⚡ electrical engineering

Research on Resistance Prediction Model for Coal Mining Face Supports Based on Machine Learning

This paper proposes an immuno-particle swarm hybrid algorithm optimized BP neural network (IA-PSO-BP) model to overcome the limitations of traditional prediction methods, achieving high-precision support resistance forecasting for shallow-buried coal seams in northern Shaanxi with a correlation coefficient of 0.966.

Original authors: Yang Lei, Miao Yanping, Li Guowei, Zhang Kun, Xu Baojun, Hou Pengfei, Zhang Pengfei

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang Lei, Miao Yanping, Li Guowei, Zhang Kun, Xu Baojun, Hou Pengfei, Zhang Pengfei

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep beneath the surface of the earth, where the rock is thin and the coal seams lie close to the top, mining operations face a unique and dangerous challenge. Unlike deep mines where the weight of the earth is distributed slowly and predictably, shallow coal seams in regions like northern China experience sudden, violent shifts in pressure. This phenomenon, known as mine pressure, can cause the roof of a tunnel to collapse without warning, threatening the lives of workers and the stability of heavy machinery. For decades, engineers have tried to predict these shifts using mathematical formulas based on rock mechanics, but the ground often behaves in ways that are too complex and changeable for simple equations to capture. When the pressure spikes, it can crush support structures, leading to accidents that have historically accounted for the majority of fatalities in coal mining. The need for a system that can anticipate these dangerous surges with high precision is not just a technical goal; it is a matter of safety and survival for the thousands of people working underground every day.

To address this, a team of researchers from a major mining company and a university in Shaanxi has developed a new way to forecast these pressure changes using a type of computer learning known as machine learning. Instead of relying solely on fixed physical laws, they built a digital model that learns from the actual behavior of the mine. They focused on a specific working face, a long tunnel where coal is extracted, which is equipped with hundreds of hydraulic supports—massive, piston-driven columns that hold up the roof. These supports are fitted with sensors that constantly measure the pressure they are bearing. The researchers collected over two million data points from these sensors over a period of five months, creating a detailed record of how the ground pushes and pulls against the mine's structure. However, raw data from the deep earth is often messy, containing gaps where sensors failed or spikes caused by equipment glitches rather than actual geological shifts. The team first cleaned this data, removing the obvious errors and filling in the missing pieces by looking at the patterns of nearby supports and previous moments in time, ensuring the computer had a complete and accurate picture to learn from.

The core of their work involved teaching a computer program to recognize the complex, non-linear patterns of mine pressure. They started with a standard neural network, a computing system modeled after the human brain that is good at finding patterns, but they knew this system alone often gets stuck guessing the same wrong answer over and over again. To fix this, they combined it with two other intelligent systems. One system, inspired by how birds flock together, helps the computer search for the best possible solution by sharing information across a group of virtual "particles." The second system mimics the human immune system, which is excellent at recognizing threats and remembering them to fight them off faster next time. By blending these three approaches, they created a hybrid model that could explore a vast range of possibilities without getting trapped in local dead ends, allowing it to find the true, global best way to predict the pressure.

The researchers tested this new model against older, simpler methods using three different sets of data, ranging from small samples to a massive dataset covering nearly five months of mining activity. In every test, the new hybrid model outperformed the others. While the traditional methods struggled to predict the sudden, sharp spikes in pressure that occur when the roof shifts, the new model tracked these dangerous peaks with remarkable accuracy. On the largest dataset, the model's predictions matched the actual measurements so closely that the correlation was nearly perfect, reaching a score of 0.966. This means the computer could accurately forecast both the quiet, steady periods of pressure and the violent, periodic surges that signal a potential roof collapse. The model worked consistently across different sections of the mine, proving it could generalize its learning to new situations rather than just memorizing the past.

The study concludes that this new approach offers a reliable technical foundation for preventing roof disasters in shallow coal seams. By accurately predicting when and where the pressure will spike, mine operators can take proactive measures, such as reinforcing supports or evacuating areas before a collapse occurs. While the model was tested on a single mine and could eventually be improved by including more geological factors like the depth of the coal seam, the results demonstrate a significant leap forward in safety technology. It shows that by combining the pattern-recognition power of machine learning with the adaptive strategies of biological systems, engineers can finally tame the unpredictable forces of the earth, turning a major safety risk into a manageable, predictable variable.

Technical Summary: Research on Resistance Prediction Model for Coal Mining Face Supports Based on Machine Learning

Problem Statement
Coal mining in shallow-buried seams in northern Shaanxi, China, presents unique geological challenges characterized by complex mine pressure manifestations and a high frequency of roof accidents. Traditional prediction methods, ranging from empirical statistics to theoretical calculations (e.g., "transmission rock beam theory"), struggle to handle the nonlinear, time-varying, and multi-factor coupling nature of mine pressure data. While recent advancements in artificial intelligence (AI) have introduced machine learning models like Random Forest, LSTM, and Transformer-based architectures, existing optimized neural network models (specifically Particle Swarm Optimization-BP, or PSO-BP) suffer from critical limitations. These include a tendency to become trapped in local optima and the degradation of population diversity during later iteration stages, leading to premature convergence and reduced prediction accuracy for complex engineering scenarios.

Methodology
To address these deficiencies, the authors constructed a support resistance prediction model based on an Immune Algorithm (IA) optimized Particle Swarm Optimization (PSO) and Backpropagation (BP) neural network, termed IA-PSO-BP.

Algorithmic Framework: The model integrates the immune algorithm into the PSO process to optimize the initial weights and thresholds of the BP neural network. The BP network's prediction error serves as the fitness function. The immune algorithm introduces three specific mechanisms to enhance the PSO:
1. Cloning Selection: Maps PSO particles to antibodies, retaining high-affinity particles while suppressing the excessive proliferation of similar ones to maintain population diversity.
2. High-Frequency Mutation: Applies perturbation via cloning and mutation operations to enhance global search capabilities and avoid local optima.
3. Concentration Regulation & Immune Memory: Records the history of optimal particles to guide the population toward the global optimal direction.
Data Preprocessing: The study utilized real-time monitoring data from the 43104 fully mechanized mining face (April–September 2025). The data underwent rigorous preprocessing:
- Outlier Removal: The 3σ criterion was applied to eliminate extreme outliers, zero values, and negative values.
- Missing Value Imputation: The K-Nearest Neighbors (KNN) algorithm was used to fill missing data points based on spatiotemporal correlations.
- Normalization: Data were normalized to the [0,1] interval, and time features were encoded using periodic sine-cosine functions.
Experimental Design: Three datasets with gradient differences in sample capacity (Dataset 1: 5,400 items; Dataset 2: 8,856 items; Dataset 3: 17,499 items) were constructed. The models (BP, PSO-BP, and IA-PSO-BP) were trained and tested on these datasets across three measurement zones (upper, middle, and lower) of the mining face.

Key Contributions

Hybrid Algorithm Development: The paper proposes a novel IA-PSO-BP architecture that specifically targets the premature convergence and diversity loss issues inherent in standard PSO-BP models when applied to mine pressure prediction.
Comprehensive Data Processing Pipeline: The study establishes a robust preprocessing workflow for mining pressure data, combining statistical outlier detection (3σ) with machine learning-based imputation (KNN) and time-series feature encoding.
Multi-Condition Verification: Unlike studies relying on single datasets, this research validates the model across three distinct sample sizes to systematically evaluate generalization ability and the impact of training data volume.

Results
The IA-PSO-BP model demonstrated superior performance compared to traditional BP and PSO-BP models across all evaluation metrics (Mean Absolute Error, Mean Square Error, Mean Absolute Percentage Error, and Correlation Coefficient $R^2$ ).

Accuracy: On the largest dataset (Dataset 3), the IA-PSO-BP model achieved a correlation coefficient ( $R^2$ ) of 0.966, significantly outperforming the PSO-BP model ( $R^2$ = 0.904) and the traditional BP model ( $R^2$ = 0.795 for the best-performing support).
Fitting Capability: The model successfully captured both the stable fluctuation periods of mine pressure and the peak periods of periodic roof pressure. It effectively reduced amplitude deviations and time-sequence lags observed in the other models, particularly during sudden pressure changes and cyclic overburden pressure events.
Generalization: The performance improvement was consistent across all three measurement zones (supports 20, 80, and 180) and all dataset sizes, indicating strong generalization capabilities.

Significance
The authors claim that the IA-PSO-BP model provides reliable technical support for intelligent mine pressure prediction and roof disaster prevention in shallow-buried coal seams. By achieving high-precision fitting for complex pressure manifestations, the model addresses the urgent need for accurate prediction systems in regions with high roof accident rates. The study concludes that the hybrid algorithm effectively overcomes the limitations of existing neural network optimization methods, offering a robust solution for engineering applications in the specific geological context of northern Shaanxi.

Limitations
The authors acknowledge that the study was validated using data from a single working face (43104), and the model's universality across different geological conditions requires further expansion. Additionally, the current input features rely primarily on time-series resistance data, without fully integrating geological mining parameters such as burial depth or mining height. Future work is suggested to incorporate these multi-factor influences and validate the model across a broader range of working faces.

More like this