation coefficients for the vibration energies in frequency ranges of (0-1), (8)(9) and (9-10) kHz were low, with values of 0.666, −0.379 and −0.126, respectively (see supplementary Fig. S2a, S2e, and S2f). This suggests that the viscosity of the internal slurry can be determined from the vibration energy in a specific frequency range. In addition, a high correlation was observed between the vibration energy and viscosity regardless of the shear rate; moreover, when the shear rate was 10s −1 or higher where nearly Newtonian behavior is visible, the correlation coefficient was approximately 0.95 (see supplementary Fig. S3). To determine whether the energy change was due to the slurry, energy data corresponding to the (3-4) kHz range were obtained from additional milling experiments with only balls or balls + DIW at a similar rotation speed of 40 rpm for 12 h (Figure 9(b)). Because the absolute energy values are different in terms of the scale, the data were normalized by dividing them by the initial value. In the case of operation with only the ball or ball + DIW, the energy value slightly decreased, which was assumed to be due to the presence of alumina particles generated by the friction between the balls. However, there was no sudden change, as in the case involving the slurry, which indicates that the change in the vibration energy generated during the ball milling process is due to the slurry. Additionally, although the normalized vibration energies for the three repeated experiments under the same conditions were not completely consistent (Figure 9(c)), the results showed similar trends, in which the energy decreased rapidly and then decreased at a lower rate. These results indicate that the data are reproducible and can be used to determine the slurry viscosity during milling processes.

Conclusions
This study analyzed the correlation between vibration data collected in real time and characteristic changes of a slurry with ball milling time. The results can be summarized as follows: (1) During ball milling, the particle size decreased and slurry viscosity increased. The properties of the slurry significantly influenced the granule production in the spray drying process. (2) By collecting vibration data in real time and calculating the energy for each frequency band, a specific frequency range that was strongly correlated with the viscosity data was identified.
(3) Three additional experiments were conducted under similar conditions, and the data in the aforementioned frequency range, which had a strong correlation with the viscosity, showed a similar trend.
(4) The characteristics of the slurry can be monitored during ball milling using vibration data, which enables the determination of the optimal milling conditions and time. (5) The results of this study show that the viscosity of the slurry can be inferred during the ball milling process without directly measuring it. It is expected that the same analysis will be possible for other properties, such as particle size and slurry load, which will form the basis of our future work in this area.