The article explores the critical role of data collection and processing in optimising the operation of mobile impact crushers, with a focus on enhancing energy efficiency and product quality. It introduces a collaborative research project that leverages machine learning models to predict particle size distribution (PSD) and support automated control strategies. The research involves the SBM REMAX 600 mobile impact crusher, equipped with advanced sensor systems to capture key performance parameters such as rotor speed, gap width, throughput, and power consumption. The study delves into the complexities of data preprocessing, including value-based filtering, outlier detection, and data smoothing techniques, to ensure the reliability and accuracy of the datasets used for modelling. A notable finding is the significant impact of pre-screening modes on PSD, highlighting the need to treat this parameter as a categorical feature in predictive models. The article also discusses the implementation of feature engineering to enhance model performance and reduce dataset variance. The established data processing pipeline paves the way for future developments in real-time predictive tools and autonomous adjustment systems, marking a significant step towards fully data-driven, self-optimising crushing operations.
AI Generated
This summary of the content was generated with the help of AI.
Abstract
This contribution presents the current results of a PhD research project aimed at enabling an energy- and product-optimised operation of mobile impact crushers through advanced data collection and processing. The study focuses on systematic sensor data acquisition, data cleaning, and correlation analysis to support predictive modelling of the particle size distribution (PSD) of crusher products. A structured methodology involving parameter selection, statistical filtering, and time-series smoothing was applied to real-world data collected from SBM’s REMAX 600 mobile impact crusher. This work serves as a foundation for further development of machine learning models and a future autonomous control agent for real-time optimisation of crushing operations.
Notes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
1 Introduction
SBM Mineral Processing GmbH, a leading manufacturer of mobile and stationary crushing technology, has initiated a collaborative research project with the Chair of Mineral Processing at Montanuniversität Leoben to investigate data-driven optimisation strategies for mobile impact crushers. The task of improving the energy efficiency of crushers is a relevant issue today as they are critical components in mining and processing plants [1]. In response to increasing demands for energy efficiency and product quality, the project explores the potential of machine learning models trained on operational data for predicting a product’s particle size distribution (PSD) and supporting automated control strategies.
This article presents the data-related aspects of the PhD project conducted within this collaboration. The focus lies on the acquisition and processing of sensor data collected during the operation of a mobile impact crusher. By understanding correlations between machine parameters and PSD, this research lays the groundwork for developing predictive models that can support energy-efficient and product-optimised operations.
Advertisement
2 Scope and Objectives
Mobile crushing represents a complex system where multiple factors influence efficiency and product quality. The aim of the data processing work is to prepare the foundation for developing a real-time, model-driven optimisation framework for mobile crushing. In particular, the project targets the following:
Collection and cleaning of sensor data from the crusher
Exploration of parameter dependencies and correlations
Implementation of noise-reducing preprocessing techniques
Extraction of relevant features for machine learning
Identification of optimal preprocessing strategies for PSD prediction
3 Crusher Setup and Process Parameters
The test machine used throughout this research is the SBM REMAX 600—a mobile impact crusher equipped with adjustable rotor speed, gap width control, and a sensor system [2]. The PSD of the output material is captured via an optical camera system positioned above the discharge belt. Key performance parameters include product size indices (k30, k80, D1, D4, and Main Flank), gap width, rotor speed, throughput, and power consumption, which are explained in more detail in the internal FFG Year 3 Research Report [3]. Table 1 shows the operational parameters that were collected.
TABLE 1
Input and Output Parameters
Input parameters
Output parameters
Timestamp
Rotor Speed
Gap Width
Power Consumption
Throughput
Pre-screening Mode (1–3)
k80-to-gap ratio
Main Flank
k80, k30
D1, D4
These parameters were systematically logged and synchronised at approximately 30-second intervals, providing sufficient resolution for time-series analysis and identification of operational trends.
4 Data Cleaning and Preprocessing
Data preprocessing refers to a set of techniques for enhancing the quality of raw data, which is crucial for a reliable data analysis [4]. When working with real-world datasets, extreme values that deviate significantly from the rest of the data can distort the analysis and negatively impact model performance. These are called outliers, which can arise due to measurement errors, variability in data, or rare occurrences [5]. Removing or handling outliers appropriately can improve the accuracy and generalisation of machine learning models [6]. Therefore, the robust preprocessing workflow was implemented, which included the following steps:
1.
Value-Based Filtering:
Removal of data points recorded during idle states, signal dropouts, or startup phases, which may appear as zero or negative values in throughput and rotor speed due to unit errors or communication delays
Removal of invalid PSD outputs (e.g. k80 > 6000 μm or D1 < 0%)
2.
Time Window Selection:
Selection of valid production windows based on machine logs
Removal of idle or non-production intervals
3.
Outlier Detection:
Statistical checks (Altman’s Z‑score, interquartile range (IQR))
Visual inspections of key PSD parameters
4.
Data Averaging (Smoothing):
Simple Moving Average (SMA), e.g. over 10 and 20 points (corresponding to 5–10 min of operation time)
Used to reduce noise and capture meaningful trends
Advertisement
This preprocessing was applied to datasets recorded over several days, resulting in a clean, synchronised dataset suitable for modelling [7]. We propose a data-cleaning framework for real-world data research with a focus on outlier detection to serve as a reference for the application of such technologies in future studies.
5 Parameter Dependencies and Correlation Analysis
To understand the relationships between operational settings and PSD, a correlation matrix was constructed based on Pearson correlation coefficients (Fig. 1). In the dataset provided, these dependencies, while present, are generally not extremely strong, with most correlations lying in moderate ranges. This suggests that, while the parameters do influence each other, the relationships are not purely linear, highlighting the complexity of the crushing process.
Fig. 1
The Feature Correlation Matrix
The most relevant findings include:
Gap Width and Rotor Speed showed a moderate correlation with PSD indicators (e.g. k80, Main Flank);
Throughput exhibited weaker but visible trends toward a coarser product (higher k80) under increased load;
Power Consumption showed variable correlation, indicating non-linear dependencies and potential interaction effects.
These results confirm the necessity of using advanced regression models rather than relying on simple linear approximations. Despite the presence of several visible trends, most correlations between machine parameters and PSD indicators were relatively weak. This suggests that other hidden factors may be influencing the output characteristics. A further analysis revealed that the pre-screening mode (with values 1–3) plays a significant role in influencing particle size distribution. Different modes represent different material flows (as shown in Fig. 2) and result in different particle compositions entering the crushing chamber. When datasets are grouped by pre-screening mode, the strength of correlations, particularly for k80, D1, and Main Flank, should theoretically increase. The pre-screening parameter in the dataset is a key factor that determines what happens to the material which is not processed by the crusher. The values for pre-screening can be 1, 2, or 3, each representing a different material flow scenario, as explained below:
Pre-screening Value 3: All particle size classes 0/45 mm bypass the crusher and are directed straight to the final product (discharge). Particles > 45 mm are processed by the crusher. Thus, the output of the crusher (commuted > 45 mm fraction) as well as all the bypassed fractions (< 45 mm) are discharged into the final product and thus captured by the camera, which calculates the PSD.
Pre-screening Value 2: The 0/22 mm fraction is screened out and removed to a stockpile, thus not present in the final product. However, the 22/45 mm bypasses the crusher and goes into the final product. Particles > 45 mm are processed by the crusher.
Pre-screening Value 1: Both the 0/22 and 22/45 mm fractions are removed to a stockpile, meaning only particles > 45 mm pass the crusher. This is the most restrictive mode, where only the particles passing the crusher are discharged as the final product and thus captured by the camera, calculating the PSD.
Fig. 2
Material flow with different pre-screening parameters
This pre-screening parameter is critical because it determines the material flow through the crusher and affects the resulting particle size distribution, throughput, and energy consumption. Understanding the impact of pre-screening settings is essential for optimising the overall crushing process.
As a result, it was concluded that pre-screening must be treated as a categorical feature during modelling. Segmenting or normalising datasets using this parameter is recommended. This insight is crucial for the future development of predictive models and the structuring of training data for supervised machine learning algorithms.
6 Feature Engineering for Modelling
To support machine learning applications, several engineered features were introduced:
Aggregated time-window values (averages over 5–10 min)
Normalised throughput-to-power ratios
Rotor speed categories (high/low operation modes)
Pre-screening as categorical input to control for material flow
These features aim to enhance model performance by making key relationships more explicit and reducing variance in the dataset. The use of the smoothing technique not only improves the clarity of the data but may also have implications for understanding the crusher’s reaction time. In the context of this project, smoothing over a 10-point window (5 min) may correspond to the machine’s response time to changes in its operating environment. The averaged data is not only valuable for the current analysis and modelling but could also play a crucial role in future developments, particularly in the design of a decision agent for autonomous crusher operation.
The implementation of the SMA technique for the dataset representation is depicted in Figs. 3 and 4. Fig. 3 shows the trend of the Main Flank over the index: the initial data is rather noisy, which makes it difficult to interpret. In comparison, Fig. 4 shows the trend after averaging over 10 points, making the trend clearer.
Fig. 3
Trend of the Main Flank over index (initial dataset)
Fig. 4
Trend of the Main Flank over index (after averaging over 10 points)
7 Discussion and Outlook
The current work has established a robust pipeline for preprocessing operational crusher data. Through data filtering, smoothing, and feature engineering, a clean and consistent dataset was produced. However, the correlation analysis demonstrated that the direct relationships between machine parameters (such as rotor speed and gap width) and PSD values are generally weak.
The root cause was identified as the influence of pre-screening mode, which alters the input stream before crushing. Once this factor is isolated, stronger and more meaningful parameter dependencies are to be observed. This confirms that future predictive modelling must consider material flow configuration as a critical context variable.
The established data processing flow is complete and operational, enabling systematic evaluation of new datasets. As more tests are conducted under different operating conditions and with multiple material types, the correlation strength between machine inputs and PSD will improve. This will allow:
Accurate training of machine learning models tailored to each mode/material type
Validation of PSD predictions under laboratory and industrial conditions
Deployment of real-time predictive tools and autonomous adjustment systems
This represents a significant step toward achieving fully data-driven, self-optimising crushing systems.
Open Access Dieser Artikel wird unter der Creative Commons Namensnennung 4.0 International Lizenz veröffentlicht, welche die Nutzung, Vervielfältigung, Bearbeitung, Verbreitung und Wiedergabe in jeglichem Medium und Format erlaubt, sofern Sie den/die ursprünglichen Autor(en) und die Quelle ordnungsgemäß nennen, einen Link zur Creative Commons Lizenz beifügen und angeben, ob Änderungen vorgenommen wurden. Die in diesem Artikel enthaltenen Bilder und sonstiges Drittmaterial unterliegen ebenfalls der genannten Creative Commons Lizenz, sofern sich aus der Abbildungslegende nichts anderes ergibt. Sofern das betreffende Material nicht unter der genannten Creative Commons Lizenz steht und die betreffende Handlung nicht nach gesetzlichen Vorschriften erlaubt ist, ist für die oben aufgeführten Weiterverwendungen des Materials die Einwilligung des jeweiligen Rechteinhabers einzuholen. Weitere Details zur Lizenz entnehmen Sie bitte der Lizenzinformation auf http://creativecommons.org/licenses/by/4.0/deed.de.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
