Assessing load in ball mill using instrumented grinding media

Highlights

• An instrumented grinding media was designed to mimic the motion state of ordinary grinding balls.

• Grinding efficiency index was proposed to evaluate the grinding effect.

• Time-domain features were combined with sample entropy in feature extraction.

• Recognition accuracy of ball mill load based on PSO-SVM classification model achieved as 96.67%.

Abstract

Monitoring mill load is vital for the optimization and control of grinding process. This study proposed the use of an instrumented grinding media to assess solid loading inside a ball mill, with size and density of the instrumented ball comparable to that of the ordinary grinding media. The acceleration signal was captured by an embedded triaxial accelerometer. The signal was first detrended by a complete ensemble empirical mode decomposition and then reconstructed using a correlation coefficient method. The filling ratio, particle to ball ratio, time domain features and sample entropy are features extracted from the signal, providing input to a support vector machine (SVM) learning model. Grinding experiments with different loads were conducted. The typical loading level was classified according to grinding efficiency index and associated power consumption. Different methods were adopted to determine the optimal values of parameters in the SVM model, including particle swarm optimizer (PSO), genetic algorithm (GA), and grid search (GS). The results showed that the accuracy of particle swarm optimizer can reach 96.67%. This study demonstrates the potential of using instrumented grinding media for real-time characterization of mill feed and operation monitoring.

Introduction

Grinding is an important unit operation in mineral processing, which aims for sufficient particle size reduction and mineral liberation for downstream separation processes. Grinding is a low efficiency and energy-intensive process, accounting for the majority of the operational cost in a mineral processing plant. Maintaining an optimum mill load is key to improve energy utilization and hence processing efficiency. However, the grinding process is featured with opaque, unstable and highly complex flow conditions (Tang et al., 2018a), making it difficult to achieve in-situ observation of flow behavior inside a ball mill. A load monitoring method which correlates well with operating state of grinding is thus desired to ensure grinding devices operated at an optimum condition (Tang et al., 2010a, Hilden et al., 2021).

Several approaches have been proposed to detect mill load, including active power, vibration (Davey et al., 2012, Xie et al., 2013, Zhao et al., 2013), acoustic sensors (Jackson et al., 2014, Pax, 2001), toe and shoulder angles (Clermont et al., 2008, Millsense, 2021). Vibration method is the most widely used one for in-situ measurement, it can accurately detect material changes in the mill, the device is convenient to install and maintain. However, the measurement results are easily affected by the wear of steel balls,other grinding media and liners. Acoustic method can predict the approximate filling ratio, but it is easily affected by grinding concentration and environmental noise. Toe and shoulder angles method was used in Sensomag® (Keshav et al., 2011) to detect mill load and achieved excellent results. The above-mentioned detection methods are all based on the relationship between the external response signal of ball mill and the charge. Consequently, they have limitations in exploring the behaviour of grinding media and particles inside the mill.

Apart from measuring external response signal, instrumented ball has been used to mimic the behavior of grinding ball inside ball mill. Dunn and Martin (1978) used accelerometers to detect the maximum impact force of steel ball. However, insufficient sampling frequency and unreasonable calculation of material contact coefficient largely limit the accuracy of measurement. Rolf and Vongluekiet (1982) used instrumented ball to measure impact energies in a 0.8 × 0.4 m mill, showing that impact frequencies are higher at 55% and 75% critical speed. Gao and Thelen (1994) developed the grinding balls which integrated piezo-electric sensors and data-processing electronics to analyze load movement from distributions of impact frequency and impact energy. Martins et al., 2008, Martins et al., 2013 developed an instrumented ball and a camera system to measure the state of the charge within a laboratory mill. However, this instrumented ball does not have the same effective density as the ordinary grinding ball, consequently the accuracy of measurement cannot be guaranteed. Due to the interaction between pulp, ball and mineral particles, the motion of charge is difficult to investigate under wet grinding, Yin et al. (2019) designed an instrumented ball with micro-electro-mechanical system to explore the influence of rotation speed on energy consumption. Moreover, instrumented particles embedded with micro-electro-mechanical sensors has also been widely used in structural health monitoring (Gu et al., 2020, Al-Obaidi and Valyrakis, 2021).

The signals captured by instrumented grinding media is similar to vibration signals, therefore, the vibration signal is used as an example here to illustrate the identification of loading state. Due to the interference of noise, much attention has been devoted to extract valuable features from vibration signal. For example, Behera et al. (2007) converted the acquired vibration signals from time domain to frequency domain by combining Hanning window function with fast Fourier transformation, thereby improving resolution in frequency domain. For the non-stationary signals, wavelet transform can better combine time domain signals with frequency-domain signals, thus enabling a more comprehensive description of the signal characteristics. However, it is difficult to select appropriate wavelet bases (Zhang and Zhang, 2020; Cai et al., 2019). Shi et al. (2019) proposed a feature extraction method based on fractional Fourier transform, in which adaptive filtering and feature extraction of vibration signals from both mill bearing and cylinder surface were performed to assess mill load. A reduced weighted soft sensor model was established based on least squares support vector machine (Shi et al., 2019, Si et al., 2008). However, the signals from two sensors are difficult to fuse to obtain an accurate prediction of the mill load. To reduce noise interference, empirical mode decomposition (EMD) can be applied to decompose vibration signals into a series of intrinsic mode functions (IMFs) and residual components. The useful intrinsic mode functions can then be synthesized to achieve noise reduction. However, EMD suffers from the problem of mode aliasing. Improved algorithms, such as ensemble empirical mode decomposition (EEMD), complete ensemble empirical mode decomposition (CEEMD) and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) were derived to assist signals decomposition by adding white noise independently. The reconstructed signals by CEEMD and CEEMDAN have less noise and better convergence than that of EMD and EEMD (Tang et al., 2020, Yang and Cai, 2021, Luo et al., 2020, Gao et al., 2020).

The extracted feature parameters are then trained as vector input classifiers to estimate mill load. Support vector machine (SVM) and artificial neural network are the most commonly used classifiers. SVM requires a small number of samples and has good robustness. However, it does not perform very well as the extremum problem is solved quadratically (Tang et al., 2010b). Consequently, various optimizations of SVM algorithm have been proposed, such as PSO-SVM and GA-SVM (Wang et al., 2019, Xue et al., 2020). Feedforward neural network demonstrated faster convergence speed and better prediction than back propagation network (Zhang et al., 2007). Deep learning network was applied in the field of flaw detection (Zhang and Su, 2020). Convolutional neural networks were applied to analyze the vibration signals on the bearing to judge the health of the bearing (Pinedo et al., 2020). Vibration was also integrated with other single source signals. However, the high-dimensional mechanical signals of multi-source modes are difficult to map the characteristic parameters of the ball mill, which leads to low fusion efficiency, slow convergence speed and poor generalization of the established model (Tang et al., 2018a, Tang et al., 2018b, Cai et al., 2021).

In this work, an instrumented grinding media employing triaxial accelerometer and microprocessor was designed to mimic the behaviour of ordinary grinding media in ball mill. Grinding experiments with typical feed loading states were conducted to obtain acceleration signals from the instrumented grinding media. The obtained signal data was preprocessed with detrending combined with the CEEMD algorithm. The signal was reconstructed using the correlation coefficient method and then evaluated in terms of the sample entropy. The feature vectors including filling ratio, particle to ball ratio, time domain features and sample entropy were selected as inputs to GA-SVM, GS-SVM, PSO-SVM models for classification, leading to a soft-sensing method for classification, identification, and prediction of load state in ball mill.