Machine learning approach for the prediction of mining-induced stress in underground mines to mitigate ground control disasters and accidents

The bord and pillar method is commonly employed in Indian underground coal mines, and the extraction rate varies between 30 and 65%. During pillar extraction, pillars are subjected to severe stress conditions. Due to this, the natural state of stress equilibrium is disturbed, resulting in severe strata control problems leading to sudden, unpredictable failure such as a premature collapse of pillars, severe roof or side fall, and sometimes leading to serious/fatal injury or burial of machinery. This paper deals with the prediction of mining-induced stress during pillar extraction using Machine Learning (ML) techniques like Random Forest and Multilayer Perceptron. The various factors used for the formulation of the models for predicting the mining-induced stresses are Depth of the workings (H), Panel width/length (W/L), Pillar width/working height (w/h), Goaf length, and Area of extraction. This paper highlights the importance of operational parameters rather than geological parameters. The Correlation coefficient (R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}^{2}$$\end{document}) of mining-induced stresses for the case studies discussed in the paper is 0.85 for Random Forest and 0.76 for Multilayer Perceptron, which shows Random Forest results have a comparative edge over Multilayer perceptron. With this developed prediction models, the stress conditions of pillars can be predicted. A comprehensive review on ground fall disasters in different countries like India, Australia, and the United States of America from 1900 to 2020. A comprehensive review on the present status of ground fall accidents in India. Identification of parameters influencing stability of Continuous Miner Panel. Development of models using Random Forest and Multilayer Perceptron for prediction of mining-induced stress w.r.t various pillar extraction stages and validation of models with actual field results. Exploring the future directions in underground coal mining towards Zero accidents mission. To increase underground coal mining safety and productivity. To reduce idle time of Continuous Miners due to strata control problem. Use of Machine Learning for prediction of mining-induced stresses in place of numerical modelling for faster analysis and less resources. A comprehensive review on ground fall disasters in different countries like India, Australia, and the United States of America from 1900 to 2020. A comprehensive review on the present status of ground fall accidents in India. Identification of parameters influencing stability of Continuous Miner Panel. Development of models using Random Forest and Multilayer Perceptron for prediction of mining-induced stress w.r.t various pillar extraction stages and validation of models with actual field results. Exploring the future directions in underground coal mining towards Zero accidents mission. To increase underground coal mining safety and productivity. To reduce idle time of Continuous Miners due to strata control problem. Use of Machine Learning for prediction of mining-induced stresses in place of numerical modelling for faster analysis and less resources.

Perceptron.The various factors used for the formulation of the models for predicting the mininginduced stresses are Depth of the workings (H), Panel width/length (W/L), Pillar width/working height (w/h), Goaf length, and Area of extraction.This paper highlights the importance of operational parameters rather than geological parameters.The Correlation coefficient ( R 2 ) of mining-induced stresses for the case studies discussed in the paper is 0.85 for Random Forest and 0.76 for Multilayer Perceptron, which shows Random Forest results have a comparative edge over Multilayer perceptron.With this developed prediction models, the stress conditions of pillars can be predicted.

Graphic abstract Article highlights
• A comprehensive review on ground fall disasters in different countries like India, Australia, and the United States of America from 1900 to 2020.Though there is significant global pressure on the immediate reduction of fossil fuels as an energy source, coal is and will remain one of the primary sources of energy in India in the future.In India, coal Page 3 of 19 159 Vol.: (0123456789) production from opencast mining is 96.10%, whereas underground mining contributes 3.90% (Provisional Coal Statistics 2022-2023).However, the thrust on increasing underground coal production is being felt by the coal companies because of many critical issues like difficulty in land acquisition for opencast mining under the current socio-political environment, the severe threat of irreparable environmental damage due to opencast mining vis-à-vis the growing concern for mitigating the environmental impact due to opencast mining and its associated cost, depletion of shallower deposits, and at the same time, increasing demand for quality coal production to maintain economic sustainability of the country.
There is no other option but to go for the exploitation of deeper deposits by underground mining.Presently, most underground coal mines in India are semi-mechanized using SDL/LHD with the chronic problem of very low productivity, leading to huge losses per ton of coal production and the lack of capability of mass production.To improve the production and productivity from underground mining, 'it's high time to go for mass production by adopting suitable technologies like Longwall mining, Shortwall/Short longwall mining for developed coal pillars, Room and Pillar mining, and Wongawilli or Rib pillar extraction by Continuous Miner and Shuttle cars (Bhattacharjee and Vinay 2019).Even though Longwall is highly successful in other countries, it still had only limited success in India due to adverse strata conditions, lack of suitable locale for the introduction of high-capacity longwall mining, halfhearted approach to go for all-out mechanization right from face to surface, and most importantly, depending solely on importing technology without creating in house capabilities for absorption of the technology.
In India, the Bord and Pillar Method contributes 3-5% of total coal production and 65-75% of total underground production, whereas Room and Pillar mining using continuous miners contributes less than 1-2% of total production and 20-25% of underground coal production.The main problems associated with low production from B&P and R&P mining are the stability of the underground workings, resulting in roof and side falls leading to the burial of machinery and fatalities or serious injuries in many cases.The main challenge in underground coal mining is the stability of underground workings, mainly during pillar extraction (Galvin 2016;Vinay et al. 2022).The pillars are designed to ensure the stability of the pillars for a longer period till extraction of the pillars in the later stage.During the pillar extraction, the overlying roof is allowed to cave in, or the pillars are allowed to yield depending on the caveability of the roof or the strength of pillars.However, pillar extraction has been historically the most critical activity, resulting in the loss of many lives and the burial of machines deployed.
During the extraction of pillars, strata monitoring plays a vital role.A good understanding of mininginduced stresses and their impact on stability of the pillars and workings helps the mining engineer design working dimensions of panels, pillars, and design of support systems and eventually ensures safety by preventing strata failure or safe withdrawal of man machines before actual failure.Even though the pillar extraction process is being practiced for more than 150 years in India, there has not been a properly scientific understanding of underground strata behavior and estimation of mining-induced stresses.Still, some thumb rules based on previous experiences are mostly being followed, which have failed, resulting in serious consequences.
Theoretically, a pillar is designed with a safety factor depending on the strength of the coal and the stress acting on the pillar, where stress is calculated based on the tributary area and pressure arch theory, whereas the strength is calculated numerically using different pillar strength formulae.However, as extraction progresses, the above factor of safety does not hold good.
Even after using high-capacity roof bolts, cable bolts, and mobile roof supports, failure of the roof during pillar extraction could not be arrested in many cases.The theory to predict the mininginduced stresses with respect to the area of extraction could not be established due to the complex interaction of many geo-mining and operational parameters.
The state of the stability of workings after the formation of pillars is largely disrupted by the openings created by splitting or slicing or lifting, or fendering while extracting the pillars.The area of excavation in a pillar increases with the progress of the extraction process, eventually resulting in an increase in stress acting over the adjacent pillars due to the transfer of stresses of the extracted area on the pillars under extraction or adjacent pillars.
Failure in understanding the stresses acting on surrounding pillars under extraction becomes very critical on most occasions which eventually results in catastrophic failures like sudden collapses, a large number of pillars, or large area collapses.There are many examples of such failures like the Coalbrook disaster (South Africa), Sago disaster and Crandall Canyon disaster (USA), Topa, Khottadih, Churcha, and GDK-8A disaster (India), due to failure in the prediction of such failures and timely withdrawal of men and machines.It is expected that there will be a shift in production technology of coal in India in the recent future when the share of underground coal production will be increased to a large extent, and strata control problems will be more critical due to its huge potential of severe consequences.

Necessity of the study
Detailed literature survey of Ground fall disasters in countries like India, Australia and United States of America from 1900 to 2020 (CDC 2022; DGMS 2022; MSHA 2022) has been done, where it is observed that Australia is highly successful in reducing the ground fall fatalities by adopting mechanized underground coal mining technologies and improving the safety culture through risk-based safety management.Similarly, there has been significant decline in ground fall fatalities in United States of America.
However, India has not been successful in bringing significant reduction in ground fall fatalities.Statistical data of 10 yearly average ground fall fatalities per year in India, Australia and United States of America in major accidents (killing four or more in India, three or more in Australia, Five or more in United States of America is shown in Fig. 1. The Fig. 2 shows the cause-wise fatal accidents in underground coal mines from 2001 to 2020 killing one or more persons per event (DGMS 2022).
Underground coal mining has always been recognized as one of the most hazardous occupations across the globe.In the United States, since 1910, more than 85,000 underground miners have lost their lives while mining coal.Approximately 47% of these fatalities, involving 40,000 miners, have occurred because of falls of roof and rib, which is a greater proportion than for any other type of incident classification (Mark et al. 1995(Mark et al. , 2020;;Mark and Molinda 1901-10 1911-20 1921-30 1931-40 1941-50 1951-60 1961-70 1971-80 1981-90 1991-00 2001-10 2011-20 India 4.7 3.7 11.3 5.9 12 13.7 12.7 9.9 6.1 Vol.: (0123456789) 2003; Molinda et al. 2000).Ground fall (roof and ribs) is one of the major problems in underground coal mines, which results not only loss of lives or serious injuries leading to the high cost of medical expenses and compensation but also loss of equipment, idle time of machinery, delay in reaching the normal level of production.
In recent decades, there has been a significant decline in accidents due to ground failure because of improved technologies.The progress is mainly due to three factors, first, reduction in the number of miners employed in underground workings.Second, advances in underground technology like improved bolting techniques.Blasting has largely been replaced by cutting technology with Continuous Miner and Bolter Miner.And the third factor is the change in the "safety culture", where the level of risk that miners take or are exposed to during underground mining operations are determined, and mitigating measures are taken to reduce the level of risk to an acceptable level.Safety culture had changed significantly.In the early 1900s, as the old saying went, "it was cheaper to lose a man than a donkey because the company always hired a new man, but hard to buy a new mule".The new safety culture through the introduction of a risk-based safety management system requires prioritizing high-risk activities or conditions in mining and its appropriate and effective mitigation.
Today, mining companies are taking steps to change the culture of safety by systematically adopting advanced technologies to prevent such collapses and reduce the risk to miners.The recent roof fall accidents on 23-02-2021 took the lives of two coal mine workers in the Moonidih coal mine of Coal India Limited (CIL), India.Another roof fall accident on 07-04-2021 took the lives of two coal mine workers in Kakatiya Khani-6 coal mine of Singareni Collieries Company Limited (SCCL), India.These recent roof fall accidents remind us that the age-old problem is yet to be addressed appropriately by identifying the causes of the mechanism of failure and addressing it with engineering controls.
Such ground fall accidents can only be prevented by properly understanding the influence of mininginduced stresses on the stability of workings and by adopting advanced rock reinforcement technologies, reducing the human exposure by mechanization, automation, remote operations, and continuous strata monitoring and its analysis using Artificial Intelligence or Machine Learning tools for the prediction of such failure.
Room and Pillar mining technology using Continuous Miner or Bolter Miner and Shuttle car combination will be the future technology of underground coal mining in India primarily due to the introduction of continuous mining system  From the detailed literature survey, critical parameters influencing mining-induced stresses of underground workings have been identified as shown in Table 1.

Application of machine learning
Strata behavior and stability analysis of underground strata is a very complex phenomenon due to the complex interaction of multiple parameters.Many factors contribute to the mining-induced stresses.However, all the factors do not have same effect or influence on stress.Assessment of the impact of all individual factors and their combined effect on induced stress levels is a very complex task.So far, such analysis was primarily done by statistical observations of failure cases and corresponding parametric evaluation.
Machine learning enables machines to learn from past data or experiences without being explicitly programmed.Machine learning uses a massive amount of structured and semi-structured data so that a machine learning model can generate accurate results or give predictions based on that data.
Machine learning can be mainly classified into Supervised learning, Unsupervised learning, Semisupervised learning, and Reinforcement learning.Classification of Machine learning is shown in Fig. 3.
The main strength of Machine learning lies in its basic algorithm, where it develops a generalized solution to most complex problems from past data trained to it.,Machine learning has been a vital area of research in all fields: engineering, science, education, medicine, business, accounting, finance, marketing, economics, stock market, and law, etc. Machine learning will be the driving force of the fourth industrial revolution (Mitchell 1997; Praveen Kumar et al. 2019).

Random forest
Leo Breiman is the first person who projected Random Forests during 2000s to build a predictor ensemble with a collection of random trees that develop in willy-nilly chosen subspaces of knowledge.The primary work of Amit and Geman based on Breiman's concepts (Amit and Geman 1997) on geometric feature choice, the random topological space technique of atomic number 67, and therefore the random split choice approach of (Dietterich 1998).As recommended by numerous research studies (Breiman 2001a(Breiman , 2001b;;Langroodi et al. 2021), Random Forest is a mix of tree predictors.Every tree depends on an arbitrary vector's values sampled severally and with a similar distribution for all trees within the forest.Since the variety of trees generated within random forest becomes massive, therefore generalization error of forest concurs to a limit.
A hypothetical architecture for the Random Forest is shown in Fig. 6A hypothetical architecture for the Random Forest is shown in Fig. 4.

Multilayer perceptron
Werbos was the first person to propose the Multilayer Perceptron (MLP) for purpose of solving nonlinear problems (Werbos 1974), later on modified by Rumelhart et al. (1986).It is one of the most advanced machine learning techniques adopted worldwide.
The Multilayer perceptron works on the principle of feed-forward ANN, consisting of three layers, i.e., input layer, output layer, and hidden layer.The input data or signal which needs to be processed is received by input layer.The outcome such as prediction or classification is done by output layer by necessary tasks.The number of hidden layers developed between the input and output layer is the basic computational mechanism behind the concept of Multilayer perceptron (Abirami and Chitra 2020;Gowd et al. 2018;Jain et al. 1996).A hypothetical architecture for the Multilayer Perceptron is shown in Fig. 5.Many researchers have also used the Multilayer Perceptron technique, Back Propagation theory and for prediction of tunnel boring machine penetration rate and other applications in Mining industry (Javad and Narges 2010;Kapageridis 2002;Zeng et al. 2016).

Study site
An underground coal mine Venkateshkhani-7 (VK-7), located in Kothagudem Coal belt of Godavari Valley Coal Field (GVCF) has been selected for this study, which has Four coal seams: Top seam, Index seam, King seam, and Bottom seam with a thickness varying 1.2-11 m and depth varying 62-426 m and is located at Telangana state in India as shown in Fig. 6.The proven workable coal reserves are 65MT in-situ.
The mine was started in 1954, and the Continuous Miner technology was deployed in 2006.In this mine, four Continuous Miner Panels (CMP) are studied and observed throughout the depillaring process.The geo-mining conditions of Continuous Miner Panels are shown in Table 2.The split and fender method is adopted to extract the pillars.

Continuous miner panel 5A
The size of the Continuous Miner Panel was 170 m × 231 m with a depth cover of 337.2 m.For the study of the Continuous Miner Panel stability, different geotechnical instruments were installed at different stations, as shown in Fig. 7.
The installed vibrating wire stress cells were monitored throughout the extraction.The maximum stress of 7.2 MPa was observed in Pillar-4 while   stage (A/D).As usual, the maximum value of induced stress is observed close to the goaf edge, and the development of stress is found to be rapid during the caving of the overlying roof strata.

Continuous miner panel 5B
The size of the Continuous Miner Panel was 179 m × 239 m with a depth cover of 339.6 m.For the study of the Continuous Miner Panel stability, different geotechnical instruments were installed at different selected stations, as shown in Fig. 9.The installed vibrating wire stress cells were monitored until they went into the goaf.Throughout extraction, Maximum stress of 5.8 MPa was observed in Pillar-10 while working pillar No.9.Thus, Fig. 10 shows all the vibrating wire stress cells readings w.r.t Area of Extraction/Distance of strata monitoring instrument from the extraction stage (A/D).The maximum induced stress is observed close to the goaf edge, and the development of stress is found to be rapid during the caving of the overlying roof strata.

Continuous miner panel 6A1
The

Continuous miner panel 6A2
The size of the Continuous Miner Panel was 185 m × 230 m with a depth cover of 351.9 m.For the study of the Continuous Miner Panel stability,        3.

Results and discussions
The Random Forest and Multilayer Perceptron models are developed using open-source software Waikato Environment for Knowledge Analysis (Weka) with different input parameters monitored in the fields, as shown in Table 2. 90% of the data was used for training the model and 10% for

Conclusions
The following conclusions can be drawn from this research.
• Machine learning tools like Random Forest or Multilayer Perceptron can be highly useful in predicting the strata behavior of the extraction panel.• From the prediction model, the stress condition of pillars under different operational conditions is predicted with acceptable accuracy.• Threshold limits of stress values due to different values of A/D ratio may be decided to make decisions of withdrawal of persons or machinery beyond certain threshold values.Such models will give a better understanding of the futuristic strata behavior for certain operational conditions.• Escalation of stress levels on pillars may be predicted, and a Triggered Action response Plan (TARP) for different geo-mining conditions may be developed to manage the safety risk due to strata control problems and helps mining engineers take immediate actions that avoid fatal accidents and loss of machinery, etc. • Although both models gave encouraging results, the Mean Absolute Error was high in Multilayer perceptron.• Numerical modeling tools like Finite Element Methods (FEM), Finite Difference Methods (FDM) can also be useful for understanding strata behavior based on certain input parameters, but they require high-speed computers and large memory.• The performance of the machine learning models was found to be satisfactory and encouraging.In future, more studies can be done in underground coal mines for improving safety and productivity.• This work is first of its kind to predict mininginduced stress in underground mines with the help of Machine Learning.Although much work has done earlier, it has been limited only to opencast workings.

Limitations of this research
Despite many benefits, there are some limitations in applying ML in any field.ML techniques do not produce instant, accurate predictions initially.They need to be adequately trained from the historical data available.The accuracy of the developed model depends on the amount of historical data.If the data size is large and accurate, the model will be developed with much accuracy The limitation of Machine learning is, it works only for specific domains, such as if we are creating a machine learning model to detect pictures of dogs, it will only give results for dog images.If we provide new data like cat image, then it will become unresponsive.
In this research, only four Continuous Miner panels stress readings are evaluated considering data collection constraints and limited to a single coal mine.In the future, more CM panels and varied geo-mining data conditions may be considered, and the model can be improved.As we know, the prediction of underground mining stress is still the biggest mystery to mining engineers, even after much automation, advanced instrumentation and numerical simulation techniques.
• A comprehensive review on the present status of ground fall accidents in India.• Identification of parameters influencing stability of Continuous Miner Panel.• Development of models using Random Forest and Multilayer Perceptron for prediction of mininginduced stress w.r.t various pillar extraction stages and validation of models with actual field results.• Exploring the future directions in underground coal mining towards Zero accidents mission.• To increase underground coal mining safety and productivity.• To reduce idle time of Continuous Miners due to strata control problem.• Use of Machine Learning for prediction of mining-induced stresses in place of numerical modelling for faster analysis and less resources.Keywords Bord and Pillar • Continuous Miner • Ground fall • Machine Learning • Random Forest • Multilayer Perceptron 1 Introduction

Fig. 1
Fig. 1 Comparison of decade-wise trend of ground fall fatalities per year in India, Australia and United States of America

Fig. 2
Fig. 2 Comparison of various cause-wise fatalities in Indian Coal Mines

Fig.
Fig. Classification of machine learning

Fig. 6
Fig. 6 Location map of Field Site

Fig. 7
Fig. 7 Strata monitoring plan CMP5A size of the Continuous Miner Panel was 216 m × 221.5 m with a depth cover of 372.55 m.For the study of the Continuous Miner Panel stability, different geotechnical instruments were installed at different selected stations, as shown in Fig. 11.The installed vibrating wire stress cells were monitored until they went into the goaf.Throughout extraction, Maximum stress of 23.7 MPa was observed in Pillar-4 while working pillar No.3.During extraction, all the vibrating wire stress cells were thoroughly monitored three shifts/day until they went into the goaf.Thus, Fig. 12 shows all the vibrating wire stress cells readings w.r.t Area of Extraction/Distance of strata monitoring instrument from the extraction stage (A/D).The maximum induced stress is observed close to the goaf edge, and the development of stress is found to be rapid during the caving of the overlying roof strata.